1 //===- lib/MC/WasmObjectWriter.cpp - Wasm File Writer ---------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements Wasm object file writer information.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "llvm/ADT/STLExtras.h"
15 #include "llvm/ADT/SmallPtrSet.h"
16 #include "llvm/BinaryFormat/Wasm.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/ErrorHandling.h"
32 #include "llvm/Support/LEB128.h"
33 #include "llvm/Support/StringSaver.h"
34 #include <vector>
35
36 using namespace llvm;
37
38 #define DEBUG_TYPE "mc"
39
40 namespace {
41
42 // Went we ceate the indirect function table we start at 1, so that there is
43 // and emtpy slot at 0 and therefore calling a null function pointer will trap.
44 static const uint32_t kInitialTableOffset = 1;
45
46 // For patching purposes, we need to remember where each section starts, both
47 // for patching up the section size field, and for patching up references to
48 // locations within the section.
49 struct SectionBookkeeping {
50 // Where the size of the section is written.
51 uint64_t SizeOffset;
52 // Where the section header ends (without custom section name).
53 uint64_t PayloadOffset;
54 // Where the contents of the section starts.
55 uint64_t ContentsOffset;
56 uint32_t Index;
57 };
58
59 // The signature of a wasm function, in a struct capable of being used as a
60 // DenseMap key.
61 struct WasmFunctionType {
62 // Support empty and tombstone instances, needed by DenseMap.
63 enum { Plain, Empty, Tombstone } State;
64
65 // The return types of the function.
66 SmallVector<wasm::ValType, 1> Returns;
67
68 // The parameter types of the function.
69 SmallVector<wasm::ValType, 4> Params;
70
WasmFunctionType__anone61d87550111::WasmFunctionType71 WasmFunctionType() : State(Plain) {}
72
operator ==__anone61d87550111::WasmFunctionType73 bool operator==(const WasmFunctionType &Other) const {
74 return State == Other.State && Returns == Other.Returns &&
75 Params == Other.Params;
76 }
77 };
78
79 // Traits for using WasmFunctionType in a DenseMap.
80 struct WasmFunctionTypeDenseMapInfo {
getEmptyKey__anone61d87550111::WasmFunctionTypeDenseMapInfo81 static WasmFunctionType getEmptyKey() {
82 WasmFunctionType FuncTy;
83 FuncTy.State = WasmFunctionType::Empty;
84 return FuncTy;
85 }
getTombstoneKey__anone61d87550111::WasmFunctionTypeDenseMapInfo86 static WasmFunctionType getTombstoneKey() {
87 WasmFunctionType FuncTy;
88 FuncTy.State = WasmFunctionType::Tombstone;
89 return FuncTy;
90 }
getHashValue__anone61d87550111::WasmFunctionTypeDenseMapInfo91 static unsigned getHashValue(const WasmFunctionType &FuncTy) {
92 uintptr_t Value = FuncTy.State;
93 for (wasm::ValType Ret : FuncTy.Returns)
94 Value += DenseMapInfo<int32_t>::getHashValue(int32_t(Ret));
95 for (wasm::ValType Param : FuncTy.Params)
96 Value += DenseMapInfo<int32_t>::getHashValue(int32_t(Param));
97 return Value;
98 }
isEqual__anone61d87550111::WasmFunctionTypeDenseMapInfo99 static bool isEqual(const WasmFunctionType &LHS,
100 const WasmFunctionType &RHS) {
101 return LHS == RHS;
102 }
103 };
104
105 // A wasm data segment. A wasm binary contains only a single data section
106 // but that can contain many segments, each with their own virtual location
107 // in memory. Each MCSection data created by llvm is modeled as its own
108 // wasm data segment.
109 struct WasmDataSegment {
110 MCSectionWasm *Section;
111 StringRef Name;
112 uint32_t Offset;
113 uint32_t Alignment;
114 uint32_t Flags;
115 SmallVector<char, 4> Data;
116 };
117
118 // A wasm function to be written into the function section.
119 struct WasmFunction {
120 int32_t Type;
121 const MCSymbolWasm *Sym;
122 };
123
124 // A wasm global to be written into the global section.
125 struct WasmGlobal {
126 wasm::WasmGlobalType Type;
127 uint64_t InitialValue;
128 };
129
130 // Information about a single item which is part of a COMDAT. For each data
131 // segment or function which is in the COMDAT, there is a corresponding
132 // WasmComdatEntry.
133 struct WasmComdatEntry {
134 unsigned Kind;
135 uint32_t Index;
136 };
137
138 // Information about a single relocation.
139 struct WasmRelocationEntry {
140 uint64_t Offset; // Where is the relocation.
141 const MCSymbolWasm *Symbol; // The symbol to relocate with.
142 int64_t Addend; // A value to add to the symbol.
143 unsigned Type; // The type of the relocation.
144 const MCSectionWasm *FixupSection;// The section the relocation is targeting.
145
WasmRelocationEntry__anone61d87550111::WasmRelocationEntry146 WasmRelocationEntry(uint64_t Offset, const MCSymbolWasm *Symbol,
147 int64_t Addend, unsigned Type,
148 const MCSectionWasm *FixupSection)
149 : Offset(Offset), Symbol(Symbol), Addend(Addend), Type(Type),
150 FixupSection(FixupSection) {}
151
hasAddend__anone61d87550111::WasmRelocationEntry152 bool hasAddend() const {
153 switch (Type) {
154 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_LEB:
155 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_SLEB:
156 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_I32:
157 case wasm::R_WEBASSEMBLY_FUNCTION_OFFSET_I32:
158 case wasm::R_WEBASSEMBLY_SECTION_OFFSET_I32:
159 return true;
160 default:
161 return false;
162 }
163 }
164
print__anone61d87550111::WasmRelocationEntry165 void print(raw_ostream &Out) const {
166 Out << wasm::relocTypetoString(Type)
167 << " Off=" << Offset << ", Sym=" << *Symbol << ", Addend=" << Addend
168 << ", FixupSection=" << FixupSection->getSectionName();
169 }
170
171 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
dump__anone61d87550111::WasmRelocationEntry172 LLVM_DUMP_METHOD void dump() const { print(dbgs()); }
173 #endif
174 };
175
176 static const uint32_t INVALID_INDEX = -1;
177
178 struct WasmCustomSection {
179
180 StringRef Name;
181 MCSectionWasm *Section;
182
183 uint32_t OutputContentsOffset;
184 uint32_t OutputIndex;
185
WasmCustomSection__anone61d87550111::WasmCustomSection186 WasmCustomSection(StringRef Name, MCSectionWasm *Section)
187 : Name(Name), Section(Section), OutputContentsOffset(0),
188 OutputIndex(INVALID_INDEX) {}
189 };
190
191 #if !defined(NDEBUG)
operator <<(raw_ostream & OS,const WasmRelocationEntry & Rel)192 raw_ostream &operator<<(raw_ostream &OS, const WasmRelocationEntry &Rel) {
193 Rel.print(OS);
194 return OS;
195 }
196 #endif
197
198 class WasmObjectWriter : public MCObjectWriter {
199 support::endian::Writer W;
200
201 /// The target specific Wasm writer instance.
202 std::unique_ptr<MCWasmObjectTargetWriter> TargetObjectWriter;
203
204 // Relocations for fixing up references in the code section.
205 std::vector<WasmRelocationEntry> CodeRelocations;
206 uint32_t CodeSectionIndex;
207
208 // Relocations for fixing up references in the data section.
209 std::vector<WasmRelocationEntry> DataRelocations;
210 uint32_t DataSectionIndex;
211
212 // Index values to use for fixing up call_indirect type indices.
213 // Maps function symbols to the index of the type of the function
214 DenseMap<const MCSymbolWasm *, uint32_t> TypeIndices;
215 // Maps function symbols to the table element index space. Used
216 // for TABLE_INDEX relocation types (i.e. address taken functions).
217 DenseMap<const MCSymbolWasm *, uint32_t> TableIndices;
218 // Maps function/global symbols to the function/global/section index space.
219 DenseMap<const MCSymbolWasm *, uint32_t> WasmIndices;
220 // Maps data symbols to the Wasm segment and offset/size with the segment.
221 DenseMap<const MCSymbolWasm *, wasm::WasmDataReference> DataLocations;
222
223 // Stores output data (index, relocations, content offset) for custom
224 // section.
