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 // Went we ceate the indirect function table we start at 1, so that there is 42 // and emtpy slot at 0 and therefore calling a null function pointer will trap. 43 static const uint32_t kInitialTableOffset = 1; 44 45 // For patching purposes, we need to remember where each section starts, both 46 // for patching up the section size field, and for patching up references to 47 // locations within the section. 48 struct SectionBookkeeping { 49 // Where the size of the section is written. 50 uint64_t SizeOffset; 51 // Where the contents of the section starts (after the header). 52 uint64_t ContentsOffset; 53 }; 54 55 // The signature of a wasm function, in a struct capable of being used as a 56 // DenseMap key. 57 struct WasmFunctionType { 58 // Support empty and tombstone instances, needed by DenseMap. 59 enum { Plain, Empty, Tombstone } State; 60 61 // The return types of the function. 62 SmallVector<wasm::ValType, 1> Returns; 63 64 // The parameter types of the function. 65 SmallVector<wasm::ValType, 4> Params; 66 67 WasmFunctionType() : State(Plain) {} 68 69 bool operator==(const WasmFunctionType &Other) const { 70 return State == Other.State && Returns == Other.Returns && 71 Params == Other.Params; 72 } 73 }; 74 75 // Traits for using WasmFunctionType in a DenseMap. 76 struct WasmFunctionTypeDenseMapInfo { 77 static WasmFunctionType getEmptyKey() { 78 WasmFunctionType FuncTy; 79 FuncTy.State = WasmFunctionType::Empty; 80 return FuncTy; 81 } 82 static WasmFunctionType getTombstoneKey() { 83 WasmFunctionType FuncTy; 84 FuncTy.State = WasmFunctionType::Tombstone; 85 return FuncTy; 86 } 87 static unsigned getHashValue(const WasmFunctionType &FuncTy) { 88 uintptr_t Value = FuncTy.State; 89 for (wasm::ValType Ret : FuncTy.Returns) 90 Value += DenseMapInfo<int32_t>::getHashValue(int32_t(Ret)); 91 for (wasm::ValType Param : FuncTy.Params) 92 Value += DenseMapInfo<int32_t>::getHashValue(int32_t(Param)); 93 return Value; 94 } 95 static bool isEqual(const WasmFunctionType &LHS, 96 const WasmFunctionType &RHS) { 97 return LHS == RHS; 98 } 99 }; 100 101 // A wasm data segment. A wasm binary contains only a single data section 102 // but that can contain many segments, each with their own virtual location 103 // in memory. Each MCSection data created by llvm is modeled as its own 104 // wasm data segment. 105 struct WasmDataSegment { 106 MCSectionWasm *Section; 107 StringRef Name; 108 uint32_t Offset; 109 uint32_t Alignment; 110 uint32_t Flags; 111 SmallVector<char, 4> Data; 112 }; 113 114 // A wasm import to be written into the import section. 115 struct WasmImport { 116 StringRef ModuleName; 117 StringRef FieldName; 118 unsigned Kind; 119 int32_t Type; 120 bool IsMutable; 121 }; 122 123 // A wasm function to be written into the function section. 124 struct WasmFunction { 125 int32_t Type; 126 const MCSymbolWasm *Sym; 127 }; 128 129 // A wasm export to be written into the export section. 130 struct WasmExport { 131 StringRef FieldName; 132 unsigned Kind; 133 uint32_t Index; 134 }; 135 136 // A wasm global to be written into the global section. 137 struct WasmGlobal { 138 wasm::ValType Type; 139 bool IsMutable; 140 bool HasImport; 141 uint64_t InitialValue; 142 uint32_t ImportIndex; 143 }; 144 145 // Information about a single item which is part of a COMDAT. For each data 146 // segment or function which is in the COMDAT, there is a corresponding 147 // WasmComdatEntry. 148 struct WasmComdatEntry { 149 unsigned Kind; 150 uint32_t Index; 151 }; 152 153 // Information about a single relocation. 154 struct WasmRelocationEntry { 155 uint64_t Offset; // Where is the relocation. 156 const MCSymbolWasm *Symbol; // The symbol to relocate with. 157 int64_t Addend; // A value to add to the symbol. 158 unsigned Type; // The type of the relocation. 159 const MCSectionWasm *FixupSection;// The section the relocation is targeting. 160 161 WasmRelocationEntry(uint64_t Offset, const MCSymbolWasm *Symbol, 162 int64_t Addend, unsigned Type, 163 const MCSectionWasm *FixupSection) 164 : Offset(Offset), Symbol(Symbol), Addend(Addend), Type(Type), 165 FixupSection(FixupSection) {} 166 167 bool hasAddend() const { 168 switch (Type) { 169 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_LEB: 170 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_SLEB: 171 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_I32: 172 return true; 173 default: 174 return false; 175 } 176 } 177 178 void print(raw_ostream &Out) const { 179 Out << "Off=" << Offset << ", Sym=" << *Symbol << ", Addend=" << Addend 180 << ", Type=" << Type 181 << ", FixupSection=" << FixupSection->getSectionName(); 182 } 183 184 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 185 LLVM_DUMP_METHOD void dump() const { print(dbgs()); } 186 #endif 187 }; 188 189 #if !defined(NDEBUG) 190 raw_ostream &operator<<(raw_ostream &OS, const WasmRelocationEntry &Rel) { 191 Rel.print(OS); 192 return OS; 193 } 194 #endif 195 196 class WasmObjectWriter : public MCObjectWriter { 197 /// Helper struct for containing some precomputed information on symbols. 198 struct WasmSymbolData { 199 const MCSymbolWasm *Symbol; 200 StringRef Name; 201 202 // Support lexicographic sorting. 203 bool operator<(const WasmSymbolData &RHS) const { return Name < RHS.Name; } 204 }; 205 206 /// The target specific Wasm writer instance. 207 std::unique_ptr<MCWasmObjectTargetWriter> TargetObjectWriter; 208 209 // Relocations for fixing up references in the code section. 210 std::vector<WasmRelocationEntry> CodeRelocations; 211 212 // Relocations for fixing up references in the data section. 213 std::vector<WasmRelocationEntry> DataRelocations; 214 215 // Index values to use for fixing up call_indirect type indices. 216 // Maps function symbols to the index of the type of the function 217 DenseMap<const MCSymbolWasm *, uint32_t> TypeIndices; 218 // Maps function symbols to the table element index space. Used 219 // for TABLE_INDEX relocation types (i.e. address taken functions). 