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/StringExtras.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/DWARFDataExtractor.h"
35#include "llvm/DebugInfo/DWARF/LowLevel/DWARFExpression.h"
36#include "llvm/IR/Constants.h"
37#include "llvm/IR/DebugInfoMetadata.h"
38#include "llvm/IR/Function.h"
39#include "llvm/IR/GlobalVariable.h"
40#include "llvm/IR/Module.h"
41#include "llvm/MC/MCAsmInfo.h"
42#include "llvm/MC/MCContext.h"
43#include "llvm/MC/MCSection.h"
44#include "llvm/MC/MCStreamer.h"
45#include "llvm/MC/MCSymbol.h"
46#include "llvm/MC/MCTargetOptions.h"
47#include "llvm/MC/MachineLocation.h"
48#include "llvm/Support/Casting.h"
49#include "llvm/Support/CommandLine.h"
50#include "llvm/Support/Debug.h"
51#include "llvm/Support/ErrorHandling.h"
52#include "llvm/Support/MD5.h"
53#include "llvm/Support/MathExtras.h"
54#include "llvm/Support/raw_ostream.h"
55#include "llvm/Target/TargetLoweringObjectFile.h"
56#include "llvm/Target/TargetMachine.h"
57#include "llvm/TargetParser/Triple.h"
58#include <cstddef>
59#include <iterator>
60#include <optional>
61#include <string>
62
63using namespace llvm;
64
65#define DEBUG_TYPE "dwarfdebug"
66
67STATISTIC(NumCSParams, "Number of dbg call site params created");
68
69static cl::opt<bool> UseDwarfRangesBaseAddressSpecifier(
70 "use-dwarf-ranges-base-address-specifier", cl::Hidden,
71 cl::desc("Use base address specifiers in debug_ranges"), cl::init(Val: false));
72
73static cl::opt<bool> GenerateARangeSection("generate-arange-section",
74 cl::Hidden,
75 cl::desc("Generate dwarf aranges"),
76 cl::init(Val: false));
77
78static cl::opt<bool>
79 GenerateDwarfTypeUnits("generate-type-units", cl::Hidden,
80 cl::desc("Generate DWARF4 type units."),
81 cl::init(Val: false));
82
83static cl::opt<bool> SplitDwarfCrossCuReferences(
84 "split-dwarf-cross-cu-references", cl::Hidden,
85 cl::desc("Enable cross-cu references in DWO files"), cl::init(Val: false));
86
87enum DefaultOnOff { Default, Enable, Disable };
88
89static cl::opt<DefaultOnOff> UnknownLocations(
90 "use-unknown-locations", cl::Hidden,
91 cl::desc("Make an absence of debug location information explicit."),
92 cl::values(clEnumVal(Default, "At top of block or after label"),
93 clEnumVal(Enable, "In all cases"), clEnumVal(Disable, "Never")),
94 cl::init(Val: Default));
95
96static cl::opt<AccelTableKind> AccelTables(
97 "accel-tables", cl::Hidden, cl::desc("Output dwarf accelerator tables."),
98 cl::values(clEnumValN(AccelTableKind::Default, "Default",
99 "Default for platform"),
100 clEnumValN(AccelTableKind::None, "Disable", "Disabled."),
101 clEnumValN(AccelTableKind::Apple, "Apple", "Apple"),
102 clEnumValN(AccelTableKind::Dwarf, "Dwarf", "DWARF")),
103 cl::init(Val: AccelTableKind::Default));
104
105static cl::opt<DefaultOnOff>
106DwarfInlinedStrings("dwarf-inlined-strings", cl::Hidden,
107 cl::desc("Use inlined strings rather than string section."),
108 cl::values(clEnumVal(Default, "Default for platform"),
109 clEnumVal(Enable, "Enabled"),
110 clEnumVal(Disable, "Disabled")),
111 cl::init(Val: Default));
112
113static cl::opt<bool>
114 NoDwarfRangesSection("no-dwarf-ranges-section", cl::Hidden,
115 cl::desc("Disable emission .debug_ranges section."),
116 cl::init(Val: false));
117
118static cl::opt<DefaultOnOff> DwarfSectionsAsReferences(
119 "dwarf-sections-as-references", cl::Hidden,
120 cl::desc("Use sections+offset as references rather than labels."),
121 cl::values(clEnumVal(Default, "Default for platform"),
122 clEnumVal(Enable, "Enabled"), clEnumVal(Disable, "Disabled")),
123 cl::init(Val: Default));
124
125static cl::opt<bool>
126 UseGNUDebugMacro("use-gnu-debug-macro", cl::Hidden,
127 cl::desc("Emit the GNU .debug_macro format with DWARF <5"),
128 cl::init(Val: false));
129
130static cl::opt<DefaultOnOff> DwarfOpConvert(
131 "dwarf-op-convert", cl::Hidden,
132 cl::desc("Enable use of the DWARFv5 DW_OP_convert operator"),
133 cl::values(clEnumVal(Default, "Default for platform"),
134 clEnumVal(Enable, "Enabled"), clEnumVal(Disable, "Disabled")),
135 cl::init(Val: Default));
136
137enum LinkageNameOption {
138 DefaultLinkageNames,
139 AllLinkageNames,
140 AbstractLinkageNames
141};
142
143static cl::opt<LinkageNameOption>
144 DwarfLinkageNames("dwarf-linkage-names", cl::Hidden,
145 cl::desc("Which DWARF linkage-name attributes to emit."),
146 cl::values(clEnumValN(DefaultLinkageNames, "Default",
147 "Default for platform"),
148 clEnumValN(AllLinkageNames, "All", "All"),
149 clEnumValN(AbstractLinkageNames, "Abstract",
150 "Abstract subprograms")),
151 cl::init(Val: DefaultLinkageNames));
152
153static cl::opt<DwarfDebug::MinimizeAddrInV5> MinimizeAddrInV5Option(
154 "minimize-addr-in-v5", cl::Hidden,
155 cl::desc("Always use DW_AT_ranges in DWARFv5 whenever it could allow more "
156 "address pool entry sharing to reduce relocations/object size"),
157 cl::values(clEnumValN(DwarfDebug::MinimizeAddrInV5::Default, "Default",
158 "Default address minimization strategy"),
159 clEnumValN(DwarfDebug::MinimizeAddrInV5::Ranges, "Ranges",
160 "Use rnglists for contiguous ranges if that allows "
161 "using a pre-existing base address"),
162 clEnumValN(DwarfDebug::MinimizeAddrInV5::Expressions,
163 "Expressions",
164 "Use exprloc addrx+offset expressions for any "
165 "address with a prior base address"),
166 clEnumValN(DwarfDebug::MinimizeAddrInV5::Form, "Form",
167 "Use addrx+offset extension form for any address "
168 "with a prior base address"),
169 clEnumValN(DwarfDebug::MinimizeAddrInV5::Disabled, "Disabled",
170 "Stuff")),
171 cl::init(Val: DwarfDebug::MinimizeAddrInV5::Default));
172
173/// Set to false to ignore Key Instructions metadata.
174static cl::opt<bool> KeyInstructionsAreStmts(
175 "dwarf-use-key-instructions", cl::Hidden, cl::init(Val: true),
176 cl::desc("Set to false to ignore Key Instructions metadata"));
177
178static constexpr unsigned ULEB128PadSize = 4;
179
180void DebugLocDwarfExpression::emitOp(uint8_t Op, const char *Comment) {
181 getActiveStreamer().emitInt8(
182 Byte: Op, Comment: Comment ? Twine(Comment) + " " + dwarf::OperationEncodingString(Encoding: Op)
183 : dwarf::OperationEncodingString(Encoding: Op));
184}
185
186void DebugLocDwarfExpression::emitSigned(int64_t Value) {
187 getActiveStreamer().emitSLEB128(DWord: Value, Comment: Twine(Value));
188}
189
190void DebugLocDwarfExpression::emitUnsigned(uint64_t Value) {
191 getActiveStreamer().emitULEB128(DWord: Value, Comment: Twine(Value));
192}
193
194void DebugLocDwarfExpression::emitData1(uint8_t Value) {
195 getActiveStreamer().emitInt8(Byte: Value, Comment: Twine(Value));
196}
197
198void DebugLocDwarfExpression::emitBaseTypeRef(uint64_t Idx) {
199 assert(Idx < (1ULL << (ULEB128PadSize * 7)) && "Idx wont fit");
200 getActiveStreamer().emitULEB128(DWord: Idx, Comment: Twine(Idx), PadTo: ULEB128PadSize);
201}
202
203bool DebugLocDwarfExpression::isFrameRegister(const TargetRegisterInfo &TRI,
204 llvm::Register MachineReg) {
205 // This information is not available while emitting .debug_loc entries.
206 return false;
207}
208
209void DebugLocDwarfExpression::enableTemporaryBuffer() {
210 assert(!IsBuffering && "Already buffering?");
211 if (!TmpBuf)
212 TmpBuf = std::make_unique<TempBuffer>(args: OutBS.GenerateComments);
213 IsBuffering = true;
214}
215
216void DebugLocDwarfExpression::disableTemporaryBuffer() { IsBuffering = false; }
217
218unsigned DebugLocDwarfExpression::getTemporaryBufferSize() {
219 return TmpBuf ? TmpBuf->Bytes.size() : 0;
220}
221
222void DebugLocDwarfExpression::commitTemporaryBuffer() {
223 if (!TmpBuf)
224 return;
225 for (auto Byte : enumerate(First&: TmpBuf->Bytes)) {
226 const char *Comment = (Byte.index() < TmpBuf->Comments.size())
227 ? TmpBuf->Comments[Byte.index()].c_str()
228 : "";
229 OutBS.emitInt8(Byte: Byte.value(), Comment);
230 }
231 TmpBuf->Bytes.clear();
232 TmpBuf->Comments.clear();
233}
234
235const DIType *DbgVariable::getType() const {
236 return getVariable()->getType();
237}
238
239/// Get .debug_loc entry for the instruction range starting at MI.
240static DbgValueLoc getDebugLocValue(const MachineInstr *MI) {
241 const DIExpression *Expr = MI->getDebugExpression();
242 auto SingleLocExprOpt = DIExpression::convertToNonVariadicExpression(Expr);
243 const bool IsVariadic = !SingleLocExprOpt;
244 // If we have a variadic debug value instruction that is equivalent to a
245 // non-variadic instruction, then convert it to non-variadic form here.
246 if (!IsVariadic && !MI->isNonListDebugValue()) {
247 assert(MI->getNumDebugOperands() == 1 &&
248 "Mismatched DIExpression and debug operands for debug instruction.");
249 Expr = *SingleLocExprOpt;
250 }
251 assert(MI->getNumOperands() >= 3);
252 SmallVector<DbgValueLocEntry, 4> DbgValueLocEntries;
253 for (const MachineOperand &Op : MI->debug_operands()) {
254 if (Op.isReg()) {
255 MachineLocation MLoc(Op.getReg(),
256 MI->isNonListDebugValue() && MI->isDebugOffsetImm());
257 DbgValueLocEntries.push_back(Elt: DbgValueLocEntry(MLoc));
258 } else if (Op.isTargetIndex()) {
259 DbgValueLocEntries.push_back(
260 Elt: DbgValueLocEntry(TargetIndexLocation(Op.getIndex(), Op.getOffset())));
261 } else if (Op.isImm())
262 DbgValueLocEntries.push_back(Elt: DbgValueLocEntry(Op.getImm()));
263 else if (Op.isFPImm())
264 DbgValueLocEntries.push_back(Elt: DbgValueLocEntry(Op.getFPImm()));
265 else if (Op.isCImm())
266 DbgValueLocEntries.push_back(Elt: DbgValueLocEntry(Op.getCImm()));
267 else
268 llvm_unreachable("Unexpected debug operand in DBG_VALUE* instruction!");
269 }
270 return DbgValueLoc(Expr, DbgValueLocEntries, IsVariadic);
271}
272
273static uint64_t getFragmentOffsetInBits(const DIExpression &Expr) {
274 std::optional<DIExpression::FragmentInfo> Fragment = Expr.getFragmentInfo();
275 return Fragment ? Fragment->OffsetInBits : 0;
276}
277
278bool llvm::operator<(const FrameIndexExpr &LHS, const FrameIndexExpr &RHS) {
279 return getFragmentOffsetInBits(Expr: *LHS.Expr) <
280 getFragmentOffsetInBits(Expr: *RHS.Expr);
281}
282
283bool llvm::operator<(const EntryValueInfo &LHS, const EntryValueInfo &RHS) {
284 return getFragmentOffsetInBits(Expr: LHS.Expr) < getFragmentOffsetInBits(Expr: RHS.Expr);
285}
286
287Loc::Single::Single(DbgValueLoc ValueLoc)
288 : ValueLoc(std::make_unique<DbgValueLoc>(args&: ValueLoc)),
289 Expr(ValueLoc.getExpression()) {
290 if (!Expr->getNumElements())
291 Expr = nullptr;
292}
293
294Loc::Single::Single(const MachineInstr *DbgValue)
295 : Single(getDebugLocValue(MI: DbgValue)) {}
296
297const std::set<FrameIndexExpr> &Loc::MMI::getFrameIndexExprs() const {
298 return FrameIndexExprs;
299}
300
301void Loc::MMI::addFrameIndexExpr(const DIExpression *Expr, int FI) {
302 FrameIndexExprs.insert(x: {.FI: FI, .Expr: Expr});
303 assert((FrameIndexExprs.size() == 1 ||
304 llvm::all_of(FrameIndexExprs,
305 [](const FrameIndexExpr &FIE) {
306 return FIE.Expr && FIE.Expr->isFragment();
307 })) &&
308 "conflicting locations for variable");
309}
310
311static AccelTableKind computeAccelTableKind(unsigned DwarfVersion,
312 bool GenerateTypeUnits,
313 DebuggerKind Tuning,
314 const Triple &TT) {
315 // Honor an explicit request.
316 if (AccelTables != AccelTableKind::Default)
317 return AccelTables;
318
319 // Generating DWARF5 acceleration table.
320 // Currently Split dwarf and non ELF format is not supported.
321 if (GenerateTypeUnits && (DwarfVersion < 5 || !TT.isOSBinFormatELF()))
322 return AccelTableKind::None;
323
324 // Accelerator tables get emitted if targetting DWARF v5 or LLDB. DWARF v5
325 // always implies debug_names. For lower standard versions we use apple
326 // accelerator tables on apple platforms and debug_names elsewhere.
327 if (DwarfVersion >= 5)
328 return AccelTableKind::Dwarf;
329 if (Tuning == DebuggerKind::LLDB)
330 return TT.isOSBinFormatMachO() ? AccelTableKind::Apple
331 : AccelTableKind::Dwarf;
332 return AccelTableKind::None;
333}
334
335DwarfDebug::DwarfDebug(AsmPrinter *A)
336 : DebugHandlerBase(A), DebugLocs(A->OutStreamer->isVerboseAsm()),
337 SkeletonHolder(A, "skel_string", DIEValueAllocator),
338 IsDarwin(A->TM.getTargetTriple().isOSDarwin()),
339 InfoHolder(A, "info_string", DIEValueAllocator) {
340 const Triple &TT = Asm->TM.getTargetTriple();
341
342 // Make sure we know our "debugger tuning". The target option takes
343 // precedence; fall back to triple-based defaults.
344 if (Asm->TM.Options.DebuggerTuning != DebuggerKind::Default)
345 DebuggerTuning = Asm->TM.Options.DebuggerTuning;
346 else if (IsDarwin)
347 DebuggerTuning = DebuggerKind::LLDB;
348 else if (TT.isPS())
349 DebuggerTuning = DebuggerKind::SCE;
350 else if (TT.isOSAIX())
351 DebuggerTuning = DebuggerKind::DBX;
352 else
353 DebuggerTuning = DebuggerKind::GDB;
354
355 if (DwarfInlinedStrings == Default)
356 UseInlineStrings = tuneForDBX();
357 else
358 UseInlineStrings = DwarfInlinedStrings == Enable;
359
360 // Always emit .debug_aranges for SCE tuning.
361 UseARangesSection = GenerateARangeSection || tuneForSCE();
362
363 HasAppleExtensionAttributes = tuneForLLDB();
364
365 // Handle split DWARF.
366 HasSplitDwarf = !Asm->TM.Options.MCOptions.SplitDwarfFile.empty();
367
368 // SCE defaults to linkage names only for abstract subprograms.
369 if (DwarfLinkageNames == DefaultLinkageNames)
370 UseAllLinkageNames = !tuneForSCE();
371 else
372 UseAllLinkageNames = DwarfLinkageNames == AllLinkageNames;
373
374 unsigned DwarfVersionNumber = Asm->TM.Options.MCOptions.DwarfVersion;
375 unsigned DwarfVersion = DwarfVersionNumber ? DwarfVersionNumber
376 : MMI->getModule()->getDwarfVersion();
377 if (!DwarfVersion)
378 DwarfVersion = dwarf::DWARF_VERSION;
379
380 bool Dwarf64 = DwarfVersion >= 3 && // DWARF64 was introduced in DWARFv3.
381 TT.isArch64Bit(); // DWARF64 requires 64-bit relocations.
382
383 // Support DWARF64
384 // 1: For ELF when requested.
385 // 2: For XCOFF64: the AIX assembler will fill in debug section lengths
386 // according to the DWARF64 format for 64-bit assembly, so we must use
387 // DWARF64 in the compiler too for 64-bit mode.
388 Dwarf64 &=
389 ((Asm->TM.Options.MCOptions.Dwarf64 || MMI->getModule()->isDwarf64()) &&
390 TT.isOSBinFormatELF()) ||
391 TT.isOSBinFormatXCOFF();
392
393 if (!Dwarf64 && TT.isArch64Bit() && TT.isOSBinFormatXCOFF())
394 report_fatal_error(reason: "XCOFF requires DWARF64 for 64-bit mode!");
395
396 UseRangesSection = !NoDwarfRangesSection;
397
398 if (DwarfSectionsAsReferences != Default)
399 UseSectionsAsReferences = DwarfSectionsAsReferences == Enable;
400
401 // Don't generate type units for unsupported object file formats.
402 GenerateTypeUnits = (A->TM.getTargetTriple().isOSBinFormatELF() ||
403 A->TM.getTargetTriple().isOSBinFormatWasm()) &&
404 GenerateDwarfTypeUnits;
405
406 TheAccelTableKind = computeAccelTableKind(
407 DwarfVersion, GenerateTypeUnits, Tuning: DebuggerTuning, TT: A->TM.getTargetTriple());
408
409 // Work around a GDB bug. GDB doesn't support the standard opcode;
410 // SCE doesn't support GNU's; LLDB prefers the standard opcode, which
411 // is defined as of DWARF 3.
412 // See GDB bug 11616 - DW_OP_form_tls_address is unimplemented
413 // https://sourceware.org/bugzilla/show_bug.cgi?id=11616
414 UseGNUTLSOpcode = tuneForGDB() || DwarfVersion < 3;
415
416 UseDWARF2Bitfields = DwarfVersion < 4;
417
418 // The DWARF v5 string offsets table has - possibly shared - contributions
419 // from each compile and type unit each preceded by a header. The string
420 // offsets table used by the pre-DWARF v5 split-DWARF implementation uses
421 // a monolithic string offsets table without any header.
422 UseSegmentedStringOffsetsTable = DwarfVersion >= 5;
423
424 // Emit call-site-param debug info for GDB and LLDB, if the target supports
425 // the debug entry values feature. It can also be enabled explicitly.
426 EmitDebugEntryValues = Asm->TM.Options.ShouldEmitDebugEntryValues();
427
428 // It is unclear if the GCC .debug_macro extension is well-specified
429 // for split DWARF. For now, do not allow LLVM to emit it.
430 UseDebugMacroSection =
431 DwarfVersion >= 5 || (UseGNUDebugMacro && !useSplitDwarf());
432 if (DwarfOpConvert == Default)
433 EnableOpConvert = !((tuneForGDB() && useSplitDwarf()) || (tuneForLLDB() && !TT.isOSBinFormatMachO()));
434 else
435 EnableOpConvert = (DwarfOpConvert == Enable);
436
437 // Split DWARF would benefit object size significantly by trading reductions
438 // in address pool usage for slightly increased range list encodings.
439 if (DwarfVersion >= 5)
440 MinimizeAddr = MinimizeAddrInV5Option;
441
442 Asm->OutStreamer->getContext().setDwarfVersion(DwarfVersion);
443 Asm->OutStreamer->getContext().setDwarfFormat(Dwarf64 ? dwarf::DWARF64
444 : dwarf::DWARF32);
445}
446
447// Define out of line so we don't have to include DwarfUnit.h in DwarfDebug.h.
448DwarfDebug::~DwarfDebug() = default;
449
450static bool isObjCClass(StringRef Name) {
451 return Name.starts_with(Prefix: "+") || Name.starts_with(Prefix: "-");
452}
453
454static bool hasObjCCategory(StringRef Name) {
455 if (!isObjCClass(Name))
456 return false;
457
458 return Name.contains(Other: ") ");
459}
460
461static void getObjCClassCategory(StringRef In, StringRef &Class,
462 StringRef &Category) {
463 if (!hasObjCCategory(Name: In)) {
464 Class = In.slice(Start: In.find(C: '[') + 1, End: In.find(C: ' '));
465 Category = "";
466 return;
467 }
468
469 Class = In.slice(Start: In.find(C: '[') + 1, End: In.find(C: '('));
470 Category = In.slice(Start: In.find(C: '[') + 1, End: In.find(C: ' '));
471}
472
473static StringRef getObjCMethodName(StringRef In) {
474 return In.slice(Start: In.find(C: ' ') + 1, End: In.find(C: ']'));
475}
476
477// Add the various names to the Dwarf accelerator table names.
478void DwarfDebug::addSubprogramNames(
479 const DwarfUnit &Unit,
480 const DICompileUnit::DebugNameTableKind NameTableKind,
481 const DISubprogram *SP, DIE &Die) {
482 if (getAccelTableKind() != AccelTableKind::Apple &&
483 NameTableKind != DICompileUnit::DebugNameTableKind::Apple &&
484 NameTableKind == DICompileUnit::DebugNameTableKind::None)
485 return;
486
487 if (!SP->isDefinition())
488 return;
489
490 if (SP->getName() != "")
491 addAccelName(Unit, NameTableKind, Name: SP->getName(), Die);
492
493 // We drop the mangling escape prefix when emitting the DW_AT_linkage_name. So
494 // ensure we don't include it when inserting into the accelerator tables.
495 llvm::StringRef LinkageName =
496 GlobalValue::dropLLVMManglingEscape(Name: SP->getLinkageName());
497
498 // If the linkage name is different than the name, go ahead and output that as
499 // well into the name table. Only do that if we are going to actually emit
500 // that name.
501 if (LinkageName != "" && SP->getName() != LinkageName &&
502 (useAllLinkageNames() || InfoHolder.getAbstractScopeDIEs().lookup(Val: SP)))
503 addAccelName(Unit, NameTableKind, Name: LinkageName, Die);
504
505 // If this is an Objective-C selector name add it to the ObjC accelerator
506 // too.
507 if (isObjCClass(Name: SP->getName())) {
508 StringRef Class, Category;
509 getObjCClassCategory(In: SP->getName(), Class, Category);
510 addAccelObjC(Unit, NameTableKind, Name: Class, Die);
511 if (Category != "")
512 addAccelObjC(Unit, NameTableKind, Name: Category, Die);
513 // Also add the base method name to the name table.
514 addAccelName(Unit, NameTableKind, Name: getObjCMethodName(In: SP->getName()), Die);
515 }
516}
517
518/// Check whether we should create a DIE for the given Scope, return true
519/// if we don't create a DIE (the corresponding DIE is null).
520bool DwarfDebug::isLexicalScopeDIENull(LexicalScope *Scope) {
521 if (Scope->isAbstractScope())
522 return false;
523
524 // We don't create a DIE if there is no Range.
525 const SmallVectorImpl<InsnRange> &Ranges = Scope->getRanges();
526 if (Ranges.empty())
527 return true;
528
529 if (Ranges.size() > 1)
530 return false;
531
532 // We don't create a DIE if we have a single Range and the end label
533 // is null.
534 return !getLabelAfterInsn(MI: Ranges.front().second);
535}
536
537template <typename Func> static void forBothCUs(DwarfCompileUnit &CU, Func F) {
538 F(CU);
539 if (auto *SkelCU = CU.getSkeleton())
540 if (CU.getCUNode()->getSplitDebugInlining())
541 F(*SkelCU);
542}
543
544bool DwarfDebug::shareAcrossDWOCUs() const {
545 return SplitDwarfCrossCuReferences;
546}
547
548DwarfCompileUnit &
549DwarfDebug::getOrCreateAbstractSubprogramCU(const DISubprogram *SP,
550 DwarfCompileUnit &SrcCU) {
551 auto &CU = getOrCreateDwarfCompileUnit(DIUnit: SP->getUnit());
552 if (CU.getSkeleton())
553 return shareAcrossDWOCUs() ? CU : SrcCU;
554
555 return CU;
556}
557
558void DwarfDebug::constructAbstractSubprogramScopeDIE(DwarfCompileUnit &SrcCU,
559 LexicalScope *Scope) {
560 assert(Scope && Scope->getScopeNode());
561 assert(Scope->isAbstractScope());
562 assert(!Scope->getInlinedAt());
563
564 auto *SP = cast<DISubprogram>(Val: Scope->getScopeNode());
565
566 // Find the subprogram's DwarfCompileUnit in the SPMap in case the subprogram
567 // was inlined from another compile unit.
568 auto &CU = getOrCreateDwarfCompileUnit(DIUnit: SP->getUnit());
569 auto &TargetCU = getOrCreateAbstractSubprogramCU(SP, SrcCU);
570 TargetCU.constructAbstractSubprogramScopeDIE(Scope);
571 if (auto *SkelCU = CU.getSkeleton())
572 if (CU.getCUNode()->getSplitDebugInlining())
573 SkelCU->constructAbstractSubprogramScopeDIE(Scope);
574}
575
576/// Represents a parameter whose call site value can be described by applying a
577/// debug expression to a register in the forwarded register worklist.
578struct FwdRegParamInfo {
579 /// The described parameter register.
580 uint64_t ParamReg;
581
582 /// Debug expression that has been built up when walking through the
583 /// instruction chain that produces the parameter's value.
584 const DIExpression *Expr;
585};
586
587/// Register worklist for finding call site values.
588using FwdRegWorklist = MapVector<uint64_t, SmallVector<FwdRegParamInfo, 2>>;
589/// Container for the set of register units known to be clobbered on the path
590/// to a call site.
