1 //===- lib/MC/WasmObjectWriter.cpp - Wasm File Writer ---------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements Wasm object file writer information. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/ADT/STLExtras.h" 15 #include "llvm/ADT/SmallPtrSet.h" 16 #include "llvm/BinaryFormat/Wasm.h" 17 #include "llvm/Config/llvm-config.h" 18 #include "llvm/MC/MCAsmBackend.h" 19 #include "llvm/MC/MCAsmLayout.h" 20 #include "llvm/MC/MCAssembler.h" 21 #include "llvm/MC/MCContext.h" 22 #include "llvm/MC/MCExpr.h" 23 #include "llvm/MC/MCFixupKindInfo.h" 24 #include "llvm/MC/MCObjectWriter.h" 25 #include "llvm/MC/MCSectionWasm.h" 26 #include "llvm/MC/MCSymbolWasm.h" 27 #include "llvm/MC/MCValue.h" 28 #include "llvm/MC/MCWasmObjectWriter.h" 29 #include "llvm/Support/Casting.h" 30 #include "llvm/Support/Debug.h" 31 #include "llvm/Support/ErrorHandling.h" 32 #include "llvm/Support/LEB128.h" 33 #include "llvm/Support/StringSaver.h" 34 #include <vector> 35 36 using namespace llvm; 37 38 #define DEBUG_TYPE "mc" 39 40 namespace { 41 42 // Went we ceate the indirect function table we start at 1, so that there is 43 // and emtpy slot at 0 and therefore calling a null function pointer will trap. 44 static const uint32_t kInitialTableOffset = 1; 45 46 // For patching purposes, we need to remember where each section starts, both 47 // for patching up the section size field, and for patching up references to 48 // locations within the section. 49 struct SectionBookkeeping { 50 // Where the size of the section is written. 51 uint64_t SizeOffset; 52 // Where the section header ends (without custom section name). 53 uint64_t PayloadOffset; 54 // Where the contents of the section starts. 55 uint64_t ContentsOffset; 56 uint32_t Index; 57 }; 58 59 // The signature of a wasm function, in a struct capable of being used as a 60 // DenseMap key. 61 // TODO: Consider using WasmSignature directly instead. 62 struct WasmFunctionType { 63 // Support empty and tombstone instances, needed by DenseMap. 64 enum { Plain, Empty, Tombstone } State; 65 66 // The return types of the function. 67 SmallVector<wasm::ValType, 1> Returns; 68 69 // The parameter types of the function. 70 SmallVector<wasm::ValType, 4> Params; 71 72 WasmFunctionType() : State(Plain) {} 73 74 bool operator==(const WasmFunctionType &Other) const { 75 return State == Other.State && Returns == Other.Returns && 76 Params == Other.Params; 77 } 78 }; 79 80 // Traits for using WasmFunctionType in a DenseMap. 81 struct WasmFunctionTypeDenseMapInfo { 82 static WasmFunctionType getEmptyKey() { 83 WasmFunctionType FuncTy; 84 FuncTy.State = WasmFunctionType::Empty; 85 return FuncTy; 86 } 87 static WasmFunctionType getTombstoneKey() { 88 WasmFunctionType FuncTy; 89 FuncTy.State = WasmFunctionType::Tombstone; 90 return FuncTy; 91 } 92 static unsigned getHashValue(const WasmFunctionType &FuncTy) { 93 uintptr_t Value = FuncTy.State; 94 for (wasm::ValType Ret : FuncTy.Returns) 95 Value += DenseMapInfo<uint32_t>::getHashValue(uint32_t(Ret)); 96 for (wasm::ValType Param : FuncTy.Params) 97 Value += DenseMapInfo<uint32_t>::getHashValue(uint32_t(Param)); 98 return Value; 99 } 100 static bool isEqual(const WasmFunctionType &LHS, 101 const WasmFunctionType &RHS) { 102 return LHS == RHS; 103 } 104 }; 105 106 // A wasm data segment. A wasm binary contains only a single data section 107 // but that can contain many segments, each with their own virtual location 108 // in memory. Each MCSection data created by llvm is modeled as its own 109 // wasm data segment. 110 struct WasmDataSegment { 111 MCSectionWasm *Section; 112 StringRef Name; 113 uint32_t Offset; 114 uint32_t Alignment; 115 uint32_t Flags; 116 SmallVector<char, 4> Data; 117 }; 118 119 // A wasm function to be written into the function section. 120 struct WasmFunction { 121 uint32_t Type; 122 const MCSymbolWasm *Sym; 123 }; 124 125 // A wasm global to be written into the global section. 126 struct WasmGlobal { 127 wasm::WasmGlobalType Type; 128 uint64_t InitialValue; 129 }; 130 131 // Information about a single item which is part of a COMDAT. For each data 132 // segment or function which is in the COMDAT, there is a corresponding 133 // WasmComdatEntry. 134 struct WasmComdatEntry { 135 unsigned Kind; 136 uint32_t Index; 137 }; 138 139 // Information about a single relocation. 140 struct WasmRelocationEntry { 141 uint64_t Offset; // Where is the relocation. 142 const MCSymbolWasm *Symbol; // The symbol to relocate with. 143 int64_t Addend; // A value to add to the symbol. 144 unsigned Type; // The type of the relocation. 145 const MCSectionWasm *FixupSection; // The section the relocation is targeting. 146 147 WasmRelocationEntry(uint64_t Offset, const MCSymbolWasm *Symbol, 148 int64_t Addend, unsigned Type, 149 const MCSectionWasm *FixupSection) 150 : Offset(Offset), Symbol(Symbol), Addend(Addend), Type(Type), 151 FixupSection(FixupSection) {} 152 153 bool hasAddend() const { 154 switch (Type) { 155 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_LEB: 156 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_SLEB: 157 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_I32: 158 case wasm::R_WEBASSEMBLY_FUNCTION_OFFSET_I32: 159 case wasm::R_WEBASSEMBLY_SECTION_OFFSET_I32: 160 return true; 161 default: 162 return false; 163 } 164 } 165 166 void print(raw_ostream &Out) const { 167 Out << wasm::relocTypetoString(Type) << " Off=" << Offset 168 << ", Sym=" << *Symbol << ", Addend=" << Addend 169 << ", FixupSection=" << FixupSection->getSectionName(); 170 } 171 172 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 173 LLVM_DUMP_METHOD void dump() const { print(dbgs()); } 174 #endif 175 }; 176 177 static const uint32_t INVALID_INDEX = -1; 178 179 struct WasmCustomSection { 180 181 StringRef Name; 182 MCSectionWasm *Section; 183 184 uint32_t OutputContentsOffset; 185 uint32_t OutputIndex; 186 187 WasmCustomSection(StringRef Name, MCSectionWasm *Section) 188 : Name(Name), Section(Section), OutputContentsOffset(0), 189 OutputIndex(INVALID_INDEX) {} 190 }; 191 192 #if !defined(NDEBUG) 193 raw_ostream &operator<<(raw_ostream &OS, const WasmRelocationEntry &Rel) { 194 Rel.print(OS); 195 return OS; 196 } 197 #endif 198 199 class WasmObjectWriter : public MCObjectWriter { 200 support::endian::Writer W; 201 202 /// The target specific Wasm writer instance. 203 std::unique_ptr<MCWasmObjectTargetWriter> TargetObjectWriter; 204 205 // Relocations for fixing up references in the code section. 206 std::vector<WasmRelocationEntry> CodeRelocations; 207 uint32_t CodeSectionIndex; 208 209 // Relocations for fixing up references in the data section. 