1 //===- bolt/Core/Exceptions.cpp - Helpers for C++ exceptions --------------===// 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 implements functions for handling C++ exception meta data. 10 // 11 // Some of the code is taken from examples/ExceptionDemo 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "bolt/Core/Exceptions.h" 16 #include "bolt/Core/BinaryFunction.h" 17 #include "llvm/ADT/ArrayRef.h" 18 #include "llvm/ADT/Twine.h" 19 #include "llvm/BinaryFormat/Dwarf.h" 20 #include "llvm/DebugInfo/DWARF/DWARFDebugFrame.h" 21 #include "llvm/Support/Casting.h" 22 #include "llvm/Support/CommandLine.h" 23 #include "llvm/Support/Debug.h" 24 #include "llvm/Support/Errc.h" 25 #include "llvm/Support/LEB128.h" 26 #include "llvm/Support/MathExtras.h" 27 #include "llvm/Support/raw_ostream.h" 28 #include <map> 29 30 #undef DEBUG_TYPE 31 #define DEBUG_TYPE "bolt-exceptions" 32 33 using namespace llvm::dwarf; 34 35 namespace opts { 36 37 extern llvm::cl::OptionCategory BoltCategory; 38 39 extern llvm::cl::opt<unsigned> Verbosity; 40 41 static llvm::cl::opt<bool> 42 PrintExceptions("print-exceptions", 43 llvm::cl::desc("print exception handling data"), 44 llvm::cl::Hidden, llvm::cl::cat(BoltCategory)); 45 46 } // namespace opts 47 48 namespace llvm { 49 namespace bolt { 50 51 // Read and dump the .gcc_exception_table section entry. 52 // 53 // .gcc_except_table section contains a set of Language-Specific Data Areas - 54 // a fancy name for exception handling tables. There's one LSDA entry per 55 // function. However, we can't actually tell which function LSDA refers to 56 // unless we parse .eh_frame entry that refers to the LSDA. 57 // Then inside LSDA most addresses are encoded relative to the function start, 58 // so we need the function context in order to get to real addresses. 59 // 60 // The best visual representation of the tables comprising LSDA and 61 // relationships between them is illustrated at: 62 // https://github.com/itanium-cxx-abi/cxx-abi/blob/master/exceptions.pdf 63 // Keep in mind that GCC implementation deviates slightly from that document. 64 // 65 // To summarize, there are 4 tables in LSDA: call site table, actions table, 66 // types table, and types index table (for indirection). The main table contains 67 // call site entries. Each call site includes a PC range that can throw an 68 // exception, a handler (landing pad), and a reference to an entry in the action 69 // table. The handler and/or action could be 0. The action entry is a head 70 // of a list of actions associated with a call site. The action table contains 71 // all such lists (it could be optimized to share list tails). Each action could 72 // be either to catch an exception of a given type, to perform a cleanup, or to 73 // propagate the exception after filtering it out (e.g. to make sure function 74 // exception specification is not violated). Catch action contains a reference 75 // to an entry in the type table, and filter action refers to an entry in the 76 // type index table to encode a set of types to filter. 77 // 78 // Call site table follows LSDA header. Action table immediately follows the 79 // call site table. 80 // 81 // Both types table and type index table start at the same location, but they 82 // grow in opposite directions (types go up, indices go down). The beginning of 83 // these tables is encoded in LSDA header. Sizes for both of the tables are not 84 // included anywhere. 85 // 86 // We have to parse all of the tables to determine their sizes. Then we have 87 // to parse the call site table and associate discovered information with 88 // actual call instructions and landing pad blocks. 89 // 90 // For the purpose of rewriting exception handling tables, we can reuse action, 91 // and type index tables in their original binary format. 92 // 93 // Type table could be encoded using position-independent references, and thus 94 // may require relocation. 