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() = default; 23 24 Expected<std::unique_ptr<LinkGraph>> MachOLinkGraphBuilder::buildGraph() { 25 26 // 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( 49 const object::MachOObjectFile &Obj, Triple TT, 50 LinkGraph::GetEdgeKindNameFunction GetEdgeKindName) 51 : Obj(Obj), 52 G(std::make_unique<LinkGraph>( 53 std::string(Obj.getFileName()), std::move(TT), getPointerSize(Obj), 54 getEndianness(Obj), std::move(GetEdgeKindName))) { 55 auto &MachHeader = Obj.getHeader64(); 56 SubsectionsViaSymbols = MachHeader.flags & MachO::MH_SUBSECTIONS_VIA_SYMBOLS; 57 } 58 59 void MachOLinkGraphBuilder::addCustomSectionParser( 60 StringRef SectionName, SectionParserFunction Parser) { 61 assert(!CustomSectionParserFunctions.count(SectionName) && 62 "Custom parser for this section already exists"); 63 CustomSectionParserFunctions[SectionName] = std::move(Parser); 64 } 65 66 Linkage MachOLinkGraphBuilder::getLinkage(uint16_t Desc) { 67 if ((Desc & MachO::N_WEAK_DEF) || (Desc & MachO::N_WEAK_REF)) 68 return Linkage::Weak; 69 return Linkage::Strong; 70 } 71 72 Scope MachOLinkGraphBuilder::getScope(StringRef Name, uint8_t Type) { 73 if (Type & MachO::N_EXT) { 74 if ((Type & MachO::N_PEXT) || Name.startswith("l")) 75 return Scope::Hidden; 76 else 77 return Scope::Default; 78 } 79 return Scope::Local; 80 } 81 82 bool MachOLinkGraphBuilder::isAltEntry(const NormalizedSymbol &NSym) { 83 return NSym.Desc & MachO::N_ALT_ENTRY; 84 } 85 86 bool MachOLinkGraphBuilder::isDebugSection(const NormalizedSection &NSec) { 87 return (NSec.Flags & MachO::S_ATTR_DEBUG && 88 strcmp(NSec.SegName, "__DWARF") == 0); 89 } 90 91 bool MachOLinkGraphBuilder::isZeroFillSection(const NormalizedSection &NSec) { 92 switch (NSec.Flags & MachO::SECTION_TYPE) { 93 case MachO::S_ZEROFILL: 94 case MachO::S_GB_ZEROFILL: 95 case MachO::S_THREAD_LOCAL_ZEROFILL: 96 return true; 97 default: 98 return false; 99 } 100 } 101 102 unsigned 103 MachOLinkGraphBuilder::getPointerSize(const object::MachOObjectFile &Obj) { 104 return Obj.is64Bit() ? 8 : 4; 105 } 106 107 support::endianness 108 MachOLinkGraphBuilder::getEndianness(const object::MachOObjectFile &Obj) { 109 return Obj.isLittleEndian() ? support::little : support::big; 110 } 111 112 Section &MachOLinkGraphBuilder::getCommonSection() { 113 if (!CommonSection) 114 CommonSection = 115 &G->createSection(CommonSectionName, MemProt::Read | MemProt::Write); 116 return *CommonSection; 117 } 118 119 Error MachOLinkGraphBuilder::createNormalizedSections() { 120 // Build normalized sections. Verifies that section data is in-range (for 121 // sections with content) and that address ranges are non-overlapping. 122 123 LLVM_DEBUG(dbgs() << "Creating normalized sections...\n"); 124 125 for (auto &SecRef : Obj.sections()) { 126 NormalizedSection NSec; 127 uint32_t DataOffset = 0; 128 129 auto SecIndex = Obj.getSectionIndex(SecRef.getRawDataRefImpl()); 130 131 if (Obj.is64Bit()) { 132 const MachO::section_64 &Sec64 = 133 Obj.getSection64(SecRef.getRawDataRefImpl()); 134 135 memcpy(&NSec.SectName, &Sec64.sectname, 16); 136 NSec.SectName[16] = '\0'; 137 memcpy(&NSec.SegName, Sec64.segname, 16); 138 NSec.SegName[16] = '\0'; 139 140 NSec.Address = orc::ExecutorAddr(Sec64.addr); 141 NSec.Size = Sec64.size; 142 NSec.Alignment = 1ULL << Sec64.align; 143 