| 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 | // This file contains functions to parse Mach-O object files. In this comment, |
| 10 | // we describe the Mach-O file structure and how we parse it. |
| 11 | // |
| 12 | // Mach-O is not very different from ELF or COFF. The notion of symbols, |
| 13 | // sections and relocations exists in Mach-O as it does in ELF and COFF. |
| 14 | // |
| 15 | // Perhaps the notion that is new to those who know ELF/COFF is "subsections". |
| 16 | // In ELF/COFF, sections are an atomic unit of data copied from input files to |
| 17 | // output files. When we merge or garbage-collect sections, we treat each |
| 18 | // section as an atomic unit. In Mach-O, that's not the case. Sections can |
| 19 | // consist of multiple subsections, and subsections are a unit of merging and |
| 20 | // garbage-collecting. Therefore, Mach-O's subsections are more similar to |
| 21 | // ELF/COFF's sections than Mach-O's sections are. |
| 22 | // |
| 23 | // A section can have multiple symbols. A symbol that does not have the |
| 24 | // N_ALT_ENTRY attribute indicates a beginning of a subsection. Therefore, by |
| 25 | // definition, a symbol is always present at the beginning of each subsection. A |
| 26 | // symbol with N_ALT_ENTRY attribute does not start a new subsection and can |
| 27 | // point to a middle of a subsection. |
| 28 | // |
| 29 | // The notion of subsections also affects how relocations are represented in |
| 30 | // Mach-O. All references within a section need to be explicitly represented as |
| 31 | // relocations if they refer to different subsections, because we obviously need |
| 32 | // to fix up addresses if subsections are laid out in an output file differently |
| 33 | // than they were in object files. To represent that, Mach-O relocations can |
| 34 | // refer to an unnamed location via its address. Scattered relocations (those |
| 35 | // with the R_SCATTERED bit set) always refer to unnamed locations. |
| 36 | // Non-scattered relocations refer to an unnamed location if r_extern is not set |
| 37 | // and r_symbolnum is zero. |
| 38 | // |
| 39 | // Without the above differences, I think you can use your knowledge about ELF |
| 40 | // and COFF for Mach-O. |
| 41 | // |
| 42 | //===----------------------------------------------------------------------===// |
| 43 | |
| 44 | #include "InputFiles.h" |
| 45 | #include "Config.h" |
| 46 | #include "Driver.h" |
| 47 | #include "Dwarf.h" |
| 48 | #include "EhFrame.h" |
| 49 | #include "ExportTrie.h" |
| 50 | #include "InputSection.h" |
| 51 | #include "ObjC.h" |
| 52 | #include "OutputSection.h" |
| 53 | #include "OutputSegment.h" |
| 54 | #include "SymbolTable.h" |
| 55 | #include "Symbols.h" |
| 56 | #include "SyntheticSections.h" |
| 57 | #include "Target.h" |
| 58 | |
| 59 | #include "lld/Common/CommonLinkerContext.h" |
| 60 | #include "lld/Common/DWARF.h" |
| 61 | #include "lld/Common/Reproduce.h" |
| 62 | #include "llvm/ADT/iterator.h" |
| 63 | #include "llvm/BinaryFormat/MachO.h" |
| 64 | #include "llvm/LTO/LTO.h" |
| 65 | #include "llvm/Support/BinaryStreamReader.h" |
| 66 | #include "llvm/Support/Endian.h" |
| 67 | #include "llvm/Support/MemoryBuffer.h" |
| 68 | #include "llvm/Support/Parallel.h" |
| 69 | #include "llvm/Support/Path.h" |
| 70 | #include "llvm/Support/TarWriter.h" |
| 71 | #include "llvm/Support/TimeProfiler.h" |
| 72 | #include "llvm/TextAPI/Architecture.h" |
| 73 | #include "llvm/TextAPI/InterfaceFile.h" |
| 74 | |
| 75 | #include <optional> |
| 76 | #include <type_traits> |
| 77 | |
| 78 | using namespace llvm; |
| 79 | using namespace llvm::MachO; |
| 80 | using namespace llvm::support::endian; |
| 81 | using namespace llvm::sys; |
| 82 | using namespace lld; |
| 83 | using namespace lld::macho; |
| 84 | |
| 85 | // Returns "<internal>", "foo.a(bar.o)", or "baz.o". |
| 86 | std::string lld::toString(const InputFile *f) { |
| 87 | if (!f) |
| 88 | return "<internal>"; |
| 89 | |
| 90 | // Multiple dylibs can be defined in one .tbd file. |
| 91 | if (const auto *dylibFile = dyn_cast<DylibFile>(Val: f)) |
| 92 | if (f->getName().ends_with(Suffix: ".tbd")) |
| 93 | return (f->getName() + "("+ dylibFile->installName + ")").str(); |
| 94 | |
| 95 | if (f->archiveName.empty()) |
| 96 | return std::string(f->getName()); |
| 97 | return (f->archiveName + "("+ path::filename(path: f->getName()) + ")").str(); |
| 98 | } |
| 99 | |
| 100 | std::string lld::toString(const Section &sec) { |
| 101 | return (toString(f: sec.file) + ":("+ sec.name + ")").str(); |
| 102 | } |
| 103 | |
| 104 | SetVector<InputFile *> macho::inputFiles; |
| 105 | std::unique_ptr<TarWriter> macho::tar; |
| 106 | int InputFile::idCount = 0; |
| 107 | |
| 108 | static VersionTuple decodeVersion(uint32_t version) { |
| 109 | unsigned major = version >> 16; |
| 110 | unsigned minor = (version >> 8) & 0xffu; |
| 111 | unsigned subMinor = version & 0xffu; |
| 112 | return VersionTuple(major, minor, subMinor); |
| 113 | } |
| 114 | |
| 115 | static std::vector<PlatformInfo> getPlatformInfos(const InputFile *input) { |
| 116 | if (!isa<ObjFile>(Val: input) && !isa<DylibFile>(Val: input)) |
| 117 | return {}; |
| 118 | |
| 119 | const char *hdr = input->mb.getBufferStart(); |
| 120 | |
| 121 | // "Zippered" object files can have multiple LC_BUILD_VERSION load commands. |
| 122 | std::vector<PlatformInfo> platformInfos; |
| 123 | for (auto *cmd : findCommands<build_version_command>(anyHdr: hdr, types: LC_BUILD_VERSION)) { |
| 124 | PlatformInfo info; |
| 125 | info.target.Platform = static_cast<PlatformType>(cmd->platform); |
| 126 | info.target.MinDeployment = decodeVersion(version: cmd->minos); |
| 127 | platformInfos.emplace_back(args: std::move(info)); |
| 128 | } |
| 129 | for (auto *cmd : findCommands<version_min_command>( |
| 130 | anyHdr: hdr, types: LC_VERSION_MIN_MACOSX, types: LC_VERSION_MIN_IPHONEOS, |
| 131 | types: LC_VERSION_MIN_TVOS, types: LC_VERSION_MIN_WATCHOS)) { |
| 132 | PlatformInfo info; |
| 133 | switch (cmd->cmd) { |
| 134 | case LC_VERSION_MIN_MACOSX: |
| 135 | info.target.Platform = PLATFORM_MACOS; |
| 136 | break; |
| 137 | case LC_VERSION_MIN_IPHONEOS: |
| 138 | info.target.Platform = PLATFORM_IOS; |
| 139 | break; |
| 140 | case LC_VERSION_MIN_TVOS: |
| 141 | info.target.Platform = PLATFORM_TVOS; |
| 142 | break; |
| 143 | case LC_VERSION_MIN_WATCHOS: |
| 144 | info.target.Platform = PLATFORM_WATCHOS; |
| 145 | break; |
| 146 | } |
| 147 | info.target.MinDeployment = decodeVersion(version: cmd->version); |
| 148 | platformInfos.emplace_back(args: std::move(info)); |
| 149 | } |
| 150 | |
| 151 | return platformInfos; |
| 152 | } |
| 153 | |
| 154 | static bool checkCompatibility(const InputFile *input) { |
| 155 | std::vector<PlatformInfo> platformInfos = getPlatformInfos(input); |
| 156 | if (platformInfos.empty()) |
| 157 | return true; |
| 158 | |
| 159 | auto it = find_if(Range&: platformInfos, P: [&](const PlatformInfo &info) { |
| 160 | return removeSimulator(platform: info.target.Platform) == |
| 161 | removeSimulator(platform: config->platform()); |
| 162 | }); |
| 163 | if (it == platformInfos.end()) { |
| 164 | std::string platformNames; |
| 165 | raw_string_ostream os(platformNames); |
| 166 | interleave( |
| 167 | c: platformInfos, os, |
| 168 | each_fn: [&](const PlatformInfo &info) { |
| 169 | os << getPlatformName(Platform: info.target.Platform); |
| 170 | }, |
| 171 | separator: "/"); |
| 172 | error(msg: toString(f: input) + " has platform "+ platformNames + |
| 173 | Twine(", which is different from target platform ") + |
| 174 | getPlatformName(Platform: config->platform())); |
| 175 | return false; |
| 176 | } |
| 177 | |
| 178 | if (it->target.MinDeployment > config->platformInfo.target.MinDeployment) |
| 179 | warn(msg: toString(f: input) + " has version "+ |
| 180 | it->target.MinDeployment.getAsString() + |
| 181 | ", which is newer than target minimum of "+ |
| 182 | config->platformInfo.target.MinDeployment.getAsString()); |
| 183 | |
| 184 | return true; |
| 185 | } |
| 186 | |
| 187 | template <class Header> |
| 188 | static bool compatWithTargetArch(const InputFile *file, const Header *hdr) { |
| 189 | uint32_t cpuType; |
| 190 | std::tie(args&: cpuType, args: std::ignore) = getCPUTypeFromArchitecture(Arch: config->arch()); |
| 191 | |
| 192 | if (hdr->cputype != cpuType) { |
| 193 | Architecture arch = |
| 194 | getArchitectureFromCpuType(hdr->cputype, hdr->cpusubtype); |
| 195 | auto msg = config->errorForArchMismatch |
| 196 | ? static_cast<void (*)(const Twine &)>(error) |
| 197 | : warn; |
| 198 | |
| 199 | msg(toString(f: file) + " has architecture "+ getArchitectureName(Arch: arch) + |
| 200 | " which is incompatible with target architecture "+ |
| 201 | getArchitectureName(Arch: config->arch())); |
| 202 | return false; |
| 203 | } |
| 204 | |
| 205 | return checkCompatibility(input: file); |
| 206 | } |
| 207 | |
| 208 | // This cache mostly exists to store system libraries (and .tbds) as they're |
| 209 | // loaded, rather than the input archives, which are already cached at a higher |
| 210 | // level, and other files like the filelist that are only read once. |
| 211 | // Theoretically this caching could be more efficient by hoisting it, but that |
| 212 | // would require altering many callers to track the state. |
| 213 | DenseMap<CachedHashStringRef, MemoryBufferRef> macho::cachedReads; |
| 214 | // Open a given file path and return it as a memory-mapped file. |
| 215 | std::optional<MemoryBufferRef> macho::readFile(StringRef path) { |
| 216 | CachedHashStringRef key(path); |
| 217 | auto entry = cachedReads.find(Val: key); |
| 218 | if (entry != cachedReads.end()) |
| 219 | return entry->second; |
| 220 | |
| 221 | ErrorOr<std::unique_ptr<MemoryBuffer>> mbOrErr = |
| 222 | MemoryBuffer::getFile(Filename: path, IsText: false, /*RequiresNullTerminator=*/false); |
| 223 | if (std::error_code ec = mbOrErr.getError()) { |
| 224 | error(msg: "cannot open "+ path + ": "+ ec.message()); |
| 225 | return std::nullopt; |
| 226 | } |
| 227 | |
| 228 | std::unique_ptr<MemoryBuffer> &mb = *mbOrErr; |
| 229 | MemoryBufferRef mbref = mb->getMemBufferRef(); |
| 230 | make<std::unique_ptr<MemoryBuffer>>(args: std::move(mb)); // take mb ownership |
| 231 | |
| 232 | // If this is a regular non-fat file, return it. |
| 233 | const char *buf = mbref.getBufferStart(); |
| 234 | const auto *hdr = reinterpret_cast<const fat_header *>(buf); |
| 235 | if (mbref.getBufferSize() < sizeof(uint32_t) || |
| 236 | read32be(P: &hdr->magic) != FAT_MAGIC) { |
| 237 | if (tar) |
| 238 | tar->append(Path: relativeToRoot(path), Data: mbref.getBuffer()); |
| 239 | return cachedReads[key] = mbref; |
| 240 | } |
| 241 | |
| 242 | llvm::BumpPtrAllocator &bAlloc = lld::bAlloc(); |
| 243 | |
| 244 | // Object files and archive files may be fat files, which contain multiple |
| 245 | // real files for different CPU ISAs. Here, we search for a file that matches |
| 246 | // with the current link target and returns it as a MemoryBufferRef. |
| 247 | const auto *arch = reinterpret_cast<const fat_arch *>(buf + sizeof(*hdr)); |
| 248 | auto getArchName = [](uint32_t cpuType, uint32_t cpuSubtype) { |
| 249 | return getArchitectureName(Arch: getArchitectureFromCpuType(CPUType: cpuType, CPUSubType: cpuSubtype)); |
| 250 | }; |
| 251 | |
| 252 | std::vector<StringRef> archs; |
| 253 | for (uint32_t i = 0, n = read32be(P: &hdr->nfat_arch); i < n; ++i) { |
| 254 | if (reinterpret_cast<const char *>(arch + i + 1) > |
| 255 | buf + mbref.getBufferSize()) { |
| 256 | error(msg: path + ": fat_arch struct extends beyond end of file"); |
| 257 | return std::nullopt; |
| 258 | } |
| 259 | |
| 260 | uint32_t cpuType = read32be(P: &arch[i].cputype); |
| 261 | uint32_t cpuSubtype = |
| 262 | read32be(P: &arch[i].cpusubtype) & ~MachO::CPU_SUBTYPE_MASK; |
| 263 | |
| 264 | // FIXME: LD64 has a more complex fallback logic here. |
| 265 | // Consider implementing that as well? |
| 266 | if (cpuType != static_cast<uint32_t>(target->cpuType) || |
| 267 | cpuSubtype != target->cpuSubtype) { |
| 268 | archs.emplace_back(args: getArchName(cpuType, cpuSubtype)); |
| 269 | continue; |
| 270 | } |
| 271 | |
| 272 | uint32_t offset = read32be(P: &arch[i].offset); |
| 273 | uint32_t size = read32be(P: &arch[i].size); |
| 274 | if (offset + size > mbref.getBufferSize()) |
| 275 | error(msg: path + ": slice extends beyond end of file"); |
| 276 | if (tar) |
| 277 | tar->append(Path: relativeToRoot(path), Data: mbref.getBuffer()); |
| 278 | return cachedReads[key] = MemoryBufferRef(StringRef(buf + offset, size), |
| 279 | path.copy(A&: bAlloc)); |
| 280 | } |
| 281 | |
| 282 | auto targetArchName = getArchName(target->cpuType, target->cpuSubtype); |
| 283 | warn(msg: path + ": ignoring file because it is universal ("+ join(R&: archs, Separator: ",") + |
| 284 | ") but does not contain the "+ targetArchName + " architecture"); |
| 285 | return std::nullopt; |
| 286 | } |
| 287 | |
| 288 | InputFile::InputFile(Kind kind, const InterfaceFile &interface) |
| 289 | : id(idCount++), fileKind(kind), name(saver().save(S: interface.getPath())) {} |
| 290 | |
| 291 | // Some sections comprise of fixed-size records, so instead of splitting them at |
| 292 | // symbol boundaries, we split them based on size. Records are distinct from |
| 293 | // literals in that they may contain references to other sections, instead of |
| 294 | // being leaf nodes in the InputSection graph. |
| 295 | // |
| 296 | // Note that "record" is a term I came up with. In contrast, "literal" is a term |
| 297 | // used by the Mach-O format. |
| 298 | static std::optional<size_t> getRecordSize(StringRef segname, StringRef name) { |
| 299 | if (name == section_names::compactUnwind) { |
| 300 | if (segname == segment_names::ld) |
| 301 | return target->wordSize == 8 ? 32 : 20; |
| 302 | } |
| 303 | if (!config->dedupStrings) |
| 304 | return {}; |
| 305 | |
| 306 | if (name == section_names::cfString && segname == segment_names::data) |
| 307 | return target->wordSize == 8 ? 32 : 16; |
| 308 | |
| 309 | if (config->icfLevel == ICFLevel::none) |
| 310 | return {}; |
| 311 | |
| 312 | if (name == section_names::objcClassRefs && segname == segment_names::data) |
| 313 | return target->wordSize; |
| 314 | |
| 315 | if (name == section_names::objcSelrefs && segname == segment_names::data) |
| 316 | return target->wordSize; |
| 317 | return {}; |
| 318 | } |
| 319 | |
| 320 | static Error parseCallGraph(ArrayRef<uint8_t> data, |
| 321 | std::vector<CallGraphEntry> &callGraph) { |
| 322 | TimeTraceScope timeScope("Parsing call graph section"); |
| 323 | BinaryStreamReader reader(data, llvm::endianness::little); |
| 324 | while (!reader.empty()) { |
| 325 | uint32_t fromIndex, toIndex; |
| 326 | uint64_t count; |
| 327 | if (Error err = reader.readInteger(Dest&: fromIndex)) |
| 328 | return err; |
| 329 | if (Error err = reader.readInteger(Dest&: toIndex)) |
| 330 | return err; |
| 331 | if (Error err = reader.readInteger(Dest&: count)) |
| 332 | return err; |
| 333 | callGraph.emplace_back(args&: fromIndex, args&: toIndex, args&: count); |
| 334 | } |
| 335 | return Error::success(); |
| 336 | } |
| 337 | |
| 338 | // Parse the sequence of sections within a single LC_SEGMENT(_64). |
