1 //===- llvm/CodeGen/DwarfDebug.cpp - Dwarf Debug Framework ----------------===// 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 dwarf debug info into asm files. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "DwarfDebug.h" 14 #include "ByteStreamer.h" 15 #include "DIEHash.h" 16 #include "DwarfCompileUnit.h" 17 #include "DwarfExpression.h" 18 #include "DwarfUnit.h" 19 #include "llvm/ADT/APInt.h" 20 #include "llvm/ADT/Statistic.h" 21 #include "llvm/ADT/Triple.h" 22 #include "llvm/ADT/Twine.h" 23 #include "llvm/CodeGen/AsmPrinter.h" 24 #include "llvm/CodeGen/DIE.h" 25 #include "llvm/CodeGen/LexicalScopes.h" 26 #include "llvm/CodeGen/MachineBasicBlock.h" 27 #include "llvm/CodeGen/MachineFunction.h" 28 #include "llvm/CodeGen/MachineModuleInfo.h" 29 #include "llvm/CodeGen/MachineOperand.h" 30 #include "llvm/CodeGen/TargetInstrInfo.h" 31 #include "llvm/CodeGen/TargetLowering.h" 32 #include "llvm/CodeGen/TargetRegisterInfo.h" 33 #include "llvm/CodeGen/TargetSubtargetInfo.h" 34 #include "llvm/DebugInfo/DWARF/DWARFExpression.h" 35 #include "llvm/DebugInfo/DWARF/DWARFDataExtractor.h" 36 #include "llvm/IR/Constants.h" 37 #include "llvm/IR/Function.h" 38 #include "llvm/IR/GlobalVariable.h" 39 #include "llvm/IR/Module.h" 40 #include "llvm/MC/MCAsmInfo.h" 41 #include "llvm/MC/MCContext.h" 42 #include "llvm/MC/MCSection.h" 43 #include "llvm/MC/MCStreamer.h" 44 #include "llvm/MC/MCSymbol.h" 45 #include "llvm/MC/MCTargetOptions.h" 46 #include "llvm/MC/MachineLocation.h" 47 #include "llvm/MC/SectionKind.h" 48 #include "llvm/Pass.h" 49 #include "llvm/Support/Casting.h" 50 #include "llvm/Support/CommandLine.h" 51 #include "llvm/Support/Debug.h" 52 #include "llvm/Support/ErrorHandling.h" 53 #include "llvm/Support/MD5.h" 54 #include "llvm/Support/MathExtras.h" 55 #include "llvm/Support/Timer.h" 56 #include "llvm/Support/raw_ostream.h" 57 #include "llvm/Target/TargetLoweringObjectFile.h" 58 #include "llvm/Target/TargetMachine.h" 59 #include <algorithm> 60 #include <cstddef> 61 #include <iterator> 62 #include <string> 63 64 using namespace llvm; 65 66 #define DEBUG_TYPE "dwarfdebug" 67 68 STATISTIC(NumCSParams, "Number of dbg call site params created"); 69 70 static cl::opt<bool> UseDwarfRangesBaseAddressSpecifier( 71 "use-dwarf-ranges-base-address-specifier", cl::Hidden, 72 cl::desc("Use base address specifiers in debug_ranges"), cl::init(false)); 73 74 static cl::opt<bool> GenerateARangeSection("generate-arange-section", 75 cl::Hidden, 76 cl::desc("Generate dwarf aranges"), 77 cl::init(false)); 78 79 static cl::opt<bool> 80 GenerateDwarfTypeUnits("generate-type-units", cl::Hidden, 81 cl::desc("Generate DWARF4 type units."), 82 cl::init(false)); 83 84 static cl::opt<bool> SplitDwarfCrossCuReferences( 85 "split-dwarf-cross-cu-references", cl::Hidden, 86 cl::desc("Enable cross-cu references in DWO files"), cl::init(false)); 87 88 enum DefaultOnOff { Default, Enable, Disable }; 89 90 static cl::opt<DefaultOnOff> UnknownLocations( 91 "use-unknown-locations", cl::Hidden, 92 cl::desc("Make an absence of debug location information explicit."), 93 cl::values(clEnumVal(Default, "At top of block or after label"), 94 clEnumVal(Enable, "In all cases"), clEnumVal(Disable, "Never")), 95 cl::init(Default)); 96 97 static cl::opt<AccelTableKind> AccelTables( 98 "accel-tables", cl::Hidden, cl::desc("Output dwarf accelerator tables."), 99 cl::values(clEnumValN(AccelTableKind::Default, "Default", 100 "Default for platform"), 101 clEnumValN(AccelTableKind::None, "Disable", "Disabled."), 102 clEnumValN(AccelTableKind::Apple, "Apple", "Apple"), 103 clEnumValN(AccelTableKind::Dwarf, "Dwarf", "DWARF")), 104 cl::init(AccelTableKind::Default)); 105 106 static cl::opt<DefaultOnOff> 107 DwarfInlinedStrings("dwarf-inlined-strings", cl::Hidden, 108 cl::desc("Use inlined strings rather than string section."), 109 cl::values(clEnumVal(Default, "Default for platform"), 110 clEnumVal(Enable, "Enabled"), 111 clEnumVal(Disable, "Disabled")), 112 cl::init(Default)); 113 114 static cl::opt<bool> 115 NoDwarfRangesSection("no-dwarf-ranges-section", cl::Hidden, 116 cl::desc("Disable emission .debug_ranges section."), 117 cl::init(false)); 118 119 static cl::opt<DefaultOnOff> DwarfSectionsAsReferences( 120 "dwarf-sections-as-references", cl::Hidden, 121 cl::desc("Use sections+offset as references rather than labels."), 122 cl::values(clEnumVal(Default, "Default for platform"), 123 clEnumVal(Enable, "Enabled"), clEnumVal(Disable, "Disabled")), 124 cl::init(Default)); 125 126 static cl::opt<bool> 127 UseGNUDebugMacro("use-gnu-debug-macro", cl::Hidden, 128 cl::desc("Emit the GNU .debug_macro format with DWARF <5"), 129 cl::init(false)); 130 131 static cl::opt<DefaultOnOff> DwarfOpConvert( 132 "dwarf-op-convert", cl::Hidden, 133 cl::desc("Enable use of the DWARFv5 DW_OP_convert operator"), 134 cl::values(clEnumVal(Default, "Default for platform"), 135 clEnumVal(Enable, "Enabled"), clEnumVal(Disable, "Disabled")), 136 cl::init(Default)); 137 138 enum LinkageNameOption { 139 DefaultLinkageNames, 140 AllLinkageNames, 141 AbstractLinkageNames 142 }; 143 144 static cl::opt<LinkageNameOption> 145 DwarfLinkageNames("dwarf-linkage-names", cl::Hidden, 146 cl::desc("Which DWARF linkage-name attributes to emit."), 147 cl::values(clEnumValN(DefaultLinkageNames, "Default", 148 "Default for platform"), 149 clEnumValN(AllLinkageNames, "All", "All"), 150 clEnumValN(AbstractLinkageNames, "Abstract", 151 "Abstract subprograms")), 152 cl::init(DefaultLinkageNames)); 153 154 static cl::opt<DwarfDebug::MinimizeAddrInV5> MinimizeAddrInV5Option( 155 "minimize-addr-in-v5", cl::Hidden, 156 cl::desc("Always use DW_AT_ranges in DWARFv5 whenever it could allow more " 157 "address pool entry sharing to reduce relocations/object size"), 158 cl::values(clEnumValN(DwarfDebug::MinimizeAddrInV5::Default, "Default", 159 "Default address minimization strategy"), 160 clEnumValN(DwarfDebug::MinimizeAddrInV5::Ranges, "Ranges", 161 "Use rnglists for contiguous ranges if that allows " 162 "using a pre-existing base address"), 163 clEnumValN(DwarfDebug::MinimizeAddrInV5::Expressions, 164 "Expressions", 165 "Use exprloc addrx+offset expressions for any " 166 "address with a prior base address"), 167 clEnumValN(DwarfDebug::MinimizeAddrInV5::Form, "Form", 168 "Use addrx+offset extension form for any address " 169 "with a prior base address"), 170 clEnumValN(DwarfDebug::MinimizeAddrInV5::Disabled, "Disabled", 171 "Stuff")), 172 cl::init(DwarfDebug::MinimizeAddrInV5::Default)); 173 174 static constexpr unsigned ULEB128PadSize = 4; 175 176 void DebugLocDwarfExpression::emitOp(uint8_t Op, const char *Comment) { 177 getActiveStreamer().emitInt8( 178 Op, Comment ? Twine(Comment) + " " + dwarf::OperationEncodingString(Op) 179 : dwarf::OperationEncodingString(Op)); 180 } 181 182 void DebugLocDwarfExpression::emitSigned(int64_t Value) { 183 getActiveStreamer().emitSLEB128(Value, Twine(Value)); 184 } 185 186 void DebugLocDwarfExpression::emitUnsigned(uint64_t Value) { 187 getActiveStreamer().emitULEB128(Value, Twine(Value)); 188 } 189 190 void DebugLocDwarfExpression::emitData1(uint8_t Value) { 191 getActiveStreamer().emitInt8(Value, Twine(Value)); 192 } 193 194 void DebugLocDwarfExpression::emitBaseTypeRef(uint64_t Idx) { 195 assert(Idx < (1ULL << (ULEB128PadSize * 7)) && "Idx wont fit"); 196 getActiveStreamer().emitULEB128(Idx, Twine(Idx), ULEB128PadSize); 197 } 198 199 bool DebugLocDwarfExpression::isFrameRegister(const TargetRegisterInfo &TRI, 200 llvm::Register MachineReg) { 201 // This information is not available while emitting .debug_loc entries. 202 return false; 203 } 204 205 void DebugLocDwarfExpression::enableTemporaryBuffer() { 206 assert(!IsBuffering && "Already buffering?"); 207 if (!TmpBuf) 208 TmpBuf = std::make_unique<TempBuffer>(OutBS.GenerateComments); 209 IsBuffering = true; 210 } 211 212 void DebugLocDwarfExpression::disableTemporaryBuffer() { IsBuffering = false; } 213 214 unsigned DebugLocDwarfExpression::getTemporaryBufferSize() { 215 return TmpBuf ? TmpBuf->Bytes.size() : 0; 216 } 217 218 void DebugLocDwarfExpression::commitTemporaryBuffer() { 219 if (!TmpBuf) 220 return; 221 for (auto Byte : enumerate(TmpBuf->Bytes)) { 222 const char *Comment = (Byte.index() < TmpBuf->Comments.size()) 223 ? TmpBuf->Comments[Byte.index()].c_str() 224 : ""; 225 OutBS.emitInt8(Byte.value(), Comment); 226 } 227 TmpBuf->Bytes.clear(); 228 TmpBuf->Comments.clear(); 229 } 230 231 const DIType *DbgVariable::getType() const { 232 return getVariable()->getType(); 233 } 234 235 /// Get .debug_loc entry for the instruction range starting at MI. 236 static DbgValueLoc getDebugLocValue(const MachineInstr *MI) { 237 const DIExpression *Expr = MI->getDebugExpression(); 238 const bool IsVariadic = MI->isDebugValueList(); 239 assert(MI->getNumOperands() >= 3); 240 SmallVector<DbgValueLocEntry, 4> DbgValueLocEntries; 241 for (const MachineOperand &Op : MI->debug_operands()) { 242 if (Op.isReg()) { 243 MachineLocation MLoc(Op.getReg(), 244 MI->isNonListDebugValue() && MI->isDebugOffsetImm()); 245 DbgValueLocEntries.push_back(DbgValueLocEntry(MLoc)); 246 } else if (Op.isTargetIndex()) { 247 DbgValueLocEntries.push_back( 248 DbgValueLocEntry(TargetIndexLocation(Op.getIndex(), Op.getOffset()))); 249 } else if (Op.isImm()) 250 DbgValueLocEntries.push_back(DbgValueLocEntry(Op.getImm())); 251 else if (Op.isFPImm()) 252 DbgValueLocEntries.push_back(DbgValueLocEntry(Op.getFPImm())); 253 else if (Op.isCImm()) 254 DbgValueLocEntries.push_back(DbgValueLocEntry(Op.getCImm())); 255 else 256 llvm_unreachable("Unexpected debug operand in DBG_VALUE* instruction!"); 257 } 258 return DbgValueLoc(Expr, DbgValueLocEntries, IsVariadic); 259 } 260 261 void DbgVariable::initializeDbgValue(const MachineInstr *DbgValue) { 262 assert(FrameIndexExprs.empty() && "Already initialized?"); 263 assert(!ValueLoc.get() && "Already initialized?"); 264 265 assert(getVariable() == DbgValue->getDebugVariable() && "Wrong variable"); 266 assert(getInlinedAt() == DbgValue->getDebugLoc()->getInlinedAt() && 267 "Wrong inlined-at"); 268 269 ValueLoc = std::make_unique<DbgValueLoc>(getDebugLocValue(DbgValue)); 270 if (auto *E = DbgValue->getDebugExpression()) 271 if (E->getNumElements()) 272 FrameIndexExprs.push_back({0, E}); 273 } 274 275 ArrayRef<DbgVariable::FrameIndexExpr> DbgVariable::getFrameIndexExprs() const { 276 if (FrameIndexExprs.size() == 1) 277 return FrameIndexExprs; 278 279 assert(llvm::all_of(FrameIndexExprs, 280 [](const FrameIndexExpr &A) { 281 return A.Expr->isFragment(); 282 }) && 283 "multiple FI expressions without DW_OP_LLVM_fragment"); 284 llvm::sort(FrameIndexExprs, 285 [](const FrameIndexExpr &A, const FrameIndexExpr &B) -> bool { 286 return A.Expr->getFragmentInfo()->OffsetInBits < 287 B.Expr->getFragmentInfo()->OffsetInBits; 288 }); 289 290 return FrameIndexExprs; 291 } 292 293 void DbgVariable::addMMIEntry(const DbgVariable &V) { 294 assert(DebugLocListIndex == ~0U && !ValueLoc.get() && "not an MMI entry"); 295 assert(V.DebugLocListIndex == ~0U && !V.ValueLoc.get() && "not an MMI entry"); 296 assert(V.getVariable() == getVariable() && "conflicting variable"); 297 assert(V.getInlinedAt() == getInlinedAt() && "conflicting inlined-at location"); 298 299 assert(!FrameIndexExprs.empty() && "Expected an MMI entry"); 300 assert(!V.FrameIndexExprs.empty() && "Expected an MMI entry"); 301 302 // FIXME: This logic should not be necessary anymore, as we now have proper 303 // deduplication. However, without it, we currently run into the assertion 304 // below, which means that we are likely dealing with broken input, i.e. two 305 // non-fragment entries for the same variable at different frame indices. 306 if (FrameIndexExprs.size()) { 307 auto *Expr = FrameIndexExprs.back().Expr; 308 if (!Expr || !Expr->isFragment()) 309 return; 310 } 311 312 for (const auto &FIE : V.FrameIndexExprs) 313 // Ignore duplicate entries. 314 if (llvm::none_of(FrameIndexExprs, [&](const FrameIndexExpr &Other) { 315 return FIE.FI == Other.FI && FIE.Expr == Other.Expr; 316 })) 317 FrameIndexExprs.push_back(FIE); 318 319 assert((FrameIndexExprs.size() == 1 || 320 llvm::all_of(FrameIndexExprs, 321 [](FrameIndexExpr &FIE) { 322 return FIE.Expr && FIE.Expr->isFragment(); 323 })) && 324 "conflicting locations for variable"); 325 } 326 327 static AccelTableKind computeAccelTableKind(unsigned DwarfVersion, 328 bool GenerateTypeUnits, 329 DebuggerKind Tuning, 330 const Triple &TT) { 331 // Honor an explicit request. 332 if (AccelTables != AccelTableKind::Default) 333 return AccelTables; 334 335 // Accelerator tables with type units are currently not supported. 336 if (GenerateTypeUnits) 337 return AccelTableKind::None; 338 339 // Accelerator tables get emitted if targetting DWARF v5 or LLDB. DWARF v5 340 // always implies debug_names. For lower standard versions we use apple 341 // accelerator tables on apple platforms and debug_names elsewhere. 342 if (DwarfVersion >= 5) 343 return AccelTableKind::Dwarf; 344 if (Tuning == DebuggerKind::LLDB) 345 return TT.isOSBinFormatMachO() ? AccelTableKind::Apple 346 : AccelTableKind::Dwarf; 347 return AccelTableKind::None; 348 } 349 350 DwarfDebug::DwarfDebug(AsmPrinter *A) 351 : DebugHandlerBase(A), DebugLocs(A->OutStreamer->isVerboseAsm()), 352 InfoHolder(A, "info_string", DIEValueAllocator), 353 SkeletonHolder(A, "skel_string", DIEValueAllocator), 354 IsDarwin(A->TM.getTargetTriple().isOSDarwin()) { 355 const Triple &TT = Asm->TM.getTargetTriple(); 356 357 // Make sure we know our "debugger tuning". The target option takes 358 // precedence; fall back to triple-based defaults. 359 if (Asm->TM.Options.DebuggerTuning != DebuggerKind::Default) 360 DebuggerTuning = Asm->TM.Options.DebuggerTuning; 361 else if (IsDarwin) 362 DebuggerTuning = DebuggerKind::LLDB; 363 else if (TT.isPS4CPU()) 364 DebuggerTuning = DebuggerKind::SCE; 365 else if (TT.isOSAIX()) 366 DebuggerTuning = DebuggerKind::DBX; 367 else 368 DebuggerTuning = DebuggerKind::GDB; 369 370 if (DwarfInlinedStrings == Default) 371 UseInlineStrings = TT.isNVPTX() || tuneForDBX(); 372 else 373 UseInlineStrings = DwarfInlinedStrings == Enable; 374 375 UseLocSection = !TT.isNVPTX(); 376 377 HasAppleExtensionAttributes = tuneForLLDB(); 378 379 // Handle split DWARF. 380 HasSplitDwarf = !Asm->TM.Options.MCOptions.SplitDwarfFile.empty(); 381 382 // SCE defaults to linkage names only for abstract subprograms. 383 if (DwarfLinkageNames == DefaultLinkageNames) 384 UseAllLinkageNames = !tuneForSCE(); 385 else 386 UseAllLinkageNames = DwarfLinkageNames == AllLinkageNames; 387 388 unsigned DwarfVersionNumber = Asm->TM.Options.MCOptions.DwarfVersion; 389 unsigned DwarfVersion = DwarfVersionNumber ? DwarfVersionNumber 390 : MMI->getModule()->getDwarfVersion(); 391 // Use dwarf 4 by default if nothing is requested. For NVPTX, use dwarf 2. 392 DwarfVersion = 393 TT.isNVPTX() ? 2 : (DwarfVersion ? DwarfVersion : dwarf::DWARF_VERSION); 394 395 bool Dwarf64 = DwarfVersion >= 3 && // DWARF64 was introduced in DWARFv3. 396 TT.isArch64Bit(); // DWARF64 requires 64-bit relocations. 397 398 // Support DWARF64 399 // 1: For ELF when requested. 400 // 2: For XCOFF64: the AIX assembler will fill in debug section lengths 401 // according to the DWARF64 format for 64-bit assembly, so we must use 402 // DWARF64 in the compiler too for 64-bit mode. 403 Dwarf64 &= 404 ((Asm->TM.Options.MCOptions.Dwarf64 || MMI->getModule()->isDwarf64()) && 405 TT.isOSBinFormatELF()) || 406 TT.isOSBinFormatXCOFF(); 407 408 if (!Dwarf64 && TT.isArch64Bit() && TT.isOSBinFormatXCOFF()) 409 report_fatal_error("XCOFF requires DWARF64 for 64-bit mode!"); 410 411 UseRangesSection = !NoDwarfRangesSection && !TT.isNVPTX(); 412 413 // Use sections as references. Force for NVPTX. 414 if (DwarfSectionsAsReferences == Default) 415 UseSectionsAsReferences = TT.isNVPTX(); 416 else 417 UseSectionsAsReferences = DwarfSectionsAsReferences == Enable; 418 419 // Don't generate type units for unsupported object file formats. 420 GenerateTypeUnits = (A->TM.getTargetTriple().isOSBinFormatELF() || 421 A->TM.getTargetTriple().isOSBinFormatWasm()) && 422 GenerateDwarfTypeUnits; 423 424 TheAccelTableKind = computeAccelTableKind( 425 DwarfVersion, GenerateTypeUnits, DebuggerTuning, A->TM.getTargetTriple()); 426 427 // Work around a GDB bug. GDB doesn't support the standard opcode; 428 // SCE doesn't support GNU's; LLDB prefers the standard opcode, which 429 // is defined as of DWARF 3. 430 // See GDB bug 11616 - DW_OP_form_tls_address is unimplemented 431 // https://sourceware.org/bugzilla/show_bug.cgi?id=11616 432 UseGNUTLSOpcode = tuneForGDB() || DwarfVersion < 3; 433 434 // GDB does not fully support the DWARF 4 representation for bitfields. 435 UseDWARF2Bitfields = (DwarfVersion < 4) || tuneForGDB(); 436 437 // The DWARF v5 string offsets table has - possibly shared - contributions 438 // from each compile and type unit each preceded by a header. The string 439 // offsets table used by the pre-DWARF v5 split-DWARF implementation uses 440 // a monolithic string offsets table without any header. 441 UseSegmentedStringOffsetsTable = DwarfVersion >= 5; 442 443 // Emit call-site-param debug info for GDB and LLDB, if the target supports 444 // the debug entry values feature. It can also be enabled explicitly. 