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 assert(isParentOrChildOf(*Fragment) && 192 "BOLT-ERROR: cannot have landing pads in different functions"); 193 setHasIndirectTargetToSplitFragment(true); 194 BC.addFragmentsToSkip(this); 195 return; 196 } 197 198 if (opts::PrintExceptions) { 199 outs() << "Call Site: [0x" << Twine::utohexstr(RangeBase + Start) 200 << ", 0x" << Twine::utohexstr(RangeBase + Start + Length) 201 << "); landing pad: 0x" << Twine::utohexstr(LPOffset) 202 << "; action entry: 0x" << Twine::utohexstr(ActionEntry) << "\n"; 203 outs() << " current offset is " << (CallSitePtr - CallSiteTableStart) 204 << '\n'; 205 } 206 207 // Create a handler entry if necessary. 208 MCSymbol *LPSymbol = nullptr; 209 if (LPOffset) { 210 if (!getInstructionAtOffset(LPOffset)) { 211 if (opts::Verbosity >= 1) 212 errs() << "BOLT-WARNING: landing pad " << Twine::utohexstr(LPOffset) 213 << " not pointing to an instruction in function " << *this 214 << " - ignoring.\n"; 215 } else { 216 auto Label = Labels.find(LPOffset); 217 if (Label != Labels.end()) { 218 LPSymbol = Label->second; 219 } else { 220 LPSymbol = BC.Ctx->createNamedTempSymbol("LP"); 221 Labels[LPOffset] = LPSymbol; 222 } 223 } 224 } 225 226 // Mark all call instructions in the range. 227 auto II = Instructions.find(Start); 228 auto IE = Instructions.end(); 229 assert(II != IE && "exception range not pointing to an instruction"); 230 do { 231 MCInst &Instruction = II->second; 232 if (BC.MIB->isCall(Instruction) && 233 !BC.MIB->getConditionalTailCall(Instruction)) { 234 assert(!BC.MIB->isInvoke(Instruction) && 235 "overlapping exception ranges detected"); 236 // Add extra operands to a call instruction making it an invoke from 237 // now on. 238 BC.MIB->addEHInfo(Instruction, 239 MCPlus::MCLandingPad(LPSymbol, ActionEntry)); 240 } 241 ++II; 242 } while (II != IE && II->first < Start + Length); 243 244 if (ActionEntry != 0) { 245 auto printType = [&](int Index, raw_ostream &OS) { 246 assert(Index > 0 && "only positive indices are valid"); 247 uint64_t TTEntry = TypeTableStart - Index * TTypeEncodingSize; 248 const uint64_t TTEntryAddress = TTEntry + LSDASectionAddress; 249 uint64_t TypeAddress = 250 *Data.getEncodedPointer(&TTEntry, TTypeEncoding, TTEntryAddress); 251 if ((TTypeEncoding & DW_EH_PE_pcrel) && TypeAddress == TTEntryAddress) 252 TypeAddress = 0; 253 if (TypeAddress == 0) { 254 OS << "<all>"; 255 return; 256 } 257 if (TTypeEncoding & DW_EH_PE_indirect) { 258 ErrorOr<uint64_t> PointerOrErr = BC.getPointerAtAddress(TypeAddress); 259 assert(PointerOrErr && "failed to decode indirect address"); 260 TypeAddress = *PointerOrErr; 261 } 262 if (BinaryData *TypeSymBD = BC.getBinaryDataAtAddress(TypeAddress)) 263 OS << TypeSymBD->getName(); 264 else 265 OS << "0x" << Twine::utohexstr(TypeAddress); 266 }; 267 if (opts::PrintExceptions) 268 outs() << " actions: "; 269 uint64_t ActionPtr = ActionTableStart + ActionEntry - 1; 270 int64_t ActionType; 271 int64_t ActionNext; 272 const char *Sep = ""; 273 do { 274 ActionType = Data.getSLEB128(&ActionPtr); 275 const uint32_t Self = ActionPtr; 276 ActionNext = Data.getSLEB128(&ActionPtr); 277 if (opts::PrintExceptions) 278 outs() << Sep << "(" << ActionType << ", " << ActionNext << ") "; 279 if (ActionType == 0) { 280 if (opts::PrintExceptions) 281 outs() << "cleanup"; 282 } else if (ActionType > 0) { 283 // It's an index into a type table. 