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.getAbstractScopeDIEs().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 1071 // Collect local scopes that contain any local declarations 1072 // (excluding local variables) to be sure they will be emitted 1073 // (see DwarfCompileUnit::createAndAddScopeChildren() for details). 1074 static void collectLocalScopesWithDeclsFromCU(const DICompileUnit *CUNode, 1075 DwarfCompileUnit &CU) { 1076 auto getLocalScope = [](const DIScope *S) -> const DILocalScope * { 1077 if (!S) 1078 return nullptr; 1079 if (isa<DICommonBlock>(S)) 1080 S = S->getScope(); 1081 if (const auto *LScope = dyn_cast_or_null<DILocalScope>(S)) 1082 return LScope->getNonLexicalBlockFileScope(); 1083 return nullptr; 1084 }; 1085 1086 for (auto *GVE : CUNode->getGlobalVariables()) 1087 if (auto *LScope = getLocalScope(GVE->getVariable()->getScope())) 1088 CU.recordLocalScopeWithDecls(LScope); 1089 1090 for (auto *Ty : CUNode->getEnumTypes()) 1091 if (auto *LScope = getLocalScope(Ty->getScope())) 1092 CU.recordLocalScopeWithDecls(LScope); 1093 1094 for (auto *Ty : CUNode->getRetainedTypes()) 1095 if (DIType *RT = dyn_cast<DIType>(Ty)) 1096 if (auto *LScope = getLocalScope(RT->getScope())) 1097 CU.recordLocalScopeWithDecls(LScope); 1098 1099 for (auto *IE : CUNode->getImportedEntities()) 1100 if (auto *LScope = getLocalScope(IE->getScope())) 1101 CU.recordLocalScopeWithDecls(LScope); 1102 1103 // FIXME: We know nothing about local records and typedefs here. 1104 // since nothing but local variables (and members of local records) 1105 // references them. So that they will be emitted in a first available 1106 // parent scope DIE. 1107 } 1108 1109 // Create new DwarfCompileUnit for the given metadata node with tag 1110 // DW_TAG_compile_unit. 1111 DwarfCompileUnit & 1112 DwarfDebug::getOrCreateDwarfCompileUnit(const DICompileUnit *DIUnit) { 1113 if (auto *CU = CUMap.lookup(DIUnit)) 1114 return *CU; 1115 1116 CompilationDir = DIUnit->getDirectory(); 1117 1118 auto OwnedUnit = std::make_unique<DwarfCompileUnit>( 1119 InfoHolder.getUnits().size(), DIUnit, Asm, this, &InfoHolder); 1120 DwarfCompileUnit &NewCU = *OwnedUnit; 1121 InfoHolder.addUnit(std::move(OwnedUnit)); 1122 1123 // LTO with assembly output shares a single line table amongst multiple CUs. 1124 // To avoid the compilation directory being ambiguous, let the line table 1125 // explicitly describe the directory of all files, never relying on the 1126 // compilation directory. 1127 if (!Asm->OutStreamer->hasRawTextSupport() || SingleCU) 1128 Asm->OutStreamer->emitDwarfFile0Directive( 1129 CompilationDir, DIUnit->getFilename(), getMD5AsBytes(DIUnit->getFile()), 1130 DIUnit->getSource(), NewCU.getUniqueID()); 1131 1132 if (useSplitDwarf()) { 1133 NewCU.setSkeleton(constructSkeletonCU(NewCU)); 1134 NewCU.setSection(Asm->getObjFileLowering().getDwarfInfoDWOSection()); 1135 } else { 1136 finishUnitAttributes(DIUnit, NewCU); 1137 NewCU.setSection(Asm->getObjFileLowering().getDwarfInfoSection()); 1138 } 1139 1140 CUMap.insert({DIUnit, &NewCU}); 1141 CUDieMap.insert({&NewCU.getUnitDie(), &NewCU}); 1142 1143 // Record local scopes, that have some globals (static locals), 1144 // imports or types declared within. 1145 collectLocalScopesWithDeclsFromCU(DIUnit, NewCU); 1146 1147 return NewCU; 1148 } 1149 1150 // Emit all Dwarf sections that should come prior to the content. Create 1151 // global DIEs and emit initial debug info sections. This is invoked by 1152 // the target AsmPrinter. 1153 void DwarfDebug::beginModule(Module *M) { 1154 DebugHandlerBase::beginModule(M); 1155 1156 if (!Asm || !MMI->hasDebugInfo()) 1157 return; 1158 1159 unsigned NumDebugCUs = std::distance(M->debug_compile_units_begin(), 1160 M->debug_compile_units_end()); 1161 assert(NumDebugCUs > 0 && "Asm unexpectedly initialized"); 1162 assert(MMI->hasDebugInfo() && 1163 "DebugInfoAvailabilty unexpectedly not initialized"); 1164 SingleCU = NumDebugCUs == 1; 1165 1166 // Create the symbol that designates the start of the unit's contribution 1167 // to the string offsets table. In a split DWARF scenario, only the skeleton 1168 // unit has the DW_AT_str_offsets_base attribute (and hence needs the symbol). 1169 if (useSegmentedStringOffsetsTable()) 1170 (useSplitDwarf() ? SkeletonHolder : InfoHolder) 1171 .setStringOffsetsStartSym(Asm->createTempSymbol("str_offsets_base")); 1172 1173 1174 // Create the symbols that designates the start of the DWARF v5 range list 1175 // and locations list tables. They are located past the table headers. 1176 if (getDwarfVersion() >= 5) { 1177 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 1178 Holder.setRnglistsTableBaseSym( 1179 Asm->createTempSymbol("rnglists_table_base")); 1180 1181 if (useSplitDwarf()) 1182 InfoHolder.setRnglistsTableBaseSym( 1183 Asm->createTempSymbol("rnglists_dwo_table_base")); 1184 } 1185 1186 // Create the symbol that points to the first entry following the debug 1187 // address table (.debug_addr) header. 1188 AddrPool.setLabel(Asm->createTempSymbol("addr_table_base")); 1189 DebugLocs.setSym(Asm->createTempSymbol("loclists_table_base")); 1190 } 1191 1192 void DwarfDebug::finishEntityDefinitions() { 1193 for (const auto &Entity : ConcreteEntities) { 1194 DIE *Die = Entity->getDIE(); 1195 assert(Die); 1196 // FIXME: Consider the time-space tradeoff of just storing the unit pointer 1197 // in the ConcreteEntities list, rather than looking it up again here. 1198 // DIE::getUnit isn't simple - it walks parent pointers, etc. 1199 DwarfCompileUnit *Unit = CUDieMap.lookup(Die->getUnitDie()); 1200 assert(Unit); 1201 Unit->finishEntityDefinition(Entity.get()); 1202 } 1203 } 1204 1205 void DwarfDebug::finishSubprogramDefinitions() { 1206 for (const DISubprogram *SP : ProcessedSPNodes) { 1207 assert(SP->getUnit()->getEmissionKind() != DICompileUnit::NoDebug); 1208 forBothCUs( 1209 getOrCreateDwarfCompileUnit(SP->getUnit()), 1210 [&](DwarfCompileUnit &CU) { CU.finishSubprogramDefinition(SP); }); 1211 } 1212 } 1213 1214 void DwarfDebug::finalizeModuleInfo() { 1215 const TargetLoweringObjectFile &TLOF = Asm->getObjFileLowering(); 1216 1217 finishSubprogramDefinitions(); 1218 1219 finishEntityDefinitions(); 1220 1221 // Include the DWO file name in the hash if there's more than one CU. 1222 // This handles ThinLTO's situation where imported CUs may very easily be 1223 // duplicate with the same CU partially imported into another ThinLTO unit. 1224 StringRef DWOName; 1225 if (CUMap.size() > 1) 1226 DWOName = Asm->TM.Options.MCOptions.SplitDwarfFile; 1227 1228 // Handle anything that needs to be done on a per-unit basis after 1229 // all other generation. 1230 for (const auto &P : CUMap) { 1231 auto &TheCU = *P.second; 1232 if (TheCU.getCUNode()->isDebugDirectivesOnly()) 1233 continue; 1234 // Emit DW_AT_containing_type attribute to connect types with their 1235 // vtable holding type. 1236 TheCU.constructContainingTypeDIEs(); 1237 1238 // Add CU specific attributes if we need to add any. 1239 // If we're splitting the dwarf out now that we've got the entire 1240 // CU then add the dwo id to it. 1241 auto *SkCU = TheCU.getSkeleton(); 1242 1243 bool HasSplitUnit = SkCU && !TheCU.getUnitDie().children().empty(); 1244 1245 if (HasSplitUnit) { 1246 dwarf::Attribute attrDWOName = getDwarfVersion() >= 5 1247 ? dwarf::DW_AT_dwo_name 1248 : dwarf::DW_AT_GNU_dwo_name; 1249 finishUnitAttributes(TheCU.getCUNode(), TheCU); 1250 TheCU.addString(TheCU.getUnitDie(), attrDWOName, 1251 Asm->TM.Options.MCOptions.SplitDwarfFile); 1252 SkCU->addString(SkCU->getUnitDie(), attrDWOName, 1253 Asm->TM.Options.MCOptions.SplitDwarfFile); 1254 // Emit a unique identifier for this CU. 1255 uint64_t ID = 1256 DIEHash(Asm, &TheCU).computeCUSignature(DWOName, TheCU.getUnitDie()); 1257 if (getDwarfVersion() >= 5) { 1258 TheCU.setDWOId(ID); 1259 SkCU->setDWOId(ID); 1260 } else { 1261 TheCU.addUInt(TheCU.getUnitDie(), dwarf::DW_AT_GNU_dwo_id, 1262 dwarf::DW_FORM_data8, ID); 1263 SkCU->addUInt(SkCU->getUnitDie(), dwarf::DW_AT_GNU_dwo_id, 1264 dwarf::DW_FORM_data8, ID); 1265 } 1266 1267 if (getDwarfVersion() < 5 && !SkeletonHolder.getRangeLists().empty()) { 1268 const MCSymbol *Sym = TLOF.getDwarfRangesSection()->getBeginSymbol(); 1269 SkCU->addSectionLabel(SkCU->getUnitDie(), dwarf::DW_AT_GNU_ranges_base, 1270 Sym, Sym); 1271 } 1272 } else if (SkCU) { 1273 finishUnitAttributes(SkCU->getCUNode(), *SkCU); 1274 } 1275 1276 // If we have code split among multiple sections or non-contiguous 1277 // ranges of code then emit a DW_AT_ranges attribute on the unit that will 1278 // remain in the .o file, otherwise add a DW_AT_low_pc. 1279 // FIXME: We should use ranges allow reordering of code ala 1280 // .subsections_via_symbols in mach-o. This would mean turning on 1281 // ranges for all subprogram DIEs for mach-o. 1282 DwarfCompileUnit &U = SkCU ? *SkCU : TheCU; 1283 1284 if (unsigned NumRanges = TheCU.getRanges().size()) { 1285 if (NumRanges > 1 && useRangesSection()) 1286 // A DW_AT_low_pc attribute may also be specified in combination with 1287 // DW_AT_ranges to specify the default base address for use in 1288 // location lists (see Section 2.6.2) and range lists (see Section 1289 // 2.17.3). 1290 U.addUInt(U.getUnitDie(), dwarf::DW_AT_low_pc, dwarf::DW_FORM_addr, 0); 1291 else 1292 U.setBaseAddress(TheCU.getRanges().front().Begin); 1293 U.attachRangesOrLowHighPC(U.getUnitDie(), TheCU.takeRanges()); 1294 } 1295 1296 // We don't keep track of which addresses are used in which CU so this 1297 // is a bit pessimistic under LTO. 1298 if ((HasSplitUnit || getDwarfVersion() >= 5) && !AddrPool.isEmpty()) 1299 U.addAddrTableBase(); 1300 1301 if (getDwarfVersion() >= 5) { 1302 if (U.hasRangeLists()) 1303 U.addRnglistsBase(); 1304 1305 if (!DebugLocs.getLists().empty()) { 1306 if (!useSplitDwarf()) 1307 U.addSectionLabel(U.getUnitDie(), dwarf::DW_AT_loclists_base, 1308 DebugLocs.getSym(), 1309 TLOF.getDwarfLoclistsSection()->getBeginSymbol()); 1310 } 1311 } 1312 1313 auto *CUNode = cast<DICompileUnit>(P.first); 1314 // If compile Unit has macros, emit "DW_AT_macro_info/DW_AT_macros" 1315 // attribute. 1316 if (CUNode->getMacros()) { 1317 if (UseDebugMacroSection) { 1318 if (useSplitDwarf()) 1319 TheCU.addSectionDelta( 1320 TheCU.getUnitDie(), dwarf::DW_AT_macros, U.getMacroLabelBegin(), 1321 TLOF.getDwarfMacroDWOSection()->getBeginSymbol()); 1322 else { 1323 dwarf::Attribute MacrosAttr = getDwarfVersion() >= 5 1324 ? dwarf::DW_AT_macros 1325 : dwarf::DW_AT_GNU_macros; 1326 U.addSectionLabel(U.getUnitDie(), MacrosAttr, U.getMacroLabelBegin(), 1327 TLOF.getDwarfMacroSection()->getBeginSymbol()); 1328 } 1329 } else { 1330 if (useSplitDwarf()) 1331 TheCU.addSectionDelta( 1332 TheCU.getUnitDie(), dwarf::DW_AT_macro_info, 1333 U.getMacroLabelBegin(), 1334 TLOF.getDwarfMacinfoDWOSection()->getBeginSymbol()); 1335 else 1336 U.addSectionLabel(U.getUnitDie(), dwarf::DW_AT_macro_info, 1337 U.getMacroLabelBegin(), 1338 TLOF.getDwarfMacinfoSection()->getBeginSymbol()); 1339 } 1340 } 1341 } 1342 1343 // Emit all frontend-produced Skeleton CUs, i.e., Clang modules. 1344 for (auto *CUNode : MMI->getModule()->debug_compile_units()) 1345 if (CUNode->getDWOId()) 1346 getOrCreateDwarfCompileUnit(CUNode); 1347 1348 // Compute DIE offsets and sizes. 1349 InfoHolder.computeSizeAndOffsets(); 1350 if (useSplitDwarf()) 1351 SkeletonHolder.computeSizeAndOffsets(); 1352 } 1353 1354 /// Sort and unique GVEs by comparing their fragment offset. 1355 static SmallVectorImpl<DwarfCompileUnit::GlobalExpr> & 1356 sortGlobalExprs(SmallVectorImpl<DwarfCompileUnit::GlobalExpr> &GVEs) { 1357 llvm::sort( 1358 GVEs, [](DwarfCompileUnit::GlobalExpr A, DwarfCompileUnit::GlobalExpr B) { 1359 // Sort order: first null exprs, then exprs without fragment 1360 // info, then sort by fragment offset in bits. 1361 // FIXME: Come up with a more comprehensive comparator so 1362 // the sorting isn't non-deterministic, and so the following 1363 // std::unique call works correctly. 1364 if (!A.Expr || !B.Expr) 1365 return !!B.Expr; 1366 auto FragmentA = A.Expr->getFragmentInfo(); 1367 auto FragmentB = B.Expr->getFragmentInfo(); 1368 if (!FragmentA || !FragmentB) 1369 return !!FragmentB; 1370 return FragmentA->OffsetInBits < FragmentB->OffsetInBits; 1371 }); 1372 GVEs.erase(std::unique(GVEs.begin(), GVEs.end(), 1373 [](DwarfCompileUnit::GlobalExpr A, 1374 DwarfCompileUnit::GlobalExpr B) { 1375 return A.Expr == B.Expr; 1376 }), 1377 GVEs.end()); 1378 return GVEs; 1379 } 1380 1381 // Emit all Dwarf sections that should come after the content. 