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/BinaryFormat/WasmTraits.h" 17 #include "llvm/Config/llvm-config.h" 18 #include "llvm/MC/MCAsmBackend.h" 19 #include "llvm/MC/MCAsmLayout.h" 20 #include "llvm/MC/MCAssembler.h" 21 #include "llvm/MC/MCContext.h" 22 #include "llvm/MC/MCExpr.h" 23 #include "llvm/MC/MCFixupKindInfo.h" 24 #include "llvm/MC/MCObjectWriter.h" 25 #include "llvm/MC/MCSectionWasm.h" 26 #include "llvm/MC/MCSymbolWasm.h" 27 #include "llvm/MC/MCValue.h" 28 #include "llvm/MC/MCWasmObjectWriter.h" 29 #include "llvm/Support/Casting.h" 30 #include "llvm/Support/Debug.h" 31 #include "llvm/Support/EndianStream.h" 32 #include "llvm/Support/ErrorHandling.h" 33 #include "llvm/Support/LEB128.h" 34 #include "llvm/Support/StringSaver.h" 35 #include <vector> 36 37 using namespace llvm; 38 39 #define DEBUG_TYPE "mc" 40 41 namespace { 42 43 // When we create the indirect function table we start at 1, so that there is 44 // and empty slot at 0 and therefore calling a null function pointer will trap. 45 static const uint32_t InitialTableOffset = 1; 46 47 // For patching purposes, we need to remember where each section starts, both 48 // for patching up the section size field, and for patching up references to 49 // locations within the section. 50 struct SectionBookkeeping { 51 // Where the size of the section is written. 52 uint64_t SizeOffset; 53 // Where the section header ends (without custom section name). 54 uint64_t PayloadOffset; 55 // Where the contents of the section starts. 56 uint64_t ContentsOffset; 57 uint32_t Index; 58 }; 59 60 // A wasm data segment. A wasm binary contains only a single data section 61 // but that can contain many segments, each with their own virtual location 62 // in memory. Each MCSection data created by llvm is modeled as its own 63 // wasm data segment. 64 struct WasmDataSegment { 65 MCSectionWasm *Section; 66 StringRef Name; 67 uint32_t InitFlags; 68 uint64_t Offset; 69 uint32_t Alignment; 70 uint32_t LinkerFlags; 71 SmallVector<char, 4> Data; 72 }; 73 74 // A wasm function to be written into the function section. 75 struct WasmFunction { 76 uint32_t SigIndex; 77 const MCSymbolWasm *Sym; 78 }; 79 80 // A wasm global to be written into the global section. 81 struct WasmGlobal { 82 wasm::WasmGlobalType Type; 83 uint64_t InitialValue; 84 }; 85 86 // Information about a single item which is part of a COMDAT. For each data 87 // segment or function which is in the COMDAT, there is a corresponding 88 // WasmComdatEntry. 89 struct WasmComdatEntry { 90 unsigned Kind; 91 uint32_t Index; 92 }; 93 94 // Information about a single relocation. 95 struct WasmRelocationEntry { 96 uint64_t Offset; // Where is the relocation. 97 const MCSymbolWasm *Symbol; // The symbol to relocate with. 98 int64_t Addend; // A value to add to the symbol. 99 unsigned Type; // The type of the relocation. 100 const MCSectionWasm *FixupSection; // The section the relocation is targeting. 101 102 WasmRelocationEntry(uint64_t Offset, const MCSymbolWasm *Symbol, 103 int64_t Addend, unsigned Type, 104 const MCSectionWasm *FixupSection) 105 : Offset(Offset), Symbol(Symbol), Addend(Addend), Type(Type), 106 FixupSection(FixupSection) {} 107 108 bool hasAddend() const { return wasm::relocTypeHasAddend(Type); } 109 110 void print(raw_ostream &Out) const { 111 Out << wasm::relocTypetoString(Type) << " Off=" << Offset 112 << ", Sym=" << *Symbol << ", Addend=" << Addend 113 << ", FixupSection=" << FixupSection->getName(); 114 } 115 116 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 117 LLVM_DUMP_METHOD void dump() const { print(dbgs()); } 118 #endif 119 }; 120 121 static const uint32_t InvalidIndex = -1; 122 123 struct WasmCustomSection { 124 125 StringRef Name; 126 MCSectionWasm *Section; 127 128 uint32_t OutputContentsOffset; 129 uint32_t OutputIndex; 130 131 WasmCustomSection(StringRef Name, MCSectionWasm *Section) 132 : Name(Name), Section(Section), OutputContentsOffset(0), 133 OutputIndex(InvalidIndex) {} 134 }; 135 136 #if !defined(NDEBUG) 137 raw_ostream &operator<<(raw_ostream &OS, const WasmRelocationEntry &Rel) { 138 Rel.print(OS); 139 return OS; 140 } 141 #endif 142 143 // Write X as an (unsigned) LEB value at offset Offset in Stream, padded 144 // to allow patching. 145 template <int W> 146 void writePatchableLEB(raw_pwrite_stream &Stream, uint64_t X, uint64_t Offset) { 147 uint8_t Buffer[W]; 148 unsigned SizeLen = encodeULEB128(X, Buffer, W); 149 assert(SizeLen == W); 150 Stream.pwrite((char *)Buffer, SizeLen, Offset); 151 } 152 153 // Write X as an signed LEB value at offset Offset in Stream, padded 154 // to allow patching. 155 template <int W> 156 void writePatchableSLEB(raw_pwrite_stream &Stream, int64_t X, uint64_t Offset) { 157 uint8_t Buffer[W]; 158 unsigned SizeLen = encodeSLEB128(X, Buffer, W); 159 assert(SizeLen == W); 160 Stream.pwrite((char *)Buffer, SizeLen, Offset); 161 } 162 163 // Write X as a plain integer value at offset Offset in Stream. 164 static void patchI32(raw_pwrite_stream &Stream, uint32_t X, uint64_t Offset) { 165 uint8_t Buffer[4]; 166 support::endian::write32le(Buffer, X); 167 Stream.pwrite((char *)Buffer, sizeof(Buffer), Offset); 168 } 169 170 static void patchI64(raw_pwrite_stream &Stream, uint64_t X, uint64_t Offset) { 171 uint8_t Buffer[8]; 172 support::endian::write64le(Buffer, X); 173 Stream.pwrite((char *)Buffer, sizeof(Buffer), Offset); 174 } 175 176 bool isDwoSection(const MCSection &Sec) { 177 return Sec.getName().endswith(".dwo"); 178 } 179 180 class WasmObjectWriter : public MCObjectWriter { 181 support::endian::Writer *W; 182 183 /// The target specific Wasm writer instance. 184 std::unique_ptr<MCWasmObjectTargetWriter> TargetObjectWriter; 185 186 // Relocations for fixing up references in the code section. 187 std::vector<WasmRelocationEntry> CodeRelocations; 188 // Relocations for fixing up references in the data section. 189 std::vector<WasmRelocationEntry> DataRelocations; 190 191 // Index values to use for fixing up call_indirect type indices. 192 // Maps function symbols to the index of the type of the function 193 DenseMap<const MCSymbolWasm *, uint32_t> TypeIndices; 194 // Maps function symbols to the table element index space. Used 195 // for TABLE_INDEX relocation types (i.e. address taken functions). 196 DenseMap<const MCSymbolWasm *, uint32_t> TableIndices; 197 // Maps function/global/table symbols to the 198 // function/global/table/event/section index space. 199 DenseMap<const MCSymbolWasm *, uint32_t> WasmIndices; 200 DenseMap<const MCSymbolWasm *, uint32_t> GOTIndices; 201 // Maps data symbols to the Wasm segment and offset/size with the segment. 202 DenseMap<const MCSymbolWasm *, wasm::WasmDataReference> DataLocations; 203 204 // Stores output data (index, relocations, content offset) for custom 205 // section. 206 std::vector<WasmCustomSection> CustomSections; 207 std::unique_ptr<WasmCustomSection> ProducersSection; 208 std::unique_ptr<WasmCustomSection> TargetFeaturesSection; 209 // Relocations for fixing up references in the custom sections. 210 DenseMap<const MCSectionWasm *, std::vector<WasmRelocationEntry>> 211 CustomSectionsRelocations; 212 213 // Map from section to defining function symbol. 214 DenseMap<const MCSection *, const MCSymbol *> SectionFunctions; 215 216 DenseMap<wasm::WasmSignature, uint32_t> SignatureIndices; 217 SmallVector<wasm::WasmSignature, 4> Signatures; 218 SmallVector<WasmDataSegment, 4> DataSegments; 219 unsigned NumFunctionImports = 0; 220 unsigned NumGlobalImports = 0; 221 unsigned NumTableImports = 0; 222 unsigned NumEventImports = 0; 223 uint32_t SectionCount = 0; 224 225 enum class DwoMode { 226 AllSections, 227 NonDwoOnly, 228 DwoOnly, 229 }; 230 bool IsSplitDwarf = false; 231 raw_pwrite_stream *OS = nullptr; 232 raw_pwrite_stream *DwoOS = nullptr; 233 234 // TargetObjectWriter wranppers. 