| 339 | // Split each section into subsections. |
| 340 | template <class SectionHeader> |
| 341 | void ObjFile::parseSections(ArrayRef<SectionHeader> sectionHeaders) { |
| 342 | sections.reserve(n: sectionHeaders.size()); |
| 343 | auto *buf = reinterpret_cast<const uint8_t *>(mb.getBufferStart()); |
| 344 | |
| 345 | for (const SectionHeader &sec : sectionHeaders) { |
| 346 | StringRef name = |
| 347 | StringRef(sec.sectname, strnlen(sec.sectname, sizeof(sec.sectname))); |
| 348 | StringRef segname = |
| 349 | StringRef(sec.segname, strnlen(sec.segname, sizeof(sec.segname))); |
| 350 | sections.push_back(make<Section>(this, segname, name, sec.flags, sec.addr)); |
| 351 | if (sec.align >= 32) { |
| 352 | error("alignment "+ std::to_string(sec.align) + " of section "+ name + |
| 353 | " is too large"); |
| 354 | continue; |
| 355 | } |
| 356 | Section §ion = *sections.back(); |
| 357 | uint32_t align = 1 << sec.align; |
| 358 | ArrayRef<uint8_t> data = {isZeroFill(sec.flags) ? nullptr |
| 359 | : buf + sec.offset, |
| 360 | static_cast<size_t>(sec.size)}; |
| 361 | |
| 362 | auto splitRecords = [&](size_t recordSize) -> void { |
| 363 | if (data.empty()) |
| 364 | return; |
| 365 | Subsections &subsections = section.subsections; |
| 366 | subsections.reserve(n: data.size() / recordSize); |
| 367 | for (uint64_t off = 0; off < data.size(); off += recordSize) { |
| 368 | auto *isec = make<ConcatInputSection>( |
| 369 | args&: section, args: data.slice(N: off, M: std::min(a: data.size(), b: recordSize)), args&: align); |
| 370 | subsections.push_back(x: {.offset: off, .isec: isec}); |
| 371 | } |
| 372 | section.doneSplitting = true; |
| 373 | }; |
| 374 | |
| 375 | if (sectionType(sec.flags) == S_CSTRING_LITERALS) { |
| 376 | if (sec.nreloc) |
| 377 | fatal(toString(f: this) + ": "+ sec.segname + ","+ sec.sectname + |
| 378 | " contains relocations, which is unsupported"); |
| 379 | bool dedupLiterals = |
| 380 | name == section_names::objcMethname || config->dedupStrings; |
| 381 | InputSection *isec = |
| 382 | make<CStringInputSection>(args&: section, args&: data, args&: align, args&: dedupLiterals); |
| 383 | // FIXME: parallelize this? |
| 384 | cast<CStringInputSection>(Val: isec)->splitIntoPieces(); |
| 385 | section.subsections.push_back(x: {.offset: 0, .isec: isec}); |
| 386 | } else if (isWordLiteralSection(sec.flags)) { |
| 387 | if (sec.nreloc) |
| 388 | fatal(toString(f: this) + ": "+ sec.segname + ","+ sec.sectname + |
| 389 | " contains relocations, which is unsupported"); |
| 390 | InputSection *isec = make<WordLiteralInputSection>(args&: section, args&: data, args&: align); |
| 391 | section.subsections.push_back(x: {.offset: 0, .isec: isec}); |
| 392 | } else if (auto recordSize = getRecordSize(segname, name)) { |
| 393 | splitRecords(*recordSize); |
| 394 | } else if (name == section_names::ehFrame && |
| 395 | segname == segment_names::text) { |
| 396 | splitEhFrames(dataArr: data, ehFrameSection&: *sections.back()); |
| 397 | } else if (segname == segment_names::llvm) { |
| 398 | if (config->callGraphProfileSort && name == section_names::cgProfile) |
| 399 | checkError(e: parseCallGraph(data, callGraph)); |
| 400 | // ld64 does not appear to emit contents from sections within the __LLVM |
| 401 | // segment. Symbols within those sections point to bitcode metadata |
| 402 | // instead of actual symbols. Global symbols within those sections could |
| 403 | // have the same name without causing duplicate symbol errors. To avoid |
| 404 | // spurious duplicate symbol errors, we do not parse these sections. |
| 405 | // TODO: Evaluate whether the bitcode metadata is needed. |
| 406 | } else if (name == section_names::objCImageInfo && |
| 407 | segname == segment_names::data) { |
| 408 | objCImageInfo = data; |
| 409 | } else { |
| 410 | if (name == section_names::addrSig) |
| 411 | addrSigSection = sections.back(); |
| 412 | |
| 413 | auto *isec = make<ConcatInputSection>(args&: section, args&: data, args&: align); |
| 414 | if (isDebugSection(flags: isec->getFlags()) && |
| 415 | isec->getSegName() == segment_names::dwarf) { |
| 416 | // Instead of emitting DWARF sections, we emit STABS symbols to the |
| 417 | // object files that contain them. We filter them out early to avoid |
| 418 | // parsing their relocations unnecessarily. |
| 419 | debugSections.push_back(x: isec); |
| 420 | } else { |
| 421 | section.subsections.push_back(x: {.offset: 0, .isec: isec}); |
| 422 | } |
| 423 | } |
| 424 | } |
| 425 | } |
| 426 | |
| 427 | void ObjFile::splitEhFrames(ArrayRef<uint8_t> data, Section &ehFrameSection) { |
| 428 | EhReader reader(this, data, /*dataOff=*/0); |
| 429 | size_t off = 0; |
| 430 | while (off < reader.size()) { |
| 431 | uint64_t frameOff = off; |
| 432 | uint64_t length = reader.readLength(off: &off); |
| 433 | if (length == 0) |
| 434 | break; |
| 435 | uint64_t fullLength = length + (off - frameOff); |
| 436 | off += length; |
| 437 | // We hard-code an alignment of 1 here because we don't actually want our |
| 438 | // EH frames to be aligned to the section alignment. EH frame decoders don't |
| 439 | // expect this alignment. Moreover, each EH frame must start where the |
| 440 | // previous one ends, and where it ends is indicated by the length field. |
| 441 | // Unless we update the length field (troublesome), we should keep the |
| 442 | // alignment to 1. |
| 443 | // Note that we still want to preserve the alignment of the overall section, |
| 444 | // just not of the individual EH frames. |
| 445 | ehFrameSection.subsections.push_back( |
| 446 | x: {.offset: frameOff, .isec: make<ConcatInputSection>(args&: ehFrameSection, |
| 447 | args: data.slice(N: frameOff, M: fullLength), |
| 448 | /*align=*/args: 1)}); |
| 449 | } |
| 450 | ehFrameSection.doneSplitting = true; |
| 451 | } |
| 452 | |
| 453 | template <class T> |
| 454 | static Section *findContainingSection(const std::vector<Section *> §ions, |
| 455 | T *offset) { |
| 456 | static_assert(std::is_same<uint64_t, T>::value || |
| 457 | std::is_same<uint32_t, T>::value, |
| 458 | "unexpected type for offset"); |
| 459 | auto it = std::prev(llvm::upper_bound( |
| 460 | sections, *offset, |
| 461 | [](uint64_t value, const Section *sec) { return value < sec->addr; })); |
| 462 | *offset -= (*it)->addr; |
| 463 | return *it; |
| 464 | } |
| 465 | |
| 466 | // Find the subsection corresponding to the greatest section offset that is <= |
| 467 | // that of the given offset. |
| 468 | // |
| 469 | // offset: an offset relative to the start of the original InputSection (before |
| 470 | // any subsection splitting has occurred). It will be updated to represent the |
| 471 | // same location as an offset relative to the start of the containing |
| 472 | // subsection. |
| 473 | template <class T> |
| 474 | static InputSection *findContainingSubsection(const Section §ion, |
| 475 | T *offset) { |
| 476 | static_assert(std::is_same<uint64_t, T>::value || |
| 477 | std::is_same<uint32_t, T>::value, |
| 478 | "unexpected type for offset"); |
| 479 | auto it = std::prev(llvm::upper_bound( |
| 480 | section.subsections, *offset, |
| 481 | [](uint64_t value, Subsection subsec) { return value < subsec.offset; })); |
| 482 | *offset -= it->offset; |
| 483 | return it->isec; |
| 484 | } |
| 485 | |
| 486 | // Try to find a symbol at offset `off` within `isec`. |
| 487 | // Returns nullptr if no symbol exists at that offset. |
| 488 | static Defined *tryFindSymbolAtOffset(const ConcatInputSection *isec, |
| 489 | uint64_t off) { |
| 490 | auto it = llvm::lower_bound(Range: isec->symbols, Value&: off, C: [](Defined *d, uint64_t off) { |
| 491 | return d->value < off; |
| 492 | }); |
| 493 | if (it == isec->symbols.end() || (*it)->value != off) |
| 494 | return nullptr; |
| 495 | return *it; |
| 496 | } |
| 497 | |
| 498 | // Find a symbol at offset `off` within `isec`. |
| 499 | // If no symbol is found, assume the section must have been coalesced. |
| 500 | static Defined *findSymbolAtOffset(const ConcatInputSection *isec, |
| 501 | uint64_t off) { |
| 502 | Defined *d = tryFindSymbolAtOffset(isec, off); |
| 503 | // The offset should point at the exact address of a symbol (with no addend.) |
| 504 | assert(d || isec->wasCoalesced); |
| 505 | return d; |
| 506 | } |
| 507 | |
| 508 | template <class SectionHeader> |
| 509 | static bool validateRelocationInfo(InputFile *file, const SectionHeader &sec, |
| 510 | relocation_info rel) { |
| 511 | const RelocAttrs &relocAttrs = target->getRelocAttrs(type: rel.r_type); |
| 512 | bool valid = true; |
| 513 | auto message = [relocAttrs, file, sec, rel, &valid](const Twine &diagnostic) { |
| 514 | valid = false; |
| 515 | return (relocAttrs.name + " relocation "+ diagnostic + " at offset "+ |
| 516 | std::to_string(val: rel.r_address) + " of "+ sec.segname + ","+ |
| 517 | sec.sectname + " in "+ toString(f: file)) |
| 518 | .str(); |
| 519 | }; |
| 520 | |
| 521 | if (!relocAttrs.hasAttr(b: RelocAttrBits::LOCAL) && !rel.r_extern) |
| 522 | error(message("must be extern")); |
| 523 | if (relocAttrs.hasAttr(b: RelocAttrBits::PCREL) != rel.r_pcrel) |
| 524 | error(message(Twine("must ") + (rel.r_pcrel ? "not ": "") + |
| 525 | "be PC-relative")); |
| 526 | if (isThreadLocalVariables(sec.flags) && |
| 527 | !relocAttrs.hasAttr(b: RelocAttrBits::UNSIGNED)) |
| 528 | error(message("not allowed in thread-local section, must be UNSIGNED")); |
| 529 | if (!relocAttrs.hasAttr(b: static_cast<RelocAttrBits>(1 << rel.r_length))) { |
| 530 | error(message("has invalid width of "+ std::to_string(val: 1 << rel.r_length) + |
| 531 | " bytes")); |
| 532 | } |
| 533 | return valid; |
| 534 | } |
| 535 | |
| 536 | template <class SectionHeader> |
| 537 | void ObjFile::parseRelocations(ArrayRef<SectionHeader> sectionHeaders, |
| 538 | const SectionHeader &sec, Section §ion) { |
| 539 | auto *buf = reinterpret_cast<const uint8_t *>(mb.getBufferStart()); |
| 540 | ArrayRef<relocation_info> relInfos( |
| 541 | reinterpret_cast<const relocation_info *>(buf + sec.reloff), sec.nreloc); |
| 542 | |
| 543 | Subsections &subsections = section.subsections; |
| 544 | auto subsecIt = subsections.rbegin(); |
| 545 | for (size_t i = 0; i < relInfos.size(); i++) { |
| 546 | // Paired relocations serve as Mach-O's method for attaching a |
| 547 | // supplemental datum to a primary relocation record. ELF does not |
| 548 | // need them because the *_RELOC_RELA records contain the extra |
| 549 | // addend field, vs. *_RELOC_REL which omit the addend. |
| 550 | // |
| 551 | // The {X86_64,ARM64}_RELOC_SUBTRACTOR record holds the subtrahend, |
| 552 | // and the paired *_RELOC_UNSIGNED record holds the minuend. The |
| 553 | // datum for each is a symbolic address. The result is the offset |
| 554 | // between two addresses. |
| 555 | // |
| 556 | // The ARM64_RELOC_ADDEND record holds the addend, and the paired |
| 557 | // ARM64_RELOC_BRANCH26 or ARM64_RELOC_PAGE21/PAGEOFF12 holds the |
| 558 | // base symbolic address. |
| 559 | // |
| 560 | // Note: X86 does not use *_RELOC_ADDEND because it can embed an addend into |
| 561 | // the instruction stream. On X86, a relocatable address field always |
| 562 | // occupies an entire contiguous sequence of byte(s), so there is no need to |
| 563 | // merge opcode bits with address bits. Therefore, it's easy and convenient |
| 564 | // to store addends in the instruction-stream bytes that would otherwise |
| 565 | // contain zeroes. By contrast, RISC ISAs such as ARM64 mix opcode bits with |
| 566 | // address bits so that bitwise arithmetic is necessary to extract and |
| 567 | // insert them. Storing addends in the instruction stream is possible, but |
| 568 | // inconvenient and more costly at link time. |
| 569 | |
| 570 | relocation_info relInfo = relInfos[i]; |
| 571 | bool isSubtrahend = |
| 572 | target->hasAttr(type: relInfo.r_type, bit: RelocAttrBits::SUBTRAHEND); |
| 573 | int64_t pairedAddend = 0; |
| 574 | if (target->hasAttr(type: relInfo.r_type, bit: RelocAttrBits::ADDEND)) { |
| 575 | pairedAddend = SignExtend64<24>(x: relInfo.r_symbolnum); |
| 576 | relInfo = relInfos[++i]; |
| 577 | } |
| 578 | assert(i < relInfos.size()); |
| 579 | if (!validateRelocationInfo(this, sec, relInfo)) |
| 580 | continue; |
| 581 | if (relInfo.r_address & R_SCATTERED) |
| 582 | fatal(msg: "TODO: Scattered relocations not supported"); |
| 583 | |
| 584 | int64_t embeddedAddend = target->getEmbeddedAddend(mb, offset: sec.offset, relInfo); |
| 585 | assert(!(embeddedAddend && pairedAddend)); |
| 586 | int64_t totalAddend = pairedAddend + embeddedAddend; |
| 587 | Relocation r; |
| 588 | r.type = relInfo.r_type; |
| 589 | r.pcrel = relInfo.r_pcrel; |
| 590 | r.length = relInfo.r_length; |
| 591 | r.offset = relInfo.r_address; |
| 592 | if (relInfo.r_extern) { |
| 593 | r.referent = symbols[relInfo.r_symbolnum]; |
| 594 | r.addend = isSubtrahend ? 0 : totalAddend; |
| 595 | } else { |
| 596 | assert(!isSubtrahend); |
| 597 | const SectionHeader &referentSecHead = |
| 598 | sectionHeaders[relInfo.r_symbolnum - 1]; |
| 599 | uint64_t referentOffset; |
| 600 | if (relInfo.r_pcrel) { |
| 601 | // The implicit addend for pcrel section relocations is the pcrel offset |
| 602 | // in terms of the addresses in the input file. Here we adjust it so |
| 603 | // that it describes the offset from the start of the referent section. |
| 604 | // FIXME This logic was written around x86_64 behavior -- ARM64 doesn't |
| 605 | // have pcrel section relocations. We may want to factor this out into |
| 606 | // the arch-specific .cpp file. |
| 607 | referentOffset = sec.addr + relInfo.r_address + |
| 608 | (1ull << relInfo.r_length) + totalAddend - |
| 609 | referentSecHead.addr; |
| 610 | } else { |
| 611 | // The addend for a non-pcrel relocation is its absolute address. |
| 612 | referentOffset = totalAddend - referentSecHead.addr; |
| 613 | } |
| 614 | r.referent = findContainingSubsection(section: *sections[relInfo.r_symbolnum - 1], |
| 615 | offset: &referentOffset); |
| 616 | r.addend = referentOffset; |
| 617 | } |
| 618 | |
| 619 | // Find the subsection that this relocation belongs to. |
| 620 | // Though not required by the Mach-O format, clang and gcc seem to emit |
| 621 | // relocations in order, so let's take advantage of it. However, ld64 emits |
| 622 | // unsorted relocations (in `-r` mode), so we have a fallback for that |
| 623 | // uncommon case. |
| 624 | InputSection *subsec; |
| 625 | while (subsecIt != subsections.rend() && subsecIt->offset > r.offset) |
| 626 | ++subsecIt; |
| 627 | if (subsecIt == subsections.rend() || |
| 628 | subsecIt->offset + subsecIt->isec->getSize() <= r.offset) { |
| 629 | subsec = findContainingSubsection(section, offset: &r.offset); |
| 630 | // Now that we know the relocs are unsorted, avoid trying the 'fast path' |
| 631 | // for the other relocations. |
| 632 | subsecIt = subsections.rend(); |
| 633 | } else { |
| 634 | subsec = subsecIt->isec; |
| 635 | r.offset -= subsecIt->offset; |
| 636 | } |
| 637 | subsec->relocs.push_back(x: r); |
| 638 | |
| 639 | if (isSubtrahend) { |
| 640 | relocation_info minuendInfo = relInfos[++i]; |
| 641 | // SUBTRACTOR relocations should always be followed by an UNSIGNED one |
| 642 | // attached to the same address. |