1382 void DwarfDebug::endModule() { 1383 // Terminate the pending line table. 1384 if (PrevCU) 1385 terminateLineTable(PrevCU); 1386 PrevCU = nullptr; 1387 assert(CurFn == nullptr); 1388 assert(CurMI == nullptr); 1389 1390 // Collect global variables info. 1391 DenseMap<DIGlobalVariable *, SmallVector<DwarfCompileUnit::GlobalExpr, 1>> 1392 GVMap; 1393 for (const GlobalVariable &Global : MMI->getModule()->globals()) { 1394 SmallVector<DIGlobalVariableExpression *, 1> GVs; 1395 Global.getDebugInfo(GVs); 1396 for (auto *GVE : GVs) 1397 GVMap[GVE->getVariable()].push_back({&Global, GVE->getExpression()}); 1398 } 1399 1400 for (DICompileUnit *CUNode : MMI->getModule()->debug_compile_units()) { 1401 auto *CU = CUMap.lookup(CUNode); 1402 1403 // If this CU hasn't been emitted yet, create it here unless it is empty. 1404 if (!CU) { 1405 // FIXME: Move local imported entities into a list attached to the 1406 // subprogram, then this search won't be needed and a 1407 // getImportedEntities().empty() test should go below with the rest. 1408 bool HasNonLocalImportedEntities = llvm::any_of( 1409 CUNode->getImportedEntities(), [](const DIImportedEntity *IE) { 1410 return !isa<DILocalScope>(IE->getScope()); 1411 }); 1412 1413 if (!HasNonLocalImportedEntities && CUNode->getEnumTypes().empty() && 1414 CUNode->getRetainedTypes().empty() && 1415 CUNode->getGlobalVariables().empty() && CUNode->getMacros().empty()) 1416 continue; 1417 1418 CU = &getOrCreateDwarfCompileUnit(CUNode); 1419 } 1420 1421 // Global Variables. 1422 for (auto *GVE : CUNode->getGlobalVariables()) { 1423 // Don't bother adding DIGlobalVariableExpressions listed in the CU if we 1424 // already know about the variable and it isn't adding a constant 1425 // expression. 1426 auto &GVMapEntry = GVMap[GVE->getVariable()]; 1427 auto *Expr = GVE->getExpression(); 1428 if (!GVMapEntry.size() || (Expr && Expr->isConstant())) 1429 GVMapEntry.push_back({nullptr, Expr}); 1430 } 1431 1432 DenseSet<DIGlobalVariable *> Processed; 1433 for (auto *GVE : CUNode->getGlobalVariables()) { 1434 DIGlobalVariable *GV = GVE->getVariable(); 1435 if (Processed.insert(GV).second) 1436 CU->getOrCreateGlobalVariableDIE(GV, sortGlobalExprs(GVMap[GV])); 1437 } 1438 1439 for (auto *Ty : CUNode->getEnumTypes()) 1440 CU->getOrCreateTypeDIE(cast<DIType>(Ty)); 1441 1442 for (auto *Ty : CUNode->getRetainedTypes()) 1443 if (DIType *RT = dyn_cast<DIType>(Ty)) 1444 // There is no point in force-emitting a forward declaration. 1445 CU->getOrCreateTypeDIE(RT); 1446 1447 // Emit imported entities last so that the relevant context 1448 // is already available. 1449 for (auto *IE : CUNode->getImportedEntities()) 1450 CU->createAndAddImportedEntityDIE(IE); 1451 1452 CU->createBaseTypeDIEs(); 1453 } 1454 1455 // If we aren't actually generating debug info (check beginModule - 1456 // conditionalized on the presence of the llvm.dbg.cu metadata node) 1457 if (!Asm || !MMI->hasDebugInfo()) 1458 return; 1459 1460 // Finalize the debug info for the module. 1461 finalizeModuleInfo(); 1462 1463 if (useSplitDwarf()) 1464 // Emit debug_loc.dwo/debug_loclists.dwo section. 1465 emitDebugLocDWO(); 1466 else 1467 // Emit debug_loc/debug_loclists section. 1468 emitDebugLoc(); 1469 1470 // Corresponding abbreviations into a abbrev section. 1471 emitAbbreviations(); 1472 1473 // Emit all the DIEs into a debug info section. 1474 emitDebugInfo(); 1475 1476 // Emit info into a debug aranges section. 1477 if (GenerateARangeSection) 1478 emitDebugARanges(); 1479 1480 // Emit info into a debug ranges section. 1481 emitDebugRanges(); 1482 1483 if (useSplitDwarf()) 1484 // Emit info into a debug macinfo.dwo section. 1485 emitDebugMacinfoDWO(); 1486 else 1487 // Emit info into a debug macinfo/macro section. 1488 emitDebugMacinfo(); 1489 1490 emitDebugStr(); 1491 1492 if (useSplitDwarf()) { 1493 emitDebugStrDWO(); 1494 emitDebugInfoDWO(); 1495 emitDebugAbbrevDWO(); 1496 emitDebugLineDWO(); 1497 emitDebugRangesDWO(); 1498 } 1499 1500 emitDebugAddr(); 1501 1502 // Emit info into the dwarf accelerator table sections. 1503 switch (getAccelTableKind()) { 1504 case AccelTableKind::Apple: 1505 emitAccelNames(); 1506 emitAccelObjC(); 1507 emitAccelNamespaces(); 1508 emitAccelTypes(); 1509 break; 1510 case AccelTableKind::Dwarf: 1511 emitAccelDebugNames(); 1512 break; 1513 case AccelTableKind::None: 1514 break; 1515 case AccelTableKind::Default: 1516 llvm_unreachable("Default should have already been resolved."); 1517 } 1518 1519 // Emit the pubnames and pubtypes sections if requested. 1520 emitDebugPubSections(); 1521 1522 // clean up. 1523 // FIXME: AbstractVariables.clear(); 1524 } 1525 1526 void DwarfDebug::ensureAbstractEntityIsCreated(DwarfCompileUnit &CU, 1527 const DINode *Node, 1528 const MDNode *ScopeNode) { 1529 if (CU.getExistingAbstractEntity(Node)) 1530 return; 1531 1532 CU.createAbstractEntity(Node, LScopes.getOrCreateAbstractScope( 1533 cast<DILocalScope>(ScopeNode))); 1534 } 1535 1536 void DwarfDebug::ensureAbstractEntityIsCreatedIfScoped(DwarfCompileUnit &CU, 1537 const DINode *Node, const MDNode *ScopeNode) { 1538 if (CU.getExistingAbstractEntity(Node)) 1539 return; 1540 1541 if (LexicalScope *Scope = 1542 LScopes.findAbstractScope(cast_or_null<DILocalScope>(ScopeNode))) 1543 CU.createAbstractEntity(Node, Scope); 1544 } 1545 1546 // Collect variable information from side table maintained by MF. 1547 void DwarfDebug::collectVariableInfoFromMFTable( 1548 DwarfCompileUnit &TheCU, DenseSet<InlinedEntity> &Processed) { 1549 SmallDenseMap<InlinedEntity, DbgVariable *> MFVars; 1550 LLVM_DEBUG(dbgs() << "DwarfDebug: collecting variables from MF side table\n"); 1551 for (const auto &VI : Asm->MF->getVariableDbgInfo()) { 1552 if (!VI.Var) 1553 continue; 1554 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1555 "Expected inlined-at fields to agree"); 1556 1557 InlinedEntity Var(VI.Var, VI.Loc->getInlinedAt()); 1558 Processed.insert(Var); 1559 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1560 1561 // If variable scope is not found then skip this variable. 1562 if (!Scope) { 1563 LLVM_DEBUG(dbgs() << "Dropping debug info for " << VI.Var->getName() 1564 << ", no variable scope found\n"); 1565 continue; 1566 } 1567 1568 ensureAbstractEntityIsCreatedIfScoped(TheCU, Var.first, Scope->getScopeNode()); 1569 auto RegVar = std::make_unique<DbgVariable>( 1570 cast<DILocalVariable>(Var.first), Var.second); 1571 RegVar->initializeMMI(VI.Expr, VI.Slot); 1572 LLVM_DEBUG(dbgs() << "Created DbgVariable for " << VI.Var->getName() 1573 << "\n"); 1574 1575 if (DbgVariable *DbgVar = MFVars.lookup(Var)) 1576 DbgVar->addMMIEntry(*RegVar); 1577 else if (InfoHolder.addScopeVariable(Scope, RegVar.get())) { 1578 MFVars.insert({Var, RegVar.get()}); 1579 ConcreteEntities.push_back(std::move(RegVar)); 1580 } 1581 } 1582 } 1583 1584 /// Determine whether a *singular* DBG_VALUE is valid for the entirety of its 1585 /// enclosing lexical scope. The check ensures there are no other instructions 1586 /// in the same lexical scope preceding the DBG_VALUE and that its range is 1587 /// either open or otherwise rolls off the end of the scope. 1588 static bool validThroughout(LexicalScopes &LScopes, 1589 const MachineInstr *DbgValue, 1590 const MachineInstr *RangeEnd, 1591 const InstructionOrdering &Ordering) { 1592 assert(DbgValue->getDebugLoc() && "DBG_VALUE without a debug location"); 1593 auto MBB = DbgValue->getParent(); 1594 auto DL = DbgValue->getDebugLoc(); 1595 auto *LScope = LScopes.findLexicalScope(DL); 1596 // Scope doesn't exist; this is a dead DBG_VALUE. 1597 if (!LScope) 1598 return false; 1599 auto &LSRange = LScope->getRanges(); 1600 if (LSRange.size() == 0) 1601 return false; 1602 1603 const MachineInstr *LScopeBegin = LSRange.front().first; 1604 // If the scope starts before the DBG_VALUE then we may have a negative 1605 // result. Otherwise the location is live coming into the scope and we 1606 // can skip the following checks. 1607 if (!Ordering.isBefore(DbgValue, LScopeBegin)) { 1608 // Exit if the lexical scope begins outside of the current block. 1609 if (LScopeBegin->getParent() != MBB) 1610 return false; 1611 1612 MachineBasicBlock::const_reverse_iterator Pred(DbgValue); 1613 for (++Pred; Pred != MBB->rend(); ++Pred) { 1614 if (Pred->getFlag(MachineInstr::FrameSetup)) 1615 break; 1616 auto PredDL = Pred->getDebugLoc(); 1617 if (!PredDL || Pred->isMetaInstruction()) 1618 continue; 1619 // Check whether the instruction preceding the DBG_VALUE is in the same 1620 // (sub)scope as the DBG_VALUE. 1621 if (DL->getScope() == PredDL->getScope()) 1622 return false; 1623 auto *PredScope = LScopes.findLexicalScope(PredDL); 1624 if (!PredScope || LScope->dominates(PredScope)) 1625 return false; 1626 } 1627 } 1628 1629 // If the range of the DBG_VALUE is open-ended, report success. 1630 if (!RangeEnd) 1631 return true; 1632 1633 // Single, constant DBG_VALUEs in the prologue are promoted to be live 1634 // throughout the function. This is a hack, presumably for DWARF v2 and not 1635 // necessarily correct. It would be much better to use a dbg.declare instead 1636 // if we know the constant is live throughout the scope. 1637 if (MBB->pred_empty() && 1638 all_of(DbgValue->debug_operands(), 1639 [](const MachineOperand &Op) { return Op.isImm(); })) 1640 return true; 1641 1642 // Test if the location terminates before the end of the scope. 1643 const MachineInstr *LScopeEnd = LSRange.back().second; 1644 if (Ordering.isBefore(RangeEnd, LScopeEnd)) 1645 return false; 1646 1647 // There's a single location which starts at the scope start, and ends at or 1648 // after the scope end. 1649 return true; 1650 } 1651 1652 /// Build the location list for all DBG_VALUEs in the function that 1653 /// describe the same variable. The resulting DebugLocEntries will have 1654 /// strict monotonically increasing begin addresses and will never 1655 /// overlap. If the resulting list has only one entry that is valid 1656 /// throughout variable's scope return true. 1657 // 1658 // See the definition of DbgValueHistoryMap::Entry for an explanation of the 1659 // different kinds of history map entries. One thing to be aware of is that if 1660 // a debug value is ended by another entry (rather than being valid until the 1661 // end of the function), that entry's instruction may or may not be included in 1662 // the range, depending on if the entry is a clobbering entry (it has an 1663 // instruction that clobbers one or more preceding locations), or if it is an 1664 // (overlapping) debug value entry. This distinction can be seen in the example 1665 // below. The first debug value is ended by the clobbering entry 2, and the 1666 // second and third debug values are ended by the overlapping debug value entry 1667 // 4. 1668 // 1669 // Input: 1670 // 1671 // History map entries [type, end index, mi] 1672 // 1673 // 0 | [DbgValue, 2, DBG_VALUE $reg0, [...] (fragment 0, 32)] 1674 // 1 | | [DbgValue, 4, DBG_VALUE $reg1, [...] (fragment 32, 32)] 1675 // 2 | | [Clobber, $reg0 = [...], -, -] 1676 // 3 | | [DbgValue, 4, DBG_VALUE 123, [...] (fragment 64, 32)] 1677 // 4 [DbgValue, ~0, DBG_VALUE @g, [...] (fragment 0, 96)] 1678 // 1679 // Output [start, end) [Value...]: 1680 // 1681 // [0-1) [(reg0, fragment 0, 32)] 1682 // [1-3) [(reg0, fragment 0, 32), (reg1, fragment 32, 32)] 1683 // [3-4) [(reg1, fragment 32, 32), (123, fragment 64, 32)] 1684 // [4-) [(@g, fragment 0, 96)] 1685 bool DwarfDebug::buildLocationList(SmallVectorImpl<DebugLocEntry> &DebugLoc, 1686 const DbgValueHistoryMap::Entries &Entries) { 1687 using OpenRange = 1688 std::pair<DbgValueHistoryMap::EntryIndex, DbgValueLoc>; 1689 SmallVector<OpenRange, 4> OpenRanges; 1690 bool isSafeForSingleLocation = true; 1691 const MachineInstr *StartDebugMI = nullptr; 1692 const MachineInstr *EndMI = nullptr; 1693 1694 for (auto EB = Entries.begin(), EI = EB, EE = Entries.end(); EI != EE; ++EI) { 1695 const MachineInstr *Instr = EI->getInstr(); 1696 1697 // Remove all values that are no longer live. 1698 size_t Index = std::distance(EB, EI); 1699 erase_if(OpenRanges, [&](OpenRange &R) { return R.first <= Index; }); 1700 1701 // If we are dealing with a clobbering entry, this iteration will result in 1702 // a location list entry starting after the clobbering instruction. 