xref: /freebsd/contrib/llvm-project/lld/COFF/Writer.cpp (revision d4eeb02986980bf33dd56c41ceb9fc5f180c0d47)
1 //===- Writer.cpp ---------------------------------------------------------===//
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 #include "Writer.h"
10 #include "COFFLinkerContext.h"
11 #include "CallGraphSort.h"
12 #include "Config.h"
13 #include "DLL.h"
14 #include "InputFiles.h"
15 #include "LLDMapFile.h"
16 #include "MapFile.h"
17 #include "PDB.h"
18 #include "SymbolTable.h"
19 #include "Symbols.h"
20 #include "lld/Common/ErrorHandler.h"
21 #include "lld/Common/Memory.h"
22 #include "lld/Common/Timer.h"
23 #include "llvm/ADT/DenseMap.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/ADT/StringSet.h"
26 #include "llvm/ADT/StringSwitch.h"
27 #include "llvm/Support/BinaryStreamReader.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/Endian.h"
30 #include "llvm/Support/FileOutputBuffer.h"
31 #include "llvm/Support/Parallel.h"
32 #include "llvm/Support/Path.h"
33 #include "llvm/Support/RandomNumberGenerator.h"
34 #include "llvm/Support/xxhash.h"
35 #include <algorithm>
36 #include <cstdio>
37 #include <map>
38 #include <memory>
39 #include <utility>
40 
41 using namespace llvm;
42 using namespace llvm::COFF;
43 using namespace llvm::object;
44 using namespace llvm::support;
45 using namespace llvm::support::endian;
46 using namespace lld;
47 using namespace lld::coff;
48 
49 /* To re-generate DOSProgram:
50 $ cat > /tmp/DOSProgram.asm
51 org 0
52         ; Copy cs to ds.
53         push cs
54         pop ds
55         ; Point ds:dx at the $-terminated string.
56         mov dx, str
57         ; Int 21/AH=09h: Write string to standard output.
58         mov ah, 0x9
59         int 0x21
60         ; Int 21/AH=4Ch: Exit with return code (in AL).
61         mov ax, 0x4C01
62         int 0x21
63 str:
64         db 'This program cannot be run in DOS mode.$'
65 align 8, db 0
66 $ nasm -fbin /tmp/DOSProgram.asm -o /tmp/DOSProgram.bin
67 $ xxd -i /tmp/DOSProgram.bin
68 */
69 static unsigned char dosProgram[] = {
70   0x0e, 0x1f, 0xba, 0x0e, 0x00, 0xb4, 0x09, 0xcd, 0x21, 0xb8, 0x01, 0x4c,
71   0xcd, 0x21, 0x54, 0x68, 0x69, 0x73, 0x20, 0x70, 0x72, 0x6f, 0x67, 0x72,
72   0x61, 0x6d, 0x20, 0x63, 0x61, 0x6e, 0x6e, 0x6f, 0x74, 0x20, 0x62, 0x65,
73   0x20, 0x72, 0x75, 0x6e, 0x20, 0x69, 0x6e, 0x20, 0x44, 0x4f, 0x53, 0x20,
74   0x6d, 0x6f, 0x64, 0x65, 0x2e, 0x24, 0x00, 0x00
75 };
76 static_assert(sizeof(dosProgram) % 8 == 0,
77               "DOSProgram size must be multiple of 8");
78 
79 static const int dosStubSize = sizeof(dos_header) + sizeof(dosProgram);
80 static_assert(dosStubSize % 8 == 0, "DOSStub size must be multiple of 8");
81 
82 static const int numberOfDataDirectory = 16;
83 
84 namespace {
85 
86 class DebugDirectoryChunk : public NonSectionChunk {
87 public:
88   DebugDirectoryChunk(COFFLinkerContext &c,
89                       const std::vector<std::pair<COFF::DebugType, Chunk *>> &r,
90                       bool writeRepro)
91       : records(r), writeRepro(writeRepro), ctx(c) {}
92 
93   size_t getSize() const override {
94     return (records.size() + int(writeRepro)) * sizeof(debug_directory);
95   }
96 
97   void writeTo(uint8_t *b) const override {
98     auto *d = reinterpret_cast<debug_directory *>(b);
99 
100     for (const std::pair<COFF::DebugType, Chunk *>& record : records) {
101       Chunk *c = record.second;
102       OutputSection *os = ctx.getOutputSection(c);
103       uint64_t offs = os->getFileOff() + (c->getRVA() - os->getRVA());
104       fillEntry(d, record.first, c->getSize(), c->getRVA(), offs);
105       ++d;
106     }
107 
108     if (writeRepro) {
109       // FIXME: The COFF spec allows either a 0-sized entry to just say
110       // "the timestamp field is really a hash", or a 4-byte size field
111       // followed by that many bytes containing a longer hash (with the
112       // lowest 4 bytes usually being the timestamp in little-endian order).
113       // Consider storing the full 8 bytes computed by xxHash64 here.
114       fillEntry(d, COFF::IMAGE_DEBUG_TYPE_REPRO, 0, 0, 0);
115     }
116   }
117 
118   void setTimeDateStamp(uint32_t timeDateStamp) {
119     for (support::ulittle32_t *tds : timeDateStamps)
120       *tds = timeDateStamp;
121   }
122 
123 private:
124   void fillEntry(debug_directory *d, COFF::DebugType debugType, size_t size,
125                  uint64_t rva, uint64_t offs) const {
126     d->Characteristics = 0;
127     d->TimeDateStamp = 0;
128     d->MajorVersion = 0;
129     d->MinorVersion = 0;
130     d->Type = debugType;
131     d->SizeOfData = size;
132     d->AddressOfRawData = rva;
133     d->PointerToRawData = offs;
134 
135     timeDateStamps.push_back(&d->TimeDateStamp);
136   }
137 
138   mutable std::vector<support::ulittle32_t *> timeDateStamps;
139   const std::vector<std::pair<COFF::DebugType, Chunk *>> &records;
140   bool writeRepro;
141 
142   COFFLinkerContext &ctx;
143 };
144 
145 class CVDebugRecordChunk : public NonSectionChunk {
146 public:
147   size_t getSize() const override {
148     return sizeof(codeview::DebugInfo) + config->pdbAltPath.size() + 1;
149   }
150 
151   void writeTo(uint8_t *b) const override {
152     // Save off the DebugInfo entry to backfill the file signature (build id)
153     // in Writer::writeBuildId
154     buildId = reinterpret_cast<codeview::DebugInfo *>(b);
155 
156     // variable sized field (PDB Path)
157     char *p = reinterpret_cast<char *>(b + sizeof(*buildId));
158     if (!config->pdbAltPath.empty())
159       memcpy(p, config->pdbAltPath.data(), config->pdbAltPath.size());
160     p[config->pdbAltPath.size()] = '\0';
161   }
162 
163   mutable codeview::DebugInfo *buildId = nullptr;
164 };
165 
166 class ExtendedDllCharacteristicsChunk : public NonSectionChunk {
167 public:
168   ExtendedDllCharacteristicsChunk(uint32_t c) : characteristics(c) {}
169 
170   size_t getSize() const override { return 4; }
171 
172   void writeTo(uint8_t *buf) const override { write32le(buf, characteristics); }
173 
174   uint32_t characteristics = 0;
175 };
176 
177 // PartialSection represents a group of chunks that contribute to an
178 // OutputSection. Collating a collection of PartialSections of same name and
179 // characteristics constitutes the OutputSection.
180 class PartialSectionKey {
181 public:
182   StringRef name;
183   unsigned characteristics;
184 
185   bool operator<(const PartialSectionKey &other) const {
186     int c = name.compare(other.name);
187     if (c == 1)
188       return false;
189     if (c == 0)
190       return characteristics < other.characteristics;
191     return true;
192   }
193 };
194 
195 // The writer writes a SymbolTable result to a file.
196 class Writer {
197 public:
198   Writer(COFFLinkerContext &c) : buffer(errorHandler().outputBuffer), ctx(c) {}
199   void run();
200 
201 private:
202   void createSections();
203   void createMiscChunks();
204   void createImportTables();
205   void appendImportThunks();
206   void locateImportTables();
207   void createExportTable();
208   void mergeSections();
209   void removeUnusedSections();
210   void assignAddresses();
211   void finalizeAddresses();
212   void removeEmptySections();
213   void assignOutputSectionIndices();
214   void createSymbolAndStringTable();
215   void openFile(StringRef outputPath);
216   template <typename PEHeaderTy> void writeHeader();
217   void createSEHTable();
218   void createRuntimePseudoRelocs();
219   void insertCtorDtorSymbols();
220   void createGuardCFTables();
221   void markSymbolsForRVATable(ObjFile *file,
222                               ArrayRef<SectionChunk *> symIdxChunks,
223                               SymbolRVASet &tableSymbols);
224   void getSymbolsFromSections(ObjFile *file,
225                               ArrayRef<SectionChunk *> symIdxChunks,
226                               std::vector<Symbol *> &symbols);
227   void maybeAddRVATable(SymbolRVASet tableSymbols, StringRef tableSym,
228                         StringRef countSym, bool hasFlag=false);
229   void setSectionPermissions();
230   void writeSections();
231   void writeBuildId();
232   void sortSections();
233   void sortExceptionTable();
234   void sortCRTSectionChunks(std::vector<Chunk *> &chunks);
235   void addSyntheticIdata();
236   void fixPartialSectionChars(StringRef name, uint32_t chars);
237   bool fixGnuImportChunks();
238   void fixTlsAlignment();
239   PartialSection *createPartialSection(StringRef name, uint32_t outChars);
240   PartialSection *findPartialSection(StringRef name, uint32_t outChars);
241 
242   llvm::Optional<coff_symbol16> createSymbol(Defined *d);
243   size_t addEntryToStringTable(StringRef str);
244 
245   OutputSection *findSection(StringRef name);
246   void addBaserels();
247   void addBaserelBlocks(std::vector<Baserel> &v);
248 
249   uint32_t getSizeOfInitializedData();
250 
251   std::unique_ptr<FileOutputBuffer> &buffer;
252   std::map<PartialSectionKey, PartialSection *> partialSections;
253   std::vector<char> strtab;
254   std::vector<llvm::object::coff_symbol16> outputSymtab;
255   IdataContents idata;
256   Chunk *importTableStart = nullptr;
257   uint64_t importTableSize = 0;
258   Chunk *edataStart = nullptr;
259   Chunk *edataEnd = nullptr;
260   Chunk *iatStart = nullptr;
261   uint64_t iatSize = 0;
262   DelayLoadContents delayIdata;
263   EdataContents edata;
264   bool setNoSEHCharacteristic = false;
265   uint32_t tlsAlignment = 0;
266 
267   DebugDirectoryChunk *debugDirectory = nullptr;
268   std::vector<std::pair<COFF::DebugType, Chunk *>> debugRecords;
269   CVDebugRecordChunk *buildId = nullptr;
270   ArrayRef<uint8_t> sectionTable;
271 
272   uint64_t fileSize;
273   uint32_t pointerToSymbolTable = 0;
274   uint64_t sizeOfImage;
275   uint64_t sizeOfHeaders;
276 
277   OutputSection *textSec;
278   OutputSection *rdataSec;
279   OutputSection *buildidSec;
280   OutputSection *dataSec;
281   OutputSection *pdataSec;
282   OutputSection *idataSec;
283   OutputSection *edataSec;
284   OutputSection *didatSec;
285   OutputSection *rsrcSec;
286   OutputSection *relocSec;
287   OutputSection *ctorsSec;
288   OutputSection *dtorsSec;
289 
290   // The first and last .pdata sections in the output file.