210 std::vector<WasmRelocationEntry> DataRelocations; 211 uint32_t DataSectionIndex; 212 213 // Index values to use for fixing up call_indirect type indices. 214 // Maps function symbols to the index of the type of the function 215 DenseMap<const MCSymbolWasm *, uint32_t> TypeIndices; 216 // Maps function symbols to the table element index space. Used 217 // for TABLE_INDEX relocation types (i.e. address taken functions). 218 DenseMap<const MCSymbolWasm *, uint32_t> TableIndices; 219 // Maps function/global symbols to the function/global/section index space. 220 DenseMap<const MCSymbolWasm *, uint32_t> WasmIndices; 221 // Maps data symbols to the Wasm segment and offset/size with the segment. 222 DenseMap<const MCSymbolWasm *, wasm::WasmDataReference> DataLocations; 223 224 // Stores output data (index, relocations, content offset) for custom 225 // section. 226 std::vector<WasmCustomSection> CustomSections; 227 // Relocations for fixing up references in the custom sections. 228 DenseMap<const MCSectionWasm *, std::vector<WasmRelocationEntry>> 229 CustomSectionsRelocations; 230 231 // Map from section to defining function symbol. 232 DenseMap<const MCSection *, const MCSymbol *> SectionFunctions; 233 234 DenseMap<WasmFunctionType, uint32_t, WasmFunctionTypeDenseMapInfo> 235 FunctionTypeIndices; 236 SmallVector<WasmFunctionType, 4> FunctionTypes; 237 SmallVector<WasmGlobal, 4> Globals; 238 SmallVector<WasmDataSegment, 4> DataSegments; 239 unsigned NumFunctionImports = 0; 240 unsigned NumGlobalImports = 0; 241 uint32_t SectionCount = 0; 242 243 // TargetObjectWriter wrappers. 244 bool is64Bit() const { return TargetObjectWriter->is64Bit(); } 245 unsigned getRelocType(const MCValue &Target, const MCFixup &Fixup) const { 246 return TargetObjectWriter->getRelocType(Target, Fixup); 247 } 248 249 void startSection(SectionBookkeeping &Section, unsigned SectionId); 250 void startCustomSection(SectionBookkeeping &Section, StringRef Name); 251 void endSection(SectionBookkeeping &Section); 252 253 public: 254 WasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 255 raw_pwrite_stream &OS) 256 : W(OS, support::little), TargetObjectWriter(std::move(MOTW)) {} 257 258 ~WasmObjectWriter() override; 259 260 private: 261 void reset() override { 262 CodeRelocations.clear(); 263 DataRelocations.clear(); 264 TypeIndices.clear(); 265 WasmIndices.clear(); 266 TableIndices.clear(); 267 DataLocations.clear(); 268 CustomSectionsRelocations.clear(); 269 FunctionTypeIndices.clear(); 270 FunctionTypes.clear(); 271 Globals.clear(); 272 DataSegments.clear(); 273 SectionFunctions.clear(); 274 NumFunctionImports = 0; 275 NumGlobalImports = 0; 276 MCObjectWriter::reset(); 277 } 278 279 void writeHeader(const MCAssembler &Asm); 280 281 void recordRelocation(MCAssembler &Asm, const MCAsmLayout &Layout, 282 const MCFragment *Fragment, const MCFixup &Fixup, 283 MCValue Target, uint64_t &FixedValue) override; 284 285 void executePostLayoutBinding(MCAssembler &Asm, 286 const MCAsmLayout &Layout) override; 287 288 uint64_t writeObject(MCAssembler &Asm, const MCAsmLayout &Layout) override; 289 290 void writeString(const StringRef Str) { 291 encodeULEB128(Str.size(), W.OS); 292 W.OS << Str; 293 } 294 295 void writeValueType(wasm::ValType Ty) { W.OS << static_cast<char>(Ty); } 296 297 void writeTypeSection(ArrayRef<WasmFunctionType> FunctionTypes); 298 void writeImportSection(ArrayRef<wasm::WasmImport> Imports, uint32_t DataSize, 299 uint32_t NumElements); 300 void writeFunctionSection(ArrayRef<WasmFunction> Functions); 301 void writeGlobalSection(); 302 void writeExportSection(ArrayRef<wasm::WasmExport> Exports); 303 void writeElemSection(ArrayRef<uint32_t> TableElems); 304 void writeCodeSection(const MCAssembler &Asm, const MCAsmLayout &Layout, 305 ArrayRef<WasmFunction> Functions); 306 void writeDataSection(); 307 void writeRelocSection(uint32_t SectionIndex, StringRef Name, 308 std::vector<WasmRelocationEntry> &Relocations); 309 void writeLinkingMetaDataSection( 310 ArrayRef<wasm::WasmSymbolInfo> SymbolInfos, 311 ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs, 312 const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats); 313 void writeCustomSections(const MCAssembler &Asm, const MCAsmLayout &Layout); 314 void writeCustomRelocSections(); 315 void 316 updateCustomSectionRelocations(const SmallVector<WasmFunction, 4> &Functions, 317 const MCAsmLayout &Layout); 318 319 uint32_t getProvisionalValue(const WasmRelocationEntry &RelEntry); 320 void applyRelocations(ArrayRef<WasmRelocationEntry> Relocations, 321 uint64_t ContentsOffset); 322 323 uint32_t getRelocationIndexValue(const WasmRelocationEntry &RelEntry); 324 uint32_t getFunctionType(const MCSymbolWasm &Symbol); 325 uint32_t registerFunctionType(const MCSymbolWasm &Symbol); 326 }; 327 328 } // end anonymous namespace 329 330 WasmObjectWriter::~WasmObjectWriter() {} 331 332 // Write out a section header and a patchable section size field. 333 void WasmObjectWriter::startSection(SectionBookkeeping &Section, 334 unsigned SectionId) { 335 LLVM_DEBUG(dbgs() << "startSection " << SectionId << "\n"); 336 W.OS << char(SectionId); 337 338 Section.SizeOffset = W.OS.tell(); 339 340 // The section size. We don't know the size yet, so reserve enough space 341 // for any 32-bit value; we'll patch it later. 342 encodeULEB128(UINT32_MAX, W.OS); 343 344 // The position where the section starts, for measuring its size. 345 Section.ContentsOffset = W.OS.tell(); 346 Section.PayloadOffset = W.OS.tell(); 347 Section.Index = SectionCount++; 348 } 349 350 void WasmObjectWriter::startCustomSection(SectionBookkeeping &Section, 351 StringRef Name) { 352 LLVM_DEBUG(dbgs() << "startCustomSection " << Name << "\n"); 353 startSection(Section, wasm::WASM_SEC_CUSTOM); 354 355 // The position where the section header ends, for measuring its size. 356 Section.PayloadOffset = W.OS.tell(); 357 358 // Custom sections in wasm also have a string identifier. 359 writeString(Name); 360 361 // The position where the custom section starts. 362 Section.ContentsOffset = W.OS.tell(); 363 } 364 365 // Now that the section is complete and we know how big it is, patch up the 366 // section size field at the start of the section. 367 void WasmObjectWriter::endSection(SectionBookkeeping &Section) { 368 uint64_t Size = W.OS.tell() - Section.PayloadOffset; 369 if (uint32_t(Size) != Size) 370 report_fatal_error("section size does not fit in a uint32_t"); 371 372 LLVM_DEBUG(dbgs() << "endSection size=" << Size << "\n"); 373 374 // Write the final section size to the payload_len field, which follows 375 // the section id byte. 376 uint8_t Buffer[16]; 377 unsigned SizeLen = encodeULEB128(Size, Buffer, 5); 378 assert(SizeLen == 5); 379 static_cast<raw_pwrite_stream &>(W.OS).pwrite((char *)Buffer, SizeLen, 380 Section.SizeOffset); 381 } 382 383 // Emit the Wasm header. 