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/CodeViewRecordIO.h" 45 #include "llvm/DebugInfo/CodeView/ContinuationRecordBuilder.h" 46 #include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h" 47 #include "llvm/DebugInfo/CodeView/EnumTables.h" 48 #include "llvm/DebugInfo/CodeView/Line.h" 49 #include "llvm/DebugInfo/CodeView/SymbolRecord.h" 50 #include "llvm/DebugInfo/CodeView/TypeDumpVisitor.h" 51 #include "llvm/DebugInfo/CodeView/TypeIndex.h" 52 #include "llvm/DebugInfo/CodeView/TypeRecord.h" 53 #include "llvm/DebugInfo/CodeView/TypeTableCollection.h" 54 #include "llvm/IR/Constants.h" 55 #include "llvm/IR/DataLayout.h" 56 #include "llvm/IR/DebugInfoMetadata.h" 57 #include "llvm/IR/DebugLoc.h" 58 #include "llvm/IR/Function.h" 59 #include "llvm/IR/GlobalValue.h" 60 #include "llvm/IR/GlobalVariable.h" 61 #include "llvm/IR/Metadata.h" 62 #include "llvm/IR/Module.h" 63 #include "llvm/MC/MCAsmInfo.h" 64 #include "llvm/MC/MCContext.h" 65 #include "llvm/MC/MCSectionCOFF.h" 66 #include "llvm/MC/MCStreamer.h" 67 #include "llvm/MC/MCSymbol.h" 68 #include "llvm/Support/BinaryByteStream.h" 69 #include "llvm/Support/BinaryStreamReader.h" 70 #include "llvm/Support/BinaryStreamWriter.h" 71 #include "llvm/Support/Casting.h" 72 #include "llvm/Support/CommandLine.h" 73 #include "llvm/Support/Compiler.h" 74 #include "llvm/Support/Endian.h" 75 #include "llvm/Support/Error.h" 76 #include "llvm/Support/ErrorHandling.h" 77 #include "llvm/Support/FormatVariadic.h" 78 #include "llvm/Support/Path.h" 79 #include "llvm/Support/SMLoc.h" 80 #include "llvm/Support/ScopedPrinter.h" 81 #include "llvm/Target/TargetLoweringObjectFile.h" 82 #include "llvm/Target/TargetMachine.h" 83 #include <algorithm> 84 #include <cassert> 85 #include <cctype> 86 #include <cstddef> 87 #include <cstdint> 88 #include <iterator> 89 #include <limits> 90 #include <string> 91 #include <utility> 92 #include <vector> 93 94 using namespace llvm; 95 using namespace llvm::codeview; 96 97 static CPUType mapArchToCVCPUType(Triple::ArchType Type) { 98 switch (Type) { 99 case Triple::ArchType::x86: 100 return CPUType::Pentium3; 101 case Triple::ArchType::x86_64: 102 return CPUType::X64; 103 case Triple::ArchType::thumb: 104 return CPUType::Thumb; 105 case Triple::ArchType::aarch64: 106 return CPUType::ARM64; 107 default: 108 report_fatal_error("target architecture doesn't map to a CodeView CPUType"); 109 } 110 } 111 112 CodeViewDebug::CodeViewDebug(AsmPrinter *AP) 113 : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) { 114 // If module doesn't have named metadata anchors or COFF debug section 115 // is not available, skip any debug info related stuff. 116 if (!MMI->getModule()->getNamedMetadata("llvm.dbg.cu") || 117 !AP->getObjFileLowering().getCOFFDebugSymbolsSection()) { 118 Asm = nullptr; 119 MMI->setDebugInfoAvailability(false); 120 return; 121 } 122 // Tell MMI that we have debug info. 123 MMI->setDebugInfoAvailability(true); 124 125 TheCPU = 126 mapArchToCVCPUType(Triple(MMI->getModule()->getTargetTriple()).getArch()); 127 128 collectGlobalVariableInfo(); 129 130 // Check if we should emit type record hashes. 131 ConstantInt *GH = mdconst::extract_or_null<ConstantInt>( 132 MMI->getModule()->getModuleFlag("CodeViewGHash")); 133 EmitDebugGlobalHashes = GH && !GH->isZero(); 134 } 135 136 StringRef CodeViewDebug::getFullFilepath(const DIFile *File) { 137 std::string &Filepath = FileToFilepathMap[File]; 138 if (!Filepath.empty()) 139 return Filepath; 140 141 StringRef Dir = File->getDirectory(), Filename = File->getFilename(); 142 143 // If this is a Unix-style path, just use it as is. Don't try to canonicalize 144 // it textually because one of the path components could be a symlink. 145 if (Dir.startswith("/") || Filename.startswith("/")) { 146 if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix)) 147 return Filename; 148 Filepath = Dir; 149 if (Dir.back() != '/') 150 Filepath += '/'; 151 Filepath += Filename; 152 return Filepath; 153 } 154 155 // Clang emits directory and relative filename info into the IR, but CodeView 156 // operates on full paths. We could change Clang to emit full paths too, but 157 // that would increase the IR size and probably not needed for other users. 158 // For now, just concatenate and canonicalize the path here. 159 if (Filename.find(':') == 1) 160 Filepath = Filename; 161 else 162 Filepath = (Dir + "\\" + Filename).str(); 163 164 // Canonicalize the path. We have to do it textually because we may no longer 165 // have access the file in the filesystem. 166 // First, replace all slashes with backslashes. 167 std::replace(Filepath.begin(), Filepath.end(), '/', '\\'); 168 169 // Remove all "\.\" with "\". 170 size_t Cursor = 0; 171 while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos) 172 Filepath.erase(Cursor, 2); 173 174 // Replace all "\XXX\..\" with "\". Don't try too hard though as the original 175 // path should be well-formatted, e.g. start with a drive letter, etc. 176 Cursor = 0; 177 while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) { 178 // Something's wrong if the path starts with "\..\", abort. 179 if (Cursor == 0) 180 break; 181 182 size_t PrevSlash = Filepath.rfind('\\', Cursor - 1); 183 if (PrevSlash == std::string::npos) 184 // Something's wrong, abort. 185 break; 186 187 Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash); 188 // The next ".." might be following the one we've just erased. 189 Cursor = PrevSlash; 190 } 191 192 // Remove all duplicate backslashes. 193 Cursor = 0; 194 while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos) 195 Filepath.erase(Cursor, 1); 196 197 return Filepath; 198 } 199 200 unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) { 201 StringRef FullPath = getFullFilepath(F); 202 unsigned NextId = FileIdMap.size() + 1; 203 auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId)); 204 if (Insertion.second) { 205 // We have to compute the full filepath and emit a .cv_file directive. 206 ArrayRef<uint8_t> ChecksumAsBytes; 207 FileChecksumKind CSKind = FileChecksumKind::None; 208 if (F->getChecksum()) { 209 std::string Checksum = fromHex(F->getChecksum()->Value); 210 void *CKMem = OS.getContext().allocate(Checksum.size(), 1); 211 memcpy(CKMem, Checksum.data(), Checksum.size()); 212 ChecksumAsBytes = ArrayRef<uint8_t>( 213 reinterpret_cast<const uint8_t *>(CKMem), Checksum.size()); 214 switch (F->getChecksum()->Kind) { 215 case DIFile::CSK_MD5: CSKind = FileChecksumKind::MD5; break; 216 case DIFile::CSK_SHA1: CSKind = FileChecksumKind::SHA1; break; 217 } 218 } 219 bool Success = OS.EmitCVFileDirective(NextId, FullPath, ChecksumAsBytes, 220 static_cast<unsigned>(CSKind)); 221 (void)Success; 222 assert(Success && ".cv_file directive failed"); 223 } 224 return Insertion.first->second; 225 } 226 227 CodeViewDebug::InlineSite & 228 CodeViewDebug::getInlineSite(const DILocation *InlinedAt, 229 const DISubprogram *Inlinee) { 230 auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()}); 231 InlineSite *Site = &SiteInsertion.first->second; 232 if (SiteInsertion.second) { 233 unsigned ParentFuncId = CurFn->FuncId; 234 if (const DILocation *OuterIA = InlinedAt->getInlinedAt()) 235 ParentFuncId = 236 getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram()) 237 .SiteFuncId; 238 239 Site->SiteFuncId = NextFuncId++; 240 OS.EmitCVInlineSiteIdDirective( 241 Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()), 242 InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc()); 243 Site->Inlinee = Inlinee; 244 InlinedSubprograms.insert(Inlinee); 245 getFuncIdForSubprogram(Inlinee); 246 } 247 return *Site; 248 } 249 250 static StringRef getPrettyScopeName(const DIScope *Scope) { 251 StringRef ScopeName = Scope->getName(); 252 if (!ScopeName.empty()) 253 return ScopeName; 254 255 switch (Scope->getTag()) { 256 case dwarf::DW_TAG_enumeration_type: 257 case dwarf::DW_TAG_class_type: 258 case dwarf::DW_TAG_structure_type: 259 case dwarf::DW_TAG_union_type: 260 return "<unnamed-tag>"; 261 case dwarf::DW_TAG_namespace: 262 return "`anonymous namespace'"; 263 } 264 265 return StringRef(); 266 } 267 268 static const DISubprogram *getQualifiedNameComponents( 269 const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) { 270 const DISubprogram *ClosestSubprogram = nullptr; 271 while (Scope != nullptr) { 272 if (ClosestSubprogram == nullptr) 273 ClosestSubprogram = dyn_cast<DISubprogram>(Scope); 274 StringRef ScopeName = getPrettyScopeName(Scope); 275 if (!ScopeName.empty()) 276 QualifiedNameComponents.push_back(ScopeName); 277 Scope = Scope->getScope(); 278 } 279 return ClosestSubprogram; 280 } 281 282 static std::string getQualifiedName(ArrayRef<StringRef> QualifiedNameComponents, 283 StringRef TypeName) { 284 std::string FullyQualifiedName; 285 for (StringRef QualifiedNameComponent : 286 llvm::reverse(QualifiedNameComponents)) { 287 FullyQualifiedName.append(QualifiedNameComponent); 288 FullyQualifiedName.append("::"); 289 } 290 FullyQualifiedName.append(TypeName); 291 return FullyQualifiedName; 292 } 293 294 static std::string getFullyQualifiedName(const DIScope *Scope, StringRef Name) { 295 SmallVector<StringRef, 5> QualifiedNameComponents; 296 getQualifiedNameComponents(Scope, QualifiedNameComponents); 297 return getQualifiedName(QualifiedNameComponents, Name); 298 } 299 300 struct CodeViewDebug::TypeLoweringScope { 301 TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; } 302 ~TypeLoweringScope() { 303 // Don't decrement TypeEmissionLevel until after emitting deferred types, so 304 // inner TypeLoweringScopes don't attempt to emit deferred types. 305 if (CVD.TypeEmissionLevel == 1) 306 CVD.emitDeferredCompleteTypes(); 307 --CVD.TypeEmissionLevel; 308 } 309 CodeViewDebug &CVD; 310 }; 311 312 static std::string getFullyQualifiedName(const DIScope *Ty) { 313 const DIScope *Scope = Ty->getScope(); 314 return getFullyQualifiedName(Scope, getPrettyScopeName(Ty)); 315 } 316 317 TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) { 318 // No scope means global scope and that uses the zero index. 319 if (!Scope || isa<DIFile>(Scope)) 320 return TypeIndex(); 321 322 assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type"); 323 324 // Check if we've already translated this scope. 325 auto I = TypeIndices.find({Scope, nullptr}); 326 if (I != TypeIndices.end()) 327 return I->second; 328 329 // Build the fully qualified name of the scope. 330 std::string ScopeName = getFullyQualifiedName(Scope); 331 StringIdRecord SID(TypeIndex(), ScopeName); 332 auto TI = TypeTable.writeLeafType(SID); 333 return recordTypeIndexForDINode(Scope, TI); 334 } 335 336 TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) { 337 assert(SP); 338 339 // Check if we've already translated this subprogram. 340 auto I = TypeIndices.find({SP, nullptr}); 341 if (I != TypeIndices.end()) 342 return I->second; 343 344 // The display name includes function template arguments. Drop them to match 345 // MSVC. 346 StringRef DisplayName = SP->getName().split('<').first; 347 348 const DIScope *Scope = SP->getScope(); 349 TypeIndex TI; 350 if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) { 351 // If the scope is a DICompositeType, then this must be a method. Member 352 // function types take some special handling, and require access to the 353 // subprogram. 354 TypeIndex ClassType = getTypeIndex(Class); 355 MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class), 356 DisplayName); 357 TI = TypeTable.writeLeafType(MFuncId); 358 } else { 359 // Otherwise, this must be a free function. 360 TypeIndex ParentScope = getScopeIndex(Scope); 361 FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName); 362 TI = TypeTable.writeLeafType(FuncId); 363 } 364 365 return recordTypeIndexForDINode(SP, TI); 366 } 367 368 static bool isNonTrivial(const DICompositeType *DCTy) { 369 return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial); 370 } 371 372 static FunctionOptions 373 getFunctionOptions(const DISubroutineType *Ty, 374 const DICompositeType *ClassTy = nullptr, 375 StringRef SPName = StringRef("")) { 376 FunctionOptions FO = FunctionOptions::None; 377 const DIType *ReturnTy = nullptr; 378 if (auto TypeArray = Ty->getTypeArray()) { 379 if (TypeArray.size()) 380 ReturnTy = TypeArray[0]; 381 } 382 383 if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) { 384 if (isNonTrivial(ReturnDCTy)) 385 FO |= FunctionOptions::CxxReturnUdt; 386 } 387 388 // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison. 389 if (ClassTy && isNonTrivial(ClassTy) && SPName == ClassTy->getName()) { 390 FO |= FunctionOptions::Constructor; 391 392 // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag. 393 394 } 395 return FO; 396 } 397 398 TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP, 399 const DICompositeType *Class) { 400 // Always use the method declaration as the key for the function type. The 401 // method declaration contains the this adjustment. 402 if (SP->getDeclaration()) 403 SP = SP->getDeclaration(); 404 assert(!SP->getDeclaration() && "should use declaration as key"); 405 406 // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide 407 // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}. 408 auto I = TypeIndices.find({SP, Class}); 409 if (I != TypeIndices.end()) 410 return I->second; 411 412 // Make sure complete type info for the class is emitted *after* the member 413 // function type, as the complete class type is likely to reference this 414 // member function type. 415 TypeLoweringScope S(*this); 416 const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0; 417 418 FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName()); 419 TypeIndex TI = lowerTypeMemberFunction( 420 SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO); 421 return recordTypeIndexForDINode(SP, TI, Class); 422 } 423 424 TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node, 425 TypeIndex TI, 426 const DIType *ClassTy) { 427 auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI}); 428 (void)InsertResult; 429 assert(InsertResult.second && "DINode was already assigned a type index"); 430 return TI; 431 } 432 433 unsigned CodeViewDebug::getPointerSizeInBytes() { 434 return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8; 435 } 436 437 void CodeViewDebug::recordLocalVariable(LocalVariable &&Var, 438 const LexicalScope *LS) { 439 if (const DILocation *InlinedAt = LS->getInlinedAt()) { 440 // This variable was inlined. Associate it with the InlineSite. 