1 //===- lib/MC/WasmObjectWriter.cpp - Wasm File Writer ---------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements Wasm object file writer information.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/ADT/STLExtras.h"
14 #include "llvm/ADT/SmallPtrSet.h"
15 #include "llvm/BinaryFormat/Wasm.h"
16 #include "llvm/BinaryFormat/WasmTraits.h"
17 #include "llvm/Config/llvm-config.h"
18 #include "llvm/MC/MCAsmBackend.h"
19 #include "llvm/MC/MCAsmLayout.h"
20 #include "llvm/MC/MCAssembler.h"
21 #include "llvm/MC/MCContext.h"
22 #include "llvm/MC/MCExpr.h"
23 #include "llvm/MC/MCFixupKindInfo.h"
24 #include "llvm/MC/MCObjectWriter.h"
25 #include "llvm/MC/MCSectionWasm.h"
26 #include "llvm/MC/MCSymbolWasm.h"
27 #include "llvm/MC/MCValue.h"
28 #include "llvm/MC/MCWasmObjectWriter.h"
29 #include "llvm/Support/Casting.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/EndianStream.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/LEB128.h"
34 #include "llvm/Support/StringSaver.h"
35 #include <vector>
36 
37 using namespace llvm;
38 
39 #define DEBUG_TYPE "mc"
40 
41 namespace {
42 
43 // When we create the indirect function table we start at 1, so that there is
44 // and empty slot at 0 and therefore calling a null function pointer will trap.
45 static const uint32_t InitialTableOffset = 1;
46 
47 // For patching purposes, we need to remember where each section starts, both
48 // for patching up the section size field, and for patching up references to
49 // locations within the section.
50 struct SectionBookkeeping {
51   // Where the size of the section is written.
52   uint64_t SizeOffset;
53   // Where the section header ends (without custom section name).
54   uint64_t PayloadOffset;
55   // Where the contents of the section starts.
56   uint64_t ContentsOffset;
57   uint32_t Index;
58 };
59 
60 // A wasm data segment.  A wasm binary contains only a single data section
61 // but that can contain many segments, each with their own virtual location
62 // in memory.  Each MCSection data created by llvm is modeled as its own
63 // wasm data segment.
64 struct WasmDataSegment {
65   MCSectionWasm *Section;
66   StringRef Name;
67   uint32_t InitFlags;
68   uint64_t Offset;
69   uint32_t Alignment;
70   uint32_t LinkerFlags;
71   SmallVector<char, 4> Data;
72 };
73 
74 // A wasm function to be written into the function section.
75 struct WasmFunction {
76   uint32_t SigIndex;
77   const MCSymbolWasm *Sym;
78 };
79 
80 // A wasm global to be written into the global section.
81 struct WasmGlobal {
82   wasm::WasmGlobalType Type;
83   uint64_t InitialValue;
84 };
85 
86 // Information about a single item which is part of a COMDAT.  For each data
87 // segment or function which is in the COMDAT, there is a corresponding
88 // WasmComdatEntry.
89 struct WasmComdatEntry {
90   unsigned Kind;
91   uint32_t Index;
92 };
93 
94 // Information about a single relocation.
95 struct WasmRelocationEntry {
96   uint64_t Offset;                   // Where is the relocation.
97   const MCSymbolWasm *Symbol;        // The symbol to relocate with.
98   int64_t Addend;                    // A value to add to the symbol.
99   unsigned Type;                     // The type of the relocation.
100   const MCSectionWasm *FixupSection; // The section the relocation is targeting.
101 
WasmRelocationEntry__anonbf91e1730111::WasmRelocationEntry102   WasmRelocationEntry(uint64_t Offset, const MCSymbolWasm *Symbol,
103                       int64_t Addend, unsigned Type,
104                       const MCSectionWasm *FixupSection)
105       : Offset(Offset), Symbol(Symbol), Addend(Addend), Type(Type),
106         FixupSection(FixupSection) {}
107 
hasAddend__anonbf91e1730111::WasmRelocationEntry108   bool hasAddend() const { return wasm::relocTypeHasAddend(Type); }
109 
print__anonbf91e1730111::WasmRelocationEntry110   void print(raw_ostream &Out) const {
111     Out << wasm::relocTypetoString(Type) << " Off=" << Offset
112         << ", Sym=" << *Symbol << ", Addend=" << Addend
113         << ", FixupSection=" << FixupSection->getName();
114   }
115 
116 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
dump__anonbf91e1730111::WasmRelocationEntry117   LLVM_DUMP_METHOD void dump() const { print(dbgs()); }
118 #endif
119 };
120 
121 static const uint32_t InvalidIndex = -1;
122 
123 struct WasmCustomSection {
124 
125   StringRef Name;
126   MCSectionWasm *Section;
127 
128   uint32_t OutputContentsOffset;
129   uint32_t OutputIndex;
130 
WasmCustomSection__anonbf91e1730111::WasmCustomSection131   WasmCustomSection(StringRef Name, MCSectionWasm *Section)
132       : Name(Name), Section(Section), OutputContentsOffset(0),
133         OutputIndex(InvalidIndex) {}
134 };
135 
136 #if !defined(NDEBUG)
operator <<(raw_ostream & OS,const WasmRelocationEntry & Rel)137 raw_ostream &operator<<(raw_ostream &OS, const WasmRelocationEntry &Rel) {
138   Rel.print(OS);
139   return OS;
140 }
141 #endif
142 
143 // Write X as an (unsigned) LEB value at offset Offset in Stream, padded
144 // to allow patching.
145 template <int W>
writePatchableLEB(raw_pwrite_stream & Stream,uint64_t X,uint64_t Offset)146 void writePatchableLEB(raw_pwrite_stream &Stream, uint64_t X, uint64_t Offset) {
147   uint8_t Buffer[W];
148   unsigned SizeLen = encodeULEB128(X, Buffer, W);
149   assert(SizeLen == W);
150   Stream.pwrite((char *)Buffer, SizeLen, Offset);
151 }
152 
153 // Write X as an signed LEB value at offset Offset in Stream, padded
154 // to allow patching.
155 template <int W>
writePatchableSLEB(raw_pwrite_stream & Stream,int64_t X,uint64_t Offset)156 void writePatchableSLEB(raw_pwrite_stream &Stream, int64_t X, uint64_t Offset) {
157   uint8_t Buffer[W];
158   unsigned SizeLen = encodeSLEB128(X, Buffer, W);
159   assert(SizeLen == W);
160   Stream.pwrite((char *)Buffer, SizeLen, Offset);
161 }
162 
163 // Write X as a plain integer value at offset Offset in Stream.
patchI32(raw_pwrite_stream & Stream,uint32_t X,uint64_t Offset)164 static void patchI32(raw_pwrite_stream &Stream, uint32_t X, uint64_t Offset) {
165   uint8_t Buffer[4];
166   support::endian::write32le(Buffer, X);
167   Stream.pwrite((char *)Buffer, sizeof(Buffer), Offset);
168 }
169 
patchI64(raw_pwrite_stream & Stream,uint64_t X,uint64_t Offset)170 static void patchI64(raw_pwrite_stream &Stream, uint64_t X, uint64_t Offset) {
171   uint8_t Buffer[8];
172   support::endian::write64le(Buffer, X);
173   Stream.pwrite((char *)Buffer, sizeof(Buffer), Offset);
174 }
175 
isDwoSection(const MCSection & Sec)176 bool isDwoSection(const MCSection &Sec) {
177   return Sec.getName().endswith(".dwo");
178 }
179 
180 class WasmObjectWriter : public MCObjectWriter {
181   support::endian::Writer *W;
182 
183   /// The target specific Wasm writer instance.
184   std::unique_ptr<MCWasmObjectTargetWriter> TargetObjectWriter;
185 
186   // Relocations for fixing up references in the code section.
187   std::vector<WasmRelocationEntry> CodeRelocations;
188   // Relocations for fixing up references in the data section.
189   std::vector<WasmRelocationEntry> DataRelocations;
190 
191   // Index values to use for fixing up call_indirect type indices.
192   // Maps function symbols to the index of the type of the function
193   DenseMap<const MCSymbolWasm *, uint32_t> TypeIndices;
194   // Maps function symbols to the table element index space. Used
195   // for TABLE_INDEX relocation types (i.e. address taken functions).
196   DenseMap<const MCSymbolWasm *, uint32_t> TableIndices;
197   // Maps function/global/table symbols to the
198   // function/global/table/event/section index space.
199   DenseMap<const MCSymbolWasm *, uint32_t> WasmIndices;
200   DenseMap<const MCSymbolWasm *, uint32_t> GOTIndices;
201   // Maps data symbols to the Wasm segment and offset/size with the segment.
202   DenseMap<const MCSymbolWasm *, wasm::WasmDataReference> DataLocations;
203 
204   // Stores output data (index, relocations, content offset) for custom
205   // section.
206   std::vector<WasmCustomSection> CustomSections;
207   std::unique_ptr<WasmCustomSection> ProducersSection;
208   std::unique_ptr<WasmCustomSection> TargetFeaturesSection;
209   // Relocations for fixing up references in the custom sections.
210   DenseMap<const MCSectionWasm *, std::vector<WasmRelocationEntry>>
211       CustomSectionsRelocations;
212 
213   // Map from section to defining function symbol.
214   DenseMap<const MCSection *, const MCSymbol *> SectionFunctions;
215 
216   DenseMap<wasm::WasmSignature, uint32_t> SignatureIndices;
217   SmallVector<wasm::WasmSignature, 4> Signatures;
218   SmallVector<WasmDataSegment, 4> DataSegments;
219   unsigned NumFunctionImports = 0;
220   unsigned NumGlobalImports = 0;
221   unsigned NumTableImports = 0;
222   unsigned NumEventImports = 0;
223   uint32_t SectionCount = 0;
224 
225   enum class DwoMode {
226     AllSections,
227     NonDwoOnly,
228     DwoOnly,
229   };
230   bool IsSplitDwarf = false;
231   raw_pwrite_stream *OS = nullptr;
232   raw_pwrite_stream *DwoOS = nullptr;
233 
234   // TargetObjectWriter wranppers.
