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 /// Sort and unique GVEs by comparing their fragment offset. 1118 static SmallVectorImpl<DwarfCompileUnit::GlobalExpr> & 1119 sortGlobalExprs(SmallVectorImpl<DwarfCompileUnit::GlobalExpr> &GVEs) { 1120 llvm::sort( 1121 GVEs, [](DwarfCompileUnit::GlobalExpr A, DwarfCompileUnit::GlobalExpr B) { 1122 // Sort order: first null exprs, then exprs without fragment 1123 // info, then sort by fragment offset in bits. 1124 // FIXME: Come up with a more comprehensive comparator so 1125 // the sorting isn't non-deterministic, and so the following 1126 // std::unique call works correctly. 1127 if (!A.Expr || !B.Expr) 1128 return !!B.Expr; 1129 auto FragmentA = A.Expr->getFragmentInfo(); 1130 auto FragmentB = B.Expr->getFragmentInfo(); 1131 if (!FragmentA || !FragmentB) 1132 return !!FragmentB; 1133 return FragmentA->OffsetInBits < FragmentB->OffsetInBits; 1134 }); 1135 GVEs.erase(std::unique(GVEs.begin(), GVEs.end(), 1136 [](DwarfCompileUnit::GlobalExpr A, 1137 DwarfCompileUnit::GlobalExpr B) { 1138 return A.Expr == B.Expr; 1139 }), 1140 GVEs.end()); 1141 return GVEs; 1142 } 1143 1144 /// Create a DIE for \p Ty if it doesn't already exist. If type units are 1145 /// enabled, try to emit a type unit without a CU skeleton DIE. 1146 static void createMaybeUnusedType(DwarfDebug &DD, DwarfCompileUnit &CU, 1147 DIType &Ty) { 1148 // Try to generate a type unit without creating a skeleton DIE in this CU. 1149 if (DICompositeType const *CTy = dyn_cast<DICompositeType>(&Ty)) { 1150 MDString const *TypeId = CTy->getRawIdentifier(); 1151 if (DD.generateTypeUnits() && TypeId && !Ty.isForwardDecl()) 1152 if (DD.getOrCreateDwarfTypeUnit(CU, TypeId->getString(), CTy)) 1153 return; 1154 } 1155 // We couldn't or shouldn't add a type unit so create the DIE normally. 1156 CU.getOrCreateTypeDIE(&Ty); 1157 } 1158 1159 // Emit all Dwarf sections that should come prior to the content. Create 1160 // global DIEs and emit initial debug info sections. This is invoked by 1161 // the target AsmPrinter. 1162 void DwarfDebug::beginModule(Module *M) { 1163 DebugHandlerBase::beginModule(M); 1164 1165 if (!Asm || !MMI->hasDebugInfo()) 1166 return; 1167 1168 unsigned NumDebugCUs = std::distance(M->debug_compile_units_begin(), 1169 M->debug_compile_units_end()); 1170 assert(NumDebugCUs > 0 && "Asm unexpectedly initialized"); 1171 assert(MMI->hasDebugInfo() && 1172 "DebugInfoAvailabilty unexpectedly not initialized"); 1173 SingleCU = NumDebugCUs == 1; 1174 DenseMap<DIGlobalVariable *, SmallVector<DwarfCompileUnit::GlobalExpr, 1>> 1175 GVMap; 1176 for (const GlobalVariable &Global : M->globals()) { 1177 SmallVector<DIGlobalVariableExpression *, 1> GVs; 1178 Global.getDebugInfo(GVs); 1179 for (auto *GVE : GVs) 1180 GVMap[GVE->getVariable()].push_back({&Global, GVE->getExpression()}); 1181 } 1182 1183 // Create the symbol that designates the start of the unit's contribution 1184 // to the string offsets table. In a split DWARF scenario, only the skeleton 1185 // unit has the DW_AT_str_offsets_base attribute (and hence needs the symbol). 1186 if (useSegmentedStringOffsetsTable()) 1187 (useSplitDwarf() ? SkeletonHolder : InfoHolder) 1188 .setStringOffsetsStartSym(Asm->createTempSymbol("str_offsets_base")); 1189 1190 1191 // Create the symbols that designates the start of the DWARF v5 range list 1192 // and locations list tables. They are located past the table headers. 1193 if (getDwarfVersion() >= 5) { 1194 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 1195 Holder.setRnglistsTableBaseSym( 1196 Asm->createTempSymbol("rnglists_table_base")); 1197 1198 if (useSplitDwarf()) 1199 InfoHolder.setRnglistsTableBaseSym( 1200 Asm->createTempSymbol("rnglists_dwo_table_base")); 1201 } 1202 1203 // Create the symbol that points to the first entry following the debug 1204 // address table (.debug_addr) header. 1205 AddrPool.setLabel(Asm->createTempSymbol("addr_table_base")); 1206 DebugLocs.setSym(Asm->createTempSymbol("loclists_table_base")); 1207 1208 for (DICompileUnit *CUNode : M->debug_compile_units()) { 1209 // FIXME: Move local imported entities into a list attached to the 1210 // subprogram, then this search won't be needed and a 1211 // getImportedEntities().empty() test should go below with the rest. 1212 bool HasNonLocalImportedEntities = llvm::any_of( 1213 CUNode->getImportedEntities(), [](const DIImportedEntity *IE) { 1214 return !isa<DILocalScope>(IE->getScope()); 1215 }); 1216 1217 if (!HasNonLocalImportedEntities && CUNode->getEnumTypes().empty() && 1218 CUNode->getRetainedTypes().empty() && 1219 CUNode->getGlobalVariables().empty() && CUNode->getMacros().empty()) 1220 continue; 1221 1222 DwarfCompileUnit &CU = getOrCreateDwarfCompileUnit(CUNode); 1223 1224 // Global Variables. 1225 for (auto *GVE : CUNode->getGlobalVariables()) { 1226 // Don't bother adding DIGlobalVariableExpressions listed in the CU if we 1227 // already know about the variable and it isn't adding a constant 1228 // expression. 1229 auto &GVMapEntry = GVMap[GVE->getVariable()]; 1230 auto *Expr = GVE->getExpression(); 1231 if (!GVMapEntry.size() || (Expr && Expr->isConstant())) 1232 GVMapEntry.push_back({nullptr, Expr}); 1233 } 1234 1235 DenseSet<DIGlobalVariable *> Processed; 1236 for (auto *GVE : CUNode->getGlobalVariables()) { 1237 DIGlobalVariable *GV = GVE->getVariable(); 1238 if (Processed.insert(GV).second) 1239 CU.getOrCreateGlobalVariableDIE(GV, sortGlobalExprs(GVMap[GV])); 1240 } 1241 1242 for (auto *Ty : CUNode->getEnumTypes()) { 1243 // The enum types array by design contains pointers to 1244 // MDNodes rather than DIRefs. Unique them here. 1245 createMaybeUnusedType(*this, CU, *Ty); 1246 } 1247 for (auto *Ty : CUNode->getRetainedTypes()) { 1248 // The retained types array by design contains pointers to 1249 // MDNodes rather than DIRefs. Unique them here. 1250 if (DIType *RT = dyn_cast<DIType>(Ty)) 1251 // There is no point in force-emitting a forward declaration. 1252 createMaybeUnusedType(*this, CU, *RT); 1253 } 1254 // Emit imported_modules last so that the relevant context is already 1255 // available. 1256 for (auto *IE : CUNode->getImportedEntities()) 1257 constructAndAddImportedEntityDIE(CU, IE); 1258 } 1259 } 1260 1261 void DwarfDebug::finishEntityDefinitions() { 1262 for (const auto &Entity : ConcreteEntities) { 1263 DIE *Die = Entity->getDIE(); 1264 assert(Die); 1265 // FIXME: Consider the time-space tradeoff of just storing the unit pointer 1266 // in the ConcreteEntities list, rather than looking it up again here. 1267 // DIE::getUnit isn't simple - it walks parent pointers, etc. 1268 DwarfCompileUnit *Unit = CUDieMap.lookup(Die->getUnitDie()); 1269 assert(Unit); 1270 Unit->finishEntityDefinition(Entity.get()); 1271 } 1272 } 1273 1274 void DwarfDebug::finishSubprogramDefinitions() { 1275 for (const DISubprogram *SP : ProcessedSPNodes) { 1276 assert(SP->getUnit()->getEmissionKind() != DICompileUnit::NoDebug); 1277 forBothCUs( 1278 getOrCreateDwarfCompileUnit(SP->getUnit()), 1279 [&](DwarfCompileUnit &CU) { CU.finishSubprogramDefinition(SP); }); 1280 } 1281 } 1282 1283 void DwarfDebug::finalizeModuleInfo() { 1284 const TargetLoweringObjectFile &TLOF = Asm->getObjFileLowering(); 1285 1286 finishSubprogramDefinitions(); 1287 1288 finishEntityDefinitions(); 1289 1290 // Include the DWO file name in the hash if there's more than one CU. 1291 // This handles ThinLTO's situation where imported CUs may very easily be 1292 // duplicate with the same CU partially imported into another ThinLTO unit. 1293 StringRef DWOName; 1294 if (CUMap.size() > 1) 1295 DWOName = Asm->TM.Options.MCOptions.SplitDwarfFile; 1296 1297 // Handle anything that needs to be done on a per-unit basis after 1298 // all other generation. 1299 for (const auto &P : CUMap) { 1300 auto &TheCU = *P.second; 1301 if (TheCU.getCUNode()->isDebugDirectivesOnly()) 1302 continue; 1303 // Emit DW_AT_containing_type attribute to connect types with their 1304 // vtable holding type. 1305 TheCU.constructContainingTypeDIEs(); 1306 1307 // Add CU specific attributes if we need to add any. 1308 // If we're splitting the dwarf out now that we've got the entire 1309 // CU then add the dwo id to it. 1310 auto *SkCU = TheCU.getSkeleton(); 1311 1312 bool HasSplitUnit = SkCU && !TheCU.getUnitDie().children().empty(); 1313 1314 if (HasSplitUnit) { 1315 dwarf::Attribute attrDWOName = getDwarfVersion() >= 5 1316 ? dwarf::DW_AT_dwo_name 1317 : dwarf::DW_AT_GNU_dwo_name; 1318 finishUnitAttributes(TheCU.getCUNode(), TheCU); 1319 TheCU.addString(TheCU.getUnitDie(), attrDWOName, 1320 Asm->TM.Options.MCOptions.SplitDwarfFile); 1321 SkCU->addString(SkCU->getUnitDie(), attrDWOName, 1322 Asm->TM.Options.MCOptions.SplitDwarfFile); 1323 // Emit a unique identifier for this CU. 1324 uint64_t ID = 1325 DIEHash(Asm, &TheCU).computeCUSignature(DWOName, TheCU.getUnitDie()); 1326 if (getDwarfVersion() >= 5) { 1327 TheCU.setDWOId(ID); 1328 SkCU->setDWOId(ID); 1329 } else { 1330 TheCU.addUInt(TheCU.getUnitDie(), dwarf::DW_AT_GNU_dwo_id, 1331 dwarf::DW_FORM_data8, ID); 1332 SkCU->addUInt(SkCU->getUnitDie(), dwarf::DW_AT_GNU_dwo_id, 1333 dwarf::DW_FORM_data8, ID); 1334 } 1335 1336 if (getDwarfVersion() < 5 && !SkeletonHolder.getRangeLists().empty()) { 1337 const MCSymbol *Sym = TLOF.getDwarfRangesSection()->getBeginSymbol(); 1338 SkCU->addSectionLabel(SkCU->getUnitDie(), dwarf::DW_AT_GNU_ranges_base, 1339 Sym, Sym); 1340 } 1341 } else if (SkCU) { 1342 finishUnitAttributes(SkCU->getCUNode(), *SkCU); 1343 } 1344 1345 // If we have code split among multiple sections or non-contiguous 1346 // ranges of code then emit a DW_AT_ranges attribute on the unit that will 1347 // remain in the .o file, otherwise add a DW_AT_low_pc. 1348 // FIXME: We should use ranges allow reordering of code ala 1349 // .subsections_via_symbols in mach-o. This would mean turning on 1350 // ranges for all subprogram DIEs for mach-o. 1351 DwarfCompileUnit &U = SkCU ? *SkCU : TheCU; 1352 1353 if (unsigned NumRanges = TheCU.getRanges().size()) { 1354 if (NumRanges > 1 && useRangesSection()) 1355 // A DW_AT_low_pc attribute may also be specified in combination with 1356 // DW_AT_ranges to specify the default base address for use in 1357 // location lists (see Section 2.6.2) and range lists (see Section 1358 // 2.17.3). 1359 U.addUInt(U.getUnitDie(), dwarf::DW_AT_low_pc, dwarf::DW_FORM_addr, 0); 1360 else 1361 U.setBaseAddress(TheCU.getRanges().front().Begin); 1362 U.attachRangesOrLowHighPC(U.getUnitDie(), TheCU.takeRanges()); 1363 } 1364 1365 // We don't keep track of which addresses are used in which CU so this 1366 // is a bit pessimistic under LTO. 1367 if ((HasSplitUnit || getDwarfVersion() >= 5) && !AddrPool.isEmpty()) 1368 U.addAddrTableBase(); 1369 1370 if (getDwarfVersion() >= 5) { 1371 if (U.hasRangeLists()) 1372 U.addRnglistsBase(); 1373 1374 if (!DebugLocs.getLists().empty()) { 1375 if (!useSplitDwarf()) 1376 U.addSectionLabel(U.getUnitDie(), dwarf::DW_AT_loclists_base, 1377 DebugLocs.getSym(), 1378 TLOF.getDwarfLoclistsSection()->getBeginSymbol()); 1379 } 1380 } 1381 1382 auto *CUNode = cast<DICompileUnit>(P.first); 1383 // If compile Unit has macros, emit "DW_AT_macro_info/DW_AT_macros" 1384 // attribute. 1385 if (CUNode->getMacros()) { 1386 if (UseDebugMacroSection) { 1387 if (useSplitDwarf()) 1388 TheCU.addSectionDelta( 1389 TheCU.getUnitDie(), dwarf::DW_AT_macros, U.getMacroLabelBegin(), 1390 TLOF.getDwarfMacroDWOSection()->getBeginSymbol()); 1391 else { 1392 dwarf::Attribute MacrosAttr = getDwarfVersion() >= 5 1393 ? dwarf::DW_AT_macros 1394 : dwarf::DW_AT_GNU_macros; 1395 U.addSectionLabel(U.getUnitDie(), MacrosAttr, U.getMacroLabelBegin(), 1396 TLOF.getDwarfMacroSection()->getBeginSymbol()); 1397 } 1398 } else { 1399 if (useSplitDwarf()) 1400 TheCU.addSectionDelta( 1401 TheCU.getUnitDie(), dwarf::DW_AT_macro_info, 1402 U.getMacroLabelBegin(), 1403 TLOF.getDwarfMacinfoDWOSection()->getBeginSymbol()); 1404 else 1405 U.addSectionLabel(U.getUnitDie(), dwarf::DW_AT_macro_info, 1406 U.getMacroLabelBegin(), 1407 TLOF.getDwarfMacinfoSection()->getBeginSymbol()); 1408 } 1409 } 1410 } 1411 1412 // Emit all frontend-produced Skeleton CUs, i.e., Clang modules. 