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 LinkingFlags;
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__anoned0fb1560111::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__anoned0fb1560111::WasmRelocationEntry108   bool hasAddend() const { return wasm::relocTypeHasAddend(Type); }
109 
print__anoned0fb1560111::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__anoned0fb1560111::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__anoned0fb1560111::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/tag/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 NumTagImports = 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(const MCSymbolWasm *IndirectFunctionTable,
315                         ArrayRef<uint32_t> TableElems);
316   void writeDataCountSection();
317   uint32_t writeCodeSection(const MCAssembler &Asm, const MCAsmLayout &Layout,
318                             ArrayRef<WasmFunction> Functions);
319   uint32_t writeDataSection(const MCAsmLayout &Layout);
320   void writeTagSection(ArrayRef<wasm::WasmTagType> Tags);
321   void writeGlobalSection(ArrayRef<wasm::WasmGlobal> Globals);
322   void writeTableSection(ArrayRef<wasm::WasmTable> Tables);
323   void writeRelocSection(uint32_t SectionIndex, StringRef Name,
324                          std::vector<WasmRelocationEntry> &Relocations);
325   void writeLinkingMetaDataSection(
326       ArrayRef<wasm::WasmSymbolInfo> SymbolInfos,
327       ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs,
328       const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats);
329   void writeCustomSection(WasmCustomSection &CustomSection,
330                           const MCAssembler &Asm, const MCAsmLayout &Layout);
331   void writeCustomRelocSections();
332 
333   uint64_t getProvisionalValue(const WasmRelocationEntry &RelEntry,
334                                const MCAsmLayout &Layout);
335   void applyRelocations(ArrayRef<WasmRelocationEntry> Relocations,
336                         uint64_t ContentsOffset, const MCAsmLayout &Layout);
337 
338   uint32_t getRelocationIndexValue(const WasmRelocationEntry &RelEntry);
339   uint32_t getFunctionType(const MCSymbolWasm &Symbol);
340   uint32_t getTagType(const MCSymbolWasm &Symbol);
341   void registerFunctionType(const MCSymbolWasm &Symbol);
342   void registerTagType(const MCSymbolWasm &Symbol);
343 };
344 
345 } // end anonymous namespace
346 
347 // Write out a section header and a patchable section size field.
startSection(SectionBookkeeping & Section,unsigned SectionId)348 void WasmObjectWriter::startSection(SectionBookkeeping &Section,
349                                     unsigned SectionId) {
350   LLVM_DEBUG(dbgs() << "startSection " << SectionId << "\n");
351   W->OS << char(SectionId);
352 
353   Section.SizeOffset = W->OS.tell();
354 
355   // The section size. We don't know the size yet, so reserve enough space
356   // for any 32-bit value; we'll patch it later.
357   encodeULEB128(0, W->OS, 5);
358 
359   // The position where the section starts, for measuring its size.
360   Section.ContentsOffset = W->OS.tell();
361   Section.PayloadOffset = W->OS.tell();
362   Section.Index = SectionCount++;
363 }
364 
startCustomSection(SectionBookkeeping & Section,StringRef Name)365 void WasmObjectWriter::startCustomSection(SectionBookkeeping &Section,
366                                           StringRef Name) {
367   LLVM_DEBUG(dbgs() << "startCustomSection " << Name << "\n");
368   startSection(Section, wasm::WASM_SEC_CUSTOM);
369 
370   // The position where the section header ends, for measuring its size.
371   Section.PayloadOffset = W->OS.tell();
372 
373   // Custom sections in wasm also have a string identifier.
374   writeString(Name);
375 
376   // The position where the custom section starts.
377   Section.ContentsOffset = W->OS.tell();
378 }
379 
380 // Now that the section is complete and we know how big it is, patch up the
381 // section size field at the start of the section.
endSection(SectionBookkeeping & Section)382 void WasmObjectWriter::endSection(SectionBookkeeping &Section) {
383   uint64_t Size = W->OS.tell();
384   // /dev/null doesn't support seek/tell and can report offset of 0.
385   // Simply skip this patching in that case.
386   if (!Size)
387     return;
388 
389   Size -= Section.PayloadOffset;
390   if (uint32_t(Size) != Size)
391     report_fatal_error("section size does not fit in a uint32_t");
392 
393   LLVM_DEBUG(dbgs() << "endSection size=" << Size << "\n");
394 
395   // Write the final section size to the payload_len field, which follows
396   // the section id byte.
397   writePatchableLEB<5>(static_cast<raw_pwrite_stream &>(W->OS), Size,
398                        Section.SizeOffset);
399 }
400 
401 // Emit the Wasm header.
writeHeader(const MCAssembler & Asm)402 void WasmObjectWriter::writeHeader(const MCAssembler &Asm) {
403   W->OS.write(wasm::WasmMagic, sizeof(wasm::WasmMagic));
404   W->write<uint32_t>(wasm::WasmVersion);
405 }
406 
executePostLayoutBinding(MCAssembler & Asm,const MCAsmLayout & Layout)407 void WasmObjectWriter::executePostLayoutBinding(MCAssembler &Asm,
408                                                 const MCAsmLayout &Layout) {
409   // Some compilation units require the indirect function table to be present
410   // but don't explicitly reference it.  This is the case for call_indirect
411   // without the reference-types feature, and also function bitcasts in all
412   // cases.  In those cases the __indirect_function_table has the
413   // WASM_SYMBOL_NO_STRIP attribute.  Here we make sure this symbol makes it to
414   // the assembler, if needed.
415   if (auto *Sym = Asm.getContext().lookupSymbol("__indirect_function_table")) {
416     const auto *WasmSym = static_cast<const MCSymbolWasm *>(Sym);
417     if (WasmSym->isNoStrip())
418       Asm.registerSymbol(*Sym);
419   }
420 
421   // Build a map of sections to the function that defines them, for use
422   // in recordRelocation.
423   for (const MCSymbol &S : Asm.symbols()) {
424     const auto &WS = static_cast<const MCSymbolWasm &>(S);
425     if (WS.isDefined() && WS.isFunction() && !WS.isVariable()) {
426       const auto &Sec = static_cast<const MCSectionWasm &>(S.getSection());
427       auto Pair = SectionFunctions.insert(std::make_pair(&Sec, &S));
428       if (!Pair.second)
429         report_fatal_error("section already has a defining function: " +
430                            Sec.getName());
431     }
432   }
433 }
434 
recordRelocation(MCAssembler & Asm,const MCAsmLayout & Layout,const MCFragment * Fragment,const MCFixup & Fixup,MCValue Target,uint64_t & FixedValue)435 void WasmObjectWriter::recordRelocation(MCAssembler &Asm,
436                                         const MCAsmLayout &Layout,
437                                         const MCFragment *Fragment,
438                                         const MCFixup &Fixup, MCValue Target,
439                                         uint64_t &FixedValue) {
440   // The WebAssembly backend should never generate FKF_IsPCRel fixups
441   assert(!(Asm.getBackend().getFixupKindInfo(Fixup.getKind()).Flags &
442            MCFixupKindInfo::FKF_IsPCRel));
443 
444   const auto &FixupSection = cast<MCSectionWasm>(*Fragment->getParent());
445   uint64_t C = Target.getConstant();
446   uint64_t FixupOffset = Layout.getFragmentOffset(Fragment) + Fixup.getOffset();
447   MCContext &Ctx = Asm.getContext();
448   bool IsLocRel = false;
449 
450   if (const MCSymbolRefExpr *RefB = Target.getSymB()) {
451 
452     const auto &SymB = cast<MCSymbolWasm>(RefB->getSymbol());
453 
454     if (FixupSection.getKind().isText()) {
455       Ctx.reportError(Fixup.getLoc(),
456                       Twine("symbol '") + SymB.getName() +
457                           "' unsupported subtraction expression used in "
458                           "relocation in code section.");
459       return;
460     }
461 
462     if (SymB.isUndefined()) {
463       Ctx.reportError(Fixup.getLoc(),
464                       Twine("symbol '") + SymB.getName() +
465                           "' can not be undefined in a subtraction expression");
466       return;
467     }
468     const MCSection &SecB = SymB.getSection();
469     if (&SecB != &FixupSection) {
470       Ctx.reportError(Fixup.getLoc(),
471                       Twine("symbol '") + SymB.getName() +
472                           "' can not be placed in a different section");
473       return;
474     }
475     IsLocRel = true;
476     C += FixupOffset - Layout.getSymbolOffset(SymB);
477   }
478 
479   // We either rejected the fixup or folded B into C at this point.
480   const MCSymbolRefExpr *RefA = Target.getSymA();
481   const auto *SymA = cast<MCSymbolWasm>(&RefA->getSymbol());
482 
483   // The .init_array isn't translated as data, so don't do relocations in it.
484   if (FixupSection.getName().startswith(".init_array")) {
485     SymA->setUsedInInitArray();
486     return;
487   }
488 
489   if (SymA->isVariable()) {
490     const MCExpr *Expr = SymA->getVariableValue();
491     if (const auto *Inner = dyn_cast<MCSymbolRefExpr>(Expr))
492       if (Inner->getKind() == MCSymbolRefExpr::VK_WEAKREF)
493         llvm_unreachable("weakref used in reloc not yet implemented");
494   }
495 
496   // Put any constant offset in an addend. Offsets can be negative, and
497   // LLVM expects wrapping, in contrast to wasm's immediates which can't
498   // be negative and don't wrap.
499   FixedValue = 0;
500 
501   unsigned Type =
502       TargetObjectWriter->getRelocType(Target, Fixup, FixupSection, IsLocRel);
503 
504   // Absolute offset within a section or a function.
505   // Currently only supported for for metadata sections.