is64Bit() const235   bool is64Bit() const { return TargetObjectWriter->is64Bit(); }
isEmscripten() const236   bool isEmscripten() const { return TargetObjectWriter->isEmscripten(); }
237 
238   void startSection(SectionBookkeeping &Section, unsigned SectionId);
239   void startCustomSection(SectionBookkeeping &Section, StringRef Name);
240   void endSection(SectionBookkeeping &Section);
241 
242 public:
WasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW,raw_pwrite_stream & OS_)243   WasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW,
244                    raw_pwrite_stream &OS_)
245       : TargetObjectWriter(std::move(MOTW)), OS(&OS_) {}
246 
WasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW,raw_pwrite_stream & OS_,raw_pwrite_stream & DwoOS_)247   WasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW,
248                    raw_pwrite_stream &OS_, raw_pwrite_stream &DwoOS_)
249       : TargetObjectWriter(std::move(MOTW)), IsSplitDwarf(true), OS(&OS_),
250         DwoOS(&DwoOS_) {}
251 
252 private:
reset()253   void reset() override {
254     CodeRelocations.clear();
255     DataRelocations.clear();
256     TypeIndices.clear();
257     WasmIndices.clear();
258     GOTIndices.clear();
259     TableIndices.clear();
260     DataLocations.clear();
261     CustomSections.clear();
262     ProducersSection.reset();
263     TargetFeaturesSection.reset();
264     CustomSectionsRelocations.clear();
265     SignatureIndices.clear();
266     Signatures.clear();
267     DataSegments.clear();
268     SectionFunctions.clear();
269     NumFunctionImports = 0;
270     NumGlobalImports = 0;
271     NumTableImports = 0;
272     MCObjectWriter::reset();
273   }
274 
275   void writeHeader(const MCAssembler &Asm);
276 
277   void recordRelocation(MCAssembler &Asm, const MCAsmLayout &Layout,
278                         const MCFragment *Fragment, const MCFixup &Fixup,
279                         MCValue Target, uint64_t &FixedValue) override;
280 
281   void executePostLayoutBinding(MCAssembler &Asm,
282                                 const MCAsmLayout &Layout) override;
283   void prepareImports(SmallVectorImpl<wasm::WasmImport> &Imports,
284                       MCAssembler &Asm, const MCAsmLayout &Layout);
285   uint64_t writeObject(MCAssembler &Asm, const MCAsmLayout &Layout) override;
286 
287   uint64_t writeOneObject(MCAssembler &Asm, const MCAsmLayout &Layout,
288                           DwoMode Mode);
289 
writeString(const StringRef Str)290   void writeString(const StringRef Str) {
291     encodeULEB128(Str.size(), W->OS);
292     W->OS << Str;
293   }
294 
writeI32(int32_t val)295   void writeI32(int32_t val) {
296     char Buffer[4];
297     support::endian::write32le(Buffer, val);
298     W->OS.write(Buffer, sizeof(Buffer));
299   }
300 
writeI64(int64_t val)301   void writeI64(int64_t val) {
302     char Buffer[8];
303     support::endian::write64le(Buffer, val);
304     W->OS.write(Buffer, sizeof(Buffer));
305   }
306 
writeValueType(wasm::ValType Ty)307   void writeValueType(wasm::ValType Ty) { W->OS << static_cast<char>(Ty); }
308 
309   void writeTypeSection(ArrayRef<wasm::WasmSignature> Signatures);
310   void writeImportSection(ArrayRef<wasm::WasmImport> Imports, uint64_t DataSize,
311                           uint32_t NumElements);
312   void writeFunctionSection(ArrayRef<WasmFunction> Functions);
313   void writeExportSection(ArrayRef<wasm::WasmExport> Exports);
314   void writeElemSection(ArrayRef<uint32_t> TableElems);
315   void writeDataCountSection();
316   uint32_t writeCodeSection(const MCAssembler &Asm, const MCAsmLayout &Layout,
317                             ArrayRef<WasmFunction> Functions);
318   uint32_t writeDataSection(const MCAsmLayout &Layout);
319   void writeEventSection(ArrayRef<wasm::WasmEventType> Events);
320   void writeGlobalSection(ArrayRef<wasm::WasmGlobal> Globals);
321   void writeTableSection(ArrayRef<wasm::WasmTable> Tables);
322   void writeRelocSection(uint32_t SectionIndex, StringRef Name,
323                          std::vector<WasmRelocationEntry> &Relocations);
324   void writeLinkingMetaDataSection(
325       ArrayRef<wasm::WasmSymbolInfo> SymbolInfos,
326       ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs,
327       const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats);
328   void writeCustomSection(WasmCustomSection &CustomSection,
329                           const MCAssembler &Asm, const MCAsmLayout &Layout);
330   void writeCustomRelocSections();
331 
332   uint64_t getProvisionalValue(const WasmRelocationEntry &RelEntry,
333                                const MCAsmLayout &Layout);
334   void applyRelocations(ArrayRef<WasmRelocationEntry> Relocations,
335                         uint64_t ContentsOffset, const MCAsmLayout &Layout);
336 
337   uint32_t getRelocationIndexValue(const WasmRelocationEntry &RelEntry);
338   uint32_t getFunctionType(const MCSymbolWasm &Symbol);
339   uint32_t getEventType(const MCSymbolWasm &Symbol);
340   void registerFunctionType(const MCSymbolWasm &Symbol);
341   void registerEventType(const MCSymbolWasm &Symbol);
342 };
343 
344 } // end anonymous namespace
345 
346 // Write out a section header and a patchable section size field.
startSection(SectionBookkeeping & Section,unsigned SectionId)347 void WasmObjectWriter::startSection(SectionBookkeeping &Section,
348                                     unsigned SectionId) {
349   LLVM_DEBUG(dbgs() << "startSection " << SectionId << "\n");
350   W->OS << char(SectionId);
351 
352   Section.SizeOffset = W->OS.tell();
353 
354   // The section size. We don't know the size yet, so reserve enough space
355   // for any 32-bit value; we'll patch it later.
356   encodeULEB128(0, W->OS, 5);
357 
358   // The position where the section starts, for measuring its size.
359   Section.ContentsOffset = W->OS.tell();
360   Section.PayloadOffset = W->OS.tell();
361   Section.Index = SectionCount++;
362 }
363 
startCustomSection(SectionBookkeeping & Section,StringRef Name)364 void WasmObjectWriter::startCustomSection(SectionBookkeeping &Section,
365                                           StringRef Name) {
366   LLVM_DEBUG(dbgs() << "startCustomSection " << Name << "\n");
367   startSection(Section, wasm::WASM_SEC_CUSTOM);
368 
369   // The position where the section header ends, for measuring its size.
370   Section.PayloadOffset = W->OS.tell();
371 
372   // Custom sections in wasm also have a string identifier.
373   writeString(Name);
374 
375   // The position where the custom section starts.
376   Section.ContentsOffset = W->OS.tell();
377 }
378 
379 // Now that the section is complete and we know how big it is, patch up the
380 // section size field at the start of the section.
endSection(SectionBookkeeping & Section)381 void WasmObjectWriter::endSection(SectionBookkeeping &Section) {
382   uint64_t Size = W->OS.tell();
383   // /dev/null doesn't support seek/tell and can report offset of 0.
384   // Simply skip this patching in that case.
385   if (!Size)
386     return;
387 
388   Size -= Section.PayloadOffset;
389   if (uint32_t(Size) != Size)
390     report_fatal_error("section size does not fit in a uint32_t");
391 
392   LLVM_DEBUG(dbgs() << "endSection size=" << Size << "\n");
393 
394   // Write the final section size to the payload_len field, which follows
395   // the section id byte.
396   writePatchableLEB<5>(static_cast<raw_pwrite_stream &>(W->OS), Size,
397                        Section.SizeOffset);
398 }
399 
400 // Emit the Wasm header.
writeHeader(const MCAssembler & Asm)401 void WasmObjectWriter::writeHeader(const MCAssembler &Asm) {
402   W->OS.write(wasm::WasmMagic, sizeof(wasm::WasmMagic));
403   W->write<uint32_t>(wasm::WasmVersion);
404 }
405 
executePostLayoutBinding(MCAssembler & Asm,const MCAsmLayout & Layout)406 void WasmObjectWriter::executePostLayoutBinding(MCAssembler &Asm,
407                                                 const MCAsmLayout &Layout) {
408   // As a stopgap measure until call_indirect instructions start explicitly
409   // referencing the indirect function table via TABLE_NUMBER relocs, ensure
410   // that the indirect function table import makes it to the output if anything
411   // in the compilation unit has caused it to be present.
412   if (auto *Sym = Asm.getContext().lookupSymbol("__indirect_function_table"))
413     Asm.registerSymbol(*Sym);
414 
415   // Build a map of sections to the function that defines them, for use
416   // in recordRelocation.
417   for (const MCSymbol &S : Asm.symbols()) {
418     const auto &WS = static_cast<const MCSymbolWasm &>(S);
419     if (WS.isDefined() && WS.isFunction() && !WS.isVariable()) {
420       const auto &Sec = static_cast<const MCSectionWasm &>(S.getSection());
421       auto Pair = SectionFunctions.insert(std::make_pair(&Sec, &S));
422       if (!Pair.second)
423         report_fatal_error("section already has a defining function: " +
424                            Sec.getName());
425     }
426   }
427 }
428 
recordRelocation(MCAssembler & Asm,const MCAsmLayout & Layout,const MCFragment * Fragment,const MCFixup & Fixup,MCValue Target,uint64_t & FixedValue)429 void WasmObjectWriter::recordRelocation(MCAssembler &Asm,
430                                         const MCAsmLayout &Layout,
431                                         const MCFragment *Fragment,
432                                         const MCFixup &Fixup, MCValue Target,
433                                         uint64_t &FixedValue) {
434   // The WebAssembly backend should never generate FKF_IsPCRel fixups
435   assert(!(Asm.getBackend().getFixupKindInfo(Fixup.getKind()).Flags &
436            MCFixupKindInfo::FKF_IsPCRel));
437 
438   const auto &FixupSection = cast<MCSectionWasm>(*Fragment->getParent());
439   uint64_t C = Target.getConstant();
440   uint64_t FixupOffset = Layout.getFragmentOffset(Fragment) + Fixup.getOffset();
441   MCContext &Ctx = Asm.getContext();
442 
443   if (const MCSymbolRefExpr *RefB = Target.getSymB()) {
444     // To get here the A - B expression must have failed evaluateAsRelocatable.
445     // This means either A or B must be undefined and in WebAssembly we can't
446     // support either of those cases.
447     const auto &SymB = cast<MCSymbolWasm>(RefB->getSymbol());
448     Ctx.reportError(
449         Fixup.getLoc(),
450         Twine("symbol '") + SymB.getName() +
451             "': unsupported subtraction expression used in relocation.");
452     return;
453   }
454 
455   // We either rejected the fixup or folded B into C at this point.
456   const MCSymbolRefExpr *RefA = Target.getSymA();
457   const auto *SymA = cast<MCSymbolWasm>(&RefA->getSymbol());
458 
459   // The .init_array isn't translated as data, so don't do relocations in it.
460   if (FixupSection.getName().startswith(".init_array")) {
461     SymA->setUsedInInitArray();
462     return;
463   }
464 
465   if (SymA->isVariable()) {
466     const MCExpr *Expr = SymA->getVariableValue();
467     if (const auto *Inner = dyn_cast<MCSymbolRefExpr>(Expr))
468       if (Inner->getKind() == MCSymbolRefExpr::VK_WEAKREF)
469         llvm_unreachable("weakref used in reloc not yet implemented");
470   }
471 
472   // Put any constant offset in an addend. Offsets can be negative, and
473   // LLVM expects wrapping, in contrast to wasm's immediates which can't
474   // be negative and don't wrap.
