xref: /netbsd-src/external/apache2/llvm/dist/llvm/lib/MC/WasmObjectWriter.cpp (revision 82d56013d7b633d116a93943de88e08335357a7c)
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__anon668be8e60111::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__anon668be8e60111::WasmRelocationEntry108   bool hasAddend() const { return wasm::relocTypeHasAddend(Type); }
109 
print__anon668be8e60111::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__anon668be8e60111::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__anon668be8e60111::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(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 writeEventSection(ArrayRef<wasm::WasmEventType> Events);
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 getEventType(const MCSymbolWasm &Symbol);
341   void registerFunctionType(const MCSymbolWasm &Symbol);
342   void registerEventType(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 = TargetObjectWriter->getRelocType(Target, Fixup, IsLocRel);
502 
503   // Absolute offset within a section or a function.
504   // Currently only supported for for metadata sections.
505   // See: test/MC/WebAssembly/blockaddress.ll
506   if (Type == wasm::R_WASM_FUNCTION_OFFSET_I32 ||
507       Type == wasm::R_WASM_FUNCTION_OFFSET_I64 ||
508       Type == wasm::R_WASM_SECTION_OFFSET_I32) {
509     if (!FixupSection.getKind().isMetadata())
510       report_fatal_error("relocations for function or section offsets are "
511                          "only supported in metadata sections");
512 
513     const MCSymbol *SectionSymbol = nullptr;
514     const MCSection &SecA = SymA->getSection();
515     if (SecA.getKind().isText()) {
516       auto SecSymIt = SectionFunctions.find(&SecA);
517       if (SecSymIt == SectionFunctions.end())
518         report_fatal_error("section doesn\'t have defining symbol");
519       SectionSymbol = SecSymIt->second;
520     } else {
521       SectionSymbol = SecA.getBeginSymbol();
522     }
523     if (!SectionSymbol)
524       report_fatal_error("section symbol is required for relocation");
525 
526     C += Layout.getSymbolOffset(*SymA);
527     SymA = cast<MCSymbolWasm>(SectionSymbol);
528   }
529 
530   if (Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB ||
531       Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB64 ||
532       Type == wasm::R_WASM_TABLE_INDEX_SLEB ||
533       Type == wasm::R_WASM_TABLE_INDEX_SLEB64 ||
534       Type == wasm::R_WASM_TABLE_INDEX_I32 ||
535       Type == wasm::R_WASM_TABLE_INDEX_I64) {
536     // TABLE_INDEX relocs implicitly use the default indirect function table.
537     // We require the function table to have already been defined.
538     auto TableName = "__indirect_function_table";
539     MCSymbolWasm *Sym = cast_or_null<MCSymbolWasm>(Ctx.lookupSymbol(TableName));
540     if (!Sym) {
541       report_fatal_error("missing indirect function table symbol");
542     } else {
543       if (!Sym->isFunctionTable())
544         report_fatal_error("__indirect_function_table symbol has wrong type");
545       // Ensure that __indirect_function_table reaches the output.
546       Sym->setNoStrip();
547       Asm.registerSymbol(*Sym);
548     }
549   }
550 
551   // Relocation other than R_WASM_TYPE_INDEX_LEB are required to be
552   // against a named symbol.
553   if (Type != wasm::R_WASM_TYPE_INDEX_LEB) {
554     if (SymA->getName().empty())
555       report_fatal_error("relocations against un-named temporaries are not yet "
556                          "supported by wasm");
557 
558     SymA->setUsedInReloc();
559   }
560 
561   if (RefA->getKind() == MCSymbolRefExpr::VK_GOT)
562     SymA->setUsedInGOT();
563 
564   WasmRelocationEntry Rec(FixupOffset, SymA, C, Type, &FixupSection);
565   LLVM_DEBUG(dbgs() << "WasmReloc: " << Rec << "\n");
566 
567   if (FixupSection.isWasmData()) {
568     DataRelocations.push_back(Rec);
569   } else if (FixupSection.getKind().isText()) {
570     CodeRelocations.push_back(Rec);
571   } else if (FixupSection.getKind().isMetadata()) {
572     CustomSectionsRelocations[&FixupSection].push_back(Rec);
573   } else {
574     llvm_unreachable("unexpected section type");
575   }
576 }
577 
578 // Compute a value to write into the code at the location covered
579 // by RelEntry. This value isn't used by the static linker; it just serves
580 // to make the object format more readable and more likely to be directly
581 // useable.
582 uint64_t
getProvisionalValue(const WasmRelocationEntry & RelEntry,const MCAsmLayout & Layout)583 WasmObjectWriter::getProvisionalValue(const WasmRelocationEntry &RelEntry,
584                                       const MCAsmLayout &Layout) {
585   if ((RelEntry.Type == wasm::R_WASM_GLOBAL_INDEX_LEB ||
586        RelEntry.Type == wasm::R_WASM_GLOBAL_INDEX_I32) &&
587       !RelEntry.Symbol->isGlobal()) {
588     assert(GOTIndices.count(RelEntry.Symbol) > 0 && "symbol not found in GOT index space");
589     return GOTIndices[RelEntry.Symbol];
590   }
591 
592   switch (RelEntry.Type) {
593   case wasm::R_WASM_TABLE_INDEX_REL_SLEB:
594   case wasm::R_WASM_TABLE_INDEX_REL_SLEB64:
595   case wasm::R_WASM_TABLE_INDEX_SLEB:
596   case wasm::R_WASM_TABLE_INDEX_SLEB64:
597   case wasm::R_WASM_TABLE_INDEX_I32:
598   case wasm::R_WASM_TABLE_INDEX_I64: {
599     // Provisional value is table address of the resolved symbol itself
600     const MCSymbolWasm *Base =
601         cast<MCSymbolWasm>(Layout.getBaseSymbol(*RelEntry.Symbol));
602     assert(Base->isFunction());
603     if (RelEntry.Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB ||
604         RelEntry.Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB64)
605       return TableIndices[Base] - InitialTableOffset;
606     else
607       return TableIndices[Base];
608   }
609   case wasm::R_WASM_TYPE_INDEX_LEB:
610     // Provisional value is same as the index
611     return getRelocationIndexValue(RelEntry);
612   case wasm::R_WASM_FUNCTION_INDEX_LEB:
613   case wasm::R_WASM_GLOBAL_INDEX_LEB:
614   case wasm::R_WASM_GLOBAL_INDEX_I32:
615   case wasm::R_WASM_EVENT_INDEX_LEB:
616   case wasm::R_WASM_TABLE_NUMBER_LEB:
617     // Provisional value is function/global/event Wasm index
618     assert(WasmIndices.count(RelEntry.Symbol) > 0 && "symbol not found in wasm index space");
619     return WasmIndices[RelEntry.Symbol];
620   case wasm::R_WASM_FUNCTION_OFFSET_I32:
621   case wasm::R_WASM_FUNCTION_OFFSET_I64:
622   case wasm::R_WASM_SECTION_OFFSET_I32: {
623     const auto &Section =
624         static_cast<const MCSectionWasm &>(RelEntry.Symbol->getSection());
625     return Section.getSectionOffset() + RelEntry.Addend;
626   }
627   case wasm::R_WASM_MEMORY_ADDR_LEB:
628   case wasm::R_WASM_MEMORY_ADDR_LEB64:
629   case wasm::R_WASM_MEMORY_ADDR_SLEB:
630   case wasm::R_WASM_MEMORY_ADDR_SLEB64:
631   case wasm::R_WASM_MEMORY_ADDR_REL_SLEB:
632   case wasm::R_WASM_MEMORY_ADDR_REL_SLEB64:
633   case wasm::R_WASM_MEMORY_ADDR_I32:
634   case wasm::R_WASM_MEMORY_ADDR_I64:
635   case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB:
636   case wasm::R_WASM_MEMORY_ADDR_LOCREL_I32: {
637     // Provisional value is address of the global plus the offset
638     // For undefined symbols, use zero
639     if (!RelEntry.Symbol->isDefined())
640       return 0;
641     const wasm::WasmDataReference &SymRef = DataLocations[RelEntry.Symbol];
642     const WasmDataSegment &Segment = DataSegments[SymRef.Segment];
643     // Ignore overflow. LLVM allows address arithmetic to silently wrap.