220 DenseMap<const MCSymbolWasm *, uint32_t> IndirectSymbolIndices; 221 // Maps function/global symbols to the function/global index space. 222 DenseMap<const MCSymbolWasm *, uint32_t> SymbolIndices; 223 224 DenseMap<WasmFunctionType, int32_t, WasmFunctionTypeDenseMapInfo> 225 FunctionTypeIndices; 226 SmallVector<WasmFunctionType, 4> FunctionTypes; 227 SmallVector<WasmGlobal, 4> Globals; 228 unsigned NumFunctionImports = 0; 229 unsigned NumGlobalImports = 0; 230 231 // TargetObjectWriter wrappers. 232 bool is64Bit() const { return TargetObjectWriter->is64Bit(); } 233 unsigned getRelocType(const MCValue &Target, const MCFixup &Fixup) const { 234 return TargetObjectWriter->getRelocType(Target, Fixup); 235 } 236 237 void startSection(SectionBookkeeping &Section, unsigned SectionId, 238 const char *Name = nullptr); 239 void endSection(SectionBookkeeping &Section); 240 241 public: 242 WasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 243 raw_pwrite_stream &OS) 244 : MCObjectWriter(OS, /*IsLittleEndian=*/true), 245 TargetObjectWriter(std::move(MOTW)) {} 246 247 ~WasmObjectWriter() override; 248 249 private: 250 void reset() override { 251 CodeRelocations.clear(); 252 DataRelocations.clear(); 253 TypeIndices.clear(); 254 SymbolIndices.clear(); 255 IndirectSymbolIndices.clear(); 256 FunctionTypeIndices.clear(); 257 FunctionTypes.clear(); 258 Globals.clear(); 259 MCObjectWriter::reset(); 260 NumFunctionImports = 0; 261 NumGlobalImports = 0; 262 } 263 264 void writeHeader(const MCAssembler &Asm); 265 266 void recordRelocation(MCAssembler &Asm, const MCAsmLayout &Layout, 267 const MCFragment *Fragment, const MCFixup &Fixup, 268 MCValue Target, uint64_t &FixedValue) override; 269 270 void executePostLayoutBinding(MCAssembler &Asm, 271 const MCAsmLayout &Layout) override; 272 273 void writeObject(MCAssembler &Asm, const MCAsmLayout &Layout) override; 274 275 void writeString(const StringRef Str) { 276 encodeULEB128(Str.size(), getStream()); 277 writeBytes(Str); 278 } 279 280 void writeValueType(wasm::ValType Ty) { 281 encodeSLEB128(int32_t(Ty), getStream()); 282 } 283 284 void writeTypeSection(ArrayRef<WasmFunctionType> FunctionTypes); 285 void writeImportSection(ArrayRef<WasmImport> Imports, uint32_t DataSize, 286 uint32_t NumElements); 287 void writeFunctionSection(ArrayRef<WasmFunction> Functions); 288 void writeGlobalSection(); 289 void writeExportSection(ArrayRef<WasmExport> Exports); 290 void writeElemSection(ArrayRef<uint32_t> TableElems); 291 void writeCodeSection(const MCAssembler &Asm, const MCAsmLayout &Layout, 292 ArrayRef<WasmFunction> Functions); 293 void writeDataSection(ArrayRef<WasmDataSegment> Segments); 294 void writeCodeRelocSection(); 295 void writeDataRelocSection(); 296 void writeLinkingMetaDataSection( 297 ArrayRef<WasmDataSegment> Segments, uint32_t DataSize, 298 ArrayRef<std::pair<StringRef, uint32_t>> SymbolFlags, 299 ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs, 300 const std::map<StringRef, std::vector<WasmComdatEntry>>& Comdats); 301 302 uint32_t getProvisionalValue(const WasmRelocationEntry &RelEntry); 303 void applyRelocations(ArrayRef<WasmRelocationEntry> Relocations, 304 uint64_t ContentsOffset); 305 306 void writeRelocations(ArrayRef<WasmRelocationEntry> Relocations); 307 uint32_t getRelocationIndexValue(const WasmRelocationEntry &RelEntry); 308 uint32_t getFunctionType(const MCSymbolWasm& Symbol); 309 uint32_t registerFunctionType(const MCSymbolWasm& Symbol); 310 }; 311 312 } // end anonymous namespace 313 314 WasmObjectWriter::~WasmObjectWriter() {} 315 316 // Write out a section header and a patchable section size field. 317 void WasmObjectWriter::startSection(SectionBookkeeping &Section, 318 unsigned SectionId, 319 const char *Name) { 320 assert((Name != nullptr) == (SectionId == wasm::WASM_SEC_CUSTOM) && 321 "Only custom sections can have names"); 322 323 DEBUG(dbgs() << "startSection " << SectionId << ": " << Name << "\n"); 324 encodeULEB128(SectionId, getStream()); 325 326 Section.SizeOffset = getStream().tell(); 327 328 // The section size. We don't know the size yet, so reserve enough space 329 // for any 32-bit value; we'll patch it later. 330 encodeULEB128(UINT32_MAX, getStream()); 331 332 // The position where the section starts, for measuring its size. 333 Section.ContentsOffset = getStream().tell(); 334 335 // Custom sections in wasm also have a string identifier. 336 if (SectionId == wasm::WASM_SEC_CUSTOM) { 337 assert(Name); 338 writeString(StringRef(Name)); 339 } 340 } 341 342 // Now that the section is complete and we know how big it is, patch up the 343 // section size field at the start of the section. 344 void WasmObjectWriter::endSection(SectionBookkeeping &Section) { 345 uint64_t Size = getStream().tell() - Section.ContentsOffset; 346 if (uint32_t(Size) != Size) 347 report_fatal_error("section size does not fit in a uint32_t"); 348 349 DEBUG(dbgs() << "endSection size=" << Size << "\n"); 350 351 // Write the final section size to the payload_len field, which follows 352 // the section id byte. 353 uint8_t Buffer[16]; 354 unsigned SizeLen = encodeULEB128(Size, Buffer, 5); 355 assert(SizeLen == 5); 356 getStream().pwrite((char *)Buffer, SizeLen, Section.SizeOffset); 357 } 358 359 // Emit the Wasm header. 360 void WasmObjectWriter::writeHeader(const MCAssembler &Asm) { 361 writeBytes(StringRef(wasm::WasmMagic, sizeof(wasm::WasmMagic))); 362 writeLE32(wasm::WasmVersion); 363 } 364 365 void WasmObjectWriter::executePostLayoutBinding(MCAssembler &Asm, 366 const MCAsmLayout &Layout) { 367 } 368 369 void WasmObjectWriter::recordRelocation(MCAssembler &Asm, 370 const MCAsmLayout &Layout, 371 const MCFragment *Fragment, 372 const MCFixup &Fixup, MCValue Target, 373 uint64_t &FixedValue) { 374 MCAsmBackend &Backend = Asm.getBackend(); 375 bool IsPCRel = Backend.getFixupKindInfo(Fixup.getKind()).Flags & 376 MCFixupKindInfo::FKF_IsPCRel; 377 const auto &FixupSection = cast<MCSectionWasm>(*Fragment->getParent()); 378 uint64_t C = Target.getConstant(); 379 uint64_t FixupOffset = Layout.getFragmentOffset(Fragment) + Fixup.getOffset(); 380 MCContext &Ctx = Asm.getContext(); 381 382 // The .init_array isn't translated as data, so don't do relocations in it. 383 if (FixupSection.getSectionName().startswith(".init_array")) 384 return; 385 386 if (const MCSymbolRefExpr *RefB = Target.getSymB()) { 387 assert(RefB->getKind() == MCSymbolRefExpr::VK_None && 388 "Should not have constructed this"); 389 390 // Let A, B and C being the components of Target and R be the location of 391 // the fixup. If the fixup is not pcrel, we want to compute (A - B + C). 