1703 const MCSymbol *StartLabel = 1704 EI->isClobber() ? getLabelAfterInsn(Instr) : getLabelBeforeInsn(Instr); 1705 assert(StartLabel && 1706 "Forgot label before/after instruction starting a range!"); 1707 1708 const MCSymbol *EndLabel; 1709 if (std::next(EI) == Entries.end()) { 1710 const MachineBasicBlock &EndMBB = Asm->MF->back(); 1711 EndLabel = Asm->MBBSectionRanges[EndMBB.getSectionIDNum()].EndLabel; 1712 if (EI->isClobber()) 1713 EndMI = EI->getInstr(); 1714 } 1715 else if (std::next(EI)->isClobber()) 1716 EndLabel = getLabelAfterInsn(std::next(EI)->getInstr()); 1717 else 1718 EndLabel = getLabelBeforeInsn(std::next(EI)->getInstr()); 1719 assert(EndLabel && "Forgot label after instruction ending a range!"); 1720 1721 if (EI->isDbgValue()) 1722 LLVM_DEBUG(dbgs() << "DotDebugLoc: " << *Instr << "\n"); 1723 1724 // If this history map entry has a debug value, add that to the list of 1725 // open ranges and check if its location is valid for a single value 1726 // location. 1727 if (EI->isDbgValue()) { 1728 // Do not add undef debug values, as they are redundant information in 1729 // the location list entries. An undef debug results in an empty location 1730 // description. If there are any non-undef fragments then padding pieces 1731 // with empty location descriptions will automatically be inserted, and if 1732 // all fragments are undef then the whole location list entry is 1733 // redundant. 1734 if (!Instr->isUndefDebugValue()) { 1735 auto Value = getDebugLocValue(Instr); 1736 OpenRanges.emplace_back(EI->getEndIndex(), Value); 1737 1738 // TODO: Add support for single value fragment locations. 1739 if (Instr->getDebugExpression()->isFragment()) 1740 isSafeForSingleLocation = false; 1741 1742 if (!StartDebugMI) 1743 StartDebugMI = Instr; 1744 } else { 1745 isSafeForSingleLocation = false; 1746 } 1747 } 1748 1749 // Location list entries with empty location descriptions are redundant 1750 // information in DWARF, so do not emit those. 1751 if (OpenRanges.empty()) 1752 continue; 1753 1754 // Omit entries with empty ranges as they do not have any effect in DWARF. 1755 if (StartLabel == EndLabel) { 1756 LLVM_DEBUG(dbgs() << "Omitting location list entry with empty range.\n"); 1757 continue; 1758 } 1759 1760 SmallVector<DbgValueLoc, 4> Values; 1761 for (auto &R : OpenRanges) 1762 Values.push_back(R.second); 1763 1764 // With Basic block sections, it is posssible that the StartLabel and the 1765 // Instr are not in the same section. This happens when the StartLabel is 1766 // the function begin label and the dbg value appears in a basic block 1767 // that is not the entry. In this case, the range needs to be split to 1768 // span each individual section in the range from StartLabel to EndLabel. 1769 if (Asm->MF->hasBBSections() && StartLabel == Asm->getFunctionBegin() && 1770 !Instr->getParent()->sameSection(&Asm->MF->front())) { 1771 const MCSymbol *BeginSectionLabel = StartLabel; 1772 1773 for (const MachineBasicBlock &MBB : *Asm->MF) { 1774 if (MBB.isBeginSection() && &MBB != &Asm->MF->front()) 1775 BeginSectionLabel = MBB.getSymbol(); 1776 1777 if (MBB.sameSection(Instr->getParent())) { 1778 DebugLoc.emplace_back(BeginSectionLabel, EndLabel, Values); 1779 break; 1780 } 1781 if (MBB.isEndSection()) 1782 DebugLoc.emplace_back(BeginSectionLabel, MBB.getEndSymbol(), Values); 1783 } 1784 } else { 1785 DebugLoc.emplace_back(StartLabel, EndLabel, Values); 1786 } 1787 1788 // Attempt to coalesce the ranges of two otherwise identical 1789 // DebugLocEntries. 1790 auto CurEntry = DebugLoc.rbegin(); 1791 LLVM_DEBUG({ 1792 dbgs() << CurEntry->getValues().size() << " Values:\n"; 1793 for (auto &Value : CurEntry->getValues()) 1794 Value.dump(); 1795 dbgs() << "-----\n"; 1796 }); 1797 1798 auto PrevEntry = std::next(CurEntry); 1799 if (PrevEntry != DebugLoc.rend() && PrevEntry->MergeRanges(*CurEntry)) 1800 DebugLoc.pop_back(); 1801 } 1802 1803 if (!isSafeForSingleLocation || 1804 !validThroughout(LScopes, StartDebugMI, EndMI, getInstOrdering())) 1805 return false; 1806 1807 if (DebugLoc.size() == 1) 1808 return true; 1809 1810 if (!Asm->MF->hasBBSections()) 1811 return false; 1812 1813 // Check here to see if loclist can be merged into a single range. If not, 1814 // we must keep the split loclists per section. This does exactly what 1815 // MergeRanges does without sections. We don't actually merge the ranges 1816 // as the split ranges must be kept intact if this cannot be collapsed 1817 // into a single range. 1818 const MachineBasicBlock *RangeMBB = nullptr; 1819 if (DebugLoc[0].getBeginSym() == Asm->getFunctionBegin()) 1820 RangeMBB = &Asm->MF->front(); 1821 else 1822 RangeMBB = Entries.begin()->getInstr()->getParent(); 1823 auto *CurEntry = DebugLoc.begin(); 1824 auto *NextEntry = std::next(CurEntry); 1825 while (NextEntry != DebugLoc.end()) { 1826 // Get the last machine basic block of this section. 1827 while (!RangeMBB->isEndSection()) 1828 RangeMBB = RangeMBB->getNextNode(); 1829 if (!RangeMBB->getNextNode()) 1830 return false; 1831 // CurEntry should end the current section and NextEntry should start 1832 // the next section and the Values must match for these two ranges to be 1833 // merged. 1834 if (CurEntry->getEndSym() != RangeMBB->getEndSymbol() || 1835 NextEntry->getBeginSym() != RangeMBB->getNextNode()->getSymbol() || 1836 CurEntry->getValues() != NextEntry->getValues()) 1837 return false; 1838 RangeMBB = RangeMBB->getNextNode(); 1839 CurEntry = NextEntry; 1840 NextEntry = std::next(CurEntry); 1841 } 1842 return true; 1843 } 1844 1845 DbgEntity *DwarfDebug::createConcreteEntity(DwarfCompileUnit &TheCU, 1846 LexicalScope &Scope, 1847 const DINode *Node, 1848 const DILocation *Location, 1849 const MCSymbol *Sym) { 1850 ensureAbstractEntityIsCreatedIfScoped(TheCU, Node, Scope.getScopeNode()); 1851 if (isa<const DILocalVariable>(Node)) { 1852 ConcreteEntities.push_back( 1853 std::make_unique<DbgVariable>(cast<const DILocalVariable>(Node), 1854 Location)); 1855 InfoHolder.addScopeVariable(&Scope, 1856 cast<DbgVariable>(ConcreteEntities.back().get())); 1857 } else if (isa<const DILabel>(Node)) { 1858 ConcreteEntities.push_back( 1859 std::make_unique<DbgLabel>(cast<const DILabel>(Node), 1860 Location, Sym)); 1861 InfoHolder.addScopeLabel(&Scope, 1862 cast<DbgLabel>(ConcreteEntities.back().get())); 1863 } 1864 return ConcreteEntities.back().get(); 1865 } 1866 1867 // Find variables for each lexical scope. 1868 void DwarfDebug::collectEntityInfo(DwarfCompileUnit &TheCU, 1869 const DISubprogram *SP, 1870 DenseSet<InlinedEntity> &Processed) { 1871 // Grab the variable info that was squirreled away in the MMI side-table. 1872 collectVariableInfoFromMFTable(TheCU, Processed); 1873 1874 for (const auto &I : DbgValues) { 1875 InlinedEntity IV = I.first; 1876 if (Processed.count(IV)) 1877 continue; 1878 1879 // Instruction ranges, specifying where IV is accessible. 1880 const auto &HistoryMapEntries = I.second; 1881 1882 // Try to find any non-empty variable location. Do not create a concrete 1883 // entity if there are no locations. 1884 if (!DbgValues.hasNonEmptyLocation(HistoryMapEntries)) 1885 continue; 1886 1887 LexicalScope *Scope = nullptr; 1888 const DILocalVariable *LocalVar = cast<DILocalVariable>(IV.first); 1889 if (const DILocation *IA = IV.second) 1890 Scope = LScopes.findInlinedScope(LocalVar->getScope(), IA); 1891 else 1892 Scope = LScopes.findLexicalScope(LocalVar->getScope()); 1893 // If variable scope is not found then skip this variable. 1894 if (!Scope) 1895 continue; 1896 1897 Processed.insert(IV); 1898 DbgVariable *RegVar = cast<DbgVariable>(createConcreteEntity(TheCU, 1899 *Scope, LocalVar, IV.second)); 1900 1901 const MachineInstr *MInsn = HistoryMapEntries.front().getInstr(); 1902 assert(MInsn->isDebugValue() && "History must begin with debug value"); 1903 1904 // Check if there is a single DBG_VALUE, valid throughout the var's scope. 1905 // If the history map contains a single debug value, there may be an 1906 // additional entry which clobbers the debug value. 1907 size_t HistSize = HistoryMapEntries.size(); 1908 bool SingleValueWithClobber = 1909 HistSize == 2 && HistoryMapEntries[1].isClobber(); 1910 if (HistSize == 1 || SingleValueWithClobber) { 1911 const auto *End = 1912 SingleValueWithClobber ? HistoryMapEntries[1].getInstr() : nullptr; 1913 if (validThroughout(LScopes, MInsn, End, getInstOrdering())) { 1914 RegVar->initializeDbgValue(MInsn); 1915 continue; 1916 } 1917 } 1918 1919 // Do not emit location lists if .debug_loc secton is disabled. 1920 if (!useLocSection()) 1921 continue; 1922 1923 // Handle multiple DBG_VALUE instructions describing one variable. 1924 DebugLocStream::ListBuilder List(DebugLocs, TheCU, *Asm, *RegVar, *MInsn); 1925 1926 // Build the location list for this variable. 1927 SmallVector<DebugLocEntry, 8> Entries; 1928 bool isValidSingleLocation = buildLocationList(Entries, HistoryMapEntries); 1929 1930 // Check whether buildLocationList managed to merge all locations to one 1931 // that is valid throughout the variable's scope. If so, produce single 1932 // value location. 1933 if (isValidSingleLocation) { 1934 RegVar->initializeDbgValue(Entries[0].getValues()[0]); 1935 continue; 1936 } 1937 1938 // If the variable has a DIBasicType, extract it. Basic types cannot have 1939 // unique identifiers, so don't bother resolving the type with the 1940 // identifier map. 1941 const DIBasicType *BT = dyn_cast<DIBasicType>( 1942 static_cast<const Metadata *>(LocalVar->getType())); 1943 1944 // Finalize the entry by lowering it into a DWARF bytestream. 1945 for (auto &Entry : Entries) 1946 Entry.finalize(*Asm, List, BT, TheCU); 1947 } 1948 1949 // For each InlinedEntity collected from DBG_LABEL instructions, convert to 1950 // DWARF-related DbgLabel. 1951 for (const auto &I : DbgLabels) { 1952 InlinedEntity IL = I.first; 1953 const MachineInstr *MI = I.second; 1954 if (MI == nullptr) 1955 continue; 1956 1957 LexicalScope *Scope = nullptr; 1958 const DILabel *Label = cast<DILabel>(IL.first); 1959 // The scope could have an extra lexical block file. 1960 const DILocalScope *LocalScope = 1961 Label->getScope()->getNonLexicalBlockFileScope(); 1962 // Get inlined DILocation if it is inlined label. 1963 if (const DILocation *IA = IL.second) 1964 Scope = LScopes.findInlinedScope(LocalScope, IA); 1965 else 1966 Scope = LScopes.findLexicalScope(LocalScope); 1967 // If label scope is not found then skip this label. 1968 if (!Scope) 1969 continue; 1970 1971 Processed.insert(IL); 1972 /// At this point, the temporary label is created. 1973 /// Save the temporary label to DbgLabel entity to get the 1974 /// actually address when generating Dwarf DIE. 1975 MCSymbol *Sym = getLabelBeforeInsn(MI); 1976 createConcreteEntity(TheCU, *Scope, Label, IL.second, Sym); 1977 } 1978 1979 // Collect info for variables/labels that were optimized out. 1980 for (const DINode *DN : SP->getRetainedNodes()) { 1981 if (!Processed.insert(InlinedEntity(DN, nullptr)).second) 1982 continue; 1983 LexicalScope *Scope = nullptr; 1984 if (auto *DV = dyn_cast<DILocalVariable>(DN)) { 1985 Scope = LScopes.findLexicalScope(DV->getScope()); 1986 } else if (auto *DL = dyn_cast<DILabel>(DN)) { 1987 Scope = LScopes.findLexicalScope(DL->getScope()); 1988 } 1989 1990 if (Scope) 1991 createConcreteEntity(TheCU, *Scope, DN, nullptr); 1992 } 1993 } 1994 1995 // Process beginning of an instruction. 1996 void DwarfDebug::beginInstruction(const MachineInstr *MI) { 1997 const MachineFunction &MF = *MI->getMF(); 1998 const auto *SP = MF.getFunction().getSubprogram(); 1999 bool NoDebug = 2000 !SP || SP->getUnit()->getEmissionKind() == DICompileUnit::NoDebug; 2001 2002 // Delay slot support check. 2003 auto delaySlotSupported = [](const MachineInstr &MI) { 2004 if (!MI.isBundledWithSucc()) 2005 return false; 2006 auto Suc = std::next(MI.getIterator()); 2007 (void)Suc; 2008 // Ensure that delay slot instruction is successor of the call instruction. 2009 // Ex. CALL_INSTRUCTION { 2010 // DELAY_SLOT_INSTRUCTION } 2011 assert(Suc->isBundledWithPred() && 2012 "Call bundle instructions are out of order"); 2013 return true; 2014 }; 2015 2016 // When describing calls, we need a label for the call instruction. 