1 | //===- InputFiles.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 | #include "InputFiles.h" |
10 | #include "Config.h" |
11 | #include "DWARF.h" |
12 | #include "Driver.h" |
13 | #include "InputSection.h" |
14 | #include "LinkerScript.h" |
15 | #include "SymbolTable.h" |
16 | #include "Symbols.h" |
17 | #include "SyntheticSections.h" |
18 | #include "Target.h" |
19 | #include "lld/Common/CommonLinkerContext.h" |
20 | #include "lld/Common/DWARF.h" |
21 | #include "llvm/ADT/CachedHashString.h" |
22 | #include "llvm/ADT/STLExtras.h" |
23 | #include "llvm/LTO/LTO.h" |
24 | #include "llvm/Object/IRObjectFile.h" |
25 | #include "llvm/Support/ARMAttributeParser.h" |
26 | #include "llvm/Support/ARMBuildAttributes.h" |
27 | #include "llvm/Support/Endian.h" |
28 | #include "llvm/Support/FileSystem.h" |
29 | #include "llvm/Support/Path.h" |
30 | #include "llvm/Support/RISCVAttributeParser.h" |
31 | #include "llvm/Support/TarWriter.h" |
32 | #include "llvm/Support/raw_ostream.h" |
33 | #include <optional> |
34 | |
35 | using namespace llvm; |
36 | using namespace llvm::ELF; |
37 | using namespace llvm::object; |
38 | using namespace llvm::sys; |
39 | using namespace llvm::sys::fs; |
40 | using namespace llvm::support::endian; |
41 | using namespace lld; |
42 | using namespace lld::elf; |
43 | |
44 | // This function is explicitly instantiated in ARM.cpp, don't do it here to |
45 | // avoid warnings with MSVC. |
46 | extern template void ObjFile<ELF32LE>::importCmseSymbols(); |
47 | extern template void ObjFile<ELF32BE>::importCmseSymbols(); |
48 | extern template void ObjFile<ELF64LE>::importCmseSymbols(); |
49 | extern template void ObjFile<ELF64BE>::importCmseSymbols(); |
50 | |
51 | bool InputFile::isInGroup; |
52 | uint32_t InputFile::nextGroupId; |
53 | |
54 | std::unique_ptr<TarWriter> elf::tar; |
55 | |
56 | // Returns "<internal>", "foo.a(bar.o)" or "baz.o". |
57 | std::string lld::toString(const InputFile *f) { |
58 | static std::mutex mu; |
59 | if (!f) |
60 | return "<internal>" ; |
61 | |
62 | { |
63 | std::lock_guard<std::mutex> lock(mu); |
64 | if (f->toStringCache.empty()) { |
65 | if (f->archiveName.empty()) |
66 | f->toStringCache = f->getName(); |
67 | else |
68 | (f->archiveName + "(" + f->getName() + ")" ).toVector(Out&: f->toStringCache); |
69 | } |
70 | } |
71 | return std::string(f->toStringCache); |
72 | } |
73 | |
74 | static ELFKind getELFKind(MemoryBufferRef mb, StringRef archiveName) { |
75 | unsigned char size; |
76 | unsigned char endian; |
77 | std::tie(args&: size, args&: endian) = getElfArchType(Object: mb.getBuffer()); |
78 | |
79 | auto report = [&](StringRef msg) { |
80 | StringRef filename = mb.getBufferIdentifier(); |
81 | if (archiveName.empty()) |
82 | fatal(msg: filename + ": " + msg); |
83 | else |
84 | fatal(msg: archiveName + "(" + filename + "): " + msg); |
85 | }; |
86 | |
87 | if (!mb.getBuffer().starts_with(Prefix: ElfMagic)) |
88 | report("not an ELF file" ); |
89 | if (endian != ELFDATA2LSB && endian != ELFDATA2MSB) |
90 | report("corrupted ELF file: invalid data encoding" ); |
91 | if (size != ELFCLASS32 && size != ELFCLASS64) |
92 | report("corrupted ELF file: invalid file class" ); |
93 | |
94 | size_t bufSize = mb.getBuffer().size(); |
95 | if ((size == ELFCLASS32 && bufSize < sizeof(Elf32_Ehdr)) || |
96 | (size == ELFCLASS64 && bufSize < sizeof(Elf64_Ehdr))) |
97 | report("corrupted ELF file: file is too short" ); |
98 | |
99 | if (size == ELFCLASS32) |
100 | return (endian == ELFDATA2LSB) ? ELF32LEKind : ELF32BEKind; |
101 | return (endian == ELFDATA2LSB) ? ELF64LEKind : ELF64BEKind; |
102 | } |
103 | |
104 | // For ARM only, to set the EF_ARM_ABI_FLOAT_SOFT or EF_ARM_ABI_FLOAT_HARD |
105 | // flag in the ELF Header we need to look at Tag_ABI_VFP_args to find out how |
106 | // the input objects have been compiled. |
107 | static void updateARMVFPArgs(const ARMAttributeParser &attributes, |
108 | const InputFile *f) { |
109 | std::optional<unsigned> attr = |
110 | attributes.getAttributeValue(tag: ARMBuildAttrs::ABI_VFP_args); |
111 | if (!attr) |
112 | // If an ABI tag isn't present then it is implicitly given the value of 0 |
113 | // which maps to ARMBuildAttrs::BaseAAPCS. However many assembler files, |
114 | // including some in glibc that don't use FP args (and should have value 3) |
115 | // don't have the attribute so we do not consider an implicit value of 0 |
116 | // as a clash. |
117 | return; |
118 | |
119 | unsigned vfpArgs = *attr; |
120 | ARMVFPArgKind arg; |
121 | switch (vfpArgs) { |
122 | case ARMBuildAttrs::BaseAAPCS: |
123 | arg = ARMVFPArgKind::Base; |
124 | break; |
125 | case ARMBuildAttrs::HardFPAAPCS: |
126 | arg = ARMVFPArgKind::VFP; |
127 | break; |
128 | case ARMBuildAttrs::ToolChainFPPCS: |
129 | // Tool chain specific convention that conforms to neither AAPCS variant. |
130 | arg = ARMVFPArgKind::ToolChain; |
131 | break; |
132 | case ARMBuildAttrs::CompatibleFPAAPCS: |
133 | // Object compatible with all conventions. |
134 | return; |
135 | default: |
136 | error(msg: toString(f) + ": unknown Tag_ABI_VFP_args value: " + Twine(vfpArgs)); |
137 | return; |
138 | } |
139 | // Follow ld.bfd and error if there is a mix of calling conventions. |
140 | if (config->armVFPArgs != arg && config->armVFPArgs != ARMVFPArgKind::Default) |
141 | error(msg: toString(f) + ": incompatible Tag_ABI_VFP_args" ); |
142 | else |
143 | config->armVFPArgs = arg; |
144 | } |
145 | |
146 | // The ARM support in lld makes some use of instructions that are not available |
147 | // on all ARM architectures. Namely: |
148 | // - Use of BLX instruction for interworking between ARM and Thumb state. |
149 | // - Use of the extended Thumb branch encoding in relocation. |
150 | // - Use of the MOVT/MOVW instructions in Thumb Thunks. |
151 | // The ARM Attributes section contains information about the architecture chosen |
152 | // at compile time. We follow the convention that if at least one input object |
153 | // is compiled with an architecture that supports these features then lld is |
154 | // permitted to use them. |
155 | static void updateSupportedARMFeatures(const ARMAttributeParser &attributes) { |
156 | std::optional<unsigned> attr = |
157 | attributes.getAttributeValue(tag: ARMBuildAttrs::CPU_arch); |
158 | if (!attr) |
159 | return; |
160 | auto arch = *attr; |
161 | switch (arch) { |
162 | case ARMBuildAttrs::Pre_v4: |
163 | case ARMBuildAttrs::v4: |
164 | case ARMBuildAttrs::v4T: |
165 | // Architectures prior to v5 do not support BLX instruction |
166 | break; |
167 | case ARMBuildAttrs::v5T: |
168 | case ARMBuildAttrs::v5TE: |
169 | case ARMBuildAttrs::v5TEJ: |
170 | case ARMBuildAttrs::v6: |
171 | case ARMBuildAttrs::v6KZ: |
172 | case ARMBuildAttrs::v6K: |
173 | config->armHasBlx = true; |
174 | // Architectures used in pre-Cortex processors do not support |
175 | // The J1 = 1 J2 = 1 Thumb branch range extension, with the exception |
176 | // of Architecture v6T2 (arm1156t2-s and arm1156t2f-s) that do. |
177 | break; |
178 | default: |
179 | // All other Architectures have BLX and extended branch encoding |
180 | config->armHasBlx = true; |
181 | config->armJ1J2BranchEncoding = true; |
182 | if (arch != ARMBuildAttrs::v6_M && arch != ARMBuildAttrs::v6S_M) |
183 | // All Architectures used in Cortex processors with the exception |
184 | // of v6-M and v6S-M have the MOVT and MOVW instructions. |
185 | config->armHasMovtMovw = true; |
186 | break; |
187 | } |
188 | |
189 | // Only ARMv8-M or later architectures have CMSE support. |
190 | std::optional<unsigned> profile = |
191 | attributes.getAttributeValue(tag: ARMBuildAttrs::CPU_arch_profile); |
192 | if (!profile) |
193 | return; |
194 | if (arch >= ARMBuildAttrs::CPUArch::v8_M_Base && |
195 | profile == ARMBuildAttrs::MicroControllerProfile) |
196 | config->armCMSESupport = true; |
197 | } |
198 | |
199 | InputFile::InputFile(Kind k, MemoryBufferRef m) |
200 | : mb(m), groupId(nextGroupId), fileKind(k) { |
201 | // All files within the same --{start,end}-group get the same group ID. |
202 | // Otherwise, a new file will get a new group ID. |
203 | if (!isInGroup) |
204 | ++nextGroupId; |
205 | } |
206 | |
207 | std::optional<MemoryBufferRef> elf::readFile(StringRef path) { |
208 | llvm::TimeTraceScope timeScope("Load input files" , path); |
209 | |
210 | // The --chroot option changes our virtual root directory. |
211 | // This is useful when you are dealing with files created by --reproduce. |
212 | if (!config->chroot.empty() && path.starts_with(Prefix: "/" )) |
213 | path = saver().save(S: config->chroot + path); |
214 | |
215 | bool remapped = false; |
216 | auto it = config->remapInputs.find(Val: path); |
217 | if (it != config->remapInputs.end()) { |
218 | path = it->second; |
219 | remapped = true; |
220 | } else { |
221 | for (const auto &[pat, toFile] : config->remapInputsWildcards) { |
222 | if (pat.match(S: path)) { |
223 | path = toFile; |
224 | remapped = true; |
225 | break; |
226 | } |
227 | } |
228 | } |
229 | if (remapped) { |
230 | // Use /dev/null to indicate an input file that should be ignored. Change |
231 | // the path to NUL on Windows. |
232 | #ifdef _WIN32 |
233 | if (path == "/dev/null" ) |
234 | path = "NUL" ; |
235 | #endif |
236 | } |
237 | |
238 | log(msg: path); |
239 | config->dependencyFiles.insert(X: llvm::CachedHashString(path)); |
240 | |
241 | auto mbOrErr = MemoryBuffer::getFile(Filename: path, /*IsText=*/false, |
242 | /*RequiresNullTerminator=*/false); |
243 | if (auto ec = mbOrErr.getError()) { |
244 | error(msg: "cannot open " + path + ": " + ec.message()); |
245 | return std::nullopt; |
246 | } |
247 | |
248 | MemoryBufferRef mbref = (*mbOrErr)->getMemBufferRef(); |
249 | ctx.memoryBuffers.push_back(Elt: std::move(*mbOrErr)); // take MB ownership |
250 | |
251 | if (tar) |
252 | tar->append(Path: relativeToRoot(path), Data: mbref.getBuffer()); |
253 | return mbref; |
254 | } |
255 | |
256 | // All input object files must be for the same architecture |
257 | // (e.g. it does not make sense to link x86 object files with |
258 | // MIPS object files.) This function checks for that error. |