392 // If it is pcrel, we want to compute (A - B + C - R). 393 394 // In general, Wasm has no relocations for -B. It can only represent (A + C) 395 // or (A + C - R). If B = R + K and the relocation is not pcrel, we can 396 // replace B to implement it: (A - R - K + C) 397 if (IsPCRel) { 398 Ctx.reportError( 399 Fixup.getLoc(), 400 "No relocation available to represent this relative expression"); 401 return; 402 } 403 404 const auto &SymB = cast<MCSymbolWasm>(RefB->getSymbol()); 405 406 if (SymB.isUndefined()) { 407 Ctx.reportError(Fixup.getLoc(), 408 Twine("symbol '") + SymB.getName() + 409 "' can not be undefined in a subtraction expression"); 410 return; 411 } 412 413 assert(!SymB.isAbsolute() && "Should have been folded"); 414 const MCSection &SecB = SymB.getSection(); 415 if (&SecB != &FixupSection) { 416 Ctx.reportError(Fixup.getLoc(), 417 "Cannot represent a difference across sections"); 418 return; 419 } 420 421 uint64_t SymBOffset = Layout.getSymbolOffset(SymB); 422 uint64_t K = SymBOffset - FixupOffset; 423 IsPCRel = true; 424 C -= K; 425 } 426 427 // We either rejected the fixup or folded B into C at this point. 428 const MCSymbolRefExpr *RefA = Target.getSymA(); 429 const auto *SymA = RefA ? cast<MCSymbolWasm>(&RefA->getSymbol()) : nullptr; 430 431 if (SymA && SymA->isVariable()) { 432 const MCExpr *Expr = SymA->getVariableValue(); 433 const auto *Inner = cast<MCSymbolRefExpr>(Expr); 434 if (Inner->getKind() == MCSymbolRefExpr::VK_WEAKREF) 435 llvm_unreachable("weakref used in reloc not yet implemented"); 436 } 437 438 // Put any constant offset in an addend. Offsets can be negative, and 439 // LLVM expects wrapping, in contrast to wasm's immediates which can't 440 // be negative and don't wrap. 441 FixedValue = 0; 442 443 if (SymA) 444 SymA->setUsedInReloc(); 445 446 assert(!IsPCRel); 447 assert(SymA); 448 449 unsigned Type = getRelocType(Target, Fixup); 450 451 WasmRelocationEntry Rec(FixupOffset, SymA, C, Type, &FixupSection); 452 DEBUG(dbgs() << "WasmReloc: " << Rec << "\n"); 453 454 if (FixupSection.isWasmData()) 455 DataRelocations.push_back(Rec); 456 else if (FixupSection.getKind().isText()) 457 CodeRelocations.push_back(Rec); 458 else if (!FixupSection.getKind().isMetadata()) 459 // TODO(sbc): Add support for debug sections. 460 llvm_unreachable("unexpected section type"); 461 } 462 463 // Write X as an (unsigned) LEB value at offset Offset in Stream, padded 464 // to allow patching. 465 static void 466 WritePatchableLEB(raw_pwrite_stream &Stream, uint32_t X, uint64_t Offset) { 467 uint8_t Buffer[5]; 468 unsigned SizeLen = encodeULEB128(X, Buffer, 5); 469 assert(SizeLen == 5); 470 Stream.pwrite((char *)Buffer, SizeLen, Offset); 471 } 472 473 // Write X as an signed LEB value at offset Offset in Stream, padded 474 // to allow patching. 475 static void 476 WritePatchableSLEB(raw_pwrite_stream &Stream, int32_t X, uint64_t Offset) { 477 uint8_t Buffer[5]; 478 unsigned SizeLen = encodeSLEB128(X, Buffer, 5); 479 assert(SizeLen == 5); 480 Stream.pwrite((char *)Buffer, SizeLen, Offset); 481 } 482 483 // Write X as a plain integer value at offset Offset in Stream. 484 static void WriteI32(raw_pwrite_stream &Stream, uint32_t X, uint64_t Offset) { 485 uint8_t Buffer[4]; 486 support::endian::write32le(Buffer, X); 487 Stream.pwrite((char *)Buffer, sizeof(Buffer), Offset); 488 } 489 490 static const MCSymbolWasm* ResolveSymbol(const MCSymbolWasm& Symbol) { 491 if (Symbol.isVariable()) { 492 const MCExpr *Expr = Symbol.getVariableValue(); 493 auto *Inner = cast<MCSymbolRefExpr>(Expr); 494 return cast<MCSymbolWasm>(&Inner->getSymbol()); 495 } 496 return &Symbol; 497 } 498 499 // Compute a value to write into the code at the location covered 500 // by RelEntry. This value isn't used by the static linker; it just serves 501 // to make the object format more readable and more likely to be directly 502 // useable. 503 uint32_t 504 WasmObjectWriter::getProvisionalValue(const WasmRelocationEntry &RelEntry) { 505 506 switch (RelEntry.Type) { 507 case wasm::R_WEBASSEMBLY_TABLE_INDEX_SLEB: 508 case wasm::R_WEBASSEMBLY_TABLE_INDEX_I32: 509 // Provitional value is the indirect symbol index 510 if (!IndirectSymbolIndices.count(RelEntry.Symbol)) 511 report_fatal_error("symbol not found in table index space: " + 512 RelEntry.Symbol->getName()); 513 return IndirectSymbolIndices[RelEntry.Symbol]; 514 case wasm::R_WEBASSEMBLY_FUNCTION_INDEX_LEB: 515 case wasm::R_WEBASSEMBLY_TYPE_INDEX_LEB: 516 case wasm::R_WEBASSEMBLY_GLOBAL_INDEX_LEB: 517 // Provitional value is function/type/global index itself 518 return getRelocationIndexValue(RelEntry); 519 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_LEB: 520 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_I32: 521 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_SLEB: { 522 // Provitional value is address of the global 523 const MCSymbolWasm *Sym = ResolveSymbol(*RelEntry.Symbol); 524 // For undefined symbols, use zero 525 if (!Sym->isDefined()) 526 return 0; 527 528 uint32_t GlobalIndex = SymbolIndices[Sym]; 529 const WasmGlobal& Global = Globals[GlobalIndex - NumGlobalImports]; 530 uint64_t Address = Global.InitialValue + RelEntry.Addend; 531 532 // Ignore overflow. LLVM allows address arithmetic to silently wrap. 533 return Address; 534 } 535 default: 536 llvm_unreachable("invalid relocation type"); 537 } 538 } 539 540 static void addData(SmallVectorImpl<char> &DataBytes, 541 MCSectionWasm &DataSection) { 542 DEBUG(errs() << "addData: " << DataSection.getSectionName() << "\n"); 543 544 DataBytes.resize(alignTo(DataBytes.size(), DataSection.getAlignment())); 545 546 size_t LastFragmentSize = 0; 547 for (const MCFragment &Frag : DataSection) { 548 if (Frag.hasInstructions()) 549 report_fatal_error("only data supported in data sections"); 550 551 if (auto *Align = dyn_cast<MCAlignFragment>(&Frag)) { 552 if (Align->getValueSize() != 1) 553 report_fatal_error("only byte values supported for alignment"); 554 // If nops are requested, use zeros, as this is the data section. 