591using ClobberedRegUnitSet = SmallSet<MCRegUnit, 16>;
592
593/// Append the expression \p Addition to \p Original and return the result.
594static const DIExpression *combineDIExpressions(const DIExpression *Original,
595 const DIExpression *Addition) {
596 std::vector<uint64_t> Elts = Addition->getElements().vec();
597 // Avoid multiple DW_OP_stack_values.
598 if (Original->isImplicit() && Addition->isImplicit())
599 llvm::erase(C&: Elts, V: dwarf::DW_OP_stack_value);
600 const DIExpression *CombinedExpr =
601 (Elts.size() > 0) ? DIExpression::append(Expr: Original, Ops: Elts) : Original;
602 return CombinedExpr;
603}
604
605/// Emit call site parameter entries that are described by the given value and
606/// debug expression.
607template <typename ValT>
608static void finishCallSiteParams(ValT Val, const DIExpression *Expr,
609 ArrayRef<FwdRegParamInfo> DescribedParams,
610 ParamSet &Params) {
611 for (auto Param : DescribedParams) {
612 bool ShouldCombineExpressions = Expr && Param.Expr->getNumElements() > 0;
613
614 // If a parameter's call site value is produced by a chain of
615 // instructions we may have already created an expression for the
616 // parameter when walking through the instructions. Append that to the
617 // base expression.
618 const DIExpression *CombinedExpr =
619 ShouldCombineExpressions ? combineDIExpressions(Original: Expr, Addition: Param.Expr)
620 : Expr;
621 assert((!CombinedExpr || CombinedExpr->isValid()) &&
622 "Combined debug expression is invalid");
623
624 DbgValueLoc DbgLocVal(CombinedExpr, DbgValueLocEntry(Val));
625 DbgCallSiteParam CSParm(Param.ParamReg, DbgLocVal);
626 Params.push_back(Elt: CSParm);
627 ++NumCSParams;
628 }
629}
630
631/// Add \p Reg to the worklist, if it's not already present, and mark that the
632/// given parameter registers' values can (potentially) be described using
633/// that register and an debug expression.
634static void addToFwdRegWorklist(FwdRegWorklist &Worklist, unsigned Reg,
635 const DIExpression *Expr,
636 ArrayRef<FwdRegParamInfo> ParamsToAdd) {
637 auto &ParamsForFwdReg = Worklist[Reg];
638 for (auto Param : ParamsToAdd) {
639 assert(none_of(ParamsForFwdReg,
640 [Param](const FwdRegParamInfo &D) {
641 return D.ParamReg == Param.ParamReg;
642 }) &&
643 "Same parameter described twice by forwarding reg");
644
645 // If a parameter's call site value is produced by a chain of
646 // instructions we may have already created an expression for the
647 // parameter when walking through the instructions. Append that to the
648 // new expression.
649 const DIExpression *CombinedExpr = combineDIExpressions(Original: Expr, Addition: Param.Expr);
650 ParamsForFwdReg.push_back(Elt: {.ParamReg: Param.ParamReg, .Expr: CombinedExpr});
651 }
652}
653
654/// Interpret values loaded into registers by \p CurMI.
655static void interpretValues(const MachineInstr *CurMI,
656 FwdRegWorklist &ForwardedRegWorklist,
657 ParamSet &Params,
658 ClobberedRegUnitSet &ClobberedRegUnits) {
659
660 const MachineFunction *MF = CurMI->getMF();
661 const DIExpression *EmptyExpr =
662 DIExpression::get(Context&: MF->getFunction().getContext(), Elements: {});
663 const auto &TRI = *MF->getSubtarget().getRegisterInfo();
664 const auto &TII = *MF->getSubtarget().getInstrInfo();
665 const auto &TLI = *MF->getSubtarget().getTargetLowering();
666
667 // It's possible that we find a copy from a non-volatile register to the param
668 // register, which is clobbered in the meantime. Test for clobbered reg unit
669 // overlaps before completing.
670 auto IsRegClobberedInMeantime = [&](Register Reg) -> bool {
671 for (auto &RegUnit : ClobberedRegUnits)
672 if (TRI.hasRegUnit(Reg, RegUnit))
673 return true;
674 return false;
675 };
676
677 auto DescribeFwdRegsByCalleeSavedCopy = [&](const DestSourcePair &CopyInst) {
678 Register CopyDestReg = CopyInst.Destination->getReg();
679 Register CopySrcReg = CopyInst.Source->getReg();
680 if (IsRegClobberedInMeantime(CopyDestReg))
681 return;
682 // FIXME: This may be incorrect in cases where the caller and callee use
683 // different calling conventions.
684 if (!TRI.isCalleeSavedPhysReg(PhysReg: CopyDestReg, MF: *MF))
685 return;
686 // Describe any forward registers matching the source register. If the
687 // forward register is a sub-register of the source, we describe it using
688 // the corresponding sub-register in the destination, if such a
689 // sub-register exists. The end iterator in the MapVector is invalidated at
690 // erase(), so it needs to be evaluated at each iteration.
691 for (auto FwdRegIt = ForwardedRegWorklist.begin();
692 FwdRegIt != ForwardedRegWorklist.end();) {
693 Register CalleeSavedReg = MCRegister::NoRegister;
694 if (FwdRegIt->first == CopySrcReg)
695 CalleeSavedReg = CopyDestReg;
696 else if (unsigned SubRegIdx =
697 TRI.getSubRegIndex(RegNo: CopySrcReg, SubRegNo: FwdRegIt->first))
698 if (Register CopyDestSubReg = TRI.getSubReg(Reg: CopyDestReg, Idx: SubRegIdx))
699 CalleeSavedReg = CopyDestSubReg;
700
701 if (CalleeSavedReg == MCRegister::NoRegister) {
702 ++FwdRegIt;
703 continue;
704 }
705
706 MachineLocation MLoc(CalleeSavedReg, /*Indirect=*/false);
707 finishCallSiteParams(Val: MLoc, Expr: EmptyExpr, DescribedParams: FwdRegIt->second, Params);
708 FwdRegIt = ForwardedRegWorklist.erase(Iterator: FwdRegIt);
709 }
710 };
711
712 // Detect if this is a copy instruction. If this saves any of the forward
713 // registers in callee-saved registers, we can finalize those parameters
714 // directly.
715 // TODO: Can we do something similar for stack saves?
716 if (auto CopyInst = TII.isCopyInstr(MI: *CurMI))
717 DescribeFwdRegsByCalleeSavedCopy(*CopyInst);
718
719 // If an instruction defines more than one item in the worklist, we may run
720 // into situations where a worklist register's value is (potentially)
721 // described by the previous value of another register that is also defined
722 // by that instruction.
723 //
724 // This can for example occur in cases like this:
725 //
726 // $r1 = mov 123
727 // $r0, $r1 = mvrr $r1, 456
728 // call @foo, $r0, $r1
729 //
730 // When describing $r1's value for the mvrr instruction, we need to make sure
731 // that we don't finalize an entry value for $r0, as that is dependent on the
732 // previous value of $r1 (123 rather than 456).
733 //
734 // In order to not have to distinguish between those cases when finalizing
735 // entry values, we simply postpone adding new parameter registers to the
736 // worklist, by first keeping them in this temporary container until the
737 // instruction has been handled.
738 FwdRegWorklist TmpWorklistItems;
739
740 // If the MI is an instruction defining one or more parameters' forwarding
741 // registers, add those defines.
742 ClobberedRegUnitSet NewClobberedRegUnits;
743 auto getForwardingRegsDefinedByMI = [&](const MachineInstr &MI,
744 SmallSetVector<unsigned, 4> &Defs) {
745 if (MI.isDebugInstr())
746 return;
747
748 for (const MachineOperand &MO : MI.all_defs()) {
749 if (MO.getReg().isPhysical()) {
750 for (auto &FwdReg : ForwardedRegWorklist)
751 if (TRI.regsOverlap(RegA: FwdReg.first, RegB: MO.getReg()))
752 Defs.insert(X: FwdReg.first);
753 NewClobberedRegUnits.insert_range(R: TRI.regunits(Reg: MO.getReg()));
754 }
755 }
756 };
757
758 // Set of worklist registers that are defined by this instruction.
759 SmallSetVector<unsigned, 4> FwdRegDefs;
760
761 getForwardingRegsDefinedByMI(*CurMI, FwdRegDefs);
762 if (FwdRegDefs.empty()) {
763 // Any definitions by this instruction will clobber earlier reg movements.
764 ClobberedRegUnits.insert_range(R&: NewClobberedRegUnits);
765 return;
766 }
767
768 for (auto ParamFwdReg : FwdRegDefs) {
769 if (auto ParamValue = TII.describeLoadedValue(MI: *CurMI, Reg: ParamFwdReg)) {
770 if (ParamValue->first.isImm()) {
771 int64_t Val = ParamValue->first.getImm();
772 finishCallSiteParams(Val, Expr: ParamValue->second,
773 DescribedParams: ForwardedRegWorklist[ParamFwdReg], Params);
774 } else if (ParamValue->first.isReg()) {
775 Register RegLoc = ParamValue->first.getReg();
776 Register SP = TLI.getStackPointerRegisterToSaveRestore();
777 Register FP = TRI.getFrameRegister(MF: *MF);
778 bool IsSPorFP = (RegLoc == SP) || (RegLoc == FP);
779 // FIXME: This may be incorrect in cases where the caller and callee use
780 // different calling conventions.
781 if (!IsRegClobberedInMeantime(RegLoc) &&
782 (TRI.isCalleeSavedPhysReg(PhysReg: RegLoc, MF: *MF) || IsSPorFP)) {
783 MachineLocation MLoc(RegLoc, /*Indirect=*/IsSPorFP);
784 finishCallSiteParams(Val: MLoc, Expr: ParamValue->second,
785 DescribedParams: ForwardedRegWorklist[ParamFwdReg], Params);
786 } else {
787 // ParamFwdReg was described by the non-callee saved register
788 // RegLoc. Mark that the call site values for the parameters are
789 // dependent on that register instead of ParamFwdReg. Since RegLoc
790 // may be a register that will be handled in this iteration, we
791 // postpone adding the items to the worklist, and instead keep them
792 // in a temporary container.
793 addToFwdRegWorklist(Worklist&: TmpWorklistItems, Reg: RegLoc, Expr: ParamValue->second,
794 ParamsToAdd: ForwardedRegWorklist[ParamFwdReg]);
795 }
796 }
797 }
798 }
799
800 // Remove all registers that this instruction defines from the worklist.
801 for (auto ParamFwdReg : FwdRegDefs)
802 ForwardedRegWorklist.erase(Key: ParamFwdReg);
803
804 // Any definitions by this instruction will clobber earlier reg movements.
805 ClobberedRegUnits.insert_range(R&: NewClobberedRegUnits);
806
807 // Now that we are done handling this instruction, add items from the
808 // temporary worklist to the real one.
809 for (auto &New : TmpWorklistItems)
810 addToFwdRegWorklist(Worklist&: ForwardedRegWorklist, Reg: New.first, Expr: EmptyExpr, ParamsToAdd: New.second);
811 TmpWorklistItems.clear();
812}
813
814static bool interpretNextInstr(const MachineInstr *CurMI,
815 FwdRegWorklist &ForwardedRegWorklist,
816 ParamSet &Params,
817 ClobberedRegUnitSet &ClobberedRegUnits) {
818 // Skip bundle headers.
819 if (CurMI->isBundle())
820 return true;
821
822 // If the next instruction is a call we can not interpret parameter's
823 // forwarding registers or we finished the interpretation of all
824 // parameters.
825 if (CurMI->isCall())
826 return false;
827
828 if (ForwardedRegWorklist.empty())
829 return false;
830
831 // Avoid NOP description.
832 if (CurMI->getNumOperands() == 0)
833 return true;
834
835 interpretValues(CurMI, ForwardedRegWorklist, Params, ClobberedRegUnits);
836
837 return true;
838}
839
840/// Try to interpret values loaded into registers that forward parameters
841/// for \p CallMI. Store parameters with interpreted value into \p Params.
842static void collectCallSiteParameters(const MachineInstr *CallMI,
843 ParamSet &Params) {
844 const MachineFunction *MF = CallMI->getMF();
845 const auto &CalleesMap = MF->getCallSitesInfo();
846 auto CSInfo = CalleesMap.find(Val: CallMI);
847
848 // There is no information for the call instruction.
849 if (CSInfo == CalleesMap.end())
850 return;
851
852 const MachineBasicBlock *MBB = CallMI->getParent();
853
854 // Skip the call instruction.
855 auto I = std::next(x: CallMI->getReverseIterator());
856
857 FwdRegWorklist ForwardedRegWorklist;
858
859 const DIExpression *EmptyExpr =
860 DIExpression::get(Context&: MF->getFunction().getContext(), Elements: {});
861
862 // Add all the forwarding registers into the ForwardedRegWorklist.
863 for (const auto &ArgReg : CSInfo->second.ArgRegPairs) {
864 bool InsertedReg =
865 ForwardedRegWorklist.insert(KV: {ArgReg.Reg, {{.ParamReg: ArgReg.Reg, .Expr: EmptyExpr}}})
866 .second;
867 assert(InsertedReg && "Single register used to forward two arguments?");
868 (void)InsertedReg;
869 }
870
871 // Do not emit CSInfo for undef forwarding registers.
872 for (const auto &MO : CallMI->uses())
873 if (MO.isReg() && MO.isUndef())
874 ForwardedRegWorklist.erase(Key: MO.getReg());
875
876 // We erase, from the ForwardedRegWorklist, those forwarding registers for
877 // which we successfully describe a loaded value (by using
878 // the describeLoadedValue()). For those remaining arguments in the working
879 // list, for which we do not describe a loaded value by
880 // the describeLoadedValue(), we try to generate an entry value expression
881 // for their call site value description, if the call is within the entry MBB.
882 // TODO: Handle situations when call site parameter value can be described
883 // as the entry value within basic blocks other than the first one.
884 bool ShouldTryEmitEntryVals = MBB->getIterator() == MF->begin();
885
886 // Search for a loading value in forwarding registers inside call delay slot.
887 ClobberedRegUnitSet ClobberedRegUnits;
888 if (CallMI->hasDelaySlot()) {
889 auto Suc = std::next(x: CallMI->getIterator());
890 // Only one-instruction delay slot is supported.
891 auto BundleEnd = llvm::getBundleEnd(I: CallMI->getIterator());
892 (void)BundleEnd;
893 assert(std::next(Suc) == BundleEnd &&
894 "More than one instruction in call delay slot");
895 // Try to interpret value loaded by instruction.
896 if (!interpretNextInstr(CurMI: &*Suc, ForwardedRegWorklist, Params, ClobberedRegUnits))
897 return;
898 }
899
900 // Search for a loading value in forwarding registers.
901 for (; I != MBB->rend(); ++I) {
902 // Try to interpret values loaded by instruction.
903 if (!interpretNextInstr(CurMI: &*I, ForwardedRegWorklist, Params, ClobberedRegUnits))
904 return;
905 }
906
907 // Emit the call site parameter's value as an entry value.
908 if (ShouldTryEmitEntryVals) {
909 // Create an expression where the register's entry value is used.
910 DIExpression *EntryExpr = DIExpression::get(
911 Context&: MF->getFunction().getContext(), Elements: {dwarf::DW_OP_LLVM_entry_value, 1});
912 for (auto &RegEntry : ForwardedRegWorklist) {
913 MachineLocation MLoc(RegEntry.first);
914 finishCallSiteParams(Val: MLoc, Expr: EntryExpr, DescribedParams: RegEntry.second, Params);
915 }
916 }
917}
918
919void DwarfDebug::constructCallSiteEntryDIEs(const DISubprogram &SP,
920 DwarfCompileUnit &CU, DIE &ScopeDIE,
921 const MachineFunction &MF) {
922 // Add a call site-related attribute (DWARF5, Sec. 3.3.1.3). Do this only if
923 // the subprogram is required to have one.
924 if (!SP.areAllCallsDescribed() || !SP.isDefinition())
925 return;
926
927 // Use DW_AT_call_all_calls to express that call site entries are present
928 // for both tail and non-tail calls. Don't use DW_AT_call_all_source_calls
929 // because one of its requirements is not met: call site entries for
930 // optimized-out calls are elided.
931 CU.addFlag(Die&: ScopeDIE, Attribute: CU.getDwarf5OrGNUAttr(Attr: dwarf::DW_AT_call_all_calls));
932
933 const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo();
934 assert(TII && "TargetInstrInfo not found: cannot label tail calls");
935
936 // Delay slot support check.
937 auto delaySlotSupported = [&](const MachineInstr &MI) {
938 if (!MI.isBundledWithSucc())
939 return false;
940 auto Suc = std::next(x: MI.getIterator());
941 auto CallInstrBundle = getBundleStart(I: MI.getIterator());
942 (void)CallInstrBundle;
943 auto DelaySlotBundle = getBundleStart(I: Suc);
944 (void)DelaySlotBundle;
945 // Ensure that label after call is following delay slot instruction.
946 // Ex. CALL_INSTRUCTION {
947 // DELAY_SLOT_INSTRUCTION }
948 // LABEL_AFTER_CALL
949 assert(getLabelAfterInsn(&*CallInstrBundle) ==
950 getLabelAfterInsn(&*DelaySlotBundle) &&
951 "Call and its successor instruction don't have same label after.");
952 return true;
953 };
954
955 // Create call_target connections for indirect calls.
956 auto addCallSiteTargetForIndirectCalls = [&](const MachineInstr *MI,
957 DIE &CallSiteDIE) {
958 const MachineFunction *MF = MI->getMF();
959 const auto &CalleesMap = MF->getCallSitesInfo();
960 auto CSInfo = CalleesMap.find(Val: MI);
961 // Get the information for the call instruction.
962 if (CSInfo == CalleesMap.end() || !CSInfo->second.CallTarget)
963 return;
964
965 MDNode *CallTarget = CSInfo->second.CallTarget;
966 // Add DW_AT_LLVM_virtual_call_origin with the 'call_target' metadata.
967 assert(!CallSiteDIE.findAttribute(dwarf::DW_AT_LLVM_virtual_call_origin) &&
968 "DW_AT_LLVM_virtual_call_origin already exists");
969 const DISubprogram *CalleeSP = dyn_cast<DISubprogram>(Val: CallTarget);
970 DIE *CalleeDIE = CU.getOrCreateSubprogramDIE(SP: CalleeSP, F: nullptr);
971 assert(CalleeDIE && "Could not create DIE for call site entry origin");
972 CU.addDIEEntry(Die&: CallSiteDIE,
973 Attribute: CU.getDwarf5OrGNUAttr(Attr: dwarf::DW_AT_LLVM_virtual_call_origin),
974 Entry&: *CalleeDIE);
975 // Add DW_AT_linkage_name to the method declaration if needed.
976 CU.addLinkageNamesToDeclarations(DD: *this, CalleeSP: *CalleeSP, CalleeDIE&: *CalleeDIE);
977 };
978
979 // Emit call site entries for each call or tail call in the function.
980 for (const MachineBasicBlock &MBB : MF) {
981 for (const MachineInstr &MI : MBB.instrs()) {
982 // Bundles with call in them will pass the isCall() test below but do not
983 // have callee operand information so skip them here. Iterator will
984 // eventually reach the call MI.
985 if (MI.isBundle())
986 continue;
987
988 // Skip instructions which aren't calls. Both calls and tail-calling jump
989 // instructions (e.g TAILJMPd64) are classified correctly here.
990 if (!MI.isCandidateForAdditionalCallInfo())
991 continue;
992
993 // Skip instructions marked as frame setup, as they are not interesting to
994 // the user.
995 if (MI.getFlag(Flag: MachineInstr::FrameSetup))
996 continue;
997
998 // Check if delay slot support is enabled.
999 if (MI.hasDelaySlot() && !delaySlotSupported(*&MI))
1000 return;
1001
1002 DIType *AllocSiteTy = dyn_cast_or_null<DIType>(Val: MI.getHeapAllocMarker());
1003
1004 // If this is a direct call, find the callee's subprogram.
1005 // In the case of an indirect call find the register or memory location
1006 // that holds the callee address.
1007 const MachineOperand &CalleeOp = TII->getCalleeOperand(MI);
1008 bool PhysRegCalleeOperand =
1009 CalleeOp.isReg() && CalleeOp.getReg().isPhysical();
1010 MachineLocation CallTarget{0};
1011 int64_t Offset = 0;
1012 const DISubprogram *CalleeSP = nullptr;
1013 const Function *CalleeDecl = nullptr;
1014 if (PhysRegCalleeOperand) {
1015 bool Scalable = false;
1016 const MachineOperand *BaseOp = nullptr;
1017 const TargetRegisterInfo &TRI =
1018 *Asm->MF->getSubtarget().getRegisterInfo();
1019 if (TII->getMemOperandWithOffset(MI, BaseOp, Offset, OffsetIsScalable&: Scalable, TRI: &TRI)) {
1020 if (BaseOp && BaseOp->isReg() && !Scalable)
1021 CallTarget = MachineLocation(BaseOp->getReg(), /*Indirect*/ true);
1022 }
1023
1024 if (!CallTarget.isIndirect())
1025 CallTarget = MachineLocation(CalleeOp.getReg()); // Might be zero.
1026 } else if (CalleeOp.isGlobal()) {
1027 CalleeDecl = dyn_cast<Function>(Val: CalleeOp.getGlobal());
1028 if (CalleeDecl)
1029 CalleeSP = CalleeDecl->getSubprogram(); // might be nullptr
1030 }
1031
1032 // Omit DIE if we can't tell where the call goes *and* we don't want to
1033 // add metadata to it.
1034 if (CalleeSP == nullptr && CallTarget.getReg() == 0 &&
1035 AllocSiteTy == nullptr)
1036 continue;
1037
1038 // TODO: Omit call site entries for runtime calls (objc_msgSend, etc).
1039
1040 bool IsTail = TII->isTailCall(Inst: MI);
1041
1042 // If MI is in a bundle, the label was created after the bundle since
1043 // EmitFunctionBody iterates over top-level MIs. Get that top-level MI
1044 // to search for that label below.
1045 const MachineInstr *TopLevelCallMI =
1046 MI.isInsideBundle() ? &*getBundleStart(I: MI.getIterator()) : &MI;
1047
1048 // For non-tail calls, the return PC is needed to disambiguate paths in
1049 // the call graph which could lead to some target function. For tail
1050 // calls, no return PC information is needed, unless tuning for GDB in
1051 // DWARF4 mode in which case we fake a return PC for compatibility.
1052 const MCSymbol *PCAddr = (!IsTail || CU.useGNUAnalogForDwarf5Feature())
1053 ? getLabelAfterInsn(MI: TopLevelCallMI)
1054 : nullptr;
1055
1056 // For tail calls, it's necessary to record the address of the branch
1057 // instruction so that the debugger can show where the tail call occurred.
1058 const MCSymbol *CallAddr =
1059 IsTail ? getLabelBeforeInsn(MI: TopLevelCallMI) : nullptr;
1060
1061 assert((IsTail || PCAddr) && "Non-tail call without return PC");
1062
1063 LLVM_DEBUG(
1064 dbgs() << "CallSiteEntry: " << MF.getName() << " -> "
1065 << (CalleeDecl
1066 ? CalleeDecl->getName()
1067 : StringRef(
1068 MF.getSubtarget().getRegisterInfo()->getName(
1069 CallTarget.getReg())))
1070 << (IsTail ? " [IsTail]" : "") << "\n");
1071
1072 DIE &CallSiteDIE = CU.constructCallSiteEntryDIE(
1073 ScopeDIE, CalleeSP, CalleeF: CalleeDecl, IsTail, PCAddr, CallAddr, CallTarget,
1074 Offset, AllocSiteTy);
1075
1076 if (CallTarget.getReg())
1077 addCallSiteTargetForIndirectCalls(TopLevelCallMI, CallSiteDIE);
1078
1079 // Optionally emit call-site-param debug info.
1080 if (emitDebugEntryValues()) {
1081 ParamSet Params;
1082 // Try to interpret values of call site parameters.
1083 collectCallSiteParameters(CallMI: &MI, Params);
1084 CU.constructCallSiteParmEntryDIEs(CallSiteDIE, Params);
1085 }
1086 }
1087 }
1088}
1089
1090void DwarfDebug::addGnuPubAttributes(DwarfCompileUnit &U, DIE &D) const {
1091 if (!U.hasDwarfPubSections())
1092 return;
1093
1094 U.addFlag(Die&: D, Attribute: dwarf::DW_AT_GNU_pubnames);
1095}
1096
1097void DwarfDebug::finishUnitAttributes(const DICompileUnit *DIUnit,
1098 DwarfCompileUnit &NewCU) {
1099 DIE &Die = NewCU.getUnitDie();
1100 StringRef FN = DIUnit->getFilename();
1101
1102 StringRef Producer = DIUnit->getProducer();
1103 StringRef Flags = DIUnit->getFlags();
1104 if (!Flags.empty() && !useAppleExtensionAttributes()) {
1105 std::string ProducerWithFlags = Producer.str() + " " + Flags.str();
1106 NewCU.addString(Die, Attribute: dwarf::DW_AT_producer, Str: ProducerWithFlags);
1107 } else
1108 NewCU.addString(Die, Attribute: dwarf::DW_AT_producer, Str: Producer);
1109
1110 if (auto Lang = DIUnit->getSourceLanguage(); Lang.hasVersionedName()) {
1111 NewCU.addUInt(Die, Attribute: dwarf::DW_AT_language_name, Form: dwarf::DW_FORM_data2,
1112 Integer: Lang.getName());
1113
1114 if (uint32_t LangVersion = Lang.getVersion(); LangVersion != 0)
1115 NewCU.addUInt(Die, Attribute: dwarf::DW_AT_language_version, /*Form=*/std::nullopt,
1116 Integer: LangVersion);
1117 } else {
1118 NewCU.addUInt(Die, Attribute: dwarf::DW_AT_language, Form: dwarf::DW_FORM_data2,
1119 Integer: Lang.getName());
1120 }
1121
1122 NewCU.addString(Die, Attribute: dwarf::DW_AT_name, Str: FN);
1123
1124 finishTargetUnitAttributes(DIUnit: *DIUnit, NewCU);
1125
1126 StringRef SysRoot = DIUnit->getSysRoot();
1127 if (!SysRoot.empty())
1128 NewCU.addString(Die, Attribute: dwarf::DW_AT_LLVM_sysroot, Str: SysRoot);
1129 StringRef SDK = DIUnit->getSDK();
1130 if (!SDK.empty())
1131 NewCU.addString(Die, Attribute: dwarf::DW_AT_APPLE_sdk, Str: SDK);
1132
1133 if (!useSplitDwarf()) {
1134 // Add DW_str_offsets_base to the unit DIE, except for split units.