384 void WasmObjectWriter::writeHeader(const MCAssembler &Asm) { 385 W.OS.write(wasm::WasmMagic, sizeof(wasm::WasmMagic)); 386 W.write<uint32_t>(wasm::WasmVersion); 387 } 388 389 void WasmObjectWriter::executePostLayoutBinding(MCAssembler &Asm, 390 const MCAsmLayout &Layout) { 391 // Build a map of sections to the function that defines them, for use 392 // in recordRelocation. 393 for (const MCSymbol &S : Asm.symbols()) { 394 const auto &WS = static_cast<const MCSymbolWasm &>(S); 395 if (WS.isDefined() && WS.isFunction() && !WS.isVariable()) { 396 const auto &Sec = static_cast<const MCSectionWasm &>(S.getSection()); 397 auto Pair = SectionFunctions.insert(std::make_pair(&Sec, &S)); 398 if (!Pair.second) 399 report_fatal_error("section already has a defining function: " + 400 Sec.getSectionName()); 401 } 402 } 403 } 404 405 void WasmObjectWriter::recordRelocation(MCAssembler &Asm, 406 const MCAsmLayout &Layout, 407 const MCFragment *Fragment, 408 const MCFixup &Fixup, MCValue Target, 409 uint64_t &FixedValue) { 410 MCAsmBackend &Backend = Asm.getBackend(); 411 bool IsPCRel = Backend.getFixupKindInfo(Fixup.getKind()).Flags & 412 MCFixupKindInfo::FKF_IsPCRel; 413 const auto &FixupSection = cast<MCSectionWasm>(*Fragment->getParent()); 414 uint64_t C = Target.getConstant(); 415 uint64_t FixupOffset = Layout.getFragmentOffset(Fragment) + Fixup.getOffset(); 416 MCContext &Ctx = Asm.getContext(); 417 418 // The .init_array isn't translated as data, so don't do relocations in it. 419 if (FixupSection.getSectionName().startswith(".init_array")) 420 return; 421 422 if (const MCSymbolRefExpr *RefB = Target.getSymB()) { 423 assert(RefB->getKind() == MCSymbolRefExpr::VK_None && 424 "Should not have constructed this"); 425 426 // Let A, B and C being the components of Target and R be the location of 427 // the fixup. If the fixup is not pcrel, we want to compute (A - B + C). 428 // If it is pcrel, we want to compute (A - B + C - R). 429 430 // In general, Wasm has no relocations for -B. It can only represent (A + C) 431 // or (A + C - R). If B = R + K and the relocation is not pcrel, we can 432 // replace B to implement it: (A - R - K + C) 433 if (IsPCRel) { 434 Ctx.reportError( 435 Fixup.getLoc(), 436 "No relocation available to represent this relative expression"); 437 return; 438 } 439 440 const auto &SymB = cast<MCSymbolWasm>(RefB->getSymbol()); 441 442 if (SymB.isUndefined()) { 443 Ctx.reportError(Fixup.getLoc(), 444 Twine("symbol '") + SymB.getName() + 445 "' can not be undefined in a subtraction expression"); 446 return; 447 } 448 449 assert(!SymB.isAbsolute() && "Should have been folded"); 450 const MCSection &SecB = SymB.getSection(); 451 if (&SecB != &FixupSection) { 452 Ctx.reportError(Fixup.getLoc(), 453 "Cannot represent a difference across sections"); 454 return; 455 } 456 457 uint64_t SymBOffset = Layout.getSymbolOffset(SymB); 458 uint64_t K = SymBOffset - FixupOffset; 459 IsPCRel = true; 460 C -= K; 461 } 462 463 // We either rejected the fixup or folded B into C at this point. 464 const MCSymbolRefExpr *RefA = Target.getSymA(); 465 const auto *SymA = RefA ? cast<MCSymbolWasm>(&RefA->getSymbol()) : nullptr; 466 467 if (SymA && SymA->isVariable()) { 468 const MCExpr *Expr = SymA->getVariableValue(); 469 const auto *Inner = cast<MCSymbolRefExpr>(Expr); 470 if (Inner->getKind() == MCSymbolRefExpr::VK_WEAKREF) 471 llvm_unreachable("weakref used in reloc not yet implemented"); 472 } 473 474 // Put any constant offset in an addend. Offsets can be negative, and 475 // LLVM expects wrapping, in contrast to wasm's immediates which can't 476 // be negative and don't wrap. 477 FixedValue = 0; 478 479 unsigned Type = getRelocType(Target, Fixup); 480 assert(!IsPCRel); 481 assert(SymA); 482 483 // Absolute offset within a section or a function. 484 // Currently only supported for for metadata sections. 485 // See: test/MC/WebAssembly/blockaddress.ll 486 if (Type == wasm::R_WEBASSEMBLY_FUNCTION_OFFSET_I32 || 487 Type == wasm::R_WEBASSEMBLY_SECTION_OFFSET_I32) { 488 if (!FixupSection.getKind().isMetadata()) 489 report_fatal_error("relocations for function or section offsets are " 490 "only supported in metadata sections"); 491 492 const MCSymbol *SectionSymbol = nullptr; 493 const MCSection &SecA = SymA->getSection(); 494 if (SecA.getKind().isText()) 495 SectionSymbol = SectionFunctions.find(&SecA)->second; 496 else 497 SectionSymbol = SecA.getBeginSymbol(); 498 if (!SectionSymbol) 499 report_fatal_error("section symbol is required for relocation"); 500 501 C += Layout.getSymbolOffset(*SymA); 502 SymA = cast<MCSymbolWasm>(SectionSymbol); 503 } 504 505 // Relocation other than R_WEBASSEMBLY_TYPE_INDEX_LEB are required to be 506 // against a named symbol. 507 if (Type != wasm::R_WEBASSEMBLY_TYPE_INDEX_LEB) { 508 if (SymA->getName().empty()) 509 report_fatal_error("relocations against un-named temporaries are not yet " 510 "supported by wasm"); 511 512 SymA->setUsedInReloc(); 513 } 514 515 WasmRelocationEntry Rec(FixupOffset, SymA, C, Type, &FixupSection); 516 LLVM_DEBUG(dbgs() << "WasmReloc: " << Rec << "\n"); 517 518 if (FixupSection.isWasmData()) { 519 DataRelocations.push_back(Rec); 520 } else if (FixupSection.getKind().isText()) { 521 CodeRelocations.push_back(Rec); 522 } else if (FixupSection.getKind().isMetadata()) { 523 CustomSectionsRelocations[&FixupSection].push_back(Rec); 524 } else { 525 llvm_unreachable("unexpected section type"); 526 } 527 } 528 529 // Write X as an (unsigned) LEB value at offset Offset in Stream, padded 530 // to allow patching. 531 static void WritePatchableLEB(raw_pwrite_stream &Stream, uint32_t X, 532 uint64_t Offset) { 533 uint8_t Buffer[5]; 534 unsigned SizeLen = encodeULEB128(X, Buffer, 5); 535 assert(SizeLen == 5); 536 Stream.pwrite((char *)Buffer, SizeLen, Offset); 537 } 538 539 // Write X as an signed LEB value at offset Offset in Stream, padded 540 // to allow patching. 541 static void WritePatchableSLEB(raw_pwrite_stream &Stream, int32_t X, 542 uint64_t Offset) { 543 uint8_t Buffer[5]; 544 unsigned SizeLen = encodeSLEB128(X, Buffer, 5); 545 assert(SizeLen == 5); 546 Stream.pwrite((char *)Buffer, SizeLen, Offset); 547 } 548 549 // Write X as a plain integer value at offset Offset in Stream. 550 static void WriteI32(raw_pwrite_stream &Stream, uint32_t X, uint64_t Offset) { 551 uint8_t Buffer[4]; 552 support::endian::write32le(Buffer, X); 553 Stream.pwrite((char *)Buffer, sizeof(Buffer), Offset); 554 } 555 556 static const MCSymbolWasm *ResolveSymbol(const MCSymbolWasm &Symbol) { 557 if (Symbol.isVariable()) { 558 const MCExpr *Expr = Symbol.getVariableValue(); 559 auto *Inner = cast<MCSymbolRefExpr>(Expr); 560 return cast<MCSymbolWasm>(&Inner->getSymbol()); 561 } 562 return &Symbol; 563 } 564 565 // Compute a value to write into the code at the location covered 566 // by RelEntry. This value isn't used by the static linker; it just serves 567 // to make the object format more readable and more likely to be directly 568 // useable. 