1 //===- llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp ----------------------===// 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 contains support for writing Microsoft CodeView debug info. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "CodeViewDebug.h" 14 #include "DwarfExpression.h" 15 #include "llvm/ADT/APSInt.h" 16 #include "llvm/ADT/ArrayRef.h" 17 #include "llvm/ADT/DenseMap.h" 18 #include "llvm/ADT/DenseSet.h" 19 #include "llvm/ADT/MapVector.h" 20 #include "llvm/ADT/None.h" 21 #include "llvm/ADT/Optional.h" 22 #include "llvm/ADT/STLExtras.h" 23 #include "llvm/ADT/SmallString.h" 24 #include "llvm/ADT/SmallVector.h" 25 #include "llvm/ADT/StringRef.h" 26 #include "llvm/ADT/TinyPtrVector.h" 27 #include "llvm/ADT/Triple.h" 28 #include "llvm/ADT/Twine.h" 29 #include "llvm/BinaryFormat/COFF.h" 30 #include "llvm/BinaryFormat/Dwarf.h" 31 #include "llvm/CodeGen/AsmPrinter.h" 32 #include "llvm/CodeGen/LexicalScopes.h" 33 #include "llvm/CodeGen/MachineFrameInfo.h" 34 #include "llvm/CodeGen/MachineFunction.h" 35 #include "llvm/CodeGen/MachineInstr.h" 36 #include "llvm/CodeGen/MachineModuleInfo.h" 37 #include "llvm/CodeGen/MachineOperand.h" 38 #include "llvm/CodeGen/TargetFrameLowering.h" 39 #include "llvm/CodeGen/TargetRegisterInfo.h" 40 #include "llvm/CodeGen/TargetSubtargetInfo.h" 41 #include "llvm/Config/llvm-config.h" 42 #include "llvm/DebugInfo/CodeView/CVTypeVisitor.h" 43 #include "llvm/DebugInfo/CodeView/CodeView.h" 44 #include "llvm/DebugInfo/CodeView/ContinuationRecordBuilder.h" 45 #include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h" 46 #include "llvm/DebugInfo/CodeView/EnumTables.h" 47 #include "llvm/DebugInfo/CodeView/Line.h" 48 #include "llvm/DebugInfo/CodeView/SymbolRecord.h" 49 #include "llvm/DebugInfo/CodeView/TypeDumpVisitor.h" 50 #include "llvm/DebugInfo/CodeView/TypeIndex.h" 51 #include "llvm/DebugInfo/CodeView/TypeRecord.h" 52 #include "llvm/DebugInfo/CodeView/TypeTableCollection.h" 53 #include "llvm/IR/Constants.h" 54 #include "llvm/IR/DataLayout.h" 55 #include "llvm/IR/DebugInfoMetadata.h" 56 #include "llvm/IR/DebugLoc.h" 57 #include "llvm/IR/Function.h" 58 #include "llvm/IR/GlobalValue.h" 59 #include "llvm/IR/GlobalVariable.h" 60 #include "llvm/IR/Metadata.h" 61 #include "llvm/IR/Module.h" 62 #include "llvm/MC/MCAsmInfo.h" 63 #include "llvm/MC/MCContext.h" 64 #include "llvm/MC/MCSectionCOFF.h" 65 #include "llvm/MC/MCStreamer.h" 66 #include "llvm/MC/MCSymbol.h" 67 #include "llvm/Support/BinaryByteStream.h" 68 #include "llvm/Support/BinaryStreamReader.h" 69 #include "llvm/Support/Casting.h" 70 #include "llvm/Support/CommandLine.h" 71 #include "llvm/Support/Compiler.h" 72 #include "llvm/Support/Endian.h" 73 #include "llvm/Support/Error.h" 74 #include "llvm/Support/ErrorHandling.h" 75 #include "llvm/Support/FormatVariadic.h" 76 #include "llvm/Support/Path.h" 77 #include "llvm/Support/SMLoc.h" 78 #include "llvm/Support/ScopedPrinter.h" 79 #include "llvm/Target/TargetLoweringObjectFile.h" 80 #include "llvm/Target/TargetMachine.h" 81 #include <algorithm> 82 #include <cassert> 83 #include <cctype> 84 #include <cstddef> 85 #include <cstdint> 86 #include <iterator> 87 #include <limits> 88 #include <string> 89 #include <utility> 90 #include <vector> 91 92 using namespace llvm; 93 using namespace llvm::codeview; 94 95 static CPUType mapArchToCVCPUType(Triple::ArchType Type) { 96 switch (Type) { 97 case Triple::ArchType::x86: 98 return CPUType::Pentium3; 99 case Triple::ArchType::x86_64: 100 return CPUType::X64; 101 case Triple::ArchType::thumb: 102 return CPUType::Thumb; 103 case Triple::ArchType::aarch64: 104 return CPUType::ARM64; 105 default: 106 report_fatal_error("target architecture doesn't map to a CodeView CPUType"); 107 } 108 } 109 110 CodeViewDebug::CodeViewDebug(AsmPrinter *AP) 111 : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) { 112 // If module doesn't have named metadata anchors or COFF debug section 113 // is not available, skip any debug info related stuff. 114 if (!MMI->getModule()->getNamedMetadata("llvm.dbg.cu") || 115 !AP->getObjFileLowering().getCOFFDebugSymbolsSection()) { 116 Asm = nullptr; 117 MMI->setDebugInfoAvailability(false); 118 return; 119 } 120 // Tell MMI that we have debug info. 121 MMI->setDebugInfoAvailability(true); 122 123 TheCPU = 124 mapArchToCVCPUType(Triple(MMI->getModule()->getTargetTriple()).getArch()); 125 126 collectGlobalVariableInfo(); 127 128 // Check if we should emit type record hashes. 129 ConstantInt *GH = mdconst::extract_or_null<ConstantInt>( 130 MMI->getModule()->getModuleFlag("CodeViewGHash")); 131 EmitDebugGlobalHashes = GH && !GH->isZero(); 132 } 133 134 StringRef CodeViewDebug::getFullFilepath(const DIFile *File) { 135 std::string &Filepath = FileToFilepathMap[File]; 136 if (!Filepath.empty()) 137 return Filepath; 138 139 StringRef Dir = File->getDirectory(), Filename = File->getFilename(); 140 141 // If this is a Unix-style path, just use it as is. Don't try to canonicalize 142 // it textually because one of the path components could be a symlink. 143 if (Dir.startswith("/") || Filename.startswith("/")) { 144 if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix)) 145 return Filename; 146 Filepath = Dir; 147 if (Dir.back() != '/') 148 Filepath += '/'; 149 Filepath += Filename; 150 return Filepath; 151 } 152 153 // Clang emits directory and relative filename info into the IR, but CodeView 154 // operates on full paths. We could change Clang to emit full paths too, but 155 // that would increase the IR size and probably not needed for other users. 156 // For now, just concatenate and canonicalize the path here. 157 if (Filename.find(':') == 1) 158 Filepath = Filename; 159 else 160 Filepath = (Dir + "\\" + Filename).str(); 161 162 // Canonicalize the path. We have to do it textually because we may no longer 163 // have access the file in the filesystem. 164 // First, replace all slashes with backslashes. 165 std::replace(Filepath.begin(), Filepath.end(), '/', '\\'); 166 167 // Remove all "\.\" with "\". 168 size_t Cursor = 0; 169 while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos) 170 Filepath.erase(Cursor, 2); 171 172 // Replace all "\XXX\..\" with "\". Don't try too hard though as the original 173 // path should be well-formatted, e.g. start with a drive letter, etc. 174 Cursor = 0; 175 while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) { 176 // Something's wrong if the path starts with "\..\", abort. 177 if (Cursor == 0) 178 break; 179 180 size_t PrevSlash = Filepath.rfind('\\', Cursor - 1); 181 if (PrevSlash == std::string::npos) 182 // Something's wrong, abort. 183 break; 184 185 Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash); 186 // The next ".." might be following the one we've just erased. 187 Cursor = PrevSlash; 188 } 189 190 // Remove all duplicate backslashes. 191 Cursor = 0; 192 while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos) 193 Filepath.erase(Cursor, 1); 194 195 return Filepath; 196 } 197 198 unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) { 199 StringRef FullPath = getFullFilepath(F); 200 unsigned NextId = FileIdMap.size() + 1; 201 auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId)); 202 if (Insertion.second) { 203 // We have to compute the full filepath and emit a .cv_file directive. 204 ArrayRef<uint8_t> ChecksumAsBytes; 205 FileChecksumKind CSKind = FileChecksumKind::None; 206 if (F->getChecksum()) { 207 std::string Checksum = fromHex(F->getChecksum()->Value); 208 void *CKMem = OS.getContext().allocate(Checksum.size(), 1); 209 memcpy(CKMem, Checksum.data(), Checksum.size()); 210 ChecksumAsBytes = ArrayRef<uint8_t>( 211 reinterpret_cast<const uint8_t *>(CKMem), Checksum.size()); 212 switch (F->getChecksum()->Kind) { 213 case DIFile::CSK_MD5: CSKind = FileChecksumKind::MD5; break; 214 case DIFile::CSK_SHA1: CSKind = FileChecksumKind::SHA1; break; 215 } 216 } 217 bool Success = OS.EmitCVFileDirective(NextId, FullPath, ChecksumAsBytes, 218 static_cast<unsigned>(CSKind)); 219 (void)Success; 220 assert(Success && ".cv_file directive failed"); 221 } 222 return Insertion.first->second; 223 } 224 225 CodeViewDebug::InlineSite & 226 CodeViewDebug::getInlineSite(const DILocation *InlinedAt, 227 const DISubprogram *Inlinee) { 228 auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()}); 229 InlineSite *Site = &SiteInsertion.first->second; 230 if (SiteInsertion.second) { 231 unsigned ParentFuncId = CurFn->FuncId; 232 if (const DILocation *OuterIA = InlinedAt->getInlinedAt()) 233 ParentFuncId = 234 getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram()) 235 .SiteFuncId; 236 237 Site->SiteFuncId = NextFuncId++; 238 OS.EmitCVInlineSiteIdDirective( 239 Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()), 240 InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc()); 241 Site->Inlinee = Inlinee; 242 InlinedSubprograms.insert(Inlinee); 243 getFuncIdForSubprogram(Inlinee); 244 } 245 return *Site; 246 } 247 248 static StringRef getPrettyScopeName(const DIScope *Scope) { 249 StringRef ScopeName = Scope->getName(); 250 if (!ScopeName.empty()) 251 return ScopeName; 252 253 switch (Scope->getTag()) { 254 case dwarf::DW_TAG_enumeration_type: 255 case dwarf::DW_TAG_class_type: 256 case dwarf::DW_TAG_structure_type: 257 case dwarf::DW_TAG_union_type: 258 return "<unnamed-tag>"; 259 case dwarf::DW_TAG_namespace: 260 return "`anonymous namespace'"; 261 } 262 263 return StringRef(); 264 } 265 266 static const DISubprogram *getQualifiedNameComponents( 267 const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) { 268 const DISubprogram *ClosestSubprogram = nullptr; 269 while (Scope != nullptr) { 270 if (ClosestSubprogram == nullptr) 271 ClosestSubprogram = dyn_cast<DISubprogram>(Scope); 272 StringRef ScopeName = getPrettyScopeName(Scope); 273 if (!ScopeName.empty()) 274 QualifiedNameComponents.push_back(ScopeName); 275 Scope = Scope->getScope().resolve(); 276 } 277 return ClosestSubprogram; 278 } 279 280 static std::string getQualifiedName(ArrayRef<StringRef> QualifiedNameComponents, 281 StringRef TypeName) { 282 std::string FullyQualifiedName; 283 for (StringRef QualifiedNameComponent : 284 llvm::reverse(QualifiedNameComponents)) { 285 FullyQualifiedName.append(QualifiedNameComponent); 286 FullyQualifiedName.append("::"); 287 } 288 FullyQualifiedName.append(TypeName); 289 return FullyQualifiedName; 290 } 291 292 static std::string getFullyQualifiedName(const DIScope *Scope, StringRef Name) { 293 SmallVector<StringRef, 5> QualifiedNameComponents; 294 getQualifiedNameComponents(Scope, QualifiedNameComponents); 295 return getQualifiedName(QualifiedNameComponents, Name); 296 } 297 298 struct CodeViewDebug::TypeLoweringScope { 299 TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; } 300 ~TypeLoweringScope() { 301 // Don't decrement TypeEmissionLevel until after emitting deferred types, so 302 // inner TypeLoweringScopes don't attempt to emit deferred types. 303 if (CVD.TypeEmissionLevel == 1) 304 CVD.emitDeferredCompleteTypes(); 305 --CVD.TypeEmissionLevel; 306 } 307 CodeViewDebug &CVD; 308 }; 309 310 static std::string getFullyQualifiedName(const DIScope *Ty) { 311 const DIScope *Scope = Ty->getScope().resolve(); 312 return getFullyQualifiedName(Scope, getPrettyScopeName(Ty)); 313 } 314 315 TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) { 316 // No scope means global scope and that uses the zero index. 317 if (!Scope || isa<DIFile>(Scope)) 318 return TypeIndex(); 319 320 assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type"); 321 322 // Check if we've already translated this scope. 323 auto I = TypeIndices.find({Scope, nullptr}); 324 if (I != TypeIndices.end()) 325 return I->second; 326 327 // Build the fully qualified name of the scope. 328 std::string ScopeName = getFullyQualifiedName(Scope); 329 StringIdRecord SID(TypeIndex(), ScopeName); 330 auto TI = TypeTable.writeLeafType(SID); 331 return recordTypeIndexForDINode(Scope, TI); 332 } 333 334 TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) { 335 assert(SP); 336 337 // Check if we've already translated this subprogram. 338 auto I = TypeIndices.find({SP, nullptr}); 339 if (I != TypeIndices.end()) 340 return I->second; 341 342 // The display name includes function template arguments. Drop them to match 343 // MSVC. 344 StringRef DisplayName = SP->getName().split('<').first; 345 346 const DIScope *Scope = SP->getScope().resolve(); 347 TypeIndex TI; 348 if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) { 349 // If the scope is a DICompositeType, then this must be a method. Member 350 // function types take some special handling, and require access to the 351 // subprogram. 352 TypeIndex ClassType = getTypeIndex(Class); 353 MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class), 354 DisplayName); 355 TI = TypeTable.writeLeafType(MFuncId); 356 } else { 357 // Otherwise, this must be a free function. 358 TypeIndex ParentScope = getScopeIndex(Scope); 359 FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName); 360 TI = TypeTable.writeLeafType(FuncId); 361 } 362 363 return recordTypeIndexForDINode(SP, TI); 364 } 365 366 static bool isNonTrivial(const DICompositeType *DCTy) { 367 return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial); 368 } 369 370 static FunctionOptions 371 getFunctionOptions(const DISubroutineType *Ty, 372 const DICompositeType *ClassTy = nullptr, 373 StringRef SPName = StringRef("")) { 374 FunctionOptions FO = FunctionOptions::None; 375 const DIType *ReturnTy = nullptr; 376 if (auto TypeArray = Ty->getTypeArray()) { 377 if (TypeArray.size()) 378 ReturnTy = TypeArray[0].resolve(); 379 } 380 381 if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) { 382 if (isNonTrivial(ReturnDCTy)) 383 FO |= FunctionOptions::CxxReturnUdt; 384 } 385 386 // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison. 387 if (ClassTy && isNonTrivial(ClassTy) && SPName == ClassTy->getName()) { 388 FO |= FunctionOptions::Constructor; 389 390 // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag. 391 392 } 393 return FO; 394 } 395 396 TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP, 397 const DICompositeType *Class) { 398 // Always use the method declaration as the key for the function type. The 399 // method declaration contains the this adjustment. 