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 #include <optional> 15 16 #define DEBUG_TYPE "jitlink" 17 18 static const char *CommonSectionName = "__common"; 19 20 namespace llvm { 21 namespace jitlink { 22 23 MachOLinkGraphBuilder::~MachOLinkGraphBuilder() = default; 24 25 Expected<std::unique_ptr<LinkGraph>> MachOLinkGraphBuilder::buildGraph() { 26 27 // We only operate on relocatable objects. 28 if (!Obj.isRelocatableObject()) 29 return make_error<JITLinkError>("Object is not a relocatable MachO"); 30 31 if (auto Err = createNormalizedSections()) 32 return std::move(Err); 33 34 if (auto Err = createNormalizedSymbols()) 35 return std::move(Err); 36 37 if (auto Err = graphifyRegularSymbols()) 38 return std::move(Err); 39 40 if (auto Err = graphifySectionsWithCustomParsers()) 41 return std::move(Err); 42 43 if (auto Err = addRelocations()) 44 return std::move(Err); 45 46 return std::move(G); 47 } 48 49 MachOLinkGraphBuilder::MachOLinkGraphBuilder( 50 const object::MachOObjectFile &Obj, Triple TT, 51 LinkGraph::GetEdgeKindNameFunction GetEdgeKindName) 52 : Obj(Obj), 53 G(std::make_unique<LinkGraph>( 54 std::string(Obj.getFileName()), std::move(TT), getPointerSize(Obj), 55 getEndianness(Obj), std::move(GetEdgeKindName))) { 56 auto &MachHeader = Obj.getHeader64(); 57 SubsectionsViaSymbols = MachHeader.flags & MachO::MH_SUBSECTIONS_VIA_SYMBOLS; 58 } 59 60 void MachOLinkGraphBuilder::addCustomSectionParser( 61 StringRef SectionName, SectionParserFunction Parser) { 62 assert(!CustomSectionParserFunctions.count(SectionName) && 63 "Custom parser for this section already exists"); 64 CustomSectionParserFunctions[SectionName] = std::move(Parser); 65 } 66 67 Linkage MachOLinkGraphBuilder::getLinkage(uint16_t Desc) { 68 if ((Desc & MachO::N_WEAK_DEF) || (Desc & MachO::N_WEAK_REF)) 69 return Linkage::Weak; 70 return Linkage::Strong; 71 } 72 73 Scope MachOLinkGraphBuilder::getScope(StringRef Name, uint8_t Type) { 74 if (Type & MachO::N_EXT) { 75 if ((Type & MachO::N_PEXT) || Name.startswith("l")) 76 return Scope::Hidden; 77 else 78 return Scope::Default; 79 } 80 return Scope::Local; 81 } 82 83 bool MachOLinkGraphBuilder::isAltEntry(const NormalizedSymbol &NSym) { 84 return NSym.Desc & MachO::N_ALT_ENTRY; 85 } 86 87 bool MachOLinkGraphBuilder::isDebugSection(const NormalizedSection &NSec) { 88 return (NSec.Flags & MachO::S_ATTR_DEBUG && 89 strcmp(NSec.SegName, "__DWARF") == 0); 90 } 91 92 bool MachOLinkGraphBuilder::isZeroFillSection(const NormalizedSection &NSec) { 93 switch (NSec.Flags & MachO::SECTION_TYPE) { 94 case MachO::S_ZEROFILL: 95 case MachO::S_GB_ZEROFILL: 96 case MachO::S_THREAD_LOCAL_ZEROFILL: 97 return true; 98 default: 99 return false; 100 } 101 } 102 103 unsigned 104 MachOLinkGraphBuilder::getPointerSize(const object::MachOObjectFile &Obj) { 105 return Obj.is64Bit() ? 8 : 4; 106 } 107 108 support::endianness 109 MachOLinkGraphBuilder::getEndianness(const object::MachOObjectFile &Obj) { 110 return Obj.isLittleEndian() ? support::little : support::big; 111 } 112 113 Section &MachOLinkGraphBuilder::getCommonSection() { 114 if (!CommonSection) 115 CommonSection = &G->createSection(CommonSectionName, 116 orc::MemProt::Read | orc::MemProt::Write); 117 return *CommonSection; 118 } 119 120 Error MachOLinkGraphBuilder::createNormalizedSections() { 121 // Build normalized sections. Verifies that section data is in-range (for 122 // sections with content) and that address ranges are non-overlapping. 