| 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/DWARF.h" |
| 20 | #include "llvm/ADT/CachedHashString.h" |
| 21 | #include "llvm/ADT/STLExtras.h" |
| 22 | #include "llvm/LTO/LTO.h" |
| 23 | #include "llvm/Object/IRObjectFile.h" |
| 24 | #include "llvm/Support/ARMAttributeParser.h" |
| 25 | #include "llvm/Support/ARMBuildAttributes.h" |
| 26 | #include "llvm/Support/Endian.h" |
| 27 | #include "llvm/Support/FileSystem.h" |
| 28 | #include "llvm/Support/Path.h" |
| 29 | #include "llvm/Support/TimeProfiler.h" |
| 30 | #include "llvm/Support/raw_ostream.h" |
| 31 | #include <optional> |
| 32 | |
| 33 | using namespace llvm; |
| 34 | using namespace llvm::ELF; |
| 35 | using namespace llvm::object; |
| 36 | using namespace llvm::sys; |
| 37 | using namespace llvm::sys::fs; |
| 38 | using namespace llvm::support::endian; |
| 39 | using namespace lld; |
| 40 | using namespace lld::elf; |
| 41 | |
| 42 | // This function is explicitly instantiated in ARM.cpp, don't do it here to |
| 43 | // avoid warnings with MSVC. |
| 44 | extern template void ObjFile<ELF32LE>::importCmseSymbols(); |
| 45 | extern template void ObjFile<ELF32BE>::importCmseSymbols(); |
| 46 | extern template void ObjFile<ELF64LE>::importCmseSymbols(); |
| 47 | extern template void ObjFile<ELF64BE>::importCmseSymbols(); |
| 48 | |
| 49 | // Returns "<internal>", "foo.a(bar.o)" or "baz.o". |
| 50 | std::string elf::toStr(Ctx &ctx, const InputFile *f) { |
| 51 | static std::mutex mu; |
| 52 | if (!f) |
| 53 | return "<internal>" ; |
| 54 | |
| 55 | { |
| 56 | std::lock_guard<std::mutex> lock(mu); |
| 57 | if (f->toStringCache.empty()) { |
| 58 | if (f->archiveName.empty()) |
| 59 | f->toStringCache = f->getName(); |
| 60 | else |
| 61 | (f->archiveName + "(" + f->getName() + ")" ).toVector(Out&: f->toStringCache); |
| 62 | } |
| 63 | } |
| 64 | return std::string(f->toStringCache); |
| 65 | } |
| 66 | |
| 67 | const ELFSyncStream &elf::operator<<(const ELFSyncStream &s, |
| 68 | const InputFile *f) { |
| 69 | return s << toStr(ctx&: s.ctx, f); |
| 70 | } |
| 71 | |
| 72 | static ELFKind getELFKind(Ctx &ctx, MemoryBufferRef mb, StringRef archiveName) { |
| 73 | unsigned char size; |
| 74 | unsigned char endian; |
| 75 | std::tie(args&: size, args&: endian) = getElfArchType(Object: mb.getBuffer()); |
| 76 | |
| 77 | auto report = [&](StringRef msg) { |
| 78 | StringRef filename = mb.getBufferIdentifier(); |
| 79 | if (archiveName.empty()) |
| 80 | Fatal(ctx) << filename << ": " << msg; |
| 81 | else |
| 82 | Fatal(ctx) << archiveName << "(" << filename << "): " << msg; |
| 83 | }; |
| 84 | |
| 85 | if (!mb.getBuffer().starts_with(Prefix: ElfMagic)) |
| 86 | report("not an ELF file" ); |
| 87 | if (endian != ELFDATA2LSB && endian != ELFDATA2MSB) |
| 88 | report("corrupted ELF file: invalid data encoding" ); |
| 89 | if (size != ELFCLASS32 && size != ELFCLASS64) |
| 90 | report("corrupted ELF file: invalid file class" ); |
| 91 | |
| 92 | size_t bufSize = mb.getBuffer().size(); |
| 93 | if ((size == ELFCLASS32 && bufSize < sizeof(Elf32_Ehdr)) || |
| 94 | (size == ELFCLASS64 && bufSize < sizeof(Elf64_Ehdr))) |
| 95 | report("corrupted ELF file: file is too short" ); |
| 96 | |
| 97 | if (size == ELFCLASS32) |
| 98 | return (endian == ELFDATA2LSB) ? ELF32LEKind : ELF32BEKind; |
| 99 | return (endian == ELFDATA2LSB) ? ELF64LEKind : ELF64BEKind; |
| 100 | } |
| 101 | |
| 102 | // For ARM only, to set the EF_ARM_ABI_FLOAT_SOFT or EF_ARM_ABI_FLOAT_HARD |
| 103 | // flag in the ELF Header we need to look at Tag_ABI_VFP_args to find out how |
| 104 | // the input objects have been compiled. |
| 105 | static void updateARMVFPArgs(Ctx &ctx, const ARMAttributeParser &attributes, |
| 106 | const InputFile *f) { |
| 107 | std::optional<unsigned> attr = |
| 108 | attributes.getAttributeValue(tag: ARMBuildAttrs::ABI_VFP_args); |
| 109 | if (!attr) |
| 110 | // If an ABI tag isn't present then it is implicitly given the value of 0 |
| 111 | // which maps to ARMBuildAttrs::BaseAAPCS. However many assembler files, |
| 112 | // including some in glibc that don't use FP args (and should have value 3) |
| 113 | // don't have the attribute so we do not consider an implicit value of 0 |
| 114 | // as a clash. |
| 115 | return; |
| 116 | |
| 117 | unsigned vfpArgs = *attr; |
| 118 | ARMVFPArgKind arg; |
| 119 | switch (vfpArgs) { |
| 120 | case ARMBuildAttrs::BaseAAPCS: |
| 121 | arg = ARMVFPArgKind::Base; |
| 122 | break; |
| 123 | case ARMBuildAttrs::HardFPAAPCS: |
| 124 | arg = ARMVFPArgKind::VFP; |
| 125 | break; |
| 126 | case ARMBuildAttrs::ToolChainFPPCS: |
| 127 | // Tool chain specific convention that conforms to neither AAPCS variant. |
| 128 | arg = ARMVFPArgKind::ToolChain; |
| 129 | break; |
| 130 | case ARMBuildAttrs::CompatibleFPAAPCS: |
| 131 | // Object compatible with all conventions. |
| 132 | return; |
| 133 | default: |
| 134 | ErrAlways(ctx) << f << ": unknown Tag_ABI_VFP_args value: " << vfpArgs; |
| 135 | return; |
| 136 | } |
| 137 | // Follow ld.bfd and error if there is a mix of calling conventions. |
| 138 | if (ctx.arg.armVFPArgs != arg && ctx.arg.armVFPArgs != ARMVFPArgKind::Default) |
| 139 | ErrAlways(ctx) << f << ": incompatible Tag_ABI_VFP_args" ; |
| 140 | else |
| 141 | ctx.arg.armVFPArgs = arg; |
| 142 | } |
| 143 | |
| 144 | // The ARM support in lld makes some use of instructions that are not available |
| 145 | // on all ARM architectures. Namely: |
| 146 | // - Use of BLX instruction for interworking between ARM and Thumb state. |
| 147 | // - Use of the extended Thumb branch encoding in relocation. |
| 148 | // - Use of the MOVT/MOVW instructions in Thumb Thunks. |
| 149 | // The ARM Attributes section contains information about the architecture chosen |
| 150 | // at compile time. We follow the convention that if at least one input object |
| 151 | // is compiled with an architecture that supports these features then lld is |
| 152 | // permitted to use them. |
| 153 | static void updateSupportedARMFeatures(Ctx &ctx, |
| 154 | const ARMAttributeParser &attributes) { |
| 155 | std::optional<unsigned> attr = |
| 156 | attributes.getAttributeValue(tag: ARMBuildAttrs::CPU_arch); |
| 157 | if (!attr) |
| 158 | return; |
| 159 | auto arch = *attr; |
| 160 | switch (arch) { |
| 161 | case ARMBuildAttrs::Pre_v4: |
| 162 | case ARMBuildAttrs::v4: |
| 163 | case ARMBuildAttrs::v4T: |
| 164 | // Architectures prior to v5 do not support BLX instruction |
| 165 | break; |
| 166 | case ARMBuildAttrs::v5T: |
| 167 | case ARMBuildAttrs::v5TE: |
| 168 | case ARMBuildAttrs::v5TEJ: |
| 169 | case ARMBuildAttrs::v6: |
| 170 | case ARMBuildAttrs::v6KZ: |
| 171 | case ARMBuildAttrs::v6K: |
| 172 | ctx.arg.armHasBlx = true; |
| 173 | // Architectures used in pre-Cortex processors do not support |
| 174 | // The J1 = 1 J2 = 1 Thumb branch range extension, with the exception |
| 175 | // of Architecture v6T2 (arm1156t2-s and arm1156t2f-s) that do. |
| 176 | break; |
| 177 | default: |
| 178 | // All other Architectures have BLX and extended branch encoding |
| 179 | ctx.arg.armHasBlx = true; |
| 180 | ctx.arg.armJ1J2BranchEncoding = true; |
| 181 | if (arch != ARMBuildAttrs::v6_M && arch != ARMBuildAttrs::v6S_M) |
| 182 | // All Architectures used in Cortex processors with the exception |
| 183 | // of v6-M and v6S-M have the MOVT and MOVW instructions. |
| 184 | ctx.arg.armHasMovtMovw = true; |
| 185 | break; |
| 186 | } |
| 187 | |
| 188 | // Only ARMv8-M or later architectures have CMSE support. |
| 189 | std::optional<unsigned> profile = |
| 190 | attributes.getAttributeValue(tag: ARMBuildAttrs::CPU_arch_profile); |
| 191 | if (!profile) |
| 192 | return; |
| 193 | if (arch >= ARMBuildAttrs::CPUArch::v8_M_Base && |
| 194 | profile == ARMBuildAttrs::MicroControllerProfile) |
| 195 | ctx.arg.armCMSESupport = true; |
| 196 | |
| 197 | // The thumb PLT entries require Thumb2 which can be used on multiple archs. |
| 198 | // For now, let's limit it to ones where ARM isn't available and we know have |
| 199 | // Thumb2. |
| 200 | std::optional<unsigned> armISA = |
| 201 | attributes.getAttributeValue(tag: ARMBuildAttrs::ARM_ISA_use); |
| 202 | std::optional<unsigned> thumb = |
| 203 | attributes.getAttributeValue(tag: ARMBuildAttrs::THUMB_ISA_use); |
| 204 | ctx.arg.armHasArmISA |= armISA && *armISA >= ARMBuildAttrs::Allowed; |
| 205 | ctx.arg.armHasThumb2ISA |= thumb && *thumb >= ARMBuildAttrs::AllowThumb32; |
| 206 | } |
| 207 | |
| 208 | InputFile::InputFile(Ctx &ctx, Kind k, MemoryBufferRef m) |
| 209 | : ctx(ctx), mb(m), groupId(ctx.driver.nextGroupId), fileKind(k) { |
| 210 | // All files within the same --{start,end}-group get the same group ID. |
| 211 | // Otherwise, a new file will get a new group ID. |
| 212 | if (!ctx.driver.isInGroup) |
| 213 | ++ctx.driver.nextGroupId; |
| 214 | } |
| 215 | |
| 216 | InputFile::~InputFile() {} |
| 217 | |
| 218 | std::optional<MemoryBufferRef> elf::readFile(Ctx &ctx, StringRef path) { |
| 219 | llvm::TimeTraceScope timeScope("Load input files" , path); |
| 220 | |
| 221 | // The --chroot option changes our virtual root directory. |
| 222 | // This is useful when you are dealing with files created by --reproduce. |
| 223 | if (!ctx.arg.chroot.empty() && path.starts_with(Prefix: "/" )) |
| 224 | path = ctx.saver.save(S: ctx.arg.chroot + path); |
| 225 | |
| 226 | bool remapped = false; |
| 227 | auto it = ctx.arg.remapInputs.find(Val: path); |
| 228 | if (it != ctx.arg.remapInputs.end()) { |
| 229 | path = it->second; |
| 230 | remapped = true; |
| 231 | } else { |
| 232 | for (const auto &[pat, toFile] : ctx.arg.remapInputsWildcards) { |
| 233 | if (pat.match(S: path)) { |
| 234 | path = toFile; |
| 235 | remapped = true; |
| 236 | break; |
| 237 | } |
| 238 | } |
| 239 | } |
| 240 | if (remapped) { |
| 241 | // Use /dev/null to indicate an input file that should be ignored. Change |
| 242 | // the path to NUL on Windows. |
| 243 | #ifdef _WIN32 |
| 244 | if (path == "/dev/null" ) |
| 245 | path = "NUL" ; |
| 246 | #endif |
| 247 | } |
| 248 | |
| 249 | Log(ctx) << path; |
| 250 | ctx.arg.dependencyFiles.insert(X: llvm::CachedHashString(path)); |
| 251 | |
| 252 | auto mbOrErr = MemoryBuffer::getFile(Filename: path, /*IsText=*/false, |
| 253 | /*RequiresNullTerminator=*/false); |
| 254 | if (auto ec = mbOrErr.getError()) { |
| 255 | ErrAlways(ctx) << "cannot open " << path << ": " << ec.message(); |
| 256 | return std::nullopt; |
| 257 | } |
| 258 | |
| 259 | MemoryBufferRef mbref = (*mbOrErr)->getMemBufferRef(); |
| 260 | ctx.memoryBuffers.push_back(Elt: std::move(*mbOrErr)); // take MB ownership |
| 261 | |
| 262 | if (ctx.tar) |
| 263 | ctx.tar->append(Path: relativeToRoot(path), Data: mbref.getBuffer()); |
| 264 | return mbref; |
| 265 | } |
| 266 | |
| 267 | // All input object files must be for the same architecture |
| 268 | // (e.g. it does not make sense to link x86 object files with |
| 269 | // MIPS object files.) This function checks for that error. |
| 270 | static bool isCompatible(Ctx &ctx, InputFile *file) { |
| 271 | if (!file->isElf() && !isa<BitcodeFile>(Val: file)) |
