1 //===- lib/MC/WasmObjectWriter.cpp - Wasm File Writer ---------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements Wasm object file writer information. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "llvm/ADT/STLExtras.h" 14 #include "llvm/ADT/SmallPtrSet.h" 15 #include "llvm/BinaryFormat/Wasm.h" 16 #include "llvm/Config/llvm-config.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/EndianStream.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 InitialTableOffset = 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 or event, in a struct capable of being used 60 // as a DenseMap key. 61 // TODO: Consider using wasm::WasmSignature directly instead. 62 struct WasmSignature { 63 // Support empty and tombstone instances, needed by DenseMap. 64 enum { Plain, Empty, Tombstone } State = Plain; 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 bool operator==(const WasmSignature &Other) const { 73 return State == Other.State && Returns == Other.Returns && 74 Params == Other.Params; 75 } 76 }; 77 78 // Traits for using WasmSignature in a DenseMap. 79 struct WasmSignatureDenseMapInfo { 80 static WasmSignature getEmptyKey() { 81 WasmSignature Sig; 82 Sig.State = WasmSignature::Empty; 83 return Sig; 84 } 85 static WasmSignature getTombstoneKey() { 86 WasmSignature Sig; 87 Sig.State = WasmSignature::Tombstone; 88 return Sig; 89 } 90 static unsigned getHashValue(const WasmSignature &Sig) { 91 uintptr_t Value = Sig.State; 92 for (wasm::ValType Ret : Sig.Returns) 93 Value += DenseMapInfo<uint32_t>::getHashValue(uint32_t(Ret)); 94 for (wasm::ValType Param : Sig.Params) 95 Value += DenseMapInfo<uint32_t>::getHashValue(uint32_t(Param)); 96 return Value; 97 } 98 static bool isEqual(const WasmSignature &LHS, const WasmSignature &RHS) { 99 return LHS == RHS; 100 } 101 }; 102 103 // A wasm data segment. A wasm binary contains only a single data section 104 // but that can contain many segments, each with their own virtual location 105 // in memory. Each MCSection data created by llvm is modeled as its own 106 // wasm data segment. 107 struct WasmDataSegment { 108 MCSectionWasm *Section; 109 StringRef Name; 110 uint32_t InitFlags; 111 uint32_t Offset; 112 uint32_t Alignment; 113 uint32_t LinkerFlags; 114 SmallVector<char, 4> Data; 115 }; 116 117 // A wasm function to be written into the function section. 118 struct WasmFunction { 119 uint32_t SigIndex; 120 const MCSymbolWasm *Sym; 121 }; 122 123 // A wasm global to be written into the global section. 124 struct WasmGlobal { 125 wasm::WasmGlobalType Type; 126 uint64_t InitialValue; 127 }; 128 129 // Information about a single item which is part of a COMDAT. For each data 130 // segment or function which is in the COMDAT, there is a corresponding 131 // WasmComdatEntry. 132 struct WasmComdatEntry { 133 unsigned Kind; 134 uint32_t Index; 135 }; 136 137 // Information about a single relocation. 138 struct WasmRelocationEntry { 139 uint64_t Offset; // Where is the relocation. 140 const MCSymbolWasm *Symbol; // The symbol to relocate with. 141 int64_t Addend; // A value to add to the symbol. 142 unsigned Type; // The type of the relocation. 143 const MCSectionWasm *FixupSection; // The section the relocation is targeting. 144 145 WasmRelocationEntry(uint64_t Offset, const MCSymbolWasm *Symbol, 146 int64_t Addend, unsigned Type, 147 const MCSectionWasm *FixupSection) 148 : Offset(Offset), Symbol(Symbol), Addend(Addend), Type(Type), 149 FixupSection(FixupSection) {} 150 151 bool hasAddend() const { return wasm::relocTypeHasAddend(Type); } 152 153 void print(raw_ostream &Out) const { 154 Out << wasm::relocTypetoString(Type) << " Off=" << Offset 155 << ", Sym=" << *Symbol << ", Addend=" << Addend 156 << ", FixupSection=" << FixupSection->getName(); 157 } 158 159 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 160 LLVM_DUMP_METHOD void dump() const { print(dbgs()); } 161 #endif 162 }; 163 164 static const uint32_t InvalidIndex = -1; 165 166 struct WasmCustomSection { 167 168 StringRef Name; 169 MCSectionWasm *Section; 170 171 uint32_t OutputContentsOffset; 172 uint32_t OutputIndex; 173 174 WasmCustomSection(StringRef Name, MCSectionWasm *Section) 175 : Name(Name), Section(Section), OutputContentsOffset(0), 176 OutputIndex(InvalidIndex) {} 177 }; 178 179 #if !defined(NDEBUG) 180 raw_ostream &operator<<(raw_ostream &OS, const WasmRelocationEntry &Rel) { 181 Rel.print(OS); 182 return OS; 183 } 184 #endif 185 186 // Write X as an (unsigned) LEB value at offset Offset in Stream, padded 187 // to allow patching. 188 static void writePatchableLEB(raw_pwrite_stream &Stream, uint32_t X, 189 uint64_t Offset) { 190 uint8_t Buffer[5]; 191 unsigned SizeLen = encodeULEB128(X, Buffer, 5); 192 assert(SizeLen == 5); 193 Stream.pwrite((char *)Buffer, SizeLen, Offset); 194 } 195 196 // Write X as an signed LEB value at offset Offset in Stream, padded 197 // to allow patching. 198 static void writePatchableSLEB(raw_pwrite_stream &Stream, int32_t X, 199 uint64_t Offset) { 200 uint8_t Buffer[5]; 201 unsigned SizeLen = encodeSLEB128(X, Buffer, 5); 202 assert(SizeLen == 5); 203 Stream.pwrite((char *)Buffer, SizeLen, Offset); 204 } 205 206 // Write X as a plain integer value at offset Offset in Stream. 207 static void patchI32(raw_pwrite_stream &Stream, uint32_t X, uint64_t Offset) { 208 uint8_t Buffer[4]; 209 support::endian::write32le(Buffer, X); 210 Stream.pwrite((char *)Buffer, sizeof(Buffer), Offset); 211 } 212 213 class WasmObjectWriter : public MCObjectWriter { 214 support::endian::Writer W; 215 216 /// The target specific Wasm writer instance. 217 std::unique_ptr<MCWasmObjectTargetWriter> TargetObjectWriter; 218 219 // Relocations for fixing up references in the code section. 220 std::vector<WasmRelocationEntry> CodeRelocations; 221 uint32_t CodeSectionIndex; 222 223 // Relocations for fixing up references in the data section. 224 std::vector<WasmRelocationEntry> DataRelocations; 225 uint32_t DataSectionIndex; 226 227 // Index values to use for fixing up call_indirect type indices. 228 // Maps function symbols to the index of the type of the function 229 DenseMap<const MCSymbolWasm *, uint32_t> TypeIndices; 230 // Maps function symbols to the table element index space. Used 231 // for TABLE_INDEX relocation types (i.e. address taken functions). 232 DenseMap<const MCSymbolWasm *, uint32_t> TableIndices; 233 // Maps function/global symbols to the function/global/event/section index 234 // space. 235 DenseMap<const MCSymbolWasm *, uint32_t> WasmIndices; 236 DenseMap<const MCSymbolWasm *, uint32_t> GOTIndices; 237 // Maps data symbols to the Wasm segment and offset/size with the segment. 238 DenseMap<const MCSymbolWasm *, wasm::WasmDataReference> DataLocations; 239 240 // Stores output data (index, relocations, content offset) for custom 241 // section. 242 std::vector<WasmCustomSection> CustomSections; 243 std::unique_ptr<WasmCustomSection> ProducersSection; 244 std::unique_ptr<WasmCustomSection> TargetFeaturesSection; 245 // Relocations for fixing up references in the custom sections. 246 DenseMap<const MCSectionWasm *, std::vector<WasmRelocationEntry>> 247 CustomSectionsRelocations; 248 249 // Map from section to defining function symbol. 