1135 if (useSegmentedStringOffsetsTable())
1136 NewCU.addStringOffsetsStart();
1137
1138 NewCU.initStmtList();
1139
1140 // If we're using split dwarf the compilation dir is going to be in the
1141 // skeleton CU and so we don't need to duplicate it here.
1142 if (!CompilationDir.empty())
1143 NewCU.addString(Die, Attribute: dwarf::DW_AT_comp_dir, Str: CompilationDir);
1144 addGnuPubAttributes(U&: NewCU, D&: Die);
1145 }
1146
1147 if (useAppleExtensionAttributes()) {
1148 if (DIUnit->isOptimized())
1149 NewCU.addFlag(Die, Attribute: dwarf::DW_AT_APPLE_optimized);
1150
1151 StringRef Flags = DIUnit->getFlags();
1152 if (!Flags.empty())
1153 NewCU.addString(Die, Attribute: dwarf::DW_AT_APPLE_flags, Str: Flags);
1154
1155 if (unsigned RVer = DIUnit->getRuntimeVersion())
1156 NewCU.addUInt(Die, Attribute: dwarf::DW_AT_APPLE_major_runtime_vers,
1157 Form: dwarf::DW_FORM_data1, Integer: RVer);
1158 }
1159
1160 if (DIUnit->getDWOId()) {
1161 // This CU is either a clang module DWO or a skeleton CU.
1162 NewCU.addUInt(Die, Attribute: dwarf::DW_AT_GNU_dwo_id, Form: dwarf::DW_FORM_data8,
1163 Integer: DIUnit->getDWOId());
1164 if (!DIUnit->getSplitDebugFilename().empty()) {
1165 // This is a prefabricated skeleton CU.
1166 dwarf::Attribute attrDWOName = getDwarfVersion() >= 5
1167 ? dwarf::DW_AT_dwo_name
1168 : dwarf::DW_AT_GNU_dwo_name;
1169 NewCU.addString(Die, Attribute: attrDWOName, Str: DIUnit->getSplitDebugFilename());
1170 }
1171 }
1172}
1173
1174DwarfCompileUnit *DwarfDebug::getDwarfCompileUnit(const DICompileUnit *DIUnit) {
1175 if (auto *CU = CUMap.lookup(Key: DIUnit))
1176 return CU;
1177
1178 if (useSplitDwarf() && !shareAcrossDWOCUs() &&
1179 (!DIUnit->getSplitDebugInlining() ||
1180 DIUnit->getEmissionKind() == DICompileUnit::FullDebug) &&
1181 !CUMap.empty())
1182 return CUMap.begin()->second;
1183
1184 return nullptr;
1185}
1186
1187// Create new DwarfCompileUnit for the given metadata node with tag
1188// DW_TAG_compile_unit.
1189DwarfCompileUnit &
1190DwarfDebug::getOrCreateDwarfCompileUnit(const DICompileUnit *DIUnit) {
1191 if (auto *CU = getDwarfCompileUnit(DIUnit))
1192 return *CU;
1193
1194 CompilationDir = DIUnit->getDirectory();
1195
1196 auto OwnedUnit = std::make_unique<DwarfCompileUnit>(
1197 args: InfoHolder.getUnits().size(), args&: DIUnit, args&: Asm, args: this, args: &InfoHolder);
1198 DwarfCompileUnit &NewCU = *OwnedUnit;
1199 InfoHolder.addUnit(U: std::move(OwnedUnit));
1200
1201 // LTO with assembly output shares a single line table amongst multiple CUs.
1202 // To avoid the compilation directory being ambiguous, let the line table
1203 // explicitly describe the directory of all files, never relying on the
1204 // compilation directory.
1205 if (!Asm->OutStreamer->hasRawTextSupport() || SingleCU)
1206 Asm->OutStreamer->emitDwarfFile0Directive(
1207 Directory: CompilationDir, Filename: DIUnit->getFilename(), Checksum: getMD5AsBytes(File: DIUnit->getFile()),
1208 Source: DIUnit->getSource(), CUID: NewCU.getUniqueID());
1209
1210 if (useSplitDwarf()) {
1211 NewCU.setSkeleton(constructSkeletonCU(CU: NewCU));
1212 NewCU.setSection(Asm->getObjFileLowering().getDwarfInfoDWOSection());
1213 } else {
1214 finishUnitAttributes(DIUnit, NewCU);
1215 NewCU.setSection(Asm->getObjFileLowering().getDwarfInfoSection());
1216 }
1217
1218 CUMap.insert(KV: {DIUnit, &NewCU});
1219 CUDieMap.insert(KV: {&NewCU.getUnitDie(), &NewCU});
1220 return NewCU;
1221}
1222
1223/// Sort and unique GVEs by comparing their fragment offset.
1224static SmallVectorImpl<DwarfCompileUnit::GlobalExpr> &
1225sortGlobalExprs(SmallVectorImpl<DwarfCompileUnit::GlobalExpr> &GVEs) {
1226 llvm::sort(
1227 C&: GVEs, Comp: [](DwarfCompileUnit::GlobalExpr A, DwarfCompileUnit::GlobalExpr B) {
1228 // Sort order: first null exprs, then exprs without fragment
1229 // info, then sort by fragment offset in bits.
1230 // FIXME: Come up with a more comprehensive comparator so
1231 // the sorting isn't non-deterministic, and so the following
1232 // std::unique call works correctly.
1233 if (!A.Expr || !B.Expr)
1234 return !!B.Expr;
1235 auto FragmentA = A.Expr->getFragmentInfo();
1236 auto FragmentB = B.Expr->getFragmentInfo();
1237 if (!FragmentA || !FragmentB)
1238 return !!FragmentB;
1239 return FragmentA->OffsetInBits < FragmentB->OffsetInBits;
1240 });
1241 GVEs.erase(CS: llvm::unique(R&: GVEs,
1242 P: [](DwarfCompileUnit::GlobalExpr A,
1243 DwarfCompileUnit::GlobalExpr B) {
1244 return A.Expr == B.Expr;
1245 }),
1246 CE: GVEs.end());
1247 return GVEs;
1248}
1249
1250// Emit all Dwarf sections that should come prior to the content. Create
1251// global DIEs and emit initial debug info sections. This is invoked by
1252// the target AsmPrinter.
1253void DwarfDebug::beginModule(Module *M) {
1254 DebugHandlerBase::beginModule(M);
1255
1256 if (!Asm)
1257 return;
1258
1259 unsigned NumDebugCUs = std::distance(first: M->debug_compile_units_begin(),
1260 last: M->debug_compile_units_end());
1261 if (NumDebugCUs == 0)
1262 return;
1263
1264 assert(NumDebugCUs > 0 && "Asm unexpectedly initialized");
1265 SingleCU = NumDebugCUs == 1;
1266
1267 // Create the symbol that designates the start of the unit's contribution
1268 // to the string offsets table. In a split DWARF scenario, only the skeleton
1269 // unit has the DW_AT_str_offsets_base attribute (and hence needs the symbol).
1270 if (useSegmentedStringOffsetsTable())
1271 (useSplitDwarf() ? SkeletonHolder : InfoHolder)
1272 .setStringOffsetsStartSym(Asm->createTempSymbol(Name: "str_offsets_base"));
1273
1274
1275 // Create the symbols that designates the start of the DWARF v5 range list
1276 // and locations list tables. They are located past the table headers.
1277 if (getDwarfVersion() >= 5) {
1278 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder;
1279 Holder.setRnglistsTableBaseSym(
1280 Asm->createTempSymbol(Name: "rnglists_table_base"));
1281
1282 if (useSplitDwarf())
1283 InfoHolder.setRnglistsTableBaseSym(
1284 Asm->createTempSymbol(Name: "rnglists_dwo_table_base"));
1285 }
1286
1287 // Create the symbol that points to the first entry following the debug
1288 // address table (.debug_addr) header.
1289 AddrPool.setLabel(Asm->createTempSymbol(Name: "addr_table_base"));
1290 DebugLocs.setSym(Asm->createTempSymbol(Name: "loclists_table_base"));
1291
1292 for (DICompileUnit *CUNode : M->debug_compile_units()) {
1293 if (CUNode->getImportedEntities().empty() &&
1294 CUNode->getEnumTypes().empty() && CUNode->getRetainedTypes().empty() &&
1295 CUNode->getGlobalVariables().empty() && CUNode->getMacros().empty())
1296 continue;
1297
1298 DwarfCompileUnit &CU = getOrCreateDwarfCompileUnit(DIUnit: CUNode);
1299
1300 for (auto *Ty : CUNode->getEnumTypes()) {
1301 assert(!isa_and_nonnull<DILocalScope>(Ty->getScope()) &&
1302 "Unexpected function-local entity in 'enums' CU field.");
1303 CU.getOrCreateTypeDIE(TyNode: cast<DIType>(Val: Ty));
1304 }
1305
1306 for (auto *Ty : CUNode->getRetainedTypes()) {
1307 if (DIType *RT = dyn_cast<DIType>(Val: Ty))
1308 // There is no point in force-emitting a forward declaration.
1309 CU.getOrCreateTypeDIE(TyNode: RT);
1310 }
1311 }
1312}
1313
1314void DwarfDebug::finishEntityDefinitions() {
1315 for (const auto &Entity : ConcreteEntities) {
1316 DIE *Die = Entity->getDIE();
1317 assert(Die);
1318 // FIXME: Consider the time-space tradeoff of just storing the unit pointer
1319 // in the ConcreteEntities list, rather than looking it up again here.
1320 // DIE::getUnit isn't simple - it walks parent pointers, etc.
1321 DwarfCompileUnit *Unit = CUDieMap.lookup(Val: Die->getUnitDie());
1322 assert(Unit);
1323 Unit->finishEntityDefinition(Entity: Entity.get());
1324 }
1325}
1326
1327void DwarfDebug::finishSubprogramDefinitions() {
1328 for (const DISubprogram *SP : ProcessedSPNodes) {
1329 assert(SP->getUnit()->getEmissionKind() != DICompileUnit::NoDebug);
1330 forBothCUs(
1331 CU&: getOrCreateDwarfCompileUnit(DIUnit: SP->getUnit()),
1332 F: [&](DwarfCompileUnit &CU) { CU.finishSubprogramDefinition(SP); });
1333 }
1334}
1335
1336void DwarfDebug::finalizeModuleInfo() {
1337 const TargetLoweringObjectFile &TLOF = Asm->getObjFileLowering();
1338
1339 finishSubprogramDefinitions();
1340
1341 finishEntityDefinitions();
1342
1343 bool HasEmittedSplitCU = false;
1344
1345 // Handle anything that needs to be done on a per-unit basis after
1346 // all other generation.
1347 for (const auto &P : CUMap) {
1348 auto &TheCU = *P.second;
1349 if (TheCU.getCUNode()->isDebugDirectivesOnly())
1350 continue;
1351 TheCU.attachLexicalScopesAbstractOrigins();
1352 // Emit DW_AT_containing_type attribute to connect types with their
1353 // vtable holding type.
1354 TheCU.constructContainingTypeDIEs();
1355
1356 // Add CU specific attributes if we need to add any.
1357 // If we're splitting the dwarf out now that we've got the entire
1358 // CU then add the dwo id to it.
1359 auto *SkCU = TheCU.getSkeleton();
1360
1361 bool HasSplitUnit = SkCU && !TheCU.getUnitDie().children().empty();
1362
1363 if (HasSplitUnit) {
1364 (void)HasEmittedSplitCU;
1365 assert((shareAcrossDWOCUs() || !HasEmittedSplitCU) &&
1366 "Multiple CUs emitted into a single dwo file");
1367 HasEmittedSplitCU = true;
1368 dwarf::Attribute attrDWOName = getDwarfVersion() >= 5
1369 ? dwarf::DW_AT_dwo_name
1370 : dwarf::DW_AT_GNU_dwo_name;
1371 finishUnitAttributes(DIUnit: TheCU.getCUNode(), NewCU&: TheCU);
1372 StringRef DWOName = Asm->TM.Options.MCOptions.SplitDwarfFile;
1373 TheCU.addString(Die&: TheCU.getUnitDie(), Attribute: attrDWOName, Str: DWOName);
1374 SkCU->addString(Die&: SkCU->getUnitDie(), Attribute: attrDWOName, Str: DWOName);
1375 // Emit a unique identifier for this CU. Include the DWO file name in the
1376 // hash to avoid the case where two (almost) empty compile units have the
1377 // same contents. This can happen if link-time optimization removes nearly
1378 // all (unused) code from a CU.
1379 uint64_t ID =
1380 DIEHash(Asm, &TheCU).computeCUSignature(DWOName, Die: TheCU.getUnitDie());
1381 if (getDwarfVersion() >= 5) {
1382 TheCU.setDWOId(ID);
1383 SkCU->setDWOId(ID);
1384 } else {
1385 TheCU.addUInt(Die&: TheCU.getUnitDie(), Attribute: dwarf::DW_AT_GNU_dwo_id,
1386 Form: dwarf::DW_FORM_data8, Integer: ID);
1387 SkCU->addUInt(Die&: SkCU->getUnitDie(), Attribute: dwarf::DW_AT_GNU_dwo_id,
1388 Form: dwarf::DW_FORM_data8, Integer: ID);
1389 }
1390
1391 if (getDwarfVersion() < 5 && !SkeletonHolder.getRangeLists().empty()) {
1392 const MCSymbol *Sym = TLOF.getDwarfRangesSection()->getBeginSymbol();
1393 SkCU->addSectionLabel(Die&: SkCU->getUnitDie(), Attribute: dwarf::DW_AT_GNU_ranges_base,
1394 Label: Sym, Sec: Sym);
1395 }
1396 } else if (SkCU) {
1397 finishUnitAttributes(DIUnit: SkCU->getCUNode(), NewCU&: *SkCU);
1398 }
1399
1400 // If we have code split among multiple sections or non-contiguous
1401 // ranges of code then emit a DW_AT_ranges attribute on the unit that will
1402 // remain in the .o file, otherwise add a DW_AT_low_pc.
1403 // FIXME: We should use ranges allow reordering of code ala
1404 // .subsections_via_symbols in mach-o. This would mean turning on
1405 // ranges for all subprogram DIEs for mach-o.
1406 DwarfCompileUnit &U = SkCU ? *SkCU : TheCU;
1407
1408 if (unsigned NumRanges = TheCU.getRanges().size()) {
1409 if (shouldAttachCompileUnitRanges()) {
1410 if (NumRanges > 1 && useRangesSection())
1411 // A DW_AT_low_pc attribute may also be specified in combination with
1412 // DW_AT_ranges to specify the default base address for use in
1413 // location lists (see Section 2.6.2) and range lists (see Section
1414 // 2.17.3).
1415 U.addUInt(Die&: U.getUnitDie(), Attribute: dwarf::DW_AT_low_pc, Form: dwarf::DW_FORM_addr,
1416 Integer: 0);
1417 else
1418 U.setBaseAddress(TheCU.getRanges().front().Begin);
1419 U.attachRangesOrLowHighPC(D&: U.getUnitDie(), Ranges: TheCU.takeRanges());
1420 }
1421 }
1422
1423 // We don't keep track of which addresses are used in which CU so this
1424 // is a bit pessimistic under LTO.
1425 if ((HasSplitUnit || getDwarfVersion() >= 5) && !AddrPool.isEmpty())
1426 U.addAddrTableBase();
1427
1428 if (getDwarfVersion() >= 5) {
1429 if (U.hasRangeLists())
1430 U.addRnglistsBase();
1431
1432 if (!DebugLocs.getLists().empty() && !useSplitDwarf()) {
1433 U.addSectionLabel(Die&: U.getUnitDie(), Attribute: dwarf::DW_AT_loclists_base,
1434 Label: DebugLocs.getSym(),
1435 Sec: TLOF.getDwarfLoclistsSection()->getBeginSymbol());
1436 }
1437 }
1438
1439 auto *CUNode = cast<DICompileUnit>(Val: P.first);
1440 // If compile Unit has macros, emit "DW_AT_macro_info/DW_AT_macros"
1441 // attribute.
1442 if (CUNode->getMacros()) {
1443 if (UseDebugMacroSection) {
1444 if (useSplitDwarf())
1445 TheCU.addSectionDelta(
1446 Die&: TheCU.getUnitDie(), Attribute: dwarf::DW_AT_macros, Hi: U.getMacroLabelBegin(),
1447 Lo: TLOF.getDwarfMacroDWOSection()->getBeginSymbol());
1448 else {
1449 dwarf::Attribute MacrosAttr = getDwarfVersion() >= 5
1450 ? dwarf::DW_AT_macros
1451 : dwarf::DW_AT_GNU_macros;
1452 U.addSectionLabel(Die&: U.getUnitDie(), Attribute: MacrosAttr, Label: U.getMacroLabelBegin(),
1453 Sec: TLOF.getDwarfMacroSection()->getBeginSymbol());
1454 }
1455 } else {
1456 if (useSplitDwarf())
1457 TheCU.addSectionDelta(
1458 Die&: TheCU.getUnitDie(), Attribute: dwarf::DW_AT_macro_info,
1459 Hi: U.getMacroLabelBegin(),
1460 Lo: TLOF.getDwarfMacinfoDWOSection()->getBeginSymbol());
1461 else
1462 U.addSectionLabel(Die&: U.getUnitDie(), Attribute: dwarf::DW_AT_macro_info,
1463 Label: U.getMacroLabelBegin(),
1464 Sec: TLOF.getDwarfMacinfoSection()->getBeginSymbol());
1465 }
1466 }
1467 }
1468
1469 // Emit all frontend-produced Skeleton CUs, i.e., Clang modules.
1470 for (auto *CUNode : MMI->getModule()->debug_compile_units())
1471 if (CUNode->getDWOId())
1472 getOrCreateDwarfCompileUnit(DIUnit: CUNode);
1473
1474 // Compute DIE offsets and sizes.
1475 InfoHolder.computeSizeAndOffsets();
1476 if (useSplitDwarf())
1477 SkeletonHolder.computeSizeAndOffsets();
1478
1479 // Now that offsets are computed, can replace DIEs in debug_names Entry with
1480 // an actual offset.
1481 AccelDebugNames.convertDieToOffset();
1482}
1483
1484// Emit all Dwarf sections that should come after the content.
1485void DwarfDebug::endModule() {
1486 // Terminate the pending line table.
1487 if (PrevCU)
1488 terminateLineTable(CU: PrevCU);
1489 PrevCU = nullptr;
1490 assert(CurFn == nullptr);
1491 assert(CurMI == nullptr);
1492
1493 const Module *M = MMI->getModule();
1494
1495 // Collect global variables info.
1496 DenseMap<DIGlobalVariable *, SmallVector<DwarfCompileUnit::GlobalExpr, 1>>
1497 GVMap;
1498 for (const GlobalVariable &Global : M->globals()) {
1499 SmallVector<DIGlobalVariableExpression *, 1> GVs;
1500 Global.getDebugInfo(GVs);
1501 for (auto *GVE : GVs)
1502 GVMap[GVE->getVariable()].push_back(Elt: {.Var: &Global, .Expr: GVE->getExpression()});
1503 }
1504
1505 for (DICompileUnit *CUNode : M->debug_compile_units()) {
1506 DwarfCompileUnit *CU = getDwarfCompileUnit(DIUnit: CUNode);
1507
1508 // If the CU hasn't been emitted yet, it must be empty. Skip it.
1509 if (!CU)
1510 continue;
1511
1512 // Emit Global Variables.
1513 for (auto *GVE : CUNode->getGlobalVariables()) {
1514 // Don't bother adding DIGlobalVariableExpressions listed in the CU if we
1515 // already know about the variable and it isn't adding a constant
1516 // expression.
1517 auto &GVMapEntry = GVMap[GVE->getVariable()];
1518 auto *Expr = GVE->getExpression();
1519 if (!GVMapEntry.size() || (Expr && Expr->isConstant()))
1520 GVMapEntry.push_back(Elt: {.Var: nullptr, .Expr: Expr});
1521 }
1522 DenseSet<DIGlobalVariable *> Processed;
1523 for (auto *GVE : CUNode->getGlobalVariables()) {
1524 DIGlobalVariable *GV = GVE->getVariable();
1525 if (Processed.insert(V: GV).second)
1526 CU->getOrCreateGlobalVariableDIE(GV, GlobalExprs: sortGlobalExprs(GVEs&: GVMap[GV]));
1527 }
1528
1529 // Emit imported entities.
1530 for (auto *IE : CUNode->getImportedEntities()) {
1531 assert(!isa_and_nonnull<DILocalScope>(IE->getScope()) &&
1532 "Unexpected function-local entity in 'imports' CU field.");
1533 CU->getOrCreateImportedEntityDIE(IE);
1534 }
1535
1536 // Emit function-local entities.
1537 for (const auto *D : CU->getDeferredLocalDecls()) {
1538 if (auto *IE = dyn_cast<DIImportedEntity>(Val: D))
1539 CU->getOrCreateImportedEntityDIE(IE);
1540 else if (auto *Ty = dyn_cast<DIType>(Val: D))
1541 CU->getOrCreateTypeDIE(TyNode: Ty);
1542 else
1543 llvm_unreachable("Unexpected local retained node!");
1544 }
1545
1546 // Emit base types.
1547 CU->createBaseTypeDIEs();
1548 }
1549
1550 // If we aren't actually generating debug info (check beginModule -
1551 // conditionalized on the presence of the llvm.dbg.cu metadata node)
1552 if (!Asm || !Asm->hasDebugInfo())
1553 return;
1554
1555 // Finalize the debug info for the module.
1556 finalizeModuleInfo();
1557
1558 if (useSplitDwarf())
1559 // Emit debug_loc.dwo/debug_loclists.dwo section.
1560 emitDebugLocDWO();
1561 else
1562 // Emit debug_loc/debug_loclists section.
1563 emitDebugLoc();
1564
1565 // Corresponding abbreviations into a abbrev section.
1566 emitAbbreviations();
1567
1568 // Emit all the DIEs into a debug info section.
1569 emitDebugInfo();
1570
1571 // Emit info into a debug aranges section.
1572 if (UseARangesSection)
1573 emitDebugARanges();
1574
1575 // Emit info into a debug ranges section.
1576 emitDebugRanges();
1577
1578 if (useSplitDwarf())
1579 // Emit info into a debug macinfo.dwo section.
1580 emitDebugMacinfoDWO();
1581 else
1582 // Emit info into a debug macinfo/macro section.
1583 emitDebugMacinfo();
1584
1585 emitDebugStr();
1586
1587 if (useSplitDwarf()) {
1588 emitDebugStrDWO();
1589 emitDebugInfoDWO();
1590 emitDebugAbbrevDWO();
1591 emitDebugLineDWO();
1592 emitDebugRangesDWO();
1593 }
1594
1595 emitDebugAddr();
1596
1597 // Emit info into the dwarf accelerator table sections.
1598 switch (getAccelTableKind()) {
1599 case AccelTableKind::Apple:
1600 emitAccelNames();
1601 emitAccelObjC();
1602 emitAccelNamespaces();
1603 emitAccelTypes();
1604 break;
1605 case AccelTableKind::Dwarf:
1606 emitAccelDebugNames();
1607 break;
1608 case AccelTableKind::None:
1609 break;
1610 case AccelTableKind::Default:
1611 llvm_unreachable("Default should have already been resolved.");
1612 }
1613
1614 // Emit the pubnames and pubtypes sections if requested.
1615 emitDebugPubSections();
1616
1617 // clean up.
1618 // FIXME: AbstractVariables.clear();
1619}
1620
1621void DwarfDebug::ensureAbstractEntityIsCreatedIfScoped(DwarfCompileUnit &CU,
1622 const DINode *Node, const MDNode *ScopeNode) {
1623 if (CU.getExistingAbstractEntity(Node))
1624 return;
1625
1626 if (LexicalScope *Scope =
1627 LScopes.findAbstractScope(N: cast_or_null<DILocalScope>(Val: ScopeNode)))
1628 CU.createAbstractEntity(Node, Scope);
1629}
1630
1631static const DILocalScope *getRetainedNodeScope(const MDNode *N) {
1632 // Ensure the scope is not a DILexicalBlockFile.
1633 return DISubprogram::getRetainedNodeScope(N)->getNonLexicalBlockFileScope();
1634}
1635
1636// Collect variable information from side table maintained by MF.
1637void DwarfDebug::collectVariableInfoFromMFTable(
1638 DwarfCompileUnit &TheCU, DenseSet<InlinedEntity> &Processed) {
1639 SmallDenseMap<InlinedEntity, DbgVariable *> MFVars;
1640 LLVM_DEBUG(dbgs() << "DwarfDebug: collecting variables from MF side table\n");
1641 for (const auto &VI : Asm->MF->getVariableDbgInfo()) {
1642 if (!VI.Var)
1643 continue;
1644 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) &&
1645 "Expected inlined-at fields to agree");
1646
1647 InlinedEntity Var(VI.Var, VI.Loc->getInlinedAt());
1648 Processed.insert(V: Var);
1649 LexicalScope *Scope = LScopes.findLexicalScope(DL: VI.Loc);
1650
1651 // If variable scope is not found then skip this variable.
1652 if (!Scope) {
1653 LLVM_DEBUG(dbgs() << "Dropping debug info for " << VI.Var->getName()
1654 << ", no variable scope found\n");
1655 continue;
1656 }
1657
1658 ensureAbstractEntityIsCreatedIfScoped(CU&: TheCU, Node: Var.first, ScopeNode: Scope->getScopeNode());
1659
1660 // If we have already seen information for this variable, add to what we
1661 // already know.
1662 if (DbgVariable *PreviousLoc = MFVars.lookup(Val: Var)) {
1663 auto *PreviousMMI = std::get_if<Loc::MMI>(ptr: PreviousLoc);
1664 auto *PreviousEntryValue = std::get_if<Loc::EntryValue>(ptr: PreviousLoc);
1665 // Previous and new locations are both stack slots (MMI).
1666 if (PreviousMMI && VI.inStackSlot())
1667 PreviousMMI->addFrameIndexExpr(Expr: VI.Expr, FI: VI.getStackSlot());
1668 // Previous and new locations are both entry values.