445 EmitDebugEntryValues = Asm->TM.Options.ShouldEmitDebugEntryValues(); 446 447 // It is unclear if the GCC .debug_macro extension is well-specified 448 // for split DWARF. For now, do not allow LLVM to emit it. 449 UseDebugMacroSection = 450 DwarfVersion >= 5 || (UseGNUDebugMacro && !useSplitDwarf()); 451 if (DwarfOpConvert == Default) 452 EnableOpConvert = !((tuneForGDB() && useSplitDwarf()) || (tuneForLLDB() && !TT.isOSBinFormatMachO())); 453 else 454 EnableOpConvert = (DwarfOpConvert == Enable); 455 456 // Split DWARF would benefit object size significantly by trading reductions 457 // in address pool usage for slightly increased range list encodings. 458 if (DwarfVersion >= 5) { 459 MinimizeAddr = MinimizeAddrInV5Option; 460 // FIXME: In the future, enable this by default for Split DWARF where the 461 // tradeoff is more pronounced due to being able to offload the range 462 // lists to the dwo file and shrink object files/reduce relocations there. 463 if (MinimizeAddr == MinimizeAddrInV5::Default) 464 MinimizeAddr = MinimizeAddrInV5::Disabled; 465 } 466 467 Asm->OutStreamer->getContext().setDwarfVersion(DwarfVersion); 468 Asm->OutStreamer->getContext().setDwarfFormat(Dwarf64 ? dwarf::DWARF64 469 : dwarf::DWARF32); 470 } 471 472 // Define out of line so we don't have to include DwarfUnit.h in DwarfDebug.h. 473 DwarfDebug::~DwarfDebug() = default; 474 475 static bool isObjCClass(StringRef Name) { 476 return Name.startswith("+") || Name.startswith("-"); 477 } 478 479 static bool hasObjCCategory(StringRef Name) { 480 if (!isObjCClass(Name)) 481 return false; 482 483 return Name.contains(") "); 484 } 485 486 static void getObjCClassCategory(StringRef In, StringRef &Class, 487 StringRef &Category) { 488 if (!hasObjCCategory(In)) { 489 Class = In.slice(In.find('[') + 1, In.find(' ')); 490 Category = ""; 491 return; 492 } 493 494 Class = In.slice(In.find('[') + 1, In.find('(')); 495 Category = In.slice(In.find('[') + 1, In.find(' ')); 496 } 497 498 static StringRef getObjCMethodName(StringRef In) { 499 return In.slice(In.find(' ') + 1, In.find(']')); 500 } 501 502 // Add the various names to the Dwarf accelerator table names. 503 void DwarfDebug::addSubprogramNames(const DICompileUnit &CU, 504 const DISubprogram *SP, DIE &Die) { 505 if (getAccelTableKind() != AccelTableKind::Apple && 506 CU.getNameTableKind() == DICompileUnit::DebugNameTableKind::None) 507 return; 508 509 if (!SP->isDefinition()) 510 return; 511 512 if (SP->getName() != "") 513 addAccelName(CU, SP->getName(), Die); 514 515 // If the linkage name is different than the name, go ahead and output that as 516 // well into the name table. Only do that if we are going to actually emit 517 // that name. 518 if (SP->getLinkageName() != "" && SP->getName() != SP->getLinkageName() && 519 (useAllLinkageNames() || InfoHolder.getAbstractSPDies().lookup(SP))) 520 addAccelName(CU, SP->getLinkageName(), Die); 521 522 // If this is an Objective-C selector name add it to the ObjC accelerator 523 // too. 524 if (isObjCClass(SP->getName())) { 525 StringRef Class, Category; 526 getObjCClassCategory(SP->getName(), Class, Category); 527 addAccelObjC(CU, Class, Die); 528 if (Category != "") 529 addAccelObjC(CU, Category, Die); 530 // Also add the base method name to the name table. 531 addAccelName(CU, getObjCMethodName(SP->getName()), Die); 532 } 533 } 534 535 /// Check whether we should create a DIE for the given Scope, return true 536 /// if we don't create a DIE (the corresponding DIE is null). 537 bool DwarfDebug::isLexicalScopeDIENull(LexicalScope *Scope) { 538 if (Scope->isAbstractScope()) 539 return false; 540 541 // We don't create a DIE if there is no Range. 542 const SmallVectorImpl<InsnRange> &Ranges = Scope->getRanges(); 543 if (Ranges.empty()) 544 return true; 545 546 if (Ranges.size() > 1) 547 return false; 548 549 // We don't create a DIE if we have a single Range and the end label 550 // is null. 551 return !getLabelAfterInsn(Ranges.front().second); 552 } 553 554 template <typename Func> static void forBothCUs(DwarfCompileUnit &CU, Func F) { 555 F(CU); 556 if (auto *SkelCU = CU.getSkeleton()) 557 if (CU.getCUNode()->getSplitDebugInlining()) 558 F(*SkelCU); 559 } 560 561 bool DwarfDebug::shareAcrossDWOCUs() const { 562 return SplitDwarfCrossCuReferences; 563 } 564 565 void DwarfDebug::constructAbstractSubprogramScopeDIE(DwarfCompileUnit &SrcCU, 566 LexicalScope *Scope) { 567 assert(Scope && Scope->getScopeNode()); 568 assert(Scope->isAbstractScope()); 569 assert(!Scope->getInlinedAt()); 570 571 auto *SP = cast<DISubprogram>(Scope->getScopeNode()); 572 573 // Find the subprogram's DwarfCompileUnit in the SPMap in case the subprogram 574 // was inlined from another compile unit. 575 if (useSplitDwarf() && !shareAcrossDWOCUs() && !SP->getUnit()->getSplitDebugInlining()) 576 // Avoid building the original CU if it won't be used 577 SrcCU.constructAbstractSubprogramScopeDIE(Scope); 578 else { 579 auto &CU = getOrCreateDwarfCompileUnit(SP->getUnit()); 580 if (auto *SkelCU = CU.getSkeleton()) { 581 (shareAcrossDWOCUs() ? CU : SrcCU) 582 .constructAbstractSubprogramScopeDIE(Scope); 583 if (CU.getCUNode()->getSplitDebugInlining()) 584 SkelCU->constructAbstractSubprogramScopeDIE(Scope); 585 } else 586 CU.constructAbstractSubprogramScopeDIE(Scope); 587 } 588 } 589 590 /// Represents a parameter whose call site value can be described by applying a 591 /// debug expression to a register in the forwarded register worklist. 592 struct FwdRegParamInfo { 593 /// The described parameter register. 594 unsigned ParamReg; 595 596 /// Debug expression that has been built up when walking through the 597 /// instruction chain that produces the parameter's value. 598 const DIExpression *Expr; 599 }; 600 601 /// Register worklist for finding call site values. 602 using FwdRegWorklist = MapVector<unsigned, SmallVector<FwdRegParamInfo, 2>>; 603 604 /// Append the expression \p Addition to \p Original and return the result. 605 static const DIExpression *combineDIExpressions(const DIExpression *Original, 606 const DIExpression *Addition) { 607 std::vector<uint64_t> Elts = Addition->getElements().vec(); 608 // Avoid multiple DW_OP_stack_values. 609 if (Original->isImplicit() && Addition->isImplicit()) 610 erase_value(Elts, dwarf::DW_OP_stack_value); 611 const DIExpression *CombinedExpr = 612 (Elts.size() > 0) ? DIExpression::append(Original, Elts) : Original; 613 return CombinedExpr; 614 } 615 616 /// Emit call site parameter entries that are described by the given value and 617 /// debug expression. 618 template <typename ValT> 619 static void finishCallSiteParams(ValT Val, const DIExpression *Expr, 620 ArrayRef<FwdRegParamInfo> DescribedParams, 621 ParamSet &Params) { 622 for (auto Param : DescribedParams) { 623 bool ShouldCombineExpressions = Expr && Param.Expr->getNumElements() > 0; 624 625 // TODO: Entry value operations can currently not be combined with any 626 // other expressions, so we can't emit call site entries in those cases. 627 if (ShouldCombineExpressions && Expr->isEntryValue()) 628 continue; 629 630 // If a parameter's call site value is produced by a chain of 631 // instructions we may have already created an expression for the 632 // parameter when walking through the instructions. Append that to the 633 // base expression. 634 const DIExpression *CombinedExpr = 635 ShouldCombineExpressions ? combineDIExpressions(Expr, Param.Expr) 636 : Expr; 637 assert((!CombinedExpr || CombinedExpr->isValid()) && 638 "Combined debug expression is invalid"); 639 640 DbgValueLoc DbgLocVal(CombinedExpr, DbgValueLocEntry(Val)); 641 DbgCallSiteParam CSParm(Param.ParamReg, DbgLocVal); 642 Params.push_back(CSParm); 643 ++NumCSParams; 644 } 645 } 646 647 /// Add \p Reg to the worklist, if it's not already present, and mark that the 648 /// given parameter registers' values can (potentially) be described using 649 /// that register and an debug expression. 650 static void addToFwdRegWorklist(FwdRegWorklist &Worklist, unsigned Reg, 651 const DIExpression *Expr, 652 ArrayRef<FwdRegParamInfo> ParamsToAdd) { 653 auto I = Worklist.insert({Reg, {}}); 654 auto &ParamsForFwdReg = I.first->second; 655 for (auto Param : ParamsToAdd) { 656 assert(none_of(ParamsForFwdReg, 657 [Param](const FwdRegParamInfo &D) { 658 return D.ParamReg == Param.ParamReg; 659 }) && 660 "Same parameter described twice by forwarding reg"); 661 662 // If a parameter's call site value is produced by a chain of 663 // instructions we may have already created an expression for the 664 // parameter when walking through the instructions. Append that to the 665 // new expression. 666 const DIExpression *CombinedExpr = combineDIExpressions(Expr, Param.Expr); 667 ParamsForFwdReg.push_back({Param.ParamReg, CombinedExpr}); 668 } 669 } 670 671 /// Interpret values loaded into registers by \p CurMI. 672 static void interpretValues(const MachineInstr *CurMI, 673 FwdRegWorklist &ForwardedRegWorklist, 674 ParamSet &Params) { 675 676 const MachineFunction *MF = CurMI->getMF(); 677 const DIExpression *EmptyExpr = 678 DIExpression::get(MF->getFunction().getContext(), {}); 679 const auto &TRI = *MF->getSubtarget().getRegisterInfo(); 680 const auto &TII = *MF->getSubtarget().getInstrInfo(); 681 const auto &TLI = *MF->getSubtarget().getTargetLowering(); 682 683 // If an instruction defines more than one item in the worklist, we may run 684 // into situations where a worklist register's value is (potentially) 685 // described by the previous value of another register that is also defined 686 // by that instruction. 687 // 688 // This can for example occur in cases like this: 689 // 690 // $r1 = mov 123 691 // $r0, $r1 = mvrr $r1, 456 692 // call @foo, $r0, $r1 693 // 694 // When describing $r1's value for the mvrr instruction, we need to make sure 695 // that we don't finalize an entry value for $r0, as that is dependent on the 696 // previous value of $r1 (123 rather than 456). 697 // 698 // In order to not have to distinguish between those cases when finalizing 699 // entry values, we simply postpone adding new parameter registers to the 700 // worklist, by first keeping them in this temporary container until the 701 // instruction has been handled. 702 FwdRegWorklist TmpWorklistItems; 703 704 // If the MI is an instruction defining one or more parameters' forwarding 705 // registers, add those defines. 706 auto getForwardingRegsDefinedByMI = [&](const MachineInstr &MI, 707 SmallSetVector<unsigned, 4> &Defs) { 708 if (MI.isDebugInstr()) 709 return; 710 711 for (const MachineOperand &MO : MI.operands()) { 712 if (MO.isReg() && MO.isDef() && 713 Register::isPhysicalRegister(MO.getReg())) { 714 for (auto &FwdReg : ForwardedRegWorklist) 715 if (TRI.regsOverlap(FwdReg.first, MO.getReg())) 716 Defs.insert(FwdReg.first); 717 } 718 } 719 }; 720 721 // Set of worklist registers that are defined by this instruction. 722 SmallSetVector<unsigned, 4> FwdRegDefs; 723 724 getForwardingRegsDefinedByMI(*CurMI, FwdRegDefs); 725 if (FwdRegDefs.empty()) 726 return; 727 728 for (auto ParamFwdReg : FwdRegDefs) { 729 if (auto ParamValue = TII.describeLoadedValue(*CurMI, ParamFwdReg)) { 730 if (ParamValue->first.isImm()) { 731 int64_t Val = ParamValue->first.getImm(); 732 finishCallSiteParams(Val, ParamValue->second, 733 ForwardedRegWorklist[ParamFwdReg], Params); 734 } else if (ParamValue->first.isReg()) { 735 Register RegLoc = ParamValue->first.getReg(); 736 Register SP = TLI.getStackPointerRegisterToSaveRestore(); 737 Register FP = TRI.getFrameRegister(*MF); 738 bool IsSPorFP = (RegLoc == SP) || (RegLoc == FP); 739 if (TRI.isCalleeSavedPhysReg(RegLoc, *MF) || IsSPorFP) { 740 MachineLocation MLoc(RegLoc, /*Indirect=*/IsSPorFP); 741 finishCallSiteParams(MLoc, ParamValue->second, 742 ForwardedRegWorklist[ParamFwdReg], Params); 743 } else { 744 // ParamFwdReg was described by the non-callee saved register 745 // RegLoc. Mark that the call site values for the parameters are 746 // dependent on that register instead of ParamFwdReg. Since RegLoc 747 // may be a register that will be handled in this iteration, we 748 // postpone adding the items to the worklist, and instead keep them 749 // in a temporary container. 750 addToFwdRegWorklist(TmpWorklistItems, RegLoc, ParamValue->second, 751 ForwardedRegWorklist[ParamFwdReg]); 752 } 753 } 754 } 755 } 756 757 // Remove all registers that this instruction defines from the worklist. 758 for (auto ParamFwdReg : FwdRegDefs) 759 ForwardedRegWorklist.erase(ParamFwdReg); 760 761 // Now that we are done handling this instruction, add items from the 762 // temporary worklist to the real one. 763 for (auto &New : TmpWorklistItems) 764 addToFwdRegWorklist(ForwardedRegWorklist, New.first, EmptyExpr, New.second); 765 TmpWorklistItems.clear(); 766 } 767 768 static bool interpretNextInstr(const MachineInstr *CurMI, 769 FwdRegWorklist &ForwardedRegWorklist, 770 ParamSet &Params) { 771 // Skip bundle headers. 772 if (CurMI->isBundle()) 773 return true; 774 775 // If the next instruction is a call we can not interpret parameter's 776 // forwarding registers or we finished the interpretation of all 777 // parameters. 778 if (CurMI->isCall()) 779 return false; 780 781 if (ForwardedRegWorklist.empty()) 782 return false; 783 784 // Avoid NOP description. 785 if (CurMI->getNumOperands() == 0) 786 return true; 787 788 interpretValues(CurMI, ForwardedRegWorklist, Params); 789 790 return true; 791 } 792 793 /// Try to interpret values loaded into registers that forward parameters 794 /// for \p CallMI. Store parameters with interpreted value into \p Params. 795 static void collectCallSiteParameters(const MachineInstr *CallMI, 796 ParamSet &Params) { 797 const MachineFunction *MF = CallMI->getMF(); 798 const auto &CalleesMap = MF->getCallSitesInfo(); 799 auto CallFwdRegsInfo = CalleesMap.find(CallMI); 800 801 // There is no information for the call instruction. 802 if (CallFwdRegsInfo == CalleesMap.end()) 803 return; 804 805 const MachineBasicBlock *MBB = CallMI->getParent(); 806 807 // Skip the call instruction. 808 auto I = std::next(CallMI->getReverseIterator()); 809 810 FwdRegWorklist ForwardedRegWorklist; 811 812 const DIExpression *EmptyExpr = 813 DIExpression::get(MF->getFunction().getContext(), {}); 814 815 // Add all the forwarding registers into the ForwardedRegWorklist. 816 for (const auto &ArgReg : CallFwdRegsInfo->second) { 817 bool InsertedReg = 818 ForwardedRegWorklist.insert({ArgReg.Reg, {{ArgReg.Reg, EmptyExpr}}}) 819 .second; 820 assert(InsertedReg && "Single register used to forward two arguments?"); 821 (void)InsertedReg; 822 } 823 824 // Do not emit CSInfo for undef forwarding registers. 825 for (auto &MO : CallMI->uses()) 826 if (MO.isReg() && MO.isUndef()) 827 ForwardedRegWorklist.erase(MO.getReg()); 828 829 // We erase, from the ForwardedRegWorklist, those forwarding registers for 830 // which we successfully describe a loaded value (by using 831 // the describeLoadedValue()). For those remaining arguments in the working 832 // list, for which we do not describe a loaded value by 833 // the describeLoadedValue(), we try to generate an entry value expression 834 // for their call site value description, if the call is within the entry MBB. 835 // TODO: Handle situations when call site parameter value can be described 836 // as the entry value within basic blocks other than the first one. 837 bool ShouldTryEmitEntryVals = MBB->getIterator() == MF->begin(); 838 839 // Search for a loading value in forwarding registers inside call delay slot. 840 if (CallMI->hasDelaySlot()) { 841 auto Suc = std::next(CallMI->getIterator()); 842 // Only one-instruction delay slot is supported. 843 auto BundleEnd = llvm::getBundleEnd(CallMI->getIterator()); 844 (void)BundleEnd; 845 assert(std::next(Suc) == BundleEnd && 846 "More than one instruction in call delay slot"); 847 // Try to interpret value loaded by instruction. 848 if (!interpretNextInstr(&*Suc, ForwardedRegWorklist, Params)) 849 return; 850 } 851 852 // Search for a loading value in forwarding registers. 853 for (; I != MBB->rend(); ++I) { 854 // Try to interpret values loaded by instruction. 855 if (!interpretNextInstr(&*I, ForwardedRegWorklist, Params)) 856 return; 857 } 858 859 // Emit the call site parameter's value as an entry value. 860 if (ShouldTryEmitEntryVals) { 861 // Create an expression where the register's entry value is used. 862 DIExpression *EntryExpr = DIExpression::get( 863 MF->getFunction().getContext(), {dwarf::DW_OP_LLVM_entry_value, 1}); 864 for (auto &RegEntry : ForwardedRegWorklist) { 865 MachineLocation MLoc(RegEntry.first); 866 finishCallSiteParams(MLoc, EntryExpr, RegEntry.second, Params); 867 } 868 } 869 } 870 871 void DwarfDebug::constructCallSiteEntryDIEs(const DISubprogram &SP, 872 DwarfCompileUnit &CU, DIE &ScopeDIE, 873 const MachineFunction &MF) { 874 // Add a call site-related attribute (DWARF5, Sec. 3.3.1.3). Do this only if 875 // the subprogram is required to have one. 876 if (!SP.areAllCallsDescribed() || !SP.isDefinition()) 877 return; 878 879 // Use DW_AT_call_all_calls to express that call site entries are present 880 // for both tail and non-tail calls. Don't use DW_AT_call_all_source_calls 881 // because one of its requirements is not met: call site entries for 882 // optimized-out calls are elided. 883 CU.addFlag(ScopeDIE, CU.getDwarf5OrGNUAttr(dwarf::DW_AT_call_all_calls)); 884 885 const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo(); 886 assert(TII && "TargetInstrInfo not found: cannot label tail calls"); 887 888 // Delay slot support check. 