235 bool is64Bit() const { return TargetObjectWriter->is64Bit(); } 236 bool isEmscripten() const { return TargetObjectWriter->isEmscripten(); } 237 238 void startSection(SectionBookkeeping &Section, unsigned SectionId); 239 void startCustomSection(SectionBookkeeping &Section, StringRef Name); 240 void endSection(SectionBookkeeping &Section); 241 242 public: 243 WasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 244 raw_pwrite_stream &OS_) 245 : TargetObjectWriter(std::move(MOTW)), OS(&OS_) {} 246 247 WasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 248 raw_pwrite_stream &OS_, raw_pwrite_stream &DwoOS_) 249 : TargetObjectWriter(std::move(MOTW)), IsSplitDwarf(true), OS(&OS_), 250 DwoOS(&DwoOS_) {} 251 252 private: 253 void reset() override { 254 CodeRelocations.clear(); 255 DataRelocations.clear(); 256 TypeIndices.clear(); 257 WasmIndices.clear(); 258 GOTIndices.clear(); 259 TableIndices.clear(); 260 DataLocations.clear(); 261 CustomSections.clear(); 262 ProducersSection.reset(); 263 TargetFeaturesSection.reset(); 264 CustomSectionsRelocations.clear(); 265 SignatureIndices.clear(); 266 Signatures.clear(); 267 DataSegments.clear(); 268 SectionFunctions.clear(); 269 NumFunctionImports = 0; 270 NumGlobalImports = 0; 271 NumTableImports = 0; 272 MCObjectWriter::reset(); 273 } 274 275 void writeHeader(const MCAssembler &Asm); 276 277 void recordRelocation(MCAssembler &Asm, const MCAsmLayout &Layout, 278 const MCFragment *Fragment, const MCFixup &Fixup, 279 MCValue Target, uint64_t &FixedValue) override; 280 281 void executePostLayoutBinding(MCAssembler &Asm, 282 const MCAsmLayout &Layout) override; 283 void prepareImports(SmallVectorImpl<wasm::WasmImport> &Imports, 284 MCAssembler &Asm, const MCAsmLayout &Layout); 285 uint64_t writeObject(MCAssembler &Asm, const MCAsmLayout &Layout) override; 286 287 uint64_t writeOneObject(MCAssembler &Asm, const MCAsmLayout &Layout, 288 DwoMode Mode); 289 290 void writeString(const StringRef Str) { 291 encodeULEB128(Str.size(), W->OS); 292 W->OS << Str; 293 } 294 295 void writeI32(int32_t val) { 296 char Buffer[4]; 297 support::endian::write32le(Buffer, val); 298 W->OS.write(Buffer, sizeof(Buffer)); 299 } 300 301 void writeI64(int64_t val) { 302 char Buffer[8]; 303 support::endian::write64le(Buffer, val); 304 W->OS.write(Buffer, sizeof(Buffer)); 305 } 306 307 void writeValueType(wasm::ValType Ty) { W->OS << static_cast<char>(Ty); } 308 309 void writeTypeSection(ArrayRef<wasm::WasmSignature> Signatures); 310 void writeImportSection(ArrayRef<wasm::WasmImport> Imports, uint64_t DataSize, 311 uint32_t NumElements); 312 void writeFunctionSection(ArrayRef<WasmFunction> Functions); 313 void writeExportSection(ArrayRef<wasm::WasmExport> Exports); 314 void writeElemSection(const MCSymbolWasm *IndirectFunctionTable, 315 ArrayRef<uint32_t> TableElems); 316 void writeDataCountSection(); 317 uint32_t writeCodeSection(const MCAssembler &Asm, const MCAsmLayout &Layout, 318 ArrayRef<WasmFunction> Functions); 319 uint32_t writeDataSection(const MCAsmLayout &Layout); 320 void writeEventSection(ArrayRef<wasm::WasmEventType> Events); 321 void writeGlobalSection(ArrayRef<wasm::WasmGlobal> Globals); 322 void writeTableSection(ArrayRef<wasm::WasmTable> Tables); 323 void writeRelocSection(uint32_t SectionIndex, StringRef Name, 324 std::vector<WasmRelocationEntry> &Relocations); 325 void writeLinkingMetaDataSection( 326 ArrayRef<wasm::WasmSymbolInfo> SymbolInfos, 327 ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs, 328 const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats); 329 void writeCustomSection(WasmCustomSection &CustomSection, 330 const MCAssembler &Asm, const MCAsmLayout &Layout); 331 void writeCustomRelocSections(); 332 333 uint64_t getProvisionalValue(const WasmRelocationEntry &RelEntry, 334 const MCAsmLayout &Layout); 335 void applyRelocations(ArrayRef<WasmRelocationEntry> Relocations, 336 uint64_t ContentsOffset, const MCAsmLayout &Layout); 337 338 uint32_t getRelocationIndexValue(const WasmRelocationEntry &RelEntry); 339 uint32_t getFunctionType(const MCSymbolWasm &Symbol); 340 uint32_t getEventType(const MCSymbolWasm &Symbol); 341 void registerFunctionType(const MCSymbolWasm &Symbol); 342 void registerEventType(const MCSymbolWasm &Symbol); 343 }; 344 345 } // end anonymous namespace 346 347 // Write out a section header and a patchable section size field. 348 void WasmObjectWriter::startSection(SectionBookkeeping &Section, 349 unsigned SectionId) { 350 LLVM_DEBUG(dbgs() << "startSection " << SectionId << "\n"); 351 W->OS << char(SectionId); 352 353 Section.SizeOffset = W->OS.tell(); 354 355 // The section size. We don't know the size yet, so reserve enough space 356 // for any 32-bit value; we'll patch it later. 357 encodeULEB128(0, W->OS, 5); 358 359 // The position where the section starts, for measuring its size. 360 Section.ContentsOffset = W->OS.tell(); 361 Section.PayloadOffset = W->OS.tell(); 362 Section.Index = SectionCount++; 363 } 364 365 void WasmObjectWriter::startCustomSection(SectionBookkeeping &Section, 366 StringRef Name) { 367 LLVM_DEBUG(dbgs() << "startCustomSection " << Name << "\n"); 368 startSection(Section, wasm::WASM_SEC_CUSTOM); 369 370 // The position where the section header ends, for measuring its size. 371 Section.PayloadOffset = W->OS.tell(); 372 373 // Custom sections in wasm also have a string identifier. 374 writeString(Name); 375 376 // The position where the custom section starts. 377 Section.ContentsOffset = W->OS.tell(); 378 } 379 380 // Now that the section is complete and we know how big it is, patch up the 381 // section size field at the start of the section. 382 void WasmObjectWriter::endSection(SectionBookkeeping &Section) { 383 uint64_t Size = W->OS.tell(); 384 // /dev/null doesn't support seek/tell and can report offset of 0. 385 // Simply skip this patching in that case. 386 if (!Size) 387 return; 388 389 Size -= Section.PayloadOffset; 390 if (uint32_t(Size) != Size) 391 report_fatal_error("section size does not fit in a uint32_t"); 392 393 LLVM_DEBUG(dbgs() << "endSection size=" << Size << "\n"); 394 395 // Write the final section size to the payload_len field, which follows 396 // the section id byte. 397 writePatchableLEB<5>(static_cast<raw_pwrite_stream &>(W->OS), Size, 398 Section.SizeOffset); 399 } 400 401 // Emit the Wasm header. 402 void WasmObjectWriter::writeHeader(const MCAssembler &Asm) { 403 W->OS.write(wasm::WasmMagic, sizeof(wasm::WasmMagic)); 404 W->write<uint32_t>(wasm::WasmVersion); 405 } 406 407 void WasmObjectWriter::executePostLayoutBinding(MCAssembler &Asm, 408 const MCAsmLayout &Layout) { 409 // Some compilation units require the indirect function table to be present 410 // but don't explicitly reference it. This is the case for call_indirect 411 // without the reference-types feature, and also function bitcasts in all 412 // cases. In those cases the __indirect_function_table has the 413 // WASM_SYMBOL_NO_STRIP attribute. Here we make sure this symbol makes it to 414 // the assembler, if needed. 415 if (auto *Sym = Asm.getContext().lookupSymbol("__indirect_function_table")) { 416 const auto *WasmSym = static_cast<const MCSymbolWasm *>(Sym); 417 if (WasmSym->isNoStrip()) 418 Asm.registerSymbol(*Sym); 419 } 420 421 // Build a map of sections to the function that defines them, for use 422 // in recordRelocation. 423 for (const MCSymbol &S : Asm.symbols()) { 424 const auto &WS = static_cast<const MCSymbolWasm &>(S); 425 if (WS.isDefined() && WS.isFunction() && !WS.isVariable()) { 426 const auto &Sec = static_cast<const MCSectionWasm &>(S.getSection()); 427 auto Pair = SectionFunctions.insert(std::make_pair(&Sec, &S)); 428 if (!Pair.second) 429 report_fatal_error("section already has a defining function: " + 430 Sec.getName()); 431 } 432 } 433 } 434 435 void WasmObjectWriter::recordRelocation(MCAssembler &Asm, 436 const MCAsmLayout &Layout, 437 const MCFragment *Fragment, 438 const MCFixup &Fixup, MCValue Target, 439 uint64_t &FixedValue) { 440 // The WebAssembly backend should never generate FKF_IsPCRel fixups 441 assert(!(Asm.getBackend().getFixupKindInfo(Fixup.getKind()).Flags & 442 MCFixupKindInfo::FKF_IsPCRel)); 443 444 const auto &FixupSection = cast<MCSectionWasm>(*Fragment->getParent()); 445 uint64_t C = Target.getConstant(); 446 uint64_t FixupOffset = Layout.getFragmentOffset(Fragment) + Fixup.getOffset(); 447 MCContext &Ctx = Asm.getContext(); 448 449 if (const MCSymbolRefExpr *RefB = Target.getSymB()) { 450 // To get here the A - B expression must have failed evaluateAsRelocatable. 451 // This means either A or B must be undefined and in WebAssembly we can't 452 // support either of those cases. 453 const auto &SymB = cast<MCSymbolWasm>(RefB->getSymbol()); 454 Ctx.reportError( 455 Fixup.getLoc(), 456 Twine("symbol '") + SymB.getName() + 457 "': unsupported subtraction expression used in relocation."); 458 return; 459 } 460 461 // We either rejected the fixup or folded B into C at this point. 