225 std::vector<WasmCustomSection> CustomSections;
226 // Relocations for fixing up references in the custom sections.
227 DenseMap<const MCSectionWasm *, std::vector<WasmRelocationEntry>>
228 CustomSectionsRelocations;
229
230 // Map from section to defining function symbol.
231 DenseMap<const MCSection *, const MCSymbol *> SectionFunctions;
232
233 DenseMap<WasmFunctionType, int32_t, WasmFunctionTypeDenseMapInfo>
234 FunctionTypeIndices;
235 SmallVector<WasmFunctionType, 4> FunctionTypes;
236 SmallVector<WasmGlobal, 4> Globals;
237 SmallVector<WasmDataSegment, 4> DataSegments;
238 unsigned NumFunctionImports = 0;
239 unsigned NumGlobalImports = 0;
240 uint32_t SectionCount = 0;
241
242 // TargetObjectWriter wrappers.
is64Bit() const243 bool is64Bit() const { return TargetObjectWriter->is64Bit(); }
getRelocType(const MCValue & Target,const MCFixup & Fixup) const244 unsigned getRelocType(const MCValue &Target, const MCFixup &Fixup) const {
245 return TargetObjectWriter->getRelocType(Target, Fixup);
246 }
247
248 void startSection(SectionBookkeeping &Section, unsigned SectionId);
249 void startCustomSection(SectionBookkeeping &Section, StringRef Name);
250 void endSection(SectionBookkeeping &Section);
251
252 public:
WasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW,raw_pwrite_stream & OS)253 WasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW,
254 raw_pwrite_stream &OS)
255 : W(OS, support::little), TargetObjectWriter(std::move(MOTW)) {}
256
257 ~WasmObjectWriter() override;
258
259 private:
reset()260 void reset() override {
261 CodeRelocations.clear();
262 DataRelocations.clear();
263 TypeIndices.clear();
264 WasmIndices.clear();
265 TableIndices.clear();
266 DataLocations.clear();
267 CustomSectionsRelocations.clear();
268 FunctionTypeIndices.clear();
269 FunctionTypes.clear();
270 Globals.clear();
271 DataSegments.clear();
272 SectionFunctions.clear();
273 NumFunctionImports = 0;
274 NumGlobalImports = 0;
275 MCObjectWriter::reset();
276 }
277
278 void writeHeader(const MCAssembler &Asm);
279
280 void recordRelocation(MCAssembler &Asm, const MCAsmLayout &Layout,
281 const MCFragment *Fragment, const MCFixup &Fixup,
282 MCValue Target, uint64_t &FixedValue) override;
283
284 void executePostLayoutBinding(MCAssembler &Asm,
285 const MCAsmLayout &Layout) override;
286
287 uint64_t writeObject(MCAssembler &Asm, const MCAsmLayout &Layout) override;
288
writeString(const StringRef Str)289 void writeString(const StringRef Str) {
290 encodeULEB128(Str.size(), W.OS);
291 W.OS << Str;
292 }
293
writeValueType(wasm::ValType Ty)294 void writeValueType(wasm::ValType Ty) {
295 W.OS << static_cast<char>(Ty);
296 }
297
298 void writeTypeSection(ArrayRef<WasmFunctionType> FunctionTypes);
299 void writeImportSection(ArrayRef<wasm::WasmImport> Imports, uint32_t DataSize,
300 uint32_t NumElements);
301 void writeFunctionSection(ArrayRef<WasmFunction> Functions);
302 void writeGlobalSection();
303 void writeExportSection(ArrayRef<wasm::WasmExport> Exports);
304 void writeElemSection(ArrayRef<uint32_t> TableElems);
305 void writeCodeSection(const MCAssembler &Asm, const MCAsmLayout &Layout,
306 ArrayRef<WasmFunction> Functions);
307 void writeDataSection();
308 void writeRelocSection(uint32_t SectionIndex, StringRef Name,
309 ArrayRef<WasmRelocationEntry> Relocations);
310 void writeLinkingMetaDataSection(
311 ArrayRef<wasm::WasmSymbolInfo> SymbolInfos,
312 ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs,
313 const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats);
314 void writeCustomSections(const MCAssembler &Asm, const MCAsmLayout &Layout);
315 void writeCustomRelocSections();
316 void
317 updateCustomSectionRelocations(const SmallVector<WasmFunction, 4> &Functions,
318 const MCAsmLayout &Layout);
319
320 uint32_t getProvisionalValue(const WasmRelocationEntry &RelEntry);
321 void applyRelocations(ArrayRef<WasmRelocationEntry> Relocations,
322 uint64_t ContentsOffset);
323
324 uint32_t getRelocationIndexValue(const WasmRelocationEntry &RelEntry);
325 uint32_t getFunctionType(const MCSymbolWasm &Symbol);
326 uint32_t registerFunctionType(const MCSymbolWasm &Symbol);
327 };
328
329 } // end anonymous namespace
330
~WasmObjectWriter()331 WasmObjectWriter::~WasmObjectWriter() {}
332
333 // Write out a section header and a patchable section size field.
startSection(SectionBookkeeping & Section,unsigned SectionId)334 void WasmObjectWriter::startSection(SectionBookkeeping &Section,
335 unsigned SectionId) {
336 LLVM_DEBUG(dbgs() << "startSection " << SectionId << "\n");
337 W.OS << char(SectionId);
338
339 Section.SizeOffset = W.OS.tell();
340
341 // The section size. We don't know the size yet, so reserve enough space
342 // for any 32-bit value; we'll patch it later.
343 encodeULEB128(UINT32_MAX, W.OS);
344
345 // The position where the section starts, for measuring its size.
346 Section.ContentsOffset = W.OS.tell();
347 Section.PayloadOffset = W.OS.tell();
348 Section.Index = SectionCount++;
349 }
350
startCustomSection(SectionBookkeeping & Section,StringRef Name)351 void WasmObjectWriter::startCustomSection(SectionBookkeeping &Section,
352 StringRef Name) {
353 LLVM_DEBUG(dbgs() << "startCustomSection " << Name << "\n");
354 startSection(Section, wasm::WASM_SEC_CUSTOM);
355
356 // The position where the section header ends, for measuring its size.
357 Section.PayloadOffset = W.OS.tell();
358
359 // Custom sections in wasm also have a string identifier.
360 writeString(Name);
361
362 // The position where the custom section starts.
363 Section.ContentsOffset = W.OS.tell();
364 }
365
366 // Now that the section is complete and we know how big it is, patch up the
367 // section size field at the start of the section.
endSection(SectionBookkeeping & Section)368 void WasmObjectWriter::endSection(SectionBookkeeping &Section) {
369 uint64_t Size = W.OS.tell() - Section.PayloadOffset;
370 if (uint32_t(Size) != Size)
371 report_fatal_error("section size does not fit in a uint32_t");
372
373 LLVM_DEBUG(dbgs() << "endSection size=" << Size << "\n");
374
375 // Write the final section size to the payload_len field, which follows
376 // the section id byte.
377 uint8_t Buffer[16];
378 unsigned SizeLen = encodeULEB128(Size, Buffer, 5);
379 assert(SizeLen == 5);
380 static_cast<raw_pwrite_stream &>(W.OS).pwrite((char *)Buffer, SizeLen,
381 Section.SizeOffset);
382 }
383
384 // Emit the Wasm header.
writeHeader(const MCAssembler & Asm)385 void WasmObjectWriter::writeHeader(const MCAssembler &Asm) {
386 W.OS.write(wasm::WasmMagic, sizeof(wasm::WasmMagic));
387 W.write<uint32_t>(wasm::WasmVersion);
388 }
389
executePostLayoutBinding(MCAssembler & Asm,const MCAsmLayout & Layout)390 void WasmObjectWriter::executePostLayoutBinding(MCAssembler &Asm,
391 const MCAsmLayout &Layout) {
392 // Build a map of sections to the function that defines them, for use
393 // in recordRelocation.