441 const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram(); 442 InlineSite &Site = getInlineSite(InlinedAt, Inlinee); 443 Site.InlinedLocals.emplace_back(Var); 444 } else { 445 // This variable goes into the corresponding lexical scope. 446 ScopeVariables[LS].emplace_back(Var); 447 } 448 } 449 450 static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs, 451 const DILocation *Loc) { 452 auto B = Locs.begin(), E = Locs.end(); 453 if (std::find(B, E, Loc) == E) 454 Locs.push_back(Loc); 455 } 456 457 void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL, 458 const MachineFunction *MF) { 459 // Skip this instruction if it has the same location as the previous one. 460 if (!DL || DL == PrevInstLoc) 461 return; 462 463 const DIScope *Scope = DL.get()->getScope(); 464 if (!Scope) 465 return; 466 467 // Skip this line if it is longer than the maximum we can record. 468 LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true); 469 if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() || 470 LI.isNeverStepInto()) 471 return; 472 473 ColumnInfo CI(DL.getCol(), /*EndColumn=*/0); 474 if (CI.getStartColumn() != DL.getCol()) 475 return; 476 477 if (!CurFn->HaveLineInfo) 478 CurFn->HaveLineInfo = true; 479 unsigned FileId = 0; 480 if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile()) 481 FileId = CurFn->LastFileId; 482 else 483 FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile()); 484 PrevInstLoc = DL; 485 486 unsigned FuncId = CurFn->FuncId; 487 if (const DILocation *SiteLoc = DL->getInlinedAt()) { 488 const DILocation *Loc = DL.get(); 489 490 // If this location was actually inlined from somewhere else, give it the ID 491 // of the inline call site. 492 FuncId = 493 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId; 494 495 // Ensure we have links in the tree of inline call sites. 496 bool FirstLoc = true; 497 while ((SiteLoc = Loc->getInlinedAt())) { 498 InlineSite &Site = 499 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()); 500 if (!FirstLoc) 501 addLocIfNotPresent(Site.ChildSites, Loc); 502 FirstLoc = false; 503 Loc = SiteLoc; 504 } 505 addLocIfNotPresent(CurFn->ChildSites, Loc); 506 } 507 508 OS.EmitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(), 509 /*PrologueEnd=*/false, /*IsStmt=*/false, 510 DL->getFilename(), SMLoc()); 511 } 512 513 void CodeViewDebug::emitCodeViewMagicVersion() { 514 OS.EmitValueToAlignment(4); 515 OS.AddComment("Debug section magic"); 516 OS.EmitIntValue(COFF::DEBUG_SECTION_MAGIC, 4); 517 } 518 519 void CodeViewDebug::endModule() { 520 if (!Asm || !MMI->hasDebugInfo()) 521 return; 522 523 assert(Asm != nullptr); 524 525 // The COFF .debug$S section consists of several subsections, each starting 526 // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length 527 // of the payload followed by the payload itself. The subsections are 4-byte 528 // aligned. 529 530 // Use the generic .debug$S section, and make a subsection for all the inlined 531 // subprograms. 532 switchToDebugSectionForSymbol(nullptr); 533 534 MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols); 535 emitCompilerInformation(); 536 endCVSubsection(CompilerInfo); 537 538 emitInlineeLinesSubsection(); 539 540 // Emit per-function debug information. 541 for (auto &P : FnDebugInfo) 542 if (!P.first->isDeclarationForLinker()) 543 emitDebugInfoForFunction(P.first, *P.second); 544 545 // Emit global variable debug information. 546 setCurrentSubprogram(nullptr); 547 emitDebugInfoForGlobals(); 548 549 // Emit retained types. 550 emitDebugInfoForRetainedTypes(); 551 552 // Switch back to the generic .debug$S section after potentially processing 553 // comdat symbol sections. 554 switchToDebugSectionForSymbol(nullptr); 555 556 // Emit UDT records for any types used by global variables. 557 if (!GlobalUDTs.empty()) { 558 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 559 emitDebugInfoForUDTs(GlobalUDTs); 560 endCVSubsection(SymbolsEnd); 561 } 562 563 // This subsection holds a file index to offset in string table table. 564 OS.AddComment("File index to string table offset subsection"); 565 OS.EmitCVFileChecksumsDirective(); 566 567 // This subsection holds the string table. 568 OS.AddComment("String table"); 569 OS.EmitCVStringTableDirective(); 570 571 // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol 572 // subsection in the generic .debug$S section at the end. There is no 573 // particular reason for this ordering other than to match MSVC. 574 emitBuildInfo(); 575 576 // Emit type information and hashes last, so that any types we translate while 577 // emitting function info are included. 578 emitTypeInformation(); 579 580 if (EmitDebugGlobalHashes) 581 emitTypeGlobalHashes(); 582 583 clear(); 584 } 585 586 static void emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S, 587 unsigned MaxFixedRecordLength = 0xF00) { 588 // The maximum CV record length is 0xFF00. Most of the strings we emit appear 589 // after a fixed length portion of the record. The fixed length portion should 590 // always be less than 0xF00 (3840) bytes, so truncate the string so that the 591 // overall record size is less than the maximum allowed. 592 SmallString<32> NullTerminatedString( 593 S.take_front(MaxRecordLength - MaxFixedRecordLength - 1)); 594 NullTerminatedString.push_back('\0'); 595 OS.EmitBytes(NullTerminatedString); 596 } 597 598 void CodeViewDebug::emitTypeInformation() { 599 if (TypeTable.empty()) 600 return; 601 602 // Start the .debug$T or .debug$P section with 0x4. 603 OS.SwitchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection()); 604 emitCodeViewMagicVersion(); 605 606 SmallString<8> CommentPrefix; 607 if (OS.isVerboseAsm()) { 608 CommentPrefix += '\t'; 609 CommentPrefix += Asm->MAI->getCommentString(); 610 CommentPrefix += ' '; 611 } 612 613 TypeTableCollection Table(TypeTable.records()); 614 Optional<TypeIndex> B = Table.getFirst(); 615 while (B) { 616 // This will fail if the record data is invalid. 617 CVType Record = Table.getType(*B); 618 619 if (OS.isVerboseAsm()) { 620 // Emit a block comment describing the type record for readability. 621 SmallString<512> CommentBlock; 622 raw_svector_ostream CommentOS(CommentBlock); 623 ScopedPrinter SP(CommentOS); 624 SP.setPrefix(CommentPrefix); 625 TypeDumpVisitor TDV(Table, &SP, false); 626 627 Error E = codeview::visitTypeRecord(Record, *B, TDV); 628 if (E) { 629 logAllUnhandledErrors(std::move(E), errs(), "error: "); 630 llvm_unreachable("produced malformed type record"); 631 } 632 // emitRawComment will insert its own tab and comment string before 633 // the first line, so strip off our first one. It also prints its own 634 // newline. 635 OS.emitRawComment( 636 CommentOS.str().drop_front(CommentPrefix.size() - 1).rtrim()); 637 } 638 OS.EmitBinaryData(Record.str_data()); 639 B = Table.getNext(*B); 640 } 641 } 642 643 void CodeViewDebug::emitTypeGlobalHashes() { 644 if (TypeTable.empty()) 645 return; 646 647 // Start the .debug$H section with the version and hash algorithm, currently 648 // hardcoded to version 0, SHA1. 649 OS.SwitchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection()); 650 651 OS.EmitValueToAlignment(4); 652 OS.AddComment("Magic"); 653 OS.EmitIntValue(COFF::DEBUG_HASHES_SECTION_MAGIC, 4); 654 OS.AddComment("Section Version"); 655 OS.EmitIntValue(0, 2); 656 OS.AddComment("Hash Algorithm"); 657 OS.EmitIntValue(uint16_t(GlobalTypeHashAlg::SHA1_8), 2); 658 659 TypeIndex TI(TypeIndex::FirstNonSimpleIndex); 660 for (const auto &GHR : TypeTable.hashes()) { 661 if (OS.isVerboseAsm()) { 662 // Emit an EOL-comment describing which TypeIndex this hash corresponds 663 // to, as well as the stringified SHA1 hash. 664 SmallString<32> Comment; 665 raw_svector_ostream CommentOS(Comment); 666 CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR); 667 OS.AddComment(Comment); 668 ++TI; 669 } 670 assert(GHR.Hash.size() == 8); 671 StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()), 672 GHR.Hash.size()); 673 OS.EmitBinaryData(S); 674 } 675 } 676 677 static SourceLanguage MapDWLangToCVLang(unsigned DWLang) { 678 switch (DWLang) { 679 case dwarf::DW_LANG_C: 680 case dwarf::DW_LANG_C89: 681 case dwarf::DW_LANG_C99: 682 case dwarf::DW_LANG_C11: 683 case dwarf::DW_LANG_ObjC: 684 return SourceLanguage::C; 685 case dwarf::DW_LANG_C_plus_plus: 686 case dwarf::DW_LANG_C_plus_plus_03: 687 case dwarf::DW_LANG_C_plus_plus_11: 688 case dwarf::DW_LANG_C_plus_plus_14: 689 return SourceLanguage::Cpp; 690 case dwarf::DW_LANG_Fortran77: 691 case dwarf::DW_LANG_Fortran90: 692 case dwarf::DW_LANG_Fortran03: 693 case dwarf::DW_LANG_Fortran08: 694 return SourceLanguage::Fortran; 695 case dwarf::DW_LANG_Pascal83: 696 return SourceLanguage::Pascal; 697 case dwarf::DW_LANG_Cobol74: 698 case dwarf::DW_LANG_Cobol85: 699 return SourceLanguage::Cobol; 700 case dwarf::DW_LANG_Java: 701 return SourceLanguage::Java; 702 case dwarf::DW_LANG_D: 703 return SourceLanguage::D; 704 case dwarf::DW_LANG_Swift: 705 return SourceLanguage::Swift; 706 default: 707 // There's no CodeView representation for this language, and CV doesn't 708 // have an "unknown" option for the language field, so we'll use MASM, 709 // as it's very low level. 710 return SourceLanguage::Masm; 711 } 712 } 713 714 namespace { 715 struct Version { 716 int Part[4]; 717 }; 718 } // end anonymous namespace 719 720 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out 721 // the version number. 722 static Version parseVersion(StringRef Name) { 723 Version V = {{0}}; 724 int N = 0; 725 for (const char C : Name) { 726 if (isdigit(C)) { 727 V.Part[N] *= 10; 728 V.Part[N] += C - '0'; 729 } else if (C == '.') { 730 ++N; 731 if (N >= 4) 732 return V; 733 } else if (N > 0) 734 return V; 735 } 736 return V; 737 } 738 739 void CodeViewDebug::emitCompilerInformation() { 740 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3); 741 uint32_t Flags = 0; 742 743 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 744 const MDNode *Node = *CUs->operands().begin(); 745 const auto *CU = cast<DICompileUnit>(Node); 746 747 // The low byte of the flags indicates the source language. 748 Flags = MapDWLangToCVLang(CU->getSourceLanguage()); 749 // TODO: Figure out which other flags need to be set. 750 751 OS.AddComment("Flags and language"); 752 OS.EmitIntValue(Flags, 4); 753 754 OS.AddComment("CPUType"); 755 OS.EmitIntValue(static_cast<uint64_t>(TheCPU), 2); 756 757 StringRef CompilerVersion = CU->getProducer(); 758 Version FrontVer = parseVersion(CompilerVersion); 759 OS.AddComment("Frontend version"); 760 for (int N = 0; N < 4; ++N) 761 OS.EmitIntValue(FrontVer.Part[N], 2); 762 763 // Some Microsoft tools, like Binscope, expect a backend version number of at 764 // least 8.something, so we'll coerce the LLVM version into a form that 765 // guarantees it'll be big enough without really lying about the version. 766 int Major = 1000 * LLVM_VERSION_MAJOR + 767 10 * LLVM_VERSION_MINOR + 768 LLVM_VERSION_PATCH; 769 // Clamp it for builds that use unusually large version numbers. 770 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 771 Version BackVer = {{ Major, 0, 0, 0 }}; 772 OS.AddComment("Backend version"); 773 for (int N = 0; N < 4; ++N) 774 OS.EmitIntValue(BackVer.Part[N], 2); 775 776 OS.AddComment("Null-terminated compiler version string"); 777 emitNullTerminatedSymbolName(OS, CompilerVersion); 778 779 endSymbolRecord(CompilerEnd); 780 } 781 782 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 783 StringRef S) { 784 StringIdRecord SIR(TypeIndex(0x0), S); 785 return TypeTable.writeLeafType(SIR); 786 } 787 788 void CodeViewDebug::emitBuildInfo() { 789 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 790 // build info. The known prefix is: 791 // - Absolute path of current directory 792 // - Compiler path 793 // - Main source file path, relative to CWD or absolute 794 // - Type server PDB file 795 // - Canonical compiler command line 796 // If frontend and backend compilation are separated (think llc or LTO), it's 797 // not clear if the compiler path should refer to the executable for the 798 // frontend or the backend. Leave it blank for now. 799 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 800 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 801 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 802 const auto *CU = cast<DICompileUnit>(Node); 803 const DIFile *MainSourceFile = CU->getFile(); 804 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 805 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 806 BuildInfoArgs[BuildInfoRecord::SourceFile] = 807 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 808 // FIXME: Path to compiler and command line. PDB is intentionally blank unless 809 // we implement /Zi type servers. 810 BuildInfoRecord BIR(BuildInfoArgs); 811 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 812 813 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 814 // from the module symbols into the type stream. 815 MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 816 MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO); 817 OS.AddComment("LF_BUILDINFO index"); 818 OS.EmitIntValue(BuildInfoIndex.getIndex(), 4); 819 endSymbolRecord(BIEnd); 820 endCVSubsection(BISubsecEnd); 821 } 822 823 void CodeViewDebug::emitInlineeLinesSubsection() { 824 if (InlinedSubprograms.empty()) 825 return; 826 827 OS.AddComment("Inlinee lines subsection"); 828 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 829 830 // We emit the checksum info for files. This is used by debuggers to 831 // determine if a pdb matches the source before loading it. Visual Studio, 832 // for instance, will display a warning that the breakpoints are not valid if 833 // the pdb does not match the source. 