569 uint32_t 570 WasmObjectWriter::getProvisionalValue(const WasmRelocationEntry &RelEntry) { 571 switch (RelEntry.Type) { 572 case wasm::R_WEBASSEMBLY_TABLE_INDEX_SLEB: 573 case wasm::R_WEBASSEMBLY_TABLE_INDEX_I32: { 574 // Provisional value is table address of the resolved symbol itself 575 const MCSymbolWasm *Sym = ResolveSymbol(*RelEntry.Symbol); 576 assert(Sym->isFunction()); 577 return TableIndices[Sym]; 578 } 579 case wasm::R_WEBASSEMBLY_TYPE_INDEX_LEB: 580 // Provisional value is same as the index 581 return getRelocationIndexValue(RelEntry); 582 case wasm::R_WEBASSEMBLY_FUNCTION_INDEX_LEB: 583 case wasm::R_WEBASSEMBLY_GLOBAL_INDEX_LEB: 584 // Provisional value is function/global Wasm index 585 if (!WasmIndices.count(RelEntry.Symbol)) 586 report_fatal_error("symbol not found in wasm index space: " + 587 RelEntry.Symbol->getName()); 588 return WasmIndices[RelEntry.Symbol]; 589 case wasm::R_WEBASSEMBLY_FUNCTION_OFFSET_I32: 590 case wasm::R_WEBASSEMBLY_SECTION_OFFSET_I32: { 591 const auto &Section = 592 static_cast<const MCSectionWasm &>(RelEntry.Symbol->getSection()); 593 return Section.getSectionOffset() + RelEntry.Addend; 594 } 595 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_LEB: 596 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_I32: 597 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_SLEB: { 598 // Provisional value is address of the global 599 const MCSymbolWasm *Sym = ResolveSymbol(*RelEntry.Symbol); 600 // For undefined symbols, use zero 601 if (!Sym->isDefined()) 602 return 0; 603 const wasm::WasmDataReference &Ref = DataLocations[Sym]; 604 const WasmDataSegment &Segment = DataSegments[Ref.Segment]; 605 // Ignore overflow. LLVM allows address arithmetic to silently wrap. 606 return Segment.Offset + Ref.Offset + RelEntry.Addend; 607 } 608 default: 609 llvm_unreachable("invalid relocation type"); 610 } 611 } 612 613 static void addData(SmallVectorImpl<char> &DataBytes, 614 MCSectionWasm &DataSection) { 615 LLVM_DEBUG(errs() << "addData: " << DataSection.getSectionName() << "\n"); 616 617 DataBytes.resize(alignTo(DataBytes.size(), DataSection.getAlignment())); 618 619 for (const MCFragment &Frag : DataSection) { 620 if (Frag.hasInstructions()) 621 report_fatal_error("only data supported in data sections"); 622 623 if (auto *Align = dyn_cast<MCAlignFragment>(&Frag)) { 624 if (Align->getValueSize() != 1) 625 report_fatal_error("only byte values supported for alignment"); 626 // If nops are requested, use zeros, as this is the data section. 627 uint8_t Value = Align->hasEmitNops() ? 0 : Align->getValue(); 628 uint64_t Size = 629 std::min<uint64_t>(alignTo(DataBytes.size(), Align->getAlignment()), 630 DataBytes.size() + Align->getMaxBytesToEmit()); 631 DataBytes.resize(Size, Value); 632 } else if (auto *Fill = dyn_cast<MCFillFragment>(&Frag)) { 633 int64_t NumValues; 634 if (!Fill->getNumValues().evaluateAsAbsolute(NumValues)) 635 llvm_unreachable("The fill should be an assembler constant"); 636 DataBytes.insert(DataBytes.end(), Fill->getValueSize() * NumValues, 637 Fill->getValue()); 638 } else if (auto *LEB = dyn_cast<MCLEBFragment>(&Frag)) { 639 const SmallVectorImpl<char> &Contents = LEB->getContents(); 640 DataBytes.insert(DataBytes.end(), Contents.begin(), Contents.end()); 641 } else { 642 const auto &DataFrag = cast<MCDataFragment>(Frag); 643 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 644 DataBytes.insert(DataBytes.end(), Contents.begin(), Contents.end()); 645 } 646 } 647 648 LLVM_DEBUG(dbgs() << "addData -> " << DataBytes.size() << "\n"); 649 } 650 651 uint32_t 652 WasmObjectWriter::getRelocationIndexValue(const WasmRelocationEntry &RelEntry) { 653 if (RelEntry.Type == wasm::R_WEBASSEMBLY_TYPE_INDEX_LEB) { 654 if (!TypeIndices.count(RelEntry.Symbol)) 655 report_fatal_error("symbol not found in type index space: " + 656 RelEntry.Symbol->getName()); 657 return TypeIndices[RelEntry.Symbol]; 658 } 659 660 return RelEntry.Symbol->getIndex(); 661 } 662 663 // Apply the portions of the relocation records that we can handle ourselves 664 // directly. 665 void WasmObjectWriter::applyRelocations( 666 ArrayRef<WasmRelocationEntry> Relocations, uint64_t ContentsOffset) { 667 auto &Stream = static_cast<raw_pwrite_stream &>(W.OS); 668 for (const WasmRelocationEntry &RelEntry : Relocations) { 669 uint64_t Offset = ContentsOffset + 670 RelEntry.FixupSection->getSectionOffset() + 671 RelEntry.Offset; 672 673 LLVM_DEBUG(dbgs() << "applyRelocation: " << RelEntry << "\n"); 674 uint32_t Value = getProvisionalValue(RelEntry); 675 676 switch (RelEntry.Type) { 677 case wasm::R_WEBASSEMBLY_FUNCTION_INDEX_LEB: 678 case wasm::R_WEBASSEMBLY_TYPE_INDEX_LEB: 679 case wasm::R_WEBASSEMBLY_GLOBAL_INDEX_LEB: 680 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_LEB: 681 WritePatchableLEB(Stream, Value, Offset); 682 break; 683 case wasm::R_WEBASSEMBLY_TABLE_INDEX_I32: 684 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_I32: 685 case wasm::R_WEBASSEMBLY_FUNCTION_OFFSET_I32: 686 case wasm::R_WEBASSEMBLY_SECTION_OFFSET_I32: 687 WriteI32(Stream, Value, Offset); 688 break; 689 case wasm::R_WEBASSEMBLY_TABLE_INDEX_SLEB: 690 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_SLEB: 691 WritePatchableSLEB(Stream, Value, Offset); 692 break; 693 default: 694 llvm_unreachable("invalid relocation type"); 695 } 696 } 697 } 698 699 void WasmObjectWriter::writeTypeSection( 700 ArrayRef<WasmFunctionType> FunctionTypes) { 701 if (FunctionTypes.empty()) 702 return; 703 704 SectionBookkeeping Section; 705 startSection(Section, wasm::WASM_SEC_TYPE); 706 707 encodeULEB128(FunctionTypes.size(), W.OS); 708 709 for (const WasmFunctionType &FuncTy : FunctionTypes) { 710 W.OS << char(wasm::WASM_TYPE_FUNC); 711 encodeULEB128(FuncTy.Params.size(), W.OS); 712 for (wasm::ValType Ty : FuncTy.Params) 713 writeValueType(Ty); 714 encodeULEB128(FuncTy.Returns.size(), W.OS); 715 for (wasm::ValType Ty : FuncTy.Returns) 716 writeValueType(Ty); 717 } 718 719 endSection(Section); 720 } 721 722 void WasmObjectWriter::writeImportSection(ArrayRef<wasm::WasmImport> Imports, 723 uint32_t DataSize, 724 uint32_t NumElements) { 725 if (Imports.empty()) 726 return; 727 728 uint32_t NumPages = (DataSize + wasm::WasmPageSize - 1) / wasm::WasmPageSize; 729 730 SectionBookkeeping Section; 731 startSection(Section, wasm::WASM_SEC_IMPORT); 732 733 encodeULEB128(Imports.size(), W.OS); 734 for (const wasm::WasmImport &Import : Imports) { 735 writeString(Import.Module); 736 writeString(Import.Field); 737 W.OS << char(Import.Kind); 738 739 switch (Import.Kind) { 740 case wasm::WASM_EXTERNAL_FUNCTION: 741 encodeULEB128(Import.SigIndex, W.OS); 742 break; 743 case wasm::WASM_EXTERNAL_GLOBAL: 744 W.OS << char(Import.Global.Type); 745 W.OS << char(Import.Global.Mutable ? 