475   FixedValue = 0;
476 
477   unsigned Type = TargetObjectWriter->getRelocType(Target, Fixup);
478 
479   // Absolute offset within a section or a function.
480   // Currently only supported for for metadata sections.
481   // See: test/MC/WebAssembly/blockaddress.ll
482   if (Type == wasm::R_WASM_FUNCTION_OFFSET_I32 ||
483       Type == wasm::R_WASM_FUNCTION_OFFSET_I64 ||
484       Type == wasm::R_WASM_SECTION_OFFSET_I32) {
485     if (!FixupSection.getKind().isMetadata())
486       report_fatal_error("relocations for function or section offsets are "
487                          "only supported in metadata sections");
488 
489     const MCSymbol *SectionSymbol = nullptr;
490     const MCSection &SecA = SymA->getSection();
491     if (SecA.getKind().isText())
492       SectionSymbol = SectionFunctions.find(&SecA)->second;
493     else
494       SectionSymbol = SecA.getBeginSymbol();
495     if (!SectionSymbol)
496       report_fatal_error("section symbol is required for relocation");
497 
498     C += Layout.getSymbolOffset(*SymA);
499     SymA = cast<MCSymbolWasm>(SectionSymbol);
500   }
501 
502   if (Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB ||
503       Type == wasm::R_WASM_TABLE_INDEX_SLEB ||
504       Type == wasm::R_WASM_TABLE_INDEX_SLEB64 ||
505       Type == wasm::R_WASM_TABLE_INDEX_I32 ||
506       Type == wasm::R_WASM_TABLE_INDEX_I64) {
507     // TABLE_INDEX relocs implicitly use the default indirect function table.
508     auto TableName = "__indirect_function_table";
509     MCSymbolWasm *Sym = cast_or_null<MCSymbolWasm>(Ctx.lookupSymbol(TableName));
510     if (Sym) {
511       if (!Sym->isFunctionTable())
512         Ctx.reportError(
513             Fixup.getLoc(),
514             "symbol '__indirect_function_table' is not a function table");
515     } else {
516       Sym = cast<MCSymbolWasm>(Ctx.getOrCreateSymbol(TableName));
517       Sym->setFunctionTable();
518       // The default function table is synthesized by the linker.
519       Sym->setUndefined();
520     }
521     Sym->setUsedInReloc();
522     Asm.registerSymbol(*Sym);
523   }
524 
525   // Relocation other than R_WASM_TYPE_INDEX_LEB are required to be
526   // against a named symbol.
527   if (Type != wasm::R_WASM_TYPE_INDEX_LEB) {
528     if (SymA->getName().empty())
529       report_fatal_error("relocations against un-named temporaries are not yet "
530                          "supported by wasm");
531 
532     SymA->setUsedInReloc();
533   }
534 
535   if (RefA->getKind() == MCSymbolRefExpr::VK_GOT)
536     SymA->setUsedInGOT();
537 
538   WasmRelocationEntry Rec(FixupOffset, SymA, C, Type, &FixupSection);
539   LLVM_DEBUG(dbgs() << "WasmReloc: " << Rec << "\n");
540 
541   if (FixupSection.isWasmData()) {
542     DataRelocations.push_back(Rec);
543   } else if (FixupSection.getKind().isText()) {
544     CodeRelocations.push_back(Rec);
545   } else if (FixupSection.getKind().isMetadata()) {
546     CustomSectionsRelocations[&FixupSection].push_back(Rec);
547   } else {
548     llvm_unreachable("unexpected section type");
549   }
550 }
551 
552 // Compute a value to write into the code at the location covered
553 // by RelEntry. This value isn't used by the static linker; it just serves
554 // to make the object format more readable and more likely to be directly
555 // useable.
556 uint64_t
getProvisionalValue(const WasmRelocationEntry & RelEntry,const MCAsmLayout & Layout)557 WasmObjectWriter::getProvisionalValue(const WasmRelocationEntry &RelEntry,
558                                       const MCAsmLayout &Layout) {
559   if ((RelEntry.Type == wasm::R_WASM_GLOBAL_INDEX_LEB ||
560        RelEntry.Type == wasm::R_WASM_GLOBAL_INDEX_I32) &&
561       !RelEntry.Symbol->isGlobal()) {
562     assert(GOTIndices.count(RelEntry.Symbol) > 0 && "symbol not found in GOT index space");
563     return GOTIndices[RelEntry.Symbol];
564   }
565 
566   switch (RelEntry.Type) {
567   case wasm::R_WASM_TABLE_INDEX_REL_SLEB:
568   case wasm::R_WASM_TABLE_INDEX_SLEB:
569   case wasm::R_WASM_TABLE_INDEX_SLEB64:
570   case wasm::R_WASM_TABLE_INDEX_I32:
571   case wasm::R_WASM_TABLE_INDEX_I64: {
572     // Provisional value is table address of the resolved symbol itself
573     const MCSymbolWasm *Base =
574         cast<MCSymbolWasm>(Layout.getBaseSymbol(*RelEntry.Symbol));
575     assert(Base->isFunction());
576     if (RelEntry.Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB)
577       return TableIndices[Base] - InitialTableOffset;
578     else
579       return TableIndices[Base];
580   }
581   case wasm::R_WASM_TYPE_INDEX_LEB:
582     // Provisional value is same as the index
583     return getRelocationIndexValue(RelEntry);
584   case wasm::R_WASM_FUNCTION_INDEX_LEB:
585   case wasm::R_WASM_GLOBAL_INDEX_LEB:
586   case wasm::R_WASM_GLOBAL_INDEX_I32:
587   case wasm::R_WASM_EVENT_INDEX_LEB:
588   case wasm::R_WASM_TABLE_NUMBER_LEB:
589     // Provisional value is function/global/event Wasm index
590     assert(WasmIndices.count(RelEntry.Symbol) > 0 && "symbol not found in wasm index space");
591     return WasmIndices[RelEntry.Symbol];
592   case wasm::R_WASM_FUNCTION_OFFSET_I32:
593   case wasm::R_WASM_FUNCTION_OFFSET_I64:
594   case wasm::R_WASM_SECTION_OFFSET_I32: {
595     const auto &Section =
596         static_cast<const MCSectionWasm &>(RelEntry.Symbol->getSection());
597     return Section.getSectionOffset() + RelEntry.Addend;
598   }
599   case wasm::R_WASM_MEMORY_ADDR_LEB:
600   case wasm::R_WASM_MEMORY_ADDR_LEB64:
601   case wasm::R_WASM_MEMORY_ADDR_SLEB:
602   case wasm::R_WASM_MEMORY_ADDR_SLEB64:
603   case wasm::R_WASM_MEMORY_ADDR_REL_SLEB:
604   case wasm::R_WASM_MEMORY_ADDR_REL_SLEB64:
605   case wasm::R_WASM_MEMORY_ADDR_I32:
606   case wasm::R_WASM_MEMORY_ADDR_I64:
607   case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB: {
608     // Provisional value is address of the global plus the offset
609     const MCSymbolWasm *Base =
610         cast<MCSymbolWasm>(Layout.getBaseSymbol(*RelEntry.Symbol));
611     // For undefined symbols, use zero
612     if (!Base->isDefined())
613       return 0;
614     const wasm::WasmDataReference &BaseRef = DataLocations[Base],
615                                   &SymRef = DataLocations[RelEntry.Symbol];
616     const WasmDataSegment &Segment = DataSegments[BaseRef.Segment];
617     // Ignore overflow. LLVM allows address arithmetic to silently wrap.
618     return Segment.Offset + BaseRef.Offset + SymRef.Offset + RelEntry.Addend;
619   }
620   default:
621     llvm_unreachable("invalid relocation type");
622   }
623 }
624 
addData(SmallVectorImpl<char> & DataBytes,MCSectionWasm & DataSection)625 static void addData(SmallVectorImpl<char> &DataBytes,
626                     MCSectionWasm &DataSection) {
627   LLVM_DEBUG(errs() << "addData: " << DataSection.getName() << "\n");
628 
629   DataBytes.resize(alignTo(DataBytes.size(), DataSection.getAlignment()));
630 
631   for (const MCFragment &Frag : DataSection) {
632     if (Frag.hasInstructions())
633       report_fatal_error("only data supported in data sections");
634 
635     if (auto *Align = dyn_cast<MCAlignFragment>(&Frag)) {
636       if (Align->getValueSize() != 1)
637         report_fatal_error("only byte values supported for alignment");
638       // If nops are requested, use zeros, as this is the data section.
639       uint8_t Value = Align->hasEmitNops() ? 0 : Align->getValue();
640       uint64_t Size =
641           std::min<uint64_t>(alignTo(DataBytes.size(), Align->getAlignment()),
642                              DataBytes.size() + Align->getMaxBytesToEmit());
643       DataBytes.resize(Size, Value);
644     } else if (auto *Fill = dyn_cast<MCFillFragment>(&Frag)) {
645       int64_t NumValues;
646       if (!Fill->getNumValues().evaluateAsAbsolute(NumValues))
647         llvm_unreachable("The fill should be an assembler constant");
648       DataBytes.insert(DataBytes.end(), Fill->getValueSize() * NumValues,
649                        Fill->getValue());
650     } else if (auto *LEB = dyn_cast<MCLEBFragment>(&Frag)) {
651       const SmallVectorImpl<char> &Contents = LEB->getContents();
652       llvm::append_range(DataBytes, Contents);
653     } else {
654       const auto &DataFrag = cast<MCDataFragment>(Frag);
655       const SmallVectorImpl<char> &Contents = DataFrag.getContents();
656       llvm::append_range(DataBytes, Contents);
657     }
658   }
659 
660   LLVM_DEBUG(dbgs() << "addData -> " << DataBytes.size() << "\n");
661 }
662 
663 uint32_t
getRelocationIndexValue(const WasmRelocationEntry & RelEntry)664 WasmObjectWriter::getRelocationIndexValue(const WasmRelocationEntry &RelEntry) {
665   if (RelEntry.Type == wasm::R_WASM_TYPE_INDEX_LEB) {
666     if (!TypeIndices.count(RelEntry.Symbol))
667       report_fatal_error("symbol not found in type index space: " +
668                          RelEntry.Symbol->getName());
669     return TypeIndices[RelEntry.Symbol];
670   }
671 
672   return RelEntry.Symbol->getIndex();
673 }
674 
675 // Apply the portions of the relocation records that we can handle ourselves
676 // directly.
