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/MC/MCAsmBackend.h" 18 #include "llvm/MC/MCAsmLayout.h" 19 #include "llvm/MC/MCAssembler.h" 20 #include "llvm/MC/MCContext.h" 21 #include "llvm/MC/MCExpr.h" 22 #include "llvm/MC/MCFixupKindInfo.h" 23 #include "llvm/MC/MCObjectWriter.h" 24 #include "llvm/MC/MCSectionWasm.h" 25 #include "llvm/MC/MCSymbolWasm.h" 26 #include "llvm/MC/MCValue.h" 27 #include "llvm/MC/MCWasmObjectWriter.h" 28 #include "llvm/Support/Casting.h" 29 #include "llvm/Support/Debug.h" 30 #include "llvm/Support/ErrorHandling.h" 31 #include "llvm/Support/LEB128.h" 32 #include "llvm/Support/StringSaver.h" 33 #include <vector> 34 35 using namespace llvm; 36 37 #define DEBUG_TYPE "mc" 38 39 namespace { 40 41 // For patching purposes, we need to remember where each section starts, both 42 // for patching up the section size field, and for patching up references to 43 // locations within the section. 44 struct SectionBookkeeping { 45 // Where the size of the section is written. 46 uint64_t SizeOffset; 47 // Where the contents of the section starts (after the header). 48 uint64_t ContentsOffset; 49 }; 50 51 // The signature of a wasm function, in a struct capable of being used as a 52 // DenseMap key. 53 struct WasmFunctionType { 54 // Support empty and tombstone instances, needed by DenseMap. 55 enum { Plain, Empty, Tombstone } State; 56 57 // The return types of the function. 58 SmallVector<wasm::ValType, 1> Returns; 59 60 // The parameter types of the function. 61 SmallVector<wasm::ValType, 4> Params; 62 63 WasmFunctionType() : State(Plain) {} 64 65 bool operator==(const WasmFunctionType &Other) const { 66 return State == Other.State && Returns == Other.Returns && 67 Params == Other.Params; 68 } 69 }; 70 71 // Traits for using WasmFunctionType in a DenseMap. 72 struct WasmFunctionTypeDenseMapInfo { 73 static WasmFunctionType getEmptyKey() { 74 WasmFunctionType FuncTy; 75 FuncTy.State = WasmFunctionType::Empty; 76 return FuncTy; 77 } 78 static WasmFunctionType getTombstoneKey() { 79 WasmFunctionType FuncTy; 80 FuncTy.State = WasmFunctionType::Tombstone; 81 return FuncTy; 82 } 83 static unsigned getHashValue(const WasmFunctionType &FuncTy) { 84 uintptr_t Value = FuncTy.State; 85 for (wasm::ValType Ret : FuncTy.Returns) 86 Value += DenseMapInfo<int32_t>::getHashValue(int32_t(Ret)); 87 for (wasm::ValType Param : FuncTy.Params) 88 Value += DenseMapInfo<int32_t>::getHashValue(int32_t(Param)); 89 return Value; 90 } 91 static bool isEqual(const WasmFunctionType &LHS, 92 const WasmFunctionType &RHS) { 93 return LHS == RHS; 94 } 95 }; 96 97 // A wasm data segment. A wasm binary contains only a single data section 98 // but that can contain many segments, each with their own virtual location 99 // in memory. Each MCSection data created by llvm is modeled as its own 100 // wasm data segment. 101 struct WasmDataSegment { 102 MCSectionWasm *Section; 103 StringRef Name; 104 uint32_t Offset; 105 uint32_t Alignment; 106 uint32_t Flags; 107 SmallVector<char, 4> Data; 108 }; 109 110 // A wasm import to be written into the import section. 111 struct WasmImport { 112 StringRef ModuleName; 113 StringRef FieldName; 114 unsigned Kind; 115 int32_t Type; 116 bool IsMutable; 117 }; 118 119 // A wasm function to be written into the function section. 120 struct WasmFunction { 121 int32_t Type; 122 const MCSymbolWasm *Sym; 123 }; 124 125 // A wasm export to be written into the export section. 126 struct WasmExport { 127 StringRef FieldName; 128 unsigned Kind; 129 uint32_t Index; 130 }; 131 132 // A wasm global to be written into the global section. 133 struct WasmGlobal { 134 wasm::ValType Type; 135 bool IsMutable; 136 bool HasImport; 137 uint64_t InitialValue; 138 uint32_t ImportIndex; 139 }; 140 141 // Information about a single relocation. 142 struct WasmRelocationEntry { 143 uint64_t Offset; // Where is the relocation. 144 const MCSymbolWasm *Symbol; // The symbol to relocate with. 145 int64_t Addend; // A value to add to the symbol. 146 unsigned Type; // The type of the relocation. 147 const MCSectionWasm *FixupSection;// The section the relocation is targeting. 148 149 WasmRelocationEntry(uint64_t Offset, const MCSymbolWasm *Symbol, 150 int64_t Addend, unsigned Type, 151 const MCSectionWasm *FixupSection) 152 : Offset(Offset), Symbol(Symbol), Addend(Addend), Type(Type), 153 FixupSection(FixupSection) {} 154 155 bool hasAddend() const { 156 switch (Type) { 157 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_LEB: 158 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_SLEB: 159 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_I32: 160 return true; 161 default: 162 return false; 163 } 164 } 165 166 void print(raw_ostream &Out) const { 167 Out << "Off=" << Offset << ", Sym=" << *Symbol << ", Addend=" << Addend 168 << ", Type=" << Type 169 << ", FixupSection=" << FixupSection->getSectionName(); 170 } 171 172 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 173 LLVM_DUMP_METHOD void dump() const { print(dbgs()); } 174 #endif 175 }; 176 177 #if !defined(NDEBUG) 178 raw_ostream &operator<<(raw_ostream &OS, const WasmRelocationEntry &Rel) { 179 Rel.print(OS); 180 return OS; 181 } 182 #endif 183 184 class WasmObjectWriter : public MCObjectWriter { 185 /// Helper struct for containing some precomputed information on symbols. 186 struct WasmSymbolData { 187 const MCSymbolWasm *Symbol; 188 StringRef Name; 189 190 // Support lexicographic sorting. 191 bool operator<(const WasmSymbolData &RHS) const { return Name < RHS.Name; } 192 }; 193 194 /// The target specific Wasm writer instance. 195 std::unique_ptr<MCWasmObjectTargetWriter> TargetObjectWriter; 196 197 // Relocations for fixing up references in the code section. 198 std::vector<WasmRelocationEntry> CodeRelocations; 199 200 // Relocations for fixing up references in the data section. 201 std::vector<WasmRelocationEntry> DataRelocations; 202 203 // Index values to use for fixing up call_indirect type indices. 