394 for (const MCSymbol &S : Asm.symbols()) {
395 const auto &WS = static_cast<const MCSymbolWasm &>(S);
396 if (WS.isDefined() && WS.isFunction() && !WS.isVariable()) {
397 const auto &Sec = static_cast<const MCSectionWasm &>(S.getSection());
398 auto Pair = SectionFunctions.insert(std::make_pair(&Sec, &S));
399 if (!Pair.second)
400 report_fatal_error("section already has a defining function: " +
401 Sec.getSectionName());
402 }
403 }
404 }
405
recordRelocation(MCAssembler & Asm,const MCAsmLayout & Layout,const MCFragment * Fragment,const MCFixup & Fixup,MCValue Target,uint64_t & FixedValue)406 void WasmObjectWriter::recordRelocation(MCAssembler &Asm,
407 const MCAsmLayout &Layout,
408 const MCFragment *Fragment,
409 const MCFixup &Fixup, MCValue Target,
410 uint64_t &FixedValue) {
411 MCAsmBackend &Backend = Asm.getBackend();
412 bool IsPCRel = Backend.getFixupKindInfo(Fixup.getKind()).Flags &
413 MCFixupKindInfo::FKF_IsPCRel;
414 const auto &FixupSection = cast<MCSectionWasm>(*Fragment->getParent());
415 uint64_t C = Target.getConstant();
416 uint64_t FixupOffset = Layout.getFragmentOffset(Fragment) + Fixup.getOffset();
417 MCContext &Ctx = Asm.getContext();
418
419 // The .init_array isn't translated as data, so don't do relocations in it.
420 if (FixupSection.getSectionName().startswith(".init_array"))
421 return;
422
423 if (const MCSymbolRefExpr *RefB = Target.getSymB()) {
424 assert(RefB->getKind() == MCSymbolRefExpr::VK_None &&
425 "Should not have constructed this");
426
427 // Let A, B and C being the components of Target and R be the location of
428 // the fixup. If the fixup is not pcrel, we want to compute (A - B + C).
429 // If it is pcrel, we want to compute (A - B + C - R).
430
431 // In general, Wasm has no relocations for -B. It can only represent (A + C)
432 // or (A + C - R). If B = R + K and the relocation is not pcrel, we can
433 // replace B to implement it: (A - R - K + C)
434 if (IsPCRel) {
435 Ctx.reportError(
436 Fixup.getLoc(),
437 "No relocation available to represent this relative expression");
438 return;
439 }
440
441 const auto &SymB = cast<MCSymbolWasm>(RefB->getSymbol());
442
443 if (SymB.isUndefined()) {
444 Ctx.reportError(Fixup.getLoc(),
445 Twine("symbol '") + SymB.getName() +
446 "' can not be undefined in a subtraction expression");
447 return;
448 }
449
450 assert(!SymB.isAbsolute() && "Should have been folded");
451 const MCSection &SecB = SymB.getSection();
452 if (&SecB != &FixupSection) {
453 Ctx.reportError(Fixup.getLoc(),
454 "Cannot represent a difference across sections");
455 return;
456 }
457
458 uint64_t SymBOffset = Layout.getSymbolOffset(SymB);
459 uint64_t K = SymBOffset - FixupOffset;
460 IsPCRel = true;
461 C -= K;
462 }
463
464 // We either rejected the fixup or folded B into C at this point.
465 const MCSymbolRefExpr *RefA = Target.getSymA();
466 const auto *SymA = RefA ? cast<MCSymbolWasm>(&RefA->getSymbol()) : nullptr;
467
468 if (SymA && SymA->isVariable()) {
469 const MCExpr *Expr = SymA->getVariableValue();
470 const auto *Inner = cast<MCSymbolRefExpr>(Expr);
471 if (Inner->getKind() == MCSymbolRefExpr::VK_WEAKREF)
472 llvm_unreachable("weakref used in reloc not yet implemented");
473 }
474
475 // Put any constant offset in an addend. Offsets can be negative, and
476 // LLVM expects wrapping, in contrast to wasm's immediates which can't
477 // be negative and don't wrap.
478 FixedValue = 0;
479
480 unsigned Type = getRelocType(Target, Fixup);
481 assert(!IsPCRel);
482 assert(SymA);
483
484 // Absolute offset within a section or a function.
485 // Currently only supported for for metadata sections.
486 // See: test/MC/WebAssembly/blockaddress.ll
487 if (Type == wasm::R_WEBASSEMBLY_FUNCTION_OFFSET_I32 ||
488 Type == wasm::R_WEBASSEMBLY_SECTION_OFFSET_I32) {
489 if (!FixupSection.getKind().isMetadata())
490 report_fatal_error("relocations for function or section offsets are "
491 "only supported in metadata sections");
492
493 const MCSymbol *SectionSymbol = nullptr;
494 const MCSection &SecA = SymA->getSection();
495 if (SecA.getKind().isText())
496 SectionSymbol = SectionFunctions.find(&SecA)->second;
497 else
498 SectionSymbol = SecA.getBeginSymbol();
499 if (!SectionSymbol)
500 report_fatal_error("section symbol is required for relocation");
501
502 C += Layout.getSymbolOffset(*SymA);
503 SymA = cast<MCSymbolWasm>(SectionSymbol);
504 }
505
506 // Relocation other than R_WEBASSEMBLY_TYPE_INDEX_LEB are required to be
507 // against a named symbol.
508 if (Type != wasm::R_WEBASSEMBLY_TYPE_INDEX_LEB) {
509 if (SymA->getName().empty())
510 report_fatal_error("relocations against un-named temporaries are not yet "
511 "supported by wasm");
512
513 SymA->setUsedInReloc();
514 }
515
516 WasmRelocationEntry Rec(FixupOffset, SymA, C, Type, &FixupSection);
517 LLVM_DEBUG(dbgs() << "WasmReloc: " << Rec << "\n");
518
519 if (FixupSection.isWasmData()) {
520 DataRelocations.push_back(Rec);
521 } else if (FixupSection.getKind().isText()) {
522 CodeRelocations.push_back(Rec);
523 } else if (FixupSection.getKind().isMetadata()) {
524 CustomSectionsRelocations[&FixupSection].push_back(Rec);
525 } else {
526 llvm_unreachable("unexpected section type");
527 }
528 }
529
530 // Write X as an (unsigned) LEB value at offset Offset in Stream, padded
531 // to allow patching.
532 static void
WritePatchableLEB(raw_pwrite_stream & Stream,uint32_t X,uint64_t Offset)533 WritePatchableLEB(raw_pwrite_stream &Stream, uint32_t X, uint64_t Offset) {
534 uint8_t Buffer[5];
535 unsigned SizeLen = encodeULEB128(X, Buffer, 5);
536 assert(SizeLen == 5);
537 Stream.pwrite((char *)Buffer, SizeLen, Offset);
538 }
539
540 // Write X as an signed LEB value at offset Offset in Stream, padded
541 // to allow patching.
542 static void
WritePatchableSLEB(raw_pwrite_stream & Stream,int32_t X,uint64_t Offset)543 WritePatchableSLEB(raw_pwrite_stream &Stream, int32_t X, uint64_t Offset) {
544 uint8_t Buffer[5];
545 unsigned SizeLen = encodeSLEB128(X, Buffer, 5);
546 assert(SizeLen == 5);
547 Stream.pwrite((char *)Buffer, SizeLen, Offset);
548 }
549
550 // Write X as a plain integer value at offset Offset in Stream.
WriteI32(raw_pwrite_stream & Stream,uint32_t X,uint64_t Offset)551 static void WriteI32(raw_pwrite_stream &Stream, uint32_t X, uint64_t Offset) {
552 uint8_t Buffer[4];
553 support::endian::write32le(Buffer, X);
554 Stream.pwrite((char *)Buffer, sizeof(Buffer), Offset);
555 }
556
ResolveSymbol(const MCSymbolWasm & Symbol)557 static const MCSymbolWasm* ResolveSymbol(const MCSymbolWasm& Symbol) {
558 if (Symbol.isVariable()) {
559 const MCExpr *Expr = Symbol.getVariableValue();
560 auto *Inner = cast<MCSymbolRefExpr>(Expr);
561 return cast<MCSymbolWasm>(&Inner->getSymbol());
562 }
563 return &Symbol;
564 }
565
566 // Compute a value to write into the code at the location covered
567 // by RelEntry. This value isn't used by the static linker; it just serves
568 // to make the object format more readable and more likely to be directly
569 // useable.