123 124 LLVM_DEBUG(dbgs() << "Creating normalized sections...\n"); 125 126 for (auto &SecRef : Obj.sections()) { 127 NormalizedSection NSec; 128 uint32_t DataOffset = 0; 129 130 auto SecIndex = Obj.getSectionIndex(SecRef.getRawDataRefImpl()); 131 132 if (Obj.is64Bit()) { 133 const MachO::section_64 &Sec64 = 134 Obj.getSection64(SecRef.getRawDataRefImpl()); 135 136 memcpy(&NSec.SectName, &Sec64.sectname, 16); 137 NSec.SectName[16] = '\0'; 138 memcpy(&NSec.SegName, Sec64.segname, 16); 139 NSec.SegName[16] = '\0'; 140 141 NSec.Address = orc::ExecutorAddr(Sec64.addr); 142 NSec.Size = Sec64.size; 143 NSec.Alignment = 1ULL << Sec64.align; 144 NSec.Flags = Sec64.flags; 145 DataOffset = Sec64.offset; 146 } else { 147 const MachO::section &Sec32 = Obj.getSection(SecRef.getRawDataRefImpl()); 148 149 memcpy(&NSec.SectName, &Sec32.sectname, 16); 150 NSec.SectName[16] = '\0'; 151 memcpy(&NSec.SegName, Sec32.segname, 16); 152 NSec.SegName[16] = '\0'; 153 154 NSec.Address = orc::ExecutorAddr(Sec32.addr); 155 NSec.Size = Sec32.size; 156 NSec.Alignment = 1ULL << Sec32.align; 157 NSec.Flags = Sec32.flags; 158 DataOffset = Sec32.offset; 159 } 160 161 LLVM_DEBUG({ 162 dbgs() << " " << NSec.SegName << "," << NSec.SectName << ": " 163 << formatv("{0:x16}", NSec.Address) << " -- " 164 << formatv("{0:x16}", NSec.Address + NSec.Size) 165 << ", align: " << NSec.Alignment << ", index: " << SecIndex 166 << "\n"; 167 }); 168 169 // Get the section data if any. 170 if (!isZeroFillSection(NSec)) { 171 if (DataOffset + NSec.Size > Obj.getData().size()) 172 return make_error<JITLinkError>( 173 "Section data extends past end of file"); 174 175 NSec.Data = Obj.getData().data() + DataOffset; 176 } 177 178 // Get prot flags. 179 // FIXME: Make sure this test is correct (it's probably missing cases 180 // as-is). 181 orc::MemProt Prot; 182 if (NSec.Flags & MachO::S_ATTR_PURE_INSTRUCTIONS) 183 Prot = orc::MemProt::Read | orc::MemProt::Exec; 184 else 185 Prot = orc::MemProt::Read | orc::MemProt::Write; 186 187 auto FullyQualifiedName = 188 G->allocateContent(StringRef(NSec.SegName) + "," + NSec.SectName); 189 NSec.GraphSection = &G->createSection( 190 StringRef(FullyQualifiedName.data(), FullyQualifiedName.size()), Prot); 191 192 IndexToSection.insert(std::make_pair(SecIndex, std::move(NSec))); 193 } 194 195 std::vector<NormalizedSection *> Sections; 196 Sections.reserve(IndexToSection.size()); 197 for (auto &KV : IndexToSection) 198 Sections.push_back(&KV.second); 199 200 // If we didn't end up creating any sections then bail out. The code below 201 // assumes that we have at least one section. 202 if (Sections.empty()) 203 return Error::success(); 204 205 llvm::sort(Sections, 206 [](const NormalizedSection *LHS, const NormalizedSection *RHS) { 207 assert(LHS && RHS && "Null section?"); 208 if (LHS->Address != RHS->Address) 209 return LHS->Address < RHS->Address; 210 return LHS->Size < RHS->Size; 211 }); 212 213 for (unsigned I = 0, E = Sections.size() - 1; I != E; ++I) { 214 auto &Cur = *Sections[I]; 215 auto &Next = *Sections[I + 1]; 216 if (Next.Address < Cur.Address + Cur.Size) 217 return make_error<JITLinkError>( 218 "Address range for section " + 219 formatv("\"{0}/{1}\" [ {2:x16} -- {3:x16} ] ", Cur.SegName, 220 Cur.SectName, Cur.Address, Cur.Address + Cur.Size) + 221 "overlaps section \"" + Next.SegName + "/" + Next.SectName + "\"" + 222 formatv("\"{0}/{1}\" [ {2:x16} -- {3:x16} ] ", Next.SegName, 223 Next.SectName, Next.Address, Next.Address + Next.Size)); 224 } 225 226 return Error::success(); 227 } 228 229 Error MachOLinkGraphBuilder::createNormalizedSymbols() { 230 LLVM_DEBUG(dbgs() << "Creating normalized symbols...