204 // Maps function symbols to the index of the type of the function 205 DenseMap<const MCSymbolWasm *, uint32_t> TypeIndices; 206 // Maps function symbols to the table element index space. Used 207 // for TABLE_INDEX relocation types (i.e. address taken functions). 208 DenseMap<const MCSymbolWasm *, uint32_t> IndirectSymbolIndices; 209 // Maps function/global symbols to the function/global index space. 210 DenseMap<const MCSymbolWasm *, uint32_t> SymbolIndices; 211 212 DenseMap<WasmFunctionType, int32_t, WasmFunctionTypeDenseMapInfo> 213 FunctionTypeIndices; 214 SmallVector<WasmFunctionType, 4> FunctionTypes; 215 SmallVector<WasmGlobal, 4> Globals; 216 unsigned NumGlobalImports = 0; 217 218 // TargetObjectWriter wrappers. 219 bool is64Bit() const { return TargetObjectWriter->is64Bit(); } 220 unsigned getRelocType(const MCValue &Target, const MCFixup &Fixup) const { 221 return TargetObjectWriter->getRelocType(Target, Fixup); 222 } 223 224 void startSection(SectionBookkeeping &Section, unsigned SectionId, 225 const char *Name = nullptr); 226 void endSection(SectionBookkeeping &Section); 227 228 public: 229 WasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 230 raw_pwrite_stream &OS) 231 : MCObjectWriter(OS, /*IsLittleEndian=*/true), 232 TargetObjectWriter(std::move(MOTW)) {} 233 234 private: 235 ~WasmObjectWriter() override; 236 237 void reset() override { 238 CodeRelocations.clear(); 239 DataRelocations.clear(); 240 TypeIndices.clear(); 241 SymbolIndices.clear(); 242 IndirectSymbolIndices.clear(); 243 FunctionTypeIndices.clear(); 244 FunctionTypes.clear(); 245 Globals.clear(); 246 MCObjectWriter::reset(); 247 NumGlobalImports = 0; 248 } 249 250 void writeHeader(const MCAssembler &Asm); 251 252 void recordRelocation(MCAssembler &Asm, const MCAsmLayout &Layout, 253 const MCFragment *Fragment, const MCFixup &Fixup, 254 MCValue Target, uint64_t &FixedValue) override; 255 256 void executePostLayoutBinding(MCAssembler &Asm, 257 const MCAsmLayout &Layout) override; 258 259 void writeObject(MCAssembler &Asm, const MCAsmLayout &Layout) override; 260 261 void writeString(const StringRef Str) { 262 encodeULEB128(Str.size(), getStream()); 263 writeBytes(Str); 264 } 265 266 void writeValueType(wasm::ValType Ty) { 267 encodeSLEB128(int32_t(Ty), getStream()); 268 } 269 270 void writeTypeSection(ArrayRef<WasmFunctionType> FunctionTypes); 271 void writeImportSection(ArrayRef<WasmImport> Imports, uint32_t DataSize, 272 uint32_t NumElements); 273 void writeFunctionSection(ArrayRef<WasmFunction> Functions); 274 void writeGlobalSection(); 275 void writeExportSection(ArrayRef<WasmExport> Exports); 276 void writeElemSection(ArrayRef<uint32_t> TableElems); 277 void writeCodeSection(const MCAssembler &Asm, const MCAsmLayout &Layout, 278 ArrayRef<WasmFunction> Functions); 279 void writeDataSection(ArrayRef<WasmDataSegment> Segments); 280 void writeNameSection(ArrayRef<WasmFunction> Functions, 281 ArrayRef<WasmImport> Imports, 282 uint32_t NumFuncImports); 283 void writeCodeRelocSection(); 284 void writeDataRelocSection(); 285 void writeLinkingMetaDataSection( 286 ArrayRef<WasmDataSegment> Segments, uint32_t DataSize, 287 const SmallVector<std::pair<StringRef, uint32_t>, 4> &SymbolFlags, 288 const SmallVector<std::pair<uint16_t, uint32_t>, 2> &InitFuncs); 289 290 uint32_t getProvisionalValue(const WasmRelocationEntry &RelEntry); 291 void applyRelocations(ArrayRef<WasmRelocationEntry> Relocations, 292 uint64_t ContentsOffset); 293 294 void writeRelocations(ArrayRef<WasmRelocationEntry> Relocations); 295 uint32_t getRelocationIndexValue(const WasmRelocationEntry &RelEntry); 296 uint32_t getFunctionType(const MCSymbolWasm& Symbol); 297 uint32_t registerFunctionType(const MCSymbolWasm& Symbol); 298 }; 299 300 } // end anonymous namespace 301 302 WasmObjectWriter::~WasmObjectWriter() {} 303 304 // Write out a section header and a patchable section size field. 305 void WasmObjectWriter::startSection(SectionBookkeeping &Section, 306 unsigned SectionId, 307 const char *Name) { 308 assert((Name != nullptr) == (SectionId == wasm::WASM_SEC_CUSTOM) && 309 "Only custom sections can have names"); 310 311 DEBUG(dbgs() << "startSection " << SectionId << ": " << Name << "\n"); 312 encodeULEB128(SectionId, getStream()); 313 314 Section.SizeOffset = getStream().tell(); 315 316 // The section size. We don't know the size yet, so reserve enough space 317 // for any 32-bit value; we'll patch it later. 318 encodeULEB128(UINT32_MAX, getStream()); 319 320 // The position where the section starts, for measuring its size. 321 Section.ContentsOffset = getStream().tell(); 322 323 // Custom sections in wasm also have a string identifier. 324 if (SectionId == wasm::WASM_SEC_CUSTOM) { 325 assert(Name); 326 writeString(StringRef(Name)); 327 } 328 } 329 330 // Now that the section is complete and we know how big it is, patch up the 331 // section size field at the start of the section. 332 void WasmObjectWriter::endSection(SectionBookkeeping &Section) { 333 uint64_t Size = getStream().tell() - Section.ContentsOffset; 334 if (uint32_t(Size) != Size) 335 report_fatal_error("section size does not fit in a uint32_t"); 336 337 DEBUG(dbgs() << "endSection size=" << Size << "\n"); 338 339 // Write the final section size to the payload_len field, which follows 340 // the section id byte. 341 uint8_t Buffer[16]; 342 unsigned SizeLen = encodeULEB128(Size, Buffer, 5); 343 assert(SizeLen == 5); 344 getStream().pwrite((char *)Buffer, SizeLen, Section.SizeOffset); 345 } 346 347 // Emit the Wasm header. 