834 OS.AddComment("Inlinee lines signature"); 835 OS.EmitIntValue(unsigned(InlineeLinesSignature::Normal), 4); 836 837 for (const DISubprogram *SP : InlinedSubprograms) { 838 assert(TypeIndices.count({SP, nullptr})); 839 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 840 841 OS.AddBlankLine(); 842 unsigned FileId = maybeRecordFile(SP->getFile()); 843 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 844 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 845 OS.AddBlankLine(); 846 OS.AddComment("Type index of inlined function"); 847 OS.EmitIntValue(InlineeIdx.getIndex(), 4); 848 OS.AddComment("Offset into filechecksum table"); 849 OS.EmitCVFileChecksumOffsetDirective(FileId); 850 OS.AddComment("Starting line number"); 851 OS.EmitIntValue(SP->getLine(), 4); 852 } 853 854 endCVSubsection(InlineEnd); 855 } 856 857 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 858 const DILocation *InlinedAt, 859 const InlineSite &Site) { 860 assert(TypeIndices.count({Site.Inlinee, nullptr})); 861 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 862 863 // SymbolRecord 864 MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE); 865 866 OS.AddComment("PtrParent"); 867 OS.EmitIntValue(0, 4); 868 OS.AddComment("PtrEnd"); 869 OS.EmitIntValue(0, 4); 870 OS.AddComment("Inlinee type index"); 871 OS.EmitIntValue(InlineeIdx.getIndex(), 4); 872 873 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 874 unsigned StartLineNum = Site.Inlinee->getLine(); 875 876 OS.EmitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 877 FI.Begin, FI.End); 878 879 endSymbolRecord(InlineEnd); 880 881 emitLocalVariableList(FI, Site.InlinedLocals); 882 883 // Recurse on child inlined call sites before closing the scope. 884 for (const DILocation *ChildSite : Site.ChildSites) { 885 auto I = FI.InlineSites.find(ChildSite); 886 assert(I != FI.InlineSites.end() && 887 "child site not in function inline site map"); 888 emitInlinedCallSite(FI, ChildSite, I->second); 889 } 890 891 // Close the scope. 892 emitEndSymbolRecord(SymbolKind::S_INLINESITE_END); 893 } 894 895 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 896 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 897 // comdat key. A section may be comdat because of -ffunction-sections or 898 // because it is comdat in the IR. 899 MCSectionCOFF *GVSec = 900 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 901 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 902 903 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 904 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 905 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 906 907 OS.SwitchSection(DebugSec); 908 909 // Emit the magic version number if this is the first time we've switched to 910 // this section. 911 if (ComdatDebugSections.insert(DebugSec).second) 912 emitCodeViewMagicVersion(); 913 } 914 915 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 916 // The only supported thunk ordinal is currently the standard type. 917 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 918 FunctionInfo &FI, 919 const MCSymbol *Fn) { 920 std::string FuncName = GlobalValue::dropLLVMManglingEscape(GV->getName()); 921 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 922 923 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 924 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 925 926 // Emit S_THUNK32 927 MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32); 928 OS.AddComment("PtrParent"); 929 OS.EmitIntValue(0, 4); 930 OS.AddComment("PtrEnd"); 931 OS.EmitIntValue(0, 4); 932 OS.AddComment("PtrNext"); 933 OS.EmitIntValue(0, 4); 934 OS.AddComment("Thunk section relative address"); 935 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 936 OS.AddComment("Thunk section index"); 937 OS.EmitCOFFSectionIndex(Fn); 938 OS.AddComment("Code size"); 939 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 940 OS.AddComment("Ordinal"); 941 OS.EmitIntValue(unsigned(ordinal), 1); 942 OS.AddComment("Function name"); 943 emitNullTerminatedSymbolName(OS, FuncName); 944 // Additional fields specific to the thunk ordinal would go here. 945 endSymbolRecord(ThunkRecordEnd); 946 947 // Local variables/inlined routines are purposely omitted here. The point of 948 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 949 950 // Emit S_PROC_ID_END 951 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 952 953 endCVSubsection(SymbolsEnd); 954 } 955 956 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 957 FunctionInfo &FI) { 958 // For each function there is a separate subsection which holds the PC to 959 // file:line table. 960 const MCSymbol *Fn = Asm->getSymbol(GV); 961 assert(Fn); 962 963 // Switch to the to a comdat section, if appropriate. 964 switchToDebugSectionForSymbol(Fn); 965 966 std::string FuncName; 967 auto *SP = GV->getSubprogram(); 968 assert(SP); 969 setCurrentSubprogram(SP); 970 971 if (SP->isThunk()) { 972 emitDebugInfoForThunk(GV, FI, Fn); 973 return; 974 } 975 976 // If we have a display name, build the fully qualified name by walking the 977 // chain of scopes. 978 if (!SP->getName().empty()) 979 FuncName = getFullyQualifiedName(SP->getScope(), SP->getName()); 980 981 // If our DISubprogram name is empty, use the mangled name. 982 if (FuncName.empty()) 983 FuncName = GlobalValue::dropLLVMManglingEscape(GV->getName()); 984 985 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 986 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 987 OS.EmitCVFPOData(Fn); 988 989 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 990 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 991 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 992 { 993 SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID 994 : SymbolKind::S_GPROC32_ID; 995 MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind); 996 997 // These fields are filled in by tools like CVPACK which run after the fact. 998 OS.AddComment("PtrParent"); 999 OS.EmitIntValue(0, 4); 1000 OS.AddComment("PtrEnd"); 1001 OS.EmitIntValue(0, 4); 1002 OS.AddComment("PtrNext"); 1003 OS.EmitIntValue(0, 4); 1004 // This is the important bit that tells the debugger where the function 1005 // code is located and what's its size: 1006 OS.AddComment("Code size"); 1007 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4); 1008 OS.AddComment("Offset after prologue"); 1009 OS.EmitIntValue(0, 4); 1010 OS.AddComment("Offset before epilogue"); 1011 OS.EmitIntValue(0, 4); 1012 OS.AddComment("Function type index"); 1013 OS.EmitIntValue(getFuncIdForSubprogram(GV->getSubprogram()).getIndex(), 4); 1014 OS.AddComment("Function section relative address"); 1015 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1016 OS.AddComment("Function section index"); 1017 OS.EmitCOFFSectionIndex(Fn); 1018 OS.AddComment("Flags"); 1019 OS.EmitIntValue(0, 1); 1020 // Emit the function display name as a null-terminated string. 1021 OS.AddComment("Function name"); 1022 // Truncate the name so we won't overflow the record length field. 1023 emitNullTerminatedSymbolName(OS, FuncName); 1024 endSymbolRecord(ProcRecordEnd); 1025 1026 MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC); 1027 // Subtract out the CSR size since MSVC excludes that and we include it. 1028 OS.AddComment("FrameSize"); 1029 OS.EmitIntValue(FI.FrameSize - FI.CSRSize, 4); 1030 OS.AddComment("Padding"); 1031 OS.EmitIntValue(0, 4); 1032 OS.AddComment("Offset of padding"); 1033 OS.EmitIntValue(0, 4); 1034 OS.AddComment("Bytes of callee saved registers"); 1035 OS.EmitIntValue(FI.CSRSize, 4); 1036 OS.AddComment("Exception handler offset"); 1037 OS.EmitIntValue(0, 4); 1038 OS.AddComment("Exception handler section"); 1039 OS.EmitIntValue(0, 2); 1040 OS.AddComment("Flags (defines frame register)"); 1041 OS.EmitIntValue(uint32_t(FI.FrameProcOpts), 4); 1042 endSymbolRecord(FrameProcEnd); 1043 1044 emitLocalVariableList(FI, FI.Locals); 1045 emitGlobalVariableList(FI.Globals); 1046 emitLexicalBlockList(FI.ChildBlocks, FI); 1047 1048 // Emit inlined call site information. Only emit functions inlined directly 1049 // into the parent function. We'll emit the other sites recursively as part 1050 // of their parent inline site. 1051 for (const DILocation *InlinedAt : FI.ChildSites) { 1052 auto I = FI.InlineSites.find(InlinedAt); 1053 assert(I != FI.InlineSites.end() && 1054 "child site not in function inline site map"); 1055 emitInlinedCallSite(FI, InlinedAt, I->second); 1056 } 1057 1058 for (auto Annot : FI.Annotations) { 1059 MCSymbol *Label = Annot.first; 1060 MDTuple *Strs = cast<MDTuple>(Annot.second); 1061 MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION); 1062 OS.EmitCOFFSecRel32(Label, /*Offset=*/0); 1063 // FIXME: Make sure we don't overflow the max record size. 1064 OS.EmitCOFFSectionIndex(Label); 1065 OS.EmitIntValue(Strs->getNumOperands(), 2); 1066 for (Metadata *MD : Strs->operands()) { 1067 // MDStrings are null terminated, so we can do EmitBytes and get the 1068 // nice .asciz directive. 1069 StringRef Str = cast<MDString>(MD)->getString(); 1070 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1071 OS.EmitBytes(StringRef(Str.data(), Str.size() + 1)); 1072 } 1073 endSymbolRecord(AnnotEnd); 1074 } 1075 1076 for (auto HeapAllocSite : FI.HeapAllocSites) { 1077 MCSymbol *BeginLabel = std::get<0>(HeapAllocSite); 1078 MCSymbol *EndLabel = std::get<1>(HeapAllocSite); 1079 1080 // The labels might not be defined if the instruction was replaced 1081 // somewhere in the codegen pipeline. 1082 if (!BeginLabel->isDefined() || !EndLabel->isDefined()) 1083 continue; 1084 1085 DIType *DITy = std::get<2>(HeapAllocSite); 1086 MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE); 1087 OS.AddComment("Call site offset"); 1088 OS.EmitCOFFSecRel32(BeginLabel, /*Offset=*/0); 1089 OS.AddComment("Call site section index"); 1090 OS.EmitCOFFSectionIndex(BeginLabel); 1091 OS.AddComment("Call instruction length"); 1092 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 1093 OS.AddComment("Type index"); 1094 OS.EmitIntValue(getCompleteTypeIndex(DITy).getIndex(), 4); 1095 endSymbolRecord(HeapAllocEnd); 1096 } 1097 1098 if (SP != nullptr) 1099 emitDebugInfoForUDTs(LocalUDTs); 1100 1101 // We're done with this function. 1102 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1103 } 1104 endCVSubsection(SymbolsEnd); 1105 1106 // We have an assembler directive that takes care of the whole line table. 1107 OS.EmitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1108 } 1109 1110 CodeViewDebug::LocalVarDefRange 1111 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1112 LocalVarDefRange DR; 1113 DR.InMemory = -1; 1114 DR.DataOffset = Offset; 1115 assert(DR.DataOffset == Offset && "truncation"); 1116 DR.IsSubfield = 0; 1117 DR.StructOffset = 0; 1118 DR.CVRegister = CVRegister; 1119 return DR; 1120 } 1121 1122 void CodeViewDebug::collectVariableInfoFromMFTable( 1123 DenseSet<InlinedEntity> &Processed) { 1124 const MachineFunction &MF = *Asm->MF; 1125 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1126 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1127 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1128 1129 for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) { 1130 if (!VI.Var) 1131 continue; 1132 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1133 "Expected inlined-at fields to agree"); 1134 1135 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1136 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1137 1138 // If variable scope is not found then skip this variable. 1139 if (!Scope) 1140 continue; 1141 1142 // If the variable has an attached offset expression, extract it. 1143 // FIXME: Try to handle DW_OP_deref as well. 1144 int64_t ExprOffset = 0; 1145 if (VI.Expr) 1146 if (!VI.Expr->extractIfOffset(ExprOffset)) 1147 continue; 1148 1149 // Get the frame register used and the offset. 1150 unsigned FrameReg = 0; 1151 int FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg); 1152 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1153 1154 // Calculate the label ranges. 1155 LocalVarDefRange DefRange = 1156 createDefRangeMem(CVReg, FrameOffset + ExprOffset); 1157 for (const InsnRange &Range : Scope->getRanges()) { 1158 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1159 const MCSymbol *End = getLabelAfterInsn(Range.second); 1160 End = End ? End : Asm->getFunctionEnd(); 1161 DefRange.Ranges.emplace_back(Begin, End); 1162 } 1163 1164 LocalVariable Var; 1165 Var.DIVar = VI.Var; 1166 Var.DefRanges.emplace_back(std::move(DefRange)); 1167 recordLocalVariable(std::move(Var), Scope); 1168 } 1169 } 1170 1171 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1172 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1173 } 1174 1175 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1176 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1177 } 1178 1179 void CodeViewDebug::calculateRanges( 1180 LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { 1181 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1182 1183 // Calculate the definition ranges. 1184 for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { 1185 const auto &Entry = *I; 1186 if (!Entry.isDbgValue()) 1187 continue; 1188 const MachineInstr *DVInst = Entry.getInstr(); 1189 assert(DVInst->isDebugValue() && "Invalid History entry"); 1190 // FIXME: Find a way to represent constant variables, since they are 1191 // relatively common. 1192 Optional<DbgVariableLocation> Location = 1193 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1194 if (!Location) 1195 continue; 1196 1197 // CodeView can only express variables in register and variables in memory 1198 // at a constant offset from a register. However, for variables passed 1199 // indirectly by pointer, it is common for that pointer to be spilled to a 1200 // stack location. For the special case of one offseted load followed by a 1201 // zero offset load (a pointer spilled to the stack), we change the type of 1202 // the local variable from a value type to a reference type. This tricks the 1203 // debugger into doing the load for us. 1204 if (Var.UseReferenceType) { 1205 // We're using a reference type. Drop the last zero offset load. 1206 if (canUseReferenceType(*Location)) 1207 Location->LoadChain.pop_back(); 1208 else 1209 continue; 1210 } else if (needsReferenceType(*Location)) { 1211 // This location can't be expressed without switching to a reference type. 1212 // Start over using that. 1213 Var.UseReferenceType = true; 1214 Var.DefRanges.clear(); 1215 calculateRanges(Var, Entries); 1216 return; 1217 } 1218 1219 // We can only handle a register or an offseted load of a register. 