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/ErrorHandling.h" 31 #include "llvm/Support/LEB128.h" 32 #include "llvm/Support/StringSaver.h" 33 #include <vector> 34 35 using namespace llvm; 36 37 #define DEBUG_TYPE "mc" 38 39 namespace { 40 41 // Went we ceate the indirect function table we start at 1, so that there is 42 // and emtpy slot at 0 and therefore calling a null function pointer will trap. 43 static const uint32_t InitialTableOffset = 1; 44 45 // For patching purposes, we need to remember where each section starts, both 46 // for patching up the section size field, and for patching up references to 47 // locations within the section. 48 struct SectionBookkeeping { 49 // Where the size of the section is written. 50 uint64_t SizeOffset; 51 // Where the section header ends (without custom section name). 52 uint64_t PayloadOffset; 53 // Where the contents of the section starts. 54 uint64_t ContentsOffset; 55 uint32_t Index; 56 }; 57 58 // The signature of a wasm function or event, in a struct capable of being used 59 // as a DenseMap key. 60 // TODO: Consider using wasm::WasmSignature directly instead. 61 struct WasmSignature { 62 // Support empty and tombstone instances, needed by DenseMap. 63 enum { Plain, Empty, Tombstone } State = Plain; 64 65 // The return types of the function. 66 SmallVector<wasm::ValType, 1> Returns; 67 68 // The parameter types of the function. 69 SmallVector<wasm::ValType, 4> Params; 70 71 bool operator==(const WasmSignature &Other) const { 72 return State == Other.State && Returns == Other.Returns && 73 Params == Other.Params; 74 } 75 }; 76 77 // Traits for using WasmSignature in a DenseMap. 78 struct WasmSignatureDenseMapInfo { 79 static WasmSignature getEmptyKey() { 80 WasmSignature Sig; 81 Sig.State = WasmSignature::Empty; 82 return Sig; 83 } 84 static WasmSignature getTombstoneKey() { 85 WasmSignature Sig; 86 Sig.State = WasmSignature::Tombstone; 87 return Sig; 88 } 89 static unsigned getHashValue(const WasmSignature &Sig) { 90 uintptr_t Value = Sig.State; 91 for (wasm::ValType Ret : Sig.Returns) 92 Value += DenseMapInfo<uint32_t>::getHashValue(uint32_t(Ret)); 93 for (wasm::ValType Param : Sig.Params) 94 Value += DenseMapInfo<uint32_t>::getHashValue(uint32_t(Param)); 95 return Value; 96 } 97 static bool isEqual(const WasmSignature &LHS, const WasmSignature &RHS) { 98 return LHS == RHS; 99 } 100 }; 101 102 // A wasm data segment. A wasm binary contains only a single data section 103 // but that can contain many segments, each with their own virtual location 104 // in memory. Each MCSection data created by llvm is modeled as its own 105 // wasm data segment. 106 struct WasmDataSegment { 107 MCSectionWasm *Section; 108 StringRef Name; 109 uint32_t InitFlags; 110 uint32_t Offset; 111 uint32_t Alignment; 112 uint32_t LinkerFlags; 113 SmallVector<char, 4> Data; 114 }; 115 116 // A wasm function to be written into the function section. 117 struct WasmFunction { 118 uint32_t SigIndex; 119 const MCSymbolWasm *Sym; 120 }; 121 122 // A wasm global to be written into the global section. 123 struct WasmGlobal { 124 wasm::WasmGlobalType Type; 125 uint64_t InitialValue; 126 }; 127 128 // Information about a single item which is part of a COMDAT. For each data 129 // segment or function which is in the COMDAT, there is a corresponding 130 // WasmComdatEntry. 131 struct WasmComdatEntry { 132 unsigned Kind; 133 uint32_t Index; 134 }; 135 136 // Information about a single relocation. 137 struct WasmRelocationEntry { 138 uint64_t Offset; // Where is the relocation. 139 const MCSymbolWasm *Symbol; // The symbol to relocate with. 140 int64_t Addend; // A value to add to the symbol. 141 unsigned Type; // The type of the relocation. 142 const MCSectionWasm *FixupSection; // The section the relocation is targeting. 143 144 WasmRelocationEntry(uint64_t Offset, const MCSymbolWasm *Symbol, 145 int64_t Addend, unsigned Type, 146 const MCSectionWasm *FixupSection) 147 : Offset(Offset), Symbol(Symbol), Addend(Addend), Type(Type), 148 FixupSection(FixupSection) {} 149 150 bool hasAddend() const { return wasm::relocTypeHasAddend(Type); } 151 152 void print(raw_ostream &Out) const { 153 Out << wasm::relocTypetoString(Type) << " Off=" << Offset 154 << ", Sym=" << *Symbol << ", Addend=" << Addend 155 << ", FixupSection=" << FixupSection->getSectionName(); 156 } 157 158 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 159 LLVM_DUMP_METHOD void dump() const { print(dbgs()); } 160 #endif 161 }; 162 163 static const uint32_t InvalidIndex = -1; 164 165 struct WasmCustomSection { 166 167 StringRef Name; 168 MCSectionWasm *Section; 169 170 uint32_t OutputContentsOffset; 171 uint32_t OutputIndex; 172 173 WasmCustomSection(StringRef Name, MCSectionWasm *Section) 174 : Name(Name), Section(Section), OutputContentsOffset(0), 175 OutputIndex(InvalidIndex) {} 176 }; 177 178 #if !defined(NDEBUG) 179 raw_ostream &operator<<(raw_ostream &OS, const WasmRelocationEntry &Rel) { 180 Rel.print(OS); 181 return OS; 182 } 183 #endif 184 185 // Write X as an (unsigned) LEB value at offset Offset in Stream, padded 186 // to allow patching. 187 static void writePatchableLEB(raw_pwrite_stream &Stream, uint32_t X, 188 uint64_t Offset) { 189 uint8_t Buffer[5]; 190 unsigned SizeLen = encodeULEB128(X, Buffer, 5); 191 assert(SizeLen == 5); 192 Stream.pwrite((char *)Buffer, SizeLen, Offset); 193 } 194 195 // Write X as an signed LEB value at offset Offset in Stream, padded 196 // to allow patching. 197 static void writePatchableSLEB(raw_pwrite_stream &Stream, int32_t X, 198 uint64_t Offset) { 199 uint8_t Buffer[5]; 200 unsigned SizeLen = encodeSLEB128(X, Buffer, 5); 201 assert(SizeLen == 5); 202 Stream.pwrite((char *)Buffer, SizeLen, Offset); 203 } 204 205 // Write X as a plain integer value at offset Offset in Stream. 206 static void writeI32(raw_pwrite_stream &Stream, uint32_t X, uint64_t Offset) { 207 uint8_t Buffer[4]; 208 support::endian::write32le(Buffer, X); 209 Stream.pwrite((char *)Buffer, sizeof(Buffer), Offset); 210 } 211 212 class WasmObjectWriter : public MCObjectWriter { 213 support::endian::Writer W; 214 215 /// The target specific Wasm writer instance. 216 std::unique_ptr<MCWasmObjectTargetWriter> TargetObjectWriter; 217 218 // Relocations for fixing up references in the code section. 219 std::vector<WasmRelocationEntry> CodeRelocations; 220 uint32_t CodeSectionIndex; 221 222 // Relocations for fixing up references in the data section. 223 std::vector<WasmRelocationEntry> DataRelocations; 224 uint32_t DataSectionIndex; 225 226 // Index values to use for fixing up call_indirect type indices. 227 // Maps function symbols to the index of the type of the function 228 DenseMap<const MCSymbolWasm *, uint32_t> TypeIndices; 229 // Maps function symbols to the table element index space. Used 230 // for TABLE_INDEX relocation types (i.e. address taken functions). 231 DenseMap<const MCSymbolWasm *, uint32_t> TableIndices; 232 // Maps function/global symbols to the function/global/event/section index 233 // space. 