348 void WasmObjectWriter::writeHeader(const MCAssembler &Asm) { 349 writeBytes(StringRef(wasm::WasmMagic, sizeof(wasm::WasmMagic))); 350 writeLE32(wasm::WasmVersion); 351 } 352 353 void WasmObjectWriter::executePostLayoutBinding(MCAssembler &Asm, 354 const MCAsmLayout &Layout) { 355 } 356 357 void WasmObjectWriter::recordRelocation(MCAssembler &Asm, 358 const MCAsmLayout &Layout, 359 const MCFragment *Fragment, 360 const MCFixup &Fixup, MCValue Target, 361 uint64_t &FixedValue) { 362 MCAsmBackend &Backend = Asm.getBackend(); 363 bool IsPCRel = Backend.getFixupKindInfo(Fixup.getKind()).Flags & 364 MCFixupKindInfo::FKF_IsPCRel; 365 const auto &FixupSection = cast<MCSectionWasm>(*Fragment->getParent()); 366 uint64_t C = Target.getConstant(); 367 uint64_t FixupOffset = Layout.getFragmentOffset(Fragment) + Fixup.getOffset(); 368 MCContext &Ctx = Asm.getContext(); 369 370 // The .init_array isn't translated as data, so don't do relocations in it. 371 if (FixupSection.getSectionName().startswith(".init_array")) 372 return; 373 374 if (const MCSymbolRefExpr *RefB = Target.getSymB()) { 375 assert(RefB->getKind() == MCSymbolRefExpr::VK_None && 376 "Should not have constructed this"); 377 378 // Let A, B and C being the components of Target and R be the location of 379 // the fixup. If the fixup is not pcrel, we want to compute (A - B + C). 380 // If it is pcrel, we want to compute (A - B + C - R). 381 382 // In general, Wasm has no relocations for -B. It can only represent (A + C) 383 // or (A + C - R). If B = R + K and the relocation is not pcrel, we can 384 // replace B to implement it: (A - R - K + C) 385 if (IsPCRel) { 386 Ctx.reportError( 387 Fixup.getLoc(), 388 "No relocation available to represent this relative expression"); 389 return; 390 } 391 392 const auto &SymB = cast<MCSymbolWasm>(RefB->getSymbol()); 393 394 if (SymB.isUndefined()) { 395 Ctx.reportError(Fixup.getLoc(), 396 Twine("symbol '") + SymB.getName() + 397 "' can not be undefined in a subtraction expression"); 398 return; 399 } 400 401 assert(!SymB.isAbsolute() && "Should have been folded"); 402 const MCSection &SecB = SymB.getSection(); 403 if (&SecB != &FixupSection) { 404 Ctx.reportError(Fixup.getLoc(), 405 "Cannot represent a difference across sections"); 406 return; 407 } 408 409 uint64_t SymBOffset = Layout.getSymbolOffset(SymB); 410 uint64_t K = SymBOffset - FixupOffset; 411 IsPCRel = true; 412 C -= K; 413 } 414 415 // We either rejected the fixup or folded B into C at this point. 416 const MCSymbolRefExpr *RefA = Target.getSymA(); 417 const auto *SymA = RefA ? cast<MCSymbolWasm>(&RefA->getSymbol()) : nullptr; 418 419 if (SymA && SymA->isVariable()) { 420 const MCExpr *Expr = SymA->getVariableValue(); 421 const auto *Inner = cast<MCSymbolRefExpr>(Expr); 422 if (Inner->getKind() == MCSymbolRefExpr::VK_WEAKREF) 423 llvm_unreachable("weakref used in reloc not yet implemented"); 424 } 425 426 // Put any constant offset in an addend. Offsets can be negative, and 427 // LLVM expects wrapping, in contrast to wasm's immediates which can't 428 // be negative and don't wrap. 429 FixedValue = 0; 430 431 if (SymA) 432 SymA->setUsedInReloc(); 433 434 assert(!IsPCRel); 435 assert(SymA); 436 437 unsigned Type = getRelocType(Target, Fixup); 438 439 WasmRelocationEntry Rec(FixupOffset, SymA, C, Type, &FixupSection); 440 DEBUG(dbgs() << "WasmReloc: " << Rec << "\n"); 441 442 if (FixupSection.isWasmData()) 443 DataRelocations.push_back(Rec); 444 else if (FixupSection.getKind().isText()) 445 CodeRelocations.push_back(Rec); 446 else if (!FixupSection.getKind().isMetadata()) 447 // TODO(sbc): Add support for debug sections. 448 llvm_unreachable("unexpected section type"); 449 } 450 451 // Write X as an (unsigned) LEB value at offset Offset in Stream, padded 452 // to allow patching. 453 static void 454 WritePatchableLEB(raw_pwrite_stream &Stream, uint32_t X, uint64_t Offset) { 455 uint8_t Buffer[5]; 456 unsigned SizeLen = encodeULEB128(X, Buffer, 5); 457 assert(SizeLen == 5); 458 Stream.pwrite((char *)Buffer, SizeLen, Offset); 459 } 460 461 // Write X as an signed LEB value at offset Offset in Stream, padded 462 // to allow patching. 463 static void 464 WritePatchableSLEB(raw_pwrite_stream &Stream, int32_t X, uint64_t Offset) { 465 uint8_t Buffer[5]; 466 unsigned SizeLen = encodeSLEB128(X, Buffer, 5); 467 assert(SizeLen == 5); 468 Stream.pwrite((char *)Buffer, SizeLen, Offset); 469 } 470 471 // Write X as a plain integer value at offset Offset in Stream. 472 static void WriteI32(raw_pwrite_stream &Stream, uint32_t X, uint64_t Offset) { 473 uint8_t Buffer[4]; 474 support::endian::write32le(Buffer, X); 475 Stream.pwrite((char *)Buffer, sizeof(Buffer), Offset); 476 } 477 478 static const MCSymbolWasm* ResolveSymbol(const MCSymbolWasm& Symbol) { 479 if (Symbol.isVariable()) { 480 const MCExpr *Expr = Symbol.getVariableValue(); 481 auto *Inner = cast<MCSymbolRefExpr>(Expr); 482 return cast<MCSymbolWasm>(&Inner->getSymbol()); 483 } 484 return &Symbol; 485 } 486 487 // Compute a value to write into the code at the location covered 488 // by RelEntry. This value isn't used by the static linker, since 489 // we have addends; it just serves to make the code more readable 490 // and to make standalone wasm modules directly usable. 491 uint32_t 492 WasmObjectWriter::getProvisionalValue(const WasmRelocationEntry &RelEntry) { 493 const MCSymbolWasm *Sym = ResolveSymbol(*RelEntry.Symbol); 494 495 // For undefined symbols, use a hopefully invalid value. 496 if (!Sym->isDefined(/*SetUsed=*/false)) 497 return UINT32_MAX; 498 499 uint32_t GlobalIndex = SymbolIndices[Sym]; 500 const WasmGlobal& Global = Globals[GlobalIndex - NumGlobalImports]; 501 uint64_t Address = Global.InitialValue + RelEntry.Addend; 502 503 // Ignore overflow. LLVM allows address arithmetic to silently wrap. 504 uint32_t Value = Address; 505 506 return Value; 507 } 508 509 static void addData(SmallVectorImpl<char> &DataBytes, 510 MCSectionWasm &DataSection) { 511 DEBUG(errs() << "addData: " << DataSection.getSectionName() << "\n"); 512 513 DataBytes.resize(alignTo(DataBytes.size(), DataSection.getAlignment())); 514 515 size_t LastFragmentSize = 0; 516 for (const MCFragment &Frag : DataSection) { 517 if (Frag.hasInstructions()) 518 report_fatal_error("only data supported in data sections"); 519 520 if (auto *Align = dyn_cast<MCAlignFragment>(&Frag)) { 521 if (Align->getValueSize() != 1) 522 report_fatal_error("only byte values supported for alignment"); 523 // If nops are requested, use zeros, as this is the data section. 524 uint8_t Value = Align->hasEmitNops() ? 