1669 else if (PreviousEntryValue && VI.inEntryValueRegister())
1670 PreviousEntryValue->addExpr(Reg: VI.getEntryValueRegister(), Expr: *VI.Expr);
1671 else {
1672 // Locations differ, this should (rarely) happen in optimized async
1673 // coroutines.
1674 // Prefer whichever location has an EntryValue.
1675 if (PreviousLoc->holds<Loc::MMI>())
1676 PreviousLoc->emplace<Loc::EntryValue>(args: VI.getEntryValueRegister(),
1677 args: *VI.Expr);
1678 LLVM_DEBUG(dbgs() << "Dropping debug info for " << VI.Var->getName()
1679 << ", conflicting fragment location types\n");
1680 }
1681 continue;
1682 }
1683
1684 auto RegVar = std::make_unique<DbgVariable>(
1685 args: cast<DILocalVariable>(Val: Var.first), args&: Var.second);
1686 if (VI.inStackSlot())
1687 RegVar->emplace<Loc::MMI>(args: VI.Expr, args: VI.getStackSlot());
1688 else
1689 RegVar->emplace<Loc::EntryValue>(args: VI.getEntryValueRegister(), args: *VI.Expr);
1690 LLVM_DEBUG(dbgs() << "Created DbgVariable for " << VI.Var->getName()
1691 << "\n");
1692 InfoHolder.addScopeVariable(LS: Scope, Var: RegVar.get());
1693 MFVars.insert(KV: {Var, RegVar.get()});
1694 ConcreteEntities.push_back(Elt: std::move(RegVar));
1695 }
1696}
1697
1698/// Determine whether a *singular* DBG_VALUE is valid for the entirety of its
1699/// enclosing lexical scope. The check ensures there are no other instructions
1700/// in the same lexical scope preceding the DBG_VALUE and that its range is
1701/// either open or otherwise rolls off the end of the scope.
1702static bool validThroughout(LexicalScopes &LScopes,
1703 const MachineInstr *DbgValue,
1704 const MachineInstr *RangeEnd,
1705 const InstructionOrdering &Ordering) {
1706 assert(DbgValue->getDebugLoc() && "DBG_VALUE without a debug location");
1707 auto MBB = DbgValue->getParent();
1708 auto DL = DbgValue->getDebugLoc();
1709 auto *LScope = LScopes.findLexicalScope(DL);
1710 // Scope doesn't exist; this is a dead DBG_VALUE.
1711 if (!LScope)
1712 return false;
1713 auto &LSRange = LScope->getRanges();
1714 if (LSRange.size() == 0)
1715 return false;
1716
1717 const MachineInstr *LScopeBegin = LSRange.front().first;
1718 // If the scope starts before the DBG_VALUE then we may have a negative
1719 // result. Otherwise the location is live coming into the scope and we
1720 // can skip the following checks.
1721 if (!Ordering.isBefore(A: DbgValue, B: LScopeBegin)) {
1722 // Exit if the lexical scope begins outside of the current block.
1723 if (LScopeBegin->getParent() != MBB)
1724 return false;
1725
1726 MachineBasicBlock::const_reverse_iterator Pred(DbgValue);
1727 for (++Pred; Pred != MBB->rend(); ++Pred) {
1728 if (Pred->getFlag(Flag: MachineInstr::FrameSetup))
1729 break;
1730 auto PredDL = Pred->getDebugLoc();
1731 if (!PredDL || Pred->isMetaInstruction())
1732 continue;
1733 // Check whether the instruction preceding the DBG_VALUE is in the same
1734 // (sub)scope as the DBG_VALUE.
1735 if (DL->getScope() == PredDL->getScope())
1736 return false;
1737 auto *PredScope = LScopes.findLexicalScope(DL: PredDL);
1738 if (!PredScope || LScope->dominates(S: PredScope))
1739 return false;
1740 }
1741 }
1742
1743 // If the range of the DBG_VALUE is open-ended, report success.
1744 if (!RangeEnd)
1745 return true;
1746
1747 // Single, constant DBG_VALUEs in the prologue are promoted to be live
1748 // throughout the function. This is a hack, presumably for DWARF v2 and not
1749 // necessarily correct. It would be much better to use a dbg.declare instead
1750 // if we know the constant is live throughout the scope.
1751 if (MBB->pred_empty() &&
1752 all_of(Range: DbgValue->debug_operands(),
1753 P: [](const MachineOperand &Op) { return Op.isImm(); }))
1754 return true;
1755
1756 // Test if the location terminates before the end of the scope.
1757 const MachineInstr *LScopeEnd = LSRange.back().second;
1758 if (Ordering.isBefore(A: RangeEnd, B: LScopeEnd))
1759 return false;
1760
1761 // There's a single location which starts at the scope start, and ends at or
1762 // after the scope end.
1763 return true;
1764}
1765
1766/// Build the location list for all DBG_VALUEs in the function that
1767/// describe the same variable. The resulting DebugLocEntries will have
1768/// strict monotonically increasing begin addresses and will never
1769/// overlap. If the resulting list has only one entry that is valid
1770/// throughout variable's scope return true.
1771//
1772// See the definition of DbgValueHistoryMap::Entry for an explanation of the
1773// different kinds of history map entries. One thing to be aware of is that if
1774// a debug value is ended by another entry (rather than being valid until the
1775// end of the function), that entry's instruction may or may not be included in
1776// the range, depending on if the entry is a clobbering entry (it has an
1777// instruction that clobbers one or more preceding locations), or if it is an
1778// (overlapping) debug value entry. This distinction can be seen in the example
1779// below. The first debug value is ended by the clobbering entry 2, and the
1780// second and third debug values are ended by the overlapping debug value entry
1781// 4.
1782//
1783// Input:
1784//
1785// History map entries [type, end index, mi]
1786//
1787// 0 | [DbgValue, 2, DBG_VALUE $reg0, [...] (fragment 0, 32)]
1788// 1 | | [DbgValue, 4, DBG_VALUE $reg1, [...] (fragment 32, 32)]
1789// 2 | | [Clobber, $reg0 = [...], -, -]
1790// 3 | | [DbgValue, 4, DBG_VALUE 123, [...] (fragment 64, 32)]
1791// 4 [DbgValue, ~0, DBG_VALUE @g, [...] (fragment 0, 96)]
1792//
1793// Output [start, end) [Value...]:
1794//
1795// [0-1) [(reg0, fragment 0, 32)]
1796// [1-3) [(reg0, fragment 0, 32), (reg1, fragment 32, 32)]
1797// [3-4) [(reg1, fragment 32, 32), (123, fragment 64, 32)]
1798// [4-) [(@g, fragment 0, 96)]
1799bool DwarfDebug::buildLocationList(SmallVectorImpl<DebugLocEntry> &DebugLoc,
1800 const DbgValueHistoryMap::Entries &Entries) {
1801 using OpenRange =
1802 std::pair<DbgValueHistoryMap::EntryIndex, DbgValueLoc>;
1803 SmallVector<OpenRange, 4> OpenRanges;
1804 bool isSafeForSingleLocation = true;
1805 const MachineInstr *StartDebugMI = nullptr;
1806 const MachineInstr *EndMI = nullptr;
1807
1808 for (auto EB = Entries.begin(), EI = EB, EE = Entries.end(); EI != EE; ++EI) {
1809 const MachineInstr *Instr = EI->getInstr();
1810
1811 // Remove all values that are no longer live.
1812 size_t Index = std::distance(first: EB, last: EI);
1813 erase_if(C&: OpenRanges, P: [&](OpenRange &R) { return R.first <= Index; });
1814
1815 // If we are dealing with a clobbering entry, this iteration will result in
1816 // a location list entry starting after the clobbering instruction.
1817 const MCSymbol *StartLabel =
1818 EI->isClobber() ? getLabelAfterInsn(MI: Instr) : getLabelBeforeInsn(MI: Instr);
1819 assert(StartLabel &&
1820 "Forgot label before/after instruction starting a range!");
1821
1822 const MCSymbol *EndLabel;
1823 if (std::next(x: EI) == Entries.end()) {
1824 const MachineBasicBlock &EndMBB = Asm->MF->back();
1825 EndLabel = Asm->MBBSectionRanges[EndMBB.getSectionID()].EndLabel;
1826 if (EI->isClobber())
1827 EndMI = EI->getInstr();
1828 }
1829 else if (std::next(x: EI)->isClobber())
1830 EndLabel = getLabelAfterInsn(MI: std::next(x: EI)->getInstr());
1831 else
1832 EndLabel = getLabelBeforeInsn(MI: std::next(x: EI)->getInstr());
1833 assert(EndLabel && "Forgot label after instruction ending a range!");
1834
1835 if (EI->isDbgValue())
1836 LLVM_DEBUG(dbgs() << "DotDebugLoc: " << *Instr << "\n");
1837
1838 // If this history map entry has a debug value, add that to the list of
1839 // open ranges and check if its location is valid for a single value
1840 // location.
1841 if (EI->isDbgValue()) {
1842 // Do not add undef debug values, as they are redundant information in
1843 // the location list entries. An undef debug results in an empty location
1844 // description. If there are any non-undef fragments then padding pieces
1845 // with empty location descriptions will automatically be inserted, and if
1846 // all fragments are undef then the whole location list entry is
1847 // redundant.
1848 if (!Instr->isUndefDebugValue()) {
1849 auto Value = getDebugLocValue(MI: Instr);
1850 OpenRanges.emplace_back(Args: EI->getEndIndex(), Args&: Value);
1851
1852 // TODO: Add support for single value fragment locations.
1853 if (Instr->getDebugExpression()->isFragment())
1854 isSafeForSingleLocation = false;
1855
1856 if (!StartDebugMI)
1857 StartDebugMI = Instr;
1858 } else {
1859 isSafeForSingleLocation = false;
1860 }
1861 }
1862
1863 // Location list entries with empty location descriptions are redundant
1864 // information in DWARF, so do not emit those.
1865 if (OpenRanges.empty())
1866 continue;
1867
1868 // Omit entries with empty ranges as they do not have any effect in DWARF.
1869 if (StartLabel == EndLabel) {
1870 LLVM_DEBUG(dbgs() << "Omitting location list entry with empty range.\n");
1871 continue;
1872 }
1873
1874 SmallVector<DbgValueLoc, 4> Values;
1875 for (auto &R : OpenRanges)
1876 Values.push_back(Elt: R.second);
1877
1878 // With Basic block sections, it is posssible that the StartLabel and the
1879 // Instr are not in the same section. This happens when the StartLabel is
1880 // the function begin label and the dbg value appears in a basic block
1881 // that is not the entry. In this case, the range needs to be split to
1882 // span each individual section in the range from StartLabel to EndLabel.
1883 if (Asm->MF->hasBBSections() && StartLabel == Asm->getFunctionBegin() &&
1884 !Instr->getParent()->sameSection(MBB: &Asm->MF->front())) {
1885 for (const auto &[MBBSectionId, MBBSectionRange] :
1886 Asm->MBBSectionRanges) {
1887 if (Instr->getParent()->getSectionID() == MBBSectionId) {
1888 DebugLoc.emplace_back(Args: MBBSectionRange.BeginLabel, Args&: EndLabel, Args&: Values);
1889 break;
1890 }
1891 DebugLoc.emplace_back(Args: MBBSectionRange.BeginLabel,
1892 Args: MBBSectionRange.EndLabel, Args&: Values);
1893 }
1894 } else {
1895 DebugLoc.emplace_back(Args&: StartLabel, Args&: EndLabel, Args&: Values);
1896 }
1897
1898 // Attempt to coalesce the ranges of two otherwise identical
1899 // DebugLocEntries.
1900 auto CurEntry = DebugLoc.rbegin();
1901 LLVM_DEBUG({
1902 dbgs() << CurEntry->getValues().size() << " Values:\n";
1903 for (auto &Value : CurEntry->getValues())
1904 Value.dump();
1905 dbgs() << "-----\n";
1906 });
1907
1908 auto PrevEntry = std::next(x: CurEntry);
1909 if (PrevEntry != DebugLoc.rend() && PrevEntry->MergeRanges(Next: *CurEntry))
1910 DebugLoc.pop_back();
1911 }
1912
1913 if (!isSafeForSingleLocation ||
1914 !validThroughout(LScopes, DbgValue: StartDebugMI, RangeEnd: EndMI, Ordering: getInstOrdering()))
1915 return false;
1916
1917 if (DebugLoc.size() == 1)
1918 return true;
1919
1920 if (!Asm->MF->hasBBSections())
1921 return false;
1922
1923 // Check here to see if loclist can be merged into a single range. If not,
1924 // we must keep the split loclists per section. This does exactly what
1925 // MergeRanges does without sections. We don't actually merge the ranges
1926 // as the split ranges must be kept intact if this cannot be collapsed
1927 // into a single range.
1928 const MachineBasicBlock *RangeMBB = nullptr;
1929 if (DebugLoc[0].getBeginSym() == Asm->getFunctionBegin())
1930 RangeMBB = &Asm->MF->front();
1931 else
1932 RangeMBB = Entries.begin()->getInstr()->getParent();
1933 auto RangeIt = Asm->MBBSectionRanges.find(Key: RangeMBB->getSectionID());
1934 assert(RangeIt != Asm->MBBSectionRanges.end() &&
1935 "Range MBB not found in MBBSectionRanges!");
1936 auto *CurEntry = DebugLoc.begin();
1937 auto *NextEntry = std::next(x: CurEntry);
1938 auto NextRangeIt = std::next(x: RangeIt);
1939 while (NextEntry != DebugLoc.end()) {
1940 if (NextRangeIt == Asm->MBBSectionRanges.end())
1941 return false;
1942 // CurEntry should end the current section and NextEntry should start
1943 // the next section and the Values must match for these two ranges to be
1944 // merged. Do not match the section label end if it is the entry block
1945 // section. This is because the end label for the Debug Loc and the
1946 // Function end label could be different.
1947 if ((RangeIt->second.EndLabel != Asm->getFunctionEnd() &&
1948 CurEntry->getEndSym() != RangeIt->second.EndLabel) ||
1949 NextEntry->getBeginSym() != NextRangeIt->second.BeginLabel ||
1950 CurEntry->getValues() != NextEntry->getValues())
1951 return false;
1952 RangeIt = NextRangeIt;
1953 NextRangeIt = std::next(x: RangeIt);
1954 CurEntry = NextEntry;
1955 NextEntry = std::next(x: CurEntry);
1956 }
1957 return true;
1958}
1959
1960DbgEntity *DwarfDebug::createConcreteEntity(DwarfCompileUnit &TheCU,
1961 LexicalScope &Scope,
1962 const DINode *Node,
1963 const DILocation *Location,
1964 const MCSymbol *Sym) {
1965 ensureAbstractEntityIsCreatedIfScoped(CU&: TheCU, Node, ScopeNode: Scope.getScopeNode());
1966 if (isa<const DILocalVariable>(Val: Node)) {
1967 ConcreteEntities.push_back(
1968 Elt: std::make_unique<DbgVariable>(args: cast<const DILocalVariable>(Val: Node),
1969 args&: Location));
1970 InfoHolder.addScopeVariable(LS: &Scope,
1971 Var: cast<DbgVariable>(Val: ConcreteEntities.back().get()));
1972 } else if (isa<const DILabel>(Val: Node)) {
1973 ConcreteEntities.push_back(
1974 Elt: std::make_unique<DbgLabel>(args: cast<const DILabel>(Val: Node),
1975 args&: Location, args&: Sym));
1976 InfoHolder.addScopeLabel(LS: &Scope,
1977 Label: cast<DbgLabel>(Val: ConcreteEntities.back().get()));
1978 }
1979 return ConcreteEntities.back().get();
1980}
1981
1982// Find variables for each lexical scope.
1983void DwarfDebug::collectEntityInfo(DwarfCompileUnit &TheCU,
1984 const DISubprogram *SP,
1985 DenseSet<InlinedEntity> &Processed) {
1986 // Grab the variable info that was squirreled away in the MMI side-table.
1987 collectVariableInfoFromMFTable(TheCU, Processed);
1988
1989 for (const auto &I : DbgValues) {
1990 InlinedEntity IV = I.first;
1991 if (Processed.count(V: IV))
1992 continue;
1993
1994 // Instruction ranges, specifying where IV is accessible.
1995 const auto &HistoryMapEntries = I.second;
1996
1997 // Try to find any non-empty variable location. Do not create a concrete
1998 // entity if there are no locations.
1999 if (!DbgValues.hasNonEmptyLocation(Entries: HistoryMapEntries))
2000 continue;
2001
2002 LexicalScope *Scope = nullptr;
2003 const DILocalVariable *LocalVar = cast<DILocalVariable>(Val: IV.first);
2004 if (const DILocation *IA = IV.second)
2005 Scope = LScopes.findInlinedScope(N: LocalVar->getScope(), IA);
2006 else
2007 Scope = LScopes.findLexicalScope(N: LocalVar->getScope());
2008 // If variable scope is not found then skip this variable.
2009 if (!Scope)
2010 continue;
2011
2012 Processed.insert(V: IV);
2013 DbgVariable *RegVar = cast<DbgVariable>(Val: createConcreteEntity(TheCU,
2014 Scope&: *Scope, Node: LocalVar, Location: IV.second));
2015
2016 const MachineInstr *MInsn = HistoryMapEntries.front().getInstr();
2017 assert(MInsn->isDebugValue() && "History must begin with debug value");
2018
2019 // Check if there is a single DBG_VALUE, valid throughout the var's scope.
2020 // If the history map contains a single debug value, there may be an
2021 // additional entry which clobbers the debug value.
2022 size_t HistSize = HistoryMapEntries.size();
2023 bool SingleValueWithClobber =
2024 HistSize == 2 && HistoryMapEntries[1].isClobber();
2025 if (HistSize == 1 || SingleValueWithClobber) {
2026 const auto *End =
2027 SingleValueWithClobber ? HistoryMapEntries[1].getInstr() : nullptr;
2028 if (validThroughout(LScopes, DbgValue: MInsn, RangeEnd: End, Ordering: getInstOrdering())) {
2029 RegVar->emplace<Loc::Single>(args&: MInsn);
2030 continue;
2031 }
2032 }
2033
2034 // Handle multiple DBG_VALUE instructions describing one variable.
2035 DebugLocStream::ListBuilder List(DebugLocs, TheCU, *Asm, *RegVar);
2036
2037 // Build the location list for this variable.
2038 SmallVector<DebugLocEntry, 8> Entries;
2039 bool isValidSingleLocation = buildLocationList(DebugLoc&: Entries, Entries: HistoryMapEntries);
2040
2041 // Check whether buildLocationList managed to merge all locations to one
2042 // that is valid throughout the variable's scope. If so, produce single
2043 // value location.
2044 if (isValidSingleLocation) {
2045 RegVar->emplace<Loc::Single>(args: Entries[0].getValues()[0]);
2046 continue;
2047 }
2048
2049 // If the variable has a DIBasicType, extract it. Basic types cannot have
2050 // unique identifiers, so don't bother resolving the type with the
2051 // identifier map.
2052 const DIBasicType *BT = dyn_cast<DIBasicType>(
2053 Val: static_cast<const Metadata *>(LocalVar->getType()));
2054
2055 // Finalize the entry by lowering it into a DWARF bytestream.
2056 for (auto &Entry : Entries)
2057 Entry.finalize(AP: *Asm, List, BT, TheCU);
2058 }
2059
2060 // For each InlinedEntity collected from DBG_LABEL instructions, convert to
2061 // DWARF-related DbgLabel.
2062 for (const auto &I : DbgLabels) {
2063 InlinedEntity IL = I.first;
2064 const MachineInstr *MI = I.second;
2065 if (MI == nullptr)
2066 continue;
2067
2068 LexicalScope *Scope = nullptr;
2069 const DILabel *Label = cast<DILabel>(Val: IL.first);
2070 // The scope could have an extra lexical block file.
2071 const DILocalScope *LocalScope =
2072 Label->getScope()->getNonLexicalBlockFileScope();
2073 // Get inlined DILocation if it is inlined label.
2074 if (const DILocation *IA = IL.second)
2075 Scope = LScopes.findInlinedScope(N: LocalScope, IA);
2076 else
2077 Scope = LScopes.findLexicalScope(N: LocalScope);
2078 // If label scope is not found then skip this label.
2079 if (!Scope)
2080 continue;
2081
2082 Processed.insert(V: IL);
2083 /// At this point, the temporary label is created.
2084 /// Save the temporary label to DbgLabel entity to get the
2085 /// actually address when generating Dwarf DIE.
2086 MCSymbol *Sym = getLabelBeforeInsn(MI);
2087 createConcreteEntity(TheCU, Scope&: *Scope, Node: Label, Location: IL.second, Sym);
2088 }
2089
2090 // Collect info for retained nodes.
2091 for (const DINode *DN : SP->getRetainedNodes()) {
2092 const auto *LS = getRetainedNodeScope(N: DN);
2093 if (isa<DILocalVariable>(Val: DN) || isa<DILabel>(Val: DN)) {
2094 if (!Processed.insert(V: InlinedEntity(DN, nullptr)).second)
2095 continue;
2096 LexicalScope *LexS = LScopes.findLexicalScope(N: LS);
2097 if (LexS)
2098 createConcreteEntity(TheCU, Scope&: *LexS, Node: DN, Location: nullptr);
2099 } else {
2100 LocalDeclsPerLS[LS].insert(X: DN);
2101 }
2102 }
2103}
2104
2105// Process beginning of an instruction.
2106void DwarfDebug::beginInstruction(const MachineInstr *MI) {
2107 const MachineFunction &MF = *MI->getMF();
2108 const auto *SP = MF.getFunction().getSubprogram();
2109 bool NoDebug =
2110 !SP || SP->getUnit()->getEmissionKind() == DICompileUnit::NoDebug;
2111
2112 // Delay slot support check.
2113 auto delaySlotSupported = [](const MachineInstr &MI) {
2114 if (!MI.isBundledWithSucc())
2115 return false;
2116 auto Suc = std::next(x: MI.getIterator());
2117 (void)Suc;
2118 // Ensure that delay slot instruction is successor of the call instruction.
2119 // Ex. CALL_INSTRUCTION {
2120 // DELAY_SLOT_INSTRUCTION }
2121 assert(Suc->isBundledWithPred() &&
2122 "Call bundle instructions are out of order");
2123 return true;
2124 };
2125
2126 // When describing calls, we need a label for the call instruction.
2127 if (!NoDebug && SP->areAllCallsDescribed() &&
2128 MI->isCandidateForAdditionalCallInfo(Type: MachineInstr::AnyInBundle) &&
2129 (!MI->hasDelaySlot() || delaySlotSupported(*MI))) {
2130 const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo();
2131 bool IsTail = TII->isTailCall(Inst: *MI);
2132 // For tail calls, we need the address of the branch instruction for
2133 // DW_AT_call_pc.
2134 if (IsTail)
2135 requestLabelBeforeInsn(MI);
2136 // For non-tail calls, we need the return address for the call for
2137 // DW_AT_call_return_pc. Under GDB tuning, this information is needed for
2138 // tail calls as well.
2139 requestLabelAfterInsn(MI);
2140 }
2141
2142 DebugHandlerBase::beginInstruction(MI);
2143 if (!CurMI)
2144 return;
2145
2146 if (NoDebug)
2147 return;
2148
2149 auto RecordLineZero = [&]() {
2150 // Preserve the file and column numbers, if we can, to save space in
2151 // the encoded line table.
2152 // Do not update PrevInstLoc, it remembers the last non-0 line.
2153 const MDNode *Scope = nullptr;
2154 unsigned Column = 0;
2155 if (PrevInstLoc) {
2156 Scope = PrevInstLoc.getScope();
2157 Column = PrevInstLoc.getCol();
2158 }
2159 recordSourceLine(/*Line=*/0, Col: Column, Scope, /*Flags=*/0);
2160 };
2161
2162 // When we emit a line-0 record, we don't update PrevInstLoc; so look at
2163 // the last line number actually emitted, to see if it was line 0.
2164 unsigned LastAsmLine =
2165 Asm->OutStreamer->getContext().getCurrentDwarfLoc().getLine();
2166
2167 // Check if source location changes, but ignore DBG_VALUE and CFI locations.
2168 // If the instruction is part of the function frame setup code, do not emit
2169 // any line record, as there is no correspondence with any user code.
2170 if (MI->isMetaInstruction())
2171 return;
2172 if (MI->getFlag(Flag: MachineInstr::FrameSetup)) {
2173 // Prevent a loc from the previous block leaking into frame setup instrs.
2174 if (LastAsmLine && PrevInstBB && PrevInstBB != MI->getParent())
2175 RecordLineZero();
2176 return;
2177 }
2178
2179 const DebugLoc &DL = MI->getDebugLoc();
2180 unsigned Flags = 0;
2181
2182 if (MI->getFlag(Flag: MachineInstr::FrameDestroy) && DL) {
2183 const MachineBasicBlock *MBB = MI->getParent();
2184 if (MBB && (MBB != EpilogBeginBlock)) {
2185 // First time FrameDestroy has been seen in this basic block
2186 EpilogBeginBlock = MBB;
2187 Flags |= DWARF2_FLAG_EPILOGUE_BEGIN;
2188 }
2189 }
2190
2191 auto RecordSourceLine = [this](auto &DL, auto Flags) {
2192 SmallString<128> LocationString;
2193 if (Asm->OutStreamer->isVerboseAsm()) {
2194 raw_svector_ostream OS(LocationString);
2195 DL.print(OS);
2196 }
2197 recordSourceLine(Line: DL.getLine(), Col: DL.getCol(), Scope: DL.getScope(), Flags,
2198 Location: LocationString);
2199 };
2200
2201 // There may be a mixture of scopes using and not using Key Instructions.
2202 // Not-Key-Instructions functions inlined into Key Instructions functions
2203 // should use not-key is_stmt handling. Key Instructions functions inlined
2204 // into Not-Key-Instructions functions should use Key Instructions is_stmt
2205 // handling.
2206 bool ScopeUsesKeyInstructions =
2207 KeyInstructionsAreStmts && DL &&
2208 DL->getScope()->getSubprogram()->getKeyInstructionsEnabled();
2209
2210 bool IsKey = false;
2211 if (ScopeUsesKeyInstructions && DL && DL.getLine())
2212 IsKey = KeyInstructions.contains(V: MI);
2213
2214 if (!DL && MI == PrologEndLoc) {
2215 // In rare situations, we might want to place the end of the prologue
2216 // somewhere that doesn't have a source location already. It should be in
2217 // the entry block.