1220 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1221 continue; 1222 { 1223 LocalVarDefRange DR; 1224 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1225 DR.InMemory = !Location->LoadChain.empty(); 1226 DR.DataOffset = 1227 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1228 if (Location->FragmentInfo) { 1229 DR.IsSubfield = true; 1230 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1231 } else { 1232 DR.IsSubfield = false; 1233 DR.StructOffset = 0; 1234 } 1235 1236 if (Var.DefRanges.empty() || 1237 Var.DefRanges.back().isDifferentLocation(DR)) { 1238 Var.DefRanges.emplace_back(std::move(DR)); 1239 } 1240 } 1241 1242 // Compute the label range. 1243 const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr()); 1244 const MCSymbol *End; 1245 if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { 1246 auto &EndingEntry = Entries[Entry.getEndIndex()]; 1247 End = EndingEntry.isDbgValue() 1248 ? getLabelBeforeInsn(EndingEntry.getInstr()) 1249 : getLabelAfterInsn(EndingEntry.getInstr()); 1250 } else 1251 End = Asm->getFunctionEnd(); 1252 1253 // If the last range end is our begin, just extend the last range. 1254 // Otherwise make a new range. 1255 SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = 1256 Var.DefRanges.back().Ranges; 1257 if (!R.empty() && R.back().second == Begin) 1258 R.back().second = End; 1259 else 1260 R.emplace_back(Begin, End); 1261 1262 // FIXME: Do more range combining. 1263 } 1264 } 1265 1266 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { 1267 DenseSet<InlinedEntity> Processed; 1268 // Grab the variable info that was squirreled away in the MMI side-table. 1269 collectVariableInfoFromMFTable(Processed); 1270 1271 for (const auto &I : DbgValues) { 1272 InlinedEntity IV = I.first; 1273 if (Processed.count(IV)) 1274 continue; 1275 const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first); 1276 const DILocation *InlinedAt = IV.second; 1277 1278 // Instruction ranges, specifying where IV is accessible. 1279 const auto &Entries = I.second; 1280 1281 LexicalScope *Scope = nullptr; 1282 if (InlinedAt) 1283 Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt); 1284 else 1285 Scope = LScopes.findLexicalScope(DIVar->getScope()); 1286 // If variable scope is not found then skip this variable. 1287 if (!Scope) 1288 continue; 1289 1290 LocalVariable Var; 1291 Var.DIVar = DIVar; 1292 1293 calculateRanges(Var, Entries); 1294 recordLocalVariable(std::move(Var), Scope); 1295 } 1296 } 1297 1298 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { 1299 const TargetSubtargetInfo &TSI = MF->getSubtarget(); 1300 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1301 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1302 const Function &GV = MF->getFunction(); 1303 auto Insertion = FnDebugInfo.insert({&GV, llvm::make_unique<FunctionInfo>()}); 1304 assert(Insertion.second && "function already has info"); 1305 CurFn = Insertion.first->second.get(); 1306 CurFn->FuncId = NextFuncId++; 1307 CurFn->Begin = Asm->getFunctionBegin(); 1308 1309 // The S_FRAMEPROC record reports the stack size, and how many bytes of 1310 // callee-saved registers were used. For targets that don't use a PUSH 1311 // instruction (AArch64), this will be zero. 1312 CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); 1313 CurFn->FrameSize = MFI.getStackSize(); 1314 CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); 1315 CurFn->HasStackRealignment = TRI->needsStackRealignment(*MF); 1316 1317 // For this function S_FRAMEPROC record, figure out which codeview register 1318 // will be the frame pointer. 1319 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. 1320 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. 1321 if (CurFn->FrameSize > 0) { 1322 if (!TSI.getFrameLowering()->hasFP(*MF)) { 1323 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1324 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; 1325 } else { 1326 // If there is an FP, parameters are always relative to it. 1327 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; 1328 if (CurFn->HasStackRealignment) { 1329 // If the stack needs realignment, locals are relative to SP or VFRAME. 1330 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1331 } else { 1332 // Otherwise, locals are relative to EBP, and we probably have VLAs or 1333 // other stack adjustments. 1334 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; 1335 } 1336 } 1337 } 1338 1339 // Compute other frame procedure options. 1340 FrameProcedureOptions FPO = FrameProcedureOptions::None; 1341 if (MFI.hasVarSizedObjects()) 1342 FPO |= FrameProcedureOptions::HasAlloca; 1343 if (MF->exposesReturnsTwice()) 1344 FPO |= FrameProcedureOptions::HasSetJmp; 1345 // FIXME: Set HasLongJmp if we ever track that info. 1346 if (MF->hasInlineAsm()) 1347 FPO |= FrameProcedureOptions::HasInlineAssembly; 1348 if (GV.hasPersonalityFn()) { 1349 if (isAsynchronousEHPersonality( 1350 classifyEHPersonality(GV.getPersonalityFn()))) 1351 FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; 1352 else 1353 FPO |= FrameProcedureOptions::HasExceptionHandling; 1354 } 1355 if (GV.hasFnAttribute(Attribute::InlineHint)) 1356 FPO |= FrameProcedureOptions::MarkedInline; 1357 if (GV.hasFnAttribute(Attribute::Naked)) 1358 FPO |= FrameProcedureOptions::Naked; 1359 if (MFI.hasStackProtectorIndex()) 1360 FPO |= FrameProcedureOptions::SecurityChecks; 1361 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); 1362 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); 1363 if (Asm->TM.getOptLevel() != CodeGenOpt::None && 1364 !GV.hasOptSize() && !GV.hasOptNone()) 1365 FPO |= FrameProcedureOptions::OptimizedForSpeed; 1366 // FIXME: Set GuardCfg when it is implemented. 1367 CurFn->FrameProcOpts = FPO; 1368 1369 OS.EmitCVFuncIdDirective(CurFn->FuncId); 1370 1371 // Find the end of the function prolog. First known non-DBG_VALUE and 1372 // non-frame setup location marks the beginning of the function body. 1373 // FIXME: is there a simpler a way to do this? Can we just search 1374 // for the first instruction of the function, not the last of the prolog? 1375 DebugLoc PrologEndLoc; 1376 bool EmptyPrologue = true; 1377 for (const auto &MBB : *MF) { 1378 for (const auto &MI : MBB) { 1379 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 1380 MI.getDebugLoc()) { 1381 PrologEndLoc = MI.getDebugLoc(); 1382 break; 1383 } else if (!MI.isMetaInstruction()) { 1384 EmptyPrologue = false; 1385 } 1386 } 1387 } 1388 1389 // Record beginning of function if we have a non-empty prologue. 1390 if (PrologEndLoc && !EmptyPrologue) { 1391 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); 1392 maybeRecordLocation(FnStartDL, MF); 1393 } 1394 } 1395 1396 static bool shouldEmitUdt(const DIType *T) { 1397 if (!T) 1398 return false; 1399 1400 // MSVC does not emit UDTs for typedefs that are scoped to classes. 1401 if (T->getTag() == dwarf::DW_TAG_typedef) { 1402 if (DIScope *Scope = T->getScope()) { 1403 switch (Scope->getTag()) { 1404 case dwarf::DW_TAG_structure_type: 1405 case dwarf::DW_TAG_class_type: 1406 case dwarf::DW_TAG_union_type: 1407 return false; 1408 } 1409 } 1410 } 1411 1412 while (true) { 1413 if (!T || T->isForwardDecl()) 1414 return false; 1415 1416 const DIDerivedType *DT = dyn_cast<DIDerivedType>(T); 1417 if (!DT) 1418 return true; 1419 T = DT->getBaseType(); 1420 } 1421 return true; 1422 } 1423 1424 void CodeViewDebug::addToUDTs(const DIType *Ty) { 1425 // Don't record empty UDTs. 1426 if (Ty->getName().empty()) 1427 return; 1428 if (!shouldEmitUdt(Ty)) 1429 return; 1430 1431 SmallVector<StringRef, 5> QualifiedNameComponents; 1432 const DISubprogram *ClosestSubprogram = 1433 getQualifiedNameComponents(Ty->getScope(), QualifiedNameComponents); 1434 1435 std::string FullyQualifiedName = 1436 getQualifiedName(QualifiedNameComponents, getPrettyScopeName(Ty)); 1437 1438 if (ClosestSubprogram == nullptr) { 1439 GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1440 } else if (ClosestSubprogram == CurrentSubprogram) { 1441 LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1442 } 1443 1444 // TODO: What if the ClosestSubprogram is neither null or the current 1445 // subprogram? Currently, the UDT just gets dropped on the floor. 1446 // 1447 // The current behavior is not desirable. To get maximal fidelity, we would 1448 // need to perform all type translation before beginning emission of .debug$S 1449 // and then make LocalUDTs a member of FunctionInfo 1450 } 1451 1452 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { 1453 // Generic dispatch for lowering an unknown type. 1454 switch (Ty->getTag()) { 1455 case dwarf::DW_TAG_array_type: 1456 return lowerTypeArray(cast<DICompositeType>(Ty)); 1457 case dwarf::DW_TAG_typedef: 1458 return lowerTypeAlias(cast<DIDerivedType>(Ty)); 1459 case dwarf::DW_TAG_base_type: 1460 return lowerTypeBasic(cast<DIBasicType>(Ty)); 1461 case dwarf::DW_TAG_pointer_type: 1462 if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type") 1463 return lowerTypeVFTableShape(cast<DIDerivedType>(Ty)); 1464 LLVM_FALLTHROUGH; 1465 case dwarf::DW_TAG_reference_type: 1466 case dwarf::DW_TAG_rvalue_reference_type: 1467 return lowerTypePointer(cast<DIDerivedType>(Ty)); 1468 case dwarf::DW_TAG_ptr_to_member_type: 1469 return lowerTypeMemberPointer(cast<DIDerivedType>(Ty)); 1470 case dwarf::DW_TAG_restrict_type: 1471 case dwarf::DW_TAG_const_type: 1472 case dwarf::DW_TAG_volatile_type: 1473 // TODO: add support for DW_TAG_atomic_type here 1474 return lowerTypeModifier(cast<DIDerivedType>(Ty)); 1475 case dwarf::DW_TAG_subroutine_type: 1476 if (ClassTy) { 1477 // The member function type of a member function pointer has no 1478 // ThisAdjustment. 1479 return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy, 1480 /*ThisAdjustment=*/0, 1481 /*IsStaticMethod=*/false); 1482 } 1483 return lowerTypeFunction(cast<DISubroutineType>(Ty)); 1484 case dwarf::DW_TAG_enumeration_type: 1485 return lowerTypeEnum(cast<DICompositeType>(Ty)); 1486 case dwarf::DW_TAG_class_type: 1487 case dwarf::DW_TAG_structure_type: 1488 return lowerTypeClass(cast<DICompositeType>(Ty)); 1489 case dwarf::DW_TAG_union_type: 1490 return lowerTypeUnion(cast<DICompositeType>(Ty)); 1491 case dwarf::DW_TAG_unspecified_type: 1492 if (Ty->getName() == "decltype(nullptr)") 1493 return TypeIndex::NullptrT(); 1494 return TypeIndex::None(); 1495 default: 1496 // Use the null type index. 1497 return TypeIndex(); 1498 } 1499 } 1500 1501 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { 1502 TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty->getBaseType()); 1503 StringRef TypeName = Ty->getName(); 1504 1505 addToUDTs(Ty); 1506 1507 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && 1508 TypeName == "HRESULT") 1509 return TypeIndex(SimpleTypeKind::HResult); 1510 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && 1511 TypeName == "wchar_t") 1512 return TypeIndex(SimpleTypeKind::WideCharacter); 1513 1514 return UnderlyingTypeIndex; 1515 } 1516 1517 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { 1518 const DIType *ElementType = Ty->getBaseType(); 1519 TypeIndex ElementTypeIndex = getTypeIndex(ElementType); 1520 // IndexType is size_t, which depends on the bitness of the target. 1521 TypeIndex IndexType = getPointerSizeInBytes() == 8 1522 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1523 : TypeIndex(SimpleTypeKind::UInt32Long); 1524 1525 uint64_t ElementSize = getBaseTypeSize(ElementType) / 8; 1526 1527 // Add subranges to array type. 1528 DINodeArray Elements = Ty->getElements(); 1529 for (int i = Elements.size() - 1; i >= 0; --i) { 1530 const DINode *Element = Elements[i]; 1531 assert(Element->getTag() == dwarf::DW_TAG_subrange_type); 1532 1533 const DISubrange *Subrange = cast<DISubrange>(Element); 1534 assert(Subrange->getLowerBound() == 0 && 1535 "codeview doesn't support subranges with lower bounds"); 1536 int64_t Count = -1; 1537 if (auto *CI = Subrange->getCount().dyn_cast<ConstantInt*>()) 1538 Count = CI->getSExtValue(); 1539 1540 // Forward declarations of arrays without a size and VLAs use a count of -1. 1541 // Emit a count of zero in these cases to match what MSVC does for arrays 1542 // without a size. MSVC doesn't support VLAs, so it's not clear what we 1543 // should do for them even if we could distinguish them. 1544 if (Count == -1) 1545 Count = 0; 1546 1547 // Update the element size and element type index for subsequent subranges. 