0 : Align->getValue(); 525 uint64_t Size = std::min<uint64_t>(alignTo(DataBytes.size(), 526 Align->getAlignment()), 527 DataBytes.size() + 528 Align->getMaxBytesToEmit()); 529 DataBytes.resize(Size, Value); 530 } else if (auto *Fill = dyn_cast<MCFillFragment>(&Frag)) { 531 int64_t Size; 532 if (!Fill->getSize().evaluateAsAbsolute(Size)) 533 llvm_unreachable("The fill should be an assembler constant"); 534 DataBytes.insert(DataBytes.end(), Size, Fill->getValue()); 535 } else { 536 const auto &DataFrag = cast<MCDataFragment>(Frag); 537 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 538 539 DataBytes.insert(DataBytes.end(), Contents.begin(), Contents.end()); 540 LastFragmentSize = Contents.size(); 541 } 542 } 543 544 // Don't allow empty segments, or segments that end with zero-sized 545 // fragment, otherwise the linker cannot map symbols to a unique 546 // data segment. This can be triggered by zero-sized structs 547 // See: test/MC/WebAssembly/bss.ll 548 if (LastFragmentSize == 0) 549 DataBytes.resize(DataBytes.size() + 1); 550 DEBUG(dbgs() << "addData -> " << DataBytes.size() << "\n"); 551 } 552 553 uint32_t WasmObjectWriter::getRelocationIndexValue( 554 const WasmRelocationEntry &RelEntry) { 555 switch (RelEntry.Type) { 556 case wasm::R_WEBASSEMBLY_TABLE_INDEX_SLEB: 557 case wasm::R_WEBASSEMBLY_TABLE_INDEX_I32: 558 if (!IndirectSymbolIndices.count(RelEntry.Symbol)) 559 report_fatal_error("symbol not found in table index space: " + 560 RelEntry.Symbol->getName()); 561 return IndirectSymbolIndices[RelEntry.Symbol]; 562 case wasm::R_WEBASSEMBLY_FUNCTION_INDEX_LEB: 563 case wasm::R_WEBASSEMBLY_GLOBAL_INDEX_LEB: 564 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_LEB: 565 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_SLEB: 566 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_I32: 567 if (!SymbolIndices.count(RelEntry.Symbol)) 568 report_fatal_error("symbol not found in function/global index space: " + 569 RelEntry.Symbol->getName()); 570 return SymbolIndices[RelEntry.Symbol]; 571 case wasm::R_WEBASSEMBLY_TYPE_INDEX_LEB: 572 if (!TypeIndices.count(RelEntry.Symbol)) 573 report_fatal_error("symbol not found in type index space: " + 574 RelEntry.Symbol->getName()); 575 return TypeIndices[RelEntry.Symbol]; 576 default: 577 llvm_unreachable("invalid relocation type"); 578 } 579 } 580 581 // Apply the portions of the relocation records that we can handle ourselves 582 // directly. 583 void WasmObjectWriter::applyRelocations( 584 ArrayRef<WasmRelocationEntry> Relocations, uint64_t ContentsOffset) { 585 raw_pwrite_stream &Stream = getStream(); 586 for (const WasmRelocationEntry &RelEntry : Relocations) { 587 uint64_t Offset = ContentsOffset + 588 RelEntry.FixupSection->getSectionOffset() + 589 RelEntry.Offset; 590 591 DEBUG(dbgs() << "applyRelocation: " << RelEntry << "\n"); 592 switch (RelEntry.Type) { 593 case wasm::R_WEBASSEMBLY_TABLE_INDEX_SLEB: 594 case wasm::R_WEBASSEMBLY_FUNCTION_INDEX_LEB: 595 case wasm::R_WEBASSEMBLY_TYPE_INDEX_LEB: 596 case wasm::R_WEBASSEMBLY_GLOBAL_INDEX_LEB: { 597 uint32_t Index = getRelocationIndexValue(RelEntry); 598 WritePatchableSLEB(Stream, Index, Offset); 599 break; 600 } 601 case wasm::R_WEBASSEMBLY_TABLE_INDEX_I32: { 602 uint32_t Index = getRelocationIndexValue(RelEntry); 603 WriteI32(Stream, Index, Offset); 604 break; 605 } 606 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_SLEB: { 607 uint32_t Value = getProvisionalValue(RelEntry); 608 WritePatchableSLEB(Stream, Value, Offset); 609 break; 610 } 611 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_LEB: { 612 uint32_t Value = getProvisionalValue(RelEntry); 613 WritePatchableLEB(Stream, Value, Offset); 614 break; 615 } 616 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_I32: { 617 uint32_t Value = getProvisionalValue(RelEntry); 618 WriteI32(Stream, Value, Offset); 619 break; 620 } 621 default: 622 llvm_unreachable("invalid relocation type"); 623 } 624 } 625 } 626 627 // Write out the portions of the relocation records that the linker will 628 // need to handle. 629 void WasmObjectWriter::writeRelocations( 630 ArrayRef<WasmRelocationEntry> Relocations) { 631 raw_pwrite_stream &Stream = getStream(); 632 for (const WasmRelocationEntry& RelEntry : Relocations) { 633 634 uint64_t Offset = RelEntry.Offset + 635 RelEntry.FixupSection->getSectionOffset(); 636 uint32_t Index = getRelocationIndexValue(RelEntry); 637 638 encodeULEB128(RelEntry.Type, Stream); 639 encodeULEB128(Offset, Stream); 640 encodeULEB128(Index, Stream); 641 if (RelEntry.hasAddend()) 642 encodeSLEB128(RelEntry.Addend, Stream); 643 } 644 } 645 646 void WasmObjectWriter::writeTypeSection( 647 ArrayRef<WasmFunctionType> FunctionTypes) { 648 if (FunctionTypes.empty()) 649 return; 650 651 SectionBookkeeping Section; 652 startSection(Section, wasm::WASM_SEC_TYPE); 653 654 encodeULEB128(FunctionTypes.size(), getStream()); 655 656 for (const WasmFunctionType &FuncTy : FunctionTypes) { 657 encodeSLEB128(wasm::WASM_TYPE_FUNC, getStream()); 658 encodeULEB128(FuncTy.Params.size(), getStream()); 659 for (wasm::ValType Ty : FuncTy.Params) 660 writeValueType(Ty); 661 encodeULEB128(FuncTy.Returns.size(), getStream()); 662 for (wasm::ValType Ty : FuncTy.Returns) 663 writeValueType(Ty); 664 } 665 666 endSection(Section); 667 } 668 669 void WasmObjectWriter::writeImportSection(ArrayRef<WasmImport> Imports, 670 uint32_t DataSize, 671 uint32_t NumElements) { 672 if (Imports.empty()) 673 return; 674 675 uint32_t NumPages = (DataSize + wasm::WasmPageSize - 1) / wasm::WasmPageSize; 676 677 SectionBookkeeping Section; 678 startSection(Section, wasm::WASM_SEC_IMPORT); 679 680 encodeULEB128(Imports.size(), getStream()); 681 for (const WasmImport &Import : Imports) { 682 writeString(Import.ModuleName); 683 writeString(Import.FieldName); 684 685 encodeULEB128(Import.Kind, getStream()); 686 687 switch (Import.Kind) { 688 case wasm::WASM_EXTERNAL_FUNCTION: 689 encodeULEB128(Import.Type, getStream()); 690 break; 691 case wasm::WASM_EXTERNAL_GLOBAL: 692 encodeSLEB128(int32_t(Import.Type), getStream()); 693 encodeULEB128(int32_t(Import.IsMutable), getStream()); 694 break; 695 case wasm::WASM_EXTERNAL_MEMORY: 696 encodeULEB128(0, getStream()); // flags 697 