291   //
292   // We need to keep track of the location of .pdata in whichever section it
293   // gets merged into so that we can sort its contents and emit a correct data
294   // directory entry for the exception table. This is also the case for some
295   // other sections (such as .edata) but because the contents of those sections
296   // are entirely linker-generated we can keep track of their locations using
297   // the chunks that the linker creates. All .pdata chunks come from input
298   // files, so we need to keep track of them separately.
299   Chunk *firstPdata = nullptr;
300   Chunk *lastPdata;
301 
302   COFFLinkerContext &ctx;
303 };
304 } // anonymous namespace
305 
306 void lld::coff::writeResult(COFFLinkerContext &ctx) { Writer(ctx).run(); }
307 
308 void OutputSection::addChunk(Chunk *c) {
309   chunks.push_back(c);
310 }
311 
312 void OutputSection::insertChunkAtStart(Chunk *c) {
313   chunks.insert(chunks.begin(), c);
314 }
315 
316 void OutputSection::setPermissions(uint32_t c) {
317   header.Characteristics &= ~permMask;
318   header.Characteristics |= c;
319 }
320 
321 void OutputSection::merge(OutputSection *other) {
322   chunks.insert(chunks.end(), other->chunks.begin(), other->chunks.end());
323   other->chunks.clear();
324   contribSections.insert(contribSections.end(), other->contribSections.begin(),
325                          other->contribSections.end());
326   other->contribSections.clear();
327 }
328 
329 // Write the section header to a given buffer.
330 void OutputSection::writeHeaderTo(uint8_t *buf) {
331   auto *hdr = reinterpret_cast<coff_section *>(buf);
332   *hdr = header;
333   if (stringTableOff) {
334     // If name is too long, write offset into the string table as a name.
335     sprintf(hdr->Name, "/%d", stringTableOff);
336   } else {
337     assert(!config->debug || name.size() <= COFF::NameSize ||
338            (hdr->Characteristics & IMAGE_SCN_MEM_DISCARDABLE) == 0);
339     strncpy(hdr->Name, name.data(),
340             std::min(name.size(), (size_t)COFF::NameSize));
341   }
342 }
343 
344 void OutputSection::addContributingPartialSection(PartialSection *sec) {
345   contribSections.push_back(sec);
346 }
347 
348 // Check whether the target address S is in range from a relocation
349 // of type relType at address P.
350 static bool isInRange(uint16_t relType, uint64_t s, uint64_t p, int margin) {
351   if (config->machine == ARMNT) {
352     int64_t diff = AbsoluteDifference(s, p + 4) + margin;
353     switch (relType) {
354     case IMAGE_REL_ARM_BRANCH20T:
355       return isInt<21>(diff);
356     case IMAGE_REL_ARM_BRANCH24T:
357     case IMAGE_REL_ARM_BLX23T:
358       return isInt<25>(diff);
359     default:
360       return true;
361     }
362   } else if (config->machine == ARM64) {
363     int64_t diff = AbsoluteDifference(s, p) + margin;
364     switch (relType) {
365     case IMAGE_REL_ARM64_BRANCH26:
366       return isInt<28>(diff);
367     case IMAGE_REL_ARM64_BRANCH19:
368       return isInt<21>(diff);
369     case IMAGE_REL_ARM64_BRANCH14:
370       return isInt<16>(diff);
371     default:
372       return true;
373     }
374   } else {
375     llvm_unreachable("Unexpected architecture");
376   }
377 }
378 
379 // Return the last thunk for the given target if it is in range,
380 // or create a new one.
381 static std::pair<Defined *, bool>
382 getThunk(DenseMap<uint64_t, Defined *> &lastThunks, Defined *target, uint64_t p,
383          uint16_t type, int margin) {
384   Defined *&lastThunk = lastThunks[target->getRVA()];
385   if (lastThunk && isInRange(type, lastThunk->getRVA(), p, margin))
386     return {lastThunk, false};
387   Chunk *c;
388   switch (config->machine) {
389   case ARMNT:
390     c = make<RangeExtensionThunkARM>(target);
391     break;
392   case ARM64:
393     c = make<RangeExtensionThunkARM64>(target);
394     break;
395   default:
396     llvm_unreachable("Unexpected architecture");
397   }
398   Defined *d = make<DefinedSynthetic>("", c);
399   lastThunk = d;
400   return {d, true};
401 }
402 
403 // This checks all relocations, and for any relocation which isn't in range
404 // it adds a thunk after the section chunk that contains the relocation.
405 // If the latest thunk for the specific target is in range, that is used
406 // instead of creating a new thunk. All range checks are done with the
407 // specified margin, to make sure that relocations that originally are in
408 // range, but only barely, also get thunks - in case other added thunks makes
409 // the target go out of range.
410 //
411 // After adding thunks, we verify that all relocations are in range (with
412 // no extra margin requirements). If this failed, we restart (throwing away
413 // the previously created thunks) and retry with a wider margin.
414 static bool createThunks(OutputSection *os, int margin) {
415   bool addressesChanged = false;
416   DenseMap<uint64_t, Defined *> lastThunks;
417   DenseMap<std::pair<ObjFile *, Defined *>, uint32_t> thunkSymtabIndices;
418   size_t thunksSize = 0;
419   // Recheck Chunks.size() each iteration, since we can insert more
420   // elements into it.
421   for (size_t i = 0; i != os->chunks.size(); ++i) {
422     SectionChunk *sc = dyn_cast_or_null<SectionChunk>(os->chunks[i]);
423     if (!sc)
424       continue;
425     size_t thunkInsertionSpot = i + 1;
426 
427     // Try to get a good enough estimate of where new thunks will be placed.
428     // Offset this by the size of the new thunks added so far, to make the
429     // estimate slightly better.
430     size_t thunkInsertionRVA = sc->getRVA() + sc->getSize() + thunksSize;
431     ObjFile *file = sc->file;
432     std::vector<std::pair<uint32_t, uint32_t>> relocReplacements;
433     ArrayRef<coff_relocation> originalRelocs =
434         file->getCOFFObj()->getRelocations(sc->header);
435     for (size_t j = 0, e = originalRelocs.size(); j < e; ++j) {
436       const coff_relocation &rel = originalRelocs[j];
437       Symbol *relocTarget = file->getSymbol(rel.SymbolTableIndex);
438 
439       // The estimate of the source address P should be pretty accurate,
440       // but we don't know whether the target Symbol address should be
441       // offset by thunksSize or not (or by some of thunksSize but not all of
442       // it), giving us some uncertainty once we have added one thunk.
443       uint64_t p = sc->getRVA() + rel.VirtualAddress + thunksSize;
444 
445       Defined *sym = dyn_cast_or_null<Defined>(relocTarget);
446       if (!sym)
447         continue;
448 
449       uint64_t s = sym->getRVA();
450 
451       if (isInRange(rel.Type, s, p, margin))
452         continue;
453 
454       // If the target isn't in range, hook it up to an existing or new
455       // thunk.
456       Defined *thunk;
457       bool wasNew;
458       std::tie(thunk, wasNew) = getThunk(lastThunks, sym, p, rel.Type, margin);
459       if (wasNew) {
460         Chunk *thunkChunk = thunk->getChunk();
461         thunkChunk->setRVA(
462             thunkInsertionRVA); // Estimate of where it will be located.
463         os->chunks.insert(os->chunks.begin() + thunkInsertionSpot, thunkChunk);
464         thunkInsertionSpot++;
465         thunksSize += thunkChunk->getSize();
466         thunkInsertionRVA += thunkChunk->getSize();
467         addressesChanged = true;
468       }
469 
470       // To redirect the relocation, add a symbol to the parent object file's
471       // symbol table, and replace the relocation symbol table index with the
472       // new index.
473       auto insertion = thunkSymtabIndices.insert({{file, thunk}, ~0U});
474       uint32_t &thunkSymbolIndex = insertion.first->second;
475       if (insertion.second)
476         thunkSymbolIndex = file->addRangeThunkSymbol(thunk);
477       relocReplacements.push_back({j, thunkSymbolIndex});
478     }
479 
480     // Get a writable copy of this section's relocations so they can be
481     // modified. If the relocations point into the object file, allocate new
482     // memory. Otherwise, this must be previously allocated memory that can be
483     // modified in place.
484     ArrayRef<coff_relocation> curRelocs = sc->getRelocs();
485     MutableArrayRef<coff_relocation> newRelocs;
486     if (originalRelocs.data() == curRelocs.data()) {
487       newRelocs = makeMutableArrayRef(
488           bAlloc().Allocate<coff_relocation>(originalRelocs.size()),
489           originalRelocs.size());
490     } else {
491       newRelocs = makeMutableArrayRef(
492           const_cast<coff_relocation *>(curRelocs.data()), curRelocs.size());
493     }
494 
495     // Copy each relocation, but replace the symbol table indices which need
496     // thunks.
497     auto nextReplacement = relocReplacements.begin();
498     auto endReplacement = relocReplacements.end();
499     for (size_t i = 0, e = originalRelocs.size(); i != e; ++i) {
500       newRelocs[i] = originalRelocs[i];
501       if (nextReplacement != endReplacement && nextReplacement->first == i) {
502         newRelocs[i].SymbolTableIndex = nextReplacement->second;
503         ++nextReplacement;
504       }
505     }
506 
507     sc->setRelocs(newRelocs);
508   }
509   return addressesChanged;
510 }
511 
512 // Verify that all relocations are in range, with no extra margin requirements.
513 static bool verifyRanges(const std::vector<Chunk *> chunks) {
514   for (Chunk *c : chunks) {
515     SectionChunk *sc = dyn_cast_or_null<SectionChunk>(c);
516     if (!sc)
517       continue;
518 
519     ArrayRef<coff_relocation> relocs = sc->getRelocs();
520     for (size_t j = 0, e = relocs.size(); j < e; ++j) {
521       const coff_relocation &rel = relocs[j];
522       Symbol *relocTarget = sc->file->getSymbol(rel.SymbolTableIndex);
523 
524       Defined *sym = dyn_cast_or_null<Defined>(relocTarget);
525       if (!sym)
526         continue;
527 
528       uint64_t p = sc->getRVA() + rel.VirtualAddress;
529       uint64_t s = sym->getRVA();
530 
531       if (!isInRange(rel.Type, s, p, 0))
532         return false;
533     }
534   }
535   return true;
536 }
537 
538 // Assign addresses and add thunks if necessary.
539 void Writer::finalizeAddresses() {
540   assignAddresses();
541   if (config->machine != ARMNT && config->machine != ARM64)
542     return;
543 
544   size_t origNumChunks = 0;
545   for (OutputSection *sec : ctx.outputSections) {
546     sec->origChunks = sec->chunks;
547     origNumChunks += sec->chunks.size();
548   }
549 
550   int pass = 0;
551   int margin = 1024 * 100;
552   while (true) {
553     // First check whether we need thunks at all, or if the previous pass of
554     // adding them turned out ok.
555     bool rangesOk = true;
556     size_t numChunks = 0;
557     for (OutputSection *sec : ctx.outputSections) {
558       if (!verifyRanges(sec->chunks)) {
559         rangesOk = false;
560         break;
561       }
562       numChunks += sec->chunks.size();
563     }
564     if (rangesOk) {
565       if (pass > 0)
566         log("Added " + Twine(numChunks - origNumChunks) + " thunks with " +
567             "margin " + Twine(margin) + " in " + Twine(pass) + " passes");
568       return;
569     }
570 
571     if (pass >= 10)
572       fatal("adding thunks hasn't converged after " + Twine(pass) + " passes");
573 
574     if (pass > 0) {
575       // If the previous pass didn't work out, reset everything back to the
576       // original conditions before retrying with a wider margin. This should
577       // ideally never happen under real circumstances.