| 272 | return true; |
| 273 | |
| 274 | if (file->ekind == ctx.arg.ekind && file->emachine == ctx.arg.emachine) { |
| 275 | if (ctx.arg.emachine != EM_MIPS) |
| 276 | return true; |
| 277 | if (isMipsN32Abi(ctx, f: *file) == ctx.arg.mipsN32Abi) |
| 278 | return true; |
| 279 | } |
| 280 | |
| 281 | StringRef target = |
| 282 | !ctx.arg.bfdname.empty() ? ctx.arg.bfdname : ctx.arg.emulation; |
| 283 | if (!target.empty()) { |
| 284 | Err(ctx) << file << " is incompatible with " << target; |
| 285 | return false; |
| 286 | } |
| 287 | |
| 288 | InputFile *existing = nullptr; |
| 289 | if (!ctx.objectFiles.empty()) |
| 290 | existing = ctx.objectFiles[0]; |
| 291 | else if (!ctx.sharedFiles.empty()) |
| 292 | existing = ctx.sharedFiles[0]; |
| 293 | else if (!ctx.bitcodeFiles.empty()) |
| 294 | existing = ctx.bitcodeFiles[0]; |
| 295 | auto diag = Err(ctx); |
| 296 | diag << file << " is incompatible" ; |
| 297 | if (existing) |
| 298 | diag << " with " << existing; |
| 299 | return false; |
| 300 | } |
| 301 | |
| 302 | template <class ELFT> static void doParseFile(Ctx &ctx, InputFile *file) { |
| 303 | if (!isCompatible(ctx, file)) |
| 304 | return; |
| 305 | |
| 306 | // Lazy object file |
| 307 | if (file->lazy) { |
| 308 | if (auto *f = dyn_cast<BitcodeFile>(Val: file)) { |
| 309 | ctx.lazyBitcodeFiles.push_back(Elt: f); |
| 310 | f->parseLazy(); |
| 311 | } else { |
| 312 | cast<ObjFile<ELFT>>(file)->parseLazy(); |
| 313 | } |
| 314 | return; |
| 315 | } |
| 316 | |
| 317 | if (ctx.arg.trace) |
| 318 | Msg(ctx) << file; |
| 319 | |
| 320 | if (file->kind() == InputFile::ObjKind) { |
| 321 | ctx.objectFiles.push_back(Elt: cast<ELFFileBase>(Val: file)); |
| 322 | cast<ObjFile<ELFT>>(file)->parse(); |
| 323 | } else if (auto *f = dyn_cast<SharedFile>(Val: file)) { |
| 324 | f->parse<ELFT>(); |
| 325 | } else if (auto *f = dyn_cast<BitcodeFile>(Val: file)) { |
| 326 | ctx.bitcodeFiles.push_back(Elt: f); |
| 327 | f->parse(); |
| 328 | } else { |
| 329 | ctx.binaryFiles.push_back(Elt: cast<BinaryFile>(Val: file)); |
| 330 | cast<BinaryFile>(Val: file)->parse(); |
| 331 | } |
| 332 | } |
| 333 | |
| 334 | // Add symbols in File to the symbol table. |
| 335 | void elf::parseFile(Ctx &ctx, InputFile *file) { |
| 336 | invokeELFT(doParseFile, ctx, file); |
| 337 | } |
| 338 | |
| 339 | // This function is explicitly instantiated in ARM.cpp. Mark it extern here, |
| 340 | // to avoid warnings when building with MSVC. |
| 341 | extern template void ObjFile<ELF32LE>::importCmseSymbols(); |
| 342 | extern template void ObjFile<ELF32BE>::importCmseSymbols(); |
| 343 | extern template void ObjFile<ELF64LE>::importCmseSymbols(); |
| 344 | extern template void ObjFile<ELF64BE>::importCmseSymbols(); |
| 345 | |
| 346 | template <class ELFT> |
| 347 | static void |
| 348 | doParseFiles(Ctx &ctx, |
| 349 | const SmallVector<std::unique_ptr<InputFile>, 0> &files) { |
| 350 | // Add all files to the symbol table. This will add almost all symbols that we |
| 351 | // need to the symbol table. This process might add files to the link due to |
| 352 | // addDependentLibrary. |
| 353 | for (size_t i = 0; i < files.size(); ++i) { |
| 354 | llvm::TimeTraceScope timeScope("Parse input files" , files[i]->getName()); |
| 355 | doParseFile<ELFT>(ctx, files[i].get()); |
| 356 | } |
| 357 | if (ctx.driver.armCmseImpLib) |
| 358 | cast<ObjFile<ELFT>>(*ctx.driver.armCmseImpLib).importCmseSymbols(); |
| 359 | } |
| 360 | |
| 361 | void elf::parseFiles(Ctx &ctx, |
| 362 | const SmallVector<std::unique_ptr<InputFile>, 0> &files) { |
| 363 | llvm::TimeTraceScope timeScope("Parse input files" ); |
| 364 | invokeELFT(doParseFiles, ctx, files); |
| 365 | } |
| 366 | |
| 367 | // Concatenates arguments to construct a string representing an error location. |
| 368 | StringRef InputFile::getNameForScript() const { |
| 369 | if (archiveName.empty()) |
| 370 | return getName(); |
| 371 | |
| 372 | if (nameForScriptCache.empty()) |
| 373 | nameForScriptCache = (archiveName + Twine(':') + getName()).str(); |
| 374 | |
| 375 | return nameForScriptCache; |
| 376 | } |
| 377 | |
| 378 | // An ELF object file may contain a `.deplibs` section. If it exists, the |
| 379 | // section contains a list of library specifiers such as `m` for libm. This |
| 380 | // function resolves a given name by finding the first matching library checking |
| 381 | // the various ways that a library can be specified to LLD. This ELF extension |
| 382 | // is a form of autolinking and is called `dependent libraries`. It is currently |
| 383 | // unique to LLVM and lld. |
| 384 | static void addDependentLibrary(Ctx &ctx, StringRef specifier, |
| 385 | const InputFile *f) { |
| 386 | if (!ctx.arg.dependentLibraries) |
| 387 | return; |
| 388 | if (std::optional<std::string> s = searchLibraryBaseName(ctx, path: specifier)) |
| 389 | ctx.driver.addFile(path: ctx.saver.save(S: *s), /*withLOption=*/true); |
| 390 | else if (std::optional<std::string> s = findFromSearchPaths(ctx, path: specifier)) |
| 391 | ctx.driver.addFile(path: ctx.saver.save(S: *s), /*withLOption=*/true); |
| 392 | else if (fs::exists(Path: specifier)) |
| 393 | ctx.driver.addFile(path: specifier, /*withLOption=*/false); |
| 394 | else |
| 395 | ErrAlways(ctx) |
| 396 | << f << ": unable to find library from dependent library specifier: " |
| 397 | << specifier; |
| 398 | } |
| 399 | |
| 400 | // Record the membership of a section group so that in the garbage collection |
| 401 | // pass, section group members are kept or discarded as a unit. |
| 402 | template <class ELFT> |
| 403 | static void handleSectionGroup(ArrayRef<InputSectionBase *> sections, |
| 404 | ArrayRef<typename ELFT::Word> entries) { |
| 405 | bool hasAlloc = false; |
| 406 | for (uint32_t index : entries.slice(1)) { |
| 407 | if (index >= sections.size()) |
| 408 | return; |
| 409 | if (InputSectionBase *s = sections[index]) |
| 410 | if (s != &InputSection::discarded && s->flags & SHF_ALLOC) |
| 411 | hasAlloc = true; |
| 412 | } |
| 413 | |
| 414 | // If any member has the SHF_ALLOC flag, the whole group is subject to garbage |
| 415 | // collection. See the comment in markLive(). This rule retains .debug_types |
| 416 | // and .rela.debug_types. |
| 417 | if (!hasAlloc) |
| 418 | return; |
| 419 | |
| 420 | // Connect the members in a circular doubly-linked list via |
| 421 | // nextInSectionGroup. |
| 422 | InputSectionBase *head; |
| 423 | InputSectionBase *prev = nullptr; |
| 424 | for (uint32_t index : entries.slice(1)) { |
| 425 | InputSectionBase *s = sections[index]; |
| 426 | if (!s || s == &InputSection::discarded) |
| 427 | continue; |
| 428 | if (prev) |
| 429 | prev->nextInSectionGroup = s; |
| 430 | else |
| 431 | head = s; |
| 432 | prev = s; |
| 433 | } |
| 434 | if (prev) |
| 435 | prev->nextInSectionGroup = head; |
| 436 | } |
| 437 | |
| 438 | template <class ELFT> void ObjFile<ELFT>::initDwarf() { |
| 439 | dwarf = std::make_unique<DWARFCache>(std::make_unique<DWARFContext>( |
| 440 | std::make_unique<LLDDwarfObj<ELFT>>(this), "" , |
| 441 | [&](Error err) { Warn(ctx) << getName() + ": " << std::move(err); }, |
| 442 | [&](Error warning) { |
| 443 | Warn(ctx) << getName() << ": " << std::move(warning); |
| 444 | })); |
| 445 | } |
| 446 | |
| 447 | DWARFCache *ELFFileBase::getDwarf() { |
| 448 | assert(fileKind == ObjKind); |
| 449 | llvm::call_once(flag&: initDwarf, F: [this]() { |
| 450 | switch (ekind) { |
| 451 | default: |
| 452 | llvm_unreachable("" ); |
| 453 | case ELF32LEKind: |
| 454 | return cast<ObjFile<ELF32LE>>(Val: this)->initDwarf(); |
| 455 | case ELF32BEKind: |
| 456 | return cast<ObjFile<ELF32BE>>(Val: this)->initDwarf(); |
| 457 | case ELF64LEKind: |
| 458 | return cast<ObjFile<ELF64LE>>(Val: this)->initDwarf(); |
| 459 | case ELF64BEKind: |
| 460 | return cast<ObjFile<ELF64BE>>(Val: this)->initDwarf(); |
| 461 | } |
| 462 | }); |
| 463 | return dwarf.get(); |
| 464 | } |
| 465 | |
| 466 | ELFFileBase::ELFFileBase(Ctx &ctx, Kind k, ELFKind ekind, MemoryBufferRef mb) |
| 467 | : InputFile(ctx, k, mb) { |
| 468 | this->ekind = ekind; |
| 469 | } |
| 470 | |
| 471 | ELFFileBase::~ELFFileBase() {} |
| 472 | |
| 473 | template <typename Elf_Shdr> |
| 474 | static const Elf_Shdr *findSection(ArrayRef<Elf_Shdr> sections, uint32_t type) { |
| 475 | for (const Elf_Shdr &sec : sections) |
| 476 | if (sec.sh_type == type) |
| 477 | return &sec; |
| 478 | return nullptr; |
| 479 | } |
| 480 | |
| 481 | void ELFFileBase::init() { |
| 482 | switch (ekind) { |
| 483 | case ELF32LEKind: |
| 484 | init<ELF32LE>(k: fileKind); |
| 485 | break; |
| 486 | case ELF32BEKind: |
| 487 | init<ELF32BE>(k: fileKind); |
| 488 | break; |
| 489 | case ELF64LEKind: |
| 490 | init<ELF64LE>(k: fileKind); |
| 491 | break; |
| 492 | case ELF64BEKind: |
| 493 | init<ELF64BE>(k: fileKind); |
| 494 | break; |
| 495 | default: |
| 496 | llvm_unreachable("getELFKind" ); |
| 497 | } |
| 498 | } |
| 499 | |
| 500 | template <class ELFT> void ELFFileBase::init(InputFile::Kind k) { |
| 501 | using Elf_Shdr = typename ELFT::Shdr; |
| 502 | using Elf_Sym = typename ELFT::Sym; |
| 503 | |
| 504 | // Initialize trivial attributes. |
| 505 | const ELFFile<ELFT> &obj = getObj<ELFT>(); |
| 506 | emachine = obj.getHeader().e_machine; |
| 507 | osabi = obj.getHeader().e_ident[llvm::ELF::EI_OSABI]; |
| 508 | abiVersion = obj.getHeader().e_ident[llvm::ELF::EI_ABIVERSION]; |
| 509 | |
| 510 | ArrayRef<Elf_Shdr> sections = CHECK2(obj.sections(), this); |
| 511 | elfShdrs = sections.data(); |
| 512 | numELFShdrs = sections.size(); |
| 513 | |
| 514 | // Find a symbol table. |
| 515 | const Elf_Shdr *symtabSec = |
| 516 | findSection(sections, k == SharedKind ? SHT_DYNSYM : SHT_SYMTAB); |
| 517 | |
| 518 | if (!symtabSec) |
| 519 | return; |
| 520 | |
| 521 | // Initialize members corresponding to a symbol table. |
| 522 | firstGlobal = symtabSec->sh_info; |
| 523 | |
| 524 | ArrayRef<Elf_Sym> eSyms = CHECK2(obj.symbols(symtabSec), this); |
| 525 | if (firstGlobal == 0 || firstGlobal > eSyms.size()) |
| 526 | Fatal(ctx) << this << ": invalid sh_info in symbol table" ; |
| 527 | |
| 528 | elfSyms = reinterpret_cast<const void *>(eSyms.data()); |
| 529 | numSymbols = eSyms.size(); |
| 530 | stringTable = CHECK2(obj.getStringTableForSymtab(*symtabSec, sections), this); |
| 531 | } |
| 532 | |
| 533 | template <class ELFT> |
| 534 | uint32_t ObjFile<ELFT>::getSectionIndex(const Elf_Sym &sym) const { |
| 535 | return CHECK2( |
| 536 | this->getObj().getSectionIndex(sym, getELFSyms<ELFT>(), shndxTable), |
| 537 | this); |
| 538 | } |
| 539 | |
| 540 | template <class ELFT> void ObjFile<ELFT>::parse(bool ignoreComdats) { |
| 541 | object::ELFFile<ELFT> obj = this->getObj(); |
| 542 | // Read a section table. justSymbols is usually false. |
| 543 | if (this->justSymbols) { |
| 544 | initializeJustSymbols(); |
| 545 | initializeSymbols(obj); |
| 546 | return; |
| 547 | } |
| 548 | |
| 549 | // Handle dependent libraries and selection of section groups as these are not |
| 550 | // done in parallel. |