1 : 0); 746 break; 747 case wasm::WASM_EXTERNAL_MEMORY: 748 encodeULEB128(0, W.OS); // flags 749 encodeULEB128(NumPages, W.OS); // initial 750 break; 751 case wasm::WASM_EXTERNAL_TABLE: 752 W.OS << char(Import.Table.ElemType); 753 encodeULEB128(0, W.OS); // flags 754 encodeULEB128(NumElements, W.OS); // initial 755 break; 756 default: 757 llvm_unreachable("unsupported import kind"); 758 } 759 } 760 761 endSection(Section); 762 } 763 764 void WasmObjectWriter::writeFunctionSection(ArrayRef<WasmFunction> Functions) { 765 if (Functions.empty()) 766 return; 767 768 SectionBookkeeping Section; 769 startSection(Section, wasm::WASM_SEC_FUNCTION); 770 771 encodeULEB128(Functions.size(), W.OS); 772 for (const WasmFunction &Func : Functions) 773 encodeULEB128(Func.Type, W.OS); 774 775 endSection(Section); 776 } 777 778 void WasmObjectWriter::writeGlobalSection() { 779 if (Globals.empty()) 780 return; 781 782 SectionBookkeeping Section; 783 startSection(Section, wasm::WASM_SEC_GLOBAL); 784 785 encodeULEB128(Globals.size(), W.OS); 786 for (const WasmGlobal &Global : Globals) { 787 writeValueType(static_cast<wasm::ValType>(Global.Type.Type)); 788 W.OS << char(Global.Type.Mutable); 789 790 W.OS << char(wasm::WASM_OPCODE_I32_CONST); 791 encodeSLEB128(Global.InitialValue, W.OS); 792 W.OS << char(wasm::WASM_OPCODE_END); 793 } 794 795 endSection(Section); 796 } 797 798 void WasmObjectWriter::writeExportSection(ArrayRef<wasm::WasmExport> Exports) { 799 if (Exports.empty()) 800 return; 801 802 SectionBookkeeping Section; 803 startSection(Section, wasm::WASM_SEC_EXPORT); 804 805 encodeULEB128(Exports.size(), W.OS); 806 for (const wasm::WasmExport &Export : Exports) { 807 writeString(Export.Name); 808 W.OS << char(Export.Kind); 809 encodeULEB128(Export.Index, W.OS); 810 } 811 812 endSection(Section); 813 } 814 815 void WasmObjectWriter::writeElemSection(ArrayRef<uint32_t> TableElems) { 816 if (TableElems.empty()) 817 return; 818 819 SectionBookkeeping Section; 820 startSection(Section, wasm::WASM_SEC_ELEM); 821 822 encodeULEB128(1, W.OS); // number of "segments" 823 encodeULEB128(0, W.OS); // the table index 824 825 // init expr for starting offset 826 W.OS << char(wasm::WASM_OPCODE_I32_CONST); 827 encodeSLEB128(kInitialTableOffset, W.OS); 828 W.OS << char(wasm::WASM_OPCODE_END); 829 830 encodeULEB128(TableElems.size(), W.OS); 831 for (uint32_t Elem : TableElems) 832 encodeULEB128(Elem, W.OS); 833 834 endSection(Section); 835 } 836 837 void WasmObjectWriter::writeCodeSection(const MCAssembler &Asm, 838 const MCAsmLayout &Layout, 839 ArrayRef<WasmFunction> Functions) { 840 if (Functions.empty()) 841 return; 842 843 SectionBookkeeping Section; 844 startSection(Section, wasm::WASM_SEC_CODE); 845 CodeSectionIndex = Section.Index; 846 847 encodeULEB128(Functions.size(), W.OS); 848 849 for (const WasmFunction &Func : Functions) { 850 auto &FuncSection = static_cast<MCSectionWasm &>(Func.Sym->getSection()); 851 852 int64_t Size = 0; 853 if (!Func.Sym->getSize()->evaluateAsAbsolute(Size, Layout)) 854 report_fatal_error(".size expression must be evaluatable"); 855 856 encodeULEB128(Size, W.OS); 857 FuncSection.setSectionOffset(W.OS.tell() - Section.ContentsOffset); 858 Asm.writeSectionData(W.OS, &FuncSection, Layout); 859 } 860 861 // Apply fixups. 862 applyRelocations(CodeRelocations, Section.ContentsOffset); 863 864 endSection(Section); 865 } 866 867 void WasmObjectWriter::writeDataSection() { 868 if (DataSegments.empty()) 869 return; 870 871 SectionBookkeeping Section; 872 startSection(Section, wasm::WASM_SEC_DATA); 873 DataSectionIndex = Section.Index; 874 875 encodeULEB128(DataSegments.size(), W.OS); // count 876 877 for (const WasmDataSegment &Segment : DataSegments) { 878 encodeULEB128(0, W.OS); // memory index 879 W.OS << char(wasm::WASM_OPCODE_I32_CONST); 880 encodeSLEB128(Segment.Offset, W.OS); // offset 881 W.OS << char(wasm::WASM_OPCODE_END); 882 encodeULEB128(Segment.Data.size(), W.OS); // size 883 Segment.Section->setSectionOffset(W.OS.tell() - Section.ContentsOffset); 884 W.OS << Segment.Data; // data 885 } 886 887 // Apply fixups. 888 applyRelocations(DataRelocations, Section.ContentsOffset); 889 890 endSection(Section); 891 } 892 893 void WasmObjectWriter::writeRelocSection( 894 uint32_t SectionIndex, StringRef Name, 895 std::vector<WasmRelocationEntry> &Relocs) { 896 // See: https://github.com/WebAssembly/tool-conventions/blob/master/Linking.md 897 // for descriptions of the reloc sections. 898 899 if (Relocs.empty()) 900 return; 901 902 // First, ensure the relocations are sorted in offset order. In general they 903 // should already be sorted since `recordRelocation` is called in offset 904 // order, but for the code section we combine many MC sections into single 905 // wasm section, and this order is determined by the order of Asm.Symbols() 906 // not the sections order. 907 std::stable_sort( 908 Relocs.begin(), Relocs.end(), 909 [](const WasmRelocationEntry &A, const WasmRelocationEntry &B) { 910 return (A.Offset + A.FixupSection->getSectionOffset()) < 911 (B.Offset + B.FixupSection->getSectionOffset()); 912 }); 913 914 SectionBookkeeping Section; 915 startCustomSection(Section, std::string("reloc.") + Name.str()); 916 917 encodeULEB128(SectionIndex, W.OS); 918 encodeULEB128(Relocs.size(), W.OS); 919 for (const WasmRelocationEntry &RelEntry : Relocs) { 920 uint64_t Offset = 921 RelEntry.Offset + RelEntry.FixupSection->getSectionOffset(); 922 uint32_t Index = getRelocationIndexValue(RelEntry); 923 924 W.OS << char(RelEntry.Type); 925 encodeULEB128(Offset, W.OS); 926 encodeULEB128(Index, W.OS); 927 if (RelEntry.hasAddend()) 928 encodeSLEB128(RelEntry.Addend, W.OS); 929 } 930 931 endSection(Section); 932 } 933 934 void WasmObjectWriter::writeCustomRelocSections() { 935 for (const auto &Sec : CustomSections) { 936 auto &Relocations = CustomSectionsRelocations[Sec.Section]; 937 writeRelocSection(Sec.OutputIndex, Sec.Name, Relocations); 938 } 939 } 940 941 void WasmObjectWriter::writeLinkingMetaDataSection( 942 ArrayRef<wasm::WasmSymbolInfo> SymbolInfos, 943 ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs, 944 const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats) { 945 SectionBookkeeping Section; 946 startCustomSection(Section, "linking"); 947 encodeULEB128(wasm::WasmMetadataVersion, W.OS); 