284 MaxTypeIndex = 285 std::max(MaxTypeIndex, static_cast<unsigned>(ActionType)); 286 if (opts::PrintExceptions) { 287 outs() << "catch type "; 288 printType(ActionType, outs()); 289 } 290 } else { // ActionType < 0 291 if (opts::PrintExceptions) 292 outs() << "filter exception types "; 293 const char *TSep = ""; 294 // ActionType is a negative *byte* offset into *uleb128-encoded* table 295 // of indices with base 1. 296 // E.g. -1 means offset 0, -2 is offset 1, etc. The indices are 297 // encoded using uleb128 thus we cannot directly dereference them. 298 uint64_t TypeIndexTablePtr = TypeIndexTableStart - ActionType - 1; 299 while (uint64_t Index = Data.getULEB128(&TypeIndexTablePtr)) { 300 MaxTypeIndex = std::max(MaxTypeIndex, static_cast<unsigned>(Index)); 301 if (opts::PrintExceptions) { 302 outs() << TSep; 303 printType(Index, outs()); 304 TSep = ", "; 305 } 306 } 307 MaxTypeIndexTableOffset = std::max( 308 MaxTypeIndexTableOffset, TypeIndexTablePtr - TypeIndexTableStart); 309 } 310 311 Sep = "; "; 312 313 ActionPtr = Self + ActionNext; 314 } while (ActionNext); 315 if (opts::PrintExceptions) 316 outs() << '\n'; 317 } 318 } 319 if (opts::PrintExceptions) 320 outs() << '\n'; 321 322 assert(TypeIndexTableStart + MaxTypeIndexTableOffset <= 323 Data.getData().size() && 324 "LSDA entry has crossed section boundary"); 325 326 if (TTypeEnd) { 327 LSDAActionTable = LSDASectionData.slice( 328 ActionTableStart, TypeIndexTableStart - 329 MaxTypeIndex * TTypeEncodingSize - 330 ActionTableStart); 331 for (unsigned Index = 1; Index <= MaxTypeIndex; ++Index) { 332 uint64_t TTEntry = TypeTableStart - Index * TTypeEncodingSize; 333 const uint64_t TTEntryAddress = TTEntry + LSDASectionAddress; 334 uint64_t TypeAddress = 335 *Data.getEncodedPointer(&TTEntry, TTypeEncoding, TTEntryAddress); 336 if ((TTypeEncoding & DW_EH_PE_pcrel) && (TypeAddress == TTEntryAddress)) 337 TypeAddress = 0; 338 if (TTypeEncoding & DW_EH_PE_indirect) { 339 LSDATypeAddressTable.emplace_back(TypeAddress); 340 if (TypeAddress) { 341 ErrorOr<uint64_t> PointerOrErr = BC.getPointerAtAddress(TypeAddress); 342 assert(PointerOrErr && "failed to decode indirect address"); 343 TypeAddress = *PointerOrErr; 344 } 345 } 346 LSDATypeTable.emplace_back(TypeAddress); 347 } 348 LSDATypeIndexTable = 349 LSDASectionData.slice(TypeIndexTableStart, MaxTypeIndexTableOffset); 350 } 351 } 352 353 void BinaryFunction::updateEHRanges() { 354 if (getSize() == 0) 355 return; 356 357 assert(CurrentState == State::CFG_Finalized && "unexpected state"); 358 359 // Build call sites table. 360 struct EHInfo { 361 const MCSymbol *LP; // landing pad 362 uint64_t Action; 363 }; 364 365 // Sites to update. 366 CallSitesList Sites; 367 368 for (FunctionFragment &FF : getLayout().fragments()) { 369 // If previous call can throw, this is its exception handler. 370 EHInfo PreviousEH = {nullptr, 0}; 371 372 // Marker for the beginning of exceptions range. 373 const MCSymbol *StartRange = nullptr; 374 375 for (BinaryBasicBlock *const BB : FF) { 376 for (MCInst &Instr : *BB) { 377 if (!BC.MIB->isCall(Instr)) 378 continue; 379 380 // Instruction can throw an exception that should be handled. 381 const bool Throws = BC.MIB->isInvoke(Instr); 382 383 // Ignore the call if it's a continuation of a no-throw gap. 384 if (!Throws && !StartRange) 385 continue; 386 387 // Extract exception handling information from the instruction. 