| 551 | ArrayRef<Elf_Shdr> objSections = getELFShdrs<ELFT>(); |
| 552 | StringRef shstrtab = CHECK2(obj.getSectionStringTable(objSections), this); |
| 553 | uint64_t size = objSections.size(); |
| 554 | sections.resize(size); |
| 555 | for (size_t i = 0; i != size; ++i) { |
| 556 | const Elf_Shdr &sec = objSections[i]; |
| 557 | if (LLVM_LIKELY(sec.sh_type == SHT_PROGBITS)) |
| 558 | continue; |
| 559 | if (LLVM_LIKELY(sec.sh_type == SHT_GROUP)) { |
| 560 | StringRef signature = getShtGroupSignature(sections: objSections, sec); |
| 561 | ArrayRef<Elf_Word> entries = |
| 562 | CHECK2(obj.template getSectionContentsAsArray<Elf_Word>(sec), this); |
| 563 | if (entries.empty()) |
| 564 | Fatal(ctx) << this << ": empty SHT_GROUP" ; |
| 565 | |
| 566 | Elf_Word flag = entries[0]; |
| 567 | if (flag && flag != GRP_COMDAT) |
| 568 | Fatal(ctx) << this << ": unsupported SHT_GROUP format" ; |
| 569 | |
| 570 | bool keepGroup = !flag || ignoreComdats || |
| 571 | ctx.symtab->comdatGroups |
| 572 | .try_emplace(CachedHashStringRef(signature), this) |
| 573 | .second; |
| 574 | if (keepGroup) { |
| 575 | if (!ctx.arg.resolveGroups) |
| 576 | sections[i] = createInputSection( |
| 577 | idx: i, sec, name: check(obj.getSectionName(sec, shstrtab))); |
| 578 | } else { |
| 579 | // Otherwise, discard group members. |
| 580 | for (uint32_t secIndex : entries.slice(1)) { |
| 581 | if (secIndex >= size) |
| 582 | Fatal(ctx) << this |
| 583 | << ": invalid section index in group: " << secIndex; |
| 584 | sections[secIndex] = &InputSection::discarded; |
| 585 | } |
| 586 | } |
| 587 | continue; |
| 588 | } |
| 589 | |
| 590 | if (sec.sh_type == SHT_LLVM_DEPENDENT_LIBRARIES && !ctx.arg.relocatable) { |
| 591 | StringRef name = check(obj.getSectionName(sec, shstrtab)); |
| 592 | ArrayRef<char> data = CHECK2( |
| 593 | this->getObj().template getSectionContentsAsArray<char>(sec), this); |
| 594 | if (!data.empty() && data.back() != '\0') { |
| 595 | Err(ctx) |
| 596 | << this |
| 597 | << ": corrupted dependent libraries section (unterminated string): " |
| 598 | << name; |
| 599 | } else { |
| 600 | for (const char *d = data.begin(), *e = data.end(); d < e;) { |
| 601 | StringRef s(d); |
| 602 | addDependentLibrary(ctx, s, this); |
| 603 | d += s.size() + 1; |
| 604 | } |
| 605 | } |
| 606 | sections[i] = &InputSection::discarded; |
| 607 | continue; |
| 608 | } |
| 609 | |
| 610 | switch (ctx.arg.emachine) { |
| 611 | case EM_ARM: |
| 612 | if (sec.sh_type == SHT_ARM_ATTRIBUTES) { |
| 613 | ARMAttributeParser attributes; |
| 614 | ArrayRef<uint8_t> contents = |
| 615 | check(this->getObj().getSectionContents(sec)); |
| 616 | StringRef name = check(obj.getSectionName(sec, shstrtab)); |
| 617 | sections[i] = &InputSection::discarded; |
| 618 | if (Error e = attributes.parse(section: contents, endian: ekind == ELF32LEKind |
| 619 | ? llvm::endianness::little |
| 620 | : llvm::endianness::big)) { |
| 621 | InputSection isec(*this, sec, name); |
| 622 | Warn(ctx) << &isec << ": " << std::move(e); |
| 623 | } else { |
| 624 | updateSupportedARMFeatures(ctx, attributes); |
| 625 | updateARMVFPArgs(ctx, attributes, this); |
| 626 | |
| 627 | // FIXME: Retain the first attribute section we see. The eglibc ARM |
| 628 | // dynamic loaders require the presence of an attribute section for |
| 629 | // dlopen to work. In a full implementation we would merge all |
| 630 | // attribute sections. |
| 631 | if (ctx.in.attributes == nullptr) { |
| 632 | ctx.in.attributes = |
| 633 | std::make_unique<InputSection>(*this, sec, name); |
| 634 | sections[i] = ctx.in.attributes.get(); |
| 635 | } |
| 636 | } |
| 637 | } |
| 638 | break; |
| 639 | case EM_AARCH64: |
| 640 | // FIXME: BuildAttributes have been implemented in llvm, but not yet in |
| 641 | // lld. Remove the section so that it does not accumulate in the output |
| 642 | // file. When support is implemented we expect not to output a build |
| 643 | // attributes section in files of type ET_EXEC or ET_SHARED, but ld -r |
| 644 | // ouptut will need a single merged attributes section. |
| 645 | if (sec.sh_type == SHT_AARCH64_ATTRIBUTES) |
| 646 | sections[i] = &InputSection::discarded; |
| 647 | // Producing a static binary with MTE globals is not currently supported, |
| 648 | // remove all SHT_AARCH64_MEMTAG_GLOBALS_STATIC sections as they're unused |
| 649 | // medatada, and we don't want them to end up in the output file for |
| 650 | // static executables. |
| 651 | if (sec.sh_type == SHT_AARCH64_MEMTAG_GLOBALS_STATIC && |
| 652 | !canHaveMemtagGlobals(ctx)) |
| 653 | sections[i] = &InputSection::discarded; |
| 654 | break; |
| 655 | } |
| 656 | } |
| 657 | |
| 658 | // Read a symbol table. |
| 659 | initializeSymbols(obj); |
| 660 | } |
| 661 | |
| 662 | // Sections with SHT_GROUP and comdat bits define comdat section groups. |
| 663 | // They are identified and deduplicated by group name. This function |
| 664 | // returns a group name. |
| 665 | template <class ELFT> |
| 666 | StringRef ObjFile<ELFT>::getShtGroupSignature(ArrayRef<Elf_Shdr> sections, |
| 667 | const Elf_Shdr &sec) { |
| 668 | typename ELFT::SymRange symbols = this->getELFSyms<ELFT>(); |
| 669 | if (sec.sh_info >= symbols.size()) |
| 670 | Fatal(ctx) << this << ": invalid symbol index" ; |
| 671 | const typename ELFT::Sym &sym = symbols[sec.sh_info]; |
| 672 | return CHECK2(sym.getName(this->stringTable), this); |
| 673 | } |
| 674 | |
| 675 | template <class ELFT> |
| 676 | bool ObjFile<ELFT>::shouldMerge(const Elf_Shdr &sec, StringRef name) { |
| 677 | // On a regular link we don't merge sections if -O0 (default is -O1). This |
| 678 | // sometimes makes the linker significantly faster, although the output will |
| 679 | // be bigger. |
| 680 | // |
| 681 | // Doing the same for -r would create a problem as it would combine sections |
| 682 | // with different sh_entsize. One option would be to just copy every SHF_MERGE |
| 683 | // section as is to the output. While this would produce a valid ELF file with |
| 684 | // usable SHF_MERGE sections, tools like (llvm-)?dwarfdump get confused when |
| 685 | // they see two .debug_str. We could have separate logic for combining |
| 686 | // SHF_MERGE sections based both on their name and sh_entsize, but that seems |
| 687 | // to be more trouble than it is worth. Instead, we just use the regular (-O1) |
| 688 | // logic for -r. |
| 689 | if (ctx.arg.optimize == 0 && !ctx.arg.relocatable) |
| 690 | return false; |
| 691 | |
| 692 | // A mergeable section with size 0 is useless because they don't have |
| 693 | // any data to merge. A mergeable string section with size 0 can be |
| 694 | // argued as invalid because it doesn't end with a null character. |
| 695 | // We'll avoid a mess by handling them as if they were non-mergeable. |
| 696 | if (sec.sh_size == 0) |
| 697 | return false; |
| 698 | |
| 699 | // Check for sh_entsize. The ELF spec is not clear about the zero |
| 700 | // sh_entsize. It says that "the member [sh_entsize] contains 0 if |
| 701 | // the section does not hold a table of fixed-size entries". We know |
| 702 | // that Rust 1.13 produces a string mergeable section with a zero |
| 703 | // sh_entsize. Here we just accept it rather than being picky about it. |
| 704 | uint64_t entSize = sec.sh_entsize; |
| 705 | if (entSize == 0) |
| 706 | return false; |
| 707 | if (sec.sh_size % entSize) |
| 708 | ErrAlways(ctx) << this << ":(" << name << "): SHF_MERGE section size (" |
| 709 | << uint64_t(sec.sh_size) |
| 710 | << ") must be a multiple of sh_entsize (" << entSize << ")" ; |
| 711 | if (sec.sh_flags & SHF_WRITE) |
| 712 | Err(ctx) << this << ":(" << name |
| 713 | << "): writable SHF_MERGE section is not supported" ; |
| 714 | |
| 715 | return true; |
| 716 | } |
| 717 | |
| 718 | // This is for --just-symbols. |
| 719 | // |
| 720 | // --just-symbols is a very minor feature that allows you to link your |
| 721 | // output against other existing program, so that if you load both your |
| 722 | // program and the other program into memory, your output can refer the |
| 723 | // other program's symbols. |
| 724 | // |
| 725 | // When the option is given, we link "just symbols". The section table is |
| 726 | // initialized with null pointers. |
| 727 | template <class ELFT> void ObjFile<ELFT>::initializeJustSymbols() { |
| 728 | sections.resize(numELFShdrs); |
| 729 | } |
| 730 | |
| 731 | static bool isKnownSpecificSectionType(uint32_t t, uint32_t flags) { |
| 732 | if (SHT_LOUSER <= t && t <= SHT_HIUSER && !(flags & SHF_ALLOC)) |
| 733 | return true; |
| 734 | if (SHT_LOOS <= t && t <= SHT_HIOS && !(flags & SHF_OS_NONCONFORMING)) |
| 735 | return true; |
| 736 | // Allow all processor-specific types. This is different from GNU ld. |
| 737 | return SHT_LOPROC <= t && t <= SHT_HIPROC; |
| 738 | } |
| 739 | |
| 740 | template <class ELFT> |
| 741 | void ObjFile<ELFT>::initializeSections(bool ignoreComdats, |
| 742 | const llvm::object::ELFFile<ELFT> &obj) { |
| 743 | ArrayRef<Elf_Shdr> objSections = getELFShdrs<ELFT>(); |
| 744 | StringRef shstrtab = CHECK2(obj.getSectionStringTable(objSections), this); |
| 745 | uint64_t size = objSections.size(); |
| 746 | SmallVector<ArrayRef<Elf_Word>, 0> selectedGroups; |
| 747 | for (size_t i = 0; i != size; ++i) { |
| 748 | if (this->sections[i] == &InputSection::discarded) |
| 749 | continue; |
| 750 | const Elf_Shdr &sec = objSections[i]; |
| 751 | const uint32_t type = sec.sh_type; |
| 752 | |
| 753 | // SHF_EXCLUDE'ed sections are discarded by the linker. However, |
| 754 | // if -r is given, we'll let the final link discard such sections. |
| 755 | // This is compatible with GNU. |
| 756 | if ((sec.sh_flags & SHF_EXCLUDE) && !ctx.arg.relocatable) { |
| 757 | if (type == SHT_LLVM_CALL_GRAPH_PROFILE) |
| 758 | cgProfileSectionIndex = i; |
| 759 | if (type == SHT_LLVM_ADDRSIG) { |
| 760 | // We ignore the address-significance table if we know that the object |
| 761 | // file was created by objcopy or ld -r. This is because these tools |
| 762 | // will reorder the symbols in the symbol table, invalidating the data |
| 763 | // in the address-significance table, which refers to symbols by index. |
| 764 | if (sec.sh_link != 0) |
| 765 | this->addrsigSec = &sec; |
| 766 | else if (ctx.arg.icf == ICFLevel::Safe) |
| 767 | Warn(ctx) << this |
| 768 | << ": --icf=safe conservatively ignores " |
| 769 | "SHT_LLVM_ADDRSIG [index " |
| 770 | << i |
| 771 | << "] with sh_link=0 " |
| 772 | "(likely created using objcopy or ld -r)" ; |
| 773 | } |
| 774 | this->sections[i] = &InputSection::discarded; |
| 775 | continue; |
| 776 | } |
| 777 | |
| 778 | switch (type) { |
| 779 | case SHT_GROUP: { |
| 780 | if (!ctx.arg.relocatable) |
| 781 | sections[i] = &InputSection::discarded; |
| 782 | StringRef signature = |
| 783 | cantFail(this->getELFSyms<ELFT>()[sec.sh_info].getName(stringTable)); |
| 784 | ArrayRef<Elf_Word> entries = |
| 785 | cantFail(obj.template getSectionContentsAsArray<Elf_Word>(sec)); |
| 786 | if ((entries[0] & GRP_COMDAT) == 0 || ignoreComdats || |