259 | static bool isCompatible(InputFile *file) { |
260 | if (!file->isElf() && !isa<BitcodeFile>(Val: file)) |
261 | return true; |
262 | |
263 | if (file->ekind == config->ekind && file->emachine == config->emachine) { |
264 | if (config->emachine != EM_MIPS) |
265 | return true; |
266 | if (isMipsN32Abi(f: file) == config->mipsN32Abi) |
267 | return true; |
268 | } |
269 | |
270 | StringRef target = |
271 | !config->bfdname.empty() ? config->bfdname : config->emulation; |
272 | if (!target.empty()) { |
273 | error(msg: toString(f: file) + " is incompatible with " + target); |
274 | return false; |
275 | } |
276 | |
277 | InputFile *existing = nullptr; |
278 | if (!ctx.objectFiles.empty()) |
279 | existing = ctx.objectFiles[0]; |
280 | else if (!ctx.sharedFiles.empty()) |
281 | existing = ctx.sharedFiles[0]; |
282 | else if (!ctx.bitcodeFiles.empty()) |
283 | existing = ctx.bitcodeFiles[0]; |
284 | std::string with; |
285 | if (existing) |
286 | with = " with " + toString(f: existing); |
287 | error(msg: toString(f: file) + " is incompatible" + with); |
288 | return false; |
289 | } |
290 | |
291 | template <class ELFT> static void doParseFile(InputFile *file) { |
292 | if (!isCompatible(file)) |
293 | return; |
294 | |
295 | // Lazy object file |
296 | if (file->lazy) { |
297 | if (auto *f = dyn_cast<BitcodeFile>(Val: file)) { |
298 | ctx.lazyBitcodeFiles.push_back(Elt: f); |
299 | f->parseLazy(); |
300 | } else { |
301 | cast<ObjFile<ELFT>>(file)->parseLazy(); |
302 | } |
303 | return; |
304 | } |
305 | |
306 | if (config->trace) |
307 | message(msg: toString(f: file)); |
308 | |
309 | if (file->kind() == InputFile::ObjKind) { |
310 | ctx.objectFiles.push_back(Elt: cast<ELFFileBase>(Val: file)); |
311 | cast<ObjFile<ELFT>>(file)->parse(); |
312 | } else if (auto *f = dyn_cast<SharedFile>(Val: file)) { |
313 | f->parse<ELFT>(); |
314 | } else if (auto *f = dyn_cast<BitcodeFile>(Val: file)) { |
315 | ctx.bitcodeFiles.push_back(Elt: f); |
316 | f->parse(); |
317 | } else { |
318 | ctx.binaryFiles.push_back(Elt: cast<BinaryFile>(Val: file)); |
319 | cast<BinaryFile>(Val: file)->parse(); |
320 | } |
321 | } |
322 | |
323 | // Add symbols in File to the symbol table. |
324 | void elf::parseFile(InputFile *file) { invokeELFT(doParseFile, file); } |
325 | |
326 | // This function is explicitly instantiated in ARM.cpp. Mark it extern here, |
327 | // to avoid warnings when building with MSVC. |
328 | extern template void ObjFile<ELF32LE>::importCmseSymbols(); |
329 | extern template void ObjFile<ELF32BE>::importCmseSymbols(); |
330 | extern template void ObjFile<ELF64LE>::importCmseSymbols(); |
331 | extern template void ObjFile<ELF64BE>::importCmseSymbols(); |
332 | |
333 | template <class ELFT> |
334 | static void doParseFiles(const std::vector<InputFile *> &files, |
335 | InputFile *armCmseImpLib) { |
336 | // Add all files to the symbol table. This will add almost all symbols that we |
337 | // need to the symbol table. This process might add files to the link due to |
338 | // addDependentLibrary. |
339 | for (size_t i = 0; i < files.size(); ++i) { |
340 | llvm::TimeTraceScope timeScope("Parse input files" , files[i]->getName()); |
341 | doParseFile<ELFT>(files[i]); |
342 | } |
343 | if (armCmseImpLib) |
344 | cast<ObjFile<ELFT>>(*armCmseImpLib).importCmseSymbols(); |
345 | } |
346 | |
347 | void elf::parseFiles(const std::vector<InputFile *> &files, |
348 | InputFile *armCmseImpLib) { |
349 | llvm::TimeTraceScope timeScope("Parse input files" ); |
350 | invokeELFT(doParseFiles, files, armCmseImpLib); |
351 | } |
352 | |
353 | // Concatenates arguments to construct a string representing an error location. |
354 | static std::string createFileLineMsg(StringRef path, unsigned line) { |
355 | std::string filename = std::string(path::filename(path)); |
356 | std::string lineno = ":" + std::to_string(val: line); |
357 | if (filename == path) |
358 | return filename + lineno; |
359 | return filename + lineno + " (" + path.str() + lineno + ")" ; |
360 | } |
361 | |
362 | template <class ELFT> |
363 | static std::string getSrcMsgAux(ObjFile<ELFT> &file, const Symbol &sym, |
364 | const InputSectionBase &sec, uint64_t offset) { |
365 | // In DWARF, functions and variables are stored to different places. |
366 | // First, look up a function for a given offset. |
367 | if (std::optional<DILineInfo> info = file.getDILineInfo(&sec, offset)) |
368 | return createFileLineMsg(path: info->FileName, line: info->Line); |
369 | |
370 | // If it failed, look up again as a variable. |
371 | if (std::optional<std::pair<std::string, unsigned>> fileLine = |
372 | file.getVariableLoc(sym.getName())) |
373 | return createFileLineMsg(path: fileLine->first, line: fileLine->second); |
374 | |
375 | // File.sourceFile contains STT_FILE symbol, and that is a last resort. |
376 | return std::string(file.sourceFile); |
377 | } |
378 | |
379 | std::string InputFile::getSrcMsg(const Symbol &sym, const InputSectionBase &sec, |
380 | uint64_t offset) { |
381 | if (kind() != ObjKind) |
382 | return "" ; |
383 | switch (ekind) { |
384 | default: |
385 | llvm_unreachable("Invalid kind" ); |
386 | case ELF32LEKind: |
387 | return getSrcMsgAux(file&: cast<ObjFile<ELF32LE>>(Val&: *this), sym, sec, offset); |
388 | case ELF32BEKind: |
389 | return getSrcMsgAux(file&: cast<ObjFile<ELF32BE>>(Val&: *this), sym, sec, offset); |
390 | case ELF64LEKind: |
391 | return getSrcMsgAux(file&: cast<ObjFile<ELF64LE>>(Val&: *this), sym, sec, offset); |
392 | case ELF64BEKind: |
393 | return getSrcMsgAux(file&: cast<ObjFile<ELF64BE>>(Val&: *this), sym, sec, offset); |
394 | } |
395 | } |
396 | |
397 | StringRef InputFile::getNameForScript() const { |
398 | if (archiveName.empty()) |
399 | return getName(); |
400 | |
401 | if (nameForScriptCache.empty()) |
402 | nameForScriptCache = (archiveName + Twine(':') + getName()).str(); |
403 | |
404 | return nameForScriptCache; |
405 | } |
406 | |
407 | // An ELF object file may contain a `.deplibs` section. If it exists, the |
408 | // section contains a list of library specifiers such as `m` for libm. This |
409 | // function resolves a given name by finding the first matching library checking |
410 | // the various ways that a library can be specified to LLD. This ELF extension |
411 | // is a form of autolinking and is called `dependent libraries`. It is currently |
412 | // unique to LLVM and lld. |
413 | static void addDependentLibrary(StringRef specifier, const InputFile *f) { |
414 | if (!config->dependentLibraries) |
415 | return; |
416 | if (std::optional<std::string> s = searchLibraryBaseName(path: specifier)) |
417 | ctx.driver.addFile(path: saver().save(S: *s), /*withLOption=*/true); |
418 | else if (std::optional<std::string> s = findFromSearchPaths(path: specifier)) |
419 | ctx.driver.addFile(path: saver().save(S: *s), /*withLOption=*/true); |
420 | else if (fs::exists(Path: specifier)) |
421 | ctx.driver.addFile(path: specifier, /*withLOption=*/false); |
422 | else |
423 | error(msg: toString(f) + |
424 | ": unable to find library from dependent library specifier: " + |
425 | specifier); |
426 | } |
427 | |
428 | // Record the membership of a section group so that in the garbage collection |
429 | // pass, section group members are kept or discarded as a unit. |
430 | template <class ELFT> |
431 | static void handleSectionGroup(ArrayRef<InputSectionBase *> sections, |
432 | ArrayRef<typename ELFT::Word> entries) { |
433 | bool hasAlloc = false; |
434 | for (uint32_t index : entries.slice(1)) { |
435 | if (index >= sections.size()) |
436 | return; |
437 | if (InputSectionBase *s = sections[index]) |
438 | if (s != &InputSection::discarded && s->flags & SHF_ALLOC) |
439 | hasAlloc = true; |
440 | } |
441 | |
442 | // If any member has the SHF_ALLOC flag, the whole group is subject to garbage |
443 | // collection. See the comment in markLive(). This rule retains .debug_types |
444 | // and .rela.debug_types. |
445 | if (!hasAlloc) |
446 | return; |
447 | |
448 | // Connect the members in a circular doubly-linked list via |
449 | // nextInSectionGroup. |
450 | InputSectionBase *head; |
451 | InputSectionBase *prev = nullptr; |
452 | for (uint32_t index : entries.slice(1)) { |
453 | InputSectionBase *s = sections[index]; |
454 | if (!s || s == &InputSection::discarded) |
455 | continue; |
456 | if (prev) |
457 | prev->nextInSectionGroup = s; |
458 | else |
459 | head = s; |
460 | prev = s; |
461 | } |
462 | if (prev) |
463 | prev->nextInSectionGroup = head; |
464 | } |
465 | |
466 | template <class ELFT> DWARFCache *ObjFile<ELFT>::getDwarf() { |
467 | llvm::call_once(initDwarf, [this]() { |
468 | dwarf = std::make_unique<DWARFCache>(std::make_unique<DWARFContext>( |
469 | std::make_unique<LLDDwarfObj<ELFT>>(this), "" , |
470 | [&](Error err) { warn(getName() + ": " + toString(E: std::move(err))); }, |
471 | [&](Error warning) { |
472 | warn(getName() + ": " + toString(E: std::move(warning))); |
473 | })); |
474 | }); |
475 | |
476 | return dwarf.get(); |
477 | } |
478 | |
479 | // Returns the pair of file name and line number describing location of data |
480 | // object (variable, array, etc) definition. |
481 | template <class ELFT> |
482 | std::optional<std::pair<std::string, unsigned>> |
483 | ObjFile<ELFT>::getVariableLoc(StringRef name) { |
484 | return getDwarf()->getVariableLoc(name); |
485 | } |
486 | |
487 | // Returns source line information for a given offset |
488 | // using DWARF debug info. |
489 | template <class ELFT> |
490 | std::optional<DILineInfo> |
491 | ObjFile<ELFT>::getDILineInfo(const InputSectionBase *s, uint64_t offset) { |
492 | // Detect SectionIndex for specified section. |
493 | uint64_t sectionIndex = object::SectionedAddress::UndefSection; |
494 | ArrayRef<InputSectionBase *> sections = s->file->getSections(); |
495 | for (uint64_t curIndex = 0; curIndex < sections.size(); ++curIndex) { |
496 | if (s == sections[curIndex]) { |
497 | sectionIndex = curIndex; |
498 | break; |
499 | } |
500 | } |
501 | |
502 | return getDwarf()->getDILineInfo(offset, sectionIndex); |
503 | } |
504 | |
505 | ELFFileBase::ELFFileBase(Kind k, ELFKind ekind, MemoryBufferRef mb) |
