1 //=--------- MachOLinkGraphBuilder.cpp - MachO LinkGraph builder ----------===// 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 // Generic MachO LinkGraph buliding code. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "MachOLinkGraphBuilder.h" 14 15 #define DEBUG_TYPE "jitlink" 16 17 static const char *CommonSectionName = "__common"; 18 19 namespace llvm { 20 namespace jitlink { 21 22 MachOLinkGraphBuilder::~MachOLinkGraphBuilder() {} 23 24 Expected<std::unique_ptr<LinkGraph>> MachOLinkGraphBuilder::buildGraph() { 25 26 // Sanity check: we only operate on relocatable objects. 27 if (!Obj.isRelocatableObject()) 28 return make_error<JITLinkError>("Object is not a relocatable MachO"); 29 30 if (auto Err = createNormalizedSections()) 31 return std::move(Err); 32 33 if (auto Err = createNormalizedSymbols()) 34 return std::move(Err); 35 36 if (auto Err = graphifyRegularSymbols()) 37 return std::move(Err); 38 39 if (auto Err = graphifySectionsWithCustomParsers()) 40 return std::move(Err); 41 42 if (auto Err = addRelocations()) 43 return std::move(Err); 44 45 return std::move(G); 46 } 47 48 MachOLinkGraphBuilder::MachOLinkGraphBuilder(const object::MachOObjectFile &Obj) 49 : Obj(Obj), 50 G(std::make_unique<LinkGraph>(Obj.getFileName(), getPointerSize(Obj), 51 getEndianness(Obj))) {} 52 53 void MachOLinkGraphBuilder::addCustomSectionParser( 54 StringRef SectionName, SectionParserFunction Parser) { 55 assert(!CustomSectionParserFunctions.count(SectionName) && 56 "Custom parser for this section already exists"); 57 CustomSectionParserFunctions[SectionName] = std::move(Parser); 58 } 59 60 Linkage MachOLinkGraphBuilder::getLinkage(uint16_t Desc) { 61 if ((Desc & MachO::N_WEAK_DEF) || (Desc & MachO::N_WEAK_REF)) 62 return Linkage::Weak; 63 return Linkage::Strong; 64 } 65 66 Scope MachOLinkGraphBuilder::getScope(StringRef Name, uint8_t Type) { 67 if (Name.startswith("l")) 68 return Scope::Local; 69 if (Type & MachO::N_PEXT) 70 return Scope::Hidden; 71 if (Type & MachO::N_EXT) 72 return Scope::Default; 73 return Scope::Local; 74 } 75 76 bool MachOLinkGraphBuilder::isAltEntry(const NormalizedSymbol &NSym) { 77 return NSym.Desc & MachO::N_ALT_ENTRY; 78 } 79 80 unsigned 81 MachOLinkGraphBuilder::getPointerSize(const object::MachOObjectFile &Obj) { 82 return Obj.is64Bit() ? 8 : 4; 83 } 84 85 support::endianness 86 MachOLinkGraphBuilder::getEndianness(const object::MachOObjectFile &Obj) { 87 return Obj.isLittleEndian() ? support::little : support::big; 88 } 89 90 Section &MachOLinkGraphBuilder::getCommonSection() { 91 if (!CommonSection) { 92 auto Prot = static_cast<sys::Memory::ProtectionFlags>( 93 sys::Memory::MF_READ | sys::Memory::MF_WRITE); 94 CommonSection = &G->createSection(CommonSectionName, Prot); 95 } 96 return *CommonSection; 97 } 98 99 Error MachOLinkGraphBuilder::createNormalizedSections() { 100 // Build normalized sections. Verifies that section data is in-range (for 101 // sections with content) and that address ranges are non-overlapping. 102 103 LLVM_DEBUG(dbgs() << "Creating normalized sections...