889 auto delaySlotSupported = [&](const MachineInstr &MI) { 890 if (!MI.isBundledWithSucc()) 891 return false; 892 auto Suc = std::next(MI.getIterator()); 893 auto CallInstrBundle = getBundleStart(MI.getIterator()); 894 (void)CallInstrBundle; 895 auto DelaySlotBundle = getBundleStart(Suc); 896 (void)DelaySlotBundle; 897 // Ensure that label after call is following delay slot instruction. 898 // Ex. CALL_INSTRUCTION { 899 // DELAY_SLOT_INSTRUCTION } 900 // LABEL_AFTER_CALL 901 assert(getLabelAfterInsn(&*CallInstrBundle) == 902 getLabelAfterInsn(&*DelaySlotBundle) && 903 "Call and its successor instruction don't have same label after."); 904 return true; 905 }; 906 907 // Emit call site entries for each call or tail call in the function. 908 for (const MachineBasicBlock &MBB : MF) { 909 for (const MachineInstr &MI : MBB.instrs()) { 910 // Bundles with call in them will pass the isCall() test below but do not 911 // have callee operand information so skip them here. Iterator will 912 // eventually reach the call MI. 913 if (MI.isBundle()) 914 continue; 915 916 // Skip instructions which aren't calls. Both calls and tail-calling jump 917 // instructions (e.g TAILJMPd64) are classified correctly here. 918 if (!MI.isCandidateForCallSiteEntry()) 919 continue; 920 921 // Skip instructions marked as frame setup, as they are not interesting to 922 // the user. 923 if (MI.getFlag(MachineInstr::FrameSetup)) 924 continue; 925 926 // Check if delay slot support is enabled. 927 if (MI.hasDelaySlot() && !delaySlotSupported(*&MI)) 928 return; 929 930 // If this is a direct call, find the callee's subprogram. 931 // In the case of an indirect call find the register that holds 932 // the callee. 933 const MachineOperand &CalleeOp = TII->getCalleeOperand(MI); 934 if (!CalleeOp.isGlobal() && 935 (!CalleeOp.isReg() || 936 !Register::isPhysicalRegister(CalleeOp.getReg()))) 937 continue; 938 939 unsigned CallReg = 0; 940 const DISubprogram *CalleeSP = nullptr; 941 const Function *CalleeDecl = nullptr; 942 if (CalleeOp.isReg()) { 943 CallReg = CalleeOp.getReg(); 944 if (!CallReg) 945 continue; 946 } else { 947 CalleeDecl = dyn_cast<Function>(CalleeOp.getGlobal()); 948 if (!CalleeDecl || !CalleeDecl->getSubprogram()) 949 continue; 950 CalleeSP = CalleeDecl->getSubprogram(); 951 } 952 953 // TODO: Omit call site entries for runtime calls (objc_msgSend, etc). 954 955 bool IsTail = TII->isTailCall(MI); 956 957 // If MI is in a bundle, the label was created after the bundle since 958 // EmitFunctionBody iterates over top-level MIs. Get that top-level MI 959 // to search for that label below. 960 const MachineInstr *TopLevelCallMI = 961 MI.isInsideBundle() ? &*getBundleStart(MI.getIterator()) : &MI; 962 963 // For non-tail calls, the return PC is needed to disambiguate paths in 964 // the call graph which could lead to some target function. For tail 965 // calls, no return PC information is needed, unless tuning for GDB in 966 // DWARF4 mode in which case we fake a return PC for compatibility. 967 const MCSymbol *PCAddr = 968 (!IsTail || CU.useGNUAnalogForDwarf5Feature()) 969 ? const_cast<MCSymbol *>(getLabelAfterInsn(TopLevelCallMI)) 970 : nullptr; 971 972 // For tail calls, it's necessary to record the address of the branch 973 // instruction so that the debugger can show where the tail call occurred. 974 const MCSymbol *CallAddr = 975 IsTail ? getLabelBeforeInsn(TopLevelCallMI) : nullptr; 976 977 assert((IsTail || PCAddr) && "Non-tail call without return PC"); 978 979 LLVM_DEBUG(dbgs() << "CallSiteEntry: " << MF.getName() << " -> " 980 << (CalleeDecl ? CalleeDecl->getName() 981 : StringRef(MF.getSubtarget() 982 .getRegisterInfo() 983 ->getName(CallReg))) 984 << (IsTail ? " [IsTail]" : "") << "\n"); 985 986 DIE &CallSiteDIE = CU.constructCallSiteEntryDIE( 987 ScopeDIE, CalleeSP, IsTail, PCAddr, CallAddr, CallReg); 988 989 // Optionally emit call-site-param debug info. 990 if (emitDebugEntryValues()) { 991 ParamSet Params; 992 // Try to interpret values of call site parameters. 993 collectCallSiteParameters(&MI, Params); 994 CU.constructCallSiteParmEntryDIEs(CallSiteDIE, Params); 995 } 996 } 997 } 998 } 999 1000 void DwarfDebug::addGnuPubAttributes(DwarfCompileUnit &U, DIE &D) const { 1001 if (!U.hasDwarfPubSections()) 1002 return; 1003 1004 U.addFlag(D, dwarf::DW_AT_GNU_pubnames); 1005 } 1006 1007 void DwarfDebug::finishUnitAttributes(const DICompileUnit *DIUnit, 1008 DwarfCompileUnit &NewCU) { 1009 DIE &Die = NewCU.getUnitDie(); 1010 StringRef FN = DIUnit->getFilename(); 1011 1012 StringRef Producer = DIUnit->getProducer(); 1013 StringRef Flags = DIUnit->getFlags(); 1014 if (!Flags.empty() && !useAppleExtensionAttributes()) { 1015 std::string ProducerWithFlags = Producer.str() + " " + Flags.str(); 1016 NewCU.addString(Die, dwarf::DW_AT_producer, ProducerWithFlags); 1017 } else 1018 NewCU.addString(Die, dwarf::DW_AT_producer, Producer); 1019 1020 NewCU.addUInt(Die, dwarf::DW_AT_language, dwarf::DW_FORM_data2, 1021 DIUnit->getSourceLanguage()); 1022 NewCU.addString(Die, dwarf::DW_AT_name, FN); 1023 StringRef SysRoot = DIUnit->getSysRoot(); 1024 if (!SysRoot.empty()) 1025 NewCU.addString(Die, dwarf::DW_AT_LLVM_sysroot, SysRoot); 1026 StringRef SDK = DIUnit->getSDK(); 1027 if (!SDK.empty()) 1028 NewCU.addString(Die, dwarf::DW_AT_APPLE_sdk, SDK); 1029 1030 // Add DW_str_offsets_base to the unit DIE, except for split units. 1031 if (useSegmentedStringOffsetsTable() && !useSplitDwarf()) 1032 NewCU.addStringOffsetsStart(); 1033 1034 if (!useSplitDwarf()) { 1035 NewCU.initStmtList(); 1036 1037 // If we're using split dwarf the compilation dir is going to be in the 1038 // skeleton CU and so we don't need to duplicate it here. 1039 if (!CompilationDir.empty()) 1040 NewCU.addString(Die, dwarf::DW_AT_comp_dir, CompilationDir); 1041 addGnuPubAttributes(NewCU, Die); 1042 } 1043 1044 if (useAppleExtensionAttributes()) { 1045 if (DIUnit->isOptimized()) 1046 NewCU.addFlag(Die, dwarf::DW_AT_APPLE_optimized); 1047 1048 StringRef Flags = DIUnit->getFlags(); 1049 if (!Flags.empty()) 1050 NewCU.addString(Die, dwarf::DW_AT_APPLE_flags, Flags); 1051 1052 if (unsigned RVer = DIUnit->getRuntimeVersion()) 1053 NewCU.addUInt(Die, dwarf::DW_AT_APPLE_major_runtime_vers, 1054 dwarf::DW_FORM_data1, RVer); 1055 } 1056 1057 if (DIUnit->getDWOId()) { 1058 // This CU is either a clang module DWO or a skeleton CU. 1059 NewCU.addUInt(Die, dwarf::DW_AT_GNU_dwo_id, dwarf::DW_FORM_data8, 1060 DIUnit->getDWOId()); 1061 if (!DIUnit->getSplitDebugFilename().empty()) { 1062 // This is a prefabricated skeleton CU. 1063 dwarf::Attribute attrDWOName = getDwarfVersion() >= 5 1064 ? dwarf::DW_AT_dwo_name 1065 : dwarf::DW_AT_GNU_dwo_name; 1066 NewCU.addString(Die, attrDWOName, DIUnit->getSplitDebugFilename()); 1067 } 1068 } 1069 } 1070 // Create new DwarfCompileUnit for the given metadata node with tag 1071 // DW_TAG_compile_unit. 1072 DwarfCompileUnit & 1073 DwarfDebug::getOrCreateDwarfCompileUnit(const DICompileUnit *DIUnit) { 1074 if (auto *CU = CUMap.lookup(DIUnit)) 1075 return *CU; 1076 1077 CompilationDir = DIUnit->getDirectory(); 1078 1079 auto OwnedUnit = std::make_unique<DwarfCompileUnit>( 1080 InfoHolder.getUnits().size(), DIUnit, Asm, this, &InfoHolder); 1081 DwarfCompileUnit &NewCU = *OwnedUnit; 1082 InfoHolder.addUnit(std::move(OwnedUnit)); 1083 1084 for (auto *IE : DIUnit->getImportedEntities()) 1085 NewCU.addImportedEntity(IE); 1086 1087 // LTO with assembly output shares a single line table amongst multiple CUs. 1088 // To avoid the compilation directory being ambiguous, let the line table 1089 // explicitly describe the directory of all files, never relying on the 1090 // compilation directory. 1091 if (!Asm->OutStreamer->hasRawTextSupport() || SingleCU) 1092 Asm->OutStreamer->emitDwarfFile0Directive( 1093 CompilationDir, DIUnit->getFilename(), getMD5AsBytes(DIUnit->getFile()), 1094 DIUnit->getSource(), NewCU.getUniqueID()); 1095 1096 if (useSplitDwarf()) { 1097 NewCU.setSkeleton(constructSkeletonCU(NewCU)); 1098 NewCU.setSection(Asm->getObjFileLowering().getDwarfInfoDWOSection()); 1099 } else { 1100 finishUnitAttributes(DIUnit, NewCU); 1101 NewCU.setSection(Asm->getObjFileLowering().getDwarfInfoSection()); 1102 } 1103 1104 CUMap.insert({DIUnit, &NewCU}); 1105 CUDieMap.insert({&NewCU.getUnitDie(), &NewCU}); 1106 return NewCU; 1107 } 1108 1109 void DwarfDebug::constructAndAddImportedEntityDIE(DwarfCompileUnit &TheCU, 1110 const DIImportedEntity *N) { 1111 if (isa<DILocalScope>(N->getScope())) 1112 return; 1113 if (DIE *D = TheCU.getOrCreateContextDIE(N->getScope())) 1114 D->addChild(TheCU.constructImportedEntityDIE(N)); 1115 } 1116 1117 // Emit all Dwarf sections that should come prior to the content. Create 1118 // global DIEs and emit initial debug info sections. This is invoked by 1119 // the target AsmPrinter. 1120 void DwarfDebug::beginModule(Module *M) { 1121 DebugHandlerBase::beginModule(M); 1122 1123 if (!Asm || !MMI->hasDebugInfo()) 1124 return; 1125 1126 unsigned NumDebugCUs = std::distance(M->debug_compile_units_begin(), 1127 M->debug_compile_units_end()); 1128 assert(NumDebugCUs > 0 && "Asm unexpectedly initialized"); 1129 assert(MMI->hasDebugInfo() && 1130 "DebugInfoAvailabilty unexpectedly not initialized"); 1131 SingleCU = NumDebugCUs == 1; 1132 1133 // Create the symbol that designates the start of the unit's contribution 1134 // to the string offsets table. In a split DWARF scenario, only the skeleton 1135 // unit has the DW_AT_str_offsets_base attribute (and hence needs the symbol). 1136 if (useSegmentedStringOffsetsTable()) 1137 (useSplitDwarf() ? SkeletonHolder : InfoHolder) 1138 .setStringOffsetsStartSym(Asm->createTempSymbol("str_offsets_base")); 1139 1140 1141 // Create the symbols that designates the start of the DWARF v5 range list 1142 // and locations list tables. They are located past the table headers. 1143 if (getDwarfVersion() >= 5) { 1144 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 1145 Holder.setRnglistsTableBaseSym( 1146 Asm->createTempSymbol("rnglists_table_base")); 1147 1148 if (useSplitDwarf()) 1149 InfoHolder.setRnglistsTableBaseSym( 1150 Asm->createTempSymbol("rnglists_dwo_table_base")); 1151 } 1152 1153 // Create the symbol that points to the first entry following the debug 1154 // address table (.debug_addr) header. 1155 AddrPool.setLabel(Asm->createTempSymbol("addr_table_base")); 1156 DebugLocs.setSym(Asm->createTempSymbol("loclists_table_base")); 1157 } 1158 1159 void DwarfDebug::finishEntityDefinitions() { 1160 for (const auto &Entity : ConcreteEntities) { 1161 DIE *Die = Entity->getDIE(); 1162 assert(Die); 1163 // FIXME: Consider the time-space tradeoff of just storing the unit pointer 1164 // in the ConcreteEntities list, rather than looking it up again here. 1165 // DIE::getUnit isn't simple - it walks parent pointers, etc. 1166 DwarfCompileUnit *Unit = CUDieMap.lookup(Die->getUnitDie()); 1167 assert(Unit); 1168 Unit->finishEntityDefinition(Entity.get()); 1169 } 1170 } 1171 1172 void DwarfDebug::finishSubprogramDefinitions() { 1173 for (const DISubprogram *SP : ProcessedSPNodes) { 1174 assert(SP->getUnit()->getEmissionKind() != DICompileUnit::NoDebug); 1175 forBothCUs( 1176 getOrCreateDwarfCompileUnit(SP->getUnit()), 1177 [&](DwarfCompileUnit &CU) { CU.finishSubprogramDefinition(SP); }); 1178 } 1179 } 1180 1181 void DwarfDebug::finalizeModuleInfo() { 1182 const TargetLoweringObjectFile &TLOF = Asm->getObjFileLowering(); 1183 1184 finishSubprogramDefinitions(); 1185 1186 finishEntityDefinitions(); 1187 1188 // Include the DWO file name in the hash if there's more than one CU. 1189 // This handles ThinLTO's situation where imported CUs may very easily be 1190 // duplicate with the same CU partially imported into another ThinLTO unit. 1191 StringRef DWOName; 1192 if (CUMap.size() > 1) 1193 DWOName = Asm->TM.Options.MCOptions.SplitDwarfFile; 1194 1195 // Handle anything that needs to be done on a per-unit basis after 1196 // all other generation. 1197 for (const auto &P : CUMap) { 1198 auto &TheCU = *P.second; 1199 if (TheCU.getCUNode()->isDebugDirectivesOnly()) 1200 continue; 1201 // Emit DW_AT_containing_type attribute to connect types with their 1202 // vtable holding type. 1203 TheCU.constructContainingTypeDIEs(); 1204 1205 // Add CU specific attributes if we need to add any. 1206 // If we're splitting the dwarf out now that we've got the entire 1207 // CU then add the dwo id to it. 1208 auto *SkCU = TheCU.getSkeleton(); 1209 1210 bool HasSplitUnit = SkCU && !TheCU.getUnitDie().children().empty(); 1211 1212 if (HasSplitUnit) { 1213 dwarf::Attribute attrDWOName = getDwarfVersion() >= 5 1214 ? dwarf::DW_AT_dwo_name 1215 : dwarf::DW_AT_GNU_dwo_name; 1216 finishUnitAttributes(TheCU.getCUNode(), TheCU); 1217 TheCU.addString(TheCU.getUnitDie(), attrDWOName, 1218 Asm->TM.Options.MCOptions.SplitDwarfFile); 1219 SkCU->addString(SkCU->getUnitDie(), attrDWOName, 1220 Asm->TM.Options.MCOptions.SplitDwarfFile); 1221 // Emit a unique identifier for this CU. 1222 uint64_t ID = 1223 DIEHash(Asm, &TheCU).computeCUSignature(DWOName, TheCU.getUnitDie()); 1224 if (getDwarfVersion() >= 5) { 1225 TheCU.setDWOId(ID); 1226 SkCU->setDWOId(ID); 1227 } else { 1228 TheCU.addUInt(TheCU.getUnitDie(), dwarf::DW_AT_GNU_dwo_id, 1229 dwarf::DW_FORM_data8, ID); 1230 SkCU->addUInt(SkCU->getUnitDie(), dwarf::DW_AT_GNU_dwo_id, 1231 dwarf::DW_FORM_data8, ID); 1232 } 1233 1234 if (getDwarfVersion() < 5 && !SkeletonHolder.getRangeLists().empty()) { 1235 const MCSymbol *Sym = TLOF.getDwarfRangesSection()->getBeginSymbol(); 1236 SkCU->addSectionLabel(SkCU->getUnitDie(), dwarf::DW_AT_GNU_ranges_base, 1237 Sym, Sym); 1238 } 1239 } else if (SkCU) { 1240 finishUnitAttributes(SkCU->getCUNode(), *SkCU); 1241 } 1242 1243 // If we have code split among multiple sections or non-contiguous 1244 // ranges of code then emit a DW_AT_ranges attribute on the unit that will 1245 // remain in the .o file, otherwise add a DW_AT_low_pc. 1246 // FIXME: We should use ranges allow reordering of code ala 1247 // .subsections_via_symbols in mach-o. This would mean turning on 1248 // ranges for all subprogram DIEs for mach-o. 1249 DwarfCompileUnit &U = SkCU ? *SkCU : TheCU; 1250 1251 if (unsigned NumRanges = TheCU.getRanges().size()) { 1252 if (NumRanges > 1 && useRangesSection()) 1253 // A DW_AT_low_pc attribute may also be specified in combination with 1254 // DW_AT_ranges to specify the default base address for use in 1255 // location lists (see Section 2.6.2) and range lists (see Section 1256 // 2.17.3). 1257 U.addUInt(U.getUnitDie(), dwarf::DW_AT_low_pc, dwarf::DW_FORM_addr, 0); 1258 else 1259 U.setBaseAddress(TheCU.getRanges().front().Begin); 1260 U.attachRangesOrLowHighPC(U.getUnitDie(), TheCU.takeRanges()); 1261 } 1262 1263 // We don't keep track of which addresses are used in which CU so this 1264 // is a bit pessimistic under LTO. 1265 if ((HasSplitUnit || getDwarfVersion() >= 5) && !AddrPool.isEmpty()) 1266 U.addAddrTableBase(); 1267 1268 if (getDwarfVersion() >= 5) { 1269 if (U.hasRangeLists()) 1270 U.addRnglistsBase(); 1271 1272 if (!DebugLocs.getLists().empty()) { 1273 if (!useSplitDwarf()) 1274 U.addSectionLabel(U.getUnitDie(), dwarf::DW_AT_loclists_base, 1275 DebugLocs.getSym(), 1276 TLOF.getDwarfLoclistsSection()->getBeginSymbol()); 1277 } 1278 } 1279 1280 auto *CUNode = cast<DICompileUnit>(P.first); 1281 // If compile Unit has macros, emit "DW_AT_macro_info/DW_AT_macros" 1282 // attribute. 1283 if (CUNode->getMacros()) { 1284 if (UseDebugMacroSection) { 1285 if (useSplitDwarf()) 1286 TheCU.addSectionDelta( 1287 TheCU.getUnitDie(), dwarf::DW_AT_macros, U.getMacroLabelBegin(), 1288 TLOF.getDwarfMacroDWOSection()->getBeginSymbol()); 1289 else { 1290 dwarf::Attribute MacrosAttr = getDwarfVersion() >= 5 1291 ? dwarf::DW_AT_macros 1292 : dwarf::DW_AT_GNU_macros; 1293 U.addSectionLabel(U.getUnitDie(), MacrosAttr, U.getMacroLabelBegin(), 1294 TLOF.getDwarfMacroSection()->getBeginSymbol()); 1295 } 1296 } else { 1297 if (useSplitDwarf()) 1298 TheCU.addSectionDelta( 1299 TheCU.getUnitDie(), dwarf::DW_AT_macro_info, 1300 U.getMacroLabelBegin(), 1301 TLOF.getDwarfMacinfoDWOSection()->getBeginSymbol()); 1302 else 1303 U.addSectionLabel(U.getUnitDie(), dwarf::DW_AT_macro_info, 1304 U.getMacroLabelBegin(), 1305 TLOF.getDwarfMacinfoSection()->getBeginSymbol()); 1306 } 1307 } 1308 } 1309 1310 // Emit all frontend-produced Skeleton CUs, i.e., Clang modules. 1311 for (auto *CUNode : MMI->getModule()->debug_compile_units()) 1312 if (CUNode->getDWOId()) 1313 getOrCreateDwarfCompileUnit(CUNode); 1314 1315 // Compute DIE offsets and sizes. 1316 InfoHolder.computeSizeAndOffsets(); 1317 if (useSplitDwarf()) 1318 SkeletonHolder.computeSizeAndOffsets(); 1319 } 1320 1321 /// Sort and unique GVEs by comparing their fragment offset. 1322 static SmallVectorImpl<DwarfCompileUnit::GlobalExpr> & 1323 sortGlobalExprs(SmallVectorImpl<DwarfCompileUnit::GlobalExpr> &GVEs) { 1324 llvm::sort( 1325 GVEs, [](DwarfCompileUnit::GlobalExpr A, DwarfCompileUnit::GlobalExpr B) { 1326 // Sort order: first null exprs, then exprs without fragment 1327 // info, then sort by fragment offset in bits. 1328 // FIXME: Come up with a more comprehensive comparator so 1329 // the sorting isn't non-deterministic, and so the following 1330 // std::unique call works correctly. 