applyRelocations(ArrayRef<WasmRelocationEntry> Relocations,uint64_t ContentsOffset,const MCAsmLayout & Layout)677 void WasmObjectWriter::applyRelocations(
678     ArrayRef<WasmRelocationEntry> Relocations, uint64_t ContentsOffset,
679     const MCAsmLayout &Layout) {
680   auto &Stream = static_cast<raw_pwrite_stream &>(W->OS);
681   for (const WasmRelocationEntry &RelEntry : Relocations) {
682     uint64_t Offset = ContentsOffset +
683                       RelEntry.FixupSection->getSectionOffset() +
684                       RelEntry.Offset;
685 
686     LLVM_DEBUG(dbgs() << "applyRelocation: " << RelEntry << "\n");
687     auto Value = getProvisionalValue(RelEntry, Layout);
688 
689     switch (RelEntry.Type) {
690     case wasm::R_WASM_FUNCTION_INDEX_LEB:
691     case wasm::R_WASM_TYPE_INDEX_LEB:
692     case wasm::R_WASM_GLOBAL_INDEX_LEB:
693     case wasm::R_WASM_MEMORY_ADDR_LEB:
694     case wasm::R_WASM_EVENT_INDEX_LEB:
695     case wasm::R_WASM_TABLE_NUMBER_LEB:
696       writePatchableLEB<5>(Stream, Value, Offset);
697       break;
698     case wasm::R_WASM_MEMORY_ADDR_LEB64:
699       writePatchableLEB<10>(Stream, Value, Offset);
700       break;
701     case wasm::R_WASM_TABLE_INDEX_I32:
702     case wasm::R_WASM_MEMORY_ADDR_I32:
703     case wasm::R_WASM_FUNCTION_OFFSET_I32:
704     case wasm::R_WASM_SECTION_OFFSET_I32:
705     case wasm::R_WASM_GLOBAL_INDEX_I32:
706       patchI32(Stream, Value, Offset);
707       break;
708     case wasm::R_WASM_TABLE_INDEX_I64:
709     case wasm::R_WASM_MEMORY_ADDR_I64:
710     case wasm::R_WASM_FUNCTION_OFFSET_I64:
711       patchI64(Stream, Value, Offset);
712       break;
713     case wasm::R_WASM_TABLE_INDEX_SLEB:
714     case wasm::R_WASM_TABLE_INDEX_REL_SLEB:
715     case wasm::R_WASM_MEMORY_ADDR_SLEB:
716     case wasm::R_WASM_MEMORY_ADDR_REL_SLEB:
717     case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB:
718       writePatchableSLEB<5>(Stream, Value, Offset);
719       break;
720     case wasm::R_WASM_TABLE_INDEX_SLEB64:
721     case wasm::R_WASM_MEMORY_ADDR_SLEB64:
722     case wasm::R_WASM_MEMORY_ADDR_REL_SLEB64:
723       writePatchableSLEB<10>(Stream, Value, Offset);
724       break;
725     default:
726       llvm_unreachable("invalid relocation type");
727     }
728   }
729 }
730 
writeTypeSection(ArrayRef<wasm::WasmSignature> Signatures)731 void WasmObjectWriter::writeTypeSection(
732     ArrayRef<wasm::WasmSignature> Signatures) {
733   if (Signatures.empty())
734     return;
735 
736   SectionBookkeeping Section;
737   startSection(Section, wasm::WASM_SEC_TYPE);
738 
739   encodeULEB128(Signatures.size(), W->OS);
740 
741   for (const wasm::WasmSignature &Sig : Signatures) {
742     W->OS << char(wasm::WASM_TYPE_FUNC);
743     encodeULEB128(Sig.Params.size(), W->OS);
744     for (wasm::ValType Ty : Sig.Params)
745       writeValueType(Ty);
746     encodeULEB128(Sig.Returns.size(), W->OS);
747     for (wasm::ValType Ty : Sig.Returns)
748       writeValueType(Ty);
749   }
750 
751   endSection(Section);
752 }
753 
writeImportSection(ArrayRef<wasm::WasmImport> Imports,uint64_t DataSize,uint32_t NumElements)754 void WasmObjectWriter::writeImportSection(ArrayRef<wasm::WasmImport> Imports,
755                                           uint64_t DataSize,
756                                           uint32_t NumElements) {
757   if (Imports.empty())
758     return;
759 
760   uint64_t NumPages = (DataSize + wasm::WasmPageSize - 1) / wasm::WasmPageSize;
761 
762   SectionBookkeeping Section;
763   startSection(Section, wasm::WASM_SEC_IMPORT);
764 
765   encodeULEB128(Imports.size(), W->OS);
766   for (const wasm::WasmImport &Import : Imports) {
767     writeString(Import.Module);
768     writeString(Import.Field);
769     W->OS << char(Import.Kind);
770 
771     switch (Import.Kind) {
772     case wasm::WASM_EXTERNAL_FUNCTION:
773       encodeULEB128(Import.SigIndex, W->OS);
774       break;
775     case wasm::WASM_EXTERNAL_GLOBAL:
776       W->OS << char(Import.Global.Type);
777       W->OS << char(Import.Global.Mutable ? 1 : 0);
778       break;
779     case wasm::WASM_EXTERNAL_MEMORY:
780       encodeULEB128(Import.Memory.Flags, W->OS);
781       encodeULEB128(NumPages, W->OS); // initial
782       break;
783     case wasm::WASM_EXTERNAL_TABLE:
784       W->OS << char(Import.Table.ElemType);
785       encodeULEB128(0, W->OS);           // flags
786       encodeULEB128(NumElements, W->OS); // initial
787       break;
788     case wasm::WASM_EXTERNAL_EVENT:
789       encodeULEB128(Import.Event.Attribute, W->OS);
790       encodeULEB128(Import.Event.SigIndex, W->OS);
791       break;
792     default:
793       llvm_unreachable("unsupported import kind");
794     }
795   }
796 
797   endSection(Section);
798 }
799 
writeFunctionSection(ArrayRef<WasmFunction> Functions)800 void WasmObjectWriter::writeFunctionSection(ArrayRef<WasmFunction> Functions) {
801   if (Functions.empty())
802     return;
803 
804   SectionBookkeeping Section;
805   startSection(Section, wasm::WASM_SEC_FUNCTION);
806 
807   encodeULEB128(Functions.size(), W->OS);
808   for (const WasmFunction &Func : Functions)
809     encodeULEB128(Func.SigIndex, W->OS);
810 
811   endSection(Section);
812 }
813 
writeEventSection(ArrayRef<wasm::WasmEventType> Events)814 void WasmObjectWriter::writeEventSection(ArrayRef<wasm::WasmEventType> Events) {
815   if (Events.empty())
816     return;
817 
818   SectionBookkeeping Section;
819   startSection(Section, wasm::WASM_SEC_EVENT);
820 
821   encodeULEB128(Events.size(), W->OS);
822   for (const wasm::WasmEventType &Event : Events) {
823     encodeULEB128(Event.Attribute, W->OS);
824     encodeULEB128(Event.SigIndex, W->OS);
825   }
826 
827   endSection(Section);
828 }
829 
writeGlobalSection(ArrayRef<wasm::WasmGlobal> Globals)830 void WasmObjectWriter::writeGlobalSection(ArrayRef<wasm::WasmGlobal> Globals) {
831   if (Globals.empty())
832     return;
833 
834   SectionBookkeeping Section;
835   startSection(Section, wasm::WASM_SEC_GLOBAL);
836 
837   encodeULEB128(Globals.size(), W->OS);
838   for (const wasm::WasmGlobal &Global : Globals) {
839     encodeULEB128(Global.Type.Type, W->OS);
840     W->OS << char(Global.Type.Mutable);
841     W->OS << char(Global.InitExpr.Opcode);
842     switch (Global.Type.Type) {
843     case wasm::WASM_TYPE_I32:
844       encodeSLEB128(0, W->OS);
845       break;
846     case wasm::WASM_TYPE_I64:
847       encodeSLEB128(0, W->OS);
848       break;
849     case wasm::WASM_TYPE_F32:
850       writeI32(0);
851       break;
852     case wasm::WASM_TYPE_F64:
853       writeI64(0);
854       break;
855     case wasm::WASM_TYPE_EXTERNREF:
856       writeValueType(wasm::ValType::EXTERNREF);
857       break;
858     default:
859       llvm_unreachable("unexpected type");
860     }
861     W->OS << char(wasm::WASM_OPCODE_END);
862   }
863 
864   endSection(Section);
865 }
866 
writeTableSection(ArrayRef<wasm::WasmTable> Tables)867 void WasmObjectWriter::writeTableSection(ArrayRef<wasm::WasmTable> Tables) {
868   if (Tables.empty())
869     return;
870 
871   SectionBookkeeping Section;
872   startSection(Section, wasm::WASM_SEC_TABLE);
873 
874   encodeULEB128(Tables.size(), W->OS);
875   for (const wasm::WasmTable &Table : Tables) {
876     encodeULEB128(Table.Type.ElemType, W->OS);
877     encodeULEB128(Table.Type.Limits.Flags, W->OS);
878     encodeULEB128(Table.Type.Limits.Initial, W->OS);
879     if (Table.Type.Limits.Flags & wasm::WASM_LIMITS_FLAG_HAS_MAX)
880       encodeULEB128(Table.Type.Limits.Maximum, W->OS);
881   }
882   endSection(Section);
883 }
884 
writeExportSection(ArrayRef<wasm::WasmExport> Exports)885 void WasmObjectWriter::writeExportSection(ArrayRef<wasm::WasmExport> Exports) {
886   if (Exports.empty())
887     return;
888 
889   SectionBookkeeping Section;
890   startSection(Section, wasm::WASM_SEC_EXPORT);
891 
892   encodeULEB128(Exports.size(), W->OS);
893   for (const wasm::WasmExport &Export : Exports) {
894     writeString(Export.Name);
895     W->OS << char(Export.Kind);
896     encodeULEB128(Export.Index, W->OS);
897   }
898 
899   endSection(Section);
900 }
901 
writeElemSection(ArrayRef<uint32_t> TableElems)902 void WasmObjectWriter::writeElemSection(ArrayRef<uint32_t> TableElems) {
903   if (TableElems.empty())
904     return;
905 
906   SectionBookkeeping Section;
907   startSection(Section, wasm::WASM_SEC_ELEM);
908 
909   encodeULEB128(1, W->OS); // number of "segments"
910   encodeULEB128(0, W->OS); // the table index
911 
912   // init expr for starting offset
913   W->OS << char(wasm::WASM_OPCODE_I32_CONST);
914   encodeSLEB128(InitialTableOffset, W->OS);
915   W->OS << char(wasm::WASM_OPCODE_END);
916 
917   encodeULEB128(TableElems.size(), W->OS);
918   for (uint32_t Elem : TableElems)
919     encodeULEB128(Elem, W->OS);
920 
921   endSection(Section);
922 }
923 
writeDataCountSection()924 void WasmObjectWriter::writeDataCountSection() {
925   if (DataSegments.empty())
926     return;
927 
928   SectionBookkeeping Section;
929   startSection(Section, wasm::WASM_SEC_DATACOUNT);
930   encodeULEB128(DataSegments.size(), W->OS);
931   endSection(Section);
932 }
933 
writeCodeSection(const MCAssembler & Asm,const MCAsmLayout & Layout,ArrayRef<WasmFunction> Functions)934 uint32_t WasmObjectWriter::writeCodeSection(const MCAssembler &Asm,
935                                             const MCAsmLayout &Layout,
936                                             ArrayRef<WasmFunction> Functions) {
937   if (Functions.empty())
938     return 0;
939 
940   SectionBookkeeping Section;
941   startSection(Section, wasm::WASM_SEC_CODE);
942 
943   encodeULEB128(Functions.size(), W->OS);
944 
945   for (const WasmFunction &Func : Functions) {
946     auto &FuncSection = static_cast<MCSectionWasm &>(Func.Sym->getSection());
947 
948     int64_t Size = 0;
949     if (!Func.Sym->getSize()->evaluateAsAbsolute(Size, Layout))
950       report_fatal_error(".size expression must be evaluatable");
951 
952     encodeULEB128(Size, W->OS);
953     FuncSection.setSectionOffset(W->OS.tell() - Section.ContentsOffset);
954     Asm.writeSectionData(W->OS, &FuncSection, Layout);
955   }
956 
957   // Apply fixups.