506 | : InputFile(k, mb) { |
507 | this->ekind = ekind; |
508 | } |
509 | |
510 | template <typename Elf_Shdr> |
511 | static const Elf_Shdr *findSection(ArrayRef<Elf_Shdr> sections, uint32_t type) { |
512 | for (const Elf_Shdr &sec : sections) |
513 | if (sec.sh_type == type) |
514 | return &sec; |
515 | return nullptr; |
516 | } |
517 | |
518 | void ELFFileBase::init() { |
519 | switch (ekind) { |
520 | case ELF32LEKind: |
521 | init<ELF32LE>(k: fileKind); |
522 | break; |
523 | case ELF32BEKind: |
524 | init<ELF32BE>(k: fileKind); |
525 | break; |
526 | case ELF64LEKind: |
527 | init<ELF64LE>(k: fileKind); |
528 | break; |
529 | case ELF64BEKind: |
530 | init<ELF64BE>(k: fileKind); |
531 | break; |
532 | default: |
533 | llvm_unreachable("getELFKind" ); |
534 | } |
535 | } |
536 | |
537 | template <class ELFT> void ELFFileBase::init(InputFile::Kind k) { |
538 | using Elf_Shdr = typename ELFT::Shdr; |
539 | using Elf_Sym = typename ELFT::Sym; |
540 | |
541 | // Initialize trivial attributes. |
542 | const ELFFile<ELFT> &obj = getObj<ELFT>(); |
543 | emachine = obj.getHeader().e_machine; |
544 | osabi = obj.getHeader().e_ident[llvm::ELF::EI_OSABI]; |
545 | abiVersion = obj.getHeader().e_ident[llvm::ELF::EI_ABIVERSION]; |
546 | |
547 | ArrayRef<Elf_Shdr> sections = CHECK(obj.sections(), this); |
548 | elfShdrs = sections.data(); |
549 | numELFShdrs = sections.size(); |
550 | |
551 | // Find a symbol table. |
552 | const Elf_Shdr *symtabSec = |
553 | findSection(sections, k == SharedKind ? SHT_DYNSYM : SHT_SYMTAB); |
554 | |
555 | if (!symtabSec) |
556 | return; |
557 | |
558 | // Initialize members corresponding to a symbol table. |
559 | firstGlobal = symtabSec->sh_info; |
560 | |
561 | ArrayRef<Elf_Sym> eSyms = CHECK(obj.symbols(symtabSec), this); |
562 | if (firstGlobal == 0 || firstGlobal > eSyms.size()) |
563 | fatal(msg: toString(f: this) + ": invalid sh_info in symbol table" ); |
564 | |
565 | elfSyms = reinterpret_cast<const void *>(eSyms.data()); |
566 | numELFSyms = uint32_t(eSyms.size()); |
567 | stringTable = CHECK(obj.getStringTableForSymtab(*symtabSec, sections), this); |
568 | } |
569 | |
570 | template <class ELFT> |
571 | uint32_t ObjFile<ELFT>::getSectionIndex(const Elf_Sym &sym) const { |
572 | return CHECK( |
573 | this->getObj().getSectionIndex(sym, getELFSyms<ELFT>(), shndxTable), |
574 | this); |
575 | } |
576 | |
577 | template <class ELFT> void ObjFile<ELFT>::parse(bool ignoreComdats) { |
578 | object::ELFFile<ELFT> obj = this->getObj(); |
579 | // Read a section table. justSymbols is usually false. |
580 | if (this->justSymbols) { |
581 | initializeJustSymbols(); |
582 | initializeSymbols(obj); |
583 | return; |
584 | } |
585 | |
586 | // Handle dependent libraries and selection of section groups as these are not |
587 | // done in parallel. |
588 | ArrayRef<Elf_Shdr> objSections = getELFShdrs<ELFT>(); |
589 | StringRef shstrtab = CHECK(obj.getSectionStringTable(objSections), this); |
590 | uint64_t size = objSections.size(); |
591 | sections.resize(size); |
592 | for (size_t i = 0; i != size; ++i) { |
593 | const Elf_Shdr &sec = objSections[i]; |
594 | if (sec.sh_type == SHT_LLVM_DEPENDENT_LIBRARIES && !config->relocatable) { |
595 | StringRef name = check(obj.getSectionName(sec, shstrtab)); |
596 | ArrayRef<char> data = CHECK( |
597 | this->getObj().template getSectionContentsAsArray<char>(sec), this); |
598 | if (!data.empty() && data.back() != '\0') { |
599 | error( |
600 | toString(this) + |
601 | ": corrupted dependent libraries section (unterminated string): " + |
602 | name); |
603 | } else { |
604 | for (const char *d = data.begin(), *e = data.end(); d < e;) { |
605 | StringRef s(d); |
606 | addDependentLibrary(s, this); |
607 | d += s.size() + 1; |
608 | } |
609 | } |
610 | this->sections[i] = &InputSection::discarded; |
611 | continue; |
612 | } |
613 | |
614 | if (sec.sh_type == SHT_ARM_ATTRIBUTES && config->emachine == EM_ARM) { |
615 | ARMAttributeParser attributes; |
616 | ArrayRef<uint8_t> contents = |
617 | check(this->getObj().getSectionContents(sec)); |
618 | StringRef name = check(obj.getSectionName(sec, shstrtab)); |
619 | this->sections[i] = &InputSection::discarded; |
620 | if (Error e = attributes.parse(section: contents, endian: ekind == ELF32LEKind |
621 | ? llvm::endianness::little |
622 | : llvm::endianness::big)) { |
623 | InputSection isec(*this, sec, name); |
624 | warn(msg: toString(&isec) + ": " + llvm::toString(E: std::move(e))); |
625 | } else { |
626 | updateSupportedARMFeatures(attributes); |
627 | updateARMVFPArgs(attributes, this); |
628 | |
629 | // FIXME: Retain the first attribute section we see. The eglibc ARM |
630 | // dynamic loaders require the presence of an attribute section for |
631 | // dlopen to work. In a full implementation we would merge all attribute |
632 | // sections. |
633 | if (in.attributes == nullptr) { |
634 | in.attributes = std::make_unique<InputSection>(*this, sec, name); |
635 | this->sections[i] = in.attributes.get(); |
636 | } |
637 | } |
638 | } |
639 | |
640 | // Producing a static binary with MTE globals is not currently supported, |
641 | // remove all SHT_AARCH64_MEMTAG_GLOBALS_STATIC sections as they're unused |
642 | // medatada, and we don't want them to end up in the output file for static |
643 | // executables. |
644 | if (sec.sh_type == SHT_AARCH64_MEMTAG_GLOBALS_STATIC && |
645 | !canHaveMemtagGlobals()) { |
646 | this->sections[i] = &InputSection::discarded; |
647 | continue; |
648 | } |
649 | |
650 | if (sec.sh_type != SHT_GROUP) |
651 | continue; |
652 | StringRef signature = getShtGroupSignature(sections: objSections, sec); |
653 | ArrayRef<Elf_Word> entries = |
654 | CHECK(obj.template getSectionContentsAsArray<Elf_Word>(sec), this); |
655 | if (entries.empty()) |
656 | fatal(toString(this) + ": empty SHT_GROUP" ); |
657 | |
658 | Elf_Word flag = entries[0]; |
659 | if (flag && flag != GRP_COMDAT) |
660 | fatal(toString(this) + ": unsupported SHT_GROUP format" ); |
661 | |
662 | bool keepGroup = |
663 | (flag & GRP_COMDAT) == 0 || ignoreComdats || |
664 | symtab.comdatGroups.try_emplace(CachedHashStringRef(signature), this) |
665 | .second; |
666 | if (keepGroup) { |
667 | if (config->relocatable) |
668 | this->sections[i] = createInputSection( |
669 | idx: i, sec, name: check(obj.getSectionName(sec, shstrtab))); |
670 | continue; |
671 | } |
672 | |
673 | // Otherwise, discard group members. |
674 | for (uint32_t secIndex : entries.slice(1)) { |
675 | if (secIndex >= size) |
676 | fatal(toString(this) + |
677 | ": invalid section index in group: " + Twine(secIndex)); |
678 | this->sections[secIndex] = &InputSection::discarded; |
679 | } |
680 | } |
681 | |
682 | // Read a symbol table. |
683 | initializeSymbols(obj); |
684 | } |
685 | |
686 | // Sections with SHT_GROUP and comdat bits define comdat section groups. |
687 | // They are identified and deduplicated by group name. This function |
688 | // returns a group name. |
689 | template <class ELFT> |
690 | StringRef ObjFile<ELFT>::getShtGroupSignature(ArrayRef<Elf_Shdr> sections, |
691 | const Elf_Shdr &sec) { |
692 | typename ELFT::SymRange symbols = this->getELFSyms<ELFT>(); |
693 | if (sec.sh_info >= symbols.size()) |
694 | fatal(toString(this) + ": invalid symbol index" ); |
695 | const typename ELFT::Sym &sym = symbols[sec.sh_info]; |
696 | return CHECK(sym.getName(this->stringTable), this); |
697 | } |
698 | |
699 | template <class ELFT> |
700 | bool ObjFile<ELFT>::shouldMerge(const Elf_Shdr &sec, StringRef name) { |
701 | // On a regular link we don't merge sections if -O0 (default is -O1). This |
702 | // sometimes makes the linker significantly faster, although the output will |
703 | // be bigger. |
704 | // |
705 | // Doing the same for -r would create a problem as it would combine sections |
706 | // with different sh_entsize. One option would be to just copy every SHF_MERGE |
707 | // section as is to the output. While this would produce a valid ELF file with |
708 | // usable SHF_MERGE sections, tools like (llvm-)?dwarfdump get confused when |
709 | // they see two .debug_str. We could have separate logic for combining |
710 | // SHF_MERGE sections based both on their name and sh_entsize, but that seems |
711 | // to be more trouble than it is worth. Instead, we just use the regular (-O1) |
712 | // logic for -r. |
713 | if (config->optimize == 0 && !config->relocatable) |
714 | return false; |
715 | |
716 | // A mergeable section with size 0 is useless because they don't have |
717 | // any data to merge. A mergeable string section with size 0 can be |
718 | // argued as invalid because it doesn't end with a null character. |
719 | // We'll avoid a mess by handling them as if they were non-mergeable. |
720 | if (sec.sh_size == 0) |
721 | return false; |
722 | |
723 | // Check for sh_entsize. The ELF spec is not clear about the zero |
724 | // sh_entsize. It says that "the member [sh_entsize] contains 0 if |
725 | // the section does not hold a table of fixed-size entries". We know |
726 | // that Rust 1.13 produces a string mergeable section with a zero |
727 | // sh_entsize. Here we just accept it rather than being picky about it. |
728 | uint64_t entSize = sec.sh_entsize; |
729 | if (entSize == 0) |
730 | return false; |
731 | if (sec.sh_size % entSize) |
732 | fatal(toString(this) + ":(" + name + "): SHF_MERGE section size (" + |
733 | Twine(sec.sh_size) + ") must be a multiple of sh_entsize (" + |
734 | Twine(entSize) + ")" ); |
735 | |
736 | if (sec.sh_flags & SHF_WRITE) |
737 | fatal(toString(this) + ":(" + name + |
738 | "): writable SHF_MERGE section is not supported" ); |
739 | |
740 | return true; |
741 | } |
742 | |
743 | // This is for --just-symbols. |
744 | // |
745 | // --just-symbols is a very minor feature that allows you to link your |
746 | // output against other existing program, so that if you load both your |
747 | // program and the other program into memory, your output can refer the |
748 | // other program's symbols. |
749 | // |
750 | // When the option is given, we link "just symbols". The section table is |
751 | // initialized with null pointers. |
