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 Error BinaryFunction::parseLSDA(ArrayRef<uint8_t> LSDASectionData, 102 uint64_t LSDASectionAddress) { 103 assert(CurrentState == State::Disassembled && "unexpected function state"); 104 105 if (!getLSDAAddress()) 106 return Error::success(); 107 108 DWARFDataExtractor Data( 109 StringRef(reinterpret_cast<const char *>(LSDASectionData.data()), 110 LSDASectionData.size()), 111 BC.DwCtx->getDWARFObj().isLittleEndian(), 112 BC.DwCtx->getDWARFObj().getAddressSize()); 113 uint64_t Offset = getLSDAAddress() - LSDASectionAddress; 114 assert(Data.isValidOffset(Offset) && "wrong LSDA address"); 115 116 const uint8_t LPStartEncoding = Data.getU8(&Offset); 117 uint64_t LPStart = Address; 118 if (LPStartEncoding != dwarf::DW_EH_PE_omit) { 119 std::optional<uint64_t> MaybeLPStart = Data.getEncodedPointer( 120 &Offset, LPStartEncoding, Offset + LSDASectionAddress); 121 if (!MaybeLPStart) { 122 BC.errs() << "BOLT-ERROR: unsupported LPStartEncoding: " 123 << (unsigned)LPStartEncoding << '\n'; 124 return createFatalBOLTError(""); 125 } 126 LPStart = *MaybeLPStart; 127 } 128 129 const uint8_t TTypeEncoding = Data.getU8(&Offset); 130 LSDATypeEncoding = TTypeEncoding; 131 size_t TTypeEncodingSize = 0; 132 uintptr_t TTypeEnd = 0; 133 if (TTypeEncoding != DW_EH_PE_omit) { 134 TTypeEnd = Data.getULEB128(&Offset); 135 TTypeEncodingSize = BC.getDWARFEncodingSize(TTypeEncoding); 136 } 137 138 if (opts::PrintExceptions) { 139 BC.outs() << "[LSDA at 0x" << Twine::utohexstr(getLSDAAddress()) 140 << " for function " << *this << "]:\n"; 141 BC.outs() << "LPStart Encoding = 0x" << Twine::utohexstr(LPStartEncoding) 142 << '\n'; 143 BC.outs() << "LPStart = 0x" << Twine::utohexstr(LPStart) << '\n'; 144 BC.outs() << "TType Encoding = 0x" << Twine::utohexstr(TTypeEncoding) 145 << '\n'; 146 BC.outs() << "TType End = " << TTypeEnd << '\n'; 147 } 148 149 // Table to store list of indices in type table. Entries are uleb128 values. 150 const uint64_t TypeIndexTableStart = Offset + TTypeEnd; 151 152 // Offset past the last decoded index. 153 uint64_t MaxTypeIndexTableOffset = 0; 154 155 // Max positive index used in type table. 156 unsigned MaxTypeIndex = 0; 157 158 // The actual type info table starts at the same location, but grows in 159 // opposite direction. TTypeEncoding is used to encode stored values. 160 const uint64_t TypeTableStart = Offset + TTypeEnd; 161 162 uint8_t CallSiteEncoding = Data.getU8(&Offset); 163 uint32_t CallSiteTableLength = Data.getULEB128(&Offset); 164 uint64_t CallSiteTableStart = Offset; 165 uint64_t CallSiteTableEnd = CallSiteTableStart + CallSiteTableLength; 166 uint64_t CallSitePtr = CallSiteTableStart; 167 uint64_t ActionTableStart = CallSiteTableEnd; 168 169 if (opts::PrintExceptions) { 170 BC.outs() << "CallSite Encoding = " << (unsigned)CallSiteEncoding << '\n'; 171 BC.outs() << "CallSite table length = " << CallSiteTableLength << '\n'; 172 BC.outs() << '\n'; 173 } 174 175 this->HasEHRanges = CallSitePtr < CallSiteTableEnd; 176 const uint64_t RangeBase = getAddress(); 177 while (CallSitePtr < CallSiteTableEnd) { 178 uint64_t Start = *Data.getEncodedPointer(&CallSitePtr, CallSiteEncoding, 179 CallSitePtr + LSDASectionAddress); 180 uint64_t Length = *Data.getEncodedPointer(&CallSitePtr, CallSiteEncoding, 181 CallSitePtr + LSDASectionAddress); 182 uint64_t LandingPad = *Data.getEncodedPointer( 183 &CallSitePtr, CallSiteEncoding, CallSitePtr + LSDASectionAddress); 184 uint64_t