1413 for (auto *CUNode : MMI->getModule()->debug_compile_units()) 1414 if (CUNode->getDWOId()) 1415 getOrCreateDwarfCompileUnit(CUNode); 1416 1417 // Compute DIE offsets and sizes. 1418 InfoHolder.computeSizeAndOffsets(); 1419 if (useSplitDwarf()) 1420 SkeletonHolder.computeSizeAndOffsets(); 1421 } 1422 1423 1424 // Emit all Dwarf sections that should come after the content. 1425 void DwarfDebug::endModule() { 1426 // Terminate the pending line table. 1427 if (PrevCU) 1428 terminateLineTable(PrevCU); 1429 PrevCU = nullptr; 1430 assert(CurFn == nullptr); 1431 assert(CurMI == nullptr); 1432 1433 for (const auto &P : CUMap) { 1434 auto &CU = *P.second; 1435 CU.createBaseTypeDIEs(); 1436 } 1437 1438 // If we aren't actually generating debug info (check beginModule - 1439 // conditionalized on the presence of the llvm.dbg.cu metadata node) 1440 if (!Asm || !MMI->hasDebugInfo()) 1441 return; 1442 1443 // Finalize the debug info for the module. 1444 finalizeModuleInfo(); 1445 1446 if (useSplitDwarf()) 1447 // Emit debug_loc.dwo/debug_loclists.dwo section. 1448 emitDebugLocDWO(); 1449 else 1450 // Emit debug_loc/debug_loclists section. 1451 emitDebugLoc(); 1452 1453 // Corresponding abbreviations into a abbrev section. 1454 emitAbbreviations(); 1455 1456 // Emit all the DIEs into a debug info section. 1457 emitDebugInfo(); 1458 1459 // Emit info into a debug aranges section. 1460 if (GenerateARangeSection) 1461 emitDebugARanges(); 1462 1463 // Emit info into a debug ranges section. 1464 emitDebugRanges(); 1465 1466 if (useSplitDwarf()) 1467 // Emit info into a debug macinfo.dwo section. 1468 emitDebugMacinfoDWO(); 1469 else 1470 // Emit info into a debug macinfo/macro section. 1471 emitDebugMacinfo(); 1472 1473 emitDebugStr(); 1474 1475 if (useSplitDwarf()) { 1476 emitDebugStrDWO(); 1477 emitDebugInfoDWO(); 1478 emitDebugAbbrevDWO(); 1479 emitDebugLineDWO(); 1480 emitDebugRangesDWO(); 1481 } 1482 1483 emitDebugAddr(); 1484 1485 // Emit info into the dwarf accelerator table sections. 1486 switch (getAccelTableKind()) { 1487 case AccelTableKind::Apple: 1488 emitAccelNames(); 1489 emitAccelObjC(); 1490 emitAccelNamespaces(); 1491 emitAccelTypes(); 1492 break; 1493 case AccelTableKind::Dwarf: 1494 emitAccelDebugNames(); 1495 break; 1496 case AccelTableKind::None: 1497 break; 1498 case AccelTableKind::Default: 1499 llvm_unreachable("Default should have already been resolved."); 1500 } 1501 1502 // Emit the pubnames and pubtypes sections if requested. 1503 emitDebugPubSections(); 1504 1505 // clean up. 1506 // FIXME: AbstractVariables.clear(); 1507 } 1508 1509 void DwarfDebug::ensureAbstractEntityIsCreated(DwarfCompileUnit &CU, 1510 const DINode *Node, 1511 const MDNode *ScopeNode) { 1512 if (CU.getExistingAbstractEntity(Node)) 1513 return; 1514 1515 CU.createAbstractEntity(Node, LScopes.getOrCreateAbstractScope( 1516 cast<DILocalScope>(ScopeNode))); 1517 } 1518 1519 void DwarfDebug::ensureAbstractEntityIsCreatedIfScoped(DwarfCompileUnit &CU, 1520 const DINode *Node, const MDNode *ScopeNode) { 1521 if (CU.getExistingAbstractEntity(Node)) 1522 return; 1523 1524 if (LexicalScope *Scope = 1525 LScopes.findAbstractScope(cast_or_null<DILocalScope>(ScopeNode))) 1526 CU.createAbstractEntity(Node, Scope); 1527 } 1528 1529 // Collect variable information from side table maintained by MF. 1530 void DwarfDebug::collectVariableInfoFromMFTable( 1531 DwarfCompileUnit &TheCU, DenseSet<InlinedEntity> &Processed) { 1532 SmallDenseMap<InlinedEntity, DbgVariable *> MFVars; 1533 LLVM_DEBUG(dbgs() << "DwarfDebug: collecting variables from MF side table\n"); 1534 for (const auto &VI : Asm->MF->getVariableDbgInfo()) { 1535 if (!VI.Var) 1536 continue; 1537 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1538 "Expected inlined-at fields to agree"); 1539 1540 InlinedEntity Var(VI.Var, VI.Loc->getInlinedAt()); 1541 Processed.insert(Var); 1542 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1543 1544 // If variable scope is not found then skip this variable. 1545 if (!Scope) { 1546 LLVM_DEBUG(dbgs() << "Dropping debug info for " << VI.Var->getName() 1547 << ", no variable scope found\n"); 1548 continue; 1549 } 1550 1551 ensureAbstractEntityIsCreatedIfScoped(TheCU, Var.first, Scope->getScopeNode()); 1552 auto RegVar = std::make_unique<DbgVariable>( 1553 cast<DILocalVariable>(Var.first), Var.second); 1554 RegVar->initializeMMI(VI.Expr, VI.Slot); 1555 LLVM_DEBUG(dbgs() << "Created DbgVariable for " << VI.Var->getName() 1556 << "\n"); 1557 1558 if (DbgVariable *DbgVar = MFVars.lookup(Var)) 1559 DbgVar->addMMIEntry(*RegVar); 1560 else if (InfoHolder.addScopeVariable(Scope, RegVar.get())) { 1561 MFVars.insert({Var, RegVar.get()}); 1562 ConcreteEntities.push_back(std::move(RegVar)); 1563 } 1564 } 1565 } 1566 1567 /// Determine whether a *singular* DBG_VALUE is valid for the entirety of its 1568 /// enclosing lexical scope. The check ensures there are no other instructions 1569 /// in the same lexical scope preceding the DBG_VALUE and that its range is 1570 /// either open or otherwise rolls off the end of the scope. 1571 static bool validThroughout(LexicalScopes &LScopes, 1572 const MachineInstr *DbgValue, 1573 const MachineInstr *RangeEnd, 1574 const InstructionOrdering &Ordering) { 1575 assert(DbgValue->getDebugLoc() && "DBG_VALUE without a debug location"); 1576 auto MBB = DbgValue->getParent(); 1577 auto DL = DbgValue->getDebugLoc(); 1578 auto *LScope = LScopes.findLexicalScope(DL); 1579 // Scope doesn't exist; this is a dead DBG_VALUE. 1580 if (!LScope) 1581 return false; 1582 auto &LSRange = LScope->getRanges(); 1583 if (LSRange.size() == 0) 1584 return false; 1585 1586 const MachineInstr *LScopeBegin = LSRange.front().first; 1587 // If the scope starts before the DBG_VALUE then we may have a negative 1588 // result. Otherwise the location is live coming into the scope and we 1589 // can skip the following checks. 1590 if (!Ordering.isBefore(DbgValue, LScopeBegin)) { 1591 // Exit if the lexical scope begins outside of the current block. 1592 if (LScopeBegin->getParent() != MBB) 1593 return false; 1594 1595 MachineBasicBlock::const_reverse_iterator Pred(DbgValue); 1596 for (++Pred; Pred != MBB->rend(); ++Pred) { 1597 if (Pred->getFlag(MachineInstr::FrameSetup)) 1598 break; 1599 auto PredDL = Pred->getDebugLoc(); 1600 if (!PredDL || Pred->isMetaInstruction()) 1601 continue; 1602 // Check whether the instruction preceding the DBG_VALUE is in the same 1603 // (sub)scope as the DBG_VALUE. 1604 if (DL->getScope() == PredDL->getScope()) 1605 return false; 1606 auto *PredScope = LScopes.findLexicalScope(PredDL); 1607 if (!PredScope || LScope->dominates(PredScope)) 1608 return false; 1609 } 1610 } 1611 1612 // If the range of the DBG_VALUE is open-ended, report success. 1613 if (!RangeEnd) 1614 return true; 1615 1616 // Single, constant DBG_VALUEs in the prologue are promoted to be live 1617 // throughout the function. This is a hack, presumably for DWARF v2 and not 1618 // necessarily correct. It would be much better to use a dbg.declare instead 1619 // if we know the constant is live throughout the scope. 1620 if (MBB->pred_empty() && 1621 all_of(DbgValue->debug_operands(), 1622 [](const MachineOperand &Op) { return Op.isImm(); })) 1623 return true; 1624 1625 // Test if the location terminates before the end of the scope. 1626 const MachineInstr *LScopeEnd = LSRange.back().second; 1627 if (Ordering.isBefore(RangeEnd, LScopeEnd)) 1628 return false; 1629 1630 // There's a single location which starts at the scope start, and ends at or 1631 // after the scope end. 1632 return true; 1633 } 1634 1635 /// Build the location list for all DBG_VALUEs in the function that 1636 /// describe the same variable. The resulting DebugLocEntries will have 1637 /// strict monotonically increasing begin addresses and will never 1638 /// overlap. If the resulting list has only one entry that is valid 1639 /// throughout variable's scope return true. 1640 // 1641 // See the definition of DbgValueHistoryMap::Entry for an explanation of the 1642 // different kinds of history map entries. One thing to be aware of is that if 1643 // a debug value is ended by another entry (rather than being valid until the 1644 // end of the function), that entry's instruction may or may not be included in 1645 // the range, depending on if the entry is a clobbering entry (it has an 1646 // instruction that clobbers one or more preceding locations), or if it is an 1647 // (overlapping) debug value entry. This distinction can be seen in the example 1648 // below. The first debug value is ended by the clobbering entry 2, and the 1649 // second and third debug values are ended by the overlapping debug value entry 1650 // 4. 1651 // 1652 // Input: 1653 // 1654 // History map entries [type, end index, mi] 1655 // 1656 // 0 | [DbgValue, 2, DBG_VALUE $reg0, [...] (fragment 0, 32)] 1657 // 1 | | [DbgValue, 4, DBG_VALUE $reg1, [...] (fragment 32, 32)] 1658 // 2 | | [Clobber, $reg0 = [...], -, -] 1659 // 3 | | [DbgValue, 4, DBG_VALUE 123, [...] (fragment 64, 32)] 1660 // 4 [DbgValue, ~0, DBG_VALUE @g, [...] (fragment 0, 96)] 1661 // 1662 // Output [start, end) [Value...]: 1663 // 1664 // [0-1) [(reg0, fragment 0, 32)] 1665 // [1-3) [(reg0, fragment 0, 32), (reg1, fragment 32, 32)] 1666 // [3-4) [(reg1, fragment 32, 32), (123, fragment 64, 32)] 1667 // [4-) [(@g, fragment 0, 96)] 1668 bool DwarfDebug::buildLocationList(SmallVectorImpl<DebugLocEntry> &DebugLoc, 1669 const DbgValueHistoryMap::Entries &Entries) { 1670 using OpenRange = 1671 std::pair<DbgValueHistoryMap::EntryIndex, DbgValueLoc>; 1672 SmallVector<OpenRange, 4> OpenRanges; 1673 bool isSafeForSingleLocation = true; 1674 const MachineInstr *StartDebugMI = nullptr; 1675 const MachineInstr *EndMI = nullptr; 1676 1677 for (auto EB = Entries.begin(), EI = EB, EE = Entries.end(); EI != EE; ++EI) { 1678 const MachineInstr *Instr = EI->getInstr(); 1679 1680 // Remove all values that are no longer live. 1681 size_t Index = std::distance(EB, EI); 1682 erase_if(OpenRanges, [&](OpenRange &R) { return R.first <= Index; }); 1683 1684 // If we are dealing with a clobbering entry, this iteration will result in 1685 // a location list entry starting after the clobbering instruction. 1686 const MCSymbol *StartLabel = 1687 EI->isClobber() ? getLabelAfterInsn(Instr) : getLabelBeforeInsn(Instr); 1688 assert(StartLabel && 1689 "Forgot label before/after instruction starting a range!"); 1690 1691 const MCSymbol *EndLabel; 1692 if (std::next(EI) == Entries.end()) { 1693 const MachineBasicBlock &EndMBB = Asm->MF->back(); 1694 EndLabel = Asm->MBBSectionRanges[EndMBB.getSectionIDNum()].EndLabel; 1695 if (EI->isClobber()) 1696 EndMI = EI->getInstr(); 1697 } 1698 else if (std::next(EI)->isClobber()) 1699 EndLabel = getLabelAfterInsn(std::next(EI)->getInstr()); 1700 else 1701 EndLabel = getLabelBeforeInsn(std::next(EI)->getInstr()); 1702 assert(EndLabel && "Forgot label after instruction ending a range!"); 1703 1704 if (EI->isDbgValue()) 1705 LLVM_DEBUG(dbgs() << "DotDebugLoc: " << *Instr << "\n"); 1706 1707 // If this history map entry has a debug value, add that to the list of 1708 // open ranges and check if its location is valid for a single value 1709 // location. 1710 if (EI->isDbgValue()) { 1711 // Do not add undef debug values, as they are redundant information in 1712 // the location list entries. An undef debug results in an empty location 1713 // description. If there are any non-undef fragments then padding pieces 1714 // with empty location descriptions will automatically be inserted, and if 1715 // all fragments are undef then the whole location list entry is 1716 // redundant. 1717 if (!Instr->isUndefDebugValue()) { 1718 auto Value = getDebugLocValue(Instr); 1719 OpenRanges.emplace_back(EI->getEndIndex(), Value); 1720 1721 // TODO: Add support for single value fragment locations. 