| 643 | assert(target->hasAttr(minuendInfo.r_type, RelocAttrBits::UNSIGNED) && |
| 644 | relInfo.r_address == minuendInfo.r_address); |
| 645 | Relocation p; |
| 646 | p.type = minuendInfo.r_type; |
| 647 | p.pcrel = minuendInfo.r_pcrel; |
| 648 | p.length = minuendInfo.r_length; |
| 649 | p.offset = r.offset; |
| 650 | if (minuendInfo.r_extern) { |
| 651 | p.referent = symbols[minuendInfo.r_symbolnum]; |
| 652 | p.addend = totalAddend; |
| 653 | } else { |
| 654 | uint64_t referentOffset = |
| 655 | totalAddend - sectionHeaders[minuendInfo.r_symbolnum - 1].addr; |
| 656 | p.referent = findContainingSubsection( |
| 657 | section: *sections[minuendInfo.r_symbolnum - 1], offset: &referentOffset); |
| 658 | p.addend = referentOffset; |
| 659 | } |
| 660 | subsec->relocs.push_back(x: p); |
| 661 | } |
| 662 | } |
| 663 | } |
| 664 | |
| 665 | // ld64 never turns these labels into named atoms or symbol table entries. |
| 666 | static bool shouldIgnoreLabel(const InputSection *isec, StringRef name) { |
| 667 | if (isCfStringSection(isec) || isClassRefsSection(isec) || |
| 668 | isSelRefsSection(isec)) |
| 669 | return true; |
| 670 | if ((isa<WordLiteralInputSection>(Val: isec) || isa<CStringInputSection>(Val: isec)) && |
| 671 | isPrivateLabel(name)) |
| 672 | return true; |
| 673 | return false; |
| 674 | } |
| 675 | |
| 676 | template <class NList> |
| 677 | static macho::Symbol *createDefined(const NList &sym, StringRef name, |
| 678 | InputSection *isec, uint64_t value, |
| 679 | uint64_t size, bool forceHidden) { |
| 680 | // Symbol scope is determined by sym.n_type & (N_EXT | N_PEXT): |
| 681 | // N_EXT: Global symbols. These go in the symbol table during the link, |
| 682 | // and also in the export table of the output so that the dynamic |
| 683 | // linker sees them. |
| 684 | // N_EXT | N_PEXT: Linkage unit (think: dylib) scoped. These go in the |
| 685 | // symbol table during the link so that duplicates are |
| 686 | // either reported (for non-weak symbols) or merged |
| 687 | // (for weak symbols), but they do not go in the export |
| 688 | // table of the output. |
| 689 | // N_PEXT: llvm-mc does not emit these, but `ld -r` (wherein ld64 emits |
| 690 | // object files) may produce them. LLD does not yet support -r. |
| 691 | // These are translation-unit scoped, identical to the `0` case. |
| 692 | // 0: Translation-unit scoped. These are not in the symbol table during |
| 693 | // link, and not in the export table of the output either. |
| 694 | bool isWeakDefCanBeHidden = |
| 695 | (sym.n_desc & (N_WEAK_DEF | N_WEAK_REF)) == (N_WEAK_DEF | N_WEAK_REF); |
| 696 | |
| 697 | assert(!(sym.n_desc & N_ARM_THUMB_DEF) && "ARM32 arch is not supported"); |
| 698 | |
| 699 | bool isCold = sym.n_desc & N_COLD_FUNC; |
| 700 | |
| 701 | if ((sym.n_type & N_EXT) && !shouldIgnoreLabel(isec, name)) { |
| 702 | // -load_hidden makes us treat global symbols as linkage unit scoped. |
| 703 | // Duplicates are reported but the symbol does not go in the export trie. |
| 704 | bool isPrivateExtern = sym.n_type & N_PEXT || forceHidden; |
| 705 | |
| 706 | // lld's behavior for merging symbols is slightly different from ld64: |
| 707 | // ld64 picks the winning symbol based on several criteria (see |
| 708 | // pickBetweenRegularAtoms() in ld64's SymbolTable.cpp), while lld |
| 709 | // just merges metadata and keeps the contents of the first symbol |
| 710 | // with that name (see SymbolTable::addDefined). For: |
| 711 | // * inline function F in a TU built with -fvisibility-inlines-hidden |
| 712 | // * and inline function F in another TU built without that flag |
| 713 | // ld64 will pick the one from the file built without |
| 714 | // -fvisibility-inlines-hidden. |
| 715 | // lld will instead pick the one listed first on the link command line and |
| 716 | // give it visibility as if the function was built without |
| 717 | // -fvisibility-inlines-hidden. |
| 718 | // If both functions have the same contents, this will have the same |
| 719 | // behavior. If not, it won't, but the input had an ODR violation in |
| 720 | // that case. |
| 721 | // |
| 722 | // Similarly, merging a symbol |
| 723 | // that's isPrivateExtern and not isWeakDefCanBeHidden with one |
| 724 | // that's not isPrivateExtern but isWeakDefCanBeHidden technically |
| 725 | // should produce one |
| 726 | // that's not isPrivateExtern but isWeakDefCanBeHidden. That matters |
| 727 | // with ld64's semantics, because it means the non-private-extern |
| 728 | // definition will continue to take priority if more private extern |
| 729 | // definitions are encountered. With lld's semantics there's no observable |
| 730 | // difference between a symbol that's isWeakDefCanBeHidden(autohide) or one |
| 731 | // that's privateExtern -- neither makes it into the dynamic symbol table, |
| 732 | // unless the autohide symbol is explicitly exported. |
| 733 | // But if a symbol is both privateExtern and autohide then it can't |
| 734 | // be exported. |
| 735 | // So we nullify the autohide flag when privateExtern is present |
| 736 | // and promote the symbol to privateExtern when it is not already. |
| 737 | if (isWeakDefCanBeHidden && isPrivateExtern) |
| 738 | isWeakDefCanBeHidden = false; |
| 739 | else if (isWeakDefCanBeHidden) |
| 740 | isPrivateExtern = true; |
| 741 | return symtab->addDefined( |
| 742 | name, isec->getFile(), isec, value, size, isWeakDef: sym.n_desc & N_WEAK_DEF, |
| 743 | isPrivateExtern, isReferencedDynamically: sym.n_desc & REFERENCED_DYNAMICALLY, |
| 744 | noDeadStrip: sym.n_desc & N_NO_DEAD_STRIP, isWeakDefCanBeHidden, isCold); |
| 745 | } |
| 746 | bool includeInSymtab = !isPrivateLabel(name) && !isEhFrameSection(isec); |
| 747 | auto *defined = make<Defined>( |
| 748 | name, isec->getFile(), isec, value, size, sym.n_desc & N_WEAK_DEF, |
| 749 | /*isExternal=*/false, /*isPrivateExtern=*/false, includeInSymtab, |
| 750 | sym.n_desc & REFERENCED_DYNAMICALLY, sym.n_desc & N_NO_DEAD_STRIP); |
| 751 | defined->cold = isCold; |
| 752 | return defined; |
| 753 | } |
| 754 | |
| 755 | // Absolute symbols are defined symbols that do not have an associated |
| 756 | // InputSection. They cannot be weak. |
| 757 | template <class NList> |
| 758 | static macho::Symbol *createAbsolute(const NList &sym, InputFile *file, |
| 759 | StringRef name, bool forceHidden) { |
| 760 | bool isCold = sym.n_desc & N_COLD_FUNC; |
| 761 | assert(!(sym.n_desc & N_ARM_THUMB_DEF) && "ARM32 arch is not supported"); |
| 762 | |
| 763 | if (sym.n_type & N_EXT) { |
| 764 | bool isPrivateExtern = sym.n_type & N_PEXT || forceHidden; |
| 765 | return symtab->addDefined(name, file, nullptr, value: sym.n_value, /*size=*/0, |
| 766 | /*isWeakDef=*/false, isPrivateExtern, |
| 767 | /*isReferencedDynamically=*/false, |
| 768 | noDeadStrip: sym.n_desc & N_NO_DEAD_STRIP, |
| 769 | /*isWeakDefCanBeHidden=*/false, isCold); |
| 770 | } |
| 771 | auto *defined = make<Defined>(name, file, nullptr, sym.n_value, /*size=*/0, |
| 772 | /*isWeakDef=*/false, |
| 773 | /*isExternal=*/false, /*isPrivateExtern=*/false, |
| 774 | /*includeInSymtab=*/true, |
| 775 | /*isReferencedDynamically=*/false, |
| 776 | sym.n_desc & N_NO_DEAD_STRIP); |
| 777 | defined->cold = isCold; |
| 778 | return defined; |
| 779 | } |
| 780 | |
| 781 | template <class NList> |
| 782 | macho::Symbol *ObjFile::parseNonSectionSymbol(const NList &sym, |
| 783 | const char *strtab) { |
| 784 | StringRef name = StringRef(strtab + sym.n_strx); |
| 785 | uint8_t type = sym.n_type & N_TYPE; |
| 786 | bool isPrivateExtern = sym.n_type & N_PEXT || forceHidden; |
| 787 | switch (type) { |
| 788 | case N_UNDF: |
| 789 | return sym.n_value == 0 |
| 790 | ? symtab->addUndefined(name, this, isWeakRef: sym.n_desc & N_WEAK_REF) |
| 791 | : symtab->addCommon(name, this, size: sym.n_value, |
| 792 | align: 1 << GET_COMM_ALIGN(sym.n_desc), |
| 793 | isPrivateExtern); |
| 794 | case N_ABS: |
| 795 | return createAbsolute(sym, this, name, forceHidden); |
| 796 | case N_INDR: { |
| 797 | // Not much point in making local aliases -- relocs in the current file can |
| 798 | // just refer to the actual symbol itself. ld64 ignores these symbols too. |
| 799 | if (!(sym.n_type & N_EXT)) |
| 800 | return nullptr; |
| 801 | StringRef aliasedName = StringRef(strtab + sym.n_value); |
| 802 | // isPrivateExtern is the only symbol flag that has an impact on the final |
| 803 | // aliased symbol. |
| 804 | auto *alias = make<AliasSymbol>(args: this, args&: name, args&: aliasedName, args&: isPrivateExtern); |
| 805 | aliases.push_back(x: alias); |
| 806 | return alias; |
| 807 | } |
| 808 | case N_PBUD: |
| 809 | error(msg: "TODO: support symbols of type N_PBUD"); |
| 810 | return nullptr; |
| 811 | case N_SECT: |
| 812 | llvm_unreachable( |
| 813 | "N_SECT symbols should not be passed to parseNonSectionSymbol"); |
| 814 | default: |
| 815 | llvm_unreachable("invalid symbol type"); |
| 816 | } |
| 817 | } |
| 818 | |
| 819 | template <class NList> static bool isUndef(const NList &sym) { |
| 820 | return (sym.n_type & N_TYPE) == N_UNDF && sym.n_value == 0; |
| 821 | } |
| 822 | |
| 823 | template <class LP> |
| 824 | void ObjFile::parseSymbols(ArrayRef<typename LP::section> sectionHeaders, |
| 825 | ArrayRef<typename LP::nlist> nList, |
| 826 | const char *strtab, bool subsectionsViaSymbols) { |
| 827 | using NList = typename LP::nlist; |
| 828 | |
| 829 | // Groups indices of the symbols by the sections that contain them. |
| 830 | std::vector<std::vector<uint32_t>> symbolsBySection(sections.size()); |
| 831 | symbols.resize(nList.size()); |
| 832 | SmallVector<unsigned, 32> undefineds; |
| 833 | for (uint32_t i = 0; i < nList.size(); ++i) { |
| 834 | const NList &sym = nList[i]; |
| 835 | |
| 836 | // Ignore debug symbols for now. |
| 837 | // FIXME: may need special handling. |
| 838 | if (sym.n_type & N_STAB) |
| 839 | continue; |
| 840 | |
| 841 | if ((sym.n_type & N_TYPE) == N_SECT) { |
| 842 | if (sym.n_sect == 0) { |
| 843 | fatal(msg: "section symbol "+ StringRef(strtab + sym.n_strx) + " in "+ |
| 844 | toString(f: this) + " has an invalid section index [0]"); |
| 845 | } |
| 846 | if (sym.n_sect > sections.size()) { |
| 847 | fatal(msg: "section symbol "+ StringRef(strtab + sym.n_strx) + " in "+ |
| 848 | toString(f: this) + " has an invalid section index ["+ |
| 849 | Twine(static_cast<unsigned>(sym.n_sect)) + |
| 850 | "] greater than the total number of sections ["+ |
| 851 | Twine(sections.size()) + "]"); |
| 852 | } |
| 853 | Subsections &subsections = sections[sym.n_sect - 1]->subsections; |
| 854 | // parseSections() may have chosen not to parse this section. |
| 855 | if (subsections.empty()) |
| 856 | continue; |
| 857 | symbolsBySection[sym.n_sect - 1].push_back(i); |
| 858 | } else if (isUndef(sym)) { |
| 859 | undefineds.push_back(Elt: i); |
| 860 | } else { |
| 861 | symbols[i] = parseNonSectionSymbol(sym, strtab); |
| 862 | } |
| 863 | } |
| 864 | |
| 865 | for (size_t i = 0; i < sections.size(); ++i) { |
| 866 | Subsections &subsections = sections[i]->subsections; |
| 867 | if (subsections.empty()) |
| 868 | continue; |
| 869 | std::vector<uint32_t> &symbolIndices = symbolsBySection[i]; |
| 870 | uint64_t sectionAddr = sectionHeaders[i].addr; |
| 871 | uint32_t sectionAlign = 1u << sectionHeaders[i].align; |
| 872 | |
| 873 | // Some sections have already been split into subsections during |
| 874 | // parseSections(), so we simply need to match Symbols to the corresponding |
| 875 | // subsection here. |
| 876 | if (sections[i]->doneSplitting) { |
| 877 | for (size_t j = 0; j < symbolIndices.size(); ++j) { |
| 878 | const uint32_t symIndex = symbolIndices[j]; |
| 879 | const NList &sym = nList[symIndex]; |
| 880 | StringRef name = strtab + sym.n_strx; |
| 881 | uint64_t symbolOffset = sym.n_value - sectionAddr; |
| 882 | InputSection *isec = |
| 883 | findContainingSubsection(section: *sections[i], offset: &symbolOffset); |
| 884 | if (symbolOffset != 0) { |
| 885 | error(msg: toString(sec: *sections[i]) + ": symbol "+ name + |
| 886 | " at misaligned offset"); |
| 887 | continue; |
| 888 | } |
| 889 | symbols[symIndex] = |
| 890 | createDefined(sym, name, isec, 0, isec->getSize(), forceHidden); |
| 891 | } |
| 892 | continue; |
| 893 | } |
| 894 | sections[i]->doneSplitting = true; |
| 895 | |
| 896 | auto getSymName = [strtab](const NList& sym) -> StringRef { |
| 897 | return StringRef(strtab + sym.n_strx); |
| 898 | }; |
| 899 | |
| 900 | // Calculate symbol sizes and create subsections by splitting the sections |
| 901 | // along symbol boundaries. |
| 902 | // We populate subsections by repeatedly splitting the last (highest |
| 903 | // address) subsection. |
| 904 | llvm::stable_sort(symbolIndices, [&](uint32_t lhs, uint32_t rhs) { |
| 905 | // Put extern weak symbols after other symbols at the same address so |
| 906 | // that weak symbol coalescing works correctly. See |
| 907 | // SymbolTable::addDefined() for details. |
| 908 | if (nList[lhs].n_value == nList[rhs].n_value && |
| 909 | nList[lhs].n_type & N_EXT && nList[rhs].n_type & N_EXT) |
| 910 | return !(nList[lhs].n_desc & N_WEAK_DEF) && (nList[rhs].n_desc & N_WEAK_DEF); |
| 911 | return nList[lhs].n_value < nList[rhs].n_value; |
| 912 | }); |
| 913 | size_t sameAddrGroupIdx = 0; |
| 914 | bool sameAddrHasRegularSymbol = false; |
| 915 | for (size_t j = 0; j < symbolIndices.size(); ++j) { |
| 916 | const uint32_t symIndex = symbolIndices[j]; |
| 917 | const NList &sym = nList[symIndex]; |
| 918 | // An ordinary symbol establishes an atom boundary for every symbol at |
| 919 | // its address, even when a local alt entry precedes it in the nlist. |
| 920 | if (j == sameAddrGroupIdx) { |
| 921 | sameAddrHasRegularSymbol = false; |
| 922 | while (sameAddrGroupIdx < symbolIndices.size()) { |
| 923 | const NList &sameAddrSym = nList[symbolIndices[sameAddrGroupIdx]]; |
| 924 | if (sameAddrSym.n_value != sym.n_value) |
| 925 | break; |
| 926 | sameAddrHasRegularSymbol |= !(sameAddrSym.n_desc & N_ALT_ENTRY); |
| 927 | ++sameAddrGroupIdx; |
| 928 | } |
| 929 | } |
| 930 | StringRef name = getSymName(sym); |
| 931 | Subsection &subsec = subsections.back(); |
| 932 | InputSection *isec = subsec.isec; |
| 933 | |
| 934 | uint64_t subsecAddr = sectionAddr + subsec.offset; |
| 935 | size_t symbolOffset = sym.n_value - subsecAddr; |
| 936 | uint64_t symbolSize = |
| 937 | j + 1 < symbolIndices.size() |
| 938 | ? nList[symbolIndices[j + 1]].n_value - sym.n_value |
| 939 | : isec->data.size() - symbolOffset; |
| 940 | const bool isInteriorAltEntry = |
| 941 | (sym.n_desc & N_ALT_ENTRY) && !sameAddrHasRegularSymbol; |
| 942 | // There are 4 cases where we do not need to create a new subsection: |
| 943 | // 1. If the input file does not use subsections-via-symbols. |
| 944 | // 2. Multiple symbols at the same address only induce one subsection. |
| 945 | // (The symbolOffset == 0 check covers both this case as well as |
| 946 | // the first loop iteration.) |
| 947 | // 3. Alternative entry points without a coincident ordinary symbol do |
| 948 | // not induce new subsections. |
| 949 | // 4. If we have a literal section (e.g. __cstring and __literal4). |
| 950 | if (!subsectionsViaSymbols || symbolOffset == 0 || isInteriorAltEntry || |
| 951 | !isa<ConcatInputSection>(Val: isec)) { |
| 952 | isec->hasAltEntry = symbolOffset != 0; |
