1 //===-- llvm/lib/CodeGen/AsmPrinter/WinCodeViewLineTables.cpp --*- C++ -*--===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains support for writing line tables info into COFF files.
12 //===----------------------------------------------------------------------===//
14 #include "WinCodeViewLineTables.h"
15 #include "llvm/MC/MCExpr.h"
16 #include "llvm/MC/MCSymbol.h"
17 #include "llvm/Support/COFF.h"
21 StringRef WinCodeViewLineTables::getFullFilepath(const MDNode *S) {
23 assert((isa<DICompileUnit>(S) || isa<DIFile>(S) || isa<DISubprogram>(S) ||
24 isa<DILexicalBlockBase>(S)) &&
25 "Unexpected scope info");
27 auto *Scope = cast<DIScope>(S);
28 StringRef Dir = Scope->getDirectory(),
29 Filename = Scope->getFilename();
30 std::string &Filepath =
31 DirAndFilenameToFilepathMap[std::make_pair(Dir, Filename)];
32 if (!Filepath.empty())
35 // Clang emits directory and relative filename info into the IR, but CodeView
36 // operates on full paths. We could change Clang to emit full paths too, but
37 // that would increase the IR size and probably not needed for other users.
38 // For now, just concatenate and canonicalize the path here.
39 if (Filename.find(':') == 1)
42 Filepath = (Dir + "\\" + Filename).str();
44 // Canonicalize the path. We have to do it textually because we may no longer
45 // have access the file in the filesystem.
46 // First, replace all slashes with backslashes.
47 std::replace(Filepath.begin(), Filepath.end(), '/', '\\');
49 // Remove all "\.\" with "\".
51 while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos)
52 Filepath.erase(Cursor, 2);
54 // Replace all "\XXX\..\" with "\". Don't try too hard though as the original
55 // path should be well-formatted, e.g. start with a drive letter, etc.
57 while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) {
58 // Something's wrong if the path starts with "\..\", abort.
62 size_t PrevSlash = Filepath.rfind('\\', Cursor - 1);
63 if (PrevSlash == std::string::npos)
64 // Something's wrong, abort.
67 Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash);
68 // The next ".." might be following the one we've just erased.
72 // Remove all duplicate backslashes.
74 while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos)
75 Filepath.erase(Cursor, 1);
80 void WinCodeViewLineTables::maybeRecordLocation(DebugLoc DL,
81 const MachineFunction *MF) {
82 const MDNode *Scope = DL.getScope();
85 unsigned LineNumber = DL.getLine();
86 // Skip this line if it is longer than the maximum we can record.
87 if (LineNumber > COFF::CVL_MaxLineNumber)
90 unsigned ColumnNumber = DL.getCol();
91 // Truncate the column number if it is longer than the maximum we can record.
92 if (ColumnNumber > COFF::CVL_MaxColumnNumber)
95 StringRef Filename = getFullFilepath(Scope);
97 // Skip this instruction if it has the same file:line as the previous one.
99 if (!CurFn->Instrs.empty()) {
100 const InstrInfoTy &LastInstr = InstrInfo[CurFn->Instrs.back()];
101 if (LastInstr.Filename == Filename && LastInstr.LineNumber == LineNumber &&
102 LastInstr.ColumnNumber == ColumnNumber)
105 FileNameRegistry.add(Filename);
107 MCSymbol *MCL = Asm->MMI->getContext().createTempSymbol();
108 Asm->OutStreamer->EmitLabel(MCL);
109 CurFn->Instrs.push_back(MCL);
110 InstrInfo[MCL] = InstrInfoTy(Filename, LineNumber, ColumnNumber);
113 WinCodeViewLineTables::WinCodeViewLineTables(AsmPrinter *AP)
114 : Asm(nullptr), CurFn(nullptr) {
115 MachineModuleInfo *MMI = AP->MMI;
117 // If module doesn't have named metadata anchors or COFF debug section
118 // is not available, skip any debug info related stuff.
119 if (!MMI->getModule()->getNamedMetadata("llvm.dbg.cu") ||
120 !AP->getObjFileLowering().getCOFFDebugSymbolsSection())
123 // Tell MMI that we have debug info.