1722 if (Instr->getDebugExpression()->isFragment()) 1723 isSafeForSingleLocation = false; 1724 1725 if (!StartDebugMI) 1726 StartDebugMI = Instr; 1727 } else { 1728 isSafeForSingleLocation = false; 1729 } 1730 } 1731 1732 // Location list entries with empty location descriptions are redundant 1733 // information in DWARF, so do not emit those. 1734 if (OpenRanges.empty()) 1735 continue; 1736 1737 // Omit entries with empty ranges as they do not have any effect in DWARF. 1738 if (StartLabel == EndLabel) { 1739 LLVM_DEBUG(dbgs() << "Omitting location list entry with empty range.\n"); 1740 continue; 1741 } 1742 1743 SmallVector<DbgValueLoc, 4> Values; 1744 for (auto &R : OpenRanges) 1745 Values.push_back(R.second); 1746 1747 // With Basic block sections, it is posssible that the StartLabel and the 1748 // Instr are not in the same section. This happens when the StartLabel is 1749 // the function begin label and the dbg value appears in a basic block 1750 // that is not the entry. In this case, the range needs to be split to 1751 // span each individual section in the range from StartLabel to EndLabel. 1752 if (Asm->MF->hasBBSections() && StartLabel == Asm->getFunctionBegin() && 1753 !Instr->getParent()->sameSection(&Asm->MF->front())) { 1754 const MCSymbol *BeginSectionLabel = StartLabel; 1755 1756 for (const MachineBasicBlock &MBB : *Asm->MF) { 1757 if (MBB.isBeginSection() && &MBB != &Asm->MF->front()) 1758 BeginSectionLabel = MBB.getSymbol(); 1759 1760 if (MBB.sameSection(Instr->getParent())) { 1761 DebugLoc.emplace_back(BeginSectionLabel, EndLabel, Values); 1762 break; 1763 } 1764 if (MBB.isEndSection()) 1765 DebugLoc.emplace_back(BeginSectionLabel, MBB.getEndSymbol(), Values); 1766 } 1767 } else { 1768 DebugLoc.emplace_back(StartLabel, EndLabel, Values); 1769 } 1770 1771 // Attempt to coalesce the ranges of two otherwise identical 1772 // DebugLocEntries. 1773 auto CurEntry = DebugLoc.rbegin(); 1774 LLVM_DEBUG({ 1775 dbgs() << CurEntry->getValues().size() << " Values:\n"; 1776 for (auto &Value : CurEntry->getValues()) 1777 Value.dump(); 1778 dbgs() << "-----\n"; 1779 }); 1780 1781 auto PrevEntry = std::next(CurEntry); 1782 if (PrevEntry != DebugLoc.rend() && PrevEntry->MergeRanges(*CurEntry)) 1783 DebugLoc.pop_back(); 1784 } 1785 1786 if (!isSafeForSingleLocation || 1787 !validThroughout(LScopes, StartDebugMI, EndMI, getInstOrdering())) 1788 return false; 1789 1790 if (DebugLoc.size() == 1) 1791 return true; 1792 1793 if (!Asm->MF->hasBBSections()) 1794 return false; 1795 1796 // Check here to see if loclist can be merged into a single range. If not, 1797 // we must keep the split loclists per section. This does exactly what 1798 // MergeRanges does without sections. We don't actually merge the ranges 1799 // as the split ranges must be kept intact if this cannot be collapsed 1800 // into a single range. 1801 const MachineBasicBlock *RangeMBB = nullptr; 1802 if (DebugLoc[0].getBeginSym() == Asm->getFunctionBegin()) 1803 RangeMBB = &Asm->MF->front(); 1804 else 1805 RangeMBB = Entries.begin()->getInstr()->getParent(); 1806 auto *CurEntry = DebugLoc.begin(); 1807 auto *NextEntry = std::next(CurEntry); 1808 while (NextEntry != DebugLoc.end()) { 1809 // Get the last machine basic block of this section. 1810 while (!RangeMBB->isEndSection()) 1811 RangeMBB = RangeMBB->getNextNode(); 1812 if (!RangeMBB->getNextNode()) 1813 return false; 1814 // CurEntry should end the current section and NextEntry should start 1815 // the next section and the Values must match for these two ranges to be 1816 // merged. 1817 if (CurEntry->getEndSym() != RangeMBB->getEndSymbol() || 1818 NextEntry->getBeginSym() != RangeMBB->getNextNode()->getSymbol() || 1819 CurEntry->getValues() != NextEntry->getValues()) 1820 return false; 1821 RangeMBB = RangeMBB->getNextNode(); 1822 CurEntry = NextEntry; 1823 NextEntry = std::next(CurEntry); 1824 } 1825 return true; 1826 } 1827 1828 DbgEntity *DwarfDebug::createConcreteEntity(DwarfCompileUnit &TheCU, 1829 LexicalScope &Scope, 1830 const DINode *Node, 1831 const DILocation *Location, 1832 const MCSymbol *Sym) { 1833 ensureAbstractEntityIsCreatedIfScoped(TheCU, Node, Scope.getScopeNode()); 1834 if (isa<const DILocalVariable>(Node)) { 1835 ConcreteEntities.push_back( 1836 std::make_unique<DbgVariable>(cast<const DILocalVariable>(Node), 1837 Location)); 1838 InfoHolder.addScopeVariable(&Scope, 1839 cast<DbgVariable>(ConcreteEntities.back().get())); 1840 } else if (isa<const DILabel>(Node)) { 1841 ConcreteEntities.push_back( 1842 std::make_unique<DbgLabel>(cast<const DILabel>(Node), 1843 Location, Sym)); 1844 InfoHolder.addScopeLabel(&Scope, 1845 cast<DbgLabel>(ConcreteEntities.back().get())); 1846 } 1847 return ConcreteEntities.back().get(); 1848 } 1849 1850 // Find variables for each lexical scope. 1851 void DwarfDebug::collectEntityInfo(DwarfCompileUnit &TheCU, 1852 const DISubprogram *SP, 1853 DenseSet<InlinedEntity> &Processed) { 1854 // Grab the variable info that was squirreled away in the MMI side-table. 1855 collectVariableInfoFromMFTable(TheCU, Processed); 1856 1857 for (const auto &I : DbgValues) { 1858 InlinedEntity IV = I.first; 1859 if (Processed.count(IV)) 1860 continue; 1861 1862 // Instruction ranges, specifying where IV is accessible. 1863 const auto &HistoryMapEntries = I.second; 1864 1865 // Try to find any non-empty variable location. Do not create a concrete 1866 // entity if there are no locations. 1867 if (!DbgValues.hasNonEmptyLocation(HistoryMapEntries)) 1868 continue; 1869 1870 LexicalScope *Scope = nullptr; 1871 const DILocalVariable *LocalVar = cast<DILocalVariable>(IV.first); 1872 if (const DILocation *IA = IV.second) 1873 Scope = LScopes.findInlinedScope(LocalVar->getScope(), IA); 1874 else 1875 Scope = LScopes.findLexicalScope(LocalVar->getScope()); 1876 // If variable scope is not found then skip this variable. 1877 if (!Scope) 1878 continue; 1879 1880 Processed.insert(IV); 1881 DbgVariable *RegVar = cast<DbgVariable>(createConcreteEntity(TheCU, 1882 *Scope, LocalVar, IV.second)); 1883 1884 const MachineInstr *MInsn = HistoryMapEntries.front().getInstr(); 1885 assert(MInsn->isDebugValue() && "History must begin with debug value"); 1886 1887 // Check if there is a single DBG_VALUE, valid throughout the var's scope. 1888 // If the history map contains a single debug value, there may be an 1889 // additional entry which clobbers the debug value. 1890 size_t HistSize = HistoryMapEntries.size(); 1891 bool SingleValueWithClobber = 1892 HistSize == 2 && HistoryMapEntries[1].isClobber(); 1893 if (HistSize == 1 || SingleValueWithClobber) { 1894 const auto *End = 1895 SingleValueWithClobber ? HistoryMapEntries[1].getInstr() : nullptr; 1896 if (validThroughout(LScopes, MInsn, End, getInstOrdering())) { 1897 RegVar->initializeDbgValue(MInsn); 1898 continue; 1899 } 1900 } 1901 1902 // Do not emit location lists if .debug_loc secton is disabled. 1903 if (!useLocSection()) 1904 continue; 1905 1906 // Handle multiple DBG_VALUE instructions describing one variable. 1907 DebugLocStream::ListBuilder List(DebugLocs, TheCU, *Asm, *RegVar, *MInsn); 1908 1909 // Build the location list for this variable. 1910 SmallVector<DebugLocEntry, 8> Entries; 1911 bool isValidSingleLocation = buildLocationList(Entries, HistoryMapEntries); 1912 1913 // Check whether buildLocationList managed to merge all locations to one 1914 // that is valid throughout the variable's scope. If so, produce single 1915 // value location. 1916 if (isValidSingleLocation) { 1917 RegVar->initializeDbgValue(Entries[0].getValues()[0]); 1918 continue; 1919 } 1920 1921 // If the variable has a DIBasicType, extract it. Basic types cannot have 1922 // unique identifiers, so don't bother resolving the type with the 1923 // identifier map. 1924 const DIBasicType *BT = dyn_cast<DIBasicType>( 1925 static_cast<const Metadata *>(LocalVar->getType())); 1926 1927 // Finalize the entry by lowering it into a DWARF bytestream. 1928 for (auto &Entry : Entries) 1929 Entry.finalize(*Asm, List, BT, TheCU); 1930 } 1931 1932 // For each InlinedEntity collected from DBG_LABEL instructions, convert to 1933 // DWARF-related DbgLabel. 1934 for (const auto &I : DbgLabels) { 1935 InlinedEntity IL = I.first; 1936 const MachineInstr *MI = I.second; 1937 if (MI == nullptr) 1938 continue; 1939 1940 LexicalScope *Scope = nullptr; 1941 const DILabel *Label = cast<DILabel>(IL.first); 1942 // The scope could have an extra lexical block file. 1943 const DILocalScope *LocalScope = 1944 Label->getScope()->getNonLexicalBlockFileScope(); 1945 // Get inlined DILocation if it is inlined label. 1946 if (const DILocation *IA = IL.second) 1947 Scope = LScopes.findInlinedScope(LocalScope, IA); 1948 else 1949 Scope = LScopes.findLexicalScope(LocalScope); 1950 // If label scope is not found then skip this label. 1951 if (!Scope) 1952 continue; 1953 1954 Processed.insert(IL); 1955 /// At this point, the temporary label is created. 1956 /// Save the temporary label to DbgLabel entity to get the 1957 /// actually address when generating Dwarf DIE. 1958 MCSymbol *Sym = getLabelBeforeInsn(MI); 1959 createConcreteEntity(TheCU, *Scope, Label, IL.second, Sym); 1960 } 1961 1962 // Collect info for variables/labels that were optimized out. 1963 for (const DINode *DN : SP->getRetainedNodes()) { 1964 if (!Processed.insert(InlinedEntity(DN, nullptr)).second) 1965 continue; 1966 LexicalScope *Scope = nullptr; 1967 if (auto *DV = dyn_cast<DILocalVariable>(DN)) { 1968 Scope = LScopes.findLexicalScope(DV->getScope()); 1969 } else if (auto *DL = dyn_cast<DILabel>(DN)) { 1970 Scope = LScopes.findLexicalScope(DL->getScope()); 1971 } 1972 1973 if (Scope) 1974 createConcreteEntity(TheCU, *Scope, DN, nullptr); 1975 } 1976 } 1977 1978 // Process beginning of an instruction. 1979 void DwarfDebug::beginInstruction(const MachineInstr *MI) { 1980 const MachineFunction &MF = *MI->getMF(); 1981 const auto *SP = MF.getFunction().getSubprogram(); 1982 bool NoDebug = 1983 !SP || SP->getUnit()->getEmissionKind() == DICompileUnit::NoDebug; 1984 1985 // Delay slot support check. 1986 auto delaySlotSupported = [](const MachineInstr &MI) { 1987 if (!MI.isBundledWithSucc()) 1988 return false; 1989 auto Suc = std::next(MI.getIterator()); 1990 (void)Suc; 1991 // Ensure that delay slot instruction is successor of the call instruction. 1992 // Ex. CALL_INSTRUCTION { 1993 // DELAY_SLOT_INSTRUCTION } 1994 assert(Suc->isBundledWithPred() && 1995 "Call bundle instructions are out of order"); 1996 return true; 1997 }; 1998 1999 // When describing calls, we need a label for the call instruction. 2000 if (!NoDebug && SP->areAllCallsDescribed() && 2001 MI->isCandidateForCallSiteEntry(MachineInstr::AnyInBundle) && 2002 (!MI->hasDelaySlot() || delaySlotSupported(*MI))) { 2003 const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo(); 2004 bool IsTail = TII->isTailCall(*MI); 2005 // For tail calls, we need the address of the branch instruction for 2006 // DW_AT_call_pc. 2007 if (IsTail) 2008 requestLabelBeforeInsn(MI); 2009 // For non-tail calls, we need the return address for the call for 2010 // DW_AT_call_return_pc. Under GDB tuning, this information is needed for 2011 // tail calls as well. 2012 requestLabelAfterInsn(MI); 2013 } 2014 2015 DebugHandlerBase::beginInstruction(MI); 2016 if (!CurMI) 2017 return; 2018 2019 if (NoDebug) 2020 return; 2021 2022 // Check if source location changes, but ignore DBG_VALUE and CFI locations. 2023 // If the instruction is part of the function frame setup code, do not emit 2024 // any line record, as there is no correspondence with any user code. 2025 if (MI->isMetaInstruction() || MI->getFlag(MachineInstr::FrameSetup)) 2026 return; 2027 const DebugLoc &DL = MI->getDebugLoc(); 2028 // When we emit a line-0 record, we don't update PrevInstLoc; so look at 2029 // the last line number actually emitted, to see if it was line 0. 2030 unsigned LastAsmLine = 2031 Asm->OutStreamer->getContext().getCurrentDwarfLoc().getLine(); 2032 2033 if (DL == PrevInstLoc) { 2034 // If we have an ongoing unspecified location, nothing to do here. 2035 if (!DL) 2036 return; 2037 // We have an explicit location, same as the previous location. 