958   applyRelocations(CodeRelocations, Section.ContentsOffset, Layout);
959 
960   endSection(Section);
961   return Section.Index;
962 }
963 
writeDataSection(const MCAsmLayout & Layout)964 uint32_t WasmObjectWriter::writeDataSection(const MCAsmLayout &Layout) {
965   if (DataSegments.empty())
966     return 0;
967 
968   SectionBookkeeping Section;
969   startSection(Section, wasm::WASM_SEC_DATA);
970 
971   encodeULEB128(DataSegments.size(), W->OS); // count
972 
973   for (const WasmDataSegment &Segment : DataSegments) {
974     encodeULEB128(Segment.InitFlags, W->OS); // flags
975     if (Segment.InitFlags & wasm::WASM_DATA_SEGMENT_HAS_MEMINDEX)
976       encodeULEB128(0, W->OS); // memory index
977     if ((Segment.InitFlags & wasm::WASM_DATA_SEGMENT_IS_PASSIVE) == 0) {
978       W->OS << char(Segment.Offset > INT32_MAX ? wasm::WASM_OPCODE_I64_CONST
979                                                : wasm::WASM_OPCODE_I32_CONST);
980       encodeSLEB128(Segment.Offset, W->OS); // offset
981       W->OS << char(wasm::WASM_OPCODE_END);
982     }
983     encodeULEB128(Segment.Data.size(), W->OS); // size
984     Segment.Section->setSectionOffset(W->OS.tell() - Section.ContentsOffset);
985     W->OS << Segment.Data; // data
986   }
987 
988   // Apply fixups.
989   applyRelocations(DataRelocations, Section.ContentsOffset, Layout);
990 
991   endSection(Section);
992   return Section.Index;
993 }
994 
writeRelocSection(uint32_t SectionIndex,StringRef Name,std::vector<WasmRelocationEntry> & Relocs)995 void WasmObjectWriter::writeRelocSection(
996     uint32_t SectionIndex, StringRef Name,
997     std::vector<WasmRelocationEntry> &Relocs) {
998   // See: https://github.com/WebAssembly/tool-conventions/blob/master/Linking.md
999   // for descriptions of the reloc sections.
1000 
1001   if (Relocs.empty())
1002     return;
1003 
1004   // First, ensure the relocations are sorted in offset order.  In general they
1005   // should already be sorted since `recordRelocation` is called in offset
1006   // order, but for the code section we combine many MC sections into single
1007   // wasm section, and this order is determined by the order of Asm.Symbols()
1008   // not the sections order.
1009   llvm::stable_sort(
1010       Relocs, [](const WasmRelocationEntry &A, const WasmRelocationEntry &B) {
1011         return (A.Offset + A.FixupSection->getSectionOffset()) <
1012                (B.Offset + B.FixupSection->getSectionOffset());
1013       });
1014 
1015   SectionBookkeeping Section;
1016   startCustomSection(Section, std::string("reloc.") + Name.str());
1017 
1018   encodeULEB128(SectionIndex, W->OS);
1019   encodeULEB128(Relocs.size(), W->OS);
1020   for (const WasmRelocationEntry &RelEntry : Relocs) {
1021     uint64_t Offset =
1022         RelEntry.Offset + RelEntry.FixupSection->getSectionOffset();
1023     uint32_t Index = getRelocationIndexValue(RelEntry);
1024 
1025     W->OS << char(RelEntry.Type);
1026     encodeULEB128(Offset, W->OS);
1027     encodeULEB128(Index, W->OS);
1028     if (RelEntry.hasAddend())
1029       encodeSLEB128(RelEntry.Addend, W->OS);
1030   }
1031 
1032   endSection(Section);
1033 }
1034 
writeCustomRelocSections()1035 void WasmObjectWriter::writeCustomRelocSections() {
1036   for (const auto &Sec : CustomSections) {
1037     auto &Relocations = CustomSectionsRelocations[Sec.Section];
1038     writeRelocSection(Sec.OutputIndex, Sec.Name, Relocations);
1039   }
1040 }
1041 
writeLinkingMetaDataSection(ArrayRef<wasm::WasmSymbolInfo> SymbolInfos,ArrayRef<std::pair<uint16_t,uint32_t>> InitFuncs,const std::map<StringRef,std::vector<WasmComdatEntry>> & Comdats)1042 void WasmObjectWriter::writeLinkingMetaDataSection(
1043     ArrayRef<wasm::WasmSymbolInfo> SymbolInfos,
1044     ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs,
1045     const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats) {
1046   SectionBookkeeping Section;
1047   startCustomSection(Section, "linking");
1048   encodeULEB128(wasm::WasmMetadataVersion, W->OS);
1049 
1050   SectionBookkeeping SubSection;
1051   if (SymbolInfos.size() != 0) {
1052     startSection(SubSection, wasm::WASM_SYMBOL_TABLE);
1053     encodeULEB128(SymbolInfos.size(), W->OS);
1054     for (const wasm::WasmSymbolInfo &Sym : SymbolInfos) {
1055       encodeULEB128(Sym.Kind, W->OS);
1056       encodeULEB128(Sym.Flags, W->OS);
1057       switch (Sym.Kind) {
1058       case wasm::WASM_SYMBOL_TYPE_FUNCTION:
1059       case wasm::WASM_SYMBOL_TYPE_GLOBAL:
1060       case wasm::WASM_SYMBOL_TYPE_EVENT:
1061       case wasm::WASM_SYMBOL_TYPE_TABLE:
1062         encodeULEB128(Sym.ElementIndex, W->OS);
1063         if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0 ||
1064             (Sym.Flags & wasm::WASM_SYMBOL_EXPLICIT_NAME) != 0)
1065           writeString(Sym.Name);
1066         break;
1067       case wasm::WASM_SYMBOL_TYPE_DATA:
1068         writeString(Sym.Name);
1069         if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0) {
1070           encodeULEB128(Sym.DataRef.Segment, W->OS);
1071           encodeULEB128(Sym.DataRef.Offset, W->OS);
1072           encodeULEB128(Sym.DataRef.Size, W->OS);
1073         }
1074         break;
1075       case wasm::WASM_SYMBOL_TYPE_SECTION: {
1076         const uint32_t SectionIndex =
1077             CustomSections[Sym.ElementIndex].OutputIndex;
1078         encodeULEB128(SectionIndex, W->OS);
1079         break;
1080       }
1081       default:
1082         llvm_unreachable("unexpected kind");
1083       }
1084     }
1085     endSection(SubSection);
1086   }
1087 
1088   if (DataSegments.size()) {
1089     startSection(SubSection, wasm::WASM_SEGMENT_INFO);
1090     encodeULEB128(DataSegments.size(), W->OS);
1091     for (const WasmDataSegment &Segment : DataSegments) {
1092       writeString(Segment.Name);
1093       encodeULEB128(Segment.Alignment, W->OS);
1094       encodeULEB128(Segment.LinkerFlags, W->OS);
1095     }
1096     endSection(SubSection);
1097   }
1098 
1099   if (!InitFuncs.empty()) {
1100     startSection(SubSection, wasm::WASM_INIT_FUNCS);
1101     encodeULEB128(InitFuncs.size(), W->OS);
1102     for (auto &StartFunc : InitFuncs) {
1103       encodeULEB128(StartFunc.first, W->OS);  // priority
1104       encodeULEB128(StartFunc.second, W->OS); // function index
1105     }
1106     endSection(SubSection);
1107   }
1108 
1109   if (Comdats.size()) {
1110     startSection(SubSection, wasm::WASM_COMDAT_INFO);
1111     encodeULEB128(Comdats.size(), W->OS);
1112     for (const auto &C : Comdats) {
1113       writeString(C.first);
1114       encodeULEB128(0, W->OS); // flags for future use
1115       encodeULEB128(C.second.size(), W->OS);
1116       for (const WasmComdatEntry &Entry : C.second) {
1117         encodeULEB128(Entry.Kind, W->OS);
1118         encodeULEB128(Entry.Index, W->OS);
1119       }
1120     }
1121     endSection(SubSection);
1122   }
1123 
1124   endSection(Section);
1125 }
1126 
writeCustomSection(WasmCustomSection & CustomSection,const MCAssembler & Asm,const MCAsmLayout & Layout)1127 void WasmObjectWriter::writeCustomSection(WasmCustomSection &CustomSection,
1128                                           const MCAssembler &Asm,
1129                                           const MCAsmLayout &Layout) {
1130   SectionBookkeeping Section;
1131   auto *Sec = CustomSection.Section;
1132   startCustomSection(Section, CustomSection.Name);
1133 
1134   Sec->setSectionOffset(W->OS.tell() - Section.ContentsOffset);
1135   Asm.writeSectionData(W->OS, Sec, Layout);
1136 
1137   CustomSection.OutputContentsOffset = Section.ContentsOffset;
1138   CustomSection.OutputIndex = Section.Index;
1139 
1140   endSection(Section);
1141 
1142   // Apply fixups.
1143   auto &Relocations = CustomSectionsRelocations[CustomSection.Section];
1144   applyRelocations(Relocations, CustomSection.OutputContentsOffset, Layout);
1145 }
1146 
getFunctionType(const MCSymbolWasm & Symbol)1147 uint32_t WasmObjectWriter::getFunctionType(const MCSymbolWasm &Symbol) {
1148   assert(Symbol.isFunction());
1149   assert(TypeIndices.count(&Symbol));
1150   return TypeIndices[&Symbol];
1151 }
1152 
getEventType(const MCSymbolWasm & Symbol)1153 uint32_t WasmObjectWriter::getEventType(const MCSymbolWasm &Symbol) {
1154   assert(Symbol.isEvent());
1155   assert(TypeIndices.count(&Symbol));
1156   return TypeIndices[&Symbol];
1157 }
1158 
registerFunctionType(const MCSymbolWasm & Symbol)1159 void WasmObjectWriter::registerFunctionType(const MCSymbolWasm &Symbol) {
1160   assert(Symbol.isFunction());
1161 
1162   wasm::WasmSignature S;
1163 
1164   if (auto *Sig = Symbol.getSignature()) {
1165     S.Returns = Sig->Returns;
1166     S.Params = Sig->Params;
1167   }
1168 
1169   auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size()));
1170   if (Pair.second)
1171     Signatures.push_back(S);
1172   TypeIndices[&Symbol] = Pair.first->second;
1173 
1174   LLVM_DEBUG(dbgs() << "registerFunctionType: " << Symbol
1175                     << " new:" << Pair.second << "\n");
1176   LLVM_DEBUG(dbgs() << "  -> type index: " << Pair.first->second << "\n");
1177 }
1178 
registerEventType(const MCSymbolWasm & Symbol)1179 void WasmObjectWriter::registerEventType(const MCSymbolWasm &Symbol) {
1180   assert(Symbol.isEvent());
1181 
1182   // TODO Currently we don't generate imported exceptions, but if we do, we
1183   // should have a way of infering types of imported exceptions.