encodeULEB128(NumPages, getStream()); // initial 698 break; 699 case wasm::WASM_EXTERNAL_TABLE: 700 encodeSLEB128(int32_t(Import.Type), getStream()); 701 encodeULEB128(0, getStream()); // flags 702 encodeULEB128(NumElements, getStream()); // initial 703 break; 704 default: 705 llvm_unreachable("unsupported import kind"); 706 } 707 } 708 709 endSection(Section); 710 } 711 712 void WasmObjectWriter::writeFunctionSection(ArrayRef<WasmFunction> Functions) { 713 if (Functions.empty()) 714 return; 715 716 SectionBookkeeping Section; 717 startSection(Section, wasm::WASM_SEC_FUNCTION); 718 719 encodeULEB128(Functions.size(), getStream()); 720 for (const WasmFunction &Func : Functions) 721 encodeULEB128(Func.Type, getStream()); 722 723 endSection(Section); 724 } 725 726 void WasmObjectWriter::writeGlobalSection() { 727 if (Globals.empty()) 728 return; 729 730 SectionBookkeeping Section; 731 startSection(Section, wasm::WASM_SEC_GLOBAL); 732 733 encodeULEB128(Globals.size(), getStream()); 734 for (const WasmGlobal &Global : Globals) { 735 writeValueType(Global.Type); 736 write8(Global.IsMutable); 737 738 if (Global.HasImport) { 739 assert(Global.InitialValue == 0); 740 write8(wasm::WASM_OPCODE_GET_GLOBAL); 741 encodeULEB128(Global.ImportIndex, getStream()); 742 } else { 743 assert(Global.ImportIndex == 0); 744 write8(wasm::WASM_OPCODE_I32_CONST); 745 encodeSLEB128(Global.InitialValue, getStream()); // offset 746 } 747 write8(wasm::WASM_OPCODE_END); 748 } 749 750 endSection(Section); 751 } 752 753 void WasmObjectWriter::writeExportSection(ArrayRef<WasmExport> Exports) { 754 if (Exports.empty()) 755 return; 756 757 SectionBookkeeping Section; 758 startSection(Section, wasm::WASM_SEC_EXPORT); 759 760 encodeULEB128(Exports.size(), getStream()); 761 for (const WasmExport &Export : Exports) { 762 writeString(Export.FieldName); 763 encodeSLEB128(Export.Kind, getStream()); 764 encodeULEB128(Export.Index, getStream()); 765 } 766 767 endSection(Section); 768 } 769 770 void WasmObjectWriter::writeElemSection(ArrayRef<uint32_t> TableElems) { 771 if (TableElems.empty()) 772 return; 773 774 SectionBookkeeping Section; 775 startSection(Section, wasm::WASM_SEC_ELEM); 776 777 encodeULEB128(1, getStream()); // number of "segments" 778 encodeULEB128(0, getStream()); // the table index 779 780 // init expr for starting offset 781 write8(wasm::WASM_OPCODE_I32_CONST); 782 encodeSLEB128(0, getStream()); 783 write8(wasm::WASM_OPCODE_END); 784 785 encodeULEB128(TableElems.size(), getStream()); 786 for (uint32_t Elem : TableElems) 787 encodeULEB128(Elem, getStream()); 788 789 endSection(Section); 790 } 791 792 void WasmObjectWriter::writeCodeSection(const MCAssembler &Asm, 793 const MCAsmLayout &Layout, 794 ArrayRef<WasmFunction> Functions) { 795 if (Functions.empty()) 796 return; 797 798 SectionBookkeeping Section; 799 startSection(Section, wasm::WASM_SEC_CODE); 800 801 encodeULEB128(Functions.size(), getStream()); 802 803 for (const WasmFunction &Func : Functions) { 804 auto &FuncSection = static_cast<MCSectionWasm &>(Func.Sym->getSection()); 805 806 int64_t Size = 0; 807 if (!Func.Sym->getSize()->evaluateAsAbsolute(Size, Layout)) 808 report_fatal_error(".size expression must be evaluatable"); 809 810 encodeULEB128(Size, getStream()); 811 FuncSection.setSectionOffset(getStream().tell() - Section.ContentsOffset); 812 Asm.writeSectionData(&FuncSection, Layout); 813 } 814 815 // Apply fixups. 816 applyRelocations(CodeRelocations, Section.ContentsOffset); 817 818 endSection(Section); 819 } 820 821 void WasmObjectWriter::writeDataSection(ArrayRef<WasmDataSegment> Segments) { 822 if (Segments.empty()) 823 return; 824 825 SectionBookkeeping Section; 826 startSection(Section, wasm::WASM_SEC_DATA); 827 828 encodeULEB128(Segments.size(), getStream()); // count 829 830 for (const WasmDataSegment & Segment : Segments) { 831 encodeULEB128(0, getStream()); // memory index 832 write8(wasm::WASM_OPCODE_I32_CONST); 833 encodeSLEB128(Segment.Offset, getStream()); // offset 834 write8(wasm::WASM_OPCODE_END); 835 encodeULEB128(Segment.Data.size(), getStream()); // size 836 Segment.Section->setSectionOffset(getStream().tell() - Section.ContentsOffset); 837 writeBytes(Segment.Data); // data 838 } 839 840 // Apply fixups. 841 applyRelocations(DataRelocations, Section.ContentsOffset); 842 843 endSection(Section); 844 } 845 846 void WasmObjectWriter::writeNameSection( 847 ArrayRef<WasmFunction> Functions, 848 ArrayRef<WasmImport> Imports, 849 unsigned NumFuncImports) { 850 uint32_t TotalFunctions = NumFuncImports + Functions.size(); 851 if (TotalFunctions == 0) 852 return; 853 854 SectionBookkeeping Section; 855 startSection(Section, wasm::WASM_SEC_CUSTOM, "name"); 856 SectionBookkeeping SubSection; 857 startSection(SubSection, wasm::WASM_NAMES_FUNCTION); 858 859 encodeULEB128(TotalFunctions, getStream()); 860 uint32_t Index = 0; 861 for (const WasmImport &Import : Imports) { 862 if (Import.Kind == wasm::WASM_EXTERNAL_FUNCTION) { 863 encodeULEB128(Index, getStream()); 864 writeString(Import.FieldName); 865 ++Index; 866 } 867 } 868 for (const WasmFunction &Func : Functions) { 869 encodeULEB128(Index, getStream()); 870 writeString(Func.Sym->getName()); 871 ++Index; 872 } 873 874 endSection(SubSection); 875 endSection(Section); 876 } 877 878 void WasmObjectWriter::writeCodeRelocSection() { 879 // See: https://github.com/WebAssembly/tool-conventions/blob/master/Linking.md 880 // for descriptions of the reloc sections. 881 882 if (CodeRelocations.empty()) 883 return; 884 885 SectionBookkeeping Section; 886 startSection(Section, wasm::WASM_SEC_CUSTOM, "reloc.CODE"); 887 888 encodeULEB128(wasm::WASM_SEC_CODE, getStream()); 889 encodeULEB128(CodeRelocations.size(), getStream()); 890 891 writeRelocations(CodeRelocations); 892 893 endSection(Section); 894 } 895 896 void WasmObjectWriter::writeDataRelocSection() { 897 // See: https://github.com/WebAssembly/tool-conventions/blob/master/Linking.md 898 // for descriptions of the reloc sections. 