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