1 | //===- SyntheticSections.cpp ----------------------------------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file contains linker-synthesized sections. Currently, |
10 | // synthetic sections are created either output sections or input sections, |
11 | // but we are rewriting code so that all synthetic sections are created as |
12 | // input sections. |
13 | // |
14 | //===----------------------------------------------------------------------===// |
15 | |
16 | #include "SyntheticSections.h" |
17 | #include "Config.h" |
18 | #include "DWARF.h" |
19 | #include "EhFrame.h" |
20 | #include "InputFiles.h" |
21 | #include "LinkerScript.h" |
22 | #include "OutputSections.h" |
23 | #include "SymbolTable.h" |
24 | #include "Symbols.h" |
25 | #include "Target.h" |
26 | #include "Thunks.h" |
27 | #include "Writer.h" |
28 | #include "lld/Common/CommonLinkerContext.h" |
29 | #include "lld/Common/DWARF.h" |
30 | #include "lld/Common/Strings.h" |
31 | #include "lld/Common/Version.h" |
32 | #include "llvm/ADT/STLExtras.h" |
33 | #include "llvm/ADT/Sequence.h" |
34 | #include "llvm/ADT/SetOperations.h" |
35 | #include "llvm/ADT/StringExtras.h" |
36 | #include "llvm/BinaryFormat/Dwarf.h" |
37 | #include "llvm/BinaryFormat/ELF.h" |
38 | #include "llvm/DebugInfo/DWARF/DWARFAcceleratorTable.h" |
39 | #include "llvm/DebugInfo/DWARF/DWARFDebugPubTable.h" |
40 | #include "llvm/Support/DJB.h" |
41 | #include "llvm/Support/Endian.h" |
42 | #include "llvm/Support/LEB128.h" |
43 | #include "llvm/Support/Parallel.h" |
44 | #include "llvm/Support/TimeProfiler.h" |
45 | #include <cinttypes> |
46 | #include <cstdlib> |
47 | |
48 | using namespace llvm; |
49 | using namespace llvm::dwarf; |
50 | using namespace llvm::ELF; |
51 | using namespace llvm::object; |
52 | using namespace llvm::support; |
53 | using namespace lld; |
54 | using namespace lld::elf; |
55 | |
56 | using llvm::support::endian::read32le; |
57 | using llvm::support::endian::write32le; |
58 | using llvm::support::endian::write64le; |
59 | |
60 | constexpr size_t MergeNoTailSection::numShards; |
61 | |
62 | static uint64_t readUint(uint8_t *buf) { |
63 | return config->is64 ? read64(p: buf) : read32(p: buf); |
64 | } |
65 | |
66 | static void writeUint(uint8_t *buf, uint64_t val) { |
67 | if (config->is64) |
68 | write64(p: buf, v: val); |
69 | else |
70 | write32(p: buf, v: val); |
71 | } |
72 | |
73 | // Returns an LLD version string. |
74 | static ArrayRef<uint8_t> getVersion() { |
75 | // Check LLD_VERSION first for ease of testing. |
76 | // You can get consistent output by using the environment variable. |
77 | // This is only for testing. |
78 | StringRef s = getenv(name: "LLD_VERSION" ); |
79 | if (s.empty()) |
80 | s = saver().save(S: Twine("Linker: " ) + getLLDVersion()); |
81 | |
82 | // +1 to include the terminating '\0'. |
83 | return {(const uint8_t *)s.data(), s.size() + 1}; |
84 | } |
85 | |
86 | // Creates a .comment section containing LLD version info. |
87 | // With this feature, you can identify LLD-generated binaries easily |
88 | // by "readelf --string-dump .comment <file>". |
89 | // The returned object is a mergeable string section. |
90 | MergeInputSection *elf::() { |
91 | auto *sec = make<MergeInputSection>(args: SHF_MERGE | SHF_STRINGS, args: SHT_PROGBITS, args: 1, |
92 | args: getVersion(), args: ".comment" ); |
93 | sec->splitIntoPieces(); |
94 | return sec; |
95 | } |
96 | |
97 | // .MIPS.abiflags section. |
98 | template <class ELFT> |
99 | MipsAbiFlagsSection<ELFT>::MipsAbiFlagsSection(Elf_Mips_ABIFlags flags) |
100 | : SyntheticSection(SHF_ALLOC, SHT_MIPS_ABIFLAGS, 8, ".MIPS.abiflags" ), |
101 | flags(flags) { |
102 | this->entsize = sizeof(Elf_Mips_ABIFlags); |
103 | } |
104 | |
105 | template <class ELFT> void MipsAbiFlagsSection<ELFT>::writeTo(uint8_t *buf) { |
106 | memcpy(buf, &flags, sizeof(flags)); |
107 | } |
108 | |
109 | template <class ELFT> |
110 | std::unique_ptr<MipsAbiFlagsSection<ELFT>> MipsAbiFlagsSection<ELFT>::create() { |
111 | Elf_Mips_ABIFlags flags = {}; |
112 | bool create = false; |
113 | |
114 | for (InputSectionBase *sec : ctx.inputSections) { |
115 | if (sec->type != SHT_MIPS_ABIFLAGS) |
116 | continue; |
117 | sec->markDead(); |
118 | create = true; |
119 | |
120 | std::string filename = toString(f: sec->file); |
121 | const size_t size = sec->content().size(); |
122 | // Older version of BFD (such as the default FreeBSD linker) concatenate |
123 | // .MIPS.abiflags instead of merging. To allow for this case (or potential |
124 | // zero padding) we ignore everything after the first Elf_Mips_ABIFlags |
125 | if (size < sizeof(Elf_Mips_ABIFlags)) { |
126 | error(msg: filename + ": invalid size of .MIPS.abiflags section: got " + |
127 | Twine(size) + " instead of " + Twine(sizeof(Elf_Mips_ABIFlags))); |
128 | return nullptr; |
129 | } |
130 | auto *s = |
131 | reinterpret_cast<const Elf_Mips_ABIFlags *>(sec->content().data()); |
132 | if (s->version != 0) { |
133 | error(msg: filename + ": unexpected .MIPS.abiflags version " + |
134 | Twine(s->version)); |
135 | return nullptr; |
136 | } |
137 | |
138 | // LLD checks ISA compatibility in calcMipsEFlags(). Here we just |
139 | // select the highest number of ISA/Rev/Ext. |
140 | flags.isa_level = std::max(flags.isa_level, s->isa_level); |
141 | flags.isa_rev = std::max(flags.isa_rev, s->isa_rev); |
142 | flags.isa_ext = std::max(flags.isa_ext, s->isa_ext); |
143 | flags.gpr_size = std::max(flags.gpr_size, s->gpr_size); |
144 | flags.cpr1_size = std::max(flags.cpr1_size, s->cpr1_size); |
145 | flags.cpr2_size = std::max(flags.cpr2_size, s->cpr2_size); |
146 | flags.ases |= s->ases; |
147 | flags.flags1 |= s->flags1; |
148 | flags.flags2 |= s->flags2; |
149 | flags.fp_abi = elf::getMipsFpAbiFlag(oldFlag: flags.fp_abi, newFlag: s->fp_abi, fileName: filename); |
150 | }; |
151 | |
152 | if (create) |
153 | return std::make_unique<MipsAbiFlagsSection<ELFT>>(flags); |
154 | return nullptr; |
155 | } |
156 | |
157 | // .MIPS.options section. |
158 | template <class ELFT> |
159 | MipsOptionsSection<ELFT>::MipsOptionsSection(Elf_Mips_RegInfo reginfo) |
160 | : SyntheticSection(SHF_ALLOC, SHT_MIPS_OPTIONS, 8, ".MIPS.options" ), |
161 | reginfo(reginfo) { |
162 | this->entsize = sizeof(Elf_Mips_Options) + sizeof(Elf_Mips_RegInfo); |
163 | } |
164 | |
165 | template <class ELFT> void MipsOptionsSection<ELFT>::writeTo(uint8_t *buf) { |
166 | auto *options = reinterpret_cast<Elf_Mips_Options *>(buf); |
167 | options->kind = ODK_REGINFO; |
168 | options->size = getSize(); |
169 | |
170 | if (!config->relocatable) |
171 | reginfo.ri_gp_value = in.mipsGot->getGp(); |
172 | memcpy(buf + sizeof(Elf_Mips_Options), ®info, sizeof(reginfo)); |
173 | } |
174 | |
175 | template <class ELFT> |
176 | std::unique_ptr<MipsOptionsSection<ELFT>> MipsOptionsSection<ELFT>::create() { |
177 | // N64 ABI only. |
178 | if (!ELFT::Is64Bits) |
179 | return nullptr; |
180 | |
181 | SmallVector<InputSectionBase *, 0> sections; |
182 | for (InputSectionBase *sec : ctx.inputSections) |
183 | if (sec->type == SHT_MIPS_OPTIONS) |
184 | sections.push_back(Elt: sec); |
185 | |
186 | if (sections.empty()) |
187 | return nullptr; |
188 | |
189 | Elf_Mips_RegInfo reginfo = {}; |
190 | for (InputSectionBase *sec : sections) { |
191 | sec->markDead(); |
192 | |
193 | std::string filename = toString(f: sec->file); |
194 | ArrayRef<uint8_t> d = sec->content(); |
195 | |
196 | while (!d.empty()) { |
197 | if (d.size() < sizeof(Elf_Mips_Options)) { |
198 | error(msg: filename + ": invalid size of .MIPS.options section" ); |
199 | break; |
200 | } |
201 | |
202 | auto *opt = reinterpret_cast<const Elf_Mips_Options *>(d.data()); |
203 | if (opt->kind == ODK_REGINFO) { |
204 | reginfo.ri_gprmask |= opt->getRegInfo().ri_gprmask; |
205 | sec->getFile<ELFT>()->mipsGp0 = opt->getRegInfo().ri_gp_value; |
206 | break; |
207 | } |
208 | |
209 | if (!opt->size) |
210 | fatal(msg: filename + ": zero option descriptor size" ); |
211 | d = d.slice(opt->size); |
212 | } |
213 | }; |
214 | |
215 | return std::make_unique<MipsOptionsSection<ELFT>>(reginfo); |
216 | } |
217 | |
218 | // MIPS .reginfo section. |
219 | template <class ELFT> |
220 | MipsReginfoSection<ELFT>::MipsReginfoSection(Elf_Mips_RegInfo reginfo) |
221 | : SyntheticSection(SHF_ALLOC, SHT_MIPS_REGINFO, 4, ".reginfo" ), |
222 | reginfo(reginfo) { |
223 | this->entsize = sizeof(Elf_Mips_RegInfo); |
224 | } |
225 | |
226 | template <class ELFT> void MipsReginfoSection<ELFT>::writeTo(uint8_t *buf) { |
227 | if (!config->relocatable) |
228 | reginfo.ri_gp_value = in.mipsGot->getGp(); |
229 | memcpy(buf, ®info, sizeof(reginfo)); |
230 | } |
231 | |
232 | template <class ELFT> |
233 | std::unique_ptr<MipsReginfoSection<ELFT>> MipsReginfoSection<ELFT>::create() { |
234 | // Section should be alive for O32 and N32 ABIs only. |
235 | if (ELFT::Is64Bits) |
236 | return nullptr; |
237 | |
238 | SmallVector<InputSectionBase *, 0> sections; |
239 | for (InputSectionBase *sec : ctx.inputSections) |
240 | if (sec->type == SHT_MIPS_REGINFO) |
241 | sections.push_back(Elt: sec); |
242 | |
243 | if (sections.empty()) |
244 | return nullptr; |
245 | |
246 | Elf_Mips_RegInfo reginfo = {}; |
247 | for (InputSectionBase *sec : sections) { |
248 | sec->markDead(); |
249 | |
250 | if (sec->content().size() != sizeof(Elf_Mips_RegInfo)) { |
251 | error(msg: toString(f: sec->file) + ": invalid size of .reginfo section" ); |
252 | return nullptr; |
253 | } |
254 | |
255 | auto *r = reinterpret_cast<const Elf_Mips_RegInfo *>(sec->content().data()); |
256 | reginfo.ri_gprmask |= r->ri_gprmask; |
257 | sec->getFile<ELFT>()->mipsGp0 = r->ri_gp_value; |
258 | }; |
259 | |
260 | return std::make_unique<MipsReginfoSection<ELFT>>(reginfo); |
261 | } |
262 | |
263 | InputSection *elf::createInterpSection() { |
264 | // StringSaver guarantees that the returned string ends with '\0'. |
265 | StringRef s = saver().save(S: config->dynamicLinker); |
266 | ArrayRef<uint8_t> contents = {(const uint8_t *)s.data(), s.size() + 1}; |
267 | |
268 | return make<InputSection>(args&: ctx.internalFile, args: SHF_ALLOC, args: SHT_PROGBITS, args: 1, |
269 | args&: contents, args: ".interp" ); |
270 | } |
271 | |
272 | Defined *elf::addSyntheticLocal(StringRef name, uint8_t type, uint64_t value, |
273 | uint64_t size, InputSectionBase §ion) { |
274 | Defined *s = makeDefined(args&: section.file, args&: name, args: STB_LOCAL, args: STV_DEFAULT, args&: type, |
275 | args&: value, args&: size, args: §ion); |
276 | if (in.symTab) |
277 | in.symTab->addSymbol(sym: s); |
278 | |
279 | if (config->emachine == EM_ARM && !config->isLE && config->armBe8 && |
280 | (section.flags & SHF_EXECINSTR)) |
281 | // Adding Linker generated mapping symbols to the arm specific mapping |
282 | // symbols list. |
283 | addArmSyntheticSectionMappingSymbol(s); |
284 | |
285 | return s; |
286 | } |
287 | |
288 | static size_t getHashSize() { |
289 | switch (config->buildId) { |
290 | case BuildIdKind::Fast: |
291 | return 8; |
292 | case BuildIdKind::Md5: |
293 | case BuildIdKind::Uuid: |
294 | return 16; |
295 | case BuildIdKind::Sha1: |
296 | return 20; |
297 | case BuildIdKind::Hexstring: |
298 | return config->buildIdVector.size(); |
299 | default: |
300 | llvm_unreachable("unknown BuildIdKind" ); |
301 | } |
302 | } |
303 | |
304 | // This class represents a linker-synthesized .note.gnu.property section. |
305 | // |
306 | // In x86 and AArch64, object files may contain feature flags indicating the |
307 | // features that they have used. The flags are stored in a .note.gnu.property |
308 | // section. |
309 | // |
310 | // lld reads the sections from input files and merges them by computing AND of |
311 | // the flags. The result is written as a new .note.gnu.property section. |
312 | // |
313 | // If the flag is zero (which indicates that the intersection of the feature |
314 | // sets is empty, or some input files didn't have .note.gnu.property sections), |
315 | // we don't create this section. |
316 | GnuPropertySection::GnuPropertySection() |
317 | : SyntheticSection(llvm::ELF::SHF_ALLOC, llvm::ELF::SHT_NOTE, |
318 | config->wordsize, ".note.gnu.property" ) {} |
319 | |
320 | void GnuPropertySection::writeTo(uint8_t *buf) { |
321 | write32(p: buf, v: 4); // Name size |
322 | write32(p: buf + 4, v: getSize() - 16); // Content size |
323 | write32(p: buf + 8, v: NT_GNU_PROPERTY_TYPE_0); // Type |
324 | memcpy(dest: buf + 12, src: "GNU" , n: 4); // Name string |
325 | |
326 | uint32_t featureAndType = config->emachine == EM_AARCH64 |
327 | ? GNU_PROPERTY_AARCH64_FEATURE_1_AND |
328 | : GNU_PROPERTY_X86_FEATURE_1_AND; |
329 | |
330 | unsigned offset = 16; |
331 | if (config->andFeatures != 0) { |
332 | write32(p: buf + offset + 0, v: featureAndType); // Feature type |
333 | write32(p: buf + offset + 4, v: 4); // Feature size |
334 | write32(p: buf + offset + 8, v: config->andFeatures); // Feature flags |
335 | if (config->is64) |
336 | write32(p: buf + offset + 12, v: 0); // Padding |
337 | offset += 16; |
338 | } |
339 | |
340 | if (!ctx.aarch64PauthAbiCoreInfo.empty()) { |
341 | write32(p: buf + offset + 0, v: GNU_PROPERTY_AARCH64_FEATURE_PAUTH); |
342 | write32(p: buf + offset + 4, v: ctx.aarch64PauthAbiCoreInfo.size()); |
343 | memcpy(dest: buf + offset + 8, src: ctx.aarch64PauthAbiCoreInfo.data(), |
344 | n: ctx.aarch64PauthAbiCoreInfo.size()); |
345 | } |
346 | } |
347 | |
348 | size_t GnuPropertySection::getSize() const { |
349 | uint32_t contentSize = 0; |
350 | if (config->andFeatures != 0) |
351 | contentSize += config->is64 ? 16 : 12; |
352 | if (!ctx.aarch64PauthAbiCoreInfo.empty()) |
353 | contentSize += 4 + 4 + ctx.aarch64PauthAbiCoreInfo.size(); |
354 | assert(contentSize != 0); |
355 | return contentSize + 16; |
356 | } |
357 | |
358 | BuildIdSection::BuildIdSection() |
359 | : SyntheticSection(SHF_ALLOC, SHT_NOTE, 4, ".note.gnu.build-id" ), |
360 | hashSize(getHashSize()) {} |
361 | |
362 | void BuildIdSection::writeTo(uint8_t *buf) { |
363 | write32(p: buf, v: 4); // Name size |
364 | write32(p: buf + 4, v: hashSize); // Content size |
365 | write32(p: buf + 8, v: NT_GNU_BUILD_ID); // Type |
366 | memcpy(dest: buf + 12, src: "GNU" , n: 4); // Name string |
367 | hashBuf = buf + 16; |
368 | } |
369 | |
370 | void BuildIdSection::writeBuildId(ArrayRef<uint8_t> buf) { |
371 | assert(buf.size() == hashSize); |
372 | memcpy(dest: hashBuf, src: buf.data(), n: hashSize); |
373 | } |
374 | |
375 | BssSection::BssSection(StringRef name, uint64_t size, uint32_t alignment) |
376 | : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_NOBITS, alignment, name) { |
377 | this->bss = true; |
378 | this->size = size; |
379 | } |
380 | |
381 | EhFrameSection::EhFrameSection() |
382 | : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 1, ".eh_frame" ) {} |
383 | |
384 | // Search for an existing CIE record or create a new one. |
385 | // CIE records from input object files are uniquified by their contents |
386 | // and where their relocations point to. |
387 | template <class ELFT, class RelTy> |
388 | CieRecord *EhFrameSection::addCie(EhSectionPiece &cie, ArrayRef<RelTy> rels) { |
389 | Symbol *personality = nullptr; |
390 | unsigned firstRelI = cie.firstRelocation; |
391 | if (firstRelI != (unsigned)-1) |
392 | personality = &cie.sec->file->getRelocTargetSym(rels[firstRelI]); |
393 | |
394 | // Search for an existing CIE by CIE contents/relocation target pair. |
395 | CieRecord *&rec = cieMap[{cie.data(), personality}]; |
396 | |
397 | // If not found, create a new one. |
398 | if (!rec) { |
399 | rec = make<CieRecord>(); |
400 | rec->cie = &cie; |
401 | cieRecords.push_back(Elt: rec); |
402 | } |
403 | return rec; |
404 | } |
405 | |
406 | // There is one FDE per function. Returns a non-null pointer to the function |
407 | // symbol if the given FDE points to a live function. |
408 | template <class ELFT, class RelTy> |
409 | Defined *EhFrameSection::isFdeLive(EhSectionPiece &fde, ArrayRef<RelTy> rels) { |
410 | auto *sec = cast<EhInputSection>(Val: fde.sec); |
411 | unsigned firstRelI = fde.firstRelocation; |
412 | |
413 | // An FDE should point to some function because FDEs are to describe |
414 | // functions. That's however not always the case due to an issue of |
415 | // ld.gold with -r. ld.gold may discard only functions and leave their |
416 | // corresponding FDEs, which results in creating bad .eh_frame sections. |
417 | // To deal with that, we ignore such FDEs. |
418 | if (firstRelI == (unsigned)-1) |
419 | return nullptr; |
420 | |
421 | const RelTy &rel = rels[firstRelI]; |
422 | Symbol &b = sec->file->getRelocTargetSym(rel); |
423 | |
424 | // FDEs for garbage-collected or merged-by-ICF sections, or sections in |
425 | // another partition, are dead. |
426 | if (auto *d = dyn_cast<Defined>(Val: &b)) |
427 | if (!d->folded && d->section && d->section->partition == partition) |
428 | return d; |
429 | return nullptr; |
430 | } |
431 | |
432 | // .eh_frame is a sequence of CIE or FDE records. In general, there |
433 | // is one CIE record per input object file which is followed by |
434 | // a list of FDEs. This function searches an existing CIE or create a new |
435 | // one and associates FDEs to the CIE. |
436 | template <class ELFT, class RelTy> |
437 | void EhFrameSection::addRecords(EhInputSection *sec, ArrayRef<RelTy> rels) { |
438 | offsetToCie.clear(); |
439 | for (EhSectionPiece &cie : sec->cies) |
440 | offsetToCie[cie.inputOff] = addCie<ELFT>(cie, rels); |
441 | for (EhSectionPiece &fde : sec->fdes) { |
442 | uint32_t id = endian::read32<ELFT::Endianness>(fde.data().data() + 4); |
443 | CieRecord *rec = offsetToCie[fde.inputOff + 4 - id]; |
444 | if (!rec) |
445 | fatal(msg: toString(sec) + ": invalid CIE reference" ); |
446 | |
447 | if (!isFdeLive<ELFT>(fde, rels)) |
448 | continue; |
449 | rec->fdes.push_back(Elt: &fde); |
450 | numFdes++; |
451 | } |
452 | } |
453 | |
454 | template <class ELFT> |
455 | void EhFrameSection::addSectionAux(EhInputSection *sec) { |
456 | if (!sec->isLive()) |
457 | return; |
458 | const RelsOrRelas<ELFT> rels = |
459 | sec->template relsOrRelas<ELFT>(/*supportsCrel=*/false); |
460 | if (rels.areRelocsRel()) |
461 | addRecords<ELFT>(sec, rels.rels); |
462 | else |
463 | addRecords<ELFT>(sec, rels.relas); |
464 | } |
465 | |
466 | // Used by ICF<ELFT>::handleLSDA(). This function is very similar to |
467 | // EhFrameSection::addRecords(). |
468 | template <class ELFT, class RelTy> |
469 | void EhFrameSection::iterateFDEWithLSDAAux( |
470 | EhInputSection &sec, ArrayRef<RelTy> rels, DenseSet<size_t> &ciesWithLSDA, |
471 | llvm::function_ref<void(InputSection &)> fn) { |
472 | for (EhSectionPiece &cie : sec.cies) |
473 | if (hasLSDA(p: cie)) |
474 | ciesWithLSDA.insert(V: cie.inputOff); |
475 | for (EhSectionPiece &fde : sec.fdes) { |
476 | uint32_t id = endian::read32<ELFT::Endianness>(fde.data().data() + 4); |
477 | if (!ciesWithLSDA.contains(V: fde.inputOff + 4 - id)) |
478 | continue; |
479 | |
480 | // The CIE has a LSDA argument. Call fn with d's section. |
481 | if (Defined *d = isFdeLive<ELFT>(fde, rels)) |
482 | if (auto *s = dyn_cast_or_null<InputSection>(Val: d->section)) |
483 | fn(*s); |
484 | } |
485 | } |
486 | |
487 | template <class ELFT> |
488 | void EhFrameSection::iterateFDEWithLSDA( |
489 | llvm::function_ref<void(InputSection &)> fn) { |
490 | DenseSet<size_t> ciesWithLSDA; |
491 | for (EhInputSection *sec : sections) { |
492 | ciesWithLSDA.clear(); |
493 | const RelsOrRelas<ELFT> rels = |
494 | sec->template relsOrRelas<ELFT>(/*supportsCrel=*/false); |
495 | if (rels.areRelocsRel()) |
496 | iterateFDEWithLSDAAux<ELFT>(*sec, rels.rels, ciesWithLSDA, fn); |
497 | else |
498 | iterateFDEWithLSDAAux<ELFT>(*sec, rels.relas, ciesWithLSDA, fn); |
499 | } |
500 | } |
501 | |
502 | static void writeCieFde(uint8_t *buf, ArrayRef<uint8_t> d) { |
503 | memcpy(dest: buf, src: d.data(), n: d.size()); |
504 | // Fix the size field. -4 since size does not include the size field itself. |
505 | write32(p: buf, v: d.size() - 4); |
506 | } |
507 | |
508 | void EhFrameSection::finalizeContents() { |
509 | assert(!this->size); // Not finalized. |
510 | |
511 | switch (config->ekind) { |
512 | case ELFNoneKind: |
513 | llvm_unreachable("invalid ekind" ); |
514 | case ELF32LEKind: |
515 | for (EhInputSection *sec : sections) |
516 | addSectionAux<ELF32LE>(sec); |
517 | break; |
518 | case ELF32BEKind: |
519 | for (EhInputSection *sec : sections) |
520 | addSectionAux<ELF32BE>(sec); |
521 | break; |
522 | case ELF64LEKind: |
523 | for (EhInputSection *sec : sections) |
524 | addSectionAux<ELF64LE>(sec); |
525 | break; |
526 | case ELF64BEKind: |
527 | for (EhInputSection *sec : sections) |
528 | addSectionAux<ELF64BE>(sec); |
529 | break; |
530 | } |
531 | |
532 | size_t off = 0; |
533 | for (CieRecord *rec : cieRecords) { |
534 | rec->cie->outputOff = off; |
535 | off += rec->cie->size; |
536 | |
537 | for (EhSectionPiece *fde : rec->fdes) { |
538 | fde->outputOff = off; |
539 | off += fde->size; |
540 | } |
541 | } |
542 | |
543 | // The LSB standard does not allow a .eh_frame section with zero |
544 | // Call Frame Information records. glibc unwind-dw2-fde.c |
545 | // classify_object_over_fdes expects there is a CIE record length 0 as a |
546 | // terminator. Thus we add one unconditionally. |
547 | off += 4; |
548 | |
549 | this->size = off; |
550 | } |
551 | |
552 | // Returns data for .eh_frame_hdr. .eh_frame_hdr is a binary search table |
553 | // to get an FDE from an address to which FDE is applied. This function |
554 | // returns a list of such pairs. |
555 | SmallVector<EhFrameSection::FdeData, 0> EhFrameSection::getFdeData() const { |
556 | uint8_t *buf = Out::bufferStart + getParent()->offset + outSecOff; |
557 | SmallVector<FdeData, 0> ret; |
558 | |
559 | uint64_t va = getPartition().ehFrameHdr->getVA(); |
560 | for (CieRecord *rec : cieRecords) { |
561 | uint8_t enc = getFdeEncoding(p: rec->cie); |
562 | for (EhSectionPiece *fde : rec->fdes) { |
563 | uint64_t pc = getFdePc(buf, off: fde->outputOff, enc); |
564 | uint64_t fdeVA = getParent()->addr + fde->outputOff; |
565 | if (!isInt<32>(x: pc - va)) { |
566 | errorOrWarn(msg: toString(fde->sec) + ": PC offset is too large: 0x" + |
567 | Twine::utohexstr(Val: pc - va)); |
568 | continue; |
569 | } |
570 | ret.push_back(Elt: {.pcRel: uint32_t(pc - va), .fdeVARel: uint32_t(fdeVA - va)}); |
571 | } |
572 | } |
573 | |
574 | // Sort the FDE list by their PC and uniqueify. Usually there is only |
575 | // one FDE for a PC (i.e. function), but if ICF merges two functions |
576 | // into one, there can be more than one FDEs pointing to the address. |
577 | auto less = [](const FdeData &a, const FdeData &b) { |
578 | return a.pcRel < b.pcRel; |
579 | }; |
580 | llvm::stable_sort(Range&: ret, C: less); |
581 | auto eq = [](const FdeData &a, const FdeData &b) { |
582 | return a.pcRel == b.pcRel; |
583 | }; |
584 | ret.erase(CS: std::unique(first: ret.begin(), last: ret.end(), binary_pred: eq), CE: ret.end()); |
585 | |
586 | return ret; |
587 | } |
588 | |
589 | static uint64_t readFdeAddr(uint8_t *buf, int size) { |
590 | switch (size) { |
591 | case DW_EH_PE_udata2: |
592 | return read16(p: buf); |
593 | case DW_EH_PE_sdata2: |
594 | return (int16_t)read16(p: buf); |
595 | case DW_EH_PE_udata4: |
596 | return read32(p: buf); |
597 | case DW_EH_PE_sdata4: |
598 | return (int32_t)read32(p: buf); |
599 | case DW_EH_PE_udata8: |
600 | case DW_EH_PE_sdata8: |
601 | return read64(p: buf); |
602 | case DW_EH_PE_absptr: |
603 | return readUint(buf); |
604 | } |
605 | fatal(msg: "unknown FDE size encoding" ); |
606 | } |
607 | |
608 | // Returns the VA to which a given FDE (on a mmap'ed buffer) is applied to. |
609 | // We need it to create .eh_frame_hdr section. |
610 | uint64_t EhFrameSection::getFdePc(uint8_t *buf, size_t fdeOff, |
611 | uint8_t enc) const { |
612 | // The starting address to which this FDE applies is |
613 | // stored at FDE + 8 byte. And this offset is within |
614 | // the .eh_frame section. |
615 | size_t off = fdeOff + 8; |
616 | uint64_t addr = readFdeAddr(buf: buf + off, size: enc & 0xf); |
617 | if ((enc & 0x70) == DW_EH_PE_absptr) |
618 | return config->is64 ? addr : uint32_t(addr); |
619 | if ((enc & 0x70) == DW_EH_PE_pcrel) |
620 | return addr + getParent()->addr + off + outSecOff; |
621 | fatal(msg: "unknown FDE size relative encoding" ); |
622 | } |
623 | |
624 | void EhFrameSection::writeTo(uint8_t *buf) { |
625 | // Write CIE and FDE records. |
626 | for (CieRecord *rec : cieRecords) { |
627 | size_t cieOffset = rec->cie->outputOff; |
628 | writeCieFde(buf: buf + cieOffset, d: rec->cie->data()); |
629 | |
630 | for (EhSectionPiece *fde : rec->fdes) { |
631 | size_t off = fde->outputOff; |
632 | writeCieFde(buf: buf + off, d: fde->data()); |
633 | |
634 | // FDE's second word should have the offset to an associated CIE. |
635 | // Write it. |
636 | write32(p: buf + off + 4, v: off + 4 - cieOffset); |
637 | } |
638 | } |
639 | |
640 | // Apply relocations. .eh_frame section contents are not contiguous |
641 | // in the output buffer, but relocateAlloc() still works because |
642 | // getOffset() takes care of discontiguous section pieces. |
643 | for (EhInputSection *s : sections) |
644 | target->relocateAlloc(sec&: *s, buf); |
645 | |
646 | if (getPartition().ehFrameHdr && getPartition().ehFrameHdr->getParent()) |
647 | getPartition().ehFrameHdr->write(); |
648 | } |
649 | |
650 | GotSection::GotSection() |
651 | : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, |
652 | target->gotEntrySize, ".got" ) { |
653 | numEntries = target->gotHeaderEntriesNum; |
654 | } |
655 | |
656 | void GotSection::addConstant(const Relocation &r) { relocations.push_back(Elt: r); } |
657 | void GotSection::addEntry(const Symbol &sym) { |
658 | assert(sym.auxIdx == symAux.size() - 1); |
659 | symAux.back().gotIdx = numEntries++; |
660 | } |
661 | |
662 | bool GotSection::addTlsDescEntry(const Symbol &sym) { |
663 | assert(sym.auxIdx == symAux.size() - 1); |
664 | symAux.back().tlsDescIdx = numEntries; |
665 | numEntries += 2; |
666 | return true; |
667 | } |
668 | |
669 | bool GotSection::addDynTlsEntry(const Symbol &sym) { |
670 | assert(sym.auxIdx == symAux.size() - 1); |
671 | symAux.back().tlsGdIdx = numEntries; |
672 | // Global Dynamic TLS entries take two GOT slots. |
673 | numEntries += 2; |
674 | return true; |
675 | } |
676 | |
677 | // Reserves TLS entries for a TLS module ID and a TLS block offset. |
678 | // In total it takes two GOT slots. |
679 | bool GotSection::addTlsIndex() { |