234 DenseMap<const MCSymbolWasm *, uint32_t> WasmIndices; 235 DenseMap<const MCSymbolWasm *, uint32_t> GOTIndices; 236 // Maps data symbols to the Wasm segment and offset/size with the segment. 237 DenseMap<const MCSymbolWasm *, wasm::WasmDataReference> DataLocations; 238 239 // Stores output data (index, relocations, content offset) for custom 240 // section. 241 std::vector<WasmCustomSection> CustomSections; 242 std::unique_ptr<WasmCustomSection> ProducersSection; 243 std::unique_ptr<WasmCustomSection> TargetFeaturesSection; 244 // Relocations for fixing up references in the custom sections. 245 DenseMap<const MCSectionWasm *, std::vector<WasmRelocationEntry>> 246 CustomSectionsRelocations; 247 248 // Map from section to defining function symbol. 249 DenseMap<const MCSection *, const MCSymbol *> SectionFunctions; 250 251 DenseMap<WasmSignature, uint32_t, WasmSignatureDenseMapInfo> SignatureIndices; 252 SmallVector<WasmSignature, 4> Signatures; 253 SmallVector<WasmDataSegment, 4> DataSegments; 254 unsigned NumFunctionImports = 0; 255 unsigned NumGlobalImports = 0; 256 unsigned NumEventImports = 0; 257 uint32_t SectionCount = 0; 258 259 // TargetObjectWriter wrappers. 260 bool is64Bit() const { return TargetObjectWriter->is64Bit(); } 261 262 void startSection(SectionBookkeeping &Section, unsigned SectionId); 263 void startCustomSection(SectionBookkeeping &Section, StringRef Name); 264 void endSection(SectionBookkeeping &Section); 265 266 public: 267 WasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 268 raw_pwrite_stream &OS) 269 : W(OS, support::little), TargetObjectWriter(std::move(MOTW)) {} 270 271 private: 272 void reset() override { 273 CodeRelocations.clear(); 274 DataRelocations.clear(); 275 TypeIndices.clear(); 276 WasmIndices.clear(); 277 GOTIndices.clear(); 278 TableIndices.clear(); 279 DataLocations.clear(); 280 CustomSections.clear(); 281 ProducersSection.reset(); 282 TargetFeaturesSection.reset(); 283 CustomSectionsRelocations.clear(); 284 SignatureIndices.clear(); 285 Signatures.clear(); 286 DataSegments.clear(); 287 SectionFunctions.clear(); 288 NumFunctionImports = 0; 289 NumGlobalImports = 0; 290 MCObjectWriter::reset(); 291 } 292 293 void writeHeader(const MCAssembler &Asm); 294 295 void recordRelocation(MCAssembler &Asm, const MCAsmLayout &Layout, 296 const MCFragment *Fragment, const MCFixup &Fixup, 297 MCValue Target, uint64_t &FixedValue) override; 298 299 void executePostLayoutBinding(MCAssembler &Asm, 300 const MCAsmLayout &Layout) override; 301 302 uint64_t writeObject(MCAssembler &Asm, const MCAsmLayout &Layout) override; 303 304 void writeString(const StringRef Str) { 305 encodeULEB128(Str.size(), W.OS); 306 W.OS << Str; 307 } 308 309 void writeValueType(wasm::ValType Ty) { W.OS << static_cast<char>(Ty); } 310 311 void writeTypeSection(ArrayRef<WasmSignature> Signatures); 312 void writeImportSection(ArrayRef<wasm::WasmImport> Imports, uint32_t DataSize, 313 uint32_t NumElements); 314 void writeFunctionSection(ArrayRef<WasmFunction> Functions); 315 void writeExportSection(ArrayRef<wasm::WasmExport> Exports); 316 void writeElemSection(ArrayRef<uint32_t> TableElems); 317 void writeDataCountSection(); 318 void writeCodeSection(const MCAssembler &Asm, const MCAsmLayout &Layout, 319 ArrayRef<WasmFunction> Functions); 320 void writeDataSection(); 321 void writeEventSection(ArrayRef<wasm::WasmEventType> Events); 322 void writeRelocSection(uint32_t SectionIndex, StringRef Name, 323 std::vector<WasmRelocationEntry> &Relocations); 324 void writeLinkingMetaDataSection( 325 ArrayRef<wasm::WasmSymbolInfo> SymbolInfos, 326 ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs, 327 const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats); 328 void writeCustomSection(WasmCustomSection &CustomSection, 329 const MCAssembler &Asm, const MCAsmLayout &Layout); 330 void writeCustomRelocSections(); 331 void 332 updateCustomSectionRelocations(const SmallVector<WasmFunction, 4> &Functions, 333 const MCAsmLayout &Layout); 334 335 uint32_t getProvisionalValue(const WasmRelocationEntry &RelEntry); 336 void applyRelocations(ArrayRef<WasmRelocationEntry> Relocations, 337 uint64_t ContentsOffset); 338 339 uint32_t getRelocationIndexValue(const WasmRelocationEntry &RelEntry); 340 uint32_t getFunctionType(const MCSymbolWasm &Symbol); 341 uint32_t getEventType(const MCSymbolWasm &Symbol); 342 void registerFunctionType(const MCSymbolWasm &Symbol); 343 void registerEventType(const MCSymbolWasm &Symbol); 344 }; 345 346 } // end anonymous namespace 347 348 // Write out a section header and a patchable section size field. 349 void WasmObjectWriter::startSection(SectionBookkeeping &Section, 350 unsigned SectionId) { 351 LLVM_DEBUG(dbgs() << "startSection " << SectionId << "\n"); 352 W.OS << char(SectionId); 353 354 Section.SizeOffset = W.OS.tell(); 355 356 // The section size. We don't know the size yet, so reserve enough space 357 // for any 32-bit value; we'll patch it later. 358 encodeULEB128(0, W.OS, 5); 359 360 // The position where the section starts, for measuring its size. 361 Section.ContentsOffset = W.OS.tell(); 362 Section.PayloadOffset = W.OS.tell(); 363 Section.Index = SectionCount++; 364 } 365 366 void WasmObjectWriter::startCustomSection(SectionBookkeeping &Section, 367 StringRef Name) { 368 LLVM_DEBUG(dbgs() << "startCustomSection " << Name << "\n"); 369 startSection(Section, wasm::WASM_SEC_CUSTOM); 370 371 // The position where the section header ends, for measuring its size. 372 Section.PayloadOffset = W.OS.tell(); 373 374 // Custom sections in wasm also have a string identifier. 375 writeString(Name); 376 377 // The position where the custom section starts. 378 Section.ContentsOffset = W.OS.tell(); 379 } 380 381 // Now that the section is complete and we know how big it is, patch up the 382 // section size field at the start of the section. 383 void WasmObjectWriter::endSection(SectionBookkeeping &Section) { 384 uint64_t Size = W.OS.tell(); 385 // /dev/null doesn't support seek/tell and can report offset of 0. 386 // Simply skip this patching in that case. 387 if (!Size) 388 return; 389 390 Size -= Section.PayloadOffset; 391 if (uint32_t(Size) != Size) 392 report_fatal_error("section size does not fit in a uint32_t"); 393 394 LLVM_DEBUG(dbgs() << "endSection size=" << Size << "\n"); 395 396 // Write the final section size to the payload_len field, which follows 397 // the section id byte. 398 writePatchableLEB(static_cast<raw_pwrite_stream &>(W.OS), Size, 399 Section.SizeOffset); 400 } 401 402 // Emit the Wasm header. 403 void WasmObjectWriter::writeHeader(const MCAssembler &Asm) { 404 W.OS.write(wasm::WasmMagic, sizeof(wasm::WasmMagic)); 405 W.write<uint32_t>(wasm::WasmVersion); 406 } 407 408 void WasmObjectWriter::executePostLayoutBinding(MCAssembler &Asm, 409 const MCAsmLayout &Layout) { 410 // Build a map of sections to the function that defines them, for use 411 // in recordRelocation. 