899 900 if (DataRelocations.empty()) 901 return; 902 903 SectionBookkeeping Section; 904 startSection(Section, wasm::WASM_SEC_CUSTOM, "reloc.DATA"); 905 906 encodeULEB128(wasm::WASM_SEC_DATA, getStream()); 907 encodeULEB128(DataRelocations.size(), getStream()); 908 909 writeRelocations(DataRelocations); 910 911 endSection(Section); 912 } 913 914 void WasmObjectWriter::writeLinkingMetaDataSection( 915 ArrayRef<WasmDataSegment> Segments, uint32_t DataSize, 916 const SmallVector<std::pair<StringRef, uint32_t>, 4> &SymbolFlags, 917 const SmallVector<std::pair<uint16_t, uint32_t>, 2> &InitFuncs) { 918 SectionBookkeeping Section; 919 startSection(Section, wasm::WASM_SEC_CUSTOM, "linking"); 920 SectionBookkeeping SubSection; 921 922 if (SymbolFlags.size() != 0) { 923 startSection(SubSection, wasm::WASM_SYMBOL_INFO); 924 encodeULEB128(SymbolFlags.size(), getStream()); 925 for (auto Pair: SymbolFlags) { 926 writeString(Pair.first); 927 encodeULEB128(Pair.second, getStream()); 928 } 929 endSection(SubSection); 930 } 931 932 if (DataSize > 0) { 933 startSection(SubSection, wasm::WASM_DATA_SIZE); 934 encodeULEB128(DataSize, getStream()); 935 endSection(SubSection); 936 } 937 938 if (Segments.size()) { 939 startSection(SubSection, wasm::WASM_SEGMENT_INFO); 940 encodeULEB128(Segments.size(), getStream()); 941 for (const WasmDataSegment &Segment : Segments) { 942 writeString(Segment.Name); 943 encodeULEB128(Segment.Alignment, getStream()); 944 encodeULEB128(Segment.Flags, getStream()); 945 } 946 endSection(SubSection); 947 } 948 949 if (!InitFuncs.empty()) { 950 startSection(SubSection, wasm::WASM_INIT_FUNCS); 951 encodeULEB128(InitFuncs.size(), getStream()); 952 for (auto &StartFunc : InitFuncs) { 953 encodeULEB128(StartFunc.first, getStream()); // priority 954 encodeULEB128(StartFunc.second, getStream()); // function index 955 } 956 endSection(SubSection); 957 } 958 959 endSection(Section); 960 } 961 962 uint32_t WasmObjectWriter::getFunctionType(const MCSymbolWasm& Symbol) { 963 assert(Symbol.isFunction()); 964 assert(TypeIndices.count(&Symbol)); 965 return TypeIndices[&Symbol]; 966 } 967 968 uint32_t WasmObjectWriter::registerFunctionType(const MCSymbolWasm& Symbol) { 969 assert(Symbol.isFunction()); 970 971 WasmFunctionType F; 972 const MCSymbolWasm* ResolvedSym = ResolveSymbol(Symbol); 973 F.Returns = ResolvedSym->getReturns(); 974 F.Params = ResolvedSym->getParams(); 975 976 auto Pair = 977 FunctionTypeIndices.insert(std::make_pair(F, FunctionTypes.size())); 978 if (Pair.second) 979 FunctionTypes.push_back(F); 980 TypeIndices[&Symbol] = Pair.first->second; 981 982 DEBUG(dbgs() << "registerFunctionType: " << Symbol << " new:" << Pair.second << "\n"); 983 DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n"); 984 return Pair.first->second; 985 } 986 987 void WasmObjectWriter::writeObject(MCAssembler &Asm, 988 const MCAsmLayout &Layout) { 989 DEBUG(dbgs() << "WasmObjectWriter::writeObject\n"); 990 MCContext &Ctx = Asm.getContext(); 991 wasm::ValType PtrType = is64Bit() ? wasm::ValType::I64 : wasm::ValType::I32; 992 993 // Collect information from the available symbols. 994 SmallVector<WasmFunction, 4> Functions; 995 SmallVector<uint32_t, 4> TableElems; 996 SmallVector<WasmImport, 4> Imports; 997 SmallVector<WasmExport, 4> Exports; 998 SmallVector<std::pair<StringRef, uint32_t>, 4> SymbolFlags; 999 SmallVector<std::pair<uint16_t, uint32_t>, 2> InitFuncs; 1000 unsigned NumFuncImports = 0; 1001 SmallVector<WasmDataSegment, 4> DataSegments; 1002 uint32_t DataSize = 0; 1003 1004 // In the special .global_variables section, we've encoded global 1005 // variables used by the function. Translate them into the Globals 1006 // list. 1007 MCSectionWasm *GlobalVars = 1008 Ctx.getWasmSection(".global_variables", SectionKind::getMetadata()); 1009 if (!GlobalVars->getFragmentList().empty()) { 1010 if (GlobalVars->getFragmentList().size() != 1) 1011 report_fatal_error("only one .global_variables fragment supported"); 1012 const MCFragment &Frag = *GlobalVars->begin(); 1013 if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data) 1014 report_fatal_error("only data supported in .global_variables"); 1015 const auto &DataFrag = cast<MCDataFragment>(Frag); 1016 if (!DataFrag.getFixups().empty()) 1017 report_fatal_error("fixups not supported in .global_variables"); 1018 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 1019 for (const uint8_t *p = (const uint8_t *)Contents.data(), 1020 *end = (const uint8_t *)Contents.data() + Contents.size(); 1021 p != end; ) { 1022 WasmGlobal G; 1023 if (end - p < 3) 1024 report_fatal_error("truncated global variable encoding"); 1025 G.Type = wasm::ValType(int8_t(*p++)); 1026 G.IsMutable = bool(*p++); 1027 G.HasImport = bool(*p++); 1028 if (G.HasImport) { 1029 G.InitialValue = 0; 1030 1031 WasmImport Import; 1032 Import.ModuleName = (const char *)p; 1033 const uint8_t *nul = (const uint8_t *)memchr(p, '\0', end - p); 1034 if (!nul) 1035 report_fatal_error("global module name must be nul-terminated"); 1036 p = nul + 1; 1037 nul = (const uint8_t *)memchr(p, '\0', end - p); 1038 if (!nul) 1039 report_fatal_error("global base name must be nul-terminated"); 1040 Import.FieldName = (const char *)p; 1041 p = nul + 1; 1042 1043 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1044 Import.Type = int32_t(G.Type); 1045 1046 G.ImportIndex = NumGlobalImports; 1047 ++NumGlobalImports; 1048 1049 Imports.push_back(Import); 1050 } else { 1051 unsigned n; 1052 G.InitialValue = decodeSLEB128(p, &n); 1053 G.ImportIndex = 0; 1054 if ((ptrdiff_t)n > end - p) 1055 report_fatal_error("global initial value must be valid SLEB128"); 1056 p += n; 1057 } 1058 Globals.push_back(G); 1059 } 1060 } 1061 1062 // For now, always emit the memory import, since loads and stores are not 1063 // valid without it. In the future, we could perhaps be more clever and omit 1064 // it if there are no loads or stores. 1065 MCSymbolWasm *MemorySym = 1066 cast<MCSymbolWasm>(Ctx.getOrCreateSymbol("__linear_memory")); 1067 WasmImport MemImport; 1068 MemImport.ModuleName = MemorySym->getModuleName(); 1069 MemImport.FieldName = MemorySym->getName(); 1070 MemImport.Kind = wasm::WASM_EXTERNAL_MEMORY; 1071 Imports.push_back(MemImport); 1072 1073 // For now, always emit the table section, since indirect calls are not 1074 // valid without it. In the future, we could perhaps be more clever and omit 1075 // it if there are no indirect calls. 