570 uint32_t
getProvisionalValue(const WasmRelocationEntry & RelEntry)571 WasmObjectWriter::getProvisionalValue(const WasmRelocationEntry &RelEntry) {
572 switch (RelEntry.Type) {
573 case wasm::R_WEBASSEMBLY_TABLE_INDEX_SLEB:
574 case wasm::R_WEBASSEMBLY_TABLE_INDEX_I32: {
575 // Provisional value is table address of the resolved symbol itself
576 const MCSymbolWasm *Sym = ResolveSymbol(*RelEntry.Symbol);
577 assert(Sym->isFunction());
578 return TableIndices[Sym];
579 }
580 case wasm::R_WEBASSEMBLY_TYPE_INDEX_LEB:
581 // Provisional value is same as the index
582 return getRelocationIndexValue(RelEntry);
583 case wasm::R_WEBASSEMBLY_FUNCTION_INDEX_LEB:
584 case wasm::R_WEBASSEMBLY_GLOBAL_INDEX_LEB:
585 // Provisional value is function/global Wasm index
586 if (!WasmIndices.count(RelEntry.Symbol))
587 report_fatal_error("symbol not found in wasm index space: " +
588 RelEntry.Symbol->getName());
589 return WasmIndices[RelEntry.Symbol];
590 case wasm::R_WEBASSEMBLY_FUNCTION_OFFSET_I32:
591 case wasm::R_WEBASSEMBLY_SECTION_OFFSET_I32: {
592 const auto &Section =
593 static_cast<const MCSectionWasm &>(RelEntry.Symbol->getSection());
594 return Section.getSectionOffset() + RelEntry.Addend;
595 }
596 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_LEB:
597 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_I32:
598 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_SLEB: {
599 // Provisional value is address of the global
600 const MCSymbolWasm *Sym = ResolveSymbol(*RelEntry.Symbol);
601 // For undefined symbols, use zero
602 if (!Sym->isDefined())
603 return 0;
604 const wasm::WasmDataReference &Ref = DataLocations[Sym];
605 const WasmDataSegment &Segment = DataSegments[Ref.Segment];
606 // Ignore overflow. LLVM allows address arithmetic to silently wrap.
607 return Segment.Offset + Ref.Offset + RelEntry.Addend;
608 }
609 default:
610 llvm_unreachable("invalid relocation type");
611 }
612 }
613
addData(SmallVectorImpl<char> & DataBytes,MCSectionWasm & DataSection)614 static void addData(SmallVectorImpl<char> &DataBytes,
615 MCSectionWasm &DataSection) {
616 LLVM_DEBUG(errs() << "addData: " << DataSection.getSectionName() << "\n");
617
618 DataBytes.resize(alignTo(DataBytes.size(), DataSection.getAlignment()));
619
620 for (const MCFragment &Frag : DataSection) {
621 if (Frag.hasInstructions())
622 report_fatal_error("only data supported in data sections");
623
624 if (auto *Align = dyn_cast<MCAlignFragment>(&Frag)) {
625 if (Align->getValueSize() != 1)
626 report_fatal_error("only byte values supported for alignment");
627 // If nops are requested, use zeros, as this is the data section.
628 uint8_t Value = Align->hasEmitNops() ? 0 : Align->getValue();
629 uint64_t Size = std::min<uint64_t>(alignTo(DataBytes.size(),
630 Align->getAlignment()),
631 DataBytes.size() +
632 Align->getMaxBytesToEmit());
633 DataBytes.resize(Size, Value);
634 } else if (auto *Fill = dyn_cast<MCFillFragment>(&Frag)) {
635 int64_t NumValues;
636 if (!Fill->getNumValues().evaluateAsAbsolute(NumValues))
637 llvm_unreachable("The fill should be an assembler constant");
638 DataBytes.insert(DataBytes.end(), Fill->getValueSize() * NumValues,
639 Fill->getValue());
640 } else {
641 const auto &DataFrag = cast<MCDataFragment>(Frag);
642 const SmallVectorImpl<char> &Contents = DataFrag.getContents();
643
644 DataBytes.insert(DataBytes.end(), Contents.begin(), Contents.end());
645 }
646 }
647
648 LLVM_DEBUG(dbgs() << "addData -> " << DataBytes.size() << "\n");
649 }
650
651 uint32_t
getRelocationIndexValue(const WasmRelocationEntry & RelEntry)652 WasmObjectWriter::getRelocationIndexValue(const WasmRelocationEntry &RelEntry) {
653 if (RelEntry.Type == wasm::R_WEBASSEMBLY_TYPE_INDEX_LEB) {
654 if (!TypeIndices.count(RelEntry.Symbol))
655 report_fatal_error("symbol not found in type index space: " +
656 RelEntry.Symbol->getName());
657 return TypeIndices[RelEntry.Symbol];
658 }
659
660 return RelEntry.Symbol->getIndex();
661 }
662
663 // Apply the portions of the relocation records that we can handle ourselves
664 // directly.
applyRelocations(ArrayRef<WasmRelocationEntry> Relocations,uint64_t ContentsOffset)665 void WasmObjectWriter::applyRelocations(
666 ArrayRef<WasmRelocationEntry> Relocations, uint64_t ContentsOffset) {
667 auto &Stream = static_cast<raw_pwrite_stream &>(W.OS);
668 for (const WasmRelocationEntry &RelEntry : Relocations) {
669 uint64_t Offset = ContentsOffset +
670 RelEntry.FixupSection->getSectionOffset() +
671 RelEntry.Offset;
672
673 LLVM_DEBUG(dbgs() << "applyRelocation: " << RelEntry << "\n");
674 uint32_t Value = getProvisionalValue(RelEntry);
675
676 switch (RelEntry.Type) {
677 case wasm::R_WEBASSEMBLY_FUNCTION_INDEX_LEB:
678 case wasm::R_WEBASSEMBLY_TYPE_INDEX_LEB:
679 case wasm::R_WEBASSEMBLY_GLOBAL_INDEX_LEB:
680 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_LEB:
681 WritePatchableLEB(Stream, Value, Offset);
682 break;
683 case wasm::R_WEBASSEMBLY_TABLE_INDEX_I32:
684 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_I32:
685 case wasm::R_WEBASSEMBLY_FUNCTION_OFFSET_I32:
686 case wasm::R_WEBASSEMBLY_SECTION_OFFSET_I32:
687 WriteI32(Stream, Value, Offset);
688 break;
689 case wasm::R_WEBASSEMBLY_TABLE_INDEX_SLEB:
690 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_SLEB:
691 WritePatchableSLEB(Stream, Value, Offset);
692 break;
693 default:
694 llvm_unreachable("invalid relocation type");
695 }
696 }
697 }
698
writeTypeSection(ArrayRef<WasmFunctionType> FunctionTypes)699 void WasmObjectWriter::writeTypeSection(
700 ArrayRef<WasmFunctionType> FunctionTypes) {
701 if (FunctionTypes.empty())
702 return;
703
704 SectionBookkeeping Section;
705 startSection(Section, wasm::WASM_SEC_TYPE);
706
707 encodeULEB128(FunctionTypes.size(), W.OS);
708
709 for (const WasmFunctionType &FuncTy : FunctionTypes) {
710 W.OS << char(wasm::WASM_TYPE_FUNC);
711 encodeULEB128(FuncTy.Params.size(), W.OS);