| 787 | ctx.symtab->comdatGroups.find(Val: CachedHashStringRef(signature)) |
| 788 | ->second == this) |
| 789 | selectedGroups.push_back(entries); |
| 790 | break; |
| 791 | } |
| 792 | case SHT_SYMTAB_SHNDX: |
| 793 | shndxTable = CHECK2(obj.getSHNDXTable(sec, objSections), this); |
| 794 | break; |
| 795 | case SHT_SYMTAB: |
| 796 | case SHT_STRTAB: |
| 797 | case SHT_REL: |
| 798 | case SHT_RELA: |
| 799 | case SHT_CREL: |
| 800 | case SHT_NULL: |
| 801 | break; |
| 802 | case SHT_PROGBITS: |
| 803 | case SHT_NOTE: |
| 804 | case SHT_NOBITS: |
| 805 | case SHT_INIT_ARRAY: |
| 806 | case SHT_FINI_ARRAY: |
| 807 | case SHT_PREINIT_ARRAY: |
| 808 | this->sections[i] = |
| 809 | createInputSection(idx: i, sec, name: check(obj.getSectionName(sec, shstrtab))); |
| 810 | break; |
| 811 | case SHT_LLVM_LTO: |
| 812 | // Discard .llvm.lto in a relocatable link that does not use the bitcode. |
| 813 | // The concatenated output does not properly reflect the linking |
| 814 | // semantics. In addition, since we do not use the bitcode wrapper format, |
| 815 | // the concatenated raw bitcode would be invalid. |
| 816 | if (ctx.arg.relocatable && !ctx.arg.fatLTOObjects) { |
| 817 | sections[i] = &InputSection::discarded; |
| 818 | break; |
| 819 | } |
| 820 | [[fallthrough]]; |
| 821 | default: |
| 822 | this->sections[i] = |
| 823 | createInputSection(idx: i, sec, name: check(obj.getSectionName(sec, shstrtab))); |
| 824 | if (type == SHT_LLVM_SYMPART) |
| 825 | ctx.hasSympart.store(true, std::memory_order_relaxed); |
| 826 | else if (ctx.arg.rejectMismatch && |
| 827 | !isKnownSpecificSectionType(type, sec.sh_flags)) |
| 828 | Err(ctx) << this->sections[i] << ": unknown section type 0x" |
| 829 | << Twine::utohexstr(Val: type); |
| 830 | break; |
| 831 | } |
| 832 | } |
| 833 | |
| 834 | // We have a second loop. It is used to: |
| 835 | // 1) handle SHF_LINK_ORDER sections. |
| 836 | // 2) create relocation sections. In some cases the section header index of a |
| 837 | // relocation section may be smaller than that of the relocated section. In |
| 838 | // such cases, the relocation section would attempt to reference a target |
| 839 | // section that has not yet been created. For simplicity, delay creation of |
| 840 | // relocation sections until now. |
| 841 | for (size_t i = 0; i != size; ++i) { |
| 842 | if (this->sections[i] == &InputSection::discarded) |
| 843 | continue; |
| 844 | const Elf_Shdr &sec = objSections[i]; |
| 845 | |
| 846 | if (isStaticRelSecType(sec.sh_type)) { |
| 847 | // Find a relocation target section and associate this section with that. |
| 848 | // Target may have been discarded if it is in a different section group |
| 849 | // and the group is discarded, even though it's a violation of the spec. |
| 850 | // We handle that situation gracefully by discarding dangling relocation |
| 851 | // sections. |
| 852 | const uint32_t info = sec.sh_info; |
| 853 | InputSectionBase *s = getRelocTarget(idx: i, info); |
| 854 | if (!s) |
| 855 | continue; |
| 856 | |
| 857 | // ELF spec allows mergeable sections with relocations, but they are rare, |
| 858 | // and it is in practice hard to merge such sections by contents, because |
| 859 | // applying relocations at end of linking changes section contents. So, we |
| 860 | // simply handle such sections as non-mergeable ones. Degrading like this |
| 861 | // is acceptable because section merging is optional. |
| 862 | if (auto *ms = dyn_cast<MergeInputSection>(Val: s)) { |
| 863 | s = makeThreadLocal<InputSection>(args&: ms->file, args&: ms->name, args&: ms->type, |
| 864 | args&: ms->flags, args&: ms->addralign, args&: ms->entsize, |
| 865 | args: ms->contentMaybeDecompress()); |
| 866 | sections[info] = s; |
| 867 | } |
| 868 | |
| 869 | if (s->relSecIdx != 0) |
| 870 | ErrAlways(ctx) << s |
| 871 | << ": multiple relocation sections to one section are " |
| 872 | "not supported" ; |
| 873 | s->relSecIdx = i; |
| 874 | |
| 875 | // Relocation sections are usually removed from the output, so return |
| 876 | // `nullptr` for the normal case. However, if -r or --emit-relocs is |
| 877 | // specified, we need to copy them to the output. (Some post link analysis |
| 878 | // tools specify --emit-relocs to obtain the information.) |
| 879 | if (ctx.arg.copyRelocs) { |
| 880 | auto *isec = makeThreadLocal<InputSection>( |
| 881 | *this, sec, check(obj.getSectionName(sec, shstrtab))); |
| 882 | // If the relocated section is discarded (due to /DISCARD/ or |
| 883 | // --gc-sections), the relocation section should be discarded as well. |
| 884 | s->dependentSections.push_back(NewVal: isec); |
| 885 | sections[i] = isec; |
| 886 | } |
| 887 | continue; |
| 888 | } |
| 889 | |
| 890 | // A SHF_LINK_ORDER section with sh_link=0 is handled as if it did not have |
| 891 | // the flag. |
| 892 | if (!sec.sh_link || !(sec.sh_flags & SHF_LINK_ORDER)) |
| 893 | continue; |
| 894 | |
| 895 | InputSectionBase *linkSec = nullptr; |
| 896 | if (sec.sh_link < size) |
| 897 | linkSec = this->sections[sec.sh_link]; |
| 898 | if (!linkSec) { |
| 899 | ErrAlways(ctx) << this |
| 900 | << ": invalid sh_link index: " << uint32_t(sec.sh_link); |
| 901 | continue; |
| 902 | } |
| 903 | |
| 904 | // A SHF_LINK_ORDER section is discarded if its linked-to section is |
| 905 | // discarded. |
| 906 | InputSection *isec = cast<InputSection>(this->sections[i]); |
| 907 | linkSec->dependentSections.push_back(NewVal: isec); |
| 908 | if (!isa<InputSection>(Val: linkSec)) |
| 909 | ErrAlways(ctx) |
| 910 | << "a section " << isec->name |
| 911 | << " with SHF_LINK_ORDER should not refer a non-regular section: " |
| 912 | << linkSec; |
| 913 | } |
| 914 | |
| 915 | for (ArrayRef<Elf_Word> entries : selectedGroups) |
| 916 | handleSectionGroup<ELFT>(this->sections, entries); |
| 917 | } |
| 918 | |
| 919 | template <typename ELFT> |
| 920 | static void parseGnuPropertyNote(Ctx &ctx, ELFFileBase &f, |
| 921 | uint32_t featureAndType, |
| 922 | ArrayRef<uint8_t> &desc, const uint8_t *base, |
| 923 | ArrayRef<uint8_t> *data = nullptr) { |
| 924 | auto err = [&](const uint8_t *place) -> ELFSyncStream { |
| 925 | auto diag = Err(ctx); |
| 926 | diag << &f << ":(" << ".note.gnu.property+0x" |
| 927 | << Twine::utohexstr(Val: place - base) << "): " ; |
| 928 | return diag; |
| 929 | }; |
| 930 | |
| 931 | while (!desc.empty()) { |
| 932 | const uint8_t *place = desc.data(); |
| 933 | if (desc.size() < 8) |
| 934 | return void(err(place) << "program property is too short" ); |
| 935 | uint32_t type = read32<ELFT::Endianness>(desc.data()); |
| 936 | uint32_t size = read32<ELFT::Endianness>(desc.data() + 4); |
| 937 | desc = desc.slice(N: 8); |
| 938 | if (desc.size() < size) |
| 939 | return void(err(place) << "program property is too short" ); |
| 940 | |
| 941 | if (type == featureAndType) { |
| 942 | // We found a FEATURE_1_AND field. There may be more than one of these |
| 943 | // in a .note.gnu.property section, for a relocatable object we |
| 944 | // accumulate the bits set. |
| 945 | if (size < 4) |
| 946 | return void(err(place) << "FEATURE_1_AND entry is too short" ); |
| 947 | f.andFeatures |= read32<ELFT::Endianness>(desc.data()); |
| 948 | } else if (ctx.arg.emachine == EM_AARCH64 && |
| 949 | type == GNU_PROPERTY_AARCH64_FEATURE_PAUTH) { |
| 950 | ArrayRef<uint8_t> contents = data ? *data : desc; |
| 951 | if (f.aarch64PauthAbiCoreInfo) { |
| 952 | return void( |
| 953 | err(contents.data()) |
| 954 | << "multiple GNU_PROPERTY_AARCH64_FEATURE_PAUTH entries are " |
| 955 | "not supported" ); |
| 956 | } else if (size != 16) { |
| 957 | return void(err(contents.data()) |
| 958 | << "GNU_PROPERTY_AARCH64_FEATURE_PAUTH entry " |
| 959 | "is invalid: expected 16 bytes, but got " |
| 960 | << size); |
| 961 | } |
| 962 | f.aarch64PauthAbiCoreInfo = { |
| 963 | support::endian::read64<ELFT::Endianness>(&desc[0]), |
| 964 | support::endian::read64<ELFT::Endianness>(&desc[8])}; |
| 965 | } |
| 966 | |
| 967 | // Padding is present in the note descriptor, if necessary. |
| 968 | desc = desc.slice(alignTo<(ELFT::Is64Bits ? 8 : 4)>(size)); |
| 969 | } |
| 970 | } |
| 971 | // Read the following info from the .note.gnu.property section and write it to |
| 972 | // the corresponding fields in `ObjFile`: |
| 973 | // - Feature flags (32 bits) representing x86, AArch64 or RISC-V features for |
| 974 | // hardware-assisted call flow control; |
| 975 | // - AArch64 PAuth ABI core info (16 bytes). |
| 976 | template <class ELFT> |
| 977 | static void readGnuProperty(Ctx &ctx, const InputSection &sec, |
| 978 | ObjFile<ELFT> &f) { |
| 979 | using Elf_Nhdr = typename ELFT::Nhdr; |
| 980 | using Elf_Note = typename ELFT::Note; |
| 981 | |
| 982 | uint32_t featureAndType; |
| 983 | switch (ctx.arg.emachine) { |
| 984 | case EM_386: |
| 985 | case EM_X86_64: |
| 986 | featureAndType = GNU_PROPERTY_X86_FEATURE_1_AND; |
| 987 | break; |
| 988 | case EM_AARCH64: |
| 989 | featureAndType = GNU_PROPERTY_AARCH64_FEATURE_1_AND; |
| 990 | break; |
| 991 | case EM_RISCV: |
| 992 | featureAndType = GNU_PROPERTY_RISCV_FEATURE_1_AND; |
| 993 | break; |
| 994 | default: |
| 995 | return; |
| 996 | } |
| 997 | |
| 998 | ArrayRef<uint8_t> data = sec.content(); |
| 999 | auto err = [&](const uint8_t *place) -> ELFSyncStream { |
| 1000 | auto diag = Err(ctx); |
| 1001 | diag << sec.file << ":(" << sec.name << "+0x" |
| 1002 | << Twine::utohexstr(Val: place - sec.content().data()) << "): " ; |
| 1003 | return diag; |
| 1004 | }; |
| 1005 | while (!data.empty()) { |
| 1006 | // Read one NOTE record. |
| 1007 | auto *nhdr = reinterpret_cast<const Elf_Nhdr *>(data.data()); |
| 1008 | if (data.size() < sizeof(Elf_Nhdr) || |
| 1009 | data.size() < nhdr->getSize(sec.addralign)) |
| 1010 | return void(err(data.data()) << "data is too short" ); |
| 1011 | |
| 1012 | Elf_Note note(*nhdr); |
| 1013 | if (nhdr->n_type != NT_GNU_PROPERTY_TYPE_0 || note.getName() != "GNU" ) { |
| 1014 | data = data.slice(nhdr->getSize(sec.addralign)); |
| 1015 | continue; |
| 1016 | } |
| 1017 | |
| 1018 | // Read a body of a NOTE record, which consists of type-length-value fields. |
| 1019 | ArrayRef<uint8_t> desc = note.getDesc(sec.addralign); |
| 1020 | const uint8_t *base = sec.content().data(); |
| 1021 | parseGnuPropertyNote<ELFT>(ctx, f, featureAndType, desc, base, &data); |
| 1022 | |
| 1023 | // Go to next NOTE record to look for more FEATURE_1_AND descriptions. |
| 1024 | data = data.slice(nhdr->getSize(sec.addralign)); |
| 1025 | } |
| 1026 | } |
| 1027 | |
| 1028 | template <class ELFT> |
| 1029 | InputSectionBase *ObjFile<ELFT>::getRelocTarget(uint32_t idx, uint32_t info) { |
| 1030 | if (info < this->sections.size()) { |
| 1031 | InputSectionBase *target = this->sections[info]; |
| 1032 | |
| 1033 | // Strictly speaking, a relocation section must be included in the |
| 1034 | // group of the section it relocates. However, LLVM 3.3 and earlier |
| 1035 | // would fail to do so, so we gracefully handle that case. |
| 1036 | if (target == &InputSection::discarded) |
| 1037 | return nullptr; |
| 1038 | |