250 DenseMap<const MCSection *, const MCSymbol *> SectionFunctions; 251 252 DenseMap<WasmSignature, uint32_t, WasmSignatureDenseMapInfo> SignatureIndices; 253 SmallVector<WasmSignature, 4> Signatures; 254 SmallVector<WasmDataSegment, 4> DataSegments; 255 unsigned NumFunctionImports = 0; 256 unsigned NumGlobalImports = 0; 257 unsigned NumEventImports = 0; 258 uint32_t SectionCount = 0; 259 260 // TargetObjectWriter wrappers. 261 bool is64Bit() const { return TargetObjectWriter->is64Bit(); } 262 bool isEmscripten() const { return TargetObjectWriter->isEmscripten(); } 263 264 void startSection(SectionBookkeeping &Section, unsigned SectionId); 265 void startCustomSection(SectionBookkeeping &Section, StringRef Name); 266 void endSection(SectionBookkeeping &Section); 267 268 public: 269 WasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 270 raw_pwrite_stream &OS) 271 : W(OS, support::little), TargetObjectWriter(std::move(MOTW)) {} 272 273 private: 274 void reset() override { 275 CodeRelocations.clear(); 276 DataRelocations.clear(); 277 TypeIndices.clear(); 278 WasmIndices.clear(); 279 GOTIndices.clear(); 280 TableIndices.clear(); 281 DataLocations.clear(); 282 CustomSections.clear(); 283 ProducersSection.reset(); 284 TargetFeaturesSection.reset(); 285 CustomSectionsRelocations.clear(); 286 SignatureIndices.clear(); 287 Signatures.clear(); 288 DataSegments.clear(); 289 SectionFunctions.clear(); 290 NumFunctionImports = 0; 291 NumGlobalImports = 0; 292 MCObjectWriter::reset(); 293 } 294 295 void writeHeader(const MCAssembler &Asm); 296 297 void recordRelocation(MCAssembler &Asm, const MCAsmLayout &Layout, 298 const MCFragment *Fragment, const MCFixup &Fixup, 299 MCValue Target, uint64_t &FixedValue) override; 300 301 void executePostLayoutBinding(MCAssembler &Asm, 302 const MCAsmLayout &Layout) override; 303 304 uint64_t writeObject(MCAssembler &Asm, const MCAsmLayout &Layout) override; 305 306 void writeString(const StringRef Str) { 307 encodeULEB128(Str.size(), W.OS); 308 W.OS << Str; 309 } 310 311 void writeI32(int32_t val) { 312 char Buffer[4]; 313 support::endian::write32le(Buffer, val); 314 W.OS.write(Buffer, sizeof(Buffer)); 315 } 316 317 void writeI64(int64_t val) { 318 char Buffer[8]; 319 support::endian::write64le(Buffer, val); 320 W.OS.write(Buffer, sizeof(Buffer)); 321 } 322 323 void writeValueType(wasm::ValType Ty) { W.OS << static_cast<char>(Ty); } 324 325 void writeTypeSection(ArrayRef<WasmSignature> Signatures); 326 void writeImportSection(ArrayRef<wasm::WasmImport> Imports, uint32_t DataSize, 327 uint32_t NumElements); 328 void writeFunctionSection(ArrayRef<WasmFunction> Functions); 329 void writeExportSection(ArrayRef<wasm::WasmExport> Exports); 330 void writeElemSection(ArrayRef<uint32_t> TableElems); 331 void writeDataCountSection(); 332 void writeCodeSection(const MCAssembler &Asm, const MCAsmLayout &Layout, 333 ArrayRef<WasmFunction> Functions); 334 void writeDataSection(); 335 void writeEventSection(ArrayRef<wasm::WasmEventType> Events); 336 void writeGlobalSection(ArrayRef<wasm::WasmGlobal> Globals); 337 void writeRelocSection(uint32_t SectionIndex, StringRef Name, 338 std::vector<WasmRelocationEntry> &Relocations); 339 void writeLinkingMetaDataSection( 340 ArrayRef<wasm::WasmSymbolInfo> SymbolInfos, 341 ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs, 342 const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats); 343 void writeCustomSection(WasmCustomSection &CustomSection, 344 const MCAssembler &Asm, const MCAsmLayout &Layout); 345 void writeCustomRelocSections(); 346 void 347 updateCustomSectionRelocations(const SmallVector<WasmFunction, 4> &Functions, 348 const MCAsmLayout &Layout); 349 350 uint32_t getProvisionalValue(const WasmRelocationEntry &RelEntry); 351 void applyRelocations(ArrayRef<WasmRelocationEntry> Relocations, 352 uint64_t ContentsOffset); 353 354 uint32_t getRelocationIndexValue(const WasmRelocationEntry &RelEntry); 355 uint32_t getFunctionType(const MCSymbolWasm &Symbol); 356 uint32_t getEventType(const MCSymbolWasm &Symbol); 357 void registerFunctionType(const MCSymbolWasm &Symbol); 358 void registerEventType(const MCSymbolWasm &Symbol); 359 }; 360 361 } // end anonymous namespace 362 363 // Write out a section header and a patchable section size field. 364 void WasmObjectWriter::startSection(SectionBookkeeping &Section, 365 unsigned SectionId) { 366 LLVM_DEBUG(dbgs() << "startSection " << SectionId << "\n"); 367 W.OS << char(SectionId); 368 369 Section.SizeOffset = W.OS.tell(); 370 371 // The section size. We don't know the size yet, so reserve enough space 372 // for any 32-bit value; we'll patch it later. 373 encodeULEB128(0, W.OS, 5); 374 375 // The position where the section starts, for measuring its size. 376 Section.ContentsOffset = W.OS.tell(); 377 Section.PayloadOffset = W.OS.tell(); 378 Section.Index = SectionCount++; 379 } 380 381 void WasmObjectWriter::startCustomSection(SectionBookkeeping &Section, 382 StringRef Name) { 383 LLVM_DEBUG(dbgs() << "startCustomSection " << Name << "\n"); 384 startSection(Section, wasm::WASM_SEC_CUSTOM); 385 386 // The position where the section header ends, for measuring its size. 387 Section.PayloadOffset = W.OS.tell(); 388 389 // Custom sections in wasm also have a string identifier. 390 writeString(Name); 391 392 // The position where the custom section starts. 393 Section.ContentsOffset = W.OS.tell(); 394 } 395 396 // Now that the section is complete and we know how big it is, patch up the 397 // section size field at the start of the section. 398 void WasmObjectWriter::endSection(SectionBookkeeping &Section) { 399 uint64_t Size = W.OS.tell(); 400 // /dev/null doesn't support seek/tell and can report offset of 0. 401 // Simply skip this patching in that case. 402 if (!Size) 403 return; 404 405 Size -= Section.PayloadOffset; 406 if (uint32_t(Size) != Size) 407 report_fatal_error("section size does not fit in a uint32_t"); 408 409 LLVM_DEBUG(dbgs() << "endSection size=" << Size << "\n"); 410 411 // Write the final section size to the payload_len field, which follows 412 // the section id byte. 413 writePatchableLEB(static_cast<raw_pwrite_stream &>(W.OS), Size, 414 Section.SizeOffset); 415 } 416 417 // Emit the Wasm header. 418 void WasmObjectWriter::writeHeader(const MCAssembler &Asm) { 419 W.OS.write(wasm::WasmMagic, sizeof(wasm::WasmMagic)); 420 W.write<uint32_t>(wasm::WasmVersion); 421 } 422 423 void WasmObjectWriter::executePostLayoutBinding(MCAssembler &Asm, 424 const MCAsmLayout &Layout) { 425 // Build a map of sections to the function that defines them, for use 426 // in recordRelocation. 427 for (const MCSymbol &S : Asm.symbols()) { 428 const auto &WS = static_cast<const MCSymbolWasm &>(S); 429 if (WS.isDefined() && WS.isFunction() && !WS.isVariable()) { 430 const auto &Sec = static_cast<const MCSectionWasm &>(S.getSection()); 431 auto Pair = SectionFunctions.insert(std::make_pair(&Sec, &S)); 432 if (!Pair.second) 433 report_fatal_error("section already has a defining function: " + 434 Sec.getName()); 435 } 436 } 437 } 438 439 void WasmObjectWriter::recordRelocation(MCAssembler &Asm, 440 const MCAsmLayout &Layout, 441 const MCFragment *Fragment, 442 const MCFixup &Fixup, MCValue Target, 443 uint64_t &FixedValue) { 444 // The WebAssembly backend should never generate FKF_IsPCRel fixups 445 assert(!