948 949 SectionBookkeeping SubSection; 950 if (SymbolInfos.size() != 0) { 951 startSection(SubSection, wasm::WASM_SYMBOL_TABLE); 952 encodeULEB128(SymbolInfos.size(), W.OS); 953 for (const wasm::WasmSymbolInfo &Sym : SymbolInfos) { 954 encodeULEB128(Sym.Kind, W.OS); 955 encodeULEB128(Sym.Flags, W.OS); 956 switch (Sym.Kind) { 957 case wasm::WASM_SYMBOL_TYPE_FUNCTION: 958 case wasm::WASM_SYMBOL_TYPE_GLOBAL: 959 encodeULEB128(Sym.ElementIndex, W.OS); 960 if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0) 961 writeString(Sym.Name); 962 break; 963 case wasm::WASM_SYMBOL_TYPE_DATA: 964 writeString(Sym.Name); 965 if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0) { 966 encodeULEB128(Sym.DataRef.Segment, W.OS); 967 encodeULEB128(Sym.DataRef.Offset, W.OS); 968 encodeULEB128(Sym.DataRef.Size, W.OS); 969 } 970 break; 971 case wasm::WASM_SYMBOL_TYPE_SECTION: { 972 const uint32_t SectionIndex = 973 CustomSections[Sym.ElementIndex].OutputIndex; 974 encodeULEB128(SectionIndex, W.OS); 975 break; 976 } 977 default: 978 llvm_unreachable("unexpected kind"); 979 } 980 } 981 endSection(SubSection); 982 } 983 984 if (DataSegments.size()) { 985 startSection(SubSection, wasm::WASM_SEGMENT_INFO); 986 encodeULEB128(DataSegments.size(), W.OS); 987 for (const WasmDataSegment &Segment : DataSegments) { 988 writeString(Segment.Name); 989 encodeULEB128(Segment.Alignment, W.OS); 990 encodeULEB128(Segment.Flags, W.OS); 991 } 992 endSection(SubSection); 993 } 994 995 if (!InitFuncs.empty()) { 996 startSection(SubSection, wasm::WASM_INIT_FUNCS); 997 encodeULEB128(InitFuncs.size(), W.OS); 998 for (auto &StartFunc : InitFuncs) { 999 encodeULEB128(StartFunc.first, W.OS); // priority 1000 encodeULEB128(StartFunc.second, W.OS); // function index 1001 } 1002 endSection(SubSection); 1003 } 1004 1005 if (Comdats.size()) { 1006 startSection(SubSection, wasm::WASM_COMDAT_INFO); 1007 encodeULEB128(Comdats.size(), W.OS); 1008 for (const auto &C : Comdats) { 1009 writeString(C.first); 1010 encodeULEB128(0, W.OS); // flags for future use 1011 encodeULEB128(C.second.size(), W.OS); 1012 for (const WasmComdatEntry &Entry : C.second) { 1013 encodeULEB128(Entry.Kind, W.OS); 1014 encodeULEB128(Entry.Index, W.OS); 1015 } 1016 } 1017 endSection(SubSection); 1018 } 1019 1020 endSection(Section); 1021 } 1022 1023 void WasmObjectWriter::writeCustomSections(const MCAssembler &Asm, 1024 const MCAsmLayout &Layout) { 1025 for (auto &CustomSection : CustomSections) { 1026 SectionBookkeeping Section; 1027 auto *Sec = CustomSection.Section; 1028 startCustomSection(Section, CustomSection.Name); 1029 1030 Sec->setSectionOffset(W.OS.tell() - Section.ContentsOffset); 1031 Asm.writeSectionData(W.OS, Sec, Layout); 1032 1033 CustomSection.OutputContentsOffset = Section.ContentsOffset; 1034 CustomSection.OutputIndex = Section.Index; 1035 1036 endSection(Section); 1037 1038 // Apply fixups. 1039 auto &Relocations = CustomSectionsRelocations[CustomSection.Section]; 1040 applyRelocations(Relocations, CustomSection.OutputContentsOffset); 1041 } 1042 } 1043 1044 uint32_t WasmObjectWriter::getFunctionType(const MCSymbolWasm &Symbol) { 1045 assert(Symbol.isFunction()); 1046 assert(TypeIndices.count(&Symbol)); 1047 return TypeIndices[&Symbol]; 1048 } 1049 1050 uint32_t WasmObjectWriter::registerFunctionType(const MCSymbolWasm &Symbol) { 1051 assert(Symbol.isFunction()); 1052 1053 WasmFunctionType F; 1054 const MCSymbolWasm *ResolvedSym = ResolveSymbol(Symbol); 1055 if (auto *Sig = ResolvedSym->getSignature()) { 1056 F.Returns = Sig->Returns; 1057 F.Params = Sig->Params; 1058 } 1059 1060 auto Pair = 1061 FunctionTypeIndices.insert(std::make_pair(F, FunctionTypes.size())); 1062 if (Pair.second) 1063 FunctionTypes.push_back(F); 1064 TypeIndices[&Symbol] = Pair.first->second; 1065 1066 LLVM_DEBUG(dbgs() << "registerFunctionType: " << Symbol 1067 << " new:" << Pair.second << "\n"); 1068 LLVM_DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n"); 1069 return Pair.first->second; 1070 } 1071 1072 static bool isInSymtab(const MCSymbolWasm &Sym) { 1073 if (Sym.isUsedInReloc()) 1074 return true; 1075 1076 if (Sym.isComdat() && !Sym.isDefined()) 1077 return false; 1078 1079 if (Sym.isTemporary() && Sym.getName().empty()) 1080 return false; 1081 1082 if (Sym.isTemporary() && Sym.isData() && !Sym.getSize()) 1083 return false; 1084 1085 if (Sym.isSection()) 1086 return false; 1087 1088 return true; 1089 } 1090 1091 uint64_t WasmObjectWriter::writeObject(MCAssembler &Asm, 1092 const MCAsmLayout &Layout) { 1093 uint64_t StartOffset = W.OS.tell(); 1094 1095 LLVM_DEBUG(dbgs() << "WasmObjectWriter::writeObject\n"); 1096 MCContext &Ctx = Asm.getContext(); 1097 1098 // Collect information from the available symbols. 1099 SmallVector<WasmFunction, 4> Functions; 1100 SmallVector<uint32_t, 4> TableElems; 1101 SmallVector<wasm::WasmImport, 4> Imports; 1102 SmallVector<wasm::WasmExport, 4> Exports; 1103 SmallVector<wasm::WasmSymbolInfo, 4> SymbolInfos; 1104 SmallVector<std::pair<uint16_t, uint32_t>, 2> InitFuncs; 1105 std::map<StringRef, std::vector<WasmComdatEntry>> Comdats; 1106 uint32_t DataSize = 0; 1107 1108 // For now, always emit the memory import, since loads and stores are not 1109 // valid without it. In the future, we could perhaps be more clever and omit 1110 // it if there are no loads or stores. 1111 MCSymbolWasm *MemorySym = 1112 cast<MCSymbolWasm>(Ctx.getOrCreateSymbol("__linear_memory")); 1113 wasm::WasmImport MemImport; 1114 MemImport.Module = MemorySym->getModuleName(); 1115 MemImport.Field = MemorySym->getName(); 1116 MemImport.Kind = wasm::WASM_EXTERNAL_MEMORY; 1117 Imports.push_back(MemImport); 1118 1119 // For now, always emit the table section, since indirect calls are not 1120 // valid without it. In the future, we could perhaps be more clever and omit 1121 // it if there are no indirect calls. 1122 MCSymbolWasm *TableSym = 1123 cast<MCSymbolWasm>(Ctx.getOrCreateSymbol("__indirect_function_table")); 1124 wasm::WasmImport TableImport; 1125 TableImport.Module = TableSym->getModuleName(); 1126 TableImport.Field = TableSym->getName(); 1127 TableImport.Kind = wasm::WASM_EXTERNAL_TABLE; 1128 TableImport.Table.ElemType = wasm::WASM_TYPE_ANYFUNC; 1129 Imports.push_back(TableImport); 1130 1131 // Populate FunctionTypeIndices, and Imports and WasmIndices for undefined 1132 // symbols. This must be done before populating WasmIndices for defined 1133 // symbols. 1134 for (const MCSymbol &S : Asm.symbols()) { 1135 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1136 1137 // Register types for all functions, including those with private linkage 1138 // (because wasm always needs a type signature). 