555 uint8_t Value = Align->hasEmitNops() ? 0 : Align->getValue(); 556 uint64_t Size = std::min<uint64_t>(alignTo(DataBytes.size(), 557 Align->getAlignment()), 558 DataBytes.size() + 559 Align->getMaxBytesToEmit()); 560 DataBytes.resize(Size, Value); 561 } else if (auto *Fill = dyn_cast<MCFillFragment>(&Frag)) { 562 int64_t Size; 563 if (!Fill->getSize().evaluateAsAbsolute(Size)) 564 llvm_unreachable("The fill should be an assembler constant"); 565 DataBytes.insert(DataBytes.end(), Size, Fill->getValue()); 566 } else { 567 const auto &DataFrag = cast<MCDataFragment>(Frag); 568 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 569 570 DataBytes.insert(DataBytes.end(), Contents.begin(), Contents.end()); 571 LastFragmentSize = Contents.size(); 572 } 573 } 574 575 // Don't allow empty segments, or segments that end with zero-sized 576 // fragment, otherwise the linker cannot map symbols to a unique 577 // data segment. This can be triggered by zero-sized structs 578 // See: test/MC/WebAssembly/bss.ll 579 if (LastFragmentSize == 0) 580 DataBytes.resize(DataBytes.size() + 1); 581 DEBUG(dbgs() << "addData -> " << DataBytes.size() << "\n"); 582 } 583 584 uint32_t 585 WasmObjectWriter::getRelocationIndexValue(const WasmRelocationEntry &RelEntry) { 586 if (RelEntry.Type == wasm::R_WEBASSEMBLY_TYPE_INDEX_LEB) { 587 if (!TypeIndices.count(RelEntry.Symbol)) 588 report_fatal_error("symbol not found in type index space: " + 589 RelEntry.Symbol->getName()); 590 return TypeIndices[RelEntry.Symbol]; 591 } 592 593 if (!SymbolIndices.count(RelEntry.Symbol)) 594 report_fatal_error("symbol not found in function/global index space: " + 595 RelEntry.Symbol->getName()); 596 return SymbolIndices[RelEntry.Symbol]; 597 } 598 599 // Apply the portions of the relocation records that we can handle ourselves 600 // directly. 601 void WasmObjectWriter::applyRelocations( 602 ArrayRef<WasmRelocationEntry> Relocations, uint64_t ContentsOffset) { 603 raw_pwrite_stream &Stream = getStream(); 604 for (const WasmRelocationEntry &RelEntry : Relocations) { 605 uint64_t Offset = ContentsOffset + 606 RelEntry.FixupSection->getSectionOffset() + 607 RelEntry.Offset; 608 609 DEBUG(dbgs() << "applyRelocation: " << RelEntry << "\n"); 610 uint32_t Value = getProvisionalValue(RelEntry); 611 612 switch (RelEntry.Type) { 613 case wasm::R_WEBASSEMBLY_FUNCTION_INDEX_LEB: 614 case wasm::R_WEBASSEMBLY_TYPE_INDEX_LEB: 615 case wasm::R_WEBASSEMBLY_GLOBAL_INDEX_LEB: 616 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_LEB: 617 WritePatchableLEB(Stream, Value, Offset); 618 break; 619 case wasm::R_WEBASSEMBLY_TABLE_INDEX_I32: 620 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_I32: 621 WriteI32(Stream, Value, Offset); 622 break; 623 case wasm::R_WEBASSEMBLY_TABLE_INDEX_SLEB: 624 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_SLEB: 625 WritePatchableSLEB(Stream, Value, Offset); 626 break; 627 default: 628 llvm_unreachable("invalid relocation type"); 629 } 630 } 631 } 632 633 // Write out the portions of the relocation records that the linker will 634 // need to handle. 635 void WasmObjectWriter::writeRelocations( 636 ArrayRef<WasmRelocationEntry> Relocations) { 637 raw_pwrite_stream &Stream = getStream(); 638 for (const WasmRelocationEntry& RelEntry : Relocations) { 639 640 uint64_t Offset = RelEntry.Offset + 641 RelEntry.FixupSection->getSectionOffset(); 642 uint32_t Index = getRelocationIndexValue(RelEntry); 643 644 encodeULEB128(RelEntry.Type, Stream); 645 encodeULEB128(Offset, Stream); 646 encodeULEB128(Index, Stream); 647 if (RelEntry.hasAddend()) 648 encodeSLEB128(RelEntry.Addend, Stream); 649 } 650 } 651 652 void WasmObjectWriter::writeTypeSection( 653 ArrayRef<WasmFunctionType> FunctionTypes) { 654 if (FunctionTypes.empty()) 655 return; 656 657 SectionBookkeeping Section; 658 startSection(Section, wasm::WASM_SEC_TYPE); 659 660 encodeULEB128(FunctionTypes.size(), getStream()); 661 662 for (const WasmFunctionType &FuncTy : FunctionTypes) { 663 encodeSLEB128(wasm::WASM_TYPE_FUNC, getStream()); 664 encodeULEB128(FuncTy.Params.size(), getStream()); 665 for (wasm::ValType Ty : FuncTy.Params) 666 writeValueType(Ty); 667 encodeULEB128(FuncTy.Returns.size(), getStream()); 668 for (wasm::ValType Ty : FuncTy.Returns) 669 writeValueType(Ty); 670 } 671 672 endSection(Section); 673 } 674 675 void WasmObjectWriter::writeImportSection(ArrayRef<WasmImport> Imports, 676 uint32_t DataSize, 677 uint32_t NumElements) { 678 if (Imports.empty()) 679 return; 680 681 uint32_t NumPages = (DataSize + wasm::WasmPageSize - 1) / wasm::WasmPageSize; 682 683 SectionBookkeeping Section; 684 startSection(Section, wasm::WASM_SEC_IMPORT); 685 686 encodeULEB128(Imports.size(), getStream()); 687 for (const WasmImport &Import : Imports) { 688 writeString(Import.ModuleName); 689 writeString(Import.FieldName); 690 691 encodeULEB128(Import.Kind, getStream()); 692 693 switch (Import.Kind) { 694 case wasm::WASM_EXTERNAL_FUNCTION: 695 encodeULEB128(Import.Type, getStream()); 696 break; 697 case wasm::WASM_EXTERNAL_GLOBAL: 698 encodeSLEB128(int32_t(Import.Type), getStream()); 699 encodeULEB128(int32_t(Import.IsMutable), getStream()); 700 break; 701 case wasm::WASM_EXTERNAL_MEMORY: 702 encodeULEB128(0, getStream()); // flags 703 encodeULEB128(NumPages, getStream()); // initial 704 break; 705 case wasm::WASM_EXTERNAL_TABLE: 706 encodeSLEB128(int32_t(Import.Type), getStream()); 