2017 if (!NoDebug && SP->areAllCallsDescribed() && 2018 MI->isCandidateForCallSiteEntry(MachineInstr::AnyInBundle) && 2019 (!MI->hasDelaySlot() || delaySlotSupported(*MI))) { 2020 const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo(); 2021 bool IsTail = TII->isTailCall(*MI); 2022 // For tail calls, we need the address of the branch instruction for 2023 // DW_AT_call_pc. 2024 if (IsTail) 2025 requestLabelBeforeInsn(MI); 2026 // For non-tail calls, we need the return address for the call for 2027 // DW_AT_call_return_pc. Under GDB tuning, this information is needed for 2028 // tail calls as well. 2029 requestLabelAfterInsn(MI); 2030 } 2031 2032 DebugHandlerBase::beginInstruction(MI); 2033 if (!CurMI) 2034 return; 2035 2036 if (NoDebug) 2037 return; 2038 2039 // Check if source location changes, but ignore DBG_VALUE and CFI locations. 2040 // If the instruction is part of the function frame setup code, do not emit 2041 // any line record, as there is no correspondence with any user code. 2042 if (MI->isMetaInstruction() || MI->getFlag(MachineInstr::FrameSetup)) 2043 return; 2044 const DebugLoc &DL = MI->getDebugLoc(); 2045 // When we emit a line-0 record, we don't update PrevInstLoc; so look at 2046 // the last line number actually emitted, to see if it was line 0. 2047 unsigned LastAsmLine = 2048 Asm->OutStreamer->getContext().getCurrentDwarfLoc().getLine(); 2049 2050 if (DL == PrevInstLoc) { 2051 // If we have an ongoing unspecified location, nothing to do here. 2052 if (!DL) 2053 return; 2054 // We have an explicit location, same as the previous location. 2055 // But we might be coming back to it after a line 0 record. 2056 if (LastAsmLine == 0 && DL.getLine() != 0) { 2057 // Reinstate the source location but not marked as a statement. 2058 const MDNode *Scope = DL.getScope(); 2059 recordSourceLine(DL.getLine(), DL.getCol(), Scope, /*Flags=*/0); 2060 } 2061 return; 2062 } 2063 2064 if (!DL) { 2065 // We have an unspecified location, which might want to be line 0. 2066 // If we have already emitted a line-0 record, don't repeat it. 2067 if (LastAsmLine == 0) 2068 return; 2069 // If user said Don't Do That, don't do that. 2070 if (UnknownLocations == Disable) 2071 return; 2072 // See if we have a reason to emit a line-0 record now. 2073 // Reasons to emit a line-0 record include: 2074 // - User asked for it (UnknownLocations). 2075 // - Instruction has a label, so it's referenced from somewhere else, 2076 // possibly debug information; we want it to have a source location. 2077 // - Instruction is at the top of a block; we don't want to inherit the 2078 // location from the physically previous (maybe unrelated) block. 2079 if (UnknownLocations == Enable || PrevLabel || 2080 (PrevInstBB && PrevInstBB != MI->getParent())) { 2081 // Preserve the file and column numbers, if we can, to save space in 2082 // the encoded line table. 2083 // Do not update PrevInstLoc, it remembers the last non-0 line. 2084 const MDNode *Scope = nullptr; 2085 unsigned Column = 0; 2086 if (PrevInstLoc) { 2087 Scope = PrevInstLoc.getScope(); 2088 Column = PrevInstLoc.getCol(); 2089 } 2090 recordSourceLine(/*Line=*/0, Column, Scope, /*Flags=*/0); 2091 } 2092 return; 2093 } 2094 2095 // We have an explicit location, different from the previous location. 2096 // Don't repeat a line-0 record, but otherwise emit the new location. 2097 // (The new location might be an explicit line 0, which we do emit.) 2098 if (DL.getLine() == 0 && LastAsmLine == 0) 2099 return; 2100 unsigned Flags = 0; 2101 if (DL == PrologEndLoc) { 2102 Flags |= DWARF2_FLAG_PROLOGUE_END | DWARF2_FLAG_IS_STMT; 2103 PrologEndLoc = DebugLoc(); 2104 } 2105 // If the line changed, we call that a new statement; unless we went to 2106 // line 0 and came back, in which case it is not a new statement. 2107 unsigned OldLine = PrevInstLoc ? PrevInstLoc.getLine() : LastAsmLine; 2108 if (DL.getLine() && DL.getLine() != OldLine) 2109 Flags |= DWARF2_FLAG_IS_STMT; 2110 2111 const MDNode *Scope = DL.getScope(); 2112 recordSourceLine(DL.getLine(), DL.getCol(), Scope, Flags); 2113 2114 // If we're not at line 0, remember this location. 2115 if (DL.getLine()) 2116 PrevInstLoc = DL; 2117 } 2118 2119 static DebugLoc findPrologueEndLoc(const MachineFunction *MF) { 2120 // First known non-DBG_VALUE and non-frame setup location marks 2121 // the beginning of the function body. 2122 DebugLoc LineZeroLoc; 2123 for (const auto &MBB : *MF) { 2124 for (const auto &MI : MBB) { 2125 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 2126 MI.getDebugLoc()) { 2127 // Scan forward to try to find a non-zero line number. The prologue_end 2128 // marks the first breakpoint in the function after the frame setup, and 2129 // a compiler-generated line 0 location is not a meaningful breakpoint. 2130 // If none is found, return the first location after the frame setup. 2131 if (MI.getDebugLoc().getLine()) 2132 return MI.getDebugLoc(); 2133 LineZeroLoc = MI.getDebugLoc(); 2134 } 2135 } 2136 } 2137 return LineZeroLoc; 2138 } 2139 2140 /// Register a source line with debug info. Returns the unique label that was 2141 /// emitted and which provides correspondence to the source line list. 2142 static void recordSourceLine(AsmPrinter &Asm, unsigned Line, unsigned Col, 2143 const MDNode *S, unsigned Flags, unsigned CUID, 2144 uint16_t DwarfVersion, 2145 ArrayRef<std::unique_ptr<DwarfCompileUnit>> DCUs) { 2146 StringRef Fn; 2147 unsigned FileNo = 1; 2148 unsigned Discriminator = 0; 2149 if (auto *Scope = cast_or_null<DIScope>(S)) { 2150 Fn = Scope->getFilename(); 2151 if (Line != 0 && DwarfVersion >= 4) 2152 if (auto *LBF = dyn_cast<DILexicalBlockFile>(Scope)) 2153 Discriminator = LBF->getDiscriminator(); 2154 2155 FileNo = static_cast<DwarfCompileUnit &>(*DCUs[CUID]) 2156 .getOrCreateSourceID(Scope->getFile()); 2157 } 2158 Asm.OutStreamer->emitDwarfLocDirective(FileNo, Line, Col, Flags, 0, 2159 Discriminator, Fn); 2160 } 2161 2162 DebugLoc DwarfDebug::emitInitialLocDirective(const MachineFunction &MF, 2163 unsigned CUID) { 2164 // Get beginning of function. 2165 if (DebugLoc PrologEndLoc = findPrologueEndLoc(&MF)) { 2166 // Ensure the compile unit is created if the function is called before 2167 // beginFunction(). 2168 (void)getOrCreateDwarfCompileUnit( 2169 MF.getFunction().getSubprogram()->getUnit()); 2170 // We'd like to list the prologue as "not statements" but GDB behaves 2171 // poorly if we do that. Revisit this with caution/GDB (7.5+) testing. 2172 const DISubprogram *SP = PrologEndLoc->getInlinedAtScope()->getSubprogram(); 2173 ::recordSourceLine(*Asm, SP->getScopeLine(), 0, SP, DWARF2_FLAG_IS_STMT, 2174 CUID, getDwarfVersion(), getUnits()); 2175 return PrologEndLoc; 2176 } 2177 return DebugLoc(); 2178 } 2179 2180 // Gather pre-function debug information. Assumes being called immediately 2181 // after the function entry point has been emitted. 2182 void DwarfDebug::beginFunctionImpl(const MachineFunction *MF) { 2183 CurFn = MF; 2184 2185 auto *SP = MF->getFunction().getSubprogram(); 2186 assert(LScopes.empty() || SP == LScopes.getCurrentFunctionScope()->getScopeNode()); 2187 if (SP->getUnit()->getEmissionKind() == DICompileUnit::NoDebug) 2188 return; 2189 2190 DwarfCompileUnit &CU = getOrCreateDwarfCompileUnit(SP->getUnit()); 2191 2192 Asm->OutStreamer->getContext().setDwarfCompileUnitID( 2193 getDwarfCompileUnitIDForLineTable(CU)); 2194 2195 // Record beginning of function. 2196 PrologEndLoc = emitInitialLocDirective( 2197 *MF, Asm->OutStreamer->getContext().getDwarfCompileUnitID()); 2198 } 2199 2200 unsigned 2201 DwarfDebug::getDwarfCompileUnitIDForLineTable(const DwarfCompileUnit &CU) { 2202 // Set DwarfDwarfCompileUnitID in MCContext to the Compile Unit this function 2203 // belongs to so that we add to the correct per-cu line table in the 2204 // non-asm case. 2205 if (Asm->OutStreamer->hasRawTextSupport()) 2206 // Use a single line table if we are generating assembly. 2207 return 0; 2208 else 2209 return CU.getUniqueID(); 2210 } 2211 2212 void DwarfDebug::terminateLineTable(const DwarfCompileUnit *CU) { 2213 const auto &CURanges = CU->getRanges(); 2214 auto &LineTable = Asm->OutStreamer->getContext().getMCDwarfLineTable( 2215 getDwarfCompileUnitIDForLineTable(*CU)); 2216 // Add the last range label for the given CU. 2217 LineTable.getMCLineSections().addEndEntry( 2218 const_cast<MCSymbol *>(CURanges.back().End)); 2219 } 2220 2221 void DwarfDebug::skippedNonDebugFunction() { 2222 // If we don't have a subprogram for this function then there will be a hole 2223 // in the range information. Keep note of this by setting the previously used 2224 // section to nullptr. 2225 // Terminate the pending line table. 2226 if (PrevCU) 2227 terminateLineTable(PrevCU); 2228 PrevCU = nullptr; 2229 CurFn = nullptr; 2230 } 2231 2232 // Gather and emit post-function debug information. 2233 void DwarfDebug::endFunctionImpl(const MachineFunction *MF) { 2234 const DISubprogram *SP = MF->getFunction().getSubprogram(); 2235 2236 assert(CurFn == MF && 2237 "endFunction should be called with the same function as beginFunction"); 2238 2239 // Set DwarfDwarfCompileUnitID in MCContext to default value. 2240 Asm->OutStreamer->getContext().setDwarfCompileUnitID(0); 2241 2242 LexicalScope *FnScope = LScopes.getCurrentFunctionScope(); 2243 assert(!FnScope || SP == FnScope->getScopeNode()); 2244 DwarfCompileUnit &TheCU = *CUMap.lookup(SP->getUnit()); 2245 if (TheCU.getCUNode()->isDebugDirectivesOnly()) { 2246 PrevLabel = nullptr; 2247 CurFn = nullptr; 2248 return; 2249 } 2250 2251 DenseSet<InlinedEntity> Processed; 2252 collectEntityInfo(TheCU, SP, Processed); 2253 2254 // Add the range of this function to the list of ranges for the CU. 2255 // With basic block sections, add ranges for all basic block sections. 2256 for (const auto &R : Asm->MBBSectionRanges) 2257 TheCU.addRange({R.second.BeginLabel, R.second.EndLabel}); 2258 2259 // Under -gmlt, skip building the subprogram if there are no inlined 2260 // subroutines inside it. But with -fdebug-info-for-profiling, the subprogram 2261 // is still needed as we need its source location. 2262 if (!TheCU.getCUNode()->getDebugInfoForProfiling() && 2263 TheCU.getCUNode()->getEmissionKind() == DICompileUnit::LineTablesOnly && 2264 LScopes.getAbstractScopesList().empty() && !IsDarwin) { 2265 assert(InfoHolder.getScopeVariables().empty()); 2266 PrevLabel = nullptr; 2267 CurFn = nullptr; 2268 return; 2269 } 2270 2271 #ifndef NDEBUG 2272 size_t NumAbstractScopes = LScopes.getAbstractScopesList().size(); 2273 #endif 2274 // Construct abstract scopes. 2275 for (LexicalScope *AScope : LScopes.getAbstractScopesList()) { 2276 auto *SP = cast<DISubprogram>(AScope->getScopeNode()); 2277 for (const DINode *DN : SP->getRetainedNodes()) { 2278 if (!Processed.insert(InlinedEntity(DN, nullptr)).second) 2279 continue; 2280 2281 const MDNode *Scope = nullptr; 2282 if (auto *DV = dyn_cast<DILocalVariable>(DN)) 2283 Scope = DV->getScope(); 2284 else if (auto *DL = dyn_cast<DILabel>(DN)) 2285 Scope = DL->getScope(); 2286 else 2287 llvm_unreachable("Unexpected DI type!"); 2288 2289 // Collect info for variables/labels that were optimized out. 2290 ensureAbstractEntityIsCreated(TheCU, DN, Scope); 2291 assert(LScopes.getAbstractScopesList().size() == NumAbstractScopes 2292 && "ensureAbstractEntityIsCreated inserted abstract scopes"); 2293 } 2294 constructAbstractSubprogramScopeDIE(TheCU, AScope); 2295 } 2296 2297 ProcessedSPNodes.insert(SP); 2298 DIE &ScopeDIE = TheCU.constructSubprogramScopeDIE(SP, FnScope); 2299 if (auto *SkelCU = TheCU.getSkeleton()) 2300 if (!LScopes.getAbstractScopesList().empty() && 2301 TheCU.getCUNode()->getSplitDebugInlining()) 2302 SkelCU->constructSubprogramScopeDIE(SP, FnScope); 2303 2304 // Construct call site entries. 