NSec.Flags = Sec64.flags; 144 DataOffset = Sec64.offset; 145 } else { 146 const MachO::section &Sec32 = Obj.getSection(SecRef.getRawDataRefImpl()); 147 148 memcpy(&NSec.SectName, &Sec32.sectname, 16); 149 NSec.SectName[16] = '\0'; 150 memcpy(&NSec.SegName, Sec32.segname, 16); 151 NSec.SegName[16] = '\0'; 152 153 NSec.Address = orc::ExecutorAddr(Sec32.addr); 154 NSec.Size = Sec32.size; 155 NSec.Alignment = 1ULL << Sec32.align; 156 NSec.Flags = Sec32.flags; 157 DataOffset = Sec32.offset; 158 } 159 160 LLVM_DEBUG({ 161 dbgs() << " " << NSec.SegName << "," << NSec.SectName << ": " 162 << formatv("{0:x16}", NSec.Address) << " -- " 163 << formatv("{0:x16}", NSec.Address + NSec.Size) 164 << ", align: " << NSec.Alignment << ", index: " << SecIndex 165 << "\n"; 166 }); 167 168 // Get the section data if any. 169 if (!isZeroFillSection(NSec)) { 170 if (DataOffset + NSec.Size > Obj.getData().size()) 171 return make_error<JITLinkError>( 172 "Section data extends past end of file"); 173 174 NSec.Data = Obj.getData().data() + DataOffset; 175 } 176 177 // Get prot flags. 178 // FIXME: Make sure this test is correct (it's probably missing cases 179 // as-is). 180 MemProt Prot; 181 if (NSec.Flags & MachO::S_ATTR_PURE_INSTRUCTIONS) 182 Prot = MemProt::Read | MemProt::Exec; 183 else 184 Prot = MemProt::Read | MemProt::Write; 185 186 auto FullyQualifiedName = 187 G->allocateString(StringRef(NSec.SegName) + "," + NSec.SectName); 188 NSec.GraphSection = &G->createSection( 189 StringRef(FullyQualifiedName.data(), FullyQualifiedName.size()), Prot); 190 191 IndexToSection.insert(std::make_pair(SecIndex, std::move(NSec))); 192 } 193 194 std::vector<NormalizedSection *> Sections; 195 Sections.reserve(IndexToSection.size()); 196 for (auto &KV : IndexToSection) 197 Sections.push_back(&KV.second); 198 199 // If we didn't end up creating any sections then bail out. The code below 200 // assumes that we have at least one section. 201 if (Sections.empty()) 202 return Error::success(); 203 204 llvm::sort(Sections, 205 [](const NormalizedSection *LHS, const NormalizedSection *RHS) { 206 assert(LHS && RHS && "Null section?"); 207 if (LHS->Address != RHS->Address) 208 return LHS->Address < RHS->Address; 209 return LHS->Size < RHS->Size; 210 }); 211 212 for (unsigned I = 0, E = Sections.size() - 1; I != E; ++I) { 213 auto &Cur = *Sections[I]; 214 auto &Next = *Sections[I + 1]; 215 if (Next.Address < Cur.Address + Cur.Size) 216 return make_error<JITLinkError>( 217 "Address range for section " + 218 formatv("\"{0}/{1}\" [ {2:x16} -- {3:x16} ] ", Cur.SegName, 219 Cur.SectName, Cur.Address, Cur.Address + Cur.Size) + 220 "overlaps section \"" + Next.SegName + "/" + Next.SectName + "\"" + 221 formatv("\"{0}/{1}\" [ {2:x16} -- {3:x16} ] ", Next.SegName, 222 Next.SectName, Next.Address, Next.Address + Next.Size)); 223 } 224 225 return Error::success(); 226 } 227 228 Error MachOLinkGraphBuilder::createNormalizedSymbols() { 229 LLVM_DEBUG(dbgs() << "Creating normalized symbols...