680 | if (tlsIndexOff != uint32_t(-1)) |
681 | return false; |
682 | tlsIndexOff = numEntries * config->wordsize; |
683 | numEntries += 2; |
684 | return true; |
685 | } |
686 | |
687 | uint32_t GotSection::getTlsDescOffset(const Symbol &sym) const { |
688 | return sym.getTlsDescIdx() * config->wordsize; |
689 | } |
690 | |
691 | uint64_t GotSection::getTlsDescAddr(const Symbol &sym) const { |
692 | return getVA() + getTlsDescOffset(sym); |
693 | } |
694 | |
695 | uint64_t GotSection::getGlobalDynAddr(const Symbol &b) const { |
696 | return this->getVA() + b.getTlsGdIdx() * config->wordsize; |
697 | } |
698 | |
699 | uint64_t GotSection::getGlobalDynOffset(const Symbol &b) const { |
700 | return b.getTlsGdIdx() * config->wordsize; |
701 | } |
702 | |
703 | void GotSection::finalizeContents() { |
704 | if (config->emachine == EM_PPC64 && |
705 | numEntries <= target->gotHeaderEntriesNum && !ElfSym::globalOffsetTable) |
706 | size = 0; |
707 | else |
708 | size = numEntries * config->wordsize; |
709 | } |
710 | |
711 | bool GotSection::isNeeded() const { |
712 | // Needed if the GOT symbol is used or the number of entries is more than just |
713 | // the header. A GOT with just the header may not be needed. |
714 | return hasGotOffRel || numEntries > target->gotHeaderEntriesNum; |
715 | } |
716 | |
717 | void GotSection::writeTo(uint8_t *buf) { |
718 | // On PPC64 .got may be needed but empty. Skip the write. |
719 | if (size == 0) |
720 | return; |
721 | target->writeGotHeader(buf); |
722 | target->relocateAlloc(sec&: *this, buf); |
723 | } |
724 | |
725 | static uint64_t getMipsPageAddr(uint64_t addr) { |
726 | return (addr + 0x8000) & ~0xffff; |
727 | } |
728 | |
729 | static uint64_t getMipsPageCount(uint64_t size) { |
730 | return (size + 0xfffe) / 0xffff + 1; |
731 | } |
732 | |
733 | MipsGotSection::MipsGotSection() |
734 | : SyntheticSection(SHF_ALLOC | SHF_WRITE | SHF_MIPS_GPREL, SHT_PROGBITS, 16, |
735 | ".got" ) {} |
736 | |
737 | void MipsGotSection::addEntry(InputFile &file, Symbol &sym, int64_t addend, |
738 | RelExpr expr) { |
739 | FileGot &g = getGot(f&: file); |
740 | if (expr == R_MIPS_GOT_LOCAL_PAGE) { |
741 | if (const OutputSection *os = sym.getOutputSection()) |
742 | g.pagesMap.insert(KV: {os, {}}); |
743 | else |
744 | g.local16.insert(KV: {{nullptr, getMipsPageAddr(addr: sym.getVA(addend))}, 0}); |
745 | } else if (sym.isTls()) |
746 | g.tls.insert(KV: {&sym, 0}); |
747 | else if (sym.isPreemptible && expr == R_ABS) |
748 | g.relocs.insert(KV: {&sym, 0}); |
749 | else if (sym.isPreemptible) |
750 | g.global.insert(KV: {&sym, 0}); |
751 | else if (expr == R_MIPS_GOT_OFF32) |
752 | g.local32.insert(KV: {{&sym, addend}, 0}); |
753 | else |
754 | g.local16.insert(KV: {{&sym, addend}, 0}); |
755 | } |
756 | |
757 | void MipsGotSection::addDynTlsEntry(InputFile &file, Symbol &sym) { |
758 | getGot(f&: file).dynTlsSymbols.insert(KV: {&sym, 0}); |
759 | } |
760 | |
761 | void MipsGotSection::addTlsIndex(InputFile &file) { |
762 | getGot(f&: file).dynTlsSymbols.insert(KV: {nullptr, 0}); |
763 | } |
764 | |
765 | size_t MipsGotSection::FileGot::getEntriesNum() const { |
766 | return getPageEntriesNum() + local16.size() + global.size() + relocs.size() + |
767 | tls.size() + dynTlsSymbols.size() * 2; |
768 | } |
769 | |
770 | size_t MipsGotSection::FileGot::getPageEntriesNum() const { |
771 | size_t num = 0; |
772 | for (const std::pair<const OutputSection *, FileGot::PageBlock> &p : pagesMap) |
773 | num += p.second.count; |
774 | return num; |
775 | } |
776 | |
777 | size_t MipsGotSection::FileGot::getIndexedEntriesNum() const { |
778 | size_t count = getPageEntriesNum() + local16.size() + global.size(); |
779 | // If there are relocation-only entries in the GOT, TLS entries |
780 | // are allocated after them. TLS entries should be addressable |
781 | // by 16-bit index so count both reloc-only and TLS entries. |
782 | if (!tls.empty() || !dynTlsSymbols.empty()) |
783 | count += relocs.size() + tls.size() + dynTlsSymbols.size() * 2; |
784 | return count; |
785 | } |
786 | |
787 | MipsGotSection::FileGot &MipsGotSection::getGot(InputFile &f) { |
788 | if (f.mipsGotIndex == uint32_t(-1)) { |
789 | gots.emplace_back(); |
790 | gots.back().file = &f; |
791 | f.mipsGotIndex = gots.size() - 1; |
792 | } |
793 | return gots[f.mipsGotIndex]; |
794 | } |
795 | |
796 | uint64_t MipsGotSection::getPageEntryOffset(const InputFile *f, |
797 | const Symbol &sym, |
798 | int64_t addend) const { |
799 | const FileGot &g = gots[f->mipsGotIndex]; |
800 | uint64_t index = 0; |
801 | if (const OutputSection *outSec = sym.getOutputSection()) { |
802 | uint64_t secAddr = getMipsPageAddr(addr: outSec->addr); |
803 | uint64_t symAddr = getMipsPageAddr(addr: sym.getVA(addend)); |
804 | index = g.pagesMap.lookup(Key: outSec).firstIndex + (symAddr - secAddr) / 0xffff; |
805 | } else { |
806 | index = g.local16.lookup(Key: {nullptr, getMipsPageAddr(addr: sym.getVA(addend))}); |
807 | } |
808 | return index * config->wordsize; |
809 | } |
810 | |
811 | uint64_t MipsGotSection::getSymEntryOffset(const InputFile *f, const Symbol &s, |
812 | int64_t addend) const { |
813 | const FileGot &g = gots[f->mipsGotIndex]; |
814 | Symbol *sym = const_cast<Symbol *>(&s); |
815 | if (sym->isTls()) |
816 | return g.tls.lookup(Key: sym) * config->wordsize; |
817 | if (sym->isPreemptible) |
818 | return g.global.lookup(Key: sym) * config->wordsize; |
819 | return g.local16.lookup(Key: {sym, addend}) * config->wordsize; |
820 | } |
821 | |
822 | uint64_t MipsGotSection::getTlsIndexOffset(const InputFile *f) const { |
823 | const FileGot &g = gots[f->mipsGotIndex]; |
824 | return g.dynTlsSymbols.lookup(Key: nullptr) * config->wordsize; |
825 | } |
826 | |
827 | uint64_t MipsGotSection::getGlobalDynOffset(const InputFile *f, |
828 | const Symbol &s) const { |
829 | const FileGot &g = gots[f->mipsGotIndex]; |
830 | Symbol *sym = const_cast<Symbol *>(&s); |
831 | return g.dynTlsSymbols.lookup(Key: sym) * config->wordsize; |
832 | } |
833 | |
834 | const Symbol *MipsGotSection::getFirstGlobalEntry() const { |
835 | if (gots.empty()) |
836 | return nullptr; |
837 | const FileGot &primGot = gots.front(); |
838 | if (!primGot.global.empty()) |
839 | return primGot.global.front().first; |
840 | if (!primGot.relocs.empty()) |
841 | return primGot.relocs.front().first; |
842 | return nullptr; |
843 | } |
844 | |
845 | unsigned MipsGotSection::getLocalEntriesNum() const { |
846 | if (gots.empty()) |
847 | return headerEntriesNum; |
848 | return headerEntriesNum + gots.front().getPageEntriesNum() + |
849 | gots.front().local16.size(); |
850 | } |
851 | |
852 | bool MipsGotSection::tryMergeGots(FileGot &dst, FileGot &src, bool isPrimary) { |
853 | FileGot tmp = dst; |
854 | set_union(S1&: tmp.pagesMap, S2: src.pagesMap); |
855 | set_union(S1&: tmp.local16, S2: src.local16); |
856 | set_union(S1&: tmp.global, S2: src.global); |
857 | set_union(S1&: tmp.relocs, S2: src.relocs); |
858 | set_union(S1&: tmp.tls, S2: src.tls); |
859 | set_union(S1&: tmp.dynTlsSymbols, S2: src.dynTlsSymbols); |
860 | |
861 | size_t count = isPrimary ? headerEntriesNum : 0; |
862 | count += tmp.getIndexedEntriesNum(); |
863 | |
864 | if (count * config->wordsize > config->mipsGotSize) |
865 | return false; |
866 | |
867 | std::swap(a&: tmp, b&: dst); |
868 | return true; |
869 | } |
870 | |
871 | void MipsGotSection::finalizeContents() { updateAllocSize(); } |
872 | |
873 | bool MipsGotSection::updateAllocSize() { |
874 | size = headerEntriesNum * config->wordsize; |
875 | for (const FileGot &g : gots) |
876 | size += g.getEntriesNum() * config->wordsize; |
877 | return false; |
878 | } |
879 | |
880 | void MipsGotSection::build() { |
881 | if (gots.empty()) |
882 | return; |
883 | |
884 | std::vector<FileGot> mergedGots(1); |
885 | |
886 | // For each GOT move non-preemptible symbols from the `Global` |
887 | // to `Local16` list. Preemptible symbol might become non-preemptible |
888 | // one if, for example, it gets a related copy relocation. |
889 | for (FileGot &got : gots) { |
890 | for (auto &p: got.global) |
891 | if (!p.first->isPreemptible) |
892 | got.local16.insert(KV: {{p.first, 0}, 0}); |
893 | got.global.remove_if(Pred: [&](const std::pair<Symbol *, size_t> &p) { |
894 | return !p.first->isPreemptible; |
895 | }); |
896 | } |
897 | |
898 | // For each GOT remove "reloc-only" entry if there is "global" |
899 | // entry for the same symbol. And add local entries which indexed |
900 | // using 32-bit value at the end of 16-bit entries. |
901 | for (FileGot &got : gots) { |
902 | got.relocs.remove_if(Pred: [&](const std::pair<Symbol *, size_t> &p) { |
903 | return got.global.count(Key: p.first); |
904 | }); |
905 | set_union(S1&: got.local16, S2: got.local32); |
906 | got.local32.clear(); |
907 | } |
908 | |
909 | // Evaluate number of "reloc-only" entries in the resulting GOT. |
910 | // To do that put all unique "reloc-only" and "global" entries |
911 | // from all GOTs to the future primary GOT. |
912 | FileGot *primGot = &mergedGots.front(); |
913 | for (FileGot &got : gots) { |
914 | set_union(S1&: primGot->relocs, S2: got.global); |
915 | set_union(S1&: primGot->relocs, S2: got.relocs); |
916 | got.relocs.clear(); |
917 | } |
918 | |
919 | // Evaluate number of "page" entries in each GOT. |
920 | for (FileGot &got : gots) { |
921 | for (std::pair<const OutputSection *, FileGot::PageBlock> &p : |
922 | got.pagesMap) { |
923 | const OutputSection *os = p.first; |
924 | uint64_t secSize = 0; |
925 | for (SectionCommand *cmd : os->commands) { |
926 | if (auto *isd = dyn_cast<InputSectionDescription>(Val: cmd)) |
927 | for (InputSection *isec : isd->sections) { |
928 | uint64_t off = alignToPowerOf2(Value: secSize, Align: isec->addralign); |
929 | secSize = off + isec->getSize(); |
930 | } |
931 | } |
932 | p.second.count = getMipsPageCount(size: secSize); |
933 | } |
934 | } |
935 | |
936 | // Merge GOTs. Try to join as much as possible GOTs but do not exceed |
937 | // maximum GOT size. At first, try to fill the primary GOT because |
938 | // the primary GOT can be accessed in the most effective way. If it |
939 | // is not possible, try to fill the last GOT in the list, and finally |
940 | // create a new GOT if both attempts failed. |
941 | for (FileGot &srcGot : gots) { |
942 | InputFile *file = srcGot.file; |
943 | if (tryMergeGots(dst&: mergedGots.front(), src&: srcGot, isPrimary: true)) { |
944 | file->mipsGotIndex = 0; |
945 | } else { |
946 | // If this is the first time we failed to merge with the primary GOT, |
947 | // MergedGots.back() will also be the primary GOT. We must make sure not |
948 | // to try to merge again with isPrimary=false, as otherwise, if the |
949 | // inputs are just right, we could allow the primary GOT to become 1 or 2 |
950 | // words bigger due to ignoring the header size. |
951 | if (mergedGots.size() == 1 || |
952 | !tryMergeGots(dst&: mergedGots.back(), src&: srcGot, isPrimary: false)) { |
953 | mergedGots.emplace_back(); |
954 | std::swap(a&: mergedGots.back(), b&: srcGot); |
955 | } |
956 | file->mipsGotIndex = mergedGots.size() - 1; |
957 | } |
958 | } |
959 | std::swap(x&: gots, y&: mergedGots); |
960 | |
961 | // Reduce number of "reloc-only" entries in the primary GOT |
962 | // by subtracting "global" entries in the primary GOT. |
963 | primGot = &gots.front(); |
964 | primGot->relocs.remove_if(Pred: [&](const std::pair<Symbol *, size_t> &p) { |
965 | return primGot->global.count(Key: p.first); |
966 | }); |
967 | |
968 | // Calculate indexes for each GOT entry. |
969 | size_t index = headerEntriesNum; |
970 | for (FileGot &got : gots) { |
971 | got.startIndex = &got == primGot ? 0 : index; |
972 | for (std::pair<const OutputSection *, FileGot::PageBlock> &p : |
973 | got.pagesMap) { |
974 | // For each output section referenced by GOT page relocations calculate |
975 | // and save into pagesMap an upper bound of MIPS GOT entries required |
976 | // to store page addresses of local symbols. We assume the worst case - |
977 | // each 64kb page of the output section has at least one GOT relocation |
978 | // against it. And take in account the case when the section intersects |
979 | // page boundaries. |
980 | p.second.firstIndex = index; |
981 | index += p.second.count; |
982 | } |
983 | for (auto &p: got.local16) |
984 | p.second = index++; |
985 | for (auto &p: got.global) |
986 | p.second = index++; |
987 | for (auto &p: got.relocs) |
988 | p.second = index++; |
989 | for (auto &p: got.tls) |
990 | p.second = index++; |
991 | for (auto &p: got.dynTlsSymbols) { |
992 | p.second = index; |
993 | index += 2; |
994 | } |
995 | } |
996 | |
997 | // Update SymbolAux::gotIdx field to use this |
998 | // value later in the `sortMipsSymbols` function. |
999 | for (auto &p : primGot->global) { |
1000 | if (p.first->auxIdx == 0) |
1001 | p.first->allocateAux(); |
1002 | symAux.back().gotIdx = p.second; |
1003 | } |
1004 | for (auto &p : primGot->relocs) { |
1005 | if (p.first->auxIdx == 0) |
1006 | p.first->allocateAux(); |
1007 | symAux.back().gotIdx = p.second; |
1008 | } |
1009 | |
1010 | // Create dynamic relocations. |
1011 | for (FileGot &got : gots) { |
1012 | // Create dynamic relocations for TLS entries. |
1013 | for (std::pair<Symbol *, size_t> &p : got.tls) { |
1014 | Symbol *s = p.first; |
1015 | uint64_t offset = p.second * config->wordsize; |
1016 | // When building a shared library we still need a dynamic relocation |
1017 | // for the TP-relative offset as we don't know how much other data will |
1018 | // be allocated before us in the static TLS block. |
1019 | if (s->isPreemptible || config->shared) |
1020 | mainPart->relaDyn->addReloc(reloc: {target->tlsGotRel, this, offset, |
1021 | DynamicReloc::AgainstSymbolWithTargetVA, |
1022 | *s, 0, R_ABS}); |
1023 | } |
1024 | for (std::pair<Symbol *, size_t> &p : got.dynTlsSymbols) { |
1025 | Symbol *s = p.first; |
1026 | uint64_t offset = p.second * config->wordsize; |
1027 | if (s == nullptr) { |
1028 | if (!config->shared) |
1029 | continue; |
1030 | mainPart->relaDyn->addReloc(reloc: {target->tlsModuleIndexRel, this, offset}); |
1031 | } else { |
1032 | // When building a shared library we still need a dynamic relocation |
1033 | // for the module index. Therefore only checking for |
1034 | // S->isPreemptible is not sufficient (this happens e.g. for |
1035 | // thread-locals that have been marked as local through a linker script) |
1036 | if (!s->isPreemptible && !config->shared) |
1037 | continue; |
1038 | mainPart->relaDyn->addSymbolReloc(dynType: target->tlsModuleIndexRel, isec&: *this, |
1039 | offsetInSec: offset, sym&: *s); |
1040 | // However, we can skip writing the TLS offset reloc for non-preemptible |
1041 | // symbols since it is known even in shared libraries |
1042 | if (!s->isPreemptible) |
1043 | continue; |
1044 | offset += config->wordsize; |
1045 | mainPart->relaDyn->addSymbolReloc(dynType: target->tlsOffsetRel, isec&: *this, offsetInSec: offset, |
1046 | sym&: *s); |
1047 | } |
1048 | } |
1049 | |
1050 | // Do not create dynamic relocations for non-TLS |
1051 | // entries in the primary GOT. |
1052 | if (&got == primGot) |
1053 | continue; |
1054 | |
1055 | // Dynamic relocations for "global" entries. |
1056 | for (const std::pair<Symbol *, size_t> &p : got.global) { |
1057 | uint64_t offset = p.second * config->wordsize; |
1058 | mainPart->relaDyn->addSymbolReloc(dynType: target->relativeRel, isec&: *this, offsetInSec: offset, |
1059 | sym&: *p.first); |
1060 | } |
1061 | if (!config->isPic) |
1062 | continue; |
1063 | // Dynamic relocations for "local" entries in case of PIC. |
1064 | for (const std::pair<const OutputSection *, FileGot::PageBlock> &l : |
1065 | got.pagesMap) { |
1066 | size_t pageCount = l.second.count; |
1067 | for (size_t pi = 0; pi < pageCount; ++pi) { |
1068 | uint64_t offset = (l.second.firstIndex + pi) * config->wordsize; |
1069 | mainPart->relaDyn->addReloc(reloc: {target->relativeRel, this, offset, l.first, |
1070 | int64_t(pi * 0x10000)}); |
1071 | } |
1072 | } |
1073 | for (const std::pair<GotEntry, size_t> &p : got.local16) { |
1074 | uint64_t offset = p.second * config->wordsize; |
1075 | mainPart->relaDyn->addReloc(reloc: {target->relativeRel, this, offset, |
1076 | DynamicReloc::AddendOnlyWithTargetVA, |
1077 | *p.first.first, p.first.second, R_ABS}); |
1078 | } |
1079 | } |
1080 | } |
1081 | |
1082 | bool MipsGotSection::isNeeded() const { |
1083 | // We add the .got section to the result for dynamic MIPS target because |
1084 | // its address and properties are mentioned in the .dynamic section. |
1085 | return !config->relocatable; |
1086 | } |
1087 | |
1088 | uint64_t MipsGotSection::getGp(const InputFile *f) const { |
1089 | // For files without related GOT or files refer a primary GOT |
1090 | // returns "common" _gp value. For secondary GOTs calculate |
1091 | // individual _gp values. |
1092 | if (!f || f->mipsGotIndex == uint32_t(-1) || f->mipsGotIndex == 0) |
1093 | return ElfSym::mipsGp->getVA(addend: 0); |
1094 | return getVA() + gots[f->mipsGotIndex].startIndex * config->wordsize + 0x7ff0; |
1095 | } |
1096 | |
1097 | void MipsGotSection::writeTo(uint8_t *buf) { |
1098 | // Set the MSB of the second GOT slot. This is not required by any |
1099 | // MIPS ABI documentation, though. |
1100 | // |
1101 | // There is a comment in glibc saying that "The MSB of got[1] of a |
1102 | // gnu object is set to identify gnu objects," and in GNU gold it |
1103 | // says "the second entry will be used by some runtime loaders". |
1104 | // But how this field is being used is unclear. |
1105 | // |
1106 | // We are not really willing to mimic other linkers behaviors |
1107 | // without understanding why they do that, but because all files |
1108 | // generated by GNU tools have this special GOT value, and because |
1109 | // we've been doing this for years, it is probably a safe bet to |
1110 | // keep doing this for now. We really need to revisit this to see |
1111 | // if we had to do this. |
1112 | writeUint(buf: buf + config->wordsize, val: (uint64_t)1 << (config->wordsize * 8 - 1)); |
1113 | for (const FileGot &g : gots) { |
1114 | auto write = [&](size_t i, const Symbol *s, int64_t a) { |
1115 | uint64_t va = a; |
1116 | if (s) |
1117 | va = s->getVA(addend: a); |
1118 | writeUint(buf: buf + i * config->wordsize, val: va); |
1119 | }; |
1120 | // Write 'page address' entries to the local part of the GOT. |
1121 | for (const std::pair<const OutputSection *, FileGot::PageBlock> &l : |
1122 | g.pagesMap) { |
1123 | size_t pageCount = l.second.count; |
1124 | uint64_t firstPageAddr = getMipsPageAddr(addr: l.first->addr); |
1125 | for (size_t pi = 0; pi < pageCount; ++pi) |
1126 | write(l.second.firstIndex + pi, nullptr, firstPageAddr + pi * 0x10000); |
1127 | } |
1128 | // Local, global, TLS, reloc-only entries. |
1129 | // If TLS entry has a corresponding dynamic relocations, leave it |
1130 | // initialized by zero. Write down adjusted TLS symbol's values otherwise. |
1131 | // To calculate the adjustments use offsets for thread-local storage. |
1132 | // http://web.archive.org/web/20190324223224/https://www.linux-mips.org/wiki/NPTL |
1133 | for (const std::pair<GotEntry, size_t> &p : g.local16) |
1134 | write(p.second, p.first.first, p.first.second); |
1135 | // Write VA to the primary GOT only. For secondary GOTs that |
1136 | // will be done by REL32 dynamic relocations. |
1137 | if (&g == &gots.front()) |
1138 | for (const std::pair<Symbol *, size_t> &p : g.global) |
1139 | write(p.second, p.first, 0); |
1140 | for (const std::pair<Symbol *, size_t> &p : g.relocs) |
1141 | write(p.second, p.first, 0); |
1142 | for (const std::pair<Symbol *, size_t> &p : g.tls) |
1143 | write(p.second, p.first, |
1144 | p.first->isPreemptible || config->shared ? 0 : -0x7000); |
1145 | for (const std::pair<Symbol *, size_t> &p : g.dynTlsSymbols) { |
1146 | if (p.first == nullptr && !config->shared) |
1147 | write(p.second, nullptr, 1); |
1148 | else if (p.first && !p.first->isPreemptible) { |
1149 | // If we are emitting a shared library with relocations we mustn't write |
1150 | // anything to the GOT here. When using Elf_Rel relocations the value |
1151 | // one will be treated as an addend and will cause crashes at runtime |
1152 | if (!config->shared) |
1153 | write(p.second, nullptr, 1); |
1154 | write(p.second + 1, p.first, -0x8000); |
1155 | } |
1156 | } |
1157 | } |
1158 | } |
1159 | |
1160 | // On PowerPC the .plt section is used to hold the table of function addresses |
1161 | // instead of the .got.plt, and the type is SHT_NOBITS similar to a .bss |
1162 | // section. I don't know why we have a BSS style type for the section but it is |
1163 | // consistent across both 64-bit PowerPC ABIs as well as the 32-bit PowerPC ABI. |
1164 | GotPltSection::GotPltSection() |
1165 | : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, config->wordsize, |
1166 | ".got.plt" ) { |
1167 | if (config->emachine == EM_PPC) { |
1168 | name = ".plt" ; |
1169 | } else if (config->emachine == EM_PPC64) { |
1170 | type = SHT_NOBITS; |
1171 | name = ".plt" ; |
1172 | } |
1173 | } |
1174 | |
1175 | void GotPltSection::addEntry(Symbol &sym) { |
1176 | assert(sym.auxIdx == symAux.size() - 1 && |
1177 | symAux.back().pltIdx == entries.size()); |
1178 | entries.push_back(Elt: &sym); |
1179 | } |
1180 | |
1181 | size_t GotPltSection::getSize() const { |
1182 | return (target->gotPltHeaderEntriesNum + entries.size()) * |
1183 | target->gotEntrySize; |
1184 | } |
1185 | |
1186 | void GotPltSection::writeTo(uint8_t *buf) { |
1187 | target->writeGotPltHeader(buf); |
1188 | buf += target->gotPltHeaderEntriesNum * target->gotEntrySize; |
1189 | for (const Symbol *b : entries) { |
1190 | target->writeGotPlt(buf, s: *b); |
1191 | buf += target->gotEntrySize; |
1192 | } |
1193 | } |
1194 | |
1195 | bool GotPltSection::isNeeded() const { |
1196 | // We need to emit GOTPLT even if it's empty if there's a relocation relative |
1197 | // to it. |
1198 | return !entries.empty() || hasGotPltOffRel; |
1199 | } |
1200 | |
1201 | static StringRef getIgotPltName() { |
1202 | // On ARM the IgotPltSection is part of the GotSection. |
1203 | if (config->emachine == EM_ARM) |
1204 | return ".got" ; |
1205 | |
1206 | // On PowerPC64 the GotPltSection is renamed to '.plt' so the IgotPltSection |
1207 | // needs to be named the same. |
1208 | if (config->emachine == EM_PPC64) |
1209 | return ".plt" ; |
1210 | |
1211 | return ".got.plt" ; |
1212 | } |
1213 | |
1214 | // On PowerPC64 the GotPltSection type is SHT_NOBITS so we have to follow suit |
1215 | // with the IgotPltSection. |
1216 | IgotPltSection::IgotPltSection() |
1217 | : SyntheticSection(SHF_ALLOC | SHF_WRITE, |
1218 | config->emachine == EM_PPC64 ? SHT_NOBITS : SHT_PROGBITS, |
1219 | target->gotEntrySize, getIgotPltName()) {} |
1220 | |
1221 | void IgotPltSection::addEntry(Symbol &sym) { |
1222 | assert(symAux.back().pltIdx == entries.size()); |
1223 | entries.push_back(Elt: &sym); |
1224 | } |
1225 | |
1226 | size_t IgotPltSection::getSize() const { |
1227 | return entries.size() * target->gotEntrySize; |
1228 | } |
1229 | |
1230 | void IgotPltSection::writeTo(uint8_t *buf) { |
1231 | for (const Symbol *b : entries) { |
1232 | target->writeIgotPlt(buf, s: *b); |
1233 | buf += target->gotEntrySize; |
1234 | } |
1235 | } |
1236 | |
1237 | StringTableSection::StringTableSection(StringRef name, bool dynamic) |
1238 | : SyntheticSection(dynamic ? (uint64_t)SHF_ALLOC : 0, SHT_STRTAB, 1, name), |
1239 | dynamic(dynamic) { |
1240 | // ELF string tables start with a NUL byte. |
1241 | strings.push_back(Elt: "" ); |
1242 | stringMap.try_emplace(Key: CachedHashStringRef("" ), Args: 0); |
1243 | size = 1; |
1244 | } |
1245 | |
1246 | // Adds a string to the string table. If `hashIt` is true we hash and check for |
1247 | // duplicates. It is optional because the name of global symbols are already |
1248 | // uniqued and hashing them again has a big cost for a small value: uniquing |
1249 | // them with some other string that happens to be the same. |
1250 | unsigned StringTableSection::addString(StringRef s, bool hashIt) { |
1251 | if (hashIt) { |
1252 | auto r = stringMap.try_emplace(Key: CachedHashStringRef(s), Args&: size); |
1253 | if (!r.second) |
1254 | return r.first->second; |
1255 | } |
1256 | if (s.empty()) |
1257 | return 0; |
1258 | unsigned ret = this->size; |
1259 | this->size = this->size + s.size() + 1; |
1260 | strings.push_back(Elt: s); |
1261 | return ret; |
1262 | } |
1263 | |
1264 | void StringTableSection::writeTo(uint8_t *buf) { |
1265 | for (StringRef s : strings) { |
1266 | memcpy(dest: buf, src: s.data(), n: s.size()); |
1267 | buf[s.size()] = '\0'; |
1268 | buf += s.size() + 1; |
1269 | } |
1270 | } |
1271 | |
1272 | // Returns the number of entries in .gnu.version_d: the number of |
1273 | // non-VER_NDX_LOCAL-non-VER_NDX_GLOBAL definitions, plus 1. |
1274 | // Note that we don't support vd_cnt > 1 yet. |
1275 | static unsigned getVerDefNum() { |
1276 | return namedVersionDefs().size() + 1; |
1277 | } |
1278 | |
1279 | template <class ELFT> |
1280 | DynamicSection<ELFT>::DynamicSection() |
1281 | : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_DYNAMIC, config->wordsize, |
1282 | ".dynamic" ) { |
1283 | this->entsize = ELFT::Is64Bits ? 16 : 8; |
1284 | |
1285 | // .dynamic section is not writable on MIPS and on Fuchsia OS |
1286 | // which passes -z rodynamic. |
1287 | // See "Special Section" in Chapter 4 in the following document: |
1288 | // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf |
1289 | if (config->emachine == EM_MIPS || config->zRodynamic) |
1290 | this->flags = SHF_ALLOC; |
1291 | } |
1292 | |
1293 | // The output section .rela.dyn may include these synthetic sections: |
1294 | // |
1295 | // - part.relaDyn |
1296 | // - in.relaPlt: this is included if a linker script places .rela.plt inside |
1297 | // .rela.dyn |
1298 | // |
1299 | // DT_RELASZ is the total size of the included sections. |
1300 | static uint64_t addRelaSz(const RelocationBaseSection &relaDyn) { |
1301 | size_t size = relaDyn.getSize(); |
1302 | if (in.relaPlt->getParent() == relaDyn.getParent()) |
1303 | size += in.relaPlt->getSize(); |
1304 | return size; |
1305 | } |
1306 | |
1307 | // A Linker script may assign the RELA relocation sections to the same |
1308 | // output section. When this occurs we cannot just use the OutputSection |
1309 | // Size. Moreover the [DT_JMPREL, DT_JMPREL + DT_PLTRELSZ) is permitted to |
1310 | // overlap with the [DT_RELA, DT_RELA + DT_RELASZ). |
