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 MCAsmBackend &Backend = Asm.getBackend(); 430 bool IsPCRel = Backend.getFixupKindInfo(Fixup.getKind()).Flags & 431 MCFixupKindInfo::FKF_IsPCRel; 432 const auto &FixupSection = cast<MCSectionWasm>(*Fragment->getParent()); 433 uint64_t C = Target.getConstant(); 434 uint64_t FixupOffset = Layout.getFragmentOffset(Fragment) + Fixup.getOffset(); 435 MCContext &Ctx = Asm.getContext(); 436 437 // The .init_array isn't translated as data, so don't do relocations in it. 438 if (FixupSection.getSectionName().startswith(".init_array")) 439 return; 440 441 if (const MCSymbolRefExpr *RefB = Target.getSymB()) { 442 assert(RefB->getKind() == MCSymbolRefExpr::VK_None && 443 "Should not have constructed this"); 444 445 // Let A, B and C being the components of Target and R be the location of 446 // the fixup. If the fixup is not pcrel, we want to compute (A - B + C). 447 // If it is pcrel, we want to compute (A - B + C - R). 448 449 // In general, Wasm has no relocations for -B. It can only represent (A + C) 450 // or (A + C - R). If B = R + K and the relocation is not pcrel, we can 451 // replace B to implement it: (A - R - K + C) 452 if (IsPCRel) { 453 Ctx.reportError( 454 Fixup.getLoc(), 455 "No relocation available to represent this relative expression"); 456 return; 457 } 458 459 const auto &SymB = cast<MCSymbolWasm>(RefB->getSymbol()); 460 461 if (SymB.isUndefined()) { 462 Ctx.reportError(Fixup.getLoc(), 463 Twine("symbol '") + SymB.getName() + 464 "' can not be undefined in a subtraction expression"); 465 return; 466 } 467 468 assert(!SymB.isAbsolute() && "Should have been folded"); 469 const MCSection &SecB = SymB.getSection(); 470 if (&SecB != &FixupSection) { 471 Ctx.reportError(Fixup.getLoc(), 472 "Cannot represent a difference across sections"); 473 return; 474 } 475 476 uint64_t SymBOffset = Layout.getSymbolOffset(SymB); 477 uint64_t K = SymBOffset - FixupOffset; 478 IsPCRel = true; 479 C -= K; 480 } 481 482 // We either rejected the fixup or folded B into C at this point. 483 const MCSymbolRefExpr *RefA = Target.getSymA(); 484 const auto *SymA = RefA ? cast<MCSymbolWasm>(&RefA->getSymbol()) : nullptr; 485 486 if (SymA && SymA->isVariable()) { 487 const MCExpr *Expr = SymA->getVariableValue(); 488 const auto *Inner = cast<MCSymbolRefExpr>(Expr); 489 if (Inner->getKind() == MCSymbolRefExpr::VK_WEAKREF) 490 llvm_unreachable("weakref used in reloc not yet implemented"); 491 } 492 493 // Put any constant offset in an addend. Offsets can be negative, and 494 // LLVM expects wrapping, in contrast to wasm's immediates which can't 495 // be negative and don't wrap. 496 FixedValue = 0; 497 498 unsigned Type = TargetObjectWriter->getRelocType(Target, Fixup); 499 assert(!IsPCRel); 500 assert(SymA); 501 502 // Absolute offset within a section or a function. 503 // Currently only supported for for metadata sections. 504 // See: test/MC/WebAssembly/blockaddress.ll 505 if (Type == wasm::R_WASM_FUNCTION_OFFSET_I32 || 506 Type == wasm::R_WASM_SECTION_OFFSET_I32) { 507 if (!FixupSection.getKind().isMetadata()) 508 report_fatal_error("relocations for function or section offsets are " 509 "only supported in metadata sections"); 510 511 const MCSymbol *SectionSymbol = nullptr; 512 const MCSection &SecA = SymA->getSection(); 513 if (SecA.getKind().isText()) 514 SectionSymbol = SectionFunctions.find(&SecA)->second; 515 else 516 SectionSymbol = SecA.getBeginSymbol(); 517 if (!SectionSymbol) 518 report_fatal_error("section symbol is required for relocation"); 519 520 C += Layout.getSymbolOffset(*SymA); 521 SymA = cast<MCSymbolWasm>(SectionSymbol); 522 } 523 524 // Relocation other than R_WASM_TYPE_INDEX_LEB are required to be 525 // against a named symbol. 526 if (Type != wasm::R_WASM_TYPE_INDEX_LEB) { 527 if (SymA->getName().empty()) 528 report_fatal_error("relocations against un-named temporaries are not yet " 529 "supported by wasm"); 530 531 SymA->setUsedInReloc(); 532 } 533 534 if (RefA->getKind() == MCSymbolRefExpr::VK_GOT) 535 SymA->setUsedInGOT(); 536 537 WasmRelocationEntry Rec(FixupOffset, SymA, C, Type, &FixupSection); 538 LLVM_DEBUG(dbgs() << "WasmReloc: " << Rec << "\n"); 539 540 if (FixupSection.isWasmData()) { 541 DataRelocations.push_back(Rec); 542 } else if (FixupSection.getKind().isText()) { 543 CodeRelocations.push_back(Rec); 544 } else if (FixupSection.getKind().isMetadata()) { 545 CustomSectionsRelocations[&FixupSection].push_back(Rec); 546 } else { 547 llvm_unreachable("unexpected section type"); 548 } 549 } 550 551 static const MCSymbolWasm *resolveSymbol(const MCSymbolWasm &Symbol) { 552 const MCSymbolWasm* Ret = &Symbol; 553 while (Ret->isVariable()) { 554 const MCExpr *Expr = Ret->getVariableValue(); 555 auto *Inner = cast<MCSymbolRefExpr>(Expr); 556 Ret = cast<MCSymbolWasm>(&Inner->getSymbol()); 557 } 558 return Ret; 559 } 560 561 // Compute a value to write into the code at the location covered 562 // by RelEntry. This value isn't used by the static linker; it just serves 563 // to make the object format more readable and more likely to be directly 564 // useable. 