2305 constructCallSiteEntryDIEs(*SP, TheCU, ScopeDIE, *MF); 2306 2307 // Clear debug info 2308 // Ownership of DbgVariables is a bit subtle - ScopeVariables owns all the 2309 // DbgVariables except those that are also in AbstractVariables (since they 2310 // can be used cross-function) 2311 InfoHolder.getScopeVariables().clear(); 2312 InfoHolder.getScopeLabels().clear(); 2313 PrevLabel = nullptr; 2314 CurFn = nullptr; 2315 } 2316 2317 // Register a source line with debug info. Returns the unique label that was 2318 // emitted and which provides correspondence to the source line list. 2319 void DwarfDebug::recordSourceLine(unsigned Line, unsigned Col, const MDNode *S, 2320 unsigned Flags) { 2321 ::recordSourceLine(*Asm, Line, Col, S, Flags, 2322 Asm->OutStreamer->getContext().getDwarfCompileUnitID(), 2323 getDwarfVersion(), getUnits()); 2324 } 2325 2326 //===----------------------------------------------------------------------===// 2327 // Emit Methods 2328 //===----------------------------------------------------------------------===// 2329 2330 // Emit the debug info section. 2331 void DwarfDebug::emitDebugInfo() { 2332 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 2333 Holder.emitUnits(/* UseOffsets */ false); 2334 } 2335 2336 // Emit the abbreviation section. 2337 void DwarfDebug::emitAbbreviations() { 2338 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 2339 2340 Holder.emitAbbrevs(Asm->getObjFileLowering().getDwarfAbbrevSection()); 2341 } 2342 2343 void DwarfDebug::emitStringOffsetsTableHeader() { 2344 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 2345 Holder.getStringPool().emitStringOffsetsTableHeader( 2346 *Asm, Asm->getObjFileLowering().getDwarfStrOffSection(), 2347 Holder.getStringOffsetsStartSym()); 2348 } 2349 2350 template <typename AccelTableT> 2351 void DwarfDebug::emitAccel(AccelTableT &Accel, MCSection *Section, 2352 StringRef TableName) { 2353 Asm->OutStreamer->SwitchSection(Section); 2354 2355 // Emit the full data. 2356 emitAppleAccelTable(Asm, Accel, TableName, Section->getBeginSymbol()); 2357 } 2358 2359 void DwarfDebug::emitAccelDebugNames() { 2360 // Don't emit anything if we have no compilation units to index. 2361 if (getUnits().empty()) 2362 return; 2363 2364 emitDWARF5AccelTable(Asm, AccelDebugNames, *this, getUnits()); 2365 } 2366 2367 // Emit visible names into a hashed accelerator table section. 2368 void DwarfDebug::emitAccelNames() { 2369 emitAccel(AccelNames, Asm->getObjFileLowering().getDwarfAccelNamesSection(), 2370 "Names"); 2371 } 2372 2373 // Emit objective C classes and categories into a hashed accelerator table 2374 // section. 2375 void DwarfDebug::emitAccelObjC() { 2376 emitAccel(AccelObjC, Asm->getObjFileLowering().getDwarfAccelObjCSection(), 2377 "ObjC"); 2378 } 2379 2380 // Emit namespace dies into a hashed accelerator table. 2381 void DwarfDebug::emitAccelNamespaces() { 2382 emitAccel(AccelNamespace, 2383 Asm->getObjFileLowering().getDwarfAccelNamespaceSection(), 2384 "namespac"); 2385 } 2386 2387 // Emit type dies into a hashed accelerator table. 2388 void DwarfDebug::emitAccelTypes() { 2389 emitAccel(AccelTypes, Asm->getObjFileLowering().getDwarfAccelTypesSection(), 2390 "types"); 2391 } 2392 2393 // Public name handling. 2394 // The format for the various pubnames: 2395 // 2396 // dwarf pubnames - offset/name pairs where the offset is the offset into the CU 2397 // for the DIE that is named. 2398 // 2399 // gnu pubnames - offset/index value/name tuples where the offset is the offset 2400 // into the CU and the index value is computed according to the type of value 2401 // for the DIE that is named. 2402 // 2403 // For type units the offset is the offset of the skeleton DIE. For split dwarf 2404 // it's the offset within the debug_info/debug_types dwo section, however, the 2405 // reference in the pubname header doesn't change. 2406 2407 /// computeIndexValue - Compute the gdb index value for the DIE and CU. 2408 static dwarf::PubIndexEntryDescriptor computeIndexValue(DwarfUnit *CU, 2409 const DIE *Die) { 2410 // Entities that ended up only in a Type Unit reference the CU instead (since 2411 // the pub entry has offsets within the CU there's no real offset that can be 2412 // provided anyway). As it happens all such entities (namespaces and types, 2413 // types only in C++ at that) are rendered as TYPE+EXTERNAL. If this turns out 2414 // not to be true it would be necessary to persist this information from the 2415 // point at which the entry is added to the index data structure - since by 2416 // the time the index is built from that, the original type/namespace DIE in a 2417 // type unit has already been destroyed so it can't be queried for properties 2418 // like tag, etc. 2419 if (Die->getTag() == dwarf::DW_TAG_compile_unit) 2420 return dwarf::PubIndexEntryDescriptor(dwarf::GIEK_TYPE, 2421 dwarf::GIEL_EXTERNAL); 2422 dwarf::GDBIndexEntryLinkage Linkage = dwarf::GIEL_STATIC; 2423 2424 // We could have a specification DIE that has our most of our knowledge, 2425 // look for that now. 2426 if (DIEValue SpecVal = Die->findAttribute(dwarf::DW_AT_specification)) { 2427 DIE &SpecDIE = SpecVal.getDIEEntry().getEntry(); 2428 if (SpecDIE.findAttribute(dwarf::DW_AT_external)) 2429 Linkage = dwarf::GIEL_EXTERNAL; 2430 } else if (Die->findAttribute(dwarf::DW_AT_external)) 2431 Linkage = dwarf::GIEL_EXTERNAL; 2432 2433 switch (Die->getTag()) { 2434 case dwarf::DW_TAG_class_type: 2435 case dwarf::DW_TAG_structure_type: 2436 case dwarf::DW_TAG_union_type: 2437 case dwarf::DW_TAG_enumeration_type: 2438 return dwarf::PubIndexEntryDescriptor( 2439 dwarf::GIEK_TYPE, 2440 dwarf::isCPlusPlus((dwarf::SourceLanguage)CU->getLanguage()) 2441 ? dwarf::GIEL_EXTERNAL 2442 : dwarf::GIEL_STATIC); 2443 case dwarf::DW_TAG_typedef: 2444 case dwarf::DW_TAG_base_type: 2445 case dwarf::DW_TAG_subrange_type: 2446 return dwarf::PubIndexEntryDescriptor(dwarf::GIEK_TYPE, dwarf::GIEL_STATIC); 2447 case dwarf::DW_TAG_namespace: 2448 return dwarf::GIEK_TYPE; 2449 case dwarf::DW_TAG_subprogram: 2450 return dwarf::PubIndexEntryDescriptor(dwarf::GIEK_FUNCTION, Linkage); 2451 case dwarf::DW_TAG_variable: 2452 return dwarf::PubIndexEntryDescriptor(dwarf::GIEK_VARIABLE, Linkage); 2453 case dwarf::DW_TAG_enumerator: 2454 return dwarf::PubIndexEntryDescriptor(dwarf::GIEK_VARIABLE, 2455 dwarf::GIEL_STATIC); 2456 default: 2457 return dwarf::GIEK_NONE; 2458 } 2459 } 2460 2461 /// emitDebugPubSections - Emit visible names and types into debug pubnames and 2462 /// pubtypes sections. 2463 void DwarfDebug::emitDebugPubSections() { 2464 for (const auto &NU : CUMap) { 2465 DwarfCompileUnit *TheU = NU.second; 2466 if (!TheU->hasDwarfPubSections()) 2467 continue; 2468 2469 bool GnuStyle = TheU->getCUNode()->getNameTableKind() == 2470 DICompileUnit::DebugNameTableKind::GNU; 2471 2472 Asm->OutStreamer->SwitchSection( 2473 GnuStyle ? Asm->getObjFileLowering().getDwarfGnuPubNamesSection() 2474 : Asm->getObjFileLowering().getDwarfPubNamesSection()); 2475 emitDebugPubSection(GnuStyle, "Names", TheU, TheU->getGlobalNames()); 2476 2477 Asm->OutStreamer->SwitchSection( 2478 GnuStyle ? Asm->getObjFileLowering().getDwarfGnuPubTypesSection() 2479 : Asm->getObjFileLowering().getDwarfPubTypesSection()); 2480 emitDebugPubSection(GnuStyle, "Types", TheU, TheU->getGlobalTypes()); 2481 } 2482 } 2483 2484 void DwarfDebug::emitSectionReference(const DwarfCompileUnit &CU) { 2485 if (useSectionsAsReferences()) 2486 Asm->emitDwarfOffset(CU.getSection()->getBeginSymbol(), 2487 CU.getDebugSectionOffset()); 2488 else 2489 Asm->emitDwarfSymbolReference(CU.getLabelBegin()); 2490 } 2491 2492 void DwarfDebug::emitDebugPubSection(bool GnuStyle, StringRef Name, 2493 DwarfCompileUnit *TheU, 2494 const StringMap<const DIE *> &Globals) { 2495 if (auto *Skeleton = TheU->getSkeleton()) 2496 TheU = Skeleton; 2497 2498 // Emit the header. 2499 MCSymbol *EndLabel = Asm->emitDwarfUnitLength( 2500 "pub" + Name, "Length of Public " + Name + " Info"); 2501 2502 Asm->OutStreamer->AddComment("DWARF Version"); 2503 Asm->emitInt16(dwarf::DW_PUBNAMES_VERSION); 2504 2505 Asm->OutStreamer->AddComment("Offset of Compilation Unit Info"); 2506 emitSectionReference(*TheU); 2507 2508 Asm->OutStreamer->AddComment("Compilation Unit Length"); 2509 Asm->emitDwarfLengthOrOffset(TheU->getLength()); 2510 2511 // Emit the pubnames for this compilation unit. 2512 for (const auto &GI : Globals) { 2513 const char *Name = GI.getKeyData(); 2514 const DIE *Entity = GI.second; 2515 2516 Asm->OutStreamer->AddComment("DIE offset"); 2517 Asm->emitDwarfLengthOrOffset(Entity->getOffset()); 2518 2519 if (GnuStyle) { 2520 dwarf::PubIndexEntryDescriptor Desc = computeIndexValue(TheU, Entity); 2521 Asm->OutStreamer->AddComment( 2522 Twine("Attributes: ") + dwarf::GDBIndexEntryKindString(Desc.Kind) + 2523 ", " + dwarf::GDBIndexEntryLinkageString(Desc.Linkage)); 2524 Asm->emitInt8(Desc.toBits()); 2525 } 2526 2527 Asm->OutStreamer->AddComment("External Name"); 2528 Asm->OutStreamer->emitBytes(StringRef(Name, GI.getKeyLength() + 1)); 2529 } 2530 2531 Asm->OutStreamer->AddComment("End Mark"); 2532 Asm->emitDwarfLengthOrOffset(0); 2533 Asm->OutStreamer->emitLabel(EndLabel); 2534 } 2535 2536 /// Emit null-terminated strings into a debug str section. 2537 void DwarfDebug::emitDebugStr() { 2538 MCSection *StringOffsetsSection = nullptr; 2539 if (useSegmentedStringOffsetsTable()) { 2540 emitStringOffsetsTableHeader(); 2541 StringOffsetsSection = Asm->getObjFileLowering().getDwarfStrOffSection(); 2542 } 2543 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 2544 Holder.emitStrings(Asm->getObjFileLowering().getDwarfStrSection(), 2545 StringOffsetsSection, /* UseRelativeOffsets = */ true); 2546 } 2547 2548 void DwarfDebug::emitDebugLocEntry(ByteStreamer &Streamer, 2549 const DebugLocStream::Entry &Entry, 2550 const DwarfCompileUnit *CU) { 2551 auto &&Comments = DebugLocs.getComments(Entry); 2552 auto Comment = Comments.begin(); 2553 auto End = Comments.end(); 2554 2555 // The expressions are inserted into a byte stream rather early (see 2556 // DwarfExpression::addExpression) so for those ops (e.g. DW_OP_convert) that 2557 // need to reference a base_type DIE the offset of that DIE is not yet known. 2558 // To deal with this we instead insert a placeholder early and then extract 2559 // it here and replace it with the real reference. 2560 unsigned PtrSize = Asm->MAI->getCodePointerSize(); 2561 DWARFDataExtractor Data(StringRef(DebugLocs.getBytes(Entry).data(), 2562 DebugLocs.getBytes(Entry).size()), 2563 Asm->getDataLayout().isLittleEndian(), PtrSize); 2564 DWARFExpression Expr(Data, PtrSize, Asm->OutContext.getDwarfFormat()); 2565 2566 using Encoding = DWARFExpression::Operation::Encoding; 2567 uint64_t Offset = 0; 2568 for (auto &Op : Expr) { 2569 assert(Op.getCode() != dwarf::DW_OP_const_type && 2570 "3 operand ops not yet supported"); 2571 Streamer.emitInt8(Op.getCode(), Comment != End ? *(Comment++) : ""); 2572 Offset++; 2573 for (unsigned I = 0; I < 2; ++I) { 2574 if (Op.getDescription().Op[I] == Encoding::SizeNA) 2575 continue; 2576 if (Op.getDescription().Op[I] == Encoding::BaseTypeRef) { 2577 uint64_t Offset = 2578 CU->ExprRefedBaseTypes[Op.getRawOperand(I)].Die->getOffset(); 2579 assert(Offset < (1ULL << (ULEB128PadSize * 7)) && "Offset wont fit"); 2580 Streamer.emitULEB128(Offset, "", ULEB128PadSize); 2581 // Make sure comments stay aligned. 2582 for (unsigned J = 0; J < ULEB128PadSize; ++J) 2583 if (Comment != End) 2584 Comment++; 2585 } else { 2586 for (uint64_t J = Offset; J < Op.getOperandEndOffset(I); ++J) 2587 Streamer.emitInt8(Data.getData()[J], Comment != End ? *(Comment++) : ""); 2588 } 2589 Offset = Op.getOperandEndOffset(I); 2590 } 2591 assert(Offset == Op.getEndOffset()); 2592 } 2593 } 2594 2595 void DwarfDebug::emitDebugLocValue(const AsmPrinter &AP, const DIBasicType *BT, 2596 const DbgValueLoc &Value, 2597 DwarfExpression &DwarfExpr) { 2598 auto *DIExpr = Value.getExpression(); 2599 DIExpressionCursor ExprCursor(DIExpr); 2600 DwarfExpr.addFragmentOffset(DIExpr); 2601 2602 // If the DIExpr is is an Entry Value, we want to follow the same code path 2603 // regardless of whether the DBG_VALUE is variadic or not. 2604 if (DIExpr && DIExpr->isEntryValue()) { 2605 // Entry values can only be a single register with no additional DIExpr, 2606 // so just add it directly. 2607 assert(Value.getLocEntries().size() == 1); 2608 assert(Value.getLocEntries()[0].isLocation()); 2609 MachineLocation Location = Value.getLocEntries()[0].getLoc(); 2610 DwarfExpr.setLocation(Location, DIExpr); 2611 2612 DwarfExpr.beginEntryValueExpression(ExprCursor); 2613 2614 const TargetRegisterInfo &TRI = *AP.MF->getSubtarget().getRegisterInfo(); 2615 if (!DwarfExpr.addMachineRegExpression(TRI, ExprCursor, Location.getReg())) 2616 return; 2617 return DwarfExpr.addExpression(std::move(ExprCursor)); 2618 } 2619 2620 // Regular entry. 