1311 | static uint64_t addPltRelSz() { return in.relaPlt->getSize(); } |
1312 | |
1313 | // Add remaining entries to complete .dynamic contents. |
1314 | template <class ELFT> |
1315 | std::vector<std::pair<int32_t, uint64_t>> |
1316 | DynamicSection<ELFT>::computeContents() { |
1317 | elf::Partition &part = getPartition(); |
1318 | bool isMain = part.name.empty(); |
1319 | std::vector<std::pair<int32_t, uint64_t>> entries; |
1320 | |
1321 | auto addInt = [&](int32_t tag, uint64_t val) { |
1322 | entries.emplace_back(args&: tag, args&: val); |
1323 | }; |
1324 | auto addInSec = [&](int32_t tag, const InputSection &sec) { |
1325 | entries.emplace_back(args&: tag, args: sec.getVA()); |
1326 | }; |
1327 | |
1328 | for (StringRef s : config->filterList) |
1329 | addInt(DT_FILTER, part.dynStrTab->addString(s)); |
1330 | for (StringRef s : config->auxiliaryList) |
1331 | addInt(DT_AUXILIARY, part.dynStrTab->addString(s)); |
1332 | |
1333 | if (!config->rpath.empty()) |
1334 | addInt(config->enableNewDtags ? DT_RUNPATH : DT_RPATH, |
1335 | part.dynStrTab->addString(s: config->rpath)); |
1336 | |
1337 | for (SharedFile *file : ctx.sharedFiles) |
1338 | if (file->isNeeded) |
1339 | addInt(DT_NEEDED, part.dynStrTab->addString(s: file->soName)); |
1340 | |
1341 | if (isMain) { |
1342 | if (!config->soName.empty()) |
1343 | addInt(DT_SONAME, part.dynStrTab->addString(s: config->soName)); |
1344 | } else { |
1345 | if (!config->soName.empty()) |
1346 | addInt(DT_NEEDED, part.dynStrTab->addString(s: config->soName)); |
1347 | addInt(DT_SONAME, part.dynStrTab->addString(s: part.name)); |
1348 | } |
1349 | |
1350 | // Set DT_FLAGS and DT_FLAGS_1. |
1351 | uint32_t dtFlags = 0; |
1352 | uint32_t dtFlags1 = 0; |
1353 | if (config->bsymbolic == BsymbolicKind::All) |
1354 | dtFlags |= DF_SYMBOLIC; |
1355 | if (config->zGlobal) |
1356 | dtFlags1 |= DF_1_GLOBAL; |
1357 | if (config->zInitfirst) |
1358 | dtFlags1 |= DF_1_INITFIRST; |
1359 | if (config->zInterpose) |
1360 | dtFlags1 |= DF_1_INTERPOSE; |
1361 | if (config->zNodefaultlib) |
1362 | dtFlags1 |= DF_1_NODEFLIB; |
1363 | if (config->zNodelete) |
1364 | dtFlags1 |= DF_1_NODELETE; |
1365 | if (config->zNodlopen) |
1366 | dtFlags1 |= DF_1_NOOPEN; |
1367 | if (config->pie) |
1368 | dtFlags1 |= DF_1_PIE; |
1369 | if (config->zNow) { |
1370 | dtFlags |= DF_BIND_NOW; |
1371 | dtFlags1 |= DF_1_NOW; |
1372 | } |
1373 | if (config->zOrigin) { |
1374 | dtFlags |= DF_ORIGIN; |
1375 | dtFlags1 |= DF_1_ORIGIN; |
1376 | } |
1377 | if (!config->zText) |
1378 | dtFlags |= DF_TEXTREL; |
1379 | if (ctx.hasTlsIe && config->shared) |
1380 | dtFlags |= DF_STATIC_TLS; |
1381 | |
1382 | if (dtFlags) |
1383 | addInt(DT_FLAGS, dtFlags); |
1384 | if (dtFlags1) |
1385 | addInt(DT_FLAGS_1, dtFlags1); |
1386 | |
1387 | // DT_DEBUG is a pointer to debug information used by debuggers at runtime. We |
1388 | // need it for each process, so we don't write it for DSOs. The loader writes |
1389 | // the pointer into this entry. |
1390 | // |
1391 | // DT_DEBUG is the only .dynamic entry that needs to be written to. Some |
1392 | // systems (currently only Fuchsia OS) provide other means to give the |
1393 | // debugger this information. Such systems may choose make .dynamic read-only. |
1394 | // If the target is such a system (used -z rodynamic) don't write DT_DEBUG. |
1395 | if (!config->shared && !config->relocatable && !config->zRodynamic) |
1396 | addInt(DT_DEBUG, 0); |
1397 | |
1398 | if (part.relaDyn->isNeeded()) { |
1399 | addInSec(part.relaDyn->dynamicTag, *part.relaDyn); |
1400 | entries.emplace_back(args&: part.relaDyn->sizeDynamicTag, |
1401 | args: addRelaSz(relaDyn: *part.relaDyn)); |
1402 | |
1403 | bool isRela = config->isRela; |
1404 | addInt(isRela ? DT_RELAENT : DT_RELENT, |
1405 | isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel)); |
1406 | |
1407 | // MIPS dynamic loader does not support RELCOUNT tag. |
1408 | // The problem is in the tight relation between dynamic |
1409 | // relocations and GOT. So do not emit this tag on MIPS. |
1410 | if (config->emachine != EM_MIPS) { |
1411 | size_t numRelativeRels = part.relaDyn->getRelativeRelocCount(); |
1412 | if (config->zCombreloc && numRelativeRels) |
1413 | addInt(isRela ? DT_RELACOUNT : DT_RELCOUNT, numRelativeRels); |
1414 | } |
1415 | } |
1416 | if (part.relrDyn && part.relrDyn->getParent() && |
1417 | !part.relrDyn->relocs.empty()) { |
1418 | addInSec(config->useAndroidRelrTags ? DT_ANDROID_RELR : DT_RELR, |
1419 | *part.relrDyn); |
1420 | addInt(config->useAndroidRelrTags ? DT_ANDROID_RELRSZ : DT_RELRSZ, |
1421 | part.relrDyn->getParent()->size); |
1422 | addInt(config->useAndroidRelrTags ? DT_ANDROID_RELRENT : DT_RELRENT, |
1423 | sizeof(Elf_Relr)); |
1424 | } |
1425 | if (part.relrAuthDyn && part.relrAuthDyn->getParent() && |
1426 | !part.relrAuthDyn->relocs.empty()) { |
1427 | addInSec(DT_AARCH64_AUTH_RELR, *part.relrAuthDyn); |
1428 | addInt(DT_AARCH64_AUTH_RELRSZ, part.relrAuthDyn->getParent()->size); |
1429 | addInt(DT_AARCH64_AUTH_RELRENT, sizeof(Elf_Relr)); |
1430 | } |
1431 | if (isMain && in.relaPlt->isNeeded()) { |
1432 | addInSec(DT_JMPREL, *in.relaPlt); |
1433 | entries.emplace_back(args: DT_PLTRELSZ, args: addPltRelSz()); |
1434 | switch (config->emachine) { |
1435 | case EM_MIPS: |
1436 | addInSec(DT_MIPS_PLTGOT, *in.gotPlt); |
1437 | break; |
1438 | case EM_S390: |
1439 | addInSec(DT_PLTGOT, *in.got); |
1440 | break; |
1441 | case EM_SPARCV9: |
1442 | addInSec(DT_PLTGOT, *in.plt); |
1443 | break; |
1444 | case EM_AARCH64: |
1445 | if (llvm::find_if(in.relaPlt->relocs, [](const DynamicReloc &r) { |
1446 | return r.type == target->pltRel && |
1447 | r.sym->stOther & STO_AARCH64_VARIANT_PCS; |
1448 | }) != in.relaPlt->relocs.end()) |
1449 | addInt(DT_AARCH64_VARIANT_PCS, 0); |
1450 | addInSec(DT_PLTGOT, *in.gotPlt); |
1451 | break; |
1452 | case EM_RISCV: |
1453 | if (llvm::any_of(in.relaPlt->relocs, [](const DynamicReloc &r) { |
1454 | return r.type == target->pltRel && |
1455 | (r.sym->stOther & STO_RISCV_VARIANT_CC); |
1456 | })) |
1457 | addInt(DT_RISCV_VARIANT_CC, 0); |
1458 | [[fallthrough]]; |
1459 | default: |
1460 | addInSec(DT_PLTGOT, *in.gotPlt); |
1461 | break; |
1462 | } |
1463 | addInt(DT_PLTREL, config->isRela ? DT_RELA : DT_REL); |
1464 | } |
1465 | |
1466 | if (config->emachine == EM_AARCH64) { |
1467 | if (config->andFeatures & GNU_PROPERTY_AARCH64_FEATURE_1_BTI) |
1468 | addInt(DT_AARCH64_BTI_PLT, 0); |
1469 | if (config->zPacPlt) |
1470 | addInt(DT_AARCH64_PAC_PLT, 0); |
1471 | |
1472 | if (hasMemtag()) { |
1473 | addInt(DT_AARCH64_MEMTAG_MODE, config->androidMemtagMode == NT_MEMTAG_LEVEL_ASYNC); |
1474 | addInt(DT_AARCH64_MEMTAG_HEAP, config->androidMemtagHeap); |
1475 | addInt(DT_AARCH64_MEMTAG_STACK, config->androidMemtagStack); |
1476 | if (mainPart->memtagGlobalDescriptors->isNeeded()) { |
1477 | addInSec(DT_AARCH64_MEMTAG_GLOBALS, *mainPart->memtagGlobalDescriptors); |
1478 | addInt(DT_AARCH64_MEMTAG_GLOBALSSZ, |
1479 | mainPart->memtagGlobalDescriptors->getSize()); |
1480 | } |
1481 | } |
1482 | } |
1483 | |
1484 | addInSec(DT_SYMTAB, *part.dynSymTab); |
1485 | addInt(DT_SYMENT, sizeof(Elf_Sym)); |
1486 | addInSec(DT_STRTAB, *part.dynStrTab); |
1487 | addInt(DT_STRSZ, part.dynStrTab->getSize()); |
1488 | if (!config->zText) |
1489 | addInt(DT_TEXTREL, 0); |
1490 | if (part.gnuHashTab && part.gnuHashTab->getParent()) |
1491 | addInSec(DT_GNU_HASH, *part.gnuHashTab); |
1492 | if (part.hashTab && part.hashTab->getParent()) |
1493 | addInSec(DT_HASH, *part.hashTab); |
1494 | |
1495 | if (isMain) { |
1496 | if (Out::preinitArray) { |
1497 | addInt(DT_PREINIT_ARRAY, Out::preinitArray->addr); |
1498 | addInt(DT_PREINIT_ARRAYSZ, Out::preinitArray->size); |
1499 | } |
1500 | if (Out::initArray) { |
1501 | addInt(DT_INIT_ARRAY, Out::initArray->addr); |
1502 | addInt(DT_INIT_ARRAYSZ, Out::initArray->size); |
1503 | } |
1504 | if (Out::finiArray) { |
1505 | addInt(DT_FINI_ARRAY, Out::finiArray->addr); |
1506 | addInt(DT_FINI_ARRAYSZ, Out::finiArray->size); |
1507 | } |
1508 | |
1509 | if (Symbol *b = symtab.find(name: config->init)) |
1510 | if (b->isDefined()) |
1511 | addInt(DT_INIT, b->getVA()); |
1512 | if (Symbol *b = symtab.find(name: config->fini)) |
1513 | if (b->isDefined()) |
1514 | addInt(DT_FINI, b->getVA()); |
1515 | } |
1516 | |
1517 | if (part.verSym && part.verSym->isNeeded()) |
1518 | addInSec(DT_VERSYM, *part.verSym); |
1519 | if (part.verDef && part.verDef->isLive()) { |
1520 | addInSec(DT_VERDEF, *part.verDef); |
1521 | addInt(DT_VERDEFNUM, getVerDefNum()); |
1522 | } |
1523 | if (part.verNeed && part.verNeed->isNeeded()) { |
1524 | addInSec(DT_VERNEED, *part.verNeed); |
1525 | unsigned needNum = 0; |
1526 | for (SharedFile *f : ctx.sharedFiles) |
1527 | if (!f->vernauxs.empty()) |
1528 | ++needNum; |
1529 | addInt(DT_VERNEEDNUM, needNum); |
1530 | } |
1531 | |
1532 | if (config->emachine == EM_MIPS) { |
1533 | addInt(DT_MIPS_RLD_VERSION, 1); |
1534 | addInt(DT_MIPS_FLAGS, RHF_NOTPOT); |
1535 | addInt(DT_MIPS_BASE_ADDRESS, target->getImageBase()); |
1536 | addInt(DT_MIPS_SYMTABNO, part.dynSymTab->getNumSymbols()); |
1537 | addInt(DT_MIPS_LOCAL_GOTNO, in.mipsGot->getLocalEntriesNum()); |
1538 | |
1539 | if (const Symbol *b = in.mipsGot->getFirstGlobalEntry()) |
1540 | addInt(DT_MIPS_GOTSYM, b->dynsymIndex); |
1541 | else |
1542 | addInt(DT_MIPS_GOTSYM, part.dynSymTab->getNumSymbols()); |
1543 | addInSec(DT_PLTGOT, *in.mipsGot); |
1544 | if (in.mipsRldMap) { |
1545 | if (!config->pie) |
1546 | addInSec(DT_MIPS_RLD_MAP, *in.mipsRldMap); |
1547 | // Store the offset to the .rld_map section |
1548 | // relative to the address of the tag. |
1549 | addInt(DT_MIPS_RLD_MAP_REL, |
1550 | in.mipsRldMap->getVA() - (getVA() + entries.size() * entsize)); |
1551 | } |
1552 | } |
1553 | |
1554 | // DT_PPC_GOT indicates to glibc Secure PLT is used. If DT_PPC_GOT is absent, |
1555 | // glibc assumes the old-style BSS PLT layout which we don't support. |
1556 | if (config->emachine == EM_PPC) |
1557 | addInSec(DT_PPC_GOT, *in.got); |
1558 | |
1559 | // Glink dynamic tag is required by the V2 abi if the plt section isn't empty. |
1560 | if (config->emachine == EM_PPC64 && in.plt->isNeeded()) { |
1561 | // The Glink tag points to 32 bytes before the first lazy symbol resolution |
1562 | // stub, which starts directly after the header. |
1563 | addInt(DT_PPC64_GLINK, in.plt->getVA() + target->pltHeaderSize - 32); |
1564 | } |
1565 | |
1566 | if (config->emachine == EM_PPC64) |
1567 | addInt(DT_PPC64_OPT, getPPC64TargetInfo()->ppc64DynamicSectionOpt); |
1568 | |
1569 | addInt(DT_NULL, 0); |
1570 | return entries; |
1571 | } |
1572 | |
1573 | template <class ELFT> void DynamicSection<ELFT>::finalizeContents() { |
1574 | if (OutputSection *sec = getPartition().dynStrTab->getParent()) |
1575 | getParent()->link = sec->sectionIndex; |
1576 | this->size = computeContents().size() * this->entsize; |
1577 | } |
1578 | |
1579 | template <class ELFT> void DynamicSection<ELFT>::writeTo(uint8_t *buf) { |
1580 | auto *p = reinterpret_cast<Elf_Dyn *>(buf); |
1581 | |
1582 | for (std::pair<int32_t, uint64_t> kv : computeContents()) { |
1583 | p->d_tag = kv.first; |
1584 | p->d_un.d_val = kv.second; |
1585 | ++p; |
1586 | } |
1587 | } |
1588 | |
1589 | uint64_t DynamicReloc::getOffset() const { |
1590 | return inputSec->getVA(offset: offsetInSec); |
1591 | } |
1592 | |
1593 | int64_t DynamicReloc::computeAddend() const { |
1594 | switch (kind) { |
1595 | case AddendOnly: |
1596 | assert(sym == nullptr); |
1597 | return addend; |
1598 | case AgainstSymbol: |
1599 | assert(sym != nullptr); |
1600 | return addend; |
1601 | case AddendOnlyWithTargetVA: |
1602 | case AgainstSymbolWithTargetVA: { |
1603 | uint64_t ca = InputSection::getRelocTargetVA(File: inputSec->file, Type: type, A: addend, |
1604 | P: getOffset(), Sym: *sym, Expr: expr); |
1605 | return config->is64 ? ca : SignExtend64<32>(x: ca); |
1606 | } |
1607 | case MipsMultiGotPage: |
1608 | assert(sym == nullptr); |
1609 | return getMipsPageAddr(addr: outputSec->addr) + addend; |
1610 | } |
1611 | llvm_unreachable("Unknown DynamicReloc::Kind enum" ); |
1612 | } |
1613 | |
1614 | uint32_t DynamicReloc::getSymIndex(SymbolTableBaseSection *symTab) const { |
1615 | if (!needsDynSymIndex()) |
1616 | return 0; |
1617 | |
1618 | size_t index = symTab->getSymbolIndex(sym: *sym); |
1619 | assert((index != 0 || (type != target->gotRel && type != target->pltRel) || |
1620 | !mainPart->dynSymTab->getParent()) && |
1621 | "GOT or PLT relocation must refer to symbol in dynamic symbol table" ); |
1622 | return index; |
1623 | } |
1624 | |
1625 | RelocationBaseSection::RelocationBaseSection(StringRef name, uint32_t type, |
1626 | int32_t dynamicTag, |
1627 | int32_t sizeDynamicTag, |
1628 | bool combreloc, |
1629 | unsigned concurrency) |
1630 | : SyntheticSection(SHF_ALLOC, type, config->wordsize, name), |
1631 | dynamicTag(dynamicTag), sizeDynamicTag(sizeDynamicTag), |
1632 | relocsVec(concurrency), combreloc(combreloc) {} |
1633 | |
1634 | void RelocationBaseSection::addSymbolReloc( |
1635 | RelType dynType, InputSectionBase &isec, uint64_t offsetInSec, Symbol &sym, |
1636 | int64_t addend, std::optional<RelType> addendRelType) { |
1637 | addReloc(kind: DynamicReloc::AgainstSymbol, dynType, sec&: isec, offsetInSec, sym, addend, |
1638 | expr: R_ADDEND, addendRelType: addendRelType ? *addendRelType : target->noneRel); |
1639 | } |
1640 | |
1641 | void RelocationBaseSection::addAddendOnlyRelocIfNonPreemptible( |
1642 | RelType dynType, GotSection &sec, uint64_t offsetInSec, Symbol &sym, |
1643 | RelType addendRelType) { |
1644 | // No need to write an addend to the section for preemptible symbols. |
1645 | if (sym.isPreemptible) |
1646 | addReloc(reloc: {dynType, &sec, offsetInSec, DynamicReloc::AgainstSymbol, sym, 0, |
1647 | R_ABS}); |
1648 | else |
1649 | addReloc(kind: DynamicReloc::AddendOnlyWithTargetVA, dynType, sec, offsetInSec, |
1650 | sym, addend: 0, expr: R_ABS, addendRelType); |
1651 | } |
1652 | |
1653 | void RelocationBaseSection::mergeRels() { |
1654 | size_t newSize = relocs.size(); |
1655 | for (const auto &v : relocsVec) |
1656 | newSize += v.size(); |
1657 | relocs.reserve(N: newSize); |
1658 | for (const auto &v : relocsVec) |
1659 | llvm::append_range(C&: relocs, R: v); |
1660 | relocsVec.clear(); |
1661 | } |
1662 | |
1663 | void RelocationBaseSection::partitionRels() { |
1664 | if (!combreloc) |
1665 | return; |
1666 | const RelType relativeRel = target->relativeRel; |
1667 | numRelativeRelocs = |
1668 | std::stable_partition(first: relocs.begin(), last: relocs.end(), |
1669 | pred: [=](auto &r) { return r.type == relativeRel; }) - |
1670 | relocs.begin(); |
1671 | } |
1672 | |
1673 | void RelocationBaseSection::finalizeContents() { |
1674 | SymbolTableBaseSection *symTab = getPartition().dynSymTab.get(); |
1675 | |
1676 | // When linking glibc statically, .rel{,a}.plt contains R_*_IRELATIVE |
1677 | // relocations due to IFUNC (e.g. strcpy). sh_link will be set to 0 in that |
1678 | // case. |
1679 | if (symTab && symTab->getParent()) |
1680 | getParent()->link = symTab->getParent()->sectionIndex; |
1681 | else |
1682 | getParent()->link = 0; |
1683 | |
1684 | if (in.relaPlt.get() == this && in.gotPlt->getParent()) { |
1685 | getParent()->flags |= ELF::SHF_INFO_LINK; |
1686 | getParent()->info = in.gotPlt->getParent()->sectionIndex; |
1687 | } |
1688 | } |
1689 | |
1690 | void DynamicReloc::computeRaw(SymbolTableBaseSection *symtab) { |
1691 | r_offset = getOffset(); |
1692 | r_sym = getSymIndex(symTab: symtab); |
1693 | addend = computeAddend(); |
1694 | kind = AddendOnly; // Catch errors |
1695 | } |
1696 | |
1697 | void RelocationBaseSection::computeRels() { |
1698 | SymbolTableBaseSection *symTab = getPartition().dynSymTab.get(); |
1699 | parallelForEach(R&: relocs, |
1700 | Fn: [symTab](DynamicReloc &rel) { rel.computeRaw(symtab: symTab); }); |
1701 | |
1702 | auto irelative = std::stable_partition( |
1703 | first: relocs.begin() + numRelativeRelocs, last: relocs.end(), |
1704 | pred: [t = target->iRelativeRel](auto &r) { return r.type != t; }); |
1705 | |
1706 | // Sort by (!IsRelative,SymIndex,r_offset). DT_REL[A]COUNT requires us to |
1707 | // place R_*_RELATIVE first. SymIndex is to improve locality, while r_offset |
1708 | // is to make results easier to read. |
1709 | if (combreloc) { |
1710 | auto nonRelative = relocs.begin() + numRelativeRelocs; |
1711 | parallelSort(Start: relocs.begin(), End: nonRelative, |
1712 | Comp: [&](auto &a, auto &b) { return a.r_offset < b.r_offset; }); |
1713 | // Non-relative relocations are few, so don't bother with parallelSort. |
1714 | llvm::sort(Start: nonRelative, End: irelative, Comp: [&](auto &a, auto &b) { |
1715 | return std::tie(a.r_sym, a.r_offset) < std::tie(b.r_sym, b.r_offset); |
1716 | }); |
1717 | } |
1718 | } |
1719 | |
1720 | template <class ELFT> |
1721 | RelocationSection<ELFT>::RelocationSection(StringRef name, bool combreloc, |
1722 | unsigned concurrency) |
1723 | : RelocationBaseSection(name, config->isRela ? SHT_RELA : SHT_REL, |
1724 | config->isRela ? DT_RELA : DT_REL, |
1725 | config->isRela ? DT_RELASZ : DT_RELSZ, combreloc, |
1726 | concurrency) { |
1727 | this->entsize = config->isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel); |
1728 | } |
1729 | |
1730 | template <class ELFT> void RelocationSection<ELFT>::writeTo(uint8_t *buf) { |
1731 | computeRels(); |
1732 | for (const DynamicReloc &rel : relocs) { |
1733 | auto *p = reinterpret_cast<Elf_Rela *>(buf); |
1734 | p->r_offset = rel.r_offset; |
1735 | p->setSymbolAndType(rel.r_sym, rel.type, config->isMips64EL); |
1736 | if (config->isRela) |
1737 | p->r_addend = rel.addend; |
1738 | buf += config->isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel); |
1739 | } |
1740 | } |
1741 | |
1742 | RelrBaseSection::RelrBaseSection(unsigned concurrency, bool isAArch64Auth) |
1743 | : SyntheticSection( |
1744 | SHF_ALLOC, |
1745 | isAArch64Auth |
1746 | ? SHT_AARCH64_AUTH_RELR |
1747 | : (config->useAndroidRelrTags ? SHT_ANDROID_RELR : SHT_RELR), |
1748 | config->wordsize, isAArch64Auth ? ".relr.auth.dyn" : ".relr.dyn" ), |
1749 | relocsVec(concurrency) {} |
1750 | |
1751 | void RelrBaseSection::mergeRels() { |
1752 | size_t newSize = relocs.size(); |
1753 | for (const auto &v : relocsVec) |
1754 | newSize += v.size(); |
1755 | relocs.reserve(N: newSize); |
1756 | for (const auto &v : relocsVec) |
1757 | llvm::append_range(C&: relocs, R: v); |
1758 | relocsVec.clear(); |
1759 | } |
1760 | |
1761 | template <class ELFT> |
1762 | AndroidPackedRelocationSection<ELFT>::AndroidPackedRelocationSection( |
1763 | StringRef name, unsigned concurrency) |
1764 | : RelocationBaseSection( |
1765 | name, config->isRela ? SHT_ANDROID_RELA : SHT_ANDROID_REL, |
1766 | config->isRela ? DT_ANDROID_RELA : DT_ANDROID_REL, |
1767 | config->isRela ? DT_ANDROID_RELASZ : DT_ANDROID_RELSZ, |
1768 | /*combreloc=*/false, concurrency) { |
1769 | this->entsize = 1; |
1770 | } |
1771 | |
1772 | template <class ELFT> |
1773 | bool AndroidPackedRelocationSection<ELFT>::updateAllocSize() { |
1774 | // This function computes the contents of an Android-format packed relocation |
1775 | // section. |
1776 | // |
1777 | // This format compresses relocations by using relocation groups to factor out |
1778 | // fields that are common between relocations and storing deltas from previous |
1779 | // relocations in SLEB128 format (which has a short representation for small |
1780 | // numbers). A good example of a relocation type with common fields is |
1781 | // R_*_RELATIVE, which is normally used to represent function pointers in |
1782 | // vtables. In the REL format, each relative relocation has the same r_info |
1783 | // field, and is only different from other relative relocations in terms of |
1784 | // the r_offset field. By sorting relocations by offset, grouping them by |
1785 | // r_info and representing each relocation with only the delta from the |
1786 | // previous offset, each 8-byte relocation can be compressed to as little as 1 |
1787 | // byte (or less with run-length encoding). This relocation packer was able to |
1788 | // reduce the size of the relocation section in an Android Chromium DSO from |
1789 | // 2,911,184 bytes to 174,693 bytes, or 6% of the original size. |
1790 | // |
1791 | // A relocation section consists of a header containing the literal bytes |
1792 | // 'APS2' followed by a sequence of SLEB128-encoded integers. The first two |
1793 | // elements are the total number of relocations in the section and an initial |
1794 | // r_offset value. The remaining elements define a sequence of relocation |
1795 | // groups. Each relocation group starts with a header consisting of the |
1796 | // following elements: |
1797 | // |
1798 | // - the number of relocations in the relocation group |
1799 | // - flags for the relocation group |
1800 | // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is set) the r_offset delta |
1801 | // for each relocation in the group. |
1802 | // - (if RELOCATION_GROUPED_BY_INFO_FLAG is set) the value of the r_info |
1803 | // field for each relocation in the group. |
1804 | // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG and |
1805 | // RELOCATION_GROUPED_BY_ADDEND_FLAG are set) the r_addend delta for |
1806 | // each relocation in the group. |
1807 | // |
1808 | // Following the relocation group header are descriptions of each of the |
1809 | // relocations in the group. They consist of the following elements: |
1810 | // |
1811 | // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is not set) the r_offset |
1812 | // delta for this relocation. |
1813 | // - (if RELOCATION_GROUPED_BY_INFO_FLAG is not set) the value of the r_info |
1814 | // field for this relocation. |
1815 | // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG is set and |
1816 | // RELOCATION_GROUPED_BY_ADDEND_FLAG is not set) the r_addend delta for |
1817 | // this relocation. |
1818 | |
1819 | size_t oldSize = relocData.size(); |
1820 | |
1821 | relocData = {'A', 'P', 'S', '2'}; |
1822 | raw_svector_ostream os(relocData); |
1823 | auto add = [&](int64_t v) { encodeSLEB128(Value: v, OS&: os); }; |
1824 | |
1825 | // The format header includes the number of relocations and the initial |
1826 | // offset (we set this to zero because the first relocation group will |
1827 | // perform the initial adjustment). |
1828 | add(relocs.size()); |
1829 | add(0); |
1830 | |
1831 | std::vector<Elf_Rela> relatives, nonRelatives; |
1832 | |
1833 | for (const DynamicReloc &rel : relocs) { |
1834 | Elf_Rela r; |
1835 | r.r_offset = rel.getOffset(); |
1836 | r.setSymbolAndType(rel.getSymIndex(symTab: getPartition().dynSymTab.get()), |
1837 | rel.type, false); |
1838 | r.r_addend = config->isRela ? rel.computeAddend() : 0; |
1839 | |
1840 | if (r.getType(config->isMips64EL) == target->relativeRel) |
1841 | relatives.push_back(r); |
1842 | else |
1843 | nonRelatives.push_back(r); |
1844 | } |
1845 | |
1846 | llvm::sort(relatives, [](const Elf_Rel &a, const Elf_Rel &b) { |
1847 | return a.r_offset < b.r_offset; |
1848 | }); |
1849 | |
1850 | // Try to find groups of relative relocations which are spaced one word |
1851 | // apart from one another. These generally correspond to vtable entries. The |
1852 | // format allows these groups to be encoded using a sort of run-length |
1853 | // encoding, but each group will cost 7 bytes in addition to the offset from |
1854 | // the previous group, so it is only profitable to do this for groups of |
1855 | // size 8 or larger. |
1856 | std::vector<Elf_Rela> ungroupedRelatives; |
1857 | std::vector<std::vector<Elf_Rela>> relativeGroups; |
1858 | for (auto i = relatives.begin(), e = relatives.end(); i != e;) { |
1859 | std::vector<Elf_Rela> group; |
1860 | do { |
1861 | group.push_back(*i++); |
1862 | } while (i != e && (i - 1)->r_offset + config->wordsize == i->r_offset); |
1863 | |
1864 | if (group.size() < 8) |
1865 | ungroupedRelatives.insert(ungroupedRelatives.end(), group.begin(), |
1866 | group.end()); |
1867 | else |
1868 | relativeGroups.emplace_back(std::move(group)); |
1869 | } |
1870 | |
1871 | // For non-relative relocations, we would like to: |
1872 | // 1. Have relocations with the same symbol offset to be consecutive, so |
1873 | // that the runtime linker can speed-up symbol lookup by implementing an |
1874 | // 1-entry cache. |
1875 | // 2. Group relocations by r_info to reduce the size of the relocation |
1876 | // section. |
1877 | // Since the symbol offset is the high bits in r_info, sorting by r_info |
1878 | // allows us to do both. |
1879 | // |
1880 | // For Rela, we also want to sort by r_addend when r_info is the same. This |
1881 | // enables us to group by r_addend as well. |
1882 | llvm::sort(nonRelatives, [](const Elf_Rela &a, const Elf_Rela &b) { |
1883 | if (a.r_info != b.r_info) |
1884 | return a.r_info < b.r_info; |
1885 | if (a.r_addend != b.r_addend) |
1886 | return a.r_addend < b.r_addend; |
1887 | return a.r_offset < b.r_offset; |
1888 | }); |
1889 | |
1890 | // Group relocations with the same r_info. Note that each group emits a group |
1891 | // header and that may make the relocation section larger. It is hard to |
1892 | // estimate the size of a group header as the encoded size of that varies |
1893 | // based on r_info. However, we can approximate this trade-off by the number |
1894 | // of values encoded. Each group header contains 3 values, and each relocation |
1895 | // in a group encodes one less value, as compared to when it is not grouped. |
1896 | // Therefore, we only group relocations if there are 3 or more of them with |
1897 | // the same r_info. |
1898 | // |
1899 | // For Rela, the addend for most non-relative relocations is zero, and thus we |
1900 | // can usually get a smaller relocation section if we group relocations with 0 |
1901 | // addend as well. |
1902 | std::vector<Elf_Rela> ungroupedNonRelatives; |
1903 | std::vector<std::vector<Elf_Rela>> nonRelativeGroups; |
1904 | for (auto i = nonRelatives.begin(), e = nonRelatives.end(); i != e;) { |
1905 | auto j = i + 1; |
1906 | while (j != e && i->r_info == j->r_info && |
1907 | (!config->isRela || i->r_addend == j->r_addend)) |
1908 | ++j; |
1909 | if (j - i < 3 || (config->isRela && i->r_addend != 0)) |
1910 | ungroupedNonRelatives.insert(ungroupedNonRelatives.end(), i, j); |
1911 | else |
1912 | nonRelativeGroups.emplace_back(i, j); |
1913 | i = j; |
1914 | } |
1915 | |
1916 | // Sort ungrouped relocations by offset to minimize the encoded length. |