1548 ElementSize *= Count; 1549 1550 // If this is the outermost array, use the size from the array. It will be 1551 // more accurate if we had a VLA or an incomplete element type size. 1552 uint64_t ArraySize = 1553 (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; 1554 1555 StringRef Name = (i == 0) ? Ty->getName() : ""; 1556 ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); 1557 ElementTypeIndex = TypeTable.writeLeafType(AR); 1558 } 1559 1560 return ElementTypeIndex; 1561 } 1562 1563 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { 1564 TypeIndex Index; 1565 dwarf::TypeKind Kind; 1566 uint32_t ByteSize; 1567 1568 Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); 1569 ByteSize = Ty->getSizeInBits() / 8; 1570 1571 SimpleTypeKind STK = SimpleTypeKind::None; 1572 switch (Kind) { 1573 case dwarf::DW_ATE_address: 1574 // FIXME: Translate 1575 break; 1576 case dwarf::DW_ATE_boolean: 1577 switch (ByteSize) { 1578 case 1: STK = SimpleTypeKind::Boolean8; break; 1579 case 2: STK = SimpleTypeKind::Boolean16; break; 1580 case 4: STK = SimpleTypeKind::Boolean32; break; 1581 case 8: STK = SimpleTypeKind::Boolean64; break; 1582 case 16: STK = SimpleTypeKind::Boolean128; break; 1583 } 1584 break; 1585 case dwarf::DW_ATE_complex_float: 1586 switch (ByteSize) { 1587 case 2: STK = SimpleTypeKind::Complex16; break; 1588 case 4: STK = SimpleTypeKind::Complex32; break; 1589 case 8: STK = SimpleTypeKind::Complex64; break; 1590 case 10: STK = SimpleTypeKind::Complex80; break; 1591 case 16: STK = SimpleTypeKind::Complex128; break; 1592 } 1593 break; 1594 case dwarf::DW_ATE_float: 1595 switch (ByteSize) { 1596 case 2: STK = SimpleTypeKind::Float16; break; 1597 case 4: STK = SimpleTypeKind::Float32; break; 1598 case 6: STK = SimpleTypeKind::Float48; break; 1599 case 8: STK = SimpleTypeKind::Float64; break; 1600 case 10: STK = SimpleTypeKind::Float80; break; 1601 case 16: STK = SimpleTypeKind::Float128; break; 1602 } 1603 break; 1604 case dwarf::DW_ATE_signed: 1605 switch (ByteSize) { 1606 case 1: STK = SimpleTypeKind::SignedCharacter; break; 1607 case 2: STK = SimpleTypeKind::Int16Short; break; 1608 case 4: STK = SimpleTypeKind::Int32; break; 1609 case 8: STK = SimpleTypeKind::Int64Quad; break; 1610 case 16: STK = SimpleTypeKind::Int128Oct; break; 1611 } 1612 break; 1613 case dwarf::DW_ATE_unsigned: 1614 switch (ByteSize) { 1615 case 1: STK = SimpleTypeKind::UnsignedCharacter; break; 1616 case 2: STK = SimpleTypeKind::UInt16Short; break; 1617 case 4: STK = SimpleTypeKind::UInt32; break; 1618 case 8: STK = SimpleTypeKind::UInt64Quad; break; 1619 case 16: STK = SimpleTypeKind::UInt128Oct; break; 1620 } 1621 break; 1622 case dwarf::DW_ATE_UTF: 1623 switch (ByteSize) { 1624 case 2: STK = SimpleTypeKind::Character16; break; 1625 case 4: STK = SimpleTypeKind::Character32; break; 1626 } 1627 break; 1628 case dwarf::DW_ATE_signed_char: 1629 if (ByteSize == 1) 1630 STK = SimpleTypeKind::SignedCharacter; 1631 break; 1632 case dwarf::DW_ATE_unsigned_char: 1633 if (ByteSize == 1) 1634 STK = SimpleTypeKind::UnsignedCharacter; 1635 break; 1636 default: 1637 break; 1638 } 1639 1640 // Apply some fixups based on the source-level type name. 1641 if (STK == SimpleTypeKind::Int32 && Ty->getName() == "long int") 1642 STK = SimpleTypeKind::Int32Long; 1643 if (STK == SimpleTypeKind::UInt32 && Ty->getName() == "long unsigned int") 1644 STK = SimpleTypeKind::UInt32Long; 1645 if (STK == SimpleTypeKind::UInt16Short && 1646 (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t")) 1647 STK = SimpleTypeKind::WideCharacter; 1648 if ((STK == SimpleTypeKind::SignedCharacter || 1649 STK == SimpleTypeKind::UnsignedCharacter) && 1650 Ty->getName() == "char") 1651 STK = SimpleTypeKind::NarrowCharacter; 1652 1653 return TypeIndex(STK); 1654 } 1655 1656 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, 1657 PointerOptions PO) { 1658 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType()); 1659 1660 // Pointers to simple types without any options can use SimpleTypeMode, rather 1661 // than having a dedicated pointer type record. 1662 if (PointeeTI.isSimple() && PO == PointerOptions::None && 1663 PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && 1664 Ty->getTag() == dwarf::DW_TAG_pointer_type) { 1665 SimpleTypeMode Mode = Ty->getSizeInBits() == 64 1666 ? SimpleTypeMode::NearPointer64 1667 : SimpleTypeMode::NearPointer32; 1668 return TypeIndex(PointeeTI.getSimpleKind(), Mode); 1669 } 1670 1671 PointerKind PK = 1672 Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; 1673 PointerMode PM = PointerMode::Pointer; 1674 switch (Ty->getTag()) { 1675 default: llvm_unreachable("not a pointer tag type"); 1676 case dwarf::DW_TAG_pointer_type: 1677 PM = PointerMode::Pointer; 1678 break; 1679 case dwarf::DW_TAG_reference_type: 1680 PM = PointerMode::LValueReference; 1681 break; 1682 case dwarf::DW_TAG_rvalue_reference_type: 1683 PM = PointerMode::RValueReference; 1684 break; 1685 } 1686 1687 if (Ty->isObjectPointer()) 1688 PO |= PointerOptions::Const; 1689 1690 PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); 1691 return TypeTable.writeLeafType(PR); 1692 } 1693 1694 static PointerToMemberRepresentation 1695 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { 1696 // SizeInBytes being zero generally implies that the member pointer type was 1697 // incomplete, which can happen if it is part of a function prototype. In this 1698 // case, use the unknown model instead of the general model. 1699 if (IsPMF) { 1700 switch (Flags & DINode::FlagPtrToMemberRep) { 1701 case 0: 1702 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1703 : PointerToMemberRepresentation::GeneralFunction; 1704 case DINode::FlagSingleInheritance: 1705 return PointerToMemberRepresentation::SingleInheritanceFunction; 1706 case DINode::FlagMultipleInheritance: 1707 return PointerToMemberRepresentation::MultipleInheritanceFunction; 1708 case DINode::FlagVirtualInheritance: 1709 return PointerToMemberRepresentation::VirtualInheritanceFunction; 1710 } 1711 } else { 1712 switch (Flags & DINode::FlagPtrToMemberRep) { 1713 case 0: 1714 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1715 : PointerToMemberRepresentation::GeneralData; 1716 case DINode::FlagSingleInheritance: 1717 return PointerToMemberRepresentation::SingleInheritanceData; 1718 case DINode::FlagMultipleInheritance: 1719 return PointerToMemberRepresentation::MultipleInheritanceData; 1720 case DINode::FlagVirtualInheritance: 1721 return PointerToMemberRepresentation::VirtualInheritanceData; 1722 } 1723 } 1724 llvm_unreachable("invalid ptr to member representation"); 1725 } 1726 1727 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, 1728 PointerOptions PO) { 1729 assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); 1730 TypeIndex ClassTI = getTypeIndex(Ty->getClassType()); 1731 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType(), Ty->getClassType()); 1732 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 1733 : PointerKind::Near32; 1734 bool IsPMF = isa<DISubroutineType>(Ty->getBaseType()); 1735 PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction 1736 : PointerMode::PointerToDataMember; 1737 1738 assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big"); 1739 uint8_t SizeInBytes = Ty->getSizeInBits() / 8; 1740 MemberPointerInfo MPI( 1741 ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags())); 1742 PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); 1743 return TypeTable.writeLeafType(PR); 1744 } 1745 1746 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't 1747 /// have a translation, use the NearC convention. 1748 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { 1749 switch (DwarfCC) { 1750 case dwarf::DW_CC_normal: return CallingConvention::NearC; 1751 case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; 1752 case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; 1753 case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; 1754 case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; 1755 case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; 1756 } 1757 return CallingConvention::NearC; 1758 } 1759 1760 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { 1761 ModifierOptions Mods = ModifierOptions::None; 1762 PointerOptions PO = PointerOptions::None; 1763 bool IsModifier = true; 1764 const DIType *BaseTy = Ty; 1765 while (IsModifier && BaseTy) { 1766 // FIXME: Need to add DWARF tags for __unaligned and _Atomic 1767 switch (BaseTy->getTag()) { 1768 case dwarf::DW_TAG_const_type: 1769 Mods |= ModifierOptions::Const; 1770 PO |= PointerOptions::Const; 1771 break; 1772 case dwarf::DW_TAG_volatile_type: 1773 Mods |= ModifierOptions::Volatile; 1774 PO |= PointerOptions::Volatile; 1775 break; 1776 case dwarf::DW_TAG_restrict_type: 1777 // Only pointer types be marked with __restrict. There is no known flag 1778 // for __restrict in LF_MODIFIER records. 1779 PO |= PointerOptions::Restrict; 1780 break; 1781 default: 1782 IsModifier = false; 1783 break; 1784 } 1785 if (IsModifier) 1786 BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType(); 1787 } 1788 1789 // Check if the inner type will use an LF_POINTER record. If so, the 1790 // qualifiers will go in the LF_POINTER record. This comes up for types like 1791 // 'int *const' and 'int *__restrict', not the more common cases like 'const 1792 // char *'. 1793 if (BaseTy) { 1794 switch (BaseTy->getTag()) { 1795 case dwarf::DW_TAG_pointer_type: 1796 case dwarf::DW_TAG_reference_type: 1797 case dwarf::DW_TAG_rvalue_reference_type: 1798 return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO); 1799 case dwarf::DW_TAG_ptr_to_member_type: 1800 return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO); 1801 default: 1802 break; 1803 } 1804 } 1805 1806 TypeIndex ModifiedTI = getTypeIndex(BaseTy); 1807 1808 // Return the base type index if there aren't any modifiers. For example, the 1809 // metadata could contain restrict wrappers around non-pointer types. 1810 if (Mods == ModifierOptions::None) 1811 return ModifiedTI; 1812 1813 ModifierRecord MR(ModifiedTI, Mods); 1814 return TypeTable.writeLeafType(MR); 1815 } 1816 1817 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { 1818 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 1819 for (const DIType *ArgType : Ty->getTypeArray()) 1820 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgType)); 1821 1822 // MSVC uses type none for variadic argument. 1823 if (ReturnAndArgTypeIndices.size() > 1 && 1824 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 1825 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 1826 } 1827 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 1828 ArrayRef<TypeIndex> ArgTypeIndices = None; 1829 if (!ReturnAndArgTypeIndices.empty()) { 1830 auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices); 1831 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 1832 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 1833 } 1834 1835 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 1836 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 1837 1838 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 1839 1840 FunctionOptions FO = getFunctionOptions(Ty); 1841 ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), 1842 ArgListIndex); 1843 return TypeTable.writeLeafType(Procedure); 1844 } 1845 1846 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, 1847 const DIType *ClassTy, 1848 int ThisAdjustment, 1849 bool IsStaticMethod, 1850 FunctionOptions FO) { 1851 // Lower the containing class type. 1852 TypeIndex ClassType = getTypeIndex(ClassTy); 1853 1854 DITypeRefArray ReturnAndArgs = Ty->getTypeArray(); 1855 1856 unsigned Index = 0; 1857 SmallVector<TypeIndex, 8> ArgTypeIndices; 1858 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 1859 if (ReturnAndArgs.size() > Index) { 1860 ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]); 1861 } 1862 1863 // If the first argument is a pointer type and this isn't a static method, 1864 // treat it as the special 'this' parameter, which is encoded separately from 1865 // the arguments. 1866 TypeIndex ThisTypeIndex; 1867 if (!IsStaticMethod && ReturnAndArgs.size() > Index) { 1868 if (const DIDerivedType *PtrTy = 1869 dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index])) { 1870 if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { 1871 ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty); 1872 Index++; 1873 } 1874 } 1875 } 1876 1877 while (Index < ReturnAndArgs.size()) 1878 ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++])); 1879 1880 // MSVC uses type none for variadic argument. 1881 if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) 1882 ArgTypeIndices.back() = TypeIndex::None(); 1883 1884 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 1885 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 1886 1887 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 1888 1889 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 1890 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 1891 return TypeTable.writeLeafType(MFR); 1892 } 1893 1894 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 1895 unsigned VSlotCount = 1896 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 1897 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 1898 1899 VFTableShapeRecord VFTSR(Slots); 1900 return TypeTable.writeLeafType(VFTSR); 1901 } 1902 1903 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 1904 switch (Flags & DINode::FlagAccessibility) { 1905 case DINode::FlagPrivate: return MemberAccess::Private; 1906 case DINode::FlagPublic: return MemberAccess::Public; 1907 case DINode::FlagProtected: return MemberAccess::Protected; 1908 case 0: 1909 // If there was no explicit access control, provide the default for the tag. 1910 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 1911 : MemberAccess::Public; 1912 } 1913 llvm_unreachable("access flags are exclusive"); 1914 } 1915 1916 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 1917 if (SP->isArtificial()) 1918 return MethodOptions::CompilerGenerated; 1919 1920 // FIXME: Handle other MethodOptions. 1921 1922 return MethodOptions::None; 1923 } 1924 1925 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 1926 bool Introduced) { 1927 if (SP->getFlags() & DINode::FlagStaticMember) 1928 return MethodKind::Static; 1929 1930 switch (SP->getVirtuality()) { 1931 case dwarf::DW_VIRTUALITY_none: 1932 break; 1933 case dwarf::DW_VIRTUALITY_virtual: 1934 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 1935 case dwarf::DW_VIRTUALITY_pure_virtual: 1936 return Introduced ? MethodKind::PureIntroducingVirtual 1937 : MethodKind::PureVirtual; 1938 default: 1939 llvm_unreachable("unhandled virtuality case"); 1940 } 1941 1942 return MethodKind::Vanilla; 1943 } 1944 1945 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 1946 switch (Ty->getTag()) { 1947 case dwarf::DW_TAG_class_type: return TypeRecordKind::Class; 1948 case dwarf::DW_TAG_structure_type: return TypeRecordKind::Struct; 1949 } 1950 llvm_unreachable("unexpected tag"); 1951 } 1952 1953 /// Return ClassOptions that should be present on both the forward declaration 1954 /// and the defintion of a tag type. 1955 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 1956 ClassOptions CO = ClassOptions::None; 1957 1958 // MSVC always sets this flag, even for local types. Clang doesn't always 1959 // appear to give every type a linkage name, which may be problematic for us. 1960 // FIXME: Investigate the consequences of not following them here. 1961 if (!Ty->getIdentifier().empty()) 1962 CO |= ClassOptions::HasUniqueName; 1963 1964 // Put the Nested flag on a type if it appears immediately inside a tag type. 1965 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 1966 // here. That flag is only set on definitions, and not forward declarations. 