388 const MCSymbol *LP = nullptr; 389 uint64_t Action = 0; 390 if (const std::optional<MCPlus::MCLandingPad> EHInfo = 391 BC.MIB->getEHInfo(Instr)) 392 std::tie(LP, Action) = *EHInfo; 393 394 // No action if the exception handler has not changed. 395 if (Throws && StartRange && PreviousEH.LP == LP && 396 PreviousEH.Action == Action) 397 continue; 398 399 // Same symbol is used for the beginning and the end of the range. 400 MCSymbol *EHSymbol; 401 if (auto InstrLabel = BC.MIB->getLabel(Instr)) { 402 EHSymbol = *InstrLabel; 403 } else { 404 std::unique_lock<llvm::sys::RWMutex> Lock(BC.CtxMutex); 405 EHSymbol = BC.Ctx->createNamedTempSymbol("EH"); 406 BC.MIB->setLabel(Instr, EHSymbol); 407 } 408 409 // At this point we could be in one of the following states: 410 // 411 // I. Exception handler has changed and we need to close previous range 412 // and start a new one. 413 // 414 // II. Start a new exception range after the gap. 415 // 416 // III. Close current exception range and start a new gap. 417 const MCSymbol *EndRange; 418 if (StartRange) { 419 // I, III: 420 EndRange = EHSymbol; 421 } else { 422 // II: 423 StartRange = EHSymbol; 424 EndRange = nullptr; 425 } 426 427 // Close the previous range. 428 if (EndRange) 429 Sites.emplace_back( 430 FF.getFragmentNum(), 431 CallSite{StartRange, EndRange, PreviousEH.LP, PreviousEH.Action}); 432 433 if (Throws) { 434 // I, II: 435 StartRange = EHSymbol; 436 PreviousEH = EHInfo{LP, Action}; 437 } else { 438 StartRange = nullptr; 439 } 440 } 441 } 442 443 // Check if we need to close the range. 444 if (StartRange) { 445 const MCSymbol *EndRange = getFunctionEndLabel(FF.getFragmentNum()); 446 Sites.emplace_back( 447 FF.getFragmentNum(), 448 CallSite{StartRange, EndRange, PreviousEH.LP, PreviousEH.Action}); 449 } 450 } 451 452 addCallSites(Sites); 453 } 454 455 const uint8_t DWARF_CFI_PRIMARY_OPCODE_MASK = 0xc0; 456 457 CFIReaderWriter::CFIReaderWriter(const DWARFDebugFrame &EHFrame) { 458 // Prepare FDEs for fast lookup 459 for (const dwarf::FrameEntry &Entry : EHFrame.entries()) { 460 const auto *CurFDE = dyn_cast<dwarf::FDE>(&Entry); 461 // Skip CIEs. 462 if (!CurFDE) 463 continue; 464 // There could me multiple FDEs with the same initial address, and perhaps 465 // different sizes (address ranges). Use the first entry with non-zero size. 466 auto FDEI = FDEs.lower_bound(CurFDE->getInitialLocation()); 467 if (FDEI != FDEs.end() && FDEI->first == CurFDE->getInitialLocation()) { 468 if (CurFDE->getAddressRange()) { 469 if (FDEI->second->getAddressRange() == 0) { 470 FDEI->second = CurFDE; 471 } else if (opts::Verbosity > 0) { 472 errs() << "BOLT-WARNING: different FDEs for function at 0x" 473 << Twine::utohexstr(FDEI->first) 474 << " detected; sizes: " << FDEI->second->getAddressRange() 475 << " and " << CurFDE->getAddressRange() << '\n'; 476 } 477 } 478 } else { 479 FDEs.emplace_hint(FDEI, CurFDE->getInitialLocation(), CurFDE); 480 } 481 } 482 } 483 484 bool CFIReaderWriter::fillCFIInfoFor(BinaryFunction &Function) const { 485 uint64_t Address = Function.getAddress(); 486 auto I = FDEs.find(Address); 487 // Ignore zero-length FDE ranges. 