1917 | llvm::sort(ungroupedNonRelatives, [](const Elf_Rela &a, const Elf_Rela &b) { |
1918 | return a.r_offset < b.r_offset; |
1919 | }); |
1920 | |
1921 | unsigned hasAddendIfRela = |
1922 | config->isRela ? RELOCATION_GROUP_HAS_ADDEND_FLAG : 0; |
1923 | |
1924 | uint64_t offset = 0; |
1925 | uint64_t addend = 0; |
1926 | |
1927 | // Emit the run-length encoding for the groups of adjacent relative |
1928 | // relocations. Each group is represented using two groups in the packed |
1929 | // format. The first is used to set the current offset to the start of the |
1930 | // group (and also encodes the first relocation), and the second encodes the |
1931 | // remaining relocations. |
1932 | for (std::vector<Elf_Rela> &g : relativeGroups) { |
1933 | // The first relocation in the group. |
1934 | add(1); |
1935 | add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG | |
1936 | RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela); |
1937 | add(g[0].r_offset - offset); |
1938 | add(target->relativeRel); |
1939 | if (config->isRela) { |
1940 | add(g[0].r_addend - addend); |
1941 | addend = g[0].r_addend; |
1942 | } |
1943 | |
1944 | // The remaining relocations. |
1945 | add(g.size() - 1); |
1946 | add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG | |
1947 | RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela); |
1948 | add(config->wordsize); |
1949 | add(target->relativeRel); |
1950 | if (config->isRela) { |
1951 | for (const auto &i : llvm::drop_begin(g)) { |
1952 | add(i.r_addend - addend); |
1953 | addend = i.r_addend; |
1954 | } |
1955 | } |
1956 | |
1957 | offset = g.back().r_offset; |
1958 | } |
1959 | |
1960 | // Now the ungrouped relatives. |
1961 | if (!ungroupedRelatives.empty()) { |
1962 | add(ungroupedRelatives.size()); |
1963 | add(RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela); |
1964 | add(target->relativeRel); |
1965 | for (Elf_Rela &r : ungroupedRelatives) { |
1966 | add(r.r_offset - offset); |
1967 | offset = r.r_offset; |
1968 | if (config->isRela) { |
1969 | add(r.r_addend - addend); |
1970 | addend = r.r_addend; |
1971 | } |
1972 | } |
1973 | } |
1974 | |
1975 | // Grouped non-relatives. |
1976 | for (ArrayRef<Elf_Rela> g : nonRelativeGroups) { |
1977 | add(g.size()); |
1978 | add(RELOCATION_GROUPED_BY_INFO_FLAG); |
1979 | add(g[0].r_info); |
1980 | for (const Elf_Rela &r : g) { |
1981 | add(r.r_offset - offset); |
1982 | offset = r.r_offset; |
1983 | } |
1984 | addend = 0; |
1985 | } |
1986 | |
1987 | // Finally the ungrouped non-relative relocations. |
1988 | if (!ungroupedNonRelatives.empty()) { |
1989 | add(ungroupedNonRelatives.size()); |
1990 | add(hasAddendIfRela); |
1991 | for (Elf_Rela &r : ungroupedNonRelatives) { |
1992 | add(r.r_offset - offset); |
1993 | offset = r.r_offset; |
1994 | add(r.r_info); |
1995 | if (config->isRela) { |
1996 | add(r.r_addend - addend); |
1997 | addend = r.r_addend; |
1998 | } |
1999 | } |
2000 | } |
2001 | |
2002 | // Don't allow the section to shrink; otherwise the size of the section can |
2003 | // oscillate infinitely. |
2004 | if (relocData.size() < oldSize) |
2005 | relocData.append(NumInputs: oldSize - relocData.size(), Elt: 0); |
2006 | |
2007 | // Returns whether the section size changed. We need to keep recomputing both |
2008 | // section layout and the contents of this section until the size converges |
2009 | // because changing this section's size can affect section layout, which in |
2010 | // turn can affect the sizes of the LEB-encoded integers stored in this |
2011 | // section. |
2012 | return relocData.size() != oldSize; |
2013 | } |
2014 | |
2015 | template <class ELFT> |
2016 | RelrSection<ELFT>::RelrSection(unsigned concurrency, bool isAArch64Auth) |
2017 | : RelrBaseSection(concurrency, isAArch64Auth) { |
2018 | this->entsize = config->wordsize; |
2019 | } |
2020 | |
2021 | template <class ELFT> bool RelrSection<ELFT>::updateAllocSize() { |
2022 | // This function computes the contents of an SHT_RELR packed relocation |
2023 | // section. |
2024 | // |
2025 | // Proposal for adding SHT_RELR sections to generic-abi is here: |
2026 | // https://groups.google.com/forum/#!topic/generic-abi/bX460iggiKg |
2027 | // |
2028 | // The encoded sequence of Elf64_Relr entries in a SHT_RELR section looks |
2029 | // like [ AAAAAAAA BBBBBBB1 BBBBBBB1 ... AAAAAAAA BBBBBB1 ... ] |
2030 | // |
2031 | // i.e. start with an address, followed by any number of bitmaps. The address |
2032 | // entry encodes 1 relocation. The subsequent bitmap entries encode up to 63 |
2033 | // relocations each, at subsequent offsets following the last address entry. |
2034 | // |
2035 | // The bitmap entries must have 1 in the least significant bit. The assumption |
2036 | // here is that an address cannot have 1 in lsb. Odd addresses are not |
2037 | // supported. |
2038 | // |
2039 | // Excluding the least significant bit in the bitmap, each non-zero bit in |
2040 | // the bitmap represents a relocation to be applied to a corresponding machine |
2041 | // word that follows the base address word. The second least significant bit |
2042 | // represents the machine word immediately following the initial address, and |
2043 | // each bit that follows represents the next word, in linear order. As such, |
2044 | // a single bitmap can encode up to 31 relocations in a 32-bit object, and |
2045 | // 63 relocations in a 64-bit object. |
2046 | // |
2047 | // This encoding has a couple of interesting properties: |
2048 | // 1. Looking at any entry, it is clear whether it's an address or a bitmap: |
2049 | // even means address, odd means bitmap. |
2050 | // 2. Just a simple list of addresses is a valid encoding. |
2051 | |
2052 | size_t oldSize = relrRelocs.size(); |
2053 | relrRelocs.clear(); |
2054 | |
2055 | // Same as Config->Wordsize but faster because this is a compile-time |
2056 | // constant. |
2057 | const size_t wordsize = sizeof(typename ELFT::uint); |
2058 | |
2059 | // Number of bits to use for the relocation offsets bitmap. |
2060 | // Must be either 63 or 31. |
2061 | const size_t nBits = wordsize * 8 - 1; |
2062 | |
2063 | // Get offsets for all relative relocations and sort them. |
2064 | std::unique_ptr<uint64_t[]> offsets(new uint64_t[relocs.size()]); |
2065 | for (auto [i, r] : llvm::enumerate(relocs)) |
2066 | offsets[i] = r.getOffset(); |
2067 | llvm::sort(offsets.get(), offsets.get() + relocs.size()); |
2068 | |
2069 | // For each leading relocation, find following ones that can be folded |
2070 | // as a bitmap and fold them. |
2071 | for (size_t i = 0, e = relocs.size(); i != e;) { |
2072 | // Add a leading relocation. |
2073 | relrRelocs.push_back(Elf_Relr(offsets[i])); |
2074 | uint64_t base = offsets[i] + wordsize; |
2075 | ++i; |
2076 | |
2077 | // Find foldable relocations to construct bitmaps. |
2078 | for (;;) { |
2079 | uint64_t bitmap = 0; |
2080 | for (; i != e; ++i) { |
2081 | uint64_t d = offsets[i] - base; |
2082 | if (d >= nBits * wordsize || d % wordsize) |
2083 | break; |
2084 | bitmap |= uint64_t(1) << (d / wordsize); |
2085 | } |
2086 | if (!bitmap) |
2087 | break; |
2088 | relrRelocs.push_back(Elf_Relr((bitmap << 1) | 1)); |
2089 | base += nBits * wordsize; |
2090 | } |
2091 | } |
2092 | |
2093 | // Don't allow the section to shrink; otherwise the size of the section can |
2094 | // oscillate infinitely. Trailing 1s do not decode to more relocations. |
2095 | if (relrRelocs.size() < oldSize) { |
2096 | log(msg: ".relr.dyn needs " + Twine(oldSize - relrRelocs.size()) + |
2097 | " padding word(s)" ); |
2098 | relrRelocs.resize(oldSize, Elf_Relr(1)); |
2099 | } |
2100 | |
2101 | return relrRelocs.size() != oldSize; |
2102 | } |
2103 | |
2104 | SymbolTableBaseSection::SymbolTableBaseSection(StringTableSection &strTabSec) |
2105 | : SyntheticSection(strTabSec.isDynamic() ? (uint64_t)SHF_ALLOC : 0, |
2106 | strTabSec.isDynamic() ? SHT_DYNSYM : SHT_SYMTAB, |
2107 | config->wordsize, |
2108 | strTabSec.isDynamic() ? ".dynsym" : ".symtab" ), |
2109 | strTabSec(strTabSec) {} |
2110 | |
2111 | // Orders symbols according to their positions in the GOT, |
2112 | // in compliance with MIPS ABI rules. |
2113 | // See "Global Offset Table" in Chapter 5 in the following document |
2114 | // for detailed description: |
2115 | // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf |
2116 | static bool sortMipsSymbols(const SymbolTableEntry &l, |
2117 | const SymbolTableEntry &r) { |
2118 | // Sort entries related to non-local preemptible symbols by GOT indexes. |
2119 | // All other entries go to the beginning of a dynsym in arbitrary order. |
2120 | if (l.sym->isInGot() && r.sym->isInGot()) |
2121 | return l.sym->getGotIdx() < r.sym->getGotIdx(); |
2122 | if (!l.sym->isInGot() && !r.sym->isInGot()) |
2123 | return false; |
2124 | return !l.sym->isInGot(); |
2125 | } |
2126 | |
2127 | void SymbolTableBaseSection::finalizeContents() { |
2128 | if (OutputSection *sec = strTabSec.getParent()) |
2129 | getParent()->link = sec->sectionIndex; |
2130 | |
2131 | if (this->type != SHT_DYNSYM) { |
2132 | sortSymTabSymbols(); |
2133 | return; |
2134 | } |
2135 | |
2136 | // If it is a .dynsym, there should be no local symbols, but we need |
2137 | // to do a few things for the dynamic linker. |
2138 | |
2139 | // Section's Info field has the index of the first non-local symbol. |
2140 | // Because the first symbol entry is a null entry, 1 is the first. |
2141 | getParent()->info = 1; |
2142 | |
2143 | if (getPartition().gnuHashTab) { |
2144 | // NB: It also sorts Symbols to meet the GNU hash table requirements. |
2145 | getPartition().gnuHashTab->addSymbols(symbols); |
2146 | } else if (config->emachine == EM_MIPS) { |
2147 | llvm::stable_sort(Range&: symbols, C: sortMipsSymbols); |
2148 | } |
2149 | |
2150 | // Only the main partition's dynsym indexes are stored in the symbols |
2151 | // themselves. All other partitions use a lookup table. |
2152 | if (this == mainPart->dynSymTab.get()) { |
2153 | size_t i = 0; |
2154 | for (const SymbolTableEntry &s : symbols) |
2155 | s.sym->dynsymIndex = ++i; |
2156 | } |
2157 | } |
2158 | |
2159 | // The ELF spec requires that all local symbols precede global symbols, so we |
2160 | // sort symbol entries in this function. (For .dynsym, we don't do that because |
2161 | // symbols for dynamic linking are inherently all globals.) |
2162 | // |
2163 | // Aside from above, we put local symbols in groups starting with the STT_FILE |
2164 | // symbol. That is convenient for purpose of identifying where are local symbols |
2165 | // coming from. |
2166 | void SymbolTableBaseSection::sortSymTabSymbols() { |
2167 | // Move all local symbols before global symbols. |
2168 | auto e = std::stable_partition( |
2169 | first: symbols.begin(), last: symbols.end(), |
2170 | pred: [](const SymbolTableEntry &s) { return s.sym->isLocal(); }); |
2171 | size_t numLocals = e - symbols.begin(); |
2172 | getParent()->info = numLocals + 1; |
2173 | |
2174 | // We want to group the local symbols by file. For that we rebuild the local |
2175 | // part of the symbols vector. We do not need to care about the STT_FILE |
2176 | // symbols, they are already naturally placed first in each group. That |
2177 | // happens because STT_FILE is always the first symbol in the object and hence |
2178 | // precede all other local symbols we add for a file. |
2179 | MapVector<InputFile *, SmallVector<SymbolTableEntry, 0>> arr; |
2180 | for (const SymbolTableEntry &s : llvm::make_range(x: symbols.begin(), y: e)) |
2181 | arr[s.sym->file].push_back(Elt: s); |
2182 | |
2183 | auto i = symbols.begin(); |
2184 | for (auto &p : arr) |
2185 | for (SymbolTableEntry &entry : p.second) |
2186 | *i++ = entry; |
2187 | } |
2188 | |
2189 | void SymbolTableBaseSection::addSymbol(Symbol *b) { |
2190 | // Adding a local symbol to a .dynsym is a bug. |
2191 | assert(this->type != SHT_DYNSYM || !b->isLocal()); |
2192 | symbols.push_back(Elt: {.sym: b, .strTabOffset: strTabSec.addString(s: b->getName(), hashIt: false)}); |
2193 | } |
2194 | |
2195 | size_t SymbolTableBaseSection::getSymbolIndex(const Symbol &sym) { |
2196 | if (this == mainPart->dynSymTab.get()) |
2197 | return sym.dynsymIndex; |
2198 | |
2199 | // Initializes symbol lookup tables lazily. This is used only for -r, |
2200 | // --emit-relocs and dynsyms in partitions other than the main one. |
2201 | llvm::call_once(flag&: onceFlag, F: [&] { |
2202 | symbolIndexMap.reserve(NumEntries: symbols.size()); |
2203 | size_t i = 0; |
2204 | for (const SymbolTableEntry &e : symbols) { |
2205 | if (e.sym->type == STT_SECTION) |
2206 | sectionIndexMap[e.sym->getOutputSection()] = ++i; |
2207 | else |
2208 | symbolIndexMap[e.sym] = ++i; |
2209 | } |
2210 | }); |
2211 | |
2212 | // Section symbols are mapped based on their output sections |
2213 | // to maintain their semantics. |
2214 | if (sym.type == STT_SECTION) |
2215 | return sectionIndexMap.lookup(Val: sym.getOutputSection()); |
2216 | return symbolIndexMap.lookup(Val: &sym); |
2217 | } |
2218 | |
2219 | template <class ELFT> |
2220 | SymbolTableSection<ELFT>::SymbolTableSection(StringTableSection &strTabSec) |
2221 | : SymbolTableBaseSection(strTabSec) { |
2222 | this->entsize = sizeof(Elf_Sym); |
2223 | } |
2224 | |
2225 | static BssSection *getCommonSec(Symbol *sym) { |
2226 | if (config->relocatable) |
2227 | if (auto *d = dyn_cast<Defined>(Val: sym)) |
2228 | return dyn_cast_or_null<BssSection>(Val: d->section); |
2229 | return nullptr; |
2230 | } |
2231 | |
2232 | static uint32_t getSymSectionIndex(Symbol *sym) { |
2233 | assert(!(sym->hasFlag(NEEDS_COPY) && sym->isObject())); |
2234 | if (!isa<Defined>(Val: sym) || sym->hasFlag(bit: NEEDS_COPY)) |
2235 | return SHN_UNDEF; |
2236 | if (const OutputSection *os = sym->getOutputSection()) |
2237 | return os->sectionIndex >= SHN_LORESERVE ? (uint32_t)SHN_XINDEX |
2238 | : os->sectionIndex; |
2239 | return SHN_ABS; |
2240 | } |
2241 | |
2242 | // Write the internal symbol table contents to the output symbol table. |
2243 | template <class ELFT> void SymbolTableSection<ELFT>::writeTo(uint8_t *buf) { |
2244 | // The first entry is a null entry as per the ELF spec. |
2245 | buf += sizeof(Elf_Sym); |
2246 | |
2247 | auto *eSym = reinterpret_cast<Elf_Sym *>(buf); |
2248 | |
2249 | for (SymbolTableEntry &ent : symbols) { |
2250 | Symbol *sym = ent.sym; |
2251 | bool isDefinedHere = type == SHT_SYMTAB || sym->partition == partition; |
2252 | |
2253 | // Set st_name, st_info and st_other. |
2254 | eSym->st_name = ent.strTabOffset; |
2255 | eSym->setBindingAndType(sym->binding, sym->type); |
2256 | eSym->st_other = sym->stOther; |
2257 | |
2258 | if (BssSection *commonSec = getCommonSec(sym)) { |
2259 | // When -r is specified, a COMMON symbol is not allocated. Its st_shndx |
2260 | // holds SHN_COMMON and st_value holds the alignment. |
2261 | eSym->st_shndx = SHN_COMMON; |
2262 | eSym->st_value = commonSec->addralign; |
2263 | eSym->st_size = cast<Defined>(Val: sym)->size; |
2264 | } else { |
2265 | const uint32_t shndx = getSymSectionIndex(sym); |
2266 | if (isDefinedHere) { |
2267 | eSym->st_shndx = shndx; |
2268 | eSym->st_value = sym->getVA(); |
2269 | // Copy symbol size if it is a defined symbol. st_size is not |
2270 | // significant for undefined symbols, so whether copying it or not is up |
2271 | // to us if that's the case. We'll leave it as zero because by not |
2272 | // setting a value, we can get the exact same outputs for two sets of |
2273 | // input files that differ only in undefined symbol size in DSOs. |
2274 | eSym->st_size = shndx != SHN_UNDEF ? cast<Defined>(Val: sym)->size : 0; |
2275 | } else { |
2276 | eSym->st_shndx = 0; |
2277 | eSym->st_value = 0; |
2278 | eSym->st_size = 0; |
2279 | } |
2280 | } |
2281 | |
2282 | ++eSym; |
2283 | } |
2284 | |
2285 | // On MIPS we need to mark symbol which has a PLT entry and requires |
2286 | // pointer equality by STO_MIPS_PLT flag. That is necessary to help |
2287 | // dynamic linker distinguish such symbols and MIPS lazy-binding stubs. |
2288 | // https://sourceware.org/ml/binutils/2008-07/txt00000.txt |
2289 | if (config->emachine == EM_MIPS) { |
2290 | auto *eSym = reinterpret_cast<Elf_Sym *>(buf); |
2291 | |
2292 | for (SymbolTableEntry &ent : symbols) { |
2293 | Symbol *sym = ent.sym; |
2294 | if (sym->isInPlt() && sym->hasFlag(bit: NEEDS_COPY)) |
2295 | eSym->st_other |= STO_MIPS_PLT; |
2296 | if (isMicroMips()) { |
2297 | // We already set the less-significant bit for symbols |
2298 | // marked by the `STO_MIPS_MICROMIPS` flag and for microMIPS PLT |
2299 | // records. That allows us to distinguish such symbols in |
2300 | // the `MIPS<ELFT>::relocate()` routine. Now we should |
2301 | // clear that bit for non-dynamic symbol table, so tools |
2302 | // like `objdump` will be able to deal with a correct |
2303 | // symbol position. |
2304 | if (sym->isDefined() && |
2305 | ((sym->stOther & STO_MIPS_MICROMIPS) || sym->hasFlag(bit: NEEDS_COPY))) { |
2306 | if (!strTabSec.isDynamic()) |
2307 | eSym->st_value &= ~1; |
2308 | eSym->st_other |= STO_MIPS_MICROMIPS; |
2309 | } |
2310 | } |
2311 | if (config->relocatable) |
2312 | if (auto *d = dyn_cast<Defined>(Val: sym)) |
2313 | if (isMipsPIC<ELFT>(d)) |
2314 | eSym->st_other |= STO_MIPS_PIC; |
2315 | ++eSym; |
2316 | } |
2317 | } |
2318 | } |
2319 | |
2320 | SymtabShndxSection::SymtabShndxSection() |
2321 | : SyntheticSection(0, SHT_SYMTAB_SHNDX, 4, ".symtab_shndx" ) { |
2322 | this->entsize = 4; |
2323 | } |
2324 | |
2325 | void SymtabShndxSection::writeTo(uint8_t *buf) { |
2326 | // We write an array of 32 bit values, where each value has 1:1 association |
2327 | // with an entry in .symtab. If the corresponding entry contains SHN_XINDEX, |
2328 | // we need to write actual index, otherwise, we must write SHN_UNDEF(0). |
2329 | buf += 4; // Ignore .symtab[0] entry. |
2330 | for (const SymbolTableEntry &entry : in.symTab->getSymbols()) { |
2331 | if (!getCommonSec(sym: entry.sym) && getSymSectionIndex(sym: entry.sym) == SHN_XINDEX) |
2332 | write32(p: buf, v: entry.sym->getOutputSection()->sectionIndex); |
2333 | buf += 4; |
2334 | } |
2335 | } |
2336 | |
2337 | bool SymtabShndxSection::isNeeded() const { |
2338 | // SHT_SYMTAB can hold symbols with section indices values up to |
2339 | // SHN_LORESERVE. If we need more, we want to use extension SHT_SYMTAB_SHNDX |
2340 | // section. Problem is that we reveal the final section indices a bit too |
2341 | // late, and we do not know them here. For simplicity, we just always create |
2342 | // a .symtab_shndx section when the amount of output sections is huge. |
2343 | size_t size = 0; |
2344 | for (SectionCommand *cmd : script->sectionCommands) |
2345 | if (isa<OutputDesc>(Val: cmd)) |
2346 | ++size; |
2347 | return size >= SHN_LORESERVE; |
2348 | } |
2349 | |
2350 | void SymtabShndxSection::finalizeContents() { |
2351 | getParent()->link = in.symTab->getParent()->sectionIndex; |
2352 | } |
2353 | |
2354 | size_t SymtabShndxSection::getSize() const { |
2355 | return in.symTab->getNumSymbols() * 4; |
2356 | } |
2357 | |
2358 | // .hash and .gnu.hash sections contain on-disk hash tables that map |
2359 | // symbol names to their dynamic symbol table indices. Their purpose |
2360 | // is to help the dynamic linker resolve symbols quickly. If ELF files |
2361 | // don't have them, the dynamic linker has to do linear search on all |
2362 | // dynamic symbols, which makes programs slower. Therefore, a .hash |
2363 | // section is added to a DSO by default. |
2364 | // |
2365 | // The Unix semantics of resolving dynamic symbols is somewhat expensive. |
2366 | // Each ELF file has a list of DSOs that the ELF file depends on and a |
2367 | // list of dynamic symbols that need to be resolved from any of the |
2368 | // DSOs. That means resolving all dynamic symbols takes O(m)*O(n) |
2369 | // where m is the number of DSOs and n is the number of dynamic |
2370 | // symbols. For modern large programs, both m and n are large. So |
2371 | // making each step faster by using hash tables substantially |
2372 | // improves time to load programs. |
2373 | // |
2374 | // (Note that this is not the only way to design the shared library. |
2375 | // For instance, the Windows DLL takes a different approach. On |
2376 | // Windows, each dynamic symbol has a name of DLL from which the symbol |
2377 | // has to be resolved. That makes the cost of symbol resolution O(n). |
2378 | // This disables some hacky techniques you can use on Unix such as |
2379 | // LD_PRELOAD, but this is arguably better semantics than the Unix ones.) |
2380 | // |
2381 | // Due to historical reasons, we have two different hash tables, .hash |
2382 | // and .gnu.hash. They are for the same purpose, and .gnu.hash is a new |
2383 | // and better version of .hash. .hash is just an on-disk hash table, but |
2384 | // .gnu.hash has a bloom filter in addition to a hash table to skip |
2385 | // DSOs very quickly. If you are sure that your dynamic linker knows |
2386 | // about .gnu.hash, you want to specify --hash-style=gnu. Otherwise, a |
2387 | // safe bet is to specify --hash-style=both for backward compatibility. |
2388 | GnuHashTableSection::GnuHashTableSection() |
2389 | : SyntheticSection(SHF_ALLOC, SHT_GNU_HASH, config->wordsize, ".gnu.hash" ) { |
2390 | } |
2391 | |
2392 | void GnuHashTableSection::finalizeContents() { |
2393 | if (OutputSection *sec = getPartition().dynSymTab->getParent()) |
2394 | getParent()->link = sec->sectionIndex; |
2395 | |
2396 | // Computes bloom filter size in word size. We want to allocate 12 |
2397 | // bits for each symbol. It must be a power of two. |
2398 | if (symbols.empty()) { |
2399 | maskWords = 1; |
2400 | } else { |
2401 | uint64_t numBits = symbols.size() * 12; |
2402 | maskWords = NextPowerOf2(A: numBits / (config->wordsize * 8)); |
2403 | } |
2404 | |
2405 | size = 16; // Header |
2406 | size += config->wordsize * maskWords; // Bloom filter |
2407 | size += nBuckets * 4; // Hash buckets |
2408 | size += symbols.size() * 4; // Hash values |
2409 | } |
2410 | |
2411 | void GnuHashTableSection::writeTo(uint8_t *buf) { |
2412 | // Write a header. |
2413 | write32(p: buf, v: nBuckets); |
2414 | write32(p: buf + 4, v: getPartition().dynSymTab->getNumSymbols() - symbols.size()); |
2415 | write32(p: buf + 8, v: maskWords); |
2416 | write32(p: buf + 12, v: Shift2); |
2417 | buf += 16; |
2418 | |
2419 | // Write the 2-bit bloom filter. |
2420 | const unsigned c = config->is64 ? 64 : 32; |
2421 | for (const Entry &sym : symbols) { |
2422 | // When C = 64, we choose a word with bits [6:...] and set 1 to two bits in |
2423 | // the word using bits [0:5] and [26:31]. |
2424 | size_t i = (sym.hash / c) & (maskWords - 1); |
2425 | uint64_t val = readUint(buf: buf + i * config->wordsize); |
2426 | val |= uint64_t(1) << (sym.hash % c); |
2427 | val |= uint64_t(1) << ((sym.hash >> Shift2) % c); |
2428 | writeUint(buf: buf + i * config->wordsize, val); |
2429 | } |
2430 | buf += config->wordsize * maskWords; |
2431 | |
2432 | // Write the hash table. |
2433 | uint32_t *buckets = reinterpret_cast<uint32_t *>(buf); |
2434 | uint32_t oldBucket = -1; |
2435 | uint32_t *values = buckets + nBuckets; |
2436 | for (auto i = symbols.begin(), e = symbols.end(); i != e; ++i) { |
2437 | // Write a hash value. It represents a sequence of chains that share the |
2438 | // same hash modulo value. The last element of each chain is terminated by |
2439 | // LSB 1. |
2440 | uint32_t hash = i->hash; |
2441 | bool isLastInChain = (i + 1) == e || i->bucketIdx != (i + 1)->bucketIdx; |
2442 | hash = isLastInChain ? hash | 1 : hash & ~1; |
2443 | write32(p: values++, v: hash); |
2444 | |
2445 | if (i->bucketIdx == oldBucket) |
2446 | continue; |
2447 | // Write a hash bucket. Hash buckets contain indices in the following hash |
2448 | // value table. |
2449 | write32(p: buckets + i->bucketIdx, |
2450 | v: getPartition().dynSymTab->getSymbolIndex(sym: *i->sym)); |
2451 | oldBucket = i->bucketIdx; |
2452 | } |
2453 | } |
2454 | |
2455 | // Add symbols to this symbol hash table. Note that this function |
2456 | // destructively sort a given vector -- which is needed because |
2457 | // GNU-style hash table places some sorting requirements. |
2458 | void GnuHashTableSection::addSymbols(SmallVectorImpl<SymbolTableEntry> &v) { |
2459 | // We cannot use 'auto' for Mid because GCC 6.1 cannot deduce |
2460 | // its type correctly. |
2461 | auto mid = |
2462 | std::stable_partition(first: v.begin(), last: v.end(), pred: [&](const SymbolTableEntry &s) { |
2463 | return !s.sym->isDefined() || s.sym->partition != partition; |
2464 | }); |
2465 | |
2466 | // We chose load factor 4 for the on-disk hash table. For each hash |
2467 | // collision, the dynamic linker will compare a uint32_t hash value. |
2468 | // Since the integer comparison is quite fast, we believe we can |
2469 | // make the load factor even larger. 4 is just a conservative choice. |
2470 | // |
2471 | // Note that we don't want to create a zero-sized hash table because |
2472 | // Android loader as of 2018 doesn't like a .gnu.hash containing such |
2473 | // table. If that's the case, we create a hash table with one unused |
2474 | // dummy slot. |
2475 | nBuckets = std::max<size_t>(a: (v.end() - mid) / 4, b: 1); |
2476 | |
2477 | if (mid == v.end()) |
2478 | return; |
2479 | |
2480 | for (SymbolTableEntry &ent : llvm::make_range(x: mid, y: v.end())) { |
2481 | Symbol *b = ent.sym; |
2482 | uint32_t hash = hashGnu(Name: b->getName()); |
2483 | uint32_t bucketIdx = hash % nBuckets; |
2484 | symbols.push_back(Elt: {.sym: b, .strTabOffset: ent.strTabOffset, .hash: hash, .bucketIdx: bucketIdx}); |
2485 | } |
2486 | |
2487 | llvm::sort(C&: symbols, Comp: [](const Entry &l, const Entry &r) { |
2488 | return std::tie(args: l.bucketIdx, args: l.strTabOffset) < |
2489 | std::tie(args: r.bucketIdx, args: r.strTabOffset); |
2490 | }); |
2491 | |
2492 | v.erase(CS: mid, CE: v.end()); |
2493 | for (const Entry &ent : symbols) |
2494 | v.push_back(Elt: {.sym: ent.sym, .strTabOffset: ent.strTabOffset}); |
2495 | } |
2496 | |
2497 | HashTableSection::HashTableSection() |
2498 | : SyntheticSection(SHF_ALLOC, SHT_HASH, 4, ".hash" ) { |
2499 | this->entsize = 4; |
2500 | } |
2501 | |
2502 | void HashTableSection::finalizeContents() { |
2503 | SymbolTableBaseSection *symTab = getPartition().dynSymTab.get(); |
2504 | |
2505 | if (OutputSection *sec = symTab->getParent()) |
2506 | getParent()->link = sec->sectionIndex; |
2507 | |
2508 | unsigned numEntries = 2; // nbucket and nchain. |
2509 | numEntries += symTab->getNumSymbols(); // The chain entries. |