712 for (wasm::ValType Ty : FuncTy.Params)
713 writeValueType(Ty);
714 encodeULEB128(FuncTy.Returns.size(), W.OS);
715 for (wasm::ValType Ty : FuncTy.Returns)
716 writeValueType(Ty);
717 }
718
719 endSection(Section);
720 }
721
writeImportSection(ArrayRef<wasm::WasmImport> Imports,uint32_t DataSize,uint32_t NumElements)722 void WasmObjectWriter::writeImportSection(ArrayRef<wasm::WasmImport> Imports,
723 uint32_t DataSize,
724 uint32_t NumElements) {
725 if (Imports.empty())
726 return;
727
728 uint32_t NumPages = (DataSize + wasm::WasmPageSize - 1) / wasm::WasmPageSize;
729
730 SectionBookkeeping Section;
731 startSection(Section, wasm::WASM_SEC_IMPORT);
732
733 encodeULEB128(Imports.size(), W.OS);
734 for (const wasm::WasmImport &Import : Imports) {
735 writeString(Import.Module);
736 writeString(Import.Field);
737 W.OS << char(Import.Kind);
738
739 switch (Import.Kind) {
740 case wasm::WASM_EXTERNAL_FUNCTION:
741 encodeULEB128(Import.SigIndex, W.OS);
742 break;
743 case wasm::WASM_EXTERNAL_GLOBAL:
744 W.OS << char(Import.Global.Type);
745 W.OS << char(Import.Global.Mutable ? 1 : 0);
746 break;
747 case wasm::WASM_EXTERNAL_MEMORY:
748 encodeULEB128(0, W.OS); // flags
749 encodeULEB128(NumPages, W.OS); // initial
750 break;
751 case wasm::WASM_EXTERNAL_TABLE:
752 W.OS << char(Import.Table.ElemType);
753 encodeULEB128(0, W.OS); // flags
754 encodeULEB128(NumElements, W.OS); // initial
755 break;
756 default:
757 llvm_unreachable("unsupported import kind");
758 }
759 }
760
761 endSection(Section);
762 }
763
writeFunctionSection(ArrayRef<WasmFunction> Functions)764 void WasmObjectWriter::writeFunctionSection(ArrayRef<WasmFunction> Functions) {
765 if (Functions.empty())
766 return;
767
768 SectionBookkeeping Section;
769 startSection(Section, wasm::WASM_SEC_FUNCTION);
770
771 encodeULEB128(Functions.size(), W.OS);
772 for (const WasmFunction &Func : Functions)
773 encodeULEB128(Func.Type, W.OS);
774
775 endSection(Section);
776 }
777
writeGlobalSection()778 void WasmObjectWriter::writeGlobalSection() {
779 if (Globals.empty())
780 return;
781
782 SectionBookkeeping Section;
783 startSection(Section, wasm::WASM_SEC_GLOBAL);
784
785 encodeULEB128(Globals.size(), W.OS);
786 for (const WasmGlobal &Global : Globals) {
787 writeValueType(static_cast<wasm::ValType>(Global.Type.Type));
788 W.OS << char(Global.Type.Mutable);
789
790 W.OS << char(wasm::WASM_OPCODE_I32_CONST);
791 encodeSLEB128(Global.InitialValue, W.OS);
792 W.OS << char(wasm::WASM_OPCODE_END);
793 }
794
795 endSection(Section);
796 }
797
writeExportSection(ArrayRef<wasm::WasmExport> Exports)798 void WasmObjectWriter::writeExportSection(ArrayRef<wasm::WasmExport> Exports) {
799 if (Exports.empty())
800 return;
801
802 SectionBookkeeping Section;
803 startSection(Section, wasm::WASM_SEC_EXPORT);
804
805 encodeULEB128(Exports.size(), W.OS);
806 for (const wasm::WasmExport &Export : Exports) {
807 writeString(Export.Name);
808 W.OS << char(Export.Kind);
809 encodeULEB128(Export.Index, W.OS);
810 }
811
812 endSection(Section);
813 }
814
writeElemSection(ArrayRef<uint32_t> TableElems)815 void WasmObjectWriter::writeElemSection(ArrayRef<uint32_t> TableElems) {
816 if (TableElems.empty())
817 return;
818
819 SectionBookkeeping Section;
820 startSection(Section, wasm::WASM_SEC_ELEM);
821
822 encodeULEB128(1, W.OS); // number of "segments"
823 encodeULEB128(0, W.OS); // the table index
824
825 // init expr for starting offset
826 W.OS << char(wasm::WASM_OPCODE_I32_CONST);
827 encodeSLEB128(kInitialTableOffset, W.OS);
828 W.OS << char(wasm::WASM_OPCODE_END);
829
830 encodeULEB128(TableElems.size(), W.OS);
831 for (uint32_t Elem : TableElems)
832 encodeULEB128(Elem, W.OS);
833
834 endSection(Section);
835 }
836
writeCodeSection(const MCAssembler & Asm,const MCAsmLayout & Layout,ArrayRef<WasmFunction> Functions)837 void WasmObjectWriter::writeCodeSection(const MCAssembler &Asm,
838 const MCAsmLayout &Layout,
839 ArrayRef<WasmFunction> Functions) {
840 if (Functions.empty())
841 return;
842
843 SectionBookkeeping Section;
844 startSection(Section, wasm::WASM_SEC_CODE);
845 CodeSectionIndex = Section.Index;
846
847 encodeULEB128(Functions.size(), W.OS);
848
849 for (const WasmFunction &Func : Functions) {
850 auto &FuncSection = static_cast<MCSectionWasm &>(Func.Sym->getSection());
851
852 int64_t Size = 0;
853 if (!Func.Sym->getSize()->evaluateAsAbsolute(Size, Layout))
854 report_fatal_error(".size expression must be evaluatable");
855
856 encodeULEB128(Size, W.OS);
857 FuncSection.setSectionOffset(W.OS.tell() - Section.ContentsOffset);
858 Asm.writeSectionData(W.OS, &FuncSection, Layout);
859 }
860
861 // Apply fixups.
862 applyRelocations(CodeRelocations, Section.ContentsOffset);
863
864 endSection(Section);
865 }
866
writeDataSection()867 void WasmObjectWriter::writeDataSection() {
868 if (DataSegments.empty())
869 return;
870
871 SectionBookkeeping Section;
872 startSection(Section, wasm::WASM_SEC_DATA);
873 DataSectionIndex = Section.Index;
874
875 encodeULEB128(DataSegments.size(), W.OS); // count
876
877 for (const WasmDataSegment &Segment : DataSegments) {
878 encodeULEB128(0, W.OS); // memory index
879 W.OS << char(wasm::WASM_OPCODE_I32_CONST);
880 encodeSLEB128(Segment.Offset, W.OS); // offset
881 W.OS << char(wasm::WASM_OPCODE_END);
882 encodeULEB128(Segment.Data.size(), W.OS); // size
883 Segment.Section->setSectionOffset(W.OS.tell() - Section.ContentsOffset);
884 W.OS << Segment.Data; // data
885 }
886
887 // Apply fixups.
888 applyRelocations(DataRelocations, Section.ContentsOffset);
889
890 endSection(Section);
891 }
892
writeRelocSection(uint32_t SectionIndex,StringRef Name,ArrayRef<WasmRelocationEntry> Relocations)893 void WasmObjectWriter::writeRelocSection(
894 uint32_t SectionIndex, StringRef Name,
895 ArrayRef<WasmRelocationEntry> Relocations) {
896 // See: https://github.com/WebAssembly/tool-conventions/blob/master/Linking.md
897 // for descriptions of the reloc sections.