506   // See: test/MC/WebAssembly/blockaddress.ll
507   if ((Type == wasm::R_WASM_FUNCTION_OFFSET_I32 ||
508        Type == wasm::R_WASM_FUNCTION_OFFSET_I64 ||
509        Type == wasm::R_WASM_SECTION_OFFSET_I32) &&
510       SymA->isDefined()) {
511     // SymA can be a temp data symbol that represents a function (in which case
512     // it needs to be replaced by the section symbol), [XXX and it apparently
513     // later gets changed again to a func symbol?] or it can be a real
514     // function symbol, in which case it can be left as-is.
515 
516     if (!FixupSection.getKind().isMetadata())
517       report_fatal_error("relocations for function or section offsets are "
518                          "only supported in metadata sections");
519 
520     const MCSymbol *SectionSymbol = nullptr;
521     const MCSection &SecA = SymA->getSection();
522     if (SecA.getKind().isText()) {
523       auto SecSymIt = SectionFunctions.find(&SecA);
524       if (SecSymIt == SectionFunctions.end())
525         report_fatal_error("section doesn\'t have defining symbol");
526       SectionSymbol = SecSymIt->second;
527     } else {
528       SectionSymbol = SecA.getBeginSymbol();
529     }
530     if (!SectionSymbol)
531       report_fatal_error("section symbol is required for relocation");
532 
533     C += Layout.getSymbolOffset(*SymA);
534     SymA = cast<MCSymbolWasm>(SectionSymbol);
535   }
536 
537   if (Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB ||
538       Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB64 ||
539       Type == wasm::R_WASM_TABLE_INDEX_SLEB ||
540       Type == wasm::R_WASM_TABLE_INDEX_SLEB64 ||
541       Type == wasm::R_WASM_TABLE_INDEX_I32 ||
542       Type == wasm::R_WASM_TABLE_INDEX_I64) {
543     // TABLE_INDEX relocs implicitly use the default indirect function table.
544     // We require the function table to have already been defined.
545     auto TableName = "__indirect_function_table";
546     MCSymbolWasm *Sym = cast_or_null<MCSymbolWasm>(Ctx.lookupSymbol(TableName));
547     if (!Sym) {
548       report_fatal_error("missing indirect function table symbol");
549     } else {
550       if (!Sym->isFunctionTable())
551         report_fatal_error("__indirect_function_table symbol has wrong type");
552       // Ensure that __indirect_function_table reaches the output.
553       Sym->setNoStrip();
554       Asm.registerSymbol(*Sym);
555     }
556   }
557 
558   // Relocation other than R_WASM_TYPE_INDEX_LEB are required to be
559   // against a named symbol.
560   if (Type != wasm::R_WASM_TYPE_INDEX_LEB) {
561     if (SymA->getName().empty())
562       report_fatal_error("relocations against un-named temporaries are not yet "
563                          "supported by wasm");
564 
565     SymA->setUsedInReloc();
566   }
567 
568   if (RefA->getKind() == MCSymbolRefExpr::VK_GOT)
569     SymA->setUsedInGOT();
570 
571   WasmRelocationEntry Rec(FixupOffset, SymA, C, Type, &FixupSection);
572   LLVM_DEBUG(dbgs() << "WasmReloc: " << Rec << "\n");
573 
574   if (FixupSection.isWasmData()) {
575     DataRelocations.push_back(Rec);
576   } else if (FixupSection.getKind().isText()) {
577     CodeRelocations.push_back(Rec);
578   } else if (FixupSection.getKind().isMetadata()) {
579     CustomSectionsRelocations[&FixupSection].push_back(Rec);
580   } else {
581     llvm_unreachable("unexpected section type");
582   }
583 }
584 
585 // Compute a value to write into the code at the location covered
586 // by RelEntry. This value isn't used by the static linker; it just serves
587 // to make the object format more readable and more likely to be directly
588 // useable.
589 uint64_t
getProvisionalValue(const WasmRelocationEntry & RelEntry,const MCAsmLayout & Layout)590 WasmObjectWriter::getProvisionalValue(const WasmRelocationEntry &RelEntry,
591                                       const MCAsmLayout &Layout) {
592   if ((RelEntry.Type == wasm::R_WASM_GLOBAL_INDEX_LEB ||
593        RelEntry.Type == wasm::R_WASM_GLOBAL_INDEX_I32) &&
594       !RelEntry.Symbol->isGlobal()) {
595     assert(GOTIndices.count(RelEntry.Symbol) > 0 && "symbol not found in GOT index space");
596     return GOTIndices[RelEntry.Symbol];
597   }
598 
599   switch (RelEntry.Type) {
600   case wasm::R_WASM_TABLE_INDEX_REL_SLEB:
601   case wasm::R_WASM_TABLE_INDEX_REL_SLEB64:
602   case wasm::R_WASM_TABLE_INDEX_SLEB:
603   case wasm::R_WASM_TABLE_INDEX_SLEB64:
604   case wasm::R_WASM_TABLE_INDEX_I32:
605   case wasm::R_WASM_TABLE_INDEX_I64: {
606     // Provisional value is table address of the resolved symbol itself
607     const MCSymbolWasm *Base =
608         cast<MCSymbolWasm>(Layout.getBaseSymbol(*RelEntry.Symbol));
609     assert(Base->isFunction());
610     if (RelEntry.Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB ||
611         RelEntry.Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB64)
612       return TableIndices[Base] - InitialTableOffset;
613     else
614       return TableIndices[Base];
615   }
616   case wasm::R_WASM_TYPE_INDEX_LEB:
617     // Provisional value is same as the index
618     return getRelocationIndexValue(RelEntry);
619   case wasm::R_WASM_FUNCTION_INDEX_LEB:
620   case wasm::R_WASM_GLOBAL_INDEX_LEB:
621   case wasm::R_WASM_GLOBAL_INDEX_I32:
622   case wasm::R_WASM_TAG_INDEX_LEB:
623   case wasm::R_WASM_TABLE_NUMBER_LEB:
624     // Provisional value is function/global/tag Wasm index
625     assert(WasmIndices.count(RelEntry.Symbol) > 0 && "symbol not found in wasm index space");
626     return WasmIndices[RelEntry.Symbol];
627   case wasm::R_WASM_FUNCTION_OFFSET_I32:
628   case wasm::R_WASM_FUNCTION_OFFSET_I64:
629   case wasm::R_WASM_SECTION_OFFSET_I32: {
630     if (!RelEntry.Symbol->isDefined())
631       return 0;
632     const auto &Section =
633         static_cast<const MCSectionWasm &>(RelEntry.Symbol->getSection());
634     return Section.getSectionOffset() + RelEntry.Addend;
635   }
636   case wasm::R_WASM_MEMORY_ADDR_LEB:
637   case wasm::R_WASM_MEMORY_ADDR_LEB64:
638   case wasm::R_WASM_MEMORY_ADDR_SLEB:
639   case wasm::R_WASM_MEMORY_ADDR_SLEB64:
640   case wasm::R_WASM_MEMORY_ADDR_REL_SLEB:
641   case wasm::R_WASM_MEMORY_ADDR_REL_SLEB64:
642   case wasm::R_WASM_MEMORY_ADDR_I32:
643   case wasm::R_WASM_MEMORY_ADDR_I64:
644   case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB:
645   case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB64:
646   case wasm::R_WASM_MEMORY_ADDR_LOCREL_I32: {
647     // Provisional value is address of the global plus the offset
648     // For undefined symbols, use zero
649     if (!RelEntry.Symbol->isDefined())
650       return 0;
651     const wasm::WasmDataReference &SymRef = DataLocations[RelEntry.Symbol];
652     const WasmDataSegment &Segment = DataSegments[SymRef.Segment];
653     // Ignore overflow. LLVM allows address arithmetic to silently wrap.
654     return Segment.Offset + SymRef.Offset + RelEntry.Addend;
655   }
656   default:
657     llvm_unreachable("invalid relocation type");
658   }
659 }
660 
addData(SmallVectorImpl<char> & DataBytes,MCSectionWasm & DataSection)661 static void addData(SmallVectorImpl<char> &DataBytes,
662                     MCSectionWasm &DataSection) {
663   LLVM_DEBUG(errs() << "addData: " << DataSection.getName() << "\n");
664 
665   DataBytes.resize(alignTo(DataBytes.size(), DataSection.getAlignment()));
666 
667   for (const MCFragment &Frag : DataSection) {
668     if (Frag.hasInstructions())
669       report_fatal_error("only data supported in data sections");
670 
671     if (auto *Align = dyn_cast<MCAlignFragment>(&Frag)) {
672       if (Align->getValueSize() != 1)
673         report_fatal_error("only byte values supported for alignment");
674       // If nops are requested, use zeros, as this is the data section.