707 encodeULEB128(0, getStream()); // flags 708 encodeULEB128(NumElements, getStream()); // initial 709 break; 710 default: 711 llvm_unreachable("unsupported import kind"); 712 } 713 } 714 715 endSection(Section); 716 } 717 718 void WasmObjectWriter::writeFunctionSection(ArrayRef<WasmFunction> Functions) { 719 if (Functions.empty()) 720 return; 721 722 SectionBookkeeping Section; 723 startSection(Section, wasm::WASM_SEC_FUNCTION); 724 725 encodeULEB128(Functions.size(), getStream()); 726 for (const WasmFunction &Func : Functions) 727 encodeULEB128(Func.Type, getStream()); 728 729 endSection(Section); 730 } 731 732 void WasmObjectWriter::writeGlobalSection() { 733 if (Globals.empty()) 734 return; 735 736 SectionBookkeeping Section; 737 startSection(Section, wasm::WASM_SEC_GLOBAL); 738 739 encodeULEB128(Globals.size(), getStream()); 740 for (const WasmGlobal &Global : Globals) { 741 writeValueType(Global.Type); 742 write8(Global.IsMutable); 743 744 if (Global.HasImport) { 745 assert(Global.InitialValue == 0); 746 write8(wasm::WASM_OPCODE_GET_GLOBAL); 747 encodeULEB128(Global.ImportIndex, getStream()); 748 } else { 749 assert(Global.ImportIndex == 0); 750 write8(wasm::WASM_OPCODE_I32_CONST); 751 encodeSLEB128(Global.InitialValue, getStream()); // offset 752 } 753 write8(wasm::WASM_OPCODE_END); 754 } 755 756 endSection(Section); 757 } 758 759 void WasmObjectWriter::writeExportSection(ArrayRef<WasmExport> Exports) { 760 if (Exports.empty()) 761 return; 762 763 SectionBookkeeping Section; 764 startSection(Section, wasm::WASM_SEC_EXPORT); 765 766 encodeULEB128(Exports.size(), getStream()); 767 for (const WasmExport &Export : Exports) { 768 writeString(Export.FieldName); 769 encodeSLEB128(Export.Kind, getStream()); 770 encodeULEB128(Export.Index, getStream()); 771 } 772 773 endSection(Section); 774 } 775 776 void WasmObjectWriter::writeElemSection(ArrayRef<uint32_t> TableElems) { 777 if (TableElems.empty()) 778 return; 779 780 SectionBookkeeping Section; 781 startSection(Section, wasm::WASM_SEC_ELEM); 782 783 encodeULEB128(1, getStream()); // number of "segments" 784 encodeULEB128(0, getStream()); // the table index 785 786 // init expr for starting offset 787 write8(wasm::WASM_OPCODE_I32_CONST); 788 encodeSLEB128(kInitialTableOffset, getStream()); 789 write8(wasm::WASM_OPCODE_END); 790 791 encodeULEB128(TableElems.size(), getStream()); 792 for (uint32_t Elem : TableElems) 793 encodeULEB128(Elem, getStream()); 794 795 endSection(Section); 796 } 797 798 void WasmObjectWriter::writeCodeSection(const MCAssembler &Asm, 799 const MCAsmLayout &Layout, 800 ArrayRef<WasmFunction> Functions) { 801 if (Functions.empty()) 802 return; 803 804 SectionBookkeeping Section; 805 startSection(Section, wasm::WASM_SEC_CODE); 806 807 encodeULEB128(Functions.size(), getStream()); 808 809 for (const WasmFunction &Func : Functions) { 810 auto &FuncSection = static_cast<MCSectionWasm &>(Func.Sym->getSection()); 811 812 int64_t Size = 0; 813 if (!Func.Sym->getSize()->evaluateAsAbsolute(Size, Layout)) 814 report_fatal_error(".size expression must be evaluatable"); 815 816 encodeULEB128(Size, getStream()); 817 FuncSection.setSectionOffset(getStream().tell() - Section.ContentsOffset); 818 Asm.writeSectionData(&FuncSection, Layout); 819 } 820 821 // Apply fixups. 822 applyRelocations(CodeRelocations, Section.ContentsOffset); 823 824 endSection(Section); 825 } 826 827 void WasmObjectWriter::writeDataSection(ArrayRef<WasmDataSegment> Segments) { 828 if (Segments.empty()) 829 return; 830 831 SectionBookkeeping Section; 832 startSection(Section, wasm::WASM_SEC_DATA); 833 834 encodeULEB128(Segments.size(), getStream()); // count 835 836 for (const WasmDataSegment & Segment : Segments) { 837 encodeULEB128(0, getStream()); // memory index 838 write8(wasm::WASM_OPCODE_I32_CONST); 839 encodeSLEB128(Segment.Offset, getStream()); // offset 840 write8(wasm::WASM_OPCODE_END); 841 encodeULEB128(Segment.Data.size(), getStream()); // size 842 Segment.Section->setSectionOffset(getStream().tell() - Section.ContentsOffset); 843 writeBytes(Segment.Data); // data 844 } 845 846 // Apply fixups. 847 applyRelocations(DataRelocations, Section.ContentsOffset); 848 849 endSection(Section); 850 } 851 852 void WasmObjectWriter::writeCodeRelocSection() { 853 // See: https://github.com/WebAssembly/tool-conventions/blob/master/Linking.md 854 // for descriptions of the reloc sections. 855 856 if (CodeRelocations.empty()) 857 return; 858 859 SectionBookkeeping Section; 860 startSection(Section, wasm::WASM_SEC_CUSTOM, "reloc.CODE"); 861 862 encodeULEB128(wasm::WASM_SEC_CODE, getStream()); 863 encodeULEB128(CodeRelocations.size(), getStream()); 864 865 writeRelocations(CodeRelocations); 866 867 endSection(Section); 868 } 869 870 void WasmObjectWriter::writeDataRelocSection() { 871 // See: https://github.com/WebAssembly/tool-conventions/blob/master/Linking.md 872 // for descriptions of the reloc sections. 873 874 if (DataRelocations.empty()) 875 return; 876 877 SectionBookkeeping Section; 878 startSection(Section, wasm::WASM_SEC_CUSTOM, "reloc.DATA"); 879 880 encodeULEB128(wasm::WASM_SEC_DATA, getStream()); 881 encodeULEB128(DataRelocations.size(), getStream()); 882 883 writeRelocations(DataRelocations); 884 885 endSection(Section); 886 } 887 888 void WasmObjectWriter::writeLinkingMetaDataSection( 889 ArrayRef<WasmDataSegment> Segments, uint32_t DataSize, 890 ArrayRef<std::pair<StringRef, uint32_t>> SymbolFlags, 891 ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs, 892 const std::map<StringRef, std::vector<WasmComdatEntry>>& Comdats) { 893 SectionBookkeeping Section; 894 startSection(Section, wasm::WASM_SEC_CUSTOM, "linking"); 895 SectionBookkeeping SubSection; 896 897 if (SymbolFlags.size() != 0) { 898 startSection(SubSection, wasm::WASM_SYMBOL_INFO); 899 encodeULEB128(SymbolFlags.size(), getStream()); 900 for (auto Pair: SymbolFlags) { 901 writeString(Pair.first); 902 encodeULEB128(Pair.second, getStream()); 903 } 904 endSection(SubSection); 905 } 906 907 if (DataSize > 