644     return Segment.Offset + SymRef.Offset + RelEntry.Addend;
645   }
646   default:
647     llvm_unreachable("invalid relocation type");
648   }
649 }
650 
addData(SmallVectorImpl<char> & DataBytes,MCSectionWasm & DataSection)651 static void addData(SmallVectorImpl<char> &DataBytes,
652                     MCSectionWasm &DataSection) {
653   LLVM_DEBUG(errs() << "addData: " << DataSection.getName() << "\n");
654 
655   DataBytes.resize(alignTo(DataBytes.size(), DataSection.getAlignment()));
656 
657   for (const MCFragment &Frag : DataSection) {
658     if (Frag.hasInstructions())
659       report_fatal_error("only data supported in data sections");
660 
661     if (auto *Align = dyn_cast<MCAlignFragment>(&Frag)) {
662       if (Align->getValueSize() != 1)
663         report_fatal_error("only byte values supported for alignment");
664       // If nops are requested, use zeros, as this is the data section.
665       uint8_t Value = Align->hasEmitNops() ? 0 : Align->getValue();
666       uint64_t Size =
667           std::min<uint64_t>(alignTo(DataBytes.size(), Align->getAlignment()),
668                              DataBytes.size() + Align->getMaxBytesToEmit());
669       DataBytes.resize(Size, Value);
670     } else if (auto *Fill = dyn_cast<MCFillFragment>(&Frag)) {
671       int64_t NumValues;
672       if (!Fill->getNumValues().evaluateAsAbsolute(NumValues))
673         llvm_unreachable("The fill should be an assembler constant");
674       DataBytes.insert(DataBytes.end(), Fill->getValueSize() * NumValues,
675                        Fill->getValue());
676     } else if (auto *LEB = dyn_cast<MCLEBFragment>(&Frag)) {
677       const SmallVectorImpl<char> &Contents = LEB->getContents();
678       llvm::append_range(DataBytes, Contents);
679     } else {
680       const auto &DataFrag = cast<MCDataFragment>(Frag);
681       const SmallVectorImpl<char> &Contents = DataFrag.getContents();
682       llvm::append_range(DataBytes, Contents);
683     }
684   }
685 
686   LLVM_DEBUG(dbgs() << "addData -> " << DataBytes.size() << "\n");
687 }
688 
689 uint32_t
getRelocationIndexValue(const WasmRelocationEntry & RelEntry)690 WasmObjectWriter::getRelocationIndexValue(const WasmRelocationEntry &RelEntry) {
691   if (RelEntry.Type == wasm::R_WASM_TYPE_INDEX_LEB) {
692     if (!TypeIndices.count(RelEntry.Symbol))
693       report_fatal_error("symbol not found in type index space: " +
694                          RelEntry.Symbol->getName());
695     return TypeIndices[RelEntry.Symbol];
696   }
697 
698   return RelEntry.Symbol->getIndex();
699 }
700 
701 // Apply the portions of the relocation records that we can handle ourselves
702 // directly.
applyRelocations(ArrayRef<WasmRelocationEntry> Relocations,uint64_t ContentsOffset,const MCAsmLayout & Layout)703 void WasmObjectWriter::applyRelocations(
704     ArrayRef<WasmRelocationEntry> Relocations, uint64_t ContentsOffset,
705     const MCAsmLayout &Layout) {
706   auto &Stream = static_cast<raw_pwrite_stream &>(W->OS);
707   for (const WasmRelocationEntry &RelEntry : Relocations) {
708     uint64_t Offset = ContentsOffset +
709                       RelEntry.FixupSection->getSectionOffset() +
710                       RelEntry.Offset;
711 
712     LLVM_DEBUG(dbgs() << "applyRelocation: " << RelEntry << "\n");
713     auto Value = getProvisionalValue(RelEntry, Layout);
714 
715     switch (RelEntry.Type) {
716     case wasm::R_WASM_FUNCTION_INDEX_LEB:
717     case wasm::R_WASM_TYPE_INDEX_LEB:
718     case wasm::R_WASM_GLOBAL_INDEX_LEB:
719     case wasm::R_WASM_MEMORY_ADDR_LEB:
720     case wasm::R_WASM_EVENT_INDEX_LEB:
721     case wasm::R_WASM_TABLE_NUMBER_LEB:
722       writePatchableLEB<5>(Stream, Value, Offset);
723       break;
724     case wasm::R_WASM_MEMORY_ADDR_LEB64:
725       writePatchableLEB<10>(Stream, Value, Offset);
726       break;
727     case wasm::R_WASM_TABLE_INDEX_I32:
728     case wasm::R_WASM_MEMORY_ADDR_I32:
729     case wasm::R_WASM_FUNCTION_OFFSET_I32:
730     case wasm::R_WASM_SECTION_OFFSET_I32:
731     case wasm::R_WASM_GLOBAL_INDEX_I32:
732     case wasm::R_WASM_MEMORY_ADDR_LOCREL_I32:
733       patchI32(Stream, Value, Offset);
734       break;
735     case wasm::R_WASM_TABLE_INDEX_I64:
736     case wasm::R_WASM_MEMORY_ADDR_I64:
737     case wasm::R_WASM_FUNCTION_OFFSET_I64:
738       patchI64(Stream, Value, Offset);
739       break;
740     case wasm::R_WASM_TABLE_INDEX_SLEB:
741     case wasm::R_WASM_TABLE_INDEX_REL_SLEB:
742     case wasm::R_WASM_MEMORY_ADDR_SLEB:
743     case wasm::R_WASM_MEMORY_ADDR_REL_SLEB:
744     case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB:
745       writePatchableSLEB<5>(Stream, Value, Offset);
746       break;
747     case wasm::R_WASM_TABLE_INDEX_SLEB64:
748     case wasm::R_WASM_TABLE_INDEX_REL_SLEB64:
749     case wasm::R_WASM_MEMORY_ADDR_SLEB64:
750     case wasm::R_WASM_MEMORY_ADDR_REL_SLEB64:
751       writePatchableSLEB<10>(Stream, Value, Offset);
752       break;
753     default:
754       llvm_unreachable("invalid relocation type");
755     }
756   }
757 }
758 
writeTypeSection(ArrayRef<wasm::WasmSignature> Signatures)759 void WasmObjectWriter::writeTypeSection(
760     ArrayRef<wasm::WasmSignature> Signatures) {
761   if (Signatures.empty())
762     return;
763 
764   SectionBookkeeping Section;
765   startSection(Section, wasm::WASM_SEC_TYPE);
766 
767   encodeULEB128(Signatures.size(), W->OS);
768 
769   for (const wasm::WasmSignature &Sig : Signatures) {
770     W->OS << char(wasm::WASM_TYPE_FUNC);
771     encodeULEB128(Sig.Params.size(), W->OS);
772     for (wasm::ValType Ty : Sig.Params)
773       writeValueType(Ty);
774     encodeULEB128(Sig.Returns.size(), W->OS);
775     for (wasm::ValType Ty : Sig.Returns)
776       writeValueType(Ty);
777   }
778 
779   endSection(Section);
780 }
781 
writeImportSection(ArrayRef<wasm::WasmImport> Imports,uint64_t DataSize,uint32_t NumElements)782 void WasmObjectWriter::writeImportSection(ArrayRef<wasm::WasmImport> Imports,
783                                           uint64_t DataSize,
784                                           uint32_t NumElements) {
785   if (Imports.empty())
786     return;
787 
788   uint64_t NumPages = (DataSize + wasm::WasmPageSize - 1) / wasm::WasmPageSize;
789 
790   SectionBookkeeping Section;
791   startSection(Section, wasm::WASM_SEC_IMPORT);
792 
793   encodeULEB128(Imports.size(), W->OS);
794   for (const wasm::WasmImport &Import : Imports) {
795     writeString(Import.Module);
796     writeString(Import.Field);
797     W->OS << char(Import.Kind);
798 
799     switch (Import.Kind) {
800     case wasm::WASM_EXTERNAL_FUNCTION:
801       encodeULEB128(Import.SigIndex, W->OS);
802       break;
803     case wasm::WASM_EXTERNAL_GLOBAL:
804       W->OS << char(Import.Global.Type);
805       W->OS << char(Import.Global.Mutable ? 1 : 0);
806       break;
807     case wasm::WASM_EXTERNAL_MEMORY:
808       encodeULEB128(Import.Memory.Flags, W->OS);
809       encodeULEB128(NumPages, W->OS); // initial
810       break;
811     case wasm::WASM_EXTERNAL_TABLE:
812       W->OS << char(Import.Table.ElemType);
813       encodeULEB128(0, W->OS);           // flags
814       encodeULEB128(NumElements, W->OS); // initial
815       break;
816     case wasm::WASM_EXTERNAL_EVENT:
817       encodeULEB128(Import.Event.Attribute, W->OS);
818       encodeULEB128(Import.Event.SigIndex, W->OS);
819       break;
820     default:
821       llvm_unreachable("unsupported import kind");
822     }
823   }
824 
825   endSection(Section);
826 }
827 
writeFunctionSection(ArrayRef<WasmFunction> Functions)828 void WasmObjectWriter::writeFunctionSection(ArrayRef<WasmFunction> Functions) {
829   if (Functions.empty())
830     return;
831 
832   SectionBookkeeping Section;
833   startSection(Section, wasm::WASM_SEC_FUNCTION);
834 
835   encodeULEB128(Functions.size(), W->OS);
836   for (const WasmFunction &Func : Functions)
837     encodeULEB128(Func.SigIndex, W->OS);
838 
839   endSection(Section);
840 }
841 
writeEventSection(ArrayRef<wasm::WasmEventType> Events)842 void WasmObjectWriter::writeEventSection(ArrayRef<wasm::WasmEventType> Events) {
843   if (Events.empty())
844     return;
845 
846   SectionBookkeeping Section;
847   startSection(Section, wasm::WASM_SEC_EVENT);
848 
849   encodeULEB128(Events.size(), W->OS);
850   for (const wasm::WasmEventType &Event : Events) {
851     encodeULEB128(Event.Attribute, W->OS);
852     encodeULEB128(Event.SigIndex, W->OS);
853   }
854 
855   endSection(Section);
856 }
857 
writeGlobalSection(ArrayRef<wasm::WasmGlobal> Globals)858 void WasmObjectWriter::writeGlobalSection(ArrayRef<wasm::WasmGlobal> Globals) {
859   if (Globals.empty())
860     return;
861 
862   SectionBookkeeping Section;
863   startSection(Section, wasm::WASM_SEC_GLOBAL);
864 
865   encodeULEB128(Globals.size(), W->OS);
866   for (const wasm::WasmGlobal &Global : Globals) {
867     encodeULEB128(Global.Type.Type, W->OS);
868     W->OS << char(Global.Type.Mutable);
869     W->OS << char(Global.InitExpr.Opcode);
870     switch (Global.Type.Type) {
871     case wasm::WASM_TYPE_I32:
872       encodeSLEB128(0, W->OS);
873       break;
874     case wasm::WASM_TYPE_I64:
875       encodeSLEB128(0, W->OS);
876       break;
877     case wasm::WASM_TYPE_F32:
878       writeI32(0);
879       break;
880     case wasm::WASM_TYPE_F64:
881       writeI64(0);
882       break;
883     case wasm::WASM_TYPE_EXTERNREF:
884       writeValueType(wasm::ValType::EXTERNREF);
885       break;
886     default:
887       llvm_unreachable("unexpected type");
888     }
889     W->OS << char(wasm::WASM_OPCODE_END);
890   }
891 
892   endSection(Section);
893 }
894 
writeTableSection(ArrayRef<wasm::WasmTable> Tables)895 void WasmObjectWriter::writeTableSection(ArrayRef<wasm::WasmTable> Tables) {
896   if (Tables.empty())
897     return;
898 
899   SectionBookkeeping Section;
900   startSection(Section, wasm::WASM_SEC_TABLE);
901 
902   encodeULEB128(Tables.size(), W->OS);
903   for (const wasm::WasmTable &Table : Tables) {
904     encodeULEB128(Table.Type.ElemType, W->OS);
905     encodeULEB128(Table.Type.Limits.Flags, W->OS);
906     encodeULEB128(Table.Type.Limits.Minimum, W->OS);
907     if (Table.Type.Limits.Flags & wasm::WASM_LIMITS_FLAG_HAS_MAX)
908       encodeULEB128(Table.Type.Limits.Maximum, W->OS);
909   }
910   endSection(Section);
911 }
912 
writeExportSection(ArrayRef<wasm::WasmExport> Exports)913 void WasmObjectWriter::writeExportSection(ArrayRef<wasm::WasmExport> Exports) {
914   if (Exports.empty())
915     return;
916 
917   SectionBookkeeping Section;
918   startSection(Section, wasm::WASM_SEC_EXPORT);
919 
920   encodeULEB128(Exports.size(), W->OS);
921   for (const wasm::WasmExport &Export : Exports) {
922     writeString(Export.Name);
923     W->OS << char(Export.Kind);
924     encodeULEB128(Export.Index, W->OS);
925   }
926 
927   endSection(Section);
928 }
929 
writeElemSection(const MCSymbolWasm * IndirectFunctionTable,ArrayRef<uint32_t> TableElems)930 void WasmObjectWriter::writeElemSection(
931     const MCSymbolWasm *IndirectFunctionTable, ArrayRef<uint32_t> TableElems) {
932   if (TableElems.empty())
933     return;
934 
935   assert(IndirectFunctionTable);
936 
937   SectionBookkeeping Section;
938   startSection(Section, wasm::WASM_SEC_ELEM);
939 
940   encodeULEB128(1, W->OS); // number of "segments"
941 
942   assert(WasmIndices.count(IndirectFunctionTable));
943   uint32_t TableNumber = WasmIndices.find(IndirectFunctionTable)->second;
944   uint32_t Flags = 0;
945   if (TableNumber)
946     Flags |= wasm::WASM_ELEM_SEGMENT_HAS_TABLE_NUMBER;
947   encodeULEB128(Flags, W->OS);
948   if (Flags & wasm::WASM_ELEM_SEGMENT_HAS_TABLE_NUMBER)
949     encodeULEB128(TableNumber, W->OS); // the table number
950 
951   // init expr for starting offset
952   W->OS << char(wasm::WASM_OPCODE_I32_CONST);
953   encodeSLEB128(InitialTableOffset, W->OS);
954   W->OS << char(wasm::WASM_OPCODE_END);
955 
956   if (Flags & wasm::WASM_ELEM_SEGMENT_MASK_HAS_ELEM_KIND) {
957     // We only write active function table initializers, for which the elem kind
958     // is specified to be written as 0x00 and interpreted to mean "funcref".
959     const uint8_t ElemKind = 0;
960     W->OS << ElemKind;
961   }
962 
963   encodeULEB128(TableElems.size(), W->OS);
964   for (uint32_t Elem : TableElems)
965     encodeULEB128(Elem, W->OS);
966 
967   endSection(Section);
968 }
969 
writeDataCountSection()970 void WasmObjectWriter::writeDataCountSection() {
971   if (DataSegments.empty())
972     return;
973 
974   SectionBookkeeping Section;
975   startSection(Section, wasm::WASM_SEC_DATACOUNT);
976   encodeULEB128(DataSegments.size(), W->OS);
977   endSection(Section);
978 }
979 
writeCodeSection(const MCAssembler & Asm,const MCAsmLayout & Layout,ArrayRef<WasmFunction> Functions)980 uint32_t WasmObjectWriter::writeCodeSection(const MCAssembler &Asm,
981                                             const MCAsmLayout &Layout,
982                                             ArrayRef<WasmFunction> Functions) {
983   if (Functions.empty())
984     return 0;
985 
986   SectionBookkeeping Section;
987   startSection(Section, wasm::WASM_SEC_CODE);
988 
989   encodeULEB128(Functions.size(), W->OS);
990 
991   for (const WasmFunction &Func : Functions) {
992     auto &FuncSection = static_cast<MCSectionWasm &>(Func.Sym->getSection());
993 
994     int64_t Size = 0;
995     if (!Func.Sym->getSize()->evaluateAsAbsolute(Size, Layout))
996       report_fatal_error(".size expression must be evaluatable");
997 
998     encodeULEB128(Size, W->OS);
999     FuncSection.setSectionOffset(W->OS.tell() - Section.ContentsOffset);
1000     Asm.writeSectionData(W->OS, &FuncSection, Layout);
1001   }
1002 
1003   // Apply fixups.