\n"); 230 231 for (auto &SymRef : Obj.symbols()) { 232 233 unsigned SymbolIndex = Obj.getSymbolIndex(SymRef.getRawDataRefImpl()); 234 uint64_t Value; 235 uint32_t NStrX; 236 uint8_t Type; 237 uint8_t Sect; 238 uint16_t Desc; 239 240 if (Obj.is64Bit()) { 241 const MachO::nlist_64 &NL64 = 242 Obj.getSymbol64TableEntry(SymRef.getRawDataRefImpl()); 243 Value = NL64.n_value; 244 NStrX = NL64.n_strx; 245 Type = NL64.n_type; 246 Sect = NL64.n_sect; 247 Desc = NL64.n_desc; 248 } else { 249 const MachO::nlist &NL32 = 250 Obj.getSymbolTableEntry(SymRef.getRawDataRefImpl()); 251 Value = NL32.n_value; 252 NStrX = NL32.n_strx; 253 Type = NL32.n_type; 254 Sect = NL32.n_sect; 255 Desc = NL32.n_desc; 256 } 257 258 // Skip stabs. 259 // FIXME: Are there other symbols we should be skipping? 260 if (Type & MachO::N_STAB) 261 continue; 262 263 Optional<StringRef> Name; 264 if (NStrX) { 265 if (auto NameOrErr = SymRef.getName()) 266 Name = *NameOrErr; 267 else 268 return NameOrErr.takeError(); 269 } 270 271 LLVM_DEBUG({ 272 dbgs() << " "; 273 if (!Name) 274 dbgs() << "<anonymous symbol>"; 275 else 276 dbgs() << *Name; 277 dbgs() << ": value = " << formatv("{0:x16}", Value) 278 << ", type = " << formatv("{0:x2}", Type) 279 << ", desc = " << formatv("{0:x4}", Desc) << ", sect = "; 280 if (Sect) 281 dbgs() << static_cast<unsigned>(Sect - 1); 282 else 283 dbgs() << "none"; 284 dbgs() << "\n"; 285 }); 286 287 // If this symbol has a section, verify that the addresses line up. 288 if (Sect != 0) { 289 auto NSec = findSectionByIndex(Sect - 1); 290 if (!NSec) 291 return NSec.takeError(); 292 293 if (orc::ExecutorAddr(Value) < NSec->Address || 294 orc::ExecutorAddr(Value) > NSec->Address + NSec->Size) 295 return make_error<JITLinkError>("Address " + formatv("{0:x}", Value) + 296 " for symbol " + *Name + 297 " does not fall within section"); 298 299 if (!NSec->GraphSection) { 300 LLVM_DEBUG({ 301 dbgs() << " Skipping: Symbol is in section " << NSec->SegName << "/" 302 << NSec->SectName 303 << " which has no associated graph section.\n"; 304 }); 305 continue; 306 } 307 } 308 309 IndexToSymbol[SymbolIndex] = 310 &createNormalizedSymbol(*Name, Value, Type, Sect, Desc, 311 getLinkage(Desc), getScope(*Name, Type)); 312 } 313 314 return Error::success(); 315 } 316 317 void MachOLinkGraphBuilder::addSectionStartSymAndBlock( 318 unsigned SecIndex, Section &GraphSec, orc::ExecutorAddr Address, 319 const char *Data, orc::ExecutorAddrDiff Size, uint32_t Alignment, 320 bool IsLive) { 321 Block &B = 322 Data ? G->createContentBlock(GraphSec, ArrayRef<char>(Data, Size), 323 Address, Alignment, 0) 324 : G->createZeroFillBlock(GraphSec, Size, Address, Alignment, 0); 325 auto &Sym = G->addAnonymousSymbol(B, 0, Size, false, IsLive); 326 auto SecI = IndexToSection.find(SecIndex); 327 assert(SecI != IndexToSection.end() && "SecIndex invalid"); 328 auto &NSec = SecI->second; 329 assert(!NSec.CanonicalSymbols.count(Sym.getAddress()) && 330 "Anonymous block start symbol clashes with existing symbol address"); 331 NSec.CanonicalSymbols[Sym.getAddress()] = &Sym; 332 } 333 334 Error MachOLinkGraphBuilder::graphifyRegularSymbols() { 335 336 LLVM_DEBUG(dbgs() << "Creating graph symbols...\n"); 337 338 /// We only have 256 section indexes: Use a vector rather than a map. 339 std::vector<std::vector<NormalizedSymbol *>> SecIndexToSymbols; 340 SecIndexToSymbols.resize(256); 341 342 // Create commons, externs, and absolutes, and partition all other symbols by 343 // section. 