1331 if (!A.Expr || !B.Expr) 1332 return !!B.Expr; 1333 auto FragmentA = A.Expr->getFragmentInfo(); 1334 auto FragmentB = B.Expr->getFragmentInfo(); 1335 if (!FragmentA || !FragmentB) 1336 return !!FragmentB; 1337 return FragmentA->OffsetInBits < FragmentB->OffsetInBits; 1338 }); 1339 GVEs.erase(std::unique(GVEs.begin(), GVEs.end(), 1340 [](DwarfCompileUnit::GlobalExpr A, 1341 DwarfCompileUnit::GlobalExpr B) { 1342 return A.Expr == B.Expr; 1343 }), 1344 GVEs.end()); 1345 return GVEs; 1346 } 1347 1348 /// Create a DIE for \p Ty if it doesn't already exist. If type units are 1349 /// enabled, try to emit a type unit without a CU skeleton DIE. 1350 static void createMaybeUnusedType(DwarfDebug &DD, DwarfCompileUnit &CU, 1351 DIType &Ty) { 1352 // Try to generate a type unit without creating a skeleton DIE in this CU. 1353 if (DICompositeType const *CTy = dyn_cast<DICompositeType>(&Ty)) { 1354 MDString const *TypeId = CTy->getRawIdentifier(); 1355 if (DD.generateTypeUnits() && TypeId && !Ty.isForwardDecl()) 1356 if (DD.getOrCreateDwarfTypeUnit(CU, TypeId->getString(), CTy)) 1357 return; 1358 } 1359 // We couldn't or shouldn't add a type unit so create the DIE normally. 1360 CU.getOrCreateTypeDIE(&Ty); 1361 } 1362 1363 // Emit all Dwarf sections that should come after the content. 1364 void DwarfDebug::endModule() { 1365 // Terminate the pending line table. 1366 if (PrevCU) 1367 terminateLineTable(PrevCU); 1368 PrevCU = nullptr; 1369 assert(CurFn == nullptr); 1370 assert(CurMI == nullptr); 1371 1372 DenseMap<DIGlobalVariable *, SmallVector<DwarfCompileUnit::GlobalExpr, 1>> 1373 GVMap; 1374 for (const GlobalVariable &Global : MMI->getModule()->globals()) { 1375 SmallVector<DIGlobalVariableExpression *, 1> GVs; 1376 Global.getDebugInfo(GVs); 1377 for (auto *GVE : GVs) 1378 GVMap[GVE->getVariable()].push_back({&Global, GVE->getExpression()}); 1379 } 1380 1381 for (DICompileUnit *CUNode : MMI->getModule()->debug_compile_units()) { 1382 auto *CU = CUMap.lookup(CUNode); 1383 1384 // If this CU hasn't been emitted yet, create it here unless it is empty. 1385 if (!CU) { 1386 // FIXME: Move local imported entities into a list attached to the 1387 // subprogram, then this search won't be needed and a 1388 // getImportedEntities().empty() test should go below with the rest. 1389 bool HasNonLocalImportedEntities = llvm::any_of( 1390 CUNode->getImportedEntities(), [](const DIImportedEntity *IE) { 1391 return !isa<DILocalScope>(IE->getScope()); 1392 }); 1393 1394 if (!HasNonLocalImportedEntities && CUNode->getEnumTypes().empty() && 1395 CUNode->getRetainedTypes().empty() && 1396 CUNode->getGlobalVariables().empty() && CUNode->getMacros().empty()) 1397 continue; 1398 1399 CU = &getOrCreateDwarfCompileUnit(CUNode); 1400 } 1401 1402 // Global Variables. 1403 for (auto *GVE : CUNode->getGlobalVariables()) { 1404 // Don't bother adding DIGlobalVariableExpressions listed in the CU if we 1405 // already know about the variable and it isn't adding a constant 1406 // expression. 1407 auto &GVMapEntry = GVMap[GVE->getVariable()]; 1408 auto *Expr = GVE->getExpression(); 1409 if (!GVMapEntry.size() || (Expr && Expr->isConstant())) 1410 GVMapEntry.push_back({nullptr, Expr}); 1411 } 1412 1413 DenseSet<DIGlobalVariable *> Processed; 1414 for (auto *GVE : CUNode->getGlobalVariables()) { 1415 DIGlobalVariable *GV = GVE->getVariable(); 1416 if (Processed.insert(GV).second) 1417 CU->getOrCreateGlobalVariableDIE(GV, sortGlobalExprs(GVMap[GV])); 1418 } 1419 1420 for (auto *Ty : CUNode->getEnumTypes()) 1421 createMaybeUnusedType(*this, *CU, *Ty); 1422 1423 for (auto *Ty : CUNode->getRetainedTypes()) { 1424 if (DIType *RT = dyn_cast<DIType>(Ty)) 1425 createMaybeUnusedType(*this, *CU, *RT); 1426 } 1427 1428 // Emit imported entities last so that the relevant context 1429 // is already available. 1430 for (auto *IE : CUNode->getImportedEntities()) 1431 constructAndAddImportedEntityDIE(*CU, IE); 1432 1433 CU->createBaseTypeDIEs(); 1434 } 1435 1436 // If we aren't actually generating debug info (check beginModule - 1437 // conditionalized on the presence of the llvm.dbg.cu metadata node) 1438 if (!Asm || !MMI->hasDebugInfo()) 1439 return; 1440 1441 // Finalize the debug info for the module. 1442 finalizeModuleInfo(); 1443 1444 if (useSplitDwarf()) 1445 // Emit debug_loc.dwo/debug_loclists.dwo section. 1446 emitDebugLocDWO(); 1447 else 1448 // Emit debug_loc/debug_loclists section. 1449 emitDebugLoc(); 1450 1451 // Corresponding abbreviations into a abbrev section. 1452 emitAbbreviations(); 1453 1454 // Emit all the DIEs into a debug info section. 1455 emitDebugInfo(); 1456 1457 // Emit info into a debug aranges section. 1458 if (GenerateARangeSection) 1459 emitDebugARanges(); 1460 1461 // Emit info into a debug ranges section. 1462 emitDebugRanges(); 1463 1464 if (useSplitDwarf()) 1465 // Emit info into a debug macinfo.dwo section. 1466 emitDebugMacinfoDWO(); 1467 else 1468 // Emit info into a debug macinfo/macro section. 1469 emitDebugMacinfo(); 1470 1471 emitDebugStr(); 1472 1473 if (useSplitDwarf()) { 1474 emitDebugStrDWO(); 1475 emitDebugInfoDWO(); 1476 emitDebugAbbrevDWO(); 1477 emitDebugLineDWO(); 1478 emitDebugRangesDWO(); 1479 } 1480 1481 emitDebugAddr(); 1482 1483 // Emit info into the dwarf accelerator table sections. 1484 switch (getAccelTableKind()) { 1485 case AccelTableKind::Apple: 1486 emitAccelNames(); 1487 emitAccelObjC(); 1488 emitAccelNamespaces(); 1489 emitAccelTypes(); 1490 break; 1491 case AccelTableKind::Dwarf: 1492 emitAccelDebugNames(); 1493 break; 1494 case AccelTableKind::None: 1495 break; 1496 case AccelTableKind::Default: 1497 llvm_unreachable("Default should have already been resolved."); 1498 } 1499 1500 // Emit the pubnames and pubtypes sections if requested. 1501 emitDebugPubSections(); 1502 1503 // clean up. 1504 // FIXME: AbstractVariables.clear(); 1505 } 1506 1507 void DwarfDebug::ensureAbstractEntityIsCreated(DwarfCompileUnit &CU, 1508 const DINode *Node, 1509 const MDNode *ScopeNode) { 1510 if (CU.getExistingAbstractEntity(Node)) 1511 return; 1512 1513 CU.createAbstractEntity(Node, LScopes.getOrCreateAbstractScope( 1514 cast<DILocalScope>(ScopeNode))); 1515 } 1516 1517 void DwarfDebug::ensureAbstractEntityIsCreatedIfScoped(DwarfCompileUnit &CU, 1518 const DINode *Node, const MDNode *ScopeNode) { 1519 if (CU.getExistingAbstractEntity(Node)) 1520 return; 1521 1522 if (LexicalScope *Scope = 1523 LScopes.findAbstractScope(cast_or_null<DILocalScope>(ScopeNode))) 1524 CU.createAbstractEntity(Node, Scope); 1525 } 1526 1527 // Collect variable information from side table maintained by MF. 1528 void DwarfDebug::collectVariableInfoFromMFTable( 1529 DwarfCompileUnit &TheCU, DenseSet<InlinedEntity> &Processed) { 1530 SmallDenseMap<InlinedEntity, DbgVariable *> MFVars; 1531 LLVM_DEBUG(dbgs() << "DwarfDebug: collecting variables from MF side table\n"); 1532 for (const auto &VI : Asm->MF->getVariableDbgInfo()) { 1533 if (!VI.Var) 1534 continue; 1535 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1536 "Expected inlined-at fields to agree"); 1537 1538 InlinedEntity Var(VI.Var, VI.Loc->getInlinedAt()); 1539 Processed.insert(Var); 1540 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1541 1542 // If variable scope is not found then skip this variable. 1543 if (!Scope) { 1544 LLVM_DEBUG(dbgs() << "Dropping debug info for " << VI.Var->getName() 1545 << ", no variable scope found\n"); 1546 continue; 1547 } 1548 1549 ensureAbstractEntityIsCreatedIfScoped(TheCU, Var.first, Scope->getScopeNode()); 1550 auto RegVar = std::make_unique<DbgVariable>( 1551 cast<DILocalVariable>(Var.first), Var.second); 1552 RegVar->initializeMMI(VI.Expr, VI.Slot); 1553 LLVM_DEBUG(dbgs() << "Created DbgVariable for " << VI.Var->getName() 1554 << "\n"); 1555 1556 if (DbgVariable *DbgVar = MFVars.lookup(Var)) 1557 DbgVar->addMMIEntry(*RegVar); 1558 else if (InfoHolder.addScopeVariable(Scope, RegVar.get())) { 1559 MFVars.insert({Var, RegVar.get()}); 1560 ConcreteEntities.push_back(std::move(RegVar)); 1561 } 1562 } 1563 } 1564 1565 /// Determine whether a *singular* DBG_VALUE is valid for the entirety of its 1566 /// enclosing lexical scope. The check ensures there are no other instructions 1567 /// in the same lexical scope preceding the DBG_VALUE and that its range is 1568 /// either open or otherwise rolls off the end of the scope. 1569 static bool validThroughout(LexicalScopes &LScopes, 1570 const MachineInstr *DbgValue, 1571 const MachineInstr *RangeEnd, 1572 const InstructionOrdering &Ordering) { 1573 assert(DbgValue->getDebugLoc() && "DBG_VALUE without a debug location"); 1574 auto MBB = DbgValue->getParent(); 1575 auto DL = DbgValue->getDebugLoc(); 1576 auto *LScope = LScopes.findLexicalScope(DL); 1577 // Scope doesn't exist; this is a dead DBG_VALUE. 1578 if (!LScope) 1579 return false; 1580 auto &LSRange = LScope->getRanges(); 1581 if (LSRange.size() == 0) 1582 return false; 1583 1584 const MachineInstr *LScopeBegin = LSRange.front().first; 1585 // If the scope starts before the DBG_VALUE then we may have a negative 1586 // result. Otherwise the location is live coming into the scope and we 1587 // can skip the following checks. 1588 if (!Ordering.isBefore(DbgValue, LScopeBegin)) { 1589 // Exit if the lexical scope begins outside of the current block. 1590 if (LScopeBegin->getParent() != MBB) 1591 return false; 1592 1593 MachineBasicBlock::const_reverse_iterator Pred(DbgValue); 1594 for (++Pred; Pred != MBB->rend(); ++Pred) { 1595 if (Pred->getFlag(MachineInstr::FrameSetup)) 1596 break; 1597 auto PredDL = Pred->getDebugLoc(); 1598 if (!PredDL || Pred->isMetaInstruction()) 1599 continue; 1600 // Check whether the instruction preceding the DBG_VALUE is in the same 1601 // (sub)scope as the DBG_VALUE. 1602 if (DL->getScope() == PredDL->getScope()) 1603 return false; 1604 auto *PredScope = LScopes.findLexicalScope(PredDL); 1605 if (!PredScope || LScope->dominates(PredScope)) 1606 return false; 1607 } 1608 } 1609 1610 // If the range of the DBG_VALUE is open-ended, report success. 1611 if (!RangeEnd) 1612 return true; 1613 1614 // Single, constant DBG_VALUEs in the prologue are promoted to be live 1615 // throughout the function. This is a hack, presumably for DWARF v2 and not 1616 // necessarily correct. It would be much better to use a dbg.declare instead 1617 // if we know the constant is live throughout the scope. 1618 if (MBB->pred_empty() && 1619 all_of(DbgValue->debug_operands(), 1620 [](const MachineOperand &Op) { return Op.isImm(); })) 1621 return true; 1622 1623 // Test if the location terminates before the end of the scope. 1624 const MachineInstr *LScopeEnd = LSRange.back().second; 1625 if (Ordering.isBefore(RangeEnd, LScopeEnd)) 1626 return false; 1627 1628 // There's a single location which starts at the scope start, and ends at or 1629 // after the scope end. 1630 return true; 1631 } 1632 1633 /// Build the location list for all DBG_VALUEs in the function that 1634 /// describe the same variable. The resulting DebugLocEntries will have 1635 /// strict monotonically increasing begin addresses and will never 1636 /// overlap. If the resulting list has only one entry that is valid 1637 /// throughout variable's scope return true. 1638 // 1639 // See the definition of DbgValueHistoryMap::Entry for an explanation of the 1640 // different kinds of history map entries. One thing to be aware of is that if 1641 // a debug value is ended by another entry (rather than being valid until the 1642 // end of the function), that entry's instruction may or may not be included in 1643 // the range, depending on if the entry is a clobbering entry (it has an 1644 // instruction that clobbers one or more preceding locations), or if it is an 1645 // (overlapping) debug value entry. This distinction can be seen in the example 1646 // below. The first debug value is ended by the clobbering entry 2, and the 1647 // second and third debug values are ended by the overlapping debug value entry 1648 // 4. 1649 // 1650 // Input: 1651 // 1652 // History map entries [type, end index, mi] 1653 // 1654 // 0 | [DbgValue, 2, DBG_VALUE $reg0, [...] (fragment 0, 32)] 1655 // 1 | | [DbgValue, 4, DBG_VALUE $reg1, [...] (fragment 32, 32)] 1656 // 2 | | [Clobber, $reg0 = [...], -, -] 1657 // 3 | | [DbgValue, 4, DBG_VALUE 123, [...] (fragment 64, 32)] 1658 // 4 [DbgValue, ~0, DBG_VALUE @g, [...] (fragment 0, 96)] 1659 // 1660 // Output [start, end) [Value...]: 1661 // 1662 // [0-1) [(reg0, fragment 0, 32)] 1663 // [1-3) [(reg0, fragment 0, 32), (reg1, fragment 32, 32)] 1664 // [3-4) [(reg1, fragment 32, 32), (123, fragment 64, 32)] 1665 // [4-) [(@g, fragment 0, 96)] 1666 bool DwarfDebug::buildLocationList(SmallVectorImpl<DebugLocEntry> &DebugLoc, 1667 const DbgValueHistoryMap::Entries &Entries) { 1668 using OpenRange = 1669 std::pair<DbgValueHistoryMap::EntryIndex, DbgValueLoc>; 1670 SmallVector<OpenRange, 4> OpenRanges; 1671 bool isSafeForSingleLocation = true; 1672 const MachineInstr *StartDebugMI = nullptr; 1673 const MachineInstr *EndMI = nullptr; 1674 1675 for (auto EB = Entries.begin(), EI = EB, EE = Entries.end(); EI != EE; ++EI) { 1676 const MachineInstr *Instr = EI->getInstr(); 1677 1678 // Remove all values that are no longer live. 1679 size_t Index = std::distance(EB, EI); 1680 erase_if(OpenRanges, [&](OpenRange &R) { return R.first <= Index; }); 1681 1682 // If we are dealing with a clobbering entry, this iteration will result in 1683 // a location list entry starting after the clobbering instruction. 1684 const MCSymbol *StartLabel = 1685 EI->isClobber() ? getLabelAfterInsn(Instr) : getLabelBeforeInsn(Instr); 1686 assert(StartLabel && 1687 "Forgot label before/after instruction starting a range!"); 1688 1689 const MCSymbol *EndLabel; 1690 if (std::next(EI) == Entries.end()) { 1691 const MachineBasicBlock &EndMBB = Asm->MF->back(); 1692 EndLabel = Asm->MBBSectionRanges[EndMBB.getSectionIDNum()].EndLabel; 1693 if (EI->isClobber()) 1694 EndMI = EI->getInstr(); 1695 } 1696 else if (std::next(EI)->isClobber()) 1697 EndLabel = getLabelAfterInsn(std::next(EI)->getInstr()); 1698 else 1699 EndLabel = getLabelBeforeInsn(std::next(EI)->getInstr()); 1700 assert(EndLabel && "Forgot label after instruction ending a range!"); 1701 1702 if (EI->isDbgValue()) 1703 LLVM_DEBUG(dbgs() << "DotDebugLoc: " << *Instr << "\n"); 1704 1705 // If this history map entry has a debug value, add that to the list of 1706 // open ranges and check if its location is valid for a single value 1707 // location. 1708 if (EI->isDbgValue()) { 1709 // Do not add undef debug values, as they are redundant information in 1710 // the location list entries. An undef debug results in an empty location 1711 // description. If there are any non-undef fragments then padding pieces 1712 // with empty location descriptions will automatically be inserted, and if 1713 // all fragments are undef then the whole location list entry is 1714 // redundant. 1715 if (!Instr->isUndefDebugValue()) { 1716 auto Value = getDebugLocValue(Instr); 1717 OpenRanges.emplace_back(EI->getEndIndex(), Value); 1718 1719 // TODO: Add support for single value fragment locations. 1720 if (Instr->getDebugExpression()->isFragment()) 1721 isSafeForSingleLocation = false; 1722 1723 if (!StartDebugMI) 1724 StartDebugMI = Instr; 1725 } else { 1726 isSafeForSingleLocation = false; 1727 } 1728 } 1729 1730 // Location list entries with empty location descriptions are redundant 1731 // information in DWARF, so do not emit those. 1732 if (OpenRanges.empty()) 1733 continue; 1734 1735 // Omit entries with empty ranges as they do not have any effect in DWARF. 1736 if (StartLabel == EndLabel) { 1737 LLVM_DEBUG(dbgs() << "Omitting location list entry with empty range.\n"); 1738 continue; 1739 } 1740 1741 SmallVector<DbgValueLoc, 4> Values; 1742 for (auto &R : OpenRanges) 1743 Values.push_back(R.second); 1744 1745 // With Basic block sections, it is posssible that the StartLabel and the 1746 // Instr are not in the same section. This happens when the StartLabel is 1747 // the function begin label and the dbg value appears in a basic block 1748 // that is not the entry. In this case, the range needs to be split to 1749 // span each individual section in the range from StartLabel to EndLabel. 1750 if (Asm->MF->hasBBSections() && StartLabel == Asm->getFunctionBegin() && 1751 !Instr->getParent()->sameSection(&Asm->MF->front())) { 1752 const MCSymbol *BeginSectionLabel = StartLabel; 1753 1754 for (const MachineBasicBlock &MBB : *Asm->MF) { 1755 if (MBB.isBeginSection() && &MBB != &Asm->MF->front()) 1756 BeginSectionLabel = MBB.getSymbol(); 1757 1758 if (MBB.sameSection(Instr->getParent())) { 1759 DebugLoc.emplace_back(BeginSectionLabel, EndLabel, Values); 1760 break; 1761 } 1762 if (MBB.isEndSection()) 1763 DebugLoc.emplace_back(BeginSectionLabel, MBB.getEndSymbol(), Values); 1764 } 1765 } else { 1766 DebugLoc.emplace_back(StartLabel, EndLabel, Values); 1767 } 1768 1769 // Attempt to coalesce the ranges of two otherwise identical 1770 // DebugLocEntries. 1771 auto CurEntry = DebugLoc.rbegin(); 1772 LLVM_DEBUG({ 1773 dbgs() << CurEntry->getValues().size() << " Values:\n"; 1774 for (auto &Value : CurEntry->getValues()) 1775 Value.dump(); 1776 dbgs() << "-----\n"; 1777 }); 1778 1779 auto PrevEntry = std::next(CurEntry); 1780 if (PrevEntry != DebugLoc.rend() && PrevEntry->MergeRanges(*CurEntry)) 1781 DebugLoc.pop_back(); 1782 } 1783 1784 if (!isSafeForSingleLocation || 1785 !validThroughout(LScopes, StartDebugMI, EndMI, getInstOrdering())) 1786 return false; 1787 1788 if (DebugLoc.size() == 1) 1789 return true; 1790 1791 if (!Asm->MF->hasBBSections()) 1792 return false; 1793 1794 // Check here to see if loclist can be merged into a single range. If not, 1795 // we must keep the split loclists per section. This does exactly what 1796 // MergeRanges does without sections. We don't actually merge the ranges 1797 // as the split ranges must be kept intact if this cannot be collapsed 1798 // into a single range. 