565 uint32_t 566 WasmObjectWriter::getProvisionalValue(const WasmRelocationEntry &RelEntry) { 567 if (RelEntry.Type == wasm::R_WASM_GLOBAL_INDEX_LEB && !RelEntry.Symbol->isGlobal()) { 568 assert(GOTIndices.count(RelEntry.Symbol) > 0 && "symbol not found in GOT index space"); 569 return GOTIndices[RelEntry.Symbol]; 570 } 571 572 switch (RelEntry.Type) { 573 case wasm::R_WASM_TABLE_INDEX_REL_SLEB: 574 case wasm::R_WASM_TABLE_INDEX_SLEB: 575 case wasm::R_WASM_TABLE_INDEX_I32: { 576 // Provisional value is table address of the resolved symbol itself 577 const MCSymbolWasm *Sym = resolveSymbol(*RelEntry.Symbol); 578 assert(Sym->isFunction()); 579 return TableIndices[Sym]; 580 } 581 case wasm::R_WASM_TYPE_INDEX_LEB: 582 // Provisional value is same as the index 583 return getRelocationIndexValue(RelEntry); 584 case wasm::R_WASM_FUNCTION_INDEX_LEB: 585 case wasm::R_WASM_GLOBAL_INDEX_LEB: 586 case wasm::R_WASM_EVENT_INDEX_LEB: 587 // Provisional value is function/global/event Wasm index 588 assert(WasmIndices.count(RelEntry.Symbol) > 0 && "symbol not found in wasm index space"); 589 return WasmIndices[RelEntry.Symbol]; 590 case wasm::R_WASM_FUNCTION_OFFSET_I32: 591 case wasm::R_WASM_SECTION_OFFSET_I32: { 592 const auto &Section = 593 static_cast<const MCSectionWasm &>(RelEntry.Symbol->getSection()); 594 return Section.getSectionOffset() + RelEntry.Addend; 595 } 596 case wasm::R_WASM_MEMORY_ADDR_LEB: 597 case wasm::R_WASM_MEMORY_ADDR_I32: 598 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB: 599 case wasm::R_WASM_MEMORY_ADDR_SLEB: { 600 // Provisional value is address of the global 601 const MCSymbolWasm *Sym = resolveSymbol(*RelEntry.Symbol); 602 // For undefined symbols, use zero 603 if (!Sym->isDefined()) 604 return 0; 605 const wasm::WasmDataReference &Ref = DataLocations[Sym]; 606 const WasmDataSegment &Segment = DataSegments[Ref.Segment]; 607 // Ignore overflow. LLVM allows address arithmetic to silently wrap. 608 return Segment.Offset + Ref.Offset + RelEntry.Addend; 609 } 610 default: 611 llvm_unreachable("invalid relocation type"); 612 } 613 } 614 615 static void addData(SmallVectorImpl<char> &DataBytes, 616 MCSectionWasm &DataSection) { 617 LLVM_DEBUG(errs() << "addData: " << DataSection.getSectionName() << "\n"); 618 619 DataBytes.resize(alignTo(DataBytes.size(), DataSection.getAlignment())); 620 621 for (const MCFragment &Frag : DataSection) { 622 if (Frag.hasInstructions()) 623 report_fatal_error("only data supported in data sections"); 624 625 if (auto *Align = dyn_cast<MCAlignFragment>(&Frag)) { 626 if (Align->getValueSize() != 1) 627 report_fatal_error("only byte values supported for alignment"); 628 // If nops are requested, use zeros, as this is the data section. 629 uint8_t Value = Align->hasEmitNops() ? 0 : Align->getValue(); 630 uint64_t Size = 631 std::min<uint64_t>(alignTo(DataBytes.size(), Align->getAlignment()), 632 DataBytes.size() + Align->getMaxBytesToEmit()); 633 DataBytes.resize(Size, Value); 634 } else if (auto *Fill = dyn_cast<MCFillFragment>(&Frag)) { 635 int64_t NumValues; 636 if (!Fill->getNumValues().evaluateAsAbsolute(NumValues)) 637 llvm_unreachable("The fill should be an assembler constant"); 638 DataBytes.insert(DataBytes.end(), Fill->getValueSize() * NumValues, 639 Fill->getValue()); 640 } else if (auto *LEB = dyn_cast<MCLEBFragment>(&Frag)) { 641 const SmallVectorImpl<char> &Contents = LEB->getContents(); 642 DataBytes.insert(DataBytes.end(), Contents.begin(), Contents.end()); 643 } else { 644 const auto &DataFrag = cast<MCDataFragment>(Frag); 645 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 646 DataBytes.insert(DataBytes.end(), Contents.begin(), Contents.end()); 647 } 648 } 649 650 LLVM_DEBUG(dbgs() << "addData -> " << DataBytes.size() << "\n"); 651 } 652 653 uint32_t 654 WasmObjectWriter::getRelocationIndexValue(const WasmRelocationEntry &RelEntry) { 655 if (RelEntry.Type == wasm::R_WASM_TYPE_INDEX_LEB) { 656 if (!TypeIndices.count(RelEntry.Symbol)) 657 report_fatal_error("symbol not found in type index space: " + 658 RelEntry.Symbol->getName()); 659 return TypeIndices[RelEntry.Symbol]; 660 } 661 662 return RelEntry.Symbol->getIndex(); 663 } 664 665 // Apply the portions of the relocation records that we can handle ourselves 666 // directly. 