(Asm.getBackend().getFixupKindInfo(Fixup.getKind()).Flags & 446 MCFixupKindInfo::FKF_IsPCRel)); 447 448 const auto &FixupSection = cast<MCSectionWasm>(*Fragment->getParent()); 449 uint64_t C = Target.getConstant(); 450 uint64_t FixupOffset = Layout.getFragmentOffset(Fragment) + Fixup.getOffset(); 451 MCContext &Ctx = Asm.getContext(); 452 453 if (const MCSymbolRefExpr *RefB = Target.getSymB()) { 454 // To get here the A - B expression must have failed evaluateAsRelocatable. 455 // This means either A or B must be undefined and in WebAssembly we can't 456 // support either of those cases. 457 const auto &SymB = cast<MCSymbolWasm>(RefB->getSymbol()); 458 Ctx.reportError( 459 Fixup.getLoc(), 460 Twine("symbol '") + SymB.getName() + 461 "': unsupported subtraction expression used in relocation."); 462 return; 463 } 464 465 // We either rejected the fixup or folded B into C at this point. 466 const MCSymbolRefExpr *RefA = Target.getSymA(); 467 const auto *SymA = cast<MCSymbolWasm>(&RefA->getSymbol()); 468 469 // The .init_array isn't translated as data, so don't do relocations in it. 470 if (FixupSection.getName().startswith(".init_array")) { 471 SymA->setUsedInInitArray(); 472 return; 473 } 474 475 if (SymA->isVariable()) { 476 const MCExpr *Expr = SymA->getVariableValue(); 477 const auto *Inner = cast<MCSymbolRefExpr>(Expr); 478 if (Inner->getKind() == MCSymbolRefExpr::VK_WEAKREF) 479 llvm_unreachable("weakref used in reloc not yet implemented"); 480 } 481 482 // Put any constant offset in an addend. Offsets can be negative, and 483 // LLVM expects wrapping, in contrast to wasm's immediates which can't 484 // be negative and don't wrap. 485 FixedValue = 0; 486 487 unsigned Type = TargetObjectWriter->getRelocType(Target, Fixup); 488 489 // Absolute offset within a section or a function. 490 // Currently only supported for for metadata sections. 491 // See: test/MC/WebAssembly/blockaddress.ll 492 if (Type == wasm::R_WASM_FUNCTION_OFFSET_I32 || 493 Type == wasm::R_WASM_SECTION_OFFSET_I32) { 494 if (!FixupSection.getKind().isMetadata()) 495 report_fatal_error("relocations for function or section offsets are " 496 "only supported in metadata sections"); 497 498 const MCSymbol *SectionSymbol = nullptr; 499 const MCSection &SecA = SymA->getSection(); 500 if (SecA.getKind().isText()) 501 SectionSymbol = SectionFunctions.find(&SecA)->second; 502 else 503 SectionSymbol = SecA.getBeginSymbol(); 504 if (!SectionSymbol) 505 report_fatal_error("section symbol is required for relocation"); 506 507 C += Layout.getSymbolOffset(*SymA); 508 SymA = cast<MCSymbolWasm>(SectionSymbol); 509 } 510 511 // Relocation other than R_WASM_TYPE_INDEX_LEB are required to be 512 // against a named symbol. 513 if (Type != wasm::R_WASM_TYPE_INDEX_LEB) { 514 if (SymA->getName().empty()) 515 report_fatal_error("relocations against un-named temporaries are not yet " 516 "supported by wasm"); 517 518 SymA->setUsedInReloc(); 519 } 520 521 if (RefA->getKind() == MCSymbolRefExpr::VK_GOT) 522 SymA->setUsedInGOT(); 523 524 WasmRelocationEntry Rec(FixupOffset, SymA, C, Type, &FixupSection); 525 LLVM_DEBUG(dbgs() << "WasmReloc: " << Rec << "\n"); 526 527 if (FixupSection.isWasmData()) { 528 DataRelocations.push_back(Rec); 529 } else if (FixupSection.getKind().isText()) { 530 CodeRelocations.push_back(Rec); 531 } else if (FixupSection.getKind().isMetadata()) { 532 CustomSectionsRelocations[&FixupSection].push_back(Rec); 533 } else { 534 llvm_unreachable("unexpected section type"); 535 } 536 } 537 538 static const MCSymbolWasm *resolveSymbol(const MCSymbolWasm &Symbol) { 539 const MCSymbolWasm* Ret = &Symbol; 540 while (Ret->isVariable()) { 541 const MCExpr *Expr = Ret->getVariableValue(); 542 auto *Inner = cast<MCSymbolRefExpr>(Expr); 543 Ret = cast<MCSymbolWasm>(&Inner->getSymbol()); 544 } 545 return Ret; 546 } 547 548 // Compute a value to write into the code at the location covered 549 // by RelEntry. This value isn't used by the static linker; it just serves 550 // to make the object format more readable and more likely to be directly 551 // useable. 552 uint32_t 553 WasmObjectWriter::getProvisionalValue(const WasmRelocationEntry &RelEntry) { 554 if ((RelEntry.Type == wasm::R_WASM_GLOBAL_INDEX_LEB || 555 RelEntry.Type == wasm::R_WASM_GLOBAL_INDEX_I32) && 556 !RelEntry.Symbol->isGlobal()) { 557 assert(GOTIndices.count(RelEntry.Symbol) > 0 && "symbol not found in GOT index space"); 558 return GOTIndices[RelEntry.Symbol]; 559 } 560 561 switch (RelEntry.Type) { 562 case wasm::R_WASM_TABLE_INDEX_REL_SLEB: 563 case wasm::R_WASM_TABLE_INDEX_SLEB: 564 case wasm::R_WASM_TABLE_INDEX_I32: { 565 // Provisional value is table address of the resolved symbol itself 566 const MCSymbolWasm *Sym = resolveSymbol(*RelEntry.Symbol); 567 assert(Sym->isFunction()); 568 if (RelEntry.Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB) 569 return TableIndices[Sym] - InitialTableOffset; 570 else 571 return TableIndices[Sym]; 572 } 573 case wasm::R_WASM_TYPE_INDEX_LEB: 574 // Provisional value is same as the index 575 return getRelocationIndexValue(RelEntry); 576 case wasm::R_WASM_FUNCTION_INDEX_LEB: 577 case wasm::R_WASM_GLOBAL_INDEX_LEB: 578 case wasm::R_WASM_GLOBAL_INDEX_I32: 579 case wasm::R_WASM_EVENT_INDEX_LEB: 580 // Provisional value is function/global/event Wasm index 581 assert(WasmIndices.count(RelEntry.Symbol) > 0 && "symbol not found in wasm index space"); 582 return WasmIndices[RelEntry.Symbol]; 583 case wasm::R_WASM_FUNCTION_OFFSET_I32: 584 case wasm::R_WASM_SECTION_OFFSET_I32: { 585 const auto &Section = 586 static_cast<const MCSectionWasm &>(RelEntry.Symbol->getSection()); 587 return Section.getSectionOffset() + RelEntry.Addend; 588 } 589 case wasm::R_WASM_MEMORY_ADDR_LEB: 590 case wasm::R_WASM_MEMORY_ADDR_I32: 591 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB: 592 case wasm::R_WASM_MEMORY_ADDR_SLEB: { 593 // Provisional value is address of the global 594 const MCSymbolWasm *Sym = resolveSymbol(*RelEntry.Symbol); 595 // For undefined symbols, use zero 596 if (!Sym->isDefined()) 597 return 0; 598 const wasm::WasmDataReference &Ref = DataLocations[Sym]; 599 const WasmDataSegment &Segment = DataSegments[Ref.Segment]; 600 // Ignore overflow. LLVM allows address arithmetic to silently wrap. 601 return Segment.Offset + Ref.Offset + RelEntry.Addend; 602 } 603 default: 604 llvm_unreachable("invalid relocation type"); 605 } 606 } 607 608 static void addData(SmallVectorImpl<char> &DataBytes, 609 MCSectionWasm &DataSection) { 610 LLVM_DEBUG(errs() << "addData: " << DataSection.getName() << "\n"); 611 612 DataBytes.resize(alignTo(DataBytes.size(), DataSection.getAlignment())); 613 614 for (const MCFragment &Frag : DataSection) { 615 if (Frag.hasInstructions()) 616 report_fatal_error("only data supported in data sections"); 617 618 if (auto *Align = dyn_cast<MCAlignFragment>(&Frag)) { 619 if (Align->getValueSize() != 1) 620 report_fatal_error("only byte values supported for alignment"); 621 // If nops are requested, use zeros, as this is the data section. 622 uint8_t Value = Align->hasEmitNops() ? 