95 // 96 // Ideally we should be able to re-write LSDA in-place, without the need to 97 // allocate a new space for it. Sadly there's no guarantee that the new call 98 // site table will be the same size as GCC uses uleb encodings for PC offsets. 99 // 100 // Note: some functions have LSDA entries with 0 call site entries. 101 void BinaryFunction::parseLSDA(ArrayRef<uint8_t> LSDASectionData, 102 uint64_t LSDASectionAddress) { 103 assert(CurrentState == State::Disassembled && "unexpected function state"); 104 105 if (!getLSDAAddress()) 106 return; 107 108 DWARFDataExtractor Data( 109 StringRef(reinterpret_cast<const char *>(LSDASectionData.data()), 110 LSDASectionData.size()), 111 BC.DwCtx->getDWARFObj().isLittleEndian(), 8); 112 uint64_t Offset = getLSDAAddress() - LSDASectionAddress; 113 assert(Data.isValidOffset(Offset) && "wrong LSDA address"); 114 115 uint8_t LPStartEncoding = Data.getU8(&Offset); 116 uint64_t LPStart = 0; 117 // Convert to offset if LPStartEncoding is typed absptr DW_EH_PE_absptr 118 if (std::optional<uint64_t> MaybeLPStart = Data.getEncodedPointer( 119 &Offset, LPStartEncoding, Offset + LSDASectionAddress)) 120 LPStart = (LPStartEncoding && 0xFF == 0) ? *MaybeLPStart 121 : *MaybeLPStart - Address; 122 123 const uint8_t TTypeEncoding = Data.getU8(&Offset); 124 LSDATypeEncoding = TTypeEncoding; 125 size_t TTypeEncodingSize = 0; 126 uintptr_t TTypeEnd = 0; 127 if (TTypeEncoding != DW_EH_PE_omit) { 128 TTypeEnd = Data.getULEB128(&Offset); 129 TTypeEncodingSize = BC.getDWARFEncodingSize(TTypeEncoding); 130 } 131 132 if (opts::PrintExceptions) { 133 outs() << "[LSDA at 0x" << Twine::utohexstr(getLSDAAddress()) 134 << " for function " << *this << "]:\n"; 135 outs() << "LPStart Encoding = 0x" << Twine::utohexstr(LPStartEncoding) 136 << '\n'; 137 outs() << "LPStart = 0x" << Twine::utohexstr(LPStart) << '\n'; 138 outs() << "TType Encoding = 0x" << Twine::utohexstr(TTypeEncoding) << '\n'; 139 outs() << "TType End = " << TTypeEnd << '\n'; 140 } 141 142 // Table to store list of indices in type table. Entries are uleb128 values. 143 const uint64_t TypeIndexTableStart = Offset + TTypeEnd; 144 145 // Offset past the last decoded index. 146 uint64_t MaxTypeIndexTableOffset = 0; 147 148 // Max positive index used in type table. 149 unsigned MaxTypeIndex = 0; 150 151 // The actual type info table starts at the same location, but grows in 152 // opposite direction. TTypeEncoding is used to encode stored values. 153 const uint64_t TypeTableStart = Offset + TTypeEnd; 154 155 uint8_t CallSiteEncoding = Data.getU8(&Offset); 156 uint32_t CallSiteTableLength = Data.getULEB128(&Offset); 157 uint64_t CallSiteTableStart = Offset; 158 uint64_t CallSiteTableEnd = CallSiteTableStart + CallSiteTableLength; 159 uint64_t CallSitePtr = CallSiteTableStart; 160 uint64_t ActionTableStart = CallSiteTableEnd; 161 162 if (opts::PrintExceptions) { 163 outs() << "CallSite Encoding = " << (unsigned)CallSiteEncoding << '\n'; 164 outs() << "CallSite table length = " << CallSiteTableLength << '\n'; 165 outs() << '\n'; 166 } 167 168 this->HasEHRanges = CallSitePtr < CallSiteTableEnd; 169 const uint64_t RangeBase = getAddress(); 170 while (CallSitePtr < CallSiteTableEnd) { 171 uint64_t Start = *Data.getEncodedPointer(&CallSitePtr, CallSiteEncoding, 172 CallSitePtr + LSDASectionAddress); 173 uint64_t Length = *Data.getEncodedPointer(&CallSitePtr, CallSiteEncoding, 174 CallSitePtr + LSDASectionAddress); 175 uint64_t LandingPad = *Data.getEncodedPointer( 176 &CallSitePtr, CallSiteEncoding, CallSitePtr + LSDASectionAddress); 177 uint64_t ActionEntry = Data.getULEB128(&CallSitePtr); 178 179 uint64_t LPOffset = LPStart + LandingPad; 180 uint64_t LPAddress = Address + LPOffset; 181 182 // Verify if landing pad code is located outside current function 183 // Support landing pad to builtin_unreachable 184 if (LPAddress < Address || LPAddress > Address + getSize()) { 185 BinaryFunction *Fragment = 186 BC.getBinaryFunctionContainingAddress(LPAddress); 187 assert(Fragment != nullptr && 188 "BOLT-ERROR: cannot find landing pad fragment"); 189 