1799 const MachineBasicBlock *RangeMBB = nullptr; 1800 if (DebugLoc[0].getBeginSym() == Asm->getFunctionBegin()) 1801 RangeMBB = &Asm->MF->front(); 1802 else 1803 RangeMBB = Entries.begin()->getInstr()->getParent(); 1804 auto *CurEntry = DebugLoc.begin(); 1805 auto *NextEntry = std::next(CurEntry); 1806 while (NextEntry != DebugLoc.end()) { 1807 // Get the last machine basic block of this section. 1808 while (!RangeMBB->isEndSection()) 1809 RangeMBB = RangeMBB->getNextNode(); 1810 if (!RangeMBB->getNextNode()) 1811 return false; 1812 // CurEntry should end the current section and NextEntry should start 1813 // the next section and the Values must match for these two ranges to be 1814 // merged. 1815 if (CurEntry->getEndSym() != RangeMBB->getEndSymbol() || 1816 NextEntry->getBeginSym() != RangeMBB->getNextNode()->getSymbol() || 1817 CurEntry->getValues() != NextEntry->getValues()) 1818 return false; 1819 RangeMBB = RangeMBB->getNextNode(); 1820 CurEntry = NextEntry; 1821 NextEntry = std::next(CurEntry); 1822 } 1823 return true; 1824 } 1825 1826 DbgEntity *DwarfDebug::createConcreteEntity(DwarfCompileUnit &TheCU, 1827 LexicalScope &Scope, 1828 const DINode *Node, 1829 const DILocation *Location, 1830 const MCSymbol *Sym) { 1831 ensureAbstractEntityIsCreatedIfScoped(TheCU, Node, Scope.getScopeNode()); 1832 if (isa<const DILocalVariable>(Node)) { 1833 ConcreteEntities.push_back( 1834 std::make_unique<DbgVariable>(cast<const DILocalVariable>(Node), 1835 Location)); 1836 InfoHolder.addScopeVariable(&Scope, 1837 cast<DbgVariable>(ConcreteEntities.back().get())); 1838 } else if (isa<const DILabel>(Node)) { 1839 ConcreteEntities.push_back( 1840 std::make_unique<DbgLabel>(cast<const DILabel>(Node), 1841 Location, Sym)); 1842 InfoHolder.addScopeLabel(&Scope, 1843 cast<DbgLabel>(ConcreteEntities.back().get())); 1844 } 1845 return ConcreteEntities.back().get(); 1846 } 1847 1848 // Find variables for each lexical scope. 1849 void DwarfDebug::collectEntityInfo(DwarfCompileUnit &TheCU, 1850 const DISubprogram *SP, 1851 DenseSet<InlinedEntity> &Processed) { 1852 // Grab the variable info that was squirreled away in the MMI side-table. 1853 collectVariableInfoFromMFTable(TheCU, Processed); 1854 1855 for (const auto &I : DbgValues) { 1856 InlinedEntity IV = I.first; 1857 if (Processed.count(IV)) 1858 continue; 1859 1860 // Instruction ranges, specifying where IV is accessible. 1861 const auto &HistoryMapEntries = I.second; 1862 1863 // Try to find any non-empty variable location. Do not create a concrete 1864 // entity if there are no locations. 1865 if (!DbgValues.hasNonEmptyLocation(HistoryMapEntries)) 1866 continue; 1867 1868 LexicalScope *Scope = nullptr; 1869 const DILocalVariable *LocalVar = cast<DILocalVariable>(IV.first); 1870 if (const DILocation *IA = IV.second) 1871 Scope = LScopes.findInlinedScope(LocalVar->getScope(), IA); 1872 else 1873 Scope = LScopes.findLexicalScope(LocalVar->getScope()); 1874 // If variable scope is not found then skip this variable. 1875 if (!Scope) 1876 continue; 1877 1878 Processed.insert(IV); 1879 DbgVariable *RegVar = cast<DbgVariable>(createConcreteEntity(TheCU, 1880 *Scope, LocalVar, IV.second)); 1881 1882 const MachineInstr *MInsn = HistoryMapEntries.front().getInstr(); 1883 assert(MInsn->isDebugValue() && "History must begin with debug value"); 1884 1885 // Check if there is a single DBG_VALUE, valid throughout the var's scope. 1886 // If the history map contains a single debug value, there may be an 1887 // additional entry which clobbers the debug value. 1888 size_t HistSize = HistoryMapEntries.size(); 1889 bool SingleValueWithClobber = 1890 HistSize == 2 && HistoryMapEntries[1].isClobber(); 1891 if (HistSize == 1 || SingleValueWithClobber) { 1892 const auto *End = 1893 SingleValueWithClobber ? HistoryMapEntries[1].getInstr() : nullptr; 1894 if (validThroughout(LScopes, MInsn, End, getInstOrdering())) { 1895 RegVar->initializeDbgValue(MInsn); 1896 continue; 1897 } 1898 } 1899 1900 // Do not emit location lists if .debug_loc secton is disabled. 1901 if (!useLocSection()) 1902 continue; 1903 1904 // Handle multiple DBG_VALUE instructions describing one variable. 1905 DebugLocStream::ListBuilder List(DebugLocs, TheCU, *Asm, *RegVar, *MInsn); 1906 1907 // Build the location list for this variable. 1908 SmallVector<DebugLocEntry, 8> Entries; 1909 bool isValidSingleLocation = buildLocationList(Entries, HistoryMapEntries); 1910 1911 // Check whether buildLocationList managed to merge all locations to one 1912 // that is valid throughout the variable's scope. If so, produce single 1913 // value location. 1914 if (isValidSingleLocation) { 1915 RegVar->initializeDbgValue(Entries[0].getValues()[0]); 1916 continue; 1917 } 1918 1919 // If the variable has a DIBasicType, extract it. Basic types cannot have 1920 // unique identifiers, so don't bother resolving the type with the 1921 // identifier map. 1922 const DIBasicType *BT = dyn_cast<DIBasicType>( 1923 static_cast<const Metadata *>(LocalVar->getType())); 1924 1925 // Finalize the entry by lowering it into a DWARF bytestream. 1926 for (auto &Entry : Entries) 1927 Entry.finalize(*Asm, List, BT, TheCU); 1928 } 1929 1930 // For each InlinedEntity collected from DBG_LABEL instructions, convert to 1931 // DWARF-related DbgLabel. 1932 for (const auto &I : DbgLabels) { 1933 InlinedEntity IL = I.first; 1934 const MachineInstr *MI = I.second; 1935 if (MI == nullptr) 1936 continue; 1937 1938 LexicalScope *Scope = nullptr; 1939 const DILabel *Label = cast<DILabel>(IL.first); 1940 // The scope could have an extra lexical block file. 1941 const DILocalScope *LocalScope = 1942 Label->getScope()->getNonLexicalBlockFileScope(); 1943 // Get inlined DILocation if it is inlined label. 1944 if (const DILocation *IA = IL.second) 1945 Scope = LScopes.findInlinedScope(LocalScope, IA); 1946 else 1947 Scope = LScopes.findLexicalScope(LocalScope); 1948 // If label scope is not found then skip this label. 1949 if (!Scope) 1950 continue; 1951 1952 Processed.insert(IL); 1953 /// At this point, the temporary label is created. 1954 /// Save the temporary label to DbgLabel entity to get the 1955 /// actually address when generating Dwarf DIE. 1956 MCSymbol *Sym = getLabelBeforeInsn(MI); 1957 createConcreteEntity(TheCU, *Scope, Label, IL.second, Sym); 1958 } 1959 1960 // Collect info for variables/labels that were optimized out. 1961 for (const DINode *DN : SP->getRetainedNodes()) { 1962 if (!Processed.insert(InlinedEntity(DN, nullptr)).second) 1963 continue; 1964 LexicalScope *Scope = nullptr; 1965 if (auto *DV = dyn_cast<DILocalVariable>(DN)) { 1966 Scope = LScopes.findLexicalScope(DV->getScope()); 1967 } else if (auto *DL = dyn_cast<DILabel>(DN)) { 1968 Scope = LScopes.findLexicalScope(DL->getScope()); 1969 } 1970 1971 if (Scope) 1972 createConcreteEntity(TheCU, *Scope, DN, nullptr); 1973 } 1974 } 1975 1976 // Process beginning of an instruction. 1977 void DwarfDebug::beginInstruction(const MachineInstr *MI) { 1978 const MachineFunction &MF = *MI->getMF(); 1979 const auto *SP = MF.getFunction().getSubprogram(); 1980 bool NoDebug = 1981 !SP || SP->getUnit()->getEmissionKind() == DICompileUnit::NoDebug; 1982 1983 // Delay slot support check. 1984 auto delaySlotSupported = [](const MachineInstr &MI) { 1985 if (!MI.isBundledWithSucc()) 1986 return false; 1987 auto Suc = std::next(MI.getIterator()); 1988 (void)Suc; 1989 // Ensure that delay slot instruction is successor of the call instruction. 1990 // Ex. CALL_INSTRUCTION { 1991 // DELAY_SLOT_INSTRUCTION } 1992 assert(Suc->isBundledWithPred() && 1993 "Call bundle instructions are out of order"); 1994 return true; 1995 }; 1996 1997 // When describing calls, we need a label for the call instruction. 1998 if (!NoDebug && SP->areAllCallsDescribed() && 1999 MI->isCandidateForCallSiteEntry(MachineInstr::AnyInBundle) && 2000 (!MI->hasDelaySlot() || delaySlotSupported(*MI))) { 2001 const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo(); 2002 bool IsTail = TII->isTailCall(*MI); 2003 // For tail calls, we need the address of the branch instruction for 2004 // DW_AT_call_pc. 2005 if (IsTail) 2006 requestLabelBeforeInsn(MI); 2007 // For non-tail calls, we need the return address for the call for 2008 // DW_AT_call_return_pc. Under GDB tuning, this information is needed for 2009 // tail calls as well. 2010 requestLabelAfterInsn(MI); 2011 } 2012 2013 DebugHandlerBase::beginInstruction(MI); 2014 if (!CurMI) 2015 return; 2016 2017 if (NoDebug) 2018 return; 2019 2020 // Check if source location changes, but ignore DBG_VALUE and CFI locations. 2021 // If the instruction is part of the function frame setup code, do not emit 2022 // any line record, as there is no correspondence with any user code. 2023 if (MI->isMetaInstruction() || MI->getFlag(MachineInstr::FrameSetup)) 2024 return; 2025 const DebugLoc &DL = MI->getDebugLoc(); 2026 // When we emit a line-0 record, we don't update PrevInstLoc; so look at 2027 // the last line number actually emitted, to see if it was line 0. 2028 unsigned LastAsmLine = 2029 Asm->OutStreamer->getContext().getCurrentDwarfLoc().getLine(); 2030 2031 if (DL == PrevInstLoc) { 2032 // If we have an ongoing unspecified location, nothing to do here. 2033 if (!DL) 2034 return; 2035 // We have an explicit location, same as the previous location. 2036 // But we might be coming back to it after a line 0 record. 2037 if (LastAsmLine == 0 && DL.getLine() != 0) { 2038 // Reinstate the source location but not marked as a statement. 2039 const MDNode *Scope = DL.getScope(); 2040 recordSourceLine(DL.getLine(), DL.getCol(), Scope, /*Flags=*/0); 2041 } 2042 return; 2043 } 2044 2045 if (!DL) { 2046 // We have an unspecified location, which might want to be line 0. 2047 // If we have already emitted a line-0 record, don't repeat it. 2048 if (LastAsmLine == 0) 2049 return; 2050 // If user said Don't Do That, don't do that. 2051 if (UnknownLocations == Disable) 2052 return; 2053 // See if we have a reason to emit a line-0 record now. 2054 // Reasons to emit a line-0 record include: 2055 // - User asked for it (UnknownLocations). 2056 // - Instruction has a label, so it's referenced from somewhere else, 2057 // possibly debug information; we want it to have a source location. 2058 // - Instruction is at the top of a block; we don't want to inherit the 2059 // location from the physically previous (maybe unrelated) block. 2060 if (UnknownLocations == Enable || PrevLabel || 2061 (PrevInstBB && PrevInstBB != MI->getParent())) { 2062 // Preserve the file and column numbers, if we can, to save space in 2063 // the encoded line table. 2064 // Do not update PrevInstLoc, it remembers the last non-0 line. 2065 const MDNode *Scope = nullptr; 2066 unsigned Column = 0; 2067 if (PrevInstLoc) { 2068 Scope = PrevInstLoc.getScope(); 2069 Column = PrevInstLoc.getCol(); 2070 } 2071 recordSourceLine(/*Line=*/0, Column, Scope, /*Flags=*/0); 2072 } 2073 return; 2074 } 2075 2076 // We have an explicit location, different from the previous location. 2077 // Don't repeat a line-0 record, but otherwise emit the new location. 2078 // (The new location might be an explicit line 0, which we do emit.) 2079 if (DL.getLine() == 0 && LastAsmLine == 0) 2080 return; 2081 unsigned Flags = 0; 2082 if (DL == PrologEndLoc) { 2083 Flags |= DWARF2_FLAG_PROLOGUE_END | DWARF2_FLAG_IS_STMT; 2084 PrologEndLoc = DebugLoc(); 2085 } 2086 // If the line changed, we call that a new statement; unless we went to 2087 // line 0 and came back, in which case it is not a new statement. 2088 unsigned OldLine = PrevInstLoc ? PrevInstLoc.getLine() : LastAsmLine; 2089 if (DL.getLine() && DL.getLine() != OldLine) 2090 Flags |= DWARF2_FLAG_IS_STMT; 2091 2092 const MDNode *Scope = DL.getScope(); 2093 recordSourceLine(DL.getLine(), DL.getCol(), Scope, Flags); 2094 2095 // If we're not at line 0, remember this location. 2096 if (DL.getLine()) 2097 PrevInstLoc = DL; 2098 } 2099 2100 static DebugLoc findPrologueEndLoc(const MachineFunction *MF) { 2101 // First known non-DBG_VALUE and non-frame setup location marks 2102 // the beginning of the function body. 2103 DebugLoc LineZeroLoc; 2104 for (const auto &MBB : *MF) { 2105 for (const auto &MI : MBB) { 2106 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 2107 MI.getDebugLoc()) { 2108 // Scan forward to try to find a non-zero line number. The prologue_end 2109 // marks the first breakpoint in the function after the frame setup, and 2110 // a compiler-generated line 0 location is not a meaningful breakpoint. 2111 // If none is found, return the first location after the frame setup. 2112 if (MI.getDebugLoc().getLine()) 2113 return MI.getDebugLoc(); 2114 LineZeroLoc = MI.getDebugLoc(); 2115 } 2116 } 2117 } 2118 return LineZeroLoc; 2119 } 2120 2121 /// Register a source line with debug info. Returns the unique label that was 2122 /// emitted and which provides correspondence to the source line list. 2123 static void recordSourceLine(AsmPrinter &Asm, unsigned Line, unsigned Col, 2124 const MDNode *S, unsigned Flags, unsigned CUID, 2125 uint16_t DwarfVersion, 2126 ArrayRef<std::unique_ptr<DwarfCompileUnit>> DCUs) { 2127 StringRef Fn; 2128 unsigned FileNo = 1; 2129 unsigned Discriminator = 0; 2130 if (auto *Scope = cast_or_null<DIScope>(S)) { 2131 Fn = Scope->getFilename(); 2132 if (Line != 0 && DwarfVersion >= 4) 2133 if (auto *LBF = dyn_cast<DILexicalBlockFile>(Scope)) 2134 Discriminator = LBF->getDiscriminator(); 2135 2136 FileNo = static_cast<DwarfCompileUnit &>(*DCUs[CUID]) 2137 .getOrCreateSourceID(Scope->getFile()); 2138 } 2139 Asm.OutStreamer->emitDwarfLocDirective(FileNo, Line, Col, Flags, 0, 2140 Discriminator, Fn); 2141 } 2142 2143 DebugLoc DwarfDebug::emitInitialLocDirective(const MachineFunction &MF, 2144 unsigned CUID) { 2145 // Get beginning of function. 2146 if (DebugLoc PrologEndLoc = findPrologueEndLoc(&MF)) { 2147 // Ensure the compile unit is created if the function is called before 2148 // beginFunction(). 2149 (void)getOrCreateDwarfCompileUnit( 2150 MF.getFunction().getSubprogram()->getUnit()); 2151 // We'd like to list the prologue as "not statements" but GDB behaves 2152 // poorly if we do that. Revisit this with caution/GDB (7.5+) testing. 2153 const DISubprogram *SP = PrologEndLoc->getInlinedAtScope()->getSubprogram(); 2154 ::recordSourceLine(*Asm, SP->getScopeLine(), 0, SP, DWARF2_FLAG_IS_STMT, 2155 CUID, getDwarfVersion(), getUnits()); 2156 return PrologEndLoc; 2157 } 2158 return DebugLoc(); 2159 } 2160 2161 // Gather pre-function debug information. Assumes being called immediately 2162 // after the function entry point has been emitted. 2163 void DwarfDebug::beginFunctionImpl(const MachineFunction *MF) { 2164 CurFn = MF; 2165 2166 auto *SP = MF->getFunction().getSubprogram(); 2167 assert(LScopes.empty() || SP == LScopes.getCurrentFunctionScope()->getScopeNode()); 2168 if (SP->getUnit()->getEmissionKind() == DICompileUnit::NoDebug) 2169 return; 2170 2171 DwarfCompileUnit &CU = getOrCreateDwarfCompileUnit(SP->getUnit()); 2172 2173 Asm->OutStreamer->getContext().setDwarfCompileUnitID( 2174 getDwarfCompileUnitIDForLineTable(CU)); 2175 2176 // Record beginning of function. 2177 PrologEndLoc = emitInitialLocDirective( 2178 *MF, Asm->OutStreamer->getContext().getDwarfCompileUnitID()); 2179 } 2180 2181 unsigned 2182 DwarfDebug::getDwarfCompileUnitIDForLineTable(const DwarfCompileUnit &CU) { 2183 // Set DwarfDwarfCompileUnitID in MCContext to the Compile Unit this function 2184 // belongs to so that we add to the correct per-cu line table in the 2185 // non-asm case. 2186 if (Asm->OutStreamer->hasRawTextSupport()) 2187 // Use a single line table if we are generating assembly. 2188 return 0; 2189 else 2190 return CU.getUniqueID(); 2191 } 2192 2193 void DwarfDebug::terminateLineTable(const DwarfCompileUnit *CU) { 2194 const auto &CURanges = CU->getRanges(); 2195 auto &LineTable = Asm->OutStreamer->getContext().getMCDwarfLineTable( 2196 getDwarfCompileUnitIDForLineTable(*CU)); 2197 // Add the last range label for the given CU. 2198 LineTable.getMCLineSections().addEndEntry( 2199 const_cast<MCSymbol *>(CURanges.back().End)); 2200 } 2201 2202 void DwarfDebug::skippedNonDebugFunction() { 2203 // If we don't have a subprogram for this function then there will be a hole 2204 // in the range information. Keep note of this by setting the previously used 2205 // section to nullptr. 2206 // Terminate the pending line table. 2207 if (PrevCU) 2208 terminateLineTable(PrevCU); 2209 PrevCU = nullptr; 2210 CurFn = nullptr; 2211 } 2212 2213 // Gather and emit post-function debug information. 2214 void DwarfDebug::endFunctionImpl(const MachineFunction *MF) { 2215 const DISubprogram *SP = MF->getFunction().getSubprogram(); 2216 2217 assert(CurFn == MF && 2218 "endFunction should be called with the same function as beginFunction"); 2219 2220 // Set DwarfDwarfCompileUnitID in MCContext to default value. 