\n"); 104 105 for (auto &SecRef : Obj.sections()) { 106 NormalizedSection NSec; 107 uint32_t DataOffset = 0; 108 109 auto SecIndex = Obj.getSectionIndex(SecRef.getRawDataRefImpl()); 110 111 auto Name = SecRef.getName(); 112 if (!Name) 113 return Name.takeError(); 114 115 if (Obj.is64Bit()) { 116 const MachO::section_64 &Sec64 = 117 Obj.getSection64(SecRef.getRawDataRefImpl()); 118 119 NSec.Address = Sec64.addr; 120 NSec.Size = Sec64.size; 121 NSec.Alignment = 1ULL << Sec64.align; 122 NSec.Flags = Sec64.flags; 123 DataOffset = Sec64.offset; 124 } else { 125 const MachO::section &Sec32 = Obj.getSection(SecRef.getRawDataRefImpl()); 126 NSec.Address = Sec32.addr; 127 NSec.Size = Sec32.size; 128 NSec.Alignment = 1ULL << Sec32.align; 129 NSec.Flags = Sec32.flags; 130 DataOffset = Sec32.offset; 131 } 132 133 LLVM_DEBUG({ 134 dbgs() << " " << *Name << ": " << formatv("{0:x16}", NSec.Address) 135 << " -- " << formatv("{0:x16}", NSec.Address + NSec.Size) 136 << ", align: " << NSec.Alignment << ", index: " << SecIndex 137 << "\n"; 138 }); 139 140 // Get the section data if any. 141 { 142 unsigned SectionType = NSec.Flags & MachO::SECTION_TYPE; 143 if (SectionType != MachO::S_ZEROFILL && 144 SectionType != MachO::S_GB_ZEROFILL) { 145 146 if (DataOffset + NSec.Size > Obj.getData().size()) 147 return make_error<JITLinkError>( 148 "Section data extends past end of file"); 149 150 NSec.Data = Obj.getData().data() + DataOffset; 151 } 152 } 153 154 // Get prot flags. 155 // FIXME: Make sure this test is correct (it's probably missing cases 156 // as-is). 157 sys::Memory::ProtectionFlags Prot; 158 if (NSec.Flags & MachO::S_ATTR_PURE_INSTRUCTIONS) 159 Prot = static_cast<sys::Memory::ProtectionFlags>(sys::Memory::MF_READ | 160 sys::Memory::MF_EXEC); 161 else 162 Prot = static_cast<sys::Memory::ProtectionFlags>(sys::Memory::MF_READ | 163 sys::Memory::MF_WRITE); 164 165 NSec.GraphSection = &G->createSection(*Name, Prot); 166 IndexToSection.insert(std::make_pair(SecIndex, std::move(NSec))); 167 } 168 169 std::vector<NormalizedSection *> Sections; 170 Sections.reserve(IndexToSection.size()); 171 for (auto &KV : IndexToSection) 172 Sections.push_back(&KV.second); 173 174 // If we didn't end up creating any sections then bail out. The code below 175 // assumes that we have at least one section. 176 if (Sections.empty()) 177 return Error::success(); 178 179 llvm::sort(Sections, 180 [](const NormalizedSection *LHS, const NormalizedSection *RHS) { 181 assert(LHS && RHS && "Null section?"); 182 return LHS->Address < RHS->Address; 183 }); 184 185 for (unsigned I = 0, E = Sections.size() - 1; I != E; ++I) { 186 auto &Cur = *Sections[I]; 187 auto &Next = *Sections[I + 1]; 188 if (Next.Address < Cur.Address + Cur.Size) 189 return make_error<JITLinkError>( 190 "Address range for section " + Cur.GraphSection->getName() + 191 formatv(" [ {0:x16} -- {1:x16} ] ", Cur.Address, 192 Cur.Address + Cur.Size) + 193 "overlaps " + 194 formatv(" [ {0:x16} -- {1:x16} ] ", Next.Address, 195 Next.Address + Next.Size)); 196 } 197 198 return Error::success(); 199 } 200 201 Error MachOLinkGraphBuilder::createNormalizedSymbols() { 202 LLVM_DEBUG(dbgs() << "Creating normalized symbols...\n"); 203 204 for (auto &SymRef : Obj.symbols()) { 205 206 unsigned SymbolIndex = Obj.getSymbolIndex(SymRef.getRawDataRefImpl()); 207 uint64_t Value; 208 uint32_t NStrX; 209 uint8_t Type; 210 uint8_t Sect; 211 uint16_t Desc; 212 213 if (Obj.is64Bit()) { 214 const MachO::nlist_64 &NL64 = 215 Obj.getSymbol64TableEntry(SymRef.getRawDataRefImpl()); 216 Value = NL64.n_value; 217 NStrX = NL64.n_strx; 218 Type = NL64.n_type; 219 Sect = NL64.n_sect; 220 Desc = NL64.n_desc; 221 } else { 222 const MachO::nlist &NL32 = 223 Obj.getSymbolTableEntry(SymRef.getRawDataRefImpl()); 224 Value = NL32.n_value; 225 NStrX = NL32.n_strx; 226 Type = NL32.n_type; 227 Sect = NL32.n_sect; 228 Desc = NL32.n_desc; 229 } 230 231 // Skip stabs. 