2510 | |
2511 | // Create as many buckets as there are symbols. |
2512 | numEntries += symTab->getNumSymbols(); |
2513 | this->size = numEntries * 4; |
2514 | } |
2515 | |
2516 | void HashTableSection::writeTo(uint8_t *buf) { |
2517 | SymbolTableBaseSection *symTab = getPartition().dynSymTab.get(); |
2518 | unsigned numSymbols = symTab->getNumSymbols(); |
2519 | |
2520 | uint32_t *p = reinterpret_cast<uint32_t *>(buf); |
2521 | write32(p: p++, v: numSymbols); // nbucket |
2522 | write32(p: p++, v: numSymbols); // nchain |
2523 | |
2524 | uint32_t *buckets = p; |
2525 | uint32_t *chains = p + numSymbols; |
2526 | |
2527 | for (const SymbolTableEntry &s : symTab->getSymbols()) { |
2528 | Symbol *sym = s.sym; |
2529 | StringRef name = sym->getName(); |
2530 | unsigned i = sym->dynsymIndex; |
2531 | uint32_t hash = hashSysV(SymbolName: name) % numSymbols; |
2532 | chains[i] = buckets[hash]; |
2533 | write32(p: buckets + hash, v: i); |
2534 | } |
2535 | } |
2536 | |
2537 | PltSection::PltSection() |
2538 | : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".plt" ), |
2539 | headerSize(target->pltHeaderSize) { |
2540 | // On PowerPC, this section contains lazy symbol resolvers. |
2541 | if (config->emachine == EM_PPC64) { |
2542 | name = ".glink" ; |
2543 | addralign = 4; |
2544 | } |
2545 | |
2546 | // On x86 when IBT is enabled, this section contains the second PLT (lazy |
2547 | // symbol resolvers). |
2548 | if ((config->emachine == EM_386 || config->emachine == EM_X86_64) && |
2549 | (config->andFeatures & GNU_PROPERTY_X86_FEATURE_1_IBT)) |
2550 | name = ".plt.sec" ; |
2551 | |
2552 | // The PLT needs to be writable on SPARC as the dynamic linker will |
2553 | // modify the instructions in the PLT entries. |
2554 | if (config->emachine == EM_SPARCV9) |
2555 | this->flags |= SHF_WRITE; |
2556 | } |
2557 | |
2558 | void PltSection::writeTo(uint8_t *buf) { |
2559 | // At beginning of PLT, we have code to call the dynamic |
2560 | // linker to resolve dynsyms at runtime. Write such code. |
2561 | target->writePltHeader(buf); |
2562 | size_t off = headerSize; |
2563 | |
2564 | for (const Symbol *sym : entries) { |
2565 | target->writePlt(buf: buf + off, sym: *sym, pltEntryAddr: getVA() + off); |
2566 | off += target->pltEntrySize; |
2567 | } |
2568 | } |
2569 | |
2570 | void PltSection::addEntry(Symbol &sym) { |
2571 | assert(sym.auxIdx == symAux.size() - 1); |
2572 | symAux.back().pltIdx = entries.size(); |
2573 | entries.push_back(Elt: &sym); |
2574 | } |
2575 | |
2576 | size_t PltSection::getSize() const { |
2577 | return headerSize + entries.size() * target->pltEntrySize; |
2578 | } |
2579 | |
2580 | bool PltSection::isNeeded() const { |
2581 | // For -z retpolineplt, .iplt needs the .plt header. |
2582 | return !entries.empty() || (config->zRetpolineplt && in.iplt->isNeeded()); |
2583 | } |
2584 | |
2585 | // Used by ARM to add mapping symbols in the PLT section, which aid |
2586 | // disassembly. |
2587 | void PltSection::addSymbols() { |
2588 | target->addPltHeaderSymbols(isec&: *this); |
2589 | |
2590 | size_t off = headerSize; |
2591 | for (size_t i = 0; i < entries.size(); ++i) { |
2592 | target->addPltSymbols(isec&: *this, off); |
2593 | off += target->pltEntrySize; |
2594 | } |
2595 | } |
2596 | |
2597 | IpltSection::IpltSection() |
2598 | : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".iplt" ) { |
2599 | if (config->emachine == EM_PPC || config->emachine == EM_PPC64) { |
2600 | name = ".glink" ; |
2601 | addralign = 4; |
2602 | } |
2603 | } |
2604 | |
2605 | void IpltSection::writeTo(uint8_t *buf) { |
2606 | uint32_t off = 0; |
2607 | for (const Symbol *sym : entries) { |
2608 | target->writeIplt(buf: buf + off, sym: *sym, pltEntryAddr: getVA() + off); |
2609 | off += target->ipltEntrySize; |
2610 | } |
2611 | } |
2612 | |
2613 | size_t IpltSection::getSize() const { |
2614 | return entries.size() * target->ipltEntrySize; |
2615 | } |
2616 | |
2617 | void IpltSection::addEntry(Symbol &sym) { |
2618 | assert(sym.auxIdx == symAux.size() - 1); |
2619 | symAux.back().pltIdx = entries.size(); |
2620 | entries.push_back(Elt: &sym); |
2621 | } |
2622 | |
2623 | // ARM uses mapping symbols to aid disassembly. |
2624 | void IpltSection::addSymbols() { |
2625 | size_t off = 0; |
2626 | for (size_t i = 0, e = entries.size(); i != e; ++i) { |
2627 | target->addPltSymbols(isec&: *this, off); |
2628 | off += target->pltEntrySize; |
2629 | } |
2630 | } |
2631 | |
2632 | PPC32GlinkSection::PPC32GlinkSection() { |
2633 | name = ".glink" ; |
2634 | addralign = 4; |
2635 | } |
2636 | |
2637 | void PPC32GlinkSection::writeTo(uint8_t *buf) { |
2638 | writePPC32GlinkSection(buf, numEntries: entries.size()); |
2639 | } |
2640 | |
2641 | size_t PPC32GlinkSection::getSize() const { |
2642 | return headerSize + entries.size() * target->pltEntrySize + footerSize; |
2643 | } |
2644 | |
2645 | // This is an x86-only extra PLT section and used only when a security |
2646 | // enhancement feature called CET is enabled. In this comment, I'll explain what |
2647 | // the feature is and why we have two PLT sections if CET is enabled. |
2648 | // |
2649 | // So, what does CET do? CET introduces a new restriction to indirect jump |
2650 | // instructions. CET works this way. Assume that CET is enabled. Then, if you |
2651 | // execute an indirect jump instruction, the processor verifies that a special |
2652 | // "landing pad" instruction (which is actually a repurposed NOP instruction and |
2653 | // now called "endbr32" or "endbr64") is at the jump target. If the jump target |
2654 | // does not start with that instruction, the processor raises an exception |
2655 | // instead of continuing executing code. |
2656 | // |
2657 | // If CET is enabled, the compiler emits endbr to all locations where indirect |
2658 | // jumps may jump to. |
2659 | // |
2660 | // This mechanism makes it extremely hard to transfer the control to a middle of |
2661 | // a function that is not supporsed to be a indirect jump target, preventing |
2662 | // certain types of attacks such as ROP or JOP. |
2663 | // |
2664 | // Note that the processors in the market as of 2019 don't actually support the |
2665 | // feature. Only the spec is available at the moment. |
2666 | // |
2667 | // Now, I'll explain why we have this extra PLT section for CET. |
2668 | // |
2669 | // Since you can indirectly jump to a PLT entry, we have to make PLT entries |
2670 | // start with endbr. The problem is there's no extra space for endbr (which is 4 |
2671 | // bytes long), as the PLT entry is only 16 bytes long and all bytes are already |
2672 | // used. |
2673 | // |
2674 | // In order to deal with the issue, we split a PLT entry into two PLT entries. |
2675 | // Remember that each PLT entry contains code to jump to an address read from |
2676 | // .got.plt AND code to resolve a dynamic symbol lazily. With the 2-PLT scheme, |
2677 | // the former code is written to .plt.sec, and the latter code is written to |
2678 | // .plt. |
2679 | // |
2680 | // Lazy symbol resolution in the 2-PLT scheme works in the usual way, except |
2681 | // that the regular .plt is now called .plt.sec and .plt is repurposed to |
2682 | // contain only code for lazy symbol resolution. |
2683 | // |
2684 | // In other words, this is how the 2-PLT scheme works. Application code is |
2685 | // supposed to jump to .plt.sec to call an external function. Each .plt.sec |
2686 | // entry contains code to read an address from a corresponding .got.plt entry |
2687 | // and jump to that address. Addresses in .got.plt initially point to .plt, so |
2688 | // when an application calls an external function for the first time, the |
2689 | // control is transferred to a function that resolves a symbol name from |
2690 | // external shared object files. That function then rewrites a .got.plt entry |
2691 | // with a resolved address, so that the subsequent function calls directly jump |
2692 | // to a desired location from .plt.sec. |
2693 | // |
2694 | // There is an open question as to whether the 2-PLT scheme was desirable or |
2695 | // not. We could have simply extended the PLT entry size to 32-bytes to |
2696 | // accommodate endbr, and that scheme would have been much simpler than the |
2697 | // 2-PLT scheme. One reason to split PLT was, by doing that, we could keep hot |
2698 | // code (.plt.sec) from cold code (.plt). But as far as I know no one proved |
2699 | // that the optimization actually makes a difference. |
2700 | // |
2701 | // That said, the 2-PLT scheme is a part of the ABI, debuggers and other tools |
2702 | // depend on it, so we implement the ABI. |
2703 | IBTPltSection::IBTPltSection() |
2704 | : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".plt" ) {} |
2705 | |
2706 | void IBTPltSection::writeTo(uint8_t *buf) { |
2707 | target->writeIBTPlt(buf, numEntries: in.plt->getNumEntries()); |
2708 | } |
2709 | |
2710 | size_t IBTPltSection::getSize() const { |
2711 | // 16 is the header size of .plt. |
2712 | return 16 + in.plt->getNumEntries() * target->pltEntrySize; |
2713 | } |
2714 | |
2715 | bool IBTPltSection::isNeeded() const { return in.plt->getNumEntries() > 0; } |
2716 | |
2717 | RelroPaddingSection::RelroPaddingSection() |
2718 | : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_NOBITS, 1, ".relro_padding" ) { |
2719 | } |
2720 | |
2721 | // The string hash function for .gdb_index. |
2722 | static uint32_t computeGdbHash(StringRef s) { |
2723 | uint32_t h = 0; |
2724 | for (uint8_t c : s) |
2725 | h = h * 67 + toLower(x: c) - 113; |
2726 | return h; |
2727 | } |
2728 | |
2729 | // 4-byte alignment ensures that values in the hash lookup table and the name |
2730 | // table are aligned. |
2731 | DebugNamesBaseSection::DebugNamesBaseSection() |
2732 | : SyntheticSection(0, SHT_PROGBITS, 4, ".debug_names" ) {} |
2733 | |
2734 | // Get the size of the .debug_names section header in bytes for DWARF32: |
2735 | static uint32_t (uint32_t augmentationStringSize) { |
2736 | return /* unit length */ 4 + |
2737 | /* version */ 2 + |
2738 | /* padding */ 2 + |
2739 | /* CU count */ 4 + |
2740 | /* TU count */ 4 + |
2741 | /* Foreign TU count */ 4 + |
2742 | /* Bucket Count */ 4 + |
2743 | /* Name Count */ 4 + |
2744 | /* Abbrev table size */ 4 + |
2745 | /* Augmentation string size */ 4 + |
2746 | /* Augmentation string */ augmentationStringSize; |
2747 | } |
2748 | |
2749 | static Expected<DebugNamesBaseSection::IndexEntry *> |
2750 | (uint64_t &offset, const DWARFDebugNames::NameIndex &ni, |
2751 | uint64_t entriesBase, DWARFDataExtractor &, |
2752 | const LLDDWARFSection &namesSec) { |
2753 | auto ie = makeThreadLocal<DebugNamesBaseSection::IndexEntry>(); |
2754 | ie->poolOffset = offset; |
2755 | Error err = Error::success(); |
2756 | uint64_t ulebVal = namesExtractor.getULEB128(offset_ptr: &offset, Err: &err); |
2757 | if (err) |
2758 | return createStringError(EC: inconvertibleErrorCode(), |
2759 | Fmt: "invalid abbrev code: %s" , |
2760 | Vals: toString(E: std::move(err)).c_str()); |
2761 | if (!isUInt<32>(x: ulebVal)) |
2762 | return createStringError(EC: inconvertibleErrorCode(), |
2763 | Fmt: "abbrev code too large for DWARF32: %" PRIu64, |
2764 | Vals: ulebVal); |
2765 | ie->abbrevCode = static_cast<uint32_t>(ulebVal); |
2766 | auto it = ni.getAbbrevs().find_as(Val: ie->abbrevCode); |
2767 | if (it == ni.getAbbrevs().end()) |
2768 | return createStringError(EC: inconvertibleErrorCode(), |
2769 | Fmt: "abbrev code not found in abbrev table: %" PRIu32, |
2770 | Vals: ie->abbrevCode); |
2771 | |
2772 | DebugNamesBaseSection::AttrValue attr, cuAttr = {.attrValue: 0, .attrSize: 0}; |
2773 | for (DWARFDebugNames::AttributeEncoding a : it->Attributes) { |
2774 | if (a.Index == dwarf::DW_IDX_parent) { |
2775 | if (a.Form == dwarf::DW_FORM_ref4) { |
2776 | attr.attrValue = namesExtractor.getU32(offset_ptr: &offset, Err: &err); |
2777 | attr.attrSize = 4; |
2778 | ie->parentOffset = entriesBase + attr.attrValue; |
2779 | } else if (a.Form != DW_FORM_flag_present) |
2780 | return createStringError(EC: inconvertibleErrorCode(), |
2781 | S: "invalid form for DW_IDX_parent" ); |
2782 | } else { |
2783 | switch (a.Form) { |
2784 | case DW_FORM_data1: |
2785 | case DW_FORM_ref1: { |
2786 | attr.attrValue = namesExtractor.getU8(offset_ptr: &offset, Err: &err); |
2787 | attr.attrSize = 1; |
2788 | break; |
2789 | } |
2790 | case DW_FORM_data2: |
2791 | case DW_FORM_ref2: { |
2792 | attr.attrValue = namesExtractor.getU16(offset_ptr: &offset, Err: &err); |
2793 | attr.attrSize = 2; |
2794 | break; |
2795 | } |
2796 | case DW_FORM_data4: |
2797 | case DW_FORM_ref4: { |
2798 | attr.attrValue = namesExtractor.getU32(offset_ptr: &offset, Err: &err); |
2799 | attr.attrSize = 4; |
2800 | break; |
2801 | } |
2802 | default: |
2803 | return createStringError( |
2804 | EC: inconvertibleErrorCode(), |
2805 | Fmt: "unrecognized form encoding %d in abbrev table" , Vals: a.Form); |
2806 | } |
2807 | } |
2808 | if (err) |
2809 | return createStringError(EC: inconvertibleErrorCode(), |
2810 | Fmt: "error while reading attributes: %s" , |
2811 | Vals: toString(E: std::move(err)).c_str()); |
2812 | if (a.Index == DW_IDX_compile_unit) |
2813 | cuAttr = attr; |
2814 | else if (a.Form != DW_FORM_flag_present) |
2815 | ie->attrValues.push_back(Elt: attr); |
2816 | } |
2817 | // Canonicalize abbrev by placing the CU/TU index at the end. |
2818 | ie->attrValues.push_back(Elt: cuAttr); |
2819 | return ie; |
2820 | } |
2821 | |
2822 | void DebugNamesBaseSection::( |
2823 | InputChunk &inputChunk, OutputChunk &chunk, |
2824 | DWARFDataExtractor &, DataExtractor &, |
2825 | function_ref<SmallVector<uint32_t, 0>( |
2826 | uint32_t numCus, const DWARFDebugNames::Header &, |
2827 | const DWARFDebugNames::DWARFDebugNamesOffsets &)> |
2828 | readOffsets) { |
2829 | const LLDDWARFSection &namesSec = inputChunk.section; |
2830 | DenseMap<uint32_t, IndexEntry *> offsetMap; |
2831 | // Number of CUs seen in previous NameIndex sections within current chunk. |
2832 | uint32_t numCus = 0; |
2833 | for (const DWARFDebugNames::NameIndex &ni : *inputChunk.llvmDebugNames) { |
2834 | NameData &nd = inputChunk.nameData.emplace_back(); |
2835 | nd.hdr = ni.getHeader(); |
2836 | if (nd.hdr.Format != DwarfFormat::DWARF32) { |
2837 | errorOrWarn(msg: toString(namesSec.sec) + |
2838 | Twine(": found DWARF64, which is currently unsupported" )); |
2839 | return; |
2840 | } |
2841 | if (nd.hdr.Version != 5) { |
2842 | errorOrWarn(msg: toString(namesSec.sec) + Twine(": unsupported version: " ) + |
2843 | Twine(nd.hdr.Version)); |
2844 | return; |
2845 | } |
2846 | uint32_t dwarfSize = dwarf::getDwarfOffsetByteSize(Format: DwarfFormat::DWARF32); |
2847 | DWARFDebugNames::DWARFDebugNamesOffsets locs = ni.getOffsets(); |
2848 | if (locs.EntriesBase > namesExtractor.getData().size()) { |
2849 | errorOrWarn(msg: toString(namesSec.sec) + |
2850 | Twine(": entry pool start is beyond end of section" )); |
2851 | return; |
2852 | } |
2853 | |
2854 | SmallVector<uint32_t, 0> entryOffsets = readOffsets(numCus, nd.hdr, locs); |
2855 | |
2856 | // Read the entry pool. |
2857 | offsetMap.clear(); |
2858 | nd.nameEntries.resize(N: nd.hdr.NameCount); |
2859 | for (auto i : seq(Size: nd.hdr.NameCount)) { |
2860 | NameEntry &ne = nd.nameEntries[i]; |
2861 | uint64_t strOffset = locs.StringOffsetsBase + i * dwarfSize; |
2862 | ne.stringOffset = strOffset; |
2863 | uint64_t strp = namesExtractor.getRelocatedValue(Size: dwarfSize, Off: &strOffset); |
2864 | StringRef name = strExtractor.getCStrRef(OffsetPtr: &strp); |
2865 | ne.name = name.data(); |
2866 | ne.hashValue = caseFoldingDjbHash(Buffer: name); |
2867 | |
2868 | // Read a series of index entries that end with abbreviation code 0. |
2869 | uint64_t offset = locs.EntriesBase + entryOffsets[i]; |
2870 | while (offset < namesSec.Data.size() && namesSec.Data[offset] != 0) { |
2871 | // Read & store all entries (for the same string). |
2872 | Expected<IndexEntry *> ieOrErr = |
2873 | readEntry(offset, ni, entriesBase: locs.EntriesBase, namesExtractor, namesSec); |
2874 | if (!ieOrErr) { |
2875 | errorOrWarn(msg: toString(namesSec.sec) + ": " + |
2876 | toString(E: ieOrErr.takeError())); |
2877 | return; |
2878 | } |
2879 | ne.indexEntries.push_back(Elt: std::move(*ieOrErr)); |
2880 | } |
2881 | if (offset >= namesSec.Data.size()) |
2882 | errorOrWarn(msg: toString(namesSec.sec) + |
2883 | Twine(": index entry is out of bounds" )); |
2884 | |
2885 | for (IndexEntry &ie : ne.entries()) |
2886 | offsetMap[ie.poolOffset] = &ie; |
2887 | } |
2888 | |
2889 | // Assign parent pointers, which will be used to update DW_IDX_parent index |
2890 | // attributes. Note: offsetMap[0] does not exist, so parentOffset == 0 will |
2891 | // get parentEntry == null as well. |
2892 | for (NameEntry &ne : nd.nameEntries) |
2893 | for (IndexEntry &ie : ne.entries()) |
2894 | ie.parentEntry = offsetMap.lookup(Val: ie.parentOffset); |
2895 | numCus += nd.hdr.CompUnitCount; |
2896 | } |
2897 | } |
2898 | |
2899 | // Compute the form for output DW_IDX_compile_unit attributes, similar to |
2900 | // DIEInteger::BestForm. The input form (often DW_FORM_data1) may not hold all |
2901 | // the merged CU indices. |
2902 | std::pair<uint8_t, dwarf::Form> static getMergedCuCountForm( |
2903 | uint32_t compUnitCount) { |
2904 | if (compUnitCount > UINT16_MAX) |
2905 | return {4, DW_FORM_data4}; |
2906 | if (compUnitCount > UINT8_MAX) |
2907 | return {2, DW_FORM_data2}; |
2908 | return {1, DW_FORM_data1}; |
2909 | } |
2910 | |
2911 | void DebugNamesBaseSection::computeHdrAndAbbrevTable( |
2912 | MutableArrayRef<InputChunk> inputChunks) { |
2913 | TimeTraceScope timeScope("Merge .debug_names" , "hdr and abbrev table" ); |
2914 | size_t numCu = 0; |
2915 | hdr.Format = DwarfFormat::DWARF32; |
2916 | hdr.Version = 5; |
2917 | hdr.CompUnitCount = 0; |
2918 | hdr.LocalTypeUnitCount = 0; |
2919 | hdr.ForeignTypeUnitCount = 0; |
2920 | hdr.AugmentationStringSize = 0; |
2921 | |
2922 | // Compute CU and TU counts. |
2923 | for (auto i : seq(Size: numChunks)) { |
2924 | InputChunk &inputChunk = inputChunks[i]; |
2925 | inputChunk.baseCuIdx = numCu; |
2926 | numCu += chunks[i].compUnits.size(); |
2927 | for (const NameData &nd : inputChunk.nameData) { |
2928 | hdr.CompUnitCount += nd.hdr.CompUnitCount; |
2929 | // TODO: We don't handle type units yet, so LocalTypeUnitCount & |
2930 | // ForeignTypeUnitCount are left as 0. |
2931 | if (nd.hdr.LocalTypeUnitCount || nd.hdr.ForeignTypeUnitCount) |
2932 | warn(msg: toString(inputChunk.section.sec) + |
2933 | Twine(": type units are not implemented" )); |
2934 | // If augmentation strings are not identical, use an empty string. |
2935 | if (i == 0) { |
2936 | hdr.AugmentationStringSize = nd.hdr.AugmentationStringSize; |
2937 | hdr.AugmentationString = nd.hdr.AugmentationString; |
2938 | } else if (hdr.AugmentationString != nd.hdr.AugmentationString) { |
2939 | // There are conflicting augmentation strings, so it's best for the |
2940 | // merged index to not use an augmentation string. |
2941 | hdr.AugmentationStringSize = 0; |
2942 | hdr.AugmentationString.clear(); |
2943 | } |
2944 | } |
2945 | } |
2946 | |
2947 | // Create the merged abbrev table, uniquifyinng the input abbrev tables and |
2948 | // computing mapping from old (per-cu) abbrev codes to new (merged) abbrev |
2949 | // codes. |
2950 | FoldingSet<Abbrev> abbrevSet; |
2951 | // Determine the form for the DW_IDX_compile_unit attributes in the merged |
2952 | // index. The input form may not be big enough for all CU indices. |
2953 | dwarf::Form cuAttrForm = getMergedCuCountForm(compUnitCount: hdr.CompUnitCount).second; |
2954 | for (InputChunk &inputChunk : inputChunks) { |
2955 | for (auto [i, ni] : enumerate(First&: *inputChunk.llvmDebugNames)) { |
2956 | for (const DWARFDebugNames::Abbrev &oldAbbrev : ni.getAbbrevs()) { |
2957 | // Canonicalize abbrev by placing the CU/TU index at the end, |
2958 | // similar to 'parseDebugNames'. |
2959 | Abbrev abbrev; |
2960 | DWARFDebugNames::AttributeEncoding cuAttr(DW_IDX_compile_unit, |
2961 | cuAttrForm); |
2962 | abbrev.code = oldAbbrev.Code; |
2963 | abbrev.tag = oldAbbrev.Tag; |
2964 | for (DWARFDebugNames::AttributeEncoding a : oldAbbrev.Attributes) { |
2965 | if (a.Index == DW_IDX_compile_unit) |
2966 | cuAttr.Index = a.Index; |
2967 | else |
2968 | abbrev.attributes.push_back(Elt: {a.Index, a.Form}); |
2969 | } |
2970 | // Put the CU/TU index at the end of the attributes list. |
2971 | abbrev.attributes.push_back(Elt: cuAttr); |
2972 | |
2973 | // Profile the abbrev, get or assign a new code, then record the abbrev |
2974 | // code mapping. |
2975 | FoldingSetNodeID id; |
2976 | abbrev.Profile(id); |
2977 | uint32_t newCode; |
2978 | void *insertPos; |
2979 | if (Abbrev *existing = abbrevSet.FindNodeOrInsertPos(ID: id, InsertPos&: insertPos)) { |
2980 | // Found it; we've already seen an identical abbreviation. |
2981 | newCode = existing->code; |
2982 | } else { |
2983 | Abbrev *abbrev2 = |
2984 | new (abbrevAlloc.Allocate()) Abbrev(std::move(abbrev)); |
2985 | abbrevSet.InsertNode(N: abbrev2, InsertPos: insertPos); |
2986 | abbrevTable.push_back(Elt: abbrev2); |
2987 | newCode = abbrevTable.size(); |
2988 | abbrev2->code = newCode; |
2989 | } |
2990 | inputChunk.nameData[i].abbrevCodeMap[oldAbbrev.Code] = newCode; |
2991 | } |
2992 | } |
2993 | } |
2994 | |
2995 | // Compute the merged abbrev table. |
2996 | raw_svector_ostream os(abbrevTableBuf); |
2997 | for (Abbrev *abbrev : abbrevTable) { |
2998 | encodeULEB128(Value: abbrev->code, OS&: os); |
2999 | encodeULEB128(Value: abbrev->tag, OS&: os); |
3000 | for (DWARFDebugNames::AttributeEncoding a : abbrev->attributes) { |
3001 | encodeULEB128(Value: a.Index, OS&: os); |
3002 | encodeULEB128(Value: a.Form, OS&: os); |
3003 | } |
3004 | os.write(Ptr: "\0" , Size: 2); // attribute specification end |
3005 | } |
3006 | os.write(C: 0); // abbrev table end |
3007 | hdr.AbbrevTableSize = abbrevTableBuf.size(); |
3008 | } |
3009 | |
3010 | void DebugNamesBaseSection::Abbrev::Profile(FoldingSetNodeID &id) const { |
3011 | id.AddInteger(I: tag); |
3012 | for (const DWARFDebugNames::AttributeEncoding &attr : attributes) { |
3013 | id.AddInteger(I: attr.Index); |
3014 | id.AddInteger(I: attr.Form); |
3015 | } |
3016 | } |
3017 | |
3018 | std::pair<uint32_t, uint32_t> DebugNamesBaseSection::computeEntryPool( |
3019 | MutableArrayRef<InputChunk> inputChunks) { |
3020 | TimeTraceScope timeScope("Merge .debug_names" , "entry pool" ); |
3021 | // Collect and de-duplicate all the names (preserving all the entries). |
3022 | // Speed it up using multithreading, as the number of symbols can be in the |
3023 | // order of millions. |
3024 | const size_t concurrency = |
3025 | bit_floor(Value: std::min<size_t>(a: config->threadCount, b: numShards)); |
3026 | const size_t shift = 32 - countr_zero(Val: numShards); |
3027 | const uint8_t cuAttrSize = getMergedCuCountForm(compUnitCount: hdr.CompUnitCount).first; |
3028 | DenseMap<CachedHashStringRef, size_t> maps[numShards]; |
3029 | |
3030 | parallelFor(Begin: 0, End: concurrency, Fn: [&](size_t threadId) { |
3031 | for (auto i : seq(Size: numChunks)) { |
3032 | InputChunk &inputChunk = inputChunks[i]; |
3033 | for (auto j : seq(Size: inputChunk.nameData.size())) { |
3034 | NameData &nd = inputChunk.nameData[j]; |
3035 | // Deduplicate the NameEntry records (based on the string/name), |
3036 | // appending all IndexEntries from duplicate NameEntry records to |
3037 | // the single preserved copy. |
3038 | for (NameEntry &ne : nd.nameEntries) { |
3039 | auto shardId = ne.hashValue >> shift; |
3040 | if ((shardId & (concurrency - 1)) != threadId) |
3041 | continue; |
3042 | |
3043 | ne.chunkIdx = i; |
3044 | for (IndexEntry &ie : ne.entries()) { |
3045 | // Update the IndexEntry's abbrev code to match the merged |
3046 | // abbreviations. |
3047 | ie.abbrevCode = nd.abbrevCodeMap[ie.abbrevCode]; |
3048 | // Update the DW_IDX_compile_unit attribute (the last one after |
3049 | // canonicalization) to have correct merged offset value and size. |
3050 | auto &back = ie.attrValues.back(); |
3051 | back.attrValue += inputChunk.baseCuIdx + j; |
3052 | back.attrSize = cuAttrSize; |
3053 | } |
3054 | |
3055 | auto &nameVec = nameVecs[shardId]; |
3056 | auto [it, inserted] = maps[shardId].try_emplace( |
3057 | Key: CachedHashStringRef(ne.name, ne.hashValue), Args: nameVec.size()); |
3058 | if (inserted) |
3059 | nameVec.push_back(Elt: std::move(ne)); |
3060 | else |
3061 | nameVec[it->second].indexEntries.append(RHS: std::move(ne.indexEntries)); |
3062 | } |
3063 | } |
3064 | } |
3065 | }); |
3066 | |
3067 | // Compute entry offsets in parallel. First, compute offsets relative to the |
3068 | // current shard. |
3069 | uint32_t offsets[numShards]; |
3070 | parallelFor(Begin: 0, End: numShards, Fn: [&](size_t shard) { |
3071 | uint32_t offset = 0; |
3072 | for (NameEntry &ne : nameVecs[shard]) { |
3073 | ne.entryOffset = offset; |
3074 | for (IndexEntry &ie : ne.entries()) { |
3075 | ie.poolOffset = offset; |
3076 | offset += getULEB128Size(Value: ie.abbrevCode); |
3077 | for (AttrValue value : ie.attrValues) |
3078 | offset += value.attrSize; |
3079 | } |
3080 | ++offset; // index entry sentinel |
3081 | } |
3082 | offsets[shard] = offset; |
3083 | }); |
3084 | // Then add shard offsets. |
3085 | std::partial_sum(first: offsets, last: std::end(arr&: offsets), result: offsets); |
3086 | parallelFor(Begin: 1, End: numShards, Fn: [&](size_t shard) { |
3087 | uint32_t offset = offsets[shard - 1]; |
3088 | for (NameEntry &ne : nameVecs[shard]) { |
3089 | ne.entryOffset += offset; |
3090 | for (IndexEntry &ie : ne.entries()) |
3091 | ie.poolOffset += offset; |
3092 | } |
3093 | }); |
3094 | |
3095 | // Update the DW_IDX_parent entries that refer to real parents (have |
3096 | // DW_FORM_ref4). |
3097 | parallelFor(Begin: 0, End: numShards, Fn: [&](size_t shard) { |
3098 | for (NameEntry &ne : nameVecs[shard]) { |
3099 | for (IndexEntry &ie : ne.entries()) { |
3100 | if (!ie.parentEntry) |
3101 | continue; |
3102 | // Abbrevs are indexed starting at 1; vector starts at 0. (abbrevCode |
3103 | // corresponds to position in the merged table vector). |
3104 | const Abbrev *abbrev = abbrevTable[ie.abbrevCode - 1]; |
3105 | for (const auto &[a, v] : zip_equal(t: abbrev->attributes, u&: ie.attrValues)) |
3106 | if (a.Index == DW_IDX_parent && a.Form == DW_FORM_ref4) |
3107 | v.attrValue = ie.parentEntry->poolOffset; |
3108 | } |
3109 | } |
3110 | }); |
3111 | |
3112 | // Return (entry pool size, number of entries). |