| 953 | symbols[symIndex] = createDefined(sym, name, isec, symbolOffset, |
| 954 | symbolSize, forceHidden); |
| 955 | continue; |
| 956 | } |
| 957 | auto *concatIsec = cast<ConcatInputSection>(Val: isec); |
| 958 | |
| 959 | auto *nextIsec = make<ConcatInputSection>(args&: *concatIsec); |
| 960 | nextIsec->wasCoalesced = false; |
| 961 | if (isZeroFill(flags: isec->getFlags())) { |
| 962 | // Zero-fill sections have NULL data.data() non-zero data.size() |
| 963 | nextIsec->data = {nullptr, isec->data.size() - symbolOffset}; |
| 964 | isec->data = {nullptr, symbolOffset}; |
| 965 | } else { |
| 966 | nextIsec->data = isec->data.slice(N: symbolOffset); |
| 967 | isec->data = isec->data.slice(N: 0, M: symbolOffset); |
| 968 | } |
| 969 | |
| 970 | // By construction, the symbol will be at offset zero in the new |
| 971 | // subsection. |
| 972 | symbols[symIndex] = createDefined(sym, name, nextIsec, /*value=*/0, |
| 973 | symbolSize, forceHidden); |
| 974 | // TODO: ld64 appears to preserve the original alignment as well as each |
| 975 | // subsection's offset from the last aligned address. We should consider |
| 976 | // emulating that behavior. |
| 977 | nextIsec->align = MinAlign(sectionAlign, sym.n_value); |
| 978 | subsections.push_back({sym.n_value - sectionAddr, nextIsec}); |
| 979 | } |
| 980 | } |
| 981 | |
| 982 | // Undefined symbols can trigger recursive fetch from Archives due to |
| 983 | // LazySymbols. Process defined symbols first so that the relative order |
| 984 | // between a defined symbol and an undefined symbol does not change the |
| 985 | // symbol resolution behavior. In addition, a set of interconnected symbols |
| 986 | // will all be resolved to the same file, instead of being resolved to |
| 987 | // different files. |
| 988 | for (unsigned i : undefineds) |
| 989 | symbols[i] = parseNonSectionSymbol(nList[i], strtab); |
| 990 | } |
| 991 | |
| 992 | OpaqueFile::OpaqueFile(MemoryBufferRef mb, StringRef segName, |
| 993 | StringRef sectName) |
| 994 | : InputFile(OpaqueKind, mb) { |
| 995 | const auto *buf = reinterpret_cast<const uint8_t *>(mb.getBufferStart()); |
| 996 | ArrayRef<uint8_t> data = {buf, mb.getBufferSize()}; |
| 997 | sections.push_back(x: make<Section>(/*file=*/args: this, args: segName.take_front(N: 16), |
| 998 | args: sectName.take_front(N: 16), |
| 999 | /*flags=*/args: 0, /*addr=*/args: 0)); |
| 1000 | Section §ion = *sections.back(); |
| 1001 | ConcatInputSection *isec = make<ConcatInputSection>(args&: section, args&: data); |
| 1002 | isec->live = true; |
| 1003 | section.subsections.push_back(x: {.offset: 0, .isec: isec}); |
| 1004 | } |
| 1005 | |
| 1006 | template <class LP> |
| 1007 | void ObjFile::parseLinkerOptions(SmallVectorImpl<StringRef> &LCLinkerOptions) { |
| 1008 | using Header = typename LP::mach_header; |
| 1009 | auto *hdr = reinterpret_cast<const Header *>(mb.getBufferStart()); |
| 1010 | |
| 1011 | for (auto *cmd : findCommands<linker_option_command>(hdr, LC_LINKER_OPTION)) { |
| 1012 | StringRef data{reinterpret_cast<const char *>(cmd + 1), |
| 1013 | cmd->cmdsize - sizeof(linker_option_command)}; |
| 1014 | parseLCLinkerOption(LCLinkerOptions, this, cmd->count, data); |
| 1015 | } |
| 1016 | } |
| 1017 | |
| 1018 | SmallVector<StringRef> macho::unprocessedLCLinkerOptions; |
| 1019 | ObjFile::ObjFile(MemoryBufferRef mb, uint32_t modTime, StringRef archiveName, |
| 1020 | bool lazy, bool forceHidden, bool compatArch, |
| 1021 | bool builtFromBitcode) |
| 1022 | : InputFile(ObjKind, mb, lazy), modTime(modTime), forceHidden(forceHidden), |
| 1023 | builtFromBitcode(builtFromBitcode) { |
| 1024 | this->archiveName = std::string(archiveName); |
| 1025 | this->compatArch = compatArch; |
| 1026 | if (lazy) { |
| 1027 | if (target->wordSize == 8) |
| 1028 | parseLazy<LP64>(); |
| 1029 | else |
| 1030 | parseLazy<ILP32>(); |
| 1031 | } else { |
| 1032 | if (target->wordSize == 8) |
| 1033 | parse<LP64>(); |
| 1034 | else |
| 1035 | parse<ILP32>(); |
| 1036 | } |
| 1037 | } |
| 1038 | |
| 1039 | static bool isUnwindSection(const Section &sec) { |
| 1040 | return sec.name == section_names::compactUnwind || |
| 1041 | sec.name == section_names::ehFrame; |
| 1042 | } |
| 1043 | |
| 1044 | template <class LP> void ObjFile::parse() { |
| 1045 | using Header = typename LP::mach_header; |
| 1046 | using SegmentCommand = typename LP::segment_command; |
| 1047 | using SectionHeader = typename LP::section; |
| 1048 | using NList = typename LP::nlist; |
| 1049 | |
| 1050 | auto *buf = reinterpret_cast<const uint8_t *>(mb.getBufferStart()); |
| 1051 | auto *hdr = reinterpret_cast<const Header *>(mb.getBufferStart()); |
| 1052 | |
| 1053 | // If we've already checked the arch, then don't need to check again. |
| 1054 | if (!compatArch) |
| 1055 | return; |
| 1056 | if (!(compatArch = compatWithTargetArch(this, hdr))) |
| 1057 | return; |
| 1058 | |
| 1059 | // We will resolve LC linker options once all native objects are loaded after |
| 1060 | // LTO is finished. |
| 1061 | SmallVector<StringRef, 4> LCLinkerOptions; |
| 1062 | parseLinkerOptions<LP>(LCLinkerOptions); |
| 1063 | unprocessedLCLinkerOptions.append(RHS: LCLinkerOptions); |
| 1064 | |
| 1065 | ArrayRef<SectionHeader> sectionHeaders; |
| 1066 | if (const load_command *cmd = findCommand(hdr, LP::segmentLCType)) { |
| 1067 | auto *c = reinterpret_cast<const SegmentCommand *>(cmd); |
| 1068 | sectionHeaders = ArrayRef<SectionHeader>{ |
| 1069 | reinterpret_cast<const SectionHeader *>(c + 1), c->nsects}; |
| 1070 | parseSections(sectionHeaders); |
| 1071 | } |
| 1072 | |
| 1073 | // TODO: Error on missing LC_SYMTAB? |
| 1074 | if (const load_command *cmd = findCommand(hdr, LC_SYMTAB)) { |
| 1075 | auto *c = reinterpret_cast<const symtab_command *>(cmd); |
| 1076 | ArrayRef<NList> nList(reinterpret_cast<const NList *>(buf + c->symoff), |
| 1077 | c->nsyms); |
| 1078 | const char *strtab = reinterpret_cast<const char *>(buf) + c->stroff; |
| 1079 | bool subsectionsViaSymbols = hdr->flags & MH_SUBSECTIONS_VIA_SYMBOLS; |
| 1080 | if (config->warnMissingSubsectionsViaSymbols && !subsectionsViaSymbols && |
| 1081 | !sectionHeaders.empty()) |
| 1082 | warn(msg: toString(f: this) + ": missing MH_SUBSECTIONS_VIA_SYMBOLS"); |
| 1083 | parseSymbols<LP>(sectionHeaders, nList, strtab, subsectionsViaSymbols); |
| 1084 | } |
| 1085 | |
| 1086 | // The relocations may refer to the symbols, so we parse them after we have |
| 1087 | // parsed all the symbols. |
| 1088 | // |
| 1089 | // Parse sections that are consumed by registerCompactUnwind() and |
| 1090 | // registerEhFrames() immediately. The rest can be deferred to be done in |
| 1091 | // parallel. |
| 1092 | for (size_t i = 0, n = sections.size(); i < n; ++i) |
| 1093 | if (!sections[i]->subsections.empty() && isUnwindSection(sec: *sections[i])) |
| 1094 | parseRelocations(sectionHeaders, sectionHeaders[i], *sections[i]); |
| 1095 | |
| 1096 | parseDebugInfo(); |
| 1097 | |
| 1098 | Section *ehFrameSection = nullptr; |
| 1099 | Section *compactUnwindSection = nullptr; |
| 1100 | for (Section *sec : sections) { |
| 1101 | Section **s = StringSwitch<Section **>(sec->name) |
| 1102 | .Case(S: section_names::compactUnwind, Value: &compactUnwindSection) |
| 1103 | .Case(S: section_names::ehFrame, Value: &ehFrameSection) |
| 1104 | .Default(Value: nullptr); |
| 1105 | if (s) |
| 1106 | *s = sec; |
| 1107 | } |
| 1108 | if (compactUnwindSection) |
| 1109 | registerCompactUnwind(compactUnwindSection&: *compactUnwindSection); |
| 1110 | if (ehFrameSection) |
| 1111 | registerEhFrames(ehFrameSection&: *ehFrameSection); |
| 1112 | } |
| 1113 | |
| 1114 | template <class LP> void ObjFile::parseLazy() { |
| 1115 | using Header = typename LP::mach_header; |
| 1116 | using NList = typename LP::nlist; |
| 1117 | |
| 1118 | auto *buf = reinterpret_cast<const uint8_t *>(mb.getBufferStart()); |
| 1119 | auto *hdr = reinterpret_cast<const Header *>(mb.getBufferStart()); |
| 1120 | |
| 1121 | if (!compatArch) |
| 1122 | return; |
| 1123 | if (!(compatArch = compatWithTargetArch(this, hdr))) |
| 1124 | return; |
| 1125 | |
| 1126 | const load_command *cmd = findCommand(hdr, LC_SYMTAB); |
| 1127 | if (!cmd) |
| 1128 | return; |
| 1129 | auto *c = reinterpret_cast<const symtab_command *>(cmd); |
| 1130 | ArrayRef<NList> nList(reinterpret_cast<const NList *>(buf + c->symoff), |
| 1131 | c->nsyms); |
| 1132 | const char *strtab = reinterpret_cast<const char *>(buf) + c->stroff; |
| 1133 | symbols.resize(nList.size()); |
| 1134 | for (const auto &[i, sym] : llvm::enumerate(nList)) { |
| 1135 | if ((sym.n_type & N_EXT) && !isUndef(sym)) { |
| 1136 | // TODO: Bound checking |
| 1137 | StringRef name = strtab + sym.n_strx; |
| 1138 | symbols[i] = symtab->addLazyObject(name, file&: *this); |
| 1139 | if (!lazy) |
| 1140 | break; |
| 1141 | } |
| 1142 | } |
| 1143 | } |
| 1144 | |
| 1145 | void ObjFile::parseDebugInfo() { |
| 1146 | std::unique_ptr<DwarfObject> dObj = DwarfObject::create(this); |
| 1147 | if (!dObj) |
| 1148 | return; |
| 1149 | |
| 1150 | // We do not re-use the context from getDwarf() here as that function |
| 1151 | // constructs an expensive DWARFCache object. |
| 1152 | auto *ctx = make<DWARFContext>( |
| 1153 | args: std::move(dObj), args: "", |
| 1154 | args: [&](Error err) { |
| 1155 | warn(msg: toString(f: this) + ": "+ toString(E: std::move(err))); |
| 1156 | }, |
| 1157 | args: [&](Error warning) { |
| 1158 | warn(msg: toString(f: this) + ": "+ toString(E: std::move(warning))); |
| 1159 | }); |
| 1160 | |
| 1161 | // TODO: Since object files can contain a lot of DWARF info, we should verify |
| 1162 | // that we are parsing just the info we need |
| 1163 | const DWARFContext::compile_unit_range &units = ctx->compile_units(); |
| 1164 | // FIXME: There can be more than one compile unit per object file. See |
| 1165 | // PR48637. |
| 1166 | auto it = units.begin(); |
| 1167 | compileUnit = it != units.end() ? it->get() : nullptr; |
| 1168 | } |
| 1169 | |
| 1170 | template <class LP> void ObjFile::parseDeferredRelocationsImpl() { |
| 1171 | using Header = typename LP::mach_header; |
| 1172 | using SegmentCommand = typename LP::segment_command; |
| 1173 | using SectionHeader = typename LP::section; |
| 1174 | |
| 1175 | auto *hdr = reinterpret_cast<const Header *>(mb.getBufferStart()); |
| 1176 | const load_command *cmd = findCommand(hdr, LP::segmentLCType); |
| 1177 | if (!cmd) |
| 1178 | return; |
| 1179 | auto *c = reinterpret_cast<const SegmentCommand *>(cmd); |
| 1180 | ArrayRef<SectionHeader> sectionHeaders{ |
| 1181 | reinterpret_cast<const SectionHeader *>(c + 1), c->nsects}; |
| 1182 | |
| 1183 | // Mirroring section filter in parse() since we already parsed unwind |
| 1184 | // sections. |
| 1185 | for (size_t i = 0, n = sections.size(); i < n; ++i) |
| 1186 | if (!sections[i]->subsections.empty() && !isUnwindSection(sec: *sections[i])) |
| 1187 | parseRelocations(sectionHeaders, sectionHeaders[i], *sections[i]); |
| 1188 | } |
| 1189 | |
| 1190 | void ObjFile::parseDeferredRelocations() { |
| 1191 | if (target->wordSize == 8) |
| 1192 | parseDeferredRelocationsImpl<LP64>(); |
| 1193 | else |
| 1194 | parseDeferredRelocationsImpl<ILP32>(); |
| 1195 | } |
| 1196 | |
| 1197 | void macho::parseDeferredRelocations() { |
| 1198 | TimeTraceScope timeScope("Parse relocations"); |
| 1199 | parallelForEach(R&: inputFiles, Fn: [](InputFile *file) { |
| 1200 | if (auto *objFile = dyn_cast<ObjFile>(Val: file)) |
| 1201 | objFile->parseDeferredRelocations(); |
| 1202 | }); |
| 1203 | } |
| 1204 | |
| 1205 | ArrayRef<data_in_code_entry> ObjFile::getDataInCode() const { |
| 1206 | const auto *buf = reinterpret_cast<const uint8_t *>(mb.getBufferStart()); |
| 1207 | const load_command *cmd = findCommand(anyHdr: buf, types: LC_DATA_IN_CODE); |
| 1208 | if (!cmd) |
| 1209 | return {}; |
| 1210 | const auto *c = reinterpret_cast<const linkedit_data_command *>(cmd); |
| 1211 | return {reinterpret_cast<const data_in_code_entry *>(buf + c->dataoff), |
| 1212 | c->datasize / sizeof(data_in_code_entry)}; |
| 1213 | } |
| 1214 | |
| 1215 | ArrayRef<uint8_t> ObjFile::getOptimizationHints() const { |
| 1216 | const auto *buf = reinterpret_cast<const uint8_t *>(mb.getBufferStart()); |
| 1217 | if (auto *cmd = |
| 1218 | findCommand<linkedit_data_command>(anyHdr: buf, types: LC_LINKER_OPTIMIZATION_HINT)) |
| 1219 | return {buf + cmd->dataoff, cmd->datasize}; |
| 1220 | return {}; |
| 1221 | } |
| 1222 | |
| 1223 | // Create pointers from symbols to their associated compact unwind entries. |
| 1224 | void ObjFile::registerCompactUnwind(Section &compactUnwindSection) { |
| 1225 | for (const Subsection &subsection : compactUnwindSection.subsections) { |
| 1226 | ConcatInputSection *isec = cast<ConcatInputSection>(Val: subsection.isec); |
| 1227 | // Hack!! Each compact unwind entry (CUE) has its UNSIGNED relocations embed |
| 1228 | // their addends in its data. Thus if ICF operated naively and compared the |
| 1229 | // entire contents of each CUE, entries with identical unwind info but e.g. |
| 1230 | // belonging to different functions would never be considered equivalent. To |
| 1231 | // work around this problem, we remove some parts of the data containing the |
| 1232 | // embedded addends. In particular, we remove the function address and LSDA |
| 1233 | // pointers. Since these locations are at the start and end of the entry, |
| 1234 | // we can do this using a simple, efficient slice rather than performing a |
| 1235 | // copy. We are not losing any information here because the embedded |
| 1236 | // addends have already been parsed in the corresponding Reloc structs. |
| 1237 | // |
| 1238 | // Removing these pointers would not be safe if they were pointers to |
| 1239 | // absolute symbols. In that case, there would be no corresponding |
| 1240 | // relocation. However, (AFAIK) MC cannot emit references to absolute |
| 1241 | // symbols for either the function address or the LSDA. However, it *can* do |
| 1242 | // so for the personality pointer, so we are not slicing that field away. |
| 1243 | // |
| 1244 | // Note that we do not adjust the offsets of the corresponding relocations; |
| 1245 | // instead, we rely on `relocateCompactUnwind()` to correctly handle these |
| 1246 | // truncated input sections. |
| 1247 | isec->data = isec->data.slice(N: target->wordSize, M: 8 + target->wordSize); |
| 1248 | uint32_t encoding = read32le(P: isec->data.data() + sizeof(uint32_t)); |
| 1249 | // llvm-mc omits CU entries for functions that need DWARF encoding, but |
| 1250 | // `ld -r` doesn't. We can ignore them because we will re-synthesize these |
| 1251 | // CU entries from the DWARF info during the output phase. |
| 1252 | if ((encoding & static_cast<uint32_t>(UNWIND_MODE_MASK)) == |
| 1253 | target->modeDwarfEncoding) |
| 1254 | continue; |
| 1255 | |
| 1256 | ConcatInputSection *referentIsec; |
| 1257 | for (auto it = isec->relocs.begin(); it != isec->relocs.end();) { |
| 1258 | Relocation &r = *it; |
| 1259 | // CUE::functionAddress is at offset 0. Skip personality & LSDA relocs. |
| 1260 | if (r.offset != 0) { |
| 1261 | ++it; |
| 1262 | continue; |
| 1263 | } |
| 1264 | uint64_t add = r.addend; |
| 1265 | if (auto *sym = cast_or_null<Defined>(Val: dyn_cast<Symbol *>(Val&: r.referent))) { |