0 : Align->getValue(); 623 uint64_t Size = 624 std::min<uint64_t>(alignTo(DataBytes.size(), Align->getAlignment()), 625 DataBytes.size() + Align->getMaxBytesToEmit()); 626 DataBytes.resize(Size, Value); 627 } else if (auto *Fill = dyn_cast<MCFillFragment>(&Frag)) { 628 int64_t NumValues; 629 if (!Fill->getNumValues().evaluateAsAbsolute(NumValues)) 630 llvm_unreachable("The fill should be an assembler constant"); 631 DataBytes.insert(DataBytes.end(), Fill->getValueSize() * NumValues, 632 Fill->getValue()); 633 } else if (auto *LEB = dyn_cast<MCLEBFragment>(&Frag)) { 634 const SmallVectorImpl<char> &Contents = LEB->getContents(); 635 DataBytes.insert(DataBytes.end(), Contents.begin(), Contents.end()); 636 } else { 637 const auto &DataFrag = cast<MCDataFragment>(Frag); 638 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 639 DataBytes.insert(DataBytes.end(), Contents.begin(), Contents.end()); 640 } 641 } 642 643 LLVM_DEBUG(dbgs() << "addData -> " << DataBytes.size() << "\n"); 644 } 645 646 uint32_t 647 WasmObjectWriter::getRelocationIndexValue(const WasmRelocationEntry &RelEntry) { 648 if (RelEntry.Type == wasm::R_WASM_TYPE_INDEX_LEB) { 649 if (!TypeIndices.count(RelEntry.Symbol)) 650 report_fatal_error("symbol not found in type index space: " + 651 RelEntry.Symbol->getName()); 652 return TypeIndices[RelEntry.Symbol]; 653 } 654 655 return RelEntry.Symbol->getIndex(); 656 } 657 658 // Apply the portions of the relocation records that we can handle ourselves 659 // directly. 660 void WasmObjectWriter::applyRelocations( 661 ArrayRef<WasmRelocationEntry> Relocations, uint64_t ContentsOffset) { 662 auto &Stream = static_cast<raw_pwrite_stream &>(W.OS); 663 for (const WasmRelocationEntry &RelEntry : Relocations) { 664 uint64_t Offset = ContentsOffset + 665 RelEntry.FixupSection->getSectionOffset() + 666 RelEntry.Offset; 667 668 LLVM_DEBUG(dbgs() << "applyRelocation: " << RelEntry << "\n"); 669 uint32_t Value = getProvisionalValue(RelEntry); 670 671 switch (RelEntry.Type) { 672 case wasm::R_WASM_FUNCTION_INDEX_LEB: 673 case wasm::R_WASM_TYPE_INDEX_LEB: 674 case wasm::R_WASM_GLOBAL_INDEX_LEB: 675 case wasm::R_WASM_MEMORY_ADDR_LEB: 676 case wasm::R_WASM_EVENT_INDEX_LEB: 677 writePatchableLEB(Stream, Value, Offset); 678 break; 679 case wasm::R_WASM_TABLE_INDEX_I32: 680 case wasm::R_WASM_MEMORY_ADDR_I32: 681 case wasm::R_WASM_FUNCTION_OFFSET_I32: 682 case wasm::R_WASM_SECTION_OFFSET_I32: 683 case wasm::R_WASM_GLOBAL_INDEX_I32: 684 patchI32(Stream, Value, Offset); 685 break; 686 case wasm::R_WASM_TABLE_INDEX_SLEB: 687 case wasm::R_WASM_TABLE_INDEX_REL_SLEB: 688 case wasm::R_WASM_MEMORY_ADDR_SLEB: 689 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB: 690 writePatchableSLEB(Stream, Value, Offset); 691 break; 692 default: 693 llvm_unreachable("invalid relocation type"); 694 } 695 } 696 } 697 698 void WasmObjectWriter::writeTypeSection(ArrayRef<WasmSignature> Signatures) { 699 if (Signatures.empty()) 700 return; 701 702 SectionBookkeeping Section; 703 startSection(Section, wasm::WASM_SEC_TYPE); 704 705 encodeULEB128(Signatures.size(), W.OS); 706 707 for (const WasmSignature &Sig : Signatures) { 708 W.OS << char(wasm::WASM_TYPE_FUNC); 709 encodeULEB128(Sig.Params.size(), W.OS); 710 for (wasm::ValType Ty : Sig.Params) 711 writeValueType(Ty); 712 encodeULEB128(Sig.Returns.size(), W.OS); 713 for (wasm::ValType Ty : Sig.Returns) 714 writeValueType(Ty); 715 } 716 717 endSection(Section); 718 } 719 720 void WasmObjectWriter::writeImportSection(ArrayRef<wasm::WasmImport> Imports, 721 uint32_t DataSize, 722 uint32_t NumElements) { 723 if (Imports.empty()) 724 return; 725 726 uint32_t NumPages = (DataSize + wasm::WasmPageSize - 1) / wasm::WasmPageSize; 727 728 SectionBookkeeping Section; 729 startSection(Section, wasm::WASM_SEC_IMPORT); 730 731 encodeULEB128(Imports.size(), W.OS); 732 for (const wasm::WasmImport &Import : Imports) { 733 writeString(Import.Module); 734 writeString(Import.Field); 735 W.OS << char(Import.Kind); 736 737 switch (Import.Kind) { 738 case wasm::WASM_EXTERNAL_FUNCTION: 739 encodeULEB128(Import.SigIndex, W.OS); 740 break; 741 case wasm::WASM_EXTERNAL_GLOBAL: 742 W.OS << char(Import.Global.Type); 743 W.OS << char(Import.Global.Mutable ? 1 : 0); 744 break; 745 case wasm::WASM_EXTERNAL_MEMORY: 746 encodeULEB128(0, W.OS); // flags 747 encodeULEB128(NumPages, W.OS); // initial 748 break; 749 case wasm::WASM_EXTERNAL_TABLE: 750 W.OS << char(Import.Table.ElemType); 751 encodeULEB128(0, W.OS); // flags 752 encodeULEB128(NumElements, W.OS); // initial 753 break; 754 case wasm::WASM_EXTERNAL_EVENT: 755 encodeULEB128(Import.Event.Attribute, W.OS); 756 encodeULEB128(Import.Event.SigIndex, W.OS); 757 break; 758 default: 759 llvm_unreachable("unsupported import kind"); 760 } 761 } 762 763 endSection(Section); 764 } 765 766 void WasmObjectWriter::writeFunctionSection(ArrayRef<WasmFunction> Functions) { 767 if (Functions.empty()) 768 return; 769 770 SectionBookkeeping Section; 771 startSection(Section, wasm::WASM_SEC_FUNCTION); 772 773 encodeULEB128(Functions.size(), W.OS); 774 for (const WasmFunction &Func : Functions) 775 encodeULEB128(Func.SigIndex, W.OS); 776 777 endSection(Section); 778 } 779 780 void WasmObjectWriter::writeEventSection(ArrayRef<wasm::WasmEventType> Events) { 781 if (Events.empty()) 782 return; 783 784 SectionBookkeeping Section; 785 startSection(Section, wasm::WASM_SEC_EVENT); 786 787 encodeULEB128(Events.size(), W.OS); 788 for (const wasm::WasmEventType &Event : Events) { 789 encodeULEB128(Event.Attribute, W.OS); 790 encodeULEB128(Event.SigIndex, W.OS); 791 } 792 793 endSection(Section); 794 } 795 796 void WasmObjectWriter::writeGlobalSection(ArrayRef<wasm::WasmGlobal> Globals) { 797 if (Globals.empty()) 798 return; 799 800 SectionBookkeeping Section; 801 startSection(Section, wasm::WASM_SEC_GLOBAL); 802 803 encodeULEB128(Globals.size(), W.OS); 804 for (const wasm::WasmGlobal &Global : Globals) { 805 encodeULEB128(Global.Type.Type, W.OS); 806 W.OS << char(Global.Type.Mutable); 807 W.OS << char(Global.InitExpr.Opcode); 808 switch (Global.Type.Type) { 809 case wasm::WASM_TYPE_I32: 810 encodeSLEB128(0, W.OS); 811 break; 812 case wasm::WASM_TYPE_I64: 813 encodeSLEB128(0, W.OS); 814 break; 815 case wasm::WASM_TYPE_F32: 816 writeI32(0); 817 break; 818 case wasm::WASM_TYPE_F64: 819 writeI64(0); 820 break; 821 default: 822 llvm_unreachable("unexpected type"); 823 } 824 W.OS << char(wasm::WASM_OPCODE_END); 825 } 826 827 endSection(Section); 828 } 829 830 void WasmObjectWriter::writeExportSection(ArrayRef<wasm::WasmExport> Exports) { 831 if (Exports.empty()) 832 return; 833 834 SectionBookkeeping Section; 835 startSection(Section, wasm::WASM_SEC_EXPORT); 836 837 encodeULEB128(Exports.size(), W.OS); 838 for (const wasm::WasmExport &Export : Exports) { 839 writeString(Export.Name); 840 W.OS << char(Export.Kind); 841 encodeULEB128(Export.Index, W.OS); 842 } 843 844 endSection(Section); 845 } 846 847 void WasmObjectWriter::writeElemSection(ArrayRef<uint32_t> TableElems) { 848 if (TableElems.empty()) 849 return; 850 851 SectionBookkeeping Section; 852 startSection(Section, wasm::WASM_SEC_ELEM); 853 854 encodeULEB128(1, W.OS); // number of "segments" 855 encodeULEB128(0, W.OS); // the table index 856 857 // init expr for starting offset 858 W.OS << char(wasm::WASM_OPCODE_I32_CONST); 859 encodeSLEB128(InitialTableOffset, W.OS); 860 W.OS << char(wasm::WASM_OPCODE_END); 861 862 encodeULEB128(TableElems.size(), W.OS); 863 for (uint32_t Elem : TableElems) 864 encodeULEB128(Elem, W.OS); 865 866 endSection(Section); 867 } 868 869 void WasmObjectWriter::writeDataCountSection() { 870 if (DataSegments.empty()) 871 return; 872 873 SectionBookkeeping Section; 874 startSection(Section, wasm::WASM_SEC_DATACOUNT); 875 encodeULEB128(DataSegments.size(), W.OS); 876 endSection(Section); 877 } 878 879 void WasmObjectWriter::writeCodeSection(const MCAssembler &Asm, 880 const MCAsmLayout &Layout, 881 ArrayRef<WasmFunction> Functions) { 882 if (Functions.empty()) 883 return; 884 885 SectionBookkeeping Section; 886 startSection(Section, wasm::WASM_SEC_CODE); 887 CodeSectionIndex = Section.Index; 888 889 encodeULEB128(Functions.size(), W.OS); 890 891 for (const WasmFunction &Func : Functions) { 892 auto &FuncSection = static_cast<MCSectionWasm &>(Func.Sym->getSection()); 893 894 int64_t Size = 0; 895 if (!Func.Sym->getSize()->evaluateAsAbsolute(Size, Layout)) 896 report_fatal_error(".size expression must be evaluatable"); 897 898 encodeULEB128(Size, W.OS); 899 FuncSection.setSectionOffset(W.OS.tell() - Section.ContentsOffset); 900 Asm.writeSectionData(W.OS, &FuncSection, Layout); 901 } 902 903 // Apply fixups. 