1139 if (WS.isFunction()) 1140 registerFunctionType(WS); 1141 1142 if (WS.isTemporary()) 1143 continue; 1144 1145 // If the symbol is not defined in this translation unit, import it. 1146 if (!WS.isDefined() && !WS.isComdat()) { 1147 if (WS.isFunction()) { 1148 wasm::WasmImport Import; 1149 Import.Module = WS.getModuleName(); 1150 Import.Field = WS.getName(); 1151 Import.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1152 Import.SigIndex = getFunctionType(WS); 1153 Imports.push_back(Import); 1154 WasmIndices[&WS] = NumFunctionImports++; 1155 } else if (WS.isGlobal()) { 1156 if (WS.isWeak()) 1157 report_fatal_error("undefined global symbol cannot be weak"); 1158 1159 wasm::WasmImport Import; 1160 Import.Module = WS.getModuleName(); 1161 Import.Field = WS.getName(); 1162 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1163 Import.Global = WS.getGlobalType(); 1164 Imports.push_back(Import); 1165 WasmIndices[&WS] = NumGlobalImports++; 1166 } 1167 } 1168 } 1169 1170 // Populate DataSegments and CustomSections, which must be done before 1171 // populating DataLocations. 1172 for (MCSection &Sec : Asm) { 1173 auto &Section = static_cast<MCSectionWasm &>(Sec); 1174 StringRef SectionName = Section.getSectionName(); 1175 1176 // .init_array sections are handled specially elsewhere. 1177 if (SectionName.startswith(".init_array")) 1178 continue; 1179 1180 // Code is handled separately 1181 if (Section.getKind().isText()) 1182 continue; 1183 1184 if (Section.isWasmData()) { 1185 uint32_t SegmentIndex = DataSegments.size(); 1186 DataSize = alignTo(DataSize, Section.getAlignment()); 1187 DataSegments.emplace_back(); 1188 WasmDataSegment &Segment = DataSegments.back(); 1189 Segment.Name = SectionName; 1190 Segment.Offset = DataSize; 1191 Segment.Section = &Section; 1192 addData(Segment.Data, Section); 1193 Segment.Alignment = Section.getAlignment(); 1194 Segment.Flags = 0; 1195 DataSize += Segment.Data.size(); 1196 Section.setSegmentIndex(SegmentIndex); 1197 1198 if (const MCSymbolWasm *C = Section.getGroup()) { 1199 Comdats[C->getName()].emplace_back( 1200 WasmComdatEntry{wasm::WASM_COMDAT_DATA, SegmentIndex}); 1201 } 1202 } else { 1203 // Create custom sections 1204 assert(Sec.getKind().isMetadata()); 1205 1206 StringRef Name = SectionName; 1207 1208 // For user-defined custom sections, strip the prefix 1209 if (Name.startswith(".custom_section.")) 1210 Name = Name.substr(strlen(".custom_section.")); 1211 1212 MCSymbol *Begin = Sec.getBeginSymbol(); 1213 if (Begin) { 1214 WasmIndices[cast<MCSymbolWasm>(Begin)] = CustomSections.size(); 1215 if (SectionName != Begin->getName()) 1216 report_fatal_error("section name and begin symbol should match: " + 1217 Twine(SectionName)); 1218 } 1219 CustomSections.emplace_back(Name, &Section); 1220 } 1221 } 1222 1223 // Populate WasmIndices and DataLocations for defined symbols. 1224 for (const MCSymbol &S : Asm.symbols()) { 1225 // Ignore unnamed temporary symbols, which aren't ever exported, imported, 1226 // or used in relocations. 1227 if (S.isTemporary() && S.getName().empty()) 1228 continue; 1229 1230 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1231 LLVM_DEBUG( 1232 dbgs() << "MCSymbol: " << toString(WS.getType()) << " '" << S << "'" 1233 << " isDefined=" << S.isDefined() << " isExternal=" 1234 << S.isExternal() << " isTemporary=" << S.isTemporary() 1235 << " isWeak=" << WS.isWeak() << " isHidden=" << WS.isHidden() 1236 << " isVariable=" << WS.isVariable() << "\n"); 1237 1238 if (WS.isVariable()) 1239 continue; 1240 if (WS.isComdat() && !WS.isDefined()) 1241 continue; 1242 1243 if (WS.isFunction()) { 1244 unsigned Index; 1245 if (WS.isDefined()) { 1246 if (WS.getOffset() != 0) 1247 report_fatal_error( 1248 "function sections must contain one function each"); 1249 1250 if (WS.getSize() == 0) 1251 report_fatal_error( 1252 "function symbols must have a size set with .size"); 1253 1254 // A definition. Write out the function body. 1255 Index = NumFunctionImports + Functions.size(); 1256 WasmFunction Func; 1257 Func.Type = getFunctionType(WS); 1258 Func.Sym = &WS; 1259 WasmIndices[&WS] = Index; 1260 Functions.push_back(Func); 1261 1262 auto &Section = static_cast<MCSectionWasm &>(WS.getSection()); 1263 if (const MCSymbolWasm *C = Section.getGroup()) { 1264 Comdats[C->getName()].emplace_back( 1265 WasmComdatEntry{wasm::WASM_COMDAT_FUNCTION, Index}); 1266 } 1267 } else { 1268 // An import; the index was assigned above. 1269 Index = WasmIndices.find(&WS)->second; 1270 } 1271 1272 LLVM_DEBUG(dbgs() << " -> function index: " << Index << "\n"); 1273 } else if (WS.isData()) { 1274 if (WS.isTemporary() && !WS.getSize()) 1275 continue; 1276 1277 if (!WS.isDefined()) { 1278 LLVM_DEBUG(dbgs() << " -> segment index: -1" 1279 << "\n"); 1280 continue; 1281 } 1282 1283 if (!WS.getSize()) 1284 report_fatal_error("data symbols must have a size set with .size: " + 1285 WS.getName()); 1286 1287 int64_t Size = 0; 1288 if (!WS.getSize()->evaluateAsAbsolute(Size, Layout)) 1289 report_fatal_error(".size expression must be evaluatable"); 1290 1291 auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection()); 1292 assert(DataSection.isWasmData()); 1293 1294 // For each data symbol, export it in the symtab as a reference to the 1295 // corresponding Wasm data segment. 1296 wasm::WasmDataReference Ref = wasm::WasmDataReference{ 1297 DataSection.getSegmentIndex(), 1298 static_cast<uint32_t>(Layout.getSymbolOffset(WS)), 1299 static_cast<uint32_t>(Size)}; 1300 DataLocations[&WS] = Ref; 1301 LLVM_DEBUG(dbgs() << " -> segment index: " << Ref.Segment << "\n"); 1302 } else if (WS.isGlobal()) { 1303 // A "true" Wasm global (currently just __stack_pointer) 1304 if (WS.isDefined()) 1305 report_fatal_error("don't yet support defined globals"); 1306 1307 // An import; the index was assigned above 1308 LLVM_DEBUG(dbgs() << " -> global index: " 1309 << WasmIndices.find(&WS)->second << "\n"); 1310 } else { 1311 assert(WS.isSection()); 1312 } 1313 } 1314 1315 // Populate WasmIndices and DataLocations for aliased symbols. We need to 1316 // process these in a separate pass because we need to have processed the 1317 // target of the alias before the alias itself and the symbols are not 1318 // necessarily ordered in this way. 1319 for (const MCSymbol &S : Asm.symbols()) { 1320 if (!S.isVariable()) 1321 continue; 1322 1323 assert(S.isDefined()); 1324 1325 // Find the target symbol of this weak alias and export that index 1326 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1327 const MCSymbolWasm *ResolvedSym = ResolveSymbol(WS); 1328 LLVM_DEBUG(dbgs() << WS.getName() << ": weak alias of '" << *ResolvedSym 1329 << "'\n"); 1330 1331 if (WS.isFunction()) { 1332 assert(WasmIndices.count(ResolvedSym) > 0); 1333 uint32_t WasmIndex = WasmIndices.find(ResolvedSym)->second; 1334 WasmIndices[&WS] = WasmIndex; 1335 LLVM_DEBUG(dbgs() << " -> index:" << WasmIndex << "\n"); 1336 } else if (WS.isData()) { 1337 assert(DataLocations.count(ResolvedSym) > 0); 1338 const wasm::WasmDataReference &Ref = 1339 DataLocations.find(ResolvedSym)->second; 1340 DataLocations[&WS] = Ref; 1341 LLVM_DEBUG(dbgs() << " -> index:" << Ref.Segment << "\n"); 1342 } else { 1343 report_fatal_error("don't yet support global aliases"); 1344 } 1345 } 1346 1347 // Finally, populate the symbol table itself, in its "natural" order. 