ActionEntry = Data.getULEB128(&CallSitePtr); 185 if (LandingPad) 186 LandingPad += LPStart; 187 188 if (opts::PrintExceptions) { 189 BC.outs() << "Call Site: [0x" << Twine::utohexstr(RangeBase + Start) 190 << ", 0x" << Twine::utohexstr(RangeBase + Start + Length) 191 << "); landing pad: 0x" << Twine::utohexstr(LandingPad) 192 << "; action entry: 0x" << Twine::utohexstr(ActionEntry) 193 << "\n"; 194 BC.outs() << " current offset is " << (CallSitePtr - CallSiteTableStart) 195 << '\n'; 196 } 197 198 // Create a handler entry if necessary. 199 MCSymbol *LPSymbol = nullptr; 200 if (LandingPad) { 201 // Verify if landing pad code is located outside current function 202 // Support landing pad to builtin_unreachable 203 if (LandingPad < Address || LandingPad > Address + getSize()) { 204 BinaryFunction *Fragment = 205 BC.getBinaryFunctionContainingAddress(LandingPad); 206 assert(Fragment != nullptr && 207 "BOLT-ERROR: cannot find landing pad fragment"); 208 BC.addInterproceduralReference(this, Fragment->getAddress()); 209 BC.processInterproceduralReferences(); 210 assert(isParentOrChildOf(*Fragment) && 211 "BOLT-ERROR: cannot have landing pads in different functions"); 212 setHasIndirectTargetToSplitFragment(true); 213 BC.addFragmentsToSkip(this); 214 return Error::success(); 215 } 216 217 const uint64_t LPOffset = LandingPad - getAddress(); 218 if (!getInstructionAtOffset(LPOffset)) { 219 if (opts::Verbosity >= 1) 220 BC.errs() << "BOLT-WARNING: landing pad " 221 << Twine::utohexstr(LPOffset) 222 << " not pointing to an instruction in function " << *this 223 << " - ignoring.\n"; 224 } else { 225 auto Label = Labels.find(LPOffset); 226 if (Label != Labels.end()) { 227 LPSymbol = Label->second; 228 } else { 229 LPSymbol = BC.Ctx->createNamedTempSymbol("LP"); 230 Labels[LPOffset] = LPSymbol; 231 } 232 } 233 } 234 235 // Mark all call instructions in the range. 236 auto II = Instructions.find(Start); 237 auto IE = Instructions.end(); 238 assert(II != IE && "exception range not pointing to an instruction"); 239 do { 240 MCInst &Instruction = II->second; 241 if (BC.MIB->isCall(Instruction) && 242 !BC.MIB->getConditionalTailCall(Instruction)) { 243 assert(!BC.MIB->isInvoke(Instruction) && 244 "overlapping exception ranges detected"); 245 // Add extra operands to a call instruction making it an invoke from 246 // now on. 247 BC.MIB->addEHInfo(Instruction, 248 MCPlus::MCLandingPad(LPSymbol, ActionEntry)); 249 } 250 ++II; 251 } while (II != IE && II->first < Start + Length); 252 253 if (ActionEntry != 0) { 254 auto printType = [&](int Index, raw_ostream &OS) { 255 assert(Index > 0 && "only positive indices are valid"); 256 uint64_t TTEntry = TypeTableStart - Index * TTypeEncodingSize; 257 const uint64_t TTEntryAddress = TTEntry + LSDASectionAddress; 258 uint64_t TypeAddress = 259 *Data.getEncodedPointer(&TTEntry, TTypeEncoding, TTEntryAddress); 260 if ((TTypeEncoding & DW_EH_PE_pcrel) && TypeAddress == TTEntryAddress) 261 TypeAddress = 0; 262 if (TypeAddress == 0) { 263 OS << "<all>"; 264 return; 265 } 266 if (TTypeEncoding & DW_EH_PE_indirect) { 267 ErrorOr<uint64_t> PointerOrErr = BC.getPointerAtAddress(TypeAddress); 268 assert(PointerOrErr && "failed to decode indirect address"); 269 TypeAddress = *PointerOrErr; 270 } 271 if (BinaryData *TypeSymBD = BC.getBinaryDataAtAddress(TypeAddress)) 272 OS << TypeSymBD->getName(); 273 else 274 OS << "0x" << Twine::utohexstr(TypeAddress); 275 }; 276 if (opts::PrintExceptions) 277 BC.outs() << " actions: "; 278 