675       uint8_t Value = Align->hasEmitNops() ? 0 : Align->getValue();
676       uint64_t Size =
677           std::min<uint64_t>(alignTo(DataBytes.size(), Align->getAlignment()),
678                              DataBytes.size() + Align->getMaxBytesToEmit());
679       DataBytes.resize(Size, Value);
680     } else if (auto *Fill = dyn_cast<MCFillFragment>(&Frag)) {
681       int64_t NumValues;
682       if (!Fill->getNumValues().evaluateAsAbsolute(NumValues))
683         llvm_unreachable("The fill should be an assembler constant");
684       DataBytes.insert(DataBytes.end(), Fill->getValueSize() * NumValues,
685                        Fill->getValue());
686     } else if (auto *LEB = dyn_cast<MCLEBFragment>(&Frag)) {
687       const SmallVectorImpl<char> &Contents = LEB->getContents();
688       llvm::append_range(DataBytes, Contents);
689     } else {
690       const auto &DataFrag = cast<MCDataFragment>(Frag);
691       const SmallVectorImpl<char> &Contents = DataFrag.getContents();
692       llvm::append_range(DataBytes, Contents);
693     }
694   }
695 
696   LLVM_DEBUG(dbgs() << "addData -> " << DataBytes.size() << "\n");
697 }
698 
699 uint32_t
getRelocationIndexValue(const WasmRelocationEntry & RelEntry)700 WasmObjectWriter::getRelocationIndexValue(const WasmRelocationEntry &RelEntry) {
701   if (RelEntry.Type == wasm::R_WASM_TYPE_INDEX_LEB) {
702     if (!TypeIndices.count(RelEntry.Symbol))
703       report_fatal_error("symbol not found in type index space: " +
704                          RelEntry.Symbol->getName());
705     return TypeIndices[RelEntry.Symbol];
706   }
707 
708   return RelEntry.Symbol->getIndex();
709 }
710 
711 // Apply the portions of the relocation records that we can handle ourselves
712 // directly.
applyRelocations(ArrayRef<WasmRelocationEntry> Relocations,uint64_t ContentsOffset,const MCAsmLayout & Layout)713 void WasmObjectWriter::applyRelocations(
714     ArrayRef<WasmRelocationEntry> Relocations, uint64_t ContentsOffset,
715     const MCAsmLayout &Layout) {
716   auto &Stream = static_cast<raw_pwrite_stream &>(W->OS);
717   for (const WasmRelocationEntry &RelEntry : Relocations) {
718     uint64_t Offset = ContentsOffset +
719                       RelEntry.FixupSection->getSectionOffset() +
720                       RelEntry.Offset;
721 
722     LLVM_DEBUG(dbgs() << "applyRelocation: " << RelEntry << "\n");
723     auto Value = getProvisionalValue(RelEntry, Layout);
724 
725     switch (RelEntry.Type) {
726     case wasm::R_WASM_FUNCTION_INDEX_LEB:
727     case wasm::R_WASM_TYPE_INDEX_LEB:
728     case wasm::R_WASM_GLOBAL_INDEX_LEB:
729     case wasm::R_WASM_MEMORY_ADDR_LEB:
730     case wasm::R_WASM_TAG_INDEX_LEB:
731     case wasm::R_WASM_TABLE_NUMBER_LEB:
732       writePatchableLEB<5>(Stream, Value, Offset);
733       break;
734     case wasm::R_WASM_MEMORY_ADDR_LEB64:
735       writePatchableLEB<10>(Stream, Value, Offset);
736       break;
737     case wasm::R_WASM_TABLE_INDEX_I32:
738     case wasm::R_WASM_MEMORY_ADDR_I32:
739     case wasm::R_WASM_FUNCTION_OFFSET_I32:
740     case wasm::R_WASM_SECTION_OFFSET_I32:
741     case wasm::R_WASM_GLOBAL_INDEX_I32:
742     case wasm::R_WASM_MEMORY_ADDR_LOCREL_I32:
743       patchI32(Stream, Value, Offset);
744       break;
745     case wasm::R_WASM_TABLE_INDEX_I64:
746     case wasm::R_WASM_MEMORY_ADDR_I64:
747     case wasm::R_WASM_FUNCTION_OFFSET_I64:
748       patchI64(Stream, Value, Offset);
749       break;
750     case wasm::R_WASM_TABLE_INDEX_SLEB:
751     case wasm::R_WASM_TABLE_INDEX_REL_SLEB:
752     case wasm::R_WASM_MEMORY_ADDR_SLEB:
753     case wasm::R_WASM_MEMORY_ADDR_REL_SLEB:
754     case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB:
755       writePatchableSLEB<5>(Stream, Value, Offset);
756       break;
757     case wasm::R_WASM_TABLE_INDEX_SLEB64:
758     case wasm::R_WASM_TABLE_INDEX_REL_SLEB64:
759     case wasm::R_WASM_MEMORY_ADDR_SLEB64:
760     case wasm::R_WASM_MEMORY_ADDR_REL_SLEB64:
761     case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB64:
762       writePatchableSLEB<10>(Stream, Value, Offset);
763       break;
764     default:
765       llvm_unreachable("invalid relocation type");
766     }
767   }
768 }
769 
writeTypeSection(ArrayRef<wasm::WasmSignature> Signatures)770 void WasmObjectWriter::writeTypeSection(
771     ArrayRef<wasm::WasmSignature> Signatures) {
772   if (Signatures.empty())
773     return;
774 
775   SectionBookkeeping Section;
776   startSection(Section, wasm::WASM_SEC_TYPE);
777 
778   encodeULEB128(Signatures.size(), W->OS);
779 
780   for (const wasm::WasmSignature &Sig : Signatures) {
781     W->OS << char(wasm::WASM_TYPE_FUNC);
782     encodeULEB128(Sig.Params.size(), W->OS);
783     for (wasm::ValType Ty : Sig.Params)
784       writeValueType(Ty);
785     encodeULEB128(Sig.Returns.size(), W->OS);
786     for (wasm::ValType Ty : Sig.Returns)
787       writeValueType(Ty);
788   }
789 
790   endSection(Section);
791 }
792 
writeImportSection(ArrayRef<wasm::WasmImport> Imports,uint64_t DataSize,uint32_t NumElements)793 void WasmObjectWriter::writeImportSection(ArrayRef<wasm::WasmImport> Imports,
794                                           uint64_t DataSize,
795                                           uint32_t NumElements) {
796   if (Imports.empty())
797     return;
798 
799   uint64_t NumPages = (DataSize + wasm::WasmPageSize - 1) / wasm::WasmPageSize;
800 
801   SectionBookkeeping Section;
802   startSection(Section, wasm::WASM_SEC_IMPORT);
803 
804   encodeULEB128(Imports.size(), W->OS);
805   for (const wasm::WasmImport &Import : Imports) {
806     writeString(Import.Module);
807     writeString(Import.Field);
808     W->OS << char(Import.Kind);
809 
810     switch (Import.Kind) {
811     case wasm::WASM_EXTERNAL_FUNCTION:
812       encodeULEB128(Import.SigIndex, W->OS);
813       break;
814     case wasm::WASM_EXTERNAL_GLOBAL:
815       W->OS << char(Import.Global.Type);
816       W->OS << char(Import.Global.Mutable ? 1 : 0);
817       break;
818     case wasm::WASM_EXTERNAL_MEMORY:
819       encodeULEB128(Import.Memory.Flags, W->OS);
820       encodeULEB128(NumPages, W->OS); // initial
821       break;
822     case wasm::WASM_EXTERNAL_TABLE:
823       W->OS << char(Import.Table.ElemType);
824       encodeULEB128(0, W->OS);           // flags
825       encodeULEB128(NumElements, W->OS); // initial
826       break;
827     case wasm::WASM_EXTERNAL_TAG:
828       W->OS << char(Import.Tag.Attribute);
829       encodeULEB128(Import.Tag.SigIndex, W->OS);
830       break;
831     default:
832       llvm_unreachable("unsupported import kind");
833     }
834   }
835 
836   endSection(Section);
837 }
838 
writeFunctionSection(ArrayRef<WasmFunction> Functions)839 void WasmObjectWriter::writeFunctionSection(ArrayRef<WasmFunction> Functions) {
840   if (Functions.empty())
841     return;
842 
843   SectionBookkeeping Section;
844   startSection(Section, wasm::WASM_SEC_FUNCTION);
845 
846   encodeULEB128(Functions.size(), W->OS);
847   for (const WasmFunction &Func : Functions)
848     encodeULEB128(Func.SigIndex, W->OS);
849 
850   endSection(Section);
851 }
852 
writeTagSection(ArrayRef<wasm::WasmTagType> Tags)853 void WasmObjectWriter::writeTagSection(ArrayRef<wasm::WasmTagType> Tags) {
854   if (Tags.empty())
855     return;
856 
857   SectionBookkeeping Section;
858   startSection(Section, wasm::WASM_SEC_TAG);
859 
860   encodeULEB128(Tags.size(), W->OS);
861   for (const wasm::WasmTagType &Tag : Tags) {
862     W->OS << char(Tag.Attribute);
863     encodeULEB128(Tag.SigIndex, W->OS);