412 for (const MCSymbol &S : Asm.symbols()) { 413 const auto &WS = static_cast<const MCSymbolWasm &>(S); 414 if (WS.isDefined() && WS.isFunction() && !WS.isVariable()) { 415 const auto &Sec = static_cast<const MCSectionWasm &>(S.getSection()); 416 auto Pair = SectionFunctions.insert(std::make_pair(&Sec, &S)); 417 if (!Pair.second) 418 report_fatal_error("section already has a defining function: " + 419 Sec.getSectionName()); 420 } 421 } 422 } 423 424 void WasmObjectWriter::recordRelocation(MCAssembler &Asm, 425 const MCAsmLayout &Layout, 426 const MCFragment *Fragment, 427 const MCFixup &Fixup, MCValue Target, 428 uint64_t &FixedValue) { 429 // The WebAssembly backend should never generate FKF_IsPCRel fixups 430 assert(!(Asm.getBackend().getFixupKindInfo(Fixup.getKind()).Flags & 431 MCFixupKindInfo::FKF_IsPCRel)); 432 433 const auto &FixupSection = cast<MCSectionWasm>(*Fragment->getParent()); 434 uint64_t C = Target.getConstant(); 435 uint64_t FixupOffset = Layout.getFragmentOffset(Fragment) + Fixup.getOffset(); 436 MCContext &Ctx = Asm.getContext(); 437 438 // The .init_array isn't translated as data, so don't do relocations in it. 439 if (FixupSection.getSectionName().startswith(".init_array")) 440 return; 441 442 if (const MCSymbolRefExpr *RefB = Target.getSymB()) { 443 // To get here the A - B expression must have failed evaluateAsRelocatable. 444 // This means either A or B must be undefined and in WebAssembly we can't 445 // support either of those cases. 446 const auto &SymB = cast<MCSymbolWasm>(RefB->getSymbol()); 447 Ctx.reportError( 448 Fixup.getLoc(), 449 Twine("symbol '") + SymB.getName() + 450 "': unsupported subtraction expression used in relocation."); 451 return; 452 } 453 454 // We either rejected the fixup or folded B into C at this point. 455 const MCSymbolRefExpr *RefA = Target.getSymA(); 456 const auto *SymA = cast<MCSymbolWasm>(&RefA->getSymbol()); 457 458 if (SymA->isVariable()) { 459 const MCExpr *Expr = SymA->getVariableValue(); 460 const auto *Inner = cast<MCSymbolRefExpr>(Expr); 461 if (Inner->getKind() == MCSymbolRefExpr::VK_WEAKREF) 462 llvm_unreachable("weakref used in reloc not yet implemented"); 463 } 464 465 // Put any constant offset in an addend. Offsets can be negative, and 466 // LLVM expects wrapping, in contrast to wasm's immediates which can't 467 // be negative and don't wrap. 468 FixedValue = 0; 469 470 unsigned Type = TargetObjectWriter->getRelocType(Target, Fixup); 471 472 // Absolute offset within a section or a function. 473 // Currently only supported for for metadata sections. 474 // See: test/MC/WebAssembly/blockaddress.ll 475 if (Type == wasm::R_WASM_FUNCTION_OFFSET_I32 || 476 Type == wasm::R_WASM_SECTION_OFFSET_I32) { 477 if (!FixupSection.getKind().isMetadata()) 478 report_fatal_error("relocations for function or section offsets are " 479 "only supported in metadata sections"); 480 481 const MCSymbol *SectionSymbol = nullptr; 482 const MCSection &SecA = SymA->getSection(); 483 if (SecA.getKind().isText()) 484 SectionSymbol = SectionFunctions.find(&SecA)->second; 485 else 486 SectionSymbol = SecA.getBeginSymbol(); 487 if (!SectionSymbol) 488 report_fatal_error("section symbol is required for relocation"); 489 490 C += Layout.getSymbolOffset(*SymA); 491 SymA = cast<MCSymbolWasm>(SectionSymbol); 492 } 493 494 // Relocation other than R_WASM_TYPE_INDEX_LEB are required to be 495 // against a named symbol. 496 if (Type != wasm::R_WASM_TYPE_INDEX_LEB) { 497 if (SymA->getName().empty()) 498 report_fatal_error("relocations against un-named temporaries are not yet " 499 "supported by wasm"); 500 501 SymA->setUsedInReloc(); 502 } 503 504 if (RefA->getKind() == MCSymbolRefExpr::VK_GOT) 505 SymA->setUsedInGOT(); 506 507 WasmRelocationEntry Rec(FixupOffset, SymA, C, Type, &FixupSection); 508 LLVM_DEBUG(dbgs() << "WasmReloc: " << Rec << "\n"); 509 510 if (FixupSection.isWasmData()) { 511 DataRelocations.push_back(Rec); 512 } else if (FixupSection.getKind().isText()) { 513 CodeRelocations.push_back(Rec); 514 } else if (FixupSection.getKind().isMetadata()) { 515 CustomSectionsRelocations[&FixupSection].push_back(Rec); 516 } else { 517 llvm_unreachable("unexpected section type"); 518 } 519 } 520 521 static const MCSymbolWasm *resolveSymbol(const MCSymbolWasm &Symbol) { 522 const MCSymbolWasm* Ret = &Symbol; 523 while (Ret->isVariable()) { 524 const MCExpr *Expr = Ret->getVariableValue(); 525 auto *Inner = cast<MCSymbolRefExpr>(Expr); 526 Ret = cast<MCSymbolWasm>(&Inner->getSymbol()); 527 } 528 return Ret; 529 } 530 531 // Compute a value to write into the code at the location covered 532 // by RelEntry. This value isn't used by the static linker; it just serves 533 // to make the object format more readable and more likely to be directly 534 // useable. 535 uint32_t 536 WasmObjectWriter::getProvisionalValue(const WasmRelocationEntry &RelEntry) { 537 if (RelEntry.Type == wasm::R_WASM_GLOBAL_INDEX_LEB && !RelEntry.Symbol->isGlobal()) { 538 assert(GOTIndices.count(RelEntry.Symbol) > 0 && "symbol not found in GOT index space"); 539 return GOTIndices[RelEntry.Symbol]; 540 } 541 542 switch (RelEntry.Type) { 543 case wasm::R_WASM_TABLE_INDEX_REL_SLEB: 544 case wasm::R_WASM_TABLE_INDEX_SLEB: 545 case wasm::R_WASM_TABLE_INDEX_I32: { 546 // Provisional value is table address of the resolved symbol itself 547 const MCSymbolWasm *Sym = resolveSymbol(*RelEntry.Symbol); 548 assert(Sym->isFunction()); 549 return TableIndices[Sym]; 550 } 551 case wasm::R_WASM_TYPE_INDEX_LEB: 552 // Provisional value is same as the index 553 return getRelocationIndexValue(RelEntry); 554 case wasm::R_WASM_FUNCTION_INDEX_LEB: 555 case wasm::R_WASM_GLOBAL_INDEX_LEB: 556 case wasm::R_WASM_EVENT_INDEX_LEB: 557 // Provisional value is function/global/event Wasm index 558 assert(WasmIndices.count(RelEntry.Symbol) > 0 && "symbol not found in wasm index space"); 559 return WasmIndices[RelEntry.Symbol]; 560 case wasm::R_WASM_FUNCTION_OFFSET_I32: 561 case wasm::R_WASM_SECTION_OFFSET_I32: { 562 const auto &Section = 563 static_cast<const MCSectionWasm &>(RelEntry.Symbol->getSection()); 564 return Section.getSectionOffset() + RelEntry.Addend; 565 } 566 case wasm::R_WASM_MEMORY_ADDR_LEB: 567 case wasm::R_WASM_MEMORY_ADDR_I32: 568 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB: 569 case wasm::R_WASM_MEMORY_ADDR_SLEB: { 570 // Provisional value is address of the global 571 const MCSymbolWasm *Sym = resolveSymbol(*RelEntry.Symbol); 572 // For undefined symbols, use zero 573 if (!Sym->isDefined()) 574 return 0; 575 const wasm::WasmDataReference &Ref = DataLocations[Sym]; 576 const WasmDataSegment &Segment = DataSegments[Ref.Segment]; 577 // Ignore overflow. LLVM allows address arithmetic to silently wrap. 578 return Segment.Offset + Ref.Offset + RelEntry.Addend; 579 } 580 default: 581 llvm_unreachable("invalid relocation type"); 582 } 583 } 584 585 static void addData(SmallVectorImpl<char> &DataBytes, 586 MCSectionWasm &DataSection) { 587 LLVM_DEBUG(errs() << "addData: " << DataSection.getSectionName() << "\n"); 588 589 DataBytes.resize(alignTo(DataBytes.size(), DataSection.getAlignment())); 590 591 for (const MCFragment &Frag : DataSection) { 592 if (Frag.hasInstructions()) 593 report_fatal_error("only data supported in data sections"); 594 595 if (auto *Align = dyn_cast<MCAlignFragment>(&Frag)) { 596 if (Align->getValueSize() != 1) 597 report_fatal_error("only byte values supported for alignment"); 598 // If nops are requested, use zeros, as this is the data section. 