232 // FIXME: Are there other symbols we should be skipping? 233 if (Type & MachO::N_STAB) 234 continue; 235 236 Optional<StringRef> Name; 237 if (NStrX) { 238 if (auto NameOrErr = SymRef.getName()) 239 Name = *NameOrErr; 240 else 241 return NameOrErr.takeError(); 242 } 243 244 LLVM_DEBUG({ 245 dbgs() << " "; 246 if (!Name) 247 dbgs() << "<anonymous symbol>"; 248 else 249 dbgs() << *Name; 250 dbgs() << ": value = " << formatv("{0:x16}", Value) 251 << ", type = " << formatv("{0:x2}", Type) 252 << ", desc = " << formatv("{0:x4}", Desc) << ", sect = "; 253 if (Sect) 254 dbgs() << static_cast<unsigned>(Sect - 1); 255 else 256 dbgs() << "none"; 257 dbgs() << "\n"; 258 }); 259 260 // If this symbol has a section, sanity check that the addresses line up. 261 NormalizedSection *NSec = nullptr; 262 if (Sect != 0) { 263 if (auto NSecOrErr = findSectionByIndex(Sect - 1)) 264 NSec = &*NSecOrErr; 265 else 266 return NSecOrErr.takeError(); 267 268 if (Value < NSec->Address || Value > NSec->Address + NSec->Size) 269 return make_error<JITLinkError>("Symbol address does not fall within " 270 "section"); 271 } 272 273 IndexToSymbol[SymbolIndex] = 274 &createNormalizedSymbol(*Name, Value, Type, Sect, Desc, 275 getLinkage(Type), getScope(*Name, Type)); 276 } 277 278 return Error::success(); 279 } 280 281 void MachOLinkGraphBuilder::addSectionStartSymAndBlock( 282 Section &GraphSec, uint64_t Address, const char *Data, uint64_t Size, 283 uint32_t Alignment, bool IsLive) { 284 Block &B = 285 Data ? G->createContentBlock(GraphSec, StringRef(Data, Size), Address, 286 Alignment, 0) 287 : G->createZeroFillBlock(GraphSec, Size, Address, Alignment, 0); 288 auto &Sym = G->addAnonymousSymbol(B, 0, Size, false, IsLive); 289 assert(!AddrToCanonicalSymbol.count(Sym.getAddress()) && 290 "Anonymous block start symbol clashes with existing symbol address"); 291 AddrToCanonicalSymbol[Sym.getAddress()] = &Sym; 292 } 293 294 Error MachOLinkGraphBuilder::graphifyRegularSymbols() { 295 296 LLVM_DEBUG(dbgs() << "Creating graph symbols...\n"); 297 298 /// We only have 256 section indexes: Use a vector rather than a map. 299 std::vector<std::vector<NormalizedSymbol *>> SecIndexToSymbols; 300 SecIndexToSymbols.resize(256); 301 302 // Create commons, externs, and absolutes, and partition all other symbols by 303 // section. 304 for (auto &KV : IndexToSymbol) { 305 auto &NSym = *KV.second; 306 307 switch (NSym.Type & MachO::N_TYPE) { 308 case MachO::N_UNDF: 309 if (NSym.Value) { 310 if (!NSym.Name) 311 return make_error<JITLinkError>("Anonymous common symbol at index " + 312 Twine(KV.first)); 313 NSym.GraphSymbol = &G->addCommonSymbol( 314 *NSym.Name, NSym.S, getCommonSection(), NSym.Value, 0, 315 1ull << MachO::GET_COMM_ALIGN(NSym.Desc), 316 NSym.Desc & MachO::N_NO_DEAD_STRIP); 317 } else { 318 if (!NSym.Name) 319 return make_error<JITLinkError>("Anonymous external symbol at " 320 "index " + 321 Twine(KV.first)); 322 NSym.GraphSymbol = &G->addExternalSymbol(*NSym.Name, 0); 323 } 324 break; 325 case MachO::N_ABS: 326 if (!NSym.Name) 327 return make_error<JITLinkError>("Anonymous absolute symbol at index " + 328 Twine(KV.first)); 329 NSym.GraphSymbol = &G->addAbsoluteSymbol( 330 *NSym.Name, NSym.Value, 0, Linkage::Strong, Scope::Default, 331 NSym.Desc & MachO::N_NO_DEAD_STRIP); 332 break; 333 case MachO::N_SECT: 334 SecIndexToSymbols[NSym.Sect - 1].push_back(&NSym); 