\n"); 231 232 for (auto &SymRef : Obj.symbols()) { 233 234 unsigned SymbolIndex = Obj.getSymbolIndex(SymRef.getRawDataRefImpl()); 235 uint64_t Value; 236 uint32_t NStrX; 237 uint8_t Type; 238 uint8_t Sect; 239 uint16_t Desc; 240 241 if (Obj.is64Bit()) { 242 const MachO::nlist_64 &NL64 = 243 Obj.getSymbol64TableEntry(SymRef.getRawDataRefImpl()); 244 Value = NL64.n_value; 245 NStrX = NL64.n_strx; 246 Type = NL64.n_type; 247 Sect = NL64.n_sect; 248 Desc = NL64.n_desc; 249 } else { 250 const MachO::nlist &NL32 = 251 Obj.getSymbolTableEntry(SymRef.getRawDataRefImpl()); 252 Value = NL32.n_value; 253 NStrX = NL32.n_strx; 254 Type = NL32.n_type; 255 Sect = NL32.n_sect; 256 Desc = NL32.n_desc; 257 } 258 259 // Skip stabs. 260 // FIXME: Are there other symbols we should be skipping? 261 if (Type & MachO::N_STAB) 262 continue; 263 264 std::optional<StringRef> Name; 265 if (NStrX) { 266 if (auto NameOrErr = SymRef.getName()) 267 Name = *NameOrErr; 268 else 269 return NameOrErr.takeError(); 270 } else if (Type & MachO::N_EXT) 271 return make_error<JITLinkError>("Symbol at index " + 272 formatv("{0}", SymbolIndex) + 273 " has no name (string table index 0), " 274 "but N_EXT bit is set"); 275 276 LLVM_DEBUG({ 277 dbgs() << " "; 278 if (!Name) 279 dbgs() << "<anonymous symbol>"; 280 else 281 dbgs() << *Name; 282 dbgs() << ": value = " << formatv("{0:x16}", Value) 283 << ", type = " << formatv("{0:x2}", Type) 284 << ", desc = " << formatv("{0:x4}", Desc) << ", sect = "; 285 if (Sect) 286 dbgs() << static_cast<unsigned>(Sect - 1); 287 else 288 dbgs() << "none"; 289 dbgs() << "\n"; 290 }); 291 292 // If this symbol has a section, verify that the addresses line up. 293 if (Sect != 0) { 294 auto NSec = findSectionByIndex(Sect - 1); 295 if (!NSec) 296 return NSec.takeError(); 297 298 if (orc::ExecutorAddr(Value) < NSec->Address || 299 orc::ExecutorAddr(Value) > NSec->Address + NSec->Size) 300 return make_error<JITLinkError>("Address " + formatv("{0:x}", Value) + 301 " for symbol " + *Name + 302 " does not fall within section"); 303 304 if (!NSec->GraphSection) { 305 LLVM_DEBUG({ 306 dbgs() << " Skipping: Symbol is in section " << NSec->SegName << "/" 307 << NSec->SectName 308 << " which has no associated graph section.\n"; 309 }); 310 continue; 311 } 312 } 313 314 IndexToSymbol[SymbolIndex] = 315 &createNormalizedSymbol(*Name, Value, Type, Sect, Desc, 316 getLinkage(Desc), getScope(*Name, Type)); 317 } 318 319 return Error::success(); 320 } 321 322 void MachOLinkGraphBuilder::addSectionStartSymAndBlock( 323 unsigned SecIndex, Section &GraphSec, orc::ExecutorAddr Address, 324 const char *Data, orc::ExecutorAddrDiff Size, uint32_t Alignment, 325 bool IsLive) { 326 Block &B = 327 Data ? G->createContentBlock(GraphSec, ArrayRef<char>(Data, Size), 328 Address, Alignment, 0) 329 : G->createZeroFillBlock(GraphSec, Size, Address, Alignment, 0); 330 auto &Sym = G->addAnonymousSymbol(B, 0, Size, false, IsLive); 331 auto SecI = IndexToSection.find(SecIndex); 332 assert(SecI != IndexToSection.end() && "SecIndex invalid"); 333 auto &NSec = SecI->second; 334 assert(!NSec.CanonicalSymbols.count(Sym.getAddress()) && 335 "Anonymous block start symbol clashes with existing symbol address"); 336 NSec.CanonicalSymbols[Sym.getAddress()] = &Sym; 337 } 338 339 Error MachOLinkGraphBuilder::graphifyRegularSymbols() { 340 341 LLVM_DEBUG(dbgs() << "Creating graph symbols...