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