599 uint8_t Value = Align->hasEmitNops() ? 0 : Align->getValue(); 600 uint64_t Size = 601 std::min<uint64_t>(alignTo(DataBytes.size(), Align->getAlignment()), 602 DataBytes.size() + Align->getMaxBytesToEmit()); 603 DataBytes.resize(Size, Value); 604 } else if (auto *Fill = dyn_cast<MCFillFragment>(&Frag)) { 605 int64_t NumValues; 606 if (!Fill->getNumValues().evaluateAsAbsolute(NumValues)) 607 llvm_unreachable("The fill should be an assembler constant"); 608 DataBytes.insert(DataBytes.end(), Fill->getValueSize() * NumValues, 609 Fill->getValue()); 610 } else if (auto *LEB = dyn_cast<MCLEBFragment>(&Frag)) { 611 const SmallVectorImpl<char> &Contents = LEB->getContents(); 612 DataBytes.insert(DataBytes.end(), Contents.begin(), Contents.end()); 613 } else { 614 const auto &DataFrag = cast<MCDataFragment>(Frag); 615 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 616 DataBytes.insert(DataBytes.end(), Contents.begin(), Contents.end()); 617 } 618 } 619 620 LLVM_DEBUG(dbgs() << "addData -> " << DataBytes.size() << "\n"); 621 } 622 623 uint32_t 624 WasmObjectWriter::getRelocationIndexValue(const WasmRelocationEntry &RelEntry) { 625 if (RelEntry.Type == wasm::R_WASM_TYPE_INDEX_LEB) { 626 if (!TypeIndices.count(RelEntry.Symbol)) 627 report_fatal_error("symbol not found in type index space: " + 628 RelEntry.Symbol->getName()); 629 return TypeIndices[RelEntry.Symbol]; 630 } 631 632 return RelEntry.Symbol->getIndex(); 633 } 634 635 // Apply the portions of the relocation records that we can handle ourselves 636 // directly. 637 void WasmObjectWriter::applyRelocations( 638 ArrayRef<WasmRelocationEntry> Relocations, uint64_t ContentsOffset) { 639 auto &Stream = static_cast<raw_pwrite_stream &>(W.OS); 640 for (const WasmRelocationEntry &RelEntry : Relocations) { 641 uint64_t Offset = ContentsOffset + 642 RelEntry.FixupSection->getSectionOffset() + 643 RelEntry.Offset; 644 645 LLVM_DEBUG(dbgs() << "applyRelocation: " << RelEntry << "\n"); 646 uint32_t Value = getProvisionalValue(RelEntry); 647 648 switch (RelEntry.Type) { 649 case wasm::R_WASM_FUNCTION_INDEX_LEB: 650 case wasm::R_WASM_TYPE_INDEX_LEB: 651 case wasm::R_WASM_GLOBAL_INDEX_LEB: 652 case wasm::R_WASM_MEMORY_ADDR_LEB: 653 case wasm::R_WASM_EVENT_INDEX_LEB: 654 writePatchableLEB(Stream, Value, Offset); 655 break; 656 case wasm::R_WASM_TABLE_INDEX_I32: 657 case wasm::R_WASM_MEMORY_ADDR_I32: 658 case wasm::R_WASM_FUNCTION_OFFSET_I32: 659 case wasm::R_WASM_SECTION_OFFSET_I32: 660 writeI32(Stream, Value, Offset); 661 break; 662 case wasm::R_WASM_TABLE_INDEX_SLEB: 663 case wasm::R_WASM_TABLE_INDEX_REL_SLEB: 664 case wasm::R_WASM_MEMORY_ADDR_SLEB: 665 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB: 666 writePatchableSLEB(Stream, Value, Offset); 667 break; 668 default: 669 llvm_unreachable("invalid relocation type"); 670 } 671 } 672 } 673 674 void WasmObjectWriter::writeTypeSection(ArrayRef<WasmSignature> Signatures) { 675 if (Signatures.empty()) 676 return; 677 678 SectionBookkeeping Section; 679 startSection(Section, wasm::WASM_SEC_TYPE); 680 681 encodeULEB128(Signatures.size(), W.OS); 682 683 for (const WasmSignature &Sig : Signatures) { 684 W.OS << char(wasm::WASM_TYPE_FUNC); 685 encodeULEB128(Sig.Params.size(), W.OS); 686 for (wasm::ValType Ty : Sig.Params) 687 writeValueType(Ty); 688 encodeULEB128(Sig.Returns.size(), W.OS); 689 for (wasm::ValType Ty : Sig.Returns) 690 writeValueType(Ty); 691 } 692 693 endSection(Section); 694 } 695 696 void WasmObjectWriter::writeImportSection(ArrayRef<wasm::WasmImport> Imports, 697 uint32_t DataSize, 698 uint32_t NumElements) { 699 if (Imports.empty()) 700 return; 701 702 uint32_t NumPages = (DataSize + wasm::WasmPageSize - 1) / wasm::WasmPageSize; 703 704 SectionBookkeeping Section; 705 startSection(Section, wasm::WASM_SEC_IMPORT); 706 707 encodeULEB128(Imports.size(), W.OS); 708 for (const wasm::WasmImport &Import : Imports) { 709 writeString(Import.Module); 710 writeString(Import.Field); 711 W.OS << char(Import.Kind); 712 713 switch (Import.Kind) { 714 case wasm::WASM_EXTERNAL_FUNCTION: 715 encodeULEB128(Import.SigIndex, W.OS); 716 break; 717 case wasm::WASM_EXTERNAL_GLOBAL: 718 W.OS << char(Import.Global.Type); 719 W.OS << char(Import.Global.Mutable ? 1 : 0); 720 break; 721 case wasm::WASM_EXTERNAL_MEMORY: 722 encodeULEB128(0, W.OS); // flags 723 encodeULEB128(NumPages, W.OS); // initial 724 break; 725 case wasm::WASM_EXTERNAL_TABLE: 726 W.OS << char(Import.Table.ElemType); 727 encodeULEB128(0, W.OS); // flags 728 encodeULEB128(NumElements, W.OS); // initial 729 break; 730 case wasm::WASM_EXTERNAL_EVENT: 731 encodeULEB128(Import.Event.Attribute, W.OS); 732 encodeULEB128(Import.Event.SigIndex, W.OS); 733 break; 734 default: 735 llvm_unreachable("unsupported import kind"); 736 } 737 } 738 739 endSection(Section); 740 } 741 742 void WasmObjectWriter::writeFunctionSection(ArrayRef<WasmFunction> Functions) { 743 if (Functions.empty()) 744 return; 745 746 SectionBookkeeping Section; 747 startSection(Section, wasm::WASM_SEC_FUNCTION); 748 749 encodeULEB128(Functions.size(), W.OS); 750 for (const WasmFunction &Func : Functions) 751 encodeULEB128(Func.SigIndex, W.OS); 752 753 endSection(Section); 754 } 755 756 void WasmObjectWriter::writeEventSection(ArrayRef<wasm::WasmEventType> Events) { 757 if (Events.empty()) 758 return; 759 760 SectionBookkeeping Section; 761 startSection(Section, wasm::WASM_SEC_EVENT); 762 763 encodeULEB128(Events.size(), W.OS); 764 for (const wasm::WasmEventType &Event : Events) { 765 encodeULEB128(Event.Attribute, W.OS); 766 encodeULEB128(Event.SigIndex, W.OS); 767 } 768 769 endSection(Section); 770 } 771 772 void WasmObjectWriter::writeExportSection(ArrayRef<wasm::WasmExport> Exports) { 773 if (Exports.empty()) 774 return; 775 776 SectionBookkeeping Section; 777 startSection(Section, wasm::WASM_SEC_EXPORT); 778 779 encodeULEB128(Exports.size(), W.OS); 780 for (const wasm::WasmExport &Export : Exports) { 781 writeString(Export.Name); 782 W.OS << char(Export.Kind); 783 encodeULEB128(Export.Index, W.OS); 784 } 785 786 endSection(Section); 787 } 788 789 void WasmObjectWriter::writeElemSection(ArrayRef<uint32_t> TableElems) { 790 if (TableElems.empty()) 791 return; 792 793 SectionBookkeeping Section; 794 startSection(Section, wasm::WASM_SEC_ELEM); 795 796 encodeULEB128(1, W.OS); // number of "segments" 797 encodeULEB128(0, W.OS); // the table index 798 799 // init expr for starting offset 800 W.OS << char(wasm::WASM_OPCODE_I32_CONST); 801 encodeSLEB128(InitialTableOffset, W.OS); 802 W.OS << char(wasm::WASM_OPCODE_END); 803 804 encodeULEB128(TableElems.size(), W.OS); 805 for (uint32_t Elem : TableElems) 806 encodeULEB128(Elem, W.OS); 807 808 endSection(Section); 809 } 810 811 void WasmObjectWriter::writeDataCountSection() { 812 if (DataSegments.empty()) 813 return; 814 815 SectionBookkeeping Section; 816 startSection(Section, wasm::WASM_SEC_DATACOUNT); 817 encodeULEB128(DataSegments.size(), W.OS); 818 endSection(Section); 819 } 820 821 void WasmObjectWriter::writeCodeSection(const MCAssembler &Asm, 822 const MCAsmLayout &Layout, 823 ArrayRef<WasmFunction> Functions) { 824 if (Functions.empty()) 825 return; 826 827 SectionBookkeeping Section; 828 startSection(Section, wasm::WASM_SEC_CODE); 829 CodeSectionIndex = Section.Index; 830 831 encodeULEB128(Functions.size(), W.OS); 832 833 for (const WasmFunction &Func : Functions) { 834 auto &FuncSection = static_cast<MCSectionWasm &>(Func.Sym->getSection()); 835 836 int64_t Size = 0; 837 if (!Func.Sym->getSize()->evaluateAsAbsolute(Size, Layout)) 838 report_fatal_error(".size expression must be evaluatable"); 839 840 encodeULEB128(Size, W.OS); 841 FuncSection.setSectionOffset(W.OS.tell() - Section.ContentsOffset); 842 Asm.writeSectionData(W.OS, &FuncSection, Layout); 843 } 844 845 // Apply fixups. 846 applyRelocations(CodeRelocations, Section.ContentsOffset); 847 848 endSection(Section); 849 } 850 851 void WasmObjectWriter::writeDataSection() { 852 if (DataSegments.empty()) 853 return; 854 855 SectionBookkeeping Section; 856 startSection(Section, wasm::WASM_SEC_DATA); 857 DataSectionIndex = Section.Index; 858 859 encodeULEB128(DataSegments.size(), W.OS); // count 860 861 for (const WasmDataSegment &Segment : DataSegments) { 862 encodeULEB128(Segment.InitFlags, W.OS); // flags 863 if (Segment.InitFlags & wasm::WASM_SEGMENT_HAS_MEMINDEX) 864 encodeULEB128(0, W.OS); // memory index 865 if ((Segment.InitFlags & wasm::WASM_SEGMENT_IS_PASSIVE) == 0) { 866 W.OS << char(wasm::WASM_OPCODE_I32_CONST); 867 encodeSLEB128(Segment.Offset, W.OS); // offset 868 W.OS << char(wasm::WASM_OPCODE_END); 869 } 870 encodeULEB128(Segment.Data.size(), W.OS); // size 871 Segment.Section->setSectionOffset(W.OS.tell() - Section.ContentsOffset); 872 W.OS << Segment.Data; // data 873 } 874 875 // Apply fixups. 876 applyRelocations(DataRelocations, Section.ContentsOffset); 877 878 endSection(Section); 879 } 880 881 void WasmObjectWriter::writeRelocSection( 882 uint32_t SectionIndex, StringRef Name, 883 std::vector<WasmRelocationEntry> &Relocs) { 884 // See: https://github.com/WebAssembly/tool-conventions/blob/master/Linking.md 885 // for descriptions of the reloc sections. 886 887 if (Relocs.empty()) 888 return; 889 890 // First, ensure the relocations are sorted in offset order. In general they 891 // should already be sorted since `recordRelocation` is called in offset 892 // order, but for the code section we combine many MC sections into single 893 // wasm section, and this order is determined by the order of Asm.Symbols() 894 // not the sections order. 895 llvm::stable_sort( 896 Relocs, [](const WasmRelocationEntry &A, const WasmRelocationEntry &B) { 897 return (A.Offset + A.FixupSection->getSectionOffset()) < 898 (B.Offset + B.FixupSection->getSectionOffset()); 899 }); 900 901 SectionBookkeeping Section; 902 startCustomSection(Section, std::string("reloc.") + Name.str()); 903 904 encodeULEB128(SectionIndex, W.OS); 905 encodeULEB128(Relocs.size(), W.OS); 906 for (const WasmRelocationEntry &RelEntry : Relocs) { 907 uint64_t Offset = 908 RelEntry.Offset + RelEntry.FixupSection->getSectionOffset(); 909 uint32_t Index = getRelocationIndexValue(RelEntry); 910 911 W.OS << char(RelEntry.Type); 912 encodeULEB128(Offset, W.OS); 913 encodeULEB128(Index, W.OS); 914 if (RelEntry.hasAddend()) 915 encodeSLEB128(RelEntry.Addend, W.OS); 916 } 917 918 endSection(Section); 919 } 920 921 void WasmObjectWriter::writeCustomRelocSections() { 922 for (const auto &Sec : CustomSections) { 923 auto &Relocations = CustomSectionsRelocations[Sec.Section]; 924 writeRelocSection(Sec.OutputIndex, Sec.Name, Relocations); 925 } 926 } 927 928 void WasmObjectWriter::writeLinkingMetaDataSection( 929 ArrayRef<wasm::WasmSymbolInfo> SymbolInfos, 930 ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs, 931 const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats) { 932 SectionBookkeeping Section; 933 startCustomSection(Section, "linking"); 934 encodeULEB128(wasm::WasmMetadataVersion, W.OS); 935 936 SectionBookkeeping SubSection; 937 if (SymbolInfos.size() != 0) { 938 startSection(SubSection, wasm::WASM_SYMBOL_TABLE); 939 encodeULEB128(SymbolInfos.size(), W.OS); 940 for (const wasm::WasmSymbolInfo &Sym : SymbolInfos) { 941 encodeULEB128(Sym.Kind, W.OS); 942 encodeULEB128(Sym.Flags, W.OS); 943 switch (Sym.Kind) { 944 case wasm::WASM_SYMBOL_TYPE_FUNCTION: 945 case wasm::WASM_SYMBOL_TYPE_GLOBAL: 946 case wasm::WASM_SYMBOL_TYPE_EVENT: 947 encodeULEB128(Sym.ElementIndex, W.OS); 948 if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0 || 949 (Sym.Flags & wasm::WASM_SYMBOL_EXPLICIT_NAME) != 0) 950 writeString(Sym.Name); 951 break; 952 case wasm::WASM_SYMBOL_TYPE_DATA: 953 writeString(Sym.Name); 954 if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0) { 955 encodeULEB128(Sym.DataRef.Segment, W.OS); 956 encodeULEB128(Sym.DataRef.Offset, W.OS); 957 encodeULEB128(Sym.DataRef.Size, W.OS); 958 } 959 break; 960 case wasm::WASM_SYMBOL_TYPE_SECTION: { 961 const uint32_t SectionIndex = 962 CustomSections[Sym.ElementIndex].OutputIndex; 963 encodeULEB128(SectionIndex, W.OS); 964 break; 965 } 966 default: 967 llvm_unreachable("unexpected kind"); 968 } 969 } 970 endSection(SubSection); 971 } 972 973 if (DataSegments.size()) { 974 startSection(SubSection, wasm::WASM_SEGMENT_INFO); 975 encodeULEB128(DataSegments.size(), W.OS); 976 for (const WasmDataSegment &Segment : DataSegments) { 977 writeString(Segment.Name); 978 encodeULEB128(Segment.Alignment, W.OS); 979 encodeULEB128(Segment.LinkerFlags, W.OS); 980 } 981 endSection(SubSection); 982 } 983 984 if (!InitFuncs.empty()) { 985 startSection(SubSection, wasm::WASM_INIT_FUNCS); 986 encodeULEB128(InitFuncs.size(), W.OS); 987 for (auto &StartFunc : InitFuncs) { 988 encodeULEB128(StartFunc.first, W.OS); // priority 989 encodeULEB128(StartFunc.second, W.OS); // function index 990 } 991 endSection(SubSection); 992 } 993 994 if (Comdats.size()) { 995 startSection(SubSection, wasm::WASM_COMDAT_INFO); 996 encodeULEB128(Comdats.size(), W.OS); 997 for (const auto &C : Comdats) { 998 writeString(C.first); 999 encodeULEB128(0, W.OS); // flags for future use 1000 encodeULEB128(C.second.size(), W.OS); 1001 for (const WasmComdatEntry &Entry : C.second) { 1002 encodeULEB128(Entry.Kind, W.OS); 1003 encodeULEB128(Entry.Index, W.OS); 1004 } 1005 } 1006 endSection(SubSection); 1007 } 1008 1009 endSection(Section); 1010 } 1011 1012 void WasmObjectWriter::writeCustomSection(WasmCustomSection &CustomSection, 1013 const MCAssembler &Asm, 1014 const MCAsmLayout &Layout) { 1015 SectionBookkeeping Section; 1016 auto *Sec = CustomSection.Section; 1017 startCustomSection(Section, CustomSection.Name); 1018 1019 Sec->setSectionOffset(W.OS.tell() - Section.ContentsOffset); 1020 Asm.writeSectionData(W.OS, Sec, Layout); 1021 1022 CustomSection.OutputContentsOffset = Section.ContentsOffset; 1023 CustomSection.OutputIndex = Section.Index; 1024 1025 endSection(Section); 1026 1027 // Apply fixups. 