\n"); 342 343 /// We only have 256 section indexes: Use a vector rather than a map. 344 std::vector<std::vector<NormalizedSymbol *>> SecIndexToSymbols; 345 SecIndexToSymbols.resize(256); 346 347 // Create commons, externs, and absolutes, and partition all other symbols by 348 // section. 349 for (auto &KV : IndexToSymbol) { 350 auto &NSym = *KV.second; 351 352 switch (NSym.Type & MachO::N_TYPE) { 353 case MachO::N_UNDF: 354 if (NSym.Value) { 355 if (!NSym.Name) 356 return make_error<JITLinkError>("Anonymous common symbol at index " + 357 Twine(KV.first)); 358 NSym.GraphSymbol = &G->addDefinedSymbol( 359 G->createZeroFillBlock(getCommonSection(), 360 orc::ExecutorAddrDiff(NSym.Value), 361 orc::ExecutorAddr(), 362 1ull << MachO::GET_COMM_ALIGN(NSym.Desc), 0), 363 0, *NSym.Name, orc::ExecutorAddrDiff(NSym.Value), Linkage::Strong, 364 NSym.S, false, NSym.Desc & MachO::N_NO_DEAD_STRIP); 365 } else { 366 if (!NSym.Name) 367 return make_error<JITLinkError>("Anonymous external symbol at " 368 "index " + 369 Twine(KV.first)); 370 NSym.GraphSymbol = &G->addExternalSymbol( 371 *NSym.Name, 0, (NSym.Desc & MachO::N_WEAK_REF) != 0); 372 } 373 break; 374 case MachO::N_ABS: 375 if (!NSym.Name) 376 return make_error<JITLinkError>("Anonymous absolute symbol at index " + 377 Twine(KV.first)); 378 NSym.GraphSymbol = &G->addAbsoluteSymbol( 379 *NSym.Name, orc::ExecutorAddr(NSym.Value), 0, Linkage::Strong, 380 getScope(*NSym.Name, NSym.Type), NSym.Desc & MachO::N_NO_DEAD_STRIP); 381 break; 382 case MachO::N_SECT: 383 SecIndexToSymbols[NSym.Sect - 1].push_back(&NSym); 384 break; 385 case MachO::N_PBUD: 386 return make_error<JITLinkError>( 387 "Unupported N_PBUD symbol " + 388 (NSym.Name ? ("\"" + *NSym.Name + "\"") : Twine("<anon>")) + 389 " at index " + Twine(KV.first)); 390 case MachO::N_INDR: 391 return make_error<JITLinkError>( 392 "Unupported N_INDR symbol " + 393 (NSym.Name ? ("\"" + *NSym.Name + "\"") : Twine("<anon>")) + 394 " at index " + Twine(KV.first)); 395 default: 396 return make_error<JITLinkError>( 397 "Unrecognized symbol type " + Twine(NSym.Type & MachO::N_TYPE) + 398 " for symbol " + 399 (NSym.Name ? ("\"" + *NSym.Name + "\"") : Twine("<anon>")) + 400 " at index " + Twine(KV.first)); 401 } 402 } 403 404 // Loop over sections performing regular graphification for those that 405 // don't have custom parsers. 406 for (auto &KV : IndexToSection) { 407 auto SecIndex = KV.first; 408 auto &NSec = KV.second; 409 410 if (!NSec.GraphSection) { 411 LLVM_DEBUG({ 412 dbgs() << " " << NSec.SegName << "/" << NSec.SectName 413 << " has no graph section. Skipping.\n"; 414 }); 415 continue; 416 } 417 418 // Skip sections with custom parsers. 419 if (CustomSectionParserFunctions.count(NSec.GraphSection->getName())) { 420 LLVM_DEBUG({ 421 dbgs() << " Skipping section " << NSec.GraphSection->getName() 422 << " as it has a custom parser.\n"; 423 }); 424 continue; 425 } else if ((NSec.Flags & MachO::SECTION_TYPE) == 426 MachO::S_CSTRING_LITERALS) { 427 if (auto Err = graphifyCStringSection( 428 NSec, std::move(SecIndexToSymbols[SecIndex]))) 429 return Err; 430 continue; 431 } else 432 LLVM_DEBUG({ 433 dbgs() << " Graphifying regular section " 434 << NSec.GraphSection->getName() << "...\n"; 435 }); 436 437 bool SectionIsNoDeadStrip = NSec.Flags & MachO::S_ATTR_NO_DEAD_STRIP; 438 bool SectionIsText = NSec.Flags & MachO::S_ATTR_PURE_INSTRUCTIONS; 439 440 auto &SecNSymStack = SecIndexToSymbols[SecIndex]; 441 442 // If this section is non-empty but there are no symbols covering it then 443 // create one block and anonymous symbol to cover the entire section. 