BC.addInterproceduralReference(this, Fragment->getAddress()); 190 BC.processInterproceduralReferences(); 191 auto isFragmentOf = [](BinaryFunction *Fragment, 192 BinaryFunction *Parent) -> bool { 193 return (Fragment->isFragment() && Fragment->isParentFragment(Parent)); 194 }; 195 (void)isFragmentOf; 196 assert((isFragmentOf(this, Fragment) || isFragmentOf(Fragment, this)) && 197 "BOLT-ERROR: cannot have landing pads in different " 198 "functions"); 199 setHasIndirectTargetToSplitFragment(true); 200 BC.addFragmentsToSkip(this); 201 return; 202 } 203 204 if (opts::PrintExceptions) { 205 outs() << "Call Site: [0x" << Twine::utohexstr(RangeBase + Start) 206 << ", 0x" << Twine::utohexstr(RangeBase + Start + Length) 207 << "); landing pad: 0x" << Twine::utohexstr(LPOffset) 208 << "; action entry: 0x" << Twine::utohexstr(ActionEntry) << "\n"; 209 outs() << " current offset is " << (CallSitePtr - CallSiteTableStart) 210 << '\n'; 211 } 212 213 // Create a handler entry if necessary. 214 MCSymbol *LPSymbol = nullptr; 215 if (LPOffset) { 216 if (!getInstructionAtOffset(LPOffset)) { 217 if (opts::Verbosity >= 1) 218 errs() << "BOLT-WARNING: landing pad " << Twine::utohexstr(LPOffset) 219 << " not pointing to an instruction in function " << *this 220 << " - ignoring.\n"; 221 } else { 222 auto Label = Labels.find(LPOffset); 223 if (Label != Labels.end()) { 224 LPSymbol = Label->second; 225 } else { 226 LPSymbol = BC.Ctx->createNamedTempSymbol("LP"); 227 Labels[LPOffset] = LPSymbol; 228 } 229 } 230 } 231 232 // Mark all call instructions in the range. 233 auto II = Instructions.find(Start); 234 auto IE = Instructions.end(); 235 assert(II != IE && "exception range not pointing to an instruction"); 236 do { 237 MCInst &Instruction = II->second; 238 if (BC.MIB->isCall(Instruction) && 239 !BC.MIB->getConditionalTailCall(Instruction)) { 240 assert(!BC.MIB->isInvoke(Instruction) && 241 "overlapping exception ranges detected"); 242 // Add extra operands to a call instruction making it an invoke from 243 // now on. 244 BC.MIB->addEHInfo(Instruction, 245 MCPlus::MCLandingPad(LPSymbol, ActionEntry)); 246 } 247 ++II; 248 } while (II != IE && II->first < Start + Length); 249 250 if (ActionEntry != 0) { 251 auto printType = [&](int Index, raw_ostream &OS) { 252 assert(Index > 0 && "only positive indices are valid"); 253 uint64_t TTEntry = TypeTableStart - Index * TTypeEncodingSize; 254 const uint64_t TTEntryAddress = TTEntry + LSDASectionAddress; 255 uint64_t TypeAddress = 256 *Data.getEncodedPointer(&TTEntry, TTypeEncoding, TTEntryAddress); 257 if ((TTypeEncoding & DW_EH_PE_pcrel) && TypeAddress == TTEntryAddress) 258 TypeAddress = 0; 259 if (TypeAddress == 0) { 260 OS << "<all>"; 261 return; 262 } 263 if (TTypeEncoding & DW_EH_PE_indirect) { 264 ErrorOr<uint64_t> PointerOrErr = BC.getPointerAtAddress(TypeAddress); 265 assert(PointerOrErr && "failed to decode indirect address"); 266 TypeAddress = *PointerOrErr; 267 } 268 if (BinaryData *TypeSymBD = BC.getBinaryDataAtAddress(TypeAddress)) 269 OS << TypeSymBD->getName(); 270 else 271 OS << "0x" << Twine::utohexstr(TypeAddress); 272 }; 273 if (opts::PrintExceptions) 274 outs() << " actions: "; 275 uint64_t ActionPtr = ActionTableStart + ActionEntry - 1; 276 int64_t ActionType; 277 int64_t ActionNext; 278 const char *Sep = ""; 279 do { 280 ActionType = Data.getSLEB128(&ActionPtr); 281 const uint32_t Self = ActionPtr; 282 ActionNext = Data.getSLEB128(&ActionPtr); 283 if (opts::PrintExceptions) 284 outs() << Sep << "(" << ActionType << ", " << ActionNext << ") "; 285 if (ActionType == 0) { 286 if (opts::PrintExceptions) 287 outs() << "cleanup"; 288 } else if (ActionType > 0) { 289 // It's an index into a type table. 290 MaxTypeIndex = 291 std::max(MaxTypeIndex, static_cast<unsigned>(ActionType)); 292 if (opts::PrintExceptions) { 293 outs() << "catch type "; 294 printType(ActionType, outs()); 295 } 296 } else { // ActionType < 0 297 if (opts::PrintExceptions) 298 outs() << "filter exception types "; 299 const char *TSep = ""; 300 // ActionType is a negative *byte* offset into *uleb128-encoded* table 301 // of indices with base 1. 