864   }
865 
866   endSection(Section);
867 }
868 
writeGlobalSection(ArrayRef<wasm::WasmGlobal> Globals)869 void WasmObjectWriter::writeGlobalSection(ArrayRef<wasm::WasmGlobal> Globals) {
870   if (Globals.empty())
871     return;
872 
873   SectionBookkeeping Section;
874   startSection(Section, wasm::WASM_SEC_GLOBAL);
875 
876   encodeULEB128(Globals.size(), W->OS);
877   for (const wasm::WasmGlobal &Global : Globals) {
878     encodeULEB128(Global.Type.Type, W->OS);
879     W->OS << char(Global.Type.Mutable);
880     W->OS << char(Global.InitExpr.Opcode);
881     switch (Global.Type.Type) {
882     case wasm::WASM_TYPE_I32:
883       encodeSLEB128(0, W->OS);
884       break;
885     case wasm::WASM_TYPE_I64:
886       encodeSLEB128(0, W->OS);
887       break;
888     case wasm::WASM_TYPE_F32:
889       writeI32(0);
890       break;
891     case wasm::WASM_TYPE_F64:
892       writeI64(0);
893       break;
894     case wasm::WASM_TYPE_EXTERNREF:
895       writeValueType(wasm::ValType::EXTERNREF);
896       break;
897     default:
898       llvm_unreachable("unexpected type");
899     }
900     W->OS << char(wasm::WASM_OPCODE_END);
901   }
902 
903   endSection(Section);
904 }
905 
writeTableSection(ArrayRef<wasm::WasmTable> Tables)906 void WasmObjectWriter::writeTableSection(ArrayRef<wasm::WasmTable> Tables) {
907   if (Tables.empty())
908     return;
909 
910   SectionBookkeeping Section;
911   startSection(Section, wasm::WASM_SEC_TABLE);
912 
913   encodeULEB128(Tables.size(), W->OS);
914   for (const wasm::WasmTable &Table : Tables) {
915     encodeULEB128(Table.Type.ElemType, W->OS);
916     encodeULEB128(Table.Type.Limits.Flags, W->OS);
917     encodeULEB128(Table.Type.Limits.Minimum, W->OS);
918     if (Table.Type.Limits.Flags & wasm::WASM_LIMITS_FLAG_HAS_MAX)
919       encodeULEB128(Table.Type.Limits.Maximum, W->OS);
920   }
921   endSection(Section);
922 }
923 
writeExportSection(ArrayRef<wasm::WasmExport> Exports)924 void WasmObjectWriter::writeExportSection(ArrayRef<wasm::WasmExport> Exports) {
925   if (Exports.empty())
926     return;
927 
928   SectionBookkeeping Section;
929   startSection(Section, wasm::WASM_SEC_EXPORT);
930 
931   encodeULEB128(Exports.size(), W->OS);
932   for (const wasm::WasmExport &Export : Exports) {
933     writeString(Export.Name);
934     W->OS << char(Export.Kind);
935     encodeULEB128(Export.Index, W->OS);
936   }
937 
938   endSection(Section);
939 }
940 
writeElemSection(const MCSymbolWasm * IndirectFunctionTable,ArrayRef<uint32_t> TableElems)941 void WasmObjectWriter::writeElemSection(
942     const MCSymbolWasm *IndirectFunctionTable, ArrayRef<uint32_t> TableElems) {
943   if (TableElems.empty())
944     return;
945 
946   assert(IndirectFunctionTable);
947 
948   SectionBookkeeping Section;
949   startSection(Section, wasm::WASM_SEC_ELEM);
950 
951   encodeULEB128(1, W->OS); // number of "segments"
952 
953   assert(WasmIndices.count(IndirectFunctionTable));
954   uint32_t TableNumber = WasmIndices.find(IndirectFunctionTable)->second;
955   uint32_t Flags = 0;
956   if (TableNumber)
957     Flags |= wasm::WASM_ELEM_SEGMENT_HAS_TABLE_NUMBER;
958   encodeULEB128(Flags, W->OS);
959   if (Flags & wasm::WASM_ELEM_SEGMENT_HAS_TABLE_NUMBER)
960     encodeULEB128(TableNumber, W->OS); // the table number
961 
962   // init expr for starting offset
963   W->OS << char(wasm::WASM_OPCODE_I32_CONST);
964   encodeSLEB128(InitialTableOffset, W->OS);
965   W->OS << char(wasm::WASM_OPCODE_END);
966 
967   if (Flags & wasm::WASM_ELEM_SEGMENT_MASK_HAS_ELEM_KIND) {
968     // We only write active function table initializers, for which the elem kind
969     // is specified to be written as 0x00 and interpreted to mean "funcref".
970     const uint8_t ElemKind = 0;
971     W->OS << ElemKind;
972   }
973 
974   encodeULEB128(TableElems.size(), W->OS);
975   for (uint32_t Elem : TableElems)
976     encodeULEB128(Elem, W->OS);
977 
978   endSection(Section);
979 }
980 
writeDataCountSection()981 void WasmObjectWriter::writeDataCountSection() {
982   if (DataSegments.empty())
983     return;
984 
985   SectionBookkeeping Section;
986   startSection(Section, wasm::WASM_SEC_DATACOUNT);
987   encodeULEB128(DataSegments.size(), W->OS);
988   endSection(Section);
989 }
990 
writeCodeSection(const MCAssembler & Asm,const MCAsmLayout & Layout,ArrayRef<WasmFunction> Functions)991 uint32_t WasmObjectWriter::writeCodeSection(const MCAssembler &Asm,
992                                             const MCAsmLayout &Layout,
993                                             ArrayRef<WasmFunction> Functions) {
994   if (Functions.empty())
995     return 0;
996 
997   SectionBookkeeping Section;
998   startSection(Section, wasm::WASM_SEC_CODE);
999 
1000   encodeULEB128(Functions.size(), W->OS);
1001 
1002   for (const WasmFunction &Func : Functions) {
1003     auto &FuncSection = static_cast<MCSectionWasm &>(Func.Sym->getSection());
1004 
1005     int64_t Size = 0;
1006     if (!Func.Sym->getSize()->evaluateAsAbsolute(Size, Layout))
1007       report_fatal_error(".size expression must be evaluatable");
1008 
1009     encodeULEB128(Size, W->OS);
1010     FuncSection.setSectionOffset(W->OS.tell() - Section.ContentsOffset);
1011     Asm.writeSectionData(W->OS, &FuncSection, Layout);
1012   }
1013 
1014   // Apply fixups.
1015   applyRelocations(CodeRelocations, Section.ContentsOffset, Layout);
1016 
1017   endSection(Section);
1018   return Section.Index;
1019 }
1020 
writeDataSection(const MCAsmLayout & Layout)1021 uint32_t WasmObjectWriter::writeDataSection(const MCAsmLayout &Layout) {
1022   if (DataSegments.empty())
1023     return 0;
1024 
1025   SectionBookkeeping Section;
1026   startSection(Section, wasm::WASM_SEC_DATA);
1027 
1028   encodeULEB128(DataSegments.size(), W->OS); // count
1029 
1030   for (const WasmDataSegment &Segment : DataSegments) {
1031     encodeULEB128(Segment.InitFlags, W->OS); // flags
1032     if (Segment.InitFlags & wasm::WASM_DATA_SEGMENT_HAS_MEMINDEX)
1033       encodeULEB128(0, W->OS); // memory index
1034     if ((Segment.InitFlags & wasm::WASM_DATA_SEGMENT_IS_PASSIVE) == 0) {
1035       W->OS << char(is64Bit() ? wasm::WASM_OPCODE_I64_CONST
1036                               : wasm::WASM_OPCODE_I32_CONST);
1037       encodeSLEB128(Segment.Offset, W->OS); // offset
1038       W->OS << char(wasm::WASM_OPCODE_END);
1039     }
1040     encodeULEB128(Segment.Data.size(), W->OS); // size
1041     Segment.Section->setSectionOffset(W->OS.tell() - Section.ContentsOffset);
1042     W->OS << Segment.Data; // data
1043   }
1044 
1045   // Apply fixups.
1046   applyRelocations(DataRelocations, Section.ContentsOffset, Layout);
1047 
1048   endSection(Section);
1049   return Section.Index;
1050 }
1051 
writeRelocSection(uint32_t SectionIndex,StringRef Name,std::vector<WasmRelocationEntry> & Relocs)1052 void WasmObjectWriter::writeRelocSection(
1053     uint32_t SectionIndex, StringRef Name,
1054     std::vector<WasmRelocationEntry> &Relocs) {
1055   // See: https://github.com/WebAssembly/tool-conventions/blob/master/Linking.md
1056   // for descriptions of the reloc sections.
1057 
1058   if (Relocs.empty())
1059     return;
1060 
1061   // First, ensure the relocations are sorted in offset order.  In general they
1062   // should already be sorted since `recordRelocation` is called in offset
1063   // order, but for the code section we combine many MC sections into single
1064   // wasm section, and this order is determined by the order of Asm.Symbols()
1065   // not the sections order.