1028 auto &Relocations = CustomSectionsRelocations[CustomSection.Section]; 1029 applyRelocations(Relocations, CustomSection.OutputContentsOffset); 1030 } 1031 1032 uint32_t WasmObjectWriter::getFunctionType(const MCSymbolWasm &Symbol) { 1033 assert(Symbol.isFunction()); 1034 assert(TypeIndices.count(&Symbol)); 1035 return TypeIndices[&Symbol]; 1036 } 1037 1038 uint32_t WasmObjectWriter::getEventType(const MCSymbolWasm &Symbol) { 1039 assert(Symbol.isEvent()); 1040 assert(TypeIndices.count(&Symbol)); 1041 return TypeIndices[&Symbol]; 1042 } 1043 1044 void WasmObjectWriter::registerFunctionType(const MCSymbolWasm &Symbol) { 1045 assert(Symbol.isFunction()); 1046 1047 WasmSignature S; 1048 const MCSymbolWasm *ResolvedSym = resolveSymbol(Symbol); 1049 if (auto *Sig = ResolvedSym->getSignature()) { 1050 S.Returns = Sig->Returns; 1051 S.Params = Sig->Params; 1052 } 1053 1054 auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size())); 1055 if (Pair.second) 1056 Signatures.push_back(S); 1057 TypeIndices[&Symbol] = Pair.first->second; 1058 1059 LLVM_DEBUG(dbgs() << "registerFunctionType: " << Symbol 1060 << " new:" << Pair.second << "\n"); 1061 LLVM_DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n"); 1062 } 1063 1064 void WasmObjectWriter::registerEventType(const MCSymbolWasm &Symbol) { 1065 assert(Symbol.isEvent()); 1066 1067 // TODO Currently we don't generate imported exceptions, but if we do, we 1068 // should have a way of infering types of imported exceptions. 1069 WasmSignature S; 1070 if (auto *Sig = Symbol.getSignature()) { 1071 S.Returns = Sig->Returns; 1072 S.Params = Sig->Params; 1073 } 1074 1075 auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size())); 1076 if (Pair.second) 1077 Signatures.push_back(S); 1078 TypeIndices[&Symbol] = Pair.first->second; 1079 1080 LLVM_DEBUG(dbgs() << "registerEventType: " << Symbol << " new:" << Pair.second 1081 << "\n"); 1082 LLVM_DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n"); 1083 } 1084 1085 static bool isInSymtab(const MCSymbolWasm &Sym) { 1086 if (Sym.isUsedInReloc()) 1087 return true; 1088 1089 if (Sym.isComdat() && !Sym.isDefined()) 1090 return false; 1091 1092 if (Sym.isTemporary() && Sym.getName().empty()) 1093 return false; 1094 1095 if (Sym.isTemporary() && Sym.isData() && !Sym.getSize()) 1096 return false; 1097 1098 if (Sym.isSection()) 1099 return false; 1100 1101 return true; 1102 } 1103 1104 uint64_t WasmObjectWriter::writeObject(MCAssembler &Asm, 1105 const MCAsmLayout &Layout) { 1106 uint64_t StartOffset = W.OS.tell(); 1107 1108 LLVM_DEBUG(dbgs() << "WasmObjectWriter::writeObject\n"); 1109 1110 // Collect information from the available symbols. 1111 SmallVector<WasmFunction, 4> Functions; 1112 SmallVector<uint32_t, 4> TableElems; 1113 SmallVector<wasm::WasmImport, 4> Imports; 1114 SmallVector<wasm::WasmExport, 4> Exports; 1115 SmallVector<wasm::WasmEventType, 1> Events; 1116 SmallVector<wasm::WasmSymbolInfo, 4> SymbolInfos; 1117 SmallVector<std::pair<uint16_t, uint32_t>, 2> InitFuncs; 1118 std::map<StringRef, std::vector<WasmComdatEntry>> Comdats; 1119 uint32_t DataSize = 0; 1120 1121 // For now, always emit the memory import, since loads and stores are not 1122 // valid without it. In the future, we could perhaps be more clever and omit 1123 // it if there are no loads or stores. 1124 wasm::WasmImport MemImport; 1125 MemImport.Module = "env"; 1126 MemImport.Field = "__linear_memory"; 1127 MemImport.Kind = wasm::WASM_EXTERNAL_MEMORY; 1128 Imports.push_back(MemImport); 1129 1130 // For now, always emit the table section, since indirect calls are not 1131 // valid without it. In the future, we could perhaps be more clever and omit 1132 // it if there are no indirect calls. 1133 wasm::WasmImport TableImport; 1134 TableImport.Module = "env"; 1135 TableImport.Field = "__indirect_function_table"; 1136 TableImport.Kind = wasm::WASM_EXTERNAL_TABLE; 1137 TableImport.Table.ElemType = wasm::WASM_TYPE_FUNCREF; 1138 Imports.push_back(TableImport); 1139 1140 // Populate SignatureIndices, and Imports and WasmIndices for undefined 1141 // symbols. This must be done before populating WasmIndices for defined 1142 // symbols. 1143 for (const MCSymbol &S : Asm.symbols()) { 1144 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1145 1146 // Register types for all functions, including those with private linkage 1147 // (because wasm always needs a type signature). 1148 if (WS.isFunction()) 1149 registerFunctionType(WS); 1150 1151 if (WS.isEvent()) 1152 registerEventType(WS); 1153 1154 if (WS.isTemporary()) 1155 continue; 1156 1157 // If the symbol is not defined in this translation unit, import it. 1158 if (!WS.isDefined() && !WS.isComdat()) { 1159 if (WS.isFunction()) { 1160 wasm::WasmImport Import; 1161 Import.Module = WS.getImportModule(); 1162 Import.Field = WS.getImportName(); 1163 Import.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1164 Import.SigIndex = getFunctionType(WS); 1165 Imports.push_back(Import); 1166 assert(WasmIndices.count(&WS) == 0); 1167 WasmIndices[&WS] = NumFunctionImports++; 1168 } else if (WS.isGlobal()) { 1169 if (WS.isWeak()) 1170 report_fatal_error("undefined global symbol cannot be weak"); 1171 1172 wasm::WasmImport Import; 1173 Import.Field = WS.getImportName(); 1174 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1175 Import.Module = WS.getImportModule(); 1176 Import.Global = WS.getGlobalType(); 1177 Imports.push_back(Import); 1178 assert(WasmIndices.count(&WS) == 0); 1179 WasmIndices[&WS] = NumGlobalImports++; 1180 } else if (WS.isEvent()) { 1181 if (WS.isWeak()) 1182 report_fatal_error("undefined event symbol cannot be weak"); 1183 1184 wasm::WasmImport Import; 1185 Import.Module = WS.getImportModule(); 1186 Import.Field = WS.getImportName(); 1187 Import.Kind = wasm::WASM_EXTERNAL_EVENT; 1188 Import.Event.Attribute = wasm::WASM_EVENT_ATTRIBUTE_EXCEPTION; 1189 Import.Event.SigIndex = getEventType(WS); 1190 Imports.push_back(Import); 1191 assert(WasmIndices.count(&WS) == 0); 1192 WasmIndices[&WS] = NumEventImports++; 1193 } 1194 } 1195 } 1196 1197 // Add imports for GOT globals 1198 for (const MCSymbol &S : Asm.symbols()) { 1199 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1200 if (WS.isUsedInGOT()) { 1201 wasm::WasmImport Import; 1202 if (WS.isFunction()) 1203 Import.Module = "GOT.func"; 1204 else 1205 Import.Module = "GOT.mem"; 1206 Import.Field = WS.getName(); 1207 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1208 Import.Global = {wasm::WASM_TYPE_I32, true}; 1209 Imports.push_back(Import); 1210 assert(GOTIndices.count(&WS) == 0); 1211 GOTIndices[&WS] = NumGlobalImports++; 1212 } 1213 } 1214 1215 // Populate DataSegments and CustomSections, which must be done before 1216 // populating DataLocations. 