2621 auto EmitValueLocEntry = [&DwarfExpr, &BT, 2622 &AP](const DbgValueLocEntry &Entry, 2623 DIExpressionCursor &Cursor) -> bool { 2624 if (Entry.isInt()) { 2625 if (BT && (BT->getEncoding() == dwarf::DW_ATE_signed || 2626 BT->getEncoding() == dwarf::DW_ATE_signed_char)) 2627 DwarfExpr.addSignedConstant(Entry.getInt()); 2628 else 2629 DwarfExpr.addUnsignedConstant(Entry.getInt()); 2630 } else if (Entry.isLocation()) { 2631 MachineLocation Location = Entry.getLoc(); 2632 if (Location.isIndirect()) 2633 DwarfExpr.setMemoryLocationKind(); 2634 2635 const TargetRegisterInfo &TRI = *AP.MF->getSubtarget().getRegisterInfo(); 2636 if (!DwarfExpr.addMachineRegExpression(TRI, Cursor, Location.getReg())) 2637 return false; 2638 } else if (Entry.isTargetIndexLocation()) { 2639 TargetIndexLocation Loc = Entry.getTargetIndexLocation(); 2640 // TODO TargetIndexLocation is a target-independent. Currently only the 2641 // WebAssembly-specific encoding is supported. 2642 assert(AP.TM.getTargetTriple().isWasm()); 2643 DwarfExpr.addWasmLocation(Loc.Index, static_cast<uint64_t>(Loc.Offset)); 2644 } else if (Entry.isConstantFP()) { 2645 if (AP.getDwarfVersion() >= 4 && !AP.getDwarfDebug()->tuneForSCE() && 2646 !Cursor) { 2647 DwarfExpr.addConstantFP(Entry.getConstantFP()->getValueAPF(), AP); 2648 } else if (Entry.getConstantFP() 2649 ->getValueAPF() 2650 .bitcastToAPInt() 2651 .getBitWidth() <= 64 /*bits*/) { 2652 DwarfExpr.addUnsignedConstant( 2653 Entry.getConstantFP()->getValueAPF().bitcastToAPInt()); 2654 } else { 2655 LLVM_DEBUG( 2656 dbgs() << "Skipped DwarfExpression creation for ConstantFP of size" 2657 << Entry.getConstantFP() 2658 ->getValueAPF() 2659 .bitcastToAPInt() 2660 .getBitWidth() 2661 << " bits\n"); 2662 return false; 2663 } 2664 } 2665 return true; 2666 }; 2667 2668 if (!Value.isVariadic()) { 2669 if (!EmitValueLocEntry(Value.getLocEntries()[0], ExprCursor)) 2670 return; 2671 DwarfExpr.addExpression(std::move(ExprCursor)); 2672 return; 2673 } 2674 2675 // If any of the location entries are registers with the value 0, then the 2676 // location is undefined. 2677 if (any_of(Value.getLocEntries(), [](const DbgValueLocEntry &Entry) { 2678 return Entry.isLocation() && !Entry.getLoc().getReg(); 2679 })) 2680 return; 2681 2682 DwarfExpr.addExpression( 2683 std::move(ExprCursor), 2684 [EmitValueLocEntry, &Value](unsigned Idx, 2685 DIExpressionCursor &Cursor) -> bool { 2686 return EmitValueLocEntry(Value.getLocEntries()[Idx], Cursor); 2687 }); 2688 } 2689 2690 void DebugLocEntry::finalize(const AsmPrinter &AP, 2691 DebugLocStream::ListBuilder &List, 2692 const DIBasicType *BT, 2693 DwarfCompileUnit &TheCU) { 2694 assert(!Values.empty() && 2695 "location list entries without values are redundant"); 2696 assert(Begin != End && "unexpected location list entry with empty range"); 2697 DebugLocStream::EntryBuilder Entry(List, Begin, End); 2698 BufferByteStreamer Streamer = Entry.getStreamer(); 2699 DebugLocDwarfExpression DwarfExpr(AP.getDwarfVersion(), Streamer, TheCU); 2700 const DbgValueLoc &Value = Values[0]; 2701 if (Value.isFragment()) { 2702 // Emit all fragments that belong to the same variable and range. 2703 assert(llvm::all_of(Values, [](DbgValueLoc P) { 2704 return P.isFragment(); 2705 }) && "all values are expected to be fragments"); 2706 assert(llvm::is_sorted(Values) && "fragments are expected to be sorted"); 2707 2708 for (const auto &Fragment : Values) 2709 DwarfDebug::emitDebugLocValue(AP, BT, Fragment, DwarfExpr); 2710 2711 } else { 2712 assert(Values.size() == 1 && "only fragments may have >1 value"); 2713 DwarfDebug::emitDebugLocValue(AP, BT, Value, DwarfExpr); 2714 } 2715 DwarfExpr.finalize(); 2716 if (DwarfExpr.TagOffset) 2717 List.setTagOffset(*DwarfExpr.TagOffset); 2718 } 2719 2720 void DwarfDebug::emitDebugLocEntryLocation(const DebugLocStream::Entry &Entry, 2721 const DwarfCompileUnit *CU) { 2722 // Emit the size. 2723 Asm->OutStreamer->AddComment("Loc expr size"); 2724 if (getDwarfVersion() >= 5) 2725 Asm->emitULEB128(DebugLocs.getBytes(Entry).size()); 2726 else if (DebugLocs.getBytes(Entry).size() <= std::numeric_limits<uint16_t>::max()) 2727 Asm->emitInt16(DebugLocs.getBytes(Entry).size()); 2728 else { 2729 // The entry is too big to fit into 16 bit, drop it as there is nothing we 2730 // can do. 2731 Asm->emitInt16(0); 2732 return; 2733 } 2734 // Emit the entry. 2735 APByteStreamer Streamer(*Asm); 2736 emitDebugLocEntry(Streamer, Entry, CU); 2737 } 2738 2739 // Emit the header of a DWARF 5 range list table list table. Returns the symbol 2740 // that designates the end of the table for the caller to emit when the table is 2741 // complete. 2742 static MCSymbol *emitRnglistsTableHeader(AsmPrinter *Asm, 2743 const DwarfFile &Holder) { 2744 MCSymbol *TableEnd = mcdwarf::emitListsTableHeaderStart(*Asm->OutStreamer); 2745 2746 Asm->OutStreamer->AddComment("Offset entry count"); 2747 Asm->emitInt32(Holder.getRangeLists().size()); 2748 Asm->OutStreamer->emitLabel(Holder.getRnglistsTableBaseSym()); 2749 2750 for (const RangeSpanList &List : Holder.getRangeLists()) 2751 Asm->emitLabelDifference(List.Label, Holder.getRnglistsTableBaseSym(), 2752 Asm->getDwarfOffsetByteSize()); 2753 2754 return TableEnd; 2755 } 2756 2757 // Emit the header of a DWARF 5 locations list table. Returns the symbol that 2758 // designates the end of the table for the caller to emit when the table is 2759 // complete. 2760 static MCSymbol *emitLoclistsTableHeader(AsmPrinter *Asm, 2761 const DwarfDebug &DD) { 2762 MCSymbol *TableEnd = mcdwarf::emitListsTableHeaderStart(*Asm->OutStreamer); 2763 2764 const auto &DebugLocs = DD.getDebugLocs(); 2765 2766 Asm->OutStreamer->AddComment("Offset entry count"); 2767 Asm->emitInt32(DebugLocs.getLists().size()); 2768 Asm->OutStreamer->emitLabel(DebugLocs.getSym()); 2769 2770 for (const auto &List : DebugLocs.getLists()) 2771 Asm->emitLabelDifference(List.Label, DebugLocs.getSym(), 2772 Asm->getDwarfOffsetByteSize()); 2773 2774 return TableEnd; 2775 } 2776 2777 template <typename Ranges, typename PayloadEmitter> 2778 static void emitRangeList( 2779 DwarfDebug &DD, AsmPrinter *Asm, MCSymbol *Sym, const Ranges &R, 2780 const DwarfCompileUnit &CU, unsigned BaseAddressx, unsigned OffsetPair, 2781 unsigned StartxLength, unsigned EndOfList, 2782 StringRef (*StringifyEnum)(unsigned), 2783 bool ShouldUseBaseAddress, 2784 PayloadEmitter EmitPayload) { 2785 2786 auto Size = Asm->MAI->getCodePointerSize(); 2787 bool UseDwarf5 = DD.getDwarfVersion() >= 5; 2788 2789 // Emit our symbol so we can find the beginning of the range. 2790 Asm->OutStreamer->emitLabel(Sym); 2791 2792 // Gather all the ranges that apply to the same section so they can share 2793 // a base address entry. 2794 MapVector<const MCSection *, std::vector<decltype(&*R.begin())>> SectionRanges; 2795 2796 for (const auto &Range : R) 2797 SectionRanges[&Range.Begin->getSection()].push_back(&Range); 2798 2799 const MCSymbol *CUBase = CU.getBaseAddress(); 2800 bool BaseIsSet = false; 2801 for (const auto &P : SectionRanges) { 2802 auto *Base = CUBase; 2803 if (!Base && ShouldUseBaseAddress) { 2804 const MCSymbol *Begin = P.second.front()->Begin; 2805 const MCSymbol *NewBase = DD.getSectionLabel(&Begin->getSection()); 2806 if (!UseDwarf5) { 2807 Base = NewBase; 2808 BaseIsSet = true; 2809 Asm->OutStreamer->emitIntValue(-1, Size); 2810 Asm->OutStreamer->AddComment(" base address"); 2811 Asm->OutStreamer->emitSymbolValue(Base, Size); 2812 } else if (NewBase != Begin || P.second.size() > 1) { 2813 // Only use a base address if 2814 // * the existing pool address doesn't match (NewBase != Begin) 2815 // * or, there's more than one entry to share the base address 2816 Base = NewBase; 2817 BaseIsSet = true; 2818 Asm->OutStreamer->AddComment(StringifyEnum(BaseAddressx)); 2819 Asm->emitInt8(BaseAddressx); 2820 Asm->OutStreamer->AddComment(" base address index"); 2821 Asm->emitULEB128(DD.getAddressPool().getIndex(Base)); 2822 } 2823 } else if (BaseIsSet && !UseDwarf5) { 2824 BaseIsSet = false; 2825 assert(!Base); 2826 Asm->OutStreamer->emitIntValue(-1, Size); 2827 Asm->OutStreamer->emitIntValue(0, Size); 2828 } 2829 2830 for (const auto *RS : P.second) { 2831 const MCSymbol *Begin = RS->Begin; 2832 const MCSymbol *End = RS->End; 2833 assert(Begin && "Range without a begin symbol?"); 2834 assert(End && "Range without an end symbol?"); 2835 if (Base) { 2836 if (UseDwarf5) { 2837 // Emit offset_pair when we have a base. 2838 Asm->OutStreamer->AddComment(StringifyEnum(OffsetPair)); 2839 Asm->emitInt8(OffsetPair); 2840 Asm->OutStreamer->AddComment(" starting offset"); 2841 Asm->emitLabelDifferenceAsULEB128(Begin, Base); 2842 Asm->OutStreamer->AddComment(" ending offset"); 2843 Asm->emitLabelDifferenceAsULEB128(End, Base); 2844 } else { 2845 Asm->emitLabelDifference(Begin, Base, Size); 2846 Asm->emitLabelDifference(End, Base, Size); 2847 } 2848 } else if (UseDwarf5) { 2849 Asm->OutStreamer->AddComment(StringifyEnum(StartxLength)); 2850 Asm->emitInt8(StartxLength); 2851 Asm->OutStreamer->AddComment(" start index"); 2852 Asm->emitULEB128(DD.getAddressPool().getIndex(Begin)); 2853 Asm->OutStreamer->AddComment(" length"); 2854 Asm->emitLabelDifferenceAsULEB128(End, Begin); 2855 } else { 2856 Asm->OutStreamer->emitSymbolValue(Begin, Size); 2857 Asm->OutStreamer->emitSymbolValue(End, Size); 2858 } 2859 EmitPayload(*RS); 2860 } 2861 } 2862 2863 if (UseDwarf5) { 2864 Asm->OutStreamer->AddComment(StringifyEnum(EndOfList)); 2865 Asm->emitInt8(EndOfList); 2866 } else { 2867 // Terminate the list with two 0 values. 2868 Asm->OutStreamer->emitIntValue(0, Size); 2869 Asm->OutStreamer->emitIntValue(0, Size); 2870 } 2871 } 2872 2873 // Handles emission of both debug_loclist / debug_loclist.dwo 2874 static void emitLocList(DwarfDebug &DD, AsmPrinter *Asm, const DebugLocStream::List &List) { 2875 emitRangeList(DD, Asm, List.Label, DD.getDebugLocs().getEntries(List), 2876 *List.CU, dwarf::DW_LLE_base_addressx, 2877 dwarf::DW_LLE_offset_pair, dwarf::DW_LLE_startx_length, 2878 dwarf::DW_LLE_end_of_list, llvm::dwarf::LocListEncodingString, 2879 /* ShouldUseBaseAddress */ true, 2880 [&](const DebugLocStream::Entry &E) { 2881 DD.emitDebugLocEntryLocation(E, List.CU); 2882 }); 2883 } 2884 2885 void DwarfDebug::emitDebugLocImpl(MCSection *Sec) { 2886 if (DebugLocs.getLists().empty()) 2887 return; 2888 2889 Asm->OutStreamer->SwitchSection(Sec); 2890 2891 MCSymbol *TableEnd = nullptr; 2892 if (getDwarfVersion() >= 5) 2893 TableEnd = emitLoclistsTableHeader(Asm, *this); 2894 2895 for (const auto &List : DebugLocs.getLists()) 2896 emitLocList(*this, Asm, List); 2897 2898 if (TableEnd) 2899 Asm->OutStreamer->emitLabel(TableEnd); 2900 } 2901 2902 // Emit locations into the .debug_loc/.debug_loclists section. 2903 void DwarfDebug::emitDebugLoc() { 2904 emitDebugLocImpl( 2905 getDwarfVersion() >= 5 2906 ? Asm->getObjFileLowering().getDwarfLoclistsSection() 2907 : Asm->getObjFileLowering().getDwarfLocSection()); 2908 } 2909 2910 // Emit locations into the .debug_loc.dwo/.debug_loclists.dwo section. 2911 void DwarfDebug::emitDebugLocDWO() { 2912 if (getDwarfVersion() >= 5) { 2913 emitDebugLocImpl( 2914 Asm->getObjFileLowering().getDwarfLoclistsDWOSection()); 2915 2916 return; 2917 } 2918 2919 for (const auto &List : DebugLocs.getLists()) { 2920 Asm->OutStreamer->SwitchSection( 2921 Asm->getObjFileLowering().getDwarfLocDWOSection()); 2922 Asm->OutStreamer->emitLabel(List.Label); 2923 2924 for (const auto &Entry : DebugLocs.getEntries(List)) { 2925 // GDB only supports startx_length in pre-standard split-DWARF. 2926 // (in v5 standard loclists, it currently* /only/ supports base_address + 2927 // offset_pair, so the implementations can't really share much since they 2928 // need to use different representations) 2929 // * as of October 2018, at least 2930 // 2931 // In v5 (see emitLocList), this uses SectionLabels to reuse existing 2932 // addresses in the address pool to minimize object size/relocations. 2933 Asm->emitInt8(dwarf::DW_LLE_startx_length); 2934 unsigned idx = AddrPool.getIndex(Entry.Begin); 2935 Asm->emitULEB128(idx); 2936 // Also the pre-standard encoding is slightly different, emitting this as 2937 // an address-length entry here, but its a ULEB128 in DWARFv5 loclists. 