578       for (OutputSection *sec : ctx.outputSections)
579         sec->chunks = sec->origChunks;
580       margin *= 2;
581     }
582 
583     // Try adding thunks everywhere where it is needed, with a margin
584     // to avoid things going out of range due to the added thunks.
585     bool addressesChanged = false;
586     for (OutputSection *sec : ctx.outputSections)
587       addressesChanged |= createThunks(sec, margin);
588     // If the verification above thought we needed thunks, we should have
589     // added some.
590     assert(addressesChanged);
591     (void)addressesChanged;
592 
593     // Recalculate the layout for the whole image (and verify the ranges at
594     // the start of the next round).
595     assignAddresses();
596 
597     pass++;
598   }
599 }
600 
601 // The main function of the writer.
602 void Writer::run() {
603   ScopedTimer t1(ctx.codeLayoutTimer);
604 
605   createImportTables();
606   createSections();
607   appendImportThunks();
608   // Import thunks must be added before the Control Flow Guard tables are added.
609   createMiscChunks();
610   createExportTable();
611   mergeSections();
612   removeUnusedSections();
613   finalizeAddresses();
614   removeEmptySections();
615   assignOutputSectionIndices();
616   setSectionPermissions();
617   createSymbolAndStringTable();
618 
619   if (fileSize > UINT32_MAX)
620     fatal("image size (" + Twine(fileSize) + ") " +
621         "exceeds maximum allowable size (" + Twine(UINT32_MAX) + ")");
622 
623   openFile(config->outputFile);
624   if (config->is64()) {
625     writeHeader<pe32plus_header>();
626   } else {
627     writeHeader<pe32_header>();
628   }
629   writeSections();
630   sortExceptionTable();
631 
632   // Fix up the alignment in the TLS Directory's characteristic field,
633   // if a specific alignment value is needed
634   if (tlsAlignment)
635     fixTlsAlignment();
636 
637   t1.stop();
638 
639   if (!config->pdbPath.empty() && config->debug) {
640     assert(buildId);
641     createPDB(ctx, sectionTable, buildId->buildId);
642   }
643   writeBuildId();
644 
645   writeLLDMapFile(ctx);
646   writeMapFile(ctx);
647 
648   if (errorCount())
649     return;
650 
651   ScopedTimer t2(ctx.outputCommitTimer);
652   if (auto e = buffer->commit())
653     fatal("failed to write the output file: " + toString(std::move(e)));
654 }
655 
656 static StringRef getOutputSectionName(StringRef name) {
657   StringRef s = name.split('$').first;
658 
659   // Treat a later period as a separator for MinGW, for sections like
660   // ".ctors.01234".
661   return s.substr(0, s.find('.', 1));
662 }
663 
664 // For /order.
665 static void sortBySectionOrder(std::vector<Chunk *> &chunks) {
666   auto getPriority = [](const Chunk *c) {
667     if (auto *sec = dyn_cast<SectionChunk>(c))
668       if (sec->sym)
669         return config->order.lookup(sec->sym->getName());
670     return 0;
671   };
672 
673   llvm::stable_sort(chunks, [=](const Chunk *a, const Chunk *b) {
674     return getPriority(a) < getPriority(b);
675   });
676 }
677 
678 // Change the characteristics of existing PartialSections that belong to the
679 // section Name to Chars.
680 void Writer::fixPartialSectionChars(StringRef name, uint32_t chars) {
681   for (auto it : partialSections) {
682     PartialSection *pSec = it.second;
683     StringRef curName = pSec->name;
684     if (!curName.consume_front(name) ||
685         (!curName.empty() && !curName.startswith("$")))
686       continue;
687     if (pSec->characteristics == chars)
688       continue;
689     PartialSection *destSec = createPartialSection(pSec->name, chars);
690     destSec->chunks.insert(destSec->chunks.end(), pSec->chunks.begin(),
691                            pSec->chunks.end());
692     pSec->chunks.clear();
693   }
694 }
695 
696 // Sort concrete section chunks from GNU import libraries.
697 //
698 // GNU binutils doesn't use short import files, but instead produces import
699 // libraries that consist of object files, with section chunks for the .idata$*
700 // sections. These are linked just as regular static libraries. Each import
701 // library consists of one header object, one object file for every imported
702 // symbol, and one trailer object. In order for the .idata tables/lists to
703 // be formed correctly, the section chunks within each .idata$* section need
704 // to be grouped by library, and sorted alphabetically within each library
705 // (which makes sure the header comes first and the trailer last).
706 bool Writer::fixGnuImportChunks() {
707   uint32_t rdata = IMAGE_SCN_CNT_INITIALIZED_DATA | IMAGE_SCN_MEM_READ;
708 
709   // Make sure all .idata$* section chunks are mapped as RDATA in order to
710   // be sorted into the same sections as our own synthesized .idata chunks.
711   fixPartialSectionChars(".idata", rdata);
712 
713   bool hasIdata = false;
714   // Sort all .idata$* chunks, grouping chunks from the same library,
715   // with alphabetical ordering of the object fils within a library.
716   for (auto it : partialSections) {
717     PartialSection *pSec = it.second;
718     if (!pSec->name.startswith(".idata"))
719       continue;
720 
721     if (!pSec->chunks.empty())
722       hasIdata = true;
723     llvm::stable_sort(pSec->chunks, [&](Chunk *s, Chunk *t) {
724       SectionChunk *sc1 = dyn_cast_or_null<SectionChunk>(s);
725       SectionChunk *sc2 = dyn_cast_or_null<SectionChunk>(t);
726       if (!sc1 || !sc2) {
727         // if SC1, order them ascending. If SC2 or both null,
728         // S is not less than T.
729         return sc1 != nullptr;
730       }
731       // Make a string with "libraryname/objectfile" for sorting, achieving
732       // both grouping by library and sorting of objects within a library,
733       // at once.
734       std::string key1 =
735           (sc1->file->parentName + "/" + sc1->file->getName()).str();
736       std::string key2 =
737           (sc2->file->parentName + "/" + sc2->file->getName()).str();
738       return key1 < key2;
739     });
740   }
741   return hasIdata;
742 }
743 
744 // Add generated idata chunks, for imported symbols and DLLs, and a
745 // terminator in .idata$2.
746 void Writer::addSyntheticIdata() {
747   uint32_t rdata = IMAGE_SCN_CNT_INITIALIZED_DATA | IMAGE_SCN_MEM_READ;
748   idata.create();
749 
750   // Add the .idata content in the right section groups, to allow
751   // chunks from other linked in object files to be grouped together.
752   // See Microsoft PE/COFF spec 5.4 for details.
753   auto add = [&](StringRef n, std::vector<Chunk *> &v) {
754     PartialSection *pSec = createPartialSection(n, rdata);
755     pSec->chunks.insert(pSec->chunks.end(), v.begin(), v.end());
756   };
757 
758   // The loader assumes a specific order of data.
759   // Add each type in the correct order.
760   add(".idata$2", idata.dirs);
761   add(".idata$4", idata.lookups);
762   add(".idata$5", idata.addresses);
763   if (!idata.hints.empty())
764     add(".idata$6", idata.hints);
765   add(".idata$7", idata.dllNames);
766 }
767 
768 // Locate the first Chunk and size of the import directory list and the
769 // IAT.
770 void Writer::locateImportTables() {
771   uint32_t rdata = IMAGE_SCN_CNT_INITIALIZED_DATA | IMAGE_SCN_MEM_READ;
772 
773   if (PartialSection *importDirs = findPartialSection(".idata$2", rdata)) {
774     if (!importDirs->chunks.empty())
775       importTableStart = importDirs->chunks.front();
776     for (Chunk *c : importDirs->chunks)
777       importTableSize += c->getSize();
778   }
779 
780   if (PartialSection *importAddresses = findPartialSection(".idata$5", rdata)) {
781     if (!importAddresses->chunks.empty())
782       iatStart = importAddresses->chunks.front();
783     for (Chunk *c : importAddresses->chunks)
784       iatSize += c->getSize();
785   }
786 }
787 
788 // Return whether a SectionChunk's suffix (the dollar and any trailing
789 // suffix) should be removed and sorted into the main suffixless
790 // PartialSection.
791 static bool shouldStripSectionSuffix(SectionChunk *sc, StringRef name) {
792   // On MinGW, comdat groups are formed by putting the comdat group name
793   // after the '$' in the section name. For .eh_frame$<symbol>, that must
794   // still be sorted before the .eh_frame trailer from crtend.o, thus just
795   // strip the section name trailer. For other sections, such as
796   // .tls$$<symbol> (where non-comdat .tls symbols are otherwise stored in
797   // ".tls$"), they must be strictly sorted after .tls. And for the
798   // hypothetical case of comdat .CRT$XCU, we definitely need to keep the
799   // suffix for sorting. Thus, to play it safe, only strip the suffix for
800   // the standard sections.
801   if (!config->mingw)
802     return false;
803   if (!sc || !sc->isCOMDAT())
804     return false;
805   return name.startswith(".text$") || name.startswith(".data$") ||
806          name.startswith(".rdata$") || name.startswith(".pdata$") ||
807          name.startswith(".xdata$") || name.startswith(".eh_frame$");
808 }
809 
810 void Writer::sortSections() {
811   if (!config->callGraphProfile.empty()) {
812     DenseMap<const SectionChunk *, int> order =
813         computeCallGraphProfileOrder(ctx);
814     for (auto it : order) {
815       if (DefinedRegular *sym = it.first->sym)
816         config->order[sym->getName()] = it.second;
817     }
818   }
819   if (!config->order.empty())
820     for (auto it : partialSections)
821       sortBySectionOrder(it.second->chunks);
822 }
823 
824 // Create output section objects and add them to OutputSections.
825 void Writer::createSections() {
826   // First, create the builtin sections.
827   const uint32_t data = IMAGE_SCN_CNT_INITIALIZED_DATA;
828   const uint32_t bss = IMAGE_SCN_CNT_UNINITIALIZED_DATA;
829   const uint32_t code = IMAGE_SCN_CNT_CODE;
830   const uint32_t discardable = IMAGE_SCN_MEM_DISCARDABLE;
831   const uint32_t r = IMAGE_SCN_MEM_READ;
832   const uint32_t w = IMAGE_SCN_MEM_WRITE;
833   const uint32_t x = IMAGE_SCN_MEM_EXECUTE;
834 
835   SmallDenseMap<std::pair<StringRef, uint32_t>, OutputSection *> sections;
836   auto createSection = [&](StringRef name, uint32_t outChars) {
837     OutputSection *&sec = sections[{name, outChars}];
838     if (!sec) {
839       sec = make<OutputSection>(name, outChars);
840       ctx.outputSections.push_back(sec);
841     }
842     return sec;
843   };
844 
845   // Try to match the section order used by link.exe.
846   textSec = createSection(".text", code | r | x);
847   createSection(".bss", bss | r | w);
848   rdataSec = createSection(".rdata", data | r);
849   buildidSec = createSection(".buildid", data | r);
850   dataSec = createSection(".data", data | r | w);
851   pdataSec = createSection(".pdata", data | r);
852   idataSec = createSection(".idata", data | r);
853   edataSec = createSection(".edata", data | r);
854   didatSec = createSection(".didat", data | r);
855   rsrcSec = createSection(".rsrc", data | r);
856   relocSec = createSection(".reloc", data | discardable | r);
857   ctorsSec = createSection(".ctors", data | r | w);
858   dtorsSec = createSection(".dtors", data | r | w);
859 
860   // Then bin chunks by name and output characteristics.
