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