444 if (SecNSymStack.empty()) { 445 if (NSec.Size > 0) { 446 LLVM_DEBUG({ 447 dbgs() << " Section non-empty, but contains no symbols. " 448 "Creating anonymous block to cover " 449 << formatv("{0:x16}", NSec.Address) << " -- " 450 << formatv("{0:x16}", NSec.Address + NSec.Size) << "\n"; 451 }); 452 addSectionStartSymAndBlock(SecIndex, *NSec.GraphSection, NSec.Address, 453 NSec.Data, NSec.Size, NSec.Alignment, 454 SectionIsNoDeadStrip); 455 } else 456 LLVM_DEBUG({ 457 dbgs() << " Section empty and contains no symbols. Skipping.\n"; 458 }); 459 continue; 460 } 461 462 // Sort the symbol stack in by address, alt-entry status, scope, and name. 463 // We sort in reverse order so that symbols will be visited in the right 464 // order when we pop off the stack below. 465 llvm::sort(SecNSymStack, [](const NormalizedSymbol *LHS, 466 const NormalizedSymbol *RHS) { 467 if (LHS->Value != RHS->Value) 468 return LHS->Value > RHS->Value; 469 if (isAltEntry(*LHS) != isAltEntry(*RHS)) 470 return isAltEntry(*RHS); 471 if (LHS->S != RHS->S) 472 return static_cast<uint8_t>(LHS->S) < static_cast<uint8_t>(RHS->S); 473 return LHS->Name < RHS->Name; 474 }); 475 476 // The first symbol in a section can not be an alt-entry symbol. 477 if (!SecNSymStack.empty() && isAltEntry(*SecNSymStack.back())) 478 return make_error<JITLinkError>( 479 "First symbol in " + NSec.GraphSection->getName() + " is alt-entry"); 480 481 // If the section is non-empty but there is no symbol covering the start 482 // address then add an anonymous one. 483 if (orc::ExecutorAddr(SecNSymStack.back()->Value) != NSec.Address) { 484 auto AnonBlockSize = 485 orc::ExecutorAddr(SecNSymStack.back()->Value) - NSec.Address; 486 LLVM_DEBUG({ 487 dbgs() << " Section start not covered by symbol. " 488 << "Creating anonymous block to cover [ " << NSec.Address 489 << " -- " << (NSec.Address + AnonBlockSize) << " ]\n"; 490 }); 491 addSectionStartSymAndBlock(SecIndex, *NSec.GraphSection, NSec.Address, 492 NSec.Data, AnonBlockSize, NSec.Alignment, 493 SectionIsNoDeadStrip); 494 } 495 496 // Visit section symbols in order by popping off the reverse-sorted stack, 497 // building graph symbols as we go. 498 // 499 // If MH_SUBSECTIONS_VIA_SYMBOLS is set we'll build a block for each 500 // alt-entry chain. 501 // 502 // If MH_SUBSECTIONS_VIA_SYMBOLS is not set then we'll just build one block 503 // for the whole section. 504 while (!SecNSymStack.empty()) { 505 SmallVector<NormalizedSymbol *, 8> BlockSyms; 506 507 // Get the symbols in this alt-entry chain, or the whole section (if 508 // !SubsectionsViaSymbols). 509 BlockSyms.push_back(SecNSymStack.back()); 510 SecNSymStack.pop_back(); 511 while (!SecNSymStack.empty() && 512 (isAltEntry(*SecNSymStack.back()) || 513 SecNSymStack.back()->Value == BlockSyms.back()->Value || 514 !SubsectionsViaSymbols)) { 515 BlockSyms.push_back(SecNSymStack.back()); 516 SecNSymStack.pop_back(); 517 } 518 519 // BlockNSyms now contains the block symbols in reverse canonical order. 520 auto BlockStart = orc::ExecutorAddr(BlockSyms.front()->Value); 521 orc::ExecutorAddr BlockEnd = 522 SecNSymStack.empty() ? NSec.Address + NSec.Size 523 : orc::ExecutorAddr(SecNSymStack.back()->Value); 524 orc::ExecutorAddrDiff BlockOffset = BlockStart - NSec.Address; 525 orc::ExecutorAddrDiff BlockSize = BlockEnd - BlockStart; 526 527 LLVM_DEBUG({ 528 dbgs() << " Creating block for " << formatv("{0:x16}", BlockStart) 529 << " -- " << formatv("{0:x16}", BlockEnd) << ": " 530 << NSec.GraphSection->getName() << " + " 531 << formatv("{0:x16}", BlockOffset) << " with " 532 << BlockSyms.size() << " symbol(s)...