1004   applyRelocations(CodeRelocations, Section.ContentsOffset, Layout);
1005 
1006   endSection(Section);
1007   return Section.Index;
1008 }
1009 
writeDataSection(const MCAsmLayout & Layout)1010 uint32_t WasmObjectWriter::writeDataSection(const MCAsmLayout &Layout) {
1011   if (DataSegments.empty())
1012     return 0;
1013 
1014   SectionBookkeeping Section;
1015   startSection(Section, wasm::WASM_SEC_DATA);
1016 
1017   encodeULEB128(DataSegments.size(), W->OS); // count
1018 
1019   for (const WasmDataSegment &Segment : DataSegments) {
1020     encodeULEB128(Segment.InitFlags, W->OS); // flags
1021     if (Segment.InitFlags & wasm::WASM_DATA_SEGMENT_HAS_MEMINDEX)
1022       encodeULEB128(0, W->OS); // memory index
1023     if ((Segment.InitFlags & wasm::WASM_DATA_SEGMENT_IS_PASSIVE) == 0) {
1024       W->OS << char(is64Bit() ? wasm::WASM_OPCODE_I64_CONST
1025                               : wasm::WASM_OPCODE_I32_CONST);
1026       encodeSLEB128(Segment.Offset, W->OS); // offset
1027       W->OS << char(wasm::WASM_OPCODE_END);
1028     }
1029     encodeULEB128(Segment.Data.size(), W->OS); // size
1030     Segment.Section->setSectionOffset(W->OS.tell() - Section.ContentsOffset);
1031     W->OS << Segment.Data; // data
1032   }
1033 
1034   // Apply fixups.
1035   applyRelocations(DataRelocations, Section.ContentsOffset, Layout);
1036 
1037   endSection(Section);
1038   return Section.Index;
1039 }
1040 
writeRelocSection(uint32_t SectionIndex,StringRef Name,std::vector<WasmRelocationEntry> & Relocs)1041 void WasmObjectWriter::writeRelocSection(
1042     uint32_t SectionIndex, StringRef Name,
1043     std::vector<WasmRelocationEntry> &Relocs) {
1044   // See: https://github.com/WebAssembly/tool-conventions/blob/master/Linking.md
1045   // for descriptions of the reloc sections.
1046 
1047   if (Relocs.empty())
1048     return;
1049 
1050   // First, ensure the relocations are sorted in offset order.  In general they
1051   // should already be sorted since `recordRelocation` is called in offset
1052   // order, but for the code section we combine many MC sections into single
1053   // wasm section, and this order is determined by the order of Asm.Symbols()
1054   // not the sections order.
1055   llvm::stable_sort(
1056       Relocs, [](const WasmRelocationEntry &A, const WasmRelocationEntry &B) {
1057         return (A.Offset + A.FixupSection->getSectionOffset()) <
1058                (B.Offset + B.FixupSection->getSectionOffset());
1059       });
1060 
1061   SectionBookkeeping Section;
1062   startCustomSection(Section, std::string("reloc.") + Name.str());
1063 
1064   encodeULEB128(SectionIndex, W->OS);
1065   encodeULEB128(Relocs.size(), W->OS);
1066   for (const WasmRelocationEntry &RelEntry : Relocs) {
1067     uint64_t Offset =
1068         RelEntry.Offset + RelEntry.FixupSection->getSectionOffset();
1069     uint32_t Index = getRelocationIndexValue(RelEntry);
1070 
1071     W->OS << char(RelEntry.Type);
1072     encodeULEB128(Offset, W->OS);
1073     encodeULEB128(Index, W->OS);
1074     if (RelEntry.hasAddend())
1075       encodeSLEB128(RelEntry.Addend, W->OS);
1076   }
1077 
1078   endSection(Section);
1079 }
1080 
writeCustomRelocSections()1081 void WasmObjectWriter::writeCustomRelocSections() {
1082   for (const auto &Sec : CustomSections) {
1083     auto &Relocations = CustomSectionsRelocations[Sec.Section];
1084     writeRelocSection(Sec.OutputIndex, Sec.Name, Relocations);
1085   }
1086 }
1087 
writeLinkingMetaDataSection(ArrayRef<wasm::WasmSymbolInfo> SymbolInfos,ArrayRef<std::pair<uint16_t,uint32_t>> InitFuncs,const std::map<StringRef,std::vector<WasmComdatEntry>> & Comdats)1088 void WasmObjectWriter::writeLinkingMetaDataSection(
1089     ArrayRef<wasm::WasmSymbolInfo> SymbolInfos,
1090     ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs,
1091     const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats) {
1092   SectionBookkeeping Section;
1093   startCustomSection(Section, "linking");
1094   encodeULEB128(wasm::WasmMetadataVersion, W->OS);
1095 
1096   SectionBookkeeping SubSection;
1097   if (SymbolInfos.size() != 0) {
1098     startSection(SubSection, wasm::WASM_SYMBOL_TABLE);
1099     encodeULEB128(SymbolInfos.size(), W->OS);
1100     for (const wasm::WasmSymbolInfo &Sym : SymbolInfos) {
1101       encodeULEB128(Sym.Kind, W->OS);
1102       encodeULEB128(Sym.Flags, W->OS);
1103       switch (Sym.Kind) {
1104       case wasm::WASM_SYMBOL_TYPE_FUNCTION:
1105       case wasm::WASM_SYMBOL_TYPE_GLOBAL:
1106       case wasm::WASM_SYMBOL_TYPE_EVENT:
1107       case wasm::WASM_SYMBOL_TYPE_TABLE:
1108         encodeULEB128(Sym.ElementIndex, W->OS);
1109         if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0 ||
1110             (Sym.Flags & wasm::WASM_SYMBOL_EXPLICIT_NAME) != 0)
1111           writeString(Sym.Name);
1112         break;
1113       case wasm::WASM_SYMBOL_TYPE_DATA:
1114         writeString(Sym.Name);
1115         if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0) {
1116           encodeULEB128(Sym.DataRef.Segment, W->OS);
1117           encodeULEB128(Sym.DataRef.Offset, W->OS);
1118           encodeULEB128(Sym.DataRef.Size, W->OS);
1119         }
1120         break;
1121       case wasm::WASM_SYMBOL_TYPE_SECTION: {
1122         const uint32_t SectionIndex =
1123             CustomSections[Sym.ElementIndex].OutputIndex;
1124         encodeULEB128(SectionIndex, W->OS);
1125         break;
1126       }
1127       default:
1128         llvm_unreachable("unexpected kind");
1129       }
1130     }
1131     endSection(SubSection);
1132   }
1133 
1134   if (DataSegments.size()) {
1135     startSection(SubSection, wasm::WASM_SEGMENT_INFO);
1136     encodeULEB128(DataSegments.size(), W->OS);
1137     for (const WasmDataSegment &Segment : DataSegments) {
1138       writeString(Segment.Name);
1139       encodeULEB128(Segment.Alignment, W->OS);
1140       encodeULEB128(Segment.LinkingFlags, W->OS);
1141     }
1142     endSection(SubSection);
1143   }
1144 
1145   if (!InitFuncs.empty()) {
1146     startSection(SubSection, wasm::WASM_INIT_FUNCS);
1147     encodeULEB128(InitFuncs.size(), W->OS);
1148     for (auto &StartFunc : InitFuncs) {
1149       encodeULEB128(StartFunc.first, W->OS);  // priority
1150       encodeULEB128(StartFunc.second, W->OS); // function index
1151     }
1152     endSection(SubSection);
1153   }
1154 
1155   if (Comdats.size()) {
1156     startSection(SubSection, wasm::WASM_COMDAT_INFO);
1157     encodeULEB128(Comdats.size(), W->OS);
1158     for (const auto &C : Comdats) {
1159       writeString(C.first);
1160       encodeULEB128(0, W->OS); // flags for future use
1161       encodeULEB128(C.second.size(), W->OS);
1162       for (const WasmComdatEntry &Entry : C.second) {
1163         encodeULEB128(Entry.Kind, W->OS);
1164         encodeULEB128(Entry.Index, W->OS);
1165       }
1166     }
1167     endSection(SubSection);
1168   }
1169 
1170   endSection(Section);
1171 }
1172 
writeCustomSection(WasmCustomSection & CustomSection,const MCAssembler & Asm,const MCAsmLayout & Layout)1173 void WasmObjectWriter::writeCustomSection(WasmCustomSection &CustomSection,
1174                                           const MCAssembler &Asm,
1175                                           const MCAsmLayout &Layout) {
1176   SectionBookkeeping Section;
1177   auto *Sec = CustomSection.Section;
1178   startCustomSection(Section, CustomSection.Name);
1179 
1180   Sec->setSectionOffset(W->OS.tell() - Section.ContentsOffset);
1181   Asm.writeSectionData(W->OS, Sec, Layout);
1182 
1183   CustomSection.OutputContentsOffset = Section.ContentsOffset;
1184   CustomSection.OutputIndex = Section.Index;
1185 
1186   endSection(Section);
1187 
1188   // Apply fixups.