400 if (SP->getDeclaration()) 401 SP = SP->getDeclaration(); 402 assert(!SP->getDeclaration() && "should use declaration as key"); 403 404 // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide 405 // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}. 406 auto I = TypeIndices.find({SP, Class}); 407 if (I != TypeIndices.end()) 408 return I->second; 409 410 // Make sure complete type info for the class is emitted *after* the member 411 // function type, as the complete class type is likely to reference this 412 // member function type. 413 TypeLoweringScope S(*this); 414 const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0; 415 416 FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName()); 417 TypeIndex TI = lowerTypeMemberFunction( 418 SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO); 419 return recordTypeIndexForDINode(SP, TI, Class); 420 } 421 422 TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node, 423 TypeIndex TI, 424 const DIType *ClassTy) { 425 auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI}); 426 (void)InsertResult; 427 assert(InsertResult.second && "DINode was already assigned a type index"); 428 return TI; 429 } 430 431 unsigned CodeViewDebug::getPointerSizeInBytes() { 432 return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8; 433 } 434 435 void CodeViewDebug::recordLocalVariable(LocalVariable &&Var, 436 const LexicalScope *LS) { 437 if (const DILocation *InlinedAt = LS->getInlinedAt()) { 438 // This variable was inlined. Associate it with the InlineSite. 439 const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram(); 440 InlineSite &Site = getInlineSite(InlinedAt, Inlinee); 441 Site.InlinedLocals.emplace_back(Var); 442 } else { 443 // This variable goes into the corresponding lexical scope. 444 ScopeVariables[LS].emplace_back(Var); 445 } 446 } 447 448 static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs, 449 const DILocation *Loc) { 450 auto B = Locs.begin(), E = Locs.end(); 451 if (std::find(B, E, Loc) == E) 452 Locs.push_back(Loc); 453 } 454 455 void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL, 456 const MachineFunction *MF) { 457 // Skip this instruction if it has the same location as the previous one. 458 if (!DL || DL == PrevInstLoc) 459 return; 460 461 const DIScope *Scope = DL.get()->getScope(); 462 if (!Scope) 463 return; 464 465 // Skip this line if it is longer than the maximum we can record. 466 LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true); 467 if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() || 468 LI.isNeverStepInto()) 469 return; 470 471 ColumnInfo CI(DL.getCol(), /*EndColumn=*/0); 472 if (CI.getStartColumn() != DL.getCol()) 473 return; 474 475 if (!CurFn->HaveLineInfo) 476 CurFn->HaveLineInfo = true; 477 unsigned FileId = 0; 478 if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile()) 479 FileId = CurFn->LastFileId; 480 else 481 FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile()); 482 PrevInstLoc = DL; 483 484 unsigned FuncId = CurFn->FuncId; 485 if (const DILocation *SiteLoc = DL->getInlinedAt()) { 486 const DILocation *Loc = DL.get(); 487 488 // If this location was actually inlined from somewhere else, give it the ID 489 // of the inline call site. 490 FuncId = 491 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId; 492 493 // Ensure we have links in the tree of inline call sites. 494 bool FirstLoc = true; 495 while ((SiteLoc = Loc->getInlinedAt())) { 496 InlineSite &Site = 497 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()); 498 if (!FirstLoc) 499 addLocIfNotPresent(Site.ChildSites, Loc); 500 FirstLoc = false; 501 Loc = SiteLoc; 502 } 503 addLocIfNotPresent(CurFn->ChildSites, Loc); 504 } 505 506 OS.EmitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(), 507 /*PrologueEnd=*/false, /*IsStmt=*/false, 508 DL->getFilename(), SMLoc()); 509 } 510 511 void CodeViewDebug::emitCodeViewMagicVersion() { 512 OS.EmitValueToAlignment(4); 513 OS.AddComment("Debug section magic"); 514 OS.EmitIntValue(COFF::DEBUG_SECTION_MAGIC, 4); 515 } 516 517 void CodeViewDebug::endModule() { 518 if (!Asm || !MMI->hasDebugInfo()) 519 return; 520 521 assert(Asm != nullptr); 522 523 // The COFF .debug$S section consists of several subsections, each starting 524 // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length 525 // of the payload followed by the payload itself. The subsections are 4-byte 526 // aligned. 527 528 // Use the generic .debug$S section, and make a subsection for all the inlined 529 // subprograms. 530 switchToDebugSectionForSymbol(nullptr); 531 532 MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols); 533 emitCompilerInformation(); 534 endCVSubsection(CompilerInfo); 535 536 emitInlineeLinesSubsection(); 537 538 // Emit per-function debug information. 539 for (auto &P : FnDebugInfo) 540 if (!P.first->isDeclarationForLinker()) 541 emitDebugInfoForFunction(P.first, *P.second); 542 543 // Emit global variable debug information. 544 setCurrentSubprogram(nullptr); 545 emitDebugInfoForGlobals(); 546 547 // Emit retained types. 548 emitDebugInfoForRetainedTypes(); 549 550 // Switch back to the generic .debug$S section after potentially processing 551 // comdat symbol sections. 552 switchToDebugSectionForSymbol(nullptr); 553 554 // Emit UDT records for any types used by global variables. 555 if (!GlobalUDTs.empty()) { 556 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 557 emitDebugInfoForUDTs(GlobalUDTs); 558 endCVSubsection(SymbolsEnd); 559 } 560 561 // This subsection holds a file index to offset in string table table. 562 OS.AddComment("File index to string table offset subsection"); 563 OS.EmitCVFileChecksumsDirective(); 564 565 // This subsection holds the string table. 566 OS.AddComment("String table"); 567 OS.EmitCVStringTableDirective(); 568 569 // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol 570 // subsection in the generic .debug$S section at the end. There is no 571 // particular reason for this ordering other than to match MSVC. 572 emitBuildInfo(); 573 574 // Emit type information and hashes last, so that any types we translate while 575 // emitting function info are included. 576 emitTypeInformation(); 577 578 if (EmitDebugGlobalHashes) 579 emitTypeGlobalHashes(); 580 581 clear(); 582 } 583 584 static void emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S, 585 unsigned MaxFixedRecordLength = 0xF00) { 586 // The maximum CV record length is 0xFF00. Most of the strings we emit appear 587 // after a fixed length portion of the record. The fixed length portion should 588 // always be less than 0xF00 (3840) bytes, so truncate the string so that the 589 // overall record size is less than the maximum allowed. 590 SmallString<32> NullTerminatedString( 591 S.take_front(MaxRecordLength - MaxFixedRecordLength - 1)); 592 NullTerminatedString.push_back('\0'); 593 OS.EmitBytes(NullTerminatedString); 594 } 595 596 void CodeViewDebug::emitTypeInformation() { 597 if (TypeTable.empty()) 598 return; 599 600 // Start the .debug$T or .debug$P section with 0x4. 601 OS.SwitchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection()); 602 emitCodeViewMagicVersion(); 603 604 SmallString<8> CommentPrefix; 605 if (OS.isVerboseAsm()) { 606 CommentPrefix += '\t'; 607 CommentPrefix += Asm->MAI->getCommentString(); 608 CommentPrefix += ' '; 609 } 610 611 TypeTableCollection Table(TypeTable.records()); 612 Optional<TypeIndex> B = Table.getFirst(); 613 while (B) { 614 // This will fail if the record data is invalid. 615 CVType Record = Table.getType(*B); 616 617 if (OS.isVerboseAsm()) { 618 // Emit a block comment describing the type record for readability. 619 SmallString<512> CommentBlock; 620 raw_svector_ostream CommentOS(CommentBlock); 621 ScopedPrinter SP(CommentOS); 622 SP.setPrefix(CommentPrefix); 623 TypeDumpVisitor TDV(Table, &SP, false); 624 625 Error E = codeview::visitTypeRecord(Record, *B, TDV); 626 if (E) { 627 logAllUnhandledErrors(std::move(E), errs(), "error: "); 628 llvm_unreachable("produced malformed type record"); 629 } 630 // emitRawComment will insert its own tab and comment string before 631 // the first line, so strip off our first one. It also prints its own 632 // newline. 633 OS.emitRawComment( 634 CommentOS.str().drop_front(CommentPrefix.size() - 1).rtrim()); 635 } 636 OS.EmitBinaryData(Record.str_data()); 637 B = Table.getNext(*B); 638 } 639 } 640 641 void CodeViewDebug::emitTypeGlobalHashes() { 642 if (TypeTable.empty()) 643 return; 644 645 // Start the .debug$H section with the version and hash algorithm, currently 646 // hardcoded to version 0, SHA1. 647 OS.SwitchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection()); 648 649 OS.EmitValueToAlignment(4); 650 OS.AddComment("Magic"); 651 OS.EmitIntValue(COFF::DEBUG_HASHES_SECTION_MAGIC, 4); 652 OS.AddComment("Section Version"); 653 OS.EmitIntValue(0, 2); 654 OS.AddComment("Hash Algorithm"); 655 OS.EmitIntValue(uint16_t(GlobalTypeHashAlg::SHA1_8), 2); 656 657 TypeIndex TI(TypeIndex::FirstNonSimpleIndex); 658 for (const auto &GHR : TypeTable.hashes()) { 659 if (OS.isVerboseAsm()) { 660 // Emit an EOL-comment describing which TypeIndex this hash corresponds 661 // to, as well as the stringified SHA1 hash. 662 SmallString<32> Comment; 663 raw_svector_ostream CommentOS(Comment); 664 CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR); 665 OS.AddComment(Comment); 666 ++TI; 667 } 668 assert(GHR.Hash.size() == 8); 669 StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()), 670 GHR.Hash.size()); 671 OS.EmitBinaryData(S); 672 } 673 } 674 675 static SourceLanguage MapDWLangToCVLang(unsigned DWLang) { 676 switch (DWLang) { 677 case dwarf::DW_LANG_C: 678 case dwarf::DW_LANG_C89: 679 case dwarf::DW_LANG_C99: 680 case dwarf::DW_LANG_C11: 681 case dwarf::DW_LANG_ObjC: 682 return SourceLanguage::C; 683 case dwarf::DW_LANG_C_plus_plus: 684 case dwarf::DW_LANG_C_plus_plus_03: 685 case dwarf::DW_LANG_C_plus_plus_11: 686 case dwarf::DW_LANG_C_plus_plus_14: 687 return SourceLanguage::Cpp; 688 case dwarf::DW_LANG_Fortran77: 689 case dwarf::DW_LANG_Fortran90: 690 case dwarf::DW_LANG_Fortran03: 691 case dwarf::DW_LANG_Fortran08: 692 return SourceLanguage::Fortran; 693 case dwarf::DW_LANG_Pascal83: 694 return SourceLanguage::Pascal; 695 case dwarf::DW_LANG_Cobol74: 696 case dwarf::DW_LANG_Cobol85: 697 return SourceLanguage::Cobol; 698 case dwarf::DW_LANG_Java: 699 return SourceLanguage::Java; 700 case dwarf::DW_LANG_D: 701 return SourceLanguage::D; 702 case dwarf::DW_LANG_Swift: 703 return SourceLanguage::Swift; 704 default: 705 // There's no CodeView representation for this language, and CV doesn't 706 // have an "unknown" option for the language field, so we'll use MASM, 707 // as it's very low level. 708 return SourceLanguage::Masm; 709 } 710 } 711 712 namespace { 713 struct Version { 714 int Part[4]; 715 }; 716 } // end anonymous namespace 717 718 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out 719 // the version number. 720 static Version parseVersion(StringRef Name) { 721 Version V = {{0}}; 722 int N = 0; 723 for (const char C : Name) { 724 if (isdigit(C)) { 725 V.Part[N] *= 10; 726 V.Part[N] += C - '0'; 727 } else if (C == '.') { 728 ++N; 729 if (N >= 4) 730 return V; 731 } else if (N > 0) 732 return V; 733 } 734 return V; 735 } 736 737 void CodeViewDebug::emitCompilerInformation() { 738 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3); 739 uint32_t Flags = 0; 740 741 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 742 const MDNode *Node = *CUs->operands().begin(); 743 const auto *CU = cast<DICompileUnit>(Node); 744 745 // The low byte of the flags indicates the source language. 746 Flags = MapDWLangToCVLang(CU->getSourceLanguage()); 747 // TODO: Figure out which other flags need to be set. 748 749 OS.AddComment("Flags and language"); 750 OS.EmitIntValue(Flags, 4); 751 752 OS.AddComment("CPUType"); 753 OS.EmitIntValue(static_cast<uint64_t>(TheCPU), 2); 754 755 StringRef CompilerVersion = CU->getProducer(); 756 Version FrontVer = parseVersion(CompilerVersion); 757 OS.AddComment("Frontend version"); 758 for (int N = 0; N < 4; ++N) 759 OS.EmitIntValue(FrontVer.Part[N], 2); 760 761 // Some Microsoft tools, like Binscope, expect a backend version number of at 762 // least 8.something, so we'll coerce the LLVM version into a form that 763 // guarantees it'll be big enough without really lying about the version. 764 int Major = 1000 * LLVM_VERSION_MAJOR + 765 10 * LLVM_VERSION_MINOR + 766 LLVM_VERSION_PATCH; 767 // Clamp it for builds that use unusually large version numbers. 768 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 769 Version BackVer = {{ Major, 0, 0, 0 }}; 770 OS.AddComment("Backend version"); 771 for (int N = 0; N < 4; ++N) 772 OS.EmitIntValue(BackVer.Part[N], 2); 773 774 OS.AddComment("Null-terminated compiler version string"); 775 emitNullTerminatedSymbolName(OS, CompilerVersion); 776 777 endSymbolRecord(CompilerEnd); 778 } 779 780 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 781 StringRef S) { 782 StringIdRecord SIR(TypeIndex(0x0), S); 783 return TypeTable.writeLeafType(SIR); 784 } 785 786 void CodeViewDebug::emitBuildInfo() { 787 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 788 // build info. The known prefix is: 789 // - Absolute path of current directory 790 // - Compiler path 791 // - Main source file path, relative to CWD or absolute 792 // - Type server PDB file 793 // - Canonical compiler command line 794 // If frontend and backend compilation are separated (think llc or LTO), it's 795 // not clear if the compiler path should refer to the executable for the 796 // frontend or the backend. Leave it blank for now. 797 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 798 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 799 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 800 const auto *CU = cast<DICompileUnit>(Node); 801 const DIFile *MainSourceFile = CU->getFile(); 802 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 803 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 804 BuildInfoArgs[BuildInfoRecord::SourceFile] = 805 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 806 // FIXME: Path to compiler and command line. PDB is intentionally blank unless 807 // we implement /Zi type servers. 