861   for (Chunk *c : ctx.symtab.getChunks()) {
862     auto *sc = dyn_cast<SectionChunk>(c);
863     if (sc && !sc->live) {
864       if (config->verbose)
865         sc->printDiscardedMessage();
866       continue;
867     }
868     StringRef name = c->getSectionName();
869     if (shouldStripSectionSuffix(sc, name))
870       name = name.split('$').first;
871 
872     if (name.startswith(".tls"))
873       tlsAlignment = std::max(tlsAlignment, c->getAlignment());
874 
875     PartialSection *pSec = createPartialSection(name,
876                                                 c->getOutputCharacteristics());
877     pSec->chunks.push_back(c);
878   }
879 
880   fixPartialSectionChars(".rsrc", data | r);
881   fixPartialSectionChars(".edata", data | r);
882   // Even in non MinGW cases, we might need to link against GNU import
883   // libraries.
884   bool hasIdata = fixGnuImportChunks();
885   if (!idata.empty())
886     hasIdata = true;
887 
888   if (hasIdata)
889     addSyntheticIdata();
890 
891   sortSections();
892 
893   if (hasIdata)
894     locateImportTables();
895 
896   // Then create an OutputSection for each section.
897   // '$' and all following characters in input section names are
898   // discarded when determining output section. So, .text$foo
899   // contributes to .text, for example. See PE/COFF spec 3.2.
900   for (auto it : partialSections) {
901     PartialSection *pSec = it.second;
902     StringRef name = getOutputSectionName(pSec->name);
903     uint32_t outChars = pSec->characteristics;
904 
905     if (name == ".CRT") {
906       // In link.exe, there is a special case for the I386 target where .CRT
907       // sections are treated as if they have output characteristics DATA | R if
908       // their characteristics are DATA | R | W. This implements the same
909       // special case for all architectures.
910       outChars = data | r;
911 
912       log("Processing section " + pSec->name + " -> " + name);
913 
914       sortCRTSectionChunks(pSec->chunks);
915     }
916 
917     OutputSection *sec = createSection(name, outChars);
918     for (Chunk *c : pSec->chunks)
919       sec->addChunk(c);
920 
921     sec->addContributingPartialSection(pSec);
922   }
923 
924   // Finally, move some output sections to the end.
925   auto sectionOrder = [&](const OutputSection *s) {
926     // Move DISCARDABLE (or non-memory-mapped) sections to the end of file
927     // because the loader cannot handle holes. Stripping can remove other
928     // discardable ones than .reloc, which is first of them (created early).
929     if (s->header.Characteristics & IMAGE_SCN_MEM_DISCARDABLE) {
930       // Move discardable sections named .debug_ to the end, after other
931       // discardable sections. Stripping only removes the sections named
932       // .debug_* - thus try to avoid leaving holes after stripping.
933       if (s->name.startswith(".debug_"))
934         return 3;
935       return 2;
936     }
937     // .rsrc should come at the end of the non-discardable sections because its
938     // size may change by the Win32 UpdateResources() function, causing
939     // subsequent sections to move (see https://crbug.com/827082).
940     if (s == rsrcSec)
941       return 1;
942     return 0;
943   };
944   llvm::stable_sort(ctx.outputSections,
945                     [&](const OutputSection *s, const OutputSection *t) {
946                       return sectionOrder(s) < sectionOrder(t);
947                     });
948 }
949 
950 void Writer::createMiscChunks() {
951   for (MergeChunk *p : ctx.mergeChunkInstances) {
952     if (p) {
953       p->finalizeContents();
954       rdataSec->addChunk(p);
955     }
956   }
957 
958   // Create thunks for locally-dllimported symbols.
959   if (!ctx.symtab.localImportChunks.empty()) {
960     for (Chunk *c : ctx.symtab.localImportChunks)
961       rdataSec->addChunk(c);
962   }
963 
964   // Create Debug Information Chunks
965   OutputSection *debugInfoSec = config->mingw ? buildidSec : rdataSec;
966   if (config->debug || config->repro || config->cetCompat) {
967     debugDirectory =
968         make<DebugDirectoryChunk>(ctx, debugRecords, config->repro);
969     debugDirectory->setAlignment(4);
970     debugInfoSec->addChunk(debugDirectory);
971   }
972 
973   if (config->debug) {
974     // Make a CVDebugRecordChunk even when /DEBUG:CV is not specified.  We
975     // output a PDB no matter what, and this chunk provides the only means of
976     // allowing a debugger to match a PDB and an executable.  So we need it even
977     // if we're ultimately not going to write CodeView data to the PDB.
978     buildId = make<CVDebugRecordChunk>();
979     debugRecords.push_back({COFF::IMAGE_DEBUG_TYPE_CODEVIEW, buildId});
980   }
981 
982   if (config->cetCompat) {
983     debugRecords.push_back({COFF::IMAGE_DEBUG_TYPE_EX_DLLCHARACTERISTICS,
984                             make<ExtendedDllCharacteristicsChunk>(
985                                 IMAGE_DLL_CHARACTERISTICS_EX_CET_COMPAT)});
986   }
987 
988   // Align and add each chunk referenced by the debug data directory.
989   for (std::pair<COFF::DebugType, Chunk *> r : debugRecords) {
990     r.second->setAlignment(4);
991     debugInfoSec->addChunk(r.second);
992   }
993 
994   // Create SEH table. x86-only.
995   if (config->safeSEH)
996     createSEHTable();
997 
998   // Create /guard:cf tables if requested.
999   if (config->guardCF != GuardCFLevel::Off)
1000     createGuardCFTables();
1001 
1002   if (config->autoImport)
1003     createRuntimePseudoRelocs();
1004 
1005   if (config->mingw)
1006     insertCtorDtorSymbols();
1007 }
1008 
1009 // Create .idata section for the DLL-imported symbol table.
1010 // The format of this section is inherently Windows-specific.
1011 // IdataContents class abstracted away the details for us,
1012 // so we just let it create chunks and add them to the section.
1013 void Writer::createImportTables() {
1014   // Initialize DLLOrder so that import entries are ordered in
1015   // the same order as in the command line. (That affects DLL
1016   // initialization order, and this ordering is MSVC-compatible.)
1017   for (ImportFile *file : ctx.importFileInstances) {
1018     if (!file->live)
1019       continue;
1020 
1021     std::string dll = StringRef(file->dllName).lower();
1022     if (config->dllOrder.count(dll) == 0)
1023       config->dllOrder[dll] = config->dllOrder.size();
1024 
1025     if (file->impSym && !isa<DefinedImportData>(file->impSym))
1026       fatal(toString(*file->impSym) + " was replaced");
1027     DefinedImportData *impSym = cast_or_null<DefinedImportData>(file->impSym);
1028     if (config->delayLoads.count(StringRef(file->dllName).lower())) {
1029       if (!file->thunkSym)
1030         fatal("cannot delay-load " + toString(file) +
1031               " due to import of data: " + toString(*impSym));
1032       delayIdata.add(impSym);
1033     } else {
1034       idata.add(impSym);
1035     }
1036   }
1037 }
1038 
1039 void Writer::appendImportThunks() {
1040   if (ctx.importFileInstances.empty())
1041     return;
1042 
1043   for (ImportFile *file : ctx.importFileInstances) {
1044     if (!file->live)
1045       continue;
1046 
1047     if (!file->thunkSym)
1048       continue;
1049 
1050     if (!isa<DefinedImportThunk>(file->thunkSym))
1051       fatal(toString(*file->thunkSym) + " was replaced");
1052     DefinedImportThunk *thunk = cast<DefinedImportThunk>(file->thunkSym);
1053     if (file->thunkLive)
1054       textSec->addChunk(thunk->getChunk());
1055   }
1056 
1057   if (!delayIdata.empty()) {
1058     Defined *helper = cast<Defined>(config->delayLoadHelper);
1059     delayIdata.create(ctx, helper);
1060     for (Chunk *c : delayIdata.getChunks())
1061       didatSec->addChunk(c);
1062     for (Chunk *c : delayIdata.getDataChunks())
1063       dataSec->addChunk(c);
1064     for (Chunk *c : delayIdata.getCodeChunks())
1065       textSec->addChunk(c);
1066   }
1067 }
1068 
1069 void Writer::createExportTable() {
1070   if (!edataSec->chunks.empty()) {
1071     // Allow using a custom built export table from input object files, instead
1072     // of having the linker synthesize the tables.
1073     if (config->hadExplicitExports)
1074       warn("literal .edata sections override exports");
1075   } else if (!config->exports.empty()) {
1076     for (Chunk *c : edata.chunks)
1077       edataSec->addChunk(c);
1078   }
1079   if (!edataSec->chunks.empty()) {
1080     edataStart = edataSec->chunks.front();
1081     edataEnd = edataSec->chunks.back();
1082   }
1083   // Warn on exported deleting destructor.
1084   for (auto e : config->exports)
1085     if (e.sym && e.sym->getName().startswith("??_G"))
1086       warn("export of deleting dtor: " + toString(*e.sym));
1087 }
1088 
1089 void Writer::removeUnusedSections() {
1090   // Remove sections that we can be sure won't get content, to avoid
1091   // allocating space for their section headers.
1092   auto isUnused = [this](OutputSection *s) {
1093     if (s == relocSec)
1094       return false; // This section is populated later.
1095     // MergeChunks have zero size at this point, as their size is finalized
1096     // later. Only remove sections that have no Chunks at all.
1097     return s->chunks.empty();
1098   };
1099   llvm::erase_if(ctx.outputSections, isUnused);
1100 }
1101 
1102 // The Windows loader doesn't seem to like empty sections,
1103 // so we remove them if any.
1104 void Writer::removeEmptySections() {
1105   auto isEmpty = [](OutputSection *s) { return s->getVirtualSize() == 0; };
1106   llvm::erase_if(ctx.outputSections, isEmpty);
1107 }
1108 
1109 void Writer::assignOutputSectionIndices() {
1110   // Assign final output section indices, and assign each chunk to its output
1111   // section.
1112   uint32_t idx = 1;
1113   for (OutputSection *os : ctx.outputSections) {
1114     os->sectionIndex = idx;
1115     for (Chunk *c : os->chunks)
1116       c->setOutputSectionIdx(idx);
1117     ++idx;
1118   }
1119 
1120   // Merge chunks are containers of chunks, so assign those an output section
1121   // too.
1122   for (MergeChunk *mc : ctx.mergeChunkInstances)
1123     if (mc)
1124       for (SectionChunk *sc : mc->sections)
1125         if (sc && sc->live)
1126           sc->setOutputSectionIdx(mc->getOutputSectionIdx());
1127 }
1128 
1129 size_t Writer::addEntryToStringTable(StringRef str) {
1130   assert(str.size() > COFF::NameSize);
1131   size_t offsetOfEntry = strtab.size() + 4; // +4 for the size field
1132   strtab.insert(strtab.end(), str.begin(), str.end());
1133   strtab.push_back('\0');
1134   return offsetOfEntry;
1135 }
1136 
1137 Optional<coff_symbol16> Writer::createSymbol(Defined *def) {
1138   coff_symbol16 sym;
1139   switch (def->kind()) {
1140   case Symbol::DefinedAbsoluteKind:
1141     sym.Value = def->getRVA();
1142     sym.SectionNumber = IMAGE_SYM_ABSOLUTE;
1143     break;
1144   case Symbol::DefinedSyntheticKind:
1145     // Relative symbols are unrepresentable in a COFF symbol table.