904 applyRelocations(CodeRelocations, Section.ContentsOffset); 905 906 endSection(Section); 907 } 908 909 void WasmObjectWriter::writeDataSection() { 910 if (DataSegments.empty()) 911 return; 912 913 SectionBookkeeping Section; 914 startSection(Section, wasm::WASM_SEC_DATA); 915 DataSectionIndex = Section.Index; 916 917 encodeULEB128(DataSegments.size(), W.OS); // count 918 919 for (const WasmDataSegment &Segment : DataSegments) { 920 encodeULEB128(Segment.InitFlags, W.OS); // flags 921 if (Segment.InitFlags & wasm::WASM_SEGMENT_HAS_MEMINDEX) 922 encodeULEB128(0, W.OS); // memory index 923 if ((Segment.InitFlags & wasm::WASM_SEGMENT_IS_PASSIVE) == 0) { 924 W.OS << char(wasm::WASM_OPCODE_I32_CONST); 925 encodeSLEB128(Segment.Offset, W.OS); // offset 926 W.OS << char(wasm::WASM_OPCODE_END); 927 } 928 encodeULEB128(Segment.Data.size(), W.OS); // size 929 Segment.Section->setSectionOffset(W.OS.tell() - Section.ContentsOffset); 930 W.OS << Segment.Data; // data 931 } 932 933 // Apply fixups. 934 applyRelocations(DataRelocations, Section.ContentsOffset); 935 936 endSection(Section); 937 } 938 939 void WasmObjectWriter::writeRelocSection( 940 uint32_t SectionIndex, StringRef Name, 941 std::vector<WasmRelocationEntry> &Relocs) { 942 // See: https://github.com/WebAssembly/tool-conventions/blob/master/Linking.md 943 // for descriptions of the reloc sections. 944 945 if (Relocs.empty()) 946 return; 947 948 // First, ensure the relocations are sorted in offset order. In general they 949 // should already be sorted since `recordRelocation` is called in offset 950 // order, but for the code section we combine many MC sections into single 951 // wasm section, and this order is determined by the order of Asm.Symbols() 952 // not the sections order. 953 llvm::stable_sort( 954 Relocs, [](const WasmRelocationEntry &A, const WasmRelocationEntry &B) { 955 return (A.Offset + A.FixupSection->getSectionOffset()) < 956 (B.Offset + B.FixupSection->getSectionOffset()); 957 }); 958 959 SectionBookkeeping Section; 960 startCustomSection(Section, std::string("reloc.") + Name.str()); 961 962 encodeULEB128(SectionIndex, W.OS); 963 encodeULEB128(Relocs.size(), W.OS); 964 for (const WasmRelocationEntry &RelEntry : Relocs) { 965 uint64_t Offset = 966 RelEntry.Offset + RelEntry.FixupSection->getSectionOffset(); 967 uint32_t Index = getRelocationIndexValue(RelEntry); 968 969 W.OS << char(RelEntry.Type); 970 encodeULEB128(Offset, W.OS); 971 encodeULEB128(Index, W.OS); 972 if (RelEntry.hasAddend()) 973 encodeSLEB128(RelEntry.Addend, W.OS); 974 } 975 976 endSection(Section); 977 } 978 979 void WasmObjectWriter::writeCustomRelocSections() { 980 for (const auto &Sec : CustomSections) { 981 auto &Relocations = CustomSectionsRelocations[Sec.Section]; 982 writeRelocSection(Sec.OutputIndex, Sec.Name, Relocations); 983 } 984 } 985 986 void WasmObjectWriter::writeLinkingMetaDataSection( 987 ArrayRef<wasm::WasmSymbolInfo> SymbolInfos, 988 ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs, 989 const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats) { 990 SectionBookkeeping Section; 991 startCustomSection(Section, "linking"); 992 encodeULEB128(wasm::WasmMetadataVersion, W.OS); 993 994 SectionBookkeeping SubSection; 995 if (SymbolInfos.size() != 0) { 996 startSection(SubSection, wasm::WASM_SYMBOL_TABLE); 997 encodeULEB128(SymbolInfos.size(), W.OS); 998 for (const wasm::WasmSymbolInfo &Sym : SymbolInfos) { 999 encodeULEB128(Sym.Kind, W.OS); 1000 encodeULEB128(Sym.Flags, W.OS); 1001 switch (Sym.Kind) { 1002 case wasm::WASM_SYMBOL_TYPE_FUNCTION: 1003 case wasm::WASM_SYMBOL_TYPE_GLOBAL: 1004 case wasm::WASM_SYMBOL_TYPE_EVENT: 1005 encodeULEB128(Sym.ElementIndex, W.OS); 1006 if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0 || 1007 (Sym.Flags & wasm::WASM_SYMBOL_EXPLICIT_NAME) != 0) 1008 writeString(Sym.Name); 1009 break; 1010 case wasm::WASM_SYMBOL_TYPE_DATA: 1011 writeString(Sym.Name); 1012 if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0) { 1013 encodeULEB128(Sym.DataRef.Segment, W.OS); 1014 encodeULEB128(Sym.DataRef.Offset, W.OS); 1015 encodeULEB128(Sym.DataRef.Size, W.OS); 1016 } 1017 break; 1018 case wasm::WASM_SYMBOL_TYPE_SECTION: { 1019 const uint32_t SectionIndex = 1020 CustomSections[Sym.ElementIndex].OutputIndex; 1021 encodeULEB128(SectionIndex, W.OS); 1022 break; 1023 } 1024 default: 1025 llvm_unreachable("unexpected kind"); 1026 } 1027 } 1028 endSection(SubSection); 1029 } 1030 1031 if (DataSegments.size()) { 1032 startSection(SubSection, wasm::WASM_SEGMENT_INFO); 1033 encodeULEB128(DataSegments.size(), W.OS); 1034 for (const WasmDataSegment &Segment : DataSegments) { 1035 writeString(Segment.Name); 1036 encodeULEB128(Segment.Alignment, W.OS); 1037 encodeULEB128(Segment.LinkerFlags, W.OS); 1038 } 1039 endSection(SubSection); 1040 } 1041 1042 if (!InitFuncs.empty()) { 1043 startSection(SubSection, wasm::WASM_INIT_FUNCS); 1044 encodeULEB128(InitFuncs.size(), W.OS); 1045 for (auto &StartFunc : InitFuncs) { 1046 encodeULEB128(StartFunc.first, W.OS); // priority 1047 encodeULEB128(StartFunc.second, W.OS); // function index 1048 } 1049 endSection(SubSection); 1050 } 1051 1052 if (Comdats.size()) { 1053 startSection(SubSection, wasm::WASM_COMDAT_INFO); 1054 encodeULEB128(Comdats.size(), W.OS); 1055 for (const auto &C : Comdats) { 1056 writeString(C.first); 1057 encodeULEB128(0, W.OS); // flags for future use 1058 encodeULEB128(C.second.size(), W.OS); 1059 for (const WasmComdatEntry &Entry : C.second) { 1060 encodeULEB128(Entry.Kind, W.OS); 1061 encodeULEB128(Entry.Index, W.OS); 1062 } 1063 } 1064 endSection(SubSection); 1065 } 1066 1067 endSection(Section); 1068 } 1069 1070 void WasmObjectWriter::writeCustomSection(WasmCustomSection &CustomSection, 1071 const MCAssembler &Asm, 1072 const MCAsmLayout &Layout) { 1073 SectionBookkeeping Section; 1074 auto *Sec = CustomSection.Section; 1075 startCustomSection(Section, CustomSection.Name); 1076 1077 Sec->setSectionOffset(W.OS.tell() - Section.ContentsOffset); 1078 Asm.writeSectionData(W.OS, Sec, Layout); 1079 1080 CustomSection.OutputContentsOffset = Section.ContentsOffset; 1081 CustomSection.OutputIndex = Section.Index; 1082 1083 endSection(Section); 1084 1085 // Apply fixups. 