667 void WasmObjectWriter::applyRelocations( 668 ArrayRef<WasmRelocationEntry> Relocations, uint64_t ContentsOffset) { 669 auto &Stream = static_cast<raw_pwrite_stream &>(W.OS); 670 for (const WasmRelocationEntry &RelEntry : Relocations) { 671 uint64_t Offset = ContentsOffset + 672 RelEntry.FixupSection->getSectionOffset() + 673 RelEntry.Offset; 674 675 LLVM_DEBUG(dbgs() << "applyRelocation: " << RelEntry << "\n"); 676 uint32_t Value = getProvisionalValue(RelEntry); 677 678 switch (RelEntry.Type) { 679 case wasm::R_WASM_FUNCTION_INDEX_LEB: 680 case wasm::R_WASM_TYPE_INDEX_LEB: 681 case wasm::R_WASM_GLOBAL_INDEX_LEB: 682 case wasm::R_WASM_MEMORY_ADDR_LEB: 683 case wasm::R_WASM_EVENT_INDEX_LEB: 684 writePatchableLEB(Stream, Value, Offset); 685 break; 686 case wasm::R_WASM_TABLE_INDEX_I32: 687 case wasm::R_WASM_MEMORY_ADDR_I32: 688 case wasm::R_WASM_FUNCTION_OFFSET_I32: 689 case wasm::R_WASM_SECTION_OFFSET_I32: 690 writeI32(Stream, Value, Offset); 691 break; 692 case wasm::R_WASM_TABLE_INDEX_SLEB: 693 case wasm::R_WASM_TABLE_INDEX_REL_SLEB: 694 case wasm::R_WASM_MEMORY_ADDR_SLEB: 695 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB: 696 writePatchableSLEB(Stream, Value, Offset); 697 break; 698 default: 699 llvm_unreachable("invalid relocation type"); 700 } 701 } 702 } 703 704 void WasmObjectWriter::writeTypeSection(ArrayRef<WasmSignature> Signatures) { 705 if (Signatures.empty()) 706 return; 707 708 SectionBookkeeping Section; 709 startSection(Section, wasm::WASM_SEC_TYPE); 710 711 encodeULEB128(Signatures.size(), W.OS); 712 713 for (const WasmSignature &Sig : Signatures) { 714 W.OS << char(wasm::WASM_TYPE_FUNC); 715 encodeULEB128(Sig.Params.size(), W.OS); 716 for (wasm::ValType Ty : Sig.Params) 717 writeValueType(Ty); 718 encodeULEB128(Sig.Returns.size(), W.OS); 719 for (wasm::ValType Ty : Sig.Returns) 720 writeValueType(Ty); 721 } 722 723 endSection(Section); 724 } 725 726 void WasmObjectWriter::writeImportSection(ArrayRef<wasm::WasmImport> Imports, 727 uint32_t DataSize, 728 uint32_t NumElements) { 729 if (Imports.empty()) 730 return; 731 732 uint32_t NumPages = (DataSize + wasm::WasmPageSize - 1) / wasm::WasmPageSize; 733 734 SectionBookkeeping Section; 735 startSection(Section, wasm::WASM_SEC_IMPORT); 736 737 encodeULEB128(Imports.size(), W.OS); 738 for (const wasm::WasmImport &Import : Imports) { 739 writeString(Import.Module); 740 writeString(Import.Field); 741 W.OS << char(Import.Kind); 742 743 switch (Import.Kind) { 744 case wasm::WASM_EXTERNAL_FUNCTION: 745 encodeULEB128(Import.SigIndex, W.OS); 746 break; 747 case wasm::WASM_EXTERNAL_GLOBAL: 748 W.OS << char(Import.Global.Type); 749 W.OS << char(Import.Global.Mutable ? 1 : 0); 750 break; 751 case wasm::WASM_EXTERNAL_MEMORY: 752 encodeULEB128(0, W.OS); // flags 753 encodeULEB128(NumPages, W.OS); // initial 754 break; 755 case wasm::WASM_EXTERNAL_TABLE: 756 W.OS << char(Import.Table.ElemType); 757 encodeULEB128(0, W.OS); // flags 758 encodeULEB128(NumElements, W.OS); // initial 759 break; 760 case wasm::WASM_EXTERNAL_EVENT: 761 encodeULEB128(Import.Event.Attribute, W.OS); 762 encodeULEB128(Import.Event.SigIndex, W.OS); 763 break; 764 default: 765 llvm_unreachable("unsupported import kind"); 766 } 767 } 768 769 endSection(Section); 770 } 771 772 void WasmObjectWriter::writeFunctionSection(ArrayRef<WasmFunction> Functions) { 773 if (Functions.empty()) 774 return; 775 776 SectionBookkeeping Section; 777 startSection(Section, wasm::WASM_SEC_FUNCTION); 778 779 encodeULEB128(Functions.size(), W.OS); 780 for (const WasmFunction &Func : Functions) 781 encodeULEB128(Func.SigIndex, W.OS); 782 783 endSection(Section); 784 } 785 786 void WasmObjectWriter::writeEventSection(ArrayRef<wasm::WasmEventType> Events) { 787 if (Events.empty()) 788 return; 789 790 SectionBookkeeping Section; 791 startSection(Section, wasm::WASM_SEC_EVENT); 792 793 encodeULEB128(Events.size(), W.OS); 794 for (const wasm::WasmEventType &Event : Events) { 795 encodeULEB128(Event.Attribute, W.OS); 796 encodeULEB128(Event.SigIndex, W.OS); 797 } 798 799 endSection(Section); 800 } 801 802 void WasmObjectWriter::writeExportSection(ArrayRef<wasm::WasmExport> Exports) { 803 if (Exports.empty()) 804 return; 805 806 SectionBookkeeping Section; 807 startSection(Section, wasm::WASM_SEC_EXPORT); 808 809 encodeULEB128(Exports.size(), W.OS); 810 for (const wasm::WasmExport &Export : Exports) { 811 writeString(Export.Name); 812 W.OS << char(Export.Kind); 813 encodeULEB128(Export.Index, W.OS); 814 } 815 816 endSection(Section); 817 } 818 819 void WasmObjectWriter::writeElemSection(ArrayRef<uint32_t> TableElems) { 820 if (TableElems.empty()) 821 return; 822 823 SectionBookkeeping Section; 824 startSection(Section, wasm::WASM_SEC_ELEM); 825 826 encodeULEB128(1, W.OS); // number of "segments" 827 encodeULEB128(0, W.OS); // the table index 828 829 // init expr for starting offset 830 W.OS << char(wasm::WASM_OPCODE_I32_CONST); 831 encodeSLEB128(InitialTableOffset, W.OS); 832 W.OS << char(wasm::WASM_OPCODE_END); 833 834 encodeULEB128(TableElems.size(), W.OS); 835 for (uint32_t Elem : TableElems) 836 encodeULEB128(Elem, W.OS); 837 838 endSection(Section); 839 } 840 841 void WasmObjectWriter::writeDataCountSection() { 842 if (DataSegments.empty()) 843 return; 844 845 SectionBookkeeping Section; 846 startSection(Section, wasm::WASM_SEC_DATACOUNT); 847 encodeULEB128(DataSegments.size(), W.OS); 848 endSection(Section); 849 } 850 851 void WasmObjectWriter::writeCodeSection(const MCAssembler &Asm, 852 const MCAsmLayout &Layout, 853 ArrayRef<WasmFunction> Functions) { 854 if (Functions.empty()) 855 return; 856 857 SectionBookkeeping Section; 858 startSection(Section, wasm::WASM_SEC_CODE); 859 CodeSectionIndex = Section.Index; 860 861 encodeULEB128(Functions.size(), W.OS); 862 863 for (const WasmFunction &Func : Functions) { 864 auto &FuncSection = static_cast<MCSectionWasm &>(Func.Sym->getSection()); 865 866 int64_t Size = 0; 867 if (!Func.Sym->getSize()->evaluateAsAbsolute(Size, Layout)) 868 report_fatal_error(".size expression must be evaluatable"); 869 870 encodeULEB128(Size, W.OS); 871 FuncSection.setSectionOffset(W.OS.tell() - Section.ContentsOffset); 872 Asm.writeSectionData(W.OS, &FuncSection, Layout); 873 } 874 875 // Apply fixups. 