| 1039 | if (target != nullptr) |
| 1040 | return target; |
| 1041 | } |
| 1042 | |
| 1043 | Err(ctx) << this << ": relocation section (index " << idx |
| 1044 | << ") has invalid sh_info (" << info << ')'; |
| 1045 | return nullptr; |
| 1046 | } |
| 1047 | |
| 1048 | // The function may be called concurrently for different input files. For |
| 1049 | // allocation, prefer makeThreadLocal which does not require holding a lock. |
| 1050 | template <class ELFT> |
| 1051 | InputSectionBase *ObjFile<ELFT>::createInputSection(uint32_t idx, |
| 1052 | const Elf_Shdr &sec, |
| 1053 | StringRef name) { |
| 1054 | if (name.starts_with(Prefix: ".n" )) { |
| 1055 | // The GNU linker uses .note.GNU-stack section as a marker indicating |
| 1056 | // that the code in the object file does not expect that the stack is |
| 1057 | // executable (in terms of NX bit). If all input files have the marker, |
| 1058 | // the GNU linker adds a PT_GNU_STACK segment to tells the loader to |
| 1059 | // make the stack non-executable. Most object files have this section as |
| 1060 | // of 2017. |
| 1061 | // |
| 1062 | // But making the stack non-executable is a norm today for security |
| 1063 | // reasons. Failure to do so may result in a serious security issue. |
| 1064 | // Therefore, we make LLD always add PT_GNU_STACK unless it is |
| 1065 | // explicitly told to do otherwise (by -z execstack). Because the stack |
| 1066 | // executable-ness is controlled solely by command line options, |
| 1067 | // .note.GNU-stack sections are, with one exception, ignored. Report |
| 1068 | // an error if we encounter an executable .note.GNU-stack to force the |
| 1069 | // user to explicitly request an executable stack. |
| 1070 | if (name == ".note.GNU-stack" ) { |
| 1071 | if ((sec.sh_flags & SHF_EXECINSTR) && !ctx.arg.relocatable && |
| 1072 | ctx.arg.zGnustack != GnuStackKind::Exec) { |
| 1073 | Err(ctx) << this |
| 1074 | << ": requires an executable stack, but -z execstack is not " |
| 1075 | "specified" ; |
| 1076 | } |
| 1077 | return &InputSection::discarded; |
| 1078 | } |
| 1079 | |
| 1080 | // Object files that use processor features such as Intel Control-Flow |
| 1081 | // Enforcement (CET), AArch64 Branch Target Identification BTI or RISC-V |
| 1082 | // Zicfilp/Zicfiss extensions, use a .note.gnu.property section containing |
| 1083 | // a bitfield of feature bits like the GNU_PROPERTY_X86_FEATURE_1_IBT flag. |
| 1084 | // |
| 1085 | // Since we merge bitmaps from multiple object files to create a new |
| 1086 | // .note.gnu.property containing a single AND'ed bitmap, we discard an input |
| 1087 | // file's .note.gnu.property section. |
| 1088 | if (name == ".note.gnu.property" ) { |
| 1089 | readGnuProperty<ELFT>(ctx, InputSection(*this, sec, name), *this); |
| 1090 | return &InputSection::discarded; |
| 1091 | } |
| 1092 | |
| 1093 | // Split stacks is a feature to support a discontiguous stack, |
| 1094 | // commonly used in the programming language Go. For the details, |
| 1095 | // see https://gcc.gnu.org/wiki/SplitStacks. An object file compiled |
| 1096 | // for split stack will include a .note.GNU-split-stack section. |
| 1097 | if (name == ".note.GNU-split-stack" ) { |
| 1098 | if (ctx.arg.relocatable) { |
| 1099 | ErrAlways(ctx) << "cannot mix split-stack and non-split-stack in a " |
| 1100 | "relocatable link" ; |
| 1101 | return &InputSection::discarded; |
| 1102 | } |
| 1103 | this->splitStack = true; |
| 1104 | return &InputSection::discarded; |
| 1105 | } |
| 1106 | |
| 1107 | // An object file compiled for split stack, but where some of the |
| 1108 | // functions were compiled with the no_split_stack_attribute will |
| 1109 | // include a .note.GNU-no-split-stack section. |
| 1110 | if (name == ".note.GNU-no-split-stack" ) { |
| 1111 | this->someNoSplitStack = true; |
| 1112 | return &InputSection::discarded; |
| 1113 | } |
| 1114 | |
| 1115 | // Strip existing .note.gnu.build-id sections so that the output won't have |
| 1116 | // more than one build-id. This is not usually a problem because input |
| 1117 | // object files normally don't have .build-id sections, but you can create |
| 1118 | // such files by "ld.{bfd,gold,lld} -r --build-id", and we want to guard |
| 1119 | // against it. |
| 1120 | if (name == ".note.gnu.build-id" ) |
| 1121 | return &InputSection::discarded; |
| 1122 | } |
| 1123 | |
| 1124 | // The linker merges EH (exception handling) frames and creates a |
| 1125 | // .eh_frame_hdr section for runtime. So we handle them with a special |
| 1126 | // class. For relocatable outputs, they are just passed through. |
| 1127 | if (name == ".eh_frame" && !ctx.arg.relocatable) |
| 1128 | return makeThreadLocal<EhInputSection>(*this, sec, name); |
| 1129 | |
| 1130 | if ((sec.sh_flags & SHF_MERGE) && shouldMerge(sec, name)) |
| 1131 | return makeThreadLocal<MergeInputSection>(*this, sec, name); |
| 1132 | return makeThreadLocal<InputSection>(*this, sec, name); |
| 1133 | } |
| 1134 | |
| 1135 | // Initialize symbols. symbols is a parallel array to the corresponding ELF |
| 1136 | // symbol table. |
| 1137 | template <class ELFT> |
| 1138 | void ObjFile<ELFT>::initializeSymbols(const object::ELFFile<ELFT> &obj) { |
| 1139 | ArrayRef<Elf_Sym> eSyms = this->getELFSyms<ELFT>(); |
| 1140 | if (!symbols) |
| 1141 | symbols = std::make_unique<Symbol *[]>(numSymbols); |
| 1142 | |
| 1143 | // Some entries have been filled by LazyObjFile. |
| 1144 | auto *symtab = ctx.symtab.get(); |
| 1145 | for (size_t i = firstGlobal, end = eSyms.size(); i != end; ++i) |
| 1146 | if (!symbols[i]) |
| 1147 | symbols[i] = symtab->insert(CHECK2(eSyms[i].getName(stringTable), this)); |
| 1148 | |
| 1149 | // Perform symbol resolution on non-local symbols. |
| 1150 | SmallVector<unsigned, 32> undefineds; |
| 1151 | for (size_t i = firstGlobal, end = eSyms.size(); i != end; ++i) { |
| 1152 | const Elf_Sym &eSym = eSyms[i]; |
| 1153 | uint32_t secIdx = eSym.st_shndx; |
| 1154 | if (secIdx == SHN_UNDEF) { |
| 1155 | undefineds.push_back(Elt: i); |
| 1156 | continue; |
| 1157 | } |
| 1158 | |
| 1159 | uint8_t binding = eSym.getBinding(); |
| 1160 | uint8_t stOther = eSym.st_other; |
| 1161 | uint8_t type = eSym.getType(); |
| 1162 | uint64_t value = eSym.st_value; |
| 1163 | uint64_t size = eSym.st_size; |
| 1164 | |
| 1165 | Symbol *sym = symbols[i]; |
| 1166 | sym->isUsedInRegularObj = true; |
| 1167 | if (LLVM_UNLIKELY(eSym.st_shndx == SHN_COMMON)) { |
| 1168 | if (value == 0 || value >= UINT32_MAX) |
| 1169 | Err(ctx) << this << ": common symbol '" << sym->getName() |
| 1170 | << "' has invalid alignment: " << value; |
| 1171 | hasCommonSyms = true; |
| 1172 | sym->resolve(ctx, CommonSymbol{ctx, this, StringRef(), binding, stOther, |
| 1173 | type, value, size}); |
| 1174 | continue; |
| 1175 | } |
| 1176 | |
| 1177 | // Handle global defined symbols. Defined::section will be set in postParse. |
| 1178 | sym->resolve(ctx, Defined{ctx, this, StringRef(), binding, stOther, type, |
| 1179 | value, size, nullptr}); |
| 1180 | } |
| 1181 | |
| 1182 | // Undefined symbols (excluding those defined relative to non-prevailing |
| 1183 | // sections) can trigger recursive extract. Process defined symbols first so |
| 1184 | // that the relative order between a defined symbol and an undefined symbol |
| 1185 | // does not change the symbol resolution behavior. In addition, a set of |
| 1186 | // interconnected symbols will all be resolved to the same file, instead of |
| 1187 | // being resolved to different files. |
| 1188 | for (unsigned i : undefineds) { |
| 1189 | const Elf_Sym &eSym = eSyms[i]; |
| 1190 | Symbol *sym = symbols[i]; |
| 1191 | sym->resolve(ctx, Undefined{this, StringRef(), eSym.getBinding(), |
| 1192 | eSym.st_other, eSym.getType()}); |
| 1193 | sym->isUsedInRegularObj = true; |
| 1194 | sym->referenced = true; |
| 1195 | } |
| 1196 | } |
| 1197 | |
| 1198 | template <class ELFT> |
| 1199 | void ObjFile<ELFT>::initSectionsAndLocalSyms(bool ignoreComdats) { |
| 1200 | if (!justSymbols) |
| 1201 | initializeSections(ignoreComdats, obj: getObj()); |
| 1202 | |
| 1203 | if (!firstGlobal) |
| 1204 | return; |
| 1205 | SymbolUnion *locals = makeThreadLocalN<SymbolUnion>(firstGlobal); |
| 1206 | memset(locals, 0, sizeof(SymbolUnion) * firstGlobal); |
| 1207 | |
| 1208 | ArrayRef<Elf_Sym> eSyms = this->getELFSyms<ELFT>(); |
| 1209 | for (size_t i = 0, end = firstGlobal; i != end; ++i) { |
| 1210 | const Elf_Sym &eSym = eSyms[i]; |
| 1211 | uint32_t secIdx = eSym.st_shndx; |
| 1212 | if (LLVM_UNLIKELY(secIdx == SHN_XINDEX)) |
| 1213 | secIdx = check(getExtendedSymbolTableIndex<ELFT>(eSym, i, shndxTable)); |
| 1214 | else if (secIdx >= SHN_LORESERVE) |
| 1215 | secIdx = 0; |
| 1216 | if (LLVM_UNLIKELY(secIdx >= sections.size())) { |
| 1217 | Err(ctx) << this << ": invalid section index: " << secIdx; |
| 1218 | secIdx = 0; |
| 1219 | } |
| 1220 | if (LLVM_UNLIKELY(eSym.getBinding() != STB_LOCAL)) |
| 1221 | ErrAlways(ctx) << this << ": non-local symbol (" << i |
| 1222 | << ") found at index < .symtab's sh_info (" << end << ")" ; |
| 1223 | |
| 1224 | InputSectionBase *sec = sections[secIdx]; |
| 1225 | uint8_t type = eSym.getType(); |
| 1226 | if (type == STT_FILE) |
| 1227 | sourceFile = CHECK2(eSym.getName(stringTable), this); |
| 1228 | unsigned stName = eSym.st_name; |
| 1229 | if (LLVM_UNLIKELY(stringTable.size() <= stName)) { |
| 1230 | Err(ctx) << this << ": invalid symbol name offset" ; |
| 1231 | stName = 0; |
| 1232 | } |
| 1233 | StringRef name(stringTable.data() + stName); |
| 1234 | |
| 1235 | symbols[i] = reinterpret_cast<Symbol *>(locals + i); |
| 1236 | if (eSym.st_shndx == SHN_UNDEF || sec == &InputSection::discarded) |
| 1237 | new (symbols[i]) Undefined(this, name, STB_LOCAL, eSym.st_other, type, |
| 1238 | /*discardedSecIdx=*/secIdx); |
| 1239 | else |
| 1240 | new (symbols[i]) Defined(ctx, this, name, STB_LOCAL, eSym.st_other, type, |
| 1241 | eSym.st_value, eSym.st_size, sec); |
| 1242 | symbols[i]->partition = 1; |
| 1243 | symbols[i]->isUsedInRegularObj = true; |
| 1244 | } |
| 1245 | } |
| 1246 | |
| 1247 | // Called after all ObjFile::parse is called for all ObjFiles. This checks |
| 1248 | // duplicate symbols and may do symbol property merge in the future. |
| 1249 | template <class ELFT> void ObjFile<ELFT>::postParse() { |
| 1250 | static std::mutex mu; |
| 1251 | ArrayRef<Elf_Sym> eSyms = this->getELFSyms<ELFT>(); |
| 1252 | for (size_t i = firstGlobal, end = eSyms.size(); i != end; ++i) { |
| 1253 | const Elf_Sym &eSym = eSyms[i]; |
| 1254 | Symbol &sym = *symbols[i]; |
| 1255 | uint32_t secIdx = eSym.st_shndx; |
| 1256 | uint8_t binding = eSym.getBinding(); |
| 1257 | if (LLVM_UNLIKELY(binding != STB_GLOBAL && binding != STB_WEAK && |
| 1258 | binding != STB_GNU_UNIQUE)) |
| 1259 | Err(ctx) << this << ": symbol (" << i |
| 1260 | << ") has invalid binding: " << (int)binding; |
| 1261 | |
| 1262 | // st_value of STT_TLS represents the assigned offset, not the actual |
| 1263 | // address which is used by STT_FUNC and STT_OBJECT. STT_TLS symbols can |