488 if (I == FDEs.end() || !I->second->getAddressRange()) 489 return true; 490 491 const FDE &CurFDE = *I->second; 492 std::optional<uint64_t> LSDA = CurFDE.getLSDAAddress(); 493 Function.setLSDAAddress(LSDA ? *LSDA : 0); 494 495 uint64_t Offset = Function.getFirstInstructionOffset(); 496 uint64_t CodeAlignment = CurFDE.getLinkedCIE()->getCodeAlignmentFactor(); 497 uint64_t DataAlignment = CurFDE.getLinkedCIE()->getDataAlignmentFactor(); 498 if (CurFDE.getLinkedCIE()->getPersonalityAddress()) { 499 Function.setPersonalityFunction( 500 *CurFDE.getLinkedCIE()->getPersonalityAddress()); 501 Function.setPersonalityEncoding( 502 *CurFDE.getLinkedCIE()->getPersonalityEncoding()); 503 } 504 505 auto decodeFrameInstruction = [&Function, &Offset, Address, CodeAlignment, 506 DataAlignment]( 507 const CFIProgram::Instruction &Instr) { 508 uint8_t Opcode = Instr.Opcode; 509 if (Opcode & DWARF_CFI_PRIMARY_OPCODE_MASK) 510 Opcode &= DWARF_CFI_PRIMARY_OPCODE_MASK; 511 switch (Instr.Opcode) { 512 case DW_CFA_nop: 513 break; 514 case DW_CFA_advance_loc4: 515 case DW_CFA_advance_loc2: 516 case DW_CFA_advance_loc1: 517 case DW_CFA_advance_loc: 518 // Advance our current address 519 Offset += CodeAlignment * int64_t(Instr.Ops[0]); 520 break; 521 case DW_CFA_offset_extended_sf: 522 Function.addCFIInstruction( 523 Offset, 524 MCCFIInstruction::createOffset( 525 nullptr, Instr.Ops[0], DataAlignment * int64_t(Instr.Ops[1]))); 526 break; 527 case DW_CFA_offset_extended: 528 case DW_CFA_offset: 529 Function.addCFIInstruction( 530 Offset, MCCFIInstruction::createOffset(nullptr, Instr.Ops[0], 531 DataAlignment * Instr.Ops[1])); 532 break; 533 case DW_CFA_restore_extended: 534 case DW_CFA_restore: 535 Function.addCFIInstruction( 536 Offset, MCCFIInstruction::createRestore(nullptr, Instr.Ops[0])); 537 break; 538 case DW_CFA_set_loc: 539 assert(Instr.Ops[0] >= Address && "set_loc out of function bounds"); 540 assert(Instr.Ops[0] <= Address + Function.getSize() && 541 "set_loc out of function bounds"); 542 Offset = Instr.Ops[0] - Address; 543 break; 544 545 case DW_CFA_undefined: 546 Function.addCFIInstruction( 547 Offset, MCCFIInstruction::createUndefined(nullptr, Instr.Ops[0])); 548 break; 549 case DW_CFA_same_value: 550 Function.addCFIInstruction( 551 Offset, MCCFIInstruction::createSameValue(nullptr, Instr.Ops[0])); 552 break; 553 case DW_CFA_register: 554 Function.addCFIInstruction( 555 Offset, MCCFIInstruction::createRegister(nullptr, Instr.Ops[0], 556 Instr.Ops[1])); 557 break; 558 case DW_CFA_remember_state: 559 Function.addCFIInstruction( 560 Offset, MCCFIInstruction::createRememberState(nullptr)); 561 break; 562 case DW_CFA_restore_state: 563 Function.addCFIInstruction(Offset, 564 MCCFIInstruction::createRestoreState(nullptr)); 565 break; 566 case DW_CFA_def_cfa: 567 Function.addCFIInstruction( 568 Offset, 569 MCCFIInstruction::cfiDefCfa(nullptr, Instr.Ops[0], Instr.Ops[1])); 570 break; 571 case DW_CFA_def_cfa_sf: 572 Function.addCFIInstruction( 573 Offset, 574 MCCFIInstruction::cfiDefCfa(nullptr, Instr.Ops[0], 575 DataAlignment * int64_t(Instr.Ops[1]))); 576 break; 577 case DW_CFA_def_cfa_register: 578 Function.addCFIInstruction(Offset, MCCFIInstruction::createDefCfaRegister( 579 nullptr, Instr.Ops[0])); 580 break; 581 case DW_CFA_def_cfa_offset: 582 Function.addCFIInstruction( 583 Offset, MCCFIInstruction::cfiDefCfaOffset(nullptr, Instr.Ops[0])); 584 break; 585 case DW_CFA_def_cfa_offset_sf: 586 Function.addCFIInstruction( 587 Offset, MCCFIInstruction::cfiDefCfaOffset( 588 nullptr, DataAlignment * int64_t(Instr.Ops[0]))); 589 break; 590 case DW_CFA_GNU_args_size: 