| 1266 | // Check whether the symbol defined in this file is the prevailing one. |
| 1267 | // Skip if it is e.g. a weak def that didn't prevail. |
| 1268 | if (sym->getFile() != this) { |
| 1269 | ++it; |
| 1270 | continue; |
| 1271 | } |
| 1272 | add += sym->value; |
| 1273 | referentIsec = cast<ConcatInputSection>(Val: sym->isec()); |
| 1274 | } else { |
| 1275 | referentIsec = |
| 1276 | cast<ConcatInputSection>(Val: dyn_cast<InputSection *>(Val&: r.referent)); |
| 1277 | } |
| 1278 | // Unwind info lives in __DATA, and finalization of __TEXT will occur |
| 1279 | // before finalization of __DATA. Moreover, the finalization of unwind |
| 1280 | // info depends on the exact addresses that it references. So it is safe |
| 1281 | // for compact unwind to reference addresses in __TEXT, but not addresses |
| 1282 | // in any other segment. |
| 1283 | if (referentIsec->getSegName() != segment_names::text) |
| 1284 | error(msg: isec->getLocation(off: r.offset) + " references section "+ |
| 1285 | referentIsec->getName() + " which is not in segment __TEXT"); |
| 1286 | // The functionAddress relocations are typically section relocations. |
| 1287 | // However, unwind info operates on a per-symbol basis, so we search for |
| 1288 | // the function symbol here. |
| 1289 | Defined *d = tryFindSymbolAtOffset(isec: referentIsec, off: add); |
| 1290 | if (!d) { |
| 1291 | // If there's no symbol at the function address (e.g. for temporary |
| 1292 | // local labels that are not in the symtab), synthesize a local one so |
| 1293 | // we still emit correct unwind info. |
| 1294 | |
| 1295 | // Avoid creating symbols for coalesced sections; those functions were |
| 1296 | // folded away. |
| 1297 | if (referentIsec->wasCoalesced) { |
| 1298 | ++it; |
| 1299 | continue; |
| 1300 | } |
| 1301 | |
| 1302 | d = make<Defined>(args: saver().save(S: Twine("Lcu.") + referentIsec->getName() + |
| 1303 | "."+ Twine::utohexstr(Val: add)), |
| 1304 | args: this, args&: referentIsec, args&: add, |
| 1305 | /*size=*/args: 0, /*isWeakDef=*/args: false, |
| 1306 | /*isExternal=*/args: false, /*isPrivateExtern=*/args: false, |
| 1307 | /*includeInSymtab=*/args: false, |
| 1308 | /*isReferencedDynamically=*/args: false, |
| 1309 | /*noDeadStrip=*/args: false); |
| 1310 | // Also add to the file-level symbol list so that scanSymbols() in |
| 1311 | // Writer picks it up and registers it with UnwindInfoSection. |
| 1312 | symbols.push_back(x: d); |
| 1313 | } |
| 1314 | d->originalUnwindEntry = isec; |
| 1315 | // Now that the symbol points to the unwind entry, we can remove the reloc |
| 1316 | // that points from the unwind entry back to the symbol. |
| 1317 | // |
| 1318 | // First, the symbol keeps the unwind entry alive (and not vice versa), so |
| 1319 | // this keeps dead-stripping simple. |
| 1320 | // |
| 1321 | // Moreover, it reduces the work that ICF needs to do to figure out if |
| 1322 | // functions with unwind info are foldable. |
| 1323 | // |
| 1324 | // However, this does make it possible for ICF to fold CUEs that point to |
| 1325 | // distinct functions (if the CUEs are otherwise identical). |
| 1326 | // UnwindInfoSection takes care of this by re-duplicating the CUEs so that |
| 1327 | // each one can hold a distinct functionAddress value. |
| 1328 | // |
| 1329 | // Given that clang emits relocations in reverse order of address, this |
| 1330 | // relocation should be at the end of the vector for most of our input |
| 1331 | // object files, so this erase() is typically an O(1) operation. |
| 1332 | it = isec->relocs.erase(position: it); |
| 1333 | } |
| 1334 | } |
| 1335 | } |
| 1336 | |
| 1337 | struct CIE { |
| 1338 | macho::Symbol *personalitySymbol = nullptr; |
| 1339 | bool fdesHaveAug = false; |
| 1340 | uint8_t lsdaPtrSize = 0; // 0 => no LSDA |
| 1341 | uint8_t funcPtrSize = 0; |
| 1342 | }; |
| 1343 | |
| 1344 | static uint8_t pointerEncodingToSize(uint8_t enc) { |
| 1345 | switch (enc & 0xf) { |
| 1346 | case dwarf::DW_EH_PE_absptr: |
| 1347 | return target->wordSize; |
| 1348 | case dwarf::DW_EH_PE_sdata4: |
| 1349 | return 4; |
| 1350 | case dwarf::DW_EH_PE_sdata8: |
| 1351 | // ld64 doesn't actually support sdata8, but this seems simple enough... |
| 1352 | return 8; |
| 1353 | default: |
| 1354 | return 0; |
| 1355 | }; |
| 1356 | } |
| 1357 | |
| 1358 | static CIE parseCIE(const InputSection *isec, const EhReader &reader, |
| 1359 | size_t off) { |
| 1360 | // Handling the full generality of possible DWARF encodings would be a major |
| 1361 | // pain. We instead take advantage of our knowledge of how llvm-mc encodes |
| 1362 | // DWARF and handle just that. |
| 1363 | constexpr uint8_t expectedPersonalityEnc = |
| 1364 | dwarf::DW_EH_PE_pcrel | dwarf::DW_EH_PE_indirect | dwarf::DW_EH_PE_sdata4; |
| 1365 | |
| 1366 | CIE cie; |
| 1367 | uint8_t version = reader.readByte(off: &off); |
| 1368 | if (version != 1 && version != 3) |
| 1369 | fatal(msg: "Expected CIE version of 1 or 3, got "+ Twine(version)); |
| 1370 | StringRef aug = reader.readString(off: &off); |
| 1371 | reader.skipLeb128(off: &off); // skip code alignment |
| 1372 | reader.skipLeb128(off: &off); // skip data alignment |
| 1373 | reader.skipLeb128(off: &off); // skip return address register |
| 1374 | reader.skipLeb128(off: &off); // skip aug data length |
| 1375 | uint64_t personalityAddrOff = 0; |
| 1376 | for (char c : aug) { |
| 1377 | switch (c) { |
| 1378 | case 'z': |
| 1379 | cie.fdesHaveAug = true; |
| 1380 | break; |
| 1381 | case 'P': { |
| 1382 | uint8_t personalityEnc = reader.readByte(off: &off); |
| 1383 | if (personalityEnc != expectedPersonalityEnc) |
| 1384 | reader.failOn(errOff: off, msg: "unexpected personality encoding 0x"+ |
| 1385 | Twine::utohexstr(Val: personalityEnc)); |
| 1386 | personalityAddrOff = off; |
| 1387 | off += 4; |
| 1388 | break; |
| 1389 | } |
| 1390 | case 'L': { |
| 1391 | uint8_t lsdaEnc = reader.readByte(off: &off); |
| 1392 | cie.lsdaPtrSize = pointerEncodingToSize(enc: lsdaEnc); |
| 1393 | if (cie.lsdaPtrSize == 0) |
| 1394 | reader.failOn(errOff: off, msg: "unexpected LSDA encoding 0x"+ |
| 1395 | Twine::utohexstr(Val: lsdaEnc)); |
| 1396 | break; |
| 1397 | } |
| 1398 | case 'R': { |
| 1399 | uint8_t pointerEnc = reader.readByte(off: &off); |
| 1400 | cie.funcPtrSize = pointerEncodingToSize(enc: pointerEnc); |
| 1401 | if (cie.funcPtrSize == 0 || !(pointerEnc & dwarf::DW_EH_PE_pcrel)) |
| 1402 | reader.failOn(errOff: off, msg: "unexpected pointer encoding 0x"+ |
| 1403 | Twine::utohexstr(Val: pointerEnc)); |
| 1404 | break; |
| 1405 | } |
| 1406 | default: |
| 1407 | break; |
| 1408 | } |
| 1409 | } |
| 1410 | if (personalityAddrOff != 0) { |
| 1411 | const auto *personalityReloc = isec->getRelocAt(off: personalityAddrOff); |
| 1412 | if (!personalityReloc) |
| 1413 | reader.failOn(errOff: off, msg: "Failed to locate relocation for personality symbol"); |
| 1414 | cie.personalitySymbol = cast<macho::Symbol *>(Val: personalityReloc->referent); |
| 1415 | } |
| 1416 | return cie; |
| 1417 | } |
| 1418 | |
| 1419 | // EH frame target addresses may be encoded as pcrel offsets. However, instead |
| 1420 | // of using an actual pcrel reloc, ld64 emits subtractor relocations instead. |
| 1421 | // This function recovers the target address from the subtractors, essentially |
| 1422 | // performing the inverse operation of EhRelocator. |
| 1423 | // |
| 1424 | // Concretely, we expect our relocations to write the value of `PC - |
| 1425 | // target_addr` to `PC`. `PC` itself is denoted by a minuend relocation that |
| 1426 | // points to a symbol plus an addend. |
| 1427 | // |
| 1428 | // It is important that the minuend relocation point to a symbol within the |
| 1429 | // same section as the fixup value, since sections may get moved around. |
| 1430 | // |
| 1431 | // For example, for arm64, llvm-mc emits relocations for the target function |
| 1432 | // address like so: |
| 1433 | // |
| 1434 | // ltmp: |
| 1435 | // <CIE start> |
| 1436 | // ... |
| 1437 | // <CIE end> |
| 1438 | // ... multiple FDEs ... |
| 1439 | // <FDE start> |
| 1440 | // <target function address - (ltmp + pcrel offset)> |
| 1441 | // ... |
| 1442 | // |
| 1443 | // If any of the FDEs in `multiple FDEs` get dead-stripped, then `FDE start` |
| 1444 | // will move to an earlier address, and `ltmp + pcrel offset` will no longer |
| 1445 | // reflect an accurate pcrel value. To avoid this problem, we "canonicalize" |
| 1446 | // our relocation by adding an `EH_Frame` symbol at `FDE start`, and updating |
| 1447 | // the reloc to be `target function address - (EH_Frame + new pcrel offset)`. |
| 1448 | // |
| 1449 | // If `Invert` is set, then we instead expect `target_addr - PC` to be written |
| 1450 | // to `PC`. |
| 1451 | template <bool Invert = false> |
| 1452 | Defined * |
| 1453 | targetSymFromCanonicalSubtractor(const InputSection *isec, |
| 1454 | std::vector<Relocation>::iterator relocIt) { |
| 1455 | Relocation &subtrahend = *relocIt; |
| 1456 | Relocation &minuend = *std::next(x: relocIt); |
| 1457 | assert(target->hasAttr(subtrahend.type, RelocAttrBits::SUBTRAHEND)); |
| 1458 | assert(target->hasAttr(minuend.type, RelocAttrBits::UNSIGNED)); |
| 1459 | // Note: pcSym may *not* be exactly at the PC; there's usually a non-zero |
| 1460 | // addend. |
| 1461 | auto *pcSym = cast<Defined>(Val: cast<macho::Symbol *>(Val&: subtrahend.referent)); |
| 1462 | Defined *target = |
| 1463 | cast_or_null<Defined>(Val: minuend.referent.dyn_cast<macho::Symbol *>()); |
| 1464 | if (!pcSym) { |
| 1465 | auto *targetIsec = |
| 1466 | cast<ConcatInputSection>(Val: cast<InputSection *>(Val&: minuend.referent)); |
| 1467 | target = findSymbolAtOffset(isec: targetIsec, off: minuend.addend); |
| 1468 | } |
| 1469 | if (Invert) |
| 1470 | std::swap(a&: pcSym, b&: target); |
| 1471 | if (pcSym->isec() == isec) { |
| 1472 | if (pcSym->value - (Invert ? -1 : 1) * minuend.addend != subtrahend.offset) |
| 1473 | fatal(msg: "invalid FDE relocation in __eh_frame"); |
| 1474 | } else { |
| 1475 | // Ensure the pcReloc points to a symbol within the current EH frame. |
| 1476 | // HACK: we should really verify that the original relocation's semantics |
| 1477 | // are preserved. In particular, we should have |
| 1478 | // `oldSym->value + oldOffset == newSym + newOffset`. However, we don't |
| 1479 | // have an easy way to access the offsets from this point in the code; some |
| 1480 | // refactoring is needed for that. |
| 1481 | Relocation &pcReloc = Invert ? minuend : subtrahend; |
| 1482 | pcReloc.referent = isec->symbols[0]; |
| 1483 | assert(isec->symbols[0]->value == 0); |
| 1484 | minuend.addend = pcReloc.offset * (Invert ? 1LL : -1LL); |
| 1485 | } |
| 1486 | return target; |
| 1487 | } |
| 1488 | |
| 1489 | Defined *findSymbolAtAddress(const std::vector<Section *> §ions, |
| 1490 | uint64_t addr) { |
| 1491 | Section *sec = findContainingSection(sections, offset: &addr); |
| 1492 | auto *isec = cast<ConcatInputSection>(Val: findContainingSubsection(section: *sec, offset: &addr)); |
| 1493 | return findSymbolAtOffset(isec, off: addr); |
| 1494 | } |
| 1495 | |
| 1496 | // For symbols that don't have compact unwind info, associate them with the more |
| 1497 | // general-purpose (and verbose) DWARF unwind info found in __eh_frame. |
| 1498 | // |
| 1499 | // This requires us to parse the contents of __eh_frame. See EhFrame.h for a |
| 1500 | // description of its format. |
| 1501 | // |
| 1502 | // While parsing, we also look for what MC calls "abs-ified" relocations -- they |
| 1503 | // are relocations which are implicitly encoded as offsets in the section data. |
| 1504 | // We convert them into explicit Reloc structs so that the EH frames can be |
| 1505 | // handled just like a regular ConcatInputSection later in our output phase. |
| 1506 | // |
| 1507 | // We also need to handle the case where our input object file has explicit |
| 1508 | // relocations. This is the case when e.g. it's the output of `ld -r`. We only |
| 1509 | // look for the "abs-ified" relocation if an explicit relocation is absent. |
| 1510 | void ObjFile::registerEhFrames(Section &ehFrameSection) { |
| 1511 | DenseMap<const InputSection *, CIE> cieMap; |
| 1512 | for (const Subsection &subsec : ehFrameSection.subsections) { |
| 1513 | auto *isec = cast<ConcatInputSection>(Val: subsec.isec); |
| 1514 | uint64_t isecOff = subsec.offset; |
| 1515 | |
| 1516 | // Subtractor relocs require the subtrahend to be a symbol reloc. Ensure |
| 1517 | // that all EH frames have an associated symbol so that we can generate |
| 1518 | // subtractor relocs that reference them. |
| 1519 | if (isec->symbols.size() == 0) |
| 1520 | make<Defined>(args: "EH_Frame", args: isec->getFile(), args&: isec, /*value=*/args: 0, |
| 1521 | args: isec->getSize(), /*isWeakDef=*/args: false, /*isExternal=*/args: false, |
| 1522 | /*isPrivateExtern=*/args: false, /*includeInSymtab=*/args: false, |
| 1523 | /*isReferencedDynamically=*/args: false, |
| 1524 | /*noDeadStrip=*/args: false); |
| 1525 | else if (isec->symbols[0]->value != 0) |
| 1526 | fatal(msg: "found symbol at unexpected offset in __eh_frame"); |
| 1527 | |
| 1528 | EhReader reader(this, isec->data, subsec.offset); |
| 1529 | size_t dataOff = 0; // Offset from the start of the EH frame. |
| 1530 | reader.skipValidLength(off: &dataOff); // readLength() already validated this. |
| 1531 | // cieOffOff is the offset from the start of the EH frame to the cieOff |
| 1532 | // value, which is itself an offset from the current PC to a CIE. |
| 1533 | const size_t cieOffOff = dataOff; |
| 1534 | |
| 1535 | EhRelocator ehRelocator(isec); |
| 1536 | auto cieOffRelocIt = llvm::find_if(Range&: isec->relocs, P: [=](const Relocation &r) { |
| 1537 | return r.offset == cieOffOff; |
| 1538 | }); |
| 1539 | InputSection *cieIsec = nullptr; |
| 1540 | if (cieOffRelocIt != isec->relocs.end()) { |
| 1541 | // We already have an explicit relocation for the CIE offset. |
| 1542 | cieIsec = |
| 1543 | targetSymFromCanonicalSubtractor</*Invert=*/true>(isec, relocIt: cieOffRelocIt) |
| 1544 | ->isec(); |
| 1545 | dataOff += sizeof(uint32_t); |
| 1546 | } else { |
| 1547 | // If we haven't found a relocation, then the CIE offset is most likely |
| 1548 | // embedded in the section data (AKA an "abs-ified" reloc.). Parse that |
| 1549 | // and generate a Reloc struct. |
| 1550 | uint32_t cieMinuend = reader.readU32(off: &dataOff); |
| 1551 | if (cieMinuend == 0) { |
| 1552 | cieIsec = isec; |
| 1553 | } else { |
| 1554 | uint32_t cieOff = isecOff + dataOff - cieMinuend; |
| 1555 | cieIsec = findContainingSubsection(section: ehFrameSection, offset: &cieOff); |
| 1556 | if (cieIsec == nullptr) |
| 1557 | fatal(msg: "failed to find CIE"); |
| 1558 | } |
| 1559 | if (cieIsec != isec) |
| 1560 | ehRelocator.makeNegativePcRel(off: cieOffOff, target: cieIsec->symbols[0], |
| 1561 | /*length=*/2); |
| 1562 | } |
| 1563 | if (cieIsec == isec) { |
| 1564 | cieMap[cieIsec] = parseCIE(isec, reader, off: dataOff); |
| 1565 | continue; |
| 1566 | } |
| 1567 | |
| 1568 | assert(cieMap.contains(cieIsec)); |
| 1569 | const CIE &cie = cieMap[cieIsec]; |
| 1570 | // Offset of the function address within the EH frame. |
| 1571 | const size_t funcAddrOff = dataOff; |