302 // E.g. -1 means offset 0, -2 is offset 1, etc. The indices are 303 // encoded using uleb128 thus we cannot directly dereference them. 304 uint64_t TypeIndexTablePtr = TypeIndexTableStart - ActionType - 1; 305 while (uint64_t Index = Data.getULEB128(&TypeIndexTablePtr)) { 306 MaxTypeIndex = std::max(MaxTypeIndex, static_cast<unsigned>(Index)); 307 if (opts::PrintExceptions) { 308 outs() << TSep; 309 printType(Index, outs()); 310 TSep = ", "; 311 } 312 } 313 MaxTypeIndexTableOffset = std::max( 314 MaxTypeIndexTableOffset, TypeIndexTablePtr - TypeIndexTableStart); 315 } 316 317 Sep = "; "; 318 319 ActionPtr = Self + ActionNext; 320 } while (ActionNext); 321 if (opts::PrintExceptions) 322 outs() << '\n'; 323 } 324 } 325 if (opts::PrintExceptions) 326 outs() << '\n'; 327 328 assert(TypeIndexTableStart + MaxTypeIndexTableOffset <= 329 Data.getData().size() && 330 "LSDA entry has crossed section boundary"); 331 332 if (TTypeEnd) { 333 LSDAActionTable = LSDASectionData.slice( 334 ActionTableStart, TypeIndexTableStart - 335 MaxTypeIndex * TTypeEncodingSize - 336 ActionTableStart); 337 for (unsigned Index = 1; Index <= MaxTypeIndex; ++Index) { 338 uint64_t TTEntry = TypeTableStart - Index * TTypeEncodingSize; 339 const uint64_t TTEntryAddress = TTEntry + LSDASectionAddress; 340 uint64_t TypeAddress = 341 *Data.getEncodedPointer(&TTEntry, TTypeEncoding, TTEntryAddress); 342 if ((TTypeEncoding & DW_EH_PE_pcrel) && (TypeAddress == TTEntryAddress)) 343 TypeAddress = 0; 344 if (TTypeEncoding & DW_EH_PE_indirect) { 345 LSDATypeAddressTable.emplace_back(TypeAddress); 346 if (TypeAddress) { 347 ErrorOr<uint64_t> PointerOrErr = BC.getPointerAtAddress(TypeAddress); 348 assert(PointerOrErr && "failed to decode indirect address"); 349 TypeAddress = *PointerOrErr; 350 } 351 } 352 LSDATypeTable.emplace_back(TypeAddress); 353 } 354 LSDATypeIndexTable = 355 LSDASectionData.slice(TypeIndexTableStart, MaxTypeIndexTableOffset); 356 } 357 } 358 359 void BinaryFunction::updateEHRanges() { 360 if (getSize() == 0) 361 return; 362 363 assert(CurrentState == State::CFG_Finalized && "unexpected state"); 364 365 // Build call sites table. 366 struct EHInfo { 367 const MCSymbol *LP; // landing pad 368 uint64_t Action; 369 }; 370 371 // Sites to update. 372 CallSitesList Sites; 373 374 for (FunctionFragment &FF : getLayout().fragments()) { 375 // If previous call can throw, this is its exception handler. 376 EHInfo PreviousEH = {nullptr, 0}; 377 378 // Marker for the beginning of exceptions range. 379 const MCSymbol *StartRange = nullptr; 380 381 for (BinaryBasicBlock *const BB : FF) { 382 for (auto II = BB->begin(); II != BB->end(); ++II) { 383 if (!BC.MIB->isCall(*II)) 384 continue; 385 386 // Instruction can throw an exception that should be handled. 387 const bool Throws = BC.MIB->isInvoke(*II); 388 389 // Ignore the call if it's a continuation of a no-throw gap. 390 if (!Throws && !StartRange) 391 continue; 392 393 // Extract exception handling information from the instruction. 394 const MCSymbol *LP = nullptr; 395 uint64_t Action = 0; 396 if (const std::optional<MCPlus::MCLandingPad> EHInfo = 397 BC.MIB->getEHInfo(*II)) 398 std::tie(LP, Action) = *EHInfo; 399 400 // No action if the exception handler has not changed. 401 if (Throws && StartRange && PreviousEH.LP == LP && 402 PreviousEH.Action == Action) 403 continue; 404 405 // Same symbol is used for the beginning and the end of the range. 406 const MCSymbol *EHSymbol; 407 MCInst EHLabel; 408 { 409 std::unique_lock<llvm::sys::RWMutex> Lock(BC.CtxMutex); 410 EHSymbol = BC.Ctx->createNamedTempSymbol("EH"); 411 BC.MIB->createEHLabel(EHLabel, EHSymbol, BC.Ctx.get()); 412 } 413 414 II = std::next(BB->insertPseudoInstr(II, EHLabel)); 415 416 // At this point we could be in one of the following states: 417 // 418 // I. Exception handler has changed and we need to close previous range 419 // and start a new one. 420 // 421 // II. Start a new exception range after the gap. 422 // 423 // III. Close current exception range and start a new gap. 424 const MCSymbol *EndRange; 425 if (StartRange) { 426 // I, III: 427 EndRange = EHSymbol; 428 } else { 429 // II: 430 StartRange = EHSymbol; 431 EndRange = nullptr; 432 } 433 434 // Close the previous range. 