462 const MCSymbolRefExpr *RefA = Target.getSymA(); 463 const auto *SymA = cast<MCSymbolWasm>(&RefA->getSymbol()); 464 465 // The .init_array isn't translated as data, so don't do relocations in it. 466 if (FixupSection.getName().startswith(".init_array")) { 467 SymA->setUsedInInitArray(); 468 return; 469 } 470 471 if (SymA->isVariable()) { 472 const MCExpr *Expr = SymA->getVariableValue(); 473 if (const auto *Inner = dyn_cast<MCSymbolRefExpr>(Expr)) 474 if (Inner->getKind() == MCSymbolRefExpr::VK_WEAKREF) 475 llvm_unreachable("weakref used in reloc not yet implemented"); 476 } 477 478 // Put any constant offset in an addend. Offsets can be negative, and 479 // LLVM expects wrapping, in contrast to wasm's immediates which can't 480 // be negative and don't wrap. 481 FixedValue = 0; 482 483 unsigned Type = TargetObjectWriter->getRelocType(Target, Fixup); 484 485 // Absolute offset within a section or a function. 486 // Currently only supported for for metadata sections. 487 // See: test/MC/WebAssembly/blockaddress.ll 488 if (Type == wasm::R_WASM_FUNCTION_OFFSET_I32 || 489 Type == wasm::R_WASM_FUNCTION_OFFSET_I64 || 490 Type == wasm::R_WASM_SECTION_OFFSET_I32) { 491 if (!FixupSection.getKind().isMetadata()) 492 report_fatal_error("relocations for function or section offsets are " 493 "only supported in metadata sections"); 494 495 const MCSymbol *SectionSymbol = nullptr; 496 const MCSection &SecA = SymA->getSection(); 497 if (SecA.getKind().isText()) { 498 auto SecSymIt = SectionFunctions.find(&SecA); 499 if (SecSymIt == SectionFunctions.end()) 500 report_fatal_error("section doesn\'t have defining symbol"); 501 SectionSymbol = SecSymIt->second; 502 } else { 503 SectionSymbol = SecA.getBeginSymbol(); 504 } 505 if (!SectionSymbol) 506 report_fatal_error("section symbol is required for relocation"); 507 508 C += Layout.getSymbolOffset(*SymA); 509 SymA = cast<MCSymbolWasm>(SectionSymbol); 510 } 511 512 if (Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB || 513 Type == wasm::R_WASM_TABLE_INDEX_SLEB || 514 Type == wasm::R_WASM_TABLE_INDEX_SLEB64 || 515 Type == wasm::R_WASM_TABLE_INDEX_I32 || 516 Type == wasm::R_WASM_TABLE_INDEX_I64) { 517 // TABLE_INDEX relocs implicitly use the default indirect function table. 518 // We require the function table to have already been defined. 519 auto TableName = "__indirect_function_table"; 520 MCSymbolWasm *Sym = cast_or_null<MCSymbolWasm>(Ctx.lookupSymbol(TableName)); 521 if (!Sym || !Sym->isFunctionTable()) { 522 Ctx.reportError( 523 Fixup.getLoc(), 524 "symbol '__indirect_function_table' is not a function table"); 525 } else { 526 // Ensure that __indirect_function_table reaches the output. 527 Sym->setNoStrip(); 528 Asm.registerSymbol(*Sym); 529 } 530 } 531 532 // Relocation other than R_WASM_TYPE_INDEX_LEB are required to be 533 // against a named symbol. 534 if (Type != wasm::R_WASM_TYPE_INDEX_LEB) { 535 if (SymA->getName().empty()) 536 report_fatal_error("relocations against un-named temporaries are not yet " 537 "supported by wasm"); 538 539 SymA->setUsedInReloc(); 540 } 541 542 if (RefA->getKind() == MCSymbolRefExpr::VK_GOT) 543 SymA->setUsedInGOT(); 544 545 WasmRelocationEntry Rec(FixupOffset, SymA, C, Type, &FixupSection); 546 LLVM_DEBUG(dbgs() << "WasmReloc: " << Rec << "\n"); 547 548 if (FixupSection.isWasmData()) { 549 DataRelocations.push_back(Rec); 550 } else if (FixupSection.getKind().isText()) { 551 CodeRelocations.push_back(Rec); 552 } else if (FixupSection.getKind().isMetadata()) { 553 CustomSectionsRelocations[&FixupSection].push_back(Rec); 554 } else { 555 llvm_unreachable("unexpected section type"); 556 } 557 } 558 559 // Compute a value to write into the code at the location covered 560 // by RelEntry. This value isn't used by the static linker; it just serves 561 // to make the object format more readable and more likely to be directly 562 // useable. 563 uint64_t 564 WasmObjectWriter::getProvisionalValue(const WasmRelocationEntry &RelEntry, 565 const MCAsmLayout &Layout) { 566 if ((RelEntry.Type == wasm::R_WASM_GLOBAL_INDEX_LEB || 567 RelEntry.Type == wasm::R_WASM_GLOBAL_INDEX_I32) && 568 !RelEntry.Symbol->isGlobal()) { 569 assert(GOTIndices.count(RelEntry.Symbol) > 0 && "symbol not found in GOT index space"); 570 return GOTIndices[RelEntry.Symbol]; 571 } 572 573 switch (RelEntry.Type) { 574 case wasm::R_WASM_TABLE_INDEX_REL_SLEB: 575 case wasm::R_WASM_TABLE_INDEX_SLEB: 576 case wasm::R_WASM_TABLE_INDEX_SLEB64: 577 case wasm::R_WASM_TABLE_INDEX_I32: 578 case wasm::R_WASM_TABLE_INDEX_I64: { 579 // Provisional value is table address of the resolved symbol itself 580 const MCSymbolWasm *Base = 581 cast<MCSymbolWasm>(Layout.getBaseSymbol(*RelEntry.Symbol)); 582 assert(Base->isFunction()); 583 if (RelEntry.Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB) 584 return TableIndices[Base] - InitialTableOffset; 585 else 586 return TableIndices[Base]; 587 } 588 case wasm::R_WASM_TYPE_INDEX_LEB: 589 // Provisional value is same as the index 590 return getRelocationIndexValue(RelEntry); 591 case wasm::R_WASM_FUNCTION_INDEX_LEB: 592 case wasm::R_WASM_GLOBAL_INDEX_LEB: 593 case wasm::R_WASM_GLOBAL_INDEX_I32: 594 case wasm::R_WASM_EVENT_INDEX_LEB: 595 case wasm::R_WASM_TABLE_NUMBER_LEB: 596 // Provisional value is function/global/event Wasm index 597 assert(WasmIndices.count(RelEntry.Symbol) > 0 && "symbol not found in wasm index space"); 598 return WasmIndices[RelEntry.Symbol]; 599 case wasm::R_WASM_FUNCTION_OFFSET_I32: 600 case wasm::R_WASM_FUNCTION_OFFSET_I64: 601 case wasm::R_WASM_SECTION_OFFSET_I32: { 602 const auto &Section = 603 static_cast<const MCSectionWasm &>(RelEntry.Symbol->getSection()); 604 return Section.getSectionOffset() + RelEntry.Addend; 605 } 606 case wasm::R_WASM_MEMORY_ADDR_LEB: 607 case wasm::R_WASM_MEMORY_ADDR_LEB64: 608 case wasm::R_WASM_MEMORY_ADDR_SLEB: 609 case wasm::R_WASM_MEMORY_ADDR_SLEB64: 610 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB: 611 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB64: 612 case wasm::R_WASM_MEMORY_ADDR_I32: 613 case wasm::R_WASM_MEMORY_ADDR_I64: 614 case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB: { 615 // Provisional value is address of the global plus the offset 616 // For undefined symbols, use zero 617 if (!RelEntry.Symbol->isDefined()) 618 return 0; 619 const wasm::WasmDataReference &SymRef = DataLocations[RelEntry.Symbol]; 620 const WasmDataSegment &Segment = DataSegments[SymRef.Segment]; 621 // Ignore overflow. LLVM allows address arithmetic to silently wrap. 622 return Segment.Offset + SymRef.Offset + RelEntry.Addend; 623 } 624 default: 625 llvm_unreachable("invalid relocation type"); 626 } 627 } 628 629 static void addData(SmallVectorImpl<char> &DataBytes, 630 MCSectionWasm &DataSection) { 631 LLVM_DEBUG(errs() << "addData: " << DataSection.getName() << "\n"); 632 633 DataBytes.resize(alignTo(DataBytes.size(), DataSection.getAlignment())); 634 635 for (const MCFragment &Frag : DataSection) { 636 if (Frag.hasInstructions()) 637 report_fatal_error("only data supported in data sections"); 638 639 if (auto *Align = dyn_cast<MCAlignFragment>(&Frag)) { 640 if (Align->getValueSize() != 1) 641 report_fatal_error("only byte values supported for alignment"); 642 // If nops are requested, use zeros, as this is the data section. 643 uint8_t Value = Align->hasEmitNops() ? 0 : Align->getValue(); 644 uint64_t Size = 645 std::min<uint64_t>(alignTo(DataBytes.size(), Align->getAlignment()), 646 DataBytes.size() + Align->getMaxBytesToEmit()); 647 DataBytes.resize(Size, Value); 648 } else if (auto *Fill = dyn_cast<MCFillFragment>(&Frag)) { 649 int64_t NumValues; 650 if (!Fill->getNumValues().evaluateAsAbsolute(NumValues)) 651 llvm_unreachable("The fill should be an assembler constant"); 652 DataBytes.insert(DataBytes.end(), Fill->getValueSize() * NumValues, 653 Fill->getValue()); 654 } else if (auto *LEB = dyn_cast<MCLEBFragment>(&Frag)) { 655 const SmallVectorImpl<char> &Contents = LEB->getContents(); 656 llvm::append_range(DataBytes, Contents); 657 } else { 658 const auto &DataFrag = cast<MCDataFragment>(Frag); 659 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 660 llvm::append_range(DataBytes, Contents); 661 } 662 } 663 664 LLVM_DEBUG(dbgs() << "addData -> " << DataBytes.size() << "\n"); 665 } 666 667 uint32_t 668 WasmObjectWriter::getRelocationIndexValue(const WasmRelocationEntry &RelEntry) { 669 if (RelEntry.Type == wasm::R_WASM_TYPE_INDEX_LEB) { 670 if (!TypeIndices.count(RelEntry.Symbol)) 671 report_fatal_error("symbol not found in type index space: " + 672 RelEntry.Symbol->getName()); 673 return TypeIndices[RelEntry.Symbol]; 674 } 675 676 return RelEntry.Symbol->getIndex(); 677 } 678 679 // Apply the portions of the relocation records that we can handle ourselves 680 // directly. 