1086 auto &Relocations = CustomSectionsRelocations[CustomSection.Section]; 1087 applyRelocations(Relocations, CustomSection.OutputContentsOffset); 1088 } 1089 1090 uint32_t WasmObjectWriter::getFunctionType(const MCSymbolWasm &Symbol) { 1091 assert(Symbol.isFunction()); 1092 assert(TypeIndices.count(&Symbol)); 1093 return TypeIndices[&Symbol]; 1094 } 1095 1096 uint32_t WasmObjectWriter::getEventType(const MCSymbolWasm &Symbol) { 1097 assert(Symbol.isEvent()); 1098 assert(TypeIndices.count(&Symbol)); 1099 return TypeIndices[&Symbol]; 1100 } 1101 1102 void WasmObjectWriter::registerFunctionType(const MCSymbolWasm &Symbol) { 1103 assert(Symbol.isFunction()); 1104 1105 WasmSignature S; 1106 const MCSymbolWasm *ResolvedSym = resolveSymbol(Symbol); 1107 if (auto *Sig = ResolvedSym->getSignature()) { 1108 S.Returns = Sig->Returns; 1109 S.Params = Sig->Params; 1110 } 1111 1112 auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size())); 1113 if (Pair.second) 1114 Signatures.push_back(S); 1115 TypeIndices[&Symbol] = Pair.first->second; 1116 1117 LLVM_DEBUG(dbgs() << "registerFunctionType: " << Symbol 1118 << " new:" << Pair.second << "\n"); 1119 LLVM_DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n"); 1120 } 1121 1122 void WasmObjectWriter::registerEventType(const MCSymbolWasm &Symbol) { 1123 assert(Symbol.isEvent()); 1124 1125 // TODO Currently we don't generate imported exceptions, but if we do, we 1126 // should have a way of infering types of imported exceptions. 1127 WasmSignature S; 1128 if (auto *Sig = Symbol.getSignature()) { 1129 S.Returns = Sig->Returns; 1130 S.Params = Sig->Params; 1131 } 1132 1133 auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size())); 1134 if (Pair.second) 1135 Signatures.push_back(S); 1136 TypeIndices[&Symbol] = Pair.first->second; 1137 1138 LLVM_DEBUG(dbgs() << "registerEventType: " << Symbol << " new:" << Pair.second 1139 << "\n"); 1140 LLVM_DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n"); 1141 } 1142 1143 static bool isInSymtab(const MCSymbolWasm &Sym) { 1144 if (Sym.isUsedInReloc() || Sym.isUsedInInitArray()) 1145 return true; 1146 1147 if (Sym.isComdat() && !Sym.isDefined()) 1148 return false; 1149 1150 if (Sym.isTemporary()) 1151 return false; 1152 1153 if (Sym.isSection()) 1154 return false; 1155 1156 return true; 1157 } 1158 1159 uint64_t WasmObjectWriter::writeObject(MCAssembler &Asm, 1160 const MCAsmLayout &Layout) { 1161 uint64_t StartOffset = W.OS.tell(); 1162 1163 LLVM_DEBUG(dbgs() << "WasmObjectWriter::writeObject\n"); 1164 1165 // Collect information from the available symbols. 1166 SmallVector<WasmFunction, 4> Functions; 1167 SmallVector<uint32_t, 4> TableElems; 1168 SmallVector<wasm::WasmImport, 4> Imports; 1169 SmallVector<wasm::WasmExport, 4> Exports; 1170 SmallVector<wasm::WasmEventType, 1> Events; 1171 SmallVector<wasm::WasmGlobal, 1> Globals; 1172 SmallVector<wasm::WasmSymbolInfo, 4> SymbolInfos; 1173 SmallVector<std::pair<uint16_t, uint32_t>, 2> InitFuncs; 1174 std::map<StringRef, std::vector<WasmComdatEntry>> Comdats; 1175 uint32_t DataSize = 0; 1176 1177 // For now, always emit the memory import, since loads and stores are not 1178 // valid without it. In the future, we could perhaps be more clever and omit 1179 // it if there are no loads or stores. 1180 wasm::WasmImport MemImport; 1181 MemImport.Module = "env"; 1182 MemImport.Field = "__linear_memory"; 1183 MemImport.Kind = wasm::WASM_EXTERNAL_MEMORY; 1184 Imports.push_back(MemImport); 1185 1186 // For now, always emit the table section, since indirect calls are not 1187 // valid without it. In the future, we could perhaps be more clever and omit 1188 // it if there are no indirect calls. 1189 wasm::WasmImport TableImport; 1190 TableImport.Module = "env"; 1191 TableImport.Field = "__indirect_function_table"; 1192 TableImport.Kind = wasm::WASM_EXTERNAL_TABLE; 1193 TableImport.Table.ElemType = wasm::WASM_TYPE_FUNCREF; 1194 Imports.push_back(TableImport); 1195 1196 // Populate SignatureIndices, and Imports and WasmIndices for undefined 1197 // symbols. This must be done before populating WasmIndices for defined 1198 // symbols. 1199 for (const MCSymbol &S : Asm.symbols()) { 1200 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1201 1202 // Register types for all functions, including those with private linkage 1203 // (because wasm always needs a type signature). 1204 if (WS.isFunction()) 1205 registerFunctionType(WS); 1206 1207 if (WS.isEvent()) 1208 registerEventType(WS); 1209 1210 if (WS.isTemporary()) 1211 continue; 1212 1213 // If the symbol is not defined in this translation unit, import it. 1214 if (!WS.isDefined() && !WS.isComdat()) { 1215 if (WS.isFunction()) { 1216 wasm::WasmImport Import; 1217 Import.Module = WS.getImportModule(); 1218 Import.Field = WS.getImportName(); 1219 Import.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1220 Import.SigIndex = getFunctionType(WS); 1221 Imports.push_back(Import); 1222 assert(WasmIndices.count(&WS) == 0); 1223 WasmIndices[&WS] = NumFunctionImports++; 1224 } else if (WS.isGlobal()) { 1225 if (WS.isWeak()) 1226 report_fatal_error("undefined global symbol cannot be weak"); 1227 1228 wasm::WasmImport Import; 1229 Import.Field = WS.getImportName(); 1230 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1231 Import.Module = WS.getImportModule(); 1232 Import.Global = WS.getGlobalType(); 1233 Imports.push_back(Import); 1234 assert(WasmIndices.count(&WS) == 0); 1235 WasmIndices[&WS] = NumGlobalImports++; 1236 } else if (WS.isEvent()) { 1237 if (WS.isWeak()) 1238 report_fatal_error("undefined event symbol cannot be weak"); 1239 1240 wasm::WasmImport Import; 1241 Import.Module = WS.getImportModule(); 1242 Import.Field = WS.getImportName(); 1243 Import.Kind = wasm::WASM_EXTERNAL_EVENT; 1244 Import.Event.Attribute = wasm::WASM_EVENT_ATTRIBUTE_EXCEPTION; 1245 Import.Event.SigIndex = getEventType(WS); 1246 Imports.push_back(Import); 1247 assert(WasmIndices.count(&WS) == 0); 1248 WasmIndices[&WS] = NumEventImports++; 1249 } 1250 } 1251 } 1252 1253 // Add imports for GOT globals 1254 for (const MCSymbol &S : Asm.symbols()) { 1255 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1256 if (WS.isUsedInGOT()) { 1257 wasm::WasmImport Import; 1258 if (WS.isFunction()) 1259 Import.Module = "GOT.func"; 1260 else 1261 Import.Module = "GOT.mem"; 1262 Import.Field = WS.getName(); 1263 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1264 Import.Global = {wasm::WASM_TYPE_I32, true}; 1265 Imports.push_back(Import); 1266 assert(GOTIndices.count(&WS) == 0); 1267 GOTIndices[&WS] = NumGlobalImports++; 1268 } 1269 } 1270 1271 // Populate DataSegments and CustomSections, which must be done before 1272 // populating DataLocations. 1273 for (MCSection &Sec : Asm) { 1274 auto &Section = static_cast<MCSectionWasm &>(Sec); 1275 StringRef SectionName = Section.getName(); 1276 1277 // .init_array sections are handled specially elsewhere. 