1348 for (const MCSymbol &S : Asm.symbols()) { 1349 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1350 if (!isInSymtab(WS)) { 1351 WS.setIndex(INVALID_INDEX); 1352 continue; 1353 } 1354 LLVM_DEBUG(dbgs() << "adding to symtab: " << WS << "\n"); 1355 1356 uint32_t Flags = 0; 1357 if (WS.isWeak()) 1358 Flags |= wasm::WASM_SYMBOL_BINDING_WEAK; 1359 if (WS.isHidden()) 1360 Flags |= wasm::WASM_SYMBOL_VISIBILITY_HIDDEN; 1361 if (!WS.isExternal() && WS.isDefined()) 1362 Flags |= wasm::WASM_SYMBOL_BINDING_LOCAL; 1363 if (WS.isUndefined()) 1364 Flags |= wasm::WASM_SYMBOL_UNDEFINED; 1365 1366 wasm::WasmSymbolInfo Info; 1367 Info.Name = WS.getName(); 1368 Info.Kind = WS.getType(); 1369 Info.Flags = Flags; 1370 if (!WS.isData()) { 1371 assert(WasmIndices.count(&WS) > 0); 1372 Info.ElementIndex = WasmIndices.find(&WS)->second; 1373 } else if (WS.isDefined()) { 1374 assert(DataLocations.count(&WS) > 0); 1375 Info.DataRef = DataLocations.find(&WS)->second; 1376 } 1377 WS.setIndex(SymbolInfos.size()); 1378 SymbolInfos.emplace_back(Info); 1379 } 1380 1381 { 1382 auto HandleReloc = [&](const WasmRelocationEntry &Rel) { 1383 // Functions referenced by a relocation need to put in the table. This is 1384 // purely to make the object file's provisional values readable, and is 1385 // ignored by the linker, which re-calculates the relocations itself. 1386 if (Rel.Type != wasm::R_WEBASSEMBLY_TABLE_INDEX_I32 && 1387 Rel.Type != wasm::R_WEBASSEMBLY_TABLE_INDEX_SLEB) 1388 return; 1389 assert(Rel.Symbol->isFunction()); 1390 const MCSymbolWasm &WS = *ResolveSymbol(*Rel.Symbol); 1391 uint32_t FunctionIndex = WasmIndices.find(&WS)->second; 1392 uint32_t TableIndex = TableElems.size() + kInitialTableOffset; 1393 if (TableIndices.try_emplace(&WS, TableIndex).second) { 1394 LLVM_DEBUG(dbgs() << " -> adding " << WS.getName() 1395 << " to table: " << TableIndex << "\n"); 1396 TableElems.push_back(FunctionIndex); 1397 registerFunctionType(WS); 1398 } 1399 }; 1400 1401 for (const WasmRelocationEntry &RelEntry : CodeRelocations) 1402 HandleReloc(RelEntry); 1403 for (const WasmRelocationEntry &RelEntry : DataRelocations) 1404 HandleReloc(RelEntry); 1405 } 1406 1407 // Translate .init_array section contents into start functions. 1408 for (const MCSection &S : Asm) { 1409 const auto &WS = static_cast<const MCSectionWasm &>(S); 1410 if (WS.getSectionName().startswith(".fini_array")) 1411 report_fatal_error(".fini_array sections are unsupported"); 1412 if (!WS.getSectionName().startswith(".init_array")) 1413 continue; 1414 if (WS.getFragmentList().empty()) 1415 continue; 1416 1417 // init_array is expected to contain a single non-empty data fragment 1418 if (WS.getFragmentList().size() != 3) 1419 report_fatal_error("only one .init_array section fragment supported"); 1420 1421 auto IT = WS.begin(); 1422 const MCFragment &EmptyFrag = *IT; 1423 if (EmptyFrag.getKind() != MCFragment::FT_Data) 1424 report_fatal_error(".init_array section should be aligned"); 1425 1426 IT = std::next(IT); 1427 const MCFragment &AlignFrag = *IT; 1428 if (AlignFrag.getKind() != MCFragment::FT_Align) 1429 report_fatal_error(".init_array section should be aligned"); 1430 if (cast<MCAlignFragment>(AlignFrag).getAlignment() != (is64Bit() ? 8 : 4)) 1431 report_fatal_error(".init_array section should be aligned for pointers"); 1432 1433 const MCFragment &Frag = *std::next(IT); 1434 if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data) 1435 report_fatal_error("only data supported in .init_array section"); 1436 1437 uint16_t Priority = UINT16_MAX; 1438 unsigned PrefixLength = strlen(".init_array"); 1439 if (WS.getSectionName().size() > PrefixLength) { 1440 if (WS.getSectionName()[PrefixLength] != '.') 1441 report_fatal_error( 1442 ".init_array section priority should start with '.'"); 1443 if (WS.getSectionName() 1444 .substr(PrefixLength + 1) 1445 .getAsInteger(10, Priority)) 1446 report_fatal_error("invalid .init_array section priority"); 1447 } 1448 const auto &DataFrag = cast<MCDataFragment>(Frag); 1449 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 1450 for (const uint8_t * 1451 p = (const uint8_t *)Contents.data(), 1452 *end = (const uint8_t *)Contents.data() + Contents.size(); 1453 p != end; ++p) { 1454 if (*p != 0) 1455 report_fatal_error("non-symbolic data in .init_array section"); 1456 } 1457 for (const MCFixup &Fixup : DataFrag.getFixups()) { 1458 assert(Fixup.getKind() == 1459 MCFixup::getKindForSize(is64Bit() ? 8 : 4, false)); 1460 const MCExpr *Expr = Fixup.getValue(); 1461 auto *Sym = dyn_cast<MCSymbolRefExpr>(Expr); 1462 if (!Sym) 1463 report_fatal_error("fixups in .init_array should be symbol references"); 1464 if (Sym->getKind() != MCSymbolRefExpr::VK_WebAssembly_FUNCTION) 1465 report_fatal_error("symbols in .init_array should be for functions"); 1466 if (Sym->getSymbol().getIndex() == INVALID_INDEX) 1467 report_fatal_error("symbols in .init_array should exist in symbtab"); 1468 InitFuncs.push_back( 1469 std::make_pair(Priority, Sym->getSymbol().getIndex())); 1470 } 1471 } 1472 1473 // Write out the Wasm header. 1474 writeHeader(Asm); 1475 1476 writeTypeSection(FunctionTypes); 1477 writeImportSection(Imports, DataSize, TableElems.size()); 1478 writeFunctionSection(Functions); 1479 // Skip the "table" section; we import the table instead. 1480 // Skip the "memory" section; we import the memory instead. 1481 writeGlobalSection(); 1482 writeExportSection(Exports); 1483 writeElemSection(TableElems); 1484 writeCodeSection(Asm, Layout, Functions); 1485 writeDataSection(); 1486 writeCustomSections(Asm, Layout); 1487 writeLinkingMetaDataSection(SymbolInfos, InitFuncs, Comdats); 1488 writeRelocSection(CodeSectionIndex, "CODE", CodeRelocations); 1489 writeRelocSection(DataSectionIndex, "DATA", DataRelocations); 1490 writeCustomRelocSections(); 1491 1492 // TODO: Translate the .comment section to the output. 1493 return W.OS.tell() - StartOffset; 1494 } 1495 1496 std::unique_ptr<MCObjectWriter> 1497 llvm::createWasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 1498 raw_pwrite_stream &OS) { 1499 return llvm::make_unique<WasmObjectWriter>(std::move(MOTW), OS); 1500 } 1501