752 | template <class ELFT> void ObjFile<ELFT>::initializeJustSymbols() { |
753 | sections.resize(numELFShdrs); |
754 | } |
755 | |
756 | static bool isKnownSpecificSectionType(uint32_t t, uint32_t flags) { |
757 | if (SHT_LOUSER <= t && t <= SHT_HIUSER && !(flags & SHF_ALLOC)) |
758 | return true; |
759 | if (SHT_LOOS <= t && t <= SHT_HIOS && !(flags & SHF_OS_NONCONFORMING)) |
760 | return true; |
761 | // Allow all processor-specific types. This is different from GNU ld. |
762 | return SHT_LOPROC <= t && t <= SHT_HIPROC; |
763 | } |
764 | |
765 | template <class ELFT> |
766 | void ObjFile<ELFT>::initializeSections(bool ignoreComdats, |
767 | const llvm::object::ELFFile<ELFT> &obj) { |
768 | ArrayRef<Elf_Shdr> objSections = getELFShdrs<ELFT>(); |
769 | StringRef shstrtab = CHECK(obj.getSectionStringTable(objSections), this); |
770 | uint64_t size = objSections.size(); |
771 | SmallVector<ArrayRef<Elf_Word>, 0> selectedGroups; |
772 | for (size_t i = 0; i != size; ++i) { |
773 | if (this->sections[i] == &InputSection::discarded) |
774 | continue; |
775 | const Elf_Shdr &sec = objSections[i]; |
776 | const uint32_t type = sec.sh_type; |
777 | |
778 | // SHF_EXCLUDE'ed sections are discarded by the linker. However, |
779 | // if -r is given, we'll let the final link discard such sections. |
780 | // This is compatible with GNU. |
781 | if ((sec.sh_flags & SHF_EXCLUDE) && !config->relocatable) { |
782 | if (type == SHT_LLVM_CALL_GRAPH_PROFILE) |
783 | cgProfileSectionIndex = i; |
784 | if (type == SHT_LLVM_ADDRSIG) { |
785 | // We ignore the address-significance table if we know that the object |
786 | // file was created by objcopy or ld -r. This is because these tools |
787 | // will reorder the symbols in the symbol table, invalidating the data |
788 | // in the address-significance table, which refers to symbols by index. |
789 | if (sec.sh_link != 0) |
790 | this->addrsigSec = &sec; |
791 | else if (config->icf == ICFLevel::Safe) |
792 | warn(toString(this) + |
793 | ": --icf=safe conservatively ignores " |
794 | "SHT_LLVM_ADDRSIG [index " + |
795 | Twine(i) + |
796 | "] with sh_link=0 " |
797 | "(likely created using objcopy or ld -r)" ); |
798 | } |
799 | this->sections[i] = &InputSection::discarded; |
800 | continue; |
801 | } |
802 | |
803 | switch (type) { |
804 | case SHT_GROUP: { |
805 | if (!config->relocatable) |
806 | sections[i] = &InputSection::discarded; |
807 | StringRef signature = |
808 | cantFail(this->getELFSyms<ELFT>()[sec.sh_info].getName(stringTable)); |
809 | ArrayRef<Elf_Word> entries = |
810 | cantFail(obj.template getSectionContentsAsArray<Elf_Word>(sec)); |
811 | if ((entries[0] & GRP_COMDAT) == 0 || ignoreComdats || |
812 | symtab.comdatGroups.find(Val: CachedHashStringRef(signature))->second == |
813 | this) |
814 | selectedGroups.push_back(entries); |
815 | break; |
816 | } |
817 | case SHT_SYMTAB_SHNDX: |
818 | shndxTable = CHECK(obj.getSHNDXTable(sec, objSections), this); |
819 | break; |
820 | case SHT_SYMTAB: |
821 | case SHT_STRTAB: |
822 | case SHT_REL: |
823 | case SHT_RELA: |
824 | case SHT_NULL: |
825 | break; |
826 | case SHT_PROGBITS: |
827 | case SHT_NOTE: |
828 | case SHT_NOBITS: |
829 | case SHT_INIT_ARRAY: |
830 | case SHT_FINI_ARRAY: |
831 | case SHT_PREINIT_ARRAY: |
832 | this->sections[i] = |
833 | createInputSection(idx: i, sec, name: check(obj.getSectionName(sec, shstrtab))); |
834 | break; |
835 | default: |
836 | this->sections[i] = |
837 | createInputSection(idx: i, sec, name: check(obj.getSectionName(sec, shstrtab))); |
838 | if (type == SHT_LLVM_SYMPART) |
839 | ctx.hasSympart.store(i: true, m: std::memory_order_relaxed); |
840 | else if (config->rejectMismatch && |
841 | !isKnownSpecificSectionType(type, sec.sh_flags)) |
842 | errorOrWarn(toString(this->sections[i]) + ": unknown section type 0x" + |
843 | Twine::utohexstr(Val: type)); |
844 | break; |
845 | } |
846 | } |
847 | |
848 | // We have a second loop. It is used to: |
849 | // 1) handle SHF_LINK_ORDER sections. |
850 | // 2) create relocation sections. In some cases the section header index of a |
851 | // relocation section may be smaller than that of the relocated section. In |
852 | // such cases, the relocation section would attempt to reference a target |
853 | // section that has not yet been created. For simplicity, delay creation of |
854 | // relocation sections until now. |
855 | for (size_t i = 0; i != size; ++i) { |
856 | if (this->sections[i] == &InputSection::discarded) |
857 | continue; |
858 | const Elf_Shdr &sec = objSections[i]; |
859 | |
860 | if (isStaticRelSecType(sec.sh_type)) { |
861 | // Find a relocation target section and associate this section with that. |
862 | // Target may have been discarded if it is in a different section group |
863 | // and the group is discarded, even though it's a violation of the spec. |
864 | // We handle that situation gracefully by discarding dangling relocation |
865 | // sections. |
866 | const uint32_t info = sec.sh_info; |
867 | InputSectionBase *s = getRelocTarget(idx: i, sec, info); |
868 | if (!s) |
869 | continue; |
870 | |
871 | // ELF spec allows mergeable sections with relocations, but they are rare, |
872 | // and it is in practice hard to merge such sections by contents, because |
873 | // applying relocations at end of linking changes section contents. So, we |
874 | // simply handle such sections as non-mergeable ones. Degrading like this |
875 | // is acceptable because section merging is optional. |
876 | if (auto *ms = dyn_cast<MergeInputSection>(Val: s)) { |
877 | s = makeThreadLocal<InputSection>( |
878 | args&: ms->file, args&: ms->flags, args&: ms->type, args&: ms->addralign, |
879 | args: ms->contentMaybeDecompress(), args&: ms->name); |
880 | sections[info] = s; |
881 | } |
882 | |
883 | if (s->relSecIdx != 0) |
884 | error( |
885 | msg: toString(s) + |
886 | ": multiple relocation sections to one section are not supported" ); |
887 | s->relSecIdx = i; |
888 | |
889 | // Relocation sections are usually removed from the output, so return |
890 | // `nullptr` for the normal case. However, if -r or --emit-relocs is |
891 | // specified, we need to copy them to the output. (Some post link analysis |
892 | // tools specify --emit-relocs to obtain the information.) |
893 | if (config->copyRelocs) { |
894 | auto *isec = makeThreadLocal<InputSection>( |
895 | *this, sec, check(obj.getSectionName(sec, shstrtab))); |
896 | // If the relocated section is discarded (due to /DISCARD/ or |
897 | // --gc-sections), the relocation section should be discarded as well. |
898 | s->dependentSections.push_back(NewVal: isec); |
899 | sections[i] = isec; |
900 | } |
901 | continue; |
902 | } |
903 | |
904 | // A SHF_LINK_ORDER section with sh_link=0 is handled as if it did not have |
905 | // the flag. |
906 | if (!sec.sh_link || !(sec.sh_flags & SHF_LINK_ORDER)) |
907 | continue; |
908 | |
909 | InputSectionBase *linkSec = nullptr; |
910 | if (sec.sh_link < size) |
911 | linkSec = this->sections[sec.sh_link]; |
912 | if (!linkSec) |
913 | fatal(toString(this) + ": invalid sh_link index: " + Twine(sec.sh_link)); |
914 | |
915 | // A SHF_LINK_ORDER section is discarded if its linked-to section is |
916 | // discarded. |
917 | InputSection *isec = cast<InputSection>(this->sections[i]); |
918 | linkSec->dependentSections.push_back(NewVal: isec); |
919 | if (!isa<InputSection>(Val: linkSec)) |
920 | error(msg: "a section " + isec->name + |
921 | " with SHF_LINK_ORDER should not refer a non-regular section: " + |
922 | toString(linkSec)); |
923 | } |
924 | |
925 | for (ArrayRef<Elf_Word> entries : selectedGroups) |
926 | handleSectionGroup<ELFT>(this->sections, entries); |
927 | } |
928 | |
929 | // Read the following info from the .note.gnu.property section and write it to |
930 | // the corresponding fields in `ObjFile`: |
931 | // - Feature flags (32 bits) representing x86 or AArch64 features for |
932 | // hardware-assisted call flow control; |
933 | // - AArch64 PAuth ABI core info (16 bytes). |
934 | template <class ELFT> |
935 | void readGnuProperty(const InputSection &sec, ObjFile<ELFT> &f) { |
936 | using Elf_Nhdr = typename ELFT::Nhdr; |
937 | using Elf_Note = typename ELFT::Note; |
938 | |
939 | ArrayRef<uint8_t> data = sec.content(); |
940 | auto reportFatal = [&](const uint8_t *place, const Twine &msg) { |
941 | fatal(msg: toString(f: sec.file) + ":(" + sec.name + "+0x" + |
942 | Twine::utohexstr(Val: place - sec.content().data()) + "): " + msg); |
943 | }; |
944 | while (!data.empty()) { |
945 | // Read one NOTE record. |
946 | auto *nhdr = reinterpret_cast<const Elf_Nhdr *>(data.data()); |
947 | if (data.size() < sizeof(Elf_Nhdr) || |
948 | data.size() < nhdr->getSize(sec.addralign)) |
949 | reportFatal(data.data(), "data is too short" ); |
950 | |
951 | Elf_Note note(*nhdr); |
952 | if (nhdr->n_type != NT_GNU_PROPERTY_TYPE_0 || note.getName() != "GNU" ) { |
953 | data = data.slice(nhdr->getSize(sec.addralign)); |
954 | continue; |
955 | } |
956 | |
957 | uint32_t featureAndType = config->emachine == EM_AARCH64 |
958 | ? GNU_PROPERTY_AARCH64_FEATURE_1_AND |
959 | : GNU_PROPERTY_X86_FEATURE_1_AND; |
960 | |
961 | // Read a body of a NOTE record, which consists of type-length-value fields. |
962 | ArrayRef<uint8_t> desc = note.getDesc(sec.addralign); |
963 | while (!desc.empty()) { |
964 | const uint8_t *place = desc.data(); |
965 | if (desc.size() < 8) |
966 | reportFatal(place, "program property is too short" ); |
967 | uint32_t type = read32<ELFT::Endianness>(desc.data()); |
968 | uint32_t size = read32<ELFT::Endianness>(desc.data() + 4); |
969 | desc = desc.slice(N: 8); |
970 | if (desc.size() < size) |
971 | reportFatal(place, "program property is too short" ); |
972 | |
973 | if (type == featureAndType) { |
974 | // We found a FEATURE_1_AND field. There may be more than one of these |
975 | // in a .note.gnu.property section, for a relocatable object we |
976 | // accumulate the bits set. |
977 | if (size < 4) |
978 | reportFatal(place, "FEATURE_1_AND entry is too short" ); |
979 | f.andFeatures |= read32<ELFT::Endianness>(desc.data()); |
980 | } else if (config->emachine == EM_AARCH64 && |