uint64_t ActionPtr = ActionTableStart + ActionEntry - 1; 279 int64_t ActionType; 280 int64_t ActionNext; 281 const char *Sep = ""; 282 do { 283 ActionType = Data.getSLEB128(&ActionPtr); 284 const uint32_t Self = ActionPtr; 285 ActionNext = Data.getSLEB128(&ActionPtr); 286 if (opts::PrintExceptions) 287 BC.outs() << Sep << "(" << ActionType << ", " << ActionNext << ") "; 288 if (ActionType == 0) { 289 if (opts::PrintExceptions) 290 BC.outs() << "cleanup"; 291 } else if (ActionType > 0) { 292 // It's an index into a type table. 293 MaxTypeIndex = 294 std::max(MaxTypeIndex, static_cast<unsigned>(ActionType)); 295 if (opts::PrintExceptions) { 296 BC.outs() << "catch type "; 297 printType(ActionType, BC.outs()); 298 } 299 } else { // ActionType < 0 300 if (opts::PrintExceptions) 301 BC.outs() << "filter exception types "; 302 const char *TSep = ""; 303 // ActionType is a negative *byte* offset into *uleb128-encoded* table 304 // of indices with base 1. 305 // E.g. -1 means offset 0, -2 is offset 1, etc. The indices are 306 // encoded using uleb128 thus we cannot directly dereference them. 307 uint64_t TypeIndexTablePtr = TypeIndexTableStart - ActionType - 1; 308 while (uint64_t Index = Data.getULEB128(&TypeIndexTablePtr)) { 309 MaxTypeIndex = std::max(MaxTypeIndex, static_cast<unsigned>(Index)); 310 if (opts::PrintExceptions) { 311 BC.outs() << TSep; 312 printType(Index, BC.outs()); 313 TSep = ", "; 314 } 315 } 316 MaxTypeIndexTableOffset = std::max( 317 MaxTypeIndexTableOffset, TypeIndexTablePtr - TypeIndexTableStart); 318 } 319 320 Sep = "; "; 321 322 ActionPtr = Self + ActionNext; 323 } while (ActionNext); 324 if (opts::PrintExceptions) 325 BC.outs() << '\n'; 326 } 327 } 328 if (opts::PrintExceptions) 329 BC.outs() << '\n'; 330 331 assert(TypeIndexTableStart + MaxTypeIndexTableOffset <= 332 Data.getData().size() && 333 "LSDA entry has crossed section boundary"); 334 335 if (TTypeEnd) { 336 LSDAActionTable = LSDASectionData.slice( 337 ActionTableStart, TypeIndexTableStart - 338 MaxTypeIndex * TTypeEncodingSize - 339 ActionTableStart); 340 for (unsigned Index = 1; Index <= MaxTypeIndex; ++Index) { 341 uint64_t TTEntry = TypeTableStart - Index * TTypeEncodingSize; 342 const uint64_t TTEntryAddress = TTEntry + LSDASectionAddress; 343 uint64_t TypeAddress = 344 *Data.getEncodedPointer(&TTEntry, TTypeEncoding, TTEntryAddress); 345 if ((TTypeEncoding & DW_EH_PE_pcrel) && (TypeAddress == TTEntryAddress)) 346 TypeAddress = 0; 347 if (TTypeEncoding & DW_EH_PE_indirect) { 348 LSDATypeAddressTable.emplace_back(TypeAddress); 349 if (TypeAddress) { 350 ErrorOr<uint64_t> PointerOrErr = BC.getPointerAtAddress(TypeAddress); 351 assert(PointerOrErr && "failed to decode indirect address"); 352 TypeAddress = *PointerOrErr; 353 } 354 } 355 LSDATypeTable.emplace_back(TypeAddress); 356 } 357 LSDATypeIndexTable = 358 LSDASectionData.slice(TypeIndexTableStart, MaxTypeIndexTableOffset); 359 } 360 return Error::success(); 361 } 362 363 void BinaryFunction::updateEHRanges() { 364 if (getSize() == 0) 365 return; 366 367 assert(CurrentState == State::CFG_Finalized && "unexpected state"); 368 369 // Build call sites table. 370 struct EHInfo { 371 const MCSymbol *LP; // landing pad 372 uint64_t Action; 373 }; 374 375 // Sites to update. 376 CallSitesList Sites; 377 378 for (FunctionFragment &FF : getLayout().fragments()) { 379 // If previous call can throw, this is its exception handler. 380 EHInfo PreviousEH = {nullptr, 0}; 381 382 // Marker for the beginning of exceptions range. 