3113 | uint32_t num = 0; |
3114 | for (auto &map : maps) |
3115 | num += map.size(); |
3116 | return {offsets[numShards - 1], num}; |
3117 | } |
3118 | |
3119 | void DebugNamesBaseSection::init( |
3120 | function_ref<void(InputFile *, InputChunk &, OutputChunk &)> parseFile) { |
3121 | TimeTraceScope timeScope("Merge .debug_names" ); |
3122 | // Collect and remove input .debug_names sections. Save InputSection pointers |
3123 | // to relocate string offsets in `writeTo`. |
3124 | SetVector<InputFile *> files; |
3125 | for (InputSectionBase *s : ctx.inputSections) { |
3126 | InputSection *isec = dyn_cast<InputSection>(Val: s); |
3127 | if (!isec) |
3128 | continue; |
3129 | if (!(s->flags & SHF_ALLOC) && s->name == ".debug_names" ) { |
3130 | s->markDead(); |
3131 | inputSections.push_back(Elt: isec); |
3132 | files.insert(X: isec->file); |
3133 | } |
3134 | } |
3135 | |
3136 | // Parse input .debug_names sections and extract InputChunk and OutputChunk |
3137 | // data. OutputChunk contains CU information, which will be needed by |
3138 | // `writeTo`. |
3139 | auto inputChunksPtr = std::make_unique<InputChunk[]>(num: files.size()); |
3140 | MutableArrayRef<InputChunk> inputChunks(inputChunksPtr.get(), files.size()); |
3141 | numChunks = files.size(); |
3142 | chunks = std::make_unique<OutputChunk[]>(num: files.size()); |
3143 | { |
3144 | TimeTraceScope timeScope("Merge .debug_names" , "parse" ); |
3145 | parallelFor(Begin: 0, End: files.size(), Fn: [&](size_t i) { |
3146 | parseFile(files[i], inputChunks[i], chunks[i]); |
3147 | }); |
3148 | } |
3149 | |
3150 | // Compute section header (except unit_length), abbrev table, and entry pool. |
3151 | computeHdrAndAbbrevTable(inputChunks); |
3152 | uint32_t entryPoolSize; |
3153 | std::tie(args&: entryPoolSize, args&: hdr.NameCount) = computeEntryPool(inputChunks); |
3154 | hdr.BucketCount = dwarf::getDebugNamesBucketCount(UniqueHashCount: hdr.NameCount); |
3155 | |
3156 | // Compute the section size. Subtract 4 to get the unit_length for DWARF32. |
3157 | uint32_t hdrSize = getDebugNamesHeaderSize(augmentationStringSize: hdr.AugmentationStringSize); |
3158 | size = findDebugNamesOffsets(EndOfHeaderOffset: hdrSize, Hdr: hdr).EntriesBase + entryPoolSize; |
3159 | hdr.UnitLength = size - 4; |
3160 | } |
3161 | |
3162 | template <class ELFT> DebugNamesSection<ELFT>::DebugNamesSection() { |
3163 | init(parseFile: [](InputFile *f, InputChunk &inputChunk, OutputChunk &chunk) { |
3164 | auto *file = cast<ObjFile<ELFT>>(f); |
3165 | DWARFContext dwarf(std::make_unique<LLDDwarfObj<ELFT>>(file)); |
3166 | auto &dobj = static_cast<const LLDDwarfObj<ELFT> &>(dwarf.getDWARFObj()); |
3167 | chunk.infoSec = dobj.getInfoSection(); |
3168 | DWARFDataExtractor (dobj, dobj.getNamesSection(), |
3169 | ELFT::Endianness == endianness::little, |
3170 | ELFT::Is64Bits ? 8 : 4); |
3171 | // .debug_str is needed to get symbol names from string offsets. |
3172 | DataExtractor (dobj.getStrSection(), |
3173 | ELFT::Endianness == endianness::little, |
3174 | ELFT::Is64Bits ? 8 : 4); |
3175 | inputChunk.section = dobj.getNamesSection(); |
3176 | |
3177 | inputChunk.llvmDebugNames.emplace(args&: namesExtractor, args&: strExtractor); |
3178 | if (Error e = inputChunk.llvmDebugNames->extract()) { |
3179 | errorOrWarn(toString(dobj.getNamesSection().sec) + Twine(": " ) + |
3180 | toString(E: std::move(e))); |
3181 | } |
3182 | parseDebugNames( |
3183 | inputChunk, chunk, namesExtractor, strExtractor, |
3184 | readOffsets: [&chunk, namesData = dobj.getNamesSection().Data.data()]( |
3185 | uint32_t numCus, const DWARFDebugNames::Header &hdr, |
3186 | const DWARFDebugNames::DWARFDebugNamesOffsets &locs) { |
3187 | // Read CU offsets, which are relocated by .debug_info + X |
3188 | // relocations. Record the section offset to be relocated by |
3189 | // `finalizeContents`. |
3190 | chunk.compUnits.resize_for_overwrite(N: numCus + hdr.CompUnitCount); |
3191 | for (auto i : seq(Size: hdr.CompUnitCount)) |
3192 | chunk.compUnits[numCus + i] = locs.CUsBase + i * 4; |
3193 | |
3194 | // Read entry offsets. |
3195 | const char *p = namesData + locs.EntryOffsetsBase; |
3196 | SmallVector<uint32_t, 0> entryOffsets; |
3197 | entryOffsets.resize_for_overwrite(N: hdr.NameCount); |
3198 | for (uint32_t &offset : entryOffsets) |
3199 | offset = endian::readNext<uint32_t, ELFT::Endianness, unaligned>(p); |
3200 | return entryOffsets; |
3201 | }); |
3202 | }); |
3203 | } |
3204 | |
3205 | template <class ELFT> |
3206 | template <class RelTy> |
3207 | void DebugNamesSection<ELFT>::getNameRelocs( |
3208 | const InputFile &file, DenseMap<uint32_t, uint32_t> &relocs, |
3209 | Relocs<RelTy> rels) { |
3210 | for (const RelTy &rel : rels) { |
3211 | Symbol &sym = file.getRelocTargetSym(rel); |
3212 | relocs[rel.r_offset] = sym.getVA(addend: getAddend<ELFT>(rel)); |
3213 | } |
3214 | } |
3215 | |
3216 | template <class ELFT> void DebugNamesSection<ELFT>::finalizeContents() { |
3217 | // Get relocations of .debug_names sections. |
3218 | auto relocs = std::make_unique<DenseMap<uint32_t, uint32_t>[]>(numChunks); |
3219 | parallelFor(0, numChunks, [&](size_t i) { |
3220 | InputSection *sec = inputSections[i]; |
3221 | invokeOnRelocs(*sec, getNameRelocs, *sec->file, relocs.get()[i]); |
3222 | |
3223 | // Relocate CU offsets with .debug_info + X relocations. |
3224 | OutputChunk &chunk = chunks.get()[i]; |
3225 | for (auto [j, cuOffset] : enumerate(First&: chunk.compUnits)) |
3226 | cuOffset = relocs.get()[i].lookup(cuOffset); |
3227 | }); |
3228 | |
3229 | // Relocate string offsets in the name table with .debug_str + X relocations. |
3230 | parallelForEach(nameVecs, [&](auto &nameVec) { |
3231 | for (NameEntry &ne : nameVec) |
3232 | ne.stringOffset = relocs.get()[ne.chunkIdx].lookup(ne.stringOffset); |
3233 | }); |
3234 | } |
3235 | |
3236 | template <class ELFT> void DebugNamesSection<ELFT>::writeTo(uint8_t *buf) { |
3237 | [[maybe_unused]] const uint8_t *const beginBuf = buf; |
3238 | // Write the header. |
3239 | endian::writeNext<uint32_t, ELFT::Endianness>(buf, hdr.UnitLength); |
3240 | endian::writeNext<uint16_t, ELFT::Endianness>(buf, hdr.Version); |
3241 | buf += 2; // padding |
3242 | endian::writeNext<uint32_t, ELFT::Endianness>(buf, hdr.CompUnitCount); |
3243 | endian::writeNext<uint32_t, ELFT::Endianness>(buf, hdr.LocalTypeUnitCount); |
3244 | endian::writeNext<uint32_t, ELFT::Endianness>(buf, hdr.ForeignTypeUnitCount); |
3245 | endian::writeNext<uint32_t, ELFT::Endianness>(buf, hdr.BucketCount); |
3246 | endian::writeNext<uint32_t, ELFT::Endianness>(buf, hdr.NameCount); |
3247 | endian::writeNext<uint32_t, ELFT::Endianness>(buf, hdr.AbbrevTableSize); |
3248 | endian::writeNext<uint32_t, ELFT::Endianness>(buf, |
3249 | hdr.AugmentationStringSize); |
3250 | memcpy(buf, hdr.AugmentationString.c_str(), hdr.AugmentationString.size()); |
3251 | buf += hdr.AugmentationStringSize; |
3252 | |
3253 | // Write the CU list. |
3254 | for (auto &chunk : getChunks()) |
3255 | for (uint32_t cuOffset : chunk.compUnits) |
3256 | endian::writeNext<uint32_t, ELFT::Endianness>(buf, cuOffset); |
3257 | |
3258 | // TODO: Write the local TU list, then the foreign TU list.. |
3259 | |
3260 | // Write the hash lookup table. |
3261 | SmallVector<SmallVector<NameEntry *, 0>, 0> buckets(hdr.BucketCount); |
3262 | // Symbols enter into a bucket whose index is the hash modulo bucket_count. |
3263 | for (auto &nameVec : nameVecs) |
3264 | for (NameEntry &ne : nameVec) |
3265 | buckets[ne.hashValue % hdr.BucketCount].push_back(&ne); |
3266 | |
3267 | // Write buckets (accumulated bucket counts). |
3268 | uint32_t bucketIdx = 1; |
3269 | for (const SmallVector<NameEntry *, 0> &bucket : buckets) { |
3270 | if (!bucket.empty()) |
3271 | endian::write32<ELFT::Endianness>(buf, bucketIdx); |
3272 | buf += 4; |
3273 | bucketIdx += bucket.size(); |
3274 | } |
3275 | // Write the hashes. |
3276 | for (const SmallVector<NameEntry *, 0> &bucket : buckets) |
3277 | for (const NameEntry *e : bucket) |
3278 | endian::writeNext<uint32_t, ELFT::Endianness>(buf, e->hashValue); |
3279 | |
3280 | // Write the name table. The name entries are ordered by bucket_idx and |
3281 | // correspond one-to-one with the hash lookup table. |
3282 | // |
3283 | // First, write the relocated string offsets. |
3284 | for (const SmallVector<NameEntry *, 0> &bucket : buckets) |
3285 | for (const NameEntry *ne : bucket) |
3286 | endian::writeNext<uint32_t, ELFT::Endianness>(buf, ne->stringOffset); |
3287 | |
3288 | // Then write the entry offsets. |
3289 | for (const SmallVector<NameEntry *, 0> &bucket : buckets) |
3290 | for (const NameEntry *ne : bucket) |
3291 | endian::writeNext<uint32_t, ELFT::Endianness>(buf, ne->entryOffset); |
3292 | |
3293 | // Write the abbrev table. |
3294 | buf = llvm::copy(abbrevTableBuf, buf); |
3295 | |
3296 | // Write the entry pool. Unlike the name table, the name entries follow the |
3297 | // nameVecs order computed by `computeEntryPool`. |
3298 | for (auto &nameVec : nameVecs) { |
3299 | for (NameEntry &ne : nameVec) { |
3300 | // Write all the entries for the string. |
3301 | for (const IndexEntry &ie : ne.entries()) { |
3302 | buf += encodeULEB128(Value: ie.abbrevCode, p: buf); |
3303 | for (AttrValue value : ie.attrValues) { |
3304 | switch (value.attrSize) { |
3305 | case 1: |
3306 | *buf++ = value.attrValue; |
3307 | break; |
3308 | case 2: |
3309 | endian::writeNext<uint16_t, ELFT::Endianness>(buf, value.attrValue); |
3310 | break; |
3311 | case 4: |
3312 | endian::writeNext<uint32_t, ELFT::Endianness>(buf, value.attrValue); |
3313 | break; |
3314 | default: |
3315 | llvm_unreachable("invalid attrSize" ); |
3316 | } |
3317 | } |
3318 | } |
3319 | ++buf; // index entry sentinel |
3320 | } |
3321 | } |
3322 | assert(uint64_t(buf - beginBuf) == size); |
3323 | } |
3324 | |
3325 | GdbIndexSection::GdbIndexSection() |
3326 | : SyntheticSection(0, SHT_PROGBITS, 1, ".gdb_index" ) {} |
3327 | |
3328 | // Returns the desired size of an on-disk hash table for a .gdb_index section. |
3329 | // There's a tradeoff between size and collision rate. We aim 75% utilization. |
3330 | size_t GdbIndexSection::computeSymtabSize() const { |
3331 | return std::max<size_t>(a: NextPowerOf2(A: symbols.size() * 4 / 3), b: 1024); |
3332 | } |
3333 | |
3334 | static SmallVector<GdbIndexSection::CuEntry, 0> |
3335 | readCuList(DWARFContext &dwarf) { |
3336 | SmallVector<GdbIndexSection::CuEntry, 0> ret; |
3337 | for (std::unique_ptr<DWARFUnit> &cu : dwarf.compile_units()) |
3338 | ret.push_back(Elt: {.cuOffset: cu->getOffset(), .cuLength: cu->getLength() + 4}); |
3339 | return ret; |
3340 | } |
3341 | |
3342 | static SmallVector<GdbIndexSection::AddressEntry, 0> |
3343 | readAddressAreas(DWARFContext &dwarf, InputSection *sec) { |
3344 | SmallVector<GdbIndexSection::AddressEntry, 0> ret; |
3345 | |
3346 | uint32_t cuIdx = 0; |
3347 | for (std::unique_ptr<DWARFUnit> &cu : dwarf.compile_units()) { |
3348 | if (Error e = cu->tryExtractDIEsIfNeeded(CUDieOnly: false)) { |
3349 | warn(msg: toString(sec) + ": " + toString(E: std::move(e))); |
3350 | return {}; |
3351 | } |
3352 | Expected<DWARFAddressRangesVector> ranges = cu->collectAddressRanges(); |
3353 | if (!ranges) { |
3354 | warn(msg: toString(sec) + ": " + toString(E: ranges.takeError())); |
3355 | return {}; |
3356 | } |
3357 | |
3358 | ArrayRef<InputSectionBase *> sections = sec->file->getSections(); |
3359 | for (DWARFAddressRange &r : *ranges) { |
3360 | if (r.SectionIndex == -1ULL) |
3361 | continue; |
3362 | // Range list with zero size has no effect. |
3363 | InputSectionBase *s = sections[r.SectionIndex]; |
3364 | if (s && s != &InputSection::discarded && s->isLive()) |
3365 | if (r.LowPC != r.HighPC) |
3366 | ret.push_back(Elt: {.section: cast<InputSection>(Val: s), .lowAddress: r.LowPC, .highAddress: r.HighPC, .cuIndex: cuIdx}); |
3367 | } |
3368 | ++cuIdx; |
3369 | } |
3370 | |
3371 | return ret; |
3372 | } |
3373 | |
3374 | template <class ELFT> |
3375 | static SmallVector<GdbIndexSection::NameAttrEntry, 0> |
3376 | readPubNamesAndTypes(const LLDDwarfObj<ELFT> &obj, |
3377 | const SmallVectorImpl<GdbIndexSection::CuEntry> &cus) { |
3378 | const LLDDWARFSection &pubNames = obj.getGnuPubnamesSection(); |
3379 | const LLDDWARFSection &pubTypes = obj.getGnuPubtypesSection(); |
3380 | |
3381 | SmallVector<GdbIndexSection::NameAttrEntry, 0> ret; |
3382 | for (const LLDDWARFSection *pub : {&pubNames, &pubTypes}) { |
3383 | DWARFDataExtractor data(obj, *pub, ELFT::Endianness == endianness::little, |
3384 | ELFT::Is64Bits ? 8 : 4); |
3385 | DWARFDebugPubTable table; |
3386 | table.extract(Data: data, /*GnuStyle=*/true, RecoverableErrorHandler: [&](Error e) { |
3387 | warn(msg: toString(pub->sec) + ": " + toString(E: std::move(e))); |
3388 | }); |
3389 | for (const DWARFDebugPubTable::Set &set : table.getData()) { |
3390 | // The value written into the constant pool is kind << 24 | cuIndex. As we |
3391 | // don't know how many compilation units precede this object to compute |
3392 | // cuIndex, we compute (kind << 24 | cuIndexInThisObject) instead, and add |
3393 | // the number of preceding compilation units later. |
3394 | uint32_t i = llvm::partition_point(cus, |
3395 | [&](GdbIndexSection::CuEntry cu) { |
3396 | return cu.cuOffset < set.Offset; |
3397 | }) - |
3398 | cus.begin(); |
3399 | for (const DWARFDebugPubTable::Entry &ent : set.Entries) |
3400 | ret.push_back(Elt: {.name: {ent.Name, computeGdbHash(s: ent.Name)}, |
3401 | .cuIndexAndAttrs: (ent.Descriptor.toBits() << 24) | i}); |
3402 | } |
3403 | } |
3404 | return ret; |
3405 | } |
3406 | |
3407 | // Create a list of symbols from a given list of symbol names and types |
3408 | // by uniquifying them by name. |
3409 | static std::pair<SmallVector<GdbIndexSection::GdbSymbol, 0>, size_t> |
3410 | createSymbols( |
3411 | ArrayRef<SmallVector<GdbIndexSection::NameAttrEntry, 0>> nameAttrs, |
3412 | const SmallVector<GdbIndexSection::GdbChunk, 0> &chunks) { |
3413 | using GdbSymbol = GdbIndexSection::GdbSymbol; |
3414 | using NameAttrEntry = GdbIndexSection::NameAttrEntry; |
3415 | |
3416 | // For each chunk, compute the number of compilation units preceding it. |
3417 | uint32_t cuIdx = 0; |
3418 | std::unique_ptr<uint32_t[]> cuIdxs(new uint32_t[chunks.size()]); |
3419 | for (uint32_t i = 0, e = chunks.size(); i != e; ++i) { |
3420 | cuIdxs[i] = cuIdx; |
3421 | cuIdx += chunks[i].compilationUnits.size(); |
3422 | } |
3423 | |
3424 | // Collect the compilation unitss for each unique name. Speed it up using |
3425 | // multi-threading as the number of symbols can be in the order of millions. |
3426 | // Shard GdbSymbols by hash's high bits. |
3427 | constexpr size_t numShards = 32; |
3428 | const size_t concurrency = |
3429 | llvm::bit_floor(Value: std::min<size_t>(a: config->threadCount, b: numShards)); |
3430 | const size_t shift = 32 - llvm::countr_zero(Val: numShards); |
3431 | auto map = |
3432 | std::make_unique<DenseMap<CachedHashStringRef, size_t>[]>(num: numShards); |
3433 | auto symbols = std::make_unique<SmallVector<GdbSymbol, 0>[]>(num: numShards); |
3434 | parallelFor(Begin: 0, End: concurrency, Fn: [&](size_t threadId) { |
3435 | uint32_t i = 0; |
3436 | for (ArrayRef<NameAttrEntry> entries : nameAttrs) { |
3437 | for (const NameAttrEntry &ent : entries) { |
3438 | size_t shardId = ent.name.hash() >> shift; |
3439 | if ((shardId & (concurrency - 1)) != threadId) |
3440 | continue; |
3441 | |
3442 | uint32_t v = ent.cuIndexAndAttrs + cuIdxs[i]; |
3443 | auto [it, inserted] = |
3444 | map[shardId].try_emplace(Key: ent.name, Args: symbols[shardId].size()); |
3445 | if (inserted) |
3446 | symbols[shardId].push_back(Elt: {.name: ent.name, .cuVector: {v}, .nameOff: 0, .cuVectorOff: 0}); |
3447 | else |
3448 | symbols[shardId][it->second].cuVector.push_back(Elt: v); |
3449 | } |
3450 | ++i; |
3451 | } |
3452 | }); |
3453 | |
3454 | size_t numSymbols = 0; |
3455 | for (ArrayRef<GdbSymbol> v : ArrayRef(symbols.get(), numShards)) |
3456 | numSymbols += v.size(); |
3457 | |
3458 | // The return type is a flattened vector, so we'll copy each vector |
3459 | // contents to Ret. |
3460 | SmallVector<GdbSymbol, 0> ret; |
3461 | ret.reserve(N: numSymbols); |
3462 | for (SmallVector<GdbSymbol, 0> &vec : |
3463 | MutableArrayRef(symbols.get(), numShards)) |
3464 | for (GdbSymbol &sym : vec) |
3465 | ret.push_back(Elt: std::move(sym)); |
3466 | |
3467 | // CU vectors and symbol names are adjacent in the output file. |
3468 | // We can compute their offsets in the output file now. |
3469 | size_t off = 0; |
3470 | for (GdbSymbol &sym : ret) { |
3471 | sym.cuVectorOff = off; |
3472 | off += (sym.cuVector.size() + 1) * 4; |
3473 | } |
3474 | for (GdbSymbol &sym : ret) { |
3475 | sym.nameOff = off; |
3476 | off += sym.name.size() + 1; |
3477 | } |
3478 | // If off overflows, the last symbol's nameOff likely overflows. |
3479 | if (!isUInt<32>(x: off)) |
3480 | errorOrWarn(msg: "--gdb-index: constant pool size (" + Twine(off) + |
3481 | ") exceeds UINT32_MAX" ); |
3482 | |
3483 | return {ret, off}; |
3484 | } |
3485 | |
3486 | // Returns a newly-created .gdb_index section. |
3487 | template <class ELFT> |
3488 | std::unique_ptr<GdbIndexSection> GdbIndexSection::create() { |
3489 | llvm::TimeTraceScope timeScope("Create gdb index" ); |
3490 | |
3491 | // Collect InputFiles with .debug_info. See the comment in |
3492 | // LLDDwarfObj<ELFT>::LLDDwarfObj. If we do lightweight parsing in the future, |
3493 | // note that isec->data() may uncompress the full content, which should be |
3494 | // parallelized. |
3495 | SetVector<InputFile *> files; |
3496 | for (InputSectionBase *s : ctx.inputSections) { |
3497 | InputSection *isec = dyn_cast<InputSection>(Val: s); |
3498 | if (!isec) |
3499 | continue; |
3500 | // .debug_gnu_pub{names,types} are useless in executables. |
3501 | // They are present in input object files solely for creating |
3502 | // a .gdb_index. So we can remove them from the output. |
3503 | if (s->name == ".debug_gnu_pubnames" || s->name == ".debug_gnu_pubtypes" ) |
3504 | s->markDead(); |
3505 | else if (isec->name == ".debug_info" ) |
3506 | files.insert(X: isec->file); |
3507 | } |
3508 | // Drop .rel[a].debug_gnu_pub{names,types} for --emit-relocs. |
3509 | llvm::erase_if(ctx.inputSections, [](InputSectionBase *s) { |
3510 | if (auto *isec = dyn_cast<InputSection>(Val: s)) |
3511 | if (InputSectionBase *rel = isec->getRelocatedSection()) |
3512 | return !rel->isLive(); |
3513 | return !s->isLive(); |
3514 | }); |
3515 | |
3516 | SmallVector<GdbChunk, 0> chunks(files.size()); |
3517 | SmallVector<SmallVector<NameAttrEntry, 0>, 0> nameAttrs(files.size()); |
3518 | |
3519 | parallelFor(0, files.size(), [&](size_t i) { |
3520 | // To keep memory usage low, we don't want to keep cached DWARFContext, so |
3521 | // avoid getDwarf() here. |
3522 | ObjFile<ELFT> *file = cast<ObjFile<ELFT>>(files[i]); |
3523 | DWARFContext dwarf(std::make_unique<LLDDwarfObj<ELFT>>(file)); |
3524 | auto &dobj = static_cast<const LLDDwarfObj<ELFT> &>(dwarf.getDWARFObj()); |
3525 | |
3526 | // If the are multiple compile units .debug_info (very rare ld -r --unique), |
3527 | // this only picks the last one. Other address ranges are lost. |
3528 | chunks[i].sec = dobj.getInfoSection(); |
3529 | chunks[i].compilationUnits = readCuList(dwarf); |
3530 | chunks[i].addressAreas = readAddressAreas(dwarf, sec: chunks[i].sec); |
3531 | nameAttrs[i] = readPubNamesAndTypes<ELFT>(dobj, chunks[i].compilationUnits); |
3532 | }); |
3533 | |
3534 | auto ret = std::make_unique<GdbIndexSection>(); |
3535 | ret->chunks = std::move(chunks); |
3536 | std::tie(args&: ret->symbols, args&: ret->size) = createSymbols(nameAttrs, chunks: ret->chunks); |
3537 | |
3538 | // Count the areas other than the constant pool. |
3539 | ret->size += sizeof(GdbIndexHeader) + ret->computeSymtabSize() * 8; |
3540 | for (GdbChunk &chunk : ret->chunks) |
3541 | ret->size += |
3542 | chunk.compilationUnits.size() * 16 + chunk.addressAreas.size() * 20; |
3543 | |
3544 | return ret; |
3545 | } |
3546 | |
3547 | void GdbIndexSection::writeTo(uint8_t *buf) { |
3548 | // Write the header. |
3549 | auto *hdr = reinterpret_cast<GdbIndexHeader *>(buf); |
3550 | uint8_t *start = buf; |
3551 | hdr->version = 7; |
3552 | buf += sizeof(*hdr); |
3553 | |
3554 | // Write the CU list. |
3555 | hdr->cuListOff = buf - start; |
3556 | for (GdbChunk &chunk : chunks) { |
3557 | for (CuEntry &cu : chunk.compilationUnits) { |
3558 | write64le(P: buf, V: chunk.sec->outSecOff + cu.cuOffset); |
3559 | write64le(P: buf + 8, V: cu.cuLength); |
3560 | buf += 16; |
3561 | } |
3562 | } |
3563 | |
3564 | // Write the address area. |
3565 | hdr->cuTypesOff = buf - start; |
3566 | hdr->addressAreaOff = buf - start; |
3567 | uint32_t cuOff = 0; |
3568 | for (GdbChunk &chunk : chunks) { |
3569 | for (AddressEntry &e : chunk.addressAreas) { |
3570 | // In the case of ICF there may be duplicate address range entries. |
3571 | const uint64_t baseAddr = e.section->repl->getVA(offset: 0); |
3572 | write64le(P: buf, V: baseAddr + e.lowAddress); |
3573 | write64le(P: buf + 8, V: baseAddr + e.highAddress); |
3574 | write32le(P: buf + 16, V: e.cuIndex + cuOff); |
3575 | buf += 20; |
3576 | } |
3577 | cuOff += chunk.compilationUnits.size(); |
3578 | } |
3579 | |
3580 | // Write the on-disk open-addressing hash table containing symbols. |
3581 | hdr->symtabOff = buf - start; |
3582 | size_t symtabSize = computeSymtabSize(); |
3583 | uint32_t mask = symtabSize - 1; |
3584 | |
3585 | for (GdbSymbol &sym : symbols) { |
3586 | uint32_t h = sym.name.hash(); |
3587 | uint32_t i = h & mask; |
3588 | uint32_t step = ((h * 17) & mask) | 1; |
3589 | |
3590 | while (read32le(P: buf + i * 8)) |
3591 | i = (i + step) & mask; |
3592 | |
3593 | write32le(P: buf + i * 8, V: sym.nameOff); |
3594 | write32le(P: buf + i * 8 + 4, V: sym.cuVectorOff); |
3595 | } |
3596 | |
3597 | buf += symtabSize * 8; |
3598 | |
3599 | // Write the string pool. |
3600 | hdr->constantPoolOff = buf - start; |
3601 | parallelForEach(R&: symbols, Fn: [&](GdbSymbol &sym) { |
3602 | memcpy(dest: buf + sym.nameOff, src: sym.name.data(), n: sym.name.size()); |
3603 | }); |
3604 | |
3605 | // Write the CU vectors. |
3606 | for (GdbSymbol &sym : symbols) { |
3607 | write32le(P: buf, V: sym.cuVector.size()); |
3608 | buf += 4; |
3609 | for (uint32_t val : sym.cuVector) { |
3610 | write32le(P: buf, V: val); |
3611 | buf += 4; |
3612 | } |
3613 | } |
3614 | } |
3615 | |
3616 | bool GdbIndexSection::isNeeded() const { return !chunks.empty(); } |
3617 | |
3618 | EhFrameHeader::() |
3619 | : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 4, ".eh_frame_hdr" ) {} |
3620 | |
3621 | void EhFrameHeader::(uint8_t *buf) { |
3622 | // Unlike most sections, the EhFrameHeader section is written while writing |
3623 | // another section, namely EhFrameSection, which calls the write() function |
3624 | // below from its writeTo() function. This is necessary because the contents |
3625 | // of EhFrameHeader depend on the relocated contents of EhFrameSection and we |
3626 | // don't know which order the sections will be written in. |
3627 | } |
3628 | |
3629 | // .eh_frame_hdr contains a binary search table of pointers to FDEs. |
3630 | // Each entry of the search table consists of two values, |
3631 | // the starting PC from where FDEs covers, and the FDE's address. |
3632 | // It is sorted by PC. |
3633 | void EhFrameHeader::() { |
3634 | uint8_t *buf = Out::bufferStart + getParent()->offset + outSecOff; |
3635 | using FdeData = EhFrameSection::FdeData; |
3636 | SmallVector<FdeData, 0> fdes = getPartition().ehFrame->getFdeData(); |
3637 | |
3638 | buf[0] = 1; |
3639 | buf[1] = DW_EH_PE_pcrel | DW_EH_PE_sdata4; |
3640 | buf[2] = DW_EH_PE_udata4; |
3641 | buf[3] = DW_EH_PE_datarel | DW_EH_PE_sdata4; |
3642 | write32(p: buf + 4, |
3643 | v: getPartition().ehFrame->getParent()->addr - this->getVA() - 4); |
3644 | write32(p: buf + 8, v: fdes.size()); |
3645 | buf += 12; |
3646 | |
3647 | for (FdeData &fde : fdes) { |
3648 | write32(p: buf, v: fde.pcRel); |
3649 | write32(p: buf + 4, v: fde.fdeVARel); |
3650 | buf += 8; |
3651 | } |
3652 | } |
3653 | |
3654 | size_t EhFrameHeader::() const { |
3655 | // .eh_frame_hdr has a 12 bytes header followed by an array of FDEs. |
3656 | return 12 + getPartition().ehFrame->numFdes * 8; |
3657 | } |
3658 | |
3659 | bool EhFrameHeader::() const { |
3660 | return isLive() && getPartition().ehFrame->isNeeded(); |
3661 | } |
3662 | |
3663 | VersionDefinitionSection::VersionDefinitionSection() |
3664 | : SyntheticSection(SHF_ALLOC, SHT_GNU_verdef, sizeof(uint32_t), |
3665 | ".gnu.version_d" ) {} |
3666 | |
3667 | StringRef VersionDefinitionSection::getFileDefName() { |
3668 | if (!getPartition().name.empty()) |
3669 | return getPartition().name; |
3670 | if (!config->soName.empty()) |
3671 | return config->soName; |
3672 | return config->outputFile; |
3673 | } |
3674 | |
3675 | void VersionDefinitionSection::finalizeContents() { |
3676 | fileDefNameOff = getPartition().dynStrTab->addString(s: getFileDefName()); |
3677 | for (const VersionDefinition &v : namedVersionDefs()) |
3678 | verDefNameOffs.push_back(Elt: getPartition().dynStrTab->addString(s: v.name)); |
3679 | |
3680 | if (OutputSection *sec = getPartition().dynStrTab->getParent()) |
3681 | getParent()->link = sec->sectionIndex; |
3682 | |
3683 | // sh_info should be set to the number of definitions. This fact is missed in |
3684 | // documentation, but confirmed by binutils community: |
3685 | // https://sourceware.org/ml/binutils/2014-11/msg00355.html |
3686 | getParent()->info = getVerDefNum(); |
3687 | } |
3688 | |
3689 | void VersionDefinitionSection::writeOne(uint8_t *buf, uint32_t index, |
3690 | StringRef name, size_t nameOff) { |
3691 | uint16_t flags = index == 1 ? VER_FLG_BASE : 0; |
3692 | |
3693 | // Write a verdef. |
3694 | write16(p: buf, v: 1); // vd_version |
3695 | write16(p: buf + 2, v: flags); // vd_flags |
3696 | write16(p: buf + 4, v: index); // vd_ndx |
3697 | write16(p: buf + 6, v: 1); // vd_cnt |
3698 | write32(p: buf + 8, v: hashSysV(SymbolName: name)); // vd_hash |
3699 | write32(p: buf + 12, v: 20); // vd_aux |
3700 | write32(p: buf + 16, v: 28); // vd_next |
3701 | |
3702 | // Write a veraux. |
3703 | write32(p: buf + 20, v: nameOff); // vda_name |
3704 | write32(p: buf + 24, v: 0); // vda_next |
3705 | } |
3706 | |
3707 | void VersionDefinitionSection::writeTo(uint8_t *buf) { |
3708 | writeOne(buf, index: 1, name: getFileDefName(), nameOff: fileDefNameOff); |
3709 | |
3710 | auto nameOffIt = verDefNameOffs.begin(); |