1967 const DIScope *ImmediateScope = Ty->getScope(); 1968 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 1969 CO |= ClassOptions::Nested; 1970 1971 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 1972 // type only when it has an immediate function scope. Clang never puts enums 1973 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 1974 // always in function, class, or file scopes. 1975 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 1976 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 1977 CO |= ClassOptions::Scoped; 1978 } else { 1979 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 1980 Scope = Scope->getScope()) { 1981 if (isa<DISubprogram>(Scope)) { 1982 CO |= ClassOptions::Scoped; 1983 break; 1984 } 1985 } 1986 } 1987 1988 return CO; 1989 } 1990 1991 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 1992 switch (Ty->getTag()) { 1993 case dwarf::DW_TAG_class_type: 1994 case dwarf::DW_TAG_structure_type: 1995 case dwarf::DW_TAG_union_type: 1996 case dwarf::DW_TAG_enumeration_type: 1997 break; 1998 default: 1999 return; 2000 } 2001 2002 if (const auto *File = Ty->getFile()) { 2003 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 2004 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 2005 2006 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 2007 TypeTable.writeLeafType(USLR); 2008 } 2009 } 2010 2011 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 2012 ClassOptions CO = getCommonClassOptions(Ty); 2013 TypeIndex FTI; 2014 unsigned EnumeratorCount = 0; 2015 2016 if (Ty->isForwardDecl()) { 2017 CO |= ClassOptions::ForwardReference; 2018 } else { 2019 ContinuationRecordBuilder ContinuationBuilder; 2020 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2021 for (const DINode *Element : Ty->getElements()) { 2022 // We assume that the frontend provides all members in source declaration 2023 // order, which is what MSVC does. 2024 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2025 EnumeratorRecord ER(MemberAccess::Public, 2026 APSInt::getUnsigned(Enumerator->getValue()), 2027 Enumerator->getName()); 2028 ContinuationBuilder.writeMemberType(ER); 2029 EnumeratorCount++; 2030 } 2031 } 2032 FTI = TypeTable.insertRecord(ContinuationBuilder); 2033 } 2034 2035 std::string FullName = getFullyQualifiedName(Ty); 2036 2037 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2038 getTypeIndex(Ty->getBaseType())); 2039 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2040 2041 addUDTSrcLine(Ty, EnumTI); 2042 2043 return EnumTI; 2044 } 2045 2046 //===----------------------------------------------------------------------===// 2047 // ClassInfo 2048 //===----------------------------------------------------------------------===// 2049 2050 struct llvm::ClassInfo { 2051 struct MemberInfo { 2052 const DIDerivedType *MemberTypeNode; 2053 uint64_t BaseOffset; 2054 }; 2055 // [MemberInfo] 2056 using MemberList = std::vector<MemberInfo>; 2057 2058 using MethodsList = TinyPtrVector<const DISubprogram *>; 2059 // MethodName -> MethodsList 2060 using MethodsMap = MapVector<MDString *, MethodsList>; 2061 2062 /// Base classes. 2063 std::vector<const DIDerivedType *> Inheritance; 2064 2065 /// Direct members. 2066 MemberList Members; 2067 // Direct overloaded methods gathered by name. 2068 MethodsMap Methods; 2069 2070 TypeIndex VShapeTI; 2071 2072 std::vector<const DIType *> NestedTypes; 2073 }; 2074 2075 void CodeViewDebug::clear() { 2076 assert(CurFn == nullptr); 2077 FileIdMap.clear(); 2078 FnDebugInfo.clear(); 2079 FileToFilepathMap.clear(); 2080 LocalUDTs.clear(); 2081 GlobalUDTs.clear(); 2082 TypeIndices.clear(); 2083 CompleteTypeIndices.clear(); 2084 ScopeGlobals.clear(); 2085 } 2086 2087 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2088 const DIDerivedType *DDTy) { 2089 if (!DDTy->getName().empty()) { 2090 Info.Members.push_back({DDTy, 0}); 2091 return; 2092 } 2093 2094 // An unnamed member may represent a nested struct or union. Attempt to 2095 // interpret the unnamed member as a DICompositeType possibly wrapped in 2096 // qualifier types. Add all the indirect fields to the current record if that 2097 // succeeds, and drop the member if that fails. 2098 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2099 uint64_t Offset = DDTy->getOffsetInBits(); 2100 const DIType *Ty = DDTy->getBaseType(); 2101 bool FullyResolved = false; 2102 while (!FullyResolved) { 2103 switch (Ty->getTag()) { 2104 case dwarf::DW_TAG_const_type: 2105 case dwarf::DW_TAG_volatile_type: 2106 // FIXME: we should apply the qualifier types to the indirect fields 2107 // rather than dropping them. 2108 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2109 break; 2110 default: 2111 FullyResolved = true; 2112 break; 2113 } 2114 } 2115 2116 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2117 if (!DCTy) 2118 return; 2119 2120 ClassInfo NestedInfo = collectClassInfo(DCTy); 2121 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2122 Info.Members.push_back( 2123 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2124 } 2125 2126 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2127 ClassInfo Info; 2128 // Add elements to structure type. 2129 DINodeArray Elements = Ty->getElements(); 2130 for (auto *Element : Elements) { 2131 // We assume that the frontend provides all members in source declaration 2132 // order, which is what MSVC does. 2133 if (!Element) 2134 continue; 2135 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2136 Info.Methods[SP->getRawName()].push_back(SP); 2137 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2138 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2139 collectMemberInfo(Info, DDTy); 2140 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2141 Info.Inheritance.push_back(DDTy); 2142 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2143 DDTy->getName() == "__vtbl_ptr_type") { 2144 Info.VShapeTI = getTypeIndex(DDTy); 2145 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2146 Info.NestedTypes.push_back(DDTy); 2147 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2148 // Ignore friend members. It appears that MSVC emitted info about 2149 // friends in the past, but modern versions do not. 2150 } 2151 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2152 Info.NestedTypes.push_back(Composite); 2153 } 2154 // Skip other unrecognized kinds of elements. 2155 } 2156 return Info; 2157 } 2158 2159 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2160 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2161 // if a complete type should be emitted instead of a forward reference. 2162 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2163 !Ty->isForwardDecl(); 2164 } 2165 2166 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2167 // Emit the complete type for unnamed structs. C++ classes with methods 2168 // which have a circular reference back to the class type are expected to 2169 // be named by the front-end and should not be "unnamed". C unnamed 2170 // structs should not have circular references. 2171 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2172 // If this unnamed complete type is already in the process of being defined 2173 // then the description of the type is malformed and cannot be emitted 2174 // into CodeView correctly so report a fatal error. 2175 auto I = CompleteTypeIndices.find(Ty); 2176 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2177 report_fatal_error("cannot debug circular reference to unnamed type"); 2178 return getCompleteTypeIndex(Ty); 2179 } 2180 2181 // First, construct the forward decl. Don't look into Ty to compute the 2182 // forward decl options, since it might not be available in all TUs. 2183 TypeRecordKind Kind = getRecordKind(Ty); 2184 ClassOptions CO = 2185 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2186 std::string FullName = getFullyQualifiedName(Ty); 2187 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2188 FullName, Ty->getIdentifier()); 2189 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2190 if (!Ty->isForwardDecl()) 2191 DeferredCompleteTypes.push_back(Ty); 2192 return FwdDeclTI; 2193 } 2194 2195 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2196 // Construct the field list and complete type record. 2197 TypeRecordKind Kind = getRecordKind(Ty); 2198 ClassOptions CO = getCommonClassOptions(Ty); 2199 TypeIndex FieldTI; 2200 TypeIndex VShapeTI; 2201 unsigned FieldCount; 2202 bool ContainsNestedClass; 2203 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2204 lowerRecordFieldList(Ty); 2205 2206 if (ContainsNestedClass) 2207 CO |= ClassOptions::ContainsNestedClass; 2208 2209 // MSVC appears to set this flag by searching any destructor or method with 2210 // FunctionOptions::Constructor among the emitted members. Clang AST has all 2211 // the members, however special member functions are not yet emitted into 2212 // debug information. For now checking a class's non-triviality seems enough. 2213 // FIXME: not true for a nested unnamed struct. 2214 if (isNonTrivial(Ty)) 2215 CO |= ClassOptions::HasConstructorOrDestructor; 2216 2217 std::string FullName = getFullyQualifiedName(Ty); 2218 2219 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2220 2221 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2222 SizeInBytes, FullName, Ty->getIdentifier()); 2223 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2224 2225 addUDTSrcLine(Ty, ClassTI); 2226 2227 addToUDTs(Ty); 2228 2229 return ClassTI; 2230 } 2231 2232 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2233 // Emit the complete type for unnamed unions. 2234 if (shouldAlwaysEmitCompleteClassType(Ty)) 2235 return getCompleteTypeIndex(Ty); 2236 2237 ClassOptions CO = 2238 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2239 std::string FullName = getFullyQualifiedName(Ty); 2240 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2241 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2242 if (!Ty->isForwardDecl()) 2243 DeferredCompleteTypes.push_back(Ty); 2244 return FwdDeclTI; 2245 } 2246 2247 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2248 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2249 TypeIndex FieldTI; 2250 unsigned FieldCount; 2251 bool ContainsNestedClass; 2252 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2253 lowerRecordFieldList(Ty); 2254 2255 if (ContainsNestedClass) 2256 CO |= ClassOptions::ContainsNestedClass; 2257 2258 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2259 std::string FullName = getFullyQualifiedName(Ty); 2260 2261 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2262 Ty->getIdentifier()); 2263 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2264 2265 addUDTSrcLine(Ty, UnionTI); 2266 2267 addToUDTs(Ty); 2268 2269 return UnionTI; 2270 } 2271 2272 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2273 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2274 // Manually count members. MSVC appears to count everything that generates a 2275 // field list record. Each individual overload in a method overload group 2276 // contributes to this count, even though the overload group is a single field 2277 // list record. 2278 unsigned MemberCount = 0; 2279 ClassInfo Info = collectClassInfo(Ty); 2280 ContinuationRecordBuilder ContinuationBuilder; 2281 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2282 2283 // Create base classes. 2284 for (const DIDerivedType *I : Info.Inheritance) { 2285 if (I->getFlags() & DINode::FlagVirtual) { 2286 // Virtual base. 2287 unsigned VBPtrOffset = I->getVBPtrOffset(); 2288 // FIXME: Despite the accessor name, the offset is really in bytes. 2289 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2290 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2291 ? TypeRecordKind::IndirectVirtualBaseClass 2292 : TypeRecordKind::VirtualBaseClass; 2293 VirtualBaseClassRecord VBCR( 2294 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2295 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2296 VBTableIndex); 2297 2298 ContinuationBuilder.writeMemberType(VBCR); 2299 MemberCount++; 2300 } else { 2301 assert(I->getOffsetInBits() % 8 == 0 && 2302 "bases must be on byte boundaries"); 2303 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2304 getTypeIndex(I->getBaseType()), 2305 I->getOffsetInBits() / 8); 2306 ContinuationBuilder.writeMemberType(BCR); 2307 MemberCount++; 2308 } 2309 } 2310 2311 // Create members. 2312 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2313 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2314 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2315 StringRef MemberName = Member->getName(); 2316 MemberAccess Access = 2317 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2318 2319 if (Member->isStaticMember()) { 2320 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2321 ContinuationBuilder.writeMemberType(SDMR); 2322 MemberCount++; 2323 continue; 2324 } 2325 2326 // Virtual function pointer member. 2327 if ((Member->getFlags() & DINode::FlagArtificial) && 2328 Member->getName().startswith("_vptr$")) { 2329 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2330 ContinuationBuilder.writeMemberType(VFPR); 2331 MemberCount++; 2332 continue; 2333 } 2334 2335 // Data member. 