591 Function.addCFIInstruction( 592 Offset, MCCFIInstruction::createGnuArgsSize(nullptr, Instr.Ops[0])); 593 Function.setUsesGnuArgsSize(); 594 break; 595 case DW_CFA_val_offset_sf: 596 case DW_CFA_val_offset: 597 if (opts::Verbosity >= 1) { 598 errs() << "BOLT-WARNING: DWARF val_offset() unimplemented\n"; 599 } 600 return false; 601 case DW_CFA_def_cfa_expression: 602 case DW_CFA_val_expression: 603 case DW_CFA_expression: { 604 StringRef ExprBytes = Instr.Expression->getData(); 605 std::string Str; 606 raw_string_ostream OS(Str); 607 // Manually encode this instruction using CFI escape 608 OS << Opcode; 609 if (Opcode != DW_CFA_def_cfa_expression) 610 encodeULEB128(Instr.Ops[0], OS); 611 encodeULEB128(ExprBytes.size(), OS); 612 OS << ExprBytes; 613 Function.addCFIInstruction( 614 Offset, MCCFIInstruction::createEscape(nullptr, OS.str())); 615 break; 616 } 617 case DW_CFA_MIPS_advance_loc8: 618 if (opts::Verbosity >= 1) 619 errs() << "BOLT-WARNING: DW_CFA_MIPS_advance_loc unimplemented\n"; 620 return false; 621 case DW_CFA_GNU_window_save: 622 // DW_CFA_GNU_window_save and DW_CFA_GNU_NegateRAState just use the same 623 // id but mean different things. The latter is used in AArch64. 624 if (Function.getBinaryContext().isAArch64()) { 625 Function.addCFIInstruction( 626 Offset, MCCFIInstruction::createNegateRAState(nullptr)); 627 break; 628 } 629 if (opts::Verbosity >= 1) 630 errs() << "BOLT-WARNING: DW_CFA_GNU_window_save unimplemented\n"; 631 return false; 632 case DW_CFA_lo_user: 633 case DW_CFA_hi_user: 634 if (opts::Verbosity >= 1) 635 errs() << "BOLT-WARNING: DW_CFA_*_user unimplemented\n"; 636 return false; 637 default: 638 if (opts::Verbosity >= 1) 639 errs() << "BOLT-WARNING: Unrecognized CFI instruction: " << Instr.Opcode 640 << '\n'; 641 return false; 642 } 643 644 return true; 645 }; 646 647 for (const CFIProgram::Instruction &Instr : CurFDE.getLinkedCIE()->cfis()) 648 if (!decodeFrameInstruction(Instr)) 649 return false; 650 651 for (const CFIProgram::Instruction &Instr : CurFDE.cfis()) 652 if (!decodeFrameInstruction(Instr)) 653 return false; 654 655 return true; 656 } 657 658 std::vector<char> CFIReaderWriter::generateEHFrameHeader( 659 const DWARFDebugFrame &OldEHFrame, const DWARFDebugFrame &NewEHFrame, 660 uint64_t EHFrameHeaderAddress, 661 std::vector<uint64_t> &FailedAddresses) const { 662 // Common PC -> FDE map to be written into .eh_frame_hdr. 663 std::map<uint64_t, uint64_t> PCToFDE; 664 665 // Presort array for binary search. 666 llvm::sort(FailedAddresses); 667 668 // Initialize PCToFDE using NewEHFrame. 669 for (dwarf::FrameEntry &Entry : NewEHFrame.entries()) { 670 const dwarf::FDE *FDE = dyn_cast<dwarf::FDE>(&Entry); 671 if (FDE == nullptr) 672 continue; 673 const uint64_t FuncAddress = FDE->getInitialLocation(); 674 const uint64_t FDEAddress = 675 NewEHFrame.getEHFrameAddress() + FDE->getOffset(); 676 677 // Ignore unused FDEs. 678 if (FuncAddress == 0) 679 continue; 680 681 // Add the address to the map unless we failed to write it. 682 if (!std::binary_search(FailedAddresses.begin(), FailedAddresses.end(), 683 FuncAddress)) { 684 LLVM_DEBUG(dbgs() << "BOLT-DEBUG: FDE for function at 0x" 685 << Twine::utohexstr(FuncAddress) << " is at 0x" 686 << Twine::utohexstr(FDEAddress) << '\n'); 687 PCToFDE[FuncAddress] = FDEAddress; 688 } 689 }; 690 691 LLVM_DEBUG(dbgs() << "BOLT-DEBUG: new .eh_frame contains " 692 << llvm::size(NewEHFrame.entries()) << " entries\n"); 693 694 // Add entries from the original .eh_frame corresponding to the functions 695 // that we did not update. 