435 if (EndRange) 436 Sites.emplace_back( 437 FF.getFragmentNum(), 438 CallSite{StartRange, EndRange, PreviousEH.LP, PreviousEH.Action}); 439 440 if (Throws) { 441 // I, II: 442 StartRange = EHSymbol; 443 PreviousEH = EHInfo{LP, Action}; 444 } else { 445 StartRange = nullptr; 446 } 447 } 448 } 449 450 // Check if we need to close the range. 451 if (StartRange) { 452 const MCSymbol *EndRange = getFunctionEndLabel(FF.getFragmentNum()); 453 Sites.emplace_back( 454 FF.getFragmentNum(), 455 CallSite{StartRange, EndRange, PreviousEH.LP, PreviousEH.Action}); 456 } 457 } 458 459 addCallSites(Sites); 460 } 461 462 const uint8_t DWARF_CFI_PRIMARY_OPCODE_MASK = 0xc0; 463 464 CFIReaderWriter::CFIReaderWriter(const DWARFDebugFrame &EHFrame) { 465 // Prepare FDEs for fast lookup 466 for (const dwarf::FrameEntry &Entry : EHFrame.entries()) { 467 const auto *CurFDE = dyn_cast<dwarf::FDE>(&Entry); 468 // Skip CIEs. 469 if (!CurFDE) 470 continue; 471 // There could me multiple FDEs with the same initial address, and perhaps 472 // different sizes (address ranges). Use the first entry with non-zero size. 473 auto FDEI = FDEs.lower_bound(CurFDE->getInitialLocation()); 474 if (FDEI != FDEs.end() && FDEI->first == CurFDE->getInitialLocation()) { 475 if (CurFDE->getAddressRange()) { 476 if (FDEI->second->getAddressRange() == 0) { 477 FDEI->second = CurFDE; 478 } else if (opts::Verbosity > 0) { 479 errs() << "BOLT-WARNING: different FDEs for function at 0x" 480 << Twine::utohexstr(FDEI->first) 481 << " detected; sizes: " << FDEI->second->getAddressRange() 482 << " and " << CurFDE->getAddressRange() << '\n'; 483 } 484 } 485 } else { 486 FDEs.emplace_hint(FDEI, CurFDE->getInitialLocation(), CurFDE); 487 } 488 } 489 } 490 491 bool CFIReaderWriter::fillCFIInfoFor(BinaryFunction &Function) const { 492 uint64_t Address = Function.getAddress(); 493 auto I = FDEs.find(Address); 494 // Ignore zero-length FDE ranges. 495 if (I == FDEs.end() || !I->second->getAddressRange()) 496 return true; 497 498 const FDE &CurFDE = *I->second; 499 std::optional<uint64_t> LSDA = CurFDE.getLSDAAddress(); 500 Function.setLSDAAddress(LSDA ? *LSDA : 0); 501 502 uint64_t Offset = Function.getFirstInstructionOffset(); 503 uint64_t CodeAlignment = CurFDE.getLinkedCIE()->getCodeAlignmentFactor(); 504 uint64_t DataAlignment = CurFDE.getLinkedCIE()->getDataAlignmentFactor(); 505 if (CurFDE.getLinkedCIE()->getPersonalityAddress()) { 506 Function.setPersonalityFunction( 507 *CurFDE.getLinkedCIE()->getPersonalityAddress()); 508 Function.setPersonalityEncoding( 509 *CurFDE.getLinkedCIE()->getPersonalityEncoding()); 510 } 511 512 auto decodeFrameInstruction = [&Function, &Offset, Address, CodeAlignment, 513 DataAlignment]( 514 const CFIProgram::Instruction &Instr) { 515 uint8_t Opcode = Instr.Opcode; 516 if (Opcode & DWARF_CFI_PRIMARY_OPCODE_MASK) 517 Opcode &= DWARF_CFI_PRIMARY_OPCODE_MASK; 518 switch (Instr.Opcode) { 519 case DW_CFA_nop: 520 break; 521 case DW_CFA_advance_loc4: 522 case DW_CFA_advance_loc2: 523 case DW_CFA_advance_loc1: 524 case DW_CFA_advance_loc: 525 // Advance our current address 526 Offset += CodeAlignment * int64_t(Instr.Ops[0]); 527 break; 528 case DW_CFA_offset_extended_sf: 529 Function.addCFIInstruction( 530 Offset, 531 MCCFIInstruction::createOffset( 532 nullptr, Instr.Ops[0], DataAlignment * int64_t(Instr.Ops[1]))); 533 break; 534 case DW_CFA_offset_extended: 535 case DW_CFA_offset: 536 Function.addCFIInstruction( 537 Offset, MCCFIInstruction::createOffset(nullptr, Instr.Ops[0], 538 DataAlignment * Instr.Ops[1])); 539 break; 540 case DW_CFA_restore_extended: 541 case DW_CFA_restore: 542 Function.addCFIInstruction( 543 