981 | type == GNU_PROPERTY_AARCH64_FEATURE_PAUTH) { |
982 | if (!f.aarch64PauthAbiCoreInfo.empty()) { |
983 | reportFatal(data.data(), |
984 | "multiple GNU_PROPERTY_AARCH64_FEATURE_PAUTH entries are " |
985 | "not supported" ); |
986 | } else if (size != 16) { |
987 | reportFatal(data.data(), "GNU_PROPERTY_AARCH64_FEATURE_PAUTH entry " |
988 | "is invalid: expected 16 bytes, but got " + |
989 | Twine(size)); |
990 | } |
991 | f.aarch64PauthAbiCoreInfo = desc; |
992 | } |
993 | |
994 | // Padding is present in the note descriptor, if necessary. |
995 | desc = desc.slice(alignTo<(ELFT::Is64Bits ? 8 : 4)>(size)); |
996 | } |
997 | |
998 | // Go to next NOTE record to look for more FEATURE_1_AND descriptions. |
999 | data = data.slice(nhdr->getSize(sec.addralign)); |
1000 | } |
1001 | } |
1002 | |
1003 | template <class ELFT> |
1004 | InputSectionBase *ObjFile<ELFT>::getRelocTarget(uint32_t idx, |
1005 | const Elf_Shdr &sec, |
1006 | uint32_t info) { |
1007 | if (info < this->sections.size()) { |
1008 | InputSectionBase *target = this->sections[info]; |
1009 | |
1010 | // Strictly speaking, a relocation section must be included in the |
1011 | // group of the section it relocates. However, LLVM 3.3 and earlier |
1012 | // would fail to do so, so we gracefully handle that case. |
1013 | if (target == &InputSection::discarded) |
1014 | return nullptr; |
1015 | |
1016 | if (target != nullptr) |
1017 | return target; |
1018 | } |
1019 | |
1020 | error(toString(this) + Twine(": relocation section (index " ) + Twine(idx) + |
1021 | ") has invalid sh_info (" + Twine(info) + ")" ); |
1022 | return nullptr; |
1023 | } |
1024 | |
1025 | // The function may be called concurrently for different input files. For |
1026 | // allocation, prefer makeThreadLocal which does not require holding a lock. |
1027 | template <class ELFT> |
1028 | InputSectionBase *ObjFile<ELFT>::createInputSection(uint32_t idx, |
1029 | const Elf_Shdr &sec, |
1030 | StringRef name) { |
1031 | if (name.starts_with(Prefix: ".n" )) { |
1032 | // The GNU linker uses .note.GNU-stack section as a marker indicating |
1033 | // that the code in the object file does not expect that the stack is |
1034 | // executable (in terms of NX bit). If all input files have the marker, |
1035 | // the GNU linker adds a PT_GNU_STACK segment to tells the loader to |
1036 | // make the stack non-executable. Most object files have this section as |
1037 | // of 2017. |
1038 | // |
1039 | // But making the stack non-executable is a norm today for security |
1040 | // reasons. Failure to do so may result in a serious security issue. |
1041 | // Therefore, we make LLD always add PT_GNU_STACK unless it is |
1042 | // explicitly told to do otherwise (by -z execstack). Because the stack |
1043 | // executable-ness is controlled solely by command line options, |
1044 | // .note.GNU-stack sections are simply ignored. |
1045 | if (name == ".note.GNU-stack" ) |
1046 | return &InputSection::discarded; |
1047 | |
1048 | // Object files that use processor features such as Intel Control-Flow |
1049 | // Enforcement (CET) or AArch64 Branch Target Identification BTI, use a |
1050 | // .note.gnu.property section containing a bitfield of feature bits like the |
1051 | // GNU_PROPERTY_X86_FEATURE_1_IBT flag. Read a bitmap containing the flag. |
1052 | // |
1053 | // Since we merge bitmaps from multiple object files to create a new |
1054 | // .note.gnu.property containing a single AND'ed bitmap, we discard an input |
1055 | // file's .note.gnu.property section. |
1056 | if (name == ".note.gnu.property" ) { |
1057 | readGnuProperty<ELFT>(InputSection(*this, sec, name), *this); |
1058 | return &InputSection::discarded; |
1059 | } |
1060 | |
1061 | // Split stacks is a feature to support a discontiguous stack, |
1062 | // commonly used in the programming language Go. For the details, |
1063 | // see https://gcc.gnu.org/wiki/SplitStacks. An object file compiled |
1064 | // for split stack will include a .note.GNU-split-stack section. |
1065 | if (name == ".note.GNU-split-stack" ) { |
1066 | if (config->relocatable) { |
1067 | error( |
1068 | msg: "cannot mix split-stack and non-split-stack in a relocatable link" ); |
1069 | return &InputSection::discarded; |
1070 | } |
1071 | this->splitStack = true; |
1072 | return &InputSection::discarded; |
1073 | } |
1074 | |
1075 | // An object file compiled for split stack, but where some of the |
1076 | // functions were compiled with the no_split_stack_attribute will |
1077 | // include a .note.GNU-no-split-stack section. |
1078 | if (name == ".note.GNU-no-split-stack" ) { |
1079 | this->someNoSplitStack = true; |
1080 | return &InputSection::discarded; |
1081 | } |
1082 | |
1083 | // Strip existing .note.gnu.build-id sections so that the output won't have |
1084 | // more than one build-id. This is not usually a problem because input |
1085 | // object files normally don't have .build-id sections, but you can create |
1086 | // such files by "ld.{bfd,gold,lld} -r --build-id", and we want to guard |
1087 | // against it. |
1088 | if (name == ".note.gnu.build-id" ) |
1089 | return &InputSection::discarded; |
1090 | } |
1091 | |
1092 | // The linker merges EH (exception handling) frames and creates a |
1093 | // .eh_frame_hdr section for runtime. So we handle them with a special |
1094 | // class. For relocatable outputs, they are just passed through. |
1095 | if (name == ".eh_frame" && !config->relocatable) |
1096 | return makeThreadLocal<EhInputSection>(*this, sec, name); |
1097 | |
1098 | if ((sec.sh_flags & SHF_MERGE) && shouldMerge(sec, name)) |
1099 | return makeThreadLocal<MergeInputSection>(*this, sec, name); |
1100 | return makeThreadLocal<InputSection>(*this, sec, name); |
1101 | } |
1102 | |
1103 | // Initialize symbols. symbols is a parallel array to the corresponding ELF |
1104 | // symbol table. |
1105 | template <class ELFT> |
1106 | void ObjFile<ELFT>::initializeSymbols(const object::ELFFile<ELFT> &obj) { |
1107 | ArrayRef<Elf_Sym> eSyms = this->getELFSyms<ELFT>(); |
1108 | if (numSymbols == 0) { |
1109 | numSymbols = eSyms.size(); |
1110 | symbols = std::make_unique<Symbol *[]>(numSymbols); |
1111 | } |
1112 | |
1113 | // Some entries have been filled by LazyObjFile. |
1114 | for (size_t i = firstGlobal, end = eSyms.size(); i != end; ++i) |
1115 | if (!symbols[i]) |
1116 | symbols[i] = symtab.insert(CHECK(eSyms[i].getName(stringTable), this)); |
1117 | |
1118 | // Perform symbol resolution on non-local symbols. |
1119 | SmallVector<unsigned, 32> undefineds; |
1120 | for (size_t i = firstGlobal, end = eSyms.size(); i != end; ++i) { |
1121 | const Elf_Sym &eSym = eSyms[i]; |
1122 | uint32_t secIdx = eSym.st_shndx; |
1123 | if (secIdx == SHN_UNDEF) { |
1124 | undefineds.push_back(Elt: i); |
1125 | continue; |
1126 | } |
1127 | |
1128 | uint8_t binding = eSym.getBinding(); |
1129 | uint8_t stOther = eSym.st_other; |
1130 | uint8_t type = eSym.getType(); |
1131 | uint64_t value = eSym.st_value; |
1132 | uint64_t size = eSym.st_size; |
1133 | |
1134 | Symbol *sym = symbols[i]; |
1135 | sym->isUsedInRegularObj = true; |
1136 | if (LLVM_UNLIKELY(eSym.st_shndx == SHN_COMMON)) { |
1137 | if (value == 0 || value >= UINT32_MAX) |
1138 | fatal(toString(this) + ": common symbol '" + sym->getName() + |
1139 | "' has invalid alignment: " + Twine(value)); |
1140 | hasCommonSyms = true; |
1141 | sym->resolve( |
1142 | other: CommonSymbol{this, StringRef(), binding, stOther, type, value, size}); |
1143 | continue; |
1144 | } |
1145 | |
1146 | // Handle global defined symbols. Defined::section will be set in postParse. |
1147 | sym->resolve(other: Defined{this, StringRef(), binding, stOther, type, value, size, |
1148 | nullptr}); |
1149 | } |
1150 | |
1151 | // Undefined symbols (excluding those defined relative to non-prevailing |
1152 | // sections) can trigger recursive extract. Process defined symbols first so |
1153 | // that the relative order between a defined symbol and an undefined symbol |
1154 | // does not change the symbol resolution behavior. In addition, a set of |
1155 | // interconnected symbols will all be resolved to the same file, instead of |
1156 | // being resolved to different files. |
1157 | for (unsigned i : undefineds) { |
1158 | const Elf_Sym &eSym = eSyms[i]; |
1159 | Symbol *sym = symbols[i]; |
1160 | sym->resolve(other: Undefined{this, StringRef(), eSym.getBinding(), eSym.st_other, |
1161 | eSym.getType()}); |
1162 | sym->isUsedInRegularObj = true; |
1163 | sym->referenced = true; |
1164 | } |
1165 | } |
1166 | |
1167 | template <class ELFT> |
1168 | void ObjFile<ELFT>::initSectionsAndLocalSyms(bool ignoreComdats) { |
1169 | if (!justSymbols) |
1170 | initializeSections(ignoreComdats, obj: getObj()); |
1171 | |
1172 | if (!firstGlobal) |
1173 | return; |
1174 | SymbolUnion *locals = makeThreadLocalN<SymbolUnion>(firstGlobal); |
1175 | memset(locals, 0, sizeof(SymbolUnion) * firstGlobal); |
1176 | |
1177 | ArrayRef<Elf_Sym> eSyms = this->getELFSyms<ELFT>(); |
1178 | for (size_t i = 0, end = firstGlobal; i != end; ++i) { |
1179 | const Elf_Sym &eSym = eSyms[i]; |
1180 | uint32_t secIdx = eSym.st_shndx; |
1181 | if (LLVM_UNLIKELY(secIdx == SHN_XINDEX)) |
1182 | secIdx = check(getExtendedSymbolTableIndex<ELFT>(eSym, i, shndxTable)); |
1183 | else if (secIdx >= SHN_LORESERVE) |
1184 | secIdx = 0; |
1185 | if (LLVM_UNLIKELY(secIdx >= sections.size())) |
1186 | fatal(toString(this) + ": invalid section index: " + Twine(secIdx)); |
1187 | if (LLVM_UNLIKELY(eSym.getBinding() != STB_LOCAL)) |
1188 | error(toString(this) + ": non-local symbol (" + Twine(i) + |
1189 | ") found at index < .symtab's sh_info (" + Twine(end) + ")" ); |
1190 | |
1191 | InputSectionBase *sec = sections[secIdx]; |
1192 | uint8_t type = eSym.getType(); |
1193 | if (type == STT_FILE) |
1194 | sourceFile = CHECK(eSym.getName(stringTable), this); |
1195 | if (LLVM_UNLIKELY(stringTable.size() <= eSym.st_name)) |
1196 | fatal(toString(this) + ": invalid symbol name offset" ); |
1197 | StringRef name(stringTable.data() + eSym.st_name); |
1198 | |
1199 | symbols[i] = reinterpret_cast<Symbol *>(locals + i); |
1200 | if (eSym.st_shndx == SHN_UNDEF || sec == &InputSection::discarded) |