1189   auto &Relocations = CustomSectionsRelocations[CustomSection.Section];
1190   applyRelocations(Relocations, CustomSection.OutputContentsOffset, Layout);
1191 }
1192 
getFunctionType(const MCSymbolWasm & Symbol)1193 uint32_t WasmObjectWriter::getFunctionType(const MCSymbolWasm &Symbol) {
1194   assert(Symbol.isFunction());
1195   assert(TypeIndices.count(&Symbol));
1196   return TypeIndices[&Symbol];
1197 }
1198 
getEventType(const MCSymbolWasm & Symbol)1199 uint32_t WasmObjectWriter::getEventType(const MCSymbolWasm &Symbol) {
1200   assert(Symbol.isEvent());
1201   assert(TypeIndices.count(&Symbol));
1202   return TypeIndices[&Symbol];
1203 }
1204 
registerFunctionType(const MCSymbolWasm & Symbol)1205 void WasmObjectWriter::registerFunctionType(const MCSymbolWasm &Symbol) {
1206   assert(Symbol.isFunction());
1207 
1208   wasm::WasmSignature S;
1209 
1210   if (auto *Sig = Symbol.getSignature()) {
1211     S.Returns = Sig->Returns;
1212     S.Params = Sig->Params;
1213   }
1214 
1215   auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size()));
1216   if (Pair.second)
1217     Signatures.push_back(S);
1218   TypeIndices[&Symbol] = Pair.first->second;
1219 
1220   LLVM_DEBUG(dbgs() << "registerFunctionType: " << Symbol
1221                     << " new:" << Pair.second << "\n");
1222   LLVM_DEBUG(dbgs() << "  -> type index: " << Pair.first->second << "\n");
1223 }
1224 
registerEventType(const MCSymbolWasm & Symbol)1225 void WasmObjectWriter::registerEventType(const MCSymbolWasm &Symbol) {
1226   assert(Symbol.isEvent());
1227 
1228   // TODO Currently we don't generate imported exceptions, but if we do, we
1229   // should have a way of infering types of imported exceptions.
1230   wasm::WasmSignature S;
1231   if (auto *Sig = Symbol.getSignature()) {
1232     S.Returns = Sig->Returns;
1233     S.Params = Sig->Params;
1234   }
1235 
1236   auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size()));
1237   if (Pair.second)
1238     Signatures.push_back(S);
1239   TypeIndices[&Symbol] = Pair.first->second;
1240 
1241   LLVM_DEBUG(dbgs() << "registerEventType: " << Symbol << " new:" << Pair.second
1242                     << "\n");
1243   LLVM_DEBUG(dbgs() << "  -> type index: " << Pair.first->second << "\n");
1244 }
1245 
isInSymtab(const MCSymbolWasm & Sym)1246 static bool isInSymtab(const MCSymbolWasm &Sym) {
1247   if (Sym.isUsedInReloc() || Sym.isUsedInInitArray())
1248     return true;
1249 
1250   if (Sym.isComdat() && !Sym.isDefined())
1251     return false;
1252 
1253   if (Sym.isTemporary())
1254     return false;
1255 
1256   if (Sym.isSection())
1257     return false;
1258 
1259   if (Sym.omitFromLinkingSection())
1260     return false;
1261 
1262   return true;
1263 }
1264 
prepareImports(SmallVectorImpl<wasm::WasmImport> & Imports,MCAssembler & Asm,const MCAsmLayout & Layout)1265 void WasmObjectWriter::prepareImports(
1266     SmallVectorImpl<wasm::WasmImport> &Imports, MCAssembler &Asm,
1267     const MCAsmLayout &Layout) {
1268   // For now, always emit the memory import, since loads and stores are not
1269   // valid without it. In the future, we could perhaps be more clever and omit
1270   // it if there are no loads or stores.
1271   wasm::WasmImport MemImport;
1272   MemImport.Module = "env";
1273   MemImport.Field = "__linear_memory";
1274   MemImport.Kind = wasm::WASM_EXTERNAL_MEMORY;
1275   MemImport.Memory.Flags = is64Bit() ? wasm::WASM_LIMITS_FLAG_IS_64
1276                                      : wasm::WASM_LIMITS_FLAG_NONE;
1277   Imports.push_back(MemImport);
1278 
1279   // Populate SignatureIndices, and Imports and WasmIndices for undefined
1280   // symbols.  This must be done before populating WasmIndices for defined
1281   // symbols.
1282   for (const MCSymbol &S : Asm.symbols()) {
1283     const auto &WS = static_cast<const MCSymbolWasm &>(S);
1284 
1285     // Register types for all functions, including those with private linkage
1286     // (because wasm always needs a type signature).
1287     if (WS.isFunction()) {
1288       const auto *BS = Layout.getBaseSymbol(S);
1289       if (!BS)
1290         report_fatal_error(Twine(S.getName()) +
1291                            ": absolute addressing not supported!");
1292       registerFunctionType(*cast<MCSymbolWasm>(BS));
1293     }
1294 
1295     if (WS.isEvent())
1296       registerEventType(WS);
1297 
1298     if (WS.isTemporary())
1299       continue;
1300 
1301     // If the symbol is not defined in this translation unit, import it.
1302     if (!WS.isDefined() && !WS.isComdat()) {
1303       if (WS.isFunction()) {
1304         wasm::WasmImport Import;
1305         Import.Module = WS.getImportModule();
1306         Import.Field = WS.getImportName();
1307         Import.Kind = wasm::WASM_EXTERNAL_FUNCTION;
1308         Import.SigIndex = getFunctionType(WS);
1309         Imports.push_back(Import);
1310         assert(WasmIndices.count(&WS) == 0);
1311         WasmIndices[&WS] = NumFunctionImports++;
1312       } else if (WS.isGlobal()) {
1313         if (WS.isWeak())
1314           report_fatal_error("undefined global symbol cannot be weak");
1315 
1316         wasm::WasmImport Import;
1317         Import.Field = WS.getImportName();
1318         Import.Kind = wasm::WASM_EXTERNAL_GLOBAL;
1319         Import.Module = WS.getImportModule();
1320         Import.Global = WS.getGlobalType();
1321         Imports.push_back(Import);
1322         assert(WasmIndices.count(&WS) == 0);
1323         WasmIndices[&WS] = NumGlobalImports++;
1324       } else if (WS.isEvent()) {
1325         if (WS.isWeak())
1326           report_fatal_error("undefined event symbol cannot be weak");
1327 
1328         wasm::WasmImport Import;
1329         Import.Module = WS.getImportModule();
1330         Import.Field = WS.getImportName();
1331         Import.Kind = wasm::WASM_EXTERNAL_EVENT;
1332         Import.Event.Attribute = wasm::WASM_EVENT_ATTRIBUTE_EXCEPTION;
1333         Import.Event.SigIndex = getEventType(WS);
1334         Imports.push_back(Import);
1335         assert(WasmIndices.count(&WS) == 0);
1336         WasmIndices[&WS] = NumEventImports++;
1337       } else if (WS.isTable()) {
1338         if (WS.isWeak())
1339           report_fatal_error("undefined table symbol cannot be weak");
1340 
1341         wasm::WasmImport Import;
1342         Import.Module = WS.getImportModule();
1343         Import.Field = WS.getImportName();
1344         Import.Kind = wasm::WASM_EXTERNAL_TABLE;
1345         Import.Table = WS.getTableType();
1346         Imports.push_back(Import);
1347         assert(WasmIndices.count(&WS) == 0);
1348         WasmIndices[&WS] = NumTableImports++;
1349       }
1350     }
1351   }
1352 
1353   // Add imports for GOT globals
1354   for (const MCSymbol &S : Asm.symbols()) {
1355     const auto &WS = static_cast<const MCSymbolWasm &>(S);
1356     if (WS.isUsedInGOT()) {
1357       wasm::WasmImport Import;
1358       if (WS.isFunction())
1359         Import.Module = "GOT.func";
1360       else
1361         Import.Module = "GOT.mem";
1362       Import.Field = WS.getName();
1363       Import.Kind = wasm::WASM_EXTERNAL_GLOBAL;
1364       Import.Global = {wasm::WASM_TYPE_I32, true};
1365       Imports.push_back(Import);
1366       assert(GOTIndices.count(&WS) == 0);
1367       GOTIndices[&WS] = NumGlobalImports++;
1368     }
1369   }
1370 }
1371 
writeObject(MCAssembler & Asm,const MCAsmLayout & Layout)1372 uint64_t WasmObjectWriter::writeObject(MCAssembler &Asm,
1373                                        const MCAsmLayout &Layout) {
1374   support::endian::Writer MainWriter(*OS, support::little);
1375   W = &MainWriter;
1376   if (IsSplitDwarf) {
1377     uint64_t TotalSize = writeOneObject(Asm, Layout, DwoMode::NonDwoOnly);
1378     assert(DwoOS);
1379     support::endian::Writer DwoWriter(*DwoOS, support::little);
1380     W = &DwoWriter;
1381     return TotalSize + writeOneObject(Asm, Layout, DwoMode::DwoOnly);
1382   } else {
1383     return writeOneObject(Asm, Layout, DwoMode::AllSections);
1384   }
1385 }
1386 
writeOneObject(MCAssembler & Asm,const MCAsmLayout & Layout,DwoMode Mode)1387 uint64_t WasmObjectWriter::writeOneObject(MCAssembler &Asm,
1388                                           const MCAsmLayout &Layout,
1389                                           DwoMode Mode) {
1390   uint64_t StartOffset = W->OS.tell();
1391   SectionCount = 0;
1392   CustomSections.clear();
1393 
1394   LLVM_DEBUG(dbgs() << "WasmObjectWriter::writeObject\n");
1395 
1396   // Collect information from the available symbols.