1184   wasm::WasmSignature S;
1185   if (auto *Sig = Symbol.getSignature()) {
1186     S.Returns = Sig->Returns;
1187     S.Params = Sig->Params;
1188   }
1189 
1190   auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size()));
1191   if (Pair.second)
1192     Signatures.push_back(S);
1193   TypeIndices[&Symbol] = Pair.first->second;
1194 
1195   LLVM_DEBUG(dbgs() << "registerEventType: " << Symbol << " new:" << Pair.second
1196                     << "\n");
1197   LLVM_DEBUG(dbgs() << "  -> type index: " << Pair.first->second << "\n");
1198 }
1199 
isInSymtab(const MCSymbolWasm & Sym)1200 static bool isInSymtab(const MCSymbolWasm &Sym) {
1201   if (Sym.isUsedInReloc() || Sym.isUsedInInitArray())
1202     return true;
1203 
1204   if (Sym.isComdat() && !Sym.isDefined())
1205     return false;
1206 
1207   if (Sym.isTemporary())
1208     return false;
1209 
1210   if (Sym.isSection())
1211     return false;
1212 
1213   return true;
1214 }
1215 
prepareImports(SmallVectorImpl<wasm::WasmImport> & Imports,MCAssembler & Asm,const MCAsmLayout & Layout)1216 void WasmObjectWriter::prepareImports(
1217     SmallVectorImpl<wasm::WasmImport> &Imports, MCAssembler &Asm,
1218     const MCAsmLayout &Layout) {
1219   // For now, always emit the memory import, since loads and stores are not
1220   // valid without it. In the future, we could perhaps be more clever and omit
1221   // it if there are no loads or stores.
1222   wasm::WasmImport MemImport;
1223   MemImport.Module = "env";
1224   MemImport.Field = "__linear_memory";
1225   MemImport.Kind = wasm::WASM_EXTERNAL_MEMORY;
1226   MemImport.Memory.Flags = is64Bit() ? wasm::WASM_LIMITS_FLAG_IS_64
1227                                      : wasm::WASM_LIMITS_FLAG_NONE;
1228   Imports.push_back(MemImport);
1229 
1230   // Populate SignatureIndices, and Imports and WasmIndices for undefined
1231   // symbols.  This must be done before populating WasmIndices for defined
1232   // symbols.
1233   for (const MCSymbol &S : Asm.symbols()) {
1234     const auto &WS = static_cast<const MCSymbolWasm &>(S);
1235 
1236     // Register types for all functions, including those with private linkage
1237     // (because wasm always needs a type signature).
1238     if (WS.isFunction()) {
1239       const auto *BS = Layout.getBaseSymbol(S);
1240       if (!BS)
1241         report_fatal_error(Twine(S.getName()) +
1242                            ": absolute addressing not supported!");
1243       registerFunctionType(*cast<MCSymbolWasm>(BS));
1244     }
1245 
1246     if (WS.isEvent())
1247       registerEventType(WS);
1248 
1249     if (WS.isTemporary())
1250       continue;
1251 
1252     // If the symbol is not defined in this translation unit, import it.
1253     if (!WS.isDefined() && !WS.isComdat()) {
1254       if (WS.isFunction()) {
1255         wasm::WasmImport Import;
1256         Import.Module = WS.getImportModule();
1257         Import.Field = WS.getImportName();
1258         Import.Kind = wasm::WASM_EXTERNAL_FUNCTION;
1259         Import.SigIndex = getFunctionType(WS);
1260         Imports.push_back(Import);
1261         assert(WasmIndices.count(&WS) == 0);
1262         WasmIndices[&WS] = NumFunctionImports++;
1263       } else if (WS.isGlobal()) {
1264         if (WS.isWeak())
1265           report_fatal_error("undefined global symbol cannot be weak");
1266 
1267         wasm::WasmImport Import;
1268         Import.Field = WS.getImportName();
1269         Import.Kind = wasm::WASM_EXTERNAL_GLOBAL;
1270         Import.Module = WS.getImportModule();
1271         Import.Global = WS.getGlobalType();
1272         Imports.push_back(Import);
1273         assert(WasmIndices.count(&WS) == 0);
1274         WasmIndices[&WS] = NumGlobalImports++;
1275       } else if (WS.isEvent()) {
1276         if (WS.isWeak())
1277           report_fatal_error("undefined event symbol cannot be weak");
1278 
1279         wasm::WasmImport Import;
1280         Import.Module = WS.getImportModule();
1281         Import.Field = WS.getImportName();
1282         Import.Kind = wasm::WASM_EXTERNAL_EVENT;
1283         Import.Event.Attribute = wasm::WASM_EVENT_ATTRIBUTE_EXCEPTION;
1284         Import.Event.SigIndex = getEventType(WS);
1285         Imports.push_back(Import);
1286         assert(WasmIndices.count(&WS) == 0);
1287         WasmIndices[&WS] = NumEventImports++;
1288       } else if (WS.isTable()) {
1289         if (WS.isWeak())
1290           report_fatal_error("undefined table symbol cannot be weak");
1291 
1292         wasm::WasmImport Import;
1293         Import.Module = WS.getImportModule();
1294         Import.Field = WS.getImportName();
1295         Import.Kind = wasm::WASM_EXTERNAL_TABLE;
1296         wasm::ValType ElemType = WS.getTableType();
1297         Import.Table.ElemType = uint8_t(ElemType);
1298         // FIXME: Extend table type to include limits? For now we don't specify
1299         // a min or max which does not place any restrictions on the size of the
1300         // imported table.
1301         Import.Table.Limits = {wasm::WASM_LIMITS_FLAG_NONE, 0, 0};
1302         Imports.push_back(Import);
1303         assert(WasmIndices.count(&WS) == 0);
1304         WasmIndices[&WS] = NumTableImports++;
1305       }
1306     }
1307   }
1308 
1309   // Add imports for GOT globals
1310   for (const MCSymbol &S : Asm.symbols()) {
1311     const auto &WS = static_cast<const MCSymbolWasm &>(S);
1312     if (WS.isUsedInGOT()) {
1313       wasm::WasmImport Import;
1314       if (WS.isFunction())
1315         Import.Module = "GOT.func";
1316       else
1317         Import.Module = "GOT.mem";
1318       Import.Field = WS.getName();
1319       Import.Kind = wasm::WASM_EXTERNAL_GLOBAL;
1320       Import.Global = {wasm::WASM_TYPE_I32, true};
1321       Imports.push_back(Import);
1322       assert(GOTIndices.count(&WS) == 0);
1323       GOTIndices[&WS] = NumGlobalImports++;
1324     }
1325   }
1326 }
1327 
writeObject(MCAssembler & Asm,const MCAsmLayout & Layout)1328 uint64_t WasmObjectWriter::writeObject(MCAssembler &Asm,
1329                                        const MCAsmLayout &Layout) {
1330   support::endian::Writer MainWriter(*OS, support::little);
1331   W = &MainWriter;
1332   if (IsSplitDwarf) {
1333     uint64_t TotalSize = writeOneObject(Asm, Layout, DwoMode::NonDwoOnly);
1334     assert(DwoOS);
1335     support::endian::Writer DwoWriter(*DwoOS, support::little);
1336     W = &DwoWriter;
1337     return TotalSize + writeOneObject(Asm, Layout, DwoMode::DwoOnly);
1338   } else {
1339     return writeOneObject(Asm, Layout, DwoMode::AllSections);
1340   }
1341 }
1342 
writeOneObject(MCAssembler & Asm,const MCAsmLayout & Layout,DwoMode Mode)1343 uint64_t WasmObjectWriter::writeOneObject(MCAssembler &Asm,
1344                                           const MCAsmLayout &Layout,
1345                                           DwoMode Mode) {
1346   uint64_t StartOffset = W->OS.tell();
1347   SectionCount = 0;
1348   CustomSections.clear();
1349 
1350   LLVM_DEBUG(dbgs() << "WasmObjectWriter::writeObject\n");
1351 
1352   // Collect information from the available symbols.
1353   SmallVector<WasmFunction, 4> Functions;
1354   SmallVector<uint32_t, 4> TableElems;
1355   SmallVector<wasm::WasmImport, 4> Imports;
1356   SmallVector<wasm::WasmExport, 4> Exports;
1357   SmallVector<wasm::WasmEventType, 1> Events;
1358   SmallVector<wasm::WasmGlobal, 1> Globals;
1359   SmallVector<wasm::WasmTable, 1> Tables;
1360   SmallVector<wasm::WasmSymbolInfo, 4> SymbolInfos;
1361   SmallVector<std::pair<uint16_t, uint32_t>, 2> InitFuncs;
1362   std::map<StringRef, std::vector<WasmComdatEntry>> Comdats;
1363   uint64_t DataSize = 0;
1364   if (Mode != DwoMode::DwoOnly) {
1365     prepareImports(Imports, Asm, Layout);
1366   }
1367 
1368   // Populate DataSegments and CustomSections, which must be done before
1369   // populating DataLocations.
1370   for (MCSection &Sec : Asm) {
1371     auto &Section = static_cast<MCSectionWasm &>(Sec);
1372     StringRef SectionName = Section.getName();
1373 
1374     if (Mode == DwoMode::NonDwoOnly && isDwoSection(Sec))
1375       continue;
1376     if (Mode == DwoMode::DwoOnly && !isDwoSection(Sec))
1377       continue;
1378 
1379     LLVM_DEBUG(dbgs() << "Processing Section " << SectionName << "  group "
1380                       << Section.getGroup() << "\n";);
1381 
1382     // .init_array sections are handled specially elsewhere.