383 const MCSymbol *StartRange = nullptr; 384 385 for (BinaryBasicBlock *const BB : FF) { 386 for (MCInst &Instr : *BB) { 387 if (!BC.MIB->isCall(Instr)) 388 continue; 389 390 // Instruction can throw an exception that should be handled. 391 const bool Throws = BC.MIB->isInvoke(Instr); 392 393 // Ignore the call if it's a continuation of a no-throw gap. 394 if (!Throws && !StartRange) 395 continue; 396 397 // Extract exception handling information from the instruction. 398 const MCSymbol *LP = nullptr; 399 uint64_t Action = 0; 400 if (const std::optional<MCPlus::MCLandingPad> EHInfo = 401 BC.MIB->getEHInfo(Instr)) 402 std::tie(LP, Action) = *EHInfo; 403 404 // No action if the exception handler has not changed. 405 if (Throws && StartRange && PreviousEH.LP == LP && 406 PreviousEH.Action == Action) 407 continue; 408 409 // Same symbol is used for the beginning and the end of the range. 410 MCSymbol *EHSymbol; 411 if (MCSymbol *InstrLabel = BC.MIB->getInstLabel(Instr)) { 412 EHSymbol = InstrLabel; 413 } else { 414 std::unique_lock<llvm::sys::RWMutex> Lock(BC.CtxMutex); 415 EHSymbol = BC.MIB->getOrCreateInstLabel(Instr, "EH", BC.Ctx.get()); 416 } 417 418 // At this point we could be in one of the following states: 419 // 420 // I. Exception handler has changed and we need to close previous range 421 // and start a new one. 422 // 423 // II. Start a new exception range after the gap. 424 // 425 // III. Close current exception range and start a new gap. 426 const MCSymbol *EndRange; 427 if (StartRange) { 428 // I, III: 429 EndRange = EHSymbol; 430 } else { 431 // II: 432 StartRange = EHSymbol; 433 EndRange = nullptr; 434 } 435 436 // Close the previous range. 437 if (EndRange) 438 Sites.emplace_back( 439 FF.getFragmentNum(), 440 CallSite{StartRange, EndRange, PreviousEH.LP, PreviousEH.Action}); 441 442 if (Throws) { 443 // I, II: 444 StartRange = EHSymbol; 445 PreviousEH = EHInfo{LP, Action}; 446 } else { 447 StartRange = nullptr; 448 } 449 } 450 } 451 452 // Check if we need to close the range. 453 if (StartRange) { 454 const MCSymbol *EndRange = getFunctionEndLabel(FF.getFragmentNum()); 455 Sites.emplace_back( 456 FF.getFragmentNum(), 457 CallSite{StartRange, EndRange, PreviousEH.LP, PreviousEH.Action}); 458 } 459 } 460 461 addCallSites(Sites); 462 } 463 464 const uint8_t DWARF_CFI_PRIMARY_OPCODE_MASK = 0xc0; 465 466 CFIReaderWriter::CFIReaderWriter(BinaryContext &BC, 467 const DWARFDebugFrame &EHFrame) 468 : BC(BC) { 469 // Prepare FDEs for fast lookup 470 for (const dwarf::FrameEntry &Entry : EHFrame.entries()) { 471 const auto *CurFDE = dyn_cast<dwarf::FDE>(&Entry); 472 // Skip CIEs. 473 if (!CurFDE) 474 continue; 475 // There could me multiple FDEs with the same initial address, and perhaps 476 // different sizes (address ranges). Use the first entry with non-zero size. 477 auto FDEI = FDEs.lower_bound(CurFDE->getInitialLocation()); 478 if (FDEI != FDEs.end() && FDEI->first == CurFDE->getInitialLocation()) { 479 if (CurFDE->getAddressRange()) { 480 if (FDEI->second->getAddressRange() == 0) { 481 FDEI->second = CurFDE; 482 } else if (opts::Verbosity > 0) { 483 BC.errs() << "BOLT-WARNING: different FDEs for function at 0x" 484 << Twine::utohexstr(FDEI->first) 485 << " detected; sizes: " << FDEI->second->getAddressRange() 486 << " and " << CurFDE->getAddressRange() << '\n'; 487 } 488 } 489 } else { 490 FDEs.emplace_hint(FDEI, CurFDE->getInitialLocation(), CurFDE); 491 } 492 } 493 } 494 495 bool CFIReaderWriter::fillCFIInfoFor(BinaryFunction &Function) const { 496 uint64_t Address = Function.getAddress(); 497 auto I = FDEs.find(Address); 498 // Ignore zero-length FDE ranges. 