1217 for (MCSection &Sec : Asm) { 1218 auto &Section = static_cast<MCSectionWasm &>(Sec); 1219 StringRef SectionName = Section.getSectionName(); 1220 1221 // .init_array sections are handled specially elsewhere. 1222 if (SectionName.startswith(".init_array")) 1223 continue; 1224 1225 // Code is handled separately 1226 if (Section.getKind().isText()) 1227 continue; 1228 1229 if (Section.isWasmData()) { 1230 uint32_t SegmentIndex = DataSegments.size(); 1231 DataSize = alignTo(DataSize, Section.getAlignment()); 1232 DataSegments.emplace_back(); 1233 WasmDataSegment &Segment = DataSegments.back(); 1234 Segment.Name = SectionName; 1235 Segment.InitFlags = 1236 Section.getPassive() ? (uint32_t)wasm::WASM_SEGMENT_IS_PASSIVE : 0; 1237 Segment.Offset = DataSize; 1238 Segment.Section = &Section; 1239 addData(Segment.Data, Section); 1240 Segment.Alignment = Log2_32(Section.getAlignment()); 1241 Segment.LinkerFlags = 0; 1242 DataSize += Segment.Data.size(); 1243 Section.setSegmentIndex(SegmentIndex); 1244 1245 if (const MCSymbolWasm *C = Section.getGroup()) { 1246 Comdats[C->getName()].emplace_back( 1247 WasmComdatEntry{wasm::WASM_COMDAT_DATA, SegmentIndex}); 1248 } 1249 } else { 1250 // Create custom sections 1251 assert(Sec.getKind().isMetadata()); 1252 1253 StringRef Name = SectionName; 1254 1255 // For user-defined custom sections, strip the prefix 1256 if (Name.startswith(".custom_section.")) 1257 Name = Name.substr(strlen(".custom_section.")); 1258 1259 MCSymbol *Begin = Sec.getBeginSymbol(); 1260 if (Begin) { 1261 WasmIndices[cast<MCSymbolWasm>(Begin)] = CustomSections.size(); 1262 if (SectionName != Begin->getName()) 1263 report_fatal_error("section name and begin symbol should match: " + 1264 Twine(SectionName)); 1265 } 1266 1267 // Separate out the producers and target features sections 1268 if (Name == "producers") { 1269 ProducersSection = std::make_unique<WasmCustomSection>(Name, &Section); 1270 continue; 1271 } 1272 if (Name == "target_features") { 1273 TargetFeaturesSection = 1274 std::make_unique<WasmCustomSection>(Name, &Section); 1275 continue; 1276 } 1277 1278 CustomSections.emplace_back(Name, &Section); 1279 } 1280 } 1281 1282 // Populate WasmIndices and DataLocations for defined symbols. 1283 for (const MCSymbol &S : Asm.symbols()) { 1284 // Ignore unnamed temporary symbols, which aren't ever exported, imported, 1285 // or used in relocations. 1286 if (S.isTemporary() && S.getName().empty()) 1287 continue; 1288 1289 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1290 LLVM_DEBUG( 1291 dbgs() << "MCSymbol: " << toString(WS.getType()) << " '" << S << "'" 1292 << " isDefined=" << S.isDefined() << " isExternal=" 1293 << S.isExternal() << " isTemporary=" << S.isTemporary() 1294 << " isWeak=" << WS.isWeak() << " isHidden=" << WS.isHidden() 1295 << " isVariable=" << WS.isVariable() << "\n"); 1296 1297 if (WS.isVariable()) 1298 continue; 1299 if (WS.isComdat() && !WS.isDefined()) 1300 continue; 1301 1302 if (WS.isFunction()) { 1303 unsigned Index; 1304 if (WS.isDefined()) { 1305 if (WS.getOffset() != 0) 1306 report_fatal_error( 1307 "function sections must contain one function each"); 1308 1309 if (WS.getSize() == nullptr) 1310 report_fatal_error( 1311 "function symbols must have a size set with .size"); 1312 1313 // A definition. Write out the function body. 1314 Index = NumFunctionImports + Functions.size(); 1315 WasmFunction Func; 1316 Func.SigIndex = getFunctionType(WS); 1317 Func.Sym = &WS; 1318 WasmIndices[&WS] = Index; 1319 Functions.push_back(Func); 1320 1321 auto &Section = static_cast<MCSectionWasm &>(WS.getSection()); 1322 if (const MCSymbolWasm *C = Section.getGroup()) { 1323 Comdats[C->getName()].emplace_back( 1324 WasmComdatEntry{wasm::WASM_COMDAT_FUNCTION, Index}); 1325 } 1326 } else { 1327 // An import; the index was assigned above. 1328 Index = WasmIndices.find(&WS)->second; 1329 } 1330 1331 LLVM_DEBUG(dbgs() << " -> function index: " << Index << "\n"); 1332 1333 } else if (WS.isData()) { 1334 if (!isInSymtab(WS)) 1335 continue; 1336 1337 if (!WS.isDefined()) { 1338 LLVM_DEBUG(dbgs() << " -> segment index: -1" 1339 << "\n"); 1340 continue; 1341 } 1342 1343 if (!WS.getSize()) 1344 report_fatal_error("data symbols must have a size set with .size: " + 1345 WS.getName()); 1346 1347 int64_t Size = 0; 1348 if (!WS.getSize()->evaluateAsAbsolute(Size, Layout)) 1349 report_fatal_error(".size expression must be evaluatable"); 1350 1351 auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection()); 1352 assert(DataSection.isWasmData()); 1353 1354 // For each data symbol, export it in the symtab as a reference to the 1355 // corresponding Wasm data segment. 1356 wasm::WasmDataReference Ref = wasm::WasmDataReference{ 1357 DataSection.getSegmentIndex(), 1358 static_cast<uint32_t>(Layout.getSymbolOffset(WS)), 1359 static_cast<uint32_t>(Size)}; 1360 DataLocations[&WS] = Ref; 1361 LLVM_DEBUG(dbgs() << " -> segment index: " << Ref.Segment << "\n"); 1362 1363 } else if (WS.isGlobal()) { 1364 // A "true" Wasm global (currently just __stack_pointer) 1365 if (WS.isDefined()) 1366 report_fatal_error("don't yet support defined globals"); 1367 1368 // An import; the index was assigned above 1369 LLVM_DEBUG(dbgs() << " -> global index: " 1370 << WasmIndices.find(&WS)->second << "\n"); 1371 1372 } else if (WS.isEvent()) { 1373 // C++ exception symbol (__cpp_exception) 1374 unsigned Index; 1375 if (WS.isDefined()) { 1376 Index = NumEventImports + Events.size(); 1377 wasm::WasmEventType Event; 1378 Event.SigIndex = getEventType(WS); 1379 Event.Attribute = wasm::WASM_EVENT_ATTRIBUTE_EXCEPTION; 1380 assert(WasmIndices.count(&WS) == 0); 1381 WasmIndices[&WS] = Index; 1382 Events.push_back(Event); 1383 } else { 1384 // An import; the index was assigned above. 1385 assert(WasmIndices.count(&WS) > 0); 1386 } 1387 LLVM_DEBUG(dbgs() << " -> event index: " << WasmIndices.find(&WS)->second 1388 << "\n"); 1389 1390 } else { 1391 assert(WS.isSection()); 1392 } 1393 } 1394 1395 // Populate WasmIndices and DataLocations for aliased symbols. We need to 1396 // process these in a separate pass because we need to have processed the 1397 // target of the alias before the alias itself and the symbols are not 1398 // necessarily ordered in this way. 1399 for (const MCSymbol &S : Asm.symbols()) { 1400 if (!S.isVariable()) 1401 continue; 1402 1403 assert(S.isDefined()); 1404 1405 // Find the target symbol of this weak alias and export that index 1406 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1407 const MCSymbolWasm *ResolvedSym = resolveSymbol(WS); 1408 LLVM_DEBUG(dbgs() << WS.getName() << ": weak alias of '" << *ResolvedSym 1409 << "'\n"); 1410 1411 if (ResolvedSym->isFunction()) { 1412 assert(WasmIndices.count(ResolvedSym) > 0); 1413 uint32_t WasmIndex = WasmIndices.find(ResolvedSym)->second; 1414 assert(WasmIndices.count(&WS) == 0); 1415 WasmIndices[&WS] = WasmIndex; 1416 LLVM_DEBUG(dbgs() << " -> index:" << WasmIndex << "\n"); 1417 } else if (ResolvedSym->isData()) { 1418 assert(DataLocations.count(ResolvedSym) > 0); 1419 const wasm::WasmDataReference &Ref = 1420 DataLocations.find(ResolvedSym)->second; 1421 DataLocations[&WS] = Ref; 1422 LLVM_DEBUG(dbgs() << " -> index:" << Ref.Segment << "\n"); 1423 } else { 1424 report_fatal_error("don't yet support global/event aliases"); 1425 } 1426 } 1427 1428 // Finally, populate the symbol table itself, in its "natural" order. 