2038 // But we might be coming back to it after a line 0 record. 2039 if (LastAsmLine == 0 && DL.getLine() != 0) { 2040 // Reinstate the source location but not marked as a statement. 2041 const MDNode *Scope = DL.getScope(); 2042 recordSourceLine(DL.getLine(), DL.getCol(), Scope, /*Flags=*/0); 2043 } 2044 return; 2045 } 2046 2047 if (!DL) { 2048 // We have an unspecified location, which might want to be line 0. 2049 // If we have already emitted a line-0 record, don't repeat it. 2050 if (LastAsmLine == 0) 2051 return; 2052 // If user said Don't Do That, don't do that. 2053 if (UnknownLocations == Disable) 2054 return; 2055 // See if we have a reason to emit a line-0 record now. 2056 // Reasons to emit a line-0 record include: 2057 // - User asked for it (UnknownLocations). 2058 // - Instruction has a label, so it's referenced from somewhere else, 2059 // possibly debug information; we want it to have a source location. 2060 // - Instruction is at the top of a block; we don't want to inherit the 2061 // location from the physically previous (maybe unrelated) block. 2062 if (UnknownLocations == Enable || PrevLabel || 2063 (PrevInstBB && PrevInstBB != MI->getParent())) { 2064 // Preserve the file and column numbers, if we can, to save space in 2065 // the encoded line table. 2066 // Do not update PrevInstLoc, it remembers the last non-0 line. 2067 const MDNode *Scope = nullptr; 2068 unsigned Column = 0; 2069 if (PrevInstLoc) { 2070 Scope = PrevInstLoc.getScope(); 2071 Column = PrevInstLoc.getCol(); 2072 } 2073 recordSourceLine(/*Line=*/0, Column, Scope, /*Flags=*/0); 2074 } 2075 return; 2076 } 2077 2078 // We have an explicit location, different from the previous location. 2079 // Don't repeat a line-0 record, but otherwise emit the new location. 2080 // (The new location might be an explicit line 0, which we do emit.) 2081 if (DL.getLine() == 0 && LastAsmLine == 0) 2082 return; 2083 unsigned Flags = 0; 2084 if (DL == PrologEndLoc) { 2085 Flags |= DWARF2_FLAG_PROLOGUE_END | DWARF2_FLAG_IS_STMT; 2086 PrologEndLoc = DebugLoc(); 2087 } 2088 // If the line changed, we call that a new statement; unless we went to 2089 // line 0 and came back, in which case it is not a new statement. 2090 unsigned OldLine = PrevInstLoc ? PrevInstLoc.getLine() : LastAsmLine; 2091 if (DL.getLine() && DL.getLine() != OldLine) 2092 Flags |= DWARF2_FLAG_IS_STMT; 2093 2094 const MDNode *Scope = DL.getScope(); 2095 recordSourceLine(DL.getLine(), DL.getCol(), Scope, Flags); 2096 2097 // If we're not at line 0, remember this location. 2098 if (DL.getLine()) 2099 PrevInstLoc = DL; 2100 } 2101 2102 static DebugLoc findPrologueEndLoc(const MachineFunction *MF) { 2103 // First known non-DBG_VALUE and non-frame setup location marks 2104 // the beginning of the function body. 2105 DebugLoc LineZeroLoc; 2106 for (const auto &MBB : *MF) { 2107 for (const auto &MI : MBB) { 2108 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 2109 MI.getDebugLoc()) { 2110 // Scan forward to try to find a non-zero line number. The prologue_end 2111 // marks the first breakpoint in the function after the frame setup, and 2112 // a compiler-generated line 0 location is not a meaningful breakpoint. 2113 // If none is found, return the first location after the frame setup. 2114 if (MI.getDebugLoc().getLine()) 2115 return MI.getDebugLoc(); 2116 LineZeroLoc = MI.getDebugLoc(); 2117 } 2118 } 2119 } 2120 return LineZeroLoc; 2121 } 2122 2123 /// Register a source line with debug info. Returns the unique label that was 2124 /// emitted and which provides correspondence to the source line list. 2125 static void recordSourceLine(AsmPrinter &Asm, unsigned Line, unsigned Col, 2126 const MDNode *S, unsigned Flags, unsigned CUID, 2127 uint16_t DwarfVersion, 2128 ArrayRef<std::unique_ptr<DwarfCompileUnit>> DCUs) { 2129 StringRef Fn; 2130 unsigned FileNo = 1; 2131 unsigned Discriminator = 0; 2132 if (auto *Scope = cast_or_null<DIScope>(S)) { 2133 Fn = Scope->getFilename(); 2134 if (Line != 0 && DwarfVersion >= 4) 2135 if (auto *LBF = dyn_cast<DILexicalBlockFile>(Scope)) 2136 Discriminator = LBF->getDiscriminator(); 2137 2138 FileNo = static_cast<DwarfCompileUnit &>(*DCUs[CUID]) 2139 .getOrCreateSourceID(Scope->getFile()); 2140 } 2141 Asm.OutStreamer->emitDwarfLocDirective(FileNo, Line, Col, Flags, 0, 2142 Discriminator, Fn); 2143 } 2144 2145 DebugLoc DwarfDebug::emitInitialLocDirective(const MachineFunction &MF, 2146 unsigned CUID) { 2147 // Get beginning of function. 2148 if (DebugLoc PrologEndLoc = findPrologueEndLoc(&MF)) { 2149 // Ensure the compile unit is created if the function is called before 2150 // beginFunction(). 2151 (void)getOrCreateDwarfCompileUnit( 2152 MF.getFunction().getSubprogram()->getUnit()); 2153 // We'd like to list the prologue as "not statements" but GDB behaves 2154 // poorly if we do that. Revisit this with caution/GDB (7.5+) testing. 2155 const DISubprogram *SP = PrologEndLoc->getInlinedAtScope()->getSubprogram(); 2156 ::recordSourceLine(*Asm, SP->getScopeLine(), 0, SP, DWARF2_FLAG_IS_STMT, 2157 CUID, getDwarfVersion(), getUnits()); 2158 return PrologEndLoc; 2159 } 2160 return DebugLoc(); 2161 } 2162 2163 // Gather pre-function debug information. Assumes being called immediately 2164 // after the function entry point has been emitted. 2165 void DwarfDebug::beginFunctionImpl(const MachineFunction *MF) { 2166 CurFn = MF; 2167 2168 auto *SP = MF->getFunction().getSubprogram(); 2169 assert(LScopes.empty() || SP == LScopes.getCurrentFunctionScope()->getScopeNode()); 2170 if (SP->getUnit()->getEmissionKind() == DICompileUnit::NoDebug) 2171 return; 2172 2173 DwarfCompileUnit &CU = getOrCreateDwarfCompileUnit(SP->getUnit()); 2174 2175 Asm->OutStreamer->getContext().setDwarfCompileUnitID( 2176 getDwarfCompileUnitIDForLineTable(CU)); 2177 2178 // Record beginning of function. 2179 PrologEndLoc = emitInitialLocDirective( 2180 *MF, Asm->OutStreamer->getContext().getDwarfCompileUnitID()); 2181 } 2182 2183 unsigned 2184 DwarfDebug::getDwarfCompileUnitIDForLineTable(const DwarfCompileUnit &CU) { 2185 // Set DwarfDwarfCompileUnitID in MCContext to the Compile Unit this function 2186 // belongs to so that we add to the correct per-cu line table in the 2187 // non-asm case. 2188 if (Asm->OutStreamer->hasRawTextSupport()) 2189 // Use a single line table if we are generating assembly. 2190 return 0; 2191 else 2192 return CU.getUniqueID(); 2193 } 2194 2195 void DwarfDebug::terminateLineTable(const DwarfCompileUnit *CU) { 2196 const auto &CURanges = CU->getRanges(); 2197 auto &LineTable = Asm->OutStreamer->getContext().getMCDwarfLineTable( 2198 getDwarfCompileUnitIDForLineTable(*CU)); 2199 // Add the last range label for the given CU. 2200 LineTable.getMCLineSections().addEndEntry( 2201 const_cast<MCSymbol *>(CURanges.back().End)); 2202 } 2203 2204 void DwarfDebug::skippedNonDebugFunction() { 2205 // If we don't have a subprogram for this function then there will be a hole 2206 // in the range information. Keep note of this by setting the previously used 2207 // section to nullptr. 2208 // Terminate the pending line table. 2209 if (PrevCU) 2210 terminateLineTable(PrevCU); 2211 PrevCU = nullptr; 2212 CurFn = nullptr; 2213 } 2214 2215 // Gather and emit post-function debug information. 2216 void DwarfDebug::endFunctionImpl(const MachineFunction *MF) { 2217 const DISubprogram *SP = MF->getFunction().getSubprogram(); 2218 2219 assert(CurFn == MF && 2220 "endFunction should be called with the same function as beginFunction"); 2221 2222 // Set DwarfDwarfCompileUnitID in MCContext to default value. 2223 Asm->OutStreamer->getContext().setDwarfCompileUnitID(0); 2224 2225 LexicalScope *FnScope = LScopes.getCurrentFunctionScope(); 2226 assert(!FnScope || SP == FnScope->getScopeNode()); 2227 DwarfCompileUnit &TheCU = *CUMap.lookup(SP->getUnit()); 2228 if (TheCU.getCUNode()->isDebugDirectivesOnly()) { 2229 PrevLabel = nullptr; 2230 CurFn = nullptr; 2231 return; 2232 } 2233 2234 DenseSet<InlinedEntity> Processed; 2235 collectEntityInfo(TheCU, SP, Processed); 2236 2237 // Add the range of this function to the list of ranges for the CU. 2238 // With basic block sections, add ranges for all basic block sections. 2239 for (const auto &R : Asm->MBBSectionRanges) 2240 TheCU.addRange({R.second.BeginLabel, R.second.EndLabel}); 2241 2242 // Under -gmlt, skip building the subprogram if there are no inlined 2243 // subroutines inside it. But with -fdebug-info-for-profiling, the subprogram 2244 // is still needed as we need its source location. 2245 if (!TheCU.getCUNode()->getDebugInfoForProfiling() && 2246 TheCU.getCUNode()->getEmissionKind() == DICompileUnit::LineTablesOnly && 2247 LScopes.getAbstractScopesList().empty() && !IsDarwin) { 2248 assert(InfoHolder.getScopeVariables().empty()); 2249 PrevLabel = nullptr; 2250 CurFn = nullptr; 2251 return; 2252 } 2253 2254 #ifndef NDEBUG 2255 size_t NumAbstractScopes = LScopes.getAbstractScopesList().size(); 2256 #endif 2257 // Construct abstract scopes. 2258 for (LexicalScope *AScope : LScopes.getAbstractScopesList()) { 2259 auto *SP = cast<DISubprogram>(AScope->getScopeNode()); 2260 for (const DINode *DN : SP->getRetainedNodes()) { 2261 if (!Processed.insert(InlinedEntity(DN, nullptr)).second) 2262 continue; 2263 2264 const MDNode *Scope = nullptr; 2265 if (auto *DV = dyn_cast<DILocalVariable>(DN)) 2266 Scope = DV->getScope(); 2267 else if (auto *DL = dyn_cast<DILabel>(DN)) 2268 Scope = DL->getScope(); 2269 else 2270 llvm_unreachable("Unexpected DI type!"); 2271 2272 // Collect info for variables/labels that were optimized out. 2273 ensureAbstractEntityIsCreated(TheCU, DN, Scope); 2274 assert(LScopes.getAbstractScopesList().size() == NumAbstractScopes 2275 && "ensureAbstractEntityIsCreated inserted abstract scopes"); 2276 } 2277 constructAbstractSubprogramScopeDIE(TheCU, AScope); 2278 } 2279 2280 ProcessedSPNodes.insert(SP); 2281 DIE &ScopeDIE = TheCU.constructSubprogramScopeDIE(SP, FnScope); 2282 if (auto *SkelCU = TheCU.getSkeleton()) 2283 if (!LScopes.getAbstractScopesList().empty() && 2284 TheCU.getCUNode()->getSplitDebugInlining()) 2285 SkelCU->constructSubprogramScopeDIE(SP, FnScope); 2286 2287 // Construct call site entries. 2288 constructCallSiteEntryDIEs(*SP, TheCU, ScopeDIE, *MF); 2289 2290 // Clear debug info 2291 // Ownership of DbgVariables is a bit subtle - ScopeVariables owns all the 2292 // DbgVariables except those that are also in AbstractVariables (since they 2293 // can be used cross-function) 2294 InfoHolder.getScopeVariables().clear(); 2295 InfoHolder.getScopeLabels().clear(); 2296 PrevLabel = nullptr; 2297 CurFn = nullptr; 2298 } 2299 2300 // Register a source line with debug info. Returns the unique label that was 2301 // emitted and which provides correspondence to the source line list. 2302 void DwarfDebug::recordSourceLine(unsigned Line, unsigned Col, const MDNode *S, 2303 unsigned Flags) { 2304 ::recordSourceLine(*Asm, Line, Col, S, Flags, 2305 Asm->OutStreamer->getContext().getDwarfCompileUnitID(), 2306 getDwarfVersion(), getUnits()); 2307 } 2308 2309 //===----------------------------------------------------------------------===// 2310 // Emit Methods 2311 //===----------------------------------------------------------------------===// 2312 2313 // Emit the debug info section. 2314 void DwarfDebug::emitDebugInfo() { 2315 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 2316 Holder.emitUnits(/* UseOffsets */ false); 2317 } 2318 2319 // Emit the abbreviation section. 2320 void DwarfDebug::emitAbbreviations() { 2321 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 2322 2323 Holder.emitAbbrevs(Asm->getObjFileLowering().getDwarfAbbrevSection()); 2324 } 2325 2326 void DwarfDebug::emitStringOffsetsTableHeader() { 2327 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 2328 Holder.getStringPool().emitStringOffsetsTableHeader( 2329 *Asm, Asm->getObjFileLowering().getDwarfStrOffSection(), 2330 Holder.getStringOffsetsStartSym()); 2331 } 2332 2333 template <typename AccelTableT> 2334 void DwarfDebug::emitAccel(AccelTableT &Accel, MCSection *Section, 2335 StringRef TableName) { 2336 Asm->OutStreamer->SwitchSection(Section); 2337 2338 // Emit the full data. 2339 emitAppleAccelTable(Asm, Accel, TableName, Section->getBeginSymbol()); 2340 } 2341 2342 void DwarfDebug::emitAccelDebugNames() { 2343 // Don't emit anything if we have no compilation units to index. 