808 BuildInfoRecord BIR(BuildInfoArgs); 809 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 810 811 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 812 // from the module symbols into the type stream. 813 MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 814 MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO); 815 OS.AddComment("LF_BUILDINFO index"); 816 OS.EmitIntValue(BuildInfoIndex.getIndex(), 4); 817 endSymbolRecord(BIEnd); 818 endCVSubsection(BISubsecEnd); 819 } 820 821 void CodeViewDebug::emitInlineeLinesSubsection() { 822 if (InlinedSubprograms.empty()) 823 return; 824 825 OS.AddComment("Inlinee lines subsection"); 826 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 827 828 // We emit the checksum info for files. This is used by debuggers to 829 // determine if a pdb matches the source before loading it. Visual Studio, 830 // for instance, will display a warning that the breakpoints are not valid if 831 // the pdb does not match the source. 832 OS.AddComment("Inlinee lines signature"); 833 OS.EmitIntValue(unsigned(InlineeLinesSignature::Normal), 4); 834 835 for (const DISubprogram *SP : InlinedSubprograms) { 836 assert(TypeIndices.count({SP, nullptr})); 837 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 838 839 OS.AddBlankLine(); 840 unsigned FileId = maybeRecordFile(SP->getFile()); 841 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 842 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 843 OS.AddBlankLine(); 844 OS.AddComment("Type index of inlined function"); 845 OS.EmitIntValue(InlineeIdx.getIndex(), 4); 846 OS.AddComment("Offset into filechecksum table"); 847 OS.EmitCVFileChecksumOffsetDirective(FileId); 848 OS.AddComment("Starting line number"); 849 OS.EmitIntValue(SP->getLine(), 4); 850 } 851 852 endCVSubsection(InlineEnd); 853 } 854 855 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 856 const DILocation *InlinedAt, 857 const InlineSite &Site) { 858 assert(TypeIndices.count({Site.Inlinee, nullptr})); 859 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 860 861 // SymbolRecord 862 MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE); 863 864 OS.AddComment("PtrParent"); 865 OS.EmitIntValue(0, 4); 866 OS.AddComment("PtrEnd"); 867 OS.EmitIntValue(0, 4); 868 OS.AddComment("Inlinee type index"); 869 OS.EmitIntValue(InlineeIdx.getIndex(), 4); 870 871 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 872 unsigned StartLineNum = Site.Inlinee->getLine(); 873 874 OS.EmitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 875 FI.Begin, FI.End); 876 877 endSymbolRecord(InlineEnd); 878 879 emitLocalVariableList(FI, Site.InlinedLocals); 880 881 // Recurse on child inlined call sites before closing the scope. 882 for (const DILocation *ChildSite : Site.ChildSites) { 883 auto I = FI.InlineSites.find(ChildSite); 884 assert(I != FI.InlineSites.end() && 885 "child site not in function inline site map"); 886 emitInlinedCallSite(FI, ChildSite, I->second); 887 } 888 889 // Close the scope. 890 emitEndSymbolRecord(SymbolKind::S_INLINESITE_END); 891 } 892 893 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 894 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 895 // comdat key. A section may be comdat because of -ffunction-sections or 896 // because it is comdat in the IR. 897 MCSectionCOFF *GVSec = 898 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 899 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 900 901 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 902 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 903 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 904 905 OS.SwitchSection(DebugSec); 906 907 // Emit the magic version number if this is the first time we've switched to 908 // this section. 909 if (ComdatDebugSections.insert(DebugSec).second) 910 emitCodeViewMagicVersion(); 911 } 912 913 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 914 // The only supported thunk ordinal is currently the standard type. 915 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 916 FunctionInfo &FI, 917 const MCSymbol *Fn) { 918 std::string FuncName = GlobalValue::dropLLVMManglingEscape(GV->getName()); 919 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 920 921 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 922 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 923 924 // Emit S_THUNK32 925 MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32); 926 OS.AddComment("PtrParent"); 927 OS.EmitIntValue(0, 4); 928 OS.AddComment("PtrEnd"); 929 OS.EmitIntValue(0, 4); 930 OS.AddComment("PtrNext"); 931 OS.EmitIntValue(0, 4); 932 OS.AddComment("Thunk section relative address"); 933 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 934 OS.AddComment("Thunk section index"); 935 OS.EmitCOFFSectionIndex(Fn); 936 OS.AddComment("Code size"); 937 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 938 OS.AddComment("Ordinal"); 939 OS.EmitIntValue(unsigned(ordinal), 1); 940 OS.AddComment("Function name"); 941 emitNullTerminatedSymbolName(OS, FuncName); 942 // Additional fields specific to the thunk ordinal would go here. 943 endSymbolRecord(ThunkRecordEnd); 944 945 // Local variables/inlined routines are purposely omitted here. The point of 946 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 947 948 // Emit S_PROC_ID_END 949 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 950 951 endCVSubsection(SymbolsEnd); 952 } 953 954 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 955 FunctionInfo &FI) { 956 // For each function there is a separate subsection which holds the PC to 957 // file:line table. 958 const MCSymbol *Fn = Asm->getSymbol(GV); 959 assert(Fn); 960 961 // Switch to the to a comdat section, if appropriate. 962 switchToDebugSectionForSymbol(Fn); 963 964 std::string FuncName; 965 auto *SP = GV->getSubprogram(); 966 assert(SP); 967 setCurrentSubprogram(SP); 968 969 if (SP->isThunk()) { 970 emitDebugInfoForThunk(GV, FI, Fn); 971 return; 972 } 973 974 // If we have a display name, build the fully qualified name by walking the 975 // chain of scopes. 976 if (!SP->getName().empty()) 977 FuncName = 978 getFullyQualifiedName(SP->getScope().resolve(), SP->getName()); 979 980 // If our DISubprogram name is empty, use the mangled name. 981 if (FuncName.empty()) 982 FuncName = GlobalValue::dropLLVMManglingEscape(GV->getName()); 983 984 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 985 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 986 OS.EmitCVFPOData(Fn); 987 988 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 989 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 990 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 991 { 992 SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID 993 : SymbolKind::S_GPROC32_ID; 994 MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind); 995 996 // These fields are filled in by tools like CVPACK which run after the fact. 997 OS.AddComment("PtrParent"); 998 OS.EmitIntValue(0, 4); 999 OS.AddComment("PtrEnd"); 1000 OS.EmitIntValue(0, 4); 1001 OS.AddComment("PtrNext"); 1002 OS.EmitIntValue(0, 4); 1003 // This is the important bit that tells the debugger where the function 1004 // code is located and what's its size: 1005 OS.AddComment("Code size"); 1006 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4); 1007 OS.AddComment("Offset after prologue"); 1008 OS.EmitIntValue(0, 4); 1009 OS.AddComment("Offset before epilogue"); 1010 OS.EmitIntValue(0, 4); 1011 OS.AddComment("Function type index"); 1012 OS.EmitIntValue(getFuncIdForSubprogram(GV->getSubprogram()).getIndex(), 4); 1013 OS.AddComment("Function section relative address"); 1014 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1015 OS.AddComment("Function section index"); 1016 OS.EmitCOFFSectionIndex(Fn); 1017 OS.AddComment("Flags"); 1018 OS.EmitIntValue(0, 1); 1019 // Emit the function display name as a null-terminated string. 1020 OS.AddComment("Function name"); 1021 // Truncate the name so we won't overflow the record length field. 1022 emitNullTerminatedSymbolName(OS, FuncName); 1023 endSymbolRecord(ProcRecordEnd); 1024 1025 MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC); 1026 // Subtract out the CSR size since MSVC excludes that and we include it. 1027 OS.AddComment("FrameSize"); 1028 OS.EmitIntValue(FI.FrameSize - FI.CSRSize, 4); 1029 OS.AddComment("Padding"); 1030 OS.EmitIntValue(0, 4); 1031 OS.AddComment("Offset of padding"); 1032 OS.EmitIntValue(0, 4); 1033 OS.AddComment("Bytes of callee saved registers"); 1034 OS.EmitIntValue(FI.CSRSize, 4); 1035 OS.AddComment("Exception handler offset"); 1036 OS.EmitIntValue(0, 4); 1037 OS.AddComment("Exception handler section"); 1038 OS.EmitIntValue(0, 2); 1039 OS.AddComment("Flags (defines frame register)"); 1040 OS.EmitIntValue(uint32_t(FI.FrameProcOpts), 4); 1041 endSymbolRecord(FrameProcEnd); 1042 1043 emitLocalVariableList(FI, FI.Locals); 1044 emitGlobalVariableList(FI.Globals); 1045 emitLexicalBlockList(FI.ChildBlocks, FI); 1046 1047 // Emit inlined call site information. Only emit functions inlined directly 1048 // into the parent function. We'll emit the other sites recursively as part 1049 // of their parent inline site. 1050 for (const DILocation *InlinedAt : FI.ChildSites) { 1051 auto I = FI.InlineSites.find(InlinedAt); 1052 assert(I != FI.InlineSites.end() && 1053 "child site not in function inline site map"); 1054 emitInlinedCallSite(FI, InlinedAt, I->second); 1055 } 1056 1057 for (auto Annot : FI.Annotations) { 1058 MCSymbol *Label = Annot.first; 1059 MDTuple *Strs = cast<MDTuple>(Annot.second); 1060 MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION); 1061 OS.EmitCOFFSecRel32(Label, /*Offset=*/0); 1062 // FIXME: Make sure we don't overflow the max record size. 1063 OS.EmitCOFFSectionIndex(Label); 1064 OS.EmitIntValue(Strs->getNumOperands(), 2); 1065 for (Metadata *MD : Strs->operands()) { 1066 // MDStrings are null terminated, so we can do EmitBytes and get the 1067 // nice .asciz directive. 1068 StringRef Str = cast<MDString>(MD)->getString(); 1069 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1070 OS.EmitBytes(StringRef(Str.data(), Str.size() + 1)); 1071 } 1072 endSymbolRecord(AnnotEnd); 1073 } 1074 1075 for (auto HeapAllocSite : FI.HeapAllocSites) { 1076 MCSymbol *BeginLabel = std::get<0>(HeapAllocSite); 1077 MCSymbol *EndLabel = std::get<1>(HeapAllocSite); 1078 DIType *DITy = std::get<2>(HeapAllocSite); 1079 1080 MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE); 1081 OS.AddComment("Call site offset"); 1082 OS.EmitCOFFSecRel32(BeginLabel, /*Offset=*/0); 1083 OS.AddComment("Call site section index"); 1084 OS.EmitCOFFSectionIndex(BeginLabel); 1085 OS.AddComment("Call instruction length"); 1086 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 1087 OS.AddComment("Type index"); 1088 OS.EmitIntValue(getCompleteTypeIndex(DITy).getIndex(), 4); 1089 endSymbolRecord(HeapAllocEnd); 1090 } 1091 1092 if (SP != nullptr) 1093 emitDebugInfoForUDTs(LocalUDTs); 1094 1095 // We're done with this function. 1096 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1097 } 1098 endCVSubsection(SymbolsEnd); 1099 1100 // We have an assembler directive that takes care of the whole line table. 1101 OS.EmitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1102 } 1103 1104 CodeViewDebug::LocalVarDefRange 1105 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1106 LocalVarDefRange DR; 1107 DR.InMemory = -1; 1108 DR.DataOffset = Offset; 1109 assert(DR.DataOffset == Offset && "truncation"); 1110 DR.IsSubfield = 0; 1111 DR.StructOffset = 0; 1112 DR.CVRegister = CVRegister; 1113 return DR; 1114 } 1115 1116 void CodeViewDebug::collectVariableInfoFromMFTable( 1117 DenseSet<InlinedEntity> &Processed) { 1118 const MachineFunction &MF = *Asm->MF; 1119 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1120 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1121 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1122 1123 for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) { 1124 if (!VI.Var) 1125 continue; 1126 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1127 "Expected inlined-at fields to agree"); 1128 1129 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1130 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1131 1132 // If variable scope is not found then skip this variable. 1133 if (!Scope) 1134 continue; 1135 1136 // If the variable has an attached offset expression, extract it. 1137 // FIXME: Try to handle DW_OP_deref as well. 1138 int64_t ExprOffset = 0; 1139 if (VI.Expr) 1140 if (!VI.Expr->extractIfOffset(ExprOffset)) 1141 continue; 1142 1143 // Get the frame register used and the offset. 1144 unsigned FrameReg = 0; 1145 int FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg); 1146 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1147 1148 // Calculate the label ranges. 1149 LocalVarDefRange DefRange = 1150 createDefRangeMem(CVReg, FrameOffset + ExprOffset); 1151 for (const InsnRange &Range : Scope->getRanges()) { 1152 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1153 const MCSymbol *End = getLabelAfterInsn(Range.second); 1154 End = End ? End : Asm->getFunctionEnd(); 1155 DefRange.Ranges.emplace_back(Begin, End); 1156 } 1157 1158 LocalVariable Var; 1159 Var.DIVar = VI.Var; 1160 Var.DefRanges.emplace_back(std::move(DefRange)); 1161 recordLocalVariable(std::move(Var), Scope); 1162 } 1163 } 1164 1165 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1166 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1167 } 1168 1169 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1170 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1171 } 1172 1173 void CodeViewDebug::calculateRanges( 1174 LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { 1175 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1176 1177 // Calculate the definition ranges. 