681 void WasmObjectWriter::applyRelocations( 682 ArrayRef<WasmRelocationEntry> Relocations, uint64_t ContentsOffset, 683 const MCAsmLayout &Layout) { 684 auto &Stream = static_cast<raw_pwrite_stream &>(W->OS); 685 for (const WasmRelocationEntry &RelEntry : Relocations) { 686 uint64_t Offset = ContentsOffset + 687 RelEntry.FixupSection->getSectionOffset() + 688 RelEntry.Offset; 689 690 LLVM_DEBUG(dbgs() << "applyRelocation: " << RelEntry << "\n"); 691 auto Value = getProvisionalValue(RelEntry, Layout); 692 693 switch (RelEntry.Type) { 694 case wasm::R_WASM_FUNCTION_INDEX_LEB: 695 case wasm::R_WASM_TYPE_INDEX_LEB: 696 case wasm::R_WASM_GLOBAL_INDEX_LEB: 697 case wasm::R_WASM_MEMORY_ADDR_LEB: 698 case wasm::R_WASM_EVENT_INDEX_LEB: 699 case wasm::R_WASM_TABLE_NUMBER_LEB: 700 writePatchableLEB<5>(Stream, Value, Offset); 701 break; 702 case wasm::R_WASM_MEMORY_ADDR_LEB64: 703 writePatchableLEB<10>(Stream, Value, Offset); 704 break; 705 case wasm::R_WASM_TABLE_INDEX_I32: 706 case wasm::R_WASM_MEMORY_ADDR_I32: 707 case wasm::R_WASM_FUNCTION_OFFSET_I32: 708 case wasm::R_WASM_SECTION_OFFSET_I32: 709 case wasm::R_WASM_GLOBAL_INDEX_I32: 710 patchI32(Stream, Value, Offset); 711 break; 712 case wasm::R_WASM_TABLE_INDEX_I64: 713 case wasm::R_WASM_MEMORY_ADDR_I64: 714 case wasm::R_WASM_FUNCTION_OFFSET_I64: 715 patchI64(Stream, Value, Offset); 716 break; 717 case wasm::R_WASM_TABLE_INDEX_SLEB: 718 case wasm::R_WASM_TABLE_INDEX_REL_SLEB: 719 case wasm::R_WASM_MEMORY_ADDR_SLEB: 720 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB: 721 case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB: 722 writePatchableSLEB<5>(Stream, Value, Offset); 723 break; 724 case wasm::R_WASM_TABLE_INDEX_SLEB64: 725 case wasm::R_WASM_MEMORY_ADDR_SLEB64: 726 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB64: 727 writePatchableSLEB<10>(Stream, Value, Offset); 728 break; 729 default: 730 llvm_unreachable("invalid relocation type"); 731 } 732 } 733 } 734 735 void WasmObjectWriter::writeTypeSection( 736 ArrayRef<wasm::WasmSignature> Signatures) { 737 if (Signatures.empty()) 738 return; 739 740 SectionBookkeeping Section; 741 startSection(Section, wasm::WASM_SEC_TYPE); 742 743 encodeULEB128(Signatures.size(), W->OS); 744 745 for (const wasm::WasmSignature &Sig : Signatures) { 746 W->OS << char(wasm::WASM_TYPE_FUNC); 747 encodeULEB128(Sig.Params.size(), W->OS); 748 for (wasm::ValType Ty : Sig.Params) 749 writeValueType(Ty); 750 encodeULEB128(Sig.Returns.size(), W->OS); 751 for (wasm::ValType Ty : Sig.Returns) 752 writeValueType(Ty); 753 } 754 755 endSection(Section); 756 } 757 758 void WasmObjectWriter::writeImportSection(ArrayRef<wasm::WasmImport> Imports, 759 uint64_t DataSize, 760 uint32_t NumElements) { 761 if (Imports.empty()) 762 return; 763 764 uint64_t NumPages = (DataSize + wasm::WasmPageSize - 1) / wasm::WasmPageSize; 765 766 SectionBookkeeping Section; 767 startSection(Section, wasm::WASM_SEC_IMPORT); 768 769 encodeULEB128(Imports.size(), W->OS); 770 for (const wasm::WasmImport &Import : Imports) { 771 writeString(Import.Module); 772 writeString(Import.Field); 773 W->OS << char(Import.Kind); 774 775 switch (Import.Kind) { 776 case wasm::WASM_EXTERNAL_FUNCTION: 777 encodeULEB128(Import.SigIndex, W->OS); 778 break; 779 case wasm::WASM_EXTERNAL_GLOBAL: 780 W->OS << char(Import.Global.Type); 781 W->OS << char(Import.Global.Mutable ? 1 : 0); 782 break; 783 case wasm::WASM_EXTERNAL_MEMORY: 784 encodeULEB128(Import.Memory.Flags, W->OS); 785 encodeULEB128(NumPages, W->OS); // initial 786 break; 787 case wasm::WASM_EXTERNAL_TABLE: 788 W->OS << char(Import.Table.ElemType); 789 encodeULEB128(0, W->OS); // flags 790 encodeULEB128(NumElements, W->OS); // initial 791 break; 792 case wasm::WASM_EXTERNAL_EVENT: 793 encodeULEB128(Import.Event.Attribute, W->OS); 794 encodeULEB128(Import.Event.SigIndex, W->OS); 795 break; 796 default: 797 llvm_unreachable("unsupported import kind"); 798 } 799 } 800 801 endSection(Section); 802 } 803 804 void WasmObjectWriter::writeFunctionSection(ArrayRef<WasmFunction> Functions) { 805 if (Functions.empty()) 806 return; 807 808 SectionBookkeeping Section; 809 startSection(Section, wasm::WASM_SEC_FUNCTION); 810 811 encodeULEB128(Functions.size(), W->OS); 812 for (const WasmFunction &Func : Functions) 813 encodeULEB128(Func.SigIndex, W->OS); 814 815 endSection(Section); 816 } 817 818 void WasmObjectWriter::writeEventSection(ArrayRef<wasm::WasmEventType> Events) { 819 if (Events.empty()) 820 return; 821 822 SectionBookkeeping Section; 823 startSection(Section, wasm::WASM_SEC_EVENT); 824 825 encodeULEB128(Events.size(), W->OS); 826 for (const wasm::WasmEventType &Event : Events) { 827 encodeULEB128(Event.Attribute, W->OS); 828 encodeULEB128(Event.SigIndex, W->OS); 829 } 830 831 endSection(Section); 832 } 833 834 void WasmObjectWriter::writeGlobalSection(ArrayRef<wasm::WasmGlobal> Globals) { 835 if (Globals.empty()) 836 return; 837 838 SectionBookkeeping Section; 839 startSection(Section, wasm::WASM_SEC_GLOBAL); 840 841 encodeULEB128(Globals.size(), W->OS); 842 for (const wasm::WasmGlobal &Global : Globals) { 843 encodeULEB128(Global.Type.Type, W->OS); 844 W->OS << char(Global.Type.Mutable); 845 W->OS << char(Global.InitExpr.Opcode); 846 switch (Global.Type.Type) { 847 case wasm::WASM_TYPE_I32: 848 encodeSLEB128(0, W->OS); 849 break; 850 case wasm::WASM_TYPE_I64: 851 encodeSLEB128(0, W->OS); 852 break; 853 case wasm::WASM_TYPE_F32: 854 writeI32(0); 855 break; 856 case wasm::WASM_TYPE_F64: 857 writeI64(0); 858 break; 859 case wasm::WASM_TYPE_EXTERNREF: 860 writeValueType(wasm::ValType::EXTERNREF); 861 break; 862 default: 863 llvm_unreachable("unexpected type"); 864 } 865 W->OS << char(wasm::WASM_OPCODE_END); 866 } 867 868 endSection(Section); 869 } 870 871 void WasmObjectWriter::writeTableSection(ArrayRef<wasm::WasmTable> Tables) { 872 if (Tables.empty()) 873 return; 874 875 SectionBookkeeping Section; 876 startSection(Section, wasm::WASM_SEC_TABLE); 877 878 encodeULEB128(Tables.size(), W->OS); 879 for (const wasm::WasmTable &Table : Tables) { 880 encodeULEB128(Table.Type.ElemType, W->OS); 881 encodeULEB128(Table.Type.Limits.Flags, W->OS); 882 encodeULEB128(Table.Type.Limits.Initial, W->OS); 883 if (Table.Type.Limits.Flags & wasm::WASM_LIMITS_FLAG_HAS_MAX) 884 encodeULEB128(Table.Type.Limits.Maximum, W->OS); 885 } 886 endSection(Section); 887 } 888 889 void WasmObjectWriter::writeExportSection(ArrayRef<wasm::WasmExport> Exports) { 890 if (Exports.empty()) 891 return; 892 893 SectionBookkeeping Section; 894 startSection(Section, wasm::WASM_SEC_EXPORT); 895 896 encodeULEB128(Exports.size(), W->OS); 897 for (const wasm::WasmExport &Export : Exports) { 898 writeString(Export.Name); 899 W->OS << char(Export.Kind); 900 encodeULEB128(Export.Index, W->OS); 901 } 902 903 endSection(Section); 904 } 905 906 void WasmObjectWriter::writeElemSection( 907 const MCSymbolWasm *IndirectFunctionTable, ArrayRef<uint32_t> TableElems) { 908 if (TableElems.empty()) 909 return; 910 911 assert(IndirectFunctionTable); 912 913 SectionBookkeeping Section; 914 startSection(Section, wasm::WASM_SEC_ELEM); 915 916 encodeULEB128(1, W->OS); // number of "segments" 917 918 assert(WasmIndices.count(IndirectFunctionTable)); 919 uint32_t TableNumber = WasmIndices.find(IndirectFunctionTable)->second; 920 uint32_t Flags = 0; 921 if (TableNumber) 922 Flags |= wasm::WASM_ELEM_SEGMENT_HAS_TABLE_NUMBER; 923 encodeULEB128(Flags, W->OS); 924 if (Flags & wasm::WASM_ELEM_SEGMENT_HAS_TABLE_NUMBER) 925 encodeULEB128(TableNumber, W->OS); // the table number 926 927 // init expr for starting offset 928 W->OS << char(wasm::WASM_OPCODE_I32_CONST); 929 encodeSLEB128(InitialTableOffset, W->OS); 930 W->OS << char(wasm::WASM_OPCODE_END); 931 932 if (Flags & wasm::WASM_ELEM_SEGMENT_MASK_HAS_ELEM_KIND) { 933 // We only write active function table initializers, for which the elem kind 934 // is specified to be written as 0x00 and interpreted to mean "funcref". 