1066   llvm::stable_sort(
1067       Relocs, [](const WasmRelocationEntry &A, const WasmRelocationEntry &B) {
1068         return (A.Offset + A.FixupSection->getSectionOffset()) <
1069                (B.Offset + B.FixupSection->getSectionOffset());
1070       });
1071 
1072   SectionBookkeeping Section;
1073   startCustomSection(Section, std::string("reloc.") + Name.str());
1074 
1075   encodeULEB128(SectionIndex, W->OS);
1076   encodeULEB128(Relocs.size(), W->OS);
1077   for (const WasmRelocationEntry &RelEntry : Relocs) {
1078     uint64_t Offset =
1079         RelEntry.Offset + RelEntry.FixupSection->getSectionOffset();
1080     uint32_t Index = getRelocationIndexValue(RelEntry);
1081 
1082     W->OS << char(RelEntry.Type);
1083     encodeULEB128(Offset, W->OS);
1084     encodeULEB128(Index, W->OS);
1085     if (RelEntry.hasAddend())
1086       encodeSLEB128(RelEntry.Addend, W->OS);
1087   }
1088 
1089   endSection(Section);
1090 }
1091 
writeCustomRelocSections()1092 void WasmObjectWriter::writeCustomRelocSections() {
1093   for (const auto &Sec : CustomSections) {
1094     auto &Relocations = CustomSectionsRelocations[Sec.Section];
1095     writeRelocSection(Sec.OutputIndex, Sec.Name, Relocations);
1096   }
1097 }
1098 
writeLinkingMetaDataSection(ArrayRef<wasm::WasmSymbolInfo> SymbolInfos,ArrayRef<std::pair<uint16_t,uint32_t>> InitFuncs,const std::map<StringRef,std::vector<WasmComdatEntry>> & Comdats)1099 void WasmObjectWriter::writeLinkingMetaDataSection(
1100     ArrayRef<wasm::WasmSymbolInfo> SymbolInfos,
1101     ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs,
1102     const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats) {
1103   SectionBookkeeping Section;
1104   startCustomSection(Section, "linking");
1105   encodeULEB128(wasm::WasmMetadataVersion, W->OS);
1106 
1107   SectionBookkeeping SubSection;
1108   if (SymbolInfos.size() != 0) {
1109     startSection(SubSection, wasm::WASM_SYMBOL_TABLE);
1110     encodeULEB128(SymbolInfos.size(), W->OS);
1111     for (const wasm::WasmSymbolInfo &Sym : SymbolInfos) {
1112       encodeULEB128(Sym.Kind, W->OS);
1113       encodeULEB128(Sym.Flags, W->OS);
1114       switch (Sym.Kind) {
1115       case wasm::WASM_SYMBOL_TYPE_FUNCTION:
1116       case wasm::WASM_SYMBOL_TYPE_GLOBAL:
1117       case wasm::WASM_SYMBOL_TYPE_TAG:
1118       case wasm::WASM_SYMBOL_TYPE_TABLE:
1119         encodeULEB128(Sym.ElementIndex, W->OS);
1120         if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0 ||
1121             (Sym.Flags & wasm::WASM_SYMBOL_EXPLICIT_NAME) != 0)
1122           writeString(Sym.Name);
1123         break;
1124       case wasm::WASM_SYMBOL_TYPE_DATA:
1125         writeString(Sym.Name);
1126         if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0) {
1127           encodeULEB128(Sym.DataRef.Segment, W->OS);
1128           encodeULEB128(Sym.DataRef.Offset, W->OS);
1129           encodeULEB128(Sym.DataRef.Size, W->OS);
1130         }
1131         break;
1132       case wasm::WASM_SYMBOL_TYPE_SECTION: {
1133         const uint32_t SectionIndex =
1134             CustomSections[Sym.ElementIndex].OutputIndex;
1135         encodeULEB128(SectionIndex, W->OS);
1136         break;
1137       }
1138       default:
1139         llvm_unreachable("unexpected kind");
1140       }
1141     }
1142     endSection(SubSection);
1143   }
1144 
1145   if (DataSegments.size()) {
1146     startSection(SubSection, wasm::WASM_SEGMENT_INFO);
1147     encodeULEB128(DataSegments.size(), W->OS);
1148     for (const WasmDataSegment &Segment : DataSegments) {
1149       writeString(Segment.Name);
1150       encodeULEB128(Segment.Alignment, W->OS);
1151       encodeULEB128(Segment.LinkingFlags, W->OS);
1152     }
1153     endSection(SubSection);
1154   }
1155 
1156   if (!InitFuncs.empty()) {
1157     startSection(SubSection, wasm::WASM_INIT_FUNCS);
1158     encodeULEB128(InitFuncs.size(), W->OS);
1159     for (auto &StartFunc : InitFuncs) {
1160       encodeULEB128(StartFunc.first, W->OS);  // priority
1161       encodeULEB128(StartFunc.second, W->OS); // function index
1162     }
1163     endSection(SubSection);
1164   }
1165 
1166   if (Comdats.size()) {
1167     startSection(SubSection, wasm::WASM_COMDAT_INFO);
1168     encodeULEB128(Comdats.size(), W->OS);
1169     for (const auto &C : Comdats) {
1170       writeString(C.first);
1171       encodeULEB128(0, W->OS); // flags for future use
1172       encodeULEB128(C.second.size(), W->OS);
1173       for (const WasmComdatEntry &Entry : C.second) {
1174         encodeULEB128(Entry.Kind, W->OS);
1175         encodeULEB128(Entry.Index, W->OS);
1176       }
1177     }
1178     endSection(SubSection);
1179   }
1180 
1181   endSection(Section);
1182 }
1183 
writeCustomSection(WasmCustomSection & CustomSection,const MCAssembler & Asm,const MCAsmLayout & Layout)1184 void WasmObjectWriter::writeCustomSection(WasmCustomSection &CustomSection,
1185                                           const MCAssembler &Asm,
1186                                           const MCAsmLayout &Layout) {
1187   SectionBookkeeping Section;
1188   auto *Sec = CustomSection.Section;
1189   startCustomSection(Section, CustomSection.Name);
1190 
1191   Sec->setSectionOffset(W->OS.tell() - Section.ContentsOffset);
1192   Asm.writeSectionData(W->OS, Sec, Layout);
1193 
1194   CustomSection.OutputContentsOffset = Section.ContentsOffset;
1195   CustomSection.OutputIndex = Section.Index;
1196 
1197   endSection(Section);
1198 
1199   // Apply fixups.
1200   auto &Relocations = CustomSectionsRelocations[CustomSection.Section];
1201   applyRelocations(Relocations, CustomSection.OutputContentsOffset, Layout);
1202 }
1203 
getFunctionType(const MCSymbolWasm & Symbol)1204 uint32_t WasmObjectWriter::getFunctionType(const MCSymbolWasm &Symbol) {
1205   assert(Symbol.isFunction());
1206   assert(TypeIndices.count(&Symbol));
1207   return TypeIndices[&Symbol];
1208 }
1209 
getTagType(const MCSymbolWasm & Symbol)1210 uint32_t WasmObjectWriter::getTagType(const MCSymbolWasm &Symbol) {
1211   assert(Symbol.isTag());
1212   assert(TypeIndices.count(&Symbol));
1213   return TypeIndices[&Symbol];
1214 }
1215 
registerFunctionType(const MCSymbolWasm & Symbol)1216 void WasmObjectWriter::registerFunctionType(const MCSymbolWasm &Symbol) {
1217   assert(Symbol.isFunction());
1218 
1219   wasm::WasmSignature S;
1220 
1221   if (auto *Sig = Symbol.getSignature()) {
1222     S.Returns = Sig->Returns;
1223     S.Params = Sig->Params;
1224   }
1225 
1226   auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size()));
1227   if (Pair.second)
1228     Signatures.push_back(S);
1229   TypeIndices[&Symbol] = Pair.first->second;
1230 
1231   LLVM_DEBUG(dbgs() << "registerFunctionType: " << Symbol
1232                     << " new:" << Pair.second << "\n");
1233   LLVM_DEBUG(dbgs() << "  -> type index: " << Pair.first->second << "\n");
1234 }
1235 
registerTagType(const MCSymbolWasm & Symbol)1236 void WasmObjectWriter::registerTagType(const MCSymbolWasm &Symbol) {
1237   assert(Symbol.isTag());
1238 
1239   // TODO Currently we don't generate imported exceptions, but if we do, we
1240   // should have a way of infering types of imported exceptions.
1241   wasm::WasmSignature S;
1242   if (auto *Sig = Symbol.getSignature()) {
1243     S.Returns = Sig->Returns;
1244     S.Params = Sig->Params;
1245   }
1246 
1247   auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size()));
1248   if (Pair.second)
1249     Signatures.push_back(S);
1250   TypeIndices[&Symbol] = Pair.first->second;
1251 
1252   LLVM_DEBUG(dbgs() << "registerTagType: " << Symbol << " new:" << Pair.second
1253                     << "\n");
1254   LLVM_DEBUG(dbgs() << "  -> type index: " << Pair.first->second << "\n");
1255 }
1256 
isInSymtab(const MCSymbolWasm & Sym)1257 static bool isInSymtab(const MCSymbolWasm &Sym) {
1258   if (Sym.isUsedInReloc() || Sym.isUsedInInitArray())
1259     return true;
1260 
1261   if (Sym.isComdat() && !Sym.isDefined())
1262     return false;
1263 
1264   if (Sym.isTemporary())
1265     return false;
1266 
1267   if (Sym.isSection())
1268     return false;
1269 
1270   if (Sym.omitFromLinkingSection())
1271     return false;
1272 
1273   return true;
1274 }
1275 
prepareImports(SmallVectorImpl<wasm::WasmImport> & Imports,MCAssembler & Asm,const MCAsmLayout & Layout)1276 void WasmObjectWriter::prepareImports(
1277     SmallVectorImpl<wasm::WasmImport> &Imports, MCAssembler &Asm,
1278     const MCAsmLayout &Layout) {
1279   // For now, always emit the memory import, since loads and stores are not
1280   // valid without it. In the future, we could perhaps be more clever and omit
1281   // it if there are no loads or stores.
1282   wasm::WasmImport MemImport;
1283   MemImport.Module = "env";
1284   MemImport.Field = "__linear_memory";
1285   MemImport.Kind = wasm::WASM_EXTERNAL_MEMORY;
1286   MemImport.Memory.Flags = is64Bit() ? wasm::WASM_LIMITS_FLAG_IS_64
1287                                      : wasm::WASM_LIMITS_FLAG_NONE;
1288   Imports.push_back(MemImport);
1289 
1290   // Populate SignatureIndices, and Imports and WasmIndices for undefined
1291   // symbols.  This must be done before populating WasmIndices for defined
1292   // symbols.
1293   for (const MCSymbol &S : Asm.symbols()) {
1294     const auto &WS = static_cast<const MCSymbolWasm &>(S);
1295 
1296     // Register types for all functions, including those with private linkage
1297     // (because wasm always needs a type signature).
1298     if (WS.isFunction()) {
1299       const auto *BS = Layout.getBaseSymbol(S);
1300       if (!BS)
1301         report_fatal_error(Twine(S.getName()) +
1302                            ": absolute addressing not supported!");
1303       registerFunctionType(*cast<MCSymbolWasm>(BS));
1304     }
1305 
1306     if (WS.isTag())
1307       registerTagType(WS);
1308 
1309     if (WS.isTemporary())
1310       continue;
1311 
1312     // If the symbol is not defined in this translation unit, import it.