0) { 908 startSection(SubSection, wasm::WASM_DATA_SIZE); 909 encodeULEB128(DataSize, getStream()); 910 endSection(SubSection); 911 } 912 913 if (Segments.size()) { 914 startSection(SubSection, wasm::WASM_SEGMENT_INFO); 915 encodeULEB128(Segments.size(), getStream()); 916 for (const WasmDataSegment &Segment : Segments) { 917 writeString(Segment.Name); 918 encodeULEB128(Segment.Alignment, getStream()); 919 encodeULEB128(Segment.Flags, getStream()); 920 } 921 endSection(SubSection); 922 } 923 924 if (!InitFuncs.empty()) { 925 startSection(SubSection, wasm::WASM_INIT_FUNCS); 926 encodeULEB128(InitFuncs.size(), getStream()); 927 for (auto &StartFunc : InitFuncs) { 928 encodeULEB128(StartFunc.first, getStream()); // priority 929 encodeULEB128(StartFunc.second, getStream()); // function index 930 } 931 endSection(SubSection); 932 } 933 934 if (Comdats.size()) { 935 startSection(SubSection, wasm::WASM_COMDAT_INFO); 936 encodeULEB128(Comdats.size(), getStream()); 937 for (const auto &C : Comdats) { 938 writeString(C.first); 939 encodeULEB128(0, getStream()); // flags for future use 940 encodeULEB128(C.second.size(), getStream()); 941 for (const WasmComdatEntry &Entry : C.second) { 942 encodeULEB128(Entry.Kind, getStream()); 943 encodeULEB128(Entry.Index, getStream()); 944 } 945 } 946 endSection(SubSection); 947 } 948 949 endSection(Section); 950 } 951 952 uint32_t WasmObjectWriter::getFunctionType(const MCSymbolWasm& Symbol) { 953 assert(Symbol.isFunction()); 954 assert(TypeIndices.count(&Symbol)); 955 return TypeIndices[&Symbol]; 956 } 957 958 uint32_t WasmObjectWriter::registerFunctionType(const MCSymbolWasm& Symbol) { 959 assert(Symbol.isFunction()); 960 961 WasmFunctionType F; 962 const MCSymbolWasm* ResolvedSym = ResolveSymbol(Symbol); 963 F.Returns = ResolvedSym->getReturns(); 964 F.Params = ResolvedSym->getParams(); 965 966 auto Pair = 967 FunctionTypeIndices.insert(std::make_pair(F, FunctionTypes.size())); 968 if (Pair.second) 969 FunctionTypes.push_back(F); 970 TypeIndices[&Symbol] = Pair.first->second; 971 972 DEBUG(dbgs() << "registerFunctionType: " << Symbol << " new:" << Pair.second << "\n"); 973 DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n"); 974 return Pair.first->second; 975 } 976 977 void WasmObjectWriter::writeObject(MCAssembler &Asm, 978 const MCAsmLayout &Layout) { 979 DEBUG(dbgs() << "WasmObjectWriter::writeObject\n"); 980 MCContext &Ctx = Asm.getContext(); 981 wasm::ValType PtrType = is64Bit() ? wasm::ValType::I64 : wasm::ValType::I32; 982 983 // Collect information from the available symbols. 984 SmallVector<WasmFunction, 4> Functions; 985 SmallVector<uint32_t, 4> TableElems; 986 SmallVector<WasmImport, 4> Imports; 987 SmallVector<WasmExport, 4> Exports; 988 SmallVector<std::pair<StringRef, uint32_t>, 4> SymbolFlags; 989 SmallVector<std::pair<uint16_t, uint32_t>, 2> InitFuncs; 990 std::map<StringRef, std::vector<WasmComdatEntry>> Comdats; 991 SmallVector<WasmDataSegment, 4> DataSegments; 992 uint32_t DataSize = 0; 993 994 // In the special .global_variables section, we've encoded global 995 // variables used by the function. Translate them into the Globals 996 // list. 997 MCSectionWasm *GlobalVars = 998 Ctx.getWasmSection(".global_variables", SectionKind::getMetadata()); 999 if (!GlobalVars->getFragmentList().empty()) { 1000 if (GlobalVars->getFragmentList().size() != 1) 1001 report_fatal_error("only one .global_variables fragment supported"); 1002 const MCFragment &Frag = *GlobalVars->begin(); 1003 if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data) 1004 report_fatal_error("only data supported in .global_variables"); 1005 const auto &DataFrag = cast<MCDataFragment>(Frag); 1006 if (!DataFrag.getFixups().empty()) 1007 report_fatal_error("fixups not supported in .global_variables"); 1008 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 1009 for (const uint8_t *p = (const uint8_t *)Contents.data(), 1010 *end = (const uint8_t *)Contents.data() + Contents.size(); 1011 p != end; ) { 1012 WasmGlobal G; 1013 if (end - p < 3) 1014 report_fatal_error("truncated global variable encoding"); 1015 G.Type = wasm::ValType(int8_t(*p++)); 1016 G.IsMutable = bool(*p++); 1017 G.HasImport = bool(*p++); 1018 if (G.HasImport) { 1019 G.InitialValue = 0; 1020 1021 WasmImport Import; 1022 Import.ModuleName = (const char *)p; 1023 const uint8_t *nul = (const uint8_t *)memchr(p, '\0', end - p); 1024 if (!nul) 1025 report_fatal_error("global module name must be nul-terminated"); 1026 p = nul + 1; 1027 nul = (const uint8_t *)memchr(p, '\0', end - p); 1028 if (!nul) 1029 report_fatal_error("global base name must be nul-terminated"); 1030 Import.FieldName = (const char *)p; 1031 p = nul + 1; 1032 1033 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1034 Import.Type = int32_t(G.Type); 1035 1036 G.ImportIndex = NumGlobalImports; 1037 ++NumGlobalImports; 1038 1039 Imports.push_back(Import); 1040 } else { 1041 unsigned n; 1042 G.InitialValue = decodeSLEB128(p, &n); 1043 G.ImportIndex = 0; 1044 if ((ptrdiff_t)n > end - p) 1045 report_fatal_error("global initial value must be valid SLEB128"); 1046 p += n; 1047 } 1048 Globals.push_back(G); 1049 } 1050 } 1051 1052 // For now, always emit the memory import, since loads and stores are not 1053 // valid without it. In the future, we could perhaps be more clever and omit 1054 // it if there are no loads or stores. 1055 MCSymbolWasm *MemorySym = 1056 cast<MCSymbolWasm>(Ctx.getOrCreateSymbol("__linear_memory")); 1057 WasmImport MemImport; 1058 MemImport.ModuleName = MemorySym->getModuleName(); 1059 MemImport.FieldName = MemorySym->getName(); 1060 MemImport.Kind = wasm::WASM_EXTERNAL_MEMORY; 1061 Imports.push_back(MemImport); 1062 1063 // For now, always emit the table section, since indirect calls are not 1064 // valid without it. In the future, we could perhaps be more clever and omit 1065 // it if there are no indirect calls. 