935 const uint8_t ElemKind = 0; 936 W->OS << ElemKind; 937 } 938 939 encodeULEB128(TableElems.size(), W->OS); 940 for (uint32_t Elem : TableElems) 941 encodeULEB128(Elem, W->OS); 942 943 endSection(Section); 944 } 945 946 void WasmObjectWriter::writeDataCountSection() { 947 if (DataSegments.empty()) 948 return; 949 950 SectionBookkeeping Section; 951 startSection(Section, wasm::WASM_SEC_DATACOUNT); 952 encodeULEB128(DataSegments.size(), W->OS); 953 endSection(Section); 954 } 955 956 uint32_t WasmObjectWriter::writeCodeSection(const MCAssembler &Asm, 957 const MCAsmLayout &Layout, 958 ArrayRef<WasmFunction> Functions) { 959 if (Functions.empty()) 960 return 0; 961 962 SectionBookkeeping Section; 963 startSection(Section, wasm::WASM_SEC_CODE); 964 965 encodeULEB128(Functions.size(), W->OS); 966 967 for (const WasmFunction &Func : Functions) { 968 auto &FuncSection = static_cast<MCSectionWasm &>(Func.Sym->getSection()); 969 970 int64_t Size = 0; 971 if (!Func.Sym->getSize()->evaluateAsAbsolute(Size, Layout)) 972 report_fatal_error(".size expression must be evaluatable"); 973 974 encodeULEB128(Size, W->OS); 975 FuncSection.setSectionOffset(W->OS.tell() - Section.ContentsOffset); 976 Asm.writeSectionData(W->OS, &FuncSection, Layout); 977 } 978 979 // Apply fixups. 980 applyRelocations(CodeRelocations, Section.ContentsOffset, Layout); 981 982 endSection(Section); 983 return Section.Index; 984 } 985 986 uint32_t WasmObjectWriter::writeDataSection(const MCAsmLayout &Layout) { 987 if (DataSegments.empty()) 988 return 0; 989 990 SectionBookkeeping Section; 991 startSection(Section, wasm::WASM_SEC_DATA); 992 993 encodeULEB128(DataSegments.size(), W->OS); // count 994 995 for (const WasmDataSegment &Segment : DataSegments) { 996 encodeULEB128(Segment.InitFlags, W->OS); // flags 997 if (Segment.InitFlags & wasm::WASM_DATA_SEGMENT_HAS_MEMINDEX) 998 encodeULEB128(0, W->OS); // memory index 999 if ((Segment.InitFlags & wasm::WASM_DATA_SEGMENT_IS_PASSIVE) == 0) { 1000 W->OS << char(is64Bit() ? wasm::WASM_OPCODE_I64_CONST 1001 : wasm::WASM_OPCODE_I32_CONST); 1002 encodeSLEB128(Segment.Offset, W->OS); // offset 1003 W->OS << char(wasm::WASM_OPCODE_END); 1004 } 1005 encodeULEB128(Segment.Data.size(), W->OS); // size 1006 Segment.Section->setSectionOffset(W->OS.tell() - Section.ContentsOffset); 1007 W->OS << Segment.Data; // data 1008 } 1009 1010 // Apply fixups. 1011 applyRelocations(DataRelocations, Section.ContentsOffset, Layout); 1012 1013 endSection(Section); 1014 return Section.Index; 1015 } 1016 1017 void WasmObjectWriter::writeRelocSection( 1018 uint32_t SectionIndex, StringRef Name, 1019 std::vector<WasmRelocationEntry> &Relocs) { 1020 // See: https://github.com/WebAssembly/tool-conventions/blob/master/Linking.md 1021 // for descriptions of the reloc sections. 1022 1023 if (Relocs.empty()) 1024 return; 1025 1026 // First, ensure the relocations are sorted in offset order. In general they 1027 // should already be sorted since `recordRelocation` is called in offset 1028 // order, but for the code section we combine many MC sections into single 1029 // wasm section, and this order is determined by the order of Asm.Symbols() 1030 // not the sections order. 1031 llvm::stable_sort( 1032 Relocs, [](const WasmRelocationEntry &A, const WasmRelocationEntry &B) { 1033 return (A.Offset + A.FixupSection->getSectionOffset()) < 1034 (B.Offset + B.FixupSection->getSectionOffset()); 1035 }); 1036 1037 SectionBookkeeping Section; 1038 startCustomSection(Section, std::string("reloc.") + Name.str()); 1039 1040 encodeULEB128(SectionIndex, W->OS); 1041 encodeULEB128(Relocs.size(), W->OS); 1042 for (const WasmRelocationEntry &RelEntry : Relocs) { 1043 uint64_t Offset = 1044 RelEntry.Offset + RelEntry.FixupSection->getSectionOffset(); 1045 uint32_t Index = getRelocationIndexValue(RelEntry); 1046 1047 W->OS << char(RelEntry.Type); 1048 encodeULEB128(Offset, W->OS); 1049 encodeULEB128(Index, W->OS); 1050 if (RelEntry.hasAddend()) 1051 encodeSLEB128(RelEntry.Addend, W->OS); 1052 } 1053 1054 endSection(Section); 1055 } 1056 1057 void WasmObjectWriter::writeCustomRelocSections() { 1058 for (const auto &Sec : CustomSections) { 1059 auto &Relocations = CustomSectionsRelocations[Sec.Section]; 1060 writeRelocSection(Sec.OutputIndex, Sec.Name, Relocations); 1061 } 1062 } 1063 1064 void WasmObjectWriter::writeLinkingMetaDataSection( 1065 ArrayRef<wasm::WasmSymbolInfo> SymbolInfos, 1066 ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs, 1067 const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats) { 1068 SectionBookkeeping Section; 1069 startCustomSection(Section, "linking"); 1070 encodeULEB128(wasm::WasmMetadataVersion, W->OS); 1071 1072 SectionBookkeeping SubSection; 1073 if (SymbolInfos.size() != 0) { 1074 startSection(SubSection, wasm::WASM_SYMBOL_TABLE); 1075 encodeULEB128(SymbolInfos.size(), W->OS); 1076 for (const wasm::WasmSymbolInfo &Sym : SymbolInfos) { 1077 encodeULEB128(Sym.Kind, W->OS); 1078 encodeULEB128(Sym.Flags, W->OS); 1079 switch (Sym.Kind) { 1080 case wasm::WASM_SYMBOL_TYPE_FUNCTION: 1081 case wasm::WASM_SYMBOL_TYPE_GLOBAL: 1082 case wasm::WASM_SYMBOL_TYPE_EVENT: 1083 case wasm::WASM_SYMBOL_TYPE_TABLE: 1084 encodeULEB128(Sym.ElementIndex, W->OS); 1085 if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0 || 1086 (Sym.Flags & wasm::WASM_SYMBOL_EXPLICIT_NAME) != 0) 1087 writeString(Sym.Name); 1088 break; 1089 case wasm::WASM_SYMBOL_TYPE_DATA: 1090 writeString(Sym.Name); 1091 if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0) { 1092 encodeULEB128(Sym.DataRef.Segment, W->OS); 1093 encodeULEB128(Sym.DataRef.Offset, W->OS); 1094 encodeULEB128(Sym.DataRef.Size, W->OS); 1095 } 1096 break; 1097 case wasm::WASM_SYMBOL_TYPE_SECTION: { 1098 const uint32_t SectionIndex = 1099 CustomSections[Sym.ElementIndex].OutputIndex; 1100 encodeULEB128(SectionIndex, W->OS); 1101 break; 1102 } 1103 default: 1104 llvm_unreachable("unexpected kind"); 1105 } 1106 } 1107 endSection(SubSection); 1108 } 1109 1110 if (DataSegments.size()) { 1111 startSection(SubSection, wasm::WASM_SEGMENT_INFO); 1112 encodeULEB128(DataSegments.size(), W->OS); 1113 for (const WasmDataSegment &Segment : DataSegments) { 1114 writeString(Segment.Name); 1115 encodeULEB128(Segment.Alignment, W->OS); 1116 encodeULEB128(Segment.LinkerFlags, W->OS); 1117 } 1118 endSection(SubSection); 1119 } 1120 1121 if (!InitFuncs.empty()) { 1122 startSection(SubSection, wasm::WASM_INIT_FUNCS); 1123 encodeULEB128(InitFuncs.size(), W->OS); 1124 for (auto &StartFunc : InitFuncs) { 1125 encodeULEB128(StartFunc.first, W->OS); // priority 1126 encodeULEB128(StartFunc.second, W->OS); // function index 1127 } 1128 endSection(SubSection); 1129 } 1130 1131 if (Comdats.size()) { 1132 startSection(SubSection, wasm::WASM_COMDAT_INFO); 1133 encodeULEB128(Comdats.size(), W->OS); 1134 for (const auto &C : Comdats) { 1135 writeString(C.first); 1136 encodeULEB128(0, W->OS); // flags for future use 1137 encodeULEB128(C.second.size(), W->OS); 1138 for (const WasmComdatEntry &Entry : C.second) { 1139 encodeULEB128(Entry.Kind, W->OS); 1140 encodeULEB128(Entry.Index, W->OS); 1141 } 1142 } 1143 endSection(SubSection); 1144 } 1145 1146 endSection(Section); 1147 } 1148 1149 void WasmObjectWriter::writeCustomSection(WasmCustomSection &CustomSection, 1150 const MCAssembler &Asm, 1151 const MCAsmLayout &Layout) { 1152 SectionBookkeeping Section; 1153 auto *Sec = CustomSection.Section; 1154 startCustomSection(Section, CustomSection.Name); 1155 1156 Sec->setSectionOffset(W->OS.tell() - Section.ContentsOffset); 1157 Asm.writeSectionData(W->OS, Sec, Layout); 1158 1159 CustomSection.OutputContentsOffset = Section.ContentsOffset; 1160 CustomSection.OutputIndex = Section.Index; 1161 1162 endSection(Section); 1163 1164 // Apply fixups. 1165 auto &Relocations = CustomSectionsRelocations[CustomSection.Section]; 1166 applyRelocations(Relocations, CustomSection.OutputContentsOffset, Layout); 1167 } 1168 1169 uint32_t WasmObjectWriter::getFunctionType(const MCSymbolWasm &Symbol) { 1170 assert(Symbol.isFunction()); 1171 assert(TypeIndices.count(&Symbol)); 1172 return TypeIndices[&Symbol]; 1173 } 1174 1175 uint32_t WasmObjectWriter::getEventType(const MCSymbolWasm &Symbol) { 1176 assert(Symbol.isEvent()); 1177 assert(TypeIndices.count(&Symbol)); 1178 return TypeIndices[&Symbol]; 1179 } 1180 1181 void WasmObjectWriter::registerFunctionType(const MCSymbolWasm &Symbol) { 1182 assert(Symbol.isFunction()); 1183 1184 wasm::WasmSignature S; 1185 1186 if (auto *Sig = Symbol.getSignature()) { 1187 S.Returns = Sig->Returns; 1188 S.Params = Sig->Params; 1189 } 1190 1191 auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size())); 1192 if (Pair.second) 1193 Signatures.push_back(S); 1194 TypeIndices[&Symbol] = Pair.first->second; 1195 1196 LLVM_DEBUG(dbgs() << "registerFunctionType: " << Symbol 1197 << " new:" << Pair.second << "\n"); 1198 LLVM_DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n"); 1199 } 1200 1201 void WasmObjectWriter::registerEventType(const MCSymbolWasm &Symbol) { 1202 assert(Symbol.isEvent()); 1203 1204 // TODO Currently we don't generate imported exceptions, but if we do, we 1205 // should have a way of infering types of imported exceptions. 