1178 for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { 1179 const auto &Entry = *I; 1180 if (!Entry.isDbgValue()) 1181 continue; 1182 const MachineInstr *DVInst = Entry.getInstr(); 1183 assert(DVInst->isDebugValue() && "Invalid History entry"); 1184 // FIXME: Find a way to represent constant variables, since they are 1185 // relatively common. 1186 Optional<DbgVariableLocation> Location = 1187 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1188 if (!Location) 1189 continue; 1190 1191 // CodeView can only express variables in register and variables in memory 1192 // at a constant offset from a register. However, for variables passed 1193 // indirectly by pointer, it is common for that pointer to be spilled to a 1194 // stack location. For the special case of one offseted load followed by a 1195 // zero offset load (a pointer spilled to the stack), we change the type of 1196 // the local variable from a value type to a reference type. This tricks the 1197 // debugger into doing the load for us. 1198 if (Var.UseReferenceType) { 1199 // We're using a reference type. Drop the last zero offset load. 1200 if (canUseReferenceType(*Location)) 1201 Location->LoadChain.pop_back(); 1202 else 1203 continue; 1204 } else if (needsReferenceType(*Location)) { 1205 // This location can't be expressed without switching to a reference type. 1206 // Start over using that. 1207 Var.UseReferenceType = true; 1208 Var.DefRanges.clear(); 1209 calculateRanges(Var, Entries); 1210 return; 1211 } 1212 1213 // We can only handle a register or an offseted load of a register. 1214 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1215 continue; 1216 { 1217 LocalVarDefRange DR; 1218 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1219 DR.InMemory = !Location->LoadChain.empty(); 1220 DR.DataOffset = 1221 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1222 if (Location->FragmentInfo) { 1223 DR.IsSubfield = true; 1224 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1225 } else { 1226 DR.IsSubfield = false; 1227 DR.StructOffset = 0; 1228 } 1229 1230 if (Var.DefRanges.empty() || 1231 Var.DefRanges.back().isDifferentLocation(DR)) { 1232 Var.DefRanges.emplace_back(std::move(DR)); 1233 } 1234 } 1235 1236 // Compute the label range. 1237 const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr()); 1238 const MCSymbol *End; 1239 if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { 1240 auto &EndingEntry = Entries[Entry.getEndIndex()]; 1241 End = EndingEntry.isDbgValue() 1242 ? getLabelBeforeInsn(EndingEntry.getInstr()) 1243 : getLabelAfterInsn(EndingEntry.getInstr()); 1244 } else 1245 End = Asm->getFunctionEnd(); 1246 1247 // If the last range end is our begin, just extend the last range. 1248 // Otherwise make a new range. 1249 SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = 1250 Var.DefRanges.back().Ranges; 1251 if (!R.empty() && R.back().second == Begin) 1252 R.back().second = End; 1253 else 1254 R.emplace_back(Begin, End); 1255 1256 // FIXME: Do more range combining. 1257 } 1258 } 1259 1260 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { 1261 DenseSet<InlinedEntity> Processed; 1262 // Grab the variable info that was squirreled away in the MMI side-table. 1263 collectVariableInfoFromMFTable(Processed); 1264 1265 for (const auto &I : DbgValues) { 1266 InlinedEntity IV = I.first; 1267 if (Processed.count(IV)) 1268 continue; 1269 const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first); 1270 const DILocation *InlinedAt = IV.second; 1271 1272 // Instruction ranges, specifying where IV is accessible. 1273 const auto &Entries = I.second; 1274 1275 LexicalScope *Scope = nullptr; 1276 if (InlinedAt) 1277 Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt); 1278 else 1279 Scope = LScopes.findLexicalScope(DIVar->getScope()); 1280 // If variable scope is not found then skip this variable. 1281 if (!Scope) 1282 continue; 1283 1284 LocalVariable Var; 1285 Var.DIVar = DIVar; 1286 1287 calculateRanges(Var, Entries); 1288 recordLocalVariable(std::move(Var), Scope); 1289 } 1290 } 1291 1292 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { 1293 const TargetSubtargetInfo &TSI = MF->getSubtarget(); 1294 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1295 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1296 const Function &GV = MF->getFunction(); 1297 auto Insertion = FnDebugInfo.insert({&GV, llvm::make_unique<FunctionInfo>()}); 1298 assert(Insertion.second && "function already has info"); 1299 CurFn = Insertion.first->second.get(); 1300 CurFn->FuncId = NextFuncId++; 1301 CurFn->Begin = Asm->getFunctionBegin(); 1302 1303 // The S_FRAMEPROC record reports the stack size, and how many bytes of 1304 // callee-saved registers were used. For targets that don't use a PUSH 1305 // instruction (AArch64), this will be zero. 1306 CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); 1307 CurFn->FrameSize = MFI.getStackSize(); 1308 CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); 1309 CurFn->HasStackRealignment = TRI->needsStackRealignment(*MF); 1310 1311 // For this function S_FRAMEPROC record, figure out which codeview register 1312 // will be the frame pointer. 1313 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. 1314 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. 1315 if (CurFn->FrameSize > 0) { 1316 if (!TSI.getFrameLowering()->hasFP(*MF)) { 1317 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1318 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; 1319 } else { 1320 // If there is an FP, parameters are always relative to it. 1321 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; 1322 if (CurFn->HasStackRealignment) { 1323 // If the stack needs realignment, locals are relative to SP or VFRAME. 1324 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1325 } else { 1326 // Otherwise, locals are relative to EBP, and we probably have VLAs or 1327 // other stack adjustments. 1328 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; 1329 } 1330 } 1331 } 1332 1333 // Compute other frame procedure options. 1334 FrameProcedureOptions FPO = FrameProcedureOptions::None; 1335 if (MFI.hasVarSizedObjects()) 1336 FPO |= FrameProcedureOptions::HasAlloca; 1337 if (MF->exposesReturnsTwice()) 1338 FPO |= FrameProcedureOptions::HasSetJmp; 1339 // FIXME: Set HasLongJmp if we ever track that info. 1340 if (MF->hasInlineAsm()) 1341 FPO |= FrameProcedureOptions::HasInlineAssembly; 1342 if (GV.hasPersonalityFn()) { 1343 if (isAsynchronousEHPersonality( 1344 classifyEHPersonality(GV.getPersonalityFn()))) 1345 FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; 1346 else 1347 FPO |= FrameProcedureOptions::HasExceptionHandling; 1348 } 1349 if (GV.hasFnAttribute(Attribute::InlineHint)) 1350 FPO |= FrameProcedureOptions::MarkedInline; 1351 if (GV.hasFnAttribute(Attribute::Naked)) 1352 FPO |= FrameProcedureOptions::Naked; 1353 if (MFI.hasStackProtectorIndex()) 1354 FPO |= FrameProcedureOptions::SecurityChecks; 1355 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); 1356 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); 1357 if (Asm->TM.getOptLevel() != CodeGenOpt::None && 1358 !GV.hasOptSize() && !GV.hasOptNone()) 1359 FPO |= FrameProcedureOptions::OptimizedForSpeed; 1360 // FIXME: Set GuardCfg when it is implemented. 1361 CurFn->FrameProcOpts = FPO; 1362 1363 OS.EmitCVFuncIdDirective(CurFn->FuncId); 1364 1365 // Find the end of the function prolog. First known non-DBG_VALUE and 1366 // non-frame setup location marks the beginning of the function body. 1367 // FIXME: is there a simpler a way to do this? Can we just search 1368 // for the first instruction of the function, not the last of the prolog? 1369 DebugLoc PrologEndLoc; 1370 bool EmptyPrologue = true; 1371 for (const auto &MBB : *MF) { 1372 for (const auto &MI : MBB) { 1373 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 1374 MI.getDebugLoc()) { 1375 PrologEndLoc = MI.getDebugLoc(); 1376 break; 1377 } else if (!MI.isMetaInstruction()) { 1378 EmptyPrologue = false; 1379 } 1380 } 1381 } 1382 1383 // Record beginning of function if we have a non-empty prologue. 1384 if (PrologEndLoc && !EmptyPrologue) { 1385 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); 1386 maybeRecordLocation(FnStartDL, MF); 1387 } 1388 } 1389 1390 static bool shouldEmitUdt(const DIType *T) { 1391 if (!T) 1392 return false; 1393 1394 // MSVC does not emit UDTs for typedefs that are scoped to classes. 1395 if (T->getTag() == dwarf::DW_TAG_typedef) { 1396 if (DIScope *Scope = T->getScope().resolve()) { 1397 switch (Scope->getTag()) { 1398 case dwarf::DW_TAG_structure_type: 1399 case dwarf::DW_TAG_class_type: 1400 case dwarf::DW_TAG_union_type: 1401 return false; 1402 } 1403 } 1404 } 1405 1406 while (true) { 1407 if (!T || T->isForwardDecl()) 1408 return false; 1409 1410 const DIDerivedType *DT = dyn_cast<DIDerivedType>(T); 1411 if (!DT) 1412 return true; 1413 T = DT->getBaseType().resolve(); 1414 } 1415 return true; 1416 } 1417 1418 void CodeViewDebug::addToUDTs(const DIType *Ty) { 1419 // Don't record empty UDTs. 1420 if (Ty->getName().empty()) 1421 return; 1422 if (!shouldEmitUdt(Ty)) 1423 return; 1424 1425 SmallVector<StringRef, 5> QualifiedNameComponents; 1426 const DISubprogram *ClosestSubprogram = getQualifiedNameComponents( 1427 Ty->getScope().resolve(), QualifiedNameComponents); 1428 1429 std::string FullyQualifiedName = 1430 getQualifiedName(QualifiedNameComponents, getPrettyScopeName(Ty)); 1431 1432 if (ClosestSubprogram == nullptr) { 1433 GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1434 } else if (ClosestSubprogram == CurrentSubprogram) { 1435 LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1436 } 1437 1438 // TODO: What if the ClosestSubprogram is neither null or the current 1439 // subprogram? Currently, the UDT just gets dropped on the floor. 1440 // 1441 // The current behavior is not desirable. To get maximal fidelity, we would 1442 // need to perform all type translation before beginning emission of .debug$S 1443 // and then make LocalUDTs a member of FunctionInfo 1444 } 1445 1446 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { 1447 // Generic dispatch for lowering an unknown type. 1448 switch (Ty->getTag()) { 1449 case dwarf::DW_TAG_array_type: 1450 return lowerTypeArray(cast<DICompositeType>(Ty)); 1451 case dwarf::DW_TAG_typedef: 1452 return lowerTypeAlias(cast<DIDerivedType>(Ty)); 1453 case dwarf::DW_TAG_base_type: 1454 return lowerTypeBasic(cast<DIBasicType>(Ty)); 1455 case dwarf::DW_TAG_pointer_type: 1456 if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type") 1457 return lowerTypeVFTableShape(cast<DIDerivedType>(Ty)); 1458 LLVM_FALLTHROUGH; 1459 case dwarf::DW_TAG_reference_type: 1460 case dwarf::DW_TAG_rvalue_reference_type: 1461 return lowerTypePointer(cast<DIDerivedType>(Ty)); 1462 case dwarf::DW_TAG_ptr_to_member_type: 1463 return lowerTypeMemberPointer(cast<DIDerivedType>(Ty)); 1464 case dwarf::DW_TAG_restrict_type: 1465 case dwarf::DW_TAG_const_type: 1466 case dwarf::DW_TAG_volatile_type: 1467 // TODO: add support for DW_TAG_atomic_type here 1468 return lowerTypeModifier(cast<DIDerivedType>(Ty)); 1469 case dwarf::DW_TAG_subroutine_type: 1470 if (ClassTy) { 1471 // The member function type of a member function pointer has no 1472 // ThisAdjustment. 1473 return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy, 1474 /*ThisAdjustment=*/0, 1475 /*IsStaticMethod=*/false); 1476 } 1477 return lowerTypeFunction(cast<DISubroutineType>(Ty)); 1478 case dwarf::DW_TAG_enumeration_type: 1479 return lowerTypeEnum(cast<DICompositeType>(Ty)); 1480 case dwarf::DW_TAG_class_type: 1481 case dwarf::DW_TAG_structure_type: 1482 return lowerTypeClass(cast<DICompositeType>(Ty)); 1483 case dwarf::DW_TAG_union_type: 1484 return lowerTypeUnion(cast<DICompositeType>(Ty)); 1485 case dwarf::DW_TAG_unspecified_type: 1486 if (Ty->getName() == "decltype(nullptr)") 1487 return TypeIndex::NullptrT(); 1488 return TypeIndex::None(); 1489 default: 1490 // Use the null type index. 1491 return TypeIndex(); 1492 } 1493 } 1494 1495 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { 1496 DITypeRef UnderlyingTypeRef = Ty->getBaseType(); 1497 TypeIndex UnderlyingTypeIndex = getTypeIndex(UnderlyingTypeRef); 1498 StringRef TypeName = Ty->getName(); 1499 1500 addToUDTs(Ty); 1501 1502 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && 1503 TypeName == "HRESULT") 1504 return TypeIndex(SimpleTypeKind::HResult); 1505 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && 1506 TypeName == "wchar_t") 1507 return TypeIndex(SimpleTypeKind::WideCharacter); 1508 1509 return UnderlyingTypeIndex; 1510 } 1511 1512 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { 1513 DITypeRef ElementTypeRef = Ty->getBaseType(); 1514 TypeIndex ElementTypeIndex = getTypeIndex(ElementTypeRef); 1515 // IndexType is size_t, which depends on the bitness of the target. 1516 TypeIndex IndexType = getPointerSizeInBytes() == 8 1517 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1518 : TypeIndex(SimpleTypeKind::UInt32Long); 1519 1520 uint64_t ElementSize = getBaseTypeSize(ElementTypeRef) / 8; 1521 1522 // Add subranges to array type. 1523 DINodeArray Elements = Ty->getElements(); 1524 for (int i = Elements.size() - 1; i >= 0; --i) { 1525 const DINode *Element = Elements[i]; 1526 assert(Element->getTag() == dwarf::DW_TAG_subrange_type); 1527 1528 const DISubrange *Subrange = cast<DISubrange>(Element); 1529 assert(Subrange->getLowerBound() == 0 && 1530 "codeview doesn't support subranges with lower bounds"); 1531 int64_t Count = -1; 1532 if (auto *CI = Subrange->getCount().dyn_cast<ConstantInt*>()) 1533 Count = CI->getSExtValue(); 1534 1535 // Forward declarations of arrays without a size and VLAs use a count of -1. 1536 // Emit a count of zero in these cases to match what MSVC does for arrays 1537 // without a size. MSVC doesn't support VLAs, so it's not clear what we 1538 // should do for them even if we could distinguish them. 1539 if (Count == -1) 1540 Count = 0; 1541 1542 // Update the element size and element type index for subsequent subranges. 