1397   SmallVector<WasmFunction, 4> Functions;
1398   SmallVector<uint32_t, 4> TableElems;
1399   SmallVector<wasm::WasmImport, 4> Imports;
1400   SmallVector<wasm::WasmExport, 4> Exports;
1401   SmallVector<wasm::WasmEventType, 1> Events;
1402   SmallVector<wasm::WasmGlobal, 1> Globals;
1403   SmallVector<wasm::WasmTable, 1> Tables;
1404   SmallVector<wasm::WasmSymbolInfo, 4> SymbolInfos;
1405   SmallVector<std::pair<uint16_t, uint32_t>, 2> InitFuncs;
1406   std::map<StringRef, std::vector<WasmComdatEntry>> Comdats;
1407   uint64_t DataSize = 0;
1408   if (Mode != DwoMode::DwoOnly) {
1409     prepareImports(Imports, Asm, Layout);
1410   }
1411 
1412   // Populate DataSegments and CustomSections, which must be done before
1413   // populating DataLocations.
1414   for (MCSection &Sec : Asm) {
1415     auto &Section = static_cast<MCSectionWasm &>(Sec);
1416     StringRef SectionName = Section.getName();
1417 
1418     if (Mode == DwoMode::NonDwoOnly && isDwoSection(Sec))
1419       continue;
1420     if (Mode == DwoMode::DwoOnly && !isDwoSection(Sec))
1421       continue;
1422 
1423     LLVM_DEBUG(dbgs() << "Processing Section " << SectionName << "  group "
1424                       << Section.getGroup() << "\n";);
1425 
1426     // .init_array sections are handled specially elsewhere.
1427     if (SectionName.startswith(".init_array"))
1428       continue;
1429 
1430     // Code is handled separately
1431     if (Section.getKind().isText())
1432       continue;
1433 
1434     if (Section.isWasmData()) {
1435       uint32_t SegmentIndex = DataSegments.size();
1436       DataSize = alignTo(DataSize, Section.getAlignment());
1437       DataSegments.emplace_back();
1438       WasmDataSegment &Segment = DataSegments.back();
1439       Segment.Name = SectionName;
1440       Segment.InitFlags = Section.getPassive()
1441                               ? (uint32_t)wasm::WASM_DATA_SEGMENT_IS_PASSIVE
1442                               : 0;
1443       Segment.Offset = DataSize;
1444       Segment.Section = &Section;
1445       addData(Segment.Data, Section);
1446       Segment.Alignment = Log2_32(Section.getAlignment());
1447       Segment.LinkingFlags = Section.getSegmentFlags();
1448       DataSize += Segment.Data.size();
1449       Section.setSegmentIndex(SegmentIndex);
1450 
1451       if (const MCSymbolWasm *C = Section.getGroup()) {
1452         Comdats[C->getName()].emplace_back(
1453             WasmComdatEntry{wasm::WASM_COMDAT_DATA, SegmentIndex});
1454       }
1455     } else {
1456       // Create custom sections
1457       assert(Sec.getKind().isMetadata());
1458 
1459       StringRef Name = SectionName;
1460 
1461       // For user-defined custom sections, strip the prefix
1462       if (Name.startswith(".custom_section."))
1463         Name = Name.substr(strlen(".custom_section."));
1464 
1465       MCSymbol *Begin = Sec.getBeginSymbol();
1466       if (Begin) {
1467         assert(WasmIndices.count(cast<MCSymbolWasm>(Begin)) == 0);
1468         WasmIndices[cast<MCSymbolWasm>(Begin)] = CustomSections.size();
1469       }
1470 
1471       // Separate out the producers and target features sections
1472       if (Name == "producers") {
1473         ProducersSection = std::make_unique<WasmCustomSection>(Name, &Section);
1474         continue;
1475       }
1476       if (Name == "target_features") {
1477         TargetFeaturesSection =
1478             std::make_unique<WasmCustomSection>(Name, &Section);
1479         continue;
1480       }
1481 
1482       // Custom sections can also belong to COMDAT groups. In this case the
1483       // decriptor's "index" field is the section index (in the final object
1484       // file), but that is not known until after layout, so it must be fixed up
1485       // later
1486       if (const MCSymbolWasm *C = Section.getGroup()) {
1487         Comdats[C->getName()].emplace_back(
1488             WasmComdatEntry{wasm::WASM_COMDAT_SECTION,
1489                             static_cast<uint32_t>(CustomSections.size())});
1490       }
1491 
1492       CustomSections.emplace_back(Name, &Section);
1493     }
1494   }
1495 
1496   if (Mode != DwoMode::DwoOnly) {
1497     // Populate WasmIndices and DataLocations for defined symbols.
1498     for (const MCSymbol &S : Asm.symbols()) {
1499       // Ignore unnamed temporary symbols, which aren't ever exported, imported,
1500       // or used in relocations.
1501       if (S.isTemporary() && S.getName().empty())
1502         continue;
1503 
1504       const auto &WS = static_cast<const MCSymbolWasm &>(S);
1505       LLVM_DEBUG(dbgs()
1506                  << "MCSymbol: "
1507                  << toString(WS.getType().getValueOr(wasm::WASM_SYMBOL_TYPE_DATA))
1508                  << " '" << S << "'"
1509                  << " isDefined=" << S.isDefined() << " isExternal="
1510                  << S.isExternal() << " isTemporary=" << S.isTemporary()
1511                  << " isWeak=" << WS.isWeak() << " isHidden=" << WS.isHidden()
1512                  << " isVariable=" << WS.isVariable() << "\n");
1513 
1514       if (WS.isVariable())
1515         continue;
1516       if (WS.isComdat() && !WS.isDefined())
1517         continue;
1518 
1519       if (WS.isFunction()) {
1520         unsigned Index;
1521         if (WS.isDefined()) {
1522           if (WS.getOffset() != 0)
1523             report_fatal_error(
1524                 "function sections must contain one function each");
1525 
1526           if (WS.getSize() == nullptr)
1527             report_fatal_error(
1528                 "function symbols must have a size set with .size");
1529 
1530           // A definition. Write out the function body.
1531           Index = NumFunctionImports + Functions.size();
1532           WasmFunction Func;
1533           Func.SigIndex = getFunctionType(WS);
1534           Func.Sym = &WS;
1535           assert(WasmIndices.count(&WS) == 0);
1536           WasmIndices[&WS] = Index;
1537           Functions.push_back(Func);
1538 
1539           auto &Section = static_cast<MCSectionWasm &>(WS.getSection());
1540           if (const MCSymbolWasm *C = Section.getGroup()) {
1541             Comdats[C->getName()].emplace_back(
1542                 WasmComdatEntry{wasm::WASM_COMDAT_FUNCTION, Index});
1543           }
1544 
1545           if (WS.hasExportName()) {
1546             wasm::WasmExport Export;
1547             Export.Name = WS.getExportName();
1548             Export.Kind = wasm::WASM_EXTERNAL_FUNCTION;
1549             Export.Index = Index;
1550             Exports.push_back(Export);
1551           }
1552         } else {
1553           // An import; the index was assigned above.