\n"; 533 }); 534 535 Block &B = 536 NSec.Data 537 ? G->createContentBlock( 538 *NSec.GraphSection, 539 ArrayRef<char>(NSec.Data + BlockOffset, BlockSize), 540 BlockStart, NSec.Alignment, BlockStart % NSec.Alignment) 541 : G->createZeroFillBlock(*NSec.GraphSection, BlockSize, 542 BlockStart, NSec.Alignment, 543 BlockStart % NSec.Alignment); 544 545 std::optional<orc::ExecutorAddr> LastCanonicalAddr; 546 auto SymEnd = BlockEnd; 547 while (!BlockSyms.empty()) { 548 auto &NSym = *BlockSyms.back(); 549 BlockSyms.pop_back(); 550 551 bool SymLive = 552 (NSym.Desc & MachO::N_NO_DEAD_STRIP) || SectionIsNoDeadStrip; 553 554 auto &Sym = createStandardGraphSymbol( 555 NSym, B, SymEnd - orc::ExecutorAddr(NSym.Value), SectionIsText, 556 SymLive, LastCanonicalAddr != orc::ExecutorAddr(NSym.Value)); 557 558 if (LastCanonicalAddr != Sym.getAddress()) { 559 if (LastCanonicalAddr) 560 SymEnd = *LastCanonicalAddr; 561 LastCanonicalAddr = Sym.getAddress(); 562 } 563 } 564 } 565 } 566 567 return Error::success(); 568 } 569 570 Symbol &MachOLinkGraphBuilder::createStandardGraphSymbol(NormalizedSymbol &NSym, 571 Block &B, size_t Size, 572 bool IsText, 573 bool IsNoDeadStrip, 574 bool IsCanonical) { 575 576 LLVM_DEBUG({ 577 dbgs() << " " << formatv("{0:x16}", NSym.Value) << " -- " 578 << formatv("{0:x16}", NSym.Value + Size) << ": "; 579 if (!NSym.Name) 580 dbgs() << "<anonymous symbol>"; 581 else 582 dbgs() << NSym.Name; 583 if (IsText) 584 dbgs() << " [text]"; 585 if (IsNoDeadStrip) 586 dbgs() << " [no-dead-strip]"; 587 if (!IsCanonical) 588 dbgs() << " [non-canonical]"; 589 dbgs() << "\n"; 590 }); 591 592 auto SymOffset = orc::ExecutorAddr(NSym.Value) - B.getAddress(); 593 auto &Sym = 594 NSym.Name 595 ? G->addDefinedSymbol(B, SymOffset, *NSym.Name, Size, NSym.L, NSym.S, 596 IsText, IsNoDeadStrip) 597 : G->addAnonymousSymbol(B, SymOffset, Size, IsText, IsNoDeadStrip); 598 NSym.GraphSymbol = &Sym; 599 600 if (IsCanonical) 601 setCanonicalSymbol(getSectionByIndex(NSym.Sect - 1), Sym); 602 603 return Sym; 604 } 605 606 Error MachOLinkGraphBuilder::graphifySectionsWithCustomParsers() { 607 // Graphify special sections. 608 for (auto &KV : IndexToSection) { 609 auto &NSec = KV.second; 610 611 // Skip non-graph sections. 612 if (!NSec.GraphSection) 613 continue; 614 615 auto HI = CustomSectionParserFunctions.find(NSec.GraphSection->getName()); 616 if (HI != CustomSectionParserFunctions.end()) { 617 auto &Parse = HI->second; 618 if (auto Err = Parse(NSec)) 619 return Err; 620 } 621 } 622 623 return Error::success(); 624 } 625 626 Error MachOLinkGraphBuilder::graphifyCStringSection( 627 NormalizedSection &NSec, std::vector<NormalizedSymbol *> NSyms) { 628 assert(NSec.GraphSection && "C string literal section missing graph section"); 629 assert(NSec.Data && "C string literal section has no data"); 630 631 LLVM_DEBUG({ 632 dbgs() << " Graphifying C-string literal section " 633 << NSec.GraphSection->getName() << "\n"; 634 }); 635 636 if (NSec.Data[NSec.Size - 1] != '\0') 637 return make_error<JITLinkError>("C string literal section " + 638 NSec.GraphSection->getName() + 639 " does not end with null terminator"); 640 641 /// Sort into reverse order to use as a stack. 