| 1572 | uint64_t funcAddr = reader.readPointer(off: &dataOff, size: cie.funcPtrSize) + |
| 1573 | ehFrameSection.addr + isecOff + funcAddrOff; |
| 1574 | uint32_t funcLength = reader.readPointer(off: &dataOff, size: cie.funcPtrSize); |
| 1575 | size_t lsdaAddrOff = 0; // Offset of the LSDA address within the EH frame. |
| 1576 | std::optional<uint64_t> lsdaAddrOpt; |
| 1577 | if (cie.fdesHaveAug) { |
| 1578 | reader.skipLeb128(off: &dataOff); |
| 1579 | lsdaAddrOff = dataOff; |
| 1580 | if (cie.lsdaPtrSize != 0) { |
| 1581 | uint64_t lsdaOff = reader.readPointer(off: &dataOff, size: cie.lsdaPtrSize); |
| 1582 | if (lsdaOff != 0) // FIXME possible to test this? |
| 1583 | lsdaAddrOpt = ehFrameSection.addr + isecOff + lsdaAddrOff + lsdaOff; |
| 1584 | } |
| 1585 | } |
| 1586 | |
| 1587 | auto funcAddrRelocIt = isec->relocs.end(); |
| 1588 | auto lsdaAddrRelocIt = isec->relocs.end(); |
| 1589 | for (auto it = isec->relocs.begin(); it != isec->relocs.end(); ++it) { |
| 1590 | if (it->offset == funcAddrOff) |
| 1591 | funcAddrRelocIt = it++; // Found subtrahend; skip over minuend reloc |
| 1592 | else if (lsdaAddrOpt && it->offset == lsdaAddrOff) |
| 1593 | lsdaAddrRelocIt = it++; // Found subtrahend; skip over minuend reloc |
| 1594 | } |
| 1595 | |
| 1596 | Defined *funcSym; |
| 1597 | if (funcAddrRelocIt != isec->relocs.end()) { |
| 1598 | funcSym = targetSymFromCanonicalSubtractor(isec, relocIt: funcAddrRelocIt); |
| 1599 | // Canonicalize the symbol. If there are multiple symbols at the same |
| 1600 | // address, we want both `registerEhFrame` and `registerCompactUnwind` |
| 1601 | // to register the unwind entry under same symbol. |
| 1602 | // This is not particularly efficient, but we should run into this case |
| 1603 | // infrequently (only when handling the output of `ld -r`). |
| 1604 | if (funcSym->isec()) |
| 1605 | funcSym = findSymbolAtOffset(isec: cast<ConcatInputSection>(Val: funcSym->isec()), |
| 1606 | off: funcSym->value); |
| 1607 | } else { |
| 1608 | funcSym = findSymbolAtAddress(sections, addr: funcAddr); |
| 1609 | ehRelocator.makePcRel(off: funcAddrOff, target: funcSym, length: target->p2WordSize); |
| 1610 | } |
| 1611 | // The symbol has been coalesced, or already has a compact unwind entry. |
| 1612 | if (!funcSym || funcSym->getFile() != this || funcSym->unwindEntry()) { |
| 1613 | // We must prune unused FDEs for correctness, so we cannot rely on |
| 1614 | // -dead_strip being enabled. |
| 1615 | isec->live = false; |
| 1616 | continue; |
| 1617 | } |
| 1618 | |
| 1619 | InputSection *lsdaIsec = nullptr; |
| 1620 | if (lsdaAddrRelocIt != isec->relocs.end()) { |
| 1621 | lsdaIsec = |
| 1622 | targetSymFromCanonicalSubtractor(isec, relocIt: lsdaAddrRelocIt)->isec(); |
| 1623 | } else if (lsdaAddrOpt) { |
| 1624 | uint64_t lsdaAddr = *lsdaAddrOpt; |
| 1625 | Section *sec = findContainingSection(sections, offset: &lsdaAddr); |
| 1626 | lsdaIsec = |
| 1627 | cast<ConcatInputSection>(Val: findContainingSubsection(section: *sec, offset: &lsdaAddr)); |
| 1628 | ehRelocator.makePcRel(off: lsdaAddrOff, target: lsdaIsec, length: target->p2WordSize); |
| 1629 | } |
| 1630 | |
| 1631 | fdes[isec] = {.funcLength: funcLength, .personality: cie.personalitySymbol, .lsda: lsdaIsec}; |
| 1632 | funcSym->originalUnwindEntry = isec; |
| 1633 | ehRelocator.commit(); |
| 1634 | } |
| 1635 | |
| 1636 | // __eh_frame is marked as S_ATTR_LIVE_SUPPORT in input files, because FDEs |
| 1637 | // are normally required to be kept alive if they reference a live symbol. |
| 1638 | // However, we've explicitly created a dependency from a symbol to its FDE, so |
| 1639 | // dead-stripping will just work as usual, and S_ATTR_LIVE_SUPPORT will only |
| 1640 | // serve to incorrectly prevent us from dead-stripping duplicate FDEs for a |
| 1641 | // live symbol (e.g. if there were multiple weak copies). Remove this flag to |
| 1642 | // let dead-stripping proceed correctly. |
| 1643 | ehFrameSection.flags &= ~S_ATTR_LIVE_SUPPORT; |
| 1644 | } |
| 1645 | |
| 1646 | std::string ObjFile::sourceFile() const { |
| 1647 | const char *unitName = compileUnit->getUnitDIE().getShortName(); |
| 1648 | // DWARF allows DW_AT_name to be absolute, in which case nothing should be |
| 1649 | // prepended. As for the styles, debug info can contain paths from any OS, not |
| 1650 | // necessarily an OS we're currently running on. Moreover different |
| 1651 | // compilation units can be compiled on different operating systems and linked |
| 1652 | // together later. |
| 1653 | if (sys::path::is_absolute(path: unitName, style: llvm::sys::path::Style::posix) || |
| 1654 | sys::path::is_absolute(path: unitName, style: llvm::sys::path::Style::windows)) |
| 1655 | return unitName; |
| 1656 | SmallString<261> dir(compileUnit->getCompilationDir()); |
| 1657 | StringRef sep = sys::path::get_separator(); |
| 1658 | // We don't use `path::append` here because we want an empty `dir` to result |
| 1659 | // in an absolute path. `append` would give us a relative path for that case. |
| 1660 | if (!dir.ends_with(Suffix: sep)) |
| 1661 | dir += sep; |
| 1662 | return (dir + unitName).str(); |
| 1663 | } |
| 1664 | |
| 1665 | lld::DWARFCache *ObjFile::getDwarf() { |
| 1666 | llvm::call_once(flag&: initDwarf, F: [this]() { |
| 1667 | auto dwObj = DwarfObject::create(this); |
| 1668 | if (!dwObj) |
| 1669 | return; |
| 1670 | dwarfCache = std::make_unique<DWARFCache>(args: std::make_unique<DWARFContext>( |
| 1671 | args: std::move(dwObj), args: "", |
| 1672 | args: [&](Error err) { warn(msg: getName() + ": "+ toString(E: std::move(err))); }, |
| 1673 | args: [&](Error warning) { |
| 1674 | warn(msg: getName() + ": "+ toString(E: std::move(warning))); |
| 1675 | })); |
| 1676 | }); |
| 1677 | |
| 1678 | return dwarfCache.get(); |
| 1679 | } |
| 1680 | // The path can point to either a dylib or a .tbd file. |
| 1681 | static DylibFile *loadDylib(StringRef path, DylibFile *umbrella) { |
| 1682 | std::optional<MemoryBufferRef> mbref = readFile(path); |
| 1683 | if (!mbref) { |
| 1684 | error(msg: "could not read dylib file at "+ path); |
| 1685 | return nullptr; |
| 1686 | } |
| 1687 | return loadDylib(mbref: *mbref, umbrella); |
| 1688 | } |
| 1689 | |
| 1690 | // TBD files are parsed into a series of TAPI documents (InterfaceFiles), with |
| 1691 | // the first document storing child pointers to the rest of them. When we are |
| 1692 | // processing a given TBD file, we store that top-level document in |
| 1693 | // currentTopLevelTapi. When processing re-exports, we search its children for |
| 1694 | // potentially matching documents in the same TBD file. Note that the children |
| 1695 | // themselves don't point to further documents, i.e. this is a two-level tree. |
| 1696 | // |
| 1697 | // Re-exports can either refer to on-disk files, or to documents within .tbd |
| 1698 | // files. |
| 1699 | static DylibFile *findDylib(StringRef path, DylibFile *umbrella, |
| 1700 | const InterfaceFile *currentTopLevelTapi) { |
| 1701 | // Search order: |
| 1702 | // 1. Install name basename in -F / -L directories. |
| 1703 | { |
| 1704 | // Framework names can be in multiple formats: |
| 1705 | // - Foo.framework/Foo |
| 1706 | // - Foo.framework/Versions/A/Foo |
| 1707 | StringRef stem = path::stem(path); |
| 1708 | SmallString<128> frameworkName("/"); |
| 1709 | frameworkName += stem; |
| 1710 | frameworkName += ".framework/"; |
| 1711 | size_t i = path.rfind(Str: frameworkName); |
| 1712 | if (i != StringRef::npos) { |
| 1713 | StringRef frameworkPath = path.substr(Start: i + 1); |
| 1714 | for (StringRef dir : config->frameworkSearchPaths) { |
| 1715 | SmallString<128> candidate = dir; |
| 1716 | path::append(path&: candidate, a: frameworkPath); |
| 1717 | if (std::optional<StringRef> dylibPath = |
| 1718 | resolveDylibPath(path: candidate.str())) |
| 1719 | return loadDylib(path: *dylibPath, umbrella); |
| 1720 | } |
| 1721 | } else if (std::optional<StringRef> dylibPath = findPathCombination( |
| 1722 | name: stem, roots: config->librarySearchPaths, extensions: {".tbd", ".dylib", ".so"})) |
| 1723 | return loadDylib(path: *dylibPath, umbrella); |
| 1724 | } |
| 1725 | |
| 1726 | // 2. As absolute path. |
| 1727 | if (path::is_absolute(path, style: path::Style::posix)) |
| 1728 | for (StringRef root : config->systemLibraryRoots) |
| 1729 | if (std::optional<StringRef> dylibPath = |
| 1730 | resolveDylibPath(path: (root + path).str())) |
| 1731 | return loadDylib(path: *dylibPath, umbrella); |
| 1732 | |
| 1733 | // 3. As relative path. |
| 1734 | |
| 1735 | // TODO: Handle -dylib_file |
| 1736 | |
| 1737 | // Replace @executable_path, @loader_path, @rpath prefixes in install name. |
| 1738 | SmallString<128> newPath; |
| 1739 | if (config->outputType == MH_EXECUTE && |
| 1740 | path.consume_front(Prefix: "@executable_path/")) { |
| 1741 | // ld64 allows overriding this with the undocumented flag -executable_path. |
| 1742 | // lld doesn't currently implement that flag. |
| 1743 | // FIXME: Consider using finalOutput instead of outputFile. |
| 1744 | path::append(path&: newPath, a: path::parent_path(path: config->outputFile), b: path); |
| 1745 | path = newPath; |
| 1746 | } else if (path.consume_front(Prefix: "@loader_path/")) { |
| 1747 | fs::real_path(path: umbrella->getName(), output&: newPath); |
| 1748 | path::remove_filename(path&: newPath); |
| 1749 | path::append(path&: newPath, a: path); |
| 1750 | path = newPath; |
| 1751 | } else if (path.starts_with(Prefix: "@rpath/")) { |
| 1752 | for (StringRef rpath : umbrella->rpaths) { |
| 1753 | newPath.clear(); |
| 1754 | if (rpath.consume_front(Prefix: "@loader_path/")) { |
| 1755 | fs::real_path(path: umbrella->getName(), output&: newPath); |
| 1756 | path::remove_filename(path&: newPath); |
| 1757 | } |
| 1758 | path::append(path&: newPath, a: rpath, b: path.drop_front(N: strlen(s: "@rpath/"))); |
| 1759 | if (std::optional<StringRef> dylibPath = resolveDylibPath(path: newPath.str())) |
| 1760 | return loadDylib(path: *dylibPath, umbrella); |
| 1761 | } |
| 1762 | // If not found in umbrella, try the rpaths specified via -rpath too. |
| 1763 | for (StringRef rpath : config->runtimePaths) { |
| 1764 | newPath.clear(); |
| 1765 | if (rpath.consume_front(Prefix: "@loader_path/")) { |
| 1766 | fs::real_path(path: umbrella->getName(), output&: newPath); |
| 1767 | path::remove_filename(path&: newPath); |
| 1768 | } |
| 1769 | path::append(path&: newPath, a: rpath, b: path.drop_front(N: strlen(s: "@rpath/"))); |
| 1770 | if (std::optional<StringRef> dylibPath = resolveDylibPath(path: newPath.str())) |
| 1771 | return loadDylib(path: *dylibPath, umbrella); |
| 1772 | } |
| 1773 | } |
| 1774 | |
| 1775 | // FIXME: Should this be further up? |
| 1776 | if (currentTopLevelTapi) { |
| 1777 | for (InterfaceFile &child : |
| 1778 | make_pointee_range(Range: currentTopLevelTapi->documents())) { |
| 1779 | assert(child.documents().empty()); |
| 1780 | if (path == child.getInstallName()) { |
| 1781 | auto *file = make<DylibFile>(args&: child, args&: umbrella, /*isBundleLoader=*/args: false, |
| 1782 | /*explicitlyLinked=*/args: false); |
| 1783 | file->parseReexports(interface: child); |
| 1784 | return file; |
| 1785 | } |
| 1786 | } |
| 1787 | } |
| 1788 | |
| 1789 | if (std::optional<StringRef> dylibPath = resolveDylibPath(path)) |
| 1790 | return loadDylib(path: *dylibPath, umbrella); |
| 1791 | |
| 1792 | return nullptr; |
| 1793 | } |
| 1794 | |
| 1795 | // If a re-exported dylib is public (lives in /usr/lib or |
| 1796 | // /System/Library/Frameworks), then it is considered implicitly linked: we |
| 1797 | // should bind to its symbols directly instead of via the re-exporting umbrella |
| 1798 | // library. |
| 1799 | static bool isImplicitlyLinked(StringRef path) { |
| 1800 | if (!config->implicitDylibs) |
| 1801 | return false; |
| 1802 | |
| 1803 | if (path::parent_path(path) == "/usr/lib") |
| 1804 | return true; |
| 1805 | |
| 1806 | // Match /System/Library/Frameworks/$FOO.framework/**/$FOO |
| 1807 | if (path.consume_front(Prefix: "/System/Library/Frameworks/")) { |
| 1808 | StringRef frameworkName = path.take_until(F: [](char c) { return c == '.'; }); |
| 1809 | return path::filename(path) == frameworkName; |
| 1810 | } |
| 1811 | |
| 1812 | return false; |
| 1813 | } |
| 1814 | |
| 1815 | void DylibFile::loadReexport(StringRef path, DylibFile *umbrella, |
| 1816 | const InterfaceFile *currentTopLevelTapi) { |
| 1817 | DylibFile *reexport = findDylib(path, umbrella, currentTopLevelTapi); |
| 1818 | if (!reexport) { |
| 1819 | // If not found in umbrella, retry since some rpaths might have been |
| 1820 | // defined in "this" dylib (which contains the LC_REEXPORT_DYLIB cmd) and |
| 1821 | // not in the umbrella. |
| 1822 | DylibFile *reexport2 = findDylib(path, umbrella: this, currentTopLevelTapi); |
| 1823 | if (!reexport2) { |
| 1824 | error(msg: toString(f: this) + ": unable to locate re-export with install name "+ |
| 1825 | path); |
| 1826 | } |
| 1827 | } |
| 1828 | } |
| 1829 | |
| 1830 | DylibFile::DylibFile(MemoryBufferRef mb, DylibFile *umbrella, |
| 1831 | bool isBundleLoader, bool explicitlyLinked) |
| 1832 | : InputFile(DylibKind, mb), refState(RefState::Unreferenced), |
| 1833 | explicitlyLinked(explicitlyLinked), isBundleLoader(isBundleLoader) { |
| 1834 | assert(!isBundleLoader || !umbrella); |
| 1835 | if (umbrella == nullptr) |
| 1836 | umbrella = this; |
| 1837 | this->umbrella = umbrella; |
| 1838 | |
| 1839 | auto *hdr = reinterpret_cast<const mach_header *>(mb.getBufferStart()); |
| 1840 | |
| 1841 | // Initialize installName. |
| 1842 | if (const load_command *cmd = findCommand(anyHdr: hdr, types: LC_ID_DYLIB)) { |
| 1843 | auto *c = reinterpret_cast<const dylib_command *>(cmd); |
| 1844 | currentVersion = read32le(P: &c->dylib.current_version); |
| 1845 | compatibilityVersion = read32le(P: &c->dylib.compatibility_version); |
| 1846 | installName = |
| 1847 | reinterpret_cast<const char *>(cmd) + read32le(P: &c->dylib.name); |
| 1848 | } else if (!isBundleLoader) { |
| 1849 | // macho_executable and macho_bundle don't have LC_ID_DYLIB, |
| 1850 | // so it's OK. |
| 1851 | error(msg: toString(f: this) + ": dylib missing LC_ID_DYLIB load command"); |
| 1852 | return; |
| 1853 | } |
| 1854 | |
| 1855 | if (config->printEachFile) |
| 1856 | message(msg: toString(f: this)); |
| 1857 | inputFiles.insert(X: this); |
| 1858 | |
| 1859 | deadStrippable = hdr->flags & MH_DEAD_STRIPPABLE_DYLIB; |
| 1860 | |
| 1861 | if (!checkCompatibility(input: this)) |
| 1862 | return; |
| 1863 | |
| 1864 | checkAppExtensionSafety(dylibIsAppExtensionSafe: hdr->flags & MH_APP_EXTENSION_SAFE); |
| 1865 | |
| 1866 | for (auto *cmd : findCommands<rpath_command>(anyHdr: hdr, types: LC_RPATH)) { |
| 1867 | StringRef rpath{reinterpret_cast<const char *>(cmd) + cmd->path}; |
| 1868 | rpaths.push_back(Elt: rpath); |
| 1869 | } |
| 1870 | |
| 1871 | // Initialize symbols. |
| 1872 | bool canBeImplicitlyLinked = findCommand(anyHdr: hdr, types: LC_SUB_CLIENT) == nullptr; |
| 1873 | exportingFile = (canBeImplicitlyLinked && isImplicitlyLinked(path: installName)) |
| 1874 | ? this |
| 1875 | : this->umbrella; |
| 1876 | |
| 1877 | if (!canBeImplicitlyLinked) { |
| 1878 | for (auto *cmd : findCommands<sub_client_command>(anyHdr: hdr, types: LC_SUB_CLIENT)) { |