898
899 if (Relocations.empty())
900 return;
901
902 SectionBookkeeping Section;
903 startCustomSection(Section, std::string("reloc.") + Name.str());
904
905 encodeULEB128(SectionIndex, W.OS);
906 encodeULEB128(Relocations.size(), W.OS);
907 for (const WasmRelocationEntry& RelEntry : Relocations) {
908 uint64_t Offset = RelEntry.Offset +
909 RelEntry.FixupSection->getSectionOffset();
910 uint32_t Index = getRelocationIndexValue(RelEntry);
911
912 W.OS << char(RelEntry.Type);
913 encodeULEB128(Offset, W.OS);
914 encodeULEB128(Index, W.OS);
915 if (RelEntry.hasAddend())
916 encodeSLEB128(RelEntry.Addend, W.OS);
917 }
918
919 endSection(Section);
920 }
921
writeCustomRelocSections()922 void WasmObjectWriter::writeCustomRelocSections() {
923 for (const auto &Sec : CustomSections) {
924 auto &Relocations = CustomSectionsRelocations[Sec.Section];
925 writeRelocSection(Sec.OutputIndex, Sec.Name, Relocations);
926 }
927 }
928
writeLinkingMetaDataSection(ArrayRef<wasm::WasmSymbolInfo> SymbolInfos,ArrayRef<std::pair<uint16_t,uint32_t>> InitFuncs,const std::map<StringRef,std::vector<WasmComdatEntry>> & Comdats)929 void WasmObjectWriter::writeLinkingMetaDataSection(
930 ArrayRef<wasm::WasmSymbolInfo> SymbolInfos,
931 ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs,
932 const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats) {
933 SectionBookkeeping Section;
934 startCustomSection(Section, "linking");
935 encodeULEB128(wasm::WasmMetadataVersion, W.OS);
936
937 SectionBookkeeping SubSection;
938 if (SymbolInfos.size() != 0) {
939 startSection(SubSection, wasm::WASM_SYMBOL_TABLE);
940 encodeULEB128(SymbolInfos.size(), W.OS);
941 for (const wasm::WasmSymbolInfo &Sym : SymbolInfos) {
942 encodeULEB128(Sym.Kind, W.OS);
943 encodeULEB128(Sym.Flags, W.OS);
944 switch (Sym.Kind) {
945 case wasm::WASM_SYMBOL_TYPE_FUNCTION:
946 case wasm::WASM_SYMBOL_TYPE_GLOBAL:
947 encodeULEB128(Sym.ElementIndex, W.OS);
948 if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0)
949 writeString(Sym.Name);
950 break;
951 case wasm::WASM_SYMBOL_TYPE_DATA:
952 writeString(Sym.Name);
953 if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0) {
954 encodeULEB128(Sym.DataRef.Segment, W.OS);
955 encodeULEB128(Sym.DataRef.Offset, W.OS);
956 encodeULEB128(Sym.DataRef.Size, W.OS);
957 }
958 break;
959 case wasm::WASM_SYMBOL_TYPE_SECTION: {
960 const uint32_t SectionIndex =
961 CustomSections[Sym.ElementIndex].OutputIndex;
962 encodeULEB128(SectionIndex, W.OS);
963 break;
964 }
965 default:
966 llvm_unreachable("unexpected kind");
967 }
968 }
969 endSection(SubSection);
970 }
971
972 if (DataSegments.size()) {
973 startSection(SubSection, wasm::WASM_SEGMENT_INFO);
974 encodeULEB128(DataSegments.size(), W.OS);
975 for (const WasmDataSegment &Segment : DataSegments) {
976 writeString(Segment.Name);
977 encodeULEB128(Segment.Alignment, W.OS);
978 encodeULEB128(Segment.Flags, W.OS);
979 }
980 endSection(SubSection);
981 }
982
983 if (!InitFuncs.empty()) {
984 startSection(SubSection, wasm::WASM_INIT_FUNCS);
985 encodeULEB128(InitFuncs.size(), W.OS);
986 for (auto &StartFunc : InitFuncs) {
987 encodeULEB128(StartFunc.first, W.OS); // priority
988 encodeULEB128(StartFunc.second, W.OS); // function index
989 }
990 endSection(SubSection);
991 }
992
993 if (Comdats.size()) {
994 startSection(SubSection, wasm::WASM_COMDAT_INFO);
995 encodeULEB128(Comdats.size(), W.OS);
996 for (const auto &C : Comdats) {
997 writeString(C.first);
998 encodeULEB128(0, W.OS); // flags for future use
999 encodeULEB128(C.second.size(), W.OS);
1000 for (const WasmComdatEntry &Entry : C.second) {
1001 encodeULEB128(Entry.Kind, W.OS);
1002 encodeULEB128(Entry.Index, W.OS);
1003 }
1004 }
1005 endSection(SubSection);
1006 }
1007
1008 endSection(Section);
1009 }
1010
writeCustomSections(const MCAssembler & Asm,const MCAsmLayout & Layout)1011 void WasmObjectWriter::writeCustomSections(const MCAssembler &Asm,
1012 const MCAsmLayout &Layout) {
1013 for (auto &CustomSection : CustomSections) {
1014 SectionBookkeeping Section;
1015 auto *Sec = CustomSection.Section;
1016 startCustomSection(Section, CustomSection.Name);
1017
1018 Sec->setSectionOffset(W.OS.tell() - Section.ContentsOffset);
1019 Asm.writeSectionData(W.OS, Sec, Layout);
1020
1021 CustomSection.OutputContentsOffset = Section.ContentsOffset;
1022 CustomSection.OutputIndex = Section.Index;
1023
1024 endSection(Section);
1025
1026 // Apply fixups.
1027 auto &Relocations = CustomSectionsRelocations[CustomSection.Section];
1028 applyRelocations(Relocations, CustomSection.OutputContentsOffset);
1029 }
1030 }
1031
getFunctionType(const MCSymbolWasm & Symbol)1032 uint32_t WasmObjectWriter::getFunctionType(const MCSymbolWasm& Symbol) {
1033 assert(Symbol.isFunction());
1034 assert(TypeIndices.count(&Symbol));
1035 return TypeIndices[&Symbol];
1036 }
1037
registerFunctionType(const MCSymbolWasm & Symbol)1038 uint32_t WasmObjectWriter::registerFunctionType(const MCSymbolWasm& Symbol) {
1039 assert(Symbol.isFunction());
1040
1041 WasmFunctionType F;
1042 const MCSymbolWasm* ResolvedSym = ResolveSymbol(Symbol);
1043 F.Returns = ResolvedSym->getReturns();
1044 F.Params = ResolvedSym->getParams();
1045
1046 auto Pair =
1047 FunctionTypeIndices.insert(std::make_pair(F, FunctionTypes.size()));
1048 if (Pair.second)
1049 FunctionTypes.push_back(F);
1050 TypeIndices[&Symbol] = Pair.first->second;
1051
1052 LLVM_DEBUG(dbgs() << "registerFunctionType: " << Symbol
1053 << " new:" << Pair.second << "\n");
1054 LLVM_DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n");
1055 return Pair.first->second;
1056 }
1057
isInSymtab(const MCSymbolWasm & Sym)1058 static bool isInSymtab(const MCSymbolWasm &Sym) {
1059 if (Sym.isUsedInReloc())
1060 return true;
1061
1062 if (Sym.isComdat() && !Sym.isDefined())
1063 return false;
1064
1065 if (Sym.isTemporary() && Sym.getName().empty())
1066 return false;
1067
1068 if (Sym.isTemporary() && Sym.isData() && !Sym.getSize())
1069 return false;
1070
1071 if (Sym.isSection())
1072 return false;
1073
1074 return true;
1075 }
1076
writeObject(MCAssembler & Asm,const MCAsmLayout & Layout)1077 uint64_t WasmObjectWriter::writeObject(MCAssembler &Asm,
1078 const MCAsmLayout &Layout) {
1079 uint64_t StartOffset = W.OS.tell();
1080
1081 LLVM_DEBUG(dbgs() << "WasmObjectWriter::writeObject\n");
1082 MCContext &Ctx = Asm.getContext();
1083
1084 // Collect information from the available symbols.
1085 SmallVector<WasmFunction, 4> Functions;
1086 SmallVector<uint32_t, 4> TableElems;
1087 SmallVector<wasm::WasmImport, 4> Imports;
1088 SmallVector<wasm::WasmExport, 4> Exports;
1089 SmallVector<wasm::WasmSymbolInfo, 4> SymbolInfos;
1090 SmallVector<std::pair<uint16_t, uint32_t>, 2> InitFuncs;
1091 std::map<StringRef, std::vector<WasmComdatEntry>> Comdats;
1092 uint32_t DataSize = 0;
1093
1094 // For now, always emit the memory import, since loads and stores are not
1095 // valid without it. In the future, we could perhaps be more clever and omit
1096 // it if there are no loads or stores.
1097 MCSymbolWasm *MemorySym =
1098 cast<MCSymbolWasm>(Ctx.getOrCreateSymbol("__linear_memory"));
1099 wasm::WasmImport MemImport;
1100 MemImport.Module = MemorySym->getModuleName();
1101 MemImport.Field = MemorySym->getName();
1102 MemImport.Kind = wasm::WASM_EXTERNAL_MEMORY;
1103 Imports.push_back(MemImport);
1104
1105 // For now, always emit the table section, since indirect calls are not
1106 // valid without it. In the future, we could perhaps be more clever and omit
1107 // it if there are no indirect calls.
1108 MCSymbolWasm *TableSym =
1109 cast<MCSymbolWasm>(Ctx.getOrCreateSymbol("__indirect_function_table"));
1110 wasm::WasmImport TableImport;
1111 TableImport.Module = TableSym->getModuleName();
1112 TableImport.Field = TableSym->getName();
1113 TableImport.Kind = wasm::WASM_EXTERNAL_TABLE;
1114 TableImport.Table.ElemType = wasm::WASM_TYPE_ANYFUNC;
1115 Imports.push_back(TableImport);
1116
1117 // Populate FunctionTypeIndices, and Imports and WasmIndices for undefined
1118 // symbols. This must be done before populating WasmIndices for defined
1119 // symbols.