876 applyRelocations(CodeRelocations, Section.ContentsOffset); 877 878 endSection(Section); 879 } 880 881 void WasmObjectWriter::writeDataSection() { 882 if (DataSegments.empty()) 883 return; 884 885 SectionBookkeeping Section; 886 startSection(Section, wasm::WASM_SEC_DATA); 887 DataSectionIndex = Section.Index; 888 889 encodeULEB128(DataSegments.size(), W.OS); // count 890 891 for (const WasmDataSegment &Segment : DataSegments) { 892 encodeULEB128(Segment.InitFlags, W.OS); // flags 893 if (Segment.InitFlags & wasm::WASM_SEGMENT_HAS_MEMINDEX) 894 encodeULEB128(0, W.OS); // memory index 895 if ((Segment.InitFlags & wasm::WASM_SEGMENT_IS_PASSIVE) == 0) { 896 W.OS << char(wasm::WASM_OPCODE_I32_CONST); 897 encodeSLEB128(Segment.Offset, W.OS); // offset 898 W.OS << char(wasm::WASM_OPCODE_END); 899 } 900 encodeULEB128(Segment.Data.size(), W.OS); // size 901 Segment.Section->setSectionOffset(W.OS.tell() - Section.ContentsOffset); 902 W.OS << Segment.Data; // data 903 } 904 905 // Apply fixups. 906 applyRelocations(DataRelocations, Section.ContentsOffset); 907 908 endSection(Section); 909 } 910 911 void WasmObjectWriter::writeRelocSection( 912 uint32_t SectionIndex, StringRef Name, 913 std::vector<WasmRelocationEntry> &Relocs) { 914 // See: https://github.com/WebAssembly/tool-conventions/blob/master/Linking.md 915 // for descriptions of the reloc sections. 916 917 if (Relocs.empty()) 918 return; 919 920 // First, ensure the relocations are sorted in offset order. In general they 921 // should already be sorted since `recordRelocation` is called in offset 922 // order, but for the code section we combine many MC sections into single 923 // wasm section, and this order is determined by the order of Asm.Symbols() 924 // not the sections order. 925 llvm::stable_sort( 926 Relocs, [](const WasmRelocationEntry &A, const WasmRelocationEntry &B) { 927 return (A.Offset + A.FixupSection->getSectionOffset()) < 928 (B.Offset + B.FixupSection->getSectionOffset()); 929 }); 930 931 SectionBookkeeping Section; 932 startCustomSection(Section, std::string("reloc.") + Name.str()); 933 934 encodeULEB128(SectionIndex, W.OS); 935 encodeULEB128(Relocs.size(), W.OS); 936 for (const WasmRelocationEntry &RelEntry : Relocs) { 937 uint64_t Offset = 938 RelEntry.Offset + RelEntry.FixupSection->getSectionOffset(); 939 uint32_t Index = getRelocationIndexValue(RelEntry); 940 941 W.OS << char(RelEntry.Type); 942 encodeULEB128(Offset, W.OS); 943 encodeULEB128(Index, W.OS); 944 if (RelEntry.hasAddend()) 945 encodeSLEB128(RelEntry.Addend, W.OS); 946 } 947 948 endSection(Section); 949 } 950 951 void WasmObjectWriter::writeCustomRelocSections() { 952 for (const auto &Sec : CustomSections) { 953 auto &Relocations = CustomSectionsRelocations[Sec.Section]; 954 writeRelocSection(Sec.OutputIndex, Sec.Name, Relocations); 955 } 956 } 957 958 void WasmObjectWriter::writeLinkingMetaDataSection( 959 ArrayRef<wasm::WasmSymbolInfo> SymbolInfos, 960 ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs, 961 const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats) { 962 SectionBookkeeping Section; 963 startCustomSection(Section, "linking"); 964 encodeULEB128(wasm::WasmMetadataVersion, W.OS); 965 966 SectionBookkeeping SubSection; 967 if (SymbolInfos.size() != 0) { 968 startSection(SubSection, wasm::WASM_SYMBOL_TABLE); 969 encodeULEB128(SymbolInfos.size(), W.OS); 970 for (const wasm::WasmSymbolInfo &Sym : SymbolInfos) { 971 encodeULEB128(Sym.Kind, W.OS); 972 encodeULEB128(Sym.Flags, W.OS); 973 switch (Sym.Kind) { 974 case wasm::WASM_SYMBOL_TYPE_FUNCTION: 975 case wasm::WASM_SYMBOL_TYPE_GLOBAL: 976 case wasm::WASM_SYMBOL_TYPE_EVENT: 977 encodeULEB128(Sym.ElementIndex, W.OS); 978 if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0 || 979 (Sym.Flags & wasm::WASM_SYMBOL_EXPLICIT_NAME) != 0) 980 writeString(Sym.Name); 981 break; 982 case wasm::WASM_SYMBOL_TYPE_DATA: 983 writeString(Sym.Name); 984 if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0) { 985 encodeULEB128(Sym.DataRef.Segment, W.OS); 986 encodeULEB128(Sym.DataRef.Offset, W.OS); 987 encodeULEB128(Sym.DataRef.Size, W.OS); 988 } 989 break; 990 case wasm::WASM_SYMBOL_TYPE_SECTION: { 991 const uint32_t SectionIndex = 992 CustomSections[Sym.ElementIndex].OutputIndex; 993 encodeULEB128(SectionIndex, W.OS); 994 break; 995 } 996 default: 997 llvm_unreachable("unexpected kind"); 998 } 999 } 1000 endSection(SubSection); 1001 } 1002 1003 if (DataSegments.size()) { 1004 startSection(SubSection, wasm::WASM_SEGMENT_INFO); 1005 