2221 Asm->OutStreamer->getContext().setDwarfCompileUnitID(0); 2222 2223 LexicalScope *FnScope = LScopes.getCurrentFunctionScope(); 2224 assert(!FnScope || SP == FnScope->getScopeNode()); 2225 DwarfCompileUnit &TheCU = *CUMap.lookup(SP->getUnit()); 2226 if (TheCU.getCUNode()->isDebugDirectivesOnly()) { 2227 PrevLabel = nullptr; 2228 CurFn = nullptr; 2229 return; 2230 } 2231 2232 DenseSet<InlinedEntity> Processed; 2233 collectEntityInfo(TheCU, SP, Processed); 2234 2235 // Add the range of this function to the list of ranges for the CU. 2236 // With basic block sections, add ranges for all basic block sections. 2237 for (const auto &R : Asm->MBBSectionRanges) 2238 TheCU.addRange({R.second.BeginLabel, R.second.EndLabel}); 2239 2240 // Under -gmlt, skip building the subprogram if there are no inlined 2241 // subroutines inside it. But with -fdebug-info-for-profiling, the subprogram 2242 // is still needed as we need its source location. 2243 if (!TheCU.getCUNode()->getDebugInfoForProfiling() && 2244 TheCU.getCUNode()->getEmissionKind() == DICompileUnit::LineTablesOnly && 2245 LScopes.getAbstractScopesList().empty() && !IsDarwin) { 2246 assert(InfoHolder.getScopeVariables().empty()); 2247 PrevLabel = nullptr; 2248 CurFn = nullptr; 2249 return; 2250 } 2251 2252 #ifndef NDEBUG 2253 size_t NumAbstractScopes = LScopes.getAbstractScopesList().size(); 2254 #endif 2255 // Construct abstract scopes. 2256 for (LexicalScope *AScope : LScopes.getAbstractScopesList()) { 2257 auto *SP = cast<DISubprogram>(AScope->getScopeNode()); 2258 for (const DINode *DN : SP->getRetainedNodes()) { 2259 if (!Processed.insert(InlinedEntity(DN, nullptr)).second) 2260 continue; 2261 2262 const MDNode *Scope = nullptr; 2263 if (auto *DV = dyn_cast<DILocalVariable>(DN)) 2264 Scope = DV->getScope(); 2265 else if (auto *DL = dyn_cast<DILabel>(DN)) 2266 Scope = DL->getScope(); 2267 else 2268 llvm_unreachable("Unexpected DI type!"); 2269 2270 // Collect info for variables/labels that were optimized out. 2271 ensureAbstractEntityIsCreated(TheCU, DN, Scope); 2272 assert(LScopes.getAbstractScopesList().size() == NumAbstractScopes 2273 && "ensureAbstractEntityIsCreated inserted abstract scopes"); 2274 } 2275 constructAbstractSubprogramScopeDIE(TheCU, AScope); 2276 } 2277 2278 ProcessedSPNodes.insert(SP); 2279 DIE &ScopeDIE = TheCU.constructSubprogramScopeDIE(SP, FnScope); 2280 if (auto *SkelCU = TheCU.getSkeleton()) 2281 if (!LScopes.getAbstractScopesList().empty() && 2282 TheCU.getCUNode()->getSplitDebugInlining()) 2283 SkelCU->constructSubprogramScopeDIE(SP, FnScope); 2284 2285 // Construct call site entries. 2286 constructCallSiteEntryDIEs(*SP, TheCU, ScopeDIE, *MF); 2287 2288 // Clear debug info 2289 // Ownership of DbgVariables is a bit subtle - ScopeVariables owns all the 2290 // DbgVariables except those that are also in AbstractVariables (since they 2291 // can be used cross-function) 2292 InfoHolder.getScopeVariables().clear(); 2293 InfoHolder.getScopeLabels().clear(); 2294 PrevLabel = nullptr; 2295 CurFn = nullptr; 2296 } 2297 2298 // Register a source line with debug info. Returns the unique label that was 2299 // emitted and which provides correspondence to the source line list. 2300 void DwarfDebug::recordSourceLine(unsigned Line, unsigned Col, const MDNode *S, 2301 unsigned Flags) { 2302 ::recordSourceLine(*Asm, Line, Col, S, Flags, 2303 Asm->OutStreamer->getContext().getDwarfCompileUnitID(), 2304 getDwarfVersion(), getUnits()); 2305 } 2306 2307 //===----------------------------------------------------------------------===// 2308 // Emit Methods 2309 //===----------------------------------------------------------------------===// 2310 2311 // Emit the debug info section. 2312 void DwarfDebug::emitDebugInfo() { 2313 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 2314 Holder.emitUnits(/* UseOffsets */ false); 2315 } 2316 2317 // Emit the abbreviation section. 2318 void DwarfDebug::emitAbbreviations() { 2319 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 2320 2321 Holder.emitAbbrevs(Asm->getObjFileLowering().getDwarfAbbrevSection()); 2322 } 2323 2324 void DwarfDebug::emitStringOffsetsTableHeader() { 2325 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 2326 Holder.getStringPool().emitStringOffsetsTableHeader( 2327 *Asm, Asm->getObjFileLowering().getDwarfStrOffSection(), 2328 Holder.getStringOffsetsStartSym()); 2329 } 2330 2331 template <typename AccelTableT> 2332 void DwarfDebug::emitAccel(AccelTableT &Accel, MCSection *Section, 2333 StringRef TableName) { 2334 Asm->OutStreamer->SwitchSection(Section); 2335 2336 // Emit the full data. 2337 emitAppleAccelTable(Asm, Accel, TableName, Section->getBeginSymbol()); 2338 } 2339 2340 void DwarfDebug::emitAccelDebugNames() { 2341 // Don't emit anything if we have no compilation units to index. 2342 if (getUnits().empty()) 2343 return; 2344 2345 emitDWARF5AccelTable(Asm, AccelDebugNames, *this, getUnits()); 2346 } 2347 2348 // Emit visible names into a hashed accelerator table section. 2349 void DwarfDebug::emitAccelNames() { 2350 emitAccel(AccelNames, Asm->getObjFileLowering().getDwarfAccelNamesSection(), 2351 "Names"); 2352 } 2353 2354 // Emit objective C classes and categories into a hashed accelerator table 2355 // section. 2356 void DwarfDebug::emitAccelObjC() { 2357 emitAccel(AccelObjC, Asm->getObjFileLowering().getDwarfAccelObjCSection(), 2358 "ObjC"); 2359 } 2360 2361 // Emit namespace dies into a hashed accelerator table. 2362 void DwarfDebug::emitAccelNamespaces() { 2363 emitAccel(AccelNamespace, 2364 Asm->getObjFileLowering().getDwarfAccelNamespaceSection(), 2365 "namespac"); 2366 } 2367 2368 // Emit type dies into a hashed accelerator table. 2369 void DwarfDebug::emitAccelTypes() { 2370 emitAccel(AccelTypes, Asm->getObjFileLowering().getDwarfAccelTypesSection(), 2371 "types"); 2372 } 2373 2374 // Public name handling. 2375 // The format for the various pubnames: 2376 // 2377 // dwarf pubnames - offset/name pairs where the offset is the offset into the CU 2378 // for the DIE that is named. 2379 // 2380 // gnu pubnames - offset/index value/name tuples where the offset is the offset 2381 // into the CU and the index value is computed according to the type of value 2382 // for the DIE that is named. 2383 // 2384 // For type units the offset is the offset of the skeleton DIE. For split dwarf 2385 // it's the offset within the debug_info/debug_types dwo section, however, the 2386 // reference in the pubname header doesn't change. 2387 2388 /// computeIndexValue - Compute the gdb index value for the DIE and CU. 2389 static dwarf::PubIndexEntryDescriptor computeIndexValue(DwarfUnit *CU, 2390 const DIE *Die) { 2391 // Entities that ended up only in a Type Unit reference the CU instead (since 2392 // the pub entry has offsets within the CU there's no real offset that can be 2393 // provided anyway). As it happens all such entities (namespaces and types, 2394 // types only in C++ at that) are rendered as TYPE+EXTERNAL. If this turns out 2395 // not to be true it would be necessary to persist this information from the 2396 // point at which the entry is added to the index data structure - since by 2397 // the time the index is built from that, the original type/namespace DIE in a 2398 // type unit has already been destroyed so it can't be queried for properties 2399 // like tag, etc. 2400 if (Die->getTag() == dwarf::DW_TAG_compile_unit) 2401 return dwarf::PubIndexEntryDescriptor(dwarf::GIEK_TYPE, 2402 dwarf::GIEL_EXTERNAL); 2403 dwarf::GDBIndexEntryLinkage Linkage = dwarf::GIEL_STATIC; 2404 2405 // We could have a specification DIE that has our most of our knowledge, 2406 // look for that now. 2407 if (DIEValue SpecVal = Die->findAttribute(dwarf::DW_AT_specification)) { 2408 DIE &SpecDIE = SpecVal.getDIEEntry().getEntry(); 2409 if (SpecDIE.findAttribute(dwarf::DW_AT_external)) 2410 Linkage = dwarf::GIEL_EXTERNAL; 2411 } else if (Die->findAttribute(dwarf::DW_AT_external)) 2412 Linkage = dwarf::GIEL_EXTERNAL; 2413 2414 switch (Die->getTag()) { 2415 case dwarf::DW_TAG_class_type: 2416 case dwarf::DW_TAG_structure_type: 2417 case dwarf::DW_TAG_union_type: 2418 case dwarf::DW_TAG_enumeration_type: 2419 return dwarf::PubIndexEntryDescriptor( 2420 dwarf::GIEK_TYPE, 2421 dwarf::isCPlusPlus((dwarf::SourceLanguage)CU->getLanguage()) 2422 ? dwarf::GIEL_EXTERNAL 2423 : dwarf::GIEL_STATIC); 2424 case dwarf::DW_TAG_typedef: 2425 case dwarf::DW_TAG_base_type: 2426 case dwarf::DW_TAG_subrange_type: 2427 return dwarf::PubIndexEntryDescriptor(dwarf::GIEK_TYPE, dwarf::GIEL_STATIC); 2428 case dwarf::DW_TAG_namespace: 2429 return dwarf::GIEK_TYPE; 2430 case dwarf::DW_TAG_subprogram: 2431 return dwarf::PubIndexEntryDescriptor(dwarf::GIEK_FUNCTION, Linkage); 2432 case dwarf::DW_TAG_variable: 2433 return dwarf::PubIndexEntryDescriptor(dwarf::GIEK_VARIABLE, Linkage); 2434 case dwarf::DW_TAG_enumerator: 2435 return dwarf::PubIndexEntryDescriptor(dwarf::GIEK_VARIABLE, 2436 dwarf::GIEL_STATIC); 2437 default: 2438 return dwarf::GIEK_NONE; 2439 } 2440 } 2441 2442 /// emitDebugPubSections - Emit visible names and types into debug pubnames and 2443 /// pubtypes sections. 2444 void DwarfDebug::emitDebugPubSections() { 2445 for (const auto &NU : CUMap) { 2446 DwarfCompileUnit *TheU = NU.second; 2447 if (!TheU->hasDwarfPubSections()) 2448 continue; 2449 2450 bool GnuStyle = TheU->getCUNode()->getNameTableKind() == 2451 DICompileUnit::DebugNameTableKind::GNU; 2452 2453 Asm->OutStreamer->SwitchSection( 2454 GnuStyle ? Asm->getObjFileLowering().getDwarfGnuPubNamesSection() 2455 : Asm->getObjFileLowering().getDwarfPubNamesSection()); 2456 emitDebugPubSection(GnuStyle, "Names", TheU, TheU->getGlobalNames()); 2457 2458 Asm->OutStreamer->SwitchSection( 2459 GnuStyle ? Asm->getObjFileLowering().getDwarfGnuPubTypesSection() 2460 : Asm->getObjFileLowering().getDwarfPubTypesSection()); 2461 emitDebugPubSection(GnuStyle, "Types", TheU, TheU->getGlobalTypes()); 2462 } 2463 } 2464 2465 void DwarfDebug::emitSectionReference(const DwarfCompileUnit &CU) { 2466 if (useSectionsAsReferences()) 2467 Asm->emitDwarfOffset(CU.getSection()->getBeginSymbol(), 2468 CU.getDebugSectionOffset()); 2469 else 2470 Asm->emitDwarfSymbolReference(CU.getLabelBegin()); 2471 } 2472 2473 void DwarfDebug::emitDebugPubSection(bool GnuStyle, StringRef Name, 2474 DwarfCompileUnit *TheU, 2475 const StringMap<const DIE *> &Globals) { 2476 if (auto *Skeleton = TheU->getSkeleton()) 2477 TheU = Skeleton; 2478 2479 // Emit the header. 2480 MCSymbol *EndLabel = Asm->emitDwarfUnitLength( 2481 "pub" + Name, "Length of Public " + Name + " Info"); 2482 2483 Asm->OutStreamer->AddComment("DWARF Version"); 2484 Asm->emitInt16(dwarf::DW_PUBNAMES_VERSION); 2485 2486 Asm->OutStreamer->AddComment("Offset of Compilation Unit Info"); 2487 emitSectionReference(*TheU); 2488 2489 Asm->OutStreamer->AddComment("Compilation Unit Length"); 2490 Asm->emitDwarfLengthOrOffset(TheU->getLength()); 2491 2492 // Emit the pubnames for this compilation unit. 2493 for (const auto &GI : Globals) { 2494 const char *Name = GI.getKeyData(); 2495 const DIE *Entity = GI.second; 2496 2497 Asm->OutStreamer->AddComment("DIE offset"); 2498 Asm->emitDwarfLengthOrOffset(Entity->getOffset()); 2499 2500 if (GnuStyle) { 2501 dwarf::PubIndexEntryDescriptor Desc = computeIndexValue(TheU, Entity); 2502 Asm->OutStreamer->AddComment( 2503 Twine("Attributes: ") + dwarf::GDBIndexEntryKindString(Desc.Kind) + 2504 ", " + dwarf::GDBIndexEntryLinkageString(Desc.Linkage)); 2505 Asm->emitInt8(Desc.toBits()); 2506 } 2507 2508 Asm->OutStreamer->AddComment("External Name"); 2509 Asm->OutStreamer->emitBytes(StringRef(Name, GI.getKeyLength() + 1)); 2510 } 2511 2512 Asm->OutStreamer->AddComment("End Mark"); 2513 Asm->emitDwarfLengthOrOffset(0); 2514 Asm->OutStreamer->emitLabel(EndLabel); 2515 } 2516 2517 /// Emit null-terminated strings into a debug str section. 2518 void DwarfDebug::emitDebugStr() { 2519 MCSection *StringOffsetsSection = nullptr; 2520 if (useSegmentedStringOffsetsTable()) { 2521 emitStringOffsetsTableHeader(); 2522 StringOffsetsSection = Asm->getObjFileLowering().getDwarfStrOffSection(); 2523 } 2524 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 2525 Holder.emitStrings(Asm->getObjFileLowering().getDwarfStrSection(), 2526 StringOffsetsSection, /* UseRelativeOffsets = */ true); 2527 } 2528 2529 void DwarfDebug::emitDebugLocEntry(ByteStreamer &Streamer, 2530 const DebugLocStream::Entry &Entry, 2531 const DwarfCompileUnit *CU) { 2532 auto &&Comments = DebugLocs.getComments(Entry); 2533 auto Comment = Comments.begin(); 2534 auto End = Comments.end(); 2535 2536 // The expressions are inserted into a byte stream rather early (see 2537 // DwarfExpression::addExpression) so for those ops (e.g. DW_OP_convert) that 2538 // need to reference a base_type DIE the offset of that DIE is not yet known. 2539 // To deal with this we instead insert a placeholder early and then extract 2540 // it here and replace it with the real reference. 2541 unsigned PtrSize = Asm->MAI->getCodePointerSize(); 2542 DWARFDataExtractor Data(StringRef(DebugLocs.getBytes(Entry).data(), 2543 DebugLocs.getBytes(Entry).size()), 2544 Asm->getDataLayout().isLittleEndian(), PtrSize); 2545 DWARFExpression Expr(Data, PtrSize, Asm->OutContext.getDwarfFormat()); 2546 2547 using Encoding = DWARFExpression::Operation::Encoding; 2548 uint64_t Offset = 0; 2549 for (auto &Op : Expr) { 2550 assert(Op.getCode() != dwarf::DW_OP_const_type && 2551 "3 operand ops not yet supported"); 2552 Streamer.emitInt8(Op.getCode(), Comment != End ? *(Comment++) : ""); 2553 Offset++; 2554 for (unsigned I = 0; I < 2; ++I) { 2555 if (Op.getDescription().Op[I] == Encoding::SizeNA) 2556 continue; 2557 if (Op.getDescription().Op[I] == Encoding::BaseTypeRef) { 2558 uint64_t Offset = 2559 CU->ExprRefedBaseTypes[Op.getRawOperand(I)].Die->getOffset(); 2560 assert(Offset < (1ULL << (ULEB128PadSize * 7)) && "Offset wont fit"); 2561 Streamer.emitULEB128(Offset, "", ULEB128PadSize); 2562 // Make sure comments stay aligned. 2563 for (unsigned J = 0; J < ULEB128PadSize; ++J) 2564 if (Comment != End) 2565 Comment++; 2566 } else { 2567 for (uint64_t J = Offset; J < Op.getOperandEndOffset(I); ++J) 2568 Streamer.emitInt8(Data.getData()[J], Comment != End ? *(Comment++) : ""); 2569 } 2570 Offset = Op.getOperandEndOffset(I); 2571 } 2572 assert(Offset == Op.getEndOffset()); 2573 } 2574 } 2575 2576 void DwarfDebug::emitDebugLocValue(const AsmPrinter &AP, const DIBasicType *BT, 2577 const DbgValueLoc &Value, 2578 DwarfExpression &DwarfExpr) { 2579 auto *DIExpr = Value.getExpression(); 2580 DIExpressionCursor ExprCursor(DIExpr); 2581 DwarfExpr.addFragmentOffset(DIExpr); 2582 2583 // If the DIExpr is is an Entry Value, we want to follow the same code path 2584 // regardless of whether the DBG_VALUE is variadic or not. 2585 if (DIExpr && DIExpr->isEntryValue()) { 2586 // Entry values can only be a single register with no additional DIExpr, 2587 // so just add it directly. 2588 assert(Value.getLocEntries().size() == 1); 2589 assert(Value.getLocEntries()[0].isLocation()); 2590 MachineLocation Location = Value.getLocEntries()[0].getLoc(); 2591 DwarfExpr.setLocation(Location, DIExpr); 2592 2593 DwarfExpr.beginEntryValueExpression(ExprCursor); 2594 2595 const TargetRegisterInfo &TRI = *AP.MF->getSubtarget().getRegisterInfo(); 2596 if (!DwarfExpr.addMachineRegExpression(TRI, ExprCursor, Location.getReg())) 2597 return; 2598 return DwarfExpr.addExpression(std::move(ExprCursor)); 2599 } 2600 2601 // Regular entry. 2602 auto EmitValueLocEntry = [&DwarfExpr, &BT, 2603 &AP](const DbgValueLocEntry &Entry, 2604 DIExpressionCursor &Cursor) -> bool { 2605 if (Entry.isInt()) { 2606 if (BT && (BT->getEncoding() == dwarf::DW_ATE_signed || 2607 BT->getEncoding() == dwarf::DW_ATE_signed_char)) 2608 DwarfExpr.addSignedConstant(Entry.getInt()); 2609 else 2610 DwarfExpr.addUnsignedConstant(Entry.getInt()); 2611 } else if (Entry.isLocation()) { 2612 MachineLocation Location = Entry.getLoc(); 2613 if (Location.isIndirect()) 2614 DwarfExpr.setMemoryLocationKind(); 2615 2616 const TargetRegisterInfo &TRI = *AP.MF->getSubtarget().getRegisterInfo(); 2617 if (!DwarfExpr.addMachineRegExpression(TRI, Cursor, Location.getReg())) 2618 return false; 2619 } else if (Entry.isTargetIndexLocation()) { 2620 TargetIndexLocation Loc = Entry.getTargetIndexLocation(); 2621 // TODO TargetIndexLocation is a target-independent. Currently only the 2622 // WebAssembly-specific encoding is supported. 2623 assert(AP.TM.getTargetTriple().isWasm()); 2624 DwarfExpr.addWasmLocation(Loc.Index, static_cast<uint64_t>(Loc.Offset)); 2625 } else if (Entry.isConstantFP()) { 2626 if (AP.getDwarfVersion() >= 4 && !AP.getDwarfDebug()->tuneForSCE() && 2627 !Cursor) { 2628 DwarfExpr.addConstantFP(Entry.getConstantFP()->getValueAPF(), AP); 2629 } else if (Entry.getConstantFP() 2630 ->getValueAPF() 2631 .bitcastToAPInt() 2632 .getBitWidth() <= 64 /*bits*/) { 2633 DwarfExpr.addUnsignedConstant( 2634 Entry.getConstantFP()->getValueAPF().bitcastToAPInt()); 2635 } else { 2636 LLVM_DEBUG( 2637 dbgs() << "Skipped DwarfExpression creation for ConstantFP of size" 2638 << Entry.getConstantFP() 2639 ->getValueAPF() 2640 .bitcastToAPInt() 2641 .getBitWidth() 2642 << " bits\n"); 2643 return false; 2644 } 2645 } 2646 return true; 2647 }; 2648 2649 if (!Value.isVariadic()) { 2650 if (!EmitValueLocEntry(Value.getLocEntries()[0], ExprCursor)) 2651 return; 2652 DwarfExpr.addExpression(std::move(ExprCursor)); 2653 return; 2654 } 2655 2656 // If any of the location entries are registers with the value 0, then the 2657 // location is undefined. 