642 llvm::sort(NSyms, 643 [](const NormalizedSymbol *LHS, const NormalizedSymbol *RHS) { 644 if (LHS->Value != RHS->Value) 645 return LHS->Value > RHS->Value; 646 if (LHS->L != RHS->L) 647 return LHS->L > RHS->L; 648 if (LHS->S != RHS->S) 649 return LHS->S > RHS->S; 650 if (RHS->Name) { 651 if (!LHS->Name) 652 return true; 653 return *LHS->Name > *RHS->Name; 654 } 655 return false; 656 }); 657 658 bool SectionIsNoDeadStrip = NSec.Flags & MachO::S_ATTR_NO_DEAD_STRIP; 659 bool SectionIsText = NSec.Flags & MachO::S_ATTR_PURE_INSTRUCTIONS; 660 orc::ExecutorAddrDiff BlockStart = 0; 661 662 // Scan section for null characters. 663 for (size_t I = 0; I != NSec.Size; ++I) { 664 if (NSec.Data[I] == '\0') { 665 size_t BlockSize = I + 1 - BlockStart; 666 // Create a block for this null terminated string. 667 auto &B = G->createContentBlock(*NSec.GraphSection, 668 {NSec.Data + BlockStart, BlockSize}, 669 NSec.Address + BlockStart, NSec.Alignment, 670 BlockStart % NSec.Alignment); 671 672 LLVM_DEBUG({ 673 dbgs() << " Created block " << B.getRange() 674 << ", align = " << B.getAlignment() 675 << ", align-ofs = " << B.getAlignmentOffset() << " for \""; 676 for (size_t J = 0; J != std::min(B.getSize(), size_t(16)); ++J) 677 switch (B.getContent()[J]) { 678 case '\0': break; 679 case '\n': dbgs() << "\\n"; break; 680 case '\t': dbgs() << "\\t"; break; 681 default: dbgs() << B.getContent()[J]; break; 682 } 683 if (B.getSize() > 16) 684 dbgs() << "..."; 685 dbgs() << "\"\n"; 686 }); 687 688 // If there's no symbol at the start of this block then create one. 689 if (NSyms.empty() || 690 orc::ExecutorAddr(NSyms.back()->Value) != B.getAddress()) { 691 auto &S = G->addAnonymousSymbol(B, 0, BlockSize, false, false); 692 setCanonicalSymbol(NSec, S); 693 LLVM_DEBUG({ 694 dbgs() << " Adding symbol for c-string block " << B.getRange() 695 << ": <anonymous symbol> at offset 0\n"; 696 }); 697 } 698 699 // Process any remaining symbols that point into this block. 700 auto LastCanonicalAddr = B.getAddress() + BlockSize; 701 while (!NSyms.empty() && orc::ExecutorAddr(NSyms.back()->Value) < 702 B.getAddress() + BlockSize) { 703 auto &NSym = *NSyms.back(); 704 size_t SymSize = (B.getAddress() + BlockSize) - 705 orc::ExecutorAddr(NSyms.back()->Value); 706 bool SymLive = 707 (NSym.Desc & MachO::N_NO_DEAD_STRIP) || SectionIsNoDeadStrip; 708 709 bool IsCanonical = false; 710 if (LastCanonicalAddr != orc::ExecutorAddr(NSym.Value)) { 711 IsCanonical = true; 712 LastCanonicalAddr = orc::ExecutorAddr(NSym.Value); 713 } 714 715 auto &Sym = createStandardGraphSymbol(NSym, B, SymSize, SectionIsText, 716 SymLive, IsCanonical); 717 (void)Sym; 718 LLVM_DEBUG({ 719 dbgs() << " Adding symbol for c-string block " << B.getRange() 720 << ": " 721 << (Sym.hasName() ? Sym.getName() : "<anonymous symbol>") 722 << " at offset " << formatv("{0:x}", Sym.getOffset()) << "\n"; 723 }); 724 725 NSyms.pop_back(); 726 } 727 728 BlockStart += BlockSize; 729 } 730 } 731 732 assert(llvm::all_of(NSec.GraphSection->blocks(), 733 [](Block *B) { return isCStringBlock(*B); }) && 734 "All blocks in section should hold single c-strings"); 735 736 return Error::success(); 737 } 738 739 Error CompactUnwindSplitter::operator()(LinkGraph &G) { 740 auto *CUSec = G.findSectionByName(CompactUnwindSectionName); 741 if (!CUSec) 742 return Error::success(); 743 744 if (!G.getTargetTriple().isOSBinFormatMachO()) 745 return make_error<JITLinkError>( 746 "Error linking " + G.getName() + 747 ": compact unwind splitting not supported on non-macho target " + 748 G.getTargetTriple().str()); 749 750 unsigned CURecordSize = 0; 751 unsigned PersonalityEdgeOffset = 0; 752 unsigned LSDAEdgeOffset = 0; 753 switch (G.getTargetTriple().getArch()) { 754 case Triple::aarch64: 755 case Triple::x86_64: 756 // 64-bit compact-unwind record format: 757 // Range start: 8 bytes. 