1146     return None;
1147   default: {
1148     // Don't write symbols that won't be written to the output to the symbol
1149     // table.
1150     Chunk *c = def->getChunk();
1151     if (!c)
1152       return None;
1153     OutputSection *os = ctx.getOutputSection(c);
1154     if (!os)
1155       return None;
1156 
1157     sym.Value = def->getRVA() - os->getRVA();
1158     sym.SectionNumber = os->sectionIndex;
1159     break;
1160   }
1161   }
1162 
1163   // Symbols that are runtime pseudo relocations don't point to the actual
1164   // symbol data itself (as they are imported), but points to the IAT entry
1165   // instead. Avoid emitting them to the symbol table, as they can confuse
1166   // debuggers.
1167   if (def->isRuntimePseudoReloc)
1168     return None;
1169 
1170   StringRef name = def->getName();
1171   if (name.size() > COFF::NameSize) {
1172     sym.Name.Offset.Zeroes = 0;
1173     sym.Name.Offset.Offset = addEntryToStringTable(name);
1174   } else {
1175     memset(sym.Name.ShortName, 0, COFF::NameSize);
1176     memcpy(sym.Name.ShortName, name.data(), name.size());
1177   }
1178 
1179   if (auto *d = dyn_cast<DefinedCOFF>(def)) {
1180     COFFSymbolRef ref = d->getCOFFSymbol();
1181     sym.Type = ref.getType();
1182     sym.StorageClass = ref.getStorageClass();
1183   } else {
1184     sym.Type = IMAGE_SYM_TYPE_NULL;
1185     sym.StorageClass = IMAGE_SYM_CLASS_EXTERNAL;
1186   }
1187   sym.NumberOfAuxSymbols = 0;
1188   return sym;
1189 }
1190 
1191 void Writer::createSymbolAndStringTable() {
1192   // PE/COFF images are limited to 8 byte section names. Longer names can be
1193   // supported by writing a non-standard string table, but this string table is
1194   // not mapped at runtime and the long names will therefore be inaccessible.
1195   // link.exe always truncates section names to 8 bytes, whereas binutils always
1196   // preserves long section names via the string table. LLD adopts a hybrid
1197   // solution where discardable sections have long names preserved and
1198   // non-discardable sections have their names truncated, to ensure that any
1199   // section which is mapped at runtime also has its name mapped at runtime.
1200   for (OutputSection *sec : ctx.outputSections) {
1201     if (sec->name.size() <= COFF::NameSize)
1202       continue;
1203     if ((sec->header.Characteristics & IMAGE_SCN_MEM_DISCARDABLE) == 0)
1204       continue;
1205     if (config->warnLongSectionNames) {
1206       warn("section name " + sec->name +
1207            " is longer than 8 characters and will use a non-standard string "
1208            "table");
1209     }
1210     sec->setStringTableOff(addEntryToStringTable(sec->name));
1211   }
1212 
1213   if (config->debugDwarf || config->debugSymtab) {
1214     for (ObjFile *file : ctx.objFileInstances) {
1215       for (Symbol *b : file->getSymbols()) {
1216         auto *d = dyn_cast_or_null<Defined>(b);
1217         if (!d || d->writtenToSymtab)
1218           continue;
1219         d->writtenToSymtab = true;
1220         if (auto *dc = dyn_cast_or_null<DefinedCOFF>(d)) {
1221           COFFSymbolRef symRef = dc->getCOFFSymbol();
1222           if (symRef.isSectionDefinition() ||
1223               symRef.getStorageClass() == COFF::IMAGE_SYM_CLASS_LABEL)
1224             continue;
1225         }
1226 
1227         if (Optional<coff_symbol16> sym = createSymbol(d))
1228           outputSymtab.push_back(*sym);
1229       }
1230     }
1231   }
1232 
1233   if (outputSymtab.empty() && strtab.empty())
1234     return;
1235 
1236   // We position the symbol table to be adjacent to the end of the last section.
1237   uint64_t fileOff = fileSize;
1238   pointerToSymbolTable = fileOff;
1239   fileOff += outputSymtab.size() * sizeof(coff_symbol16);
1240   fileOff += 4 + strtab.size();
1241   fileSize = alignTo(fileOff, config->fileAlign);
1242 }
1243 
1244 void Writer::mergeSections() {
1245   if (!pdataSec->chunks.empty()) {
1246     firstPdata = pdataSec->chunks.front();
1247     lastPdata = pdataSec->chunks.back();
1248   }
1249 
1250   for (auto &p : config->merge) {
1251     StringRef toName = p.second;
1252     if (p.first == toName)
1253       continue;
1254     StringSet<> names;
1255     while (true) {
1256       if (!names.insert(toName).second)
1257         fatal("/merge: cycle found for section '" + p.first + "'");
1258       auto i = config->merge.find(toName);
1259       if (i == config->merge.end())
1260         break;
1261       toName = i->second;
1262     }
1263     OutputSection *from = findSection(p.first);
1264     OutputSection *to = findSection(toName);
1265     if (!from)
1266       continue;
1267     if (!to) {
1268       from->name = toName;
1269       continue;
1270     }
1271     to->merge(from);
1272   }
1273 }
1274 
1275 // Visits all sections to assign incremental, non-overlapping RVAs and
1276 // file offsets.
1277 void Writer::assignAddresses() {
1278   sizeOfHeaders = dosStubSize + sizeof(PEMagic) + sizeof(coff_file_header) +
1279                   sizeof(data_directory) * numberOfDataDirectory +
1280                   sizeof(coff_section) * ctx.outputSections.size();
1281   sizeOfHeaders +=
1282       config->is64() ? sizeof(pe32plus_header) : sizeof(pe32_header);
1283   sizeOfHeaders = alignTo(sizeOfHeaders, config->fileAlign);
1284   fileSize = sizeOfHeaders;
1285 
1286   // The first page is kept unmapped.
1287   uint64_t rva = alignTo(sizeOfHeaders, config->align);
1288 
1289   for (OutputSection *sec : ctx.outputSections) {
1290     if (sec == relocSec)
1291       addBaserels();
1292     uint64_t rawSize = 0, virtualSize = 0;
1293     sec->header.VirtualAddress = rva;
1294 
1295     // If /FUNCTIONPADMIN is used, functions are padded in order to create a
1296     // hotpatchable image.
1297     const bool isCodeSection =
1298         (sec->header.Characteristics & IMAGE_SCN_CNT_CODE) &&
1299         (sec->header.Characteristics & IMAGE_SCN_MEM_READ) &&
1300         (sec->header.Characteristics & IMAGE_SCN_MEM_EXECUTE);
1301     uint32_t padding = isCodeSection ? config->functionPadMin : 0;
1302 
1303     for (Chunk *c : sec->chunks) {
1304       if (padding && c->isHotPatchable())
1305         virtualSize += padding;
1306       virtualSize = alignTo(virtualSize, c->getAlignment());
1307       c->setRVA(rva + virtualSize);
1308       virtualSize += c->getSize();
1309       if (c->hasData)
1310         rawSize = alignTo(virtualSize, config->fileAlign);
1311     }
1312     if (virtualSize > UINT32_MAX)
1313       error("section larger than 4 GiB: " + sec->name);
1314     sec->header.VirtualSize = virtualSize;
1315     sec->header.SizeOfRawData = rawSize;
1316     if (rawSize != 0)
1317       sec->header.PointerToRawData = fileSize;
1318     rva += alignTo(virtualSize, config->align);
1319     fileSize += alignTo(rawSize, config->fileAlign);
1320   }
1321   sizeOfImage = alignTo(rva, config->align);
1322 
1323   // Assign addresses to sections in MergeChunks.
1324   for (MergeChunk *mc : ctx.mergeChunkInstances)
1325     if (mc)
1326       mc->assignSubsectionRVAs();
1327 }
1328 
1329 template <typename PEHeaderTy> void Writer::writeHeader() {
1330   // Write DOS header. For backwards compatibility, the first part of a PE/COFF
1331   // executable consists of an MS-DOS MZ executable. If the executable is run
1332   // under DOS, that program gets run (usually to just print an error message).
1333   // When run under Windows, the loader looks at AddressOfNewExeHeader and uses
1334   // the PE header instead.
1335   uint8_t *buf = buffer->getBufferStart();
1336   auto *dos = reinterpret_cast<dos_header *>(buf);
1337   buf += sizeof(dos_header);
1338   dos->Magic[0] = 'M';
1339   dos->Magic[1] = 'Z';
1340   dos->UsedBytesInTheLastPage = dosStubSize % 512;
1341   dos->FileSizeInPages = divideCeil(dosStubSize, 512);
1342   dos->HeaderSizeInParagraphs = sizeof(dos_header) / 16;
1343 
1344   dos->AddressOfRelocationTable = sizeof(dos_header);
1345   dos->AddressOfNewExeHeader = dosStubSize;
1346 
1347   // Write DOS program.
1348   memcpy(buf, dosProgram, sizeof(dosProgram));
1349   buf += sizeof(dosProgram);
1350 
1351   // Write PE magic
1352   memcpy(buf, PEMagic, sizeof(PEMagic));
1353   buf += sizeof(PEMagic);
1354 
1355   // Write COFF header
1356   auto *coff = reinterpret_cast<coff_file_header *>(buf);
1357   buf += sizeof(*coff);
1358   coff->Machine = config->machine;
1359   coff->NumberOfSections = ctx.outputSections.size();
1360   coff->Characteristics = IMAGE_FILE_EXECUTABLE_IMAGE;
1361   if (config->largeAddressAware)
1362     coff->Characteristics |= IMAGE_FILE_LARGE_ADDRESS_AWARE;
1363   if (!config->is64())
1364     coff->Characteristics |= IMAGE_FILE_32BIT_MACHINE;
1365   if (config->dll)
1366     coff->Characteristics |= IMAGE_FILE_DLL;
1367   if (config->driverUponly)
1368     coff->Characteristics |= IMAGE_FILE_UP_SYSTEM_ONLY;
1369   if (!config->relocatable)
1370     coff->Characteristics |= IMAGE_FILE_RELOCS_STRIPPED;
1371   if (config->swaprunCD)
1372     coff->Characteristics |= IMAGE_FILE_REMOVABLE_RUN_FROM_SWAP;
1373   if (config->swaprunNet)
1374     coff->Characteristics |= IMAGE_FILE_NET_RUN_FROM_SWAP;
1375   coff->SizeOfOptionalHeader =
1376       sizeof(PEHeaderTy) + sizeof(data_directory) * numberOfDataDirectory;
1377 
1378   // Write PE header
1379   auto *pe = reinterpret_cast<PEHeaderTy *>(buf);
1380   buf += sizeof(*pe);
1381   pe->Magic = config->is64() ? PE32Header::PE32_PLUS : PE32Header::PE32;
1382 
1383   // If {Major,Minor}LinkerVersion is left at 0.0, then for some
1384   // reason signing the resulting PE file with Authenticode produces a
1385   // signature that fails to validate on Windows 7 (but is OK on 10).
1386   // Set it to 14.0, which is what VS2015 outputs, and which avoids
1387   // that problem.