encodeULEB128(DataSegments.size(), W.OS); 1006 for (const WasmDataSegment &Segment : DataSegments) { 1007 writeString(Segment.Name); 1008 encodeULEB128(Segment.Alignment, W.OS); 1009 encodeULEB128(Segment.LinkerFlags, W.OS); 1010 } 1011 endSection(SubSection); 1012 } 1013 1014 if (!InitFuncs.empty()) { 1015 startSection(SubSection, wasm::WASM_INIT_FUNCS); 1016 encodeULEB128(InitFuncs.size(), W.OS); 1017 for (auto &StartFunc : InitFuncs) { 1018 encodeULEB128(StartFunc.first, W.OS); // priority 1019 encodeULEB128(StartFunc.second, W.OS); // function index 1020 } 1021 endSection(SubSection); 1022 } 1023 1024 if (Comdats.size()) { 1025 startSection(SubSection, wasm::WASM_COMDAT_INFO); 1026 encodeULEB128(Comdats.size(), W.OS); 1027 for (const auto &C : Comdats) { 1028 writeString(C.first); 1029 encodeULEB128(0, W.OS); // flags for future use 1030 encodeULEB128(C.second.size(), W.OS); 1031 for (const WasmComdatEntry &Entry : C.second) { 1032 encodeULEB128(Entry.Kind, W.OS); 1033 encodeULEB128(Entry.Index, W.OS); 1034 } 1035 } 1036 endSection(SubSection); 1037 } 1038 1039 endSection(Section); 1040 } 1041 1042 void WasmObjectWriter::writeCustomSection(WasmCustomSection &CustomSection, 1043 const MCAssembler &Asm, 1044 const MCAsmLayout &Layout) { 1045 SectionBookkeeping Section; 1046 auto *Sec = CustomSection.Section; 1047 startCustomSection(Section, CustomSection.Name); 1048 1049 Sec->setSectionOffset(W.OS.tell() - Section.ContentsOffset); 1050 Asm.writeSectionData(W.OS, Sec, Layout); 1051 1052 CustomSection.OutputContentsOffset = Section.ContentsOffset; 1053 CustomSection.OutputIndex = Section.Index; 1054 1055 endSection(Section); 1056 1057 // Apply fixups. 1058 auto &Relocations = CustomSectionsRelocations[CustomSection.Section]; 1059 applyRelocations(Relocations, CustomSection.OutputContentsOffset); 1060 } 1061 1062 uint32_t WasmObjectWriter::getFunctionType(const MCSymbolWasm &Symbol) { 1063 assert(Symbol.isFunction()); 1064 assert(TypeIndices.count(&Symbol)); 1065 return TypeIndices[&Symbol]; 1066 } 1067 1068 uint32_t WasmObjectWriter::getEventType(const MCSymbolWasm &Symbol) { 1069 assert(Symbol.isEvent()); 1070 assert(TypeIndices.count(&Symbol)); 1071 return TypeIndices[&Symbol]; 1072 } 1073 1074 void WasmObjectWriter::registerFunctionType(const MCSymbolWasm &Symbol) { 1075 assert(Symbol.isFunction()); 1076 1077 WasmSignature S; 1078 const MCSymbolWasm *ResolvedSym = resolveSymbol(Symbol); 1079 if (auto *Sig = ResolvedSym->getSignature()) { 1080 S.Returns = Sig->Returns; 1081 S.Params = Sig->Params; 1082 } 1083 1084 auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size())); 1085 if (Pair.second) 1086 Signatures.push_back(S); 1087 TypeIndices[&Symbol] = Pair.first->second; 1088 1089 LLVM_DEBUG(dbgs() << "registerFunctionType: " << Symbol 1090 << " new:" << Pair.second << "\n"); 1091 LLVM_DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n"); 1092 } 1093 1094 void WasmObjectWriter::registerEventType(const MCSymbolWasm &Symbol) { 1095 assert(Symbol.isEvent()); 1096 1097 // TODO Currently we don't generate imported exceptions, but if we do, we 1098 // should have a way of infering types of imported exceptions. 1099 WasmSignature S; 1100 if (auto *Sig = Symbol.getSignature()) { 1101 S.Returns = Sig->Returns; 1102 S.Params = Sig->Params; 1103 } 1104 1105 auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size())); 1106 if (Pair.second) 1107 Signatures.push_back(S); 1108 TypeIndices[&Symbol] = Pair.first->second; 1109 1110 LLVM_DEBUG(dbgs() << "registerEventType: " << Symbol << " new:" << Pair.second 1111 << "\n"); 1112 LLVM_DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n"); 1113 } 1114 1115 static bool isInSymtab(const MCSymbolWasm &Sym) { 1116 if (Sym.isUsedInReloc()) 1117 return true; 1118 1119 if (Sym.isComdat() && !Sym.isDefined()) 1120 return false; 1121 1122 if (Sym.isTemporary() && Sym.getName().empty()) 1123 return false; 1124 1125 if (Sym.isTemporary() && Sym.isData() && !Sym.getSize()) 1126 return false; 1127 1128 if (Sym.isSection()) 1129 return false; 1130 1131 return true; 1132 } 1133 1134 uint64_t WasmObjectWriter::writeObject(MCAssembler &Asm, 1135 const MCAsmLayout &Layout) { 1136 uint64_t StartOffset = W.OS.tell(); 1137 1138 LLVM_DEBUG(dbgs() << "WasmObjectWriter::writeObject\n"); 1139 1140 // Collect information from the available symbols. 1141 SmallVector<WasmFunction, 4> Functions; 1142 SmallVector<uint32_t, 4> TableElems; 1143 SmallVector<wasm::WasmImport, 4> Imports; 1144 SmallVector<wasm::WasmExport, 4> Exports; 1145 SmallVector<wasm::WasmEventType, 1> Events; 1146 SmallVector<wasm::WasmSymbolInfo, 4> SymbolInfos; 1147 SmallVector<std::pair<uint16_t, uint32_t>, 2> InitFuncs; 1148 std::map<StringRef, std::vector<WasmComdatEntry>> Comdats; 1149 uint32_t DataSize = 0; 1150 1151 // For now, always emit the memory import, since loads and stores are not 1152 // valid without it. In the future, we could perhaps be more clever and omit 1153 // it if there are no loads or stores. 