2658 if (any_of(Value.getLocEntries(), [](const DbgValueLocEntry &Entry) { 2659 return Entry.isLocation() && !Entry.getLoc().getReg(); 2660 })) 2661 return; 2662 2663 DwarfExpr.addExpression( 2664 std::move(ExprCursor), 2665 [EmitValueLocEntry, &Value](unsigned Idx, 2666 DIExpressionCursor &Cursor) -> bool { 2667 return EmitValueLocEntry(Value.getLocEntries()[Idx], Cursor); 2668 }); 2669 } 2670 2671 void DebugLocEntry::finalize(const AsmPrinter &AP, 2672 DebugLocStream::ListBuilder &List, 2673 const DIBasicType *BT, 2674 DwarfCompileUnit &TheCU) { 2675 assert(!Values.empty() && 2676 "location list entries without values are redundant"); 2677 assert(Begin != End && "unexpected location list entry with empty range"); 2678 DebugLocStream::EntryBuilder Entry(List, Begin, End); 2679 BufferByteStreamer Streamer = Entry.getStreamer(); 2680 DebugLocDwarfExpression DwarfExpr(AP.getDwarfVersion(), Streamer, TheCU); 2681 const DbgValueLoc &Value = Values[0]; 2682 if (Value.isFragment()) { 2683 // Emit all fragments that belong to the same variable and range. 2684 assert(llvm::all_of(Values, [](DbgValueLoc P) { 2685 return P.isFragment(); 2686 }) && "all values are expected to be fragments"); 2687 assert(llvm::is_sorted(Values) && "fragments are expected to be sorted"); 2688 2689 for (const auto &Fragment : Values) 2690 DwarfDebug::emitDebugLocValue(AP, BT, Fragment, DwarfExpr); 2691 2692 } else { 2693 assert(Values.size() == 1 && "only fragments may have >1 value"); 2694 DwarfDebug::emitDebugLocValue(AP, BT, Value, DwarfExpr); 2695 } 2696 DwarfExpr.finalize(); 2697 if (DwarfExpr.TagOffset) 2698 List.setTagOffset(*DwarfExpr.TagOffset); 2699 } 2700 2701 void DwarfDebug::emitDebugLocEntryLocation(const DebugLocStream::Entry &Entry, 2702 const DwarfCompileUnit *CU) { 2703 // Emit the size. 2704 Asm->OutStreamer->AddComment("Loc expr size"); 2705 if (getDwarfVersion() >= 5) 2706 Asm->emitULEB128(DebugLocs.getBytes(Entry).size()); 2707 else if (DebugLocs.getBytes(Entry).size() <= std::numeric_limits<uint16_t>::max()) 2708 Asm->emitInt16(DebugLocs.getBytes(Entry).size()); 2709 else { 2710 // The entry is too big to fit into 16 bit, drop it as there is nothing we 2711 // can do. 2712 Asm->emitInt16(0); 2713 return; 2714 } 2715 // Emit the entry. 2716 APByteStreamer Streamer(*Asm); 2717 emitDebugLocEntry(Streamer, Entry, CU); 2718 } 2719 2720 // Emit the header of a DWARF 5 range list table list table. Returns the symbol 2721 // that designates the end of the table for the caller to emit when the table is 2722 // complete. 2723 static MCSymbol *emitRnglistsTableHeader(AsmPrinter *Asm, 2724 const DwarfFile &Holder) { 2725 MCSymbol *TableEnd = mcdwarf::emitListsTableHeaderStart(*Asm->OutStreamer); 2726 2727 Asm->OutStreamer->AddComment("Offset entry count"); 2728 Asm->emitInt32(Holder.getRangeLists().size()); 2729 Asm->OutStreamer->emitLabel(Holder.getRnglistsTableBaseSym()); 2730 2731 for (const RangeSpanList &List : Holder.getRangeLists()) 2732 Asm->emitLabelDifference(List.Label, Holder.getRnglistsTableBaseSym(), 2733 Asm->getDwarfOffsetByteSize()); 2734 2735 return TableEnd; 2736 } 2737 2738 // Emit the header of a DWARF 5 locations list table. Returns the symbol that 2739 // designates the end of the table for the caller to emit when the table is 2740 // complete. 2741 static MCSymbol *emitLoclistsTableHeader(AsmPrinter *Asm, 2742 const DwarfDebug &DD) { 2743 MCSymbol *TableEnd = mcdwarf::emitListsTableHeaderStart(*Asm->OutStreamer); 2744 2745 const auto &DebugLocs = DD.getDebugLocs(); 2746 2747 Asm->OutStreamer->AddComment("Offset entry count"); 2748 Asm->emitInt32(DebugLocs.getLists().size()); 2749 Asm->OutStreamer->emitLabel(DebugLocs.getSym()); 2750 2751 for (const auto &List : DebugLocs.getLists()) 2752 Asm->emitLabelDifference(List.Label, DebugLocs.getSym(), 2753 Asm->getDwarfOffsetByteSize()); 2754 2755 return TableEnd; 2756 } 2757 2758 template <typename Ranges, typename PayloadEmitter> 2759 static void emitRangeList( 2760 DwarfDebug &DD, AsmPrinter *Asm, MCSymbol *Sym, const Ranges &R, 2761 const DwarfCompileUnit &CU, unsigned BaseAddressx, unsigned OffsetPair, 2762 unsigned StartxLength, unsigned EndOfList, 2763 StringRef (*StringifyEnum)(unsigned), 2764 bool ShouldUseBaseAddress, 2765 PayloadEmitter EmitPayload) { 2766 2767 auto Size = Asm->MAI->getCodePointerSize(); 2768 bool UseDwarf5 = DD.getDwarfVersion() >= 5; 2769 2770 // Emit our symbol so we can find the beginning of the range. 2771 Asm->OutStreamer->emitLabel(Sym); 2772 2773 // Gather all the ranges that apply to the same section so they can share 2774 // a base address entry. 2775 MapVector<const MCSection *, std::vector<decltype(&*R.begin())>> SectionRanges; 2776 2777 for (const auto &Range : R) 2778 SectionRanges[&Range.Begin->getSection()].push_back(&Range); 2779 2780 const MCSymbol *CUBase = CU.getBaseAddress(); 2781 bool BaseIsSet = false; 2782 for (const auto &P : SectionRanges) { 2783 auto *Base = CUBase; 2784 if (!Base && ShouldUseBaseAddress) { 2785 const MCSymbol *Begin = P.second.front()->Begin; 2786 const MCSymbol *NewBase = DD.getSectionLabel(&Begin->getSection()); 2787 if (!UseDwarf5) { 2788 Base = NewBase; 2789 BaseIsSet = true; 2790 Asm->OutStreamer->emitIntValue(-1, Size); 2791 Asm->OutStreamer->AddComment(" base address"); 2792 Asm->OutStreamer->emitSymbolValue(Base, Size); 2793 } else if (NewBase != Begin || P.second.size() > 1) { 2794 // Only use a base address if 2795 // * the existing pool address doesn't match (NewBase != Begin) 2796 // * or, there's more than one entry to share the base address 2797 Base = NewBase; 2798 BaseIsSet = true; 2799 Asm->OutStreamer->AddComment(StringifyEnum(BaseAddressx)); 2800 Asm->emitInt8(BaseAddressx); 2801 Asm->OutStreamer->AddComment(" base address index"); 2802 Asm->emitULEB128(DD.getAddressPool().getIndex(Base)); 2803 } 2804 } else if (BaseIsSet && !UseDwarf5) { 2805 BaseIsSet = false; 2806 assert(!Base); 2807 Asm->OutStreamer->emitIntValue(-1, Size); 2808 Asm->OutStreamer->emitIntValue(0, Size); 2809 } 2810 2811 for (const auto *RS : P.second) { 2812 const MCSymbol *Begin = RS->Begin; 2813 const MCSymbol *End = RS->End; 2814 assert(Begin && "Range without a begin symbol?"); 2815 assert(End && "Range without an end symbol?"); 2816 if (Base) { 2817 if (UseDwarf5) { 2818 // Emit offset_pair when we have a base. 2819 Asm->OutStreamer->AddComment(StringifyEnum(OffsetPair)); 2820 Asm->emitInt8(OffsetPair); 2821 Asm->OutStreamer->AddComment(" starting offset"); 2822 Asm->emitLabelDifferenceAsULEB128(Begin, Base); 2823 Asm->OutStreamer->AddComment(" ending offset"); 2824 Asm->emitLabelDifferenceAsULEB128(End, Base); 2825 } else { 2826 Asm->emitLabelDifference(Begin, Base, Size); 2827 Asm->emitLabelDifference(End, Base, Size); 2828 } 2829 } else if (UseDwarf5) { 2830 Asm->OutStreamer->AddComment(StringifyEnum(StartxLength)); 2831 Asm->emitInt8(StartxLength); 2832 Asm->OutStreamer->AddComment(" start index"); 2833 Asm->emitULEB128(DD.getAddressPool().getIndex(Begin)); 2834 Asm->OutStreamer->AddComment(" length"); 2835 Asm->emitLabelDifferenceAsULEB128(End, Begin); 2836 } else { 2837 Asm->OutStreamer->emitSymbolValue(Begin, Size); 2838 Asm->OutStreamer->emitSymbolValue(End, Size); 2839 } 2840 EmitPayload(*RS); 2841 } 2842 } 2843 2844 if (UseDwarf5) { 2845 Asm->OutStreamer->AddComment(StringifyEnum(EndOfList)); 2846 Asm->emitInt8(EndOfList); 2847 } else { 2848 // Terminate the list with two 0 values. 2849 Asm->OutStreamer->emitIntValue(0, Size); 2850 Asm->OutStreamer->emitIntValue(0, Size); 2851 } 2852 } 2853 2854 // Handles emission of both debug_loclist / debug_loclist.dwo 2855 static void emitLocList(DwarfDebug &DD, AsmPrinter *Asm, const DebugLocStream::List &List) { 2856 emitRangeList(DD, Asm, List.Label, DD.getDebugLocs().getEntries(List), 2857 *List.CU, dwarf::DW_LLE_base_addressx, 2858 dwarf::DW_LLE_offset_pair, dwarf::DW_LLE_startx_length, 2859 dwarf::DW_LLE_end_of_list, llvm::dwarf::LocListEncodingString, 2860 /* ShouldUseBaseAddress */ true, 2861 [&](const DebugLocStream::Entry &E) { 2862 DD.emitDebugLocEntryLocation(E, List.CU); 2863 }); 2864 } 2865 2866 void DwarfDebug::emitDebugLocImpl(MCSection *Sec) { 2867 if (DebugLocs.getLists().empty()) 2868 return; 2869 2870 Asm->OutStreamer->SwitchSection(Sec); 2871 2872 MCSymbol *TableEnd = nullptr; 2873 if (getDwarfVersion() >= 5) 2874 TableEnd = emitLoclistsTableHeader(Asm, *this); 2875 2876 for (const auto &List : DebugLocs.getLists()) 2877 emitLocList(*this, Asm, List); 2878 2879 if (TableEnd) 2880 Asm->OutStreamer->emitLabel(TableEnd); 2881 } 2882 2883 // Emit locations into the .debug_loc/.debug_loclists section. 2884 void DwarfDebug::emitDebugLoc() { 2885 emitDebugLocImpl( 2886 getDwarfVersion() >= 5 2887 ? Asm->getObjFileLowering().getDwarfLoclistsSection() 2888 : Asm->getObjFileLowering().getDwarfLocSection()); 2889 } 2890 2891 // Emit locations into the .debug_loc.dwo/.debug_loclists.dwo section. 2892 void DwarfDebug::emitDebugLocDWO() { 2893 if (getDwarfVersion() >= 5) { 2894 emitDebugLocImpl( 2895 Asm->getObjFileLowering().getDwarfLoclistsDWOSection()); 2896 2897 return; 2898 } 2899 2900 for (const auto &List : DebugLocs.getLists()) { 2901 Asm->OutStreamer->SwitchSection( 2902 Asm->getObjFileLowering().getDwarfLocDWOSection()); 2903 Asm->OutStreamer->emitLabel(List.Label); 2904 2905 for (const auto &Entry : DebugLocs.getEntries(List)) { 2906 // GDB only supports startx_length in pre-standard split-DWARF. 2907 // (in v5 standard loclists, it currently* /only/ supports base_address + 2908 // offset_pair, so the implementations can't really share much since they 2909 // need to use different representations) 2910 // * as of October 2018, at least 2911 // 2912 // In v5 (see emitLocList), this uses SectionLabels to reuse existing 2913 // addresses in the address pool to minimize object size/relocations. 2914 Asm->emitInt8(dwarf::DW_LLE_startx_length); 2915 unsigned idx = AddrPool.getIndex(Entry.Begin); 2916 Asm->emitULEB128(idx); 2917 // Also the pre-standard encoding is slightly different, emitting this as 2918 // an address-length entry here, but its a ULEB128 in DWARFv5 loclists. 2919 Asm->emitLabelDifference(Entry.End, Entry.Begin, 4); 2920 emitDebugLocEntryLocation(Entry, List.CU); 2921 } 2922 Asm->emitInt8(dwarf::DW_LLE_end_of_list); 2923 } 2924 } 2925 2926 struct ArangeSpan { 2927 const MCSymbol *Start, *End; 2928 }; 2929 2930 // Emit a debug aranges section, containing a CU lookup for any 2931 // address we can tie back to a CU. 2932 void DwarfDebug::emitDebugARanges() { 2933 // Provides a unique id per text section. 2934 MapVector<MCSection *, SmallVector<SymbolCU, 8>> SectionMap; 2935 2936 // Filter labels by section. 2937 for (const SymbolCU &SCU : ArangeLabels) { 2938 if (SCU.Sym->isInSection()) { 2939 // Make a note of this symbol and it's section. 2940 MCSection *Section = &SCU.Sym->getSection(); 2941 if (!Section->getKind().isMetadata()) 2942 SectionMap[Section].push_back(SCU); 2943 } else { 2944 // Some symbols (e.g. common/bss on mach-o) can have no section but still 2945 // appear in the output. This sucks as we rely on sections to build 2946 // arange spans. We can do it without, but it's icky. 2947 SectionMap[nullptr].push_back(SCU); 2948 } 2949 } 2950 2951 DenseMap<DwarfCompileUnit *, std::vector<ArangeSpan>> Spans; 2952 2953 for (auto &I : SectionMap) { 2954 MCSection *Section = I.first; 2955 SmallVector<SymbolCU, 8> &List = I.second; 2956 if (List.size() < 1) 2957 continue; 2958 2959 // If we have no section (e.g. common), just write out 2960 // individual spans for each symbol. 2961 if (!Section) { 2962 for (const SymbolCU &Cur : List) { 2963 ArangeSpan Span; 2964 Span.Start = Cur.Sym; 2965 Span.End = nullptr; 2966 assert(Cur.CU); 2967 Spans[Cur.CU].push_back(Span); 2968 } 2969 continue; 2970 } 2971 2972 // Sort the symbols by offset within the section. 2973 llvm::stable_sort(List, [&](const SymbolCU &A, const SymbolCU &B) { 2974 unsigned IA = A.Sym ? Asm->OutStreamer->GetSymbolOrder(A.Sym) : 0; 2975 unsigned IB = B.Sym ? Asm->OutStreamer->GetSymbolOrder(B.Sym) : 0; 2976 2977 // Symbols with no order assigned should be placed at the end. 2978 // (e.g. section end labels) 2979 if (IA == 0) 2980 return false; 2981 if (IB == 0) 2982 return true; 2983 return IA < IB; 2984 }); 2985 2986 // Insert a final terminator. 2987 List.push_back(SymbolCU(nullptr, Asm->OutStreamer->endSection(Section))); 2988 2989 // Build spans between each label. 2990 const MCSymbol *StartSym = List[0].Sym; 2991 for (size_t n = 1, e = List.size(); n < e; n++) { 2992 const SymbolCU &Prev = List[n - 1]; 2993 const SymbolCU &Cur = List[n]; 2994 2995 // Try and build the longest span we can within the same CU. 2996 if (Cur.CU != Prev.CU) { 2997 ArangeSpan Span; 2998 Span.Start = StartSym; 2999 Span.End = Cur.Sym; 3000 assert(Prev.CU); 3001 Spans[Prev.CU].push_back(Span); 3002 StartSym = Cur.Sym; 3003 } 3004 } 3005 } 3006 3007 // Start the dwarf aranges section. 3008 Asm->OutStreamer->SwitchSection( 3009 Asm->getObjFileLowering().getDwarfARangesSection()); 3010 3011 unsigned PtrSize = Asm->MAI->getCodePointerSize(); 3012 3013 // Build a list of CUs used. 3014 std::vector<DwarfCompileUnit *> CUs; 3015 for (const auto &it : Spans) { 3016 DwarfCompileUnit *CU = it.first; 3017 CUs.push_back(CU); 3018 } 3019 3020 // Sort the CU list (again, to ensure consistent output order). 3021 llvm::sort(CUs, [](const DwarfCompileUnit *A, const DwarfCompileUnit *B) { 3022 return A->getUniqueID() < B->getUniqueID(); 3023 }); 3024 3025 // Emit an arange table for each CU we used. 3026 for (DwarfCompileUnit *CU : CUs) { 3027 std::vector<ArangeSpan> &List = Spans[CU]; 3028 3029 // Describe the skeleton CU's offset and length, not the dwo file's. 3030 if (auto *Skel = CU->getSkeleton()) 3031 CU = Skel; 3032 3033 // Emit size of content not including length itself. 3034 unsigned ContentSize = 3035 sizeof(int16_t) + // DWARF ARange version number 3036 Asm->getDwarfOffsetByteSize() + // Offset of CU in the .debug_info 3037 // section 3038 sizeof(int8_t) + // Pointer Size (in bytes) 3039 sizeof(int8_t); // Segment Size (in bytes) 3040 3041 unsigned TupleSize = PtrSize * 2; 3042 3043 // 7.20 in the Dwarf specs requires the table to be aligned to a tuple. 3044 unsigned Padding = offsetToAlignment( 3045 Asm->getUnitLengthFieldByteSize() + ContentSize, Align(TupleSize)); 3046 3047 ContentSize += Padding; 3048 ContentSize += (List.size() + 1) * TupleSize; 3049 3050 // For each compile unit, write the list of spans it covers. 3051 Asm->emitDwarfUnitLength(ContentSize, "Length of ARange Set"); 3052 Asm->OutStreamer->AddComment("DWARF Arange version number"); 3053 Asm->emitInt16(dwarf::DW_ARANGES_VERSION); 3054 Asm->OutStreamer->AddComment("Offset Into Debug Info Section"); 3055 emitSectionReference(*CU); 3056 Asm->OutStreamer->AddComment("Address Size (in bytes)"); 3057 Asm->emitInt8(PtrSize); 3058 Asm->OutStreamer->AddComment("Segment Size (in bytes)"); 3059 Asm->emitInt8(0); 3060 3061 Asm->OutStreamer->emitFill(Padding, 0xff); 3062 3063 for (const ArangeSpan &Span : List) { 3064 Asm->emitLabelReference(Span.Start, PtrSize); 3065 3066 // Calculate the size as being from the span start to it's end. 3067 if (Span.End) { 3068 Asm->emitLabelDifference(Span.End, Span.Start, PtrSize); 3069 } else { 3070 // For symbols without an end marker (e.g. common), we 3071 // write a single arange entry containing just that one symbol. 3072 uint64_t Size = SymSize[Span.Start]; 3073 if (Size == 0) 3074 Size = 1; 3075 3076 Asm->OutStreamer->emitIntValue(Size, PtrSize); 3077 } 3078 } 3079 3080 Asm->OutStreamer->AddComment("ARange terminator"); 3081 Asm->OutStreamer->emitIntValue(0, PtrSize); 3082 Asm->OutStreamer->emitIntValue(0, PtrSize); 3083 } 3084 } 3085 3086 /// Emit a single range list. We handle both DWARF v5 and earlier. 3087 static void emitRangeList(DwarfDebug &DD, AsmPrinter *Asm, 3088 const RangeSpanList &List) { 3089 emitRangeList(DD, Asm, List.Label, List.Ranges, *List.CU, 3090 dwarf::DW_RLE_base_addressx, dwarf::DW_RLE_offset_pair, 3091 dwarf::DW_RLE_startx_length, dwarf::DW_RLE_end_of_list, 3092 llvm::dwarf::RangeListEncodingString, 3093 List.CU->getCUNode()->getRangesBaseAddress() || 3094 DD.getDwarfVersion() >= 5, 3095 [](auto) {}); 3096 } 3097 3098 void DwarfDebug::emitDebugRangesImpl(const DwarfFile &Holder, MCSection *Section) { 3099 if (Holder.getRangeLists().empty()) 3100 return; 3101 3102 assert(useRangesSection()); 3103 assert(!CUMap.empty()); 3104 assert(llvm::any_of(CUMap, [](const decltype(CUMap)::value_type &Pair) { 3105 return !Pair.second->getCUNode()->isDebugDirectivesOnly(); 3106 })); 3107 3108 Asm->OutStreamer->SwitchSection(Section); 3109 3110 MCSymbol *TableEnd = nullptr; 3111 if (getDwarfVersion() >= 5) 3112 TableEnd = emitRnglistsTableHeader(Asm, Holder); 3113 3114 for (const RangeSpanList &List : Holder.getRangeLists()) 3115 emitRangeList(*this, Asm, List); 3116 3117 if (TableEnd) 3118 Asm->OutStreamer->emitLabel(TableEnd); 3119 } 3120 3121 /// Emit address ranges into the .debug_ranges section or into the DWARF v5 3122 /// .debug_rnglists section. 3123 void DwarfDebug::emitDebugRanges() { 3124 const auto &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 3125 3126 emitDebugRangesImpl(Holder, 3127 getDwarfVersion() >= 5 3128 ? Asm->getObjFileLowering().getDwarfRnglistsSection() 3129 : Asm->getObjFileLowering().getDwarfRangesSection()); 3130 } 3131 3132 void DwarfDebug::emitDebugRangesDWO() { 3133 emitDebugRangesImpl(InfoHolder, 3134 Asm->getObjFileLowering().getDwarfRnglistsDWOSection()); 3135 } 3136 3137 /// Emit the header of a DWARF 5 macro section, or the GNU extension for 3138 /// DWARF 4. 