1554           Index = WasmIndices.find(&WS)->second;
1555         }
1556 
1557         LLVM_DEBUG(dbgs() << "  -> function index: " << Index << "\n");
1558 
1559       } else if (WS.isData()) {
1560         if (!isInSymtab(WS))
1561           continue;
1562 
1563         if (!WS.isDefined()) {
1564           LLVM_DEBUG(dbgs() << "  -> segment index: -1"
1565                             << "\n");
1566           continue;
1567         }
1568 
1569         if (!WS.getSize())
1570           report_fatal_error("data symbols must have a size set with .size: " +
1571                              WS.getName());
1572 
1573         int64_t Size = 0;
1574         if (!WS.getSize()->evaluateAsAbsolute(Size, Layout))
1575           report_fatal_error(".size expression must be evaluatable");
1576 
1577         auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection());
1578         if (!DataSection.isWasmData())
1579           report_fatal_error("data symbols must live in a data section: " +
1580                              WS.getName());
1581 
1582         // For each data symbol, export it in the symtab as a reference to the
1583         // corresponding Wasm data segment.
1584         wasm::WasmDataReference Ref = wasm::WasmDataReference{
1585             DataSection.getSegmentIndex(), Layout.getSymbolOffset(WS),
1586             static_cast<uint64_t>(Size)};
1587         assert(DataLocations.count(&WS) == 0);
1588         DataLocations[&WS] = Ref;
1589         LLVM_DEBUG(dbgs() << "  -> segment index: " << Ref.Segment << "\n");
1590 
1591       } else if (WS.isGlobal()) {
1592         // A "true" Wasm global (currently just __stack_pointer)
1593         if (WS.isDefined()) {
1594           wasm::WasmGlobal Global;
1595           Global.Type = WS.getGlobalType();
1596           Global.Index = NumGlobalImports + Globals.size();
1597           switch (Global.Type.Type) {
1598           case wasm::WASM_TYPE_I32:
1599             Global.InitExpr.Opcode = wasm::WASM_OPCODE_I32_CONST;
1600             break;
1601           case wasm::WASM_TYPE_I64:
1602             Global.InitExpr.Opcode = wasm::WASM_OPCODE_I64_CONST;
1603             break;
1604           case wasm::WASM_TYPE_F32:
1605             Global.InitExpr.Opcode = wasm::WASM_OPCODE_F32_CONST;
1606             break;
1607           case wasm::WASM_TYPE_F64:
1608             Global.InitExpr.Opcode = wasm::WASM_OPCODE_F64_CONST;
1609             break;
1610           case wasm::WASM_TYPE_EXTERNREF:
1611             Global.InitExpr.Opcode = wasm::WASM_OPCODE_REF_NULL;
1612             break;
1613           default:
1614             llvm_unreachable("unexpected type");
1615           }
1616           assert(WasmIndices.count(&WS) == 0);
1617           WasmIndices[&WS] = Global.Index;
1618           Globals.push_back(Global);
1619         } else {
1620           // An import; the index was assigned above
1621           LLVM_DEBUG(dbgs() << "  -> global index: "
1622                             << WasmIndices.find(&WS)->second << "\n");
1623         }
1624       } else if (WS.isTable()) {
1625         if (WS.isDefined()) {
1626           wasm::WasmTable Table;
1627           Table.Index = NumTableImports + Tables.size();
1628           Table.Type = WS.getTableType();
1629           assert(WasmIndices.count(&WS) == 0);
1630           WasmIndices[&WS] = Table.Index;
1631           Tables.push_back(Table);
1632         }
1633         LLVM_DEBUG(dbgs() << " -> table index: "
1634                           << WasmIndices.find(&WS)->second << "\n");
1635       } else if (WS.isEvent()) {
1636         // C++ exception symbol (__cpp_exception)
1637         unsigned Index;
1638         if (WS.isDefined()) {
1639           Index = NumEventImports + Events.size();
1640           wasm::WasmEventType Event;
1641           Event.SigIndex = getEventType(WS);
1642           Event.Attribute = wasm::WASM_EVENT_ATTRIBUTE_EXCEPTION;
1643           assert(WasmIndices.count(&WS) == 0);
1644           WasmIndices[&WS] = Index;
1645           Events.push_back(Event);
1646         } else {
1647           // An import; the index was assigned above.
1648           assert(WasmIndices.count(&WS) > 0);
1649         }
1650         LLVM_DEBUG(dbgs() << "  -> event index: "
1651                           << WasmIndices.find(&WS)->second << "\n");
1652 
1653       } else {
1654         assert(WS.isSection());
1655       }
1656     }
1657 
1658     // Populate WasmIndices and DataLocations for aliased symbols.  We need to
1659     // process these in a separate pass because we need to have processed the
1660     // target of the alias before the alias itself and the symbols are not
1661     // necessarily ordered in this way.
1662     for (const MCSymbol &S : Asm.symbols()) {
1663       if (!S.isVariable())
1664         continue;
1665 
1666       assert(S.isDefined());
1667 
1668       const auto *BS = Layout.getBaseSymbol(S);
1669       if (!BS)
1670         report_fatal_error(Twine(S.getName()) +
1671                            ": absolute addressing not supported!");
1672       const MCSymbolWasm *Base = cast<MCSymbolWasm>(BS);
1673 
1674       // Find the target symbol of this weak alias and export that index
1675       const auto &WS = static_cast<const MCSymbolWasm &>(S);
1676       LLVM_DEBUG(dbgs() << WS.getName() << ": weak alias of '" << *Base
1677                         << "'\n");
1678 
1679       if (Base->isFunction()) {
1680         assert(WasmIndices.count(Base) > 0);
1681         uint32_t WasmIndex = WasmIndices.find(Base)->second;
1682         assert(WasmIndices.count(&WS) == 0);
1683         WasmIndices[&WS] = WasmIndex;
1684         LLVM_DEBUG(dbgs() << "  -> index:" << WasmIndex << "\n");
1685       } else if (Base->isData()) {
1686         auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection());
1687         uint64_t Offset = Layout.getSymbolOffset(S);
1688         int64_t Size = 0;
1689         // For data symbol alias we use the size of the base symbol as the
1690         // size of the alias.  When an offset from the base is involved this
1691         // can result in a offset + size goes past the end of the data section
1692         // which out object format doesn't support.  So we must clamp it.
1693         if (!Base->getSize()->evaluateAsAbsolute(Size, Layout))
1694           report_fatal_error(".size expression must be evaluatable");
1695         const WasmDataSegment &Segment =
1696             DataSegments[DataSection.getSegmentIndex()];
1697         Size =
1698             std::min(static_cast<uint64_t>(Size), Segment.Data.size() - Offset);
1699         wasm::WasmDataReference Ref = wasm::WasmDataReference{
1700             DataSection.getSegmentIndex(),
1701             static_cast<uint32_t>(Layout.getSymbolOffset(S)),
1702             static_cast<uint32_t>(Size)};
1703         DataLocations[&WS] = Ref;
1704         LLVM_DEBUG(dbgs() << "  -> index:" << Ref.Segment << "\n");
1705       } else {
1706         report_fatal_error("don't yet support global/event aliases");
1707       }
1708     }
1709   }
1710 
1711   // Finally, populate the symbol table itself, in its "natural" order.
1712   for (const MCSymbol &S : Asm.symbols()) {
1713     const auto &WS = static_cast<const MCSymbolWasm &>(S);
1714     if (!isInSymtab(WS)) {
1715       WS.setIndex(InvalidIndex);
1716       continue;
1717     }
1718     LLVM_DEBUG(dbgs() << "adding to symtab: " << WS << "\n");
1719 
1720     uint32_t Flags = 0;
1721     if (WS.isWeak())
1722       Flags |= wasm::WASM_SYMBOL_BINDING_WEAK;
1723     if (WS.isHidden())
1724       Flags |= wasm::WASM_SYMBOL_VISIBILITY_HIDDEN;
1725     if (!WS.isExternal() && WS.isDefined())
1726       Flags |= wasm::WASM_SYMBOL_BINDING_LOCAL;
1727     if (WS.isUndefined())
1728       Flags |= wasm::WASM_SYMBOL_UNDEFINED;
1729     if (WS.isNoStrip()) {
1730       Flags |= wasm::WASM_SYMBOL_NO_STRIP;
1731       if (isEmscripten()) {
1732         Flags |= wasm::WASM_SYMBOL_EXPORTED;
1733       }
1734     }
1735     if (WS.hasImportName())
1736       Flags |= wasm::WASM_SYMBOL_EXPLICIT_NAME;
1737     if (WS.hasExportName())
1738       Flags |= wasm::WASM_SYMBOL_EXPORTED;
1739 
1740     wasm::WasmSymbolInfo Info;
1741     Info.Name = WS.getName();
1742     Info.Kind = WS.getType().getValueOr(wasm::WASM_SYMBOL_TYPE_DATA);
1743     Info.Flags = Flags;
1744     if (!WS.isData()) {
1745       assert(WasmIndices.count(&WS) > 0);
1746       Info.ElementIndex = WasmIndices.find(&WS)->second;
1747     } else if (WS.isDefined()) {
1748       assert(DataLocations.count(&WS) > 0);
1749       Info.DataRef = DataLocations.find(&WS)->second;
1750     }
1751     WS.setIndex(SymbolInfos.size());
1752     SymbolInfos.emplace_back(Info);
1753   }
1754 
1755   {
1756     auto HandleReloc = [&](const WasmRelocationEntry &Rel) {
1757       // Functions referenced by a relocation need to put in the table.  This is
1758       // purely to make the object file's provisional values readable, and is
1759       // ignored by the linker, which re-calculates the relocations itself.