| 1264 | // only be referenced by special TLS relocations. It is usually an error if |
| 1265 | // a STT_TLS symbol is replaced by a non-STT_TLS symbol, vice versa. |
| 1266 | if (LLVM_UNLIKELY(sym.isTls()) && eSym.getType() != STT_TLS && |
| 1267 | eSym.getType() != STT_NOTYPE) |
| 1268 | Err(ctx) << "TLS attribute mismatch: " << &sym << "\n>>> in " << sym.file |
| 1269 | << "\n>>> in " << this; |
| 1270 | |
| 1271 | // Handle non-COMMON defined symbol below. !sym.file allows a symbol |
| 1272 | // assignment to redefine a symbol without an error. |
| 1273 | if (!sym.isDefined() || secIdx == SHN_UNDEF) |
| 1274 | continue; |
| 1275 | if (LLVM_UNLIKELY(secIdx >= SHN_LORESERVE)) { |
| 1276 | if (secIdx == SHN_COMMON) |
| 1277 | continue; |
| 1278 | if (secIdx == SHN_XINDEX) |
| 1279 | secIdx = check(getExtendedSymbolTableIndex<ELFT>(eSym, i, shndxTable)); |
| 1280 | else |
| 1281 | secIdx = 0; |
| 1282 | } |
| 1283 | |
| 1284 | if (LLVM_UNLIKELY(secIdx >= sections.size())) { |
| 1285 | Err(ctx) << this << ": invalid section index: " << secIdx; |
| 1286 | continue; |
| 1287 | } |
| 1288 | InputSectionBase *sec = sections[secIdx]; |
| 1289 | if (sec == &InputSection::discarded) { |
| 1290 | if (sym.traced) { |
| 1291 | printTraceSymbol(sym: Undefined{this, sym.getName(), sym.binding, |
| 1292 | sym.stOther, sym.type, secIdx}, |
| 1293 | name: sym.getName()); |
| 1294 | } |
| 1295 | if (sym.file == this) { |
| 1296 | std::lock_guard<std::mutex> lock(mu); |
| 1297 | ctx.nonPrevailingSyms.emplace_back(&sym, secIdx); |
| 1298 | } |
| 1299 | continue; |
| 1300 | } |
| 1301 | |
| 1302 | if (sym.file == this) { |
| 1303 | cast<Defined>(Val&: sym).section = sec; |
| 1304 | continue; |
| 1305 | } |
| 1306 | |
| 1307 | if (sym.binding == STB_WEAK || binding == STB_WEAK) |
| 1308 | continue; |
| 1309 | std::lock_guard<std::mutex> lock(mu); |
| 1310 | ctx.duplicates.push_back(Elt: {&sym, this, sec, eSym.st_value}); |
| 1311 | } |
| 1312 | } |
| 1313 | |
| 1314 | // The handling of tentative definitions (COMMON symbols) in archives is murky. |
| 1315 | // A tentative definition will be promoted to a global definition if there are |
| 1316 | // no non-tentative definitions to dominate it. When we hold a tentative |
| 1317 | // definition to a symbol and are inspecting archive members for inclusion |
| 1318 | // there are 2 ways we can proceed: |
| 1319 | // |
| 1320 | // 1) Consider the tentative definition a 'real' definition (ie promotion from |
| 1321 | // tentative to real definition has already happened) and not inspect |
| 1322 | // archive members for Global/Weak definitions to replace the tentative |
| 1323 | // definition. An archive member would only be included if it satisfies some |
| 1324 | // other undefined symbol. This is the behavior Gold uses. |
| 1325 | // |
| 1326 | // 2) Consider the tentative definition as still undefined (ie the promotion to |
| 1327 | // a real definition happens only after all symbol resolution is done). |
| 1328 | // The linker searches archive members for STB_GLOBAL definitions to |
| 1329 | // replace the tentative definition with. This is the behavior used by |
| 1330 | // GNU ld. |
| 1331 | // |
| 1332 | // The second behavior is inherited from SysVR4, which based it on the FORTRAN |
| 1333 | // COMMON BLOCK model. This behavior is needed for proper initialization in old |
| 1334 | // (pre F90) FORTRAN code that is packaged into an archive. |
| 1335 | // |
| 1336 | // The following functions search archive members for definitions to replace |
| 1337 | // tentative definitions (implementing behavior 2). |
| 1338 | static bool isBitcodeNonCommonDef(MemoryBufferRef mb, StringRef symName, |
| 1339 | StringRef archiveName) { |
| 1340 | IRSymtabFile symtabFile = check(e: readIRSymtab(MBRef: mb)); |
| 1341 | for (const irsymtab::Reader::SymbolRef &sym : |
| 1342 | symtabFile.TheReader.symbols()) { |
| 1343 | if (sym.isGlobal() && sym.getName() == symName) |
| 1344 | return !sym.isUndefined() && !sym.isWeak() && !sym.isCommon(); |
| 1345 | } |
| 1346 | return false; |
| 1347 | } |
| 1348 | |
| 1349 | template <class ELFT> |
| 1350 | static bool isNonCommonDef(Ctx &ctx, ELFKind ekind, MemoryBufferRef mb, |
| 1351 | StringRef symName, StringRef archiveName) { |
| 1352 | ObjFile<ELFT> *obj = make<ObjFile<ELFT>>(ctx, ekind, mb, archiveName); |
| 1353 | obj->init(); |
| 1354 | StringRef stringtable = obj->getStringTable(); |
| 1355 | |
| 1356 | for (auto sym : obj->template getGlobalELFSyms<ELFT>()) { |
| 1357 | Expected<StringRef> name = sym.getName(stringtable); |
| 1358 | if (name && name.get() == symName) |
| 1359 | return sym.isDefined() && sym.getBinding() == STB_GLOBAL && |
| 1360 | !sym.isCommon(); |
| 1361 | } |
| 1362 | return false; |
| 1363 | } |
| 1364 | |
| 1365 | static bool isNonCommonDef(Ctx &ctx, MemoryBufferRef mb, StringRef symName, |
| 1366 | StringRef archiveName) { |
| 1367 | switch (getELFKind(ctx, mb, archiveName)) { |
| 1368 | case ELF32LEKind: |
| 1369 | return isNonCommonDef<ELF32LE>(ctx, ekind: ELF32LEKind, mb, symName, archiveName); |
| 1370 | case ELF32BEKind: |
| 1371 | return isNonCommonDef<ELF32BE>(ctx, ekind: ELF32BEKind, mb, symName, archiveName); |
| 1372 | case ELF64LEKind: |
| 1373 | return isNonCommonDef<ELF64LE>(ctx, ekind: ELF64LEKind, mb, symName, archiveName); |
| 1374 | case ELF64BEKind: |
| 1375 | return isNonCommonDef<ELF64BE>(ctx, ekind: ELF64BEKind, mb, symName, archiveName); |
| 1376 | default: |
| 1377 | llvm_unreachable("getELFKind" ); |
| 1378 | } |
| 1379 | } |
| 1380 | |
| 1381 | SharedFile::SharedFile(Ctx &ctx, MemoryBufferRef m, StringRef defaultSoName) |
| 1382 | : ELFFileBase(ctx, SharedKind, getELFKind(ctx, mb: m, archiveName: "" ), m), |
| 1383 | soName(defaultSoName), isNeeded(!ctx.arg.asNeeded) {} |
| 1384 | |
| 1385 | // Parse the version definitions in the object file if present, and return a |
| 1386 | // vector whose nth element contains a pointer to the Elf_Verdef for version |
| 1387 | // identifier n. Version identifiers that are not definitions map to nullptr. |
| 1388 | template <typename ELFT> |
| 1389 | static SmallVector<const void *, 0> |
| 1390 | parseVerdefs(const uint8_t *base, const typename ELFT::Shdr *sec) { |
| 1391 | if (!sec) |
| 1392 | return {}; |
| 1393 | |
| 1394 | // Build the Verdefs array by following the chain of Elf_Verdef objects |
| 1395 | // from the start of the .gnu.version_d section. |
| 1396 | SmallVector<const void *, 0> verdefs; |
| 1397 | const uint8_t *verdef = base + sec->sh_offset; |
| 1398 | for (unsigned i = 0, e = sec->sh_info; i != e; ++i) { |
| 1399 | auto *curVerdef = reinterpret_cast<const typename ELFT::Verdef *>(verdef); |
| 1400 | verdef += curVerdef->vd_next; |
| 1401 | unsigned verdefIndex = curVerdef->vd_ndx; |
| 1402 | if (verdefIndex >= verdefs.size()) |
| 1403 | verdefs.resize(N: verdefIndex + 1); |
| 1404 | verdefs[verdefIndex] = curVerdef; |
| 1405 | } |
| 1406 | return verdefs; |
| 1407 | } |
| 1408 | |
| 1409 | // Parse SHT_GNU_verneed to properly set the name of a versioned undefined |
| 1410 | // symbol. We detect fatal issues which would cause vulnerabilities, but do not |
| 1411 | // implement sophisticated error checking like in llvm-readobj because the value |
| 1412 | // of such diagnostics is low. |
| 1413 | template <typename ELFT> |
| 1414 | std::vector<uint32_t> SharedFile::parseVerneed(const ELFFile<ELFT> &obj, |
| 1415 | const typename ELFT::Shdr *sec) { |
| 1416 | if (!sec) |
| 1417 | return {}; |
| 1418 | std::vector<uint32_t> verneeds; |
| 1419 | ArrayRef<uint8_t> data = CHECK2(obj.getSectionContents(*sec), this); |
| 1420 | const uint8_t *verneedBuf = data.begin(); |
| 1421 | for (unsigned i = 0; i != sec->sh_info; ++i) { |
| 1422 | if (verneedBuf + sizeof(typename ELFT::Verneed) > data.end()) { |
| 1423 | Err(ctx) << this << " has an invalid Verneed" ; |
| 1424 | break; |
| 1425 | } |
| 1426 | auto *vn = reinterpret_cast<const typename ELFT::Verneed *>(verneedBuf); |
| 1427 | const uint8_t *vernauxBuf = verneedBuf + vn->vn_aux; |
| 1428 | for (unsigned j = 0; j != vn->vn_cnt; ++j) { |
| 1429 | if (vernauxBuf + sizeof(typename ELFT::Vernaux) > data.end()) { |
| 1430 | Err(ctx) << this << " has an invalid Vernaux" ; |
| 1431 | break; |
| 1432 | } |
| 1433 | auto *aux = reinterpret_cast<const typename ELFT::Vernaux *>(vernauxBuf); |
| 1434 | if (aux->vna_name >= this->stringTable.size()) { |
| 1435 | Err(ctx) << this << " has a Vernaux with an invalid vna_name" ; |
| 1436 | break; |
| 1437 | } |
| 1438 | uint16_t version = aux->vna_other & VERSYM_VERSION; |
| 1439 | if (version >= verneeds.size()) |
| 1440 | verneeds.resize(new_size: version + 1); |
| 1441 | verneeds[version] = aux->vna_name; |
| 1442 | vernauxBuf += aux->vna_next; |
| 1443 | } |
| 1444 | verneedBuf += vn->vn_next; |
| 1445 | } |
| 1446 | return verneeds; |
| 1447 | } |
| 1448 | |
| 1449 | // Parse PT_GNU_PROPERTY segments in DSO. The process is similar to |
| 1450 | // readGnuProperty, but we don't have the InputSection information. |
| 1451 | template <typename ELFT> |
| 1452 | void SharedFile::parseGnuAndFeatures(const ELFFile<ELFT> &obj) { |
| 1453 | if (ctx.arg.emachine != EM_AARCH64) |
| 1454 | return; |
| 1455 | const uint8_t *base = obj.base(); |
| 1456 | auto phdrs = CHECK2(obj.program_headers(), this); |
| 1457 | for (auto phdr : phdrs) { |
| 1458 | if (phdr.p_type != PT_GNU_PROPERTY) |
| 1459 | continue; |
| 1460 | typename ELFT::Note note( |
| 1461 | *reinterpret_cast<const typename ELFT::Nhdr *>(base + phdr.p_offset)); |
| 1462 | if (note.getType() != NT_GNU_PROPERTY_TYPE_0 || note.getName() != "GNU" ) |
| 1463 | continue; |
| 1464 | |
| 1465 | ArrayRef<uint8_t> desc = note.getDesc(phdr.p_align); |
| 1466 | parseGnuPropertyNote<ELFT>(ctx, *this, GNU_PROPERTY_AARCH64_FEATURE_1_AND, |
| 1467 | desc, base); |
| 1468 | } |
| 1469 | } |
| 1470 | |
| 1471 | // We do not usually care about alignments of data in shared object |
| 1472 | // files because the loader takes care of it. However, if we promote a |
| 1473 | // DSO symbol to point to .bss due to copy relocation, we need to keep |
| 1474 | // the original alignment requirements. We infer it in this function. |
| 1475 | template <typename ELFT> |
| 1476 | static uint64_t getAlignment(ArrayRef<typename ELFT::Shdr> sections, |
| 1477 | const typename ELFT::Sym &sym) { |
| 1478 | uint64_t ret = UINT64_MAX; |
| 1479 | if (sym.st_value) |
| 1480 | ret = 1ULL << llvm::countr_zero(Val: (uint64_t)sym.st_value); |
| 1481 | if (0 < sym.st_shndx && sym.st_shndx < sections.size()) |
| 1482 | ret = std::min<uint64_t>(ret, sections[sym.st_shndx].sh_addralign); |
| 1483 | return (ret > UINT32_MAX) ? 0 : ret; |
| 1484 | } |
| 1485 | |
| 1486 | // Fully parse the shared object file. |
| 1487 | // |
| 1488 | // This function parses symbol versions. If a DSO has version information, |
| 1489 | // the file has a ".gnu.version_d" section which contains symbol version |
| 1490 | // definitions. Each symbol is associated to one version through a table in |
| 1491 | // ".gnu.version" section. That table is a parallel array for the symbol |