696 for (const dwarf::FrameEntry &Entry : OldEHFrame) { 697 const dwarf::FDE *FDE = dyn_cast<dwarf::FDE>(&Entry); 698 if (FDE == nullptr) 699 continue; 700 const uint64_t FuncAddress = FDE->getInitialLocation(); 701 const uint64_t FDEAddress = 702 OldEHFrame.getEHFrameAddress() + FDE->getOffset(); 703 704 // Add the address if we failed to write it. 705 if (PCToFDE.count(FuncAddress) == 0) { 706 LLVM_DEBUG(dbgs() << "BOLT-DEBUG: old FDE for function at 0x" 707 << Twine::utohexstr(FuncAddress) << " is at 0x" 708 << Twine::utohexstr(FDEAddress) << '\n'); 709 PCToFDE[FuncAddress] = FDEAddress; 710 } 711 }; 712 713 LLVM_DEBUG(dbgs() << "BOLT-DEBUG: old .eh_frame contains " 714 << llvm::size(OldEHFrame.entries()) << " entries\n"); 715 716 // Generate a new .eh_frame_hdr based on the new map. 717 718 // Header plus table of entries of size 8 bytes. 719 std::vector<char> EHFrameHeader(12 + PCToFDE.size() * 8); 720 721 // Version is 1. 722 EHFrameHeader[0] = 1; 723 // Encoding of the eh_frame pointer. 724 EHFrameHeader[1] = DW_EH_PE_pcrel | DW_EH_PE_sdata4; 725 // Encoding of the count field to follow. 726 EHFrameHeader[2] = DW_EH_PE_udata4; 727 // Encoding of the table entries - 4-byte offset from the start of the header. 728 EHFrameHeader[3] = DW_EH_PE_datarel | DW_EH_PE_sdata4; 729 730 // Address of eh_frame. Use the new one. 731 support::ulittle32_t::ref(EHFrameHeader.data() + 4) = 732 NewEHFrame.getEHFrameAddress() - (EHFrameHeaderAddress + 4); 733 734 // Number of entries in the table (FDE count). 735 support::ulittle32_t::ref(EHFrameHeader.data() + 8) = PCToFDE.size(); 736 737 // Write the table at offset 12. 738 char *Ptr = EHFrameHeader.data(); 739 uint32_t Offset = 12; 740 for (const auto &PCI : PCToFDE) { 741 int64_t InitialPCOffset = PCI.first - EHFrameHeaderAddress; 742 assert(isInt<32>(InitialPCOffset) && "PC offset out of bounds"); 743 support::ulittle32_t::ref(Ptr + Offset) = InitialPCOffset; 744 Offset += 4; 745 int64_t FDEOffset = PCI.second - EHFrameHeaderAddress; 746 assert(isInt<32>(FDEOffset) && "FDE offset out of bounds"); 747 support::ulittle32_t::ref(Ptr + Offset) = FDEOffset; 748 Offset += 4; 749 } 750 751 return EHFrameHeader; 752 } 753 754 Error EHFrameParser::parseCIE(uint64_t StartOffset) { 755 uint8_t Version = Data.getU8(&Offset); 756 const char *Augmentation = Data.getCStr(&Offset); 757 StringRef AugmentationString(Augmentation ? Augmentation : ""); 758 uint8_t AddressSize = 759 Version < 4 ? Data.getAddressSize() : Data.getU8(&Offset); 760 Data.setAddressSize(AddressSize); 761 // Skip segment descriptor size 762 if (Version >= 4) 763 Offset += 1; 764 // Skip code alignment factor 765 Data.getULEB128(&Offset); 766 // Skip data alignment 767 Data.getSLEB128(&Offset); 768 // Skip return address register 769 if (Version == 1) 770 Offset += 1; 771 else 772 Data.getULEB128(&Offset); 773 774 uint32_t FDEPointerEncoding = DW_EH_PE_absptr; 775 uint32_t LSDAPointerEncoding = DW_EH_PE_omit; 776 // Walk the augmentation string to get all the augmentation data. 777 for (unsigned i = 0, e = AugmentationString.size(); i != e; ++i) { 778 switch (AugmentationString[i]) { 779 default: 780 return createStringError( 781 errc::invalid_argument, 782 "unknown augmentation character in entry at 0x%" PRIx64, StartOffset); 783 case 'L': 784 LSDAPointerEncoding = Data.getU8(&Offset); 785 break; 786 case 'P': { 787 uint32_t PersonalityEncoding = Data.getU8(&Offset); 788 std::optional<uint64_t> Personality = 789 Data.getEncodedPointer(&Offset, PersonalityEncoding, 790 EHFrameAddress ? EHFrameAddress + Offset : 0); 791 // Patch personality address 792 if (Personality) 793 PatcherCallback(*Personality, Offset, PersonalityEncoding); 794 break; 795 } 796 case 'R': 797 FDEPointerEncoding = Data.getU8(&Offset); 798 break; 799 case 'z': 800 if (i) 801 return createStringError( 802 errc::invalid_argument, 803 "'z' must be the first character at 0x%" PRIx64, StartOffset); 804 // Skip augmentation length 805 Data.getULEB128(&Offset); 806 break; 807 case 'S': 808 case 'B': 809 break; 810 } 811 } 812 Entries.emplace_back(std::make_unique<CIEInfo>( 813 FDEPointerEncoding, LSDAPointerEncoding, AugmentationString)); 814 CIEs[StartOffset] = &*Entries.back(); 815 return Error::success(); 816 } 817 818 Error EHFrameParser::parseFDE(uint64_t CIEPointer, 819 uint64_t StartStructureOffset) { 820 std::optional<uint64_t> LSDAAddress; 821 CIEInfo *Cie = CIEs[StartStructureOffset - CIEPointer]; 822 823 // The address size is encoded in the CIE we reference. 824 if (!Cie) 825 return createStringError(errc::invalid_argument, 826 "parsing FDE data at 0x%" PRIx64 827 " failed due to missing CIE", 828 StartStructureOffset); 829 // Patch initial location 830 if (auto Val = Data.getEncodedPointer(&Offset, Cie->FDEPtrEncoding, 831 EHFrameAddress + Offset)) { 832 PatcherCallback(*Val, Offset, Cie->FDEPtrEncoding); 833 } 834 // Skip address range 835 Data.getEncodedPointer(&Offset, Cie->FDEPtrEncoding, 0); 836 837 // Process augmentation data for this FDE. 838 StringRef AugmentationString = Cie->AugmentationString; 839 if (!AugmentationString.empty() && Cie->LSDAPtrEncoding != DW_EH_PE_omit) { 840 // Skip augmentation length 841 Data.getULEB128(&Offset); 842 LSDAAddress = 843 Data.getEncodedPointer(&Offset, Cie->LSDAPtrEncoding, 844 EHFrameAddress ? Offset + EHFrameAddress : 0); 845 // Patch LSDA address 846 PatcherCallback(*LSDAAddress, Offset, Cie->LSDAPtrEncoding); 847 } 848 return Error::success(); 849 } 850 851 Error EHFrameParser::parse() { 852 while (Data.isValidOffset(Offset)) { 853 const uint64_t StartOffset = Offset; 854 855 uint64_t Length; 856 DwarfFormat Format; 857 std::tie(Length, Format) = Data.getInitialLength(&Offset); 858 859 // If the Length is 0, then this CIE is a terminator 860 if (Length == 0) 861 break; 862 863 const uint64_t StartStructureOffset = Offset; 864 const uint64_t EndStructureOffset = Offset + Length; 865 866 Error Err = Error::success(); 867 const uint64_t Id = Data.getRelocatedValue(4, &Offset, 868 /*SectionIndex=*/nullptr, &Err); 869 if (Err) 870 return Err; 871 872 if (!Id) { 873 if (Error Err = parseCIE(StartOffset)) 874 return Err; 875 } else { 876 if (Error Err = parseFDE(Id, StartStructureOffset)) 877 return Err; 878 } 879 Offset = EndStructureOffset; 880 } 881 882 return Error::success(); 883 } 884 885 Error EHFrameParser::parse(DWARFDataExtractor Data, uint64_t EHFrameAddress, 886 PatcherCallbackTy PatcherCallback) { 887 EHFrameParser Parser(Data, EHFrameAddress, PatcherCallback); 888 return Parser.parse(); 889 } 890 891 } // namespace bolt 892 } // namespace llvm 893