3711 | for (const VersionDefinition &v : namedVersionDefs()) { |
3712 | buf += EntrySize; |
3713 | writeOne(buf, index: v.id, name: v.name, nameOff: *nameOffIt++); |
3714 | } |
3715 | |
3716 | // Need to terminate the last version definition. |
3717 | write32(p: buf + 16, v: 0); // vd_next |
3718 | } |
3719 | |
3720 | size_t VersionDefinitionSection::getSize() const { |
3721 | return EntrySize * getVerDefNum(); |
3722 | } |
3723 | |
3724 | // .gnu.version is a table where each entry is 2 byte long. |
3725 | VersionTableSection::VersionTableSection() |
3726 | : SyntheticSection(SHF_ALLOC, SHT_GNU_versym, sizeof(uint16_t), |
3727 | ".gnu.version" ) { |
3728 | this->entsize = 2; |
3729 | } |
3730 | |
3731 | void VersionTableSection::finalizeContents() { |
3732 | // At the moment of june 2016 GNU docs does not mention that sh_link field |
3733 | // should be set, but Sun docs do. Also readelf relies on this field. |
3734 | getParent()->link = getPartition().dynSymTab->getParent()->sectionIndex; |
3735 | } |
3736 | |
3737 | size_t VersionTableSection::getSize() const { |
3738 | return (getPartition().dynSymTab->getSymbols().size() + 1) * 2; |
3739 | } |
3740 | |
3741 | void VersionTableSection::writeTo(uint8_t *buf) { |
3742 | buf += 2; |
3743 | for (const SymbolTableEntry &s : getPartition().dynSymTab->getSymbols()) { |
3744 | // For an unextracted lazy symbol (undefined weak), it must have been |
3745 | // converted to Undefined and have VER_NDX_GLOBAL version here. |
3746 | assert(!s.sym->isLazy()); |
3747 | write16(p: buf, v: s.sym->versionId); |
3748 | buf += 2; |
3749 | } |
3750 | } |
3751 | |
3752 | bool VersionTableSection::isNeeded() const { |
3753 | return isLive() && |
3754 | (getPartition().verDef || getPartition().verNeed->isNeeded()); |
3755 | } |
3756 | |
3757 | void elf::addVerneed(Symbol *ss) { |
3758 | auto &file = cast<SharedFile>(Val&: *ss->file); |
3759 | if (ss->versionId == VER_NDX_GLOBAL) |
3760 | return; |
3761 | |
3762 | if (file.vernauxs.empty()) |
3763 | file.vernauxs.resize(N: file.verdefs.size()); |
3764 | |
3765 | // Select a version identifier for the vernaux data structure, if we haven't |
3766 | // already allocated one. The verdef identifiers cover the range |
3767 | // [1..getVerDefNum()]; this causes the vernaux identifiers to start from |
3768 | // getVerDefNum()+1. |
3769 | if (file.vernauxs[ss->versionId] == 0) |
3770 | file.vernauxs[ss->versionId] = ++SharedFile::vernauxNum + getVerDefNum(); |
3771 | |
3772 | ss->versionId = file.vernauxs[ss->versionId]; |
3773 | } |
3774 | |
3775 | template <class ELFT> |
3776 | VersionNeedSection<ELFT>::VersionNeedSection() |
3777 | : SyntheticSection(SHF_ALLOC, SHT_GNU_verneed, sizeof(uint32_t), |
3778 | ".gnu.version_r" ) {} |
3779 | |
3780 | template <class ELFT> void VersionNeedSection<ELFT>::finalizeContents() { |
3781 | for (SharedFile *f : ctx.sharedFiles) { |
3782 | if (f->vernauxs.empty()) |
3783 | continue; |
3784 | verneeds.emplace_back(); |
3785 | Verneed &vn = verneeds.back(); |
3786 | vn.nameStrTab = getPartition().dynStrTab->addString(f->soName); |
3787 | bool isLibc = config->relrGlibc && f->soName.starts_with(Prefix: "libc.so." ); |
3788 | bool isGlibc2 = false; |
3789 | for (unsigned i = 0; i != f->vernauxs.size(); ++i) { |
3790 | if (f->vernauxs[i] == 0) |
3791 | continue; |
3792 | auto *verdef = |
3793 | reinterpret_cast<const typename ELFT::Verdef *>(f->verdefs[i]); |
3794 | StringRef ver(f->getStringTable().data() + verdef->getAux()->vda_name); |
3795 | if (isLibc && ver.starts_with(Prefix: "GLIBC_2." )) |
3796 | isGlibc2 = true; |
3797 | vn.vernauxs.push_back({verdef->vd_hash, f->vernauxs[i], |
3798 | getPartition().dynStrTab->addString(ver)}); |
3799 | } |
3800 | if (isGlibc2) { |
3801 | const char *ver = "GLIBC_ABI_DT_RELR" ; |
3802 | vn.vernauxs.push_back({hashSysV(SymbolName: ver), |
3803 | ++SharedFile::vernauxNum + getVerDefNum(), |
3804 | getPartition().dynStrTab->addString(ver)}); |
3805 | } |
3806 | } |
3807 | |
3808 | if (OutputSection *sec = getPartition().dynStrTab->getParent()) |
3809 | getParent()->link = sec->sectionIndex; |
3810 | getParent()->info = verneeds.size(); |
3811 | } |
3812 | |
3813 | template <class ELFT> void VersionNeedSection<ELFT>::writeTo(uint8_t *buf) { |
3814 | // The Elf_Verneeds need to appear first, followed by the Elf_Vernauxs. |
3815 | auto *verneed = reinterpret_cast<Elf_Verneed *>(buf); |
3816 | auto *vernaux = reinterpret_cast<Elf_Vernaux *>(verneed + verneeds.size()); |
3817 | |
3818 | for (auto &vn : verneeds) { |
3819 | // Create an Elf_Verneed for this DSO. |
3820 | verneed->vn_version = 1; |
3821 | verneed->vn_cnt = vn.vernauxs.size(); |
3822 | verneed->vn_file = vn.nameStrTab; |
3823 | verneed->vn_aux = |
3824 | reinterpret_cast<char *>(vernaux) - reinterpret_cast<char *>(verneed); |
3825 | verneed->vn_next = sizeof(Elf_Verneed); |
3826 | ++verneed; |
3827 | |
3828 | // Create the Elf_Vernauxs for this Elf_Verneed. |
3829 | for (auto &vna : vn.vernauxs) { |
3830 | vernaux->vna_hash = vna.hash; |
3831 | vernaux->vna_flags = 0; |
3832 | vernaux->vna_other = vna.verneedIndex; |
3833 | vernaux->vna_name = vna.nameStrTab; |
3834 | vernaux->vna_next = sizeof(Elf_Vernaux); |
3835 | ++vernaux; |
3836 | } |
3837 | |
3838 | vernaux[-1].vna_next = 0; |
3839 | } |
3840 | verneed[-1].vn_next = 0; |
3841 | } |
3842 | |
3843 | template <class ELFT> size_t VersionNeedSection<ELFT>::getSize() const { |
3844 | return verneeds.size() * sizeof(Elf_Verneed) + |
3845 | SharedFile::vernauxNum * sizeof(Elf_Vernaux); |
3846 | } |
3847 | |
3848 | template <class ELFT> bool VersionNeedSection<ELFT>::isNeeded() const { |
3849 | return isLive() && SharedFile::vernauxNum != 0; |
3850 | } |
3851 | |
3852 | void MergeSyntheticSection::addSection(MergeInputSection *ms) { |
3853 | ms->parent = this; |
3854 | sections.push_back(Elt: ms); |
3855 | assert(addralign == ms->addralign || !(ms->flags & SHF_STRINGS)); |
3856 | addralign = std::max(a: addralign, b: ms->addralign); |
3857 | } |
3858 | |
3859 | MergeTailSection::MergeTailSection(StringRef name, uint32_t type, |
3860 | uint64_t flags, uint32_t alignment) |
3861 | : MergeSyntheticSection(name, type, flags, alignment), |
3862 | builder(StringTableBuilder::RAW, llvm::Align(alignment)) {} |
3863 | |
3864 | size_t MergeTailSection::getSize() const { return builder.getSize(); } |
3865 | |
3866 | void MergeTailSection::writeTo(uint8_t *buf) { builder.write(Buf: buf); } |
3867 | |
3868 | void MergeTailSection::finalizeContents() { |
3869 | // Add all string pieces to the string table builder to create section |
3870 | // contents. |
3871 | for (MergeInputSection *sec : sections) |
3872 | for (size_t i = 0, e = sec->pieces.size(); i != e; ++i) |
3873 | if (sec->pieces[i].live) |
3874 | builder.add(S: sec->getData(i)); |
3875 | |
3876 | // Fix the string table content. After this, the contents will never change. |
3877 | builder.finalize(); |
3878 | |
3879 | // finalize() fixed tail-optimized strings, so we can now get |
3880 | // offsets of strings. Get an offset for each string and save it |
3881 | // to a corresponding SectionPiece for easy access. |
3882 | for (MergeInputSection *sec : sections) |
3883 | for (size_t i = 0, e = sec->pieces.size(); i != e; ++i) |
3884 | if (sec->pieces[i].live) |
3885 | sec->pieces[i].outputOff = builder.getOffset(S: sec->getData(i)); |
3886 | } |
3887 | |
3888 | void MergeNoTailSection::writeTo(uint8_t *buf) { |
3889 | parallelFor(Begin: 0, End: numShards, |
3890 | Fn: [&](size_t i) { shards[i].write(Buf: buf + shardOffsets[i]); }); |
3891 | } |
3892 | |
3893 | // This function is very hot (i.e. it can take several seconds to finish) |
3894 | // because sometimes the number of inputs is in an order of magnitude of |
3895 | // millions. So, we use multi-threading. |
3896 | // |
3897 | // For any strings S and T, we know S is not mergeable with T if S's hash |
3898 | // value is different from T's. If that's the case, we can safely put S and |
3899 | // T into different string builders without worrying about merge misses. |
3900 | // We do it in parallel. |
3901 | void MergeNoTailSection::finalizeContents() { |
3902 | // Initializes string table builders. |
3903 | for (size_t i = 0; i < numShards; ++i) |
3904 | shards.emplace_back(Args: StringTableBuilder::RAW, Args: llvm::Align(addralign)); |
3905 | |
3906 | // Concurrency level. Must be a power of 2 to avoid expensive modulo |
3907 | // operations in the following tight loop. |
3908 | const size_t concurrency = |
3909 | llvm::bit_floor(Value: std::min<size_t>(a: config->threadCount, b: numShards)); |
3910 | |
3911 | // Add section pieces to the builders. |
3912 | parallelFor(Begin: 0, End: concurrency, Fn: [&](size_t threadId) { |
3913 | for (MergeInputSection *sec : sections) { |
3914 | for (size_t i = 0, e = sec->pieces.size(); i != e; ++i) { |
3915 | if (!sec->pieces[i].live) |
3916 | continue; |
3917 | size_t shardId = getShardId(hash: sec->pieces[i].hash); |
3918 | if ((shardId & (concurrency - 1)) == threadId) |
3919 | sec->pieces[i].outputOff = shards[shardId].add(S: sec->getData(i)); |
3920 | } |
3921 | } |
3922 | }); |
3923 | |
3924 | // Compute an in-section offset for each shard. |
3925 | size_t off = 0; |
3926 | for (size_t i = 0; i < numShards; ++i) { |
3927 | shards[i].finalizeInOrder(); |
3928 | if (shards[i].getSize() > 0) |
3929 | off = alignToPowerOf2(Value: off, Align: addralign); |
3930 | shardOffsets[i] = off; |
3931 | off += shards[i].getSize(); |
3932 | } |
3933 | size = off; |
3934 | |
3935 | // So far, section pieces have offsets from beginning of shards, but |
3936 | // we want offsets from beginning of the whole section. Fix them. |
3937 | parallelForEach(R&: sections, Fn: [&](MergeInputSection *sec) { |
3938 | for (size_t i = 0, e = sec->pieces.size(); i != e; ++i) |
3939 | if (sec->pieces[i].live) |
3940 | sec->pieces[i].outputOff += |
3941 | shardOffsets[getShardId(hash: sec->pieces[i].hash)]; |
3942 | }); |
3943 | } |
3944 | |
3945 | template <class ELFT> void elf::splitSections() { |
3946 | llvm::TimeTraceScope timeScope("Split sections" ); |
3947 | // splitIntoPieces needs to be called on each MergeInputSection |
3948 | // before calling finalizeContents(). |
3949 | parallelForEach(ctx.objectFiles, [](ELFFileBase *file) { |
3950 | for (InputSectionBase *sec : file->getSections()) { |
3951 | if (!sec) |
3952 | continue; |
3953 | if (auto *s = dyn_cast<MergeInputSection>(Val: sec)) |
3954 | s->splitIntoPieces(); |
3955 | else if (auto *eh = dyn_cast<EhInputSection>(Val: sec)) |
3956 | eh->split<ELFT>(); |
3957 | } |
3958 | }); |
3959 | } |
3960 | |
3961 | void elf::combineEhSections() { |
3962 | llvm::TimeTraceScope timeScope("Combine EH sections" ); |
3963 | for (EhInputSection *sec : ctx.ehInputSections) { |
3964 | EhFrameSection &eh = *sec->getPartition().ehFrame; |
3965 | sec->parent = &eh; |
3966 | eh.addralign = std::max(a: eh.addralign, b: sec->addralign); |
3967 | eh.sections.push_back(Elt: sec); |
3968 | llvm::append_range(C&: eh.dependentSections, R&: sec->dependentSections); |
3969 | } |
3970 | |
3971 | if (!mainPart->armExidx) |
3972 | return; |
3973 | llvm::erase_if(C&: ctx.inputSections, P: [](InputSectionBase *s) { |
3974 | // Ignore dead sections and the partition end marker (.part.end), |
3975 | // whose partition number is out of bounds. |
3976 | if (!s->isLive() || s->partition == 255) |
3977 | return false; |
3978 | Partition &part = s->getPartition(); |
3979 | return s->kind() == SectionBase::Regular && part.armExidx && |
3980 | part.armExidx->addSection(isec: cast<InputSection>(Val: s)); |
3981 | }); |
3982 | } |
3983 | |
3984 | MipsRldMapSection::MipsRldMapSection() |
3985 | : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, config->wordsize, |
3986 | ".rld_map" ) {} |
3987 | |
3988 | ARMExidxSyntheticSection::ARMExidxSyntheticSection() |
3989 | : SyntheticSection(SHF_ALLOC | SHF_LINK_ORDER, SHT_ARM_EXIDX, |
3990 | config->wordsize, ".ARM.exidx" ) {} |
3991 | |
3992 | static InputSection *findExidxSection(InputSection *isec) { |
3993 | for (InputSection *d : isec->dependentSections) |
3994 | if (d->type == SHT_ARM_EXIDX && d->isLive()) |
3995 | return d; |
3996 | return nullptr; |
3997 | } |
3998 | |
3999 | static bool isValidExidxSectionDep(InputSection *isec) { |
4000 | return (isec->flags & SHF_ALLOC) && (isec->flags & SHF_EXECINSTR) && |
4001 | isec->getSize() > 0; |
4002 | } |
4003 | |
4004 | bool ARMExidxSyntheticSection::addSection(InputSection *isec) { |
4005 | if (isec->type == SHT_ARM_EXIDX) { |
4006 | if (InputSection *dep = isec->getLinkOrderDep()) |
4007 | if (isValidExidxSectionDep(isec: dep)) { |
4008 | exidxSections.push_back(Elt: isec); |
4009 | // Every exidxSection is 8 bytes, we need an estimate of |
4010 | // size before assignAddresses can be called. Final size |
4011 | // will only be known after finalize is called. |
4012 | size += 8; |
4013 | } |
4014 | return true; |
4015 | } |
4016 | |
4017 | if (isValidExidxSectionDep(isec)) { |
4018 | executableSections.push_back(Elt: isec); |
4019 | return false; |
4020 | } |
4021 | |
4022 | // FIXME: we do not output a relocation section when --emit-relocs is used |
4023 | // as we do not have relocation sections for linker generated table entries |
4024 | // and we would have to erase at a late stage relocations from merged entries. |
4025 | // Given that exception tables are already position independent and a binary |
4026 | // analyzer could derive the relocations we choose to erase the relocations. |
4027 | if (config->emitRelocs && isec->type == SHT_REL) |
4028 | if (InputSectionBase *ex = isec->getRelocatedSection()) |
4029 | if (isa<InputSection>(Val: ex) && ex->type == SHT_ARM_EXIDX) |
4030 | return true; |
4031 | |
4032 | return false; |
4033 | } |
4034 | |
4035 | // References to .ARM.Extab Sections have bit 31 clear and are not the |
4036 | // special EXIDX_CANTUNWIND bit-pattern. |
4037 | static bool isExtabRef(uint32_t unwind) { |
4038 | return (unwind & 0x80000000) == 0 && unwind != 0x1; |
4039 | } |
4040 | |
4041 | // Return true if the .ARM.exidx section Cur can be merged into the .ARM.exidx |
4042 | // section Prev, where Cur follows Prev in the table. This can be done if the |
4043 | // unwinding instructions in Cur are identical to Prev. Linker generated |
4044 | // EXIDX_CANTUNWIND entries are represented by nullptr as they do not have an |
4045 | // InputSection. |
4046 | static bool isDuplicateArmExidxSec(InputSection *prev, InputSection *cur) { |
4047 | // Get the last table Entry from the previous .ARM.exidx section. If Prev is |
4048 | // nullptr then it will be a synthesized EXIDX_CANTUNWIND entry. |
4049 | uint32_t prevUnwind = 1; |
4050 | if (prev) |
4051 | prevUnwind = read32(p: prev->content().data() + prev->content().size() - 4); |
4052 | if (isExtabRef(unwind: prevUnwind)) |
4053 | return false; |
4054 | |
4055 | // We consider the unwind instructions of an .ARM.exidx table entry |
4056 | // a duplicate if the previous unwind instructions if: |
4057 | // - Both are the special EXIDX_CANTUNWIND. |
4058 | // - Both are the same inline unwind instructions. |
4059 | // We do not attempt to follow and check links into .ARM.extab tables as |
4060 | // consecutive identical entries are rare and the effort to check that they |
4061 | // are identical is high. |
4062 | |
4063 | // If Cur is nullptr then this is synthesized EXIDX_CANTUNWIND entry. |
4064 | if (cur == nullptr) |
4065 | return prevUnwind == 1; |
4066 | |
4067 | for (uint32_t offset = 4; offset < (uint32_t)cur->content().size(); offset +=8) { |
4068 | uint32_t curUnwind = read32(p: cur->content().data() + offset); |
4069 | if (isExtabRef(unwind: curUnwind) || curUnwind != prevUnwind) |
4070 | return false; |
4071 | } |
4072 | // All table entries in this .ARM.exidx Section can be merged into the |
4073 | // previous Section. |
4074 | return true; |
4075 | } |
4076 | |
4077 | // The .ARM.exidx table must be sorted in ascending order of the address of the |
4078 | // functions the table describes. std::optionally duplicate adjacent table |
4079 | // entries can be removed. At the end of the function the executableSections |
4080 | // must be sorted in ascending order of address, Sentinel is set to the |
4081 | // InputSection with the highest address and any InputSections that have |
4082 | // mergeable .ARM.exidx table entries are removed from it. |
4083 | void ARMExidxSyntheticSection::finalizeContents() { |
4084 | // Ensure that any fixed-point iterations after the first see the original set |
4085 | // of sections. |
4086 | if (!originalExecutableSections.empty()) |
4087 | executableSections = originalExecutableSections; |
4088 | else if (config->enableNonContiguousRegions) |
4089 | originalExecutableSections = executableSections; |
4090 | |
4091 | // The executableSections and exidxSections that we use to derive the final |
4092 | // contents of this SyntheticSection are populated before |
4093 | // processSectionCommands() and ICF. A /DISCARD/ entry in SECTIONS command or |
4094 | // ICF may remove executable InputSections and their dependent .ARM.exidx |
4095 | // section that we recorded earlier. |
4096 | auto isDiscarded = [](const InputSection *isec) { return !isec->isLive(); }; |
4097 | llvm::erase_if(C&: exidxSections, P: isDiscarded); |
4098 | // We need to remove discarded InputSections and InputSections without |
4099 | // .ARM.exidx sections that if we generated the .ARM.exidx it would be out |
4100 | // of range. |
4101 | auto isDiscardedOrOutOfRange = [this](InputSection *isec) { |
4102 | if (!isec->isLive()) |
4103 | return true; |
4104 | if (findExidxSection(isec)) |
4105 | return false; |
4106 | int64_t off = static_cast<int64_t>(isec->getVA() - getVA()); |
4107 | return off != llvm::SignExtend64(X: off, B: 31); |
4108 | }; |
4109 | llvm::erase_if(C&: executableSections, P: isDiscardedOrOutOfRange); |
4110 | |
4111 | // Sort the executable sections that may or may not have associated |
4112 | // .ARM.exidx sections by order of ascending address. This requires the |
4113 | // relative positions of InputSections and OutputSections to be known. |
4114 | auto compareByFilePosition = [](const InputSection *a, |
4115 | const InputSection *b) { |
4116 | OutputSection *aOut = a->getParent(); |
4117 | OutputSection *bOut = b->getParent(); |
4118 | |
4119 | if (aOut != bOut) |
4120 | return aOut->addr < bOut->addr; |
4121 | return a->outSecOff < b->outSecOff; |
4122 | }; |
4123 | llvm::stable_sort(Range&: executableSections, C: compareByFilePosition); |
4124 | sentinel = executableSections.back(); |
4125 | // std::optionally merge adjacent duplicate entries. |
4126 | if (config->mergeArmExidx) { |
4127 | SmallVector<InputSection *, 0> selectedSections; |
4128 | selectedSections.reserve(N: executableSections.size()); |
4129 | selectedSections.push_back(Elt: executableSections[0]); |
4130 | size_t prev = 0; |
4131 | for (size_t i = 1; i < executableSections.size(); ++i) { |
4132 | InputSection *ex1 = findExidxSection(isec: executableSections[prev]); |
4133 | InputSection *ex2 = findExidxSection(isec: executableSections[i]); |
4134 | if (!isDuplicateArmExidxSec(prev: ex1, cur: ex2)) { |
4135 | selectedSections.push_back(Elt: executableSections[i]); |
4136 | prev = i; |
4137 | } |
4138 | } |
4139 | executableSections = std::move(selectedSections); |
4140 | } |
4141 | // offset is within the SyntheticSection. |
4142 | size_t offset = 0; |
4143 | size = 0; |
4144 | for (InputSection *isec : executableSections) { |
4145 | if (InputSection *d = findExidxSection(isec)) { |
4146 | d->outSecOff = offset; |
4147 | d->parent = getParent(); |
4148 | offset += d->getSize(); |
4149 | } else { |
4150 | offset += 8; |
4151 | } |
4152 | } |
4153 | // Size includes Sentinel. |
4154 | size = offset + 8; |
4155 | } |
4156 | |
4157 | InputSection *ARMExidxSyntheticSection::getLinkOrderDep() const { |
4158 | return executableSections.front(); |
4159 | } |
4160 | |
4161 | // To write the .ARM.exidx table from the ExecutableSections we have three cases |
4162 | // 1.) The InputSection has a .ARM.exidx InputSection in its dependent sections. |
4163 | // We write the .ARM.exidx section contents and apply its relocations. |
4164 | // 2.) The InputSection does not have a dependent .ARM.exidx InputSection. We |
4165 | // must write the contents of an EXIDX_CANTUNWIND directly. We use the |
4166 | // start of the InputSection as the purpose of the linker generated |
4167 | // section is to terminate the address range of the previous entry. |
4168 | // 3.) A trailing EXIDX_CANTUNWIND sentinel section is required at the end of |
4169 | // the table to terminate the address range of the final entry. |
4170 | void ARMExidxSyntheticSection::writeTo(uint8_t *buf) { |
4171 | |
4172 | // A linker generated CANTUNWIND entry is made up of two words: |
4173 | // 0x0 with R_ARM_PREL31 relocation to target. |
4174 | // 0x1 with EXIDX_CANTUNWIND. |
4175 | uint64_t offset = 0; |
4176 | for (InputSection *isec : executableSections) { |
4177 | assert(isec->getParent() != nullptr); |
4178 | if (InputSection *d = findExidxSection(isec)) { |
4179 | for (int dataOffset = 0; dataOffset != (int)d->content().size(); |
4180 | dataOffset += 4) |
4181 | write32(p: buf + offset + dataOffset, |
4182 | v: read32(p: d->content().data() + dataOffset)); |
4183 | // Recalculate outSecOff as finalizeAddressDependentContent() |
4184 | // may have altered syntheticSection outSecOff. |
4185 | d->outSecOff = offset + outSecOff; |
4186 | target->relocateAlloc(sec&: *d, buf: buf + offset); |
4187 | offset += d->getSize(); |
4188 | } else { |
4189 | // A Linker generated CANTUNWIND section. |
4190 | write32(p: buf + offset + 0, v: 0x0); |
4191 | write32(p: buf + offset + 4, v: 0x1); |
4192 | uint64_t s = isec->getVA(); |
4193 | uint64_t p = getVA() + offset; |
4194 | target->relocateNoSym(loc: buf + offset, type: R_ARM_PREL31, val: s - p); |
4195 | offset += 8; |
4196 | } |
4197 | } |
4198 | // Write Sentinel CANTUNWIND entry. |
4199 | write32(p: buf + offset + 0, v: 0x0); |
4200 | write32(p: buf + offset + 4, v: 0x1); |
4201 | uint64_t s = sentinel->getVA(offset: sentinel->getSize()); |
4202 | uint64_t p = getVA() + offset; |
4203 | target->relocateNoSym(loc: buf + offset, type: R_ARM_PREL31, val: s - p); |
4204 | assert(size == offset + 8); |
4205 | } |
4206 | |
4207 | bool ARMExidxSyntheticSection::isNeeded() const { |
4208 | return llvm::any_of(Range: exidxSections, |
4209 | P: [](InputSection *isec) { return isec->isLive(); }); |
4210 | } |
4211 | |
4212 | ThunkSection::ThunkSection(OutputSection *os, uint64_t off) |
4213 | : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, |
4214 | config->emachine == EM_PPC64 ? 16 : 4, ".text.thunk" ) { |
4215 | this->parent = os; |
4216 | this->outSecOff = off; |
4217 | } |
4218 | |
4219 | size_t ThunkSection::getSize() const { |
4220 | if (roundUpSizeForErrata) |
4221 | return alignTo(Value: size, Align: 4096); |
4222 | return size; |
4223 | } |
4224 | |
4225 | void ThunkSection::addThunk(Thunk *t) { |
4226 | thunks.push_back(Elt: t); |
4227 | t->addSymbols(isec&: *this); |
4228 | } |
4229 | |
4230 | void ThunkSection::writeTo(uint8_t *buf) { |
4231 | for (Thunk *t : thunks) |
4232 | t->writeTo(buf: buf + t->offset); |
4233 | } |
4234 | |
4235 | InputSection *ThunkSection::getTargetInputSection() const { |
4236 | if (thunks.empty()) |
4237 | return nullptr; |
4238 | const Thunk *t = thunks.front(); |
4239 | return t->getTargetInputSection(); |
4240 | } |
4241 | |
4242 | bool ThunkSection::assignOffsets() { |
4243 | uint64_t off = 0; |
4244 | for (Thunk *t : thunks) { |
4245 | off = alignToPowerOf2(Value: off, Align: t->alignment); |
4246 | t->setOffset(off); |
4247 | uint32_t size = t->size(); |
4248 | t->getThunkTargetSym()->size = size; |
4249 | off += size; |
4250 | } |
4251 | bool changed = off != size; |
4252 | size = off; |
4253 | return changed; |
4254 | } |
4255 | |
4256 | PPC32Got2Section::PPC32Got2Section() |
4257 | : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, 4, ".got2" ) {} |
4258 | |
4259 | bool PPC32Got2Section::isNeeded() const { |
4260 | // See the comment below. This is not needed if there is no other |
4261 | // InputSection. |
4262 | for (SectionCommand *cmd : getParent()->commands) |
4263 | if (auto *isd = dyn_cast<InputSectionDescription>(Val: cmd)) |
4264 | for (InputSection *isec : isd->sections) |
4265 | if (isec != this) |
4266 | return true; |
4267 | return false; |
4268 | } |
4269 | |
4270 | void PPC32Got2Section::finalizeContents() { |
4271 | // PPC32 may create multiple GOT sections for -fPIC/-fPIE, one per file in |
4272 | // .got2 . This function computes outSecOff of each .got2 to be used in |
4273 | // PPC32PltCallStub::writeTo(). The purpose of this empty synthetic section is |
4274 | // to collect input sections named ".got2". |
4275 | for (SectionCommand *cmd : getParent()->commands) |
4276 | if (auto *isd = dyn_cast<InputSectionDescription>(Val: cmd)) { |
4277 | for (InputSection *isec : isd->sections) { |
4278 | // isec->file may be nullptr for MergeSyntheticSection. |
4279 | if (isec != this && isec->file) |
4280 | isec->file->ppc32Got2 = isec; |
4281 | } |
4282 | } |
4283 | } |
4284 | |
4285 | // If linking position-dependent code then the table will store the addresses |
4286 | // directly in the binary so the section has type SHT_PROGBITS. If linking |
4287 | // position-independent code the section has type SHT_NOBITS since it will be |
4288 | // allocated and filled in by the dynamic linker. |
4289 | PPC64LongBranchTargetSection::PPC64LongBranchTargetSection() |
4290 | : SyntheticSection(SHF_ALLOC | SHF_WRITE, |
4291 | config->isPic ? SHT_NOBITS : SHT_PROGBITS, 8, |
4292 | ".branch_lt" ) {} |
4293 | |
4294 | uint64_t PPC64LongBranchTargetSection::getEntryVA(const Symbol *sym, |
4295 | int64_t addend) { |
4296 | return getVA() + entry_index.find(Val: {sym, addend})->second * 8; |
4297 | } |
4298 | |
4299 | std::optional<uint32_t> |
4300 | PPC64LongBranchTargetSection::addEntry(const Symbol *sym, int64_t addend) { |
4301 | auto res = |
4302 | entry_index.try_emplace(Key: std::make_pair(x&: sym, y&: addend), Args: entries.size()); |
4303 | if (!res.second) |
4304 | return std::nullopt; |
4305 | entries.emplace_back(Args&: sym, Args&: addend); |
4306 | return res.first->second; |
4307 | } |
4308 | |
4309 | size_t PPC64LongBranchTargetSection::getSize() const { |
4310 | return entries.size() * 8; |
4311 | } |
4312 | |
4313 | void PPC64LongBranchTargetSection::writeTo(uint8_t *buf) { |