Offset, MCCFIInstruction::createRestore(nullptr, Instr.Ops[0])); 544 break; 545 case DW_CFA_set_loc: 546 assert(Instr.Ops[0] >= Address && "set_loc out of function bounds"); 547 assert(Instr.Ops[0] <= Address + Function.getSize() && 548 "set_loc out of function bounds"); 549 Offset = Instr.Ops[0] - Address; 550 break; 551 552 case DW_CFA_undefined: 553 Function.addCFIInstruction( 554 Offset, MCCFIInstruction::createUndefined(nullptr, Instr.Ops[0])); 555 break; 556 case DW_CFA_same_value: 557 Function.addCFIInstruction( 558 Offset, MCCFIInstruction::createSameValue(nullptr, Instr.Ops[0])); 559 break; 560 case DW_CFA_register: 561 Function.addCFIInstruction( 562 Offset, MCCFIInstruction::createRegister(nullptr, Instr.Ops[0], 563 Instr.Ops[1])); 564 break; 565 case DW_CFA_remember_state: 566 Function.addCFIInstruction( 567 Offset, MCCFIInstruction::createRememberState(nullptr)); 568 break; 569 case DW_CFA_restore_state: 570 Function.addCFIInstruction(Offset, 571 MCCFIInstruction::createRestoreState(nullptr)); 572 break; 573 case DW_CFA_def_cfa: 574 Function.addCFIInstruction( 575 Offset, 576 MCCFIInstruction::cfiDefCfa(nullptr, Instr.Ops[0], Instr.Ops[1])); 577 break; 578 case DW_CFA_def_cfa_sf: 579 Function.addCFIInstruction( 580 Offset, 581 MCCFIInstruction::cfiDefCfa(nullptr, Instr.Ops[0], 582 DataAlignment * int64_t(Instr.Ops[1]))); 583 break; 584 case DW_CFA_def_cfa_register: 585 Function.addCFIInstruction(Offset, MCCFIInstruction::createDefCfaRegister( 586 nullptr, Instr.Ops[0])); 587 break; 588 case DW_CFA_def_cfa_offset: 589 Function.addCFIInstruction( 590 Offset, MCCFIInstruction::cfiDefCfaOffset(nullptr, Instr.Ops[0])); 591 break; 592 case DW_CFA_def_cfa_offset_sf: 593 Function.addCFIInstruction( 594 Offset, MCCFIInstruction::cfiDefCfaOffset( 595 nullptr, DataAlignment * int64_t(Instr.Ops[0]))); 596 break; 597 case DW_CFA_GNU_args_size: 598 Function.addCFIInstruction( 599 Offset, MCCFIInstruction::createGnuArgsSize(nullptr, Instr.Ops[0])); 600 Function.setUsesGnuArgsSize(); 601 break; 602 case DW_CFA_val_offset_sf: 603 case DW_CFA_val_offset: 604 if (opts::Verbosity >= 1) { 605 errs() << "BOLT-WARNING: DWARF val_offset() unimplemented\n"; 606 } 607 return false; 608 case DW_CFA_def_cfa_expression: 609 case DW_CFA_val_expression: 610 case DW_CFA_expression: { 611 StringRef ExprBytes = Instr.Expression->getData(); 612 std::string Str; 613 raw_string_ostream OS(Str); 614 // Manually encode this instruction using CFI escape 615 OS << Opcode; 616 if (Opcode != DW_CFA_def_cfa_expression) 617 encodeULEB128(Instr.Ops[0], OS); 618 encodeULEB128(ExprBytes.size(), OS); 619 OS << ExprBytes; 620 Function.addCFIInstruction( 621 Offset, MCCFIInstruction::createEscape(nullptr, OS.str())); 622 break; 623 } 624 case DW_CFA_MIPS_advance_loc8: 625 if (opts::Verbosity >= 1) 626 errs() << "BOLT-WARNING: DW_CFA_MIPS_advance_loc unimplemented\n"; 627 return false; 628 case DW_CFA_GNU_window_save: 629 // DW_CFA_GNU_window_save and DW_CFA_GNU_NegateRAState just use the same 630 // id but mean different things. The latter is used in AArch64. 631 if (Function.getBinaryContext().isAArch64()) { 632 Function.addCFIInstruction( 633 Offset, MCCFIInstruction::createNegateRAState(nullptr)); 634 break; 635 } 636 if (opts::Verbosity >= 1) 637 errs() << "BOLT-WARNING: DW_CFA_GNU_window_save unimplemented\n"; 638 return false; 639 case DW_CFA_lo_user: 640 case DW_CFA_hi_user: 641 if (opts::Verbosity >= 1) 642 errs() << "BOLT-WARNING: DW_CFA_*_user unimplemented\n"; 643 return false; 644 default: 645 if (opts::Verbosity >= 1) 646 errs() << "BOLT-WARNING: Unrecognized CFI instruction: " << Instr.Opcode 647 << '\n'; 648 return false; 649 } 650 651 return true; 652 }; 653 654 for (const CFIProgram::Instruction &Instr : CurFDE.getLinkedCIE()->cfis()) 655 if (!decodeFrameInstruction(Instr)) 656 return false; 657 658 for (const CFIProgram::Instruction &Instr : CurFDE.cfis()) 659 if (!decodeFrameInstruction(Instr)) 660 return false; 661 662 return true; 663 } 664 665 std::vector<char> CFIReaderWriter::generateEHFrameHeader( 666 const DWARFDebugFrame &OldEHFrame, const DWARFDebugFrame &NewEHFrame, 667 uint64_t EHFrameHeaderAddress, 668 std::vector<uint64_t> &FailedAddresses) const { 669 // Common PC -> FDE map to be written into .eh_frame_hdr. 670 std::map<uint64_t, uint64_t> PCToFDE; 671 672 // Presort array for binary search. 