1543 ElementSize *= Count; 1544 1545 // If this is the outermost array, use the size from the array. It will be 1546 // more accurate if we had a VLA or an incomplete element type size. 1547 uint64_t ArraySize = 1548 (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; 1549 1550 StringRef Name = (i == 0) ? Ty->getName() : ""; 1551 ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); 1552 ElementTypeIndex = TypeTable.writeLeafType(AR); 1553 } 1554 1555 return ElementTypeIndex; 1556 } 1557 1558 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { 1559 TypeIndex Index; 1560 dwarf::TypeKind Kind; 1561 uint32_t ByteSize; 1562 1563 Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); 1564 ByteSize = Ty->getSizeInBits() / 8; 1565 1566 SimpleTypeKind STK = SimpleTypeKind::None; 1567 switch (Kind) { 1568 case dwarf::DW_ATE_address: 1569 // FIXME: Translate 1570 break; 1571 case dwarf::DW_ATE_boolean: 1572 switch (ByteSize) { 1573 case 1: STK = SimpleTypeKind::Boolean8; break; 1574 case 2: STK = SimpleTypeKind::Boolean16; break; 1575 case 4: STK = SimpleTypeKind::Boolean32; break; 1576 case 8: STK = SimpleTypeKind::Boolean64; break; 1577 case 16: STK = SimpleTypeKind::Boolean128; break; 1578 } 1579 break; 1580 case dwarf::DW_ATE_complex_float: 1581 switch (ByteSize) { 1582 case 2: STK = SimpleTypeKind::Complex16; break; 1583 case 4: STK = SimpleTypeKind::Complex32; break; 1584 case 8: STK = SimpleTypeKind::Complex64; break; 1585 case 10: STK = SimpleTypeKind::Complex80; break; 1586 case 16: STK = SimpleTypeKind::Complex128; break; 1587 } 1588 break; 1589 case dwarf::DW_ATE_float: 1590 switch (ByteSize) { 1591 case 2: STK = SimpleTypeKind::Float16; break; 1592 case 4: STK = SimpleTypeKind::Float32; break; 1593 case 6: STK = SimpleTypeKind::Float48; break; 1594 case 8: STK = SimpleTypeKind::Float64; break; 1595 case 10: STK = SimpleTypeKind::Float80; break; 1596 case 16: STK = SimpleTypeKind::Float128; break; 1597 } 1598 break; 1599 case dwarf::DW_ATE_signed: 1600 switch (ByteSize) { 1601 case 1: STK = SimpleTypeKind::SignedCharacter; break; 1602 case 2: STK = SimpleTypeKind::Int16Short; break; 1603 case 4: STK = SimpleTypeKind::Int32; break; 1604 case 8: STK = SimpleTypeKind::Int64Quad; break; 1605 case 16: STK = SimpleTypeKind::Int128Oct; break; 1606 } 1607 break; 1608 case dwarf::DW_ATE_unsigned: 1609 switch (ByteSize) { 1610 case 1: STK = SimpleTypeKind::UnsignedCharacter; break; 1611 case 2: STK = SimpleTypeKind::UInt16Short; break; 1612 case 4: STK = SimpleTypeKind::UInt32; break; 1613 case 8: STK = SimpleTypeKind::UInt64Quad; break; 1614 case 16: STK = SimpleTypeKind::UInt128Oct; break; 1615 } 1616 break; 1617 case dwarf::DW_ATE_UTF: 1618 switch (ByteSize) { 1619 case 2: STK = SimpleTypeKind::Character16; break; 1620 case 4: STK = SimpleTypeKind::Character32; break; 1621 } 1622 break; 1623 case dwarf::DW_ATE_signed_char: 1624 if (ByteSize == 1) 1625 STK = SimpleTypeKind::SignedCharacter; 1626 break; 1627 case dwarf::DW_ATE_unsigned_char: 1628 if (ByteSize == 1) 1629 STK = SimpleTypeKind::UnsignedCharacter; 1630 break; 1631 default: 1632 break; 1633 } 1634 1635 // Apply some fixups based on the source-level type name. 1636 if (STK == SimpleTypeKind::Int32 && Ty->getName() == "long int") 1637 STK = SimpleTypeKind::Int32Long; 1638 if (STK == SimpleTypeKind::UInt32 && Ty->getName() == "long unsigned int") 1639 STK = SimpleTypeKind::UInt32Long; 1640 if (STK == SimpleTypeKind::UInt16Short && 1641 (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t")) 1642 STK = SimpleTypeKind::WideCharacter; 1643 if ((STK == SimpleTypeKind::SignedCharacter || 1644 STK == SimpleTypeKind::UnsignedCharacter) && 1645 Ty->getName() == "char") 1646 STK = SimpleTypeKind::NarrowCharacter; 1647 1648 return TypeIndex(STK); 1649 } 1650 1651 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, 1652 PointerOptions PO) { 1653 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType()); 1654 1655 // Pointers to simple types without any options can use SimpleTypeMode, rather 1656 // than having a dedicated pointer type record. 1657 if (PointeeTI.isSimple() && PO == PointerOptions::None && 1658 PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && 1659 Ty->getTag() == dwarf::DW_TAG_pointer_type) { 1660 SimpleTypeMode Mode = Ty->getSizeInBits() == 64 1661 ? SimpleTypeMode::NearPointer64 1662 : SimpleTypeMode::NearPointer32; 1663 return TypeIndex(PointeeTI.getSimpleKind(), Mode); 1664 } 1665 1666 PointerKind PK = 1667 Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; 1668 PointerMode PM = PointerMode::Pointer; 1669 switch (Ty->getTag()) { 1670 default: llvm_unreachable("not a pointer tag type"); 1671 case dwarf::DW_TAG_pointer_type: 1672 PM = PointerMode::Pointer; 1673 break; 1674 case dwarf::DW_TAG_reference_type: 1675 PM = PointerMode::LValueReference; 1676 break; 1677 case dwarf::DW_TAG_rvalue_reference_type: 1678 PM = PointerMode::RValueReference; 1679 break; 1680 } 1681 1682 if (Ty->isObjectPointer()) 1683 PO |= PointerOptions::Const; 1684 1685 PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); 1686 return TypeTable.writeLeafType(PR); 1687 } 1688 1689 static PointerToMemberRepresentation 1690 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { 1691 // SizeInBytes being zero generally implies that the member pointer type was 1692 // incomplete, which can happen if it is part of a function prototype. In this 1693 // case, use the unknown model instead of the general model. 1694 if (IsPMF) { 1695 switch (Flags & DINode::FlagPtrToMemberRep) { 1696 case 0: 1697 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1698 : PointerToMemberRepresentation::GeneralFunction; 1699 case DINode::FlagSingleInheritance: 1700 return PointerToMemberRepresentation::SingleInheritanceFunction; 1701 case DINode::FlagMultipleInheritance: 1702 return PointerToMemberRepresentation::MultipleInheritanceFunction; 1703 case DINode::FlagVirtualInheritance: 1704 return PointerToMemberRepresentation::VirtualInheritanceFunction; 1705 } 1706 } else { 1707 switch (Flags & DINode::FlagPtrToMemberRep) { 1708 case 0: 1709 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1710 : PointerToMemberRepresentation::GeneralData; 1711 case DINode::FlagSingleInheritance: 1712 return PointerToMemberRepresentation::SingleInheritanceData; 1713 case DINode::FlagMultipleInheritance: 1714 return PointerToMemberRepresentation::MultipleInheritanceData; 1715 case DINode::FlagVirtualInheritance: 1716 return PointerToMemberRepresentation::VirtualInheritanceData; 1717 } 1718 } 1719 llvm_unreachable("invalid ptr to member representation"); 1720 } 1721 1722 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, 1723 PointerOptions PO) { 1724 assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); 1725 TypeIndex ClassTI = getTypeIndex(Ty->getClassType()); 1726 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType(), Ty->getClassType()); 1727 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 1728 : PointerKind::Near32; 1729 bool IsPMF = isa<DISubroutineType>(Ty->getBaseType()); 1730 PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction 1731 : PointerMode::PointerToDataMember; 1732 1733 assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big"); 1734 uint8_t SizeInBytes = Ty->getSizeInBits() / 8; 1735 MemberPointerInfo MPI( 1736 ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags())); 1737 PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); 1738 return TypeTable.writeLeafType(PR); 1739 } 1740 1741 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't 1742 /// have a translation, use the NearC convention. 1743 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { 1744 switch (DwarfCC) { 1745 case dwarf::DW_CC_normal: return CallingConvention::NearC; 1746 case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; 1747 case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; 1748 case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; 1749 case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; 1750 case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; 1751 } 1752 return CallingConvention::NearC; 1753 } 1754 1755 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { 1756 ModifierOptions Mods = ModifierOptions::None; 1757 PointerOptions PO = PointerOptions::None; 1758 bool IsModifier = true; 1759 const DIType *BaseTy = Ty; 1760 while (IsModifier && BaseTy) { 1761 // FIXME: Need to add DWARF tags for __unaligned and _Atomic 1762 switch (BaseTy->getTag()) { 1763 case dwarf::DW_TAG_const_type: 1764 Mods |= ModifierOptions::Const; 1765 PO |= PointerOptions::Const; 1766 break; 1767 case dwarf::DW_TAG_volatile_type: 1768 Mods |= ModifierOptions::Volatile; 1769 PO |= PointerOptions::Volatile; 1770 break; 1771 case dwarf::DW_TAG_restrict_type: 1772 // Only pointer types be marked with __restrict. There is no known flag 1773 // for __restrict in LF_MODIFIER records. 1774 PO |= PointerOptions::Restrict; 1775 break; 1776 default: 1777 IsModifier = false; 1778 break; 1779 } 1780 if (IsModifier) 1781 BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType().resolve(); 1782 } 1783 1784 // Check if the inner type will use an LF_POINTER record. If so, the 1785 // qualifiers will go in the LF_POINTER record. This comes up for types like 1786 // 'int *const' and 'int *__restrict', not the more common cases like 'const 1787 // char *'. 1788 if (BaseTy) { 1789 switch (BaseTy->getTag()) { 1790 case dwarf::DW_TAG_pointer_type: 1791 case dwarf::DW_TAG_reference_type: 1792 case dwarf::DW_TAG_rvalue_reference_type: 1793 return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO); 1794 case dwarf::DW_TAG_ptr_to_member_type: 1795 return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO); 1796 default: 1797 break; 1798 } 1799 } 1800 1801 TypeIndex ModifiedTI = getTypeIndex(BaseTy); 1802 1803 // Return the base type index if there aren't any modifiers. For example, the 1804 // metadata could contain restrict wrappers around non-pointer types. 1805 if (Mods == ModifierOptions::None) 1806 return ModifiedTI; 1807 1808 ModifierRecord MR(ModifiedTI, Mods); 1809 return TypeTable.writeLeafType(MR); 1810 } 1811 1812 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { 1813 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 1814 for (DITypeRef ArgTypeRef : Ty->getTypeArray()) 1815 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgTypeRef)); 1816 1817 // MSVC uses type none for variadic argument. 1818 if (ReturnAndArgTypeIndices.size() > 1 && 1819 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 1820 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 1821 } 1822 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 1823 ArrayRef<TypeIndex> ArgTypeIndices = None; 1824 if (!ReturnAndArgTypeIndices.empty()) { 1825 auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices); 1826 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 1827 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 1828 } 1829 1830 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 1831 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 1832 1833 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 1834 1835 FunctionOptions FO = getFunctionOptions(Ty); 1836 ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), 1837 ArgListIndex); 1838 return TypeTable.writeLeafType(Procedure); 1839 } 1840 1841 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, 1842 const DIType *ClassTy, 1843 int ThisAdjustment, 1844 bool IsStaticMethod, 1845 FunctionOptions FO) { 1846 // Lower the containing class type. 1847 TypeIndex ClassType = getTypeIndex(ClassTy); 1848 1849 DITypeRefArray ReturnAndArgs = Ty->getTypeArray(); 1850 1851 unsigned Index = 0; 1852 SmallVector<TypeIndex, 8> ArgTypeIndices; 1853 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 1854 if (ReturnAndArgs.size() > Index) { 1855 ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]); 1856 } 1857 1858 // If the first argument is a pointer type and this isn't a static method, 1859 // treat it as the special 'this' parameter, which is encoded separately from 1860 // the arguments. 1861 TypeIndex ThisTypeIndex; 1862 if (!IsStaticMethod && ReturnAndArgs.size() > Index) { 1863 if (const DIDerivedType *PtrTy = 1864 dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index].resolve())) { 1865 if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { 1866 ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty); 1867 Index++; 1868 } 1869 } 1870 } 1871 1872 while (Index < ReturnAndArgs.size()) 1873 ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++])); 1874 1875 // MSVC uses type none for variadic argument. 1876 if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) 1877 ArgTypeIndices.back() = TypeIndex::None(); 1878 1879 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 1880 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 1881 1882 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 1883 1884 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 1885 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 1886 return TypeTable.writeLeafType(MFR); 1887 } 1888 1889 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 1890 unsigned VSlotCount = 1891 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 1892 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 1893 1894 VFTableShapeRecord VFTSR(Slots); 1895 return TypeTable.writeLeafType(VFTSR); 1896 } 1897 1898 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 1899 switch (Flags & DINode::FlagAccessibility) { 1900 case DINode::FlagPrivate: return MemberAccess::Private; 1901 case DINode::FlagPublic: return MemberAccess::Public; 1902 case DINode::FlagProtected: return MemberAccess::Protected; 1903 case 0: 1904 // If there was no explicit access control, provide the default for the tag. 1905 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 1906 : MemberAccess::Public; 1907 } 1908 llvm_unreachable("access flags are exclusive"); 1909 } 1910 1911 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 1912 if (SP->isArtificial()) 1913 return MethodOptions::CompilerGenerated; 1914 1915 // FIXME: Handle other MethodOptions. 