| 1492 | // table, and each table entry contains an index in ".gnu.version_d". |
| 1493 | // |
| 1494 | // The special index 0 is reserved for VERF_NDX_LOCAL and 1 is for |
| 1495 | // VER_NDX_GLOBAL. There's no table entry for these special versions in |
| 1496 | // ".gnu.version_d". |
| 1497 | // |
| 1498 | // The file format for symbol versioning is perhaps a bit more complicated |
| 1499 | // than necessary, but you can easily understand the code if you wrap your |
| 1500 | // head around the data structure described above. |
| 1501 | template <class ELFT> void SharedFile::parse() { |
| 1502 | using Elf_Dyn = typename ELFT::Dyn; |
| 1503 | using Elf_Shdr = typename ELFT::Shdr; |
| 1504 | using Elf_Sym = typename ELFT::Sym; |
| 1505 | using Elf_Verdef = typename ELFT::Verdef; |
| 1506 | using Elf_Versym = typename ELFT::Versym; |
| 1507 | |
| 1508 | ArrayRef<Elf_Dyn> dynamicTags; |
| 1509 | const ELFFile<ELFT> obj = this->getObj<ELFT>(); |
| 1510 | ArrayRef<Elf_Shdr> sections = getELFShdrs<ELFT>(); |
| 1511 | |
| 1512 | const Elf_Shdr *versymSec = nullptr; |
| 1513 | const Elf_Shdr *verdefSec = nullptr; |
| 1514 | const Elf_Shdr *verneedSec = nullptr; |
| 1515 | symbols = std::make_unique<Symbol *[]>(num: numSymbols); |
| 1516 | |
| 1517 | // Search for .dynsym, .dynamic, .symtab, .gnu.version and .gnu.version_d. |
| 1518 | for (const Elf_Shdr &sec : sections) { |
| 1519 | switch (sec.sh_type) { |
| 1520 | default: |
| 1521 | continue; |
| 1522 | case SHT_DYNAMIC: |
| 1523 | dynamicTags = |
| 1524 | CHECK2(obj.template getSectionContentsAsArray<Elf_Dyn>(sec), this); |
| 1525 | break; |
| 1526 | case SHT_GNU_versym: |
| 1527 | versymSec = &sec; |
| 1528 | break; |
| 1529 | case SHT_GNU_verdef: |
| 1530 | verdefSec = &sec; |
| 1531 | break; |
| 1532 | case SHT_GNU_verneed: |
| 1533 | verneedSec = &sec; |
| 1534 | break; |
| 1535 | } |
| 1536 | } |
| 1537 | |
| 1538 | if (versymSec && numSymbols == 0) { |
| 1539 | ErrAlways(ctx) << "SHT_GNU_versym should be associated with symbol table" ; |
| 1540 | return; |
| 1541 | } |
| 1542 | |
| 1543 | // Search for a DT_SONAME tag to initialize this->soName. |
| 1544 | for (const Elf_Dyn &dyn : dynamicTags) { |
| 1545 | if (dyn.d_tag == DT_NEEDED) { |
| 1546 | uint64_t val = dyn.getVal(); |
| 1547 | if (val >= this->stringTable.size()) { |
| 1548 | Err(ctx) << this << ": invalid DT_NEEDED entry" ; |
| 1549 | return; |
| 1550 | } |
| 1551 | dtNeeded.push_back(Elt: this->stringTable.data() + val); |
| 1552 | } else if (dyn.d_tag == DT_SONAME) { |
| 1553 | uint64_t val = dyn.getVal(); |
| 1554 | if (val >= this->stringTable.size()) { |
| 1555 | Err(ctx) << this << ": invalid DT_SONAME entry" ; |
| 1556 | return; |
| 1557 | } |
| 1558 | soName = this->stringTable.data() + val; |
| 1559 | } |
| 1560 | } |
| 1561 | |
| 1562 | // DSOs are uniquified not by filename but by soname. |
| 1563 | StringSaver &ss = ctx.saver; |
| 1564 | DenseMap<CachedHashStringRef, SharedFile *>::iterator it; |
| 1565 | bool wasInserted; |
| 1566 | std::tie(args&: it, args&: wasInserted) = |
| 1567 | ctx.symtab->soNames.try_emplace(Key: CachedHashStringRef(soName), Args: this); |
| 1568 | |
| 1569 | // If a DSO appears more than once on the command line with and without |
| 1570 | // --as-needed, --no-as-needed takes precedence over --as-needed because a |
| 1571 | // user can add an extra DSO with --no-as-needed to force it to be added to |
| 1572 | // the dependency list. |
| 1573 | it->second->isNeeded |= isNeeded; |
| 1574 | if (!wasInserted) |
| 1575 | return; |
| 1576 | |
| 1577 | ctx.sharedFiles.push_back(Elt: this); |
| 1578 | |
| 1579 | verdefs = parseVerdefs<ELFT>(obj.base(), verdefSec); |
| 1580 | std::vector<uint32_t> verneeds = parseVerneed<ELFT>(obj, verneedSec); |
| 1581 | parseGnuAndFeatures<ELFT>(obj); |
| 1582 | |
| 1583 | // Parse ".gnu.version" section which is a parallel array for the symbol |
| 1584 | // table. If a given file doesn't have a ".gnu.version" section, we use |
| 1585 | // VER_NDX_GLOBAL. |
| 1586 | size_t size = numSymbols - firstGlobal; |
| 1587 | std::vector<uint16_t> versyms(size, VER_NDX_GLOBAL); |
| 1588 | if (versymSec) { |
| 1589 | ArrayRef<Elf_Versym> versym = |
| 1590 | CHECK2(obj.template getSectionContentsAsArray<Elf_Versym>(*versymSec), |
| 1591 | this) |
| 1592 | .slice(firstGlobal); |
| 1593 | for (size_t i = 0; i < size; ++i) |
| 1594 | versyms[i] = versym[i].vs_index; |
| 1595 | } |
| 1596 | |
| 1597 | // System libraries can have a lot of symbols with versions. Using a |
| 1598 | // fixed buffer for computing the versions name (foo@ver) can save a |
| 1599 | // lot of allocations. |
| 1600 | SmallString<0> versionedNameBuffer; |
| 1601 | |
| 1602 | // Add symbols to the symbol table. |
| 1603 | ArrayRef<Elf_Sym> syms = this->getGlobalELFSyms<ELFT>(); |
| 1604 | for (size_t i = 0, e = syms.size(); i != e; ++i) { |
| 1605 | const Elf_Sym &sym = syms[i]; |
| 1606 | |
| 1607 | // ELF spec requires that all local symbols precede weak or global |
| 1608 | // symbols in each symbol table, and the index of first non-local symbol |
| 1609 | // is stored to sh_info. If a local symbol appears after some non-local |
| 1610 | // symbol, that's a violation of the spec. |
| 1611 | StringRef name = CHECK2(sym.getName(stringTable), this); |
| 1612 | if (sym.getBinding() == STB_LOCAL) { |
| 1613 | Err(ctx) << this << ": invalid local symbol '" << name |
| 1614 | << "' in global part of symbol table" ; |
| 1615 | continue; |
| 1616 | } |
| 1617 | |
| 1618 | const uint16_t ver = versyms[i], idx = ver & ~VERSYM_HIDDEN; |
| 1619 | if (sym.isUndefined()) { |
| 1620 | // For unversioned undefined symbols, VER_NDX_GLOBAL makes more sense but |
| 1621 | // as of binutils 2.34, GNU ld produces VER_NDX_LOCAL. |
| 1622 | if (ver != VER_NDX_LOCAL && ver != VER_NDX_GLOBAL) { |
| 1623 | if (idx >= verneeds.size()) { |
| 1624 | ErrAlways(ctx) << "corrupt input file: version need index " << idx |
| 1625 | << " for symbol " << name |
| 1626 | << " is out of bounds\n>>> defined in " << this; |
| 1627 | continue; |
| 1628 | } |
| 1629 | StringRef verName = stringTable.data() + verneeds[idx]; |
| 1630 | versionedNameBuffer.clear(); |
| 1631 | name = ss.save(S: (name + "@" + verName).toStringRef(Out&: versionedNameBuffer)); |
| 1632 | } |
| 1633 | Symbol *s = ctx.symtab->addSymbol( |
| 1634 | newSym: Undefined{this, name, sym.getBinding(), sym.st_other, sym.getType()}); |
| 1635 | s->isExported = true; |
| 1636 | if (sym.getBinding() != STB_WEAK && |
| 1637 | ctx.arg.unresolvedSymbolsInShlib != UnresolvedPolicy::Ignore) |
| 1638 | requiredSymbols.push_back(Elt: s); |
| 1639 | continue; |
| 1640 | } |
| 1641 | |
| 1642 | if (ver == VER_NDX_LOCAL || |
| 1643 | (ver != VER_NDX_GLOBAL && idx >= verdefs.size())) { |
| 1644 | // In GNU ld < 2.31 (before 3be08ea4728b56d35e136af4e6fd3086ade17764), the |
| 1645 | // MIPS port puts _gp_disp symbol into DSO files and incorrectly assigns |
| 1646 | // VER_NDX_LOCAL. Workaround this bug. |
| 1647 | if (ctx.arg.emachine == EM_MIPS && name == "_gp_disp" ) |
| 1648 | continue; |
| 1649 | ErrAlways(ctx) << "corrupt input file: version definition index " << idx |
| 1650 | << " for symbol " << name |
| 1651 | << " is out of bounds\n>>> defined in " << this; |
| 1652 | continue; |
| 1653 | } |
| 1654 | |
| 1655 | uint32_t alignment = getAlignment<ELFT>(sections, sym); |
| 1656 | if (ver == idx) { |
| 1657 | auto *s = ctx.symtab->addSymbol( |
| 1658 | newSym: SharedSymbol{*this, name, sym.getBinding(), sym.st_other, |
| 1659 | sym.getType(), sym.st_value, sym.st_size, alignment}); |
| 1660 | s->dsoDefined = true; |
| 1661 | if (s->file == this) |
| 1662 | s->versionId = ver; |
| 1663 | } |
| 1664 | |
| 1665 | // Also add the symbol with the versioned name to handle undefined symbols |
| 1666 | // with explicit versions. |
| 1667 | if (ver == VER_NDX_GLOBAL) |
| 1668 | continue; |
| 1669 | |
| 1670 | StringRef verName = |
| 1671 | stringTable.data() + |
| 1672 | reinterpret_cast<const Elf_Verdef *>(verdefs[idx])->getAux()->vda_name; |
| 1673 | versionedNameBuffer.clear(); |
| 1674 | name = (name + "@" + verName).toStringRef(Out&: versionedNameBuffer); |
| 1675 | auto *s = ctx.symtab->addSymbol( |
| 1676 | newSym: SharedSymbol{*this, ss.save(S: name), sym.getBinding(), sym.st_other, |
| 1677 | sym.getType(), sym.st_value, sym.st_size, alignment}); |
| 1678 | s->dsoDefined = true; |
| 1679 | if (s->file == this) |
| 1680 | s->versionId = idx; |
| 1681 | } |
| 1682 | } |
| 1683 | |
| 1684 | static ELFKind getBitcodeELFKind(const Triple &t) { |
| 1685 | if (t.isLittleEndian()) |
| 1686 | return t.isArch64Bit() ? ELF64LEKind : ELF32LEKind; |
| 1687 | return t.isArch64Bit() ? ELF64BEKind : ELF32BEKind; |
| 1688 | } |
| 1689 | |
| 1690 | static uint16_t getBitcodeMachineKind(Ctx &ctx, StringRef path, |
| 1691 | const Triple &t) { |
| 1692 | switch (t.getArch()) { |
| 1693 | case Triple::aarch64: |
| 1694 | case Triple::aarch64_be: |
| 1695 | return EM_AARCH64; |
| 1696 | case Triple::amdgcn: |
| 1697 | case Triple::r600: |
| 1698 | return EM_AMDGPU; |
| 1699 | case Triple::arm: |
| 1700 | case Triple::armeb: |
| 1701 | case Triple::thumb: |
| 1702 | case Triple::thumbeb: |
| 1703 | return EM_ARM; |
| 1704 | case Triple::avr: |
| 1705 | return EM_AVR; |
| 1706 | case Triple::hexagon: |
| 1707 | return EM_HEXAGON; |
| 1708 | case Triple::loongarch32: |
| 1709 | case Triple::loongarch64: |
| 1710 | return EM_LOONGARCH; |
| 1711 | case Triple::mips: |
| 1712 | case Triple::mipsel: |
| 1713 | case Triple::mips64: |
| 1714 | case Triple::mips64el: |
| 1715 | return EM_MIPS; |
| 1716 | case Triple::msp430: |
| 1717 | return EM_MSP430; |
| 1718 | case Triple::ppc: |
| 1719 | case Triple::ppcle: |
| 1720 | return EM_PPC; |
| 1721 | case Triple::ppc64: |
| 1722 | case Triple::ppc64le: |
| 1723 | return EM_PPC64; |
| 1724 | case Triple::riscv32: |
| 1725 | case Triple::riscv64: |
| 1726 | return EM_RISCV; |
| 1727 | case Triple::sparcv9: |
| 1728 | return EM_SPARCV9; |
| 1729 | case Triple::systemz: |
| 1730 | return EM_S390; |
| 1731 | case Triple::x86: |
| 1732 | return t.isOSIAMCU() ? EM_IAMCU : EM_386; |
| 1733 | case Triple::x86_64: |
| 1734 | return EM_X86_64; |
| 1735 | default: |
| 1736 | ErrAlways(ctx) << path |
| 1737 | << ": could not infer e_machine from bitcode target triple " |
| 1738 | << t.str(); |
| 1739 | return EM_NONE; |
| 1740 | } |
| 1741 | } |
| 1742 | |
| 1743 | static uint8_t getOsAbi(const Triple &t) { |
| 1744 | switch (t.getOS()) { |
| 1745 | case Triple::AMDHSA: |
| 1746 | return ELF::ELFOSABI_AMDGPU_HSA; |
| 1747 | case Triple::AMDPAL: |
| 1748 | return ELF::ELFOSABI_AMDGPU_PAL; |
| 1749 | case Triple::Mesa3D: |
| 1750 | return ELF::ELFOSABI_AMDGPU_MESA3D; |
| 1751 | default: |
| 1752 | return ELF::ELFOSABI_NONE; |
| 1753 | } |
| 1754 | } |
| 1755 | |
| 1756 | BitcodeFile::BitcodeFile(Ctx &ctx, MemoryBufferRef mb, StringRef archiveName, |
| 1757 | uint64_t offsetInArchive, bool lazy) |
| 1758 | : InputFile(ctx, BitcodeKind, mb) { |
| 1759 | this->archiveName = archiveName; |