499 if (I == FDEs.end() || !I->second->getAddressRange()) 500 return true; 501 502 const FDE &CurFDE = *I->second; 503 std::optional<uint64_t> LSDA = CurFDE.getLSDAAddress(); 504 Function.setLSDAAddress(LSDA ? *LSDA : 0); 505 506 uint64_t Offset = Function.getFirstInstructionOffset(); 507 uint64_t CodeAlignment = CurFDE.getLinkedCIE()->getCodeAlignmentFactor(); 508 uint64_t DataAlignment = CurFDE.getLinkedCIE()->getDataAlignmentFactor(); 509 if (CurFDE.getLinkedCIE()->getPersonalityAddress()) { 510 Function.setPersonalityFunction( 511 *CurFDE.getLinkedCIE()->getPersonalityAddress()); 512 Function.setPersonalityEncoding( 513 *CurFDE.getLinkedCIE()->getPersonalityEncoding()); 514 } 515 516 auto decodeFrameInstruction = [this, &Function, &Offset, Address, 517 CodeAlignment, DataAlignment]( 518 const CFIProgram::Instruction &Instr) { 519 uint8_t Opcode = Instr.Opcode; 520 if (Opcode & DWARF_CFI_PRIMARY_OPCODE_MASK) 521 Opcode &= DWARF_CFI_PRIMARY_OPCODE_MASK; 522 switch (Instr.Opcode) { 523 case DW_CFA_nop: 524 break; 525 case DW_CFA_advance_loc4: 526 case DW_CFA_advance_loc2: 527 case DW_CFA_advance_loc1: 528 case DW_CFA_advance_loc: 529 // Advance our current address 530 Offset += CodeAlignment * int64_t(Instr.Ops[0]); 531 break; 532 case DW_CFA_offset_extended_sf: 533 Function.addCFIInstruction( 534 Offset, 535 MCCFIInstruction::createOffset( 536 nullptr, Instr.Ops[0], DataAlignment * int64_t(Instr.Ops[1]))); 537 break; 538 case DW_CFA_offset_extended: 539 case DW_CFA_offset: 540 Function.addCFIInstruction( 541 Offset, MCCFIInstruction::createOffset(nullptr, Instr.Ops[0], 542 DataAlignment * Instr.Ops[1])); 543 break; 544 case DW_CFA_restore_extended: 545 case DW_CFA_restore: 546 Function.addCFIInstruction( 547 Offset, MCCFIInstruction::createRestore(nullptr, Instr.Ops[0])); 548 break; 549 case DW_CFA_set_loc: 550 assert(Instr.Ops[0] >= Address && "set_loc out of function bounds"); 551 assert(Instr.Ops[0] <= Address + Function.getSize() && 552 "set_loc out of function bounds"); 553 Offset = Instr.Ops[0] - Address; 554 break; 555 556 case DW_CFA_undefined: 557 Function.addCFIInstruction( 558 Offset, MCCFIInstruction::createUndefined(nullptr, Instr.Ops[0])); 559 break; 560 case DW_CFA_same_value: 561 Function.addCFIInstruction( 562 Offset, MCCFIInstruction::createSameValue(nullptr, Instr.Ops[0])); 563 break; 564 case DW_CFA_register: 565 Function.addCFIInstruction( 566 Offset, MCCFIInstruction::createRegister(nullptr, Instr.Ops[0], 567 Instr.Ops[1])); 568 break; 569 case DW_CFA_remember_state: 570 Function.addCFIInstruction( 571 Offset, MCCFIInstruction::createRememberState(nullptr)); 572 break; 573 case DW_CFA_restore_state: 574 Function.addCFIInstruction(Offset, 575 MCCFIInstruction::createRestoreState(nullptr)); 576 break; 577 case DW_CFA_def_cfa: 578 Function.addCFIInstruction( 579 Offset, 580 MCCFIInstruction::cfiDefCfa(nullptr, Instr.Ops[0], Instr.Ops[1])); 581 break; 582 case DW_CFA_def_cfa_sf: 583 Function.addCFIInstruction( 584 Offset, 585 MCCFIInstruction::cfiDefCfa(nullptr, Instr.Ops[0], 586 DataAlignment * int64_t(Instr.Ops[1]))); 587 break; 588 case DW_CFA_def_cfa_register: 589 Function.addCFIInstruction(Offset, MCCFIInstruction::createDefCfaRegister( 590 nullptr, Instr.Ops[0])); 591 break; 592 case DW_CFA_def_cfa_offset: 593 Function.addCFIInstruction( 