1076 MCSymbolWasm *TableSym = 1077 cast<MCSymbolWasm>(Ctx.getOrCreateSymbol("__indirect_function_table")); 1078 WasmImport TableImport; 1079 TableImport.ModuleName = TableSym->getModuleName(); 1080 TableImport.FieldName = TableSym->getName(); 1081 TableImport.Kind = wasm::WASM_EXTERNAL_TABLE; 1082 TableImport.Type = wasm::WASM_TYPE_ANYFUNC; 1083 Imports.push_back(TableImport); 1084 1085 // Populate FunctionTypeIndices and Imports. 1086 for (const MCSymbol &S : Asm.symbols()) { 1087 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1088 1089 // Register types for all functions, including those with private linkage 1090 // (making them 1091 // because wasm always needs a type signature. 1092 if (WS.isFunction()) 1093 registerFunctionType(WS); 1094 1095 if (WS.isTemporary()) 1096 continue; 1097 1098 // If the symbol is not defined in this translation unit, import it. 1099 if (!WS.isDefined(/*SetUsed=*/false) || WS.isVariable()) { 1100 WasmImport Import; 1101 Import.ModuleName = WS.getModuleName(); 1102 Import.FieldName = WS.getName(); 1103 1104 if (WS.isFunction()) { 1105 Import.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1106 Import.Type = getFunctionType(WS); 1107 SymbolIndices[&WS] = NumFuncImports; 1108 ++NumFuncImports; 1109 } else { 1110 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1111 Import.Type = int32_t(PtrType); 1112 Import.IsMutable = false; 1113 SymbolIndices[&WS] = NumGlobalImports; 1114 1115 // If this global is the stack pointer, make it mutable. 1116 if (WS.getName() == "__stack_pointer") 1117 Import.IsMutable = true; 1118 1119 ++NumGlobalImports; 1120 } 1121 1122 Imports.push_back(Import); 1123 } 1124 } 1125 1126 for (MCSection &Sec : Asm) { 1127 auto &Section = static_cast<MCSectionWasm &>(Sec); 1128 if (!Section.isWasmData()) 1129 continue; 1130 1131 // .init_array sections are handled specially elsewhere. 1132 if (cast<MCSectionWasm>(Sec).getSectionName().startswith(".init_array")) 1133 continue; 1134 1135 DataSize = alignTo(DataSize, Section.getAlignment()); 1136 DataSegments.emplace_back(); 1137 WasmDataSegment &Segment = DataSegments.back(); 1138 Segment.Name = Section.getSectionName(); 1139 Segment.Offset = DataSize; 1140 Segment.Section = &Section; 1141 addData(Segment.Data, Section); 1142 Segment.Alignment = Section.getAlignment(); 1143 Segment.Flags = 0; 1144 DataSize += Segment.Data.size(); 1145 Section.setMemoryOffset(Segment.Offset); 1146 } 1147 1148 // Handle regular defined and undefined symbols. 1149 for (const MCSymbol &S : Asm.symbols()) { 1150 // Ignore unnamed temporary symbols, which aren't ever exported, imported, 1151 // or used in relocations. 1152 if (S.isTemporary() && S.getName().empty()) 1153 continue; 1154 1155 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1156 DEBUG(dbgs() << "MCSymbol: '" << S << "'" 1157 << " isDefined=" << S.isDefined() << " isExternal=" 1158 << S.isExternal() << " isTemporary=" << S.isTemporary() 1159 << " isFunction=" << WS.isFunction() 1160 << " isWeak=" << WS.isWeak() 1161 << " isHidden=" << WS.isHidden() 1162 << " isVariable=" << WS.isVariable() << "\n"); 1163 1164 if (WS.isWeak() || WS.isHidden()) { 1165 uint32_t Flags = (WS.isWeak() ? wasm::WASM_SYMBOL_BINDING_WEAK : 0) | 1166 (WS.isHidden() ? wasm::WASM_SYMBOL_VISIBILITY_HIDDEN : 0); 1167 SymbolFlags.emplace_back(WS.getName(), Flags); 1168 } 1169 1170 if (WS.isVariable()) 1171 continue; 1172 1173 unsigned Index; 1174 1175 if (WS.isFunction()) { 1176 if (WS.isDefined(/*SetUsed=*/false)) { 1177 if (WS.getOffset() != 0) 1178 report_fatal_error( 1179 "function sections must contain one function each"); 1180 1181 if (WS.getSize() == 0) 1182 report_fatal_error( 1183 "function symbols must have a size set with .size"); 1184 1185 // A definition. Take the next available index. 1186 Index = NumFuncImports + Functions.size(); 1187 1188 // Prepare the function. 1189 WasmFunction Func; 1190 Func.Type = getFunctionType(WS); 1191 Func.Sym = &WS; 1192 SymbolIndices[&WS] = Index; 1193 Functions.push_back(Func); 1194 } else { 1195 // An import; the index was assigned above. 1196 Index = SymbolIndices.find(&WS)->second; 1197 } 1198 1199 DEBUG(dbgs() << " -> function index: " << Index << "\n"); 1200 } else { 1201 if (WS.isTemporary() && !WS.getSize()) 1202 continue; 1203 1204 if (!WS.isDefined(/*SetUsed=*/false)) 1205 continue; 1206 1207 if (!WS.getSize()) 1208 report_fatal_error("data symbols must have a size set with .size: " + 1209 WS.getName()); 1210 1211 int64_t Size = 0; 1212 if (!WS.getSize()->evaluateAsAbsolute(Size, Layout)) 1213 report_fatal_error(".size expression must be evaluatable"); 1214 1215 // For each global, prepare a corresponding wasm global holding its 1216 // address. For externals these will also be named exports. 1217 Index = NumGlobalImports + Globals.size(); 1218 auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection()); 1219 1220 WasmGlobal Global; 1221 Global.Type = PtrType; 1222 Global.IsMutable = false; 1223 Global.HasImport = false; 1224 Global.InitialValue = DataSection.getMemoryOffset() + Layout.getSymbolOffset(WS); 1225 Global.ImportIndex = 0; 1226 SymbolIndices[&WS] = Index; 1227 DEBUG(dbgs() << " -> global index: " << Index << "\n"); 1228 Globals.push_back(Global); 1229 } 1230 1231 // If the symbol is visible outside this translation unit, export it. 1232 if (WS.isDefined(/*SetUsed=*/false)) { 1233 WasmExport Export; 1234 Export.FieldName = WS.getName(); 1235 Export.Index = Index; 1236 if (WS.isFunction()) 1237 Export.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1238 else 1239 Export.