124 MMI->setDebugInfoAvailability(true);
128 void WinCodeViewLineTables::endModule() {
129 if (FnDebugInfo.empty())
132 assert(Asm != nullptr);
133 Asm->OutStreamer->SwitchSection(
134 Asm->getObjFileLowering().getCOFFDebugSymbolsSection());
135 Asm->EmitInt32(COFF::DEBUG_SECTION_MAGIC);
137 // The COFF .debug$S section consists of several subsections, each starting
138 // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length
139 // of the payload followed by the payload itself. The subsections are 4-byte
142 // Emit per-function debug information. This code is extracted into a
143 // separate function for readability.
144 for (size_t I = 0, E = VisitedFunctions.size(); I != E; ++I)
145 emitDebugInfoForFunction(VisitedFunctions[I]);
147 // This subsection holds a file index to offset in string table table.
148 Asm->OutStreamer->AddComment("File index to string table offset subsection");
149 Asm->EmitInt32(COFF::DEBUG_INDEX_SUBSECTION);
150 size_t NumFilenames = FileNameRegistry.Infos.size();
151 Asm->EmitInt32(8 * NumFilenames);
152 for (size_t I = 0, E = FileNameRegistry.Filenames.size(); I != E; ++I) {
153 StringRef Filename = FileNameRegistry.Filenames[I];
154 // For each unique filename, just write its offset in the string table.
155 Asm->EmitInt32(FileNameRegistry.Infos[Filename].StartOffset);
156 // The function name offset is not followed by any additional data.
160 // This subsection holds the string table.
161 Asm->OutStreamer->AddComment("String table");
162 Asm->EmitInt32(COFF::DEBUG_STRING_TABLE_SUBSECTION);
163 Asm->EmitInt32(FileNameRegistry.LastOffset);
164 // The payload starts with a null character.
167 for (size_t I = 0, E = FileNameRegistry.Filenames.size(); I != E; ++I) {
168 // Just emit unique filenames one by one, separated by a null character.
169 Asm->OutStreamer->EmitBytes(FileNameRegistry.Filenames[I]);
173 // No more subsections. Fill with zeros to align the end of the section by 4.
174 Asm->OutStreamer->EmitFill((-FileNameRegistry.LastOffset) % 4, 0);
179 static void EmitLabelDiff(MCStreamer &Streamer,
180 const MCSymbol *From, const MCSymbol *To,
181 unsigned int Size = 4) {
182 MCSymbolRefExpr::VariantKind Variant = MCSymbolRefExpr::VK_None;
183 MCContext &Context = Streamer.getContext();
184 const MCExpr *FromRef = MCSymbolRefExpr::create(From, Variant, Context),
185 *ToRef = MCSymbolRefExpr::create(To, Variant, Context);
186 const MCExpr *AddrDelta =
187 MCBinaryExpr::create(MCBinaryExpr::Sub, ToRef, FromRef, Context);
188 Streamer.EmitValue(AddrDelta, Size);
191 void WinCodeViewLineTables::emitDebugInfoForFunction(const Function *GV) {
192 // For each function there is a separate subsection
193 // which holds the PC to file:line table.
194 const MCSymbol *Fn = Asm->getSymbol(GV);
197 const FunctionInfo &FI = FnDebugInfo[GV];
198 if (FI.Instrs.empty())
200 assert(FI.End && "Don't know where the function ends?");
202 StringRef GVName = GV->getName();
204 if (auto *SP = getDISubprogram(GV))
205 FuncName = SP->getDisplayName();
207 // FIXME Clang currently sets DisplayName to "bar" for a C++
208 // "namespace_foo::bar" function, see PR21528. Luckily, dbghelp.dll is trying
209 // to demangle display names anyways, so let's just put a mangled name into
210 // the symbols subsection until Clang gives us what we need.
211 if (GVName.startswith("\01?"))
212 FuncName = GVName.substr(1);
213 // Emit a symbol subsection, required by VS2012+ to find function boundaries.
214 MCSymbol *SymbolsBegin = Asm->MMI->getContext().createTempSymbol(),
215 *SymbolsEnd = Asm->MMI->getContext().createTempSymbol();
216 Asm->OutStreamer->AddComment("Symbol subsection for " + Twine(FuncName));
217 Asm->EmitInt32(COFF::DEBUG_SYMBOL_SUBSECTION);
218 EmitLabelDiff(*Asm->OutStreamer, SymbolsBegin, SymbolsEnd);
219 Asm->OutStreamer->EmitLabel(SymbolsBegin);
221 MCSymbol *ProcSegmentBegin = Asm->MMI->getContext().createTempSymbol(),
222 *ProcSegmentEnd = Asm->MMI->getContext().createTempSymbol();
223 EmitLabelDiff(*Asm->OutStreamer, ProcSegmentBegin, ProcSegmentEnd, 2);
224 Asm->OutStreamer->EmitLabel(ProcSegmentBegin);
226 Asm->EmitInt16(COFF::DEBUG_SYMBOL_TYPE_PROC_START);
227 // Some bytes of this segment don't seem to be required for basic debugging,
228 // so just fill them with zeroes.