2938 Asm->emitLabelDifference(Entry.End, Entry.Begin, 4); 2939 emitDebugLocEntryLocation(Entry, List.CU); 2940 } 2941 Asm->emitInt8(dwarf::DW_LLE_end_of_list); 2942 } 2943 } 2944 2945 struct ArangeSpan { 2946 const MCSymbol *Start, *End; 2947 }; 2948 2949 // Emit a debug aranges section, containing a CU lookup for any 2950 // address we can tie back to a CU. 2951 void DwarfDebug::emitDebugARanges() { 2952 // Provides a unique id per text section. 2953 MapVector<MCSection *, SmallVector<SymbolCU, 8>> SectionMap; 2954 2955 // Filter labels by section. 2956 for (const SymbolCU &SCU : ArangeLabels) { 2957 if (SCU.Sym->isInSection()) { 2958 // Make a note of this symbol and it's section. 2959 MCSection *Section = &SCU.Sym->getSection(); 2960 if (!Section->getKind().isMetadata()) 2961 SectionMap[Section].push_back(SCU); 2962 } else { 2963 // Some symbols (e.g. common/bss on mach-o) can have no section but still 2964 // appear in the output. This sucks as we rely on sections to build 2965 // arange spans. We can do it without, but it's icky. 2966 SectionMap[nullptr].push_back(SCU); 2967 } 2968 } 2969 2970 DenseMap<DwarfCompileUnit *, std::vector<ArangeSpan>> Spans; 2971 2972 for (auto &I : SectionMap) { 2973 MCSection *Section = I.first; 2974 SmallVector<SymbolCU, 8> &List = I.second; 2975 if (List.size() < 1) 2976 continue; 2977 2978 // If we have no section (e.g. common), just write out 2979 // individual spans for each symbol. 2980 if (!Section) { 2981 for (const SymbolCU &Cur : List) { 2982 ArangeSpan Span; 2983 Span.Start = Cur.Sym; 2984 Span.End = nullptr; 2985 assert(Cur.CU); 2986 Spans[Cur.CU].push_back(Span); 2987 } 2988 continue; 2989 } 2990 2991 // Sort the symbols by offset within the section. 2992 llvm::stable_sort(List, [&](const SymbolCU &A, const SymbolCU &B) { 2993 unsigned IA = A.Sym ? Asm->OutStreamer->GetSymbolOrder(A.Sym) : 0; 2994 unsigned IB = B.Sym ? Asm->OutStreamer->GetSymbolOrder(B.Sym) : 0; 2995 2996 // Symbols with no order assigned should be placed at the end. 2997 // (e.g. section end labels) 2998 if (IA == 0) 2999 return false; 3000 if (IB == 0) 3001 return true; 3002 return IA < IB; 3003 }); 3004 3005 // Insert a final terminator. 3006 List.push_back(SymbolCU(nullptr, Asm->OutStreamer->endSection(Section))); 3007 3008 // Build spans between each label. 3009 const MCSymbol *StartSym = List[0].Sym; 3010 for (size_t n = 1, e = List.size(); n < e; n++) { 3011 const SymbolCU &Prev = List[n - 1]; 3012 const SymbolCU &Cur = List[n]; 3013 3014 // Try and build the longest span we can within the same CU. 3015 if (Cur.CU != Prev.CU) { 3016 ArangeSpan Span; 3017 Span.Start = StartSym; 3018 Span.End = Cur.Sym; 3019 assert(Prev.CU); 3020 Spans[Prev.CU].push_back(Span); 3021 StartSym = Cur.Sym; 3022 } 3023 } 3024 } 3025 3026 // Start the dwarf aranges section. 3027 Asm->OutStreamer->SwitchSection( 3028 Asm->getObjFileLowering().getDwarfARangesSection()); 3029 3030 unsigned PtrSize = Asm->MAI->getCodePointerSize(); 3031 3032 // Build a list of CUs used. 3033 std::vector<DwarfCompileUnit *> CUs; 3034 for (const auto &it : Spans) { 3035 DwarfCompileUnit *CU = it.first; 3036 CUs.push_back(CU); 3037 } 3038 3039 // Sort the CU list (again, to ensure consistent output order). 3040 llvm::sort(CUs, [](const DwarfCompileUnit *A, const DwarfCompileUnit *B) { 3041 return A->getUniqueID() < B->getUniqueID(); 3042 }); 3043 3044 // Emit an arange table for each CU we used. 3045 for (DwarfCompileUnit *CU : CUs) { 3046 std::vector<ArangeSpan> &List = Spans[CU]; 3047 3048 // Describe the skeleton CU's offset and length, not the dwo file's. 3049 if (auto *Skel = CU->getSkeleton()) 3050 CU = Skel; 3051 3052 // Emit size of content not including length itself. 3053 unsigned ContentSize = 3054 sizeof(int16_t) + // DWARF ARange version number 3055 Asm->getDwarfOffsetByteSize() + // Offset of CU in the .debug_info 3056 // section 3057 sizeof(int8_t) + // Pointer Size (in bytes) 3058 sizeof(int8_t); // Segment Size (in bytes) 3059 3060 unsigned TupleSize = PtrSize * 2; 3061 3062 // 7.20 in the Dwarf specs requires the table to be aligned to a tuple. 3063 unsigned Padding = offsetToAlignment( 3064 Asm->getUnitLengthFieldByteSize() + ContentSize, Align(TupleSize)); 3065 3066 ContentSize += Padding; 3067 ContentSize += (List.size() + 1) * TupleSize; 3068 3069 // For each compile unit, write the list of spans it covers. 3070 Asm->emitDwarfUnitLength(ContentSize, "Length of ARange Set"); 3071 Asm->OutStreamer->AddComment("DWARF Arange version number"); 3072 Asm->emitInt16(dwarf::DW_ARANGES_VERSION); 3073 Asm->OutStreamer->AddComment("Offset Into Debug Info Section"); 3074 emitSectionReference(*CU); 3075 Asm->OutStreamer->AddComment("Address Size (in bytes)"); 3076 Asm->emitInt8(PtrSize); 3077 Asm->OutStreamer->AddComment("Segment Size (in bytes)"); 3078 Asm->emitInt8(0); 3079 3080 Asm->OutStreamer->emitFill(Padding, 0xff); 3081 3082 for (const ArangeSpan &Span : List) { 3083 Asm->emitLabelReference(Span.Start, PtrSize); 3084 3085 // Calculate the size as being from the span start to it's end. 3086 if (Span.End) { 3087 Asm->emitLabelDifference(Span.End, Span.Start, PtrSize); 3088 } else { 3089 // For symbols without an end marker (e.g. common), we 3090 // write a single arange entry containing just that one symbol. 3091 uint64_t Size = SymSize[Span.Start]; 3092 if (Size == 0) 3093 Size = 1; 3094 3095 Asm->OutStreamer->emitIntValue(Size, PtrSize); 3096 } 3097 } 3098 3099 Asm->OutStreamer->AddComment("ARange terminator"); 3100 Asm->OutStreamer->emitIntValue(0, PtrSize); 3101 Asm->OutStreamer->emitIntValue(0, PtrSize); 3102 } 3103 } 3104 3105 /// Emit a single range list. We handle both DWARF v5 and earlier. 3106 static void emitRangeList(DwarfDebug &DD, AsmPrinter *Asm, 3107 const RangeSpanList &List) { 3108 emitRangeList(DD, Asm, List.Label, List.Ranges, *List.CU, 3109 dwarf::DW_RLE_base_addressx, dwarf::DW_RLE_offset_pair, 3110 dwarf::DW_RLE_startx_length, dwarf::DW_RLE_end_of_list, 3111 llvm::dwarf::RangeListEncodingString, 3112 List.CU->getCUNode()->getRangesBaseAddress() || 3113 DD.getDwarfVersion() >= 5, 3114 [](auto) {}); 3115 } 3116 3117 void DwarfDebug::emitDebugRangesImpl(const DwarfFile &Holder, MCSection *Section) { 3118 if (Holder.getRangeLists().empty()) 3119 return; 3120 3121 assert(useRangesSection()); 3122 assert(!CUMap.empty()); 3123 assert(llvm::any_of(CUMap, [](const decltype(CUMap)::value_type &Pair) { 3124 return !Pair.second->getCUNode()->isDebugDirectivesOnly(); 3125 })); 3126 3127 Asm->OutStreamer->SwitchSection(Section); 3128 3129 MCSymbol *TableEnd = nullptr; 3130 if (getDwarfVersion() >= 5) 3131 TableEnd = emitRnglistsTableHeader(Asm, Holder); 3132 3133 for (const RangeSpanList &List : Holder.getRangeLists()) 3134 emitRangeList(*this, Asm, List); 3135 3136 if (TableEnd) 3137 Asm->OutStreamer->emitLabel(TableEnd); 3138 } 3139 3140 /// Emit address ranges into the .debug_ranges section or into the DWARF v5 3141 /// .debug_rnglists section. 3142 void DwarfDebug::emitDebugRanges() { 3143 const auto &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 3144 3145 emitDebugRangesImpl(Holder, 3146 getDwarfVersion() >= 5 3147 ? Asm->getObjFileLowering().getDwarfRnglistsSection() 3148 : Asm->getObjFileLowering().getDwarfRangesSection()); 3149 } 3150 3151 void DwarfDebug::emitDebugRangesDWO() { 3152 emitDebugRangesImpl(InfoHolder, 3153 Asm->getObjFileLowering().getDwarfRnglistsDWOSection()); 3154 } 3155 3156 /// Emit the header of a DWARF 5 macro section, or the GNU extension for 3157 /// DWARF 4. 3158 static void emitMacroHeader(AsmPrinter *Asm, const DwarfDebug &DD, 3159 const DwarfCompileUnit &CU, uint16_t DwarfVersion) { 3160 enum HeaderFlagMask { 3161 #define HANDLE_MACRO_FLAG(ID, NAME) MACRO_FLAG_##NAME = ID, 3162 #include "llvm/BinaryFormat/Dwarf.def" 3163 }; 3164 Asm->OutStreamer->AddComment("Macro information version"); 3165 Asm->emitInt16(DwarfVersion >= 5 ? DwarfVersion : 4); 3166 // We emit the line offset flag unconditionally here, since line offset should 3167 // be mostly present. 3168 if (Asm->isDwarf64()) { 3169 Asm->OutStreamer->AddComment("Flags: 64 bit, debug_line_offset present"); 3170 Asm->emitInt8(MACRO_FLAG_OFFSET_SIZE | MACRO_FLAG_DEBUG_LINE_OFFSET); 3171 } else { 3172 Asm->OutStreamer->AddComment("Flags: 32 bit, debug_line_offset present"); 3173 Asm->emitInt8(MACRO_FLAG_DEBUG_LINE_OFFSET); 3174 } 3175 Asm->OutStreamer->AddComment("debug_line_offset"); 3176 if (DD.useSplitDwarf()) 3177 Asm->emitDwarfLengthOrOffset(0); 3178 else 3179 Asm->emitDwarfSymbolReference(CU.getLineTableStartSym()); 3180 } 3181 3182 void DwarfDebug::handleMacroNodes(DIMacroNodeArray Nodes, DwarfCompileUnit &U) { 3183 for (auto *MN : Nodes) { 3184 if (auto *M = dyn_cast<DIMacro>(MN)) 3185 emitMacro(*M); 3186 else if (auto *F = dyn_cast<DIMacroFile>(MN)) 3187 emitMacroFile(*F, U); 3188 else 3189 llvm_unreachable("Unexpected DI type!"); 3190 } 3191 } 3192 3193 void DwarfDebug::emitMacro(DIMacro &M) { 3194 StringRef Name = M.getName(); 3195 StringRef Value = M.getValue(); 3196 3197 // There should be one space between the macro name and the macro value in 3198 // define entries. In undef entries, only the macro name is emitted. 3199 std::string Str = Value.empty() ? Name.str() : (Name + " " + Value).str(); 3200 3201 if (UseDebugMacroSection) { 3202 if (getDwarfVersion() >= 5) { 3203 unsigned Type = M.getMacinfoType() == dwarf::DW_MACINFO_define 3204 ? dwarf::DW_MACRO_define_strx 3205 : dwarf::DW_MACRO_undef_strx; 3206 Asm->OutStreamer->AddComment(dwarf::MacroString(Type)); 3207 Asm->emitULEB128(Type); 3208 Asm->OutStreamer->AddComment("Line Number"); 3209 Asm->emitULEB128(M.getLine()); 3210 Asm->OutStreamer->AddComment("Macro String"); 3211 Asm->emitULEB128( 3212 InfoHolder.getStringPool().getIndexedEntry(*Asm, Str).getIndex()); 3213 } else { 3214 unsigned Type = M.getMacinfoType() == dwarf::DW_MACINFO_define 3215 ? dwarf::DW_MACRO_GNU_define_indirect 3216 : dwarf::DW_MACRO_GNU_undef_indirect; 3217 Asm->OutStreamer->AddComment(dwarf::GnuMacroString(Type)); 3218 Asm->emitULEB128(Type); 3219 Asm->OutStreamer->AddComment("Line Number"); 3220 Asm->emitULEB128(M.getLine()); 3221 Asm->OutStreamer->AddComment("Macro String"); 3222 Asm->emitDwarfSymbolReference( 3223 InfoHolder.getStringPool().getEntry(*Asm, Str).getSymbol()); 3224 } 3225 } else { 3226 Asm->OutStreamer->AddComment(dwarf::MacinfoString(M.getMacinfoType())); 3227 Asm->emitULEB128(M.getMacinfoType()); 3228 Asm->OutStreamer->AddComment("Line Number"); 3229 Asm->emitULEB128(M.getLine()); 3230 Asm->OutStreamer->AddComment("Macro String"); 3231 Asm->OutStreamer->emitBytes(Str); 3232 Asm->emitInt8('\0'); 3233 } 3234 } 3235 3236 void DwarfDebug::emitMacroFileImpl( 3237 DIMacroFile &MF, DwarfCompileUnit &U, unsigned StartFile, unsigned EndFile, 3238 StringRef (*MacroFormToString)(unsigned Form)) { 3239 3240 Asm->OutStreamer->AddComment(MacroFormToString(StartFile)); 3241 Asm->emitULEB128(StartFile); 3242 Asm->OutStreamer->AddComment("Line Number"); 3243 Asm->emitULEB128(MF.getLine()); 3244 Asm->OutStreamer->AddComment("File Number"); 3245 DIFile &F = *MF.getFile(); 3246 if (useSplitDwarf()) 3247 Asm->emitULEB128(getDwoLineTable(U)->getFile( 3248 F.getDirectory(), F.getFilename(), getMD5AsBytes(&F), 3249 Asm->OutContext.getDwarfVersion(), F.getSource())); 3250 else 3251 Asm->emitULEB128(U.getOrCreateSourceID(&F)); 3252 handleMacroNodes(MF.getElements(), U); 3253 Asm->OutStreamer->AddComment(MacroFormToString(EndFile)); 3254 Asm->emitULEB128(EndFile); 3255 } 3256 3257 void DwarfDebug::emitMacroFile(DIMacroFile &F, DwarfCompileUnit &U) { 3258 // DWARFv5 macro and DWARFv4 macinfo share some common encodings, 3259 // so for readibility/uniformity, We are explicitly emitting those. 