1916 1917 return MethodOptions::None; 1918 } 1919 1920 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 1921 bool Introduced) { 1922 if (SP->getFlags() & DINode::FlagStaticMember) 1923 return MethodKind::Static; 1924 1925 switch (SP->getVirtuality()) { 1926 case dwarf::DW_VIRTUALITY_none: 1927 break; 1928 case dwarf::DW_VIRTUALITY_virtual: 1929 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 1930 case dwarf::DW_VIRTUALITY_pure_virtual: 1931 return Introduced ? MethodKind::PureIntroducingVirtual 1932 : MethodKind::PureVirtual; 1933 default: 1934 llvm_unreachable("unhandled virtuality case"); 1935 } 1936 1937 return MethodKind::Vanilla; 1938 } 1939 1940 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 1941 switch (Ty->getTag()) { 1942 case dwarf::DW_TAG_class_type: return TypeRecordKind::Class; 1943 case dwarf::DW_TAG_structure_type: return TypeRecordKind::Struct; 1944 } 1945 llvm_unreachable("unexpected tag"); 1946 } 1947 1948 /// Return ClassOptions that should be present on both the forward declaration 1949 /// and the defintion of a tag type. 1950 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 1951 ClassOptions CO = ClassOptions::None; 1952 1953 // MSVC always sets this flag, even for local types. Clang doesn't always 1954 // appear to give every type a linkage name, which may be problematic for us. 1955 // FIXME: Investigate the consequences of not following them here. 1956 if (!Ty->getIdentifier().empty()) 1957 CO |= ClassOptions::HasUniqueName; 1958 1959 // Put the Nested flag on a type if it appears immediately inside a tag type. 1960 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 1961 // here. That flag is only set on definitions, and not forward declarations. 1962 const DIScope *ImmediateScope = Ty->getScope().resolve(); 1963 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 1964 CO |= ClassOptions::Nested; 1965 1966 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 1967 // type only when it has an immediate function scope. Clang never puts enums 1968 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 1969 // always in function, class, or file scopes. 1970 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 1971 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 1972 CO |= ClassOptions::Scoped; 1973 } else { 1974 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 1975 Scope = Scope->getScope().resolve()) { 1976 if (isa<DISubprogram>(Scope)) { 1977 CO |= ClassOptions::Scoped; 1978 break; 1979 } 1980 } 1981 } 1982 1983 return CO; 1984 } 1985 1986 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 1987 switch (Ty->getTag()) { 1988 case dwarf::DW_TAG_class_type: 1989 case dwarf::DW_TAG_structure_type: 1990 case dwarf::DW_TAG_union_type: 1991 case dwarf::DW_TAG_enumeration_type: 1992 break; 1993 default: 1994 return; 1995 } 1996 1997 if (const auto *File = Ty->getFile()) { 1998 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 1999 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 2000 2001 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 2002 TypeTable.writeLeafType(USLR); 2003 } 2004 } 2005 2006 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 2007 ClassOptions CO = getCommonClassOptions(Ty); 2008 TypeIndex FTI; 2009 unsigned EnumeratorCount = 0; 2010 2011 if (Ty->isForwardDecl()) { 2012 CO |= ClassOptions::ForwardReference; 2013 } else { 2014 ContinuationRecordBuilder ContinuationBuilder; 2015 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2016 for (const DINode *Element : Ty->getElements()) { 2017 // We assume that the frontend provides all members in source declaration 2018 // order, which is what MSVC does. 2019 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2020 EnumeratorRecord ER(MemberAccess::Public, 2021 APSInt::getUnsigned(Enumerator->getValue()), 2022 Enumerator->getName()); 2023 ContinuationBuilder.writeMemberType(ER); 2024 EnumeratorCount++; 2025 } 2026 } 2027 FTI = TypeTable.insertRecord(ContinuationBuilder); 2028 } 2029 2030 std::string FullName = getFullyQualifiedName(Ty); 2031 2032 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2033 getTypeIndex(Ty->getBaseType())); 2034 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2035 2036 addUDTSrcLine(Ty, EnumTI); 2037 2038 return EnumTI; 2039 } 2040 2041 //===----------------------------------------------------------------------===// 2042 // ClassInfo 2043 //===----------------------------------------------------------------------===// 2044 2045 struct llvm::ClassInfo { 2046 struct MemberInfo { 2047 const DIDerivedType *MemberTypeNode; 2048 uint64_t BaseOffset; 2049 }; 2050 // [MemberInfo] 2051 using MemberList = std::vector<MemberInfo>; 2052 2053 using MethodsList = TinyPtrVector<const DISubprogram *>; 2054 // MethodName -> MethodsList 2055 using MethodsMap = MapVector<MDString *, MethodsList>; 2056 2057 /// Base classes. 2058 std::vector<const DIDerivedType *> Inheritance; 2059 2060 /// Direct members. 2061 MemberList Members; 2062 // Direct overloaded methods gathered by name. 2063 MethodsMap Methods; 2064 2065 TypeIndex VShapeTI; 2066 2067 std::vector<const DIType *> NestedTypes; 2068 }; 2069 2070 void CodeViewDebug::clear() { 2071 assert(CurFn == nullptr); 2072 FileIdMap.clear(); 2073 FnDebugInfo.clear(); 2074 FileToFilepathMap.clear(); 2075 LocalUDTs.clear(); 2076 GlobalUDTs.clear(); 2077 TypeIndices.clear(); 2078 CompleteTypeIndices.clear(); 2079 ScopeGlobals.clear(); 2080 } 2081 2082 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2083 const DIDerivedType *DDTy) { 2084 if (!DDTy->getName().empty()) { 2085 Info.Members.push_back({DDTy, 0}); 2086 return; 2087 } 2088 2089 // An unnamed member may represent a nested struct or union. Attempt to 2090 // interpret the unnamed member as a DICompositeType possibly wrapped in 2091 // qualifier types. Add all the indirect fields to the current record if that 2092 // succeeds, and drop the member if that fails. 2093 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2094 uint64_t Offset = DDTy->getOffsetInBits(); 2095 const DIType *Ty = DDTy->getBaseType().resolve(); 2096 bool FullyResolved = false; 2097 while (!FullyResolved) { 2098 switch (Ty->getTag()) { 2099 case dwarf::DW_TAG_const_type: 2100 case dwarf::DW_TAG_volatile_type: 2101 // FIXME: we should apply the qualifier types to the indirect fields 2102 // rather than dropping them. 2103 Ty = cast<DIDerivedType>(Ty)->getBaseType().resolve(); 2104 break; 2105 default: 2106 FullyResolved = true; 2107 break; 2108 } 2109 } 2110 2111 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2112 if (!DCTy) 2113 return; 2114 2115 ClassInfo NestedInfo = collectClassInfo(DCTy); 2116 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2117 Info.Members.push_back( 2118 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2119 } 2120 2121 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2122 ClassInfo Info; 2123 // Add elements to structure type. 2124 DINodeArray Elements = Ty->getElements(); 2125 for (auto *Element : Elements) { 2126 // We assume that the frontend provides all members in source declaration 2127 // order, which is what MSVC does. 2128 if (!Element) 2129 continue; 2130 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2131 Info.Methods[SP->getRawName()].push_back(SP); 2132 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2133 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2134 collectMemberInfo(Info, DDTy); 2135 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2136 Info.Inheritance.push_back(DDTy); 2137 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2138 DDTy->getName() == "__vtbl_ptr_type") { 2139 Info.VShapeTI = getTypeIndex(DDTy); 2140 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2141 Info.NestedTypes.push_back(DDTy); 2142 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2143 // Ignore friend members. It appears that MSVC emitted info about 2144 // friends in the past, but modern versions do not. 2145 } 2146 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2147 Info.NestedTypes.push_back(Composite); 2148 } 2149 // Skip other unrecognized kinds of elements. 2150 } 2151 return Info; 2152 } 2153 2154 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2155 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2156 // if a complete type should be emitted instead of a forward reference. 2157 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2158 !Ty->isForwardDecl(); 2159 } 2160 2161 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2162 // Emit the complete type for unnamed structs. C++ classes with methods 2163 // which have a circular reference back to the class type are expected to 2164 // be named by the front-end and should not be "unnamed". C unnamed 2165 // structs should not have circular references. 2166 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2167 // If this unnamed complete type is already in the process of being defined 2168 // then the description of the type is malformed and cannot be emitted 2169 // into CodeView correctly so report a fatal error. 2170 auto I = CompleteTypeIndices.find(Ty); 2171 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2172 report_fatal_error("cannot debug circular reference to unnamed type"); 2173 return getCompleteTypeIndex(Ty); 2174 } 2175 2176 // First, construct the forward decl. Don't look into Ty to compute the 2177 // forward decl options, since it might not be available in all TUs. 2178 TypeRecordKind Kind = getRecordKind(Ty); 2179 ClassOptions CO = 2180 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2181 std::string FullName = getFullyQualifiedName(Ty); 2182 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2183 FullName, Ty->getIdentifier()); 2184 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2185 if (!Ty->isForwardDecl()) 2186 DeferredCompleteTypes.push_back(Ty); 2187 return FwdDeclTI; 2188 } 2189 2190 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2191 // Construct the field list and complete type record. 2192 TypeRecordKind Kind = getRecordKind(Ty); 2193 ClassOptions CO = getCommonClassOptions(Ty); 2194 TypeIndex FieldTI; 2195 TypeIndex VShapeTI; 2196 unsigned FieldCount; 2197 bool ContainsNestedClass; 2198 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2199 lowerRecordFieldList(Ty); 2200 2201 if (ContainsNestedClass) 2202 CO |= ClassOptions::ContainsNestedClass; 2203 2204 // MSVC appears to set this flag by searching any destructor or method with 2205 // FunctionOptions::Constructor among the emitted members. Clang AST has all 2206 // the members, however special member functions are not yet emitted into 2207 // debug information. For now checking a class's non-triviality seems enough. 2208 // FIXME: not true for a nested unnamed struct. 2209 if (isNonTrivial(Ty)) 2210 CO |= ClassOptions::HasConstructorOrDestructor; 2211 2212 std::string FullName = getFullyQualifiedName(Ty); 2213 2214 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2215 2216 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2217 SizeInBytes, FullName, Ty->getIdentifier()); 2218 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2219 2220 addUDTSrcLine(Ty, ClassTI); 2221 2222 addToUDTs(Ty); 2223 2224 return ClassTI; 2225 } 2226 2227 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2228 // Emit the complete type for unnamed unions. 2229 if (shouldAlwaysEmitCompleteClassType(Ty)) 2230 return getCompleteTypeIndex(Ty); 2231 2232 ClassOptions CO = 2233 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2234 std::string FullName = getFullyQualifiedName(Ty); 2235 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2236 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2237 if (!Ty->isForwardDecl()) 2238 DeferredCompleteTypes.push_back(Ty); 2239 return FwdDeclTI; 2240 } 2241 2242 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2243 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2244 TypeIndex FieldTI; 2245 unsigned FieldCount; 2246 bool ContainsNestedClass; 2247 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2248 lowerRecordFieldList(Ty); 2249 2250 if (ContainsNestedClass) 2251 CO |= ClassOptions::ContainsNestedClass; 2252 2253 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2254 std::string FullName = getFullyQualifiedName(Ty); 2255 2256 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2257 Ty->getIdentifier()); 2258 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2259 2260 addUDTSrcLine(Ty, UnionTI); 2261 2262 addToUDTs(Ty); 2263 2264 return UnionTI; 2265 } 2266 2267 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2268 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2269 // Manually count members. MSVC appears to count everything that generates a 2270 // field list record. Each individual overload in a method overload group 2271 // contributes to this count, even though the overload group is a single field 2272 // list record. 2273 unsigned MemberCount = 0; 2274 ClassInfo Info = collectClassInfo(Ty); 2275 ContinuationRecordBuilder ContinuationBuilder; 2276 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2277 2278 // Create base classes. 2279 for (const DIDerivedType *I : Info.Inheritance) { 2280 if (I->getFlags() & DINode::FlagVirtual) { 2281 // Virtual base. 2282 unsigned VBPtrOffset = I->getVBPtrOffset(); 2283 // FIXME: Despite the accessor name, the offset is really in bytes. 2284 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2285 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2286 ? TypeRecordKind::IndirectVirtualBaseClass 2287 : TypeRecordKind::VirtualBaseClass; 2288 VirtualBaseClassRecord VBCR( 2289 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2290 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2291 VBTableIndex); 2292 2293 ContinuationBuilder.writeMemberType(VBCR); 2294 MemberCount++; 2295 } else { 2296 assert(I->getOffsetInBits() % 8 == 0 && 2297 "bases must be on byte boundaries"); 2298 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2299 getTypeIndex(I->getBaseType()), 2300 I->getOffsetInBits() / 8); 2301 ContinuationBuilder.writeMemberType(BCR); 2302 MemberCount++; 2303 } 2304 } 2305 2306 // Create members. 