1120 for (const MCSymbol &S : Asm.symbols()) {
1121 const auto &WS = static_cast<const MCSymbolWasm &>(S);
1122
1123 // Register types for all functions, including those with private linkage
1124 // (because wasm always needs a type signature).
1125 if (WS.isFunction())
1126 registerFunctionType(WS);
1127
1128 if (WS.isTemporary())
1129 continue;
1130
1131 // If the symbol is not defined in this translation unit, import it.
1132 if (!WS.isDefined() && !WS.isComdat()) {
1133 if (WS.isFunction()) {
1134 wasm::WasmImport Import;
1135 Import.Module = WS.getModuleName();
1136 Import.Field = WS.getName();
1137 Import.Kind = wasm::WASM_EXTERNAL_FUNCTION;
1138 Import.SigIndex = getFunctionType(WS);
1139 Imports.push_back(Import);
1140 WasmIndices[&WS] = NumFunctionImports++;
1141 } else if (WS.isGlobal()) {
1142 if (WS.isWeak())
1143 report_fatal_error("undefined global symbol cannot be weak");
1144
1145 wasm::WasmImport Import;
1146 Import.Module = WS.getModuleName();
1147 Import.Field = WS.getName();
1148 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL;
1149 Import.Global = WS.getGlobalType();
1150 Imports.push_back(Import);
1151 WasmIndices[&WS] = NumGlobalImports++;
1152 }
1153 }
1154 }
1155
1156 // Populate DataSegments and CustomSections, which must be done before
1157 // populating DataLocations.
1158 for (MCSection &Sec : Asm) {
1159 auto &Section = static_cast<MCSectionWasm &>(Sec);
1160 StringRef SectionName = Section.getSectionName();
1161
1162 // .init_array sections are handled specially elsewhere.
1163 if (SectionName.startswith(".init_array"))
1164 continue;
1165
1166 // Code is handled separately
1167 if (Section.getKind().isText())
1168 continue;
1169
1170 if (Section.isWasmData()) {
1171 uint32_t SegmentIndex = DataSegments.size();
1172 DataSize = alignTo(DataSize, Section.getAlignment());
1173 DataSegments.emplace_back();
1174 WasmDataSegment &Segment = DataSegments.back();
1175 Segment.Name = SectionName;
1176 Segment.Offset = DataSize;
1177 Segment.Section = &Section;
1178 addData(Segment.Data, Section);
1179 Segment.Alignment = Section.getAlignment();
1180 Segment.Flags = 0;
1181 DataSize += Segment.Data.size();
1182 Section.setSegmentIndex(SegmentIndex);
1183
1184 if (const MCSymbolWasm *C = Section.getGroup()) {
1185 Comdats[C->getName()].emplace_back(
1186 WasmComdatEntry{wasm::WASM_COMDAT_DATA, SegmentIndex});
1187 }
1188 } else {
1189 // Create custom sections
1190 assert(Sec.getKind().isMetadata());
1191
1192 StringRef Name = SectionName;
1193
1194 // For user-defined custom sections, strip the prefix
1195 if (Name.startswith(".custom_section."))
1196 Name = Name.substr(strlen(".custom_section."));
1197
1198 MCSymbol* Begin = Sec.getBeginSymbol();
1199 if (Begin) {
1200 WasmIndices[cast<MCSymbolWasm>(Begin)] = CustomSections.size();
1201 if (SectionName != Begin->getName())
1202 report_fatal_error("section name and begin symbol should match: " +
1203 Twine(SectionName));
1204 }
1205 CustomSections.emplace_back(Name, &Section);
1206 }
1207 }
1208
1209 // Populate WasmIndices and DataLocations for defined symbols.
1210 for (const MCSymbol &S : Asm.symbols()) {
1211 // Ignore unnamed temporary symbols, which aren't ever exported, imported,
1212 // or used in relocations.
1213 if (S.isTemporary() && S.getName().empty())
1214 continue;
1215
1216 const auto &WS = static_cast<const MCSymbolWasm &>(S);
1217 LLVM_DEBUG(
1218 dbgs() << "MCSymbol: " << toString(WS.getType()) << " '" << S << "'"
1219 << " isDefined=" << S.isDefined() << " isExternal="
1220 << S.isExternal() << " isTemporary=" << S.isTemporary()
1221 << " isWeak=" << WS.isWeak() << " isHidden=" << WS.isHidden()
1222 << " isVariable=" << WS.isVariable() << "\n");
1223
1224 if (WS.isVariable())
1225 continue;
1226 if (WS.isComdat() && !WS.isDefined())
1227 continue;
1228
1229 if (WS.isFunction()) {
1230 unsigned Index;
1231 if (WS.isDefined()) {
1232 if (WS.getOffset() != 0)
1233 report_fatal_error(
1234 "function sections must contain one function each");
1235
1236 if (WS.getSize() == 0)
1237 report_fatal_error(
1238 "function symbols must have a size set with .size");
1239
1240 // A definition. Write out the function body.
1241 Index = NumFunctionImports + Functions.size();
1242 WasmFunction Func;
1243 Func.Type = getFunctionType(WS);
1244 Func.Sym = &WS;
1245 WasmIndices[&WS] = Index;
1246 Functions.push_back(Func);
1247
1248 auto &Section = static_cast<MCSectionWasm &>(WS.getSection());
1249 if (const MCSymbolWasm *C = Section.getGroup()) {
1250 Comdats[C->getName()].emplace_back(
1251 WasmComdatEntry{wasm::WASM_COMDAT_FUNCTION, Index});
1252 }
1253 } else {
1254 // An import; the index was assigned above.
1255 Index = WasmIndices.find(&WS)->second;
1256 }
1257
1258 LLVM_DEBUG(dbgs() << " -> function index: " << Index << "\n");
1259 } else if (WS.isData()) {
1260 if (WS.isTemporary() && !WS.getSize())
1261 continue;
1262
1263 if (!WS.isDefined()) {
1264 LLVM_DEBUG(dbgs() << " -> segment index: -1"
1265 << "\n");
1266 continue;
1267 }
1268
1269 if (!WS.getSize())
1270 report_fatal_error("data symbols must have a size set with .size: " +
1271 WS.getName());
1272
1273 int64_t Size = 0;
1274 if (!WS.getSize()->evaluateAsAbsolute(Size, Layout))
1275 report_fatal_error(".size expression must be evaluatable");
1276
1277 auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection());
1278 assert(DataSection.isWasmData());
1279
1280 // For each data symbol, export it in the symtab as a reference to the
1281 // corresponding Wasm data segment.
1282 wasm::WasmDataReference Ref = wasm::WasmDataReference{
1283 DataSection.getSegmentIndex(),
1284 static_cast<uint32_t>(Layout.getSymbolOffset(WS)),
1285 static_cast<uint32_t>(Size)};
1286 DataLocations[&WS] = Ref;
1287 LLVM_DEBUG(dbgs() << " -> segment index: " << Ref.Segment << "\n");
1288 } else if (WS.isGlobal()) {
1289 // A "true" Wasm global (currently just __stack_pointer)
1290 if (WS.isDefined())
1291 report_fatal_error("don't yet support defined globals");
1292
1293 // An import; the index was assigned above
1294 LLVM_DEBUG(dbgs() << " -> global index: "
1295 << WasmIndices.find(&WS)->second << "\n");
1296 } else {
1297 assert(WS.isSection());
1298 }
1299 }
1300
1301 // Populate WasmIndices and DataLocations for aliased symbols. We need to
1302 // process these in a separate pass because we need to have processed the
1303 // target of the alias before the alias itself and the symbols are not
1304 // necessarily ordered in this way.