2336 uint64_t MemberOffsetInBits = 2337 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2338 if (Member->isBitField()) { 2339 uint64_t StartBitOffset = MemberOffsetInBits; 2340 if (const auto *CI = 2341 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2342 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2343 } 2344 StartBitOffset -= MemberOffsetInBits; 2345 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2346 StartBitOffset); 2347 MemberBaseType = TypeTable.writeLeafType(BFR); 2348 } 2349 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2350 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2351 MemberName); 2352 ContinuationBuilder.writeMemberType(DMR); 2353 MemberCount++; 2354 } 2355 2356 // Create methods 2357 for (auto &MethodItr : Info.Methods) { 2358 StringRef Name = MethodItr.first->getString(); 2359 2360 std::vector<OneMethodRecord> Methods; 2361 for (const DISubprogram *SP : MethodItr.second) { 2362 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2363 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2364 2365 unsigned VFTableOffset = -1; 2366 if (Introduced) 2367 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2368 2369 Methods.push_back(OneMethodRecord( 2370 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2371 translateMethodKindFlags(SP, Introduced), 2372 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2373 MemberCount++; 2374 } 2375 assert(!Methods.empty() && "Empty methods map entry"); 2376 if (Methods.size() == 1) 2377 ContinuationBuilder.writeMemberType(Methods[0]); 2378 else { 2379 // FIXME: Make this use its own ContinuationBuilder so that 2380 // MethodOverloadList can be split correctly. 2381 MethodOverloadListRecord MOLR(Methods); 2382 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2383 2384 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2385 ContinuationBuilder.writeMemberType(OMR); 2386 } 2387 } 2388 2389 // Create nested classes. 2390 for (const DIType *Nested : Info.NestedTypes) { 2391 NestedTypeRecord R(getTypeIndex(Nested), Nested->getName()); 2392 ContinuationBuilder.writeMemberType(R); 2393 MemberCount++; 2394 } 2395 2396 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2397 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2398 !Info.NestedTypes.empty()); 2399 } 2400 2401 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2402 if (!VBPType.getIndex()) { 2403 // Make a 'const int *' type. 2404 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2405 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2406 2407 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2408 : PointerKind::Near32; 2409 PointerMode PM = PointerMode::Pointer; 2410 PointerOptions PO = PointerOptions::None; 2411 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2412 VBPType = TypeTable.writeLeafType(PR); 2413 } 2414 2415 return VBPType; 2416 } 2417 2418 TypeIndex CodeViewDebug::getTypeIndex(const DIType *Ty, const DIType *ClassTy) { 2419 // The null DIType is the void type. Don't try to hash it. 2420 if (!Ty) 2421 return TypeIndex::Void(); 2422 2423 // Check if we've already translated this type. Don't try to do a 2424 // get-or-create style insertion that caches the hash lookup across the 2425 // lowerType call. It will update the TypeIndices map. 2426 auto I = TypeIndices.find({Ty, ClassTy}); 2427 if (I != TypeIndices.end()) 2428 return I->second; 2429 2430 TypeLoweringScope S(*this); 2431 TypeIndex TI = lowerType(Ty, ClassTy); 2432 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2433 } 2434 2435 codeview::TypeIndex 2436 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, 2437 const DISubroutineType *SubroutineTy) { 2438 assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && 2439 "this type must be a pointer type"); 2440 2441 PointerOptions Options = PointerOptions::None; 2442 if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) 2443 Options = PointerOptions::LValueRefThisPointer; 2444 else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) 2445 Options = PointerOptions::RValueRefThisPointer; 2446 2447 // Check if we've already translated this type. If there is no ref qualifier 2448 // on the function then we look up this pointer type with no associated class 2449 // so that the TypeIndex for the this pointer can be shared with the type 2450 // index for other pointers to this class type. If there is a ref qualifier 2451 // then we lookup the pointer using the subroutine as the parent type. 2452 auto I = TypeIndices.find({PtrTy, SubroutineTy}); 2453 if (I != TypeIndices.end()) 2454 return I->second; 2455 2456 TypeLoweringScope S(*this); 2457 TypeIndex TI = lowerTypePointer(PtrTy, Options); 2458 return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy); 2459 } 2460 2461 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(const DIType *Ty) { 2462 PointerRecord PR(getTypeIndex(Ty), 2463 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2464 : PointerKind::Near32, 2465 PointerMode::LValueReference, PointerOptions::None, 2466 Ty->getSizeInBits() / 8); 2467 return TypeTable.writeLeafType(PR); 2468 } 2469 2470 TypeIndex CodeViewDebug::getCompleteTypeIndex(const DIType *Ty) { 2471 // The null DIType is the void type. Don't try to hash it. 2472 if (!Ty) 2473 return TypeIndex::Void(); 2474 2475 // Look through typedefs when getting the complete type index. Call 2476 // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are 2477 // emitted only once. 2478 if (Ty->getTag() == dwarf::DW_TAG_typedef) 2479 (void)getTypeIndex(Ty); 2480 while (Ty->getTag() == dwarf::DW_TAG_typedef) 2481 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2482 2483 // If this is a non-record type, the complete type index is the same as the 2484 // normal type index. Just call getTypeIndex. 2485 switch (Ty->getTag()) { 2486 case dwarf::DW_TAG_class_type: 2487 case dwarf::DW_TAG_structure_type: 2488 case dwarf::DW_TAG_union_type: 2489 break; 2490 default: 2491 return getTypeIndex(Ty); 2492 } 2493 2494 const auto *CTy = cast<DICompositeType>(Ty); 2495 2496 TypeLoweringScope S(*this); 2497 2498 // Make sure the forward declaration is emitted first. It's unclear if this 2499 // is necessary, but MSVC does it, and we should follow suit until we can show 2500 // otherwise. 2501 // We only emit a forward declaration for named types. 2502 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2503 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2504 2505 // Just use the forward decl if we don't have complete type info. This 2506 // might happen if the frontend is using modules and expects the complete 2507 // definition to be emitted elsewhere. 2508 if (CTy->isForwardDecl()) 2509 return FwdDeclTI; 2510 } 2511 2512 // Check if we've already translated the complete record type. 2513 // Insert the type with a null TypeIndex to signify that the type is currently 2514 // being lowered. 2515 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2516 if (!InsertResult.second) 2517 return InsertResult.first->second; 2518 2519 TypeIndex TI; 2520 switch (CTy->getTag()) { 2521 case dwarf::DW_TAG_class_type: 2522 case dwarf::DW_TAG_structure_type: 2523 TI = lowerCompleteTypeClass(CTy); 2524 break; 2525 case dwarf::DW_TAG_union_type: 2526 TI = lowerCompleteTypeUnion(CTy); 2527 break; 2528 default: 2529 llvm_unreachable("not a record"); 2530 } 2531 2532 // Update the type index associated with this CompositeType. This cannot 2533 // use the 'InsertResult' iterator above because it is potentially 2534 // invalidated by map insertions which can occur while lowering the class 2535 // type above. 2536 CompleteTypeIndices[CTy] = TI; 2537 return TI; 2538 } 2539 2540 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2541 /// and do this until fixpoint, as each complete record type typically 2542 /// references 2543 /// many other record types. 2544 void CodeViewDebug::emitDeferredCompleteTypes() { 2545 SmallVector<const DICompositeType *, 4> TypesToEmit; 2546 while (!DeferredCompleteTypes.empty()) { 2547 std::swap(DeferredCompleteTypes, TypesToEmit); 2548 for (const DICompositeType *RecordTy : TypesToEmit) 2549 getCompleteTypeIndex(RecordTy); 2550 TypesToEmit.clear(); 2551 } 2552 } 2553 2554 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2555 ArrayRef<LocalVariable> Locals) { 2556 // Get the sorted list of parameters and emit them first. 2557 SmallVector<const LocalVariable *, 6> Params; 2558 for (const LocalVariable &L : Locals) 2559 if (L.DIVar->isParameter()) 2560 Params.push_back(&L); 2561 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2562 return L->DIVar->getArg() < R->DIVar->getArg(); 2563 }); 2564 for (const LocalVariable *L : Params) 2565 emitLocalVariable(FI, *L); 2566 2567 // Next emit all non-parameters in the order that we found them. 2568 for (const LocalVariable &L : Locals) 2569 if (!L.DIVar->isParameter()) 2570 emitLocalVariable(FI, L); 2571 } 2572 2573 /// Only call this on endian-specific types like ulittle16_t and little32_t, or 2574 /// structs composed of them. 2575 template <typename T> 2576 static void copyBytesForDefRange(SmallString<20> &BytePrefix, 2577 SymbolKind SymKind, const T &DefRangeHeader) { 2578 BytePrefix.resize(2 + sizeof(T)); 2579 ulittle16_t SymKindLE = ulittle16_t(SymKind); 2580 memcpy(&BytePrefix[0], &SymKindLE, 2); 2581 memcpy(&BytePrefix[2], &DefRangeHeader, sizeof(T)); 2582 } 2583 2584 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2585 const LocalVariable &Var) { 2586 // LocalSym record, see SymbolRecord.h for more info. 2587 MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL); 2588 2589 LocalSymFlags Flags = LocalSymFlags::None; 2590 if (Var.DIVar->isParameter()) 2591 Flags |= LocalSymFlags::IsParameter; 2592 if (Var.DefRanges.empty()) 2593 Flags |= LocalSymFlags::IsOptimizedOut; 2594 2595 OS.AddComment("TypeIndex"); 2596 TypeIndex TI = Var.UseReferenceType 2597 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2598 : getCompleteTypeIndex(Var.DIVar->getType()); 2599 OS.EmitIntValue(TI.getIndex(), 4); 2600 OS.AddComment("Flags"); 2601 OS.EmitIntValue(static_cast<uint16_t>(Flags), 2); 2602 // Truncate the name so we won't overflow the record length field. 2603 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2604 endSymbolRecord(LocalEnd); 2605 2606 // Calculate the on disk prefix of the appropriate def range record. The 2607 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2608 // should be big enough to hold all forms without memory allocation. 2609 SmallString<20> BytePrefix; 2610 for (const LocalVarDefRange &DefRange : Var.DefRanges) { 2611 BytePrefix.clear(); 2612 if (DefRange.InMemory) { 2613 int Offset = DefRange.DataOffset; 2614 unsigned Reg = DefRange.CVRegister; 2615 2616 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2617 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2618 // instead. In frames without stack realignment, $T0 will be the CFA. 2619 if (RegisterId(Reg) == RegisterId::ESP) { 2620 Reg = unsigned(RegisterId::VFRAME); 2621 Offset += FI.OffsetAdjustment; 2622 } 2623 2624 // If we can use the chosen frame pointer for the frame and this isn't a 2625 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2626 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2627 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2628 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2629 (bool(Flags & LocalSymFlags::IsParameter) 2630 ? (EncFP == FI.EncodedParamFramePtrReg) 2631 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2632 little32_t FPOffset = little32_t(Offset); 2633 copyBytesForDefRange(BytePrefix, S_DEFRANGE_FRAMEPOINTER_REL, FPOffset); 2634 } else { 2635 uint16_t RegRelFlags = 0; 2636 if (DefRange.IsSubfield) { 2637 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2638 (DefRange.StructOffset 2639 << DefRangeRegisterRelSym::OffsetInParentShift); 2640 } 2641 DefRangeRegisterRelSym::Header DRHdr; 2642 DRHdr.Register = Reg; 2643 DRHdr.Flags = RegRelFlags; 2644 DRHdr.BasePointerOffset = Offset; 2645 copyBytesForDefRange(BytePrefix, S_DEFRANGE_REGISTER_REL, DRHdr); 2646 } 2647 } else { 2648 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2649 if (DefRange.IsSubfield) { 2650 DefRangeSubfieldRegisterSym::Header DRHdr; 2651 DRHdr.Register = DefRange.CVRegister; 2652 DRHdr.MayHaveNoName = 0; 2653 DRHdr.OffsetInParent = DefRange.StructOffset; 2654 copyBytesForDefRange(BytePrefix, S_DEFRANGE_SUBFIELD_REGISTER, DRHdr); 2655 } else { 2656 DefRangeRegisterSym::Header DRHdr; 2657 DRHdr.Register = DefRange.CVRegister; 2658 DRHdr.MayHaveNoName = 0; 2659 copyBytesForDefRange(BytePrefix, S_DEFRANGE_REGISTER, DRHdr); 2660 } 2661 } 2662 OS.EmitCVDefRangeDirective(DefRange.Ranges, BytePrefix); 2663 } 2664 } 2665 2666 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2667 const FunctionInfo& FI) { 2668 for (LexicalBlock *Block : Blocks) 2669 emitLexicalBlock(*Block, FI); 2670 } 2671 2672 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2673 /// lexical block scope. 2674 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2675 const FunctionInfo& FI) { 2676 MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32); 2677 OS.AddComment("PtrParent"); 2678 OS.EmitIntValue(0, 4); // PtrParent 2679 OS.AddComment("PtrEnd"); 2680 OS.EmitIntValue(0, 4); // PtrEnd 2681 OS.AddComment("Code size"); 2682 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2683 OS.AddComment("Function section relative address"); 2684 OS.EmitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2685 OS.AddComment("Function section index"); 2686 OS.EmitCOFFSectionIndex(FI.Begin); // Func Symbol 2687 OS.AddComment("Lexical block name"); 2688 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2689 endSymbolRecord(RecordEnd); 2690 2691 // Emit variables local to this lexical block. 2692 emitLocalVariableList(FI, Block.Locals); 2693 emitGlobalVariableList(Block.Globals); 2694 2695 // Emit lexical blocks contained within this block. 2696 emitLexicalBlockList(Block.Children, FI); 2697 2698 // Close the lexical block scope. 2699 emitEndSymbolRecord(SymbolKind::S_END); 2700 } 2701 2702 /// Convenience routine for collecting lexical block information for a list 2703 /// of lexical scopes. 2704 void CodeViewDebug::collectLexicalBlockInfo( 2705 SmallVectorImpl<LexicalScope *> &Scopes, 2706 SmallVectorImpl<LexicalBlock *> &Blocks, 2707 SmallVectorImpl<LocalVariable> &Locals, 2708 SmallVectorImpl<CVGlobalVariable> &Globals) { 2709 for (LexicalScope *Scope : Scopes) 2710 collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals); 2711 } 2712 2713 /// Populate the lexical blocks and local variable lists of the parent with 2714 /// information about the specified lexical scope. 