| 1879 | StringRef allowableClient{reinterpret_cast<const char *>(cmd) + |
| 1880 | cmd->client}; |
| 1881 | allowableClients.push_back(Elt: allowableClient); |
| 1882 | } |
| 1883 | } |
| 1884 | |
| 1885 | const auto *dyldInfo = findCommand<dyld_info_command>(anyHdr: hdr, types: LC_DYLD_INFO_ONLY); |
| 1886 | const auto *exportsTrie = |
| 1887 | findCommand<linkedit_data_command>(anyHdr: hdr, types: LC_DYLD_EXPORTS_TRIE); |
| 1888 | if (dyldInfo && exportsTrie) { |
| 1889 | // It's unclear what should happen in this case. Maybe we should only error |
| 1890 | // out if the two load commands refer to different data? |
| 1891 | error(msg: toString(f: this) + |
| 1892 | ": dylib has both LC_DYLD_INFO_ONLY and LC_DYLD_EXPORTS_TRIE"); |
| 1893 | return; |
| 1894 | } |
| 1895 | |
| 1896 | if (dyldInfo) { |
| 1897 | parseExportedSymbols(offset: dyldInfo->export_off, size: dyldInfo->export_size); |
| 1898 | } else if (exportsTrie) { |
| 1899 | parseExportedSymbols(offset: exportsTrie->dataoff, size: exportsTrie->datasize); |
| 1900 | } else { |
| 1901 | error(msg: "No LC_DYLD_INFO_ONLY or LC_DYLD_EXPORTS_TRIE found in "+ |
| 1902 | toString(f: this)); |
| 1903 | } |
| 1904 | } |
| 1905 | |
| 1906 | void DylibFile::parseExportedSymbols(uint32_t offset, uint32_t size) { |
| 1907 | struct TrieEntry { |
| 1908 | StringRef name; |
| 1909 | uint64_t flags; |
| 1910 | }; |
| 1911 | |
| 1912 | auto *buf = reinterpret_cast<const uint8_t *>(mb.getBufferStart()); |
| 1913 | std::vector<TrieEntry> entries; |
| 1914 | // Find all the $ld$* symbols to process first. |
| 1915 | parseTrie(fileName: toString(f: this), buf: buf + offset, size, |
| 1916 | [&](const Twine &name, uint64_t flags) { |
| 1917 | StringRef savedName = saver().save(S: name); |
| 1918 | if (handleLDSymbol(originalName: savedName)) |
| 1919 | return; |
| 1920 | entries.push_back(x: {.name: savedName, .flags: flags}); |
| 1921 | }); |
| 1922 | |
| 1923 | // Process the "normal" symbols. |
| 1924 | for (TrieEntry &entry : entries) { |
| 1925 | if (exportingFile->hiddenSymbols.contains(V: CachedHashStringRef(entry.name))) |
| 1926 | continue; |
| 1927 | |
| 1928 | bool isWeakDef = entry.flags & EXPORT_SYMBOL_FLAGS_WEAK_DEFINITION; |
| 1929 | bool isTlv = entry.flags & EXPORT_SYMBOL_FLAGS_KIND_THREAD_LOCAL; |
| 1930 | |
| 1931 | symbols.push_back( |
| 1932 | x: symtab->addDylib(name: entry.name, file: exportingFile, isWeakDef, isTlv)); |
| 1933 | } |
| 1934 | } |
| 1935 | |
| 1936 | void DylibFile::parseLoadCommands(MemoryBufferRef mb) { |
| 1937 | auto *hdr = reinterpret_cast<const mach_header *>(mb.getBufferStart()); |
| 1938 | const uint8_t *p = reinterpret_cast<const uint8_t *>(mb.getBufferStart()) + |
| 1939 | target->headerSize; |
| 1940 | for (uint32_t i = 0, n = hdr->ncmds; i < n; ++i) { |
| 1941 | auto *cmd = reinterpret_cast<const load_command *>(p); |
| 1942 | p += cmd->cmdsize; |
| 1943 | |
| 1944 | if (!(hdr->flags & MH_NO_REEXPORTED_DYLIBS) && |
| 1945 | cmd->cmd == LC_REEXPORT_DYLIB) { |
| 1946 | const auto *c = reinterpret_cast<const dylib_command *>(cmd); |
| 1947 | StringRef reexportPath = |
| 1948 | reinterpret_cast<const char *>(c) + read32le(P: &c->dylib.name); |
| 1949 | loadReexport(path: reexportPath, umbrella: exportingFile, currentTopLevelTapi: nullptr); |
| 1950 | } |
| 1951 | |
| 1952 | // FIXME: What about LC_LOAD_UPWARD_DYLIB, LC_LAZY_LOAD_DYLIB, |
| 1953 | // LC_LOAD_WEAK_DYLIB, LC_REEXPORT_DYLIB (..are reexports from dylibs with |
| 1954 | // MH_NO_REEXPORTED_DYLIBS loaded for -flat_namespace)? |
| 1955 | if (config->namespaceKind == NamespaceKind::flat && |
| 1956 | cmd->cmd == LC_LOAD_DYLIB) { |
| 1957 | const auto *c = reinterpret_cast<const dylib_command *>(cmd); |
| 1958 | StringRef dylibPath = |
| 1959 | reinterpret_cast<const char *>(c) + read32le(P: &c->dylib.name); |
| 1960 | DylibFile *dylib = findDylib(path: dylibPath, umbrella, currentTopLevelTapi: nullptr); |
| 1961 | if (!dylib) |
| 1962 | error(msg: Twine("unable to locate library '") + dylibPath + |
| 1963 | "' loaded from '"+ toString(f: this) + "' for -flat_namespace"); |
| 1964 | } |
| 1965 | } |
| 1966 | } |
| 1967 | |
| 1968 | // Some versions of Xcode ship with .tbd files that don't have the right |
| 1969 | // platform settings. |
| 1970 | constexpr std::array<StringRef, 3> skipPlatformChecks{ |
| 1971 | "/usr/lib/system/libsystem_kernel.dylib", |
| 1972 | "/usr/lib/system/libsystem_platform.dylib", |
| 1973 | "/usr/lib/system/libsystem_pthread.dylib"}; |
| 1974 | |
| 1975 | static bool isArchABICompatible(ArchitectureSet archSet, |
| 1976 | Architecture targetArch) { |
| 1977 | uint32_t cpuType; |
| 1978 | uint32_t targetCpuType; |
| 1979 | std::tie(args&: targetCpuType, args: std::ignore) = getCPUTypeFromArchitecture(Arch: targetArch); |
| 1980 | |
| 1981 | return llvm::any_of(Range&: archSet, P: [&](const auto &p) { |
| 1982 | std::tie(args&: cpuType, args: std::ignore) = getCPUTypeFromArchitecture(p); |
| 1983 | return cpuType == targetCpuType; |
| 1984 | }); |
| 1985 | } |
| 1986 | |
| 1987 | static bool skipPlatformCheckForCatalyst(const InterfaceFile &interface, |
| 1988 | bool explicitlyLinked) { |
| 1989 | // Catalyst outputs can link against implicitly linked macOS-only libraries. |
| 1990 | if (config->platform() != PLATFORM_MACCATALYST || explicitlyLinked) |
| 1991 | return false; |
| 1992 | ArchitectureSet macOSArchs; |
| 1993 | for (const auto &target : interface.targets()) |
| 1994 | if (target.Platform == PLATFORM_MACOS) |
| 1995 | macOSArchs.set(target.Arch); |
| 1996 | return isArchABICompatible(archSet: macOSArchs, targetArch: config->arch()); |
| 1997 | } |
| 1998 | |
| 1999 | static bool isTargetPlatformArchCompatible( |
| 2000 | InterfaceFile::const_target_range interfaceTargets, Target target) { |
| 2001 | if (is_contained(Range&: interfaceTargets, Element: target)) |
| 2002 | return true; |
| 2003 | |
| 2004 | if (config->forceExactCpuSubtypeMatch) |
| 2005 | return false; |
| 2006 | |
| 2007 | ArchitectureSet archSet; |
| 2008 | for (const auto &p : interfaceTargets) |
| 2009 | if (p.Platform == target.Platform) |
| 2010 | archSet.set(p.Arch); |
| 2011 | if (archSet.empty()) |
| 2012 | return false; |
| 2013 | |
| 2014 | return isArchABICompatible(archSet, targetArch: target.Arch); |
| 2015 | } |
| 2016 | |
| 2017 | DylibFile::DylibFile(const InterfaceFile &interface, DylibFile *umbrella, |
| 2018 | bool isBundleLoader, bool explicitlyLinked) |
| 2019 | : InputFile(DylibKind, interface), refState(RefState::Unreferenced), |
| 2020 | explicitlyLinked(explicitlyLinked), isBundleLoader(isBundleLoader) { |
| 2021 | // FIXME: Add test for the missing TBD code path. |
| 2022 | |
| 2023 | if (umbrella == nullptr) |
| 2024 | umbrella = this; |
| 2025 | this->umbrella = umbrella; |
| 2026 | |
| 2027 | installName = saver().save(S: interface.getInstallName()); |
| 2028 | compatibilityVersion = interface.getCompatibilityVersion().rawValue(); |
| 2029 | currentVersion = interface.getCurrentVersion().rawValue(); |
| 2030 | for (const auto &rpath : interface.rpaths()) |
| 2031 | if (rpath.first == config->platformInfo.target) |
| 2032 | rpaths.push_back(Elt: saver().save(S: rpath.second)); |
| 2033 | |
| 2034 | if (config->printEachFile) |
| 2035 | message(msg: toString(f: this)); |
| 2036 | inputFiles.insert(X: this); |
| 2037 | |
| 2038 | if (!is_contained(Range: skipPlatformChecks, Element: installName) && |
| 2039 | !isTargetPlatformArchCompatible(interfaceTargets: interface.targets(), |
| 2040 | target: config->platformInfo.target) && |
| 2041 | !skipPlatformCheckForCatalyst(interface, explicitlyLinked)) { |
| 2042 | error(msg: toString(f: this) + " is incompatible with "+ |
| 2043 | std::string(config->platformInfo.target)); |
| 2044 | return; |
| 2045 | } |
| 2046 | |
| 2047 | checkAppExtensionSafety(dylibIsAppExtensionSafe: interface.isApplicationExtensionSafe()); |
| 2048 | |
| 2049 | bool canBeImplicitlyLinked = interface.allowableClients().size() == 0; |
| 2050 | exportingFile = (canBeImplicitlyLinked && isImplicitlyLinked(path: installName)) |
| 2051 | ? this |
| 2052 | : umbrella; |
| 2053 | |
| 2054 | if (!canBeImplicitlyLinked) |
| 2055 | for (const auto &allowableClient : interface.allowableClients()) |
| 2056 | allowableClients.push_back( |
| 2057 | Elt: *make<std::string>(args: allowableClient.getInstallName().data())); |
| 2058 | |
| 2059 | auto addSymbol = [&](const llvm::MachO::Symbol &symbol, |
| 2060 | const Twine &name) -> void { |
| 2061 | StringRef savedName = saver().save(S: name); |
| 2062 | if (exportingFile->hiddenSymbols.contains(V: CachedHashStringRef(savedName))) |
| 2063 | return; |
| 2064 | |
| 2065 | symbols.push_back(x: symtab->addDylib(name: savedName, file: exportingFile, |
| 2066 | isWeakDef: symbol.isWeakDefined(), |
| 2067 | isTlv: symbol.isThreadLocalValue())); |
| 2068 | }; |
| 2069 | |
| 2070 | std::vector<const llvm::MachO::Symbol *> normalSymbols; |
| 2071 | normalSymbols.reserve(n: interface.symbolsCount()); |
| 2072 | for (const auto *symbol : interface.symbols()) { |
| 2073 | if (!isArchABICompatible(archSet: symbol->getArchitectures(), targetArch: config->arch())) |
| 2074 | continue; |
| 2075 | if (handleLDSymbol(originalName: symbol->getName())) |
| 2076 | continue; |
| 2077 | |
| 2078 | switch (symbol->getKind()) { |
| 2079 | case EncodeKind::GlobalSymbol: |
| 2080 | case EncodeKind::ObjectiveCClass: |
| 2081 | case EncodeKind::ObjectiveCClassEHType: |
| 2082 | case EncodeKind::ObjectiveCInstanceVariable: |
| 2083 | normalSymbols.push_back(x: symbol); |
| 2084 | } |
| 2085 | } |
| 2086 | // interface.symbols() order is non-deterministic. |
| 2087 | llvm::sort(C&: normalSymbols, |
| 2088 | Comp: [](auto *l, auto *r) { return l->getName() < r->getName(); }); |
| 2089 | |
| 2090 | // TODO(compnerd) filter out symbols based on the target platform |
| 2091 | for (const auto *symbol : normalSymbols) { |
| 2092 | switch (symbol->getKind()) { |
| 2093 | case EncodeKind::GlobalSymbol: |
| 2094 | addSymbol(*symbol, symbol->getName()); |
| 2095 | break; |
| 2096 | case EncodeKind::ObjectiveCClass: |
| 2097 | // XXX ld64 only creates these symbols when -ObjC is passed in. We may |
| 2098 | // want to emulate that. |
| 2099 | addSymbol(*symbol, objc::symbol_names::klass + symbol->getName()); |
| 2100 | addSymbol(*symbol, objc::symbol_names::metaclass + symbol->getName()); |
| 2101 | break; |
| 2102 | case EncodeKind::ObjectiveCClassEHType: |
| 2103 | addSymbol(*symbol, objc::symbol_names::ehtype + symbol->getName()); |
| 2104 | break; |
| 2105 | case EncodeKind::ObjectiveCInstanceVariable: |
| 2106 | addSymbol(*symbol, objc::symbol_names::ivar + symbol->getName()); |
| 2107 | break; |
| 2108 | } |
| 2109 | } |
| 2110 | } |
| 2111 | |
| 2112 | DylibFile::DylibFile(DylibFile *umbrella) |
| 2113 | : InputFile(DylibKind, MemoryBufferRef{}), refState(RefState::Unreferenced), |
| 2114 | explicitlyLinked(false), isBundleLoader(false) { |
| 2115 | if (umbrella == nullptr) |
| 2116 | umbrella = this; |
| 2117 | this->umbrella = umbrella; |
| 2118 | } |
| 2119 | |
| 2120 | void DylibFile::parseReexports(const InterfaceFile &interface) { |
| 2121 | const InterfaceFile *topLevel = |
| 2122 | interface.getParent() == nullptr ? &interface : interface.getParent(); |
| 2123 | for (const InterfaceFileRef &intfRef : interface.reexportedLibraries()) { |
| 2124 | InterfaceFile::const_target_range targets = intfRef.targets(); |
| 2125 | if (is_contained(Range: skipPlatformChecks, Element: intfRef.getInstallName()) || |
| 2126 | isTargetPlatformArchCompatible(interfaceTargets: targets, target: config->platformInfo.target)) |
| 2127 | loadReexport(path: intfRef.getInstallName(), umbrella: exportingFile, currentTopLevelTapi: topLevel); |
| 2128 | } |
| 2129 | } |
| 2130 | |
| 2131 | bool DylibFile::isExplicitlyLinked() const { |
| 2132 | if (!explicitlyLinked) |
| 2133 | return false; |
| 2134 | |
| 2135 | // If this dylib was explicitly linked, but at least one of the symbols |
| 2136 | // of the synthetic dylibs it created via $ld$previous symbols is |
| 2137 | // referenced, then that synthetic dylib fulfils the explicit linkedness |
| 2138 | // and we can deadstrip this dylib if it's unreferenced. |
| 2139 | for (const auto *dylib : extraDylibs) |
| 2140 | if (dylib->isReferenced()) |
| 2141 | return false; |
| 2142 | |
| 2143 | return true; |
| 2144 | } |
| 2145 | |
| 2146 | DylibFile *DylibFile::getSyntheticDylib(StringRef installName, |
| 2147 | uint32_t currentVersion, |
| 2148 | uint32_t compatVersion) { |
| 2149 | for (DylibFile *dylib : extraDylibs) |
| 2150 | if (dylib->installName == installName) { |
| 2151 | // FIXME: Check what to do if different $ld$previous symbols |
| 2152 | // request the same dylib, but with different versions. |
| 2153 | return dylib; |
| 2154 | } |
| 2155 | |
| 2156 | auto *dylib = make<DylibFile>(args: umbrella == this ? nullptr : umbrella); |
| 2157 | dylib->installName = saver().save(S: installName); |
| 2158 | dylib->currentVersion = currentVersion; |
| 2159 | dylib->compatibilityVersion = compatVersion; |
| 2160 | extraDylibs.push_back(Elt: dylib); |
| 2161 | return dylib; |
| 2162 | } |
| 2163 | |
| 2164 | // $ld$ symbols modify the properties/behavior of the library (e.g. its install |
| 2165 | // name, compatibility version or hide/add symbols) for specific target |
| 2166 | // versions. |
| 2167 | bool DylibFile::handleLDSymbol(StringRef originalName) { |
| 2168 | if (!originalName.starts_with(Prefix: "$ld$")) |
| 2169 | return false; |
| 2170 | |
| 2171 | StringRef action; |
| 2172 | StringRef name; |
| 2173 | std::tie(args&: action, args&: name) = originalName.drop_front(N: strlen(s: "$ld$")).split(Separator: '$'); |
| 2174 | if (action == "previous") |
| 2175 | handleLDPreviousSymbol(name, originalName); |
| 2176 | else if (action == "install_name") |
| 2177 | handleLDInstallNameSymbol(name, originalName); |
| 2178 | else if (action == "hide") |
| 2179 | handleLDHideSymbol(name, originalName); |
| 2180 | return true; |
| 2181 | } |
| 2182 | |
| 2183 | void DylibFile::handleLDPreviousSymbol(StringRef name, StringRef originalName) { |
| 2184 | // originalName: $ld$ previous $ <installname> $ <compatversion> $ |
| 2185 | // <platformstr> $ <startversion> $ <endversion> $ <symbol-name> $ |
| 2186 | StringRef installName; |
| 2187 | StringRef compatVersion; |
| 2188 | StringRef platformStr; |
| 2189 | StringRef startVersion; |
| 2190 | StringRef endVersion; |
| 2191 | StringRef symbolName; |
| 2192 | StringRef rest; |
| 2193 | |
| 2194 | std::tie(args&: installName, args&: name) = name.split(Separator: '$'); |
| 2195 | std::tie(args&: compatVersion, args&: name) = name.split(Separator: '$'); |
| 2196 | std::tie(args&: platformStr, args&: name) = name.split(Separator: '$'); |
| 2197 | std::tie(args&: startVersion, args&: name) = name.split(Separator: '$'); |
| 2198 | std::tie(args&: endVersion, args&: name) = name.split(Separator: '$'); |
| 2199 | std::tie(args&: symbolName, args&: rest) = name.rsplit(Separator: '$'); |
| 2200 | |
| 2201 | // FIXME: Does this do the right thing for zippered files? |
| 2202 | unsigned platform; |
| 2203 | if (platformStr.getAsInteger(Radix: 10, Result&: platform) || |
| 2204 | platform != static_cast<unsigned>(config->platform())) |
| 2205 | return; |