594 Offset, MCCFIInstruction::cfiDefCfaOffset(nullptr, Instr.Ops[0])); 595 break; 596 case DW_CFA_def_cfa_offset_sf: 597 Function.addCFIInstruction( 598 Offset, MCCFIInstruction::cfiDefCfaOffset( 599 nullptr, DataAlignment * int64_t(Instr.Ops[0]))); 600 break; 601 case DW_CFA_GNU_args_size: 602 Function.addCFIInstruction( 603 Offset, MCCFIInstruction::createGnuArgsSize(nullptr, Instr.Ops[0])); 604 Function.setUsesGnuArgsSize(); 605 break; 606 case DW_CFA_val_offset_sf: 607 case DW_CFA_val_offset: 608 if (opts::Verbosity >= 1) { 609 BC.errs() << "BOLT-WARNING: DWARF val_offset() unimplemented\n"; 610 } 611 return false; 612 case DW_CFA_def_cfa_expression: 613 case DW_CFA_val_expression: 614 case DW_CFA_expression: { 615 StringRef ExprBytes = Instr.Expression->getData(); 616 std::string Str; 617 raw_string_ostream OS(Str); 618 // Manually encode this instruction using CFI escape 619 OS << Opcode; 620 if (Opcode != DW_CFA_def_cfa_expression) 621 encodeULEB128(Instr.Ops[0], OS); 622 encodeULEB128(ExprBytes.size(), OS); 623 OS << ExprBytes; 624 Function.addCFIInstruction( 625 Offset, MCCFIInstruction::createEscape(nullptr, OS.str())); 626 break; 627 } 628 case DW_CFA_MIPS_advance_loc8: 629 if (opts::Verbosity >= 1) 630 BC.errs() << "BOLT-WARNING: DW_CFA_MIPS_advance_loc unimplemented\n"; 631 return false; 632 case DW_CFA_GNU_window_save: 633 // DW_CFA_GNU_window_save and DW_CFA_GNU_NegateRAState just use the same 634 // id but mean different things. The latter is used in AArch64. 635 if (Function.getBinaryContext().isAArch64()) { 636 Function.addCFIInstruction( 637 Offset, MCCFIInstruction::createNegateRAState(nullptr)); 638 break; 639 } 640 if (opts::Verbosity >= 1) 641 BC.errs() << "BOLT-WARNING: DW_CFA_GNU_window_save unimplemented\n"; 642 return false; 643 case DW_CFA_lo_user: 644 case DW_CFA_hi_user: 645 if (opts::Verbosity >= 1) 646 BC.errs() << "BOLT-WARNING: DW_CFA_*_user unimplemented\n"; 647 return false; 648 default: 649 if (opts::Verbosity >= 1) 650 BC.errs() << "BOLT-WARNING: Unrecognized CFI instruction: " 651 << Instr.Opcode << '\n'; 652 return false; 653 } 654 655 return true; 656 }; 657 658 for (const CFIProgram::Instruction &Instr : CurFDE.getLinkedCIE()->cfis()) 659 if (!decodeFrameInstruction(Instr)) 660 return false; 661 662 for (const CFIProgram::Instruction &Instr : CurFDE.cfis()) 663 if (!decodeFrameInstruction(Instr)) 664 return false; 665 666 return true; 667 } 668 669 std::vector<char> CFIReaderWriter::generateEHFrameHeader( 670 const DWARFDebugFrame &OldEHFrame, const DWARFDebugFrame &NewEHFrame, 671 uint64_t EHFrameHeaderAddress, 672 std::vector<uint64_t> &FailedAddresses) const { 673 // Common PC -> FDE map to be written into .eh_frame_hdr. 674 std::map<uint64_t, uint64_t> PCToFDE; 675 676 // Presort array for binary search. 677 llvm::sort(FailedAddresses); 678 679 // Initialize PCToFDE using NewEHFrame. 680 for (dwarf::FrameEntry &Entry : NewEHFrame.entries()) { 681 const dwarf::FDE *FDE = dyn_cast<dwarf::FDE>(&Entry); 682 if (FDE == nullptr) 683 continue; 684 const uint64_t FuncAddress = FDE->getInitialLocation(); 685 const uint64_t FDEAddress = 686 NewEHFrame.getEHFrameAddress() + FDE->getOffset(); 687 688 // Ignore unused FDEs. 689 if (FuncAddress == 0) 690 continue; 691 692 // Add the address to the map unless we failed to write it. 693 if (!std::binary_search(FailedAddresses.begin(), FailedAddresses.end(), 694 FuncAddress)) { 695 LLVM_DEBUG(dbgs() << "BOLT-DEBUG: FDE for function at 0x" 696 << Twine::utohexstr(FuncAddress) << " is at 0x" 697 << Twine::utohexstr(FDEAddress) << '\n'); 698 PCToFDE[FuncAddress] = FDEAddress; 699 } 700 }; 701 702 LLVM_DEBUG(dbgs() << "BOLT-DEBUG: new .eh_frame contains " 703 << llvm::size(NewEHFrame.entries()) << " entries\n"); 704 705 // Add entries from the original .eh_frame corresponding to the functions 706 // that we did not update. 