1313     if (!WS.isDefined() && !WS.isComdat()) {
1314       if (WS.isFunction()) {
1315         wasm::WasmImport Import;
1316         Import.Module = WS.getImportModule();
1317         Import.Field = WS.getImportName();
1318         Import.Kind = wasm::WASM_EXTERNAL_FUNCTION;
1319         Import.SigIndex = getFunctionType(WS);
1320         Imports.push_back(Import);
1321         assert(WasmIndices.count(&WS) == 0);
1322         WasmIndices[&WS] = NumFunctionImports++;
1323       } else if (WS.isGlobal()) {
1324         if (WS.isWeak())
1325           report_fatal_error("undefined global symbol cannot be weak");
1326 
1327         wasm::WasmImport Import;
1328         Import.Field = WS.getImportName();
1329         Import.Kind = wasm::WASM_EXTERNAL_GLOBAL;
1330         Import.Module = WS.getImportModule();
1331         Import.Global = WS.getGlobalType();
1332         Imports.push_back(Import);
1333         assert(WasmIndices.count(&WS) == 0);
1334         WasmIndices[&WS] = NumGlobalImports++;
1335       } else if (WS.isTag()) {
1336         if (WS.isWeak())
1337           report_fatal_error("undefined tag symbol cannot be weak");
1338 
1339         wasm::WasmImport Import;
1340         Import.Module = WS.getImportModule();
1341         Import.Field = WS.getImportName();
1342         Import.Kind = wasm::WASM_EXTERNAL_TAG;
1343         Import.Tag.Attribute = wasm::WASM_TAG_ATTRIBUTE_EXCEPTION;
1344         Import.Tag.SigIndex = getTagType(WS);
1345         Imports.push_back(Import);
1346         assert(WasmIndices.count(&WS) == 0);
1347         WasmIndices[&WS] = NumTagImports++;
1348       } else if (WS.isTable()) {
1349         if (WS.isWeak())
1350           report_fatal_error("undefined table symbol cannot be weak");
1351 
1352         wasm::WasmImport Import;
1353         Import.Module = WS.getImportModule();
1354         Import.Field = WS.getImportName();
1355         Import.Kind = wasm::WASM_EXTERNAL_TABLE;
1356         Import.Table = WS.getTableType();
1357         Imports.push_back(Import);
1358         assert(WasmIndices.count(&WS) == 0);
1359         WasmIndices[&WS] = NumTableImports++;
1360       }
1361     }
1362   }
1363 
1364   // Add imports for GOT globals
1365   for (const MCSymbol &S : Asm.symbols()) {
1366     const auto &WS = static_cast<const MCSymbolWasm &>(S);
1367     if (WS.isUsedInGOT()) {
1368       wasm::WasmImport Import;
1369       if (WS.isFunction())
1370         Import.Module = "GOT.func";
1371       else
1372         Import.Module = "GOT.mem";
1373       Import.Field = WS.getName();
1374       Import.Kind = wasm::WASM_EXTERNAL_GLOBAL;
1375       Import.Global = {wasm::WASM_TYPE_I32, true};
1376       Imports.push_back(Import);
1377       assert(GOTIndices.count(&WS) == 0);
1378       GOTIndices[&WS] = NumGlobalImports++;
1379     }
1380   }
1381 }
1382 
writeObject(MCAssembler & Asm,const MCAsmLayout & Layout)1383 uint64_t WasmObjectWriter::writeObject(MCAssembler &Asm,
1384                                        const MCAsmLayout &Layout) {
1385   support::endian::Writer MainWriter(*OS, support::little);
1386   W = &MainWriter;
1387   if (IsSplitDwarf) {
1388     uint64_t TotalSize = writeOneObject(Asm, Layout, DwoMode::NonDwoOnly);
1389     assert(DwoOS);
1390     support::endian::Writer DwoWriter(*DwoOS, support::little);
1391     W = &DwoWriter;
1392     return TotalSize + writeOneObject(Asm, Layout, DwoMode::DwoOnly);
1393   } else {
1394     return writeOneObject(Asm, Layout, DwoMode::AllSections);
1395   }
1396 }
1397 
writeOneObject(MCAssembler & Asm,const MCAsmLayout & Layout,DwoMode Mode)1398 uint64_t WasmObjectWriter::writeOneObject(MCAssembler &Asm,
1399                                           const MCAsmLayout &Layout,
1400                                           DwoMode Mode) {
1401   uint64_t StartOffset = W->OS.tell();
1402   SectionCount = 0;
1403   CustomSections.clear();
1404 
1405   LLVM_DEBUG(dbgs() << "WasmObjectWriter::writeObject\n");
1406 
1407   // Collect information from the available symbols.
1408   SmallVector<WasmFunction, 4> Functions;
1409   SmallVector<uint32_t, 4> TableElems;
1410   SmallVector<wasm::WasmImport, 4> Imports;
1411   SmallVector<wasm::WasmExport, 4> Exports;
1412   SmallVector<wasm::WasmTagType, 1> Tags;
1413   SmallVector<wasm::WasmGlobal, 1> Globals;
1414   SmallVector<wasm::WasmTable, 1> Tables;
1415   SmallVector<wasm::WasmSymbolInfo, 4> SymbolInfos;
1416   SmallVector<std::pair<uint16_t, uint32_t>, 2> InitFuncs;
1417   std::map<StringRef, std::vector<WasmComdatEntry>> Comdats;
1418   uint64_t DataSize = 0;
1419   if (Mode != DwoMode::DwoOnly) {
1420     prepareImports(Imports, Asm, Layout);
1421   }
1422 
1423   // Populate DataSegments and CustomSections, which must be done before
1424   // populating DataLocations.
1425   for (MCSection &Sec : Asm) {
1426     auto &Section = static_cast<MCSectionWasm &>(Sec);
1427     StringRef SectionName = Section.getName();
1428 
1429     if (Mode == DwoMode::NonDwoOnly && isDwoSection(Sec))
1430       continue;
1431     if (Mode == DwoMode::DwoOnly && !isDwoSection(Sec))
1432       continue;
1433 
1434     LLVM_DEBUG(dbgs() << "Processing Section " << SectionName << "  group "
1435                       << Section.getGroup() << "\n";);
1436 
1437     // .init_array sections are handled specially elsewhere.
1438     if (SectionName.startswith(".init_array"))
1439       continue;
1440 
1441     // Code is handled separately
1442     if (Section.getKind().isText())
1443       continue;
1444 
1445     if (Section.isWasmData()) {
1446       uint32_t SegmentIndex = DataSegments.size();
1447       DataSize = alignTo(DataSize, Section.getAlignment());
1448       DataSegments.emplace_back();
1449       WasmDataSegment &Segment = DataSegments.back();
1450       Segment.Name = SectionName;
1451       Segment.InitFlags = Section.getPassive()
1452                               ? (uint32_t)wasm::WASM_DATA_SEGMENT_IS_PASSIVE
1453                               : 0;
1454       Segment.Offset = DataSize;
1455       Segment.Section = &Section;
1456       addData(Segment.Data, Section);
1457       Segment.Alignment = Log2_32(Section.getAlignment());
1458       Segment.LinkingFlags = Section.getSegmentFlags();
1459       DataSize += Segment.Data.size();
1460       Section.setSegmentIndex(SegmentIndex);
1461 
1462       if (const MCSymbolWasm *C = Section.getGroup()) {
1463         Comdats[C->getName()].emplace_back(
1464             WasmComdatEntry{wasm::WASM_COMDAT_DATA, SegmentIndex});
1465       }
1466     } else {
1467       // Create custom sections
1468       assert(Sec.getKind().isMetadata());
1469 
1470       StringRef Name = SectionName;
1471 
1472       // For user-defined custom sections, strip the prefix
1473       if (Name.startswith(".custom_section."))
1474         Name = Name.substr(strlen(".custom_section."));
1475 
1476       MCSymbol *Begin = Sec.getBeginSymbol();
1477       if (Begin) {
1478         assert(WasmIndices.count(cast<MCSymbolWasm>(Begin)) == 0);
1479         WasmIndices[cast<MCSymbolWasm>(Begin)] = CustomSections.size();
1480       }
1481 
1482       // Separate out the producers and target features sections
1483       if (Name == "producers") {
1484         ProducersSection = std::make_unique<WasmCustomSection>(Name, &Section);
1485         continue;
1486       }
1487       if (Name == "target_features") {
1488         TargetFeaturesSection =
1489             std::make_unique<WasmCustomSection>(Name, &Section);
1490         continue;
1491       }
1492 
1493       // Custom sections can also belong to COMDAT groups. In this case the
1494       // decriptor's "index" field is the section index (in the final object
1495       // file), but that is not known until after layout, so it must be fixed up
1496       // later
1497       if (const MCSymbolWasm *C = Section.getGroup()) {
1498         Comdats[C->getName()].emplace_back(
1499             WasmComdatEntry{wasm::WASM_COMDAT_SECTION,
1500                             static_cast<uint32_t>(CustomSections.size())});
1501       }
1502 
1503       CustomSections.emplace_back(Name, &Section);
1504     }
1505   }
1506 
1507   if (Mode != DwoMode::DwoOnly) {
1508     // Populate WasmIndices and DataLocations for defined symbols.
1509     for (const MCSymbol &S : Asm.symbols()) {
1510       // Ignore unnamed temporary symbols, which aren't ever exported, imported,
1511       // or used in relocations.