344 for (auto &KV : IndexToSymbol) { 345 auto &NSym = *KV.second; 346 347 switch (NSym.Type & MachO::N_TYPE) { 348 case MachO::N_UNDF: 349 if (NSym.Value) { 350 if (!NSym.Name) 351 return make_error<JITLinkError>("Anonymous common symbol at index " + 352 Twine(KV.first)); 353 NSym.GraphSymbol = &G->addCommonSymbol( 354 *NSym.Name, NSym.S, getCommonSection(), orc::ExecutorAddr(), 355 orc::ExecutorAddrDiff(NSym.Value), 356 1ull << MachO::GET_COMM_ALIGN(NSym.Desc), 357 NSym.Desc & MachO::N_NO_DEAD_STRIP); 358 } else { 359 if (!NSym.Name) 360 return make_error<JITLinkError>("Anonymous external symbol at " 361 "index " + 362 Twine(KV.first)); 363 NSym.GraphSymbol = &G->addExternalSymbol( 364 *NSym.Name, 0, 365 NSym.Desc & MachO::N_WEAK_REF ? Linkage::Weak : Linkage::Strong); 366 } 367 break; 368 case MachO::N_ABS: 369 if (!NSym.Name) 370 return make_error<JITLinkError>("Anonymous absolute symbol at index " + 371 Twine(KV.first)); 372 NSym.GraphSymbol = &G->addAbsoluteSymbol( 373 *NSym.Name, orc::ExecutorAddr(NSym.Value), 0, Linkage::Strong, 374 getScope(*NSym.Name, NSym.Type), NSym.Desc & MachO::N_NO_DEAD_STRIP); 375 break; 376 case MachO::N_SECT: 377 SecIndexToSymbols[NSym.Sect - 1].push_back(&NSym); 378 break; 379 case MachO::N_PBUD: 380 return make_error<JITLinkError>( 381 "Unupported N_PBUD symbol " + 382 (NSym.Name ? ("\"" + *NSym.Name + "\"") : Twine("<anon>")) + 383 " at index " + Twine(KV.first)); 384 case MachO::N_INDR: 385 return make_error<JITLinkError>( 386 "Unupported N_INDR symbol " + 387 (NSym.Name ? ("\"" + *NSym.Name + "\"") : Twine("<anon>")) + 388 " at index " + Twine(KV.first)); 389 default: 390 return make_error<JITLinkError>( 391 "Unrecognized symbol type " + Twine(NSym.Type & MachO::N_TYPE) + 392 " for symbol " + 393 (NSym.Name ? ("\"" + *NSym.Name + "\"") : Twine("<anon>")) + 394 " at index " + Twine(KV.first)); 395 } 396 } 397 398 // Loop over sections performing regular graphification for those that 399 // don't have custom parsers. 400 for (auto &KV : IndexToSection) { 401 auto SecIndex = KV.first; 402 auto &NSec = KV.second; 403 404 if (!NSec.GraphSection) { 405 LLVM_DEBUG({ 406 dbgs() << " " << NSec.SegName << "/" << NSec.SectName 407 << " has no graph section. Skipping.\n"; 408 }); 409 continue; 410 } 411 412 // Skip sections with custom parsers. 413 if (CustomSectionParserFunctions.count(NSec.GraphSection->getName())) { 414 LLVM_DEBUG({ 415 dbgs() << " Skipping section " << NSec.GraphSection->getName() 416 << " as it has a custom parser.\n"; 417 }); 418 continue; 419 } else if ((NSec.Flags & MachO::SECTION_TYPE) == 420 MachO::S_CSTRING_LITERALS) { 421 if (auto Err = graphifyCStringSection( 422 NSec, std::move(SecIndexToSymbols[SecIndex]))) 423 return Err; 424 continue; 425 } else 426 LLVM_DEBUG({ 427 dbgs() << " Graphifying regular section " 428 << NSec.GraphSection->getName() << "...\n"; 429 }); 430 431 bool SectionIsNoDeadStrip = NSec.Flags & MachO::S_ATTR_NO_DEAD_STRIP; 432 bool SectionIsText = NSec.Flags & MachO::S_ATTR_PURE_INSTRUCTIONS; 433 434 auto &SecNSymStack = SecIndexToSymbols[SecIndex]; 435 436 // If this section is non-empty but there are no symbols covering it then 437 // create one block and anonymous symbol to cover the entire section. 438 if (SecNSymStack.empty()) { 439 if (NSec.Size > 0) { 440 LLVM_DEBUG({ 441 dbgs() << " Section non-empty, but contains no symbols. " 442 "Creating anonymous block to cover " 443 << formatv("{0:x16}", NSec.Address) << " -- " 444 << formatv("{0:x16}", NSec.Address + NSec.Size) << "\n"; 445 }); 446 addSectionStartSymAndBlock(SecIndex, *NSec.GraphSection, NSec.Address, 447 NSec.Data, NSec.Size, NSec.Alignment, 448 SectionIsNoDeadStrip); 449 } else 450 LLVM_DEBUG({ 451 dbgs() << " Section empty and contains no symbols. Skipping.