2344 if (getUnits().empty()) 2345 return; 2346 2347 emitDWARF5AccelTable(Asm, AccelDebugNames, *this, getUnits()); 2348 } 2349 2350 // Emit visible names into a hashed accelerator table section. 2351 void DwarfDebug::emitAccelNames() { 2352 emitAccel(AccelNames, Asm->getObjFileLowering().getDwarfAccelNamesSection(), 2353 "Names"); 2354 } 2355 2356 // Emit objective C classes and categories into a hashed accelerator table 2357 // section. 2358 void DwarfDebug::emitAccelObjC() { 2359 emitAccel(AccelObjC, Asm->getObjFileLowering().getDwarfAccelObjCSection(), 2360 "ObjC"); 2361 } 2362 2363 // Emit namespace dies into a hashed accelerator table. 2364 void DwarfDebug::emitAccelNamespaces() { 2365 emitAccel(AccelNamespace, 2366 Asm->getObjFileLowering().getDwarfAccelNamespaceSection(), 2367 "namespac"); 2368 } 2369 2370 // Emit type dies into a hashed accelerator table. 2371 void DwarfDebug::emitAccelTypes() { 2372 emitAccel(AccelTypes, Asm->getObjFileLowering().getDwarfAccelTypesSection(), 2373 "types"); 2374 } 2375 2376 // Public name handling. 2377 // The format for the various pubnames: 2378 // 2379 // dwarf pubnames - offset/name pairs where the offset is the offset into the CU 2380 // for the DIE that is named. 2381 // 2382 // gnu pubnames - offset/index value/name tuples where the offset is the offset 2383 // into the CU and the index value is computed according to the type of value 2384 // for the DIE that is named. 2385 // 2386 // For type units the offset is the offset of the skeleton DIE. For split dwarf 2387 // it's the offset within the debug_info/debug_types dwo section, however, the 2388 // reference in the pubname header doesn't change. 2389 2390 /// computeIndexValue - Compute the gdb index value for the DIE and CU. 2391 static dwarf::PubIndexEntryDescriptor computeIndexValue(DwarfUnit *CU, 2392 const DIE *Die) { 2393 // Entities that ended up only in a Type Unit reference the CU instead (since 2394 // the pub entry has offsets within the CU there's no real offset that can be 2395 // provided anyway). As it happens all such entities (namespaces and types, 2396 // types only in C++ at that) are rendered as TYPE+EXTERNAL. If this turns out 2397 // not to be true it would be necessary to persist this information from the 2398 // point at which the entry is added to the index data structure - since by 2399 // the time the index is built from that, the original type/namespace DIE in a 2400 // type unit has already been destroyed so it can't be queried for properties 2401 // like tag, etc. 2402 if (Die->getTag() == dwarf::DW_TAG_compile_unit) 2403 return dwarf::PubIndexEntryDescriptor(dwarf::GIEK_TYPE, 2404 dwarf::GIEL_EXTERNAL); 2405 dwarf::GDBIndexEntryLinkage Linkage = dwarf::GIEL_STATIC; 2406 2407 // We could have a specification DIE that has our most of our knowledge, 2408 // look for that now. 2409 if (DIEValue SpecVal = Die->findAttribute(dwarf::DW_AT_specification)) { 2410 DIE &SpecDIE = SpecVal.getDIEEntry().getEntry(); 2411 if (SpecDIE.findAttribute(dwarf::DW_AT_external)) 2412 Linkage = dwarf::GIEL_EXTERNAL; 2413 } else if (Die->findAttribute(dwarf::DW_AT_external)) 2414 Linkage = dwarf::GIEL_EXTERNAL; 2415 2416 switch (Die->getTag()) { 2417 case dwarf::DW_TAG_class_type: 2418 case dwarf::DW_TAG_structure_type: 2419 case dwarf::DW_TAG_union_type: 2420 case dwarf::DW_TAG_enumeration_type: 2421 return dwarf::PubIndexEntryDescriptor( 2422 dwarf::GIEK_TYPE, 2423 dwarf::isCPlusPlus((dwarf::SourceLanguage)CU->getLanguage()) 2424 ? dwarf::GIEL_EXTERNAL 2425 : dwarf::GIEL_STATIC); 2426 case dwarf::DW_TAG_typedef: 2427 case dwarf::DW_TAG_base_type: 2428 case dwarf::DW_TAG_subrange_type: 2429 return dwarf::PubIndexEntryDescriptor(dwarf::GIEK_TYPE, dwarf::GIEL_STATIC); 2430 case dwarf::DW_TAG_namespace: 2431 return dwarf::GIEK_TYPE; 2432 case dwarf::DW_TAG_subprogram: 2433 return dwarf::PubIndexEntryDescriptor(dwarf::GIEK_FUNCTION, Linkage); 2434 case dwarf::DW_TAG_variable: 2435 return dwarf::PubIndexEntryDescriptor(dwarf::GIEK_VARIABLE, Linkage); 2436 case dwarf::DW_TAG_enumerator: 2437 return dwarf::PubIndexEntryDescriptor(dwarf::GIEK_VARIABLE, 2438 dwarf::GIEL_STATIC); 2439 default: 2440 return dwarf::GIEK_NONE; 2441 } 2442 } 2443 2444 /// emitDebugPubSections - Emit visible names and types into debug pubnames and 2445 /// pubtypes sections. 2446 void DwarfDebug::emitDebugPubSections() { 2447 for (const auto &NU : CUMap) { 2448 DwarfCompileUnit *TheU = NU.second; 2449 if (!TheU->hasDwarfPubSections()) 2450 continue; 2451 2452 bool GnuStyle = TheU->getCUNode()->getNameTableKind() == 2453 DICompileUnit::DebugNameTableKind::GNU; 2454 2455 Asm->OutStreamer->SwitchSection( 2456 GnuStyle ? Asm->getObjFileLowering().getDwarfGnuPubNamesSection() 2457 : Asm->getObjFileLowering().getDwarfPubNamesSection()); 2458 emitDebugPubSection(GnuStyle, "Names", TheU, TheU->getGlobalNames()); 2459 2460 Asm->OutStreamer->SwitchSection( 2461 GnuStyle ? Asm->getObjFileLowering().getDwarfGnuPubTypesSection() 2462 : Asm->getObjFileLowering().getDwarfPubTypesSection()); 2463 emitDebugPubSection(GnuStyle, "Types", TheU, TheU->getGlobalTypes()); 2464 } 2465 } 2466 2467 void DwarfDebug::emitSectionReference(const DwarfCompileUnit &CU) { 2468 if (useSectionsAsReferences()) 2469 Asm->emitDwarfOffset(CU.getSection()->getBeginSymbol(), 2470 CU.getDebugSectionOffset()); 2471 else 2472 Asm->emitDwarfSymbolReference(CU.getLabelBegin()); 2473 } 2474 2475 void DwarfDebug::emitDebugPubSection(bool GnuStyle, StringRef Name, 2476 DwarfCompileUnit *TheU, 2477 const StringMap<const DIE *> &Globals) { 2478 if (auto *Skeleton = TheU->getSkeleton()) 2479 TheU = Skeleton; 2480 2481 // Emit the header. 2482 MCSymbol *EndLabel = Asm->emitDwarfUnitLength( 2483 "pub" + Name, "Length of Public " + Name + " Info"); 2484 2485 Asm->OutStreamer->AddComment("DWARF Version"); 2486 Asm->emitInt16(dwarf::DW_PUBNAMES_VERSION); 2487 2488 Asm->OutStreamer->AddComment("Offset of Compilation Unit Info"); 2489 emitSectionReference(*TheU); 2490 2491 Asm->OutStreamer->AddComment("Compilation Unit Length"); 2492 Asm->emitDwarfLengthOrOffset(TheU->getLength()); 2493 2494 // Emit the pubnames for this compilation unit. 2495 for (const auto &GI : Globals) { 2496 const char *Name = GI.getKeyData(); 2497 const DIE *Entity = GI.second; 2498 2499 Asm->OutStreamer->AddComment("DIE offset"); 2500 Asm->emitDwarfLengthOrOffset(Entity->getOffset()); 2501 2502 if (GnuStyle) { 2503 dwarf::PubIndexEntryDescriptor Desc = computeIndexValue(TheU, Entity); 2504 Asm->OutStreamer->AddComment( 2505 Twine("Attributes: ") + dwarf::GDBIndexEntryKindString(Desc.Kind) + 2506 ", " + dwarf::GDBIndexEntryLinkageString(Desc.Linkage)); 2507 Asm->emitInt8(Desc.toBits()); 2508 } 2509 2510 Asm->OutStreamer->AddComment("External Name"); 2511 Asm->OutStreamer->emitBytes(StringRef(Name, GI.getKeyLength() + 1)); 2512 } 2513 2514 Asm->OutStreamer->AddComment("End Mark"); 2515 Asm->emitDwarfLengthOrOffset(0); 2516 Asm->OutStreamer->emitLabel(EndLabel); 2517 } 2518 2519 /// Emit null-terminated strings into a debug str section. 2520 void DwarfDebug::emitDebugStr() { 2521 MCSection *StringOffsetsSection = nullptr; 2522 if (useSegmentedStringOffsetsTable()) { 2523 emitStringOffsetsTableHeader(); 2524 StringOffsetsSection = Asm->getObjFileLowering().getDwarfStrOffSection(); 2525 } 2526 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 2527 Holder.emitStrings(Asm->getObjFileLowering().getDwarfStrSection(), 2528 StringOffsetsSection, /* UseRelativeOffsets = */ true); 2529 } 2530 2531 void DwarfDebug::emitDebugLocEntry(ByteStreamer &Streamer, 2532 const DebugLocStream::Entry &Entry, 2533 const DwarfCompileUnit *CU) { 2534 auto &&Comments = DebugLocs.getComments(Entry); 2535 auto Comment = Comments.begin(); 2536 auto End = Comments.end(); 2537 2538 // The expressions are inserted into a byte stream rather early (see 2539 // DwarfExpression::addExpression) so for those ops (e.g. DW_OP_convert) that 2540 // need to reference a base_type DIE the offset of that DIE is not yet known. 2541 // To deal with this we instead insert a placeholder early and then extract 2542 // it here and replace it with the real reference. 2543 unsigned PtrSize = Asm->MAI->getCodePointerSize(); 2544 DWARFDataExtractor Data(StringRef(DebugLocs.getBytes(Entry).data(), 2545 DebugLocs.getBytes(Entry).size()), 2546 Asm->getDataLayout().isLittleEndian(), PtrSize); 2547 DWARFExpression Expr(Data, PtrSize, Asm->OutContext.getDwarfFormat()); 2548 2549 using Encoding = DWARFExpression::Operation::Encoding; 2550 uint64_t Offset = 0; 2551 for (auto &Op : Expr) { 2552 assert(Op.getCode() != dwarf::DW_OP_const_type && 2553 "3 operand ops not yet supported"); 2554 Streamer.emitInt8(Op.getCode(), Comment != End ? *(Comment++) : ""); 2555 Offset++; 2556 for (unsigned I = 0; I < 2; ++I) { 2557 if (Op.getDescription().Op[I] == Encoding::SizeNA) 2558 continue; 2559 if (Op.getDescription().Op[I] == Encoding::BaseTypeRef) { 2560 uint64_t Offset = 2561 CU->ExprRefedBaseTypes[Op.getRawOperand(I)].Die->getOffset(); 2562 assert(Offset < (1ULL << (ULEB128PadSize * 7)) && "Offset wont fit"); 2563 Streamer.emitULEB128(Offset, "", ULEB128PadSize); 2564 // Make sure comments stay aligned. 2565 for (unsigned J = 0; J < ULEB128PadSize; ++J) 2566 if (Comment != End) 2567 Comment++; 2568 } else { 2569 for (uint64_t J = Offset; J < Op.getOperandEndOffset(I); ++J) 2570 Streamer.emitInt8(Data.getData()[J], Comment != End ? *(Comment++) : ""); 2571 } 2572 Offset = Op.getOperandEndOffset(I); 2573 } 2574 assert(Offset == Op.getEndOffset()); 2575 } 2576 } 2577 2578 void DwarfDebug::emitDebugLocValue(const AsmPrinter &AP, const DIBasicType *BT, 2579 const DbgValueLoc &Value, 2580 DwarfExpression &DwarfExpr) { 2581 auto *DIExpr = Value.getExpression(); 2582 DIExpressionCursor ExprCursor(DIExpr); 2583 DwarfExpr.addFragmentOffset(DIExpr); 2584 2585 // If the DIExpr is is an Entry Value, we want to follow the same code path 2586 // regardless of whether the DBG_VALUE is variadic or not. 2587 if (DIExpr && DIExpr->isEntryValue()) { 2588 // Entry values can only be a single register with no additional DIExpr, 2589 // so just add it directly. 2590 assert(Value.getLocEntries().size() == 1); 2591 assert(Value.getLocEntries()[0].isLocation()); 2592 MachineLocation Location = Value.getLocEntries()[0].getLoc(); 2593 DwarfExpr.setLocation(Location, DIExpr); 2594 2595 DwarfExpr.beginEntryValueExpression(ExprCursor); 2596 2597 const TargetRegisterInfo &TRI = *AP.MF->getSubtarget().getRegisterInfo(); 2598 if (!DwarfExpr.addMachineRegExpression(TRI, ExprCursor, Location.getReg())) 2599 return; 2600 return DwarfExpr.addExpression(std::move(ExprCursor)); 2601 } 2602 2603 // Regular entry. 2604 auto EmitValueLocEntry = [&DwarfExpr, &BT, 2605 &AP](const DbgValueLocEntry &Entry, 2606 DIExpressionCursor &Cursor) -> bool { 2607 if (Entry.isInt()) { 2608 if (BT && (BT->getEncoding() == dwarf::DW_ATE_signed || 2609 BT->getEncoding() == dwarf::DW_ATE_signed_char)) 2610 DwarfExpr.addSignedConstant(Entry.getInt()); 2611 else 2612 DwarfExpr.addUnsignedConstant(Entry.getInt()); 2613 } else if (Entry.isLocation()) { 2614 MachineLocation Location = Entry.getLoc(); 2615 if (Location.isIndirect()) 2616 DwarfExpr.setMemoryLocationKind(); 2617 2618 const TargetRegisterInfo &TRI = *AP.MF->getSubtarget().getRegisterInfo(); 2619 if (!DwarfExpr.addMachineRegExpression(TRI, Cursor, Location.getReg())) 2620 return false; 2621 } else if (Entry.isTargetIndexLocation()) { 2622 TargetIndexLocation Loc = Entry.getTargetIndexLocation(); 2623 // TODO TargetIndexLocation is a target-independent. Currently only the 2624 // WebAssembly-specific encoding is supported. 