1429 for (const MCSymbol &S : Asm.symbols()) { 1430 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1431 if (!isInSymtab(WS)) { 1432 WS.setIndex(InvalidIndex); 1433 continue; 1434 } 1435 LLVM_DEBUG(dbgs() << "adding to symtab: " << WS << "\n"); 1436 1437 uint32_t Flags = 0; 1438 if (WS.isWeak()) 1439 Flags |= wasm::WASM_SYMBOL_BINDING_WEAK; 1440 if (WS.isHidden()) 1441 Flags |= wasm::WASM_SYMBOL_VISIBILITY_HIDDEN; 1442 if (!WS.isExternal() && WS.isDefined()) 1443 Flags |= wasm::WASM_SYMBOL_BINDING_LOCAL; 1444 if (WS.isUndefined()) 1445 Flags |= wasm::WASM_SYMBOL_UNDEFINED; 1446 if (WS.isExported()) 1447 Flags |= wasm::WASM_SYMBOL_EXPORTED; 1448 if (WS.getName() != WS.getImportName()) 1449 Flags |= wasm::WASM_SYMBOL_EXPLICIT_NAME; 1450 1451 wasm::WasmSymbolInfo Info; 1452 Info.Name = WS.getName(); 1453 Info.Kind = WS.getType(); 1454 Info.Flags = Flags; 1455 if (!WS.isData()) { 1456 assert(WasmIndices.count(&WS) > 0); 1457 Info.ElementIndex = WasmIndices.find(&WS)->second; 1458 } else if (WS.isDefined()) { 1459 assert(DataLocations.count(&WS) > 0); 1460 Info.DataRef = DataLocations.find(&WS)->second; 1461 } 1462 WS.setIndex(SymbolInfos.size()); 1463 SymbolInfos.emplace_back(Info); 1464 } 1465 1466 { 1467 auto HandleReloc = [&](const WasmRelocationEntry &Rel) { 1468 // Functions referenced by a relocation need to put in the table. This is 1469 // purely to make the object file's provisional values readable, and is 1470 // ignored by the linker, which re-calculates the relocations itself. 1471 if (Rel.Type != wasm::R_WASM_TABLE_INDEX_I32 && 1472 Rel.Type != wasm::R_WASM_TABLE_INDEX_SLEB) 1473 return; 1474 assert(Rel.Symbol->isFunction()); 1475 const MCSymbolWasm &WS = *resolveSymbol(*Rel.Symbol); 1476 uint32_t FunctionIndex = WasmIndices.find(&WS)->second; 1477 uint32_t TableIndex = TableElems.size() + InitialTableOffset; 1478 if (TableIndices.try_emplace(&WS, TableIndex).second) { 1479 LLVM_DEBUG(dbgs() << " -> adding " << WS.getName() 1480 << " to table: " << TableIndex << "\n"); 1481 TableElems.push_back(FunctionIndex); 1482 registerFunctionType(WS); 1483 } 1484 }; 1485 1486 for (const WasmRelocationEntry &RelEntry : CodeRelocations) 1487 HandleReloc(RelEntry); 1488 for (const WasmRelocationEntry &RelEntry : DataRelocations) 1489 HandleReloc(RelEntry); 1490 } 1491 1492 // Translate .init_array section contents into start functions. 1493 for (const MCSection &S : Asm) { 1494 const auto &WS = static_cast<const MCSectionWasm &>(S); 1495 if (WS.getSectionName().startswith(".fini_array")) 1496 report_fatal_error(".fini_array sections are unsupported"); 1497 if (!WS.getSectionName().startswith(".init_array")) 1498 continue; 1499 if (WS.getFragmentList().empty()) 1500 continue; 1501 1502 // init_array is expected to contain a single non-empty data fragment 1503 if (WS.getFragmentList().size() != 3) 1504 report_fatal_error("only one .init_array section fragment supported"); 1505 1506 auto IT = WS.begin(); 1507 const MCFragment &EmptyFrag = *IT; 1508 if (EmptyFrag.getKind() != MCFragment::FT_Data) 1509 report_fatal_error(".init_array section should be aligned"); 1510 1511 IT = std::next(IT); 1512 const MCFragment &AlignFrag = *IT; 1513 if (AlignFrag.getKind() != MCFragment::FT_Align) 1514 report_fatal_error(".init_array section should be aligned"); 1515 if (cast<MCAlignFragment>(AlignFrag).getAlignment() != (is64Bit() ? 8 : 4)) 1516 report_fatal_error(".init_array section should be aligned for pointers"); 1517 1518 const MCFragment &Frag = *std::next(IT); 1519 if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data) 1520 report_fatal_error("only data supported in .init_array section"); 1521 1522 uint16_t Priority = UINT16_MAX; 1523 unsigned PrefixLength = strlen(".init_array"); 1524 if (WS.getSectionName().size() > PrefixLength) { 1525 if (WS.getSectionName()[PrefixLength] != '.') 1526 report_fatal_error( 1527 ".init_array section priority should start with '.'"); 1528 if (WS.getSectionName() 1529 .substr(PrefixLength + 1) 1530 .getAsInteger(10, Priority)) 1531 report_fatal_error("invalid .init_array section priority"); 1532 } 1533 const auto &DataFrag = cast<MCDataFragment>(Frag); 1534 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 1535 for (const uint8_t * 1536 P = (const uint8_t *)Contents.data(), 1537 *End = (const uint8_t *)Contents.data() + Contents.size(); 1538 P != End; ++P) { 1539 if (*P != 0) 1540 report_fatal_error("non-symbolic data in .init_array section"); 1541 } 1542 for (const MCFixup &Fixup : DataFrag.getFixups()) { 1543 assert(Fixup.getKind() == 1544 MCFixup::getKindForSize(is64Bit() ? 8 : 4, false)); 1545 const MCExpr *Expr = Fixup.getValue(); 1546 auto *SymRef = dyn_cast<MCSymbolRefExpr>(Expr); 1547 if (!SymRef) 1548 report_fatal_error("fixups in .init_array should be symbol references"); 1549 const auto &TargetSym = cast<const MCSymbolWasm>(SymRef->getSymbol()); 1550 if (TargetSym.getIndex() == InvalidIndex) 1551 report_fatal_error("symbols in .init_array should exist in symbtab"); 1552 if (!TargetSym.isFunction()) 1553 report_fatal_error("symbols in .init_array should be for functions"); 1554 InitFuncs.push_back( 1555 std::make_pair(Priority, TargetSym.getIndex())); 1556 } 1557 } 1558 1559 // Write out the Wasm header. 1560 writeHeader(Asm); 1561 1562 writeTypeSection(Signatures); 1563 writeImportSection(Imports, DataSize, TableElems.size()); 1564 writeFunctionSection(Functions); 1565 // Skip the "table" section; we import the table instead. 1566 // Skip the "memory" section; we import the memory instead. 1567 writeEventSection(Events); 1568 writeExportSection(Exports); 1569 writeElemSection(TableElems); 1570 writeDataCountSection(); 1571 writeCodeSection(Asm, Layout, Functions); 1572 writeDataSection(); 1573 for (auto &CustomSection : CustomSections) 1574 writeCustomSection(CustomSection, Asm, Layout); 1575 writeLinkingMetaDataSection(SymbolInfos, InitFuncs, Comdats); 1576 writeRelocSection(CodeSectionIndex, "CODE", CodeRelocations); 1577 writeRelocSection(DataSectionIndex, "DATA", DataRelocations); 1578 writeCustomRelocSections(); 1579 if (ProducersSection) 1580 writeCustomSection(*ProducersSection, Asm, Layout); 1581 if (TargetFeaturesSection) 1582 writeCustomSection(*TargetFeaturesSection, Asm, Layout); 1583 1584 // TODO: Translate the .comment section to the output. 1585 return W.OS.tell() - StartOffset; 1586 } 1587 1588 std::unique_ptr<MCObjectWriter> 1589 llvm::createWasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 1590 raw_pwrite_stream &OS) { 1591 return std::make_unique<WasmObjectWriter>(std::move(MOTW), OS); 1592 } 1593