1206 wasm::WasmSignature S; 1207 if (auto *Sig = Symbol.getSignature()) { 1208 S.Returns = Sig->Returns; 1209 S.Params = Sig->Params; 1210 } 1211 1212 auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size())); 1213 if (Pair.second) 1214 Signatures.push_back(S); 1215 TypeIndices[&Symbol] = Pair.first->second; 1216 1217 LLVM_DEBUG(dbgs() << "registerEventType: " << Symbol << " new:" << Pair.second 1218 << "\n"); 1219 LLVM_DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n"); 1220 } 1221 1222 static bool isInSymtab(const MCSymbolWasm &Sym) { 1223 if (Sym.isUsedInReloc() || Sym.isUsedInInitArray()) 1224 return true; 1225 1226 if (Sym.isComdat() && !Sym.isDefined()) 1227 return false; 1228 1229 if (Sym.isTemporary()) 1230 return false; 1231 1232 if (Sym.isSection()) 1233 return false; 1234 1235 if (Sym.omitFromLinkingSection()) 1236 return false; 1237 1238 return true; 1239 } 1240 1241 void WasmObjectWriter::prepareImports( 1242 SmallVectorImpl<wasm::WasmImport> &Imports, MCAssembler &Asm, 1243 const MCAsmLayout &Layout) { 1244 // For now, always emit the memory import, since loads and stores are not 1245 // valid without it. In the future, we could perhaps be more clever and omit 1246 // it if there are no loads or stores. 1247 wasm::WasmImport MemImport; 1248 MemImport.Module = "env"; 1249 MemImport.Field = "__linear_memory"; 1250 MemImport.Kind = wasm::WASM_EXTERNAL_MEMORY; 1251 MemImport.Memory.Flags = is64Bit() ? wasm::WASM_LIMITS_FLAG_IS_64 1252 : wasm::WASM_LIMITS_FLAG_NONE; 1253 Imports.push_back(MemImport); 1254 1255 // Populate SignatureIndices, and Imports and WasmIndices for undefined 1256 // symbols. This must be done before populating WasmIndices for defined 1257 // symbols. 1258 for (const MCSymbol &S : Asm.symbols()) { 1259 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1260 1261 // Register types for all functions, including those with private linkage 1262 // (because wasm always needs a type signature). 1263 if (WS.isFunction()) { 1264 const auto *BS = Layout.getBaseSymbol(S); 1265 if (!BS) 1266 report_fatal_error(Twine(S.getName()) + 1267 ": absolute addressing not supported!"); 1268 registerFunctionType(*cast<MCSymbolWasm>(BS)); 1269 } 1270 1271 if (WS.isEvent()) 1272 registerEventType(WS); 1273 1274 if (WS.isTemporary()) 1275 continue; 1276 1277 // If the symbol is not defined in this translation unit, import it. 1278 if (!WS.isDefined() && !WS.isComdat()) { 1279 if (WS.isFunction()) { 1280 wasm::WasmImport Import; 1281 Import.Module = WS.getImportModule(); 1282 Import.Field = WS.getImportName(); 1283 Import.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1284 Import.SigIndex = getFunctionType(WS); 1285 Imports.push_back(Import); 1286 assert(WasmIndices.count(&WS) == 0); 1287 WasmIndices[&WS] = NumFunctionImports++; 1288 } else if (WS.isGlobal()) { 1289 if (WS.isWeak()) 1290 report_fatal_error("undefined global symbol cannot be weak"); 1291 1292 wasm::WasmImport Import; 1293 Import.Field = WS.getImportName(); 1294 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1295 Import.Module = WS.getImportModule(); 1296 Import.Global = WS.getGlobalType(); 1297 Imports.push_back(Import); 1298 assert(WasmIndices.count(&WS) == 0); 1299 WasmIndices[&WS] = NumGlobalImports++; 1300 } else if (WS.isEvent()) { 1301 if (WS.isWeak()) 1302 report_fatal_error("undefined event symbol cannot be weak"); 1303 1304 wasm::WasmImport Import; 1305 Import.Module = WS.getImportModule(); 1306 Import.Field = WS.getImportName(); 1307 Import.Kind = wasm::WASM_EXTERNAL_EVENT; 1308 Import.Event.Attribute = wasm::WASM_EVENT_ATTRIBUTE_EXCEPTION; 1309 Import.Event.SigIndex = getEventType(WS); 1310 Imports.push_back(Import); 1311 assert(WasmIndices.count(&WS) == 0); 1312 WasmIndices[&WS] = NumEventImports++; 1313 } else if (WS.isTable()) { 1314 if (WS.isWeak()) 1315 report_fatal_error("undefined table symbol cannot be weak"); 1316 1317 wasm::WasmImport Import; 1318 Import.Module = WS.getImportModule(); 1319 Import.Field = WS.getImportName(); 1320 Import.Kind = wasm::WASM_EXTERNAL_TABLE; 1321 wasm::ValType ElemType = WS.getTableType(); 1322 Import.Table.ElemType = uint8_t(ElemType); 1323 // FIXME: Extend table type to include limits? For now we don't specify 1324 // a min or max which does not place any restrictions on the size of the 1325 // imported table. 1326 Import.Table.Limits = {wasm::WASM_LIMITS_FLAG_NONE, 0, 0}; 1327 Imports.push_back(Import); 1328 assert(WasmIndices.count(&WS) == 0); 1329 WasmIndices[&WS] = NumTableImports++; 1330 } 1331 } 1332 } 1333 1334 // Add imports for GOT globals 1335 for (const MCSymbol &S : Asm.symbols()) { 1336 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1337 if (WS.isUsedInGOT()) { 1338 wasm::WasmImport Import; 1339 if (WS.isFunction()) 1340 Import.Module = "GOT.func"; 1341 else 1342 Import.Module = "GOT.mem"; 1343 Import.Field = WS.getName(); 1344 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1345 Import.Global = {wasm::WASM_TYPE_I32, true}; 1346 Imports.push_back(Import); 1347 assert(GOTIndices.count(&WS) == 0); 1348 GOTIndices[&WS] = NumGlobalImports++; 1349 } 1350 } 1351 } 1352 1353 uint64_t WasmObjectWriter::writeObject(MCAssembler &Asm, 1354 const MCAsmLayout &Layout) { 1355 support::endian::Writer MainWriter(*OS, support::little); 1356 W = &MainWriter; 1357 if (IsSplitDwarf) { 1358 uint64_t TotalSize = writeOneObject(Asm, Layout, DwoMode::NonDwoOnly); 1359 assert(DwoOS); 1360 support::endian::Writer DwoWriter(*DwoOS, support::little); 1361 W = &DwoWriter; 1362 return TotalSize + writeOneObject(Asm, Layout, DwoMode::DwoOnly); 1363 } else { 1364 return writeOneObject(Asm, Layout, DwoMode::AllSections); 1365 } 1366 } 1367 1368 uint64_t WasmObjectWriter::writeOneObject(MCAssembler &Asm, 1369 const MCAsmLayout &Layout, 1370 DwoMode Mode) { 1371 uint64_t StartOffset = W->OS.tell(); 1372 SectionCount = 0; 1373 CustomSections.clear(); 1374 1375 LLVM_DEBUG(dbgs() << "WasmObjectWriter::writeObject\n"); 1376 1377 // Collect information from the available symbols. 1378 SmallVector<WasmFunction, 4> Functions; 1379 SmallVector<uint32_t, 4> TableElems; 1380 SmallVector<wasm::WasmImport, 4> Imports; 1381 SmallVector<wasm::WasmExport, 4> Exports; 1382 SmallVector<wasm::WasmEventType, 1> Events; 1383 SmallVector<wasm::WasmGlobal, 1> Globals; 1384 SmallVector<wasm::WasmTable, 1> Tables; 1385 SmallVector<wasm::WasmSymbolInfo, 4> SymbolInfos; 1386 SmallVector<std::pair<uint16_t, uint32_t>, 2> InitFuncs; 1387 std::map<StringRef, std::vector<WasmComdatEntry>> Comdats; 1388 uint64_t DataSize = 0; 1389 if (Mode != DwoMode::DwoOnly) { 1390 prepareImports(Imports, Asm, Layout); 1391 } 1392 1393 // Populate DataSegments and CustomSections, which must be done before 1394 // populating DataLocations. 1395 for (MCSection &Sec : Asm) { 1396 auto &Section = static_cast<MCSectionWasm &>(Sec); 1397 StringRef SectionName = Section.getName(); 1398 1399 if (Mode == DwoMode::NonDwoOnly && isDwoSection(Sec)) 1400 continue; 1401 if (Mode == DwoMode::DwoOnly && !isDwoSection(Sec)) 1402 continue; 1403 1404 LLVM_DEBUG(dbgs() << "Processing Section " << SectionName << " group " 1405 << Section.getGroup() << "\n";); 1406 1407 // .init_array sections are handled specially elsewhere. 1408 if (SectionName.startswith(".init_array")) 1409 continue; 1410 1411 // Code is handled separately 1412 if (Section.getKind().isText()) 1413 continue; 1414 1415 if (Section.isWasmData()) { 1416 uint32_t SegmentIndex = DataSegments.size(); 1417 DataSize = alignTo(DataSize, Section.getAlignment()); 1418 DataSegments.emplace_back(); 1419 WasmDataSegment &Segment = DataSegments.back(); 1420 Segment.Name = SectionName; 1421 Segment.InitFlags = Section.getPassive() 1422 ? (uint32_t)wasm::WASM_DATA_SEGMENT_IS_PASSIVE 1423 : 0; 1424 Segment.Offset = DataSize; 1425 Segment.Section = &Section; 1426 addData(Segment.Data, Section); 1427 Segment.Alignment = Log2_32(Section.getAlignment()); 1428 Segment.LinkerFlags = 0; 1429 DataSize += Segment.Data.size(); 1430 Section.setSegmentIndex(SegmentIndex); 1431 1432 if (const MCSymbolWasm *C = Section.getGroup()) { 1433 Comdats[C->getName()].emplace_back( 1434 WasmComdatEntry{wasm::WASM_COMDAT_DATA, SegmentIndex}); 1435 } 1436 } else { 1437 // Create custom sections 1438 assert(Sec.getKind().isMetadata()); 1439 1440 StringRef Name = SectionName; 1441 1442 // For user-defined custom sections, strip the prefix 1443 if (Name.startswith(".custom_section.")) 