1154 wasm::WasmImport MemImport; 1155 MemImport.Module = "env"; 1156 MemImport.Field = "__linear_memory"; 1157 MemImport.Kind = wasm::WASM_EXTERNAL_MEMORY; 1158 Imports.push_back(MemImport); 1159 1160 // For now, always emit the table section, since indirect calls are not 1161 // valid without it. In the future, we could perhaps be more clever and omit 1162 // it if there are no indirect calls. 1163 wasm::WasmImport TableImport; 1164 TableImport.Module = "env"; 1165 TableImport.Field = "__indirect_function_table"; 1166 TableImport.Kind = wasm::WASM_EXTERNAL_TABLE; 1167 TableImport.Table.ElemType = wasm::WASM_TYPE_FUNCREF; 1168 Imports.push_back(TableImport); 1169 1170 // Populate SignatureIndices, and Imports and WasmIndices for undefined 1171 // symbols. This must be done before populating WasmIndices for defined 1172 // symbols. 1173 for (const MCSymbol &S : Asm.symbols()) { 1174 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1175 1176 // Register types for all functions, including those with private linkage 1177 // (because wasm always needs a type signature). 1178 if (WS.isFunction()) 1179 registerFunctionType(WS); 1180 1181 if (WS.isEvent()) 1182 registerEventType(WS); 1183 1184 if (WS.isTemporary()) 1185 continue; 1186 1187 // If the symbol is not defined in this translation unit, import it. 1188 if (!WS.isDefined() && !WS.isComdat()) { 1189 if (WS.isFunction()) { 1190 wasm::WasmImport Import; 1191 Import.Module = WS.getImportModule(); 1192 Import.Field = WS.getImportName(); 1193 Import.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1194 Import.SigIndex = getFunctionType(WS); 1195 Imports.push_back(Import); 1196 assert(WasmIndices.count(&WS) == 0); 1197 WasmIndices[&WS] = NumFunctionImports++; 1198 } else if (WS.isGlobal()) { 1199 if (WS.isWeak()) 1200 report_fatal_error("undefined global symbol cannot be weak"); 1201 1202 wasm::WasmImport Import; 1203 Import.Field = WS.getImportName(); 1204 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1205 Import.Module = WS.getImportModule(); 1206 Import.Global = WS.getGlobalType(); 1207 Imports.push_back(Import); 1208 assert(WasmIndices.count(&WS) == 0); 1209 WasmIndices[&WS] = NumGlobalImports++; 1210 } else if (WS.isEvent()) { 1211 if (WS.isWeak()) 1212 report_fatal_error("undefined event symbol cannot be weak"); 1213 1214 wasm::WasmImport Import; 1215 Import.Module = WS.getImportModule(); 1216 Import.Field = WS.getImportName(); 1217 Import.Kind = wasm::WASM_EXTERNAL_EVENT; 1218 Import.Event.Attribute = wasm::WASM_EVENT_ATTRIBUTE_EXCEPTION; 1219 Import.Event.SigIndex = getEventType(WS); 1220 Imports.push_back(Import); 1221 assert(WasmIndices.count(&WS) == 0); 1222 WasmIndices[&WS] = NumEventImports++; 1223 } 1224 } 1225 } 1226 1227 // Add imports for GOT globals 1228 for (const MCSymbol &S : Asm.symbols()) { 1229 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1230 if (WS.isUsedInGOT()) { 1231 wasm::WasmImport Import; 1232 if (WS.isFunction()) 1233 Import.Module = "GOT.func"; 1234 else 1235 Import.Module = "GOT.mem"; 1236 Import.Field = WS.getName(); 1237 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1238 Import.Global = {wasm::WASM_TYPE_I32, true}; 1239 Imports.push_back(Import); 1240 assert(GOTIndices.count(&WS) == 0); 1241 GOTIndices[&WS] = NumGlobalImports++; 1242 } 1243 } 1244 1245 // Populate DataSegments and CustomSections, which must be done before 1246 // populating DataLocations. 1247 for (MCSection &Sec : Asm) { 1248 auto &Section = static_cast<MCSectionWasm &>(Sec); 1249 StringRef SectionName = Section.getSectionName(); 1250 1251 // .init_array sections are handled specially elsewhere. 1252 if (SectionName.startswith(".init_array")) 1253 continue; 1254 1255 // Code is handled separately 1256 if (Section.getKind().isText()) 1257 continue; 1258 1259 if (Section.isWasmData()) { 1260 uint32_t SegmentIndex = DataSegments.size(); 1261 DataSize = alignTo(DataSize, Section.getAlignment()); 1262 DataSegments.emplace_back(); 1263 WasmDataSegment &Segment = DataSegments.back(); 1264 Segment.Name = SectionName; 1265 Segment.InitFlags = 1266 Section.getPassive() ? (uint32_t)wasm::WASM_SEGMENT_IS_PASSIVE : 0; 1267 Segment.Offset = DataSize; 1268 Segment.Section = &Section; 1269 addData(Segment.Data, Section); 1270 Segment.Alignment = Log2_32(Section.getAlignment()); 1271 Segment.LinkerFlags = 0; 1272 DataSize += Segment.Data.size(); 1273 Section.setSegmentIndex(SegmentIndex); 1274 1275 if (const MCSymbolWasm *C = Section.getGroup()) { 1276 Comdats[C->getName()].emplace_back( 1277 WasmComdatEntry{wasm::WASM_COMDAT_DATA, SegmentIndex}); 1278 } 1279 } else { 1280 // Create custom sections 1281 assert(Sec.getKind().isMetadata()); 1282 1283 StringRef Name = SectionName; 1284 1285 // For user-defined custom sections, strip the prefix 1286 if (Name.startswith(".custom_section.")) 