1760       if (Rel.Type != wasm::R_WASM_TABLE_INDEX_I32 &&
1761           Rel.Type != wasm::R_WASM_TABLE_INDEX_I64 &&
1762           Rel.Type != wasm::R_WASM_TABLE_INDEX_SLEB &&
1763           Rel.Type != wasm::R_WASM_TABLE_INDEX_SLEB64 &&
1764           Rel.Type != wasm::R_WASM_TABLE_INDEX_REL_SLEB &&
1765           Rel.Type != wasm::R_WASM_TABLE_INDEX_REL_SLEB64)
1766         return;
1767       assert(Rel.Symbol->isFunction());
1768       const MCSymbolWasm *Base =
1769           cast<MCSymbolWasm>(Layout.getBaseSymbol(*Rel.Symbol));
1770       uint32_t FunctionIndex = WasmIndices.find(Base)->second;
1771       uint32_t TableIndex = TableElems.size() + InitialTableOffset;
1772       if (TableIndices.try_emplace(Base, TableIndex).second) {
1773         LLVM_DEBUG(dbgs() << "  -> adding " << Base->getName()
1774                           << " to table: " << TableIndex << "\n");
1775         TableElems.push_back(FunctionIndex);
1776         registerFunctionType(*Base);
1777       }
1778     };
1779 
1780     for (const WasmRelocationEntry &RelEntry : CodeRelocations)
1781       HandleReloc(RelEntry);
1782     for (const WasmRelocationEntry &RelEntry : DataRelocations)
1783       HandleReloc(RelEntry);
1784   }
1785 
1786   // Translate .init_array section contents into start functions.
1787   for (const MCSection &S : Asm) {
1788     const auto &WS = static_cast<const MCSectionWasm &>(S);
1789     if (WS.getName().startswith(".fini_array"))
1790       report_fatal_error(".fini_array sections are unsupported");
1791     if (!WS.getName().startswith(".init_array"))
1792       continue;
1793     if (WS.getFragmentList().empty())
1794       continue;
1795 
1796     // init_array is expected to contain a single non-empty data fragment
1797     if (WS.getFragmentList().size() != 3)
1798       report_fatal_error("only one .init_array section fragment supported");
1799 
1800     auto IT = WS.begin();
1801     const MCFragment &EmptyFrag = *IT;
1802     if (EmptyFrag.getKind() != MCFragment::FT_Data)
1803       report_fatal_error(".init_array section should be aligned");
1804 
1805     IT = std::next(IT);
1806     const MCFragment &AlignFrag = *IT;
1807     if (AlignFrag.getKind() != MCFragment::FT_Align)
1808       report_fatal_error(".init_array section should be aligned");
1809     if (cast<MCAlignFragment>(AlignFrag).getAlignment() != (is64Bit() ? 8 : 4))
1810       report_fatal_error(".init_array section should be aligned for pointers");
1811 
1812     const MCFragment &Frag = *std::next(IT);
1813     if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data)
1814       report_fatal_error("only data supported in .init_array section");
1815 
1816     uint16_t Priority = UINT16_MAX;
1817     unsigned PrefixLength = strlen(".init_array");
1818     if (WS.getName().size() > PrefixLength) {
1819       if (WS.getName()[PrefixLength] != '.')
1820         report_fatal_error(
1821             ".init_array section priority should start with '.'");
1822       if (WS.getName().substr(PrefixLength + 1).getAsInteger(10, Priority))
1823         report_fatal_error("invalid .init_array section priority");
1824     }
1825     const auto &DataFrag = cast<MCDataFragment>(Frag);
1826     const SmallVectorImpl<char> &Contents = DataFrag.getContents();
1827     for (const uint8_t *
1828              P = (const uint8_t *)Contents.data(),
1829             *End = (const uint8_t *)Contents.data() + Contents.size();
1830          P != End; ++P) {
1831       if (*P != 0)
1832         report_fatal_error("non-symbolic data in .init_array section");
1833     }
1834     for (const MCFixup &Fixup : DataFrag.getFixups()) {
1835       assert(Fixup.getKind() ==
1836              MCFixup::getKindForSize(is64Bit() ? 8 : 4, false));
1837       const MCExpr *Expr = Fixup.getValue();
1838       auto *SymRef = dyn_cast<MCSymbolRefExpr>(Expr);
1839       if (!SymRef)
1840         report_fatal_error("fixups in .init_array should be symbol references");
1841       const auto &TargetSym = cast<const MCSymbolWasm>(SymRef->getSymbol());
1842       if (TargetSym.getIndex() == InvalidIndex)
1843         report_fatal_error("symbols in .init_array should exist in symtab");
1844       if (!TargetSym.isFunction())
1845         report_fatal_error("symbols in .init_array should be for functions");
1846       InitFuncs.push_back(
1847           std::make_pair(Priority, TargetSym.getIndex()));
1848     }
1849   }
1850 
1851   // Write out the Wasm header.
1852   writeHeader(Asm);
1853 
1854   uint32_t CodeSectionIndex, DataSectionIndex;
1855   if (Mode != DwoMode::DwoOnly) {
1856     writeTypeSection(Signatures);
1857     writeImportSection(Imports, DataSize, TableElems.size());
1858     writeFunctionSection(Functions);
1859     writeTableSection(Tables);
1860     // Skip the "memory" section; we import the memory instead.
1861     writeEventSection(Events);
1862     writeGlobalSection(Globals);
1863     writeExportSection(Exports);
1864     const MCSymbol *IndirectFunctionTable =
1865         Asm.getContext().lookupSymbol("__indirect_function_table");
1866     writeElemSection(cast_or_null<const MCSymbolWasm>(IndirectFunctionTable),
1867                      TableElems);
1868     writeDataCountSection();
1869 
1870     CodeSectionIndex = writeCodeSection(Asm, Layout, Functions);
1871     DataSectionIndex = writeDataSection(Layout);
1872   }
1873 
1874   // The Sections in the COMDAT list have placeholder indices (their index among
1875   // custom sections, rather than among all sections). Fix them up here.
1876   for (auto &Group : Comdats) {
1877     for (auto &Entry : Group.second) {
1878       if (Entry.Kind == wasm::WASM_COMDAT_SECTION) {
1879         Entry.Index += SectionCount;
1880       }
1881     }
1882   }
1883   for (auto &CustomSection : CustomSections)
1884     writeCustomSection(CustomSection, Asm, Layout);
1885 
1886   if (Mode != DwoMode::DwoOnly) {
1887     writeLinkingMetaDataSection(SymbolInfos, InitFuncs, Comdats);
1888 
1889     writeRelocSection(CodeSectionIndex, "CODE", CodeRelocations);
1890     writeRelocSection(DataSectionIndex, "DATA", DataRelocations);
1891   }
1892   writeCustomRelocSections();
1893   if (ProducersSection)
1894     writeCustomSection(*ProducersSection, Asm, Layout);
1895   if (TargetFeaturesSection)
1896     writeCustomSection(*TargetFeaturesSection, Asm, Layout);
1897 
1898   // TODO: Translate the .comment section to the output.
1899   return W->OS.tell() - StartOffset;
1900 }
1901 
1902 std::unique_ptr<MCObjectWriter>
createWasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW,raw_pwrite_stream & OS)1903 llvm::createWasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW,
1904                              raw_pwrite_stream &OS) {
1905   return std::make_unique<WasmObjectWriter>(std::move(MOTW), OS);
1906 }
1907 
1908 std::unique_ptr<MCObjectWriter>
createWasmDwoObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW,raw_pwrite_stream & OS,raw_pwrite_stream & DwoOS)1909 llvm::createWasmDwoObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW,
1910                                 raw_pwrite_stream &OS,
1911                                 raw_pwrite_stream &DwoOS) {
1912   return std::make_unique<WasmObjectWriter>(std::move(MOTW), OS, DwoOS);
1913 }
1914