1278 if (SectionName.startswith(".init_array")) 1279 continue; 1280 1281 // Code is handled separately 1282 if (Section.getKind().isText()) 1283 continue; 1284 1285 if (Section.isWasmData()) { 1286 uint32_t SegmentIndex = DataSegments.size(); 1287 DataSize = alignTo(DataSize, Section.getAlignment()); 1288 DataSegments.emplace_back(); 1289 WasmDataSegment &Segment = DataSegments.back(); 1290 Segment.Name = SectionName; 1291 Segment.InitFlags = 1292 Section.getPassive() ? (uint32_t)wasm::WASM_SEGMENT_IS_PASSIVE : 0; 1293 Segment.Offset = DataSize; 1294 Segment.Section = &Section; 1295 addData(Segment.Data, Section); 1296 Segment.Alignment = Log2_32(Section.getAlignment()); 1297 Segment.LinkerFlags = 0; 1298 DataSize += Segment.Data.size(); 1299 Section.setSegmentIndex(SegmentIndex); 1300 1301 if (const MCSymbolWasm *C = Section.getGroup()) { 1302 Comdats[C->getName()].emplace_back( 1303 WasmComdatEntry{wasm::WASM_COMDAT_DATA, SegmentIndex}); 1304 } 1305 } else { 1306 // Create custom sections 1307 assert(Sec.getKind().isMetadata()); 1308 1309 StringRef Name = SectionName; 1310 1311 // For user-defined custom sections, strip the prefix 1312 if (Name.startswith(".custom_section.")) 1313 Name = Name.substr(strlen(".custom_section.")); 1314 1315 MCSymbol *Begin = Sec.getBeginSymbol(); 1316 if (Begin) { 1317 WasmIndices[cast<MCSymbolWasm>(Begin)] = CustomSections.size(); 1318 if (SectionName != Begin->getName()) 1319 report_fatal_error("section name and begin symbol should match: " + 1320 Twine(SectionName)); 1321 } 1322 1323 // Separate out the producers and target features sections 1324 if (Name == "producers") { 1325 ProducersSection = std::make_unique<WasmCustomSection>(Name, &Section); 1326 continue; 1327 } 1328 if (Name == "target_features") { 1329 TargetFeaturesSection = 1330 std::make_unique<WasmCustomSection>(Name, &Section); 1331 continue; 1332 } 1333 1334 CustomSections.emplace_back(Name, &Section); 1335 } 1336 } 1337 1338 // Populate WasmIndices and DataLocations for defined symbols. 1339 for (const MCSymbol &S : Asm.symbols()) { 1340 // Ignore unnamed temporary symbols, which aren't ever exported, imported, 1341 // or used in relocations. 1342 if (S.isTemporary() && S.getName().empty()) 1343 continue; 1344 1345 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1346 LLVM_DEBUG( 1347 dbgs() << "MCSymbol: " << toString(WS.getType()) << " '" << S << "'" 1348 << " isDefined=" << S.isDefined() << " isExternal=" 1349 << S.isExternal() << " isTemporary=" << S.isTemporary() 1350 << " isWeak=" << WS.isWeak() << " isHidden=" << WS.isHidden() 1351 << " isVariable=" << WS.isVariable() << "\n"); 1352 1353 if (WS.isVariable()) 1354 continue; 1355 if (WS.isComdat() && !WS.isDefined()) 1356 continue; 1357 1358 if (WS.isFunction()) { 1359 unsigned Index; 1360 if (WS.isDefined()) { 1361 if (WS.getOffset() != 0) 1362 report_fatal_error( 1363 "function sections must contain one function each"); 1364 1365 if (WS.getSize() == nullptr) 1366 report_fatal_error( 1367 "function symbols must have a size set with .size"); 1368 1369 // A definition. Write out the function body. 1370 Index = NumFunctionImports + Functions.size(); 1371 WasmFunction Func; 1372 Func.SigIndex = getFunctionType(WS); 1373 Func.Sym = &WS; 1374 WasmIndices[&WS] = Index; 1375 Functions.push_back(Func); 1376 1377 auto &Section = static_cast<MCSectionWasm &>(WS.getSection()); 1378 if (const MCSymbolWasm *C = Section.getGroup()) { 1379 Comdats[C->getName()].emplace_back( 1380 WasmComdatEntry{wasm::WASM_COMDAT_FUNCTION, Index}); 1381 } 1382 1383 if (WS.hasExportName()) { 1384 wasm::WasmExport Export; 1385 Export.Name = WS.getExportName(); 1386 Export.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1387 Export.Index = Index; 1388 Exports.push_back(Export); 1389 } 1390 } else { 1391 // An import; the index was assigned above. 1392 Index = WasmIndices.find(&WS)->second; 1393 } 1394 1395 LLVM_DEBUG(dbgs() << " -> function index: " << Index << "\n"); 1396 1397 } else if (WS.isData()) { 1398 if (!isInSymtab(WS)) 1399 continue; 1400 1401 if (!WS.isDefined()) { 1402 LLVM_DEBUG(dbgs() << " -> segment index: -1" 1403 << "\n"); 1404 continue; 1405 } 1406 1407 if (!WS.getSize()) 1408 report_fatal_error("data symbols must have a size set with .size: " + 1409 WS.getName()); 1410 1411 int64_t Size = 0; 1412 if (!WS.getSize()->evaluateAsAbsolute(Size, Layout)) 1413 report_fatal_error(".size expression must be evaluatable"); 1414 1415 auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection()); 1416 if (!DataSection.isWasmData()) 1417 report_fatal_error("data symbols must live in a data section: " + 1418 WS.getName()); 1419 1420 // For each data symbol, export it in the symtab as a reference to the 1421 // corresponding Wasm data segment. 1422 wasm::WasmDataReference Ref = wasm::WasmDataReference{ 1423 DataSection.getSegmentIndex(), 1424 static_cast<uint32_t>(Layout.getSymbolOffset(WS)), 1425 static_cast<uint32_t>(Size)}; 1426 DataLocations[&WS] = Ref; 1427 LLVM_DEBUG(dbgs() << " -> segment index: " << Ref.Segment << "\n"); 1428 1429 } else if (WS.isGlobal()) { 1430 // A "true" Wasm global (currently just __stack_pointer) 1431 if (WS.isDefined()) { 1432 assert(WasmIndices.count(&WS) == 0); 1433 wasm::WasmGlobal Global; 1434 Global.Type = WS.getGlobalType(); 1435 Global.Index = NumGlobalImports + Globals.size(); 1436 switch (Global.Type.Type) { 1437 case wasm::WASM_TYPE_I32: 1438 Global.InitExpr.Opcode = wasm::WASM_OPCODE_I32_CONST; 1439 break; 1440 case wasm::WASM_TYPE_I64: 1441 Global.InitExpr.Opcode = wasm::WASM_OPCODE_I64_CONST; 1442 break; 1443 case wasm::WASM_TYPE_F32: 1444 Global.InitExpr.Opcode = wasm::WASM_OPCODE_F32_CONST; 1445 break; 1446 case wasm::WASM_TYPE_F64: 1447 Global.InitExpr.Opcode = wasm::WASM_OPCODE_F64_CONST; 1448 break; 1449 default: 1450 llvm_unreachable("unexpected type"); 1451 } 1452 WasmIndices[&WS] = Global.Index; 1453 Globals.push_back(Global); 1454 } else { 1455 // An import; the index was assigned above 1456 LLVM_DEBUG(dbgs() << " -> global index: " 1457 << WasmIndices.find(&WS)->second << "\n"); 1458 } 1459 } else if (WS.isEvent()) { 1460 // C++ exception symbol (__cpp_exception) 1461 unsigned Index; 1462 if (WS.isDefined()) { 1463 assert(WasmIndices.count(&WS) == 0); 1464 Index = NumEventImports + Events.size(); 1465 wasm::WasmEventType Event; 1466 Event.SigIndex = getEventType(WS); 1467 Event.Attribute = wasm::WASM_EVENT_ATTRIBUTE_EXCEPTION; 1468 WasmIndices[&WS] = Index; 1469 Events.push_back(Event); 1470 } else { 1471 // An import; the index was assigned above. 1472 assert(WasmIndices.count(&WS) > 0); 1473 } 1474 LLVM_DEBUG(dbgs() << " -> event index: " << WasmIndices.find(&WS)->second 1475 << "\n"); 1476 1477 } else { 1478 assert(WS.isSection()); 1479 } 1480 } 1481 1482 // Populate WasmIndices and DataLocations for aliased symbols. We need to 1483 // process these in a separate pass because we need to have processed the 1484 // target of the alias before the alias itself and the symbols are not 1485 // necessarily ordered in this way. 1486 for (const MCSymbol &S : Asm.symbols()) { 1487 if (!S.isVariable()) 1488 continue; 1489 1490 assert(S.isDefined()); 1491 1492 // Find the target symbol of this weak alias and export that index 1493 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1494 const MCSymbolWasm *ResolvedSym = resolveSymbol(WS); 1495 LLVM_DEBUG(dbgs() << WS.getName() << ": weak alias of '" << *ResolvedSym 1496 << "'\n"); 1497 1498 if (ResolvedSym->isFunction()) { 1499 assert(WasmIndices.count(ResolvedSym) > 0); 1500 uint32_t WasmIndex = WasmIndices.find(ResolvedSym)->second; 1501 assert(WasmIndices.count(&WS) == 0); 1502 WasmIndices[&WS] = WasmIndex; 1503 LLVM_DEBUG(dbgs() << " -> index:" << WasmIndex << "\n"); 1504 } else if (ResolvedSym->isData()) { 1505 assert(DataLocations.count(ResolvedSym) > 0); 1506 const wasm::WasmDataReference &Ref = 1507 DataLocations.find(ResolvedSym)->second; 1508 DataLocations[&WS] = Ref; 1509 LLVM_DEBUG(dbgs() << " -> index:" << Ref.Segment << "\n"); 1510 } else { 1511 report_fatal_error("don't yet support global/event aliases"); 1512 } 1513 } 1514 1515 // Finally, populate the symbol table itself, in its "natural" order. 