707 for (const dwarf::FrameEntry &Entry : OldEHFrame) { 708 const dwarf::FDE *FDE = dyn_cast<dwarf::FDE>(&Entry); 709 if (FDE == nullptr) 710 continue; 711 const uint64_t FuncAddress = FDE->getInitialLocation(); 712 const uint64_t FDEAddress = 713 OldEHFrame.getEHFrameAddress() + FDE->getOffset(); 714 715 // Add the address if we failed to write it. 716 if (PCToFDE.count(FuncAddress) == 0) { 717 LLVM_DEBUG(dbgs() << "BOLT-DEBUG: old FDE for function at 0x" 718 << Twine::utohexstr(FuncAddress) << " is at 0x" 719 << Twine::utohexstr(FDEAddress) << '\n'); 720 PCToFDE[FuncAddress] = FDEAddress; 721 } 722 }; 723 724 LLVM_DEBUG(dbgs() << "BOLT-DEBUG: old .eh_frame contains " 725 << llvm::size(OldEHFrame.entries()) << " entries\n"); 726 727 // Generate a new .eh_frame_hdr based on the new map. 728 729 // Header plus table of entries of size 8 bytes. 730 std::vector<char> EHFrameHeader(12 + PCToFDE.size() * 8); 731 732 // Version is 1. 733 EHFrameHeader[0] = 1; 734 // Encoding of the eh_frame pointer. 735 EHFrameHeader[1] = DW_EH_PE_pcrel | DW_EH_PE_sdata4; 736 // Encoding of the count field to follow. 737 EHFrameHeader[2] = DW_EH_PE_udata4; 738 // Encoding of the table entries - 4-byte offset from the start of the header. 739 EHFrameHeader[3] = DW_EH_PE_datarel | DW_EH_PE_sdata4; 740 741 // Address of eh_frame. Use the new one. 742 support::ulittle32_t::ref(EHFrameHeader.data() + 4) = 743 NewEHFrame.getEHFrameAddress() - (EHFrameHeaderAddress + 4); 744 745 // Number of entries in the table (FDE count). 746 support::ulittle32_t::ref(EHFrameHeader.data() + 8) = PCToFDE.size(); 747 748 // Write the table at offset 12. 749 char *Ptr = EHFrameHeader.data(); 750 uint32_t Offset = 12; 751 for (const auto &PCI : PCToFDE) { 752 int64_t InitialPCOffset = PCI.first - EHFrameHeaderAddress; 753 assert(isInt<32>(InitialPCOffset) && "PC offset out of bounds"); 754 support::ulittle32_t::ref(Ptr + Offset) = InitialPCOffset; 755 Offset += 4; 756 int64_t FDEOffset = PCI.second - EHFrameHeaderAddress; 757 assert(isInt<32>(FDEOffset) && "FDE offset out of bounds"); 758 support::ulittle32_t::ref(Ptr + Offset) = FDEOffset; 759 Offset += 4; 760 } 761 762 return EHFrameHeader; 763 } 764 765 Error EHFrameParser::parseCIE(uint64_t StartOffset) { 766 uint8_t Version = Data.getU8(&Offset); 767 const char *Augmentation = Data.getCStr(&Offset); 768 StringRef AugmentationString(Augmentation ? Augmentation : ""); 769 uint8_t AddressSize = 770 Version < 4 ? Data.getAddressSize() : Data.getU8(&Offset); 771 Data.setAddressSize(AddressSize); 772 // Skip segment descriptor size 773 if (Version >= 4) 774 Offset += 1; 775 // Skip code alignment factor 776 Data.getULEB128(&Offset); 777 // Skip data alignment 778 Data.getSLEB128(&Offset); 779 // Skip return address register 780 if (Version == 1) 781 Offset += 1; 782 else 783 Data.getULEB128(&Offset); 784 785 uint32_t FDEPointerEncoding = DW_EH_PE_absptr; 786 uint32_t LSDAPointerEncoding = DW_EH_PE_omit; 787 // Walk the augmentation string to get all the augmentation data. 788 for (unsigned i = 0, e = AugmentationString.size(); i != e; ++i) { 789 switch (AugmentationString[i]) { 790 default: 791 return createStringError( 792 errc::invalid_argument, 793 "unknown augmentation character in entry at 0x%" PRIx64, StartOffset); 794 case 'L': 795 LSDAPointerEncoding = Data.getU8(&Offset); 796 break; 797 case 'P': { 798 uint32_t PersonalityEncoding = Data.getU8(&Offset); 799 std::optional<uint64_t> Personality = 800 Data.getEncodedPointer(&Offset, PersonalityEncoding, 801 EHFrameAddress ? EHFrameAddress + Offset : 0); 802 // Patch personality address 803 if (Personality) 804 PatcherCallback(*Personality, Offset, PersonalityEncoding); 805 break; 806 } 807 case 'R': 808 FDEPointerEncoding = Data.getU8(&Offset); 809 break; 810 case 'z': 811 if (i) 812 return createStringError( 813 errc::invalid_argument, 814 "'z' must be the first character at 0x%" PRIx64, StartOffset); 815 // Skip augmentation length 816 Data.getULEB128(&Offset); 817 break; 818 case 'S': 819 case 'B': 820 break; 821 } 822 } 823 Entries.emplace_back(std::make_unique<CIEInfo>( 824 FDEPointerEncoding, LSDAPointerEncoding, AugmentationString)); 825 CIEs[StartOffset] = &*Entries.back(); 826 return Error::success(); 827 } 828 829 Error EHFrameParser::parseFDE(uint64_t CIEPointer, 830 uint64_t StartStructureOffset) { 831 std::optional<uint64_t> LSDAAddress; 832 CIEInfo *Cie = CIEs[StartStructureOffset - CIEPointer]; 833 834 // The address size is encoded in the CIE we reference. 835 if (!Cie) 836 return createStringError(errc::invalid_argument, 837 "parsing FDE data at 0x%" PRIx64 838 " failed due to missing CIE", 839 StartStructureOffset); 840 // Patch initial location 841 if (auto Val = Data.getEncodedPointer(&Offset, Cie->FDEPtrEncoding, 842 EHFrameAddress + Offset)) { 843 PatcherCallback(*Val, Offset, Cie->FDEPtrEncoding); 844 } 845 // Skip address range 846 Data.getEncodedPointer(&Offset, Cie->FDEPtrEncoding, 0); 847 848 // Process augmentation data for this FDE. 849 StringRef AugmentationString = Cie->AugmentationString; 850 if (!AugmentationString.empty() && Cie->LSDAPtrEncoding != DW_EH_PE_omit) { 851 // Skip augmentation length 852 Data.getULEB128(&Offset); 853 LSDAAddress = 854 Data.getEncodedPointer(&Offset, Cie->LSDAPtrEncoding, 855 EHFrameAddress ? Offset + EHFrameAddress : 0); 856 // Patch LSDA address 857 PatcherCallback(*LSDAAddress, Offset, Cie->LSDAPtrEncoding); 858 } 859 return Error::success(); 860 } 861 862 Error EHFrameParser::parse() { 863 while (Data.isValidOffset(Offset)) { 864 const uint64_t StartOffset = Offset; 865 866 uint64_t Length; 867 DwarfFormat Format; 868 std::tie(Length, Format) = Data.getInitialLength(&Offset); 869 870 // If the Length is 0, then this CIE is a terminator 871 if (Length == 0) 872 break; 873 874 const uint64_t StartStructureOffset = Offset; 875 const uint64_t EndStructureOffset = Offset + Length; 876 877 Error Err = Error::success(); 878 const uint64_t Id = Data.getRelocatedValue(4, &Offset, 879 /*SectionIndex=*/nullptr, &Err); 880 if (Err) 881 return Err; 882 883 if (!Id) { 884 if (Error Err = parseCIE(StartOffset)) 885 return Err; 886 } else { 887 if (Error Err = parseFDE(Id, StartStructureOffset)) 888 return Err; 889 } 890 Offset = EndStructureOffset; 891 } 892 893 return Error::success(); 894 } 895 896 Error EHFrameParser::parse(DWARFDataExtractor Data, uint64_t EHFrameAddress, 897 PatcherCallbackTy PatcherCallback) { 898 EHFrameParser Parser(Data, EHFrameAddress, PatcherCallback); 899 return Parser.parse(); 900 } 901 902 } // namespace bolt 903 } // namespace llvm 904