1305 for (const MCSymbol &S : Asm.symbols()) {
1306 if (!S.isVariable())
1307 continue;
1308
1309 assert(S.isDefined());
1310
1311 // Find the target symbol of this weak alias and export that index
1312 const auto &WS = static_cast<const MCSymbolWasm &>(S);
1313 const MCSymbolWasm *ResolvedSym = ResolveSymbol(WS);
1314 LLVM_DEBUG(dbgs() << WS.getName() << ": weak alias of '" << *ResolvedSym
1315 << "'\n");
1316
1317 if (WS.isFunction()) {
1318 assert(WasmIndices.count(ResolvedSym) > 0);
1319 uint32_t WasmIndex = WasmIndices.find(ResolvedSym)->second;
1320 WasmIndices[&WS] = WasmIndex;
1321 LLVM_DEBUG(dbgs() << " -> index:" << WasmIndex << "\n");
1322 } else if (WS.isData()) {
1323 assert(DataLocations.count(ResolvedSym) > 0);
1324 const wasm::WasmDataReference &Ref =
1325 DataLocations.find(ResolvedSym)->second;
1326 DataLocations[&WS] = Ref;
1327 LLVM_DEBUG(dbgs() << " -> index:" << Ref.Segment << "\n");
1328 } else {
1329 report_fatal_error("don't yet support global aliases");
1330 }
1331 }
1332
1333 // Finally, populate the symbol table itself, in its "natural" order.
1334 for (const MCSymbol &S : Asm.symbols()) {
1335 const auto &WS = static_cast<const MCSymbolWasm &>(S);
1336 if (!isInSymtab(WS)) {
1337 WS.setIndex(INVALID_INDEX);
1338 continue;
1339 }
1340 LLVM_DEBUG(dbgs() << "adding to symtab: " << WS << "\n");
1341
1342 uint32_t Flags = 0;
1343 if (WS.isWeak())
1344 Flags |= wasm::WASM_SYMBOL_BINDING_WEAK;
1345 if (WS.isHidden())
1346 Flags |= wasm::WASM_SYMBOL_VISIBILITY_HIDDEN;
1347 if (!WS.isExternal() && WS.isDefined())
1348 Flags |= wasm::WASM_SYMBOL_BINDING_LOCAL;
1349 if (WS.isUndefined())
1350 Flags |= wasm::WASM_SYMBOL_UNDEFINED;
1351
1352 wasm::WasmSymbolInfo Info;
1353 Info.Name = WS.getName();
1354 Info.Kind = WS.getType();
1355 Info.Flags = Flags;
1356 if (!WS.isData()) {
1357 assert(WasmIndices.count(&WS) > 0);
1358 Info.ElementIndex = WasmIndices.find(&WS)->second;
1359 } else if (WS.isDefined()) {
1360 assert(DataLocations.count(&WS) > 0);
1361 Info.DataRef = DataLocations.find(&WS)->second;
1362 }
1363 WS.setIndex(SymbolInfos.size());
1364 SymbolInfos.emplace_back(Info);
1365 }
1366
1367 {
1368 auto HandleReloc = [&](const WasmRelocationEntry &Rel) {
1369 // Functions referenced by a relocation need to put in the table. This is
1370 // purely to make the object file's provisional values readable, and is
1371 // ignored by the linker, which re-calculates the relocations itself.
1372 if (Rel.Type != wasm::R_WEBASSEMBLY_TABLE_INDEX_I32 &&
1373 Rel.Type != wasm::R_WEBASSEMBLY_TABLE_INDEX_SLEB)
1374 return;
1375 assert(Rel.Symbol->isFunction());
1376 const MCSymbolWasm &WS = *ResolveSymbol(*Rel.Symbol);
1377 uint32_t FunctionIndex = WasmIndices.find(&WS)->second;
1378 uint32_t TableIndex = TableElems.size() + kInitialTableOffset;
1379 if (TableIndices.try_emplace(&WS, TableIndex).second) {
1380 LLVM_DEBUG(dbgs() << " -> adding " << WS.getName()
1381 << " to table: " << TableIndex << "\n");
1382 TableElems.push_back(FunctionIndex);
1383 registerFunctionType(WS);
1384 }
1385 };
1386
1387 for (const WasmRelocationEntry &RelEntry : CodeRelocations)
1388 HandleReloc(RelEntry);
1389 for (const WasmRelocationEntry &RelEntry : DataRelocations)
1390 HandleReloc(RelEntry);
1391 }
1392
1393 // Translate .init_array section contents into start functions.
1394 for (const MCSection &S : Asm) {
1395 const auto &WS = static_cast<const MCSectionWasm &>(S);
1396 if (WS.getSectionName().startswith(".fini_array"))
1397 report_fatal_error(".fini_array sections are unsupported");
1398 if (!WS.getSectionName().startswith(".init_array"))
1399 continue;
1400 if (WS.getFragmentList().empty())
1401 continue;
1402
1403 // init_array is expected to contain a single non-empty data fragment
1404 if (WS.getFragmentList().size() != 3)
1405 report_fatal_error("only one .init_array section fragment supported");
1406
1407 auto IT = WS.begin();
1408 const MCFragment &EmptyFrag = *IT;
1409 if (EmptyFrag.getKind() != MCFragment::FT_Data)
1410 report_fatal_error(".init_array section should be aligned");
1411
1412 IT = std::next(IT);
1413 const MCFragment &AlignFrag = *IT;
1414 if (AlignFrag.getKind() != MCFragment::FT_Align)
1415 report_fatal_error(".init_array section should be aligned");
1416 if (cast<MCAlignFragment>(AlignFrag).getAlignment() != (is64Bit() ? 8 : 4))
1417 report_fatal_error(".init_array section should be aligned for pointers");
1418
1419 const MCFragment &Frag = *std::next(IT);
1420 if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data)
1421 report_fatal_error("only data supported in .init_array section");
1422
1423 uint16_t Priority = UINT16_MAX;
1424 unsigned PrefixLength = strlen(".init_array");
1425 if (WS.getSectionName().size() > PrefixLength) {
1426 if (WS.getSectionName()[PrefixLength] != '.')
1427 report_fatal_error(".init_array section priority should start with '.'");
1428 if (WS.getSectionName()
1429 .substr(PrefixLength + 1)
1430 .getAsInteger(10, Priority))
1431 report_fatal_error("invalid .init_array section priority");
1432 }
1433 const auto &DataFrag = cast<MCDataFragment>(Frag);
1434 const SmallVectorImpl<char> &Contents = DataFrag.getContents();
1435 for (const uint8_t *p = (const uint8_t *)Contents.data(),
1436 *end = (const uint8_t *)Contents.data() + Contents.size();
1437 p != end; ++p) {
1438 if (*p != 0)
1439 report_fatal_error("non-symbolic data in .init_array section");
1440 }
1441 for (const MCFixup &Fixup : DataFrag.getFixups()) {
1442 assert(Fixup.getKind() == MCFixup::getKindForSize(is64Bit() ? 8 : 4, false));
1443 const MCExpr *Expr = Fixup.getValue();
1444 auto *Sym = dyn_cast<MCSymbolRefExpr>(Expr);
1445 if (!Sym)
1446 report_fatal_error("fixups in .init_array should be symbol references");
1447 if (Sym->getKind() != MCSymbolRefExpr::VK_WebAssembly_FUNCTION)
1448 report_fatal_error("symbols in .init_array should be for functions");
1449 if (Sym->getSymbol().getIndex() == INVALID_INDEX)
1450 report_fatal_error("symbols in .init_array should exist in symbtab");
1451 InitFuncs.push_back(
1452 std::make_pair(Priority, Sym->getSymbol().getIndex()));
1453 }
1454 }
1455
1456 // Write out the Wasm header.
1457 writeHeader(Asm);
1458
1459 writeTypeSection(FunctionTypes);
1460 writeImportSection(Imports, DataSize, TableElems.size());
1461 writeFunctionSection(Functions);
1462 // Skip the "table" section; we import the table instead.
1463 // Skip the "memory" section; we import the memory instead.
1464 writeGlobalSection();
1465 writeExportSection(Exports);
1466 writeElemSection(TableElems);
1467 writeCodeSection(Asm, Layout, Functions);
1468 writeDataSection();
1469 writeCustomSections(Asm, Layout);
1470 writeLinkingMetaDataSection(SymbolInfos, InitFuncs, Comdats);
1471 writeRelocSection(CodeSectionIndex, "CODE", CodeRelocations);
1472 writeRelocSection(DataSectionIndex, "DATA", DataRelocations);
1473 writeCustomRelocSections();
1474
1475 // TODO: Translate the .comment section to the output.
1476 return W.OS.tell() - StartOffset;
1477 }
1478
1479 std::unique_ptr<MCObjectWriter>
createWasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW,raw_pwrite_stream & OS)1480 llvm::createWasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW,
1481 raw_pwrite_stream &OS) {
1482 return llvm::make_unique<WasmObjectWriter>(std::move(MOTW), OS);
1483 }
1484