2715 void CodeViewDebug::collectLexicalBlockInfo( 2716 LexicalScope &Scope, 2717 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2718 SmallVectorImpl<LocalVariable> &ParentLocals, 2719 SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { 2720 if (Scope.isAbstractScope()) 2721 return; 2722 2723 // Gather information about the lexical scope including local variables, 2724 // global variables, and address ranges. 2725 bool IgnoreScope = false; 2726 auto LI = ScopeVariables.find(&Scope); 2727 SmallVectorImpl<LocalVariable> *Locals = 2728 LI != ScopeVariables.end() ? &LI->second : nullptr; 2729 auto GI = ScopeGlobals.find(Scope.getScopeNode()); 2730 SmallVectorImpl<CVGlobalVariable> *Globals = 2731 GI != ScopeGlobals.end() ? GI->second.get() : nullptr; 2732 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2733 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2734 2735 // Ignore lexical scopes which do not contain variables. 2736 if (!Locals && !Globals) 2737 IgnoreScope = true; 2738 2739 // Ignore lexical scopes which are not lexical blocks. 2740 if (!DILB) 2741 IgnoreScope = true; 2742 2743 // Ignore scopes which have too many address ranges to represent in the 2744 // current CodeView format or do not have a valid address range. 2745 // 2746 // For lexical scopes with multiple address ranges you may be tempted to 2747 // construct a single range covering every instruction where the block is 2748 // live and everything in between. Unfortunately, Visual Studio only 2749 // displays variables from the first matching lexical block scope. If the 2750 // first lexical block contains exception handling code or cold code which 2751 // is moved to the bottom of the routine creating a single range covering 2752 // nearly the entire routine, then it will hide all other lexical blocks 2753 // and the variables they contain. 2754 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) 2755 IgnoreScope = true; 2756 2757 if (IgnoreScope) { 2758 // This scope can be safely ignored and eliminating it will reduce the 2759 // size of the debug information. Be sure to collect any variable and scope 2760 // information from the this scope or any of its children and collapse them 2761 // into the parent scope. 2762 if (Locals) 2763 ParentLocals.append(Locals->begin(), Locals->end()); 2764 if (Globals) 2765 ParentGlobals.append(Globals->begin(), Globals->end()); 2766 collectLexicalBlockInfo(Scope.getChildren(), 2767 ParentBlocks, 2768 ParentLocals, 2769 ParentGlobals); 2770 return; 2771 } 2772 2773 // Create a new CodeView lexical block for this lexical scope. If we've 2774 // seen this DILexicalBlock before then the scope tree is malformed and 2775 // we can handle this gracefully by not processing it a second time. 2776 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 2777 if (!BlockInsertion.second) 2778 return; 2779 2780 // Create a lexical block containing the variables and collect the the 2781 // lexical block information for the children. 2782 const InsnRange &Range = Ranges.front(); 2783 assert(Range.first && Range.second); 2784 LexicalBlock &Block = BlockInsertion.first->second; 2785 Block.Begin = getLabelBeforeInsn(Range.first); 2786 Block.End = getLabelAfterInsn(Range.second); 2787 assert(Block.Begin && "missing label for scope begin"); 2788 assert(Block.End && "missing label for scope end"); 2789 Block.Name = DILB->getName(); 2790 if (Locals) 2791 Block.Locals = std::move(*Locals); 2792 if (Globals) 2793 Block.Globals = std::move(*Globals); 2794 ParentBlocks.push_back(&Block); 2795 collectLexicalBlockInfo(Scope.getChildren(), 2796 Block.Children, 2797 Block.Locals, 2798 Block.Globals); 2799 } 2800 2801 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 2802 const Function &GV = MF->getFunction(); 2803 assert(FnDebugInfo.count(&GV)); 2804 assert(CurFn == FnDebugInfo[&GV].get()); 2805 2806 collectVariableInfo(GV.getSubprogram()); 2807 2808 // Build the lexical block structure to emit for this routine. 2809 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 2810 collectLexicalBlockInfo(*CFS, 2811 CurFn->ChildBlocks, 2812 CurFn->Locals, 2813 CurFn->Globals); 2814 2815 // Clear the scope and variable information from the map which will not be 2816 // valid after we have finished processing this routine. This also prepares 2817 // the map for the subsequent routine. 2818 ScopeVariables.clear(); 2819 2820 // Don't emit anything if we don't have any line tables. 2821 // Thunks are compiler-generated and probably won't have source correlation. 2822 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 2823 FnDebugInfo.erase(&GV); 2824 CurFn = nullptr; 2825 return; 2826 } 2827 2828 CurFn->Annotations = MF->getCodeViewAnnotations(); 2829 CurFn->HeapAllocSites = MF->getCodeViewHeapAllocSites(); 2830 2831 CurFn->End = Asm->getFunctionEnd(); 2832 2833 CurFn = nullptr; 2834 } 2835 2836 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 2837 DebugHandlerBase::beginInstruction(MI); 2838 2839 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 2840 if (!Asm || !CurFn || MI->isDebugInstr() || 2841 MI->getFlag(MachineInstr::FrameSetup)) 2842 return; 2843 2844 // If the first instruction of a new MBB has no location, find the first 2845 // instruction with a location and use that. 2846 DebugLoc DL = MI->getDebugLoc(); 2847 if (!DL && MI->getParent() != PrevInstBB) { 2848 for (const auto &NextMI : *MI->getParent()) { 2849 if (NextMI.isDebugInstr()) 2850 continue; 2851 DL = NextMI.getDebugLoc(); 2852 if (DL) 2853 break; 2854 } 2855 } 2856 PrevInstBB = MI->getParent(); 2857 2858 // If we still don't have a debug location, don't record a location. 2859 if (!DL) 2860 return; 2861 2862 maybeRecordLocation(DL, Asm->MF); 2863 } 2864 2865 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 2866 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 2867 *EndLabel = MMI->getContext().createTempSymbol(); 2868 OS.EmitIntValue(unsigned(Kind), 4); 2869 OS.AddComment("Subsection size"); 2870 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 2871 OS.EmitLabel(BeginLabel); 2872 return EndLabel; 2873 } 2874 2875 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 2876 OS.EmitLabel(EndLabel); 2877 // Every subsection must be aligned to a 4-byte boundary. 2878 OS.EmitValueToAlignment(4); 2879 } 2880 2881 static StringRef getSymbolName(SymbolKind SymKind) { 2882 for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) 2883 if (EE.Value == SymKind) 2884 return EE.Name; 2885 return ""; 2886 } 2887 2888 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { 2889 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 2890 *EndLabel = MMI->getContext().createTempSymbol(); 2891 OS.AddComment("Record length"); 2892 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 2893 OS.EmitLabel(BeginLabel); 2894 if (OS.isVerboseAsm()) 2895 OS.AddComment("Record kind: " + getSymbolName(SymKind)); 2896 OS.EmitIntValue(unsigned(SymKind), 2); 2897 return EndLabel; 2898 } 2899 2900 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { 2901 // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid 2902 // an extra copy of every symbol record in LLD. This increases object file 2903 // size by less than 1% in the clang build, and is compatible with the Visual 2904 // C++ linker. 2905 OS.EmitValueToAlignment(4); 2906 OS.EmitLabel(SymEnd); 2907 } 2908 2909 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { 2910 OS.AddComment("Record length"); 2911 OS.EmitIntValue(2, 2); 2912 if (OS.isVerboseAsm()) 2913 OS.AddComment("Record kind: " + getSymbolName(EndKind)); 2914 OS.EmitIntValue(unsigned(EndKind), 2); // Record Kind 2915 } 2916 2917 void CodeViewDebug::emitDebugInfoForUDTs( 2918 ArrayRef<std::pair<std::string, const DIType *>> UDTs) { 2919 for (const auto &UDT : UDTs) { 2920 const DIType *T = UDT.second; 2921 assert(shouldEmitUdt(T)); 2922 2923 MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT); 2924 OS.AddComment("Type"); 2925 OS.EmitIntValue(getCompleteTypeIndex(T).getIndex(), 4); 2926 emitNullTerminatedSymbolName(OS, UDT.first); 2927 endSymbolRecord(UDTRecordEnd); 2928 } 2929 } 2930 2931 void CodeViewDebug::collectGlobalVariableInfo() { 2932 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 2933 GlobalMap; 2934 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 2935 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 2936 GV.getDebugInfo(GVEs); 2937 for (const auto *GVE : GVEs) 2938 GlobalMap[GVE] = &GV; 2939 } 2940 2941 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 2942 for (const MDNode *Node : CUs->operands()) { 2943 const auto *CU = cast<DICompileUnit>(Node); 2944 for (const auto *GVE : CU->getGlobalVariables()) { 2945 const DIGlobalVariable *DIGV = GVE->getVariable(); 2946 const DIExpression *DIE = GVE->getExpression(); 2947 2948 // Emit constant global variables in a global symbol section. 2949 if (GlobalMap.count(GVE) == 0 && DIE->isConstant()) { 2950 CVGlobalVariable CVGV = {DIGV, DIE}; 2951 GlobalVariables.emplace_back(std::move(CVGV)); 2952 } 2953 2954 const auto *GV = GlobalMap.lookup(GVE); 2955 if (!GV || GV->isDeclarationForLinker()) 2956 continue; 2957 2958 DIScope *Scope = DIGV->getScope(); 2959 SmallVector<CVGlobalVariable, 1> *VariableList; 2960 if (Scope && isa<DILocalScope>(Scope)) { 2961 // Locate a global variable list for this scope, creating one if 2962 // necessary. 2963 auto Insertion = ScopeGlobals.insert( 2964 {Scope, std::unique_ptr<GlobalVariableList>()}); 2965 if (Insertion.second) 2966 Insertion.first->second = llvm::make_unique<GlobalVariableList>(); 2967 VariableList = Insertion.first->second.get(); 2968 } else if (GV->hasComdat()) 2969 // Emit this global variable into a COMDAT section. 2970 VariableList = &ComdatVariables; 2971 else 2972 // Emit this global variable in a single global symbol section. 2973 VariableList = &GlobalVariables; 2974 CVGlobalVariable CVGV = {DIGV, GV}; 2975 VariableList->emplace_back(std::move(CVGV)); 2976 } 2977 } 2978 } 2979 2980 void CodeViewDebug::emitDebugInfoForGlobals() { 2981 // First, emit all globals that are not in a comdat in a single symbol 2982 // substream. MSVC doesn't like it if the substream is empty, so only open 2983 // it if we have at least one global to emit. 2984 switchToDebugSectionForSymbol(nullptr); 2985 if (!GlobalVariables.empty()) { 2986 OS.AddComment("Symbol subsection for globals"); 2987 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 2988 emitGlobalVariableList(GlobalVariables); 2989 endCVSubsection(EndLabel); 2990 } 2991 2992 // Second, emit each global that is in a comdat into its own .debug$S 2993 // section along with its own symbol substream. 2994 for (const CVGlobalVariable &CVGV : ComdatVariables) { 2995 const GlobalVariable *GV = CVGV.GVInfo.get<const GlobalVariable *>(); 2996 MCSymbol *GVSym = Asm->getSymbol(GV); 2997 OS.AddComment("Symbol subsection for " + 2998 Twine(GlobalValue::dropLLVMManglingEscape(GV->getName()))); 2999 switchToDebugSectionForSymbol(GVSym); 3000 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3001 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3002 emitDebugInfoForGlobal(CVGV); 3003 endCVSubsection(EndLabel); 3004 } 3005 } 3006 3007 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 3008 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3009 for (const MDNode *Node : CUs->operands()) { 3010 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 3011 if (DIType *RT = dyn_cast<DIType>(Ty)) { 3012 getTypeIndex(RT); 3013 // FIXME: Add to global/local DTU list. 3014 } 3015 } 3016 } 3017 } 3018 3019 // Emit each global variable in the specified array. 3020 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { 3021 for (const CVGlobalVariable &CVGV : Globals) { 3022 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3023 emitDebugInfoForGlobal(CVGV); 3024 } 3025 } 3026 3027 void CodeViewDebug::emitDebugInfoForGlobal(const CVGlobalVariable &CVGV) { 3028 const DIGlobalVariable *DIGV = CVGV.DIGV; 3029 if (const GlobalVariable *GV = 3030 CVGV.GVInfo.dyn_cast<const GlobalVariable *>()) { 3031 // DataSym record, see SymbolRecord.h for more info. Thread local data 3032 // happens to have the same format as global data. 3033 MCSymbol *GVSym = Asm->getSymbol(GV); 3034 SymbolKind DataSym = GV->isThreadLocal() 3035 ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 3036 : SymbolKind::S_GTHREAD32) 3037 : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 3038 : SymbolKind::S_GDATA32); 3039 MCSymbol *DataEnd = beginSymbolRecord(DataSym); 3040 OS.AddComment("Type"); 3041 OS.EmitIntValue(getCompleteTypeIndex(DIGV->getType()).getIndex(), 4); 3042 OS.AddComment("DataOffset"); 3043 OS.EmitCOFFSecRel32(GVSym, /*Offset=*/0); 3044 OS.AddComment("Segment"); 3045 OS.EmitCOFFSectionIndex(GVSym); 3046 OS.AddComment("Name"); 3047 const unsigned LengthOfDataRecord = 12; 3048 emitNullTerminatedSymbolName(OS, DIGV->getName(), LengthOfDataRecord); 3049 endSymbolRecord(DataEnd); 3050 } else { 3051 // FIXME: Currently this only emits the global variables in the IR metadata. 3052 // This should also emit enums and static data members. 3053 const DIExpression *DIE = CVGV.GVInfo.get<const DIExpression *>(); 3054 assert(DIE->isConstant() && 3055 "Global constant variables must contain a constant expression."); 3056 uint64_t Val = DIE->getElement(1); 3057 3058 MCSymbol *SConstantEnd = beginSymbolRecord(SymbolKind::S_CONSTANT); 3059 OS.AddComment("Type"); 3060 OS.EmitIntValue(getTypeIndex(DIGV->getType()).getIndex(), 4); 3061 OS.AddComment("Value"); 3062 3063 // Encoded integers shouldn't need more than 10 bytes. 3064 uint8_t data[10]; 3065 BinaryStreamWriter Writer(data, llvm::support::endianness::little); 3066 CodeViewRecordIO IO(Writer); 3067 cantFail(IO.mapEncodedInteger(Val)); 3068 StringRef SRef((char *)data, Writer.getOffset()); 3069 OS.EmitBinaryData(SRef); 3070 3071 OS.AddComment("Name"); 3072 emitNullTerminatedSymbolName(OS, DIGV->getDisplayName()); 3073 endSymbolRecord(SConstantEnd); 3074 } 3075 } 3076