2218 assert(MI->getParent() == &*MI->getMF()->begin());
2219 recordSourceLine(Line: SP->getScopeLine(), Col: 0, Scope: SP,
2220 DWARF2_FLAG_PROLOGUE_END | DWARF2_FLAG_IS_STMT);
2221 return;
2222 }
2223
2224 bool PrevInstInSameSection =
2225 (!PrevInstBB ||
2226 PrevInstBB->getSectionID() == MI->getParent()->getSectionID());
2227 bool ForceIsStmt = ForceIsStmtInstrs.contains(V: MI);
2228 if (PrevInstInSameSection && !ForceIsStmt && DL.isSameSourceLocation(Other: PrevInstLoc)) {
2229 // If we have an ongoing unspecified location, nothing to do here.
2230 if (!DL)
2231 return;
2232
2233 // Skip this if the instruction is Key, else we might accidentally miss an
2234 // is_stmt.
2235 if (!IsKey) {
2236 // We have an explicit location, same as the previous location.
2237 // But we might be coming back to it after a line 0 record.
2238 if ((LastAsmLine == 0 && DL.getLine() != 0) || Flags) {
2239 // Reinstate the source location but not marked as a statement.
2240 RecordSourceLine(DL, Flags);
2241 }
2242 return;
2243 }
2244 }
2245
2246 if (!DL) {
2247 // FIXME: We could assert that `DL.getKind() != DebugLocKind::Temporary`
2248 // here, or otherwise record any temporary DebugLocs seen to ensure that
2249 // transient compiler-generated instructions aren't leaking their DLs to
2250 // other instructions.
2251 // We have an unspecified location, which might want to be line 0.
2252 // If we have already emitted a line-0 record, don't repeat it.
2253 if (LastAsmLine == 0)
2254 return;
2255 // If user said Don't Do That, don't do that.
2256 if (UnknownLocations == Disable)
2257 return;
2258 // See if we have a reason to emit a line-0 record now.
2259 // Reasons to emit a line-0 record include:
2260 // - User asked for it (UnknownLocations).
2261 // - Instruction has a label, so it's referenced from somewhere else,
2262 // possibly debug information; we want it to have a source location.
2263 // - Instruction is at the top of a block; we don't want to inherit the
2264 // location from the physically previous (maybe unrelated) block.
2265 if (UnknownLocations == Enable || PrevLabel ||
2266 (PrevInstBB && PrevInstBB != MI->getParent()))
2267 RecordLineZero();
2268 return;
2269 }
2270
2271 // We have an explicit location, different from the previous location.
2272 // Don't repeat a line-0 record, but otherwise emit the new location.
2273 // (The new location might be an explicit line 0, which we do emit.)
2274 if (DL.getLine() == 0 && LastAsmLine == 0)
2275 return;
2276 if (MI == PrologEndLoc) {
2277 Flags |= DWARF2_FLAG_PROLOGUE_END | DWARF2_FLAG_IS_STMT;
2278 PrologEndLoc = nullptr;
2279 }
2280
2281 if (ScopeUsesKeyInstructions) {
2282 if (IsKey)
2283 Flags |= DWARF2_FLAG_IS_STMT;
2284 } else {
2285 // If the line changed, we call that a new statement; unless we went to
2286 // line 0 and came back, in which case it is not a new statement.
2287 unsigned OldLine = PrevInstLoc ? PrevInstLoc.getLine() : LastAsmLine;
2288 if (DL.getLine() && (DL.getLine() != OldLine || ForceIsStmt))
2289 Flags |= DWARF2_FLAG_IS_STMT;
2290 }
2291
2292 // Call target-specific source line recording.
2293 recordTargetSourceLine(DL, Flags);
2294
2295 // If we're not at line 0, remember this location.
2296 if (DL.getLine())
2297 PrevInstLoc = DL;
2298}
2299
2300/// Default implementation of target-specific source line recording.
2301void DwarfDebug::recordTargetSourceLine(const DebugLoc &DL, unsigned Flags) {
2302 SmallString<128> LocationString;
2303 if (Asm->OutStreamer->isVerboseAsm()) {
2304 raw_svector_ostream OS(LocationString);
2305 DL.print(OS);
2306 }
2307 recordSourceLine(Line: DL.getLine(), Col: DL.getCol(), Scope: DL.getScope(), Flags,
2308 Location: LocationString);
2309}
2310
2311// Returns the position where we should place prologue_end, potentially nullptr,
2312// which means "no good place to put prologue_end". Returns true in the second
2313// return value if there are no setup instructions in this function at all,
2314// meaning we should not emit a start-of-function linetable entry, because it
2315// would be zero-lengthed.
2316static std::pair<const MachineInstr *, bool>
2317findPrologueEndLoc(const MachineFunction *MF) {
2318 // First known non-DBG_VALUE and non-frame setup location marks
2319 // the beginning of the function body.
2320 const auto &TII = *MF->getSubtarget().getInstrInfo();
2321 const MachineInstr *NonTrivialInst = nullptr;
2322 const Function &F = MF->getFunction();
2323 DISubprogram *SP = const_cast<DISubprogram *>(F.getSubprogram());
2324
2325 // Some instructions may be inserted into prologue after this function. Must
2326 // keep prologue for these cases.
2327 bool IsEmptyPrologue =
2328 !(F.hasPrologueData() || F.getMetadata(KindID: LLVMContext::MD_func_sanitize));
2329
2330 // Helper lambda to examine each instruction and potentially return it
2331 // as the prologue_end point.
2332 auto ExamineInst = [&](const MachineInstr &MI)
2333 -> std::optional<std::pair<const MachineInstr *, bool>> {
2334 // Is this instruction trivial data shuffling or frame-setup?
2335 bool isCopy = (TII.isCopyInstr(MI) ? true : false);
2336 bool isTrivRemat = TII.isTriviallyReMaterializable(MI);
2337 bool isFrameSetup = MI.getFlag(Flag: MachineInstr::FrameSetup);
2338
2339 if (!isFrameSetup && MI.getDebugLoc()) {
2340 // Scan forward to try to find a non-zero line number. The
2341 // prologue_end marks the first breakpoint in the function after the
2342 // frame setup, and a compiler-generated line 0 location is not a
2343 // meaningful breakpoint. If none is found, return the first
2344 // location after the frame setup.
2345 if (MI.getDebugLoc().getLine())
2346 return std::make_pair(x: &MI, y&: IsEmptyPrologue);
2347 }
2348
2349 // Keep track of the first "non-trivial" instruction seen, i.e. anything
2350 // that doesn't involve shuffling data around or is a frame-setup.
2351 if (!isCopy && !isTrivRemat && !isFrameSetup && !NonTrivialInst)
2352 NonTrivialInst = &MI;
2353
2354 IsEmptyPrologue = false;
2355 return std::nullopt;
2356 };
2357
2358 // Examine all the instructions at the start of the function. This doesn't
2359 // necessarily mean just the entry block: unoptimised code can fall-through
2360 // into an initial loop, and it makes sense to put the initial breakpoint on
2361 // the first instruction of such a loop. However, if we pass branches, we're
2362 // better off synthesising an early prologue_end.
2363 auto CurBlock = MF->begin();
2364 auto CurInst = CurBlock->begin();
2365
2366 // Find the initial instruction, we're guaranteed one by the caller, but not
2367 // which block it's in.
2368 while (CurBlock->empty())
2369 CurInst = (++CurBlock)->begin();
2370 assert(CurInst != CurBlock->end());
2371
2372 // Helper function for stepping through the initial sequence of
2373 // unconditionally executed instructions.
2374 auto getNextInst = [&CurBlock, &CurInst, MF]() -> bool {
2375 // We've reached the end of the block. Did we just look at a terminator?
2376 if (CurInst->isTerminator()) {
2377 // Some kind of "real" control flow is occurring. At the very least
2378 // we would have to start exploring the CFG, a good signal that the
2379 // prologue is over.
2380 return false;
2381 }
2382
2383 // If we've already fallen through into a loop, don't fall through
2384 // further, use a backup-location.
2385 if (CurBlock->pred_size() > 1)
2386 return false;
2387
2388 // Fall-through from entry to the next block. This is common at -O0 when
2389 // there's no initialisation in the function. Bail if we're also at the
2390 // end of the function, or the remaining blocks have no instructions.
2391 // Skip empty blocks, in rare cases the entry can be empty, and
2392 // other optimisations may add empty blocks that the control flow falls
2393 // through.
2394 do {
2395 ++CurBlock;
2396 if (CurBlock == MF->end())
2397 return false;
2398 } while (CurBlock->empty());
2399 CurInst = CurBlock->begin();
2400 return true;
2401 };
2402
2403 while (true) {
2404 // Check whether this non-meta instruction a good position for prologue_end.
2405 if (!CurInst->isMetaInstruction()) {
2406 auto FoundInst = ExamineInst(*CurInst);
2407 if (FoundInst)
2408 return *FoundInst;
2409 }
2410
2411 // In very rare scenarios function calls can have line zero, and we
2412 // shouldn't step over such a call while trying to reach prologue_end. In
2413 // these extraordinary conditions, force the call to have the scope line
2414 // and put prologue_end there. This isn't ideal, but signals that the call
2415 // is where execution in the function starts, and is less catastrophic than
2416 // stepping over the call.
2417 if (CurInst->isCall()) {
2418 if (const DILocation *Loc = CurInst->getDebugLoc().get();
2419 Loc && Loc->getLine() == 0) {
2420 // Create and assign the scope-line position.
2421 unsigned ScopeLine = SP->getScopeLine();
2422 DILocation *ScopeLineDILoc =
2423 DILocation::get(Context&: SP->getContext(), Line: ScopeLine, Column: 0, Scope: SP);
2424 const_cast<MachineInstr *>(&*CurInst)->setDebugLoc(ScopeLineDILoc);
2425
2426 // Consider this position to be where prologue_end is placed.
2427 return std::make_pair(x: &*CurInst, y: false);
2428 }
2429 }
2430
2431 // Try to continue searching, but use a backup-location if substantive
2432 // computation is happening.
2433 auto NextInst = std::next(x: CurInst);
2434 if (NextInst != CurInst->getParent()->end()) {
2435 // Continue examining the current block.
2436 CurInst = NextInst;
2437 continue;
2438 }
2439
2440 if (!getNextInst())
2441 break;
2442 }
2443
2444 // We couldn't find any source-location, suggesting all meaningful information
2445 // got optimised away. Set the prologue_end to be the first non-trivial
2446 // instruction, which will get the scope line number. This is better than
2447 // nothing.
2448 // Only do this in the entry block, as we'll be giving it the scope line for
2449 // the function. Return IsEmptyPrologue==true if we've picked the first
2450 // instruction.
2451 if (NonTrivialInst && NonTrivialInst->getParent() == &*MF->begin()) {
2452 IsEmptyPrologue = NonTrivialInst == &*MF->begin()->begin();
2453 return std::make_pair(x&: NonTrivialInst, y&: IsEmptyPrologue);
2454 }
2455
2456 // If the entry path is empty, just don't have a prologue_end at all.
2457 return std::make_pair(x: nullptr, y&: IsEmptyPrologue);
2458}
2459
2460/// Register a source line with debug info. Returns the unique label that was
2461/// emitted and which provides correspondence to the source line list.
2462static void recordSourceLine(AsmPrinter &Asm, unsigned Line, unsigned Col,
2463 const MDNode *S, unsigned Flags, unsigned CUID,
2464 uint16_t DwarfVersion,
2465 ArrayRef<std::unique_ptr<DwarfCompileUnit>> DCUs,
2466 StringRef Comment = {}) {
2467 StringRef Fn;
2468 unsigned FileNo = 1;
2469 unsigned Discriminator = 0;
2470 if (auto *Scope = cast_or_null<DIScope>(Val: S)) {
2471 Fn = Scope->getFilename();
2472 if (Line != 0 && DwarfVersion >= 4)
2473 if (auto *LBF = dyn_cast<DILexicalBlockFile>(Val: Scope))
2474 Discriminator = LBF->getDiscriminator();
2475
2476 FileNo = static_cast<DwarfCompileUnit &>(*DCUs[CUID])
2477 .getOrCreateSourceID(File: Scope->getFile());
2478 }
2479 Asm.OutStreamer->emitDwarfLocDirective(FileNo, Line, Column: Col, Flags, Isa: 0,
2480 Discriminator, FileName: Fn, Comment);
2481}
2482
2483const MachineInstr *
2484DwarfDebug::emitInitialLocDirective(const MachineFunction &MF, unsigned CUID) {
2485 // Don't deal with functions that have no instructions.
2486 if (llvm::all_of(Range: MF, P: [](const MachineBasicBlock &MBB) { return MBB.empty(); }))
2487 return nullptr;
2488
2489 std::pair<const MachineInstr *, bool> PrologEnd = findPrologueEndLoc(MF: &MF);
2490 const MachineInstr *PrologEndLoc = PrologEnd.first;
2491 bool IsEmptyPrologue = PrologEnd.second;
2492
2493 // If the prolog is empty, no need to generate scope line for the proc.
2494 if (IsEmptyPrologue) {
2495 // If there's nowhere to put a prologue_end flag, emit a scope line in case
2496 // there are simply no source locations anywhere in the function.
2497 if (PrologEndLoc) {
2498 // Avoid trying to assign prologue_end to a line-zero location.
2499 // Instructions with no DebugLoc at all are fine, they'll be given the
2500 // scope line nuumber.
2501 const DebugLoc &DL = PrologEndLoc->getDebugLoc();
2502 if (!DL || DL->getLine() != 0)
2503 return PrologEndLoc;
2504
2505 // Later, don't place the prologue_end flag on this line-zero location.
2506 PrologEndLoc = nullptr;
2507 }
2508 }
2509
2510 // Ensure the compile unit is created if the function is called before
2511 // beginFunction().
2512 DISubprogram *SP = MF.getFunction().getSubprogram();
2513 (void)getOrCreateDwarfCompileUnit(DIUnit: SP->getUnit());
2514 // We'd like to list the prologue as "not statements" but GDB behaves
2515 // poorly if we do that. Revisit this with caution/GDB (7.5+) testing.
2516 ::recordSourceLine(Asm&: *Asm, Line: SP->getScopeLine(), Col: 0, S: SP, DWARF2_FLAG_IS_STMT,
2517 CUID, DwarfVersion: getDwarfVersion(), DCUs: getUnits());
2518 return PrologEndLoc;
2519}
2520
2521void DwarfDebug::computeKeyInstructions(const MachineFunction *MF) {
2522 // New function - reset KeyInstructions.
2523 KeyInstructions.clear();
2524
2525 // The current candidate is_stmt instructions for each source atom.
2526 // Map {(InlinedAt, Group): (Rank, Instructions)}.
2527 // NOTE: Anecdotally, for a large C++ blob, 99% of the instruction
2528 // SmallVectors contain 2 or fewer elements; use 2 inline elements.
2529 DenseMap<std::pair<DILocation *, uint64_t>,
2530 std::pair<uint8_t, SmallVector<const MachineInstr *, 2>>>
2531 GroupCandidates;
2532
2533 const auto &TII = *MF->getSubtarget().getInstrInfo();
2534
2535 // For each instruction:
2536 // * Skip insts without DebugLoc, AtomGroup or AtomRank, and line zeros.
2537 // * Check if insts in this group have been seen already in GroupCandidates.
2538 // * If this instr rank is equal, add this instruction to GroupCandidates.
2539 // Remove existing instructions from GroupCandidates if they have the
2540 // same parent.
2541 // * If this instr rank is higher (lower precedence), ignore it.
2542 // * If this instr rank is lower (higher precedence), erase existing
2543 // instructions from GroupCandidates and add this one.
2544 //
2545 // Then insert each GroupCandidates instruction into KeyInstructions.
2546
2547 for (auto &MBB : *MF) {
2548 // Rather than apply is_stmt directly to Key Instructions, we "float"
2549 // is_stmt up to the 1st instruction with the same line number in a
2550 // contiguous block. That instruction is called the "buoy". The
2551 // buoy gets reset if we encouner an instruction with an atom
2552 // group.
2553 const MachineInstr *Buoy = nullptr;
2554 // The atom group number associated with Buoy which may be 0 if we haven't
2555 // encountered an atom group yet in this blob of instructions with the same
2556 // line number.
2557 uint64_t BuoyAtom = 0;
2558
2559 for (auto &MI : MBB) {
2560 if (MI.isMetaInstruction())
2561 continue;
2562
2563 const DILocation *Loc = MI.getDebugLoc().get();
2564 if (!Loc || !Loc->getLine())
2565 continue;
2566
2567 // Reset the Buoy to this instruction if it has a different line number.
2568 if (!Buoy || Buoy->getDebugLoc().getLine() != Loc->getLine()) {
2569 Buoy = &MI;
2570 BuoyAtom = 0; // Set later when we know which atom the buoy is used by.
2571 }
2572
2573 // Call instructions are handled specially - we always mark them as key
2574 // regardless of atom info.
2575 bool IsCallLike = MI.isCall() || TII.isTailCall(Inst: MI);
2576 if (IsCallLike) {
2577 // Calls are always key. Put the buoy (may not be the call) into
2578 // KeyInstructions directly rather than the candidate map to avoid it
2579 // being erased (and we may not have a group number for the call).
2580 KeyInstructions.insert(V: Buoy);
2581
2582 // Avoid floating any future is_stmts up to the call.
2583 Buoy = nullptr;
2584 BuoyAtom = 0;
2585
2586 if (!Loc->getAtomGroup() || !Loc->getAtomRank())
2587 continue;
2588 }
2589
2590 auto *InlinedAt = Loc->getInlinedAt();
2591 uint64_t Group = Loc->getAtomGroup();
2592 uint8_t Rank = Loc->getAtomRank();
2593 if (!Group || !Rank)
2594 continue;
2595
2596 // Don't let is_stmts float past instructions from different source atoms.
2597 if (BuoyAtom && BuoyAtom != Group) {
2598 Buoy = &MI;
2599 BuoyAtom = Group;
2600 }
2601
2602 auto &[CandidateRank, CandidateInsts] =
2603 GroupCandidates[{InlinedAt, Group}];
2604
2605 // If CandidateRank is zero then CandidateInsts should be empty: there
2606 // are no other candidates for this group yet. If CandidateRank is nonzero
2607 // then CandidateInsts shouldn't be empty: we've got existing candidate
2608 // instructions.
2609 assert((CandidateRank == 0 && CandidateInsts.empty()) ||
2610 (CandidateRank != 0 && !CandidateInsts.empty()));
2611
2612 assert(Rank && "expected nonzero rank");
2613 // If we've seen other instructions in this group with higher precedence
2614 // (lower nonzero rank), don't add this one as a candidate.
2615 if (CandidateRank && CandidateRank < Rank)
2616 continue;
2617
2618 // If we've seen other instructions in this group of the same rank,
2619 // discard any from this block (keeping the others). Else if we've
2620 // seen other instructions in this group of lower precedence (higher
2621 // rank), discard them all.
2622 if (CandidateRank == Rank)
2623 llvm::remove_if(Range&: CandidateInsts, P: [&MI](const MachineInstr *Candidate) {
2624 return MI.getParent() == Candidate->getParent();
2625 });
2626 else if (CandidateRank > Rank)
2627 CandidateInsts.clear();
2628
2629 if (Buoy) {
2630 // Add this candidate.
2631 CandidateInsts.push_back(Elt: Buoy);
2632 CandidateRank = Rank;
2633
2634 assert(!BuoyAtom || BuoyAtom == Loc->getAtomGroup());
2635 BuoyAtom = Loc->getAtomGroup();
2636 } else {
2637 // Don't add calls, because they've been dealt with already. This means
2638 // CandidateInsts might now be empty - handle that.
2639 assert(IsCallLike);
2640 if (CandidateInsts.empty())
2641 CandidateRank = 0;
2642 }
2643 }
2644 }
2645
2646 for (const auto &[_, Insts] : GroupCandidates.values())
2647 for (auto *I : Insts)
2648 KeyInstructions.insert(V: I);
2649}
2650
2651/// For the function \p MF, finds the set of instructions which may represent a
2652/// change in line number from one or more of the preceding MBBs. Stores the
2653/// resulting set of instructions, which should have is_stmt set, in
2654/// ForceIsStmtInstrs.
2655void DwarfDebug::findForceIsStmtInstrs(const MachineFunction *MF) {
2656 ForceIsStmtInstrs.clear();
2657
2658 // For this function, we try to find MBBs where the last source line in every
2659 // block predecessor matches the first line seen in the block itself; for
2660 // every such MBB, we set is_stmt=false on the first line in the block, and
2661 // for every other block we set is_stmt=true on the first line.
2662 // For example, if we have the block %bb.3, which has 2 predecesors %bb.1 and
2663 // %bb.2:
2664 // bb.1:
2665 // $r3 = MOV64ri 12, debug-location !DILocation(line: 4)
2666 // JMP %bb.3, debug-location !DILocation(line: 5)
2667 // bb.2:
2668 // $r3 = MOV64ri 24, debug-location !DILocation(line: 5)
2669 // JMP %bb.3
2670 // bb.3:
2671 // $r2 = MOV64ri 1
2672 // $r1 = ADD $r2, $r3, debug-location !DILocation(line: 5)
2673 // When we examine %bb.3, we first check to see if it contains any
2674 // instructions with debug locations, and select the first such instruction;
2675 // in this case, the ADD, with line=5. We then examine both of its
2676 // predecessors to see what the last debug-location in them is. For each
2677 // predecessor, if they do not contain any debug-locations, or if the last
2678 // debug-location before jumping to %bb.3 does not have line=5, then the ADD
2679 // in %bb.3 must use IsStmt. In this case, all predecessors have a
2680 // debug-location with line=5 as the last debug-location before jumping to
2681 // %bb.3, so we do not set is_stmt for the ADD instruction - we know that
2682 // whichever MBB we have arrived from, the line has not changed.
2683
2684 const auto *TII = MF->getSubtarget().getInstrInfo();
2685
2686 // We only need to the predecessors of MBBs that could have is_stmt set by
2687 // this logic.
2688 SmallDenseSet<MachineBasicBlock *, 4> PredMBBsToExamine;
2689 SmallDenseMap<MachineBasicBlock *, MachineInstr *> PotentialIsStmtMBBInstrs;
2690 // We use const_cast even though we won't actually modify MF, because some
2691 // methods we need take a non-const MBB.
2692 for (auto &MBB : *const_cast<MachineFunction *>(MF)) {
2693 if (MBB.empty() || MBB.pred_empty())
2694 continue;
2695 for (auto &MI : MBB) {
2696 if (MI.getDebugLoc() && MI.getDebugLoc()->getLine()) {
2697 PredMBBsToExamine.insert_range(R: MBB.predecessors());
2698 PotentialIsStmtMBBInstrs.insert(KV: {&MBB, &MI});
2699 break;
2700 }
2701 }
2702 }
2703
2704 // For each predecessor MBB, we examine the last line seen before each branch
2705 // or logical fallthrough. We use analyzeBranch to handle cases where
2706 // different branches have different outgoing lines (i.e. if there are
2707 // multiple branches that each have their own source location); otherwise we
2708 // just use the last line in the block.
2709 for (auto *MBB : PredMBBsToExamine) {
2710 auto CheckMBBEdge = [&](MachineBasicBlock *Succ, unsigned OutgoingLine) {
2711 auto MBBInstrIt = PotentialIsStmtMBBInstrs.find(Val: Succ);
2712 if (MBBInstrIt == PotentialIsStmtMBBInstrs.end())
2713 return;
2714 MachineInstr *MI = MBBInstrIt->second;
2715 if (MI->getDebugLoc()->getLine() == OutgoingLine)
2716 return;
2717 PotentialIsStmtMBBInstrs.erase(I: MBBInstrIt);
2718 ForceIsStmtInstrs.insert(V: MI);
2719 };
2720 // If this block is empty, we conservatively assume that its fallthrough
2721 // successor needs is_stmt; we could check MBB's predecessors to see if it
2722 // has a consistent entry line, but this seems unlikely to be worthwhile.
2723 if (MBB->empty()) {
2724 for (auto *Succ : MBB->successors())
2725 CheckMBBEdge(Succ, 0);
2726 continue;
2727 }
2728 // If MBB has no successors that are in the "potential" set, due to one or
2729 // more of them having confirmed is_stmt, we can skip this check early.
2730 if (none_of(Range: MBB->successors(), P: [&](auto *SuccMBB) {
2731 return PotentialIsStmtMBBInstrs.contains(Val: SuccMBB);
2732 }))
2733 continue;
2734 // If we can't determine what DLs this branch's successors use, just treat
2735 // all the successors as coming from the last DebugLoc.
2736 SmallVector<MachineBasicBlock *, 2> SuccessorBBs;
2737 auto MIIt = MBB->rbegin();
2738 {
2739 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
2740 SmallVector<MachineOperand, 4> Cond;
2741 bool AnalyzeFailed = TII->analyzeBranch(MBB&: *MBB, TBB, FBB, Cond);
2742 // For a conditional branch followed by unconditional branch where the
2743 // unconditional branch has a DebugLoc, that loc is the outgoing loc to
2744 // the the false destination only; otherwise, both destinations share an
2745 // outgoing loc.
2746 if (!AnalyzeFailed && !Cond.empty() && FBB != nullptr &&
2747 MBB->back().getDebugLoc() && MBB->back().getDebugLoc()->getLine()) {
2748 unsigned FBBLine = MBB->back().getDebugLoc()->getLine();
2749 assert(MIIt->isBranch() && "Bad result from analyzeBranch?");
2750 CheckMBBEdge(FBB, FBBLine);
2751 ++MIIt;
2752 SuccessorBBs.push_back(Elt: TBB);
2753 } else {
2754 // For all other cases, all successors share the last outgoing DebugLoc.
2755 SuccessorBBs.assign(in_start: MBB->succ_begin(), in_end: MBB->succ_end());
2756 }
2757 }
2758
2759 // If we don't find an outgoing loc, this block will start with a line 0.
2760 // It is possible that we have a block that has no DebugLoc, but acts as a
2761 // simple passthrough between two blocks that end and start with the same
2762 // line, e.g.:
2763 // bb.1:
2764 // JMP %bb.2, debug-location !10
2765 // bb.2:
2766 // JMP %bb.3
2767 // bb.3:
2768 // $r1 = ADD $r2, $r3, debug-location !10
2769 // If these blocks were merged into a single block, we would not attach
2770 // is_stmt to the ADD, but with this logic that only checks the immediate
2771 // predecessor, we will; we make this tradeoff because doing a full dataflow
2772 // analysis would be expensive, and these situations are probably not common
2773 // enough for this to be worthwhile.