1512       if (S.isTemporary() && S.getName().empty())
1513         continue;
1514 
1515       const auto &WS = static_cast<const MCSymbolWasm &>(S);
1516       LLVM_DEBUG(dbgs()
1517                  << "MCSymbol: "
1518                  << toString(WS.getType().getValueOr(wasm::WASM_SYMBOL_TYPE_DATA))
1519                  << " '" << S << "'"
1520                  << " isDefined=" << S.isDefined() << " isExternal="
1521                  << S.isExternal() << " isTemporary=" << S.isTemporary()
1522                  << " isWeak=" << WS.isWeak() << " isHidden=" << WS.isHidden()
1523                  << " isVariable=" << WS.isVariable() << "\n");
1524 
1525       if (WS.isVariable())
1526         continue;
1527       if (WS.isComdat() && !WS.isDefined())
1528         continue;
1529 
1530       if (WS.isFunction()) {
1531         unsigned Index;
1532         if (WS.isDefined()) {
1533           if (WS.getOffset() != 0)
1534             report_fatal_error(
1535                 "function sections must contain one function each");
1536 
1537           if (WS.getSize() == nullptr)
1538             report_fatal_error(
1539                 "function symbols must have a size set with .size");
1540 
1541           // A definition. Write out the function body.
1542           Index = NumFunctionImports + Functions.size();
1543           WasmFunction Func;
1544           Func.SigIndex = getFunctionType(WS);
1545           Func.Sym = &WS;
1546           assert(WasmIndices.count(&WS) == 0);
1547           WasmIndices[&WS] = Index;
1548           Functions.push_back(Func);
1549 
1550           auto &Section = static_cast<MCSectionWasm &>(WS.getSection());
1551           if (const MCSymbolWasm *C = Section.getGroup()) {
1552             Comdats[C->getName()].emplace_back(
1553                 WasmComdatEntry{wasm::WASM_COMDAT_FUNCTION, Index});
1554           }
1555 
1556           if (WS.hasExportName()) {
1557             wasm::WasmExport Export;
1558             Export.Name = WS.getExportName();
1559             Export.Kind = wasm::WASM_EXTERNAL_FUNCTION;
1560             Export.Index = Index;
1561             Exports.push_back(Export);
1562           }
1563         } else {
1564           // An import; the index was assigned above.
1565           Index = WasmIndices.find(&WS)->second;
1566         }
1567 
1568         LLVM_DEBUG(dbgs() << "  -> function index: " << Index << "\n");
1569 
1570       } else if (WS.isData()) {
1571         if (!isInSymtab(WS))
1572           continue;
1573 
1574         if (!WS.isDefined()) {
1575           LLVM_DEBUG(dbgs() << "  -> segment index: -1"
1576                             << "\n");
1577           continue;
1578         }
1579 
1580         if (!WS.getSize())
1581           report_fatal_error("data symbols must have a size set with .size: " +
1582                              WS.getName());
1583 
1584         int64_t Size = 0;
1585         if (!WS.getSize()->evaluateAsAbsolute(Size, Layout))
1586           report_fatal_error(".size expression must be evaluatable");
1587 
1588         auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection());
1589         if (!DataSection.isWasmData())
1590           report_fatal_error("data symbols must live in a data section: " +
1591                              WS.getName());
1592 
1593         // For each data symbol, export it in the symtab as a reference to the
1594         // corresponding Wasm data segment.
1595         wasm::WasmDataReference Ref = wasm::WasmDataReference{
1596             DataSection.getSegmentIndex(), Layout.getSymbolOffset(WS),
1597             static_cast<uint64_t>(Size)};
1598         assert(DataLocations.count(&WS) == 0);
1599         DataLocations[&WS] = Ref;
1600         LLVM_DEBUG(dbgs() << "  -> segment index: " << Ref.Segment << "\n");
1601 
1602       } else if (WS.isGlobal()) {
1603         // A "true" Wasm global (currently just __stack_pointer)
1604         if (WS.isDefined()) {
1605           wasm::WasmGlobal Global;
1606           Global.Type = WS.getGlobalType();
1607           Global.Index = NumGlobalImports + Globals.size();
1608           switch (Global.Type.Type) {
1609           case wasm::WASM_TYPE_I32:
1610             Global.InitExpr.Opcode = wasm::WASM_OPCODE_I32_CONST;
1611             break;
1612           case wasm::WASM_TYPE_I64:
1613             Global.InitExpr.Opcode = wasm::WASM_OPCODE_I64_CONST;
1614             break;
1615           case wasm::WASM_TYPE_F32:
1616             Global.InitExpr.Opcode = wasm::WASM_OPCODE_F32_CONST;
1617             break;
1618           case wasm::WASM_TYPE_F64:
1619             Global.InitExpr.Opcode = wasm::WASM_OPCODE_F64_CONST;
1620             break;
1621           case wasm::WASM_TYPE_EXTERNREF:
1622             Global.InitExpr.Opcode = wasm::WASM_OPCODE_REF_NULL;
1623             break;
1624           default:
1625             llvm_unreachable("unexpected type");
1626           }
1627           assert(WasmIndices.count(&WS) == 0);
1628           WasmIndices[&WS] = Global.Index;
1629           Globals.push_back(Global);
1630         } else {
1631           // An import; the index was assigned above
1632           LLVM_DEBUG(dbgs() << "  -> global index: "
1633                             << WasmIndices.find(&WS)->second << "\n");
1634         }
1635       } else if (WS.isTable()) {
1636         if (WS.isDefined()) {
1637           wasm::WasmTable Table;
1638           Table.Index = NumTableImports + Tables.size();
1639           Table.Type = WS.getTableType();
1640           assert(WasmIndices.count(&WS) == 0);
1641           WasmIndices[&WS] = Table.Index;
1642           Tables.push_back(Table);
1643         }
1644         LLVM_DEBUG(dbgs() << " -> table index: "
1645                           << WasmIndices.find(&WS)->second << "\n");
1646       } else if (WS.isTag()) {
1647         // C++ exception symbol (__cpp_exception)
1648         unsigned Index;
1649         if (WS.isDefined()) {
1650           Index = NumTagImports + Tags.size();
1651           wasm::WasmTagType Tag;
1652           Tag.SigIndex = getTagType(WS);
1653           Tag.Attribute = wasm::WASM_TAG_ATTRIBUTE_EXCEPTION;
1654           assert(WasmIndices.count(&WS) == 0);
1655           WasmIndices[&WS] = Index;
1656           Tags.push_back(Tag);
1657         } else {
1658           // An import; the index was assigned above.
1659           assert(WasmIndices.count(&WS) > 0);
1660         }
1661         LLVM_DEBUG(dbgs() << "  -> tag index: " << WasmIndices.find(&WS)->second
1662                           << "\n");
1663 
1664       } else {
1665         assert(WS.isSection());
1666       }
1667     }
1668 
1669     // Populate WasmIndices and DataLocations for aliased symbols.  We need to
1670     // process these in a separate pass because we need to have processed the
1671     // target of the alias before the alias itself and the symbols are not
1672     // necessarily ordered in this way.
1673     for (const MCSymbol &S : Asm.symbols()) {
1674       if (!S.isVariable())
1675         continue;
1676 
1677       assert(S.isDefined());
1678 
1679       const auto *BS = Layout.getBaseSymbol(S);
1680       if (!BS)
1681         report_fatal_error(Twine(S.getName()) +
1682                            ": absolute addressing not supported!");
1683       const MCSymbolWasm *Base = cast<MCSymbolWasm>(BS);
1684 
1685       // Find the target symbol of this weak alias and export that index
1686       const auto &WS = static_cast<const MCSymbolWasm &>(S);
1687       LLVM_DEBUG(dbgs() << WS.getName() << ": weak alias of '" << *Base
1688                         << "'\n");
1689 
1690       if (Base->isFunction()) {
1691         assert(WasmIndices.count(Base) > 0);
1692         uint32_t WasmIndex = WasmIndices.find(Base)->second;
1693         assert(WasmIndices.count(&WS) == 0);
1694         WasmIndices[&WS] = WasmIndex;
1695         LLVM_DEBUG(dbgs() << "  -> index:" << WasmIndex << "\n");
1696       } else if (Base->isData()) {
1697         auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection());
1698         uint64_t Offset = Layout.getSymbolOffset(S);
1699         int64_t Size = 0;
1700         // For data symbol alias we use the size of the base symbol as the
1701         // size of the alias.  When an offset from the base is involved this
1702         // can result in a offset + size goes past the end of the data section
1703         // which out object format doesn't support.  So we must clamp it.
1704         if (!Base->getSize()->evaluateAsAbsolute(Size, Layout))
1705           report_fatal_error(".size expression must be evaluatable");
1706         const WasmDataSegment &Segment =
1707             DataSegments[DataSection.getSegmentIndex()];
1708         Size =
1709             std::min(static_cast<uint64_t>(Size), Segment.Data.size() - Offset);
1710         wasm::WasmDataReference Ref = wasm::WasmDataReference{
1711             DataSection.getSegmentIndex(),
1712             static_cast<uint32_t>(Layout.getSymbolOffset(S)),
1713             static_cast<uint32_t>(Size)};
1714         DataLocations[&WS] = Ref;
1715         LLVM_DEBUG(dbgs() << "  -> index:" << Ref.Segment << "\n");
1716       } else {
1717         report_fatal_error("don't yet support global/tag aliases");
1718       }
1719     }
1720   }
1721 
1722   // Finally, populate the symbol table itself, in its "natural" order.