1444 Name = Name.substr(strlen(".custom_section.")); 1445 1446 MCSymbol *Begin = Sec.getBeginSymbol(); 1447 if (Begin) { 1448 assert(WasmIndices.count(cast<MCSymbolWasm>(Begin)) == 0); 1449 WasmIndices[cast<MCSymbolWasm>(Begin)] = CustomSections.size(); 1450 } 1451 1452 // Separate out the producers and target features sections 1453 if (Name == "producers") { 1454 ProducersSection = std::make_unique<WasmCustomSection>(Name, &Section); 1455 continue; 1456 } 1457 if (Name == "target_features") { 1458 TargetFeaturesSection = 1459 std::make_unique<WasmCustomSection>(Name, &Section); 1460 continue; 1461 } 1462 1463 // Custom sections can also belong to COMDAT groups. In this case the 1464 // decriptor's "index" field is the section index (in the final object 1465 // file), but that is not known until after layout, so it must be fixed up 1466 // later 1467 if (const MCSymbolWasm *C = Section.getGroup()) { 1468 Comdats[C->getName()].emplace_back( 1469 WasmComdatEntry{wasm::WASM_COMDAT_SECTION, 1470 static_cast<uint32_t>(CustomSections.size())}); 1471 } 1472 1473 CustomSections.emplace_back(Name, &Section); 1474 } 1475 } 1476 1477 if (Mode != DwoMode::DwoOnly) { 1478 // Populate WasmIndices and DataLocations for defined symbols. 1479 for (const MCSymbol &S : Asm.symbols()) { 1480 // Ignore unnamed temporary symbols, which aren't ever exported, imported, 1481 // or used in relocations. 1482 if (S.isTemporary() && S.getName().empty()) 1483 continue; 1484 1485 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1486 LLVM_DEBUG(dbgs() 1487 << "MCSymbol: " 1488 << toString(WS.getType().getValueOr(wasm::WASM_SYMBOL_TYPE_DATA)) 1489 << " '" << S << "'" 1490 << " isDefined=" << S.isDefined() << " isExternal=" 1491 << S.isExternal() << " isTemporary=" << S.isTemporary() 1492 << " isWeak=" << WS.isWeak() << " isHidden=" << WS.isHidden() 1493 << " isVariable=" << WS.isVariable() << "\n"); 1494 1495 if (WS.isVariable()) 1496 continue; 1497 if (WS.isComdat() && !WS.isDefined()) 1498 continue; 1499 1500 if (WS.isFunction()) { 1501 unsigned Index; 1502 if (WS.isDefined()) { 1503 if (WS.getOffset() != 0) 1504 report_fatal_error( 1505 "function sections must contain one function each"); 1506 1507 if (WS.getSize() == nullptr) 1508 report_fatal_error( 1509 "function symbols must have a size set with .size"); 1510 1511 // A definition. Write out the function body. 1512 Index = NumFunctionImports + Functions.size(); 1513 WasmFunction Func; 1514 Func.SigIndex = getFunctionType(WS); 1515 Func.Sym = &WS; 1516 assert(WasmIndices.count(&WS) == 0); 1517 WasmIndices[&WS] = Index; 1518 Functions.push_back(Func); 1519 1520 auto &Section = static_cast<MCSectionWasm &>(WS.getSection()); 1521 if (const MCSymbolWasm *C = Section.getGroup()) { 1522 Comdats[C->getName()].emplace_back( 1523 WasmComdatEntry{wasm::WASM_COMDAT_FUNCTION, Index}); 1524 } 1525 1526 if (WS.hasExportName()) { 1527 wasm::WasmExport Export; 1528 Export.Name = WS.getExportName(); 1529 Export.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1530 Export.Index = Index; 1531 Exports.push_back(Export); 1532 } 1533 } else { 1534 // An import; the index was assigned above. 1535 Index = WasmIndices.find(&WS)->second; 1536 } 1537 1538 LLVM_DEBUG(dbgs() << " -> function index: " << Index << "\n"); 1539 1540 } else if (WS.isData()) { 1541 if (!isInSymtab(WS)) 1542 continue; 1543 1544 if (!WS.isDefined()) { 1545 LLVM_DEBUG(dbgs() << " -> segment index: -1" 1546 << "\n"); 1547 continue; 1548 } 1549 1550 if (!WS.getSize()) 1551 report_fatal_error("data symbols must have a size set with .size: " + 1552 WS.getName()); 1553 1554 int64_t Size = 0; 1555 if (!WS.getSize()->evaluateAsAbsolute(Size, Layout)) 1556 report_fatal_error(".size expression must be evaluatable"); 1557 1558 auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection()); 1559 if (!DataSection.isWasmData()) 1560 report_fatal_error("data symbols must live in a data section: " + 1561 WS.getName()); 1562 1563 // For each data symbol, export it in the symtab as a reference to the 1564 // corresponding Wasm data segment. 1565 wasm::WasmDataReference Ref = wasm::WasmDataReference{ 1566 DataSection.getSegmentIndex(), Layout.getSymbolOffset(WS), 1567 static_cast<uint64_t>(Size)}; 1568 assert(DataLocations.count(&WS) == 0); 1569 DataLocations[&WS] = Ref; 1570 LLVM_DEBUG(dbgs() << " -> segment index: " << Ref.Segment << "\n"); 1571 1572 } else if (WS.isGlobal()) { 1573 // A "true" Wasm global (currently just __stack_pointer) 1574 if (WS.isDefined()) { 1575 wasm::WasmGlobal Global; 1576 Global.Type = WS.getGlobalType(); 1577 Global.Index = NumGlobalImports + Globals.size(); 1578 switch (Global.Type.Type) { 1579 case wasm::WASM_TYPE_I32: 1580 Global.InitExpr.Opcode = wasm::WASM_OPCODE_I32_CONST; 1581 break; 1582 case wasm::WASM_TYPE_I64: 1583 Global.InitExpr.Opcode = wasm::WASM_OPCODE_I64_CONST; 1584 break; 1585 case wasm::WASM_TYPE_F32: 1586 Global.InitExpr.Opcode = wasm::WASM_OPCODE_F32_CONST; 1587 break; 1588 case wasm::WASM_TYPE_F64: 1589 Global.InitExpr.Opcode = wasm::WASM_OPCODE_F64_CONST; 1590 break; 1591 case wasm::WASM_TYPE_EXTERNREF: 1592 Global.InitExpr.Opcode = wasm::WASM_OPCODE_REF_NULL; 1593 break; 1594 default: 1595 llvm_unreachable("unexpected type"); 1596 } 1597 assert(WasmIndices.count(&WS) == 0); 1598 WasmIndices[&WS] = Global.Index; 1599 Globals.push_back(Global); 1600 } else { 1601 // An import; the index was assigned above 1602 LLVM_DEBUG(dbgs() << " -> global index: " 1603 << WasmIndices.find(&WS)->second << "\n"); 1604 } 1605 } else if (WS.isTable()) { 1606 if (WS.isDefined()) { 1607 wasm::WasmTable Table; 1608 Table.Index = NumTableImports + Tables.size(); 1609 Table.Type.ElemType = static_cast<uint8_t>(WS.getTableType()); 1610 // FIXME: Work on custom limits is ongoing 1611 Table.Type.Limits = {wasm::WASM_LIMITS_FLAG_NONE, 0, 0}; 1612 assert(WasmIndices.count(&WS) == 0); 1613 WasmIndices[&WS] = Table.Index; 1614 Tables.push_back(Table); 1615 } 1616 LLVM_DEBUG(dbgs() << " -> table index: " 1617 << WasmIndices.find(&WS)->second << "\n"); 1618 } else if (WS.isEvent()) { 1619 // C++ exception symbol (__cpp_exception) 1620 unsigned Index; 1621 if (WS.isDefined()) { 1622 Index = NumEventImports + Events.size(); 1623 wasm::WasmEventType Event; 1624 Event.SigIndex = getEventType(WS); 1625 Event.Attribute = wasm::WASM_EVENT_ATTRIBUTE_EXCEPTION; 1626 assert(WasmIndices.count(&WS) == 0); 1627 WasmIndices[&WS] = Index; 1628 Events.push_back(Event); 1629 } else { 1630 // An import; the index was assigned above. 1631 assert(WasmIndices.count(&WS) > 0); 1632 } 1633 LLVM_DEBUG(dbgs() << " -> event index: " 1634 << WasmIndices.find(&WS)->second << "\n"); 1635 1636 } else { 1637 assert(WS.isSection()); 1638 } 1639 } 1640 1641 // Populate WasmIndices and DataLocations for aliased symbols. We need to 1642 // process these in a separate pass because we need to have processed the 1643 // target of the alias before the alias itself and the symbols are not 1644 // necessarily ordered in this way. 1645 for (const MCSymbol &S : Asm.symbols()) { 1646 if (!S.isVariable()) 1647 continue; 1648 1649 assert(S.isDefined()); 1650 1651 const auto *BS = Layout.getBaseSymbol(S); 1652 if (!BS) 1653 report_fatal_error(Twine(S.getName()) + 1654 ": absolute addressing not supported!"); 1655 const MCSymbolWasm *Base = cast<MCSymbolWasm>(BS); 1656 1657 // Find the target symbol of this weak alias and export that index 1658 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1659 LLVM_DEBUG(dbgs() << WS.getName() << ": weak alias of '" << *Base 1660 << "'\n"); 1661 1662 if (Base->isFunction()) { 1663 assert(WasmIndices.count(Base) > 0); 1664 uint32_t WasmIndex = WasmIndices.find(Base)->second; 1665 assert(WasmIndices.count(&WS) == 0); 1666 WasmIndices[&WS] = WasmIndex; 1667 LLVM_DEBUG(dbgs() << " -> index:" << WasmIndex << "\n"); 1668 } else if (Base->isData()) { 1669 auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection()); 1670 uint64_t Offset = Layout.getSymbolOffset(S); 1671 int64_t Size = 0; 1672 // For data symbol alias we use the size of the base symbol as the 1673 // size of the alias. When an offset from the base is involved this 1674 // can result in a offset + size goes past the end of the data section 1675 // which out object format doesn't support. So we must clamp it. 1676 if (!Base->getSize()->evaluateAsAbsolute(Size, Layout)) 1677 report_fatal_error(".size expression must be evaluatable"); 1678 const WasmDataSegment &Segment = 1679 DataSegments[DataSection.getSegmentIndex()]; 1680 Size = 1681 std::min(static_cast<uint64_t>(Size), Segment.Data.size() - Offset); 1682 wasm::WasmDataReference Ref = wasm::WasmDataReference{ 1683 DataSection.getSegmentIndex(), 1684 static_cast<uint32_t>(Layout.getSymbolOffset(S)), 1685 static_cast<uint32_t>(Size)}; 1686 DataLocations[&WS] = Ref; 1687 LLVM_DEBUG(dbgs() << " -> index:" << Ref.Segment << "\n"); 1688 } else { 1689 report_fatal_error("don't yet support global/event aliases"); 1690 } 1691 } 1692 } 1693 1694 // Finally, populate the symbol table itself, in its "natural" order. 