1066 MCSymbolWasm *TableSym = 1067 cast<MCSymbolWasm>(Ctx.getOrCreateSymbol("__indirect_function_table")); 1068 WasmImport TableImport; 1069 TableImport.ModuleName = TableSym->getModuleName(); 1070 TableImport.FieldName = TableSym->getName(); 1071 TableImport.Kind = wasm::WASM_EXTERNAL_TABLE; 1072 TableImport.Type = wasm::WASM_TYPE_ANYFUNC; 1073 Imports.push_back(TableImport); 1074 1075 // Populate FunctionTypeIndices and Imports. 1076 for (const MCSymbol &S : Asm.symbols()) { 1077 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1078 1079 // Register types for all functions, including those with private linkage 1080 // (because wasm always needs a type signature). 1081 if (WS.isFunction()) 1082 registerFunctionType(WS); 1083 1084 if (WS.isTemporary()) 1085 continue; 1086 1087 // If the symbol is not defined in this translation unit, import it. 1088 if ((!WS.isDefined() && !WS.isComdat()) || 1089 WS.isVariable()) { 1090 WasmImport Import; 1091 Import.ModuleName = WS.getModuleName(); 1092 Import.FieldName = WS.getName(); 1093 1094 if (WS.isFunction()) { 1095 Import.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1096 Import.Type = getFunctionType(WS); 1097 SymbolIndices[&WS] = NumFunctionImports; 1098 ++NumFunctionImports; 1099 } else { 1100 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1101 Import.Type = int32_t(PtrType); 1102 Import.IsMutable = false; 1103 SymbolIndices[&WS] = NumGlobalImports; 1104 1105 // If this global is the stack pointer, make it mutable. 1106 if (WS.getName() == "__stack_pointer") 1107 Import.IsMutable = true; 1108 1109 ++NumGlobalImports; 1110 } 1111 1112 Imports.push_back(Import); 1113 } 1114 } 1115 1116 for (MCSection &Sec : Asm) { 1117 auto &Section = static_cast<MCSectionWasm &>(Sec); 1118 if (!Section.isWasmData()) 1119 continue; 1120 1121 // .init_array sections are handled specially elsewhere. 1122 if (cast<MCSectionWasm>(Sec).getSectionName().startswith(".init_array")) 1123 continue; 1124 1125 DataSize = alignTo(DataSize, Section.getAlignment()); 1126 DataSegments.emplace_back(); 1127 WasmDataSegment &Segment = DataSegments.back(); 1128 Segment.Name = Section.getSectionName(); 1129 Segment.Offset = DataSize; 1130 Segment.Section = &Section; 1131 addData(Segment.Data, Section); 1132 Segment.Alignment = Section.getAlignment(); 1133 Segment.Flags = 0; 1134 DataSize += Segment.Data.size(); 1135 Section.setMemoryOffset(Segment.Offset); 1136 1137 if (const MCSymbolWasm *C = Section.getGroup()) { 1138 Comdats[C->getName()].emplace_back( 1139 WasmComdatEntry{wasm::WASM_COMDAT_DATA, 1140 static_cast<uint32_t>(DataSegments.size()) - 1}); 1141 } 1142 } 1143 1144 // Handle regular defined and undefined symbols. 1145 for (const MCSymbol &S : Asm.symbols()) { 1146 // Ignore unnamed temporary symbols, which aren't ever exported, imported, 1147 // or used in relocations. 1148 if (S.isTemporary() && S.getName().empty()) 1149 continue; 1150 1151 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1152 DEBUG(dbgs() << "MCSymbol: '" << S << "'" 1153 << " isDefined=" << S.isDefined() << " isExternal=" 1154 << S.isExternal() << " isTemporary=" << S.isTemporary() 1155 << " isFunction=" << WS.isFunction() 1156 << " isWeak=" << WS.isWeak() 1157 << " isHidden=" << WS.isHidden() 1158 << " isVariable=" << WS.isVariable() << "\n"); 1159 1160 if (WS.isWeak() || WS.isHidden()) { 1161 uint32_t Flags = (WS.isWeak() ? wasm::WASM_SYMBOL_BINDING_WEAK : 0) | 1162 (WS.isHidden() ? wasm::WASM_SYMBOL_VISIBILITY_HIDDEN : 0); 1163 SymbolFlags.emplace_back(WS.getName(), Flags); 1164 } 1165 1166 if (WS.isVariable()) 1167 continue; 1168 1169 unsigned Index; 1170 1171 if (WS.isFunction()) { 1172 if (WS.isDefined()) { 1173 if (WS.getOffset() != 0) 1174 report_fatal_error( 1175 "function sections must contain one function each"); 1176 1177 if (WS.getSize() == 0) 1178 report_fatal_error( 1179 "function symbols must have a size set with .size"); 1180 1181 // A definition. Take the next available index. 1182 Index = NumFunctionImports + Functions.size(); 1183 1184 // Prepare the function. 1185 WasmFunction Func; 1186 Func.Type = getFunctionType(WS); 1187 Func.Sym = &WS; 1188 SymbolIndices[&WS] = Index; 1189 Functions.push_back(Func); 1190 } else { 1191 // An import; the index was assigned above. 1192 Index = SymbolIndices.find(&WS)->second; 1193 } 1194 1195 DEBUG(dbgs() << " -> function index: " << Index << "\n"); 1196 } else { 1197 if (WS.isTemporary() && !WS.getSize()) 1198 continue; 1199 1200 if (!WS.isDefined()) 1201 continue; 1202 1203 if (!WS.getSize()) 1204 report_fatal_error("data symbols must have a size set with .size: " + 1205 WS.getName()); 1206 1207 int64_t Size = 0; 1208 if (!WS.getSize()->evaluateAsAbsolute(Size, Layout)) 1209 report_fatal_error(".size expression must be evaluatable"); 1210 1211 // For each global, prepare a corresponding wasm global holding its 1212 // address. For externals these will also be named exports. 1213 Index = NumGlobalImports + Globals.size(); 1214 auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection()); 1215 assert(DataSection.isWasmData()); 1216 1217 WasmGlobal Global; 1218 Global.Type = PtrType; 1219 Global.IsMutable = false; 1220 Global.HasImport = false; 1221 Global.InitialValue = DataSection.getMemoryOffset() + Layout.getSymbolOffset(WS); 1222 Global.ImportIndex = 0; 1223 SymbolIndices[&WS] = Index; 1224 DEBUG(dbgs() << " -> global index: " << Index << "\n"); 1225 Globals.push_back(Global); 1226 } 1227 1228 // If the symbol is visible outside this translation unit, export it. 1229 if (WS.isDefined()) { 1230 WasmExport Export; 1231 Export.FieldName = WS.getName(); 1232 Export.Index = Index; 1233 if (WS.isFunction()) 1234 Export.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1235 else 1236 Export.