1516 for (const MCSymbol &S : Asm.symbols()) { 1517 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1518 if (!isInSymtab(WS)) { 1519 WS.setIndex(InvalidIndex); 1520 continue; 1521 } 1522 LLVM_DEBUG(dbgs() << "adding to symtab: " << WS << "\n"); 1523 1524 uint32_t Flags = 0; 1525 if (WS.isWeak()) 1526 Flags |= wasm::WASM_SYMBOL_BINDING_WEAK; 1527 if (WS.isHidden()) 1528 Flags |= wasm::WASM_SYMBOL_VISIBILITY_HIDDEN; 1529 if (!WS.isExternal() && WS.isDefined()) 1530 Flags |= wasm::WASM_SYMBOL_BINDING_LOCAL; 1531 if (WS.isUndefined()) 1532 Flags |= wasm::WASM_SYMBOL_UNDEFINED; 1533 if (WS.isNoStrip()) { 1534 Flags |= wasm::WASM_SYMBOL_NO_STRIP; 1535 if (isEmscripten()) { 1536 Flags |= wasm::WASM_SYMBOL_EXPORTED; 1537 } 1538 } 1539 if (WS.hasImportName()) 1540 Flags |= wasm::WASM_SYMBOL_EXPLICIT_NAME; 1541 if (WS.hasExportName()) 1542 Flags |= wasm::WASM_SYMBOL_EXPORTED; 1543 1544 wasm::WasmSymbolInfo Info; 1545 Info.Name = WS.getName(); 1546 Info.Kind = WS.getType(); 1547 Info.Flags = Flags; 1548 if (!WS.isData()) { 1549 assert(WasmIndices.count(&WS) > 0); 1550 Info.ElementIndex = WasmIndices.find(&WS)->second; 1551 } else if (WS.isDefined()) { 1552 assert(DataLocations.count(&WS) > 0); 1553 Info.DataRef = DataLocations.find(&WS)->second; 1554 } 1555 WS.setIndex(SymbolInfos.size()); 1556 SymbolInfos.emplace_back(Info); 1557 } 1558 1559 { 1560 auto HandleReloc = [&](const WasmRelocationEntry &Rel) { 1561 // Functions referenced by a relocation need to put in the table. This is 1562 // purely to make the object file's provisional values readable, and is 1563 // ignored by the linker, which re-calculates the relocations itself. 1564 if (Rel.Type != wasm::R_WASM_TABLE_INDEX_I32 && 1565 Rel.Type != wasm::R_WASM_TABLE_INDEX_SLEB && 1566 Rel.Type != wasm::R_WASM_TABLE_INDEX_REL_SLEB) 1567 return; 1568 assert(Rel.Symbol->isFunction()); 1569 const MCSymbolWasm &WS = *resolveSymbol(*Rel.Symbol); 1570 uint32_t FunctionIndex = WasmIndices.find(&WS)->second; 1571 uint32_t TableIndex = TableElems.size() + InitialTableOffset; 1572 if (TableIndices.try_emplace(&WS, TableIndex).second) { 1573 LLVM_DEBUG(dbgs() << " -> adding " << WS.getName() 1574 << " to table: " << TableIndex << "\n"); 1575 TableElems.push_back(FunctionIndex); 1576 registerFunctionType(WS); 1577 } 1578 }; 1579 1580 for (const WasmRelocationEntry &RelEntry : CodeRelocations) 1581 HandleReloc(RelEntry); 1582 for (const WasmRelocationEntry &RelEntry : DataRelocations) 1583 HandleReloc(RelEntry); 1584 } 1585 1586 // Translate .init_array section contents into start functions. 1587 for (const MCSection &S : Asm) { 1588 const auto &WS = static_cast<const MCSectionWasm &>(S); 1589 if (WS.getName().startswith(".fini_array")) 1590 report_fatal_error(".fini_array sections are unsupported"); 1591 if (!WS.getName().startswith(".init_array")) 1592 continue; 1593 if (WS.getFragmentList().empty()) 1594 continue; 1595 1596 // init_array is expected to contain a single non-empty data fragment 1597 if (WS.getFragmentList().size() != 3) 1598 report_fatal_error("only one .init_array section fragment supported"); 1599 1600 auto IT = WS.begin(); 1601 const MCFragment &EmptyFrag = *IT; 1602 if (EmptyFrag.getKind() != MCFragment::FT_Data) 1603 report_fatal_error(".init_array section should be aligned"); 1604 1605 IT = std::next(IT); 1606 const MCFragment &AlignFrag = *IT; 1607 if (AlignFrag.getKind() != MCFragment::FT_Align) 1608 report_fatal_error(".init_array section should be aligned"); 1609 if (cast<MCAlignFragment>(AlignFrag).getAlignment() != (is64Bit() ? 8 : 4)) 1610 report_fatal_error(".init_array section should be aligned for pointers"); 1611 1612 const MCFragment &Frag = *std::next(IT); 1613 if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data) 1614 report_fatal_error("only data supported in .init_array section"); 1615 1616 uint16_t Priority = UINT16_MAX; 1617 unsigned PrefixLength = strlen(".init_array"); 1618 if (WS.getName().size() > PrefixLength) { 1619 if (WS.getName()[PrefixLength] != '.') 1620 report_fatal_error( 1621 ".init_array section priority should start with '.'"); 1622 if (WS.getName().substr(PrefixLength + 1).getAsInteger(10, Priority)) 1623 report_fatal_error("invalid .init_array section priority"); 1624 } 1625 const auto &DataFrag = cast<MCDataFragment>(Frag); 1626 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 1627 for (const uint8_t * 1628 P = (const uint8_t *)Contents.data(), 1629 *End = (const uint8_t *)Contents.data() + Contents.size(); 1630 P != End; ++P) { 1631 if (*P != 0) 1632 report_fatal_error("non-symbolic data in .init_array section"); 1633 } 1634 for (const MCFixup &Fixup : DataFrag.getFixups()) { 1635 assert(Fixup.getKind() == 1636 MCFixup::getKindForSize(is64Bit() ? 8 : 4, false)); 1637 const MCExpr *Expr = Fixup.getValue(); 1638 auto *SymRef = dyn_cast<MCSymbolRefExpr>(Expr); 1639 if (!SymRef) 1640 report_fatal_error("fixups in .init_array should be symbol references"); 1641 const auto &TargetSym = cast<const MCSymbolWasm>(SymRef->getSymbol()); 1642 if (TargetSym.getIndex() == InvalidIndex) 1643 report_fatal_error("symbols in .init_array should exist in symtab"); 1644 if (!TargetSym.isFunction()) 1645 report_fatal_error("symbols in .init_array should be for functions"); 1646 InitFuncs.push_back( 1647 std::make_pair(Priority, TargetSym.getIndex())); 1648 } 1649 } 1650 1651 // Write out the Wasm header. 1652 writeHeader(Asm); 1653 1654 writeTypeSection(Signatures); 1655 writeImportSection(Imports, DataSize, TableElems.size()); 1656 writeFunctionSection(Functions); 1657 // Skip the "table" section; we import the table instead. 1658 // Skip the "memory" section; we import the memory instead. 1659 writeEventSection(Events); 1660 writeGlobalSection(Globals); 1661 writeExportSection(Exports); 1662 writeElemSection(TableElems); 1663 writeDataCountSection(); 1664 writeCodeSection(Asm, Layout, Functions); 1665 writeDataSection(); 1666 for (auto &CustomSection : CustomSections) 1667 writeCustomSection(CustomSection, Asm, Layout); 1668 writeLinkingMetaDataSection(SymbolInfos, InitFuncs, Comdats); 1669 writeRelocSection(CodeSectionIndex, "CODE", CodeRelocations); 1670 writeRelocSection(DataSectionIndex, "DATA", DataRelocations); 1671 writeCustomRelocSections(); 1672 if (ProducersSection) 1673 writeCustomSection(*ProducersSection, Asm, Layout); 1674 if (TargetFeaturesSection) 1675 writeCustomSection(*TargetFeaturesSection, Asm, Layout); 1676 1677 // TODO: Translate the .comment section to the output. 1678 return W.OS.tell() - StartOffset; 1679 } 1680 1681 std::unique_ptr<MCObjectWriter> 1682 llvm::createWasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 1683 raw_pwrite_stream &OS) { 1684 return std::make_unique<WasmObjectWriter>(std::move(MOTW), OS); 1685 } 1686