| 1760 | this->lazy = lazy; |
| 1761 | |
| 1762 | std::string path = mb.getBufferIdentifier().str(); |
| 1763 | if (ctx.arg.thinLTOIndexOnly) |
| 1764 | path = replaceThinLTOSuffix(ctx, path: mb.getBufferIdentifier()); |
| 1765 | |
| 1766 | // ThinLTO assumes that all MemoryBufferRefs given to it have a unique |
| 1767 | // name. If two archives define two members with the same name, this |
| 1768 | // causes a collision which result in only one of the objects being taken |
| 1769 | // into consideration at LTO time (which very likely causes undefined |
| 1770 | // symbols later in the link stage). So we append file offset to make |
| 1771 | // filename unique. |
| 1772 | StringSaver &ss = ctx.saver; |
| 1773 | StringRef name = archiveName.empty() |
| 1774 | ? ss.save(S: path) |
| 1775 | : ss.save(S: archiveName + "(" + path::filename(path) + |
| 1776 | " at " + utostr(X: offsetInArchive) + ")" ); |
| 1777 | MemoryBufferRef mbref(mb.getBuffer(), name); |
| 1778 | |
| 1779 | obj = CHECK2(lto::InputFile::create(mbref), this); |
| 1780 | |
| 1781 | Triple t(obj->getTargetTriple()); |
| 1782 | ekind = getBitcodeELFKind(t); |
| 1783 | emachine = getBitcodeMachineKind(ctx, path: mb.getBufferIdentifier(), t); |
| 1784 | osabi = getOsAbi(t); |
| 1785 | } |
| 1786 | |
| 1787 | static uint8_t mapVisibility(GlobalValue::VisibilityTypes gvVisibility) { |
| 1788 | switch (gvVisibility) { |
| 1789 | case GlobalValue::DefaultVisibility: |
| 1790 | return STV_DEFAULT; |
| 1791 | case GlobalValue::HiddenVisibility: |
| 1792 | return STV_HIDDEN; |
| 1793 | case GlobalValue::ProtectedVisibility: |
| 1794 | return STV_PROTECTED; |
| 1795 | } |
| 1796 | llvm_unreachable("unknown visibility" ); |
| 1797 | } |
| 1798 | |
| 1799 | static void createBitcodeSymbol(Ctx &ctx, Symbol *&sym, |
| 1800 | const lto::InputFile::Symbol &objSym, |
| 1801 | BitcodeFile &f) { |
| 1802 | uint8_t binding = objSym.isWeak() ? STB_WEAK : STB_GLOBAL; |
| 1803 | uint8_t type = objSym.isTLS() ? STT_TLS : STT_NOTYPE; |
| 1804 | uint8_t visibility = mapVisibility(gvVisibility: objSym.getVisibility()); |
| 1805 | |
| 1806 | if (!sym) { |
| 1807 | // Symbols can be duplicated in bitcode files because of '#include' and |
| 1808 | // linkonce_odr. Use uniqueSaver to save symbol names for de-duplication. |
| 1809 | // Update objSym.Name to reference (via StringRef) the string saver's copy; |
| 1810 | // this way LTO can reference the same string saver's copy rather than |
| 1811 | // keeping copies of its own. |
| 1812 | objSym.Name = ctx.uniqueSaver.save(S: objSym.getName()); |
| 1813 | sym = ctx.symtab->insert(name: objSym.getName()); |
| 1814 | } |
| 1815 | |
| 1816 | if (objSym.isUndefined()) { |
| 1817 | Undefined newSym(&f, StringRef(), binding, visibility, type); |
| 1818 | sym->resolve(ctx, other: newSym); |
| 1819 | sym->referenced = true; |
| 1820 | return; |
| 1821 | } |
| 1822 | |
| 1823 | if (objSym.isCommon()) { |
| 1824 | sym->resolve(ctx, other: CommonSymbol{ctx, &f, StringRef(), binding, visibility, |
| 1825 | STT_OBJECT, objSym.getCommonAlignment(), |
| 1826 | objSym.getCommonSize()}); |
| 1827 | } else { |
| 1828 | Defined newSym(ctx, &f, StringRef(), binding, visibility, type, 0, 0, |
| 1829 | nullptr); |
| 1830 | // The definition can be omitted if all bitcode definitions satisfy |
| 1831 | // `canBeOmittedFromSymbolTable()` and isUsedInRegularObj is false. |
| 1832 | // The latter condition is tested in parseVersionAndComputeIsPreemptible. |
| 1833 | sym->ltoCanOmit = objSym.canBeOmittedFromSymbolTable() && |
| 1834 | (!sym->isDefined() || sym->ltoCanOmit); |
| 1835 | sym->resolve(ctx, other: newSym); |
| 1836 | } |
| 1837 | } |
| 1838 | |
| 1839 | void BitcodeFile::parse() { |
| 1840 | for (std::pair<StringRef, Comdat::SelectionKind> s : obj->getComdatTable()) { |
| 1841 | keptComdats.push_back( |
| 1842 | x: s.second == Comdat::NoDeduplicate || |
| 1843 | ctx.symtab->comdatGroups.try_emplace(Key: CachedHashStringRef(s.first), Args: this) |
| 1844 | .second); |
| 1845 | } |
| 1846 | |
| 1847 | if (numSymbols == 0) { |
| 1848 | numSymbols = obj->symbols().size(); |
| 1849 | symbols = std::make_unique<Symbol *[]>(num: numSymbols); |
| 1850 | } |
| 1851 | // Process defined symbols first. See the comment in |
| 1852 | // ObjFile<ELFT>::initializeSymbols. |
| 1853 | for (auto [i, irSym] : llvm::enumerate(First: obj->symbols())) |
| 1854 | if (!irSym.isUndefined()) |
| 1855 | createBitcodeSymbol(ctx, sym&: symbols[i], objSym: irSym, f&: *this); |
| 1856 | for (auto [i, irSym] : llvm::enumerate(First: obj->symbols())) |
| 1857 | if (irSym.isUndefined()) |
| 1858 | createBitcodeSymbol(ctx, sym&: symbols[i], objSym: irSym, f&: *this); |
| 1859 | |
| 1860 | for (auto l : obj->getDependentLibraries()) |
| 1861 | addDependentLibrary(ctx, specifier: l, f: this); |
| 1862 | } |
| 1863 | |
| 1864 | void BitcodeFile::parseLazy() { |
| 1865 | numSymbols = obj->symbols().size(); |
| 1866 | symbols = std::make_unique<Symbol *[]>(num: numSymbols); |
| 1867 | for (auto [i, irSym] : llvm::enumerate(First: obj->symbols())) { |
| 1868 | // Symbols can be duplicated in bitcode files because of '#include' and |
| 1869 | // linkonce_odr. Use uniqueSaver to save symbol names for de-duplication. |
| 1870 | // Update objSym.Name to reference (via StringRef) the string saver's copy; |
| 1871 | // this way LTO can reference the same string saver's copy rather than |
| 1872 | // keeping copies of its own. |
| 1873 | irSym.Name = ctx.uniqueSaver.save(S: irSym.getName()); |
| 1874 | if (!irSym.isUndefined()) { |
| 1875 | auto *sym = ctx.symtab->insert(name: irSym.getName()); |
| 1876 | sym->resolve(ctx, other: LazySymbol{*this}); |
| 1877 | symbols[i] = sym; |
| 1878 | } |
| 1879 | } |
| 1880 | } |
| 1881 | |
| 1882 | void BitcodeFile::postParse() { |
| 1883 | for (auto [i, irSym] : llvm::enumerate(First: obj->symbols())) { |
| 1884 | const Symbol &sym = *symbols[i]; |
| 1885 | if (sym.file == this || !sym.isDefined() || irSym.isUndefined() || |
| 1886 | irSym.isCommon() || irSym.isWeak()) |
| 1887 | continue; |
| 1888 | int c = irSym.getComdatIndex(); |
| 1889 | if (c != -1 && !keptComdats[c]) |
| 1890 | continue; |
| 1891 | reportDuplicate(ctx, sym, newFile: this, errSec: nullptr, errOffset: 0); |
| 1892 | } |
| 1893 | } |
| 1894 | |
| 1895 | void BinaryFile::parse() { |
| 1896 | ArrayRef<uint8_t> data = arrayRefFromStringRef(Input: mb.getBuffer()); |
| 1897 | auto *section = |
| 1898 | make<InputSection>(args: this, args: ".data" , args: SHT_PROGBITS, args: SHF_ALLOC | SHF_WRITE, |
| 1899 | /*addralign=*/args: 8, /*entsize=*/args: 0, args&: data); |
| 1900 | sections.push_back(Elt: section); |
| 1901 | |
| 1902 | // For each input file foo that is embedded to a result as a binary |
| 1903 | // blob, we define _binary_foo_{start,end,size} symbols, so that |
| 1904 | // user programs can access blobs by name. Non-alphanumeric |
| 1905 | // characters in a filename are replaced with underscore. |
| 1906 | std::string s = "_binary_" + mb.getBufferIdentifier().str(); |
| 1907 | for (char &c : s) |
| 1908 | if (!isAlnum(C: c)) |
| 1909 | c = '_'; |
| 1910 | |
| 1911 | llvm::StringSaver &ss = ctx.saver; |
| 1912 | ctx.symtab->addAndCheckDuplicate( |
| 1913 | ctx, newSym: Defined{ctx, this, ss.save(S: s + "_start" ), STB_GLOBAL, STV_DEFAULT, |
| 1914 | STT_OBJECT, 0, 0, section}); |
| 1915 | ctx.symtab->addAndCheckDuplicate( |
| 1916 | ctx, newSym: Defined{ctx, this, ss.save(S: s + "_end" ), STB_GLOBAL, STV_DEFAULT, |
| 1917 | STT_OBJECT, data.size(), 0, section}); |
| 1918 | ctx.symtab->addAndCheckDuplicate( |
| 1919 | ctx, newSym: Defined{ctx, this, ss.save(S: s + "_size" ), STB_GLOBAL, STV_DEFAULT, |
| 1920 | STT_OBJECT, data.size(), 0, nullptr}); |
| 1921 | } |
| 1922 | |
| 1923 | InputFile *elf::createInternalFile(Ctx &ctx, StringRef name) { |
| 1924 | auto *file = |
| 1925 | make<InputFile>(args&: ctx, args: InputFile::InternalKind, args: MemoryBufferRef("" , name)); |
| 1926 | // References from an internal file do not lead to --warn-backrefs |
| 1927 | // diagnostics. |
| 1928 | file->groupId = 0; |
| 1929 | return file; |
| 1930 | } |
| 1931 | |
| 1932 | std::unique_ptr<ELFFileBase> elf::createObjFile(Ctx &ctx, MemoryBufferRef mb, |
| 1933 | StringRef archiveName, |
| 1934 | bool lazy) { |
| 1935 | std::unique_ptr<ELFFileBase> f; |
| 1936 | switch (getELFKind(ctx, mb, archiveName)) { |
| 1937 | case ELF32LEKind: |
| 1938 | f = std::make_unique<ObjFile<ELF32LE>>(args&: ctx, args: ELF32LEKind, args&: mb, args&: archiveName); |
| 1939 | break; |
| 1940 | case ELF32BEKind: |
| 1941 | f = std::make_unique<ObjFile<ELF32BE>>(args&: ctx, args: ELF32BEKind, args&: mb, args&: archiveName); |
| 1942 | break; |
| 1943 | case ELF64LEKind: |
| 1944 | f = std::make_unique<ObjFile<ELF64LE>>(args&: ctx, args: ELF64LEKind, args&: mb, args&: archiveName); |
| 1945 | break; |
| 1946 | case ELF64BEKind: |
| 1947 | f = std::make_unique<ObjFile<ELF64BE>>(args&: ctx, args: ELF64BEKind, args&: mb, args&: archiveName); |
| 1948 | break; |
| 1949 | default: |
| 1950 | llvm_unreachable("getELFKind" ); |
| 1951 | } |
| 1952 | f->init(); |
| 1953 | f->lazy = lazy; |
| 1954 | return f; |
| 1955 | } |
| 1956 | |
| 1957 | template <class ELFT> void ObjFile<ELFT>::parseLazy() { |
| 1958 | const ArrayRef<typename ELFT::Sym> eSyms = this->getELFSyms<ELFT>(); |
| 1959 | numSymbols = eSyms.size(); |
| 1960 | symbols = std::make_unique<Symbol *[]>(numSymbols); |
| 1961 | |
| 1962 | // resolve() may trigger this->extract() if an existing symbol is an undefined |
| 1963 | // symbol. If that happens, this function has served its purpose, and we can |
| 1964 | // exit from the loop early. |
| 1965 | auto *symtab = ctx.symtab.get(); |
| 1966 | for (size_t i = firstGlobal, end = eSyms.size(); i != end; ++i) { |
| 1967 | if (eSyms[i].st_shndx == SHN_UNDEF) |
| 1968 | continue; |
| 1969 | symbols[i] = symtab->insert(CHECK2(eSyms[i].getName(stringTable), this)); |
| 1970 | symbols[i]->resolve(ctx, LazySymbol{*this}); |
| 1971 | if (!lazy) |
| 1972 | break; |
| 1973 | } |
| 1974 | } |
| 1975 | |
| 1976 | bool InputFile::(StringRef name) const { |
| 1977 | if (isa<BitcodeFile>(Val: this)) |
| 1978 | return isBitcodeNonCommonDef(mb, symName: name, archiveName); |
| 1979 | |
| 1980 | return isNonCommonDef(ctx, mb, symName: name, archiveName); |
| 1981 | } |
| 1982 | |
| 1983 | std::string elf::replaceThinLTOSuffix(Ctx &ctx, StringRef path) { |
| 1984 | auto [suffix, repl] = ctx.arg.thinLTOObjectSuffixReplace; |
| 1985 | if (path.consume_back(Suffix: suffix)) |
| 1986 | return (path + repl).str(); |
| 1987 | return std::string(path); |
| 1988 | } |
| 1989 | |
| 1990 | template class elf::ObjFile<ELF32LE>; |
| 1991 | template class elf::ObjFile<ELF32BE>; |
| 1992 | template class elf::ObjFile<ELF64LE>; |
| 1993 | template class elf::ObjFile<ELF64BE>; |
| 1994 | |
| 1995 | template void SharedFile::parse<ELF32LE>(); |
| 1996 | template void SharedFile::parse<ELF32BE>(); |
| 1997 | template void SharedFile::parse<ELF64LE>(); |
| 1998 | template void SharedFile::parse<ELF64BE>(); |
| 1999 | |