335 break; 336 case MachO::N_PBUD: 337 return make_error<JITLinkError>( 338 "Unupported N_PBUD symbol " + 339 (NSym.Name ? ("\"" + *NSym.Name + "\"") : Twine("<anon>")) + 340 " at index " + Twine(KV.first)); 341 case MachO::N_INDR: 342 return make_error<JITLinkError>( 343 "Unupported N_INDR symbol " + 344 (NSym.Name ? ("\"" + *NSym.Name + "\"") : Twine("<anon>")) + 345 " at index " + Twine(KV.first)); 346 default: 347 return make_error<JITLinkError>( 348 "Unrecognized symbol type " + Twine(NSym.Type & MachO::N_TYPE) + 349 " for symbol " + 350 (NSym.Name ? ("\"" + *NSym.Name + "\"") : Twine("<anon>")) + 351 " at index " + Twine(KV.first)); 352 } 353 } 354 355 // Loop over sections performing regular graphification for those that 356 // don't have custom parsers. 357 for (auto &KV : IndexToSection) { 358 auto SecIndex = KV.first; 359 auto &NSec = KV.second; 360 361 // Skip sections with custom parsers. 362 if (CustomSectionParserFunctions.count(NSec.GraphSection->getName())) { 363 LLVM_DEBUG({ 364 dbgs() << " Skipping section " << NSec.GraphSection->getName() 365 << " as it has a custom parser.\n"; 366 }); 367 continue; 368 } else 369 LLVM_DEBUG({ 370 dbgs() << " Processing section " << NSec.GraphSection->getName() 371 << "...\n"; 372 }); 373 374 bool SectionIsNoDeadStrip = NSec.Flags & MachO::S_ATTR_NO_DEAD_STRIP; 375 bool SectionIsText = NSec.Flags & MachO::S_ATTR_PURE_INSTRUCTIONS; 376 377 auto &SecNSymStack = SecIndexToSymbols[SecIndex]; 378 379 // If this section is non-empty but there are no symbols covering it then 380 // create one block and anonymous symbol to cover the entire section. 381 if (SecNSymStack.empty()) { 382 if (NSec.Size > 0) { 383 LLVM_DEBUG({ 384 dbgs() << " Section non-empty, but contains no symbols. " 385 "Creating anonymous block to cover " 386 << formatv("{0:x16}", NSec.Address) << " -- " 387 << formatv("{0:x16}", NSec.Address + NSec.Size) << "\n"; 388 }); 389 addSectionStartSymAndBlock(*NSec.GraphSection, NSec.Address, NSec.Data, 390 NSec.Size, NSec.Alignment, 391 SectionIsNoDeadStrip); 392 } else 393 LLVM_DEBUG({ 394 dbgs() << " Section empty and contains no symbols. Skipping.\n"; 395 }); 396 continue; 397 } 398 399 // Sort the symbol stack in by address, alt-entry status, scope, and name. 400 // We sort in reverse order so that symbols will be visited in the right 401 // order when we pop off the stack below. 402 llvm::sort(SecNSymStack, [](const NormalizedSymbol *LHS, 403 const NormalizedSymbol *RHS) { 404 if (LHS->Value != RHS->Value) 405 return LHS->Value > RHS->Value; 406 if (isAltEntry(*LHS) != isAltEntry(*RHS)) 407 return isAltEntry(*RHS); 408 if (LHS->S != RHS->S) 409 return static_cast<uint8_t>(LHS->S) < static_cast<uint8_t>(RHS->S); 410 return LHS->Name < RHS->Name; 411 }); 412 413 // The first symbol in a section can not be an alt-entry symbol. 414 if (!SecNSymStack.empty() && isAltEntry(*SecNSymStack.back())) 415 return make_error<JITLinkError>( 416 "First symbol in " + NSec.GraphSection->getName() + " is alt-entry"); 417 418 // If the section is non-empty but there is no symbol covering the start 419 // address then add an anonymous one. 420 if (SecNSymStack.back()->Value != NSec.Address) { 421 auto AnonBlockSize = SecNSymStack.back()->Value - NSec.Address; 422 LLVM_DEBUG({ 423 dbgs() << " Section start not covered by symbol. " 424 << "Creating anonymous block to cover [ " 425 << formatv("{0:x16}", NSec.Address) << " -- " 426 << formatv("{0:x16}", NSec.Address + AnonBlockSize) << " ]\n"; 427 }); 428 addSectionStartSymAndBlock(*NSec.GraphSection, NSec.Address, NSec.Data, 429 AnonBlockSize, NSec.Alignment, 430 SectionIsNoDeadStrip); 431 } 432 433 // Visit section symbols in order by popping off the reverse-sorted stack, 434 // building blocks for each alt-entry chain and creating symbols as we go. 435 while (!SecNSymStack.empty()) { 436 SmallVector<NormalizedSymbol *, 8> BlockSyms; 437 438 BlockSyms.push_back(SecNSymStack.back()); 439 SecNSymStack.pop_back(); 440 while (!SecNSymStack.empty() && 441 (isAltEntry(*SecNSymStack.back()) || 442 SecNSymStack.back()->Value == BlockSyms.back()->Value)) { 443 BlockSyms.push_back(SecNSymStack.back()); 444 SecNSymStack.pop_back(); 445 } 446 447 // BlockNSyms now contains the block symbols in reverse canonical order. 448 JITTargetAddress BlockStart = BlockSyms.front()->Value; 449 JITTargetAddress BlockEnd = SecNSymStack.empty() 450 ? NSec.Address + NSec.Size 451 : SecNSymStack.back()->Value; 452 JITTargetAddress BlockOffset = BlockStart - NSec.Address; 453 JITTargetAddress BlockSize = BlockEnd - BlockStart; 454 455 LLVM_DEBUG({ 456 dbgs() << " Creating block for " << formatv("{0:x16}", BlockStart) 457 << " -- " << formatv("{0:x16}", BlockEnd) << ": " 458 << NSec.GraphSection->getName() << " + " 459 << formatv("{0:x16}", BlockOffset) << " with " 460 << BlockSyms.size() << " symbol(s)...\n"; 461 }); 462 463 Block &B = 464 NSec.Data 465 ? G->createContentBlock( 466 *NSec.GraphSection, 467 StringRef(NSec.Data + BlockOffset, BlockSize), BlockStart, 468 NSec.Alignment, BlockStart % NSec.Alignment) 469 : G->createZeroFillBlock(*NSec.GraphSection, BlockSize, 470 BlockStart, NSec.Alignment, 471 BlockStart % NSec.Alignment); 472 473 Optional<JITTargetAddress> LastCanonicalAddr; 474 JITTargetAddress SymEnd = BlockEnd; 475 while (!BlockSyms.empty()) { 476 auto &NSym = *BlockSyms.back(); 477 BlockSyms.pop_back(); 478 479 bool SymLive = 480 (NSym.Desc & MachO::N_NO_DEAD_STRIP) || SectionIsNoDeadStrip; 481 482 LLVM_DEBUG({ 483 dbgs() << " " << formatv("{0:x16}", NSym.Value) << " -- " 484 << formatv("{0:x16}", SymEnd) << ": "; 485 if (!NSym.Name) 486 dbgs() << "<anonymous symbol>"; 487 else 488 dbgs() << NSym.Name; 489 if (SymLive) 490 dbgs() << " [no-dead-strip]"; 491 if (LastCanonicalAddr == NSym.Value) 492 dbgs() << " [non-canonical]"; 493 dbgs() << "\n"; 494 }); 495 496 auto &Sym = 497 NSym.Name 498 ? G->addDefinedSymbol(B, NSym.Value - BlockStart, *NSym.Name, 499 SymEnd - NSym.Value, NSym.L, NSym.S, 500 SectionIsText, SymLive) 501 : G->addAnonymousSymbol(B, NSym.Value - BlockStart, 502 SymEnd - NSym.Value, SectionIsText, 503 SymLive); 504 NSym.GraphSymbol = &Sym; 505 if (LastCanonicalAddr != Sym.getAddress()) { 506 if (LastCanonicalAddr) 507 SymEnd = *LastCanonicalAddr; 508 LastCanonicalAddr = Sym.getAddress(); 509 setCanonicalSymbol(Sym); 510 } 511 } 512 } 513 } 514 515 return Error::success(); 516 } 517 518 Error MachOLinkGraphBuilder::graphifySectionsWithCustomParsers() { 519 // Graphify special sections. 520 for (auto &KV : IndexToSection) { 521 auto &NSec = KV.second; 522 523 auto HI = CustomSectionParserFunctions.find(NSec.GraphSection->getName()); 524 if (HI != CustomSectionParserFunctions.end()) { 525 auto &Parse = HI->second; 526 if (auto Err = Parse(NSec)) 527 return Err; 528 } 529 } 530 531 return Error::success(); 532 } 533 534 } // end namespace jitlink 535 } // end namespace llvm 536