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1237 DEBUG(dbgs() << " -> export " << Exports.size() << "\n"); 1238 Exports.push_back(Export); 1239 1240 if (!WS.isExternal()) 1241 SymbolFlags.emplace_back(WS.getName(), wasm::WASM_SYMBOL_BINDING_LOCAL); 1242 1243 if (WS.isFunction()) { 1244 auto &Section = static_cast<MCSectionWasm &>(WS.getSection()); 1245 if (const MCSymbolWasm *C = Section.getGroup()) 1246 Comdats[C->getName()].emplace_back( 1247 WasmComdatEntry{wasm::WASM_COMDAT_FUNCTION, Index}); 1248 } 1249 } 1250 } 1251 1252 // Handle weak aliases. We need to process these in a separate pass because 1253 // we need to have processed the target of the alias before the alias itself 1254 // and the symbols are not necessarily ordered in this way. 1255 for (const MCSymbol &S : Asm.symbols()) { 1256 if (!S.isVariable()) 1257 continue; 1258 1259 assert(S.isDefined()); 1260 1261 // Find the target symbol of this weak alias and export that index 1262 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1263 const MCSymbolWasm *ResolvedSym = ResolveSymbol(WS); 1264 DEBUG(dbgs() << WS.getName() << ": weak alias of '" << *ResolvedSym << "'\n"); 1265 assert(SymbolIndices.count(ResolvedSym) > 0); 1266 uint32_t Index = SymbolIndices.find(ResolvedSym)->second; 1267 DEBUG(dbgs() << " -> index:" << Index << "\n"); 1268 1269 WasmExport Export; 1270 Export.FieldName = WS.getName(); 1271 Export.Index = Index; 1272 if (WS.isFunction()) 1273 Export.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1274 else 1275 Export.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1276 DEBUG(dbgs() << " -> export " << Exports.size() << "\n"); 1277 Exports.push_back(Export); 1278 1279 if (!WS.isExternal()) 1280 SymbolFlags.emplace_back(WS.getName(), wasm::WASM_SYMBOL_BINDING_LOCAL); 1281 } 1282 1283 { 1284 auto HandleReloc = [&](const WasmRelocationEntry &Rel) { 1285 // Functions referenced by a relocation need to prepared to be called 1286 // indirectly. 1287 const MCSymbolWasm& WS = *Rel.Symbol; 1288 if (WS.isFunction() && IndirectSymbolIndices.count(&WS) == 0) { 1289 switch (Rel.Type) { 1290 case wasm::R_WEBASSEMBLY_TABLE_INDEX_I32: 1291 case wasm::R_WEBASSEMBLY_TABLE_INDEX_SLEB: 1292 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_I32: 1293 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_SLEB: { 1294 uint32_t Index = SymbolIndices.find(&WS)->second; 1295 IndirectSymbolIndices[&WS] = TableElems.size() + kInitialTableOffset; 1296 DEBUG(dbgs() << " -> adding to table: " << TableElems.size() << "\n"); 1297 TableElems.push_back(Index); 1298 registerFunctionType(WS); 1299 break; 1300 } 1301 default: 1302 break; 1303 } 1304 } 1305 }; 1306 1307 for (const WasmRelocationEntry &RelEntry : CodeRelocations) 1308 HandleReloc(RelEntry); 1309 for (const WasmRelocationEntry &RelEntry : DataRelocations) 1310 HandleReloc(RelEntry); 1311 } 1312 1313 // Translate .init_array section contents into start functions. 1314 for (const MCSection &S : Asm) { 1315 const auto &WS = static_cast<const MCSectionWasm &>(S); 1316 if (WS.getSectionName().startswith(".fini_array")) 1317 report_fatal_error(".fini_array sections are unsupported"); 1318 if (!WS.getSectionName().startswith(".init_array")) 1319 continue; 1320 if (WS.getFragmentList().empty()) 1321 continue; 1322 if (WS.getFragmentList().size() != 2) 1323 report_fatal_error("only one .init_array section fragment supported"); 1324 const MCFragment &AlignFrag = *WS.begin(); 1325 if (AlignFrag.getKind() != MCFragment::FT_Align) 1326 report_fatal_error(".init_array section should be aligned"); 1327 if (cast<MCAlignFragment>(AlignFrag).getAlignment() != (is64Bit() ? 8 : 4)) 1328 report_fatal_error(".init_array section should be aligned for pointers"); 1329 const MCFragment &Frag = *std::next(WS.begin()); 1330 if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data) 1331 report_fatal_error("only data supported in .init_array section"); 1332 uint16_t Priority = UINT16_MAX; 1333 if (WS.getSectionName().size() != 11) { 1334 if (WS.getSectionName()[11] != '.') 1335 report_fatal_error(".init_array section priority should start with '.'"); 1336 if (WS.getSectionName().substr(12).getAsInteger(10, Priority)) 1337 report_fatal_error("invalid .init_array section priority"); 1338 } 1339 const auto &DataFrag = cast<MCDataFragment>(Frag); 1340 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 1341 for (const uint8_t *p = (const uint8_t *)Contents.data(), 1342 *end = (const uint8_t *)Contents.data() + Contents.size(); 1343 p != end; ++p) { 1344 if (*p != 0) 1345 report_fatal_error("non-symbolic data in .init_array section"); 1346 } 1347 for (const MCFixup &Fixup : DataFrag.getFixups()) { 1348 assert(Fixup.getKind() == MCFixup::getKindForSize(is64Bit() ? 8 : 4, false)); 1349 const MCExpr *Expr = Fixup.getValue(); 1350 auto *Sym = dyn_cast<MCSymbolRefExpr>(Expr); 1351 if (!Sym) 1352 report_fatal_error("fixups in .init_array should be symbol references"); 1353 if (Sym->getKind() != MCSymbolRefExpr::VK_WebAssembly_FUNCTION) 1354 report_fatal_error("symbols in .init_array should be for functions"); 1355 auto I = SymbolIndices.find(cast<MCSymbolWasm>(&Sym->getSymbol())); 1356 if (I == SymbolIndices.end()) 1357 report_fatal_error("symbols in .init_array should be defined"); 1358 uint32_t Index = I->second; 1359 InitFuncs.push_back(std::make_pair(Priority, Index)); 1360 } 1361 } 1362 1363 // Write out the Wasm header. 1364 writeHeader(Asm); 1365 1366 writeTypeSection(FunctionTypes); 1367 writeImportSection(Imports, DataSize, TableElems.size()); 1368 writeFunctionSection(Functions); 1369 // Skip the "table" section; we import the table instead. 1370 // Skip the "memory" section; we import the memory instead. 1371 writeGlobalSection(); 1372 writeExportSection(Exports); 1373 writeElemSection(TableElems); 1374 writeCodeSection(Asm, Layout, Functions); 1375 writeDataSection(DataSegments); 1376 writeCodeRelocSection(); 1377 writeDataRelocSection(); 1378 writeLinkingMetaDataSection(DataSegments, DataSize, SymbolFlags, 1379 InitFuncs, Comdats); 1380 1381 // TODO: Translate the .comment section to the output. 1382 // TODO: Translate debug sections to the output. 1383 } 1384 1385 std::unique_ptr<MCObjectWriter> 1386 llvm::createWasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 1387 raw_pwrite_stream &OS) { 1388 return llvm::make_unique<WasmObjectWriter>(std::move(MOTW), OS); 1389 } 1390