1388   pe->MajorLinkerVersion = 14;
1389   pe->MinorLinkerVersion = 0;
1390 
1391   pe->ImageBase = config->imageBase;
1392   pe->SectionAlignment = config->align;
1393   pe->FileAlignment = config->fileAlign;
1394   pe->MajorImageVersion = config->majorImageVersion;
1395   pe->MinorImageVersion = config->minorImageVersion;
1396   pe->MajorOperatingSystemVersion = config->majorOSVersion;
1397   pe->MinorOperatingSystemVersion = config->minorOSVersion;
1398   pe->MajorSubsystemVersion = config->majorSubsystemVersion;
1399   pe->MinorSubsystemVersion = config->minorSubsystemVersion;
1400   pe->Subsystem = config->subsystem;
1401   pe->SizeOfImage = sizeOfImage;
1402   pe->SizeOfHeaders = sizeOfHeaders;
1403   if (!config->noEntry) {
1404     Defined *entry = cast<Defined>(config->entry);
1405     pe->AddressOfEntryPoint = entry->getRVA();
1406     // Pointer to thumb code must have the LSB set, so adjust it.
1407     if (config->machine == ARMNT)
1408       pe->AddressOfEntryPoint |= 1;
1409   }
1410   pe->SizeOfStackReserve = config->stackReserve;
1411   pe->SizeOfStackCommit = config->stackCommit;
1412   pe->SizeOfHeapReserve = config->heapReserve;
1413   pe->SizeOfHeapCommit = config->heapCommit;
1414   if (config->appContainer)
1415     pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_APPCONTAINER;
1416   if (config->driverWdm)
1417     pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_WDM_DRIVER;
1418   if (config->dynamicBase)
1419     pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_DYNAMIC_BASE;
1420   if (config->highEntropyVA)
1421     pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_HIGH_ENTROPY_VA;
1422   if (!config->allowBind)
1423     pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_NO_BIND;
1424   if (config->nxCompat)
1425     pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_NX_COMPAT;
1426   if (!config->allowIsolation)
1427     pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_NO_ISOLATION;
1428   if (config->guardCF != GuardCFLevel::Off)
1429     pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_GUARD_CF;
1430   if (config->integrityCheck)
1431     pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_FORCE_INTEGRITY;
1432   if (setNoSEHCharacteristic || config->noSEH)
1433     pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_NO_SEH;
1434   if (config->terminalServerAware)
1435     pe->DLLCharacteristics |= IMAGE_DLL_CHARACTERISTICS_TERMINAL_SERVER_AWARE;
1436   pe->NumberOfRvaAndSize = numberOfDataDirectory;
1437   if (textSec->getVirtualSize()) {
1438     pe->BaseOfCode = textSec->getRVA();
1439     pe->SizeOfCode = textSec->getRawSize();
1440   }
1441   pe->SizeOfInitializedData = getSizeOfInitializedData();
1442 
1443   // Write data directory
1444   auto *dir = reinterpret_cast<data_directory *>(buf);
1445   buf += sizeof(*dir) * numberOfDataDirectory;
1446   if (edataStart) {
1447     dir[EXPORT_TABLE].RelativeVirtualAddress = edataStart->getRVA();
1448     dir[EXPORT_TABLE].Size =
1449         edataEnd->getRVA() + edataEnd->getSize() - edataStart->getRVA();
1450   }
1451   if (importTableStart) {
1452     dir[IMPORT_TABLE].RelativeVirtualAddress = importTableStart->getRVA();
1453     dir[IMPORT_TABLE].Size = importTableSize;
1454   }
1455   if (iatStart) {
1456     dir[IAT].RelativeVirtualAddress = iatStart->getRVA();
1457     dir[IAT].Size = iatSize;
1458   }
1459   if (rsrcSec->getVirtualSize()) {
1460     dir[RESOURCE_TABLE].RelativeVirtualAddress = rsrcSec->getRVA();
1461     dir[RESOURCE_TABLE].Size = rsrcSec->getVirtualSize();
1462   }
1463   if (firstPdata) {
1464     dir[EXCEPTION_TABLE].RelativeVirtualAddress = firstPdata->getRVA();
1465     dir[EXCEPTION_TABLE].Size =
1466         lastPdata->getRVA() + lastPdata->getSize() - firstPdata->getRVA();
1467   }
1468   if (relocSec->getVirtualSize()) {
1469     dir[BASE_RELOCATION_TABLE].RelativeVirtualAddress = relocSec->getRVA();
1470     dir[BASE_RELOCATION_TABLE].Size = relocSec->getVirtualSize();
1471   }
1472   if (Symbol *sym = ctx.symtab.findUnderscore("_tls_used")) {
1473     if (Defined *b = dyn_cast<Defined>(sym)) {
1474       dir[TLS_TABLE].RelativeVirtualAddress = b->getRVA();
1475       dir[TLS_TABLE].Size = config->is64()
1476                                 ? sizeof(object::coff_tls_directory64)
1477                                 : sizeof(object::coff_tls_directory32);
1478     }
1479   }
1480   if (debugDirectory) {
1481     dir[DEBUG_DIRECTORY].RelativeVirtualAddress = debugDirectory->getRVA();
1482     dir[DEBUG_DIRECTORY].Size = debugDirectory->getSize();
1483   }
1484   if (Symbol *sym = ctx.symtab.findUnderscore("_load_config_used")) {
1485     if (auto *b = dyn_cast<DefinedRegular>(sym)) {
1486       SectionChunk *sc = b->getChunk();
1487       assert(b->getRVA() >= sc->getRVA());
1488       uint64_t offsetInChunk = b->getRVA() - sc->getRVA();
1489       if (!sc->hasData || offsetInChunk + 4 > sc->getSize())
1490         fatal("_load_config_used is malformed");
1491 
1492       ArrayRef<uint8_t> secContents = sc->getContents();
1493       uint32_t loadConfigSize =
1494           *reinterpret_cast<const ulittle32_t *>(&secContents[offsetInChunk]);
1495       if (offsetInChunk + loadConfigSize > sc->getSize())
1496         fatal("_load_config_used is too large");
1497       dir[LOAD_CONFIG_TABLE].RelativeVirtualAddress = b->getRVA();
1498       dir[LOAD_CONFIG_TABLE].Size = loadConfigSize;
1499     }
1500   }
1501   if (!delayIdata.empty()) {
1502     dir[DELAY_IMPORT_DESCRIPTOR].RelativeVirtualAddress =
1503         delayIdata.getDirRVA();
1504     dir[DELAY_IMPORT_DESCRIPTOR].Size = delayIdata.getDirSize();
1505   }
1506 
1507   // Write section table
1508   for (OutputSection *sec : ctx.outputSections) {
1509     sec->writeHeaderTo(buf);
1510     buf += sizeof(coff_section);
1511   }
1512   sectionTable = ArrayRef<uint8_t>(
1513       buf - ctx.outputSections.size() * sizeof(coff_section), buf);
1514 
1515   if (outputSymtab.empty() && strtab.empty())
1516     return;
1517 
1518   coff->PointerToSymbolTable = pointerToSymbolTable;
1519   uint32_t numberOfSymbols = outputSymtab.size();
1520   coff->NumberOfSymbols = numberOfSymbols;
1521   auto *symbolTable = reinterpret_cast<coff_symbol16 *>(
1522       buffer->getBufferStart() + coff->PointerToSymbolTable);
1523   for (size_t i = 0; i != numberOfSymbols; ++i)
1524     symbolTable[i] = outputSymtab[i];
1525   // Create the string table, it follows immediately after the symbol table.
1526   // The first 4 bytes is length including itself.
1527   buf = reinterpret_cast<uint8_t *>(&symbolTable[numberOfSymbols]);
1528   write32le(buf, strtab.size() + 4);
1529   if (!strtab.empty())
1530     memcpy(buf + 4, strtab.data(), strtab.size());
1531 }
1532 
1533 void Writer::openFile(StringRef path) {
1534   buffer = CHECK(
1535       FileOutputBuffer::create(path, fileSize, FileOutputBuffer::F_executable),
1536       "failed to open " + path);
1537 }
1538 
1539 void Writer::createSEHTable() {
1540   SymbolRVASet handlers;
1541   for (ObjFile *file : ctx.objFileInstances) {
1542     if (!file->hasSafeSEH())
1543       error("/safeseh: " + file->getName() + " is not compatible with SEH");
1544     markSymbolsForRVATable(file, file->getSXDataChunks(), handlers);
1545   }
1546 
1547   // Set the "no SEH" characteristic if there really were no handlers, or if
1548   // there is no load config object to point to the table of handlers.
1549   setNoSEHCharacteristic =
1550       handlers.empty() || !ctx.symtab.findUnderscore("_load_config_used");
1551 
1552   maybeAddRVATable(std::move(handlers), "__safe_se_handler_table",
1553                    "__safe_se_handler_count");
1554 }
1555 
1556 // Add a symbol to an RVA set. Two symbols may have the same RVA, but an RVA set
1557 // cannot contain duplicates. Therefore, the set is uniqued by Chunk and the
1558 // symbol's offset into that Chunk.
1559 static void addSymbolToRVASet(SymbolRVASet &rvaSet, Defined *s) {
1560   Chunk *c = s->getChunk();
1561   if (auto *sc = dyn_cast<SectionChunk>(c))
1562     c = sc->repl; // Look through ICF replacement.
1563   uint32_t off = s->getRVA() - (c ? c->getRVA() : 0);
1564   rvaSet.insert({c, off});
1565 }
1566 
1567 // Given a symbol, add it to the GFIDs table if it is a live, defined, function
1568 // symbol in an executable section.
1569 static void maybeAddAddressTakenFunction(SymbolRVASet &addressTakenSyms,
1570                                          Symbol *s) {
1571   if (!s)
1572     return;
1573 
1574   switch (s->kind()) {
1575   case Symbol::DefinedLocalImportKind:
1576   case Symbol::DefinedImportDataKind:
1577     // Defines an __imp_ pointer, so it is data, so it is ignored.
1578     break;
1579   case Symbol::DefinedCommonKind:
1580     // Common is always data, so it is ignored.
1581     break;
1582   case Symbol::DefinedAbsoluteKind:
1583   case Symbol::DefinedSyntheticKind:
1584     // Absolute is never code, synthetic generally isn't and usually isn't
1585     // determinable.
1586     break;
1587   case Symbol::LazyArchiveKind:
1588   case Symbol::LazyObjectKind:
1589   case Symbol::LazyDLLSymbolKind:
1590   case Symbol::UndefinedKind:
1591     // Undefined symbols resolve to zero, so they don't have an RVA. Lazy
1592     // symbols shouldn't have relocations.
1593     break;
1594 
1595   case Symbol::DefinedImportThunkKind:
1596     // Thunks are always code, include them.
1597     addSymbolToRVASet(addressTakenSyms, cast<Defined>(s));
1598     break;
1599 
1600   case Symbol::DefinedRegularKind: {
1601     // This is a regular, defined, symbol from a COFF file. Mark the symbol as
1602     // address taken if the symbol type is function and it's in an executable
1603     // section.
1604     auto *d = cast<DefinedRegular>(s);
1605     if (d->getCOFFSymbol().getComplexType() == COFF::IMAGE_SYM_DTYPE_FUNCTION) {
1606       SectionChunk *sc = dyn_cast<SectionChunk>(d->getChunk());
1607       if (sc && sc->live &&
1608           sc->getOutputCharacteristics() & IMAGE_SCN_MEM_EXECUTE)
1609         addSymbolToRVASet(addressTakenSyms, d);
1610     }
1611     break;
1612   }
1613   }
1614 }
1615 
1616 // Visit all relocations from all section contributions of this object file and
1617 // mark the relocation target as address-taken.
1618 static void markSymbolsWithRelocations(ObjFile *file,
1619                                        SymbolRVASet &usedSymbols) {
1620   for (Chunk *c : file->getChunks()) {
1621     // We only care about live section chunks. Common chunks and other chunks
1622     // don't generally contain relocations.