229 Asm->OutStreamer->EmitFill(12, 0);
230 // This is the important bit that tells the debugger where the function
231 // code is located and what's its size:
232 EmitLabelDiff(*Asm->OutStreamer, Fn, FI.End);
233 Asm->OutStreamer->EmitFill(12, 0);
234 Asm->OutStreamer->EmitCOFFSecRel32(Fn);
235 Asm->OutStreamer->EmitCOFFSectionIndex(Fn);
237 // Emit the function display name as a null-terminated string.
238 Asm->OutStreamer->EmitBytes(FuncName);
240 Asm->OutStreamer->EmitLabel(ProcSegmentEnd);
242 // We're done with this function.
243 Asm->EmitInt16(0x0002);
244 Asm->EmitInt16(COFF::DEBUG_SYMBOL_TYPE_PROC_END);
246 Asm->OutStreamer->EmitLabel(SymbolsEnd);
247 // Every subsection must be aligned to a 4-byte boundary.
248 Asm->OutStreamer->EmitFill((-FuncName.size()) % 4, 0);
250 // PCs/Instructions are grouped into segments sharing the same filename.
251 // Pre-calculate the lengths (in instructions) of these segments and store
252 // them in a map for convenience. Each index in the map is the sequential
253 // number of the respective instruction that starts a new segment.
254 DenseMap<size_t, size_t> FilenameSegmentLengths;
255 size_t LastSegmentEnd = 0;
256 StringRef PrevFilename = InstrInfo[FI.Instrs[0]].Filename;
257 for (size_t J = 1, F = FI.Instrs.size(); J != F; ++J) {
258 if (PrevFilename == InstrInfo[FI.Instrs[J]].Filename)
260 FilenameSegmentLengths[LastSegmentEnd] = J - LastSegmentEnd;
262 PrevFilename = InstrInfo[FI.Instrs[J]].Filename;
264 FilenameSegmentLengths[LastSegmentEnd] = FI.Instrs.size() - LastSegmentEnd;
266 // Emit a line table subsection, required to do PC-to-file:line lookup.
267 Asm->OutStreamer->AddComment("Line table subsection for " + Twine(FuncName));
268 Asm->EmitInt32(COFF::DEBUG_LINE_TABLE_SUBSECTION);
269 MCSymbol *LineTableBegin = Asm->MMI->getContext().createTempSymbol(),
270 *LineTableEnd = Asm->MMI->getContext().createTempSymbol();
271 EmitLabelDiff(*Asm->OutStreamer, LineTableBegin, LineTableEnd);
272 Asm->OutStreamer->EmitLabel(LineTableBegin);
274 // Identify the function this subsection is for.
275 Asm->OutStreamer->EmitCOFFSecRel32(Fn);
276 Asm->OutStreamer->EmitCOFFSectionIndex(Fn);
277 // Insert flags after a 16-bit section index.
278 Asm->EmitInt16(COFF::DEBUG_LINE_TABLES_HAVE_COLUMN_RECORDS);
280 // Length of the function's code, in bytes.
281 EmitLabelDiff(*Asm->OutStreamer, Fn, FI.End);
283 // PC-to-linenumber lookup table:
284 MCSymbol *FileSegmentEnd = nullptr;
286 // The start of the last segment:
287 size_t LastSegmentStart = 0;
289 auto FinishPreviousChunk = [&] {
292 for (size_t ColSegI = LastSegmentStart,
293 ColSegEnd = ColSegI + FilenameSegmentLengths[LastSegmentStart];
294 ColSegI != ColSegEnd; ++ColSegI) {
295 unsigned ColumnNumber = InstrInfo[FI.Instrs[ColSegI]].ColumnNumber;
296 assert(ColumnNumber <= COFF::CVL_MaxColumnNumber);
297 Asm->EmitInt16(ColumnNumber); // Start column
298 Asm->EmitInt16(0); // End column
300 Asm->OutStreamer->EmitLabel(FileSegmentEnd);
303 for (size_t J = 0, F = FI.Instrs.size(); J != F; ++J) {
304 MCSymbol *Instr = FI.Instrs[J];
305 assert(InstrInfo.count(Instr));
307 if (FilenameSegmentLengths.count(J)) {
308 // We came to a beginning of a new filename segment.