1723   for (const MCSymbol &S : Asm.symbols()) {
1724     const auto &WS = static_cast<const MCSymbolWasm &>(S);
1725     if (!isInSymtab(WS)) {
1726       WS.setIndex(InvalidIndex);
1727       continue;
1728     }
1729     LLVM_DEBUG(dbgs() << "adding to symtab: " << WS << "\n");
1730 
1731     uint32_t Flags = 0;
1732     if (WS.isWeak())
1733       Flags |= wasm::WASM_SYMBOL_BINDING_WEAK;
1734     if (WS.isHidden())
1735       Flags |= wasm::WASM_SYMBOL_VISIBILITY_HIDDEN;
1736     if (!WS.isExternal() && WS.isDefined())
1737       Flags |= wasm::WASM_SYMBOL_BINDING_LOCAL;
1738     if (WS.isUndefined())
1739       Flags |= wasm::WASM_SYMBOL_UNDEFINED;
1740     if (WS.isNoStrip()) {
1741       Flags |= wasm::WASM_SYMBOL_NO_STRIP;
1742       if (isEmscripten()) {
1743         Flags |= wasm::WASM_SYMBOL_EXPORTED;
1744       }
1745     }
1746     if (WS.hasImportName())
1747       Flags |= wasm::WASM_SYMBOL_EXPLICIT_NAME;
1748     if (WS.hasExportName())
1749       Flags |= wasm::WASM_SYMBOL_EXPORTED;
1750 
1751     wasm::WasmSymbolInfo Info;
1752     Info.Name = WS.getName();
1753     Info.Kind = WS.getType().getValueOr(wasm::WASM_SYMBOL_TYPE_DATA);
1754     Info.Flags = Flags;
1755     if (!WS.isData()) {
1756       assert(WasmIndices.count(&WS) > 0);
1757       Info.ElementIndex = WasmIndices.find(&WS)->second;
1758     } else if (WS.isDefined()) {
1759       assert(DataLocations.count(&WS) > 0);
1760       Info.DataRef = DataLocations.find(&WS)->second;
1761     }
1762     WS.setIndex(SymbolInfos.size());
1763     SymbolInfos.emplace_back(Info);
1764   }
1765 
1766   {
1767     auto HandleReloc = [&](const WasmRelocationEntry &Rel) {
1768       // Functions referenced by a relocation need to put in the table.  This is
1769       // purely to make the object file's provisional values readable, and is
1770       // ignored by the linker, which re-calculates the relocations itself.
1771       if (Rel.Type != wasm::R_WASM_TABLE_INDEX_I32 &&
1772           Rel.Type != wasm::R_WASM_TABLE_INDEX_I64 &&
1773           Rel.Type != wasm::R_WASM_TABLE_INDEX_SLEB &&
1774           Rel.Type != wasm::R_WASM_TABLE_INDEX_SLEB64 &&
1775           Rel.Type != wasm::R_WASM_TABLE_INDEX_REL_SLEB &&
1776           Rel.Type != wasm::R_WASM_TABLE_INDEX_REL_SLEB64)
1777         return;
1778       assert(Rel.Symbol->isFunction());
1779       const MCSymbolWasm *Base =
1780           cast<MCSymbolWasm>(Layout.getBaseSymbol(*Rel.Symbol));
1781       uint32_t FunctionIndex = WasmIndices.find(Base)->second;
1782       uint32_t TableIndex = TableElems.size() + InitialTableOffset;
1783       if (TableIndices.try_emplace(Base, TableIndex).second) {
1784         LLVM_DEBUG(dbgs() << "  -> adding " << Base->getName()
1785                           << " to table: " << TableIndex << "\n");
1786         TableElems.push_back(FunctionIndex);
1787         registerFunctionType(*Base);
1788       }
1789     };
1790 
1791     for (const WasmRelocationEntry &RelEntry : CodeRelocations)
1792       HandleReloc(RelEntry);
1793     for (const WasmRelocationEntry &RelEntry : DataRelocations)
1794       HandleReloc(RelEntry);
1795   }
1796 
1797   // Translate .init_array section contents into start functions.
1798   for (const MCSection &S : Asm) {
1799     const auto &WS = static_cast<const MCSectionWasm &>(S);
1800     if (WS.getName().startswith(".fini_array"))
1801       report_fatal_error(".fini_array sections are unsupported");
1802     if (!WS.getName().startswith(".init_array"))
1803       continue;
1804     if (WS.getFragmentList().empty())
1805       continue;
1806 
1807     // init_array is expected to contain a single non-empty data fragment
1808     if (WS.getFragmentList().size() != 3)
1809       report_fatal_error("only one .init_array section fragment supported");
1810 
1811     auto IT = WS.begin();
1812     const MCFragment &EmptyFrag = *IT;
1813     if (EmptyFrag.getKind() != MCFragment::FT_Data)
1814       report_fatal_error(".init_array section should be aligned");
1815 
1816     IT = std::next(IT);
1817     const MCFragment &AlignFrag = *IT;
1818     if (AlignFrag.getKind() != MCFragment::FT_Align)
1819       report_fatal_error(".init_array section should be aligned");
1820     if (cast<MCAlignFragment>(AlignFrag).getAlignment() != (is64Bit() ? 8 : 4))
1821       report_fatal_error(".init_array section should be aligned for pointers");
1822 
1823     const MCFragment &Frag = *std::next(IT);
1824     if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data)
1825       report_fatal_error("only data supported in .init_array section");
1826 
1827     uint16_t Priority = UINT16_MAX;
1828     unsigned PrefixLength = strlen(".init_array");
1829     if (WS.getName().size() > PrefixLength) {
1830       if (WS.getName()[PrefixLength] != '.')
1831         report_fatal_error(
1832             ".init_array section priority should start with '.'");
1833       if (WS.getName().substr(PrefixLength + 1).getAsInteger(10, Priority))
1834         report_fatal_error("invalid .init_array section priority");
1835     }
1836     const auto &DataFrag = cast<MCDataFragment>(Frag);
1837     const SmallVectorImpl<char> &Contents = DataFrag.getContents();
1838     for (const uint8_t *
1839              P = (const uint8_t *)Contents.data(),
1840             *End = (const uint8_t *)Contents.data() + Contents.size();
1841          P != End; ++P) {
1842       if (*P != 0)
1843         report_fatal_error("non-symbolic data in .init_array section");
1844     }
1845     for (const MCFixup &Fixup : DataFrag.getFixups()) {
1846       assert(Fixup.getKind() ==
1847              MCFixup::getKindForSize(is64Bit() ? 8 : 4, false));
1848       const MCExpr *Expr = Fixup.getValue();
1849       auto *SymRef = dyn_cast<MCSymbolRefExpr>(Expr);
1850       if (!SymRef)
1851         report_fatal_error("fixups in .init_array should be symbol references");
1852       const auto &TargetSym = cast<const MCSymbolWasm>(SymRef->getSymbol());
1853       if (TargetSym.getIndex() == InvalidIndex)
1854         report_fatal_error("symbols in .init_array should exist in symtab");
1855       if (!TargetSym.isFunction())
1856         report_fatal_error("symbols in .init_array should be for functions");
1857       InitFuncs.push_back(
1858           std::make_pair(Priority, TargetSym.getIndex()));
1859     }
1860   }
1861 
1862   // Write out the Wasm header.
1863   writeHeader(Asm);
1864 
1865   uint32_t CodeSectionIndex, DataSectionIndex;
1866   if (Mode != DwoMode::DwoOnly) {
1867     writeTypeSection(Signatures);
1868     writeImportSection(Imports, DataSize, TableElems.size());
1869     writeFunctionSection(Functions);
1870     writeTableSection(Tables);
1871     // Skip the "memory" section; we import the memory instead.
1872     writeTagSection(Tags);
1873     writeGlobalSection(Globals);
1874     writeExportSection(Exports);
1875     const MCSymbol *IndirectFunctionTable =
1876         Asm.getContext().lookupSymbol("__indirect_function_table");
1877     writeElemSection(cast_or_null<const MCSymbolWasm>(IndirectFunctionTable),
1878                      TableElems);
1879     writeDataCountSection();
1880 
1881     CodeSectionIndex = writeCodeSection(Asm, Layout, Functions);
1882     DataSectionIndex = writeDataSection(Layout);
1883   }
1884 
1885   // The Sections in the COMDAT list have placeholder indices (their index among
1886   // custom sections, rather than among all sections). Fix them up here.
1887   for (auto &Group : Comdats) {
1888     for (auto &Entry : Group.second) {
1889       if (Entry.Kind == wasm::WASM_COMDAT_SECTION) {
1890         Entry.Index += SectionCount;
1891       }
1892     }
1893   }
1894   for (auto &CustomSection : CustomSections)
1895     writeCustomSection(CustomSection, Asm, Layout);
1896 
1897   if (Mode != DwoMode::DwoOnly) {
1898     writeLinkingMetaDataSection(SymbolInfos, InitFuncs, Comdats);
1899 
1900     writeRelocSection(CodeSectionIndex, "CODE", CodeRelocations);
1901     writeRelocSection(DataSectionIndex, "DATA", DataRelocations);
1902   }
1903   writeCustomRelocSections();
1904   if (ProducersSection)
1905     writeCustomSection(*ProducersSection, Asm, Layout);
1906   if (TargetFeaturesSection)
1907     writeCustomSection(*TargetFeaturesSection, Asm, Layout);
1908 
1909   // TODO: Translate the .comment section to the output.
1910   return W->OS.tell() - StartOffset;
1911 }
1912 
1913 std::unique_ptr<MCObjectWriter>
createWasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW,raw_pwrite_stream & OS)1914 llvm::createWasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW,
1915                              raw_pwrite_stream &OS) {
1916   return std::make_unique<WasmObjectWriter>(std::move(MOTW), OS);
1917 }
1918 
1919 std::unique_ptr<MCObjectWriter>
createWasmDwoObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW,raw_pwrite_stream & OS,raw_pwrite_stream & DwoOS)1920 llvm::createWasmDwoObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW,
1921                                 raw_pwrite_stream &OS,
1922                                 raw_pwrite_stream &DwoOS) {
1923   return std::make_unique<WasmObjectWriter>(std::move(MOTW), OS, DwoOS);
1924 }
1925