2625 assert(AP.TM.getTargetTriple().isWasm()); 2626 DwarfExpr.addWasmLocation(Loc.Index, static_cast<uint64_t>(Loc.Offset)); 2627 } else if (Entry.isConstantFP()) { 2628 if (AP.getDwarfVersion() >= 4 && !AP.getDwarfDebug()->tuneForSCE() && 2629 !Cursor) { 2630 DwarfExpr.addConstantFP(Entry.getConstantFP()->getValueAPF(), AP); 2631 } else if (Entry.getConstantFP() 2632 ->getValueAPF() 2633 .bitcastToAPInt() 2634 .getBitWidth() <= 64 /*bits*/) { 2635 DwarfExpr.addUnsignedConstant( 2636 Entry.getConstantFP()->getValueAPF().bitcastToAPInt()); 2637 } else { 2638 LLVM_DEBUG( 2639 dbgs() << "Skipped DwarfExpression creation for ConstantFP of size" 2640 << Entry.getConstantFP() 2641 ->getValueAPF() 2642 .bitcastToAPInt() 2643 .getBitWidth() 2644 << " bits\n"); 2645 return false; 2646 } 2647 } 2648 return true; 2649 }; 2650 2651 if (!Value.isVariadic()) { 2652 if (!EmitValueLocEntry(Value.getLocEntries()[0], ExprCursor)) 2653 return; 2654 DwarfExpr.addExpression(std::move(ExprCursor)); 2655 return; 2656 } 2657 2658 // If any of the location entries are registers with the value 0, then the 2659 // location is undefined. 2660 if (any_of(Value.getLocEntries(), [](const DbgValueLocEntry &Entry) { 2661 return Entry.isLocation() && !Entry.getLoc().getReg(); 2662 })) 2663 return; 2664 2665 DwarfExpr.addExpression( 2666 std::move(ExprCursor), 2667 [EmitValueLocEntry, &Value](unsigned Idx, 2668 DIExpressionCursor &Cursor) -> bool { 2669 return EmitValueLocEntry(Value.getLocEntries()[Idx], Cursor); 2670 }); 2671 } 2672 2673 void DebugLocEntry::finalize(const AsmPrinter &AP, 2674 DebugLocStream::ListBuilder &List, 2675 const DIBasicType *BT, 2676 DwarfCompileUnit &TheCU) { 2677 assert(!Values.empty() && 2678 "location list entries without values are redundant"); 2679 assert(Begin != End && "unexpected location list entry with empty range"); 2680 DebugLocStream::EntryBuilder Entry(List, Begin, End); 2681 BufferByteStreamer Streamer = Entry.getStreamer(); 2682 DebugLocDwarfExpression DwarfExpr(AP.getDwarfVersion(), Streamer, TheCU); 2683 const DbgValueLoc &Value = Values[0]; 2684 if (Value.isFragment()) { 2685 // Emit all fragments that belong to the same variable and range. 2686 assert(llvm::all_of(Values, [](DbgValueLoc P) { 2687 return P.isFragment(); 2688 }) && "all values are expected to be fragments"); 2689 assert(llvm::is_sorted(Values) && "fragments are expected to be sorted"); 2690 2691 for (const auto &Fragment : Values) 2692 DwarfDebug::emitDebugLocValue(AP, BT, Fragment, DwarfExpr); 2693 2694 } else { 2695 assert(Values.size() == 1 && "only fragments may have >1 value"); 2696 DwarfDebug::emitDebugLocValue(AP, BT, Value, DwarfExpr); 2697 } 2698 DwarfExpr.finalize(); 2699 if (DwarfExpr.TagOffset) 2700 List.setTagOffset(*DwarfExpr.TagOffset); 2701 } 2702 2703 void DwarfDebug::emitDebugLocEntryLocation(const DebugLocStream::Entry &Entry, 2704 const DwarfCompileUnit *CU) { 2705 // Emit the size. 2706 Asm->OutStreamer->AddComment("Loc expr size"); 2707 if (getDwarfVersion() >= 5) 2708 Asm->emitULEB128(DebugLocs.getBytes(Entry).size()); 2709 else if (DebugLocs.getBytes(Entry).size() <= std::numeric_limits<uint16_t>::max()) 2710 Asm->emitInt16(DebugLocs.getBytes(Entry).size()); 2711 else { 2712 // The entry is too big to fit into 16 bit, drop it as there is nothing we 2713 // can do. 2714 Asm->emitInt16(0); 2715 return; 2716 } 2717 // Emit the entry. 2718 APByteStreamer Streamer(*Asm); 2719 emitDebugLocEntry(Streamer, Entry, CU); 2720 } 2721 2722 // Emit the header of a DWARF 5 range list table list table. Returns the symbol 2723 // that designates the end of the table for the caller to emit when the table is 2724 // complete. 2725 static MCSymbol *emitRnglistsTableHeader(AsmPrinter *Asm, 2726 const DwarfFile &Holder) { 2727 MCSymbol *TableEnd = mcdwarf::emitListsTableHeaderStart(*Asm->OutStreamer); 2728 2729 Asm->OutStreamer->AddComment("Offset entry count"); 2730 Asm->emitInt32(Holder.getRangeLists().size()); 2731 Asm->OutStreamer->emitLabel(Holder.getRnglistsTableBaseSym()); 2732 2733 for (const RangeSpanList &List : Holder.getRangeLists()) 2734 Asm->emitLabelDifference(List.Label, Holder.getRnglistsTableBaseSym(), 2735 Asm->getDwarfOffsetByteSize()); 2736 2737 return TableEnd; 2738 } 2739 2740 // Emit the header of a DWARF 5 locations list table. Returns the symbol that 2741 // designates the end of the table for the caller to emit when the table is 2742 // complete. 2743 static MCSymbol *emitLoclistsTableHeader(AsmPrinter *Asm, 2744 const DwarfDebug &DD) { 2745 MCSymbol *TableEnd = mcdwarf::emitListsTableHeaderStart(*Asm->OutStreamer); 2746 2747 const auto &DebugLocs = DD.getDebugLocs(); 2748 2749 Asm->OutStreamer->AddComment("Offset entry count"); 2750 Asm->emitInt32(DebugLocs.getLists().size()); 2751 Asm->OutStreamer->emitLabel(DebugLocs.getSym()); 2752 2753 for (const auto &List : DebugLocs.getLists()) 2754 Asm->emitLabelDifference(List.Label, DebugLocs.getSym(), 2755 Asm->getDwarfOffsetByteSize()); 2756 2757 return TableEnd; 2758 } 2759 2760 template <typename Ranges, typename PayloadEmitter> 2761 static void emitRangeList( 2762 DwarfDebug &DD, AsmPrinter *Asm, MCSymbol *Sym, const Ranges &R, 2763 const DwarfCompileUnit &CU, unsigned BaseAddressx, unsigned OffsetPair, 2764 unsigned StartxLength, unsigned EndOfList, 2765 StringRef (*StringifyEnum)(unsigned), 2766 bool ShouldUseBaseAddress, 2767 PayloadEmitter EmitPayload) { 2768 2769 auto Size = Asm->MAI->getCodePointerSize(); 2770 bool UseDwarf5 = DD.getDwarfVersion() >= 5; 2771 2772 // Emit our symbol so we can find the beginning of the range. 2773 Asm->OutStreamer->emitLabel(Sym); 2774 2775 // Gather all the ranges that apply to the same section so they can share 2776 // a base address entry. 2777 MapVector<const MCSection *, std::vector<decltype(&*R.begin())>> SectionRanges; 2778 2779 for (const auto &Range : R) 2780 SectionRanges[&Range.Begin->getSection()].push_back(&Range); 2781 2782 const MCSymbol *CUBase = CU.getBaseAddress(); 2783 bool BaseIsSet = false; 2784 for (const auto &P : SectionRanges) { 2785 auto *Base = CUBase; 2786 if (!Base && ShouldUseBaseAddress) { 2787 const MCSymbol *Begin = P.second.front()->Begin; 2788 const MCSymbol *NewBase = DD.getSectionLabel(&Begin->getSection()); 2789 if (!UseDwarf5) { 2790 Base = NewBase; 2791 BaseIsSet = true; 2792 Asm->OutStreamer->emitIntValue(-1, Size); 2793 Asm->OutStreamer->AddComment(" base address"); 2794 Asm->OutStreamer->emitSymbolValue(Base, Size); 2795 } else if (NewBase != Begin || P.second.size() > 1) { 2796 // Only use a base address if 2797 // * the existing pool address doesn't match (NewBase != Begin) 2798 // * or, there's more than one entry to share the base address 2799 Base = NewBase; 2800 BaseIsSet = true; 2801 Asm->OutStreamer->AddComment(StringifyEnum(BaseAddressx)); 2802 Asm->emitInt8(BaseAddressx); 2803 Asm->OutStreamer->AddComment(" base address index"); 2804 Asm->emitULEB128(DD.getAddressPool().getIndex(Base)); 2805 } 2806 } else if (BaseIsSet && !UseDwarf5) { 2807 BaseIsSet = false; 2808 assert(!Base); 2809 Asm->OutStreamer->emitIntValue(-1, Size); 2810 Asm->OutStreamer->emitIntValue(0, Size); 2811 } 2812 2813 for (const auto *RS : P.second) { 2814 const MCSymbol *Begin = RS->Begin; 2815 const MCSymbol *End = RS->End; 2816 assert(Begin && "Range without a begin symbol?"); 2817 assert(End && "Range without an end symbol?"); 2818 if (Base) { 2819 if (UseDwarf5) { 2820 // Emit offset_pair when we have a base. 2821 Asm->OutStreamer->AddComment(StringifyEnum(OffsetPair)); 2822 Asm->emitInt8(OffsetPair); 2823 Asm->OutStreamer->AddComment(" starting offset"); 2824 Asm->emitLabelDifferenceAsULEB128(Begin, Base); 2825 Asm->OutStreamer->AddComment(" ending offset"); 2826 Asm->emitLabelDifferenceAsULEB128(End, Base); 2827 } else { 2828 Asm->emitLabelDifference(Begin, Base, Size); 2829 Asm->emitLabelDifference(End, Base, Size); 2830 } 2831 } else if (UseDwarf5) { 2832 Asm->OutStreamer->AddComment(StringifyEnum(StartxLength)); 2833 Asm->emitInt8(StartxLength); 2834 Asm->OutStreamer->AddComment(" start index"); 2835 Asm->emitULEB128(DD.getAddressPool().getIndex(Begin)); 2836 Asm->OutStreamer->AddComment(" length"); 2837 Asm->emitLabelDifferenceAsULEB128(End, Begin); 2838 } else { 2839 Asm->OutStreamer->emitSymbolValue(Begin, Size); 2840 Asm->OutStreamer->emitSymbolValue(End, Size); 2841 } 2842 EmitPayload(*RS); 2843 } 2844 } 2845 2846 if (UseDwarf5) { 2847 Asm->OutStreamer->AddComment(StringifyEnum(EndOfList)); 2848 Asm->emitInt8(EndOfList); 2849 } else { 2850 // Terminate the list with two 0 values. 2851 Asm->OutStreamer->emitIntValue(0, Size); 2852 Asm->OutStreamer->emitIntValue(0, Size); 2853 } 2854 } 2855 2856 // Handles emission of both debug_loclist / debug_loclist.dwo 2857 static void emitLocList(DwarfDebug &DD, AsmPrinter *Asm, const DebugLocStream::List &List) { 2858 emitRangeList(DD, Asm, List.Label, DD.getDebugLocs().getEntries(List), 2859 *List.CU, dwarf::DW_LLE_base_addressx, 2860 dwarf::DW_LLE_offset_pair, dwarf::DW_LLE_startx_length, 2861 dwarf::DW_LLE_end_of_list, llvm::dwarf::LocListEncodingString, 2862 /* ShouldUseBaseAddress */ true, 2863 [&](const DebugLocStream::Entry &E) { 2864 DD.emitDebugLocEntryLocation(E, List.CU); 2865 }); 2866 } 2867 2868 void DwarfDebug::emitDebugLocImpl(MCSection *Sec) { 2869 if (DebugLocs.getLists().empty()) 2870 return; 2871 2872 Asm->OutStreamer->SwitchSection(Sec); 2873 2874 MCSymbol *TableEnd = nullptr; 2875 if (getDwarfVersion() >= 5) 2876 TableEnd = emitLoclistsTableHeader(Asm, *this); 2877 2878 for (const auto &List : DebugLocs.getLists()) 2879 emitLocList(*this, Asm, List); 2880 2881 if (TableEnd) 2882 Asm->OutStreamer->emitLabel(TableEnd); 2883 } 2884 2885 // Emit locations into the .debug_loc/.debug_loclists section. 2886 void DwarfDebug::emitDebugLoc() { 2887 emitDebugLocImpl( 2888 getDwarfVersion() >= 5 2889 ? Asm->getObjFileLowering().getDwarfLoclistsSection() 2890 : Asm->getObjFileLowering().getDwarfLocSection()); 2891 } 2892 2893 // Emit locations into the .debug_loc.dwo/.debug_loclists.dwo section. 2894 void DwarfDebug::emitDebugLocDWO() { 2895 if (getDwarfVersion() >= 5) { 2896 emitDebugLocImpl( 2897 Asm->getObjFileLowering().getDwarfLoclistsDWOSection()); 2898 2899 return; 2900 } 2901 2902 for (const auto &List : DebugLocs.getLists()) { 2903 Asm->OutStreamer->SwitchSection( 2904 Asm->getObjFileLowering().getDwarfLocDWOSection()); 2905 Asm->OutStreamer->emitLabel(List.Label); 2906 2907 for (const auto &Entry : DebugLocs.getEntries(List)) { 2908 // GDB only supports startx_length in pre-standard split-DWARF. 2909 // (in v5 standard loclists, it currently* /only/ supports base_address + 2910 // offset_pair, so the implementations can't really share much since they 2911 // need to use different representations) 2912 // * as of October 2018, at least 2913 // 2914 // In v5 (see emitLocList), this uses SectionLabels to reuse existing 2915 // addresses in the address pool to minimize object size/relocations. 2916 Asm->emitInt8(dwarf::DW_LLE_startx_length); 2917 unsigned idx = AddrPool.getIndex(Entry.Begin); 2918 Asm->emitULEB128(idx); 2919 // Also the pre-standard encoding is slightly different, emitting this as 2920 // an address-length entry here, but its a ULEB128 in DWARFv5 loclists. 2921 Asm->emitLabelDifference(Entry.End, Entry.Begin, 4); 2922 emitDebugLocEntryLocation(Entry, List.CU); 2923 } 2924 Asm->emitInt8(dwarf::DW_LLE_end_of_list); 2925 } 2926 } 2927 2928 struct ArangeSpan { 2929 const MCSymbol *Start, *End; 2930 }; 2931 2932 // Emit a debug aranges section, containing a CU lookup for any 2933 // address we can tie back to a CU. 2934 void DwarfDebug::emitDebugARanges() { 2935 // Provides a unique id per text section. 2936 MapVector<MCSection *, SmallVector<SymbolCU, 8>> SectionMap; 2937 2938 // Filter labels by section. 2939 for (const SymbolCU &SCU : ArangeLabels) { 2940 if (SCU.Sym->isInSection()) { 2941 // Make a note of this symbol and it's section. 