309 FinishPreviousChunk();
310 StringRef CurFilename = InstrInfo[FI.Instrs[J]].Filename;
311 assert(FileNameRegistry.Infos.count(CurFilename));
312 size_t IndexInStringTable =
313 FileNameRegistry.Infos[CurFilename].FilenameID;
314 // Each segment starts with the offset of the filename
315 // in the string table.
316 Asm->OutStreamer->AddComment(
317 "Segment for file '" + Twine(CurFilename) + "' begins");
318 MCSymbol *FileSegmentBegin = Asm->MMI->getContext().createTempSymbol();
319 Asm->OutStreamer->EmitLabel(FileSegmentBegin);
320 Asm->EmitInt32(8 * IndexInStringTable);
322 // Number of PC records in the lookup table.
323 size_t SegmentLength = FilenameSegmentLengths[J];
324 Asm->EmitInt32(SegmentLength);
326 // Full size of the segment for this filename, including the prev two
328 FileSegmentEnd = Asm->MMI->getContext().createTempSymbol();
329 EmitLabelDiff(*Asm->OutStreamer, FileSegmentBegin, FileSegmentEnd);
330 LastSegmentStart = J;
333 // The first PC with the given linenumber and the linenumber itself.
334 EmitLabelDiff(*Asm->OutStreamer, Fn, Instr);
335 uint32_t LineNumber = InstrInfo[Instr].LineNumber;
336 assert(LineNumber <= COFF::CVL_MaxLineNumber);
337 uint32_t LineData = LineNumber | COFF::CVL_IsStatement;
338 Asm->EmitInt32(LineData);
341 FinishPreviousChunk();
342 Asm->OutStreamer->EmitLabel(LineTableEnd);
345 void WinCodeViewLineTables::beginFunction(const MachineFunction *MF) {
346 assert(!CurFn && "Can't process two functions at once!");
348 if (!Asm || !Asm->MMI->hasDebugInfo())
351 const Function *GV = MF->getFunction();
352 assert(FnDebugInfo.count(GV) == false);
353 VisitedFunctions.push_back(GV);
354 CurFn = &FnDebugInfo[GV];
356 // Find the end of the function prolog.
357 // FIXME: is there a simpler a way to do this? Can we just search
358 // for the first instruction of the function, not the last of the prolog?
359 DebugLoc PrologEndLoc;
360 bool EmptyPrologue = true;
361 for (const auto &MBB : *MF) {
364 for (const auto &MI : MBB) {
365 if (MI.isDebugValue())
368 // First known non-DBG_VALUE and non-frame setup location marks
369 // the beginning of the function body.
370 // FIXME: do we need the first subcondition?
371 if (!MI.getFlag(MachineInstr::FrameSetup) && MI.getDebugLoc()) {
372 PrologEndLoc = MI.getDebugLoc();
375 EmptyPrologue = false;
378 // Record beginning of function if we have a non-empty prologue.
379 if (PrologEndLoc && !EmptyPrologue) {
380 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc();
381 maybeRecordLocation(FnStartDL, MF);
385 void WinCodeViewLineTables::endFunction(const MachineFunction *MF) {
386 if (!Asm || !CurFn) // We haven't created any debug info for this function.
389 const Function *GV = MF->getFunction();
390 assert(FnDebugInfo.count(GV));
391 assert(CurFn == &FnDebugInfo[GV]);
393 if (CurFn->Instrs.empty()) {
394 FnDebugInfo.erase(GV);
395 VisitedFunctions.pop_back();
397 CurFn->End = Asm->getFunctionEnd();
402 void WinCodeViewLineTables::beginInstruction(const MachineInstr *MI) {
403 // Ignore DBG_VALUE locations and function prologue.
404 if (!Asm || MI->isDebugValue() || MI->getFlag(MachineInstr::FrameSetup))
406 DebugLoc DL = MI->getDebugLoc();
407 if (DL == PrevInstLoc || !DL)
409 maybeRecordLocation(DL, Asm->MF);