2307 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2308 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2309 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2310 StringRef MemberName = Member->getName(); 2311 MemberAccess Access = 2312 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2313 2314 if (Member->isStaticMember()) { 2315 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2316 ContinuationBuilder.writeMemberType(SDMR); 2317 MemberCount++; 2318 continue; 2319 } 2320 2321 // Virtual function pointer member. 2322 if ((Member->getFlags() & DINode::FlagArtificial) && 2323 Member->getName().startswith("_vptr$")) { 2324 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2325 ContinuationBuilder.writeMemberType(VFPR); 2326 MemberCount++; 2327 continue; 2328 } 2329 2330 // Data member. 2331 uint64_t MemberOffsetInBits = 2332 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2333 if (Member->isBitField()) { 2334 uint64_t StartBitOffset = MemberOffsetInBits; 2335 if (const auto *CI = 2336 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2337 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2338 } 2339 StartBitOffset -= MemberOffsetInBits; 2340 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2341 StartBitOffset); 2342 MemberBaseType = TypeTable.writeLeafType(BFR); 2343 } 2344 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2345 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2346 MemberName); 2347 ContinuationBuilder.writeMemberType(DMR); 2348 MemberCount++; 2349 } 2350 2351 // Create methods 2352 for (auto &MethodItr : Info.Methods) { 2353 StringRef Name = MethodItr.first->getString(); 2354 2355 std::vector<OneMethodRecord> Methods; 2356 for (const DISubprogram *SP : MethodItr.second) { 2357 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2358 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2359 2360 unsigned VFTableOffset = -1; 2361 if (Introduced) 2362 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2363 2364 Methods.push_back(OneMethodRecord( 2365 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2366 translateMethodKindFlags(SP, Introduced), 2367 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2368 MemberCount++; 2369 } 2370 assert(!Methods.empty() && "Empty methods map entry"); 2371 if (Methods.size() == 1) 2372 ContinuationBuilder.writeMemberType(Methods[0]); 2373 else { 2374 // FIXME: Make this use its own ContinuationBuilder so that 2375 // MethodOverloadList can be split correctly. 2376 MethodOverloadListRecord MOLR(Methods); 2377 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2378 2379 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2380 ContinuationBuilder.writeMemberType(OMR); 2381 } 2382 } 2383 2384 // Create nested classes. 2385 for (const DIType *Nested : Info.NestedTypes) { 2386 NestedTypeRecord R(getTypeIndex(DITypeRef(Nested)), Nested->getName()); 2387 ContinuationBuilder.writeMemberType(R); 2388 MemberCount++; 2389 } 2390 2391 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2392 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2393 !Info.NestedTypes.empty()); 2394 } 2395 2396 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2397 if (!VBPType.getIndex()) { 2398 // Make a 'const int *' type. 2399 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2400 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2401 2402 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2403 : PointerKind::Near32; 2404 PointerMode PM = PointerMode::Pointer; 2405 PointerOptions PO = PointerOptions::None; 2406 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2407 VBPType = TypeTable.writeLeafType(PR); 2408 } 2409 2410 return VBPType; 2411 } 2412 2413 TypeIndex CodeViewDebug::getTypeIndex(DITypeRef TypeRef, DITypeRef ClassTyRef) { 2414 const DIType *Ty = TypeRef.resolve(); 2415 const DIType *ClassTy = ClassTyRef.resolve(); 2416 2417 // The null DIType is the void type. Don't try to hash it. 2418 if (!Ty) 2419 return TypeIndex::Void(); 2420 2421 // Check if we've already translated this type. Don't try to do a 2422 // get-or-create style insertion that caches the hash lookup across the 2423 // lowerType call. It will update the TypeIndices map. 2424 auto I = TypeIndices.find({Ty, ClassTy}); 2425 if (I != TypeIndices.end()) 2426 return I->second; 2427 2428 TypeLoweringScope S(*this); 2429 TypeIndex TI = lowerType(Ty, ClassTy); 2430 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2431 } 2432 2433 codeview::TypeIndex 2434 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, 2435 const DISubroutineType *SubroutineTy) { 2436 assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && 2437 "this type must be a pointer type"); 2438 2439 PointerOptions Options = PointerOptions::None; 2440 if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) 2441 Options = PointerOptions::LValueRefThisPointer; 2442 else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) 2443 Options = PointerOptions::RValueRefThisPointer; 2444 2445 // Check if we've already translated this type. If there is no ref qualifier 2446 // on the function then we look up this pointer type with no associated class 2447 // so that the TypeIndex for the this pointer can be shared with the type 2448 // index for other pointers to this class type. If there is a ref qualifier 2449 // then we lookup the pointer using the subroutine as the parent type. 2450 auto I = TypeIndices.find({PtrTy, SubroutineTy}); 2451 if (I != TypeIndices.end()) 2452 return I->second; 2453 2454 TypeLoweringScope S(*this); 2455 TypeIndex TI = lowerTypePointer(PtrTy, Options); 2456 return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy); 2457 } 2458 2459 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(DITypeRef TypeRef) { 2460 DIType *Ty = TypeRef.resolve(); 2461 PointerRecord PR(getTypeIndex(Ty), 2462 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2463 : PointerKind::Near32, 2464 PointerMode::LValueReference, PointerOptions::None, 2465 Ty->getSizeInBits() / 8); 2466 return TypeTable.writeLeafType(PR); 2467 } 2468 2469 TypeIndex CodeViewDebug::getCompleteTypeIndex(DITypeRef TypeRef) { 2470 const DIType *Ty = TypeRef.resolve(); 2471 2472 // The null DIType is the void type. Don't try to hash it. 2473 if (!Ty) 2474 return TypeIndex::Void(); 2475 2476 // Look through typedefs when getting the complete type index. Call 2477 // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are 2478 // emitted only once. 2479 if (Ty->getTag() == dwarf::DW_TAG_typedef) 2480 (void)getTypeIndex(Ty); 2481 while (Ty->getTag() == dwarf::DW_TAG_typedef) 2482 Ty = cast<DIDerivedType>(Ty)->getBaseType().resolve(); 2483 2484 // If this is a non-record type, the complete type index is the same as the 2485 // normal type index. Just call getTypeIndex. 2486 switch (Ty->getTag()) { 2487 case dwarf::DW_TAG_class_type: 2488 case dwarf::DW_TAG_structure_type: 2489 case dwarf::DW_TAG_union_type: 2490 break; 2491 default: 2492 return getTypeIndex(Ty); 2493 } 2494 2495 // Check if we've already translated the complete record type. 2496 const auto *CTy = cast<DICompositeType>(Ty); 2497 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2498 if (!InsertResult.second) 2499 return InsertResult.first->second; 2500 2501 TypeLoweringScope S(*this); 2502 2503 // Make sure the forward declaration is emitted first. It's unclear if this 2504 // is necessary, but MSVC does it, and we should follow suit until we can show 2505 // otherwise. 2506 // We only emit a forward declaration for named types. 2507 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2508 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2509 2510 // Just use the forward decl if we don't have complete type info. This 2511 // might happen if the frontend is using modules and expects the complete 2512 // definition to be emitted elsewhere. 2513 if (CTy->isForwardDecl()) 2514 return FwdDeclTI; 2515 } 2516 2517 TypeIndex TI; 2518 switch (CTy->getTag()) { 2519 case dwarf::DW_TAG_class_type: 2520 case dwarf::DW_TAG_structure_type: 2521 TI = lowerCompleteTypeClass(CTy); 2522 break; 2523 case dwarf::DW_TAG_union_type: 2524 TI = lowerCompleteTypeUnion(CTy); 2525 break; 2526 default: 2527 llvm_unreachable("not a record"); 2528 } 2529 2530 // Update the type index associated with this CompositeType. This cannot 2531 // use the 'InsertResult' iterator above because it is potentially 2532 // invalidated by map insertions which can occur while lowering the class 2533 // type above. 2534 CompleteTypeIndices[CTy] = TI; 2535 return TI; 2536 } 2537 2538 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2539 /// and do this until fixpoint, as each complete record type typically 2540 /// references 2541 /// many other record types. 2542 void CodeViewDebug::emitDeferredCompleteTypes() { 2543 SmallVector<const DICompositeType *, 4> TypesToEmit; 2544 while (!DeferredCompleteTypes.empty()) { 2545 std::swap(DeferredCompleteTypes, TypesToEmit); 2546 for (const DICompositeType *RecordTy : TypesToEmit) 2547 getCompleteTypeIndex(RecordTy); 2548 TypesToEmit.clear(); 2549 } 2550 } 2551 2552 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2553 ArrayRef<LocalVariable> Locals) { 2554 // Get the sorted list of parameters and emit them first. 2555 SmallVector<const LocalVariable *, 6> Params; 2556 for (const LocalVariable &L : Locals) 2557 if (L.DIVar->isParameter()) 2558 Params.push_back(&L); 2559 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2560 return L->DIVar->getArg() < R->DIVar->getArg(); 2561 }); 2562 for (const LocalVariable *L : Params) 2563 emitLocalVariable(FI, *L); 2564 2565 // Next emit all non-parameters in the order that we found them. 2566 for (const LocalVariable &L : Locals) 2567 if (!L.DIVar->isParameter()) 2568 emitLocalVariable(FI, L); 2569 } 2570 2571 /// Only call this on endian-specific types like ulittle16_t and little32_t, or 2572 /// structs composed of them. 2573 template <typename T> 2574 static void copyBytesForDefRange(SmallString<20> &BytePrefix, 2575 SymbolKind SymKind, const T &DefRangeHeader) { 2576 BytePrefix.resize(2 + sizeof(T)); 2577 ulittle16_t SymKindLE = ulittle16_t(SymKind); 2578 memcpy(&BytePrefix[0], &SymKindLE, 2); 2579 memcpy(&BytePrefix[2], &DefRangeHeader, sizeof(T)); 2580 } 2581 2582 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2583 const LocalVariable &Var) { 2584 // LocalSym record, see SymbolRecord.h for more info. 2585 MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL); 2586 2587 LocalSymFlags Flags = LocalSymFlags::None; 2588 if (Var.DIVar->isParameter()) 2589 Flags |= LocalSymFlags::IsParameter; 2590 if (Var.DefRanges.empty()) 2591 Flags |= LocalSymFlags::IsOptimizedOut; 2592 2593 OS.AddComment("TypeIndex"); 2594 TypeIndex TI = Var.UseReferenceType 2595 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2596 : getCompleteTypeIndex(Var.DIVar->getType()); 2597 OS.EmitIntValue(TI.getIndex(), 4); 2598 OS.AddComment("Flags"); 2599 OS.EmitIntValue(static_cast<uint16_t>(Flags), 2); 2600 // Truncate the name so we won't overflow the record length field. 2601 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2602 endSymbolRecord(LocalEnd); 2603 2604 // Calculate the on disk prefix of the appropriate def range record. The 2605 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2606 // should be big enough to hold all forms without memory allocation. 2607 SmallString<20> BytePrefix; 2608 for (const LocalVarDefRange &DefRange : Var.DefRanges) { 2609 BytePrefix.clear(); 2610 if (DefRange.InMemory) { 2611 int Offset = DefRange.DataOffset; 2612 unsigned Reg = DefRange.CVRegister; 2613 2614 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2615 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2616 // instead. In frames without stack realignment, $T0 will be the CFA. 2617 if (RegisterId(Reg) == RegisterId::ESP) { 2618 Reg = unsigned(RegisterId::VFRAME); 2619 Offset += FI.OffsetAdjustment; 2620 } 2621 2622 // If we can use the chosen frame pointer for the frame and this isn't a 2623 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2624 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2625 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2626 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2627 (bool(Flags & LocalSymFlags::IsParameter) 2628 ? (EncFP == FI.EncodedParamFramePtrReg) 2629 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2630 little32_t FPOffset = little32_t(Offset); 2631 copyBytesForDefRange(BytePrefix, S_DEFRANGE_FRAMEPOINTER_REL, FPOffset); 2632 } else { 2633 uint16_t RegRelFlags = 0; 2634 if (DefRange.IsSubfield) { 2635 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2636 (DefRange.StructOffset 2637 << DefRangeRegisterRelSym::OffsetInParentShift); 2638 } 2639 DefRangeRegisterRelSym::Header DRHdr; 2640 DRHdr.Register = Reg; 2641 DRHdr.Flags = RegRelFlags; 2642 DRHdr.BasePointerOffset = Offset; 2643 copyBytesForDefRange(BytePrefix, S_DEFRANGE_REGISTER_REL, DRHdr); 2644 } 2645 } else { 2646 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2647 if (DefRange.IsSubfield) { 2648 DefRangeSubfieldRegisterSym::Header DRHdr; 2649 DRHdr.Register = DefRange.CVRegister; 2650 DRHdr.MayHaveNoName = 0; 2651 DRHdr.OffsetInParent = DefRange.StructOffset; 2652 copyBytesForDefRange(BytePrefix, S_DEFRANGE_SUBFIELD_REGISTER, DRHdr); 2653 } else { 2654 DefRangeRegisterSym::Header DRHdr; 2655 DRHdr.Register = DefRange.CVRegister; 2656 DRHdr.MayHaveNoName = 0; 2657 copyBytesForDefRange(BytePrefix, S_DEFRANGE_REGISTER, DRHdr); 2658 } 2659 } 2660 OS.EmitCVDefRangeDirective(DefRange.Ranges, BytePrefix); 2661 } 2662 } 2663 2664 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2665 const FunctionInfo& FI) { 2666 for (LexicalBlock *Block : Blocks) 2667 emitLexicalBlock(*Block, FI); 2668 } 2669 2670 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2671 /// lexical block scope. 