1287 Name = Name.substr(strlen(".custom_section.")); 1288 1289 MCSymbol *Begin = Sec.getBeginSymbol(); 1290 if (Begin) { 1291 WasmIndices[cast<MCSymbolWasm>(Begin)] = CustomSections.size(); 1292 if (SectionName != Begin->getName()) 1293 report_fatal_error("section name and begin symbol should match: " + 1294 Twine(SectionName)); 1295 } 1296 1297 // Separate out the producers and target features sections 1298 if (Name == "producers") { 1299 ProducersSection = llvm::make_unique<WasmCustomSection>(Name, &Section); 1300 continue; 1301 } 1302 if (Name == "target_features") { 1303 TargetFeaturesSection = 1304 llvm::make_unique<WasmCustomSection>(Name, &Section); 1305 continue; 1306 } 1307 1308 CustomSections.emplace_back(Name, &Section); 1309 } 1310 } 1311 1312 // Populate WasmIndices and DataLocations for defined symbols. 1313 for (const MCSymbol &S : Asm.symbols()) { 1314 // Ignore unnamed temporary symbols, which aren't ever exported, imported, 1315 // or used in relocations. 1316 if (S.isTemporary() && S.getName().empty()) 1317 continue; 1318 1319 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1320 LLVM_DEBUG( 1321 dbgs() << "MCSymbol: " << toString(WS.getType()) << " '" << S << "'" 1322 << " isDefined=" << S.isDefined() << " isExternal=" 1323 << S.isExternal() << " isTemporary=" << S.isTemporary() 1324 << " isWeak=" << WS.isWeak() << " isHidden=" << WS.isHidden() 1325 << " isVariable=" << WS.isVariable() << "\n"); 1326 1327 if (WS.isVariable()) 1328 continue; 1329 if (WS.isComdat() && !WS.isDefined()) 1330 continue; 1331 1332 if (WS.isFunction()) { 1333 unsigned Index; 1334 if (WS.isDefined()) { 1335 if (WS.getOffset() != 0) 1336 report_fatal_error( 1337 "function sections must contain one function each"); 1338 1339 if (WS.getSize() == nullptr) 1340 report_fatal_error( 1341 "function symbols must have a size set with .size"); 1342 1343 // A definition. Write out the function body. 1344 Index = NumFunctionImports + Functions.size(); 1345 WasmFunction Func; 1346 Func.SigIndex = getFunctionType(WS); 1347 Func.Sym = &WS; 1348 WasmIndices[&WS] = Index; 1349 Functions.push_back(Func); 1350 1351 auto &Section = static_cast<MCSectionWasm &>(WS.getSection()); 1352 if (const MCSymbolWasm *C = Section.getGroup()) { 1353 Comdats[C->getName()].emplace_back( 1354 WasmComdatEntry{wasm::WASM_COMDAT_FUNCTION, Index}); 1355 } 1356 } else { 1357 // An import; the index was assigned above. 1358 Index = WasmIndices.find(&WS)->second; 1359 } 1360 1361 LLVM_DEBUG(dbgs() << " -> function index: " << Index << "\n"); 1362 1363 } else if (WS.isData()) { 1364 if (!isInSymtab(WS)) 1365 continue; 1366 1367 if (!WS.isDefined()) { 1368 LLVM_DEBUG(dbgs() << " -> segment index: -1" 1369 << "\n"); 1370 continue; 1371 } 1372 1373 if (!WS.getSize()) 1374 report_fatal_error("data symbols must have a size set with .size: " + 1375 WS.getName()); 1376 1377 int64_t Size = 0; 1378 if (!WS.getSize()->evaluateAsAbsolute(Size, Layout)) 1379 report_fatal_error(".size expression must be evaluatable"); 1380 1381 auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection()); 1382 assert(DataSection.isWasmData()); 1383 1384 // For each data symbol, export it in the symtab as a reference to the 1385 // corresponding Wasm data segment. 1386 wasm::WasmDataReference Ref = wasm::WasmDataReference{ 1387 DataSection.getSegmentIndex(), 1388 static_cast<uint32_t>(Layout.getSymbolOffset(WS)), 1389 static_cast<uint32_t>(Size)}; 1390 DataLocations[&WS] = Ref; 1391 LLVM_DEBUG(dbgs() << " -> segment index: " << Ref.Segment << "\n"); 1392 1393 } else if (WS.isGlobal()) { 1394 // A "true" Wasm global (currently just __stack_pointer) 1395 if (WS.isDefined()) 1396 report_fatal_error("don't yet support defined globals"); 1397 1398 // An import; the index was assigned above 1399 LLVM_DEBUG(dbgs() << " -> global index: " 1400 << WasmIndices.find(&WS)->second << "\n"); 1401 1402 } else if (WS.isEvent()) { 1403 // C++ exception symbol (__cpp_exception) 1404 unsigned Index; 1405 if (WS.isDefined()) { 1406 Index = NumEventImports + Events.size(); 1407 wasm::WasmEventType Event; 1408 Event.SigIndex = getEventType(WS); 1409 Event.Attribute = wasm::WASM_EVENT_ATTRIBUTE_EXCEPTION; 1410 assert(WasmIndices.count(&WS) == 0); 1411 WasmIndices[&WS] = Index; 1412 Events.push_back(Event); 1413 } else { 1414 // An import; the index was assigned above. 1415 assert(WasmIndices.count(&WS) > 0); 1416 } 1417 LLVM_DEBUG(dbgs() << " -> event index: " << WasmIndices.find(&WS)->second 1418 << "\n"); 1419 1420 } else { 1421 assert(WS.isSection()); 1422 } 1423 } 1424 1425 // Populate WasmIndices and DataLocations for aliased symbols. We need to 1426 // process these in a separate pass because we need to have processed the 1427 // target of the alias before the alias itself and the symbols are not 1428 // necessarily ordered in this way. 1429 for (const MCSymbol &S : Asm.symbols()) { 1430 if (!S.isVariable()) 1431 continue; 1432 1433 assert(S.isDefined()); 1434 1435 // Find the target symbol of this weak alias and export that index 1436 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1437 const MCSymbolWasm *ResolvedSym = resolveSymbol(WS); 1438 LLVM_DEBUG(dbgs() << WS.getName() << ": weak alias of '" << *ResolvedSym 1439 << "'\n"); 1440 1441 if (ResolvedSym->isFunction()) { 1442 assert(WasmIndices.count(ResolvedSym) > 0); 1443 uint32_t WasmIndex = WasmIndices.find(ResolvedSym)->second; 1444 assert(WasmIndices.count(&WS) == 0); 1445 WasmIndices[&WS] = WasmIndex; 1446 LLVM_DEBUG(dbgs() << " -> index:" << WasmIndex << "\n"); 1447 } else if (ResolvedSym->isData()) { 1448 assert(DataLocations.count(ResolvedSym) > 0); 1449 const wasm::WasmDataReference &Ref = 1450 DataLocations.find(ResolvedSym)->second; 1451 DataLocations[&WS] = Ref; 1452 LLVM_DEBUG(dbgs() << " -> index:" << Ref.Segment << "\n"); 1453 } else { 1454 report_fatal_error("don't yet support global/event aliases"); 1455 } 1456 } 1457 1458 // Finally, populate the symbol table itself, in its "natural" order. 