1201 | new (symbols[i]) Undefined(this, name, STB_LOCAL, eSym.st_other, type, |
1202 | /*discardedSecIdx=*/secIdx); |
1203 | else |
1204 | new (symbols[i]) Defined(this, name, STB_LOCAL, eSym.st_other, type, |
1205 | eSym.st_value, eSym.st_size, sec); |
1206 | symbols[i]->partition = 1; |
1207 | symbols[i]->isUsedInRegularObj = true; |
1208 | } |
1209 | } |
1210 | |
1211 | // Called after all ObjFile::parse is called for all ObjFiles. This checks |
1212 | // duplicate symbols and may do symbol property merge in the future. |
1213 | template <class ELFT> void ObjFile<ELFT>::postParse() { |
1214 | static std::mutex mu; |
1215 | ArrayRef<Elf_Sym> eSyms = this->getELFSyms<ELFT>(); |
1216 | for (size_t i = firstGlobal, end = eSyms.size(); i != end; ++i) { |
1217 | const Elf_Sym &eSym = eSyms[i]; |
1218 | Symbol &sym = *symbols[i]; |
1219 | uint32_t secIdx = eSym.st_shndx; |
1220 | uint8_t binding = eSym.getBinding(); |
1221 | if (LLVM_UNLIKELY(binding != STB_GLOBAL && binding != STB_WEAK && |
1222 | binding != STB_GNU_UNIQUE)) |
1223 | errorOrWarn(toString(this) + ": symbol (" + Twine(i) + |
1224 | ") has invalid binding: " + Twine((int)binding)); |
1225 | |
1226 | // st_value of STT_TLS represents the assigned offset, not the actual |
1227 | // address which is used by STT_FUNC and STT_OBJECT. STT_TLS symbols can |
1228 | // only be referenced by special TLS relocations. It is usually an error if |
1229 | // a STT_TLS symbol is replaced by a non-STT_TLS symbol, vice versa. |
1230 | if (LLVM_UNLIKELY(sym.isTls()) && eSym.getType() != STT_TLS && |
1231 | eSym.getType() != STT_NOTYPE) |
1232 | errorOrWarn("TLS attribute mismatch: " + toString(sym) + "\n>>> in " + |
1233 | toString(f: sym.file) + "\n>>> in " + toString(this)); |
1234 | |
1235 | // Handle non-COMMON defined symbol below. !sym.file allows a symbol |
1236 | // assignment to redefine a symbol without an error. |
1237 | if (!sym.file || !sym.isDefined() || secIdx == SHN_UNDEF || |
1238 | secIdx == SHN_COMMON) |
1239 | continue; |
1240 | |
1241 | if (LLVM_UNLIKELY(secIdx == SHN_XINDEX)) |
1242 | secIdx = check(getExtendedSymbolTableIndex<ELFT>(eSym, i, shndxTable)); |
1243 | else if (secIdx >= SHN_LORESERVE) |
1244 | secIdx = 0; |
1245 | if (LLVM_UNLIKELY(secIdx >= sections.size())) |
1246 | fatal(toString(this) + ": invalid section index: " + Twine(secIdx)); |
1247 | InputSectionBase *sec = sections[secIdx]; |
1248 | if (sec == &InputSection::discarded) { |
1249 | if (sym.traced) { |
1250 | printTraceSymbol(sym: Undefined{this, sym.getName(), sym.binding, |
1251 | sym.stOther, sym.type, secIdx}, |
1252 | name: sym.getName()); |
1253 | } |
1254 | if (sym.file == this) { |
1255 | std::lock_guard<std::mutex> lock(mu); |
1256 | ctx.nonPrevailingSyms.emplace_back(Args: &sym, Args&: secIdx); |
1257 | } |
1258 | continue; |
1259 | } |
1260 | |
1261 | if (sym.file == this) { |
1262 | cast<Defined>(Val&: sym).section = sec; |
1263 | continue; |
1264 | } |
1265 | |
1266 | if (sym.binding == STB_WEAK || binding == STB_WEAK) |
1267 | continue; |
1268 | std::lock_guard<std::mutex> lock(mu); |
1269 | ctx.duplicates.push_back(Elt: {&sym, this, sec, eSym.st_value}); |
1270 | } |
1271 | } |
1272 | |
1273 | // The handling of tentative definitions (COMMON symbols) in archives is murky. |
1274 | // A tentative definition will be promoted to a global definition if there are |
1275 | // no non-tentative definitions to dominate it. When we hold a tentative |
1276 | // definition to a symbol and are inspecting archive members for inclusion |
1277 | // there are 2 ways we can proceed: |
1278 | // |
1279 | // 1) Consider the tentative definition a 'real' definition (ie promotion from |
1280 | // tentative to real definition has already happened) and not inspect |
1281 | // archive members for Global/Weak definitions to replace the tentative |
1282 | // definition. An archive member would only be included if it satisfies some |
1283 | // other undefined symbol. This is the behavior Gold uses. |
1284 | // |
1285 | // 2) Consider the tentative definition as still undefined (ie the promotion to |
1286 | // a real definition happens only after all symbol resolution is done). |
1287 | // The linker searches archive members for STB_GLOBAL definitions to |
1288 | // replace the tentative definition with. This is the behavior used by |
1289 | // GNU ld. |
1290 | // |
1291 | // The second behavior is inherited from SysVR4, which based it on the FORTRAN |
1292 | // COMMON BLOCK model. This behavior is needed for proper initialization in old |
1293 | // (pre F90) FORTRAN code that is packaged into an archive. |
1294 | // |
1295 | // The following functions search archive members for definitions to replace |
1296 | // tentative definitions (implementing behavior 2). |
1297 | static bool isBitcodeNonCommonDef(MemoryBufferRef mb, StringRef symName, |
1298 | StringRef archiveName) { |
1299 | IRSymtabFile symtabFile = check(e: readIRSymtab(MBRef: mb)); |
1300 | for (const irsymtab::Reader::SymbolRef &sym : |
1301 | symtabFile.TheReader.symbols()) { |
1302 | if (sym.isGlobal() && sym.getName() == symName) |
1303 | return !sym.isUndefined() && !sym.isWeak() && !sym.isCommon(); |
1304 | } |
1305 | return false; |
1306 | } |
1307 | |
1308 | template <class ELFT> |
1309 | static bool isNonCommonDef(ELFKind ekind, MemoryBufferRef mb, StringRef symName, |
1310 | StringRef archiveName) { |
1311 | ObjFile<ELFT> *obj = make<ObjFile<ELFT>>(ekind, mb, archiveName); |
1312 | obj->init(); |
1313 | StringRef stringtable = obj->getStringTable(); |
1314 | |
1315 | for (auto sym : obj->template getGlobalELFSyms<ELFT>()) { |
1316 | Expected<StringRef> name = sym.getName(stringtable); |
1317 | if (name && name.get() == symName) |
1318 | return sym.isDefined() && sym.getBinding() == STB_GLOBAL && |
1319 | !sym.isCommon(); |
1320 | } |
1321 | return false; |
1322 | } |
1323 | |
1324 | static bool isNonCommonDef(MemoryBufferRef mb, StringRef symName, |
1325 | StringRef archiveName) { |
1326 | switch (getELFKind(mb, archiveName)) { |
1327 | case ELF32LEKind: |
1328 | return isNonCommonDef<ELF32LE>(ekind: ELF32LEKind, mb, symName, archiveName); |
1329 | case ELF32BEKind: |
1330 | return isNonCommonDef<ELF32BE>(ekind: ELF32BEKind, mb, symName, archiveName); |
1331 | case ELF64LEKind: |
1332 | return isNonCommonDef<ELF64LE>(ekind: ELF64LEKind, mb, symName, archiveName); |
1333 | case ELF64BEKind: |
1334 | return isNonCommonDef<ELF64BE>(ekind: ELF64BEKind, mb, symName, archiveName); |
1335 | default: |
1336 | llvm_unreachable("getELFKind" ); |
1337 | } |
1338 | } |
1339 | |
1340 | unsigned SharedFile::vernauxNum; |
1341 | |
1342 | SharedFile::SharedFile(MemoryBufferRef m, StringRef defaultSoName) |
1343 | : ELFFileBase(SharedKind, getELFKind(mb: m, archiveName: "" ), m), soName(defaultSoName), |
1344 | isNeeded(!config->asNeeded) {} |
1345 | |
1346 | // Parse the version definitions in the object file if present, and return a |
1347 | // vector whose nth element contains a pointer to the Elf_Verdef for version |
1348 | // identifier n. Version identifiers that are not definitions map to nullptr. |
1349 | template <typename ELFT> |
1350 | static SmallVector<const void *, 0> |
1351 | parseVerdefs(const uint8_t *base, const typename ELFT::Shdr *sec) { |
1352 | if (!sec) |
1353 | return {}; |
1354 | |
1355 | // Build the Verdefs array by following the chain of Elf_Verdef objects |
1356 | // from the start of the .gnu.version_d section. |
1357 | SmallVector<const void *, 0> verdefs; |
1358 | const uint8_t *verdef = base + sec->sh_offset; |
1359 | for (unsigned i = 0, e = sec->sh_info; i != e; ++i) { |
1360 | auto *curVerdef = reinterpret_cast<const typename ELFT::Verdef *>(verdef); |
1361 | verdef += curVerdef->vd_next; |
1362 | unsigned verdefIndex = curVerdef->vd_ndx; |
1363 | if (verdefIndex >= verdefs.size()) |
1364 | verdefs.resize(N: verdefIndex + 1); |
1365 | verdefs[verdefIndex] = curVerdef; |
1366 | } |
1367 | return verdefs; |
1368 | } |
1369 | |
1370 | // Parse SHT_GNU_verneed to properly set the name of a versioned undefined |
1371 | // symbol. We detect fatal issues which would cause vulnerabilities, but do not |
1372 | // implement sophisticated error checking like in llvm-readobj because the value |
1373 | // of such diagnostics is low. |
1374 | template <typename ELFT> |
1375 | std::vector<uint32_t> SharedFile::parseVerneed(const ELFFile<ELFT> &obj, |
1376 | const typename ELFT::Shdr *sec) { |
1377 | if (!sec) |
1378 | return {}; |
1379 | std::vector<uint32_t> verneeds; |
1380 | ArrayRef<uint8_t> data = CHECK(obj.getSectionContents(*sec), this); |
1381 | const uint8_t *verneedBuf = data.begin(); |
1382 | for (unsigned i = 0; i != sec->sh_info; ++i) { |
1383 | if (verneedBuf + sizeof(typename ELFT::Verneed) > data.end()) |
1384 | fatal(msg: toString(f: this) + " has an invalid Verneed" ); |
1385 | auto *vn = reinterpret_cast<const typename ELFT::Verneed *>(verneedBuf); |
1386 | const uint8_t *vernauxBuf = verneedBuf + vn->vn_aux; |
1387 | for (unsigned j = 0; j != vn->vn_cnt; ++j) { |
1388 | if (vernauxBuf + sizeof(typename ELFT::Vernaux) > data.end()) |
1389 | fatal(msg: toString(f: this) + " has an invalid Vernaux" ); |
1390 | auto *aux = reinterpret_cast<const typename ELFT::Vernaux *>(vernauxBuf); |
1391 | if (aux->vna_name >= this->stringTable.size()) |
1392 | fatal(msg: toString(f: this) + " has a Vernaux with an invalid vna_name" ); |
1393 | uint16_t version = aux->vna_other & VERSYM_VERSION; |
1394 | if (version >= verneeds.size()) |
1395 | verneeds.resize(new_size: version + 1); |
1396 | verneeds[version] = aux->vna_name; |
1397 | vernauxBuf += aux->vna_next; |
1398 | } |
1399 | verneedBuf += vn->vn_next; |
1400 | } |
1401 | return verneeds; |
1402 | } |
1403 | |
1404 | // We do not usually care about alignments of data in shared object |
1405 | // files because the loader takes care of it. However, if we promote a |
1406 | // DSO symbol to point to .bss due to copy relocation, we need to keep |
1407 | // the original alignment requirements. We infer it in this function. |
1408 | template <typename ELFT> |