1695 for (const MCSymbol &S : Asm.symbols()) { 1696 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1697 if (!isInSymtab(WS)) { 1698 WS.setIndex(InvalidIndex); 1699 continue; 1700 } 1701 LLVM_DEBUG(dbgs() << "adding to symtab: " << WS << "\n"); 1702 1703 uint32_t Flags = 0; 1704 if (WS.isWeak()) 1705 Flags |= wasm::WASM_SYMBOL_BINDING_WEAK; 1706 if (WS.isHidden()) 1707 Flags |= wasm::WASM_SYMBOL_VISIBILITY_HIDDEN; 1708 if (!WS.isExternal() && WS.isDefined()) 1709 Flags |= wasm::WASM_SYMBOL_BINDING_LOCAL; 1710 if (WS.isUndefined()) 1711 Flags |= wasm::WASM_SYMBOL_UNDEFINED; 1712 if (WS.isNoStrip()) { 1713 Flags |= wasm::WASM_SYMBOL_NO_STRIP; 1714 if (isEmscripten()) { 1715 Flags |= wasm::WASM_SYMBOL_EXPORTED; 1716 } 1717 } 1718 if (WS.hasImportName()) 1719 Flags |= wasm::WASM_SYMBOL_EXPLICIT_NAME; 1720 if (WS.hasExportName()) 1721 Flags |= wasm::WASM_SYMBOL_EXPORTED; 1722 1723 wasm::WasmSymbolInfo Info; 1724 Info.Name = WS.getName(); 1725 Info.Kind = WS.getType().getValueOr(wasm::WASM_SYMBOL_TYPE_DATA); 1726 Info.Flags = Flags; 1727 if (!WS.isData()) { 1728 assert(WasmIndices.count(&WS) > 0); 1729 Info.ElementIndex = WasmIndices.find(&WS)->second; 1730 } else if (WS.isDefined()) { 1731 assert(DataLocations.count(&WS) > 0); 1732 Info.DataRef = DataLocations.find(&WS)->second; 1733 } 1734 WS.setIndex(SymbolInfos.size()); 1735 SymbolInfos.emplace_back(Info); 1736 } 1737 1738 { 1739 auto HandleReloc = [&](const WasmRelocationEntry &Rel) { 1740 // Functions referenced by a relocation need to put in the table. This is 1741 // purely to make the object file's provisional values readable, and is 1742 // ignored by the linker, which re-calculates the relocations itself. 1743 if (Rel.Type != wasm::R_WASM_TABLE_INDEX_I32 && 1744 Rel.Type != wasm::R_WASM_TABLE_INDEX_I64 && 1745 Rel.Type != wasm::R_WASM_TABLE_INDEX_SLEB && 1746 Rel.Type != wasm::R_WASM_TABLE_INDEX_SLEB64 && 1747 Rel.Type != wasm::R_WASM_TABLE_INDEX_REL_SLEB) 1748 return; 1749 assert(Rel.Symbol->isFunction()); 1750 const MCSymbolWasm *Base = 1751 cast<MCSymbolWasm>(Layout.getBaseSymbol(*Rel.Symbol)); 1752 uint32_t FunctionIndex = WasmIndices.find(Base)->second; 1753 uint32_t TableIndex = TableElems.size() + InitialTableOffset; 1754 if (TableIndices.try_emplace(Base, TableIndex).second) { 1755 LLVM_DEBUG(dbgs() << " -> adding " << Base->getName() 1756 << " to table: " << TableIndex << "\n"); 1757 TableElems.push_back(FunctionIndex); 1758 registerFunctionType(*Base); 1759 } 1760 }; 1761 1762 for (const WasmRelocationEntry &RelEntry : CodeRelocations) 1763 HandleReloc(RelEntry); 1764 for (const WasmRelocationEntry &RelEntry : DataRelocations) 1765 HandleReloc(RelEntry); 1766 } 1767 1768 // Translate .init_array section contents into start functions. 1769 for (const MCSection &S : Asm) { 1770 const auto &WS = static_cast<const MCSectionWasm &>(S); 1771 if (WS.getName().startswith(".fini_array")) 1772 report_fatal_error(".fini_array sections are unsupported"); 1773 if (!WS.getName().startswith(".init_array")) 1774 continue; 1775 if (WS.getFragmentList().empty()) 1776 continue; 1777 1778 // init_array is expected to contain a single non-empty data fragment 1779 if (WS.getFragmentList().size() != 3) 1780 report_fatal_error("only one .init_array section fragment supported"); 1781 1782 auto IT = WS.begin(); 1783 const MCFragment &EmptyFrag = *IT; 1784 if (EmptyFrag.getKind() != MCFragment::FT_Data) 1785 report_fatal_error(".init_array section should be aligned"); 1786 1787 IT = std::next(IT); 1788 const MCFragment &AlignFrag = *IT; 1789 if (AlignFrag.getKind() != MCFragment::FT_Align) 1790 report_fatal_error(".init_array section should be aligned"); 1791 if (cast<MCAlignFragment>(AlignFrag).getAlignment() != (is64Bit() ? 8 : 4)) 1792 report_fatal_error(".init_array section should be aligned for pointers"); 1793 1794 const MCFragment &Frag = *std::next(IT); 1795 if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data) 1796 report_fatal_error("only data supported in .init_array section"); 1797 1798 uint16_t Priority = UINT16_MAX; 1799 unsigned PrefixLength = strlen(".init_array"); 1800 if (WS.getName().size() > PrefixLength) { 1801 if (WS.getName()[PrefixLength] != '.') 1802 report_fatal_error( 1803 ".init_array section priority should start with '.'"); 1804 if (WS.getName().substr(PrefixLength + 1).getAsInteger(10, Priority)) 1805 report_fatal_error("invalid .init_array section priority"); 1806 } 1807 const auto &DataFrag = cast<MCDataFragment>(Frag); 1808 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 1809 for (const uint8_t * 1810 P = (const uint8_t *)Contents.data(), 1811 *End = (const uint8_t *)Contents.data() + Contents.size(); 1812 P != End; ++P) { 1813 if (*P != 0) 1814 report_fatal_error("non-symbolic data in .init_array section"); 1815 } 1816 for (const MCFixup &Fixup : DataFrag.getFixups()) { 1817 assert(Fixup.getKind() == 1818 MCFixup::getKindForSize(is64Bit() ? 8 : 4, false)); 1819 const MCExpr *Expr = Fixup.getValue(); 1820 auto *SymRef = dyn_cast<MCSymbolRefExpr>(Expr); 1821 if (!SymRef) 1822 report_fatal_error("fixups in .init_array should be symbol references"); 1823 const auto &TargetSym = cast<const MCSymbolWasm>(SymRef->getSymbol()); 1824 if (TargetSym.getIndex() == InvalidIndex) 1825 report_fatal_error("symbols in .init_array should exist in symtab"); 1826 if (!TargetSym.isFunction()) 1827 report_fatal_error("symbols in .init_array should be for functions"); 1828 InitFuncs.push_back( 1829 std::make_pair(Priority, TargetSym.getIndex())); 1830 } 1831 } 1832 1833 // Write out the Wasm header. 1834 writeHeader(Asm); 1835 1836 uint32_t CodeSectionIndex, DataSectionIndex; 1837 if (Mode != DwoMode::DwoOnly) { 1838 writeTypeSection(Signatures); 1839 writeImportSection(Imports, DataSize, TableElems.size()); 1840 writeFunctionSection(Functions); 1841 writeTableSection(Tables); 1842 // Skip the "memory" section; we import the memory instead. 1843 writeEventSection(Events); 1844 writeGlobalSection(Globals); 1845 writeExportSection(Exports); 1846 const MCSymbol *IndirectFunctionTable = 1847 Asm.getContext().lookupSymbol("__indirect_function_table"); 1848 writeElemSection(cast_or_null<const MCSymbolWasm>(IndirectFunctionTable), 1849 TableElems); 1850 writeDataCountSection(); 1851 1852 CodeSectionIndex = writeCodeSection(Asm, Layout, Functions); 1853 DataSectionIndex = writeDataSection(Layout); 1854 } 1855 1856 // The Sections in the COMDAT list have placeholder indices (their index among 1857 // custom sections, rather than among all sections). Fix them up here. 1858 for (auto &Group : Comdats) { 1859 for (auto &Entry : Group.second) { 1860 if (Entry.Kind == wasm::WASM_COMDAT_SECTION) { 1861 Entry.Index += SectionCount; 1862 } 1863 } 1864 } 1865 for (auto &CustomSection : CustomSections) 1866 writeCustomSection(CustomSection, Asm, Layout); 1867 1868 if (Mode != DwoMode::DwoOnly) { 1869 writeLinkingMetaDataSection(SymbolInfos, InitFuncs, Comdats); 1870 1871 writeRelocSection(CodeSectionIndex, "CODE", CodeRelocations); 1872 writeRelocSection(DataSectionIndex, "DATA", DataRelocations); 1873 } 1874 writeCustomRelocSections(); 1875 if (ProducersSection) 1876 writeCustomSection(*ProducersSection, Asm, Layout); 1877 if (TargetFeaturesSection) 1878 writeCustomSection(*TargetFeaturesSection, Asm, Layout); 1879 1880 // TODO: Translate the .comment section to the output. 1881 return W->OS.tell() - StartOffset; 1882 } 1883 1884 std::unique_ptr<MCObjectWriter> 1885 llvm::createWasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 1886 raw_pwrite_stream &OS) { 1887 return std::make_unique<WasmObjectWriter>(std::move(MOTW), OS); 1888 } 1889 1890 std::unique_ptr<MCObjectWriter> 1891 llvm::createWasmDwoObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 1892 raw_pwrite_stream &OS, 1893 raw_pwrite_stream &DwoOS) { 1894 return std::make_unique<WasmObjectWriter>(std::move(MOTW), OS, DwoOS); 1895 } 1896