2774 unsigned LastLine = 0;
2775 while (MIIt != MBB->rend()) {
2776 if (auto DL = MIIt->getDebugLoc(); DL && DL->getLine()) {
2777 LastLine = DL->getLine();
2778 break;
2779 }
2780 ++MIIt;
2781 }
2782 for (auto *Succ : SuccessorBBs)
2783 CheckMBBEdge(Succ, LastLine);
2784 }
2785}
2786
2787// Gather pre-function debug information. Assumes being called immediately
2788// after the function entry point has been emitted.
2789void DwarfDebug::beginFunctionImpl(const MachineFunction *MF) {
2790 CurFn = MF;
2791
2792 auto *SP = MF->getFunction().getSubprogram();
2793 assert(LScopes.empty() || SP == LScopes.getCurrentFunctionScope()->getScopeNode());
2794 if (SP->getUnit()->getEmissionKind() == DICompileUnit::NoDebug)
2795 return;
2796
2797 DwarfCompileUnit &CU = getOrCreateDwarfCompileUnit(DIUnit: SP->getUnit());
2798 FunctionLineTableLabel = CU.emitFuncLineTableOffsets()
2799 ? Asm->OutStreamer->emitLineTableLabel()
2800 : nullptr;
2801
2802 Asm->OutStreamer->getContext().setDwarfCompileUnitID(
2803 getDwarfCompileUnitIDForLineTable(CU));
2804
2805 // Call target-specific debug info initialization.
2806 initializeTargetDebugInfo(MF: *MF);
2807
2808 // Record beginning of function.
2809 PrologEndLoc = emitInitialLocDirective(
2810 MF: *MF, CUID: Asm->OutStreamer->getContext().getDwarfCompileUnitID());
2811
2812 // Run both `findForceIsStmtInstrs` and `computeKeyInstructions` because
2813 // Not-Key-Instructions functions may be inlined into Key Instructions
2814 // functions and vice versa.
2815 if (KeyInstructionsAreStmts)
2816 computeKeyInstructions(MF);
2817 findForceIsStmtInstrs(MF);
2818}
2819
2820unsigned
2821DwarfDebug::getDwarfCompileUnitIDForLineTable(const DwarfCompileUnit &CU) {
2822 // Set DwarfDwarfCompileUnitID in MCContext to the Compile Unit this function
2823 // belongs to so that we add to the correct per-cu line table in the
2824 // non-asm case.
2825 if (Asm->OutStreamer->hasRawTextSupport())
2826 // Use a single line table if we are generating assembly.
2827 return 0;
2828 else
2829 return CU.getUniqueID();
2830}
2831
2832void DwarfDebug::terminateLineTable(const DwarfCompileUnit *CU) {
2833 const auto &CURanges = CU->getRanges();
2834 auto &LineTable = Asm->OutStreamer->getContext().getMCDwarfLineTable(
2835 CUID: getDwarfCompileUnitIDForLineTable(CU: *CU));
2836 // Add the last range label for the given CU.
2837 LineTable.getMCLineSections().addEndEntry(
2838 EndLabel: const_cast<MCSymbol *>(CURanges.back().End));
2839}
2840
2841void DwarfDebug::skippedNonDebugFunction() {
2842 // If we don't have a subprogram for this function then there will be a hole
2843 // in the range information. Keep note of this by setting the previously used
2844 // section to nullptr.
2845 // Terminate the pending line table.
2846 if (PrevCU)
2847 terminateLineTable(CU: PrevCU);
2848 PrevCU = nullptr;
2849 CurFn = nullptr;
2850}
2851
2852// Gather and emit post-function debug information.
2853void DwarfDebug::endFunctionImpl(const MachineFunction *MF) {
2854 const Function &F = MF->getFunction();
2855 const DISubprogram *SP = F.getSubprogram();
2856
2857 assert(CurFn == MF &&
2858 "endFunction should be called with the same function as beginFunction");
2859
2860 // Set DwarfDwarfCompileUnitID in MCContext to default value.
2861 Asm->OutStreamer->getContext().setDwarfCompileUnitID(0);
2862
2863 LexicalScope *FnScope = LScopes.getCurrentFunctionScope();
2864 assert(!FnScope || SP == FnScope->getScopeNode());
2865 DwarfCompileUnit &TheCU = getOrCreateDwarfCompileUnit(DIUnit: SP->getUnit());
2866 if (TheCU.getCUNode()->isDebugDirectivesOnly()) {
2867 PrevLabel = nullptr;
2868 CurFn = nullptr;
2869 return;
2870 }
2871
2872 DenseSet<InlinedEntity> Processed;
2873 collectEntityInfo(TheCU, SP, Processed);
2874
2875 // Add the range of this function to the list of ranges for the CU.
2876 // With basic block sections, add ranges for all basic block sections.
2877 for (const auto &R : Asm->MBBSectionRanges)
2878 TheCU.addRange(Range: {.Begin: R.second.BeginLabel, .End: R.second.EndLabel});
2879
2880 // Under -gmlt, skip building the subprogram if there are no inlined
2881 // subroutines inside it. But with -fdebug-info-for-profiling, the subprogram
2882 // is still needed as we need its source location.
2883 if (!TheCU.getCUNode()->getDebugInfoForProfiling() &&
2884 TheCU.getCUNode()->getEmissionKind() == DICompileUnit::LineTablesOnly &&
2885 LScopes.getAbstractScopesList().empty() && !IsDarwin) {
2886 for (const auto &R : Asm->MBBSectionRanges)
2887 addArangeLabel(SCU: SymbolCU(&TheCU, R.second.BeginLabel));
2888
2889 assert(InfoHolder.getScopeVariables().empty());
2890 PrevLabel = nullptr;
2891 CurFn = nullptr;
2892 return;
2893 }
2894
2895#ifndef NDEBUG
2896 size_t NumAbstractSubprograms = LScopes.getAbstractScopesList().size();
2897#endif
2898 for (LexicalScope *AScope : LScopes.getAbstractScopesList()) {
2899 const auto *SP = cast<DISubprogram>(Val: AScope->getScopeNode());
2900 for (const DINode *DN : SP->getRetainedNodes()) {
2901 const auto *LS = getRetainedNodeScope(N: DN);
2902 // Ensure LexicalScope is created for the scope of this node.
2903 auto *LexS = LScopes.getOrCreateAbstractScope(Scope: LS);
2904 assert(LexS && "Expected the LexicalScope to be created.");
2905 if (isa<DILocalVariable>(Val: DN) || isa<DILabel>(Val: DN)) {
2906 // Collect info for variables/labels that were optimized out.
2907 if (!Processed.insert(V: InlinedEntity(DN, nullptr)).second ||
2908 TheCU.getExistingAbstractEntity(Node: DN))
2909 continue;
2910 TheCU.createAbstractEntity(Node: DN, Scope: LexS);
2911 } else {
2912 // Remember the node if this is a local declarations.
2913 LocalDeclsPerLS[LS].insert(X: DN);
2914 }
2915 assert(
2916 LScopes.getAbstractScopesList().size() == NumAbstractSubprograms &&
2917 "getOrCreateAbstractScope() inserted an abstract subprogram scope");
2918 }
2919 constructAbstractSubprogramScopeDIE(SrcCU&: TheCU, Scope: AScope);
2920 }
2921
2922 ProcessedSPNodes.insert(X: SP);
2923 DIE &ScopeDIE =
2924 TheCU.constructSubprogramScopeDIE(Sub: SP, F, Scope: FnScope, LineTableSym: FunctionLineTableLabel);
2925 if (auto *SkelCU = TheCU.getSkeleton())
2926 if (!LScopes.getAbstractScopesList().empty() &&
2927 TheCU.getCUNode()->getSplitDebugInlining())
2928 SkelCU->constructSubprogramScopeDIE(Sub: SP, F, Scope: FnScope,
2929 LineTableSym: FunctionLineTableLabel);
2930
2931 FunctionLineTableLabel = nullptr;
2932
2933 // Construct call site entries.
2934 constructCallSiteEntryDIEs(SP: *SP, CU&: TheCU, ScopeDIE, MF: *MF);
2935
2936 // Clear debug info
2937 // Ownership of DbgVariables is a bit subtle - ScopeVariables owns all the
2938 // DbgVariables except those that are also in AbstractVariables (since they
2939 // can be used cross-function)
2940 InfoHolder.getScopeVariables().clear();
2941 InfoHolder.getScopeLabels().clear();
2942 LocalDeclsPerLS.clear();
2943 PrevLabel = nullptr;
2944 CurFn = nullptr;
2945}
2946
2947// Register a source line with debug info. Returns the unique label that was
2948// emitted and which provides correspondence to the source line list.
2949void DwarfDebug::recordSourceLine(unsigned Line, unsigned Col, const MDNode *S,
2950 unsigned Flags, StringRef Location) {
2951 ::recordSourceLine(Asm&: *Asm, Line, Col, S, Flags,
2952 CUID: Asm->OutStreamer->getContext().getDwarfCompileUnitID(),
2953 DwarfVersion: getDwarfVersion(), DCUs: getUnits(), Comment: Location);
2954}
2955
2956//===----------------------------------------------------------------------===//
2957// Emit Methods
2958//===----------------------------------------------------------------------===//
2959
2960// Emit the debug info section.
2961void DwarfDebug::emitDebugInfo() {
2962 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder;
2963 Holder.emitUnits(/* UseOffsets */ false);
2964}
2965
2966// Emit the abbreviation section.
2967void DwarfDebug::emitAbbreviations() {
2968 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder;
2969
2970 Holder.emitAbbrevs(Asm->getObjFileLowering().getDwarfAbbrevSection());
2971}
2972
2973void DwarfDebug::emitStringOffsetsTableHeader() {
2974 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder;
2975 Holder.getStringPool().emitStringOffsetsTableHeader(
2976 Asm&: *Asm, OffsetSection: Asm->getObjFileLowering().getDwarfStrOffSection(),
2977 StartSym: Holder.getStringOffsetsStartSym());
2978}
2979
2980template <typename AccelTableT>
2981void DwarfDebug::emitAccel(AccelTableT &Accel, MCSection *Section,
2982 StringRef TableName) {
2983 Asm->OutStreamer->switchSection(Section);
2984
2985 // Emit the full data.
2986 emitAppleAccelTable(Asm, Accel, TableName, Section->getBeginSymbol());
2987}
2988
2989void DwarfDebug::emitAccelDebugNames() {
2990 // Don't emit anything if we have no compilation units to index.
2991 if (getUnits().empty())
2992 return;
2993
2994 emitDWARF5AccelTable(Asm, Contents&: AccelDebugNames, DD: *this, CUs: getUnits());
2995}
2996
2997// Emit visible names into a hashed accelerator table section.
2998void DwarfDebug::emitAccelNames() {
2999 emitAccel(Accel&: AccelNames, Section: Asm->getObjFileLowering().getDwarfAccelNamesSection(),
3000 TableName: "Names");
3001}
3002
3003// Emit objective C classes and categories into a hashed accelerator table
3004// section.
3005void DwarfDebug::emitAccelObjC() {
3006 emitAccel(Accel&: AccelObjC, Section: Asm->getObjFileLowering().getDwarfAccelObjCSection(),
3007 TableName: "ObjC");
3008}
3009
3010// Emit namespace dies into a hashed accelerator table.
3011void DwarfDebug::emitAccelNamespaces() {
3012 emitAccel(Accel&: AccelNamespace,
3013 Section: Asm->getObjFileLowering().getDwarfAccelNamespaceSection(),
3014 TableName: "namespac");
3015}
3016
3017// Emit type dies into a hashed accelerator table.
3018void DwarfDebug::emitAccelTypes() {
3019 emitAccel(Accel&: AccelTypes, Section: Asm->getObjFileLowering().getDwarfAccelTypesSection(),
3020 TableName: "types");
3021}
3022
3023// Public name handling.
3024// The format for the various pubnames:
3025//
3026// dwarf pubnames - offset/name pairs where the offset is the offset into the CU
3027// for the DIE that is named.
3028//
3029// gnu pubnames - offset/index value/name tuples where the offset is the offset
3030// into the CU and the index value is computed according to the type of value
3031// for the DIE that is named.
3032//
3033// For type units the offset is the offset of the skeleton DIE. For split dwarf
3034// it's the offset within the debug_info/debug_types dwo section, however, the
3035// reference in the pubname header doesn't change.
3036
3037/// computeIndexValue - Compute the gdb index value for the DIE and CU.
3038static dwarf::PubIndexEntryDescriptor computeIndexValue(DwarfUnit *CU,
3039 const DIE *Die) {
3040 // Entities that ended up only in a Type Unit reference the CU instead (since
3041 // the pub entry has offsets within the CU there's no real offset that can be
3042 // provided anyway). As it happens all such entities (namespaces and types,
3043 // types only in C++ at that) are rendered as TYPE+EXTERNAL. If this turns out
3044 // not to be true it would be necessary to persist this information from the
3045 // point at which the entry is added to the index data structure - since by
3046 // the time the index is built from that, the original type/namespace DIE in a
3047 // type unit has already been destroyed so it can't be queried for properties
3048 // like tag, etc.
3049 if (Die->getTag() == dwarf::DW_TAG_compile_unit)
3050 return dwarf::PubIndexEntryDescriptor(dwarf::GIEK_TYPE,
3051 dwarf::GIEL_EXTERNAL);
3052 dwarf::GDBIndexEntryLinkage Linkage = dwarf::GIEL_STATIC;
3053
3054 // We could have a specification DIE that has our most of our knowledge,
3055 // look for that now.
3056 if (DIEValue SpecVal = Die->findAttribute(Attribute: dwarf::DW_AT_specification)) {
3057 DIE &SpecDIE = SpecVal.getDIEEntry().getEntry();
3058 if (SpecDIE.findAttribute(Attribute: dwarf::DW_AT_external))
3059 Linkage = dwarf::GIEL_EXTERNAL;
3060 } else if (Die->findAttribute(Attribute: dwarf::DW_AT_external))
3061 Linkage = dwarf::GIEL_EXTERNAL;
3062
3063 switch (Die->getTag()) {
3064 case dwarf::DW_TAG_class_type:
3065 case dwarf::DW_TAG_structure_type:
3066 case dwarf::DW_TAG_union_type:
3067 case dwarf::DW_TAG_enumeration_type:
3068 return dwarf::PubIndexEntryDescriptor(
3069 dwarf::GIEK_TYPE, dwarf::isCPlusPlus(S: CU->getSourceLanguage())
3070 ? dwarf::GIEL_EXTERNAL
3071 : dwarf::GIEL_STATIC);
3072 case dwarf::DW_TAG_typedef:
3073 case dwarf::DW_TAG_base_type:
3074 case dwarf::DW_TAG_subrange_type:
3075 case dwarf::DW_TAG_template_alias:
3076 return dwarf::PubIndexEntryDescriptor(dwarf::GIEK_TYPE, dwarf::GIEL_STATIC);
3077 case dwarf::DW_TAG_namespace:
3078 return dwarf::GIEK_TYPE;
3079 case dwarf::DW_TAG_subprogram:
3080 return dwarf::PubIndexEntryDescriptor(dwarf::GIEK_FUNCTION, Linkage);
3081 case dwarf::DW_TAG_variable:
3082 return dwarf::PubIndexEntryDescriptor(dwarf::GIEK_VARIABLE, Linkage);
3083 case dwarf::DW_TAG_enumerator:
3084 return dwarf::PubIndexEntryDescriptor(dwarf::GIEK_VARIABLE,
3085 dwarf::GIEL_STATIC);
3086 default:
3087 return dwarf::GIEK_NONE;
3088 }
3089}
3090
3091/// emitDebugPubSections - Emit visible names and types into debug pubnames and
3092/// pubtypes sections.
3093void DwarfDebug::emitDebugPubSections() {
3094 for (const auto &NU : CUMap) {
3095 DwarfCompileUnit *TheU = NU.second;
3096 if (!TheU->hasDwarfPubSections())
3097 continue;
3098
3099 bool GnuStyle = TheU->getCUNode()->getNameTableKind() ==
3100 DICompileUnit::DebugNameTableKind::GNU;
3101
3102 Asm->OutStreamer->switchSection(
3103 Section: GnuStyle ? Asm->getObjFileLowering().getDwarfGnuPubNamesSection()
3104 : Asm->getObjFileLowering().getDwarfPubNamesSection());
3105 emitDebugPubSection(GnuStyle, Name: "Names", TheU, Globals: TheU->getGlobalNames());
3106
3107 Asm->OutStreamer->switchSection(
3108 Section: GnuStyle ? Asm->getObjFileLowering().getDwarfGnuPubTypesSection()
3109 : Asm->getObjFileLowering().getDwarfPubTypesSection());
3110 emitDebugPubSection(GnuStyle, Name: "Types", TheU, Globals: TheU->getGlobalTypes());
3111 }
3112}
3113
3114void DwarfDebug::emitSectionReference(const DwarfCompileUnit &CU) {
3115 if (useSectionsAsReferences())
3116 Asm->emitDwarfOffset(Label: CU.getSection()->getBeginSymbol(),
3117 Offset: CU.getDebugSectionOffset());
3118 else
3119 Asm->emitDwarfSymbolReference(Label: CU.getLabelBegin());
3120}
3121
3122void DwarfDebug::emitDebugPubSection(bool GnuStyle, StringRef Name,
3123 DwarfCompileUnit *TheU,
3124 const StringMap<const DIE *> &Globals) {
3125 if (auto *Skeleton = TheU->getSkeleton())
3126 TheU = Skeleton;
3127
3128 // Emit the header.
3129 MCSymbol *EndLabel = Asm->emitDwarfUnitLength(
3130 Prefix: "pub" + Name, Comment: "Length of Public " + Name + " Info");
3131
3132 Asm->OutStreamer->AddComment(T: "DWARF Version");
3133 Asm->emitInt16(Value: dwarf::DW_PUBNAMES_VERSION);
3134
3135 Asm->OutStreamer->AddComment(T: "Offset of Compilation Unit Info");
3136 emitSectionReference(CU: *TheU);
3137
3138 Asm->OutStreamer->AddComment(T: "Compilation Unit Length");
3139 Asm->emitDwarfLengthOrOffset(Value: TheU->getLength());
3140
3141 // Emit the pubnames for this compilation unit.
3142 SmallVector<std::pair<StringRef, const DIE *>, 0> Vec;
3143 for (const auto &GI : Globals)
3144 Vec.emplace_back(Args: GI.first(), Args: GI.second);
3145 llvm::sort(C&: Vec, Comp: [](auto &A, auto &B) {
3146 return A.second->getOffset() < B.second->getOffset();
3147 });
3148 for (const auto &[Name, Entity] : Vec) {
3149 Asm->OutStreamer->AddComment(T: "DIE offset");
3150 Asm->emitDwarfLengthOrOffset(Value: Entity->getOffset());
3151
3152 if (GnuStyle) {
3153 dwarf::PubIndexEntryDescriptor Desc = computeIndexValue(CU: TheU, Die: Entity);
3154 Asm->OutStreamer->AddComment(
3155 T: Twine("Attributes: ") + dwarf::GDBIndexEntryKindString(Kind: Desc.Kind) +
3156 ", " + dwarf::GDBIndexEntryLinkageString(Linkage: Desc.Linkage));
3157 Asm->emitInt8(Value: Desc.toBits());
3158 }
3159
3160 Asm->OutStreamer->AddComment(T: "External Name");
3161 Asm->OutStreamer->emitBytes(Data: StringRef(Name.data(), Name.size() + 1));
3162 }
3163
3164 Asm->OutStreamer->AddComment(T: "End Mark");
3165 Asm->emitDwarfLengthOrOffset(Value: 0);
3166 Asm->OutStreamer->emitLabel(Symbol: EndLabel);
3167}
3168
3169/// Emit null-terminated strings into a debug str section.
3170void DwarfDebug::emitDebugStr() {
3171 MCSection *StringOffsetsSection = nullptr;
3172 if (useSegmentedStringOffsetsTable()) {
3173 emitStringOffsetsTableHeader();
3174 StringOffsetsSection = Asm->getObjFileLowering().getDwarfStrOffSection();
3175 }
3176 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder;
3177 Holder.emitStrings(StrSection: Asm->getObjFileLowering().getDwarfStrSection(),
3178 OffsetSection: StringOffsetsSection, /* UseRelativeOffsets = */ true);
3179}
3180
3181void DwarfDebug::emitDebugLocEntry(ByteStreamer &Streamer,
3182 const DebugLocStream::Entry &Entry,
3183 const DwarfCompileUnit *CU) {
3184 auto &&Comments = DebugLocs.getComments(E: Entry);
3185 auto Comment = Comments.begin();
3186 auto End = Comments.end();
3187
3188 // The expressions are inserted into a byte stream rather early (see
3189 // DwarfExpression::addExpression) so for those ops (e.g. DW_OP_convert) that
3190 // need to reference a base_type DIE the offset of that DIE is not yet known.
3191 // To deal with this we instead insert a placeholder early and then extract
3192 // it here and replace it with the real reference.
3193 unsigned PtrSize = Asm->MAI.getCodePointerSize();
3194 DWARFDataExtractor Data(StringRef(DebugLocs.getBytes(E: Entry).data(),
3195 DebugLocs.getBytes(E: Entry).size()),
3196 Asm->getDataLayout().isLittleEndian(), PtrSize);
3197 DWARFExpression Expr(Data, PtrSize, Asm->OutContext.getDwarfFormat());
3198
3199 using Encoding = DWARFExpression::Operation::Encoding;
3200 uint64_t Offset = 0;
3201 for (const auto &Op : Expr) {
3202 assert(Op.getCode() != dwarf::DW_OP_const_type &&
3203 "3 operand ops not yet supported");
3204 assert(!Op.getSubCode() && "SubOps not yet supported");
3205 Streamer.emitInt8(Byte: Op.getCode(), Comment: Comment != End ? *(Comment++) : "");
3206 Offset++;
3207 for (unsigned I = 0; I < Op.getDescription().Op.size(); ++I) {
3208 if (Op.getDescription().Op[I] == Encoding::BaseTypeRef) {
3209 unsigned Length =
3210 Streamer.emitDIERef(D: *CU->ExprRefedBaseTypes[Op.getRawOperand(Idx: I)].Die);
3211 // Make sure comments stay aligned.
3212 for (unsigned J = 0; J < Length; ++J)
3213 if (Comment != End)
3214 Comment++;
3215 } else {
3216 for (uint64_t J = Offset; J < Op.getOperandEndOffset(Idx: I); ++J)
3217 Streamer.emitInt8(Byte: Data.getData()[J], Comment: Comment != End ? *(Comment++) : "");
3218 }
3219 Offset = Op.getOperandEndOffset(Idx: I);
3220 }
3221 assert(Offset == Op.getEndOffset());
3222 }
3223}
3224
3225void DwarfDebug::emitDebugLocValue(const AsmPrinter &AP, const DIBasicType *BT,
3226 const DbgValueLoc &Value,
3227 DwarfExpression &DwarfExpr) {
3228 auto *DIExpr = Value.getExpression();
3229 DIExpressionCursor ExprCursor(DIExpr);
3230 DwarfExpr.addFragmentOffset(Expr: DIExpr);
3231
3232 // If the DIExpr is an Entry Value, we want to follow the same code path
3233 // regardless of whether the DBG_VALUE is variadic or not.
3234 if (DIExpr && DIExpr->isEntryValue()) {
3235 // Entry values can only be a single register with no additional DIExpr,
3236 // so just add it directly.
3237 assert(Value.getLocEntries().size() == 1);
3238 assert(Value.getLocEntries()[0].isLocation());
3239 MachineLocation Location = Value.getLocEntries()[0].getLoc();
3240 DwarfExpr.setLocation(Loc: Location, DIExpr);
3241
3242 DwarfExpr.beginEntryValueExpression(ExprCursor);
3243
3244 const TargetRegisterInfo &TRI = *AP.MF->getSubtarget().getRegisterInfo();
3245 if (!DwarfExpr.addMachineRegExpression(TRI, Expr&: ExprCursor, MachineReg: Location.getReg()))
3246 return;
3247 return DwarfExpr.addExpression(Expr: std::move(ExprCursor));
3248 }
3249
3250 // Regular entry.
3251 auto EmitValueLocEntry = [&DwarfExpr, &BT,
3252 &AP](const DbgValueLocEntry &Entry,
3253 DIExpressionCursor &Cursor) -> bool {
3254 if (Entry.isInt()) {
3255 if (BT && (BT->getEncoding() == dwarf::DW_ATE_boolean)) {
3256 DwarfExpr.addBooleanConstant(Value: Entry.getInt());
3257 return true;
3258 }
3259
3260 bool IsSigned = BT && (BT->getEncoding() == dwarf::DW_ATE_signed ||
3261 BT->getEncoding() == dwarf::DW_ATE_signed_char);
3262 if (BT && AP.getDwarfVersion() >= 4 &&
3263 !AP.getDwarfDebug()->tuneForSCE() && !Cursor) {
3264 // DW_OP_const* pushes a generic, address-sized value. For a wider
3265 // source integer value that cannot fit in the generic type, use
3266 // DW_OP_implicit_value to preserve the source bytes instead. Keep this
3267 // limited to complete constant values: SCE tuning already avoids
3268 // DW_OP_implicit_value for compatibility, and expressions with
3269 // remaining operations may need a scalar stack value rather than an
3270 // implicit value block.
3271 unsigned GenericBitSize = AP.MAI.getCodePointerSize() * 8;
3272 uint64_t TypeBitSize = BT->getSizeInBits();
3273 bool IsByteSized = TypeBitSize % 8 == 0;
3274 bool IsOutOfRange =
3275 IsSigned ? !isIntN(N: GenericBitSize, x: Entry.getInt())
3276 : !isUIntN(N: GenericBitSize,
3277 x: static_cast<uint64_t>(Entry.getInt()));
3278 if (TypeBitSize > GenericBitSize && IsByteSized && IsOutOfRange) {
3279 DwarfExpr.addImplicitValue(
3280 Value: APInt(static_cast<unsigned>(TypeBitSize),
3281 static_cast<uint64_t>(Entry.getInt()), IsSigned),
3282 AP);
3283 return true;
3284 }
3285 }
3286
3287 if (IsSigned)
3288 DwarfExpr.addSignedConstant(Value: Entry.getInt());
3289 else
3290 DwarfExpr.addUnsignedConstant(Value: Entry.getInt());
3291 } else if (Entry.isLocation()) {
3292 MachineLocation Location = Entry.getLoc();
3293 if (Location.isIndirect())
3294 DwarfExpr.setMemoryLocationKind();
3295
3296 const TargetRegisterInfo &TRI = *AP.MF->getSubtarget().getRegisterInfo();
3297 if (!DwarfExpr.addMachineRegExpression(TRI, Expr&: Cursor, MachineReg: Location.getReg()))
3298 return false;
3299 } else if (Entry.isTargetIndexLocation()) {
3300 TargetIndexLocation Loc = Entry.getTargetIndexLocation();
3301 // TODO TargetIndexLocation is a target-independent. Currently only the
3302 // WebAssembly-specific encoding is supported.