| 2206 | |
| 2207 | VersionTuple start; |
| 2208 | if (start.tryParse(string: startVersion)) { |
| 2209 | warn(msg: toString(f: this) + ": failed to parse start version, symbol '"+ |
| 2210 | originalName + "' ignored"); |
| 2211 | return; |
| 2212 | } |
| 2213 | VersionTuple end; |
| 2214 | if (end.tryParse(string: endVersion)) { |
| 2215 | warn(msg: toString(f: this) + ": failed to parse end version, symbol '"+ |
| 2216 | originalName + "' ignored"); |
| 2217 | return; |
| 2218 | } |
| 2219 | if (config->platformInfo.target.MinDeployment < start || |
| 2220 | config->platformInfo.target.MinDeployment >= end) |
| 2221 | return; |
| 2222 | |
| 2223 | // Initialized to compatibilityVersion for the symbolName branch below. |
| 2224 | uint32_t newCompatibilityVersion = compatibilityVersion; |
| 2225 | uint32_t newCurrentVersionForSymbol = currentVersion; |
| 2226 | if (!compatVersion.empty()) { |
| 2227 | VersionTuple cVersion; |
| 2228 | if (cVersion.tryParse(string: compatVersion)) { |
| 2229 | warn(msg: toString(f: this) + |
| 2230 | ": failed to parse compatibility version, symbol '"+ originalName + |
| 2231 | "' ignored"); |
| 2232 | return; |
| 2233 | } |
| 2234 | newCompatibilityVersion = encodeVersion(version: cVersion); |
| 2235 | newCurrentVersionForSymbol = newCompatibilityVersion; |
| 2236 | } |
| 2237 | |
| 2238 | if (!symbolName.empty()) { |
| 2239 | // A $ld$previous$ symbol with symbol name adds a symbol with that name to |
| 2240 | // a dylib with given name and version. |
| 2241 | auto *dylib = getSyntheticDylib(installName, currentVersion: newCurrentVersionForSymbol, |
| 2242 | compatVersion: newCompatibilityVersion); |
| 2243 | |
| 2244 | // The tbd file usually contains the $ld$previous symbol for an old version, |
| 2245 | // and then the symbol itself later, for newer deployment targets, like so: |
| 2246 | // symbols: [ |
| 2247 | // '$ld$previous$/Another$$1$3.0$14.0$_zzz$', |
| 2248 | // _zzz, |
| 2249 | // ] |
| 2250 | // Since the symbols are sorted, adding them to the symtab in the given |
| 2251 | // order means the $ld$previous version of _zzz will prevail, as desired. |
| 2252 | dylib->symbols.push_back(x: symtab->addDylib( |
| 2253 | name: saver().save(S: symbolName), file: dylib, /*isWeakDef=*/false, /*isTlv=*/false)); |
| 2254 | return; |
| 2255 | } |
| 2256 | |
| 2257 | // A $ld$previous$ symbol without symbol name modifies the dylib it's in. |
| 2258 | this->installName = saver().save(S: installName); |
| 2259 | this->compatibilityVersion = newCompatibilityVersion; |
| 2260 | } |
| 2261 | |
| 2262 | void DylibFile::handleLDInstallNameSymbol(StringRef name, |
| 2263 | StringRef originalName) { |
| 2264 | // originalName: $ld$ install_name $ os<version> $ install_name |
| 2265 | StringRef condition, installName; |
| 2266 | std::tie(args&: condition, args&: installName) = name.split(Separator: '$'); |
| 2267 | VersionTuple version; |
| 2268 | if (!condition.consume_front(Prefix: "os") || version.tryParse(string: condition)) |
| 2269 | warn(msg: toString(f: this) + ": failed to parse os version, symbol '"+ |
| 2270 | originalName + "' ignored"); |
| 2271 | else if (version == config->platformInfo.target.MinDeployment) |
| 2272 | this->installName = saver().save(S: installName); |
| 2273 | } |
| 2274 | |
| 2275 | void DylibFile::handleLDHideSymbol(StringRef name, StringRef originalName) { |
| 2276 | StringRef symbolName; |
| 2277 | bool shouldHide = true; |
| 2278 | if (name.starts_with(Prefix: "os")) { |
| 2279 | // If it's hidden based on versions. |
| 2280 | name = name.drop_front(N: 2); |
| 2281 | StringRef minVersion; |
| 2282 | std::tie(args&: minVersion, args&: symbolName) = name.split(Separator: '$'); |
| 2283 | VersionTuple versionTup; |
| 2284 | if (versionTup.tryParse(string: minVersion)) { |
| 2285 | warn(msg: toString(f: this) + ": failed to parse hidden version, symbol `"+ originalName + |
| 2286 | "` ignored."); |
| 2287 | return; |
| 2288 | } |
| 2289 | shouldHide = versionTup == config->platformInfo.target.MinDeployment; |
| 2290 | } else { |
| 2291 | symbolName = name; |
| 2292 | } |
| 2293 | |
| 2294 | if (shouldHide) |
| 2295 | exportingFile->hiddenSymbols.insert(V: CachedHashStringRef(symbolName)); |
| 2296 | } |
| 2297 | |
| 2298 | void DylibFile::checkAppExtensionSafety(bool dylibIsAppExtensionSafe) const { |
| 2299 | if (config->applicationExtension && !dylibIsAppExtensionSafe) |
| 2300 | warn(msg: "using '-application_extension' with unsafe dylib: "+ toString(f: this)); |
| 2301 | } |
| 2302 | |
| 2303 | ArchiveFile::ArchiveFile(std::unique_ptr<object::Archive> &&f, bool forceHidden) |
| 2304 | : InputFile(ArchiveKind, f->getMemoryBufferRef()), file(std::move(f)), |
| 2305 | forceHidden(forceHidden) {} |
| 2306 | |
| 2307 | void ArchiveFile::addLazySymbols() { |
| 2308 | // Avoid calling getMemoryBufferRef() on zero-symbol archive |
| 2309 | // since that crashes. |
| 2310 | if (file->isEmpty() || |
| 2311 | (file->hasSymbolTable() && file->getNumberOfSymbols() == 0)) |
| 2312 | return; |
| 2313 | |
| 2314 | if (!file->hasSymbolTable()) { |
| 2315 | // No index, treat each child as a lazy object file. |
| 2316 | Error e = Error::success(); |
| 2317 | for (const object::Archive::Child &c : file->children(Err&: e)) { |
| 2318 | // Check `seen` but don't insert so a future eager load can still happen. |
| 2319 | if (seen.contains(V: c.getChildOffset())) |
| 2320 | continue; |
| 2321 | if (!seenLazy.insert(V: c.getChildOffset()).second) |
| 2322 | continue; |
| 2323 | auto file = childToObjectFile(c, /*lazy=*/true); |
| 2324 | if (!file) |
| 2325 | error(msg: toString(f: this) + |
| 2326 | ": couldn't process child: "+ toString(E: file.takeError())); |
| 2327 | inputFiles.insert(X: *file); |
| 2328 | } |
| 2329 | if (e) |
| 2330 | error(msg: toString(f: this) + |
| 2331 | ": Archive::children failed: "+ toString(E: std::move(e))); |
| 2332 | return; |
| 2333 | } |
| 2334 | |
| 2335 | Error err = Error::success(); |
| 2336 | auto child = file->child_begin(Err&: err); |
| 2337 | // Ignore the I/O error here - will be reported later. |
| 2338 | if (!err) { |
| 2339 | Expected<MemoryBufferRef> mbOrErr = child->getMemoryBufferRef(); |
| 2340 | if (!mbOrErr) { |
| 2341 | llvm::consumeError(Err: mbOrErr.takeError()); |
| 2342 | } else { |
| 2343 | if (identify_magic(magic: mbOrErr->getBuffer()) == file_magic::macho_object) { |
| 2344 | if (target->wordSize == 8) |
| 2345 | compatArch = compatWithTargetArch( |
| 2346 | file: this, hdr: reinterpret_cast<const LP64::mach_header *>( |
| 2347 | mbOrErr->getBufferStart())); |
| 2348 | else |
| 2349 | compatArch = compatWithTargetArch( |
| 2350 | file: this, hdr: reinterpret_cast<const ILP32::mach_header *>( |
| 2351 | mbOrErr->getBufferStart())); |
| 2352 | if (!compatArch) |
| 2353 | return; |
| 2354 | } |
| 2355 | } |
| 2356 | } |
| 2357 | |
| 2358 | for (const object::Archive::Symbol &sym : file->symbols()) |
| 2359 | symtab->addLazyArchive(name: sym.getName(), file: this, sym); |
| 2360 | } |
| 2361 | |
| 2362 | static Expected<InputFile *> |
| 2363 | loadArchiveMember(MemoryBufferRef mb, uint32_t modTime, StringRef archiveName, |
| 2364 | uint64_t offsetInArchive, bool forceHidden, bool compatArch, |
| 2365 | bool lazy) { |
| 2366 | if (config->zeroModTime) |
| 2367 | modTime = 0; |
| 2368 | |
| 2369 | switch (identify_magic(magic: mb.getBuffer())) { |
| 2370 | case file_magic::macho_object: |
| 2371 | return make<ObjFile>(args&: mb, args&: modTime, args&: archiveName, args&: lazy, args&: forceHidden, |
| 2372 | args&: compatArch); |
| 2373 | case file_magic::bitcode: |
| 2374 | return make<BitcodeFile>(args&: mb, args&: archiveName, args&: offsetInArchive, args&: lazy, |
| 2375 | args&: forceHidden, args&: compatArch); |
| 2376 | default: |
| 2377 | return createStringError(EC: inconvertibleErrorCode(), |
| 2378 | S: mb.getBufferIdentifier() + |
| 2379 | " has unhandled file type"); |
| 2380 | } |
| 2381 | } |
| 2382 | |
| 2383 | Error ArchiveFile::fetch(const object::Archive::Child &c, StringRef reason) { |
| 2384 | if (!seen.insert(V: c.getChildOffset()).second) |
| 2385 | return Error::success(); |
| 2386 | auto file = childToObjectFile(c, /*lazy=*/false); |
| 2387 | if (!file) |
| 2388 | return file.takeError(); |
| 2389 | |
| 2390 | inputFiles.insert(X: *file); |
| 2391 | printArchiveMemberLoad(reason, *file); |
| 2392 | return Error::success(); |
| 2393 | } |
| 2394 | |
| 2395 | void ArchiveFile::fetch(const object::Archive::Symbol &sym) { |
| 2396 | object::Archive::Child c = |
| 2397 | CHECK(sym.getMember(), toString(this) + |
| 2398 | ": could not get the member defining symbol "+ |
| 2399 | toMachOString(sym)); |
| 2400 | |
| 2401 | // `sym` is owned by a LazySym, which will be replace<>()d by make<ObjFile> |
| 2402 | // and become invalid after that call. Copy it to the stack so we can refer |
| 2403 | // to it later. |
| 2404 | const object::Archive::Symbol symCopy = sym; |
| 2405 | |
| 2406 | // ld64 doesn't demangle sym here even with -demangle. |
| 2407 | // Match that: intentionally don't call toMachOString(). |
| 2408 | if (Error e = fetch(c, reason: symCopy.getName())) |
| 2409 | error(msg: toString(f: this) + ": could not get the member defining symbol "+ |
| 2410 | toMachOString(symCopy) + ": "+ toString(E: std::move(e))); |
| 2411 | } |
| 2412 | |
| 2413 | Expected<InputFile *> |
| 2414 | ArchiveFile::childToObjectFile(const llvm::object::Archive::Child &c, |
| 2415 | bool lazy) { |
| 2416 | Expected<MemoryBufferRef> mb = c.getMemoryBufferRef(); |
| 2417 | if (!mb) |
| 2418 | return mb.takeError(); |
| 2419 | |
| 2420 | Expected<TimePoint<std::chrono::seconds>> modTime = c.getLastModified(); |
| 2421 | if (!modTime) |
| 2422 | return modTime.takeError(); |
| 2423 | |
| 2424 | return loadArchiveMember(mb: *mb, modTime: toTimeT(TP: *modTime), archiveName: getName(), |
| 2425 | offsetInArchive: c.getChildOffset(), forceHidden, compatArch, lazy); |
| 2426 | } |
| 2427 | |
| 2428 | static macho::Symbol *createBitcodeSymbol(const lto::InputFile::Symbol &objSym, |
| 2429 | BitcodeFile &file) { |
| 2430 | StringRef name = saver().save(S: objSym.getName()); |
| 2431 | |
| 2432 | if (objSym.isUndefined()) |
| 2433 | return symtab->addUndefined(name, &file, /*isWeakRef=*/objSym.isWeak()); |
| 2434 | |
| 2435 | // TODO: Write a test demonstrating why computing isPrivateExtern before |
| 2436 | // LTO compilation is important. |
| 2437 | bool isPrivateExtern = false; |
| 2438 | switch (objSym.getVisibility()) { |
| 2439 | case GlobalValue::HiddenVisibility: |
| 2440 | isPrivateExtern = true; |
| 2441 | break; |
| 2442 | case GlobalValue::ProtectedVisibility: |
| 2443 | error(msg: name + " has protected visibility, which is not supported by Mach-O"); |
| 2444 | break; |
| 2445 | case GlobalValue::DefaultVisibility: |
| 2446 | break; |
| 2447 | } |
| 2448 | isPrivateExtern = isPrivateExtern || objSym.canBeOmittedFromSymbolTable() || |
| 2449 | file.forceHidden; |
| 2450 | |
| 2451 | if (objSym.isCommon()) |
| 2452 | return symtab->addCommon(name, &file, size: objSym.getCommonSize(), |
| 2453 | align: objSym.getCommonAlignment(), isPrivateExtern); |
| 2454 | |
| 2455 | return symtab->addDefined(name, &file, /*isec=*/nullptr, /*value=*/0, |
| 2456 | /*size=*/0, isWeakDef: objSym.isWeak(), isPrivateExtern, |
| 2457 | /*isReferencedDynamically=*/false, |
| 2458 | /*noDeadStrip=*/false, |
| 2459 | /*isWeakDefCanBeHidden=*/false); |
| 2460 | } |
| 2461 | |
| 2462 | BitcodeFile::BitcodeFile(MemoryBufferRef mb, StringRef archiveName, |
| 2463 | uint64_t offsetInArchive, bool lazy, bool forceHidden, |
| 2464 | bool compatArch) |
| 2465 | : InputFile(BitcodeKind, mb, lazy), forceHidden(forceHidden) { |
| 2466 | this->archiveName = std::string(archiveName); |
| 2467 | this->compatArch = compatArch; |
| 2468 | std::string path = mb.getBufferIdentifier().str(); |
| 2469 | if (config->thinLTOIndexOnly) |
| 2470 | path = replaceThinLTOSuffix(path: mb.getBufferIdentifier()); |
| 2471 | |
| 2472 | // If the parent archive already determines that the arch is not compat with |
| 2473 | // target, then just return. |
| 2474 | if (!compatArch) |
| 2475 | return; |
| 2476 | |
| 2477 | // ThinLTO assumes that all MemoryBufferRefs given to it have a unique |
| 2478 | // name. If two members with the same name are provided, this causes a |
| 2479 | // collision and ThinLTO can't proceed. |
| 2480 | // So, we append the archive name to disambiguate two members with the same |
| 2481 | // name from multiple different archives, and offset within the archive to |
| 2482 | // disambiguate two members of the same name from a single archive. |
| 2483 | MemoryBufferRef mbref(mb.getBuffer(), |
| 2484 | saver().save(S: archiveName.empty() |
| 2485 | ? path |
| 2486 | : archiveName + "("+ |
| 2487 | sys::path::filename(path) + ")"+ |
| 2488 | utostr(X: offsetInArchive))); |
| 2489 | obj = check(e: lto::InputFile::create(Object: mbref)); |
| 2490 | if (lazy) |
| 2491 | parseLazy(); |
| 2492 | else |
| 2493 | parse(); |
| 2494 | } |
| 2495 | |
| 2496 | void BitcodeFile::parse() { |
| 2497 | // Convert LTO Symbols to LLD Symbols in order to perform resolution. The |
| 2498 | // "winning" symbol will then be marked as Prevailing at LTO compilation |
| 2499 | // time. |
| 2500 | symbols.resize(new_size: obj->symbols().size()); |
| 2501 | |
| 2502 | // Process defined symbols first. See the comment at the end of |
| 2503 | // ObjFile<>::parseSymbols. |
| 2504 | for (auto it : llvm::enumerate(First: obj->symbols())) |
| 2505 | if (!it.value().isUndefined()) |
| 2506 | symbols[it.index()] = createBitcodeSymbol(objSym: it.value(), file&: *this); |
| 2507 | for (auto it : llvm::enumerate(First: obj->symbols())) |
| 2508 | if (it.value().isUndefined()) |
| 2509 | symbols[it.index()] = createBitcodeSymbol(objSym: it.value(), file&: *this); |
| 2510 | } |
| 2511 | |
| 2512 | void BitcodeFile::parseLazy() { |
| 2513 | symbols.resize(new_size: obj->symbols().size()); |
| 2514 | for (const auto &[i, objSym] : llvm::enumerate(First: obj->symbols())) { |
| 2515 | if (!objSym.isUndefined()) { |
| 2516 | symbols[i] = symtab->addLazyObject(name: saver().save(S: objSym.getName()), file&: *this); |
| 2517 | if (!lazy) |
| 2518 | break; |
| 2519 | } |
| 2520 | } |
| 2521 | } |
| 2522 | |
| 2523 | std::string macho::replaceThinLTOSuffix(StringRef path) { |
| 2524 | auto [suffix, repl] = config->thinLTOObjectSuffixReplace; |
| 2525 | if (path.consume_back(Suffix: suffix)) |
| 2526 | return (path + repl).str(); |
| 2527 | return std::string(path); |
| 2528 | } |
| 2529 | |
| 2530 | void macho::extract(InputFile &file, StringRef reason) { |
| 2531 | if (!file.lazy) |
| 2532 | return; |
| 2533 | file.lazy = false; |
| 2534 | |
| 2535 | printArchiveMemberLoad(reason, &file); |
| 2536 | if (auto *bitcode = dyn_cast<BitcodeFile>(Val: &file)) { |
| 2537 | bitcode->parse(); |
| 2538 | } else { |
| 2539 | auto &f = cast<ObjFile>(Val&: file); |
| 2540 | if (target->wordSize == 8) |
| 2541 | f.parse<LP64>(); |
| 2542 | else |
| 2543 | f.parse<ILP32>(); |
| 2544 | } |
| 2545 | } |
| 2546 | |
| 2547 | template void ObjFile::parse<LP64>(); |
| 2548 |
Generated on 2026-Oct-10 from project llvm_projects revision b1dd93c6a
Powered by 
Code Browser 2.1
Generator usage only permitted with license.