1383     if (SectionName.startswith(".init_array"))
1384       continue;
1385 
1386     // Code is handled separately
1387     if (Section.getKind().isText())
1388       continue;
1389 
1390     if (Section.isWasmData()) {
1391       uint32_t SegmentIndex = DataSegments.size();
1392       DataSize = alignTo(DataSize, Section.getAlignment());
1393       DataSegments.emplace_back();
1394       WasmDataSegment &Segment = DataSegments.back();
1395       Segment.Name = SectionName;
1396       Segment.InitFlags = Section.getPassive()
1397                               ? (uint32_t)wasm::WASM_DATA_SEGMENT_IS_PASSIVE
1398                               : 0;
1399       Segment.Offset = DataSize;
1400       Segment.Section = &Section;
1401       addData(Segment.Data, Section);
1402       Segment.Alignment = Log2_32(Section.getAlignment());
1403       Segment.LinkerFlags = 0;
1404       DataSize += Segment.Data.size();
1405       Section.setSegmentIndex(SegmentIndex);
1406 
1407       if (const MCSymbolWasm *C = Section.getGroup()) {
1408         Comdats[C->getName()].emplace_back(
1409             WasmComdatEntry{wasm::WASM_COMDAT_DATA, SegmentIndex});
1410       }
1411     } else {
1412       // Create custom sections
1413       assert(Sec.getKind().isMetadata());
1414 
1415       StringRef Name = SectionName;
1416 
1417       // For user-defined custom sections, strip the prefix
1418       if (Name.startswith(".custom_section."))
1419         Name = Name.substr(strlen(".custom_section."));
1420 
1421       MCSymbol *Begin = Sec.getBeginSymbol();
1422       if (Begin) {
1423         assert(WasmIndices.count(cast<MCSymbolWasm>(Begin)) == 0);
1424         WasmIndices[cast<MCSymbolWasm>(Begin)] = CustomSections.size();
1425       }
1426 
1427       // Separate out the producers and target features sections
1428       if (Name == "producers") {
1429         ProducersSection = std::make_unique<WasmCustomSection>(Name, &Section);
1430         continue;
1431       }
1432       if (Name == "target_features") {
1433         TargetFeaturesSection =
1434             std::make_unique<WasmCustomSection>(Name, &Section);
1435         continue;
1436       }
1437 
1438       // Custom sections can also belong to COMDAT groups. In this case the
1439       // decriptor's "index" field is the section index (in the final object
1440       // file), but that is not known until after layout, so it must be fixed up
1441       // later
1442       if (const MCSymbolWasm *C = Section.getGroup()) {
1443         Comdats[C->getName()].emplace_back(
1444             WasmComdatEntry{wasm::WASM_COMDAT_SECTION,
1445                             static_cast<uint32_t>(CustomSections.size())});
1446       }
1447 
1448       CustomSections.emplace_back(Name, &Section);
1449     }
1450   }
1451 
1452   if (Mode != DwoMode::DwoOnly) {
1453     // Populate WasmIndices and DataLocations for defined symbols.
1454     for (const MCSymbol &S : Asm.symbols()) {
1455       // Ignore unnamed temporary symbols, which aren't ever exported, imported,
1456       // or used in relocations.
1457       if (S.isTemporary() && S.getName().empty())
1458         continue;
1459 
1460       const auto &WS = static_cast<const MCSymbolWasm &>(S);
1461       LLVM_DEBUG(
1462           dbgs() << "MCSymbol: " << toString(WS.getType()) << " '" << S << "'"
1463                  << " isDefined=" << S.isDefined() << " isExternal="
1464                  << S.isExternal() << " isTemporary=" << S.isTemporary()
1465                  << " isWeak=" << WS.isWeak() << " isHidden=" << WS.isHidden()
1466                  << " isVariable=" << WS.isVariable() << "\n");
1467 
1468       if (WS.isVariable())
1469         continue;
1470       if (WS.isComdat() && !WS.isDefined())
1471         continue;
1472 
1473       if (WS.isFunction()) {
1474         unsigned Index;
1475         if (WS.isDefined()) {
1476           if (WS.getOffset() != 0)
1477             report_fatal_error(
1478                 "function sections must contain one function each");
1479 
1480           if (WS.getSize() == nullptr)
1481             report_fatal_error(
1482                 "function symbols must have a size set with .size");
1483 
1484           // A definition. Write out the function body.
1485           Index = NumFunctionImports + Functions.size();
1486           WasmFunction Func;
1487           Func.SigIndex = getFunctionType(WS);
1488           Func.Sym = &WS;
1489           assert(WasmIndices.count(&WS) == 0);
1490           WasmIndices[&WS] = Index;
1491           Functions.push_back(Func);
1492 
1493           auto &Section = static_cast<MCSectionWasm &>(WS.getSection());
1494           if (const MCSymbolWasm *C = Section.getGroup()) {
1495             Comdats[C->getName()].emplace_back(
1496                 WasmComdatEntry{wasm::WASM_COMDAT_FUNCTION, Index});
1497           }
1498 
1499           if (WS.hasExportName()) {
1500             wasm::WasmExport Export;
1501             Export.Name = WS.getExportName();
1502             Export.Kind = wasm::WASM_EXTERNAL_FUNCTION;
1503             Export.Index = Index;
1504             Exports.push_back(Export);
1505           }
1506         } else {
1507           // An import; the index was assigned above.
1508           Index = WasmIndices.find(&WS)->second;
1509         }
1510 
1511         LLVM_DEBUG(dbgs() << "  -> function index: " << Index << "\n");
1512 
1513       } else if (WS.isData()) {
1514         if (!isInSymtab(WS))
1515           continue;
1516 
1517         if (!WS.isDefined()) {
1518           LLVM_DEBUG(dbgs() << "  -> segment index: -1"
1519                             << "\n");
1520           continue;
1521         }
1522 
1523         if (!WS.getSize())
1524           report_fatal_error("data symbols must have a size set with .size: " +
1525                              WS.getName());
1526 
1527         int64_t Size = 0;
1528         if (!WS.getSize()->evaluateAsAbsolute(Size, Layout))
1529           report_fatal_error(".size expression must be evaluatable");
1530 
1531         auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection());
1532         if (!DataSection.isWasmData())
1533           report_fatal_error("data symbols must live in a data section: " +
1534                              WS.getName());
1535 
1536         // For each data symbol, export it in the symtab as a reference to the
1537         // corresponding Wasm data segment.
1538         wasm::WasmDataReference Ref = wasm::WasmDataReference{
1539             DataSection.getSegmentIndex(), Layout.getSymbolOffset(WS),
1540             static_cast<uint64_t>(Size)};
1541         assert(DataLocations.count(&WS) == 0);
1542         DataLocations[&WS] = Ref;
1543         LLVM_DEBUG(dbgs() << "  -> segment index: " << Ref.Segment << "\n");
1544 
1545       } else if (WS.isGlobal()) {
1546         // A "true" Wasm global (currently just __stack_pointer)
1547         if (WS.isDefined()) {
1548           wasm::WasmGlobal Global;
1549           Global.Type = WS.getGlobalType();
1550           Global.Index = NumGlobalImports + Globals.size();
1551           switch (Global.Type.Type) {
1552           case wasm::WASM_TYPE_I32:
1553             Global.InitExpr.Opcode = wasm::WASM_OPCODE_I32_CONST;
1554             break;
1555           case wasm::WASM_TYPE_I64:
1556             Global.InitExpr.Opcode = wasm::WASM_OPCODE_I64_CONST;
1557             break;
1558           case wasm::WASM_TYPE_F32:
1559             Global.InitExpr.Opcode = wasm::WASM_OPCODE_F32_CONST;
1560             break;
1561           case wasm::WASM_TYPE_F64:
1562             Global.InitExpr.Opcode = wasm::WASM_OPCODE_F64_CONST;
1563             break;
1564           case wasm::WASM_TYPE_EXTERNREF:
1565             Global.InitExpr.Opcode = wasm::WASM_OPCODE_REF_NULL;
1566             break;
1567           default:
1568             llvm_unreachable("unexpected type");
1569           }
1570           assert(WasmIndices.count(&WS) == 0);
1571           WasmIndices[&WS] = Global.Index;
1572           Globals.push_back(Global);
1573         } else {
1574           // An import; the index was assigned above
1575           LLVM_DEBUG(dbgs() << "  -> global index: "
1576                             << WasmIndices.find(&WS)->second << "\n");
1577         }
1578       } else if (WS.isTable()) {
1579         if (WS.isDefined()) {
1580           wasm::WasmTable Table;
1581           Table.Index = NumTableImports + Tables.size();
1582           Table.Type.ElemType = static_cast<uint8_t>(WS.getTableType());
1583           // FIXME: Work on custom limits is ongoing
1584           Table.Type.Limits = {wasm::WASM_LIMITS_FLAG_NONE, 0, 0};
1585           assert(WasmIndices.count(&WS) == 0);
1586           WasmIndices[&WS] = Table.Index;
1587           Tables.push_back(Table);
1588         }
1589         LLVM_DEBUG(dbgs() << " -> table index: "
1590                           << WasmIndices.find(&WS)->second << "\n");
1591       } else if (WS.isEvent()) {
1592         // C++ exception symbol (__cpp_exception)
1593         unsigned Index;
1594         if (WS.isDefined()) {
1595           Index = NumEventImports + Events.size();
1596           wasm::WasmEventType Event;
1597           Event.SigIndex = getEventType(WS);
1598           Event.Attribute = wasm::WASM_EVENT_ATTRIBUTE_EXCEPTION;
1599           assert(WasmIndices.count(&WS) == 0);
1600           WasmIndices[&WS] = Index;
1601           Events.push_back(Event);
1602         } else {
1603           // An import; the index was assigned above.
1604           assert(WasmIndices.count(&WS) > 0);
1605         }
1606         LLVM_DEBUG(dbgs() << "  -> event index: "
1607                           << WasmIndices.find(&WS)->second << "\n");
1608 
1609       } else {
1610         assert(WS.isSection());
1611       }
1612     }
1613 
1614     // Populate WasmIndices and DataLocations for aliased symbols.  We need to
1615     // process these in a separate pass because we need to have processed the
1616     // target of the alias before the alias itself and the symbols are not
1617     // necessarily ordered in this way.
1618     for (const MCSymbol &S : Asm.symbols()) {
1619       if (!S.isVariable())
1620         continue;
1621 
1622       assert(S.isDefined());
1623 
1624       const auto *BS = Layout.getBaseSymbol(S);
1625       if (!BS)
1626         report_fatal_error(Twine(S.getName()) +
1627                            ": absolute addressing not supported!");
1628       const MCSymbolWasm *Base = cast<MCSymbolWasm>(BS);
1629 
1630       // Find the target symbol of this weak alias and export that index
1631       const auto &WS = static_cast<const MCSymbolWasm &>(S);
1632       LLVM_DEBUG(dbgs() << WS.getName() << ": weak alias of '" << *Base
1633                         << "'\n");
1634 
1635       if (Base->isFunction()) {
1636         assert(WasmIndices.count(Base) > 0);
1637         uint32_t WasmIndex = WasmIndices.find(Base)->second;
1638         assert(WasmIndices.count(&WS) == 0);
1639         WasmIndices[&WS] = WasmIndex;
1640         LLVM_DEBUG(dbgs() << "  -> index:" << WasmIndex << "\n");
1641       } else if (Base->isData()) {
1642         auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection());
1643         uint64_t Offset = Layout.getSymbolOffset(S);
1644         int64_t Size = 0;
1645         // For data symbol alias we use the size of the base symbol as the
1646         // size of the alias.  When an offset from the base is involved this
1647         // can result in a offset + size goes past the end of the data section
1648         // which out object format doesn't support.  So we must clamp it.