\n"; 452 }); 453 continue; 454 } 455 456 // Sort the symbol stack in by address, alt-entry status, scope, and name. 457 // We sort in reverse order so that symbols will be visited in the right 458 // order when we pop off the stack below. 459 llvm::sort(SecNSymStack, [](const NormalizedSymbol *LHS, 460 const NormalizedSymbol *RHS) { 461 if (LHS->Value != RHS->Value) 462 return LHS->Value > RHS->Value; 463 if (isAltEntry(*LHS) != isAltEntry(*RHS)) 464 return isAltEntry(*RHS); 465 if (LHS->S != RHS->S) 466 return static_cast<uint8_t>(LHS->S) < static_cast<uint8_t>(RHS->S); 467 return LHS->Name < RHS->Name; 468 }); 469 470 // The first symbol in a section can not be an alt-entry symbol. 471 if (!SecNSymStack.empty() && isAltEntry(*SecNSymStack.back())) 472 return make_error<JITLinkError>( 473 "First symbol in " + NSec.GraphSection->getName() + " is alt-entry"); 474 475 // If the section is non-empty but there is no symbol covering the start 476 // address then add an anonymous one. 477 if (orc::ExecutorAddr(SecNSymStack.back()->Value) != NSec.Address) { 478 auto AnonBlockSize = 479 orc::ExecutorAddr(SecNSymStack.back()->Value) - NSec.Address; 480 LLVM_DEBUG({ 481 dbgs() << " Section start not covered by symbol. " 482 << "Creating anonymous block to cover [ " << NSec.Address 483 << " -- " << (NSec.Address + AnonBlockSize) << " ]\n"; 484 }); 485 addSectionStartSymAndBlock(SecIndex, *NSec.GraphSection, NSec.Address, 486 NSec.Data, AnonBlockSize, NSec.Alignment, 487 SectionIsNoDeadStrip); 488 } 489 490 // Visit section symbols in order by popping off the reverse-sorted stack, 491 // building graph symbols as we go. 492 // 493 // If MH_SUBSECTIONS_VIA_SYMBOLS is set we'll build a block for each 494 // alt-entry chain. 495 // 496 // If MH_SUBSECTIONS_VIA_SYMBOLS is not set then we'll just build one block 497 // for the whole section. 498 while (!SecNSymStack.empty()) { 499 SmallVector<NormalizedSymbol *, 8> BlockSyms; 500 501 // Get the symbols in this alt-entry chain, or the whole section (if 502 // !SubsectionsViaSymbols). 503 BlockSyms.push_back(SecNSymStack.back()); 504 SecNSymStack.pop_back(); 505 while (!SecNSymStack.empty() && 506 (isAltEntry(*SecNSymStack.back()) || 507 SecNSymStack.back()->Value == BlockSyms.back()->Value || 508 !SubsectionsViaSymbols)) { 509 BlockSyms.push_back(SecNSymStack.back()); 510 SecNSymStack.pop_back(); 511 } 512 513 // BlockNSyms now contains the block symbols in reverse canonical order. 514 auto BlockStart = orc::ExecutorAddr(BlockSyms.front()->Value); 515 orc::ExecutorAddr BlockEnd = 516 SecNSymStack.empty() ? NSec.Address + NSec.Size 517 : orc::ExecutorAddr(SecNSymStack.back()->Value); 518 orc::ExecutorAddrDiff BlockOffset = BlockStart - NSec.Address; 519 orc::ExecutorAddrDiff BlockSize = BlockEnd - BlockStart; 520 521 LLVM_DEBUG({ 522 dbgs() << " Creating block for " << formatv("{0:x16}", BlockStart) 523 << " -- " << formatv("{0:x16}", BlockEnd) << ": " 524 << NSec.GraphSection->getName() << " + " 525 << formatv("{0:x16}", BlockOffset) << " with " 526 << BlockSyms.size() << " symbol(s)...