3260 assert(F.getMacinfoType() == dwarf::DW_MACINFO_start_file); 3261 if (UseDebugMacroSection) 3262 emitMacroFileImpl( 3263 F, U, dwarf::DW_MACRO_start_file, dwarf::DW_MACRO_end_file, 3264 (getDwarfVersion() >= 5) ? dwarf::MacroString : dwarf::GnuMacroString); 3265 else 3266 emitMacroFileImpl(F, U, dwarf::DW_MACINFO_start_file, 3267 dwarf::DW_MACINFO_end_file, dwarf::MacinfoString); 3268 } 3269 3270 void DwarfDebug::emitDebugMacinfoImpl(MCSection *Section) { 3271 for (const auto &P : CUMap) { 3272 auto &TheCU = *P.second; 3273 auto *SkCU = TheCU.getSkeleton(); 3274 DwarfCompileUnit &U = SkCU ? *SkCU : TheCU; 3275 auto *CUNode = cast<DICompileUnit>(P.first); 3276 DIMacroNodeArray Macros = CUNode->getMacros(); 3277 if (Macros.empty()) 3278 continue; 3279 Asm->OutStreamer->SwitchSection(Section); 3280 Asm->OutStreamer->emitLabel(U.getMacroLabelBegin()); 3281 if (UseDebugMacroSection) 3282 emitMacroHeader(Asm, *this, U, getDwarfVersion()); 3283 handleMacroNodes(Macros, U); 3284 Asm->OutStreamer->AddComment("End Of Macro List Mark"); 3285 Asm->emitInt8(0); 3286 } 3287 } 3288 3289 /// Emit macros into a debug macinfo/macro section. 3290 void DwarfDebug::emitDebugMacinfo() { 3291 auto &ObjLower = Asm->getObjFileLowering(); 3292 emitDebugMacinfoImpl(UseDebugMacroSection 3293 ? ObjLower.getDwarfMacroSection() 3294 : ObjLower.getDwarfMacinfoSection()); 3295 } 3296 3297 void DwarfDebug::emitDebugMacinfoDWO() { 3298 auto &ObjLower = Asm->getObjFileLowering(); 3299 emitDebugMacinfoImpl(UseDebugMacroSection 3300 ? ObjLower.getDwarfMacroDWOSection() 3301 : ObjLower.getDwarfMacinfoDWOSection()); 3302 } 3303 3304 // DWARF5 Experimental Separate Dwarf emitters. 3305 3306 void DwarfDebug::initSkeletonUnit(const DwarfUnit &U, DIE &Die, 3307 std::unique_ptr<DwarfCompileUnit> NewU) { 3308 3309 if (!CompilationDir.empty()) 3310 NewU->addString(Die, dwarf::DW_AT_comp_dir, CompilationDir); 3311 addGnuPubAttributes(*NewU, Die); 3312 3313 SkeletonHolder.addUnit(std::move(NewU)); 3314 } 3315 3316 DwarfCompileUnit &DwarfDebug::constructSkeletonCU(const DwarfCompileUnit &CU) { 3317 3318 auto OwnedUnit = std::make_unique<DwarfCompileUnit>( 3319 CU.getUniqueID(), CU.getCUNode(), Asm, this, &SkeletonHolder, 3320 UnitKind::Skeleton); 3321 DwarfCompileUnit &NewCU = *OwnedUnit; 3322 NewCU.setSection(Asm->getObjFileLowering().getDwarfInfoSection()); 3323 3324 NewCU.initStmtList(); 3325 3326 if (useSegmentedStringOffsetsTable()) 3327 NewCU.addStringOffsetsStart(); 3328 3329 initSkeletonUnit(CU, NewCU.getUnitDie(), std::move(OwnedUnit)); 3330 3331 return NewCU; 3332 } 3333 3334 // Emit the .debug_info.dwo section for separated dwarf. This contains the 3335 // compile units that would normally be in debug_info. 3336 void DwarfDebug::emitDebugInfoDWO() { 3337 assert(useSplitDwarf() && "No split dwarf debug info?"); 3338 // Don't emit relocations into the dwo file. 3339 InfoHolder.emitUnits(/* UseOffsets */ true); 3340 } 3341 3342 // Emit the .debug_abbrev.dwo section for separated dwarf. This contains the 3343 // abbreviations for the .debug_info.dwo section. 3344 void DwarfDebug::emitDebugAbbrevDWO() { 3345 assert(useSplitDwarf() && "No split dwarf?"); 3346 InfoHolder.emitAbbrevs(Asm->getObjFileLowering().getDwarfAbbrevDWOSection()); 3347 } 3348 3349 void DwarfDebug::emitDebugLineDWO() { 3350 assert(useSplitDwarf() && "No split dwarf?"); 3351 SplitTypeUnitFileTable.Emit( 3352 *Asm->OutStreamer, MCDwarfLineTableParams(), 3353 Asm->getObjFileLowering().getDwarfLineDWOSection()); 3354 } 3355 3356 void DwarfDebug::emitStringOffsetsTableHeaderDWO() { 3357 assert(useSplitDwarf() && "No split dwarf?"); 3358 InfoHolder.getStringPool().emitStringOffsetsTableHeader( 3359 *Asm, Asm->getObjFileLowering().getDwarfStrOffDWOSection(), 3360 InfoHolder.getStringOffsetsStartSym()); 3361 } 3362 3363 // Emit the .debug_str.dwo section for separated dwarf. This contains the 3364 // string section and is identical in format to traditional .debug_str 3365 // sections. 3366 void DwarfDebug::emitDebugStrDWO() { 3367 if (useSegmentedStringOffsetsTable()) 3368 emitStringOffsetsTableHeaderDWO(); 3369 assert(useSplitDwarf() && "No split dwarf?"); 3370 MCSection *OffSec = Asm->getObjFileLowering().getDwarfStrOffDWOSection(); 3371 InfoHolder.emitStrings(Asm->getObjFileLowering().getDwarfStrDWOSection(), 3372 OffSec, /* UseRelativeOffsets = */ false); 3373 } 3374 3375 // Emit address pool. 3376 void DwarfDebug::emitDebugAddr() { 3377 AddrPool.emit(*Asm, Asm->getObjFileLowering().getDwarfAddrSection()); 3378 } 3379 3380 MCDwarfDwoLineTable *DwarfDebug::getDwoLineTable(const DwarfCompileUnit &CU) { 3381 if (!useSplitDwarf()) 3382 return nullptr; 3383 const DICompileUnit *DIUnit = CU.getCUNode(); 3384 SplitTypeUnitFileTable.maybeSetRootFile( 3385 DIUnit->getDirectory(), DIUnit->getFilename(), 3386 getMD5AsBytes(DIUnit->getFile()), DIUnit->getSource()); 3387 return &SplitTypeUnitFileTable; 3388 } 3389 3390 uint64_t DwarfDebug::makeTypeSignature(StringRef Identifier) { 3391 MD5 Hash; 3392 Hash.update(Identifier); 3393 // ... take the least significant 8 bytes and return those. Our MD5 3394 // implementation always returns its results in little endian, so we actually 3395 // need the "high" word. 3396 MD5::MD5Result Result; 3397 Hash.final(Result); 3398 return Result.high(); 3399 } 3400 3401 void DwarfDebug::addDwarfTypeUnitType(DwarfCompileUnit &CU, 3402 StringRef Identifier, DIE &RefDie, 3403 const DICompositeType *CTy) { 3404 // Fast path if we're building some type units and one has already used the 3405 // address pool we know we're going to throw away all this work anyway, so 3406 // don't bother building dependent types. 3407 if (!TypeUnitsUnderConstruction.empty() && AddrPool.hasBeenUsed()) 3408 return; 3409 3410 auto Ins = TypeSignatures.insert(std::make_pair(CTy, 0)); 3411 if (!Ins.second) { 3412 CU.addDIETypeSignature(RefDie, Ins.first->second); 3413 return; 3414 } 3415 3416 bool TopLevelType = TypeUnitsUnderConstruction.empty(); 3417 AddrPool.resetUsedFlag(); 3418 3419 auto OwnedUnit = std::make_unique<DwarfTypeUnit>(CU, Asm, this, &InfoHolder, 3420 getDwoLineTable(CU)); 3421 DwarfTypeUnit &NewTU = *OwnedUnit; 3422 DIE &UnitDie = NewTU.getUnitDie(); 3423 TypeUnitsUnderConstruction.emplace_back(std::move(OwnedUnit), CTy); 3424 3425 NewTU.addUInt(UnitDie, dwarf::DW_AT_language, dwarf::DW_FORM_data2, 3426 CU.getLanguage()); 3427 3428 uint64_t Signature = makeTypeSignature(Identifier); 3429 NewTU.setTypeSignature(Signature); 3430 Ins.first->second = Signature; 3431 3432 if (useSplitDwarf()) { 3433 MCSection *Section = 3434 getDwarfVersion() <= 4 3435 ? Asm->getObjFileLowering().getDwarfTypesDWOSection() 3436 : Asm->getObjFileLowering().getDwarfInfoDWOSection(); 3437 NewTU.setSection(Section); 3438 } else { 3439 MCSection *Section = 3440 getDwarfVersion() <= 4 3441 ? Asm->getObjFileLowering().getDwarfTypesSection(Signature) 3442 : Asm->getObjFileLowering().getDwarfInfoSection(Signature); 3443 NewTU.setSection(Section); 3444 // Non-split type units reuse the compile unit's line table. 3445 CU.applyStmtList(UnitDie); 3446 } 3447 3448 // Add DW_AT_str_offsets_base to the type unit DIE, but not for split type 3449 // units. 3450 if (useSegmentedStringOffsetsTable() && !useSplitDwarf()) 3451 NewTU.addStringOffsetsStart(); 3452 3453 NewTU.setType(NewTU.createTypeDIE(CTy)); 3454 3455 if (TopLevelType) { 3456 auto TypeUnitsToAdd = std::move(TypeUnitsUnderConstruction); 3457 TypeUnitsUnderConstruction.clear(); 3458 3459 // Types referencing entries in the address table cannot be placed in type 3460 // units. 3461 if (AddrPool.hasBeenUsed()) { 3462 3463 // Remove all the types built while building this type. 3464 // This is pessimistic as some of these types might not be dependent on 3465 // the type that used an address. 3466 for (const auto &TU : TypeUnitsToAdd) 3467 TypeSignatures.erase(TU.second); 3468 3469 // Construct this type in the CU directly. 3470 // This is inefficient because all the dependent types will be rebuilt 3471 // from scratch, including building them in type units, discovering that 3472 // they depend on addresses, throwing them out and rebuilding them. 3473 CU.constructTypeDIE(RefDie, cast<DICompositeType>(CTy)); 3474 return; 3475 } 3476 3477 // If the type wasn't dependent on fission addresses, finish adding the type 3478 // and all its dependent types. 3479 for (auto &TU : TypeUnitsToAdd) { 3480 InfoHolder.computeSizeAndOffsetsForUnit(TU.first.get()); 3481 InfoHolder.emitUnit(TU.first.get(), useSplitDwarf()); 3482 } 3483 } 3484 CU.addDIETypeSignature(RefDie, Signature); 3485 } 3486 3487 DwarfDebug::NonTypeUnitContext::NonTypeUnitContext(DwarfDebug *DD) 3488 : DD(DD), 3489 TypeUnitsUnderConstruction(std::move(DD->TypeUnitsUnderConstruction)), AddrPoolUsed(DD->AddrPool.hasBeenUsed()) { 3490 DD->TypeUnitsUnderConstruction.clear(); 3491 DD->AddrPool.resetUsedFlag(); 3492 } 3493 3494 DwarfDebug::NonTypeUnitContext::~NonTypeUnitContext() { 3495 DD->TypeUnitsUnderConstruction = std::move(TypeUnitsUnderConstruction); 3496 DD->AddrPool.resetUsedFlag(AddrPoolUsed); 3497 } 3498 3499 DwarfDebug::NonTypeUnitContext DwarfDebug::enterNonTypeUnitContext() { 3500 return NonTypeUnitContext(this); 3501 } 3502 3503 // Add the Name along with its companion DIE to the appropriate accelerator 3504 // table (for AccelTableKind::Dwarf it's always AccelDebugNames, for 3505 // AccelTableKind::Apple, we use the table we got as an argument). If 3506 // accelerator tables are disabled, this function does nothing. 3507 template <typename DataT> 3508 void DwarfDebug::addAccelNameImpl(const DICompileUnit &CU, 3509 AccelTable<DataT> &AppleAccel, StringRef Name, 3510 const DIE &Die) { 3511 if (getAccelTableKind() == AccelTableKind::None) 3512 return; 3513 3514 if (getAccelTableKind() != AccelTableKind::Apple && 3515 CU.getNameTableKind() != DICompileUnit::DebugNameTableKind::Default) 3516 return; 3517 3518 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder; 3519 DwarfStringPoolEntryRef Ref = Holder.getStringPool().getEntry(*Asm, Name); 3520 3521 switch (getAccelTableKind()) { 3522 case AccelTableKind::Apple: 3523 AppleAccel.addName(Ref, Die); 3524 break; 3525 case AccelTableKind::Dwarf: 3526 AccelDebugNames.addName(Ref, Die); 3527 break; 3528 case AccelTableKind::Default: 3529 llvm_unreachable("Default should have already been resolved."); 3530 case AccelTableKind::None: 3531 llvm_unreachable("None handled above"); 3532 } 3533 } 3534 3535 void DwarfDebug::addAccelName(const DICompileUnit &CU, StringRef Name, 3536 const DIE &Die) { 3537 addAccelNameImpl(CU, AccelNames, Name, Die); 3538 } 3539 3540 void DwarfDebug::addAccelObjC(const DICompileUnit &CU, StringRef Name, 3541 const DIE &Die) { 3542 // ObjC names go only into the Apple accelerator tables. 3543 if (getAccelTableKind() == AccelTableKind::Apple) 3544 addAccelNameImpl(CU, AccelObjC, Name, Die); 3545 } 3546 3547 void DwarfDebug::addAccelNamespace(const DICompileUnit &CU, StringRef Name, 3548 const DIE &Die) { 3549 addAccelNameImpl(CU, AccelNamespace, Name, Die); 3550 } 3551 3552 void DwarfDebug::addAccelType(const DICompileUnit &CU, StringRef Name, 3553 const DIE &Die, char Flags) { 3554 addAccelNameImpl(CU, AccelTypes, Name, Die); 3555 } 3556 3557 uint16_t DwarfDebug::getDwarfVersion() const { 3558 return Asm->OutStreamer->getContext().getDwarfVersion(); 3559 } 3560 3561 dwarf::Form DwarfDebug::getDwarfSectionOffsetForm() const { 3562 if (Asm->getDwarfVersion() >= 4) 3563 return dwarf::Form::DW_FORM_sec_offset; 3564 assert((!Asm->isDwarf64() || (Asm->getDwarfVersion() == 3)) && 3565 "DWARF64 is not defined prior DWARFv3"); 3566 return Asm->isDwarf64() ? dwarf::Form::DW_FORM_data8 3567 : dwarf::Form::DW_FORM_data4; 3568 } 3569 3570 const MCSymbol *DwarfDebug::getSectionLabel(const MCSection *S) { 3571 auto I = SectionLabels.find(S); 3572 if (I == SectionLabels.end()) 3573 return nullptr; 3574 return I->second; 3575 } 3576 void DwarfDebug::insertSectionLabel(const MCSymbol *S) { 3577 if (SectionLabels.insert(std::make_pair(&S->getSection(), S)).second) 3578 if (useSplitDwarf() || getDwarfVersion() >= 5) 3579 AddrPool.getIndex(S); 3580 } 3581 3582 Optional<MD5::MD5Result> DwarfDebug::getMD5AsBytes(const DIFile *File) const { 3583 assert(File); 3584 if (getDwarfVersion() < 5) 3585 return None; 3586 Optional<DIFile::ChecksumInfo<StringRef>> Checksum = File->getChecksum(); 3587 if (!Checksum || Checksum->Kind != DIFile::CSK_MD5) 3588 return None; 3589 3590 // Convert the string checksum to an MD5Result for the streamer. 3591 // The verifier validates the checksum so we assume it's okay. 3592 // An MD5 checksum is 16 bytes. 3593 std::string ChecksumString = fromHex(Checksum->Value); 3594 MD5::MD5Result CKMem; 3595 std::copy(ChecksumString.begin(), ChecksumString.end(), CKMem.Bytes.data()); 3596 return CKMem; 3597 } 3598