3139 static void emitMacroHeader(AsmPrinter *Asm, const DwarfDebug &DD, 3140 const DwarfCompileUnit &CU, uint16_t DwarfVersion) { 3141 enum HeaderFlagMask { 3142 #define HANDLE_MACRO_FLAG(ID, NAME) MACRO_FLAG_##NAME = ID, 3143 #include "llvm/BinaryFormat/Dwarf.def" 3144 }; 3145 Asm->OutStreamer->AddComment("Macro information version"); 3146 Asm->emitInt16(DwarfVersion >= 5 ? DwarfVersion : 4); 3147 // We emit the line offset flag unconditionally here, since line offset should 3148 // be mostly present. 3149 if (Asm->isDwarf64()) { 3150 Asm->OutStreamer->AddComment("Flags: 64 bit, debug_line_offset present"); 3151 Asm->emitInt8(MACRO_FLAG_OFFSET_SIZE | MACRO_FLAG_DEBUG_LINE_OFFSET); 3152 } else { 3153 Asm->OutStreamer->AddComment("Flags: 32 bit, debug_line_offset present"); 3154 Asm->emitInt8(MACRO_FLAG_DEBUG_LINE_OFFSET); 3155 } 3156 Asm->OutStreamer->AddComment("debug_line_offset"); 3157 if (DD.useSplitDwarf()) 3158 Asm->emitDwarfLengthOrOffset(0); 3159 else 3160 Asm->emitDwarfSymbolReference(CU.getLineTableStartSym()); 3161 } 3162 3163 void DwarfDebug::handleMacroNodes(DIMacroNodeArray Nodes, DwarfCompileUnit &U) { 3164 for (auto *MN : Nodes) { 3165 if (auto *M = dyn_cast<DIMacro>(MN)) 3166 emitMacro(*M); 3167 else if (auto *F = dyn_cast<DIMacroFile>(MN)) 3168 emitMacroFile(*F, U); 3169 else 3170 llvm_unreachable("Unexpected DI type!"); 3171 } 3172 } 3173 3174 void DwarfDebug::emitMacro(DIMacro &M) { 3175 StringRef Name = M.getName(); 3176 StringRef Value = M.getValue(); 3177 3178 // There should be one space between the macro name and the macro value in 3179 // define entries. In undef entries, only the macro name is emitted. 3180 std::string Str = Value.empty() ? Name.str() : (Name + " " + Value).str(); 3181 3182 if (UseDebugMacroSection) { 3183 if (getDwarfVersion() >= 5) { 3184 unsigned Type = M.getMacinfoType() == dwarf::DW_MACINFO_define 3185 ? dwarf::DW_MACRO_define_strx 3186 : dwarf::DW_MACRO_undef_strx; 3187 Asm->OutStreamer->AddComment(dwarf::MacroString(Type)); 3188 Asm->emitULEB128(Type); 3189 Asm->OutStreamer->AddComment("Line Number"); 3190 Asm->emitULEB128(M.getLine()); 3191 Asm->OutStreamer->AddComment("Macro String"); 3192 Asm->emitULEB128( 3193 InfoHolder.getStringPool().getIndexedEntry(*Asm, Str).getIndex()); 3194 } else { 3195 unsigned Type = M.getMacinfoType() == dwarf::DW_MACINFO_define 3196 ? dwarf::DW_MACRO_GNU_define_indirect 3197 : dwarf::DW_MACRO_GNU_undef_indirect; 3198 Asm->OutStreamer->AddComment(dwarf::GnuMacroString(Type)); 3199 Asm->emitULEB128(Type); 3200 Asm->OutStreamer->AddComment("Line Number"); 3201 Asm->emitULEB128(M.getLine()); 3202 Asm->OutStreamer->AddComment("Macro String"); 3203 Asm->emitDwarfSymbolReference( 3204 InfoHolder.getStringPool().getEntry(*Asm, Str).getSymbol()); 3205 } 3206 } else { 3207 Asm->OutStreamer->AddComment(dwarf::MacinfoString(M.getMacinfoType())); 3208 Asm->emitULEB128(M.getMacinfoType()); 3209 Asm->OutStreamer->AddComment("Line Number"); 3210 Asm->emitULEB128(M.getLine()); 3211 Asm->OutStreamer->AddComment("Macro String"); 3212 Asm->OutStreamer->emitBytes(Str); 3213 Asm->emitInt8('\0'); 3214 } 3215 } 3216 3217 void DwarfDebug::emitMacroFileImpl( 3218 DIMacroFile &MF, DwarfCompileUnit &U, unsigned StartFile, unsigned EndFile, 3219 StringRef (*MacroFormToString)(unsigned Form)) { 3220 3221 Asm->OutStreamer->AddComment(MacroFormToString(StartFile)); 3222 Asm->emitULEB128(StartFile); 3223 Asm->OutStreamer->AddComment("Line Number"); 3224 Asm->emitULEB128(MF.getLine()); 3225 Asm->OutStreamer->AddComment("File Number"); 3226 DIFile &F = *MF.getFile(); 3227 if (useSplitDwarf()) 3228 Asm->emitULEB128(getDwoLineTable(U)->getFile( 3229 F.getDirectory(), F.getFilename(), getMD5AsBytes(&F), 3230 Asm->OutContext.getDwarfVersion(), F.getSource())); 3231 else 3232 Asm->emitULEB128(U.getOrCreateSourceID(&F)); 3233 handleMacroNodes(MF.getElements(), U); 3234 Asm->OutStreamer->AddComment(MacroFormToString(EndFile)); 3235 Asm->emitULEB128(EndFile); 3236 } 3237 3238 void DwarfDebug::emitMacroFile(DIMacroFile &F, DwarfCompileUnit &U) { 3239 // DWARFv5 macro and DWARFv4 macinfo share some common encodings, 3240 // so for readibility/uniformity, We are explicitly emitting those. 3241 assert(F.getMacinfoType() == dwarf::DW_MACINFO_start_file); 3242 if (UseDebugMacroSection) 3243 emitMacroFileImpl( 3244 F, U, dwarf::DW_MACRO_start_file, dwarf::DW_MACRO_end_file, 3245 (getDwarfVersion() >= 5) ? dwarf::MacroString : dwarf::GnuMacroString); 3246 else 3247 emitMacroFileImpl(F, U, dwarf::DW_MACINFO_start_file, 3248 dwarf::DW_MACINFO_end_file, dwarf::MacinfoString); 3249 } 3250 3251 void DwarfDebug::emitDebugMacinfoImpl(MCSection *Section) { 3252 for (const auto &P : CUMap) { 3253 auto &TheCU = *P.second; 3254 auto *SkCU = TheCU.getSkeleton(); 3255 DwarfCompileUnit &U = SkCU ? *SkCU : TheCU; 3256 auto *CUNode = cast<DICompileUnit>(P.first); 3257 DIMacroNodeArray Macros = CUNode->getMacros(); 3258 if (Macros.empty()) 3259 continue; 3260 Asm->OutStreamer->SwitchSection(Section); 3261 Asm->OutStreamer->emitLabel(U.getMacroLabelBegin()); 3262 if (UseDebugMacroSection) 3263 emitMacroHeader(Asm, *this, U, getDwarfVersion()); 3264 handleMacroNodes(Macros, U); 3265 Asm->OutStreamer->AddComment("End Of Macro List Mark"); 3266 Asm->emitInt8(0); 3267 } 3268 } 3269 3270 /// Emit macros into a debug macinfo/macro section. 3271 void DwarfDebug::emitDebugMacinfo() { 3272 auto &ObjLower = Asm->getObjFileLowering(); 3273 emitDebugMacinfoImpl(UseDebugMacroSection 3274 ? ObjLower.getDwarfMacroSection() 3275 : ObjLower.getDwarfMacinfoSection()); 3276 } 3277 3278 void DwarfDebug::emitDebugMacinfoDWO() { 3279 auto &ObjLower = Asm->getObjFileLowering(); 3280 emitDebugMacinfoImpl(UseDebugMacroSection 3281 ? ObjLower.getDwarfMacroDWOSection() 3282 : ObjLower.getDwarfMacinfoDWOSection()); 3283 } 3284 3285 // DWARF5 Experimental Separate Dwarf emitters. 3286 3287 void DwarfDebug::initSkeletonUnit(const DwarfUnit &U, DIE &Die, 3288 std::unique_ptr<DwarfCompileUnit> NewU) { 3289 3290 if (!CompilationDir.empty()) 3291 NewU->addString(Die, dwarf::DW_AT_comp_dir, CompilationDir); 3292 addGnuPubAttributes(*NewU, Die); 3293 3294 SkeletonHolder.addUnit(std::move(NewU)); 3295 } 3296 3297 DwarfCompileUnit &DwarfDebug::constructSkeletonCU(const DwarfCompileUnit &CU) { 3298 3299 auto OwnedUnit = std::make_unique<DwarfCompileUnit>( 3300 CU.getUniqueID(), CU.getCUNode(), Asm, this, &SkeletonHolder, 3301 UnitKind::Skeleton); 3302 DwarfCompileUnit &NewCU = *OwnedUnit; 3303 NewCU.setSection(Asm->getObjFileLowering().getDwarfInfoSection()); 3304 3305 NewCU.initStmtList(); 3306 3307 if (useSegmentedStringOffsetsTable()) 3308 NewCU.addStringOffsetsStart(); 3309 3310 initSkeletonUnit(CU, NewCU.getUnitDie(), std::move(OwnedUnit)); 3311 3312 return NewCU; 3313 } 3314 3315 // Emit the .debug_info.dwo section for separated dwarf. This contains the 3316 // compile units that would normally be in debug_info. 3317 void DwarfDebug::emitDebugInfoDWO() { 3318 assert(useSplitDwarf() && "No split dwarf debug info?"); 3319 // Don't emit relocations into the dwo file. 3320 InfoHolder.emitUnits(/* UseOffsets */ true); 3321 } 3322 3323 // Emit the .debug_abbrev.dwo section for separated dwarf. This contains the 3324 // abbreviations for the .debug_info.dwo section. 3325 void DwarfDebug::emitDebugAbbrevDWO() { 3326 assert(useSplitDwarf() && "No split dwarf?"); 3327 InfoHolder.emitAbbrevs(Asm->getObjFileLowering().getDwarfAbbrevDWOSection()); 3328 } 3329 3330 void DwarfDebug::emitDebugLineDWO() { 3331 assert(useSplitDwarf() && "No split dwarf?"); 3332 SplitTypeUnitFileTable.Emit( 3333 *Asm->OutStreamer, MCDwarfLineTableParams(), 3334 Asm->getObjFileLowering().getDwarfLineDWOSection()); 3335 } 3336 3337 void DwarfDebug::emitStringOffsetsTableHeaderDWO() { 3338 assert(useSplitDwarf() && "No split dwarf?"); 3339 InfoHolder.getStringPool().emitStringOffsetsTableHeader( 3340 *Asm, Asm->getObjFileLowering().getDwarfStrOffDWOSection(), 3341 InfoHolder.getStringOffsetsStartSym()); 3342 } 3343 3344 // Emit the .debug_str.dwo section for separated dwarf. This contains the 3345 // string section and is identical in format to traditional .debug_str 3346 // sections. 3347 void DwarfDebug::emitDebugStrDWO() { 3348 if (useSegmentedStringOffsetsTable()) 3349 emitStringOffsetsTableHeaderDWO(); 3350 assert(useSplitDwarf() && "No split dwarf?"); 3351 MCSection *OffSec = Asm->getObjFileLowering().getDwarfStrOffDWOSection(); 3352 InfoHolder.emitStrings(Asm->getObjFileLowering().getDwarfStrDWOSection(), 3353 OffSec, /* UseRelativeOffsets = */ false); 3354 } 3355 3356 // Emit address pool. 3357 void DwarfDebug::emitDebugAddr() { 3358 AddrPool.emit(*Asm, Asm->getObjFileLowering().getDwarfAddrSection()); 3359 } 3360 3361 MCDwarfDwoLineTable *DwarfDebug::getDwoLineTable(const DwarfCompileUnit &CU) { 3362 if (!useSplitDwarf()) 3363 return nullptr; 3364 const DICompileUnit *DIUnit = CU.getCUNode(); 3365 SplitTypeUnitFileTable.maybeSetRootFile( 3366 DIUnit->getDirectory(), DIUnit->getFilename(), 3367 getMD5AsBytes(DIUnit->getFile()), DIUnit->getSource()); 3368 return &SplitTypeUnitFileTable; 3369 } 3370 3371 uint64_t DwarfDebug::makeTypeSignature(StringRef Identifier) { 3372 MD5 Hash; 3373 Hash.update(Identifier); 3374 // ... take the least significant 8 bytes and return those. Our MD5 3375 // implementation always returns its results in little endian, so we actually 3376 // need the "high" word. 3377 MD5::MD5Result Result; 3378 Hash.final(Result); 3379 return Result.high(); 3380 } 3381 3382 void DwarfDebug::addDwarfTypeUnitType(DwarfCompileUnit &CU, 3383 StringRef Identifier, DIE &RefDie, 3384 const DICompositeType *CTy) { 3385 bool TopLevelType = TypeUnitsUnderConstruction.empty(); 3386 if (auto Signature = getOrCreateDwarfTypeUnit(CU, Identifier, CTy)) { 3387 CU.addDIETypeSignature(RefDie, *Signature); 3388 } else if (TopLevelType) { 3389 // Construct this type in the CU directly. 3390 // This is inefficient because all the dependent types will be rebuilt 3391 // from scratch, including building them in type units, discovering that 3392 // they depend on addresses, throwing them out and rebuilding them. 3393 CU.constructTypeDIE(RefDie, cast<DICompositeType>(CTy)); 3394 } 3395 } 3396 3397 Optional<uint64_t> 3398 DwarfDebug::getOrCreateDwarfTypeUnit(DwarfCompileUnit &CU, StringRef Identifier, 3399 const DICompositeType *CTy) { 3400 // Fast path if we're building some type units and one has already used the 3401 // address pool we know we're going to throw away all this work anyway, so 3402 // don't bother building dependent types. 3403 if (!TypeUnitsUnderConstruction.empty() && AddrPool.hasBeenUsed()) 3404 return None; 3405 3406 auto Ins = TypeSignatures.insert(std::make_pair(CTy, 0)); 3407 if (!Ins.second) 3408 return Ins.first->second; 3409 3410 bool TopLevelType = TypeUnitsUnderConstruction.empty(); 3411 AddrPool.resetUsedFlag(); 3412 3413 auto OwnedUnit = std::make_unique<DwarfTypeUnit>(CU, Asm, this, &InfoHolder, 3414 getDwoLineTable(CU)); 3415 DwarfTypeUnit &NewTU = *OwnedUnit; 3416 DIE &UnitDie = NewTU.getUnitDie(); 3417 TypeUnitsUnderConstruction.emplace_back(std::move(OwnedUnit), CTy); 3418 3419 NewTU.addUInt(UnitDie, dwarf::DW_AT_language, dwarf::DW_FORM_data2, 3420 CU.getLanguage()); 3421 3422 uint64_t Signature = makeTypeSignature(Identifier); 3423 NewTU.setTypeSignature(Signature); 3424 Ins.first->second = Signature; 3425 3426 if (useSplitDwarf()) { 3427 MCSection *Section = 3428 getDwarfVersion() <= 4 3429 ? Asm->getObjFileLowering().getDwarfTypesDWOSection() 3430 : Asm->getObjFileLowering().getDwarfInfoDWOSection(); 3431 NewTU.setSection(Section); 3432 } else { 3433 MCSection *Section = 3434 getDwarfVersion() <= 4 3435 ? Asm->getObjFileLowering().getDwarfTypesSection(Signature) 3436 : Asm->getObjFileLowering().getDwarfInfoSection(Signature); 3437 NewTU.setSection(Section); 3438 // Non-split type units reuse the compile unit's line table. 3439 CU.applyStmtList(UnitDie); 3440 } 3441 3442 // Add DW_AT_str_offsets_base to the type unit DIE, but not for split type 3443 // units. 3444 if (useSegmentedStringOffsetsTable() && !useSplitDwarf()) 3445 NewTU.addStringOffsetsStart(); 3446 3447 NewTU.setType(NewTU.createTypeDIE(CTy)); 3448 3449 if (TopLevelType) { 3450 auto TypeUnitsToAdd = std::move(TypeUnitsUnderConstruction); 3451 TypeUnitsUnderConstruction.clear(); 3452 3453 // Types referencing entries in the address table cannot be placed in type 3454 // units. 3455 if (AddrPool.hasBeenUsed()) { 3456 3457 // Remove all the types built while building this type. 3458 // This is pessimistic as some of these types might not be dependent on 3459 // the type that used an address. 3460 for (const auto &TU : TypeUnitsToAdd) 3461 TypeSignatures.erase(TU.second); 3462 return None; 3463 } 3464 3465 // If the type wasn't dependent on fission addresses, finish adding the type 3466 // and all its dependent types. 3467 for (auto &TU : TypeUnitsToAdd) { 3468 InfoHolder.computeSizeAndOffsetsForUnit(TU.first.get()); 3469 InfoHolder.emitUnit(TU.first.get(), useSplitDwarf()); 3470 } 3471 } 3472 return Signature; 3473 } 3474 3475 DwarfDebug::NonTypeUnitContext::NonTypeUnitContext(DwarfDebug *DD) 3476 : DD(DD), 3477 TypeUnitsUnderConstruction(std::move(DD->TypeUnitsUnderConstruction)), AddrPoolUsed(DD->AddrPool.hasBeenUsed()) { 3478 DD->TypeUnitsUnderConstruction.clear(); 3479 DD->AddrPool.resetUsedFlag(); 3480 } 3481 3482 DwarfDebug::NonTypeUnitContext::~NonTypeUnitContext() { 3483 DD->TypeUnitsUnderConstruction = std::move(TypeUnitsUnderConstruction); 3484 DD->AddrPool.resetUsedFlag(AddrPoolUsed); 3485 } 3486 3487 DwarfDebug::NonTypeUnitContext DwarfDebug::enterNonTypeUnitContext() { 3488 return NonTypeUnitContext(this); 3489 } 3490 3491 // Add the Name along with its companion DIE to the appropriate accelerator 3492 // table (for AccelTableKind::Dwarf it's always AccelDebugNames, for 3493 // AccelTableKind::Apple, we use the table we got as an argument). If 3494 // accelerator tables are disabled, this function does nothing. 3495 template <typename DataT> 3496 void DwarfDebug::addAccelNameImpl(const DICompileUnit &CU, 3497 AccelTable<DataT> &AppleAccel, StringRef Name, 3498 const DIE &Die) { 3499 if (getAccelTableKind() == AccelTableKind::None) 3500 return; 3501 3502 if (getAccelTableKind() != AccelTableKind::Apple && 3503 CU.getNameTableKind() != DICompileUnit::DebugNameTableKind::Default) 3504 return; 3505 3506 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 3507 DwarfStringPoolEntryRef Ref = Holder.getStringPool().getEntry(*Asm, Name); 3508 3509 switch (getAccelTableKind()) { 3510 case AccelTableKind::Apple: 3511 AppleAccel.addName(Ref, Die); 3512 break; 3513 case AccelTableKind::Dwarf: 3514 AccelDebugNames.addName(Ref, Die); 3515 break; 3516 case AccelTableKind::Default: 3517 llvm_unreachable("Default should have already been resolved."); 3518 case AccelTableKind::None: 3519 llvm_unreachable("None handled above"); 3520 } 3521 } 3522 3523 void DwarfDebug::addAccelName(const DICompileUnit &CU, StringRef Name, 3524 const DIE &Die) { 3525 addAccelNameImpl(CU, AccelNames, Name, Die); 3526 } 3527 3528 void DwarfDebug::addAccelObjC(const DICompileUnit &CU, StringRef Name, 3529 const DIE &Die) { 3530 // ObjC names go only into the Apple accelerator tables. 3531 if (getAccelTableKind() == AccelTableKind::Apple) 3532 addAccelNameImpl(CU, AccelObjC, Name, Die); 3533 } 3534 3535 void DwarfDebug::addAccelNamespace(const DICompileUnit &CU, StringRef Name, 3536 const DIE &Die) { 3537 addAccelNameImpl(CU, AccelNamespace, Name, Die); 3538 } 3539 3540 void DwarfDebug::addAccelType(const DICompileUnit &CU, StringRef Name, 3541 const DIE &Die, char Flags) { 3542 addAccelNameImpl(CU, AccelTypes, Name, Die); 3543 } 3544 3545 uint16_t DwarfDebug::getDwarfVersion() const { 3546 return Asm->OutStreamer->getContext().getDwarfVersion(); 3547 } 3548 3549 dwarf::Form DwarfDebug::getDwarfSectionOffsetForm() const { 3550 if (Asm->getDwarfVersion() >= 4) 3551 return dwarf::Form::DW_FORM_sec_offset; 3552 assert((!Asm->isDwarf64() || (Asm->getDwarfVersion() == 3)) && 3553 "DWARF64 is not defined prior DWARFv3"); 3554 return Asm->isDwarf64() ? dwarf::Form::DW_FORM_data8 3555 : dwarf::Form::DW_FORM_data4; 3556 } 3557 3558 const MCSymbol *DwarfDebug::getSectionLabel(const MCSection *S) { 3559 auto I = SectionLabels.find(S); 3560 if (I == SectionLabels.end()) 3561 return nullptr; 3562 return I->second; 3563 } 3564 void DwarfDebug::insertSectionLabel(const MCSymbol *S) { 3565 if (SectionLabels.insert(std::make_pair(&S->getSection(), S)).second) 3566 if (useSplitDwarf() || getDwarfVersion() >= 5) 3567 AddrPool.getIndex(S); 3568 } 3569 3570 Optional<MD5::MD5Result> DwarfDebug::getMD5AsBytes(const DIFile *File) const { 3571 assert(File); 3572 if (getDwarfVersion() < 5) 3573 return None; 3574 Optional<DIFile::ChecksumInfo<StringRef>> Checksum = File->getChecksum(); 3575 if (!Checksum || Checksum->Kind != DIFile::CSK_MD5) 3576 return None; 3577 3578 // Convert the string checksum to an MD5Result for the streamer. 3579 // The verifier validates the checksum so we assume it's okay. 3580 // An MD5 checksum is 16 bytes. 3581 std::string ChecksumString = fromHex(Checksum->Value); 3582 MD5::MD5Result CKMem; 3583 std::copy(ChecksumString.begin(), ChecksumString.end(), CKMem.Bytes.data()); 3584 return CKMem; 3585 } 3586