\n"; 527 }); 528 529 Block &B = 530 NSec.Data 531 ? G->createContentBlock( 532 *NSec.GraphSection, 533 ArrayRef<char>(NSec.Data + BlockOffset, BlockSize), 534 BlockStart, NSec.Alignment, BlockStart % NSec.Alignment) 535 : G->createZeroFillBlock(*NSec.GraphSection, BlockSize, 536 BlockStart, NSec.Alignment, 537 BlockStart % NSec.Alignment); 538 539 Optional<orc::ExecutorAddr> LastCanonicalAddr; 540 auto SymEnd = BlockEnd; 541 while (!BlockSyms.empty()) { 542 auto &NSym = *BlockSyms.back(); 543 BlockSyms.pop_back(); 544 545 bool SymLive = 546 (NSym.Desc & MachO::N_NO_DEAD_STRIP) || SectionIsNoDeadStrip; 547 548 auto &Sym = createStandardGraphSymbol( 549 NSym, B, SymEnd - orc::ExecutorAddr(NSym.Value), SectionIsText, 550 SymLive, LastCanonicalAddr != orc::ExecutorAddr(NSym.Value)); 551 552 if (LastCanonicalAddr != Sym.getAddress()) { 553 if (LastCanonicalAddr) 554 SymEnd = *LastCanonicalAddr; 555 LastCanonicalAddr = Sym.getAddress(); 556 } 557 } 558 } 559 } 560 561 return Error::success(); 562 } 563 564 Symbol &MachOLinkGraphBuilder::createStandardGraphSymbol(NormalizedSymbol &NSym, 565 Block &B, size_t Size, 566 bool IsText, 567 bool IsNoDeadStrip, 568 bool IsCanonical) { 569 570 LLVM_DEBUG({ 571 dbgs() << " " << formatv("{0:x16}", NSym.Value) << " -- " 572 << formatv("{0:x16}", NSym.Value + Size) << ": "; 573 if (!NSym.Name) 574 dbgs() << "<anonymous symbol>"; 575 else 576 dbgs() << NSym.Name; 577 if (IsText) 578 dbgs() << " [text]"; 579 if (IsNoDeadStrip) 580 dbgs() << " [no-dead-strip]"; 581 if (!IsCanonical) 582 dbgs() << " [non-canonical]"; 583 dbgs() << "\n"; 584 }); 585 586 auto SymOffset = orc::ExecutorAddr(NSym.Value) - B.getAddress(); 587 auto &Sym = 588 NSym.Name 589 ? G->addDefinedSymbol(B, SymOffset, *NSym.Name, Size, NSym.L, NSym.S, 590 IsText, IsNoDeadStrip) 591 : G->addAnonymousSymbol(B, SymOffset, Size, IsText, IsNoDeadStrip); 592 NSym.GraphSymbol = &Sym; 593 594 if (IsCanonical) 595 setCanonicalSymbol(getSectionByIndex(NSym.Sect - 1), Sym); 596 597 return Sym; 598 } 599 600 Error MachOLinkGraphBuilder::graphifySectionsWithCustomParsers() { 601 // Graphify special sections. 602 for (auto &KV : IndexToSection) { 603 auto &NSec = KV.second; 604 605 // Skip non-graph sections. 606 if (!NSec.GraphSection) 607 continue; 608 609 auto HI = CustomSectionParserFunctions.find(NSec.GraphSection->getName()); 610 if (HI != CustomSectionParserFunctions.end()) { 611 auto &Parse = HI->second; 612 if (auto Err = Parse(NSec)) 613 return Err; 614 } 615 } 616 617 return Error::success(); 618 } 619 620 Error MachOLinkGraphBuilder::graphifyCStringSection( 621 NormalizedSection &NSec, std::vector<NormalizedSymbol *> NSyms) { 622 assert(NSec.GraphSection && "C string literal section missing graph section"); 623 assert(NSec.Data && "C string literal section has no data"); 624 625 LLVM_DEBUG({ 626 dbgs() << " Graphifying C-string literal section " 627 << NSec.GraphSection->getName() << "\n"; 628 }); 629 630 if (NSec.Data[NSec.Size - 1] != '\0') 631 return make_error<JITLinkError>("C string literal section " + 632 NSec.GraphSection->getName() + 633 " does not end with null terminator"); 634 635 /// Sort into reverse order to use as a stack. 