1409 | static uint64_t getAlignment(ArrayRef<typename ELFT::Shdr> sections, |
1410 | const typename ELFT::Sym &sym) { |
1411 | uint64_t ret = UINT64_MAX; |
1412 | if (sym.st_value) |
1413 | ret = 1ULL << llvm::countr_zero(Val: (uint64_t)sym.st_value); |
1414 | if (0 < sym.st_shndx && sym.st_shndx < sections.size()) |
1415 | ret = std::min<uint64_t>(ret, sections[sym.st_shndx].sh_addralign); |
1416 | return (ret > UINT32_MAX) ? 0 : ret; |
1417 | } |
1418 | |
1419 | // Fully parse the shared object file. |
1420 | // |
1421 | // This function parses symbol versions. If a DSO has version information, |
1422 | // the file has a ".gnu.version_d" section which contains symbol version |
1423 | // definitions. Each symbol is associated to one version through a table in |
1424 | // ".gnu.version" section. That table is a parallel array for the symbol |
1425 | // table, and each table entry contains an index in ".gnu.version_d". |
1426 | // |
1427 | // The special index 0 is reserved for VERF_NDX_LOCAL and 1 is for |
1428 | // VER_NDX_GLOBAL. There's no table entry for these special versions in |
1429 | // ".gnu.version_d". |
1430 | // |
1431 | // The file format for symbol versioning is perhaps a bit more complicated |
1432 | // than necessary, but you can easily understand the code if you wrap your |
1433 | // head around the data structure described above. |
1434 | template <class ELFT> void SharedFile::parse() { |
1435 | using Elf_Dyn = typename ELFT::Dyn; |
1436 | using Elf_Shdr = typename ELFT::Shdr; |
1437 | using Elf_Sym = typename ELFT::Sym; |
1438 | using Elf_Verdef = typename ELFT::Verdef; |
1439 | using Elf_Versym = typename ELFT::Versym; |
1440 | |
1441 | ArrayRef<Elf_Dyn> dynamicTags; |
1442 | const ELFFile<ELFT> obj = this->getObj<ELFT>(); |
1443 | ArrayRef<Elf_Shdr> sections = getELFShdrs<ELFT>(); |
1444 | |
1445 | const Elf_Shdr *versymSec = nullptr; |
1446 | const Elf_Shdr *verdefSec = nullptr; |
1447 | const Elf_Shdr *verneedSec = nullptr; |
1448 | |
1449 | // Search for .dynsym, .dynamic, .symtab, .gnu.version and .gnu.version_d. |
1450 | for (const Elf_Shdr &sec : sections) { |
1451 | switch (sec.sh_type) { |
1452 | default: |
1453 | continue; |
1454 | case SHT_DYNAMIC: |
1455 | dynamicTags = |
1456 | CHECK(obj.template getSectionContentsAsArray<Elf_Dyn>(sec), this); |
1457 | break; |
1458 | case SHT_GNU_versym: |
1459 | versymSec = &sec; |
1460 | break; |
1461 | case SHT_GNU_verdef: |
1462 | verdefSec = &sec; |
1463 | break; |
1464 | case SHT_GNU_verneed: |
1465 | verneedSec = &sec; |
1466 | break; |
1467 | } |
1468 | } |
1469 | |
1470 | if (versymSec && numELFSyms == 0) { |
1471 | error(msg: "SHT_GNU_versym should be associated with symbol table" ); |
1472 | return; |
1473 | } |
1474 | |
1475 | // Search for a DT_SONAME tag to initialize this->soName. |
1476 | for (const Elf_Dyn &dyn : dynamicTags) { |
1477 | if (dyn.d_tag == DT_NEEDED) { |
1478 | uint64_t val = dyn.getVal(); |
1479 | if (val >= this->stringTable.size()) |
1480 | fatal(msg: toString(f: this) + ": invalid DT_NEEDED entry" ); |
1481 | dtNeeded.push_back(Elt: this->stringTable.data() + val); |
1482 | } else if (dyn.d_tag == DT_SONAME) { |
1483 | uint64_t val = dyn.getVal(); |
1484 | if (val >= this->stringTable.size()) |
1485 | fatal(msg: toString(f: this) + ": invalid DT_SONAME entry" ); |
1486 | soName = this->stringTable.data() + val; |
1487 | } |
1488 | } |
1489 | |
1490 | // DSOs are uniquified not by filename but by soname. |
1491 | DenseMap<CachedHashStringRef, SharedFile *>::iterator it; |
1492 | bool wasInserted; |
1493 | std::tie(args&: it, args&: wasInserted) = |
1494 | symtab.soNames.try_emplace(Key: CachedHashStringRef(soName), Args: this); |
1495 | |
1496 | // If a DSO appears more than once on the command line with and without |
1497 | // --as-needed, --no-as-needed takes precedence over --as-needed because a |
1498 | // user can add an extra DSO with --no-as-needed to force it to be added to |
1499 | // the dependency list. |
1500 | it->second->isNeeded |= isNeeded; |
1501 | if (!wasInserted) |
1502 | return; |
1503 | |
1504 | ctx.sharedFiles.push_back(Elt: this); |
1505 | |
1506 | verdefs = parseVerdefs<ELFT>(obj.base(), verdefSec); |
1507 | std::vector<uint32_t> verneeds = parseVerneed<ELFT>(obj, verneedSec); |
1508 | |
1509 | // Parse ".gnu.version" section which is a parallel array for the symbol |
1510 | // table. If a given file doesn't have a ".gnu.version" section, we use |
1511 | // VER_NDX_GLOBAL. |
1512 | size_t size = numELFSyms - firstGlobal; |
1513 | std::vector<uint16_t> versyms(size, VER_NDX_GLOBAL); |
1514 | if (versymSec) { |
1515 | ArrayRef<Elf_Versym> versym = |
1516 | CHECK(obj.template getSectionContentsAsArray<Elf_Versym>(*versymSec), |
1517 | this) |
1518 | .slice(firstGlobal); |
1519 | for (size_t i = 0; i < size; ++i) |
1520 | versyms[i] = versym[i].vs_index; |
1521 | } |
1522 | |
1523 | // System libraries can have a lot of symbols with versions. Using a |
1524 | // fixed buffer for computing the versions name (foo@ver) can save a |
1525 | // lot of allocations. |
1526 | SmallString<0> versionedNameBuffer; |
1527 | |
1528 | // Add symbols to the symbol table. |
1529 | ArrayRef<Elf_Sym> syms = this->getGlobalELFSyms<ELFT>(); |
1530 | for (size_t i = 0, e = syms.size(); i != e; ++i) { |
1531 | const Elf_Sym &sym = syms[i]; |
1532 | |
1533 | // ELF spec requires that all local symbols precede weak or global |
1534 | // symbols in each symbol table, and the index of first non-local symbol |
1535 | // is stored to sh_info. If a local symbol appears after some non-local |
1536 | // symbol, that's a violation of the spec. |
1537 | StringRef name = CHECK(sym.getName(stringTable), this); |
1538 | if (sym.getBinding() == STB_LOCAL) { |
1539 | errorOrWarn(msg: toString(f: this) + ": invalid local symbol '" + name + |
1540 | "' in global part of symbol table" ); |
1541 | continue; |
1542 | } |
1543 | |
1544 | const uint16_t ver = versyms[i], idx = ver & ~VERSYM_HIDDEN; |
1545 | if (sym.isUndefined()) { |
1546 | // For unversioned undefined symbols, VER_NDX_GLOBAL makes more sense but |
1547 | // as of binutils 2.34, GNU ld produces VER_NDX_LOCAL. |
1548 | if (ver != VER_NDX_LOCAL && ver != VER_NDX_GLOBAL) { |
1549 | if (idx >= verneeds.size()) { |
1550 | error(msg: "corrupt input file: version need index " + Twine(idx) + |
1551 | " for symbol " + name + " is out of bounds\n>>> defined in " + |
1552 | toString(f: this)); |
1553 | continue; |
1554 | } |
1555 | StringRef verName = stringTable.data() + verneeds[idx]; |
1556 | versionedNameBuffer.clear(); |
1557 | name = saver().save( |
1558 | S: (name + "@" + verName).toStringRef(Out&: versionedNameBuffer)); |
1559 | } |
1560 | Symbol *s = symtab.addSymbol( |
1561 | newSym: Undefined{this, name, sym.getBinding(), sym.st_other, sym.getType()}); |
1562 | s->exportDynamic = true; |
1563 | if (sym.getBinding() != STB_WEAK && |
1564 | config->unresolvedSymbolsInShlib != UnresolvedPolicy::Ignore) |
1565 | requiredSymbols.push_back(Elt: s); |
1566 | continue; |
1567 | } |
1568 | |
1569 | if (ver == VER_NDX_LOCAL || |
1570 | (ver != VER_NDX_GLOBAL && idx >= verdefs.size())) { |
1571 | // In GNU ld < 2.31 (before 3be08ea4728b56d35e136af4e6fd3086ade17764), the |
1572 | // MIPS port puts _gp_disp symbol into DSO files and incorrectly assigns |
1573 | // VER_NDX_LOCAL. Workaround this bug. |
1574 | if (config->emachine == EM_MIPS && name == "_gp_disp" ) |
1575 | continue; |
1576 | error(msg: "corrupt input file: version definition index " + Twine(idx) + |
1577 | " for symbol " + name + " is out of bounds\n>>> defined in " + |
1578 | toString(f: this)); |
1579 | continue; |
1580 | } |
1581 | |
1582 | uint32_t alignment = getAlignment<ELFT>(sections, sym); |
1583 | if (ver == idx) { |
1584 | auto *s = symtab.addSymbol( |
1585 | newSym: SharedSymbol{*this, name, sym.getBinding(), sym.st_other, |
1586 | sym.getType(), sym.st_value, sym.st_size, alignment}); |
1587 | s->dsoDefined = true; |
1588 | if (s->file == this) |
1589 | s->versionId = ver; |
1590 | } |
1591 | |
1592 | // Also add the symbol with the versioned name to handle undefined symbols |
1593 | // with explicit versions. |
1594 | if (ver == VER_NDX_GLOBAL) |
1595 | continue; |
1596 | |
1597 | StringRef verName = |
1598 | stringTable.data() + |
1599 | reinterpret_cast<const Elf_Verdef *>(verdefs[idx])->getAux()->vda_name; |
1600 | versionedNameBuffer.clear(); |
1601 | name = (name + "@" + verName).toStringRef(Out&: versionedNameBuffer); |
1602 | auto *s = symtab.addSymbol( |
1603 | newSym: SharedSymbol{*this, saver().save(S: name), sym.getBinding(), sym.st_other, |
1604 | sym.getType(), sym.st_value, sym.st_size, alignment}); |
1605 | s->dsoDefined = true; |
1606 | if (s->file == this) |
1607 | s->versionId = idx; |
1608 | } |
1609 | } |
1610 | |
1611 | static ELFKind getBitcodeELFKind(const Triple &t) { |
1612 | if (t.isLittleEndian()) |
1613 | return t.isArch64Bit() ? ELF64LEKind : ELF32LEKind; |
1614 | return t.isArch64Bit() ? ELF64BEKind : ELF32BEKind; |
1615 | } |
1616 | |
1617 | static uint16_t getBitcodeMachineKind(StringRef path, const Triple &t) { |
1618 | switch (t.getArch()) { |
1619 | case Triple::aarch64: |
1620 | case Triple::aarch64_be: |
1621 | return EM_AARCH64; |
1622 | case Triple::amdgcn: |
1623 | case Triple::r600: |
1624 | return EM_AMDGPU; |
1625 | case Triple::arm: |
1626 | case Triple::armeb: |
1627 | case Triple::thumb: |
1628 | case Triple::thumbeb: |
1629 | return EM_ARM; |
1630 | case Triple::avr: |
1631 | return EM_AVR; |
1632 | case Triple::hexagon: |
1633 | return EM_HEXAGON; |
1634 | case Triple::loongarch32: |
1635 | case Triple::loongarch64: |
1636 | return EM_LOONGARCH; |
1637 | case Triple::mips: |
1638 | case Triple::mipsel: |
1639 | case Triple::mips64: |
1640 | case Triple::mips64el: |
1641 | return EM_MIPS; |
1642 | case Triple::msp430: |
1643 | return EM_MSP430; |
1644 | case Triple::ppc: |
1645 | case Triple::ppcle: |
1646 | return EM_PPC; |
1647 | case Triple::ppc64: |
1648 | case Triple::ppc64le: |
1649 | return EM_PPC64; |
1650 | case Triple::riscv32: |
1651 | case Triple::riscv64: |
1652 | return EM_RISCV; |
1653 | case Triple::sparcv9: |
1654 | return EM_SPARCV9; |