758 // Range size: 4 bytes. 759 // CU encoding: 4 bytes. 760 // Personality: 8 bytes. 761 // LSDA: 8 bytes. 762 CURecordSize = 32; 763 PersonalityEdgeOffset = 16; 764 LSDAEdgeOffset = 24; 765 break; 766 default: 767 return make_error<JITLinkError>( 768 "Error linking " + G.getName() + 769 ": compact unwind splitting not supported on " + 770 G.getTargetTriple().getArchName()); 771 } 772 773 std::vector<Block *> OriginalBlocks(CUSec->blocks().begin(), 774 CUSec->blocks().end()); 775 LLVM_DEBUG({ 776 dbgs() << "In " << G.getName() << " splitting compact unwind section " 777 << CompactUnwindSectionName << " containing " 778 << OriginalBlocks.size() << " initial blocks...\n"; 779 }); 780 781 while (!OriginalBlocks.empty()) { 782 auto *B = OriginalBlocks.back(); 783 OriginalBlocks.pop_back(); 784 785 if (B->getSize() == 0) { 786 LLVM_DEBUG({ 787 dbgs() << " Skipping empty block at " 788 << formatv("{0:x16}", B->getAddress()) << "\n"; 789 }); 790 continue; 791 } 792 793 LLVM_DEBUG({ 794 dbgs() << " Splitting block at " << formatv("{0:x16}", B->getAddress()) 795 << " into " << (B->getSize() / CURecordSize) 796 << " compact unwind record(s)\n"; 797 }); 798 799 if (B->getSize() % CURecordSize) 800 return make_error<JITLinkError>( 801 "Error splitting compact unwind record in " + G.getName() + 802 ": block at " + formatv("{0:x}", B->getAddress()) + " has size " + 803 formatv("{0:x}", B->getSize()) + 804 " (not a multiple of CU record size of " + 805 formatv("{0:x}", CURecordSize) + ")"); 806 807 unsigned NumBlocks = B->getSize() / CURecordSize; 808 LinkGraph::SplitBlockCache C; 809 810 for (unsigned I = 0; I != NumBlocks; ++I) { 811 auto &CURec = G.splitBlock(*B, CURecordSize, &C); 812 bool AddedKeepAlive = false; 813 814 for (auto &E : CURec.edges()) { 815 if (E.getOffset() == 0) { 816 LLVM_DEBUG({ 817 dbgs() << " Updating compact unwind record at " 818 << formatv("{0:x16}", CURec.getAddress()) << " to point to " 819 << (E.getTarget().hasName() ? E.getTarget().getName() 820 : StringRef()) 821 << " (at " << formatv("{0:x16}", E.getTarget().getAddress()) 822 << ")\n"; 823 }); 824 825 if (E.getTarget().isExternal()) 826 return make_error<JITLinkError>( 827 "Error adding keep-alive edge for compact unwind record at " + 828 formatv("{0:x}", CURec.getAddress()) + ": target " + 829 E.getTarget().getName() + " is an external symbol"); 830 auto &TgtBlock = E.getTarget().getBlock(); 831 auto &CURecSym = 832 G.addAnonymousSymbol(CURec, 0, CURecordSize, false, false); 833 TgtBlock.addEdge(Edge::KeepAlive, 0, CURecSym, 0); 834 AddedKeepAlive = true; 835 } else if (E.getOffset() != PersonalityEdgeOffset && 836 E.getOffset() != LSDAEdgeOffset) 837 return make_error<JITLinkError>("Unexpected edge at offset " + 838 formatv("{0:x}", E.getOffset()) + 839 " in compact unwind record at " + 840 formatv("{0:x}", CURec.getAddress())); 841 } 842 843 if (!AddedKeepAlive) 844 return make_error<JITLinkError>( 845 "Error adding keep-alive edge for compact unwind record at " + 846 formatv("{0:x}", CURec.getAddress()) + 847 ": no outgoing target edge at offset 0"); 848 } 849 } 850 return Error::success(); 851 } 852 853 } // end namespace jitlink 854 } // end namespace llvm 855