636 llvm::sort(NSyms, 637 [](const NormalizedSymbol *LHS, const NormalizedSymbol *RHS) { 638 if (LHS->Value != RHS->Value) 639 return LHS->Value > RHS->Value; 640 if (LHS->L != RHS->L) 641 return LHS->L > RHS->L; 642 if (LHS->S != RHS->S) 643 return LHS->S > RHS->S; 644 if (RHS->Name) { 645 if (!LHS->Name) 646 return true; 647 return *LHS->Name > *RHS->Name; 648 } 649 return false; 650 }); 651 652 bool SectionIsNoDeadStrip = NSec.Flags & MachO::S_ATTR_NO_DEAD_STRIP; 653 bool SectionIsText = NSec.Flags & MachO::S_ATTR_PURE_INSTRUCTIONS; 654 orc::ExecutorAddrDiff BlockStart = 0; 655 656 // Scan section for null characters. 657 for (size_t I = 0; I != NSec.Size; ++I) 658 if (NSec.Data[I] == '\0') { 659 size_t BlockSize = I + 1 - BlockStart; 660 // Create a block for this null terminated string. 661 auto &B = G->createContentBlock(*NSec.GraphSection, 662 {NSec.Data + BlockStart, BlockSize}, 663 NSec.Address + BlockStart, NSec.Alignment, 664 BlockStart % NSec.Alignment); 665 666 LLVM_DEBUG({ 667 dbgs() << " Created block " << B.getRange() 668 << ", align = " << B.getAlignment() 669 << ", align-ofs = " << B.getAlignmentOffset() << " for \""; 670 for (size_t J = 0; J != std::min(B.getSize(), size_t(16)); ++J) 671 switch (B.getContent()[J]) { 672 case '\0': break; 673 case '\n': dbgs() << "\\n"; break; 674 case '\t': dbgs() << "\\t"; break; 675 default: dbgs() << B.getContent()[J]; break; 676 } 677 if (B.getSize() > 16) 678 dbgs() << "..."; 679 dbgs() << "\"\n"; 680 }); 681 682 // If there's no symbol at the start of this block then create one. 683 if (NSyms.empty() || 684 orc::ExecutorAddr(NSyms.back()->Value) != B.getAddress()) { 685 auto &S = G->addAnonymousSymbol(B, 0, BlockSize, false, false); 686 setCanonicalSymbol(NSec, S); 687 LLVM_DEBUG({ 688 dbgs() << " Adding symbol for c-string block " << B.getRange() 689 << ": <anonymous symbol> at offset 0\n"; 690 }); 691 } 692 693 // Process any remaining symbols that point into this block. 694 auto LastCanonicalAddr = B.getAddress() + BlockSize; 695 while (!NSyms.empty() && orc::ExecutorAddr(NSyms.back()->Value) < 696 B.getAddress() + BlockSize) { 697 auto &NSym = *NSyms.back(); 698 size_t SymSize = (B.getAddress() + BlockSize) - 699 orc::ExecutorAddr(NSyms.back()->Value); 700 bool SymLive = 701 (NSym.Desc & MachO::N_NO_DEAD_STRIP) || SectionIsNoDeadStrip; 702 703 bool IsCanonical = false; 704 if (LastCanonicalAddr != orc::ExecutorAddr(NSym.Value)) { 705 IsCanonical = true; 706 LastCanonicalAddr = orc::ExecutorAddr(NSym.Value); 707 } 708 709 auto &Sym = createStandardGraphSymbol(NSym, B, SymSize, SectionIsText, 710 SymLive, IsCanonical); 711 (void)Sym; 712 LLVM_DEBUG({ 713 dbgs() << " Adding symbol for c-string block " << B.getRange() 714 << ": " 715 << (Sym.hasName() ? Sym.getName() : "<anonymous symbol>") 716 << " at offset " << formatv("{0:x}", Sym.getOffset()) << "\n"; 717 }); 718 719 NSyms.pop_back(); 720 } 721 722 BlockStart += BlockSize; 723 } 724 725 return Error::success(); 726 } 727 728 Error CompactUnwindSplitter::operator()(LinkGraph &G) { 729 auto *CUSec = G.findSectionByName(CompactUnwindSectionName); 730 if (!CUSec) 731 return Error::success(); 732 733 if (!G.getTargetTriple().isOSBinFormatMachO()) 734 return make_error<JITLinkError>( 735 "Error linking " + G.getName() + 736 ": compact unwind splitting not supported on non-macho target " + 737 G.getTargetTriple().str()); 738 739 unsigned CURecordSize = 0; 740 unsigned PersonalityEdgeOffset = 0; 741 unsigned LSDAEdgeOffset = 0; 742 switch (G.getTargetTriple().getArch()) { 743 case Triple::aarch64: 744 case Triple::x86_64: 745 // 64-bit compact-unwind record format: 746 // Range start: 8 bytes. 