1655 | case Triple::systemz: |
1656 | return EM_S390; |
1657 | case Triple::x86: |
1658 | return t.isOSIAMCU() ? EM_IAMCU : EM_386; |
1659 | case Triple::x86_64: |
1660 | return EM_X86_64; |
1661 | default: |
1662 | error(msg: path + ": could not infer e_machine from bitcode target triple " + |
1663 | t.str()); |
1664 | return EM_NONE; |
1665 | } |
1666 | } |
1667 | |
1668 | static uint8_t getOsAbi(const Triple &t) { |
1669 | switch (t.getOS()) { |
1670 | case Triple::AMDHSA: |
1671 | return ELF::ELFOSABI_AMDGPU_HSA; |
1672 | case Triple::AMDPAL: |
1673 | return ELF::ELFOSABI_AMDGPU_PAL; |
1674 | case Triple::Mesa3D: |
1675 | return ELF::ELFOSABI_AMDGPU_MESA3D; |
1676 | default: |
1677 | return ELF::ELFOSABI_NONE; |
1678 | } |
1679 | } |
1680 | |
1681 | BitcodeFile::BitcodeFile(MemoryBufferRef mb, StringRef archiveName, |
1682 | uint64_t offsetInArchive, bool lazy) |
1683 | : InputFile(BitcodeKind, mb) { |
1684 | this->archiveName = archiveName; |
1685 | this->lazy = lazy; |
1686 | |
1687 | std::string path = mb.getBufferIdentifier().str(); |
1688 | if (config->thinLTOIndexOnly) |
1689 | path = replaceThinLTOSuffix(path: mb.getBufferIdentifier()); |
1690 | |
1691 | // ThinLTO assumes that all MemoryBufferRefs given to it have a unique |
1692 | // name. If two archives define two members with the same name, this |
1693 | // causes a collision which result in only one of the objects being taken |
1694 | // into consideration at LTO time (which very likely causes undefined |
1695 | // symbols later in the link stage). So we append file offset to make |
1696 | // filename unique. |
1697 | StringRef name = archiveName.empty() |
1698 | ? saver().save(S: path) |
1699 | : saver().save(S: archiveName + "(" + path::filename(path) + |
1700 | " at " + utostr(X: offsetInArchive) + ")" ); |
1701 | MemoryBufferRef mbref(mb.getBuffer(), name); |
1702 | |
1703 | obj = CHECK(lto::InputFile::create(mbref), this); |
1704 | |
1705 | Triple t(obj->getTargetTriple()); |
1706 | ekind = getBitcodeELFKind(t); |
1707 | emachine = getBitcodeMachineKind(path: mb.getBufferIdentifier(), t); |
1708 | osabi = getOsAbi(t); |
1709 | } |
1710 | |
1711 | static uint8_t mapVisibility(GlobalValue::VisibilityTypes gvVisibility) { |
1712 | switch (gvVisibility) { |
1713 | case GlobalValue::DefaultVisibility: |
1714 | return STV_DEFAULT; |
1715 | case GlobalValue::HiddenVisibility: |
1716 | return STV_HIDDEN; |
1717 | case GlobalValue::ProtectedVisibility: |
1718 | return STV_PROTECTED; |
1719 | } |
1720 | llvm_unreachable("unknown visibility" ); |
1721 | } |
1722 | |
1723 | static void |
1724 | createBitcodeSymbol(Symbol *&sym, const std::vector<bool> &keptComdats, |
1725 | const lto::InputFile::Symbol &objSym, BitcodeFile &f) { |
1726 | uint8_t binding = objSym.isWeak() ? STB_WEAK : STB_GLOBAL; |
1727 | uint8_t type = objSym.isTLS() ? STT_TLS : STT_NOTYPE; |
1728 | uint8_t visibility = mapVisibility(gvVisibility: objSym.getVisibility()); |
1729 | |
1730 | if (!sym) |
1731 | sym = symtab.insert(name: saver().save(S: objSym.getName())); |
1732 | |
1733 | int c = objSym.getComdatIndex(); |
1734 | if (objSym.isUndefined() || (c != -1 && !keptComdats[c])) { |
1735 | Undefined newSym(&f, StringRef(), binding, visibility, type); |
1736 | sym->resolve(other: newSym); |
1737 | sym->referenced = true; |
1738 | return; |
1739 | } |
1740 | |
1741 | if (objSym.isCommon()) { |
1742 | sym->resolve(other: CommonSymbol{&f, StringRef(), binding, visibility, STT_OBJECT, |
1743 | objSym.getCommonAlignment(), |
1744 | objSym.getCommonSize()}); |
1745 | } else { |
1746 | Defined newSym(&f, StringRef(), binding, visibility, type, 0, 0, nullptr); |
1747 | if (objSym.canBeOmittedFromSymbolTable()) |
1748 | newSym.exportDynamic = false; |
1749 | sym->resolve(other: newSym); |
1750 | } |
1751 | } |
1752 | |
1753 | void BitcodeFile::parse() { |
1754 | for (std::pair<StringRef, Comdat::SelectionKind> s : obj->getComdatTable()) { |
1755 | keptComdats.push_back( |
1756 | x: s.second == Comdat::NoDeduplicate || |
1757 | symtab.comdatGroups.try_emplace(Key: CachedHashStringRef(s.first), Args: this) |
1758 | .second); |
1759 | } |
1760 | |
1761 | if (numSymbols == 0) { |
1762 | numSymbols = obj->symbols().size(); |
1763 | symbols = std::make_unique<Symbol *[]>(num: numSymbols); |
1764 | } |
1765 | // Process defined symbols first. See the comment in |
1766 | // ObjFile<ELFT>::initializeSymbols. |
1767 | for (auto [i, irSym] : llvm::enumerate(First: obj->symbols())) |
1768 | if (!irSym.isUndefined()) |
1769 | createBitcodeSymbol(sym&: symbols[i], keptComdats, objSym: irSym, f&: *this); |
1770 | for (auto [i, irSym] : llvm::enumerate(First: obj->symbols())) |
1771 | if (irSym.isUndefined()) |
1772 | createBitcodeSymbol(sym&: symbols[i], keptComdats, objSym: irSym, f&: *this); |
1773 | |
1774 | for (auto l : obj->getDependentLibraries()) |
1775 | addDependentLibrary(specifier: l, f: this); |
1776 | } |
1777 | |
1778 | void BitcodeFile::parseLazy() { |
1779 | numSymbols = obj->symbols().size(); |
1780 | symbols = std::make_unique<Symbol *[]>(num: numSymbols); |
1781 | for (auto [i, irSym] : llvm::enumerate(First: obj->symbols())) |
1782 | if (!irSym.isUndefined()) { |
1783 | auto *sym = symtab.insert(name: saver().save(S: irSym.getName())); |
1784 | sym->resolve(other: LazySymbol{*this}); |
1785 | symbols[i] = sym; |
1786 | } |
1787 | } |
1788 | |
1789 | void BitcodeFile::postParse() { |
1790 | for (auto [i, irSym] : llvm::enumerate(First: obj->symbols())) { |
1791 | const Symbol &sym = *symbols[i]; |
1792 | if (sym.file == this || !sym.isDefined() || irSym.isUndefined() || |
1793 | irSym.isCommon() || irSym.isWeak()) |
1794 | continue; |
1795 | int c = irSym.getComdatIndex(); |
1796 | if (c != -1 && !keptComdats[c]) |
1797 | continue; |
1798 | reportDuplicate(sym, newFile: this, errSec: nullptr, errOffset: 0); |
1799 | } |
1800 | } |
1801 | |
1802 | void BinaryFile::parse() { |
1803 | ArrayRef<uint8_t> data = arrayRefFromStringRef(Input: mb.getBuffer()); |
1804 | auto *section = make<InputSection>(args: this, args: SHF_ALLOC | SHF_WRITE, args: SHT_PROGBITS, |
1805 | args: 8, args&: data, args: ".data" ); |
1806 | sections.push_back(Elt: section); |
1807 | |
1808 | // For each input file foo that is embedded to a result as a binary |
1809 | // blob, we define _binary_foo_{start,end,size} symbols, so that |
1810 | // user programs can access blobs by name. Non-alphanumeric |
1811 | // characters in a filename are replaced with underscore. |
1812 | std::string s = "_binary_" + mb.getBufferIdentifier().str(); |
1813 | for (char &c : s) |
1814 | if (!isAlnum(C: c)) |
1815 | c = '_'; |
1816 | |
1817 | llvm::StringSaver &saver = lld::saver(); |
1818 | |
1819 | symtab.addAndCheckDuplicate(newSym: Defined{this, saver.save(S: s + "_start" ), |
1820 | STB_GLOBAL, STV_DEFAULT, STT_OBJECT, 0, 0, |
1821 | section}); |
1822 | symtab.addAndCheckDuplicate(newSym: Defined{this, saver.save(S: s + "_end" ), STB_GLOBAL, |
1823 | STV_DEFAULT, STT_OBJECT, data.size(), 0, |
1824 | section}); |
1825 | symtab.addAndCheckDuplicate(newSym: Defined{this, saver.save(S: s + "_size" ), STB_GLOBAL, |
1826 | STV_DEFAULT, STT_OBJECT, data.size(), 0, |
1827 | nullptr}); |
1828 | } |
1829 | |
1830 | InputFile *elf::createInternalFile(StringRef name) { |
1831 | auto *file = |
1832 | make<InputFile>(args: InputFile::InternalKind, args: MemoryBufferRef("" , name)); |
1833 | // References from an internal file do not lead to --warn-backrefs |
1834 | // diagnostics. |
1835 | file->groupId = 0; |
1836 | return file; |
1837 | } |
1838 | |
1839 | ELFFileBase *elf::createObjFile(MemoryBufferRef mb, StringRef archiveName, |
1840 | bool lazy) { |
1841 | ELFFileBase *f; |
1842 | switch (getELFKind(mb, archiveName)) { |
1843 | case ELF32LEKind: |
1844 | f = make<ObjFile<ELF32LE>>(args: ELF32LEKind, args&: mb, args&: archiveName); |
1845 | break; |
1846 | case ELF32BEKind: |
1847 | f = make<ObjFile<ELF32BE>>(args: ELF32BEKind, args&: mb, args&: archiveName); |
1848 | break; |
1849 | case ELF64LEKind: |
1850 | f = make<ObjFile<ELF64LE>>(args: ELF64LEKind, args&: mb, args&: archiveName); |
1851 | break; |
1852 | case ELF64BEKind: |
1853 | f = make<ObjFile<ELF64BE>>(args: ELF64BEKind, args&: mb, args&: archiveName); |
1854 | break; |
1855 | default: |
1856 | llvm_unreachable("getELFKind" ); |
1857 | } |
1858 | f->init(); |
1859 | f->lazy = lazy; |
1860 | return f; |
1861 | } |
1862 | |
1863 | template <class ELFT> void ObjFile<ELFT>::parseLazy() { |
1864 | const ArrayRef<typename ELFT::Sym> eSyms = this->getELFSyms<ELFT>(); |
1865 | numSymbols = eSyms.size(); |
1866 | symbols = std::make_unique<Symbol *[]>(numSymbols); |
1867 | |
1868 | // resolve() may trigger this->extract() if an existing symbol is an undefined |
1869 | // symbol. If that happens, this function has served its purpose, and we can |
1870 | // exit from the loop early. |
1871 | for (size_t i = firstGlobal, end = eSyms.size(); i != end; ++i) { |
1872 | if (eSyms[i].st_shndx == SHN_UNDEF) |
1873 | continue; |
1874 | symbols[i] = symtab.insert(CHECK(eSyms[i].getName(stringTable), this)); |
1875 | symbols[i]->resolve(LazySymbol{*this}); |
1876 | if (!lazy) |
1877 | break; |
1878 | } |
1879 | } |
1880 | |
1881 | bool InputFile::(StringRef name) const { |
1882 | if (isa<BitcodeFile>(Val: this)) |
1883 | return isBitcodeNonCommonDef(mb, symName: name, archiveName); |
1884 | |
1885 | return isNonCommonDef(mb, symName: name, archiveName); |
1886 | } |
1887 | |
1888 | std::string elf::replaceThinLTOSuffix(StringRef path) { |
1889 | auto [suffix, repl] = config->thinLTOObjectSuffixReplace; |
1890 | if (path.consume_back(Suffix: suffix)) |
1891 | return (path + repl).str(); |
1892 | return std::string(path); |
1893 | } |
1894 | |
1895 | template class elf::ObjFile<ELF32LE>; |
1896 | template class elf::ObjFile<ELF32BE>; |
1897 | template class elf::ObjFile<ELF64LE>; |
1898 | template class elf::ObjFile<ELF64BE>; |
1899 | |
1900 | template void SharedFile::parse<ELF32LE>(); |
1901 | template void SharedFile::parse<ELF32BE>(); |
1902 | template void SharedFile::parse<ELF64LE>(); |
1903 | template void SharedFile::parse<ELF64BE>(); |
1904 | |