1623     SectionChunk *sc = dyn_cast<SectionChunk>(c);
1624     if (!sc || !sc->live)
1625       continue;
1626 
1627     for (const coff_relocation &reloc : sc->getRelocs()) {
1628       if (config->machine == I386 && reloc.Type == COFF::IMAGE_REL_I386_REL32)
1629         // Ignore relative relocations on x86. On x86_64 they can't be ignored
1630         // since they're also used to compute absolute addresses.
1631         continue;
1632 
1633       Symbol *ref = sc->file->getSymbol(reloc.SymbolTableIndex);
1634       maybeAddAddressTakenFunction(usedSymbols, ref);
1635     }
1636   }
1637 }
1638 
1639 // Create the guard function id table. This is a table of RVAs of all
1640 // address-taken functions. It is sorted and uniqued, just like the safe SEH
1641 // table.
1642 void Writer::createGuardCFTables() {
1643   SymbolRVASet addressTakenSyms;
1644   SymbolRVASet giatsRVASet;
1645   std::vector<Symbol *> giatsSymbols;
1646   SymbolRVASet longJmpTargets;
1647   SymbolRVASet ehContTargets;
1648   for (ObjFile *file : ctx.objFileInstances) {
1649     // If the object was compiled with /guard:cf, the address taken symbols
1650     // are in .gfids$y sections, the longjmp targets are in .gljmp$y sections,
1651     // and ehcont targets are in .gehcont$y sections. If the object was not
1652     // compiled with /guard:cf, we assume there were no setjmp and ehcont
1653     // targets, and that all code symbols with relocations are possibly
1654     // address-taken.
1655     if (file->hasGuardCF()) {
1656       markSymbolsForRVATable(file, file->getGuardFidChunks(), addressTakenSyms);
1657       markSymbolsForRVATable(file, file->getGuardIATChunks(), giatsRVASet);
1658       getSymbolsFromSections(file, file->getGuardIATChunks(), giatsSymbols);
1659       markSymbolsForRVATable(file, file->getGuardLJmpChunks(), longJmpTargets);
1660       markSymbolsForRVATable(file, file->getGuardEHContChunks(), ehContTargets);
1661     } else {
1662       markSymbolsWithRelocations(file, addressTakenSyms);
1663     }
1664   }
1665 
1666   // Mark the image entry as address-taken.
1667   if (config->entry)
1668     maybeAddAddressTakenFunction(addressTakenSyms, config->entry);
1669 
1670   // Mark exported symbols in executable sections as address-taken.
1671   for (Export &e : config->exports)
1672     maybeAddAddressTakenFunction(addressTakenSyms, e.sym);
1673 
1674   // For each entry in the .giats table, check if it has a corresponding load
1675   // thunk (e.g. because the DLL that defines it will be delay-loaded) and, if
1676   // so, add the load thunk to the address taken (.gfids) table.
1677   for (Symbol *s : giatsSymbols) {
1678     if (auto *di = dyn_cast<DefinedImportData>(s)) {
1679       if (di->loadThunkSym)
1680         addSymbolToRVASet(addressTakenSyms, di->loadThunkSym);
1681     }
1682   }
1683 
1684   // Ensure sections referenced in the gfid table are 16-byte aligned.
1685   for (const ChunkAndOffset &c : addressTakenSyms)
1686     if (c.inputChunk->getAlignment() < 16)
1687       c.inputChunk->setAlignment(16);
1688 
1689   maybeAddRVATable(std::move(addressTakenSyms), "__guard_fids_table",
1690                    "__guard_fids_count");
1691 
1692   // Add the Guard Address Taken IAT Entry Table (.giats).
1693   maybeAddRVATable(std::move(giatsRVASet), "__guard_iat_table",
1694                    "__guard_iat_count");
1695 
1696   // Add the longjmp target table unless the user told us not to.
1697   if (config->guardCF & GuardCFLevel::LongJmp)
1698     maybeAddRVATable(std::move(longJmpTargets), "__guard_longjmp_table",
1699                      "__guard_longjmp_count");
1700 
1701   // Add the ehcont target table unless the user told us not to.
1702   if (config->guardCF & GuardCFLevel::EHCont)
1703     maybeAddRVATable(std::move(ehContTargets), "__guard_eh_cont_table",
1704                      "__guard_eh_cont_count", true);
1705 
1706   // Set __guard_flags, which will be used in the load config to indicate that
1707   // /guard:cf was enabled.
1708   uint32_t guardFlags = uint32_t(coff_guard_flags::CFInstrumented) |
1709                         uint32_t(coff_guard_flags::HasFidTable);
1710   if (config->guardCF & GuardCFLevel::LongJmp)
1711     guardFlags |= uint32_t(coff_guard_flags::HasLongJmpTable);
1712   if (config->guardCF & GuardCFLevel::EHCont)
1713     guardFlags |= uint32_t(coff_guard_flags::HasEHContTable);
1714   Symbol *flagSym = ctx.symtab.findUnderscore("__guard_flags");
1715   cast<DefinedAbsolute>(flagSym)->setVA(guardFlags);
1716 }
1717 
1718 // Take a list of input sections containing symbol table indices and add those
1719 // symbols to a vector. The challenge is that symbol RVAs are not known and
1720 // depend on the table size, so we can't directly build a set of integers.
1721 void Writer::getSymbolsFromSections(ObjFile *file,
1722                                     ArrayRef<SectionChunk *> symIdxChunks,
1723                                     std::vector<Symbol *> &symbols) {
1724   for (SectionChunk *c : symIdxChunks) {
1725     // Skip sections discarded by linker GC. This comes up when a .gfids section
1726     // is associated with something like a vtable and the vtable is discarded.
1727     // In this case, the associated gfids section is discarded, and we don't
1728     // mark the virtual member functions as address-taken by the vtable.
1729     if (!c->live)
1730       continue;
1731 
1732     // Validate that the contents look like symbol table indices.
1733     ArrayRef<uint8_t> data = c->getContents();
1734     if (data.size() % 4 != 0) {
1735       warn("ignoring " + c->getSectionName() +
1736            " symbol table index section in object " + toString(file));
1737       continue;
1738     }
1739 
1740     // Read each symbol table index and check if that symbol was included in the
1741     // final link. If so, add it to the vector of symbols.
1742     ArrayRef<ulittle32_t> symIndices(
1743         reinterpret_cast<const ulittle32_t *>(data.data()), data.size() / 4);
1744     ArrayRef<Symbol *> objSymbols = file->getSymbols();
1745     for (uint32_t symIndex : symIndices) {
1746       if (symIndex >= objSymbols.size()) {
1747         warn("ignoring invalid symbol table index in section " +
1748              c->getSectionName() + " in object " + toString(file));
1749         continue;
1750       }
1751       if (Symbol *s = objSymbols[symIndex]) {
1752         if (s->isLive())
1753           symbols.push_back(cast<Symbol>(s));
1754       }
1755     }
1756   }
1757 }
1758 
1759 // Take a list of input sections containing symbol table indices and add those
1760 // symbols to an RVA table.
1761 void Writer::markSymbolsForRVATable(ObjFile *file,
1762                                     ArrayRef<SectionChunk *> symIdxChunks,
1763                                     SymbolRVASet &tableSymbols) {
1764   std::vector<Symbol *> syms;
1765   getSymbolsFromSections(file, symIdxChunks, syms);
1766 
1767   for (Symbol *s : syms)
1768     addSymbolToRVASet(tableSymbols, cast<Defined>(s));
1769 }
1770 
1771 // Replace the absolute table symbol with a synthetic symbol pointing to
1772 // tableChunk so that we can emit base relocations for it and resolve section
1773 // relative relocations.
1774 void Writer::maybeAddRVATable(SymbolRVASet tableSymbols, StringRef tableSym,
1775                               StringRef countSym, bool hasFlag) {
1776   if (tableSymbols.empty())
1777     return;
1778 
1779   NonSectionChunk *tableChunk;
1780   if (hasFlag)
1781     tableChunk = make<RVAFlagTableChunk>(std::move(tableSymbols));
1782   else
1783     tableChunk = make<RVATableChunk>(std::move(tableSymbols));
1784   rdataSec->addChunk(tableChunk);
1785 
1786   Symbol *t = ctx.symtab.findUnderscore(tableSym);
1787   Symbol *c = ctx.symtab.findUnderscore(countSym);
1788   replaceSymbol<DefinedSynthetic>(t, t->getName(), tableChunk);
1789   cast<DefinedAbsolute>(c)->setVA(tableChunk->getSize() / (hasFlag ? 5 : 4));
1790 }
1791 
1792 // MinGW specific. Gather all relocations that are imported from a DLL even
1793 // though the code didn't expect it to, produce the table that the runtime
1794 // uses for fixing them up, and provide the synthetic symbols that the
1795 // runtime uses for finding the table.
1796 void Writer::createRuntimePseudoRelocs() {
1797   std::vector<RuntimePseudoReloc> rels;
1798 
1799   for (Chunk *c : ctx.symtab.getChunks()) {
1800     auto *sc = dyn_cast<SectionChunk>(c);
1801     if (!sc || !sc->live)
1802       continue;
1803     sc->getRuntimePseudoRelocs(rels);
1804   }
1805 
1806   if (!config->pseudoRelocs) {
1807     // Not writing any pseudo relocs; if some were needed, error out and
1808     // indicate what required them.
1809     for (const RuntimePseudoReloc &rpr : rels)
1810       error("automatic dllimport of " + rpr.sym->getName() + " in " +
1811             toString(rpr.target->file) + " requires pseudo relocations");
1812     return;
1813   }
1814 
1815   if (!rels.empty())
1816     log("Writing " + Twine(rels.size()) + " runtime pseudo relocations");
1817   PseudoRelocTableChunk *table = make<PseudoRelocTableChunk>(rels);
1818   rdataSec->addChunk(table);
1819   EmptyChunk *endOfList = make<EmptyChunk>();
1820   rdataSec->addChunk(endOfList);
1821 
1822   Symbol *headSym = ctx.symtab.findUnderscore("__RUNTIME_PSEUDO_RELOC_LIST__");
1823   Symbol *endSym =
1824       ctx.symtab.findUnderscore("__RUNTIME_PSEUDO_RELOC_LIST_END__");
1825   replaceSymbol<DefinedSynthetic>(headSym, headSym->getName(), table);
1826   replaceSymbol<DefinedSynthetic>(endSym, endSym->getName(), endOfList);
1827 }
1828 
1829 // MinGW specific.
1830 // The MinGW .ctors and .dtors lists have sentinels at each end;
1831 // a (uintptr_t)-1 at the start and a (uintptr_t)0 at the end.
1832 // There's a symbol pointing to the start sentinel pointer, __CTOR_LIST__
1833 // and __DTOR_LIST__ respectively.
1834 void Writer::insertCtorDtorSymbols() {
1835   AbsolutePointerChunk *ctorListHead = make<AbsolutePointerChunk>(-1);
1836   AbsolutePointerChunk *ctorListEnd = make<AbsolutePointerChunk>(0);
1837   AbsolutePointerChunk *dtorListHead = make<AbsolutePointerChunk>(-1);
1838   AbsolutePointerChunk *dtorListEnd = make<AbsolutePointerChunk>(0);
1839   ctorsSec->insertChunkAtStart(ctorListHead);
1840   ctorsSec->addChunk(ctorListEnd);
1841   dtorsSec->insertChunkAtStart(dtorListHead);
1842   dtorsSec->addChunk(dtorListEnd);
1843 
1844   Symbol *ctorListSym = ctx.symtab.findUnderscore("__CTOR_LIST__");
1845   Symbol *dtorListSym = ctx.symtab.findUnderscore("__DTOR_LIST__");
1846   replaceSymbol<DefinedSynthetic>(ctorListSym, ctorListSym->getName(),
1847                                   ctorListHead);
1848   replaceSymbol<DefinedSynthetic>(dtorListSym, dtorListSym->getName(),
1849                                   dtorListHead);
1850 }
1851 
1852 // Handles /section options to allow users to overwrite
1853 // section attributes.