747 // Range size: 4 bytes. 748 // CU encoding: 4 bytes. 749 // Personality: 8 bytes. 750 // LSDA: 8 bytes. 751 CURecordSize = 32; 752 PersonalityEdgeOffset = 16; 753 LSDAEdgeOffset = 24; 754 break; 755 default: 756 return make_error<JITLinkError>( 757 "Error linking " + G.getName() + 758 ": compact unwind splitting not supported on " + 759 G.getTargetTriple().getArchName()); 760 } 761 762 std::vector<Block *> OriginalBlocks(CUSec->blocks().begin(), 763 CUSec->blocks().end()); 764 LLVM_DEBUG({ 765 dbgs() << "In " << G.getName() << " splitting compact unwind section " 766 << CompactUnwindSectionName << " containing " 767 << OriginalBlocks.size() << " initial blocks...\n"; 768 }); 769 770 while (!OriginalBlocks.empty()) { 771 auto *B = OriginalBlocks.back(); 772 OriginalBlocks.pop_back(); 773 774 if (B->getSize() == 0) { 775 LLVM_DEBUG({ 776 dbgs() << " Skipping empty block at " 777 << formatv("{0:x16}", B->getAddress()) << "\n"; 778 }); 779 continue; 780 } 781 782 LLVM_DEBUG({ 783 dbgs() << " Splitting block at " << formatv("{0:x16}", B->getAddress()) 784 << " into " << (B->getSize() / CURecordSize) 785 << " compact unwind record(s)\n"; 786 }); 787 788 if (B->getSize() % CURecordSize) 789 return make_error<JITLinkError>( 790 "Error splitting compact unwind record in " + G.getName() + 791 ": block at " + formatv("{0:x}", B->getAddress()) + " has size " + 792 formatv("{0:x}", B->getSize()) + 793 " (not a multiple of CU record size of " + 794 formatv("{0:x}", CURecordSize) + ")"); 795 796 unsigned NumBlocks = B->getSize() / CURecordSize; 797 LinkGraph::SplitBlockCache C; 798 799 for (unsigned I = 0; I != NumBlocks; ++I) { 800 auto &CURec = G.splitBlock(*B, CURecordSize, &C); 801 bool AddedKeepAlive = false; 802 803 for (auto &E : CURec.edges()) { 804 if (E.getOffset() == 0) { 805 LLVM_DEBUG({ 806 dbgs() << " Updating compact unwind record at " 807 << formatv("{0:x16}", CURec.getAddress()) << " to point to " 808 << (E.getTarget().hasName() ? E.getTarget().getName() 809 : StringRef()) 810 << " (at " << formatv("{0:x16}", E.getTarget().getAddress()) 811 << ")\n"; 812 }); 813 814 if (E.getTarget().isExternal()) 815 return make_error<JITLinkError>( 816 "Error adding keep-alive edge for compact unwind record at " + 817 formatv("{0:x}", CURec.getAddress()) + ": target " + 818 E.getTarget().getName() + " is an external symbol"); 819 auto &TgtBlock = E.getTarget().getBlock(); 820 auto &CURecSym = 821 G.addAnonymousSymbol(CURec, 0, CURecordSize, false, false); 822 TgtBlock.addEdge(Edge::KeepAlive, 0, CURecSym, 0); 823 AddedKeepAlive = true; 824 } else if (E.getOffset() != PersonalityEdgeOffset && 825 E.getOffset() != LSDAEdgeOffset) 826 return make_error<JITLinkError>("Unexpected edge at offset " + 827 formatv("{0:x}", E.getOffset()) + 828 " in compact unwind record at " + 829 formatv("{0:x}", CURec.getAddress())); 830 } 831 832 if (!AddedKeepAlive) 833 return make_error<JITLinkError>( 834 "Error adding keep-alive edge for compact unwind record at " + 835 formatv("{0:x}", CURec.getAddress()) + 836 ": no outgoing target edge at offset 0"); 837 } 838 } 839 return Error::success(); 840 } 841 842 } // end namespace jitlink 843 } // end namespace llvm 844