4314 | // If linking non-pic we have the final addresses of the targets and they get |
4315 | // written to the table directly. For pic the dynamic linker will allocate |
4316 | // the section and fill it. |
4317 | if (config->isPic) |
4318 | return; |
4319 | |
4320 | for (auto entry : entries) { |
4321 | const Symbol *sym = entry.first; |
4322 | int64_t addend = entry.second; |
4323 | assert(sym->getVA()); |
4324 | // Need calls to branch to the local entry-point since a long-branch |
4325 | // must be a local-call. |
4326 | write64(p: buf, v: sym->getVA(addend) + |
4327 | getPPC64GlobalEntryToLocalEntryOffset(stOther: sym->stOther)); |
4328 | buf += 8; |
4329 | } |
4330 | } |
4331 | |
4332 | bool PPC64LongBranchTargetSection::isNeeded() const { |
4333 | // `removeUnusedSyntheticSections()` is called before thunk allocation which |
4334 | // is too early to determine if this section will be empty or not. We need |
4335 | // Finalized to keep the section alive until after thunk creation. Finalized |
4336 | // only gets set to true once `finalizeSections()` is called after thunk |
4337 | // creation. Because of this, if we don't create any long-branch thunks we end |
4338 | // up with an empty .branch_lt section in the binary. |
4339 | return !finalized || !entries.empty(); |
4340 | } |
4341 | |
4342 | static uint8_t getAbiVersion() { |
4343 | // MIPS non-PIC executable gets ABI version 1. |
4344 | if (config->emachine == EM_MIPS) { |
4345 | if (!config->isPic && !config->relocatable && |
4346 | (config->eflags & (EF_MIPS_PIC | EF_MIPS_CPIC)) == EF_MIPS_CPIC) |
4347 | return 1; |
4348 | return 0; |
4349 | } |
4350 | |
4351 | if (config->emachine == EM_AMDGPU && !ctx.objectFiles.empty()) { |
4352 | uint8_t ver = ctx.objectFiles[0]->abiVersion; |
4353 | for (InputFile *file : ArrayRef(ctx.objectFiles).slice(N: 1)) |
4354 | if (file->abiVersion != ver) |
4355 | error(msg: "incompatible ABI version: " + toString(f: file)); |
4356 | return ver; |
4357 | } |
4358 | |
4359 | return 0; |
4360 | } |
4361 | |
4362 | template <typename ELFT> void elf::writeEhdr(uint8_t *buf, Partition &part) { |
4363 | memcpy(dest: buf, src: "\177ELF" , n: 4); |
4364 | |
4365 | auto *eHdr = reinterpret_cast<typename ELFT::Ehdr *>(buf); |
4366 | eHdr->e_ident[EI_CLASS] = ELFT::Is64Bits ? ELFCLASS64 : ELFCLASS32; |
4367 | eHdr->e_ident[EI_DATA] = |
4368 | ELFT::Endianness == endianness::little ? ELFDATA2LSB : ELFDATA2MSB; |
4369 | eHdr->e_ident[EI_VERSION] = EV_CURRENT; |
4370 | eHdr->e_ident[EI_OSABI] = config->osabi; |
4371 | eHdr->e_ident[EI_ABIVERSION] = getAbiVersion(); |
4372 | eHdr->e_machine = config->emachine; |
4373 | eHdr->e_version = EV_CURRENT; |
4374 | eHdr->e_flags = config->eflags; |
4375 | eHdr->e_ehsize = sizeof(typename ELFT::Ehdr); |
4376 | eHdr->e_phnum = part.phdrs.size(); |
4377 | eHdr->e_shentsize = sizeof(typename ELFT::Shdr); |
4378 | |
4379 | if (!config->relocatable) { |
4380 | eHdr->e_phoff = sizeof(typename ELFT::Ehdr); |
4381 | eHdr->e_phentsize = sizeof(typename ELFT::Phdr); |
4382 | } |
4383 | } |
4384 | |
4385 | template <typename ELFT> void elf::writePhdrs(uint8_t *buf, Partition &part) { |
4386 | // Write the program header table. |
4387 | auto *hBuf = reinterpret_cast<typename ELFT::Phdr *>(buf); |
4388 | for (PhdrEntry *p : part.phdrs) { |
4389 | hBuf->p_type = p->p_type; |
4390 | hBuf->p_flags = p->p_flags; |
4391 | hBuf->p_offset = p->p_offset; |
4392 | hBuf->p_vaddr = p->p_vaddr; |
4393 | hBuf->p_paddr = p->p_paddr; |
4394 | hBuf->p_filesz = p->p_filesz; |
4395 | hBuf->p_memsz = p->p_memsz; |
4396 | hBuf->p_align = p->p_align; |
4397 | ++hBuf; |
4398 | } |
4399 | } |
4400 | |
4401 | template <typename ELFT> |
4402 | PartitionElfHeaderSection<ELFT>::() |
4403 | : SyntheticSection(SHF_ALLOC, SHT_LLVM_PART_EHDR, 1, "" ) {} |
4404 | |
4405 | template <typename ELFT> |
4406 | size_t PartitionElfHeaderSection<ELFT>::() const { |
4407 | return sizeof(typename ELFT::Ehdr); |
4408 | } |
4409 | |
4410 | template <typename ELFT> |
4411 | void PartitionElfHeaderSection<ELFT>::(uint8_t *buf) { |
4412 | writeEhdr<ELFT>(buf, getPartition()); |
4413 | |
4414 | // Loadable partitions are always ET_DYN. |
4415 | auto *eHdr = reinterpret_cast<typename ELFT::Ehdr *>(buf); |
4416 | eHdr->e_type = ET_DYN; |
4417 | } |
4418 | |
4419 | template <typename ELFT> |
4420 | PartitionProgramHeadersSection<ELFT>::() |
4421 | : SyntheticSection(SHF_ALLOC, SHT_LLVM_PART_PHDR, 1, ".phdrs" ) {} |
4422 | |
4423 | template <typename ELFT> |
4424 | size_t PartitionProgramHeadersSection<ELFT>::() const { |
4425 | return sizeof(typename ELFT::Phdr) * getPartition().phdrs.size(); |
4426 | } |
4427 | |
4428 | template <typename ELFT> |
4429 | void PartitionProgramHeadersSection<ELFT>::(uint8_t *buf) { |
4430 | writePhdrs<ELFT>(buf, getPartition()); |
4431 | } |
4432 | |
4433 | PartitionIndexSection::PartitionIndexSection() |
4434 | : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 4, ".rodata" ) {} |
4435 | |
4436 | size_t PartitionIndexSection::getSize() const { |
4437 | return 12 * (partitions.size() - 1); |
4438 | } |
4439 | |
4440 | void PartitionIndexSection::finalizeContents() { |
4441 | for (size_t i = 1; i != partitions.size(); ++i) |
4442 | partitions[i].nameStrTab = mainPart->dynStrTab->addString(s: partitions[i].name); |
4443 | } |
4444 | |
4445 | void PartitionIndexSection::writeTo(uint8_t *buf) { |
4446 | uint64_t va = getVA(); |
4447 | for (size_t i = 1; i != partitions.size(); ++i) { |
4448 | write32(p: buf, v: mainPart->dynStrTab->getVA() + partitions[i].nameStrTab - va); |
4449 | write32(p: buf + 4, v: partitions[i].elfHeader->getVA() - (va + 4)); |
4450 | |
4451 | SyntheticSection *next = i == partitions.size() - 1 |
4452 | ? in.partEnd.get() |
4453 | : partitions[i + 1].elfHeader.get(); |
4454 | write32(p: buf + 8, v: next->getVA() - partitions[i].elfHeader->getVA()); |
4455 | |
4456 | va += 12; |
4457 | buf += 12; |
4458 | } |
4459 | } |
4460 | |
4461 | void InStruct::reset() { |
4462 | attributes.reset(); |
4463 | riscvAttributes.reset(); |
4464 | bss.reset(); |
4465 | bssRelRo.reset(); |
4466 | got.reset(); |
4467 | gotPlt.reset(); |
4468 | igotPlt.reset(); |
4469 | relroPadding.reset(); |
4470 | armCmseSGSection.reset(); |
4471 | ppc64LongBranchTarget.reset(); |
4472 | mipsAbiFlags.reset(); |
4473 | mipsGot.reset(); |
4474 | mipsOptions.reset(); |
4475 | mipsReginfo.reset(); |
4476 | mipsRldMap.reset(); |
4477 | partEnd.reset(); |
4478 | partIndex.reset(); |
4479 | plt.reset(); |
4480 | iplt.reset(); |
4481 | ppc32Got2.reset(); |
4482 | ibtPlt.reset(); |
4483 | relaPlt.reset(); |
4484 | debugNames.reset(); |
4485 | gdbIndex.reset(); |
4486 | shStrTab.reset(); |
4487 | strTab.reset(); |
4488 | symTab.reset(); |
4489 | symTabShndx.reset(); |
4490 | } |
4491 | |
4492 | static bool needsInterpSection() { |
4493 | return !config->relocatable && !config->shared && |
4494 | !config->dynamicLinker.empty() && script->needsInterpSection(); |
4495 | } |
4496 | |
4497 | bool elf::hasMemtag() { |
4498 | return config->emachine == EM_AARCH64 && |
4499 | config->androidMemtagMode != ELF::NT_MEMTAG_LEVEL_NONE; |
4500 | } |
4501 | |
4502 | // Fully static executables don't support MTE globals at this point in time, as |
4503 | // we currently rely on: |
4504 | // - A dynamic loader to process relocations, and |
4505 | // - Dynamic entries. |
4506 | // This restriction could be removed in future by re-using some of the ideas |
4507 | // that ifuncs use in fully static executables. |
4508 | bool elf::canHaveMemtagGlobals() { |
4509 | return hasMemtag() && |
4510 | (config->relocatable || config->shared || needsInterpSection()); |
4511 | } |
4512 | |
4513 | constexpr char kMemtagAndroidNoteName[] = "Android" ; |
4514 | void MemtagAndroidNote::writeTo(uint8_t *buf) { |
4515 | static_assert( |
4516 | sizeof(kMemtagAndroidNoteName) == 8, |
4517 | "Android 11 & 12 have an ABI that the note name is 8 bytes long. Keep it " |
4518 | "that way for backwards compatibility." ); |
4519 | |
4520 | write32(p: buf, v: sizeof(kMemtagAndroidNoteName)); |
4521 | write32(p: buf + 4, v: sizeof(uint32_t)); |
4522 | write32(p: buf + 8, v: ELF::NT_ANDROID_TYPE_MEMTAG); |
4523 | memcpy(dest: buf + 12, src: kMemtagAndroidNoteName, n: sizeof(kMemtagAndroidNoteName)); |
4524 | buf += 12 + alignTo(Value: sizeof(kMemtagAndroidNoteName), Align: 4); |
4525 | |
4526 | uint32_t value = 0; |
4527 | value |= config->androidMemtagMode; |
4528 | if (config->androidMemtagHeap) |
4529 | value |= ELF::NT_MEMTAG_HEAP; |
4530 | // Note, MTE stack is an ABI break. Attempting to run an MTE stack-enabled |
4531 | // binary on Android 11 or 12 will result in a checkfail in the loader. |
4532 | if (config->androidMemtagStack) |
4533 | value |= ELF::NT_MEMTAG_STACK; |
4534 | write32(p: buf, v: value); // note value |
4535 | } |
4536 | |
4537 | size_t MemtagAndroidNote::getSize() const { |
4538 | return sizeof(llvm::ELF::Elf64_Nhdr) + |
4539 | /*namesz=*/alignTo(Value: sizeof(kMemtagAndroidNoteName), Align: 4) + |
4540 | /*descsz=*/sizeof(uint32_t); |
4541 | } |
4542 | |
4543 | void PackageMetadataNote::writeTo(uint8_t *buf) { |
4544 | write32(p: buf, v: 4); |
4545 | write32(p: buf + 4, v: config->packageMetadata.size() + 1); |
4546 | write32(p: buf + 8, v: FDO_PACKAGING_METADATA); |
4547 | memcpy(dest: buf + 12, src: "FDO" , n: 4); |
4548 | memcpy(dest: buf + 16, src: config->packageMetadata.data(), |
4549 | n: config->packageMetadata.size()); |
4550 | } |
4551 | |
4552 | size_t PackageMetadataNote::getSize() const { |
4553 | return sizeof(llvm::ELF::Elf64_Nhdr) + 4 + |
4554 | alignTo(Value: config->packageMetadata.size() + 1, Align: 4); |
4555 | } |
4556 | |
4557 | // Helper function, return the size of the ULEB128 for 'v', optionally writing |
4558 | // it to `*(buf + offset)` if `buf` is non-null. |
4559 | static size_t computeOrWriteULEB128(uint64_t v, uint8_t *buf, size_t offset) { |
4560 | if (buf) |
4561 | return encodeULEB128(Value: v, p: buf + offset); |
4562 | return getULEB128Size(Value: v); |
4563 | } |
4564 | |
4565 | // https://github.com/ARM-software/abi-aa/blob/main/memtagabielf64/memtagabielf64.rst#83encoding-of-sht_aarch64_memtag_globals_dynamic |
4566 | constexpr uint64_t kMemtagStepSizeBits = 3; |
4567 | constexpr uint64_t kMemtagGranuleSize = 16; |
4568 | static size_t |
4569 | createMemtagGlobalDescriptors(const SmallVector<const Symbol *, 0> &symbols, |
4570 | uint8_t *buf = nullptr) { |
4571 | size_t sectionSize = 0; |
4572 | uint64_t lastGlobalEnd = 0; |
4573 | |
4574 | for (const Symbol *sym : symbols) { |
4575 | if (!includeInSymtab(b: *sym)) |
4576 | continue; |
4577 | const uint64_t addr = sym->getVA(); |
4578 | const uint64_t size = sym->getSize(); |
4579 | |
4580 | if (addr <= kMemtagGranuleSize && buf != nullptr) |
4581 | errorOrWarn(msg: "address of the tagged symbol \"" + sym->getName() + |
4582 | "\" falls in the ELF header. This is indicative of a " |
4583 | "compiler/linker bug" ); |
4584 | if (addr % kMemtagGranuleSize != 0) |
4585 | errorOrWarn(msg: "address of the tagged symbol \"" + sym->getName() + |
4586 | "\" at 0x" + Twine::utohexstr(Val: addr) + |
4587 | "\" is not granule (16-byte) aligned" ); |
4588 | if (size == 0) |
4589 | errorOrWarn(msg: "size of the tagged symbol \"" + sym->getName() + |
4590 | "\" is not allowed to be zero" ); |
4591 | if (size % kMemtagGranuleSize != 0) |
4592 | errorOrWarn(msg: "size of the tagged symbol \"" + sym->getName() + |
4593 | "\" (size 0x" + Twine::utohexstr(Val: size) + |
4594 | ") is not granule (16-byte) aligned" ); |
4595 | |
4596 | const uint64_t sizeToEncode = size / kMemtagGranuleSize; |
4597 | const uint64_t stepToEncode = ((addr - lastGlobalEnd) / kMemtagGranuleSize) |
4598 | << kMemtagStepSizeBits; |
4599 | if (sizeToEncode < (1 << kMemtagStepSizeBits)) { |
4600 | sectionSize += computeOrWriteULEB128(v: stepToEncode | sizeToEncode, buf, offset: sectionSize); |
4601 | } else { |
4602 | sectionSize += computeOrWriteULEB128(v: stepToEncode, buf, offset: sectionSize); |
4603 | sectionSize += computeOrWriteULEB128(v: sizeToEncode - 1, buf, offset: sectionSize); |
4604 | } |
4605 | lastGlobalEnd = addr + size; |
4606 | } |
4607 | |
4608 | return sectionSize; |
4609 | } |
4610 | |
4611 | bool MemtagGlobalDescriptors::updateAllocSize() { |
4612 | size_t oldSize = getSize(); |
4613 | std::stable_sort(first: symbols.begin(), last: symbols.end(), |
4614 | comp: [](const Symbol *s1, const Symbol *s2) { |
4615 | return s1->getVA() < s2->getVA(); |
4616 | }); |
4617 | return oldSize != getSize(); |
4618 | } |
4619 | |
4620 | void MemtagGlobalDescriptors::writeTo(uint8_t *buf) { |
4621 | createMemtagGlobalDescriptors(symbols, buf); |
4622 | } |
4623 | |
4624 | size_t MemtagGlobalDescriptors::getSize() const { |
4625 | return createMemtagGlobalDescriptors(symbols); |
4626 | } |
4627 | |
4628 | static OutputSection *findSection(StringRef name) { |
4629 | for (SectionCommand *cmd : script->sectionCommands) |
4630 | if (auto *osd = dyn_cast<OutputDesc>(Val: cmd)) |
4631 | if (osd->osec.name == name) |
4632 | return &osd->osec; |
4633 | return nullptr; |
4634 | } |
4635 | |
4636 | static Defined *addOptionalRegular(StringRef name, SectionBase *sec, |
4637 | uint64_t val, uint8_t stOther = STV_HIDDEN) { |
4638 | Symbol *s = symtab.find(name); |
4639 | if (!s || s->isDefined() || s->isCommon()) |
4640 | return nullptr; |
4641 | |
4642 | s->resolve(other: Defined{ctx.internalFile, StringRef(), STB_GLOBAL, stOther, |
4643 | STT_NOTYPE, val, |
4644 | /*size=*/0, sec}); |
4645 | s->isUsedInRegularObj = true; |
4646 | return cast<Defined>(Val: s); |
4647 | } |
4648 | |
4649 | template <class ELFT> void elf::createSyntheticSections() { |
4650 | // Initialize all pointers with NULL. This is needed because |
4651 | // you can call lld::elf::main more than once as a library. |
4652 | Out::tlsPhdr = nullptr; |
4653 | Out::preinitArray = nullptr; |
4654 | Out::initArray = nullptr; |
4655 | Out::finiArray = nullptr; |
4656 | |
4657 | // Add the .interp section first because it is not a SyntheticSection. |
4658 | // The removeUnusedSyntheticSections() function relies on the |
4659 | // SyntheticSections coming last. |
4660 | if (needsInterpSection()) { |
4661 | for (size_t i = 1; i <= partitions.size(); ++i) { |
4662 | InputSection *sec = createInterpSection(); |
4663 | sec->partition = i; |
4664 | ctx.inputSections.push_back(Elt: sec); |
4665 | } |
4666 | } |
4667 | |
4668 | auto add = [](SyntheticSection &sec) { ctx.inputSections.push_back(Elt: &sec); }; |
4669 | |
4670 | in.shStrTab = std::make_unique<StringTableSection>(args: ".shstrtab" , args: false); |
4671 | |
4672 | Out::programHeaders = make<OutputSection>(args: "" , args: 0, args: SHF_ALLOC); |
4673 | Out::programHeaders->addralign = config->wordsize; |
4674 | |
4675 | if (config->strip != StripPolicy::All) { |
4676 | in.strTab = std::make_unique<StringTableSection>(args: ".strtab" , args: false); |
4677 | in.symTab = std::make_unique<SymbolTableSection<ELFT>>(*in.strTab); |
4678 | in.symTabShndx = std::make_unique<SymtabShndxSection>(); |
4679 | } |
4680 | |
4681 | in.bss = std::make_unique<BssSection>(args: ".bss" , args: 0, args: 1); |
4682 | add(*in.bss); |
4683 | |
4684 | // If there is a SECTIONS command and a .data.rel.ro section name use name |
4685 | // .data.rel.ro.bss so that we match in the .data.rel.ro output section. |
4686 | // This makes sure our relro is contiguous. |
4687 | bool hasDataRelRo = script->hasSectionsCommand && findSection(name: ".data.rel.ro" ); |
4688 | in.bssRelRo = std::make_unique<BssSection>( |
4689 | args: hasDataRelRo ? ".data.rel.ro.bss" : ".bss.rel.ro" , args: 0, args: 1); |
4690 | add(*in.bssRelRo); |
4691 | |
4692 | // Add MIPS-specific sections. |
4693 | if (config->emachine == EM_MIPS) { |
4694 | if (!config->shared && config->hasDynSymTab) { |
4695 | in.mipsRldMap = std::make_unique<MipsRldMapSection>(); |
4696 | add(*in.mipsRldMap); |
4697 | } |
4698 | if ((in.mipsAbiFlags = MipsAbiFlagsSection<ELFT>::create())) |
4699 | add(*in.mipsAbiFlags); |
4700 | if ((in.mipsOptions = MipsOptionsSection<ELFT>::create())) |
4701 | add(*in.mipsOptions); |
4702 | if ((in.mipsReginfo = MipsReginfoSection<ELFT>::create())) |
4703 | add(*in.mipsReginfo); |
4704 | } |
4705 | |
4706 | StringRef relaDynName = config->isRela ? ".rela.dyn" : ".rel.dyn" ; |
4707 | |
4708 | const unsigned threadCount = config->threadCount; |
4709 | for (Partition &part : partitions) { |
4710 | auto add = [&](SyntheticSection &sec) { |
4711 | sec.partition = part.getNumber(); |
4712 | ctx.inputSections.push_back(Elt: &sec); |
4713 | }; |
4714 | |
4715 | if (!part.name.empty()) { |
4716 | part.elfHeader = std::make_unique<PartitionElfHeaderSection<ELFT>>(); |
4717 | part.elfHeader->name = part.name; |
4718 | add(*part.elfHeader); |
4719 | |
4720 | part.programHeaders = |
4721 | std::make_unique<PartitionProgramHeadersSection<ELFT>>(); |
4722 | add(*part.programHeaders); |
4723 | } |
4724 | |
4725 | if (config->buildId != BuildIdKind::None) { |
4726 | part.buildId = std::make_unique<BuildIdSection>(); |
4727 | add(*part.buildId); |
4728 | } |
4729 | |
4730 | // dynSymTab is always present to simplify sym->includeInDynsym() in |
4731 | // finalizeSections. |
4732 | part.dynStrTab = std::make_unique<StringTableSection>(args: ".dynstr" , args: true); |
4733 | part.dynSymTab = |
4734 | std::make_unique<SymbolTableSection<ELFT>>(*part.dynStrTab); |
4735 | |
4736 | if (config->relocatable) |
4737 | continue; |
4738 | part.dynamic = std::make_unique<DynamicSection<ELFT>>(); |
4739 | |
4740 | if (hasMemtag()) { |
4741 | part.memtagAndroidNote = std::make_unique<MemtagAndroidNote>(); |
4742 | add(*part.memtagAndroidNote); |
4743 | if (canHaveMemtagGlobals()) { |
4744 | part.memtagGlobalDescriptors = |
4745 | std::make_unique<MemtagGlobalDescriptors>(); |
4746 | add(*part.memtagGlobalDescriptors); |
4747 | } |
4748 | } |
4749 | |
4750 | if (config->androidPackDynRelocs) |
4751 | part.relaDyn = std::make_unique<AndroidPackedRelocationSection<ELFT>>( |
4752 | relaDynName, threadCount); |
4753 | else |
4754 | part.relaDyn = std::make_unique<RelocationSection<ELFT>>( |
4755 | relaDynName, config->zCombreloc, threadCount); |
4756 | |
4757 | if (config->hasDynSymTab) { |
4758 | add(*part.dynSymTab); |
4759 | |
4760 | part.verSym = std::make_unique<VersionTableSection>(); |
4761 | add(*part.verSym); |
4762 | |
4763 | if (!namedVersionDefs().empty()) { |
4764 | part.verDef = std::make_unique<VersionDefinitionSection>(); |
4765 | add(*part.verDef); |
4766 | } |
4767 | |
4768 | part.verNeed = std::make_unique<VersionNeedSection<ELFT>>(); |
4769 | add(*part.verNeed); |
4770 | |
4771 | if (config->gnuHash) { |
4772 | part.gnuHashTab = std::make_unique<GnuHashTableSection>(); |
4773 | add(*part.gnuHashTab); |
4774 | } |
4775 | |
4776 | if (config->sysvHash) { |
4777 | part.hashTab = std::make_unique<HashTableSection>(); |
4778 | add(*part.hashTab); |
4779 | } |
4780 | |
4781 | add(*part.dynamic); |
4782 | add(*part.dynStrTab); |
4783 | } |
4784 | add(*part.relaDyn); |
4785 | |
4786 | if (config->relrPackDynRelocs) { |
4787 | part.relrDyn = std::make_unique<RelrSection<ELFT>>(threadCount); |
4788 | add(*part.relrDyn); |
4789 | part.relrAuthDyn = std::make_unique<RelrSection<ELFT>>( |
4790 | threadCount, /*isAArch64Auth=*/true); |
4791 | add(*part.relrAuthDyn); |
4792 | } |
4793 | |
4794 | if (config->ehFrameHdr) { |
4795 | part.ehFrameHdr = std::make_unique<EhFrameHeader>(); |
4796 | add(*part.ehFrameHdr); |
4797 | } |
4798 | part.ehFrame = std::make_unique<EhFrameSection>(); |
4799 | add(*part.ehFrame); |
4800 | |
4801 | if (config->emachine == EM_ARM) { |
4802 | // This section replaces all the individual .ARM.exidx InputSections. |
4803 | part.armExidx = std::make_unique<ARMExidxSyntheticSection>(); |
4804 | add(*part.armExidx); |
4805 | } |
4806 | |
4807 | if (!config->packageMetadata.empty()) { |
4808 | part.packageMetadataNote = std::make_unique<PackageMetadataNote>(); |
4809 | add(*part.packageMetadataNote); |
4810 | } |
4811 | } |
4812 | |
4813 | if (partitions.size() != 1) { |
4814 | // Create the partition end marker. This needs to be in partition number 255 |
4815 | // so that it is sorted after all other partitions. It also has other |
4816 | // special handling (see createPhdrs() and combineEhSections()). |
4817 | in.partEnd = |
4818 | std::make_unique<BssSection>(args: ".part.end" , args&: config->maxPageSize, args: 1); |
4819 | in.partEnd->partition = 255; |
4820 | add(*in.partEnd); |
4821 | |
4822 | in.partIndex = std::make_unique<PartitionIndexSection>(); |
4823 | addOptionalRegular(name: "__part_index_begin" , sec: in.partIndex.get(), val: 0); |
4824 | addOptionalRegular(name: "__part_index_end" , sec: in.partIndex.get(), |
4825 | val: in.partIndex->getSize()); |
4826 | add(*in.partIndex); |
4827 | } |
4828 | |
4829 | // Add .got. MIPS' .got is so different from the other archs, |
4830 | // it has its own class. |
4831 | if (config->emachine == EM_MIPS) { |
4832 | in.mipsGot = std::make_unique<MipsGotSection>(); |
4833 | add(*in.mipsGot); |
4834 | } else { |
4835 | in.got = std::make_unique<GotSection>(); |
4836 | add(*in.got); |
4837 | } |
4838 | |
4839 | if (config->emachine == EM_PPC) { |
4840 | in.ppc32Got2 = std::make_unique<PPC32Got2Section>(); |
4841 | add(*in.ppc32Got2); |
4842 | } |
4843 | |
4844 | if (config->emachine == EM_PPC64) { |
4845 | in.ppc64LongBranchTarget = std::make_unique<PPC64LongBranchTargetSection>(); |
4846 | add(*in.ppc64LongBranchTarget); |
4847 | } |
4848 | |
4849 | in.gotPlt = std::make_unique<GotPltSection>(); |
4850 | add(*in.gotPlt); |
4851 | in.igotPlt = std::make_unique<IgotPltSection>(); |
4852 | add(*in.igotPlt); |
4853 | // Add .relro_padding if DATA_SEGMENT_RELRO_END is used; otherwise, add the |
4854 | // section in the absence of PHDRS/SECTIONS commands. |
4855 | if (config->zRelro && |
4856 | ((script->phdrsCommands.empty() && !script->hasSectionsCommand) || |
4857 | script->seenRelroEnd)) { |
4858 | in.relroPadding = std::make_unique<RelroPaddingSection>(); |
4859 | add(*in.relroPadding); |
4860 | } |
4861 | |
4862 | if (config->emachine == EM_ARM) { |
4863 | in.armCmseSGSection = std::make_unique<ArmCmseSGSection>(); |
4864 | add(*in.armCmseSGSection); |
4865 | } |
4866 | |
4867 | // _GLOBAL_OFFSET_TABLE_ is defined relative to either .got.plt or .got. Treat |
4868 | // it as a relocation and ensure the referenced section is created. |
4869 | if (ElfSym::globalOffsetTable && config->emachine != EM_MIPS) { |
4870 | if (target->gotBaseSymInGotPlt) |
4871 | in.gotPlt->hasGotPltOffRel = true; |
4872 | else |
4873 | in.got->hasGotOffRel = true; |
4874 | } |
4875 | |
4876 | // We always need to add rel[a].plt to output if it has entries. |
4877 | // Even for static linking it can contain R_[*]_IRELATIVE relocations. |
4878 | in.relaPlt = std::make_unique<RelocationSection<ELFT>>( |
4879 | config->isRela ? ".rela.plt" : ".rel.plt" , /*sort=*/false, |
4880 | /*threadCount=*/1); |
4881 | add(*in.relaPlt); |
4882 | |
4883 | if ((config->emachine == EM_386 || config->emachine == EM_X86_64) && |
4884 | (config->andFeatures & GNU_PROPERTY_X86_FEATURE_1_IBT)) { |
4885 | in.ibtPlt = std::make_unique<IBTPltSection>(); |
4886 | add(*in.ibtPlt); |
4887 | } |
4888 | |
4889 | if (config->emachine == EM_PPC) |
4890 | in.plt = std::make_unique<PPC32GlinkSection>(); |
4891 | else |
4892 | in.plt = std::make_unique<PltSection>(); |
4893 | add(*in.plt); |
4894 | in.iplt = std::make_unique<IpltSection>(); |
4895 | add(*in.iplt); |
4896 | |
4897 | if (config->andFeatures || !ctx.aarch64PauthAbiCoreInfo.empty()) |
4898 | add(*make<GnuPropertySection>()); |
4899 | |
4900 | if (config->debugNames) { |
4901 | in.debugNames = std::make_unique<DebugNamesSection<ELFT>>(); |
4902 | add(*in.debugNames); |
4903 | } |
4904 | |
4905 | if (config->gdbIndex) { |
4906 | in.gdbIndex = GdbIndexSection::create<ELFT>(); |
4907 | add(*in.gdbIndex); |
4908 | } |
4909 | |
4910 | // .note.GNU-stack is always added when we are creating a re-linkable |
4911 | // object file. Other linkers are using the presence of this marker |
4912 | // section to control the executable-ness of the stack area, but that |
4913 | // is irrelevant these days. Stack area should always be non-executable |
4914 | // by default. So we emit this section unconditionally. |
4915 | if (config->relocatable) |
4916 | add(*make<GnuStackSection>()); |
4917 | |
4918 | if (in.symTab) |
4919 | add(*in.symTab); |
4920 | if (in.symTabShndx) |
4921 | add(*in.symTabShndx); |
4922 | add(*in.shStrTab); |
4923 | if (in.strTab) |
4924 | add(*in.strTab); |
4925 | } |
4926 | |
4927 | InStruct elf::in; |
4928 | |
4929 | std::vector<Partition> elf::partitions; |
4930 | Partition *elf::mainPart; |
4931 | |
4932 | template void elf::splitSections<ELF32LE>(); |
4933 | template void elf::splitSections<ELF32BE>(); |
4934 | template void elf::splitSections<ELF64LE>(); |
4935 | template void elf::splitSections<ELF64BE>(); |
4936 | |
4937 | template void EhFrameSection::iterateFDEWithLSDA<ELF32LE>( |
4938 | function_ref<void(InputSection &)>); |
4939 | template void EhFrameSection::iterateFDEWithLSDA<ELF32BE>( |
4940 | function_ref<void(InputSection &)>); |
4941 | template void EhFrameSection::iterateFDEWithLSDA<ELF64LE>( |
4942 | function_ref<void(InputSection &)>); |
4943 | template void EhFrameSection::iterateFDEWithLSDA<ELF64BE>( |
4944 | function_ref<void(InputSection &)>); |
4945 | |
4946 | template class elf::SymbolTableSection<ELF32LE>; |
4947 | template class elf::SymbolTableSection<ELF32BE>; |
4948 | template class elf::SymbolTableSection<ELF64LE>; |
4949 | template class elf::SymbolTableSection<ELF64BE>; |
4950 | |
4951 | template void elf::writeEhdr<ELF32LE>(uint8_t *Buf, Partition &Part); |
4952 | template void elf::writeEhdr<ELF32BE>(uint8_t *Buf, Partition &Part); |
4953 | template void elf::writeEhdr<ELF64LE>(uint8_t *Buf, Partition &Part); |
4954 | template void elf::writeEhdr<ELF64BE>(uint8_t *Buf, Partition &Part); |
4955 | |
4956 | template void elf::writePhdrs<ELF32LE>(uint8_t *Buf, Partition &Part); |
4957 | template void elf::writePhdrs<ELF32BE>(uint8_t *Buf, Partition &Part); |
4958 | template void elf::writePhdrs<ELF64LE>(uint8_t *Buf, Partition &Part); |
4959 | template void elf::writePhdrs<ELF64BE>(uint8_t *Buf, Partition &Part); |
4960 | |
4961 | template void elf::createSyntheticSections<ELF32LE>(); |
4962 | template void elf::createSyntheticSections<ELF32BE>(); |
4963 | template void elf::createSyntheticSections<ELF64LE>(); |
4964 | template void elf::createSyntheticSections<ELF64BE>(); |
4965 | |