1649         if (!Base->getSize()->evaluateAsAbsolute(Size, Layout))
1650           report_fatal_error(".size expression must be evaluatable");
1651         const WasmDataSegment &Segment =
1652             DataSegments[DataSection.getSegmentIndex()];
1653         Size =
1654             std::min(static_cast<uint64_t>(Size), Segment.Data.size() - Offset);
1655         wasm::WasmDataReference Ref = wasm::WasmDataReference{
1656             DataSection.getSegmentIndex(),
1657             static_cast<uint32_t>(Layout.getSymbolOffset(S)),
1658             static_cast<uint32_t>(Size)};
1659         DataLocations[&WS] = Ref;
1660         LLVM_DEBUG(dbgs() << "  -> index:" << Ref.Segment << "\n");
1661       } else {
1662         report_fatal_error("don't yet support global/event aliases");
1663       }
1664     }
1665   }
1666 
1667   // Finally, populate the symbol table itself, in its "natural" order.
1668   for (const MCSymbol &S : Asm.symbols()) {
1669     const auto &WS = static_cast<const MCSymbolWasm &>(S);
1670     if (!isInSymtab(WS)) {
1671       WS.setIndex(InvalidIndex);
1672       continue;
1673     }
1674     if (WS.isTable() && WS.getName() == "__indirect_function_table") {
1675       // For the moment, don't emit table symbols -- wasm-ld can't handle them.
1676       continue;
1677     }
1678     LLVM_DEBUG(dbgs() << "adding to symtab: " << WS << "\n");
1679 
1680     uint32_t Flags = 0;
1681     if (WS.isWeak())
1682       Flags |= wasm::WASM_SYMBOL_BINDING_WEAK;
1683     if (WS.isHidden())
1684       Flags |= wasm::WASM_SYMBOL_VISIBILITY_HIDDEN;
1685     if (!WS.isExternal() && WS.isDefined())
1686       Flags |= wasm::WASM_SYMBOL_BINDING_LOCAL;
1687     if (WS.isUndefined())
1688       Flags |= wasm::WASM_SYMBOL_UNDEFINED;
1689     if (WS.isNoStrip()) {
1690       Flags |= wasm::WASM_SYMBOL_NO_STRIP;
1691       if (isEmscripten()) {
1692         Flags |= wasm::WASM_SYMBOL_EXPORTED;
1693       }
1694     }
1695     if (WS.hasImportName())
1696       Flags |= wasm::WASM_SYMBOL_EXPLICIT_NAME;
1697     if (WS.hasExportName())
1698       Flags |= wasm::WASM_SYMBOL_EXPORTED;
1699 
1700     wasm::WasmSymbolInfo Info;
1701     Info.Name = WS.getName();
1702     Info.Kind = WS.getType();
1703     Info.Flags = Flags;
1704     if (!WS.isData()) {
1705       assert(WasmIndices.count(&WS) > 0);
1706       Info.ElementIndex = WasmIndices.find(&WS)->second;
1707     } else if (WS.isDefined()) {
1708       assert(DataLocations.count(&WS) > 0);
1709       Info.DataRef = DataLocations.find(&WS)->second;
1710     }
1711     WS.setIndex(SymbolInfos.size());
1712     SymbolInfos.emplace_back(Info);
1713   }
1714 
1715   {
1716     auto HandleReloc = [&](const WasmRelocationEntry &Rel) {
1717       // Functions referenced by a relocation need to put in the table.  This is
1718       // purely to make the object file's provisional values readable, and is
1719       // ignored by the linker, which re-calculates the relocations itself.
1720       if (Rel.Type != wasm::R_WASM_TABLE_INDEX_I32 &&
1721           Rel.Type != wasm::R_WASM_TABLE_INDEX_I64 &&
1722           Rel.Type != wasm::R_WASM_TABLE_INDEX_SLEB &&
1723           Rel.Type != wasm::R_WASM_TABLE_INDEX_SLEB64 &&
1724           Rel.Type != wasm::R_WASM_TABLE_INDEX_REL_SLEB)
1725         return;
1726       assert(Rel.Symbol->isFunction());
1727       const MCSymbolWasm *Base =
1728           cast<MCSymbolWasm>(Layout.getBaseSymbol(*Rel.Symbol));
1729       uint32_t FunctionIndex = WasmIndices.find(Base)->second;
1730       uint32_t TableIndex = TableElems.size() + InitialTableOffset;
1731       if (TableIndices.try_emplace(Base, TableIndex).second) {
1732         LLVM_DEBUG(dbgs() << "  -> adding " << Base->getName()
1733                           << " to table: " << TableIndex << "\n");
1734         TableElems.push_back(FunctionIndex);
1735         registerFunctionType(*Base);
1736       }
1737     };
1738 
1739     for (const WasmRelocationEntry &RelEntry : CodeRelocations)
1740       HandleReloc(RelEntry);
1741     for (const WasmRelocationEntry &RelEntry : DataRelocations)
1742       HandleReloc(RelEntry);
1743   }
1744 
1745   // Translate .init_array section contents into start functions.
1746   for (const MCSection &S : Asm) {
1747     const auto &WS = static_cast<const MCSectionWasm &>(S);
1748     if (WS.getName().startswith(".fini_array"))
1749       report_fatal_error(".fini_array sections are unsupported");
1750     if (!WS.getName().startswith(".init_array"))
1751       continue;
1752     if (WS.getFragmentList().empty())
1753       continue;
1754 
1755     // init_array is expected to contain a single non-empty data fragment
1756     if (WS.getFragmentList().size() != 3)
1757       report_fatal_error("only one .init_array section fragment supported");
1758 
1759     auto IT = WS.begin();
1760     const MCFragment &EmptyFrag = *IT;
1761     if (EmptyFrag.getKind() != MCFragment::FT_Data)
1762       report_fatal_error(".init_array section should be aligned");
1763 
1764     IT = std::next(IT);
1765     const MCFragment &AlignFrag = *IT;
1766     if (AlignFrag.getKind() != MCFragment::FT_Align)
1767       report_fatal_error(".init_array section should be aligned");
1768     if (cast<MCAlignFragment>(AlignFrag).getAlignment() != (is64Bit() ? 8 : 4))
1769       report_fatal_error(".init_array section should be aligned for pointers");
1770 
1771     const MCFragment &Frag = *std::next(IT);
1772     if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data)
1773       report_fatal_error("only data supported in .init_array section");
1774 
1775     uint16_t Priority = UINT16_MAX;
1776     unsigned PrefixLength = strlen(".init_array");
1777     if (WS.getName().size() > PrefixLength) {
1778       if (WS.getName()[PrefixLength] != '.')
1779         report_fatal_error(
1780             ".init_array section priority should start with '.'");
1781       if (WS.getName().substr(PrefixLength + 1).getAsInteger(10, Priority))
1782         report_fatal_error("invalid .init_array section priority");
1783     }
1784     const auto &DataFrag = cast<MCDataFragment>(Frag);
1785     const SmallVectorImpl<char> &Contents = DataFrag.getContents();
1786     for (const uint8_t *
1787              P = (const uint8_t *)Contents.data(),
1788             *End = (const uint8_t *)Contents.data() + Contents.size();
1789          P != End; ++P) {
1790       if (*P != 0)
1791         report_fatal_error("non-symbolic data in .init_array section");
1792     }
1793     for (const MCFixup &Fixup : DataFrag.getFixups()) {
1794       assert(Fixup.getKind() ==
1795              MCFixup::getKindForSize(is64Bit() ? 8 : 4, false));
1796       const MCExpr *Expr = Fixup.getValue();
1797       auto *SymRef = dyn_cast<MCSymbolRefExpr>(Expr);
1798       if (!SymRef)
1799         report_fatal_error("fixups in .init_array should be symbol references");
1800       const auto &TargetSym = cast<const MCSymbolWasm>(SymRef->getSymbol());
1801       if (TargetSym.getIndex() == InvalidIndex)
1802         report_fatal_error("symbols in .init_array should exist in symtab");
1803       if (!TargetSym.isFunction())
1804         report_fatal_error("symbols in .init_array should be for functions");
1805       InitFuncs.push_back(
1806           std::make_pair(Priority, TargetSym.getIndex()));
1807     }
1808   }
1809 
1810   // Write out the Wasm header.
1811   writeHeader(Asm);
1812 
1813   uint32_t CodeSectionIndex, DataSectionIndex;
1814   if (Mode != DwoMode::DwoOnly) {
1815     writeTypeSection(Signatures);
1816     writeImportSection(Imports, DataSize, TableElems.size());
1817     writeFunctionSection(Functions);
1818     writeTableSection(Tables);
1819     // Skip the "memory" section; we import the memory instead.
1820     writeEventSection(Events);
1821     writeGlobalSection(Globals);
1822     writeExportSection(Exports);
1823     writeElemSection(TableElems);
1824     writeDataCountSection();
1825 
1826     CodeSectionIndex = writeCodeSection(Asm, Layout, Functions);
1827     DataSectionIndex = writeDataSection(Layout);
1828   }
1829 
1830   // The Sections in the COMDAT list have placeholder indices (their index among
1831   // custom sections, rather than among all sections). Fix them up here.
1832   for (auto &Group : Comdats) {
1833     for (auto &Entry : Group.second) {
1834       if (Entry.Kind == wasm::WASM_COMDAT_SECTION) {
1835         Entry.Index += SectionCount;
1836       }
1837     }
1838   }
1839   for (auto &CustomSection : CustomSections)
1840     writeCustomSection(CustomSection, Asm, Layout);
1841 
1842   if (Mode != DwoMode::DwoOnly) {
1843     writeLinkingMetaDataSection(SymbolInfos, InitFuncs, Comdats);
1844 
1845     writeRelocSection(CodeSectionIndex, "CODE", CodeRelocations);
1846     writeRelocSection(DataSectionIndex, "DATA", DataRelocations);
1847   }
1848   writeCustomRelocSections();
1849   if (ProducersSection)
1850     writeCustomSection(*ProducersSection, Asm, Layout);
1851   if (TargetFeaturesSection)
1852     writeCustomSection(*TargetFeaturesSection, Asm, Layout);
1853 
1854   // TODO: Translate the .comment section to the output.
1855   return W->OS.tell() - StartOffset;
1856 }
1857 
1858 std::unique_ptr<MCObjectWriter>
createWasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW,raw_pwrite_stream & OS)1859 llvm::createWasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW,
1860                              raw_pwrite_stream &OS) {
1861   return std::make_unique<WasmObjectWriter>(std::move(MOTW), OS);
1862 }
1863 
1864 std::unique_ptr<MCObjectWriter>
createWasmDwoObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW,raw_pwrite_stream & OS,raw_pwrite_stream & DwoOS)1865 llvm::createWasmDwoObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW,
1866                                 raw_pwrite_stream &OS,
1867                                 raw_pwrite_stream &DwoOS) {
1868   return std::make_unique<WasmObjectWriter>(std::move(MOTW), OS, DwoOS);
1869 }
1870