1854 void Writer::setSectionPermissions() {
1855   for (auto &p : config->section) {
1856     StringRef name = p.first;
1857     uint32_t perm = p.second;
1858     for (OutputSection *sec : ctx.outputSections)
1859       if (sec->name == name)
1860         sec->setPermissions(perm);
1861   }
1862 }
1863 
1864 // Write section contents to a mmap'ed file.
1865 void Writer::writeSections() {
1866   // Record the number of sections to apply section index relocations
1867   // against absolute symbols. See applySecIdx in Chunks.cpp..
1868   DefinedAbsolute::numOutputSections = ctx.outputSections.size();
1869 
1870   uint8_t *buf = buffer->getBufferStart();
1871   for (OutputSection *sec : ctx.outputSections) {
1872     uint8_t *secBuf = buf + sec->getFileOff();
1873     // Fill gaps between functions in .text with INT3 instructions
1874     // instead of leaving as NUL bytes (which can be interpreted as
1875     // ADD instructions).
1876     if (sec->header.Characteristics & IMAGE_SCN_CNT_CODE)
1877       memset(secBuf, 0xCC, sec->getRawSize());
1878     parallelForEach(sec->chunks, [&](Chunk *c) {
1879       c->writeTo(secBuf + c->getRVA() - sec->getRVA());
1880     });
1881   }
1882 }
1883 
1884 void Writer::writeBuildId() {
1885   // There are two important parts to the build ID.
1886   // 1) If building with debug info, the COFF debug directory contains a
1887   //    timestamp as well as a Guid and Age of the PDB.
1888   // 2) In all cases, the PE COFF file header also contains a timestamp.
1889   // For reproducibility, instead of a timestamp we want to use a hash of the
1890   // PE contents.
1891   if (config->debug) {
1892     assert(buildId && "BuildId is not set!");
1893     // BuildId->BuildId was filled in when the PDB was written.
1894   }
1895 
1896   // At this point the only fields in the COFF file which remain unset are the
1897   // "timestamp" in the COFF file header, and the ones in the coff debug
1898   // directory.  Now we can hash the file and write that hash to the various
1899   // timestamp fields in the file.
1900   StringRef outputFileData(
1901       reinterpret_cast<const char *>(buffer->getBufferStart()),
1902       buffer->getBufferSize());
1903 
1904   uint32_t timestamp = config->timestamp;
1905   uint64_t hash = 0;
1906   bool generateSyntheticBuildId =
1907       config->mingw && config->debug && config->pdbPath.empty();
1908 
1909   if (config->repro || generateSyntheticBuildId)
1910     hash = xxHash64(outputFileData);
1911 
1912   if (config->repro)
1913     timestamp = static_cast<uint32_t>(hash);
1914 
1915   if (generateSyntheticBuildId) {
1916     // For MinGW builds without a PDB file, we still generate a build id
1917     // to allow associating a crash dump to the executable.
1918     buildId->buildId->PDB70.CVSignature = OMF::Signature::PDB70;
1919     buildId->buildId->PDB70.Age = 1;
1920     memcpy(buildId->buildId->PDB70.Signature, &hash, 8);
1921     // xxhash only gives us 8 bytes, so put some fixed data in the other half.
1922     memcpy(&buildId->buildId->PDB70.Signature[8], "LLD PDB.", 8);
1923   }
1924 
1925   if (debugDirectory)
1926     debugDirectory->setTimeDateStamp(timestamp);
1927 
1928   uint8_t *buf = buffer->getBufferStart();
1929   buf += dosStubSize + sizeof(PEMagic);
1930   object::coff_file_header *coffHeader =
1931       reinterpret_cast<coff_file_header *>(buf);
1932   coffHeader->TimeDateStamp = timestamp;
1933 }
1934 
1935 // Sort .pdata section contents according to PE/COFF spec 5.5.
1936 void Writer::sortExceptionTable() {
1937   if (!firstPdata)
1938     return;
1939   // We assume .pdata contains function table entries only.
1940   auto bufAddr = [&](Chunk *c) {
1941     OutputSection *os = ctx.getOutputSection(c);
1942     return buffer->getBufferStart() + os->getFileOff() + c->getRVA() -
1943            os->getRVA();
1944   };
1945   uint8_t *begin = bufAddr(firstPdata);
1946   uint8_t *end = bufAddr(lastPdata) + lastPdata->getSize();
1947   if (config->machine == AMD64) {
1948     struct Entry { ulittle32_t begin, end, unwind; };
1949     if ((end - begin) % sizeof(Entry) != 0) {
1950       fatal("unexpected .pdata size: " + Twine(end - begin) +
1951             " is not a multiple of " + Twine(sizeof(Entry)));
1952     }
1953     parallelSort(
1954         MutableArrayRef<Entry>((Entry *)begin, (Entry *)end),
1955         [](const Entry &a, const Entry &b) { return a.begin < b.begin; });
1956     return;
1957   }
1958   if (config->machine == ARMNT || config->machine == ARM64) {
1959     struct Entry { ulittle32_t begin, unwind; };
1960     if ((end - begin) % sizeof(Entry) != 0) {
1961       fatal("unexpected .pdata size: " + Twine(end - begin) +
1962             " is not a multiple of " + Twine(sizeof(Entry)));
1963     }
1964     parallelSort(
1965         MutableArrayRef<Entry>((Entry *)begin, (Entry *)end),
1966         [](const Entry &a, const Entry &b) { return a.begin < b.begin; });
1967     return;
1968   }
1969   lld::errs() << "warning: don't know how to handle .pdata.\n";
1970 }
1971 
1972 // The CRT section contains, among other things, the array of function
1973 // pointers that initialize every global variable that is not trivially
1974 // constructed. The CRT calls them one after the other prior to invoking
1975 // main().
1976 //
1977 // As per C++ spec, 3.6.2/2.3,
1978 // "Variables with ordered initialization defined within a single
1979 // translation unit shall be initialized in the order of their definitions
1980 // in the translation unit"
1981 //
1982 // It is therefore critical to sort the chunks containing the function
1983 // pointers in the order that they are listed in the object file (top to
1984 // bottom), otherwise global objects might not be initialized in the
1985 // correct order.
1986 void Writer::sortCRTSectionChunks(std::vector<Chunk *> &chunks) {
1987   auto sectionChunkOrder = [](const Chunk *a, const Chunk *b) {
1988     auto sa = dyn_cast<SectionChunk>(a);
1989     auto sb = dyn_cast<SectionChunk>(b);
1990     assert(sa && sb && "Non-section chunks in CRT section!");
1991 
1992     StringRef sAObj = sa->file->mb.getBufferIdentifier();
1993     StringRef sBObj = sb->file->mb.getBufferIdentifier();
1994 
1995     return sAObj == sBObj && sa->getSectionNumber() < sb->getSectionNumber();
1996   };
1997   llvm::stable_sort(chunks, sectionChunkOrder);
1998 
1999   if (config->verbose) {
2000     for (auto &c : chunks) {
2001       auto sc = dyn_cast<SectionChunk>(c);
2002       log("  " + sc->file->mb.getBufferIdentifier().str() +
2003           ", SectionID: " + Twine(sc->getSectionNumber()));
2004     }
2005   }
2006 }
2007 
2008 OutputSection *Writer::findSection(StringRef name) {
2009   for (OutputSection *sec : ctx.outputSections)
2010     if (sec->name == name)
2011       return sec;
2012   return nullptr;
2013 }
2014 
2015 uint32_t Writer::getSizeOfInitializedData() {
2016   uint32_t res = 0;
2017   for (OutputSection *s : ctx.outputSections)
2018     if (s->header.Characteristics & IMAGE_SCN_CNT_INITIALIZED_DATA)
2019       res += s->getRawSize();
2020   return res;
2021 }
2022 
2023 // Add base relocations to .reloc section.
2024 void Writer::addBaserels() {
2025   if (!config->relocatable)
2026     return;
2027   relocSec->chunks.clear();
2028   std::vector<Baserel> v;
2029   for (OutputSection *sec : ctx.outputSections) {
2030     if (sec->header.Characteristics & IMAGE_SCN_MEM_DISCARDABLE)
2031       continue;
2032     // Collect all locations for base relocations.
2033     for (Chunk *c : sec->chunks)
2034       c->getBaserels(&v);
2035     // Add the addresses to .reloc section.
2036     if (!v.empty())
2037       addBaserelBlocks(v);
2038     v.clear();
2039   }
2040 }
2041 
2042 // Add addresses to .reloc section. Note that addresses are grouped by page.
2043 void Writer::addBaserelBlocks(std::vector<Baserel> &v) {
2044   const uint32_t mask = ~uint32_t(pageSize - 1);
2045   uint32_t page = v[0].rva & mask;
2046   size_t i = 0, j = 1;
2047   for (size_t e = v.size(); j < e; ++j) {
2048     uint32_t p = v[j].rva & mask;
2049     if (p == page)
2050       continue;
2051     relocSec->addChunk(make<BaserelChunk>(page, &v[i], &v[0] + j));
2052     i = j;
2053     page = p;
2054   }
2055   if (i == j)
2056     return;
2057   relocSec->addChunk(make<BaserelChunk>(page, &v[i], &v[0] + j));
2058 }
2059 
2060 PartialSection *Writer::createPartialSection(StringRef name,
2061                                              uint32_t outChars) {
2062   PartialSection *&pSec = partialSections[{name, outChars}];
2063   if (pSec)
2064     return pSec;
2065   pSec = make<PartialSection>(name, outChars);
2066   return pSec;
2067 }
2068 
2069 PartialSection *Writer::findPartialSection(StringRef name, uint32_t outChars) {
2070   auto it = partialSections.find({name, outChars});
2071   if (it != partialSections.end())
2072     return it->second;
2073   return nullptr;
2074 }
2075 
2076 void Writer::fixTlsAlignment() {
2077   Defined *tlsSym =
2078       dyn_cast_or_null<Defined>(ctx.symtab.findUnderscore("_tls_used"));
2079   if (!tlsSym)
2080     return;
2081 
2082   OutputSection *sec = ctx.getOutputSection(tlsSym->getChunk());
2083   assert(sec && tlsSym->getRVA() >= sec->getRVA() &&
2084          "no output section for _tls_used");
2085 
2086   uint8_t *secBuf = buffer->getBufferStart() + sec->getFileOff();
2087   uint64_t tlsOffset = tlsSym->getRVA() - sec->getRVA();
2088   uint64_t directorySize = config->is64()
2089                                ? sizeof(object::coff_tls_directory64)
2090                                : sizeof(object::coff_tls_directory32);
2091 
2092   if (tlsOffset + directorySize > sec->getRawSize())
2093     fatal("_tls_used sym is malformed");
2094 
2095   if (config->is64()) {
2096     object::coff_tls_directory64 *tlsDir =
2097         reinterpret_cast<object::coff_tls_directory64 *>(&secBuf[tlsOffset]);
2098     tlsDir->setAlignment(tlsAlignment);
2099   } else {
2100     object::coff_tls_directory32 *tlsDir =
2101         reinterpret_cast<object::coff_tls_directory32 *>(&secBuf[tlsOffset]);
2102     tlsDir->setAlignment(tlsAlignment);
2103   }
2104 }
2105