1459 for (const MCSymbol &S : Asm.symbols()) { 1460 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1461 if (!isInSymtab(WS)) { 1462 WS.setIndex(InvalidIndex); 1463 continue; 1464 } 1465 LLVM_DEBUG(dbgs() << "adding to symtab: " << WS << "\n"); 1466 1467 uint32_t Flags = 0; 1468 if (WS.isWeak()) 1469 Flags |= wasm::WASM_SYMBOL_BINDING_WEAK; 1470 if (WS.isHidden()) 1471 Flags |= wasm::WASM_SYMBOL_VISIBILITY_HIDDEN; 1472 if (!WS.isExternal() && WS.isDefined()) 1473 Flags |= wasm::WASM_SYMBOL_BINDING_LOCAL; 1474 if (WS.isUndefined()) 1475 Flags |= wasm::WASM_SYMBOL_UNDEFINED; 1476 if (WS.isExported()) 1477 Flags |= wasm::WASM_SYMBOL_EXPORTED; 1478 if (WS.getName() != WS.getImportName()) 1479 Flags |= wasm::WASM_SYMBOL_EXPLICIT_NAME; 1480 1481 wasm::WasmSymbolInfo Info; 1482 Info.Name = WS.getName(); 1483 Info.Kind = WS.getType(); 1484 Info.Flags = Flags; 1485 if (!WS.isData()) { 1486 assert(WasmIndices.count(&WS) > 0); 1487 Info.ElementIndex = WasmIndices.find(&WS)->second; 1488 } else if (WS.isDefined()) { 1489 assert(DataLocations.count(&WS) > 0); 1490 Info.DataRef = DataLocations.find(&WS)->second; 1491 } 1492 WS.setIndex(SymbolInfos.size()); 1493 SymbolInfos.emplace_back(Info); 1494 } 1495 1496 { 1497 auto HandleReloc = [&](const WasmRelocationEntry &Rel) { 1498 // Functions referenced by a relocation need to put in the table. This is 1499 // purely to make the object file's provisional values readable, and is 1500 // ignored by the linker, which re-calculates the relocations itself. 1501 if (Rel.Type != wasm::R_WASM_TABLE_INDEX_I32 && 1502 Rel.Type != wasm::R_WASM_TABLE_INDEX_SLEB) 1503 return; 1504 assert(Rel.Symbol->isFunction()); 1505 const MCSymbolWasm &WS = *resolveSymbol(*Rel.Symbol); 1506 uint32_t FunctionIndex = WasmIndices.find(&WS)->second; 1507 uint32_t TableIndex = TableElems.size() + InitialTableOffset; 1508 if (TableIndices.try_emplace(&WS, TableIndex).second) { 1509 LLVM_DEBUG(dbgs() << " -> adding " << WS.getName() 1510 << " to table: " << TableIndex << "\n"); 1511 TableElems.push_back(FunctionIndex); 1512 registerFunctionType(WS); 1513 } 1514 }; 1515 1516 for (const WasmRelocationEntry &RelEntry : CodeRelocations) 1517 HandleReloc(RelEntry); 1518 for (const WasmRelocationEntry &RelEntry : DataRelocations) 1519 HandleReloc(RelEntry); 1520 } 1521 1522 // Translate .init_array section contents into start functions. 1523 for (const MCSection &S : Asm) { 1524 const auto &WS = static_cast<const MCSectionWasm &>(S); 1525 if (WS.getSectionName().startswith(".fini_array")) 1526 report_fatal_error(".fini_array sections are unsupported"); 1527 if (!WS.getSectionName().startswith(".init_array")) 1528 continue; 1529 if (WS.getFragmentList().empty()) 1530 continue; 1531 1532 // init_array is expected to contain a single non-empty data fragment 1533 if (WS.getFragmentList().size() != 3) 1534 report_fatal_error("only one .init_array section fragment supported"); 1535 1536 auto IT = WS.begin(); 1537 const MCFragment &EmptyFrag = *IT; 1538 if (EmptyFrag.getKind() != MCFragment::FT_Data) 1539 report_fatal_error(".init_array section should be aligned"); 1540 1541 IT = std::next(IT); 1542 const MCFragment &AlignFrag = *IT; 1543 if (AlignFrag.getKind() != MCFragment::FT_Align) 1544 report_fatal_error(".init_array section should be aligned"); 1545 if (cast<MCAlignFragment>(AlignFrag).getAlignment() != (is64Bit() ? 8 : 4)) 1546 report_fatal_error(".init_array section should be aligned for pointers"); 1547 1548 const MCFragment &Frag = *std::next(IT); 1549 if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data) 1550 report_fatal_error("only data supported in .init_array section"); 1551 1552 uint16_t Priority = UINT16_MAX; 1553 unsigned PrefixLength = strlen(".init_array"); 1554 if (WS.getSectionName().size() > PrefixLength) { 1555 if (WS.getSectionName()[PrefixLength] != '.') 1556 report_fatal_error( 1557 ".init_array section priority should start with '.'"); 1558 if (WS.getSectionName() 1559 .substr(PrefixLength + 1) 1560 .getAsInteger(10, Priority)) 1561 report_fatal_error("invalid .init_array section priority"); 1562 } 1563 const auto &DataFrag = cast<MCDataFragment>(Frag); 1564 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 1565 for (const uint8_t * 1566 P = (const uint8_t *)Contents.data(), 1567 *End = (const uint8_t *)Contents.data() + Contents.size