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1240 DEBUG(dbgs() << " -> export " << Exports.size() << "\n"); 1241 Exports.push_back(Export); 1242 if (!WS.isExternal()) 1243 SymbolFlags.emplace_back(WS.getName(), wasm::WASM_SYMBOL_BINDING_LOCAL); 1244 } 1245 } 1246 1247 // Handle weak aliases. We need to process these in a separate pass because 1248 // we need to have processed the target of the alias before the alias itself 1249 // and the symbols are not necessarily ordered in this way. 1250 for (const MCSymbol &S : Asm.symbols()) { 1251 if (!S.isVariable()) 1252 continue; 1253 1254 assert(S.isDefined(/*SetUsed=*/false)); 1255 1256 // Find the target symbol of this weak alias and export that index 1257 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1258 const MCSymbolWasm *ResolvedSym = ResolveSymbol(WS); 1259 DEBUG(dbgs() << WS.getName() << ": weak alias of '" << *ResolvedSym << "'\n"); 1260 assert(SymbolIndices.count(ResolvedSym) > 0); 1261 uint32_t Index = SymbolIndices.find(ResolvedSym)->second; 1262 DEBUG(dbgs() << " -> index:" << Index << "\n"); 1263 1264 WasmExport Export; 1265 Export.FieldName = WS.getName(); 1266 Export.Index = Index; 1267 if (WS.isFunction()) 1268 Export.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1269 else 1270 Export.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1271 DEBUG(dbgs() << " -> export " << Exports.size() << "\n"); 1272 Exports.push_back(Export); 1273 1274 if (!WS.isExternal()) 1275 SymbolFlags.emplace_back(WS.getName(), wasm::WASM_SYMBOL_BINDING_LOCAL); 1276 } 1277 1278 { 1279 auto HandleReloc = [&](const WasmRelocationEntry &Rel) { 1280 // Functions referenced by a relocation need to prepared to be called 1281 // indirectly. 1282 const MCSymbolWasm& WS = *Rel.Symbol; 1283 if (WS.isFunction() && IndirectSymbolIndices.count(&WS) == 0) { 1284 switch (Rel.Type) { 1285 case wasm::R_WEBASSEMBLY_TABLE_INDEX_I32: 1286 case wasm::R_WEBASSEMBLY_TABLE_INDEX_SLEB: 1287 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_I32: 1288 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_SLEB: { 1289 uint32_t Index = SymbolIndices.find(&WS)->second; 1290 IndirectSymbolIndices[&WS] = TableElems.size(); 1291 DEBUG(dbgs() << " -> adding to table: " << TableElems.size() << "\n"); 1292 TableElems.push_back(Index); 1293 registerFunctionType(WS); 1294 break; 1295 } 1296 default: 1297 break; 1298 } 1299 } 1300 }; 1301 1302 for (const WasmRelocationEntry &RelEntry : CodeRelocations) 1303 HandleReloc(RelEntry); 1304 for (const WasmRelocationEntry &RelEntry : DataRelocations) 1305 HandleReloc(RelEntry); 1306 } 1307 1308 // Translate .init_array section contents into start functions. 1309 for (const MCSection &S : Asm) { 1310 const auto &WS = static_cast<const MCSectionWasm &>(S); 1311 if (WS.getSectionName().startswith(".fini_array")) 1312 report_fatal_error(".fini_array sections are unsupported"); 1313 if (!WS.getSectionName().startswith(".init_array")) 1314 continue; 1315 if (WS.getFragmentList().empty()) 1316 continue; 1317 if (WS.getFragmentList().size() != 2) 1318 report_fatal_error("only one .init_array section fragment supported"); 1319 const MCFragment &AlignFrag = *WS.begin(); 1320 if (AlignFrag.getKind() != MCFragment::FT_Align) 1321 report_fatal_error(".init_array section should be aligned"); 1322 if (cast<MCAlignFragment>(AlignFrag).getAlignment() != (is64Bit() ? 8 : 4)) 1323 report_fatal_error(".init_array section should be aligned for pointers"); 1324 const MCFragment &Frag = *std::next(WS.begin()); 1325 if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data) 1326 report_fatal_error("only data supported in .init_array section"); 1327 uint16_t Priority = UINT16_MAX; 1328 if (WS.getSectionName().size() != 11) { 1329 if (WS.getSectionName()[11] != '.') 1330 report_fatal_error(".init_array section priority should start with '.'"); 1331 if (WS.getSectionName().substr(12).getAsInteger(10, Priority)) 1332 report_fatal_error("invalid .init_array section priority"); 1333 } 1334 const auto &DataFrag = cast<MCDataFragment>(Frag); 1335 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 1336 for (const uint8_t *p = (const uint8_t *)Contents.data(), 1337 *end = (const uint8_t *)Contents.data() + Contents.size(); 1338 p != end; ++p) { 1339 if (*p != 0) 1340 report_fatal_error("non-symbolic data in .init_array section"); 1341 } 1342 for (const MCFixup &Fixup : DataFrag.getFixups()) { 1343 assert(Fixup.getKind() == MCFixup::getKindForSize(is64Bit() ? 8 : 4, false)); 1344 const MCExpr *Expr = Fixup.getValue(); 1345 auto *Sym = dyn_cast<MCSymbolRefExpr>(Expr); 1346 if (!Sym) 1347 report_fatal_error("fixups in .init_array should be symbol references"); 1348 if (Sym->getKind() != MCSymbolRefExpr::VK_WebAssembly_FUNCTION) 1349 report_fatal_error("symbols in .init_array should be for functions"); 1350 auto I = SymbolIndices.find(cast<MCSymbolWasm>(&Sym->getSymbol())); 1351 if (I == SymbolIndices.end()) 1352 report_fatal_error("symbols in .init_array should be defined"); 1353 uint32_t Index = I->second; 1354 InitFuncs.push_back(std::make_pair(Priority, Index)); 1355 } 1356 } 1357 1358 // Write out the Wasm header. 1359 writeHeader(Asm); 1360 1361 writeTypeSection(FunctionTypes); 1362 writeImportSection(Imports, DataSize, TableElems.size()); 1363 writeFunctionSection(Functions); 1364 // Skip the "table" section; we import the table instead. 1365 // Skip the "memory" section; we import the memory instead. 1366 writeGlobalSection(); 1367 writeExportSection(Exports); 1368 writeElemSection(TableElems); 1369 writeCodeSection(Asm, Layout, Functions); 1370 writeDataSection(DataSegments); 1371 writeNameSection(Functions, Imports, NumFuncImports); 1372 writeCodeRelocSection(); 1373 writeDataRelocSection(); 1374 writeLinkingMetaDataSection(DataSegments, DataSize, SymbolFlags, 1375 InitFuncs); 1376 1377 // TODO: Translate the .comment section to the output. 1378 // TODO: Translate debug sections to the output. 1379 } 1380 1381 std::unique_ptr<MCObjectWriter> 1382 llvm::createWasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 1383 raw_pwrite_stream &OS) { 1384 // FIXME: Can't use make_unique<WasmObjectWriter>(...) as WasmObjectWriter's 1385 // destructor is private. Is that necessary? 1386 return std::unique_ptr<MCObjectWriter>( 1387 new WasmObjectWriter(std::move(MOTW), OS)); 1388 } 1389