3303 assert(AP.TM.getTargetTriple().isWasm());
3304 DwarfExpr.addWasmLocation(Index: Loc.Index, Offset: static_cast<uint64_t>(Loc.Offset));
3305 } else if (Entry.isConstantFP()) {
3306 if (AP.getDwarfVersion() >= 4 && !AP.getDwarfDebug()->tuneForSCE() &&
3307 !Cursor) {
3308 DwarfExpr.addConstantFP(Value: Entry.getConstantFP()->getValueAPF(), AP);
3309 } else if (Entry.getConstantFP()
3310 ->getValueAPF()
3311 .bitcastToAPInt()
3312 .getBitWidth() <= 64 /*bits*/) {
3313 DwarfExpr.addUnsignedConstant(
3314 Value: Entry.getConstantFP()->getValueAPF().bitcastToAPInt());
3315 } else {
3316 LLVM_DEBUG(
3317 dbgs() << "Skipped DwarfExpression creation for ConstantFP of size"
3318 << Entry.getConstantFP()
3319 ->getValueAPF()
3320 .bitcastToAPInt()
3321 .getBitWidth()
3322 << " bits\n");
3323 return false;
3324 }
3325 }
3326 return true;
3327 };
3328
3329 if (!Value.isVariadic()) {
3330 if (!EmitValueLocEntry(Value.getLocEntries()[0], ExprCursor))
3331 return;
3332 DwarfExpr.addExpression(Expr: std::move(ExprCursor));
3333 return;
3334 }
3335
3336 // If any of the location entries are registers with the value 0, then the
3337 // location is undefined.
3338 if (any_of(Range: Value.getLocEntries(), P: [](const DbgValueLocEntry &Entry) {
3339 return Entry.isLocation() && !Entry.getLoc().getReg();
3340 }))
3341 return;
3342
3343 DwarfExpr.addExpression(
3344 Expr: std::move(ExprCursor),
3345 InsertArg: [EmitValueLocEntry, &Value](unsigned Idx,
3346 DIExpressionCursor &Cursor) -> bool {
3347 return EmitValueLocEntry(Value.getLocEntries()[Idx], Cursor);
3348 });
3349}
3350
3351void DebugLocEntry::finalize(const AsmPrinter &AP,
3352 DebugLocStream::ListBuilder &List,
3353 const DIBasicType *BT,
3354 DwarfCompileUnit &TheCU) {
3355 assert(!Values.empty() &&
3356 "location list entries without values are redundant");
3357 assert(Begin != End && "unexpected location list entry with empty range");
3358 DebugLocStream::EntryBuilder Entry(List, Begin, End);
3359 BufferByteStreamer Streamer = Entry.getStreamer();
3360 DebugLocDwarfExpression DwarfExpr(AP.getDwarfVersion(), Streamer, TheCU);
3361 const DbgValueLoc &Value = Values[0];
3362 if (Value.isFragment()) {
3363 // Emit all fragments that belong to the same variable and range.
3364 assert(llvm::all_of(Values, [](DbgValueLoc P) {
3365 return P.isFragment();
3366 }) && "all values are expected to be fragments");
3367 assert(llvm::is_sorted(Values) && "fragments are expected to be sorted");
3368
3369 for (const auto &Fragment : Values)
3370 DwarfDebug::emitDebugLocValue(AP, BT, Value: Fragment, DwarfExpr);
3371
3372 } else {
3373 assert(Values.size() == 1 && "only fragments may have >1 value");
3374 DwarfDebug::emitDebugLocValue(AP, BT, Value, DwarfExpr);
3375 }
3376 DwarfExpr.finalize();
3377 if (DwarfExpr.TagOffset)
3378 List.setTagOffset(*DwarfExpr.TagOffset);
3379}
3380
3381void DwarfDebug::emitDebugLocEntryLocation(const DebugLocStream::Entry &Entry,
3382 const DwarfCompileUnit *CU) {
3383 // Emit the size.
3384 Asm->OutStreamer->AddComment(T: "Loc expr size");
3385 if (getDwarfVersion() >= 5)
3386 Asm->emitULEB128(Value: DebugLocs.getBytes(E: Entry).size());
3387 else if (DebugLocs.getBytes(E: Entry).size() <= std::numeric_limits<uint16_t>::max())
3388 Asm->emitInt16(Value: DebugLocs.getBytes(E: Entry).size());
3389 else {
3390 // The entry is too big to fit into 16 bit, drop it as there is nothing we
3391 // can do.
3392 Asm->emitInt16(Value: 0);
3393 return;
3394 }
3395 // Emit the entry.
3396 APByteStreamer Streamer(*Asm);
3397 emitDebugLocEntry(Streamer, Entry, CU);
3398}
3399
3400// Emit the header of a DWARF 5 range list table list table. Returns the symbol
3401// that designates the end of the table for the caller to emit when the table is
3402// complete.
3403static MCSymbol *emitRnglistsTableHeader(AsmPrinter *Asm,
3404 const DwarfFile &Holder) {
3405 MCSymbol *TableEnd = mcdwarf::emitListsTableHeaderStart(S&: *Asm->OutStreamer);
3406
3407 Asm->OutStreamer->AddComment(T: "Offset entry count");
3408 Asm->emitInt32(Value: Holder.getRangeLists().size());
3409 Asm->OutStreamer->emitLabel(Symbol: Holder.getRnglistsTableBaseSym());
3410
3411 for (const RangeSpanList &List : Holder.getRangeLists())
3412 Asm->emitLabelDifference(Hi: List.Label, Lo: Holder.getRnglistsTableBaseSym(),
3413 Size: Asm->getDwarfOffsetByteSize());
3414
3415 return TableEnd;
3416}
3417
3418// Emit the header of a DWARF 5 locations list table. Returns the symbol that
3419// designates the end of the table for the caller to emit when the table is
3420// complete.
3421static MCSymbol *emitLoclistsTableHeader(AsmPrinter *Asm,
3422 const DwarfDebug &DD) {
3423 MCSymbol *TableEnd = mcdwarf::emitListsTableHeaderStart(S&: *Asm->OutStreamer);
3424
3425 const auto &DebugLocs = DD.getDebugLocs();
3426
3427 Asm->OutStreamer->AddComment(T: "Offset entry count");
3428 Asm->emitInt32(Value: DebugLocs.getLists().size());
3429 Asm->OutStreamer->emitLabel(Symbol: DebugLocs.getSym());
3430
3431 for (const auto &List : DebugLocs.getLists())
3432 Asm->emitLabelDifference(Hi: List.Label, Lo: DebugLocs.getSym(),
3433 Size: Asm->getDwarfOffsetByteSize());
3434
3435 return TableEnd;
3436}
3437
3438template <typename Ranges, typename PayloadEmitter>
3439static void
3440emitRangeList(DwarfDebug &DD, AsmPrinter *Asm, MCSymbol *Sym, const Ranges &R,
3441 const DwarfCompileUnit &CU, unsigned BaseAddressx,
3442 unsigned OffsetPair, unsigned StartxLength, unsigned StartxEndx,
3443 unsigned EndOfList, StringRef (*StringifyEnum)(unsigned),
3444 bool ShouldUseBaseAddress, PayloadEmitter EmitPayload) {
3445 auto Size = Asm->MAI.getCodePointerSize();
3446 bool UseDwarf5 = DD.getDwarfVersion() >= 5;
3447
3448 // Emit our symbol so we can find the beginning of the range.
3449 Asm->OutStreamer->emitLabel(Symbol: Sym);
3450
3451 // Gather all the ranges that apply to the same section so they can share
3452 // a base address entry.
3453 SmallMapVector<const MCSection *, std::vector<decltype(&*R.begin())>, 16>
3454 SectionRanges;
3455
3456 for (const auto &Range : R)
3457 SectionRanges[&Range.Begin->getSection()].push_back(&Range);
3458
3459 const MCSymbol *CUBase = CU.getBaseAddress();
3460 bool BaseIsSet = false;
3461 for (const auto &P : SectionRanges) {
3462 auto *Base = CUBase;
3463 if (DD.shouldResetBaseAddress(Section: *P.first) ||
3464 (DD.useSplitDwarf() && UseDwarf5 && P.first->isLinkerRelaxable())) {
3465 BaseIsSet = false;
3466 Base = nullptr;
3467 } else if (!Base && ShouldUseBaseAddress) {
3468 const MCSymbol *Begin = P.second.front()->Begin;
3469 const MCSymbol *NewBase = DD.getSectionLabel(S: &Begin->getSection());
3470 if (!UseDwarf5) {
3471 Base = NewBase;
3472 BaseIsSet = true;
3473 Asm->OutStreamer->emitIntValue(Value: -1, Size);
3474 Asm->OutStreamer->AddComment(T: " base address");
3475 Asm->OutStreamer->emitSymbolValue(Sym: Base, Size);
3476 } else if (NewBase != Begin || P.second.size() > 1) {
3477 // Only use a base address if
3478 // * the existing pool address doesn't match (NewBase != Begin)
3479 // * or, there's more than one entry to share the base address
3480 Base = NewBase;
3481 BaseIsSet = true;
3482 Asm->OutStreamer->AddComment(T: StringifyEnum(BaseAddressx));
3483 Asm->emitInt8(Value: BaseAddressx);
3484 Asm->OutStreamer->AddComment(T: " base address index");
3485 Asm->emitULEB128(Value: DD.getAddressPool().getIndex(Sym: Base));
3486 }
3487 } else if (BaseIsSet && !UseDwarf5) {
3488 BaseIsSet = false;
3489 assert(!Base);
3490 Asm->OutStreamer->emitIntValue(Value: -1, Size);
3491 Asm->OutStreamer->emitIntValue(Value: 0, Size);
3492 }
3493
3494 for (const auto *RS : P.second) {
3495 const MCSymbol *Begin = RS->Begin;
3496 const MCSymbol *End = RS->End;
3497 assert(Begin && "Range without a begin symbol?");
3498 assert(End && "Range without an end symbol?");
3499 if (Base) {
3500 if (UseDwarf5) {
3501 // Emit offset_pair when we have a base.
3502 Asm->OutStreamer->AddComment(T: StringifyEnum(OffsetPair));
3503 Asm->emitInt8(Value: OffsetPair);
3504 Asm->OutStreamer->AddComment(T: " starting offset");
3505 Asm->emitLabelDifferenceAsULEB128(Hi: Begin, Lo: Base);
3506 Asm->OutStreamer->AddComment(T: " ending offset");
3507 Asm->emitLabelDifferenceAsULEB128(Hi: End, Lo: Base);
3508 } else {
3509 Asm->emitLabelDifference(Hi: Begin, Lo: Base, Size);
3510 Asm->emitLabelDifference(Hi: End, Lo: Base, Size);
3511 }
3512 } else if (UseDwarf5) {
3513 // NOTE: We can't use absoluteSymbolDiff here instead of
3514 // isRangeRelaxable. While isRangeRelaxable only checks that the offset
3515 // between labels won't change at link time (which is exactly what we
3516 // need), absoluteSymbolDiff also requires that the offset remain
3517 // unchanged at assembly time, imposing a much stricter condition.
3518 // Consequently, this would lead to less optimal debug info emission.
3519 if (DD.useSplitDwarf() && llvm::isRangeRelaxable(Begin, End)) {
3520 Asm->OutStreamer->AddComment(T: StringifyEnum(StartxEndx));
3521 Asm->emitInt8(Value: StartxEndx);
3522 Asm->OutStreamer->AddComment(T: " start index");
3523 Asm->emitULEB128(Value: DD.getAddressPool().getIndex(Sym: Begin));
3524 Asm->OutStreamer->AddComment(T: " end index");
3525 Asm->emitULEB128(Value: DD.getAddressPool().getIndex(Sym: End));
3526 } else {
3527 Asm->OutStreamer->AddComment(T: StringifyEnum(StartxLength));
3528 Asm->emitInt8(Value: StartxLength);
3529 Asm->OutStreamer->AddComment(T: " start index");
3530 Asm->emitULEB128(Value: DD.getAddressPool().getIndex(Sym: Begin));
3531 Asm->OutStreamer->AddComment(T: " length");
3532 Asm->emitLabelDifferenceAsULEB128(Hi: End, Lo: Begin);
3533 }
3534 } else {
3535 Asm->OutStreamer->emitSymbolValue(Sym: Begin, Size);
3536 Asm->OutStreamer->emitSymbolValue(Sym: End, Size);
3537 }
3538 EmitPayload(*RS);
3539 }
3540 }
3541
3542 if (UseDwarf5) {
3543 Asm->OutStreamer->AddComment(T: StringifyEnum(EndOfList));
3544 Asm->emitInt8(Value: EndOfList);
3545 } else {
3546 // Terminate the list with two 0 values.
3547 Asm->OutStreamer->emitIntValue(Value: 0, Size);
3548 Asm->OutStreamer->emitIntValue(Value: 0, Size);
3549 }
3550}
3551
3552// Handles emission of both debug_loclist / debug_loclist.dwo
3553static void emitLocList(DwarfDebug &DD, AsmPrinter *Asm, const DebugLocStream::List &List) {
3554 emitRangeList(
3555 DD, Asm, Sym: List.Label, R: DD.getDebugLocs().getEntries(L: List), CU: *List.CU,
3556 BaseAddressx: dwarf::DW_LLE_base_addressx, OffsetPair: dwarf::DW_LLE_offset_pair,
3557 StartxLength: dwarf::DW_LLE_startx_length, StartxEndx: dwarf::DW_LLE_startx_endx,
3558 EndOfList: dwarf::DW_LLE_end_of_list, StringifyEnum: llvm::dwarf::LocListEncodingString,
3559 /* ShouldUseBaseAddress */ true, EmitPayload: [&](const DebugLocStream::Entry &E) {
3560 DD.emitDebugLocEntryLocation(Entry: E, CU: List.CU);
3561 });
3562}
3563
3564void DwarfDebug::emitDebugLocImpl(MCSection *Sec) {
3565 if (DebugLocs.getLists().empty())
3566 return;
3567
3568 Asm->OutStreamer->switchSection(Section: Sec);
3569
3570 MCSymbol *TableEnd = nullptr;
3571 if (getDwarfVersion() >= 5)
3572 TableEnd = emitLoclistsTableHeader(Asm, DD: *this);
3573
3574 for (const auto &List : DebugLocs.getLists())
3575 emitLocList(DD&: *this, Asm, List);
3576
3577 if (TableEnd)
3578 Asm->OutStreamer->emitLabel(Symbol: TableEnd);
3579}
3580
3581// Emit locations into the .debug_loc/.debug_loclists section.
3582void DwarfDebug::emitDebugLoc() {
3583 emitDebugLocImpl(
3584 Sec: getDwarfVersion() >= 5
3585 ? Asm->getObjFileLowering().getDwarfLoclistsSection()
3586 : Asm->getObjFileLowering().getDwarfLocSection());
3587}
3588
3589// Emit locations into the .debug_loc.dwo/.debug_loclists.dwo section.
3590void DwarfDebug::emitDebugLocDWO() {
3591 if (getDwarfVersion() >= 5) {
3592 emitDebugLocImpl(
3593 Sec: Asm->getObjFileLowering().getDwarfLoclistsDWOSection());
3594
3595 return;
3596 }
3597
3598 for (const auto &List : DebugLocs.getLists()) {
3599 Asm->OutStreamer->switchSection(
3600 Section: Asm->getObjFileLowering().getDwarfLocDWOSection());
3601 Asm->OutStreamer->emitLabel(Symbol: List.Label);
3602
3603 for (const auto &Entry : DebugLocs.getEntries(L: List)) {
3604 // GDB only supports startx_length in pre-standard split-DWARF.
3605 // (in v5 standard loclists, it currently* /only/ supports base_address +
3606 // offset_pair, so the implementations can't really share much since they
3607 // need to use different representations)
3608 // * as of October 2018, at least
3609 //
3610 // In v5 (see emitLocList), this uses SectionLabels to reuse existing
3611 // addresses in the address pool to minimize object size/relocations.
3612 Asm->emitInt8(Value: dwarf::DW_LLE_startx_length);
3613 unsigned idx = AddrPool.getIndex(Sym: Entry.Begin);
3614 Asm->emitULEB128(Value: idx);
3615 // Also the pre-standard encoding is slightly different, emitting this as
3616 // an address-length entry here, but its a ULEB128 in DWARFv5 loclists.
3617 Asm->emitLabelDifference(Hi: Entry.End, Lo: Entry.Begin, Size: 4);
3618 emitDebugLocEntryLocation(Entry, CU: List.CU);
3619 }
3620 Asm->emitInt8(Value: dwarf::DW_LLE_end_of_list);
3621 }
3622}
3623
3624struct ArangeSpan {
3625 const MCSymbol *Start, *End;
3626};
3627
3628// Emit a debug aranges section, containing a CU lookup for any
3629// address we can tie back to a CU.
3630void DwarfDebug::emitDebugARanges() {
3631 if (ArangeLabels.empty())
3632 return;
3633
3634 // Provides a unique id per text section.
3635 MapVector<MCSection *, SmallVector<SymbolCU, 8>> SectionMap;
3636
3637 // Filter labels by section.
3638 for (const SymbolCU &SCU : ArangeLabels) {
3639 if (SCU.Sym->isInSection()) {
3640 // Make a note of this symbol and it's section.
3641 MCSection *Section = &SCU.Sym->getSection();
3642 SectionMap[Section].push_back(Elt: SCU);
3643 } else {
3644 // Some symbols (e.g. common/bss on mach-o) can have no section but still
3645 // appear in the output. This sucks as we rely on sections to build
3646 // arange spans. We can do it without, but it's icky.
3647 SectionMap[nullptr].push_back(Elt: SCU);
3648 }
3649 }
3650
3651 DenseMap<DwarfCompileUnit *, std::vector<ArangeSpan>> Spans;
3652
3653 for (auto &I : SectionMap) {
3654 MCSection *Section = I.first;
3655 SmallVector<SymbolCU, 8> &List = I.second;
3656 assert(!List.empty());
3657
3658 // If we have no section (e.g. common), just write out
3659 // individual spans for each symbol.
3660 if (!Section) {
3661 for (const SymbolCU &Cur : List) {
3662 ArangeSpan Span;
3663 Span.Start = Cur.Sym;
3664 Span.End = nullptr;
3665 assert(Cur.CU);
3666 Spans[Cur.CU].push_back(x: Span);
3667 }
3668 continue;
3669 }
3670
3671 // Insert a final terminator.
3672 List.push_back(Elt: SymbolCU(nullptr, Asm->OutStreamer->endSection(Section)));
3673
3674 // Build spans between each label.
3675 const MCSymbol *StartSym = List[0].Sym;
3676 for (size_t n = 1, e = List.size(); n < e; n++) {
3677 const SymbolCU &Prev = List[n - 1];
3678 const SymbolCU &Cur = List[n];
3679
3680 // Try and build the longest span we can within the same CU.
3681 if (Cur.CU != Prev.CU) {
3682 ArangeSpan Span;
3683 Span.Start = StartSym;
3684 Span.End = Cur.Sym;
3685 assert(Prev.CU);
3686 Spans[Prev.CU].push_back(x: Span);
3687 StartSym = Cur.Sym;
3688 }
3689 }
3690 }
3691
3692 // Start the dwarf aranges section.
3693 Asm->OutStreamer->switchSection(
3694 Section: Asm->getObjFileLowering().getDwarfARangesSection());
3695
3696 unsigned PtrSize = Asm->MAI.getCodePointerSize();
3697
3698 // Build a list of CUs used.
3699 std::vector<DwarfCompileUnit *> CUs;
3700 for (const auto &it : Spans) {
3701 DwarfCompileUnit *CU = it.first;
3702 CUs.push_back(x: CU);
3703 }
3704
3705 // Sort the CU list (again, to ensure consistent output order).
3706 llvm::sort(C&: CUs, Comp: [](const DwarfCompileUnit *A, const DwarfCompileUnit *B) {
3707 return A->getUniqueID() < B->getUniqueID();
3708 });
3709
3710 // Emit an arange table for each CU we used.
3711 for (DwarfCompileUnit *CU : CUs) {
3712 std::vector<ArangeSpan> &List = Spans[CU];
3713
3714 // Describe the skeleton CU's offset and length, not the dwo file's.
3715 if (auto *Skel = CU->getSkeleton())
3716 CU = Skel;
3717
3718 // Emit size of content not including length itself.
3719 unsigned ContentSize =
3720 sizeof(int16_t) + // DWARF ARange version number
3721 Asm->getDwarfOffsetByteSize() + // Offset of CU in the .debug_info
3722 // section
3723 sizeof(int8_t) + // Pointer Size (in bytes)
3724 sizeof(int8_t); // Segment Size (in bytes)
3725
3726 unsigned TupleSize = PtrSize * 2;
3727
3728 // 7.20 in the Dwarf specs requires the table to be aligned to a tuple.
3729 unsigned Padding = offsetToAlignment(
3730 Value: Asm->getUnitLengthFieldByteSize() + ContentSize, Alignment: Align(TupleSize));
3731
3732 ContentSize += Padding;
3733 ContentSize += (List.size() + 1) * TupleSize;
3734
3735 // For each compile unit, write the list of spans it covers.
3736 Asm->emitDwarfUnitLength(Length: ContentSize, Comment: "Length of ARange Set");
3737 Asm->OutStreamer->AddComment(T: "DWARF Arange version number");
3738 Asm->emitInt16(Value: dwarf::DW_ARANGES_VERSION);
3739 Asm->OutStreamer->AddComment(T: "Offset Into Debug Info Section");
3740 emitSectionReference(CU: *CU);
3741 Asm->OutStreamer->AddComment(T: "Address Size (in bytes)");
3742 Asm->emitInt8(Value: PtrSize);
3743 Asm->OutStreamer->AddComment(T: "Segment Size (in bytes)");
3744 Asm->emitInt8(Value: 0);
3745
3746 Asm->OutStreamer->emitFill(NumBytes: Padding, FillValue: 0xff);
3747
3748 for (const ArangeSpan &Span : List) {
3749 Asm->emitLabelReference(Label: Span.Start, Size: PtrSize);
3750
3751 // Calculate the size as being from the span start to its end.
3752 //
3753 // If the size is zero, then round it up to one byte. The DWARF
3754 // specification requires that entries in this table have nonzero
3755 // lengths.
3756 auto SizeRef = SymSize.find(Val: Span.Start);
3757 if ((SizeRef == SymSize.end() || SizeRef->second != 0) && Span.End) {
3758 Asm->emitLabelDifference(Hi: Span.End, Lo: Span.Start, Size: PtrSize);
3759 } else {
3760 // For symbols without an end marker (e.g. common), we
3761 // write a single arange entry containing just that one symbol.
3762 uint64_t Size;
3763 if (SizeRef == SymSize.end() || SizeRef->second == 0)
3764 Size = 1;
3765 else
3766 Size = SizeRef->second;
3767
3768 Asm->OutStreamer->emitIntValue(Value: Size, Size: PtrSize);
3769 }
3770 }
3771
3772 Asm->OutStreamer->AddComment(T: "ARange terminator");
3773 Asm->OutStreamer->emitIntValue(Value: 0, Size: PtrSize);
3774 Asm->OutStreamer->emitIntValue(Value: 0, Size: PtrSize);
3775 }
3776}
3777
3778/// Emit a single range list. We handle both DWARF v5 and earlier.
3779static void emitRangeList(DwarfDebug &DD, AsmPrinter *Asm,
3780 const RangeSpanList &List) {
3781 emitRangeList(DD, Asm, Sym: List.Label, R: List.Ranges, CU: *List.CU,
3782 BaseAddressx: dwarf::DW_RLE_base_addressx, OffsetPair: dwarf::DW_RLE_offset_pair,
3783 StartxLength: dwarf::DW_RLE_startx_length, StartxEndx: dwarf::DW_RLE_startx_endx,
3784 EndOfList: dwarf::DW_RLE_end_of_list, StringifyEnum: llvm::dwarf::RangeListEncodingString,
3785 ShouldUseBaseAddress: List.CU->getCUNode()->getRangesBaseAddress() ||
3786 DD.getDwarfVersion() >= 5,
3787 EmitPayload: [](auto) {});
3788}
3789
3790void DwarfDebug::emitDebugRangesImpl(const DwarfFile &Holder, MCSection *Section) {
3791 if (Holder.getRangeLists().empty())
3792 return;
3793
3794 assert(useRangesSection());
3795 assert(!CUMap.empty());
3796 assert(llvm::any_of(CUMap, [](const decltype(CUMap)::value_type &Pair) {
3797 return !Pair.second->getCUNode()->isDebugDirectivesOnly();
3798 }));
3799
3800 Asm->OutStreamer->switchSection(Section);
3801
3802 MCSymbol *TableEnd = nullptr;
3803 if (getDwarfVersion() >= 5)
3804 TableEnd = emitRnglistsTableHeader(Asm, Holder);
3805
3806 for (const RangeSpanList &List : Holder.getRangeLists())
3807 emitRangeList(DD&: *this, Asm, List);
3808
3809 if (TableEnd)
3810 Asm->OutStreamer->emitLabel(Symbol: TableEnd);
3811}
3812
3813/// Emit address ranges into the .debug_ranges section or into the DWARF v5
3814/// .debug_rnglists section.
3815void DwarfDebug::emitDebugRanges() {
3816 const auto &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder;
3817
3818 emitDebugRangesImpl(Holder,
3819 Section: getDwarfVersion() >= 5
3820 ? Asm->getObjFileLowering().getDwarfRnglistsSection()
3821 : Asm->getObjFileLowering().getDwarfRangesSection());
3822}
3823
3824void DwarfDebug::emitDebugRangesDWO() {
3825 emitDebugRangesImpl(Holder: InfoHolder,
3826 Section: Asm->getObjFileLowering().getDwarfRnglistsDWOSection());
3827}
3828
3829/// Emit the header of a DWARF 5 macro section, or the GNU extension for
3830/// DWARF 4.