673 llvm::sort(FailedAddresses); 674 675 // Initialize PCToFDE using NewEHFrame. 676 for (dwarf::FrameEntry &Entry : NewEHFrame.entries()) { 677 const dwarf::FDE *FDE = dyn_cast<dwarf::FDE>(&Entry); 678 if (FDE == nullptr) 679 continue; 680 const uint64_t FuncAddress = FDE->getInitialLocation(); 681 const uint64_t FDEAddress = 682 NewEHFrame.getEHFrameAddress() + FDE->getOffset(); 683 684 // Ignore unused FDEs. 685 if (FuncAddress == 0) 686 continue; 687 688 // Add the address to the map unless we failed to write it. 689 if (!std::binary_search(FailedAddresses.begin(), FailedAddresses.end(), 690 FuncAddress)) { 691 LLVM_DEBUG(dbgs() << "BOLT-DEBUG: FDE for function at 0x" 692 << Twine::utohexstr(FuncAddress) << " is at 0x" 693 << Twine::utohexstr(FDEAddress) << '\n'); 694 PCToFDE[FuncAddress] = FDEAddress; 695 } 696 }; 697 698 LLVM_DEBUG(dbgs() << "BOLT-DEBUG: new .eh_frame contains " 699 << llvm::size(NewEHFrame.entries()) << " entries\n"); 700 701 // Add entries from the original .eh_frame corresponding to the functions 702 // that we did not update. 703 for (const dwarf::FrameEntry &Entry : OldEHFrame) { 704 const dwarf::FDE *FDE = dyn_cast<dwarf::FDE>(&Entry); 705 if (FDE == nullptr) 706 continue; 707 const uint64_t FuncAddress = FDE->getInitialLocation(); 708 const uint64_t FDEAddress = 709 OldEHFrame.getEHFrameAddress() + FDE->getOffset(); 710 711 // Add the address if we failed to write it. 712 if (PCToFDE.count(FuncAddress) == 0) { 713 LLVM_DEBUG(dbgs() << "BOLT-DEBUG: old FDE for function at 0x" 714 << Twine::utohexstr(FuncAddress) << " is at 0x" 715 << Twine::utohexstr(FDEAddress) << '\n'); 716 PCToFDE[FuncAddress] = FDEAddress; 717 } 718 }; 719 720 LLVM_DEBUG(dbgs() << "BOLT-DEBUG: old .eh_frame contains " 721 << llvm::size(OldEHFrame.entries()) << " entries\n"); 722 723 // Generate a new .eh_frame_hdr based on the new map. 724 725 // Header plus table of entries of size 8 bytes. 726 std::vector<char> EHFrameHeader(12 + PCToFDE.size() * 8); 727 728 // Version is 1. 729 EHFrameHeader[0] = 1; 730 // Encoding of the eh_frame pointer. 731 EHFrameHeader[1] = DW_EH_PE_pcrel | DW_EH_PE_sdata4; 732 // Encoding of the count field to follow. 733 EHFrameHeader[2] = DW_EH_PE_udata4; 734 // Encoding of the table entries - 4-byte offset from the start of the header. 735 EHFrameHeader[3] = DW_EH_PE_datarel | DW_EH_PE_sdata4; 736 737 // Address of eh_frame. Use the new one. 738 support::ulittle32_t::ref(EHFrameHeader.data() + 4) = 739 NewEHFrame.getEHFrameAddress() - (EHFrameHeaderAddress + 4); 740 741 // Number of entries in the table (FDE count). 742 support::ulittle32_t::ref(EHFrameHeader.data() + 8) = PCToFDE.size(); 743 744 // Write the table at offset 12. 745 char *Ptr = EHFrameHeader.data(); 746 uint32_t Offset = 12; 747 for (const auto &PCI : PCToFDE) { 748 int64_t InitialPCOffset = PCI.first - EHFrameHeaderAddress; 749 assert(isInt<32>(InitialPCOffset) && "PC offset out of bounds"); 750 support::ulittle32_t::ref(Ptr + Offset) = InitialPCOffset; 751 Offset += 4; 752 int64_t FDEOffset = PCI.second - EHFrameHeaderAddress; 753 assert(isInt<32>(FDEOffset) && "FDE offset out of bounds"); 754 support::ulittle32_t::ref(Ptr + Offset) = FDEOffset; 755 Offset += 4; 756 } 757 758 return EHFrameHeader; 759 } 760 761 Error EHFrameParser::parseCIE(uint64_t StartOffset) { 762 uint8_t Version = Data.getU8(&Offset); 763 const char *Augmentation = Data.getCStr(&Offset); 764 StringRef AugmentationString(Augmentation ? Augmentation : ""); 765 uint8_t AddressSize = 766 Version < 4 ? Data.getAddressSize() : Data.getU8(&Offset); 767 Data.setAddressSize(AddressSize); 768 // Skip segment descriptor size 769 if (Version >= 4) 770 Offset += 1; 771 // Skip code alignment factor 772 Data.getULEB128(&Offset); 773 // Skip data alignment 774 Data.getSLEB128(&Offset); 775 // Skip return address register 776 if (Version == 1) 777 Offset += 1; 778 else 779 Data.getULEB128(&Offset); 780 781 uint32_t FDEPointerEncoding = DW_EH_PE_absptr; 782 uint32_t LSDAPointerEncoding = DW_EH_PE_omit; 783 // Walk the augmentation string to get all the augmentation data. 784 for (unsigned i = 0, e = AugmentationString.size(); i != e; ++i) { 785 switch (AugmentationString[i]) { 786 default: 787 return createStringError( 788 errc::invalid_argument, 789 "unknown augmentation character in entry at 0x%" PRIx64, StartOffset); 790 case 'L': 791 LSDAPointerEncoding = Data.getU8(&Offset); 792 break; 793 case 'P': { 794 uint32_t PersonalityEncoding = Data.getU8(&Offset); 795 std::optional<uint64_t> Personality = 796 Data.getEncodedPointer(&Offset, PersonalityEncoding, 797 EHFrameAddress ? EHFrameAddress + Offset : 0); 798 // Patch personality address 799 if (Personality) 800 PatcherCallback(*Personality, Offset, PersonalityEncoding); 801 break; 802 } 803 case 'R': 804 FDEPointerEncoding = Data.getU8(&Offset); 805 break; 806 case 'z': 807 if (i) 808 return createStringError( 809 errc::invalid_argument, 810 "'z' must be the first character at 0x%" PRIx64, StartOffset); 811 // Skip augmentation length 812 Data.getULEB128(&Offset); 813 break; 814 case 'S': 815 case 'B': 816 break; 817 } 818 } 819 Entries.emplace_back(std::make_unique<CIEInfo>( 820 FDEPointerEncoding, LSDAPointerEncoding, AugmentationString)); 821 CIEs[StartOffset] = &*Entries.back(); 822 return Error::success(); 823 } 824 825 Error EHFrameParser::parseFDE(uint64_t CIEPointer, 826 uint64_t StartStructureOffset) { 827 std::optional<uint64_t> LSDAAddress; 828 CIEInfo *Cie = CIEs[StartStructureOffset - CIEPointer]; 829 830 // The address size is encoded in the CIE we reference. 831 if (!Cie) 832 return createStringError(errc::invalid_argument, 833 "parsing FDE data at 0x%" PRIx64 834 " failed due to missing CIE", 835 StartStructureOffset); 836 // Patch initial location 837 if (auto Val = Data.getEncodedPointer(&Offset, Cie->FDEPtrEncoding, 838 EHFrameAddress + Offset)) { 839 PatcherCallback(*Val, Offset, Cie->FDEPtrEncoding); 840 } 841 // Skip address range 842 Data.getEncodedPointer(&Offset, Cie->FDEPtrEncoding, 0); 843 844 // Process augmentation data for this FDE. 845 StringRef AugmentationString = Cie->AugmentationString; 846 if (!AugmentationString.empty() && Cie->LSDAPtrEncoding != DW_EH_PE_omit) { 847 // Skip augmentation length 848 Data.getULEB128(&Offset); 849 LSDAAddress = 850 Data.getEncodedPointer(&Offset, Cie->LSDAPtrEncoding, 851 EHFrameAddress ? Offset + EHFrameAddress : 0); 852 // Patch LSDA address 853 PatcherCallback(*LSDAAddress, Offset, Cie->LSDAPtrEncoding); 854 } 855 return Error::success(); 856 } 857 858 Error EHFrameParser::parse() { 859 while (Data.isValidOffset(Offset)) { 860 const uint64_t StartOffset = Offset; 861 862 uint64_t Length; 863 DwarfFormat Format; 864 std::tie(Length, Format) = Data.getInitialLength(&Offset); 865 866 // If the Length is 0, then this CIE is a terminator 867 if (Length == 0) 868 break; 869 870 const uint64_t StartStructureOffset = Offset; 871 const uint64_t EndStructureOffset = Offset + Length; 872 873 Error Err = Error::success(); 874 const uint64_t Id = Data.getRelocatedValue(4, &Offset, 875 /*SectionIndex=*/nullptr, &Err); 876 if (Err) 877 return Err; 878 879 if (!Id) { 880 if (Error Err = parseCIE(StartOffset)) 881 return Err; 882 } else { 883 if (Error Err = parseFDE(Id, StartStructureOffset)) 884 return Err; 885 } 886 Offset = EndStructureOffset; 887 } 888 889 return Error::success(); 890 } 891 892 Error EHFrameParser::parse(DWARFDataExtractor Data, uint64_t EHFrameAddress, 893 PatcherCallbackTy PatcherCallback) { 894 EHFrameParser Parser(Data, EHFrameAddress, PatcherCallback); 895 return Parser.parse(); 896 } 897 898 } // namespace bolt 899 } // namespace llvm 900