2672 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2673 const FunctionInfo& FI) { 2674 MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32); 2675 OS.AddComment("PtrParent"); 2676 OS.EmitIntValue(0, 4); // PtrParent 2677 OS.AddComment("PtrEnd"); 2678 OS.EmitIntValue(0, 4); // PtrEnd 2679 OS.AddComment("Code size"); 2680 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2681 OS.AddComment("Function section relative address"); 2682 OS.EmitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2683 OS.AddComment("Function section index"); 2684 OS.EmitCOFFSectionIndex(FI.Begin); // Func Symbol 2685 OS.AddComment("Lexical block name"); 2686 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2687 endSymbolRecord(RecordEnd); 2688 2689 // Emit variables local to this lexical block. 2690 emitLocalVariableList(FI, Block.Locals); 2691 emitGlobalVariableList(Block.Globals); 2692 2693 // Emit lexical blocks contained within this block. 2694 emitLexicalBlockList(Block.Children, FI); 2695 2696 // Close the lexical block scope. 2697 emitEndSymbolRecord(SymbolKind::S_END); 2698 } 2699 2700 /// Convenience routine for collecting lexical block information for a list 2701 /// of lexical scopes. 2702 void CodeViewDebug::collectLexicalBlockInfo( 2703 SmallVectorImpl<LexicalScope *> &Scopes, 2704 SmallVectorImpl<LexicalBlock *> &Blocks, 2705 SmallVectorImpl<LocalVariable> &Locals, 2706 SmallVectorImpl<CVGlobalVariable> &Globals) { 2707 for (LexicalScope *Scope : Scopes) 2708 collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals); 2709 } 2710 2711 /// Populate the lexical blocks and local variable lists of the parent with 2712 /// information about the specified lexical scope. 2713 void CodeViewDebug::collectLexicalBlockInfo( 2714 LexicalScope &Scope, 2715 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2716 SmallVectorImpl<LocalVariable> &ParentLocals, 2717 SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { 2718 if (Scope.isAbstractScope()) 2719 return; 2720 2721 // Gather information about the lexical scope including local variables, 2722 // global variables, and address ranges. 2723 bool IgnoreScope = false; 2724 auto LI = ScopeVariables.find(&Scope); 2725 SmallVectorImpl<LocalVariable> *Locals = 2726 LI != ScopeVariables.end() ? &LI->second : nullptr; 2727 auto GI = ScopeGlobals.find(Scope.getScopeNode()); 2728 SmallVectorImpl<CVGlobalVariable> *Globals = 2729 GI != ScopeGlobals.end() ? GI->second.get() : nullptr; 2730 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2731 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2732 2733 // Ignore lexical scopes which do not contain variables. 2734 if (!Locals && !Globals) 2735 IgnoreScope = true; 2736 2737 // Ignore lexical scopes which are not lexical blocks. 2738 if (!DILB) 2739 IgnoreScope = true; 2740 2741 // Ignore scopes which have too many address ranges to represent in the 2742 // current CodeView format or do not have a valid address range. 2743 // 2744 // For lexical scopes with multiple address ranges you may be tempted to 2745 // construct a single range covering every instruction where the block is 2746 // live and everything in between. Unfortunately, Visual Studio only 2747 // displays variables from the first matching lexical block scope. If the 2748 // first lexical block contains exception handling code or cold code which 2749 // is moved to the bottom of the routine creating a single range covering 2750 // nearly the entire routine, then it will hide all other lexical blocks 2751 // and the variables they contain. 2752 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) 2753 IgnoreScope = true; 2754 2755 if (IgnoreScope) { 2756 // This scope can be safely ignored and eliminating it will reduce the 2757 // size of the debug information. Be sure to collect any variable and scope 2758 // information from the this scope or any of its children and collapse them 2759 // into the parent scope. 2760 if (Locals) 2761 ParentLocals.append(Locals->begin(), Locals->end()); 2762 if (Globals) 2763 ParentGlobals.append(Globals->begin(), Globals->end()); 2764 collectLexicalBlockInfo(Scope.getChildren(), 2765 ParentBlocks, 2766 ParentLocals, 2767 ParentGlobals); 2768 return; 2769 } 2770 2771 // Create a new CodeView lexical block for this lexical scope. If we've 2772 // seen this DILexicalBlock before then the scope tree is malformed and 2773 // we can handle this gracefully by not processing it a second time. 2774 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 2775 if (!BlockInsertion.second) 2776 return; 2777 2778 // Create a lexical block containing the variables and collect the the 2779 // lexical block information for the children. 2780 const InsnRange &Range = Ranges.front(); 2781 assert(Range.first && Range.second); 2782 LexicalBlock &Block = BlockInsertion.first->second; 2783 Block.Begin = getLabelBeforeInsn(Range.first); 2784 Block.End = getLabelAfterInsn(Range.second); 2785 assert(Block.Begin && "missing label for scope begin"); 2786 assert(Block.End && "missing label for scope end"); 2787 Block.Name = DILB->getName(); 2788 if (Locals) 2789 Block.Locals = std::move(*Locals); 2790 if (Globals) 2791 Block.Globals = std::move(*Globals); 2792 ParentBlocks.push_back(&Block); 2793 collectLexicalBlockInfo(Scope.getChildren(), 2794 Block.Children, 2795 Block.Locals, 2796 Block.Globals); 2797 } 2798 2799 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 2800 const Function &GV = MF->getFunction(); 2801 assert(FnDebugInfo.count(&GV)); 2802 assert(CurFn == FnDebugInfo[&GV].get()); 2803 2804 collectVariableInfo(GV.getSubprogram()); 2805 2806 // Build the lexical block structure to emit for this routine. 2807 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 2808 collectLexicalBlockInfo(*CFS, 2809 CurFn->ChildBlocks, 2810 CurFn->Locals, 2811 CurFn->Globals); 2812 2813 // Clear the scope and variable information from the map which will not be 2814 // valid after we have finished processing this routine. This also prepares 2815 // the map for the subsequent routine. 2816 ScopeVariables.clear(); 2817 2818 // Don't emit anything if we don't have any line tables. 2819 // Thunks are compiler-generated and probably won't have source correlation. 2820 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 2821 FnDebugInfo.erase(&GV); 2822 CurFn = nullptr; 2823 return; 2824 } 2825 2826 CurFn->Annotations = MF->getCodeViewAnnotations(); 2827 CurFn->HeapAllocSites = MF->getCodeViewHeapAllocSites(); 2828 2829 CurFn->End = Asm->getFunctionEnd(); 2830 2831 CurFn = nullptr; 2832 } 2833 2834 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 2835 DebugHandlerBase::beginInstruction(MI); 2836 2837 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 2838 if (!Asm || !CurFn || MI->isDebugInstr() || 2839 MI->getFlag(MachineInstr::FrameSetup)) 2840 return; 2841 2842 // If the first instruction of a new MBB has no location, find the first 2843 // instruction with a location and use that. 2844 DebugLoc DL = MI->getDebugLoc(); 2845 if (!DL && MI->getParent() != PrevInstBB) { 2846 for (const auto &NextMI : *MI->getParent()) { 2847 if (NextMI.isDebugInstr()) 2848 continue; 2849 DL = NextMI.getDebugLoc(); 2850 if (DL) 2851 break; 2852 } 2853 } 2854 PrevInstBB = MI->getParent(); 2855 2856 // If we still don't have a debug location, don't record a location. 2857 if (!DL) 2858 return; 2859 2860 maybeRecordLocation(DL, Asm->MF); 2861 } 2862 2863 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 2864 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 2865 *EndLabel = MMI->getContext().createTempSymbol(); 2866 OS.EmitIntValue(unsigned(Kind), 4); 2867 OS.AddComment("Subsection size"); 2868 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 2869 OS.EmitLabel(BeginLabel); 2870 return EndLabel; 2871 } 2872 2873 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 2874 OS.EmitLabel(EndLabel); 2875 // Every subsection must be aligned to a 4-byte boundary. 2876 OS.EmitValueToAlignment(4); 2877 } 2878 2879 static StringRef getSymbolName(SymbolKind SymKind) { 2880 for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) 2881 if (EE.Value == SymKind) 2882 return EE.Name; 2883 return ""; 2884 } 2885 2886 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { 2887 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 2888 *EndLabel = MMI->getContext().createTempSymbol(); 2889 OS.AddComment("Record length"); 2890 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 2891 OS.EmitLabel(BeginLabel); 2892 if (OS.isVerboseAsm()) 2893 OS.AddComment("Record kind: " + getSymbolName(SymKind)); 2894 OS.EmitIntValue(unsigned(SymKind), 2); 2895 return EndLabel; 2896 } 2897 2898 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { 2899 // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid 2900 // an extra copy of every symbol record in LLD. This increases object file 2901 // size by less than 1% in the clang build, and is compatible with the Visual 2902 // C++ linker. 2903 OS.EmitValueToAlignment(4); 2904 OS.EmitLabel(SymEnd); 2905 } 2906 2907 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { 2908 OS.AddComment("Record length"); 2909 OS.EmitIntValue(2, 2); 2910 if (OS.isVerboseAsm()) 2911 OS.AddComment("Record kind: " + getSymbolName(EndKind)); 2912 OS.EmitIntValue(unsigned(EndKind), 2); // Record Kind 2913 } 2914 2915 void CodeViewDebug::emitDebugInfoForUDTs( 2916 ArrayRef<std::pair<std::string, const DIType *>> UDTs) { 2917 for (const auto &UDT : UDTs) { 2918 const DIType *T = UDT.second; 2919 assert(shouldEmitUdt(T)); 2920 2921 MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT); 2922 OS.AddComment("Type"); 2923 OS.EmitIntValue(getCompleteTypeIndex(T).getIndex(), 4); 2924 emitNullTerminatedSymbolName(OS, UDT.first); 2925 endSymbolRecord(UDTRecordEnd); 2926 } 2927 } 2928 2929 void CodeViewDebug::collectGlobalVariableInfo() { 2930 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 2931 GlobalMap; 2932 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 2933 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 2934 GV.getDebugInfo(GVEs); 2935 for (const auto *GVE : GVEs) 2936 GlobalMap[GVE] = &GV; 2937 } 2938 2939 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 2940 for (const MDNode *Node : CUs->operands()) { 2941 const auto *CU = cast<DICompileUnit>(Node); 2942 for (const auto *GVE : CU->getGlobalVariables()) { 2943 const auto *GV = GlobalMap.lookup(GVE); 2944 if (!GV || GV->isDeclarationForLinker()) 2945 continue; 2946 const DIGlobalVariable *DIGV = GVE->getVariable(); 2947 DIScope *Scope = DIGV->getScope(); 2948 SmallVector<CVGlobalVariable, 1> *VariableList; 2949 if (Scope && isa<DILocalScope>(Scope)) { 2950 // Locate a global variable list for this scope, creating one if 2951 // necessary. 2952 auto Insertion = ScopeGlobals.insert( 2953 {Scope, std::unique_ptr<GlobalVariableList>()}); 2954 if (Insertion.second) 2955 Insertion.first->second = llvm::make_unique<GlobalVariableList>(); 2956 VariableList = Insertion.first->second.get(); 2957 } else if (GV->hasComdat()) 2958 // Emit this global variable into a COMDAT section. 2959 VariableList = &ComdatVariables; 2960 else 2961 // Emit this globla variable in a single global symbol section. 2962 VariableList = &GlobalVariables; 2963 CVGlobalVariable CVGV = {DIGV, GV}; 2964 VariableList->emplace_back(std::move(CVGV)); 2965 } 2966 } 2967 } 2968 2969 void CodeViewDebug::emitDebugInfoForGlobals() { 2970 // First, emit all globals that are not in a comdat in a single symbol 2971 // substream. MSVC doesn't like it if the substream is empty, so only open 2972 // it if we have at least one global to emit. 2973 switchToDebugSectionForSymbol(nullptr); 2974 if (!GlobalVariables.empty()) { 2975 OS.AddComment("Symbol subsection for globals"); 2976 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 2977 emitGlobalVariableList(GlobalVariables); 2978 endCVSubsection(EndLabel); 2979 } 2980 2981 // Second, emit each global that is in a comdat into its own .debug$S 2982 // section along with its own symbol substream. 2983 for (const CVGlobalVariable &CVGV : ComdatVariables) { 2984 MCSymbol *GVSym = Asm->getSymbol(CVGV.GV); 2985 OS.AddComment("Symbol subsection for " + 2986 Twine(GlobalValue::dropLLVMManglingEscape(CVGV.GV->getName()))); 2987 switchToDebugSectionForSymbol(GVSym); 2988 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 2989 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 2990 emitDebugInfoForGlobal(CVGV.DIGV, CVGV.GV, GVSym); 2991 endCVSubsection(EndLabel); 2992 } 2993 } 2994 2995 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 2996 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 2997 for (const MDNode *Node : CUs->operands()) { 2998 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 2999 if (DIType *RT = dyn_cast<DIType>(Ty)) { 3000 getTypeIndex(RT); 3001 // FIXME: Add to global/local DTU list. 3002 } 3003 } 3004 } 3005 } 3006 3007 // Emit each global variable in the specified array. 3008 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { 3009 for (const CVGlobalVariable &CVGV : Globals) { 3010 MCSymbol *GVSym = Asm->getSymbol(CVGV.GV); 3011 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3012 emitDebugInfoForGlobal(CVGV.DIGV, CVGV.GV, GVSym); 3013 } 3014 } 3015 3016 void CodeViewDebug::emitDebugInfoForGlobal(const DIGlobalVariable *DIGV, 3017 const GlobalVariable *GV, 3018 MCSymbol *GVSym) { 3019 // DataSym record, see SymbolRecord.h for more info. Thread local data 3020 // happens to have the same format as global data. 3021 SymbolKind DataSym = GV->isThreadLocal() 3022 ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 3023 : SymbolKind::S_GTHREAD32) 3024 : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 3025 : SymbolKind::S_GDATA32); 3026 MCSymbol *DataEnd = beginSymbolRecord(DataSym); 3027 OS.AddComment("Type"); 3028 OS.EmitIntValue(getCompleteTypeIndex(DIGV->getType()).getIndex(), 4); 3029 OS.AddComment("DataOffset"); 3030 OS.EmitCOFFSecRel32(GVSym, /*Offset=*/0); 3031 OS.AddComment("Segment"); 3032 OS.EmitCOFFSectionIndex(GVSym); 3033 OS.AddComment("Name"); 3034 const unsigned LengthOfDataRecord = 12; 3035 emitNullTerminatedSymbolName(OS, DIGV->getName(), LengthOfDataRecord); 3036 endSymbolRecord(DataEnd); 3037 } 3038