3831static void emitMacroHeader(AsmPrinter *Asm, const DwarfDebug &DD,
3832 const DwarfCompileUnit &CU, uint16_t DwarfVersion) {
3833 enum HeaderFlagMask {
3834#define HANDLE_MACRO_FLAG(ID, NAME) MACRO_FLAG_##NAME = ID,
3835#include "llvm/BinaryFormat/Dwarf.def"
3836 };
3837 Asm->OutStreamer->AddComment(T: "Macro information version");
3838 Asm->emitInt16(Value: DwarfVersion >= 5 ? DwarfVersion : 4);
3839 // We emit the line offset flag unconditionally here, since line offset should
3840 // be mostly present.
3841 if (Asm->isDwarf64()) {
3842 Asm->OutStreamer->AddComment(T: "Flags: 64 bit, debug_line_offset present");
3843 Asm->emitInt8(Value: MACRO_FLAG_OFFSET_SIZE | MACRO_FLAG_DEBUG_LINE_OFFSET);
3844 } else {
3845 Asm->OutStreamer->AddComment(T: "Flags: 32 bit, debug_line_offset present");
3846 Asm->emitInt8(Value: MACRO_FLAG_DEBUG_LINE_OFFSET);
3847 }
3848 Asm->OutStreamer->AddComment(T: "debug_line_offset");
3849 if (DD.useSplitDwarf())
3850 Asm->emitDwarfLengthOrOffset(Value: 0);
3851 else
3852 Asm->emitDwarfSymbolReference(Label: CU.getLineTableStartSym());
3853}
3854
3855void DwarfDebug::handleMacroNodes(DIMacroNodeArray Nodes, DwarfCompileUnit &U) {
3856 for (auto *MN : Nodes) {
3857 if (auto *M = dyn_cast<DIMacro>(Val: MN))
3858 emitMacro(M&: *M);
3859 else if (auto *F = dyn_cast<DIMacroFile>(Val: MN))
3860 emitMacroFile(F&: *F, U);
3861 else
3862 llvm_unreachable("Unexpected DI type!");
3863 }
3864}
3865
3866void DwarfDebug::emitMacro(DIMacro &M) {
3867 StringRef Name = M.getName();
3868 StringRef Value = M.getValue();
3869
3870 // There should be one space between the macro name and the macro value in
3871 // define entries. In undef entries, only the macro name is emitted.
3872 std::string Str = Value.empty() ? Name.str() : (Name + " " + Value).str();
3873
3874 if (UseDebugMacroSection) {
3875 if (getDwarfVersion() >= 5) {
3876 unsigned Type = M.getMacinfoType() == dwarf::DW_MACINFO_define
3877 ? dwarf::DW_MACRO_define_strx
3878 : dwarf::DW_MACRO_undef_strx;
3879 Asm->OutStreamer->AddComment(T: dwarf::MacroString(Encoding: Type));
3880 Asm->emitULEB128(Value: Type);
3881 Asm->OutStreamer->AddComment(T: "Line Number");
3882 Asm->emitULEB128(Value: M.getLine());
3883 Asm->OutStreamer->AddComment(T: "Macro String");
3884 Asm->emitULEB128(
3885 Value: InfoHolder.getStringPool().getIndexedEntry(Asm&: *Asm, Str).getIndex());
3886 } else {
3887 unsigned Type = M.getMacinfoType() == dwarf::DW_MACINFO_define
3888 ? dwarf::DW_MACRO_GNU_define_indirect
3889 : dwarf::DW_MACRO_GNU_undef_indirect;
3890 Asm->OutStreamer->AddComment(T: dwarf::GnuMacroString(Encoding: Type));
3891 Asm->emitULEB128(Value: Type);
3892 Asm->OutStreamer->AddComment(T: "Line Number");
3893 Asm->emitULEB128(Value: M.getLine());
3894 Asm->OutStreamer->AddComment(T: "Macro String");
3895 Asm->emitDwarfSymbolReference(
3896 Label: InfoHolder.getStringPool().getEntry(Asm&: *Asm, Str).getSymbol());
3897 }
3898 } else {
3899 Asm->OutStreamer->AddComment(T: dwarf::MacinfoString(Encoding: M.getMacinfoType()));
3900 Asm->emitULEB128(Value: M.getMacinfoType());
3901 Asm->OutStreamer->AddComment(T: "Line Number");
3902 Asm->emitULEB128(Value: M.getLine());
3903 Asm->OutStreamer->AddComment(T: "Macro String");
3904 Asm->OutStreamer->emitBytes(Data: Str);
3905 Asm->emitInt8(Value: '\0');
3906 }
3907}
3908
3909void DwarfDebug::emitMacroFileImpl(
3910 DIMacroFile &MF, DwarfCompileUnit &U, unsigned StartFile, unsigned EndFile,
3911 StringRef (*MacroFormToString)(unsigned Form)) {
3912
3913 Asm->OutStreamer->AddComment(T: MacroFormToString(StartFile));
3914 Asm->emitULEB128(Value: StartFile);
3915 Asm->OutStreamer->AddComment(T: "Line Number");
3916 Asm->emitULEB128(Value: MF.getLine());
3917 Asm->OutStreamer->AddComment(T: "File Number");
3918 DIFile &F = *MF.getFile();
3919 if (useSplitDwarf())
3920 Asm->emitULEB128(Value: getDwoLineTable(U)->getFile(
3921 Directory: F.getDirectory(), FileName: F.getFilename(), Checksum: getMD5AsBytes(File: &F),
3922 DwarfVersion: Asm->OutContext.getDwarfVersion(), Source: F.getSource()));
3923 else
3924 Asm->emitULEB128(Value: U.getOrCreateSourceID(File: &F));
3925 handleMacroNodes(Nodes: MF.getElements(), U);
3926 Asm->OutStreamer->AddComment(T: MacroFormToString(EndFile));
3927 Asm->emitULEB128(Value: EndFile);
3928}
3929
3930void DwarfDebug::emitMacroFile(DIMacroFile &F, DwarfCompileUnit &U) {
3931 // DWARFv5 macro and DWARFv4 macinfo share some common encodings,
3932 // so for readibility/uniformity, We are explicitly emitting those.
3933 assert(F.getMacinfoType() == dwarf::DW_MACINFO_start_file);
3934 if (UseDebugMacroSection)
3935 emitMacroFileImpl(
3936 MF&: F, U, StartFile: dwarf::DW_MACRO_start_file, EndFile: dwarf::DW_MACRO_end_file,
3937 MacroFormToString: (getDwarfVersion() >= 5) ? dwarf::MacroString : dwarf::GnuMacroString);
3938 else
3939 emitMacroFileImpl(MF&: F, U, StartFile: dwarf::DW_MACINFO_start_file,
3940 EndFile: dwarf::DW_MACINFO_end_file, MacroFormToString: dwarf::MacinfoString);
3941}
3942
3943void DwarfDebug::emitDebugMacinfoImpl(MCSection *Section) {
3944 for (const auto &P : CUMap) {
3945 auto &TheCU = *P.second;
3946 auto *SkCU = TheCU.getSkeleton();
3947 DwarfCompileUnit &U = SkCU ? *SkCU : TheCU;
3948 auto *CUNode = cast<DICompileUnit>(Val: P.first);
3949 DIMacroNodeArray Macros = CUNode->getMacros();
3950 if (Macros.empty())
3951 continue;
3952 Asm->OutStreamer->switchSection(Section);
3953 Asm->OutStreamer->emitLabel(Symbol: U.getMacroLabelBegin());
3954 if (UseDebugMacroSection)
3955 emitMacroHeader(Asm, DD: *this, CU: U, DwarfVersion: getDwarfVersion());
3956 handleMacroNodes(Nodes: Macros, U);
3957 Asm->OutStreamer->AddComment(T: "End Of Macro List Mark");
3958 Asm->emitInt8(Value: 0);
3959 }
3960}
3961
3962/// Emit macros into a debug macinfo/macro section.
3963void DwarfDebug::emitDebugMacinfo() {
3964 auto &ObjLower = Asm->getObjFileLowering();
3965 emitDebugMacinfoImpl(Section: UseDebugMacroSection
3966 ? ObjLower.getDwarfMacroSection()
3967 : ObjLower.getDwarfMacinfoSection());
3968}
3969
3970void DwarfDebug::emitDebugMacinfoDWO() {
3971 auto &ObjLower = Asm->getObjFileLowering();
3972 emitDebugMacinfoImpl(Section: UseDebugMacroSection
3973 ? ObjLower.getDwarfMacroDWOSection()
3974 : ObjLower.getDwarfMacinfoDWOSection());
3975}
3976
3977// DWARF5 Experimental Separate Dwarf emitters.
3978
3979void DwarfDebug::initSkeletonUnit(const DwarfUnit &U, DIE &Die,
3980 std::unique_ptr<DwarfCompileUnit> NewU) {
3981
3982 if (!CompilationDir.empty())
3983 NewU->addString(Die, Attribute: dwarf::DW_AT_comp_dir, Str: CompilationDir);
3984 addGnuPubAttributes(U&: *NewU, D&: Die);
3985
3986 SkeletonHolder.addUnit(U: std::move(NewU));
3987}
3988
3989DwarfCompileUnit &DwarfDebug::constructSkeletonCU(const DwarfCompileUnit &CU) {
3990
3991 auto OwnedUnit = std::make_unique<DwarfCompileUnit>(
3992 args: CU.getUniqueID(), args: CU.getCUNode(), args&: Asm, args: this, args: &SkeletonHolder,
3993 args: UnitKind::Skeleton);
3994 DwarfCompileUnit &NewCU = *OwnedUnit;
3995 NewCU.setSection(Asm->getObjFileLowering().getDwarfInfoSection());
3996
3997 NewCU.initStmtList();
3998
3999 if (useSegmentedStringOffsetsTable())
4000 NewCU.addStringOffsetsStart();
4001
4002 initSkeletonUnit(U: CU, Die&: NewCU.getUnitDie(), NewU: std::move(OwnedUnit));
4003
4004 return NewCU;
4005}
4006
4007// Emit the .debug_info.dwo section for separated dwarf. This contains the
4008// compile units that would normally be in debug_info.
4009void DwarfDebug::emitDebugInfoDWO() {
4010 assert(useSplitDwarf() && "No split dwarf debug info?");
4011 // Don't emit relocations into the dwo file.
4012 InfoHolder.emitUnits(/* UseOffsets */ true);
4013}
4014
4015// Emit the .debug_abbrev.dwo section for separated dwarf. This contains the
4016// abbreviations for the .debug_info.dwo section.
4017void DwarfDebug::emitDebugAbbrevDWO() {
4018 assert(useSplitDwarf() && "No split dwarf?");
4019 InfoHolder.emitAbbrevs(Asm->getObjFileLowering().getDwarfAbbrevDWOSection());
4020}
4021
4022void DwarfDebug::emitDebugLineDWO() {
4023 assert(useSplitDwarf() && "No split dwarf?");
4024 SplitTypeUnitFileTable.Emit(
4025 MCOS&: *Asm->OutStreamer, Params: MCDwarfLineTableParams(),
4026 Section: Asm->getObjFileLowering().getDwarfLineDWOSection());
4027}
4028
4029void DwarfDebug::emitStringOffsetsTableHeaderDWO() {
4030 assert(useSplitDwarf() && "No split dwarf?");
4031 InfoHolder.getStringPool().emitStringOffsetsTableHeader(
4032 Asm&: *Asm, OffsetSection: Asm->getObjFileLowering().getDwarfStrOffDWOSection(),
4033 StartSym: InfoHolder.getStringOffsetsStartSym());
4034}
4035
4036// Emit the .debug_str.dwo section for separated dwarf. This contains the
4037// string section and is identical in format to traditional .debug_str
4038// sections.
4039void DwarfDebug::emitDebugStrDWO() {
4040 if (useSegmentedStringOffsetsTable())
4041 emitStringOffsetsTableHeaderDWO();
4042 assert(useSplitDwarf() && "No split dwarf?");
4043 MCSection *OffSec = Asm->getObjFileLowering().getDwarfStrOffDWOSection();
4044 InfoHolder.emitStrings(StrSection: Asm->getObjFileLowering().getDwarfStrDWOSection(),
4045 OffsetSection: OffSec, /* UseRelativeOffsets = */ false);
4046}
4047
4048// Emit address pool.
4049void DwarfDebug::emitDebugAddr() {
4050 AddrPool.emit(Asm&: *Asm, AddrSection: Asm->getObjFileLowering().getDwarfAddrSection());
4051}
4052
4053MCDwarfDwoLineTable *DwarfDebug::getDwoLineTable(const DwarfCompileUnit &CU) {
4054 if (!useSplitDwarf())
4055 return nullptr;
4056 const DICompileUnit *DIUnit = CU.getCUNode();
4057 SplitTypeUnitFileTable.maybeSetRootFile(
4058 Directory: DIUnit->getDirectory(), FileName: DIUnit->getFilename(),
4059 Checksum: getMD5AsBytes(File: DIUnit->getFile()), Source: DIUnit->getSource());
4060 return &SplitTypeUnitFileTable;
4061}
4062
4063uint64_t DwarfDebug::makeTypeSignature(StringRef Identifier) {
4064 MD5 Hash;
4065 Hash.update(Str: Identifier);
4066 // ... take the least significant 8 bytes and return those. Our MD5
4067 // implementation always returns its results in little endian, so we actually
4068 // need the "high" word.
4069 MD5::MD5Result Result;
4070 Hash.final(Result);
4071 return Result.high();
4072}
4073
4074void DwarfDebug::addDwarfTypeUnitType(DwarfCompileUnit &CU,
4075 StringRef Identifier, DIE &RefDie,
4076 const DICompositeType *CTy) {
4077 // Fast path if we're building some type units and one has already used the
4078 // address pool we know we're going to throw away all this work anyway, so
4079 // don't bother building dependent types.
4080 if (!TypeUnitsUnderConstruction.empty() && AddrPool.hasBeenUsed())
4081 return;
4082
4083 auto Ins = TypeSignatures.try_emplace(Key: CTy);
4084 if (!Ins.second) {
4085 CU.addDIETypeSignature(Die&: RefDie, Signature: Ins.first->second);
4086 return;
4087 }
4088
4089 setCurrentDWARF5AccelTable(DWARF5AccelTableKind::TU);
4090 bool TopLevelType = TypeUnitsUnderConstruction.empty();
4091 AddrPool.resetUsedFlag();
4092
4093 auto OwnedUnit = std::make_unique<DwarfTypeUnit>(
4094 args&: CU, args&: Asm, args: this, args: &InfoHolder, args: NumTypeUnitsCreated++, args: getDwoLineTable(CU));
4095 DwarfTypeUnit &NewTU = *OwnedUnit;
4096 DIE &UnitDie = NewTU.getUnitDie();
4097 TypeUnitsUnderConstruction.emplace_back(Args: std::move(OwnedUnit), Args&: CTy);
4098
4099 NewTU.addUInt(Die&: UnitDie, Attribute: dwarf::DW_AT_language, Form: dwarf::DW_FORM_data2,
4100 Integer: CU.getSourceLanguage());
4101
4102 uint64_t Signature = makeTypeSignature(Identifier);
4103 NewTU.setTypeSignature(Signature);
4104 Ins.first->second = Signature;
4105
4106 if (useSplitDwarf()) {
4107 // Although multiple type units can have the same signature, they are not
4108 // guranteed to be bit identical. When LLDB uses .debug_names it needs to
4109 // know from which CU a type unit came from. These two attrbutes help it to
4110 // figure that out.
4111 if (getDwarfVersion() >= 5) {
4112 if (!CompilationDir.empty())
4113 NewTU.addString(Die&: UnitDie, Attribute: dwarf::DW_AT_comp_dir, Str: CompilationDir);
4114 NewTU.addString(Die&: UnitDie, Attribute: dwarf::DW_AT_dwo_name,
4115 Str: Asm->TM.Options.MCOptions.SplitDwarfFile);
4116 }
4117 MCSection *Section =
4118 getDwarfVersion() <= 4
4119 ? Asm->getObjFileLowering().getDwarfTypesDWOSection()
4120 : Asm->getObjFileLowering().getDwarfInfoDWOSection();
4121 NewTU.setSection(Section);
4122 } else {
4123 MCSection *Section =
4124 getDwarfVersion() <= 4
4125 ? Asm->getObjFileLowering().getDwarfTypesSection(Hash: Signature)
4126 : Asm->getObjFileLowering().getDwarfInfoSection(Hash: Signature);
4127 NewTU.setSection(Section);
4128 // Non-split type units reuse the compile unit's line table.
4129 CU.applyStmtList(D&: UnitDie);
4130 }
4131
4132 // Add DW_AT_str_offsets_base to the type unit DIE, but not for split type
4133 // units.
4134 if (useSegmentedStringOffsetsTable() && !useSplitDwarf())
4135 NewTU.addStringOffsetsStart();
4136
4137 NewTU.setType(NewTU.createTypeDIE(Ty: CTy));
4138
4139 if (TopLevelType) {
4140 auto TypeUnitsToAdd = std::move(TypeUnitsUnderConstruction);
4141 TypeUnitsUnderConstruction.clear();
4142
4143 // Types referencing entries in the address table cannot be placed in type
4144 // units.
4145 if (AddrPool.hasBeenUsed()) {
4146 AccelTypeUnitsDebugNames.clear();
4147 // Remove all the types built while building this type.
4148 // This is pessimistic as some of these types might not be dependent on
4149 // the type that used an address.
4150 for (const auto &TU : TypeUnitsToAdd)
4151 TypeSignatures.erase(Val: TU.second);
4152
4153 // Construct this type in the CU directly.
4154 // This is inefficient because all the dependent types will be rebuilt
4155 // from scratch, including building them in type units, discovering that
4156 // they depend on addresses, throwing them out and rebuilding them.
4157 setCurrentDWARF5AccelTable(DWARF5AccelTableKind::CU);
4158 CU.constructTypeDIE(Buffer&: RefDie, CTy: cast<DICompositeType>(Val: CTy));
4159 CU.updateAcceleratorTables(Context: CTy->getScope(), Ty: CTy, TyDIE: RefDie);
4160 return;
4161 }
4162
4163 // If the type wasn't dependent on fission addresses, finish adding the type
4164 // and all its dependent types.
4165 for (auto &TU : TypeUnitsToAdd) {
4166 InfoHolder.computeSizeAndOffsetsForUnit(TheU: TU.first.get());
4167 InfoHolder.emitUnit(TheU: TU.first.get(), UseOffsets: useSplitDwarf());
4168 if (getDwarfVersion() >= 5 &&
4169 getAccelTableKind() == AccelTableKind::Dwarf) {
4170 if (useSplitDwarf())
4171 AccelDebugNames.addTypeUnitSignature(U&: *TU.first);
4172 else
4173 AccelDebugNames.addTypeUnitSymbol(U&: *TU.first);
4174 }
4175 }
4176 AccelTypeUnitsDebugNames.convertDieToOffset();
4177 AccelDebugNames.addTypeEntries(Table&: AccelTypeUnitsDebugNames);
4178 AccelTypeUnitsDebugNames.clear();
4179 setCurrentDWARF5AccelTable(DWARF5AccelTableKind::CU);
4180 }
4181 CU.addDIETypeSignature(Die&: RefDie, Signature);
4182}
4183
4184// Add the Name along with its companion DIE to the appropriate accelerator
4185// table (for AccelTableKind::Dwarf it's always AccelDebugNames, for
4186// AccelTableKind::Apple, we use the table we got as an argument). If
4187// accelerator tables are disabled, this function does nothing.
4188template <typename DataT>
4189void DwarfDebug::addAccelNameImpl(
4190 const DwarfUnit &Unit,
4191 const DICompileUnit::DebugNameTableKind NameTableKind,
4192 AccelTable<DataT> &AppleAccel, StringRef Name, const DIE &Die) {
4193 if (getAccelTableKind() == AccelTableKind::None ||
4194 Unit.getUnitDie().getTag() == dwarf::DW_TAG_skeleton_unit || Name.empty())
4195 return;
4196
4197 if (getAccelTableKind() != AccelTableKind::Apple &&
4198 NameTableKind != DICompileUnit::DebugNameTableKind::Apple &&
4199 NameTableKind != DICompileUnit::DebugNameTableKind::Default)
4200 return;
4201
4202 DwarfFile &Holder = useSplitDwarf() ? SkeletonHolder : InfoHolder;
4203 DwarfStringPoolEntryRef Ref = Holder.getStringPool().getEntry(Asm&: *Asm, Str: Name);
4204
4205 switch (getAccelTableKind()) {
4206 case AccelTableKind::Apple:
4207 AppleAccel.addName(Ref, Die);
4208 break;
4209 case AccelTableKind::Dwarf: {
4210 DWARF5AccelTable &Current = getCurrentDWARF5AccelTable();
4211 assert(((&Current == &AccelTypeUnitsDebugNames) ||
4212 ((&Current == &AccelDebugNames) &&
4213 (Unit.getUnitDie().getTag() != dwarf::DW_TAG_type_unit))) &&
4214 "Kind is CU but TU is being processed.");
4215 assert(((&Current == &AccelDebugNames) ||
4216 ((&Current == &AccelTypeUnitsDebugNames) &&
4217 (Unit.getUnitDie().getTag() == dwarf::DW_TAG_type_unit))) &&
4218 "Kind is TU but CU is being processed.");
4219 // The type unit can be discarded, so need to add references to final
4220 // acceleration table once we know it's complete and we emit it.
4221 Current.addName(Name: Ref, Args: Die, Args: Unit.getUniqueID(),
4222 Args: Unit.getUnitDie().getTag() == dwarf::DW_TAG_type_unit);
4223 break;
4224 }
4225 case AccelTableKind::Default:
4226 llvm_unreachable("Default should have already been resolved.");
4227 case AccelTableKind::None:
4228 llvm_unreachable("None handled above");
4229 }
4230}
4231
4232void DwarfDebug::addAccelName(
4233 const DwarfUnit &Unit,
4234 const DICompileUnit::DebugNameTableKind NameTableKind, StringRef Name,
4235 const DIE &Die) {
4236 addAccelNameImpl(Unit, NameTableKind, AppleAccel&: AccelNames, Name, Die);
4237}
4238
4239void DwarfDebug::addAccelObjC(
4240 const DwarfUnit &Unit,
4241 const DICompileUnit::DebugNameTableKind NameTableKind, StringRef Name,
4242 const DIE &Die) {
4243 // ObjC names go only into the Apple accelerator tables.
4244 if (getAccelTableKind() == AccelTableKind::Apple)
4245 addAccelNameImpl(Unit, NameTableKind, AppleAccel&: AccelObjC, Name, Die);
4246}
4247
4248void DwarfDebug::addAccelNamespace(
4249 const DwarfUnit &Unit,
4250 const DICompileUnit::DebugNameTableKind NameTableKind, StringRef Name,
4251 const DIE &Die) {
4252 addAccelNameImpl(Unit, NameTableKind, AppleAccel&: AccelNamespace, Name, Die);
4253}
4254
4255void DwarfDebug::addAccelType(
4256 const DwarfUnit &Unit,
4257 const DICompileUnit::DebugNameTableKind NameTableKind, StringRef Name,
4258 const DIE &Die, char Flags) {
4259 addAccelNameImpl(Unit, NameTableKind, AppleAccel&: AccelTypes, Name, Die);
4260}
4261
4262uint16_t DwarfDebug::getDwarfVersion() const {
4263 return Asm->OutStreamer->getContext().getDwarfVersion();
4264}
4265
4266dwarf::Form DwarfDebug::getDwarfSectionOffsetForm() const {
4267 if (Asm->getDwarfVersion() >= 4)
4268 return dwarf::Form::DW_FORM_sec_offset;
4269 assert((!Asm->isDwarf64() || (Asm->getDwarfVersion() == 3)) &&
4270 "DWARF64 is not defined prior DWARFv3");
4271 return Asm->isDwarf64() ? dwarf::Form::DW_FORM_data8
4272 : dwarf::Form::DW_FORM_data4;
4273}
4274
4275const MCSymbol *DwarfDebug::getSectionLabel(const MCSection *S) {
4276 return SectionLabels.lookup(Val: S);
4277}
4278
4279void DwarfDebug::insertSectionLabel(const MCSymbol *S) {
4280 if (SectionLabels.insert(KV: std::make_pair(x: &S->getSection(), y&: S)).second)
4281 if (useSplitDwarf() || getDwarfVersion() >= 5)
4282 AddrPool.getIndex(Sym: S);
4283}
4284
4285std::optional<MD5::MD5Result>
4286DwarfDebug::getMD5AsBytes(const DIFile *File) const {
4287 assert(File);
4288 if (getDwarfVersion() < 5)
4289 return std::nullopt;
4290 std::optional<DIFile::ChecksumInfo<StringRef>> Checksum = File->getChecksum();
4291 if (!Checksum || Checksum->Kind != DIFile::CSK_MD5)
4292 return std::nullopt;
4293
4294 // Convert the string checksum to an MD5Result for the streamer.
4295 // The verifier validates the checksum so we assume it's okay.
4296 // An MD5 checksum is 16 bytes.
4297 std::string ChecksumString = fromHex(Input: Checksum->Value);
4298 MD5::MD5Result CKMem;
4299 llvm::copy(Range&: ChecksumString, Out: CKMem.data());
4300 return CKMem;
4301}
4302
4303bool DwarfDebug::alwaysUseRanges(const DwarfCompileUnit &CU) const {
4304 if (MinimizeAddr == MinimizeAddrInV5::Ranges)
4305 return true;
4306 if (MinimizeAddr != MinimizeAddrInV5::Default)
4307 return false;
4308 if (useSplitDwarf())
4309 return true;
4310 return false;
4311}
4312
4313void DwarfDebug::beginCodeAlignment(const MachineBasicBlock &MBB) {
4314 if (MBB.getAlignment() == Align(1))
4315 return;
4316
4317 auto *SP = MBB.getParent()->getFunction().getSubprogram();
4318 bool NoDebug =
4319 !SP || SP->getUnit()->getEmissionKind() == DICompileUnit::NoDebug;
4320
4321 if (NoDebug)
4322 return;
4323
4324 auto PrevLoc = Asm->OutStreamer->getContext().getCurrentDwarfLoc();
4325 if (PrevLoc.getLine()) {
4326 Asm->OutStreamer->emitDwarfLocDirective(
4327 FileNo: PrevLoc.getFileNum(), Line: 0, Column: PrevLoc.getColumn(), Flags: 0, Isa: 0, Discriminator: 0, FileName: StringRef());
4328 MCDwarfLineEntry::make(MCOS: Asm->OutStreamer.get(),
4329 Section: Asm->OutStreamer->getCurrentSectionOnly());
4330 }
4331}
4332