2942 MCSection *Section = &SCU.Sym->getSection(); 2943 if (!Section->getKind().isMetadata()) 2944 SectionMap[Section].push_back(SCU); 2945 } else { 2946 // Some symbols (e.g. common/bss on mach-o) can have no section but still 2947 // appear in the output. This sucks as we rely on sections to build 2948 // arange spans. We can do it without, but it's icky. 2949 SectionMap[nullptr].push_back(SCU); 2950 } 2951 } 2952 2953 DenseMap<DwarfCompileUnit *, std::vector<ArangeSpan>> Spans; 2954 2955 for (auto &I : SectionMap) { 2956 MCSection *Section = I.first; 2957 SmallVector<SymbolCU, 8> &List = I.second; 2958 if (List.size() < 1) 2959 continue; 2960 2961 // If we have no section (e.g. common), just write out 2962 // individual spans for each symbol. 2963 if (!Section) { 2964 for (const SymbolCU &Cur : List) { 2965 ArangeSpan Span; 2966 Span.Start = Cur.Sym; 2967 Span.End = nullptr; 2968 assert(Cur.CU); 2969 Spans[Cur.CU].push_back(Span); 2970 } 2971 continue; 2972 } 2973 2974 // Sort the symbols by offset within the section. 2975 llvm::stable_sort(List, [&](const SymbolCU &A, const SymbolCU &B) { 2976 unsigned IA = A.Sym ? Asm->OutStreamer->GetSymbolOrder(A.Sym) : 0; 2977 unsigned IB = B.Sym ? Asm->OutStreamer->GetSymbolOrder(B.Sym) : 0; 2978 2979 // Symbols with no order assigned should be placed at the end. 2980 // (e.g. section end labels) 2981 if (IA == 0) 2982 return false; 2983 if (IB == 0) 2984 return true; 2985 return IA < IB; 2986 }); 2987 2988 // Insert a final terminator. 2989 List.push_back(SymbolCU(nullptr, Asm->OutStreamer->endSection(Section))); 2990 2991 // Build spans between each label. 2992 const MCSymbol *StartSym = List[0].Sym; 2993 for (size_t n = 1, e = List.size(); n < e; n++) { 2994 const SymbolCU &Prev = List[n - 1]; 2995 const SymbolCU &Cur = List[n]; 2996 2997 // Try and build the longest span we can within the same CU. 2998 if (Cur.CU != Prev.CU) { 2999 ArangeSpan Span; 3000 Span.Start = StartSym; 3001 Span.End = Cur.Sym; 3002 assert(Prev.CU); 3003 Spans[Prev.CU].push_back(Span); 3004 StartSym = Cur.Sym; 3005 } 3006 } 3007 } 3008 3009 // Start the dwarf aranges section. 3010 Asm->OutStreamer->SwitchSection( 3011 Asm->getObjFileLowering().getDwarfARangesSection()); 3012 3013 unsigned PtrSize = Asm->MAI->getCodePointerSize(); 3014 3015 // Build a list of CUs used. 3016 std::vector<DwarfCompileUnit *> CUs; 3017 for (const auto &it : Spans) { 3018 DwarfCompileUnit *CU = it.first; 3019 CUs.push_back(CU); 3020 } 3021 3022 // Sort the CU list (again, to ensure consistent output order). 3023 llvm::sort(CUs, [](const DwarfCompileUnit *A, const DwarfCompileUnit *B) { 3024 return A->getUniqueID() < B->getUniqueID(); 3025 }); 3026 3027 // Emit an arange table for each CU we used. 3028 for (DwarfCompileUnit *CU : CUs) { 3029 std::vector<ArangeSpan> &List = Spans[CU]; 3030 3031 // Describe the skeleton CU's offset and length, not the dwo file's. 3032 if (auto *Skel = CU->getSkeleton()) 3033 CU = Skel; 3034 3035 // Emit size of content not including length itself. 3036 unsigned ContentSize = 3037 sizeof(int16_t) + // DWARF ARange version number 3038 Asm->getDwarfOffsetByteSize() + // Offset of CU in the .debug_info 3039 // section 3040 sizeof(int8_t) + // Pointer Size (in bytes) 3041 sizeof(int8_t); // Segment Size (in bytes) 3042 3043 unsigned TupleSize = PtrSize * 2; 3044 3045 // 7.20 in the Dwarf specs requires the table to be aligned to a tuple. 3046 unsigned Padding = offsetToAlignment( 3047 Asm->getUnitLengthFieldByteSize() + ContentSize, Align(TupleSize)); 3048 3049 ContentSize += Padding; 3050 ContentSize += (List.size() + 1) * TupleSize; 3051 3052 // For each compile unit, write the list of spans it covers. 3053 Asm->emitDwarfUnitLength(ContentSize, "Length of ARange Set"); 3054 Asm->OutStreamer->AddComment("DWARF Arange version number"); 3055 Asm->emitInt16(dwarf::DW_ARANGES_VERSION); 3056 Asm->OutStreamer->AddComment("Offset Into Debug Info Section"); 3057 emitSectionReference(*CU); 3058 Asm->OutStreamer->AddComment("Address Size (in bytes)"); 3059 Asm->emitInt8(PtrSize); 3060 Asm->OutStreamer->AddComment("Segment Size (in bytes)"); 3061 Asm->emitInt8(0); 3062 3063 Asm->OutStreamer->emitFill(Padding, 0xff); 3064 3065 for (const ArangeSpan &Span : List) { 3066 Asm->emitLabelReference(Span.Start, PtrSize); 3067 3068 // Calculate the size as being from the span start to it's end. 3069 if (Span.End) { 3070 Asm->emitLabelDifference(Span.End, Span.Start, PtrSize); 3071 } else { 3072 // For symbols without an end marker (e.g. common), we 3073 // write a single arange entry containing just that one symbol. 3074 uint64_t Size = SymSize[Span.Start]; 3075 if (Size == 0) 3076 Size = 1; 3077 3078 Asm->OutStreamer->emitIntValue(Size, PtrSize); 3079 } 3080 } 3081 3082 Asm->OutStreamer->AddComment("ARange terminator"); 3083 Asm->OutStreamer->emitIntValue(0, PtrSize); 3084 Asm->OutStreamer->emitIntValue(0, PtrSize); 3085 } 3086 } 3087 3088 /// Emit a single range list. We handle both DWARF v5 and earlier. 3089 static void emitRangeList(DwarfDebug &DD, AsmPrinter *Asm, 3090 const RangeSpanList &List) { 3091 emitRangeList(DD, Asm, List.Label, List.Ranges, *List.CU, 3092 dwarf::DW_RLE_base_addressx, dwarf::DW_RLE_offset_pair, 3093 dwarf::DW_RLE_startx_length, dwarf::DW_RLE_end_of_list, 3094 llvm::dwarf::RangeListEncodingString, 3095 List.CU->getCUNode()->getRangesBaseAddress() || 3096 DD.getDwarfVersion() >= 5, 3097 [](auto) {}); 3098 } 3099 3100 void DwarfDebug::emitDebugRangesImpl(const DwarfFile &Holder, MCSection *Section) { 3101 if (Holder.getRangeLists().empty()) 3102 return; 3103 3104 assert(useRangesSection()); 3105 assert(!CUMap.empty()); 3106 assert(llvm::any_of(CUMap, [](const decltype(CUMap)::value_type &Pair) { 3107 return !Pair.second->getCUNode()->isDebugDirectivesOnly(); 3108 })); 3109 3110 Asm->OutStreamer->SwitchSection(Section); 3111 3112 MCSymbol *TableEnd = nullptr; 3113 if (getDwarfVersion() >= 5) 3114 TableEnd = emitRnglistsTableHeader(Asm, Holder); 3115 3116 for (const RangeSpanList &List : Holder.getRangeLists()) 3117 emitRangeList(*this, Asm, List); 3118 3119 if (TableEnd) 3120 Asm->OutStreamer->emitLabel(TableEnd); 3121 } 3122 3123 /// Emit address ranges into the .debug_ranges section or into the DWARF v5 3124 /// .debug_rnglists section. 3125 void DwarfDebug::emitDebugRanges() { 3126 const auto &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 3127 3128 emitDebugRangesImpl(Holder, 3129 getDwarfVersion() >= 5 3130 ? Asm->getObjFileLowering().getDwarfRnglistsSection() 3131 : Asm->getObjFileLowering().getDwarfRangesSection()); 3132 } 3133 3134 void DwarfDebug::emitDebugRangesDWO() { 3135 emitDebugRangesImpl(InfoHolder, 3136 Asm->getObjFileLowering().getDwarfRnglistsDWOSection()); 3137 } 3138 3139 /// Emit the header of a DWARF 5 macro section, or the GNU extension for 3140 /// DWARF 4. 3141 static void emitMacroHeader(AsmPrinter *Asm, const DwarfDebug &DD, 3142 const DwarfCompileUnit &CU, uint16_t DwarfVersion) { 3143 enum HeaderFlagMask { 3144 #define HANDLE_MACRO_FLAG(ID, NAME) MACRO_FLAG_##NAME = ID, 3145 #include "llvm/BinaryFormat/Dwarf.def" 3146 }; 3147 Asm->OutStreamer->AddComment("Macro information version"); 3148 Asm->emitInt16(DwarfVersion >= 5 ? DwarfVersion : 4); 3149 // We emit the line offset flag unconditionally here, since line offset should 3150 // be mostly present. 3151 if (Asm->isDwarf64()) { 3152 Asm->OutStreamer->AddComment("Flags: 64 bit, debug_line_offset present"); 3153 Asm->emitInt8(MACRO_FLAG_OFFSET_SIZE | MACRO_FLAG_DEBUG_LINE_OFFSET); 3154 } else { 3155 Asm->OutStreamer->AddComment("Flags: 32 bit, debug_line_offset present"); 3156 Asm->emitInt8(MACRO_FLAG_DEBUG_LINE_OFFSET); 3157 } 3158 Asm->OutStreamer->AddComment("debug_line_offset"); 3159 if (DD.useSplitDwarf()) 3160 Asm->emitDwarfLengthOrOffset(0); 3161 else 3162 Asm->emitDwarfSymbolReference(CU.getLineTableStartSym()); 3163 } 3164 3165 void DwarfDebug::handleMacroNodes(DIMacroNodeArray Nodes, DwarfCompileUnit &U) { 3166 for (auto *MN : Nodes) { 3167 if (auto *M = dyn_cast<DIMacro>(MN)) 3168 emitMacro(*M); 3169 else if (auto *F = dyn_cast<DIMacroFile>(MN)) 3170 emitMacroFile(*F, U); 3171 else 3172 llvm_unreachable("Unexpected DI type!"); 3173 } 3174 } 3175 3176 void DwarfDebug::emitMacro(DIMacro &M) { 3177 StringRef Name = M.getName(); 3178 StringRef Value = M.getValue(); 3179 3180 // There should be one space between the macro name and the macro value in 3181 // define entries. In undef entries, only the macro name is emitted. 3182 std::string Str = Value.empty() ? Name.str() : (Name + " " + Value).str(); 3183 3184 if (UseDebugMacroSection) { 3185 if (getDwarfVersion() >= 5) { 3186 unsigned Type = M.getMacinfoType() == dwarf::DW_MACINFO_define 3187 ? dwarf::DW_MACRO_define_strx 3188 : dwarf::DW_MACRO_undef_strx; 3189 Asm->OutStreamer->AddComment(dwarf::MacroString(Type)); 3190 Asm->emitULEB128(Type); 3191 Asm->OutStreamer->AddComment("Line Number"); 3192 Asm->emitULEB128(M.getLine()); 3193 Asm->OutStreamer->AddComment("Macro String"); 3194 Asm->emitULEB128( 3195 InfoHolder.getStringPool().getIndexedEntry(*Asm, Str).getIndex()); 3196 } else { 3197 unsigned Type = M.getMacinfoType() == dwarf::DW_MACINFO_define 3198 ? dwarf::DW_MACRO_GNU_define_indirect 3199 : dwarf::DW_MACRO_GNU_undef_indirect; 3200 Asm->OutStreamer->AddComment(dwarf::GnuMacroString(Type)); 3201 Asm->emitULEB128(Type); 3202 Asm->OutStreamer->AddComment("Line Number"); 3203 Asm->emitULEB128(M.getLine()); 3204 Asm->OutStreamer->AddComment("Macro String"); 3205 Asm->emitDwarfSymbolReference( 3206 InfoHolder.getStringPool().getEntry(*Asm, Str).getSymbol()); 3207 } 3208 } else { 3209 Asm->OutStreamer->AddComment(dwarf::MacinfoString(M.getMacinfoType())); 3210 Asm->emitULEB128(M.getMacinfoType()); 3211 Asm->OutStreamer->AddComment("Line Number"); 3212 Asm->emitULEB128(M.getLine()); 3213 Asm->OutStreamer->AddComment("Macro String"); 3214 Asm->OutStreamer->emitBytes(Str); 3215 Asm->emitInt8('\0'); 3216 } 3217 } 3218 3219 void DwarfDebug::emitMacroFileImpl( 3220 DIMacroFile &MF, DwarfCompileUnit &U, unsigned StartFile, unsigned EndFile, 3221 StringRef (*MacroFormToString)(unsigned Form)) { 3222 3223 Asm->OutStreamer->AddComment(MacroFormToString(StartFile)); 3224 Asm->emitULEB128(StartFile); 3225 Asm->OutStreamer->AddComment("Line Number"); 3226 Asm->emitULEB128(MF.getLine()); 3227 Asm->OutStreamer->AddComment("File Number"); 3228 DIFile &F = *MF.getFile(); 3229 if (useSplitDwarf()) 3230 Asm->emitULEB128(getDwoLineTable(U)->getFile( 3231 F.getDirectory(), F.getFilename(), getMD5AsBytes(&F), 3232 Asm->OutContext.getDwarfVersion(), F.getSource())); 3233 else 3234 Asm->emitULEB128(U.getOrCreateSourceID(&F)); 3235 handleMacroNodes(MF.getElements(), U); 3236 Asm->OutStreamer->AddComment(MacroFormToString(EndFile)); 3237 Asm->emitULEB128(EndFile); 3238 } 3239 3240 void DwarfDebug::emitMacroFile(DIMacroFile &F, DwarfCompileUnit &U) { 3241 // DWARFv5 macro and DWARFv4 macinfo share some common encodings, 3242 // so for readibility/uniformity, We are explicitly emitting those. 3243 assert(F.getMacinfoType() == dwarf::DW_MACINFO_start_file); 3244 if (UseDebugMacroSection) 3245 emitMacroFileImpl( 3246 F, U, dwarf::DW_MACRO_start_file, dwarf::DW_MACRO_end_file, 3247 (getDwarfVersion() >= 5) ? dwarf::MacroString : dwarf::GnuMacroString); 3248 else 3249 emitMacroFileImpl(F, U, dwarf::DW_MACINFO_start_file, 3250 dwarf::DW_MACINFO_end_file, dwarf::MacinfoString); 3251 } 3252 3253 void DwarfDebug::emitDebugMacinfoImpl(MCSection *Section) { 3254 for (const auto &P : CUMap) { 3255 auto &TheCU = *P.second; 3256 auto *SkCU = TheCU.getSkeleton(); 3257 DwarfCompileUnit &U = SkCU ? *SkCU : TheCU; 3258 auto *CUNode = cast<DICompileUnit>(P.first); 3259 DIMacroNodeArray Macros = CUNode->getMacros(); 3260 if (Macros.empty()) 3261 continue; 3262 Asm->OutStreamer->SwitchSection(Section); 3263 Asm->OutStreamer->emitLabel(U.getMacroLabelBegin()); 3264 if (UseDebugMacroSection) 3265 emitMacroHeader(Asm, *this, U, getDwarfVersion()); 3266 handleMacroNodes(Macros, U); 3267 Asm->OutStreamer->AddComment("End Of Macro List Mark"); 3268 Asm->emitInt8(0); 3269 } 3270 } 3271 3272 /// Emit macros into a debug macinfo/macro section. 3273 void DwarfDebug::emitDebugMacinfo() { 3274 auto &ObjLower = Asm->getObjFileLowering(); 3275 emitDebugMacinfoImpl(UseDebugMacroSection 3276 ? ObjLower.getDwarfMacroSection() 3277 : ObjLower.getDwarfMacinfoSection()); 3278 } 3279 3280 void DwarfDebug::emitDebugMacinfoDWO() { 3281 auto &ObjLower = Asm->getObjFileLowering(); 3282 emitDebugMacinfoImpl(UseDebugMacroSection 3283 ? ObjLower.getDwarfMacroDWOSection() 3284 : ObjLower.getDwarfMacinfoDWOSection()); 3285 } 3286 3287 // DWARF5 Experimental Separate Dwarf emitters. 3288 3289 void DwarfDebug::initSkeletonUnit(const DwarfUnit &U, DIE &Die, 3290 std::unique_ptr<DwarfCompileUnit> NewU) { 3291 3292 if (!CompilationDir.empty()) 3293 NewU->addString(Die, dwarf::DW_AT_comp_dir, CompilationDir); 3294 addGnuPubAttributes(*NewU, Die); 3295 3296 SkeletonHolder.addUnit(std::move(NewU)); 3297 } 3298 3299 DwarfCompileUnit &DwarfDebug::constructSkeletonCU(const DwarfCompileUnit &CU) { 3300 3301 auto OwnedUnit = std::make_unique<DwarfCompileUnit>( 3302 CU.getUniqueID(), CU.getCUNode(), Asm, this, &SkeletonHolder, 3303 UnitKind::Skeleton); 3304 DwarfCompileUnit &NewCU = *OwnedUnit; 3305 NewCU.setSection(Asm->getObjFileLowering().getDwarfInfoSection()); 3306 3307 NewCU.initStmtList(); 3308 3309 if (useSegmentedStringOffsetsTable()) 3310 NewCU.addStringOffsetsStart(); 3311 3312 initSkeletonUnit(CU, NewCU.getUnitDie(), std::move(OwnedUnit)); 3313 3314 return NewCU; 3315 } 3316 3317 // Emit the .debug_info.dwo section for separated dwarf. This contains the 3318 // compile units that would normally be in debug_info. 3319 void DwarfDebug::emitDebugInfoDWO() { 3320 assert(useSplitDwarf() && "No split dwarf debug info?"); 3321 // Don't emit relocations into the dwo file. 3322 InfoHolder.emitUnits(/* UseOffsets */ true); 3323 } 3324 3325 // Emit the .debug_abbrev.dwo section for separated dwarf. This contains the 3326 // abbreviations for the .debug_info.dwo section. 3327 void DwarfDebug::emitDebugAbbrevDWO() { 3328 assert(useSplitDwarf() && "No split dwarf?"); 3329 InfoHolder.emitAbbrevs(Asm->getObjFileLowering().getDwarfAbbrevDWOSection()); 3330 } 3331 3332 void DwarfDebug::emitDebugLineDWO() { 3333 assert(useSplitDwarf() && "No split dwarf?"); 3334 SplitTypeUnitFileTable.Emit( 3335 *Asm->OutStreamer, MCDwarfLineTableParams(), 3336 Asm->getObjFileLowering().getDwarfLineDWOSection()); 3337 } 3338 3339 void DwarfDebug::emitStringOffsetsTableHeaderDWO() { 3340 assert(useSplitDwarf() && "No split dwarf?"); 3341 InfoHolder.getStringPool().emitStringOffsetsTableHeader( 3342 *Asm, Asm->getObjFileLowering().getDwarfStrOffDWOSection(), 3343 InfoHolder.getStringOffsetsStartSym()); 3344 } 3345 3346 // Emit the .debug_str.dwo section for separated dwarf. This contains the 3347 // string section and is identical in format to traditional .debug_str 3348 // sections. 3349 void DwarfDebug::emitDebugStrDWO() { 3350 if (useSegmentedStringOffsetsTable()) 3351 emitStringOffsetsTableHeaderDWO(); 3352 assert(useSplitDwarf() && "No split dwarf?"); 3353 MCSection *OffSec = Asm->getObjFileLowering().getDwarfStrOffDWOSection(); 3354 InfoHolder.emitStrings(Asm->getObjFileLowering().getDwarfStrDWOSection(), 3355 OffSec, /* UseRelativeOffsets = */ false); 3356 } 3357 3358 // Emit address pool. 3359 void DwarfDebug::emitDebugAddr() { 3360 AddrPool.emit(*Asm, Asm->getObjFileLowering().getDwarfAddrSection()); 3361 } 3362 3363 MCDwarfDwoLineTable *DwarfDebug::getDwoLineTable(const DwarfCompileUnit &CU) { 3364 if (!useSplitDwarf()) 3365 return nullptr; 3366 const DICompileUnit *DIUnit = CU.getCUNode(); 3367 SplitTypeUnitFileTable.maybeSetRootFile( 3368 DIUnit->getDirectory(), DIUnit->getFilename(), 3369 getMD5AsBytes(DIUnit->getFile()), DIUnit->getSource()); 3370 return &SplitTypeUnitFileTable; 3371 } 3372 3373 uint64_t DwarfDebug::makeTypeSignature(StringRef Identifier) { 3374 MD5 Hash; 3375 Hash.update(Identifier); 3376 // ... take the least significant 8 bytes and return those. Our MD5 3377 // implementation always returns its results in little endian, so we actually 3378 // need the "high" word. 3379 MD5::MD5Result Result; 3380 Hash.final(Result); 3381 return Result.high(); 3382 } 3383 3384 void DwarfDebug::addDwarfTypeUnitType(DwarfCompileUnit &CU, 3385 StringRef Identifier, DIE &RefDie, 3386 const DICompositeType *CTy) { 3387 bool TopLevelType = TypeUnitsUnderConstruction.empty(); 3388 if (auto Signature = getOrCreateDwarfTypeUnit(CU, Identifier, CTy)) { 3389 CU.addDIETypeSignature(RefDie, *Signature); 3390 } else if (TopLevelType) { 3391 // Construct this type in the CU directly. 3392 // This is inefficient because all the dependent types will be rebuilt 3393 // from scratch, including building them in type units, discovering that 3394 // they depend on addresses, throwing them out and rebuilding them. 3395 CU.constructTypeDIE(RefDie, cast<DICompositeType>(CTy)); 3396 } 3397 } 3398 3399 Optional<uint64_t> 3400 DwarfDebug::getOrCreateDwarfTypeUnit(DwarfCompileUnit &CU, StringRef Identifier, 3401 const DICompositeType *CTy) { 3402 // Fast path if we're building some type units and one has already used the 3403 // address pool we know we're going to throw away all this work anyway, so 3404 // don't bother building dependent types. 3405 if (!TypeUnitsUnderConstruction.empty() && AddrPool.hasBeenUsed()) 3406 return None; 3407 3408 auto Ins = TypeSignatures.insert(std::make_pair(CTy, 0)); 3409 if (!Ins.second) 3410 return Ins.first->second; 3411 3412 bool TopLevelType = TypeUnitsUnderConstruction.empty(); 3413 AddrPool.resetUsedFlag(); 3414 3415 auto OwnedUnit = std::make_unique<DwarfTypeUnit>(CU, Asm, this, &InfoHolder, 3416 getDwoLineTable(CU)); 3417 DwarfTypeUnit &NewTU = *OwnedUnit; 3418 DIE &UnitDie = NewTU.getUnitDie(); 3419 TypeUnitsUnderConstruction.emplace_back(std::move(OwnedUnit), CTy); 3420 3421 NewTU.addUInt(UnitDie, dwarf::DW_AT_language, dwarf::DW_FORM_data2, 3422 CU.getLanguage()); 3423 3424 uint64_t Signature = makeTypeSignature(Identifier); 3425 NewTU.setTypeSignature(Signature); 3426 Ins.first->second = Signature; 3427 3428 if (useSplitDwarf()) { 3429 MCSection *Section = 3430 getDwarfVersion() <= 4 3431 ? Asm->getObjFileLowering().getDwarfTypesDWOSection() 3432 : Asm->getObjFileLowering().getDwarfInfoDWOSection(); 3433 NewTU.setSection(Section); 3434 } else { 3435 MCSection *Section = 3436 getDwarfVersion() <= 4 3437 ? Asm->getObjFileLowering().getDwarfTypesSection(Signature) 3438 : Asm->getObjFileLowering().getDwarfInfoSection(Signature); 3439 NewTU.setSection(Section); 3440 // Non-split type units reuse the compile unit's line table. 3441 CU.applyStmtList(UnitDie); 3442 } 3443 3444 // Add DW_AT_str_offsets_base to the type unit DIE, but not for split type 3445 // units. 3446 if (useSegmentedStringOffsetsTable() && !useSplitDwarf()) 3447 NewTU.addStringOffsetsStart(); 3448 3449 NewTU.setType(NewTU.createTypeDIE(CTy)); 3450 3451 if (TopLevelType) { 3452 auto TypeUnitsToAdd = std::move(TypeUnitsUnderConstruction); 3453 TypeUnitsUnderConstruction.clear(); 3454 3455 // Types referencing entries in the address table cannot be placed in type 3456 // units. 3457 if (AddrPool.hasBeenUsed()) { 3458 3459 // Remove all the types built while building this type. 3460 // This is pessimistic as some of these types might not be dependent on 3461 // the type that used an address. 3462 for (const auto &TU : TypeUnitsToAdd) 3463 TypeSignatures.erase(TU.second); 3464 return None; 3465 } 3466 3467 // If the type wasn't dependent on fission addresses, finish adding the type 3468 // and all its dependent types. 3469 for (auto &TU : TypeUnitsToAdd) { 3470 InfoHolder.computeSizeAndOffsetsForUnit(TU.first.get()); 3471 InfoHolder.emitUnit(TU.first.get(), useSplitDwarf()); 3472 } 3473 } 3474 return Signature; 3475 } 3476 3477 DwarfDebug::NonTypeUnitContext::NonTypeUnitContext(DwarfDebug *DD) 3478 : DD(DD), 3479 TypeUnitsUnderConstruction(std::move(DD->TypeUnitsUnderConstruction)), AddrPoolUsed(DD->AddrPool.hasBeenUsed()) { 3480 DD->TypeUnitsUnderConstruction.clear(); 3481 DD->AddrPool.resetUsedFlag(); 3482 } 3483 3484 DwarfDebug::NonTypeUnitContext::~NonTypeUnitContext() { 3485 DD->TypeUnitsUnderConstruction = std::move(TypeUnitsUnderConstruction); 3486 DD->AddrPool.resetUsedFlag(AddrPoolUsed); 3487 } 3488 3489 DwarfDebug::NonTypeUnitContext DwarfDebug::enterNonTypeUnitContext() { 3490 return NonTypeUnitContext(this); 3491 } 3492 3493 // Add the Name along with its companion DIE to the appropriate accelerator 3494 // table (for AccelTableKind::Dwarf it's always AccelDebugNames, for 3495 // AccelTableKind::Apple, we use the table we got as an argument). If 3496 // accelerator tables are disabled, this function does nothing. 3497 template <typename DataT> 3498 void DwarfDebug::addAccelNameImpl(const DICompileUnit &CU, 3499 AccelTable<DataT> &AppleAccel, StringRef Name, 3500 const DIE &Die) { 3501 if (getAccelTableKind() == AccelTableKind::None) 3502 return; 3503 3504 if (getAccelTableKind() != AccelTableKind::Apple && 3505 CU.getNameTableKind() != DICompileUnit::DebugNameTableKind::Default) 3506 return; 3507 3508 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 3509 DwarfStringPoolEntryRef Ref = Holder.getStringPool().getEntry(*Asm, Name); 3510 3511 switch (getAccelTableKind()) { 3512 case AccelTableKind::Apple: 3513 AppleAccel.addName(Ref, Die); 3514 break; 3515 case AccelTableKind::Dwarf: 3516 AccelDebugNames.addName(Ref, Die); 3517 break; 3518 case AccelTableKind::Default: 3519 llvm_unreachable("Default should have already been resolved."); 3520 case AccelTableKind::None: 3521 llvm_unreachable("None handled above"); 3522 } 3523 } 3524 3525 void DwarfDebug::addAccelName(const DICompileUnit &CU, StringRef Name, 3526 const DIE &Die) { 3527 addAccelNameImpl(CU, AccelNames, Name, Die); 3528 } 3529 3530 void DwarfDebug::addAccelObjC(const DICompileUnit &CU, StringRef Name, 3531 const DIE &Die) { 3532 // ObjC names go only into the Apple accelerator tables. 3533 if (getAccelTableKind() == AccelTableKind::Apple) 3534 addAccelNameImpl(CU, AccelObjC, Name, Die); 3535 } 3536 3537 void DwarfDebug::addAccelNamespace(const DICompileUnit &CU, StringRef Name, 3538 const DIE &Die) { 3539 addAccelNameImpl(CU, AccelNamespace, Name, Die); 3540 } 3541 3542 void DwarfDebug::addAccelType(const DICompileUnit &CU, StringRef Name, 3543 const DIE &Die, char Flags) { 3544 addAccelNameImpl(CU, AccelTypes, Name, Die); 3545 } 3546 3547 uint16_t DwarfDebug::getDwarfVersion() const { 3548 return Asm->OutStreamer->getContext().getDwarfVersion(); 3549 } 3550 3551 dwarf::Form DwarfDebug::getDwarfSectionOffsetForm() const { 3552 if (Asm->getDwarfVersion() >= 4) 3553 return dwarf::Form::DW_FORM_sec_offset; 3554 assert((!Asm->isDwarf64() || (Asm->getDwarfVersion() == 3)) && 3555 "DWARF64 is not defined prior DWARFv3"); 3556 return Asm->isDwarf64() ? dwarf::Form::DW_FORM_data8 3557 : dwarf::Form::DW_FORM_data4; 3558 } 3559 3560 const MCSymbol *DwarfDebug::getSectionLabel(const MCSection *S) { 3561 auto I = SectionLabels.find(S); 3562 if (I == SectionLabels.end()) 3563 return nullptr; 3564 return I->second; 3565 } 3566 void DwarfDebug::insertSectionLabel(const MCSymbol *S) { 3567 if (SectionLabels.insert(std::make_pair(&S->getSection(), S)).second) 3568 if (useSplitDwarf() || getDwarfVersion() >= 5) 3569 AddrPool.getIndex(S); 3570 } 3571 3572 Optional<MD5::MD5Result> DwarfDebug::getMD5AsBytes(const DIFile *File) const { 3573 assert(File); 3574 if (getDwarfVersion() < 5) 3575 return None; 3576 Optional<DIFile::ChecksumInfo<StringRef>> Checksum = File->getChecksum(); 3577 if (!Checksum || Checksum->Kind != DIFile::CSK_MD5) 3578 return None; 3579 3580 // Convert the string checksum to an MD5Result for the streamer. 3581 // The verifier validates the checksum so we assume it's okay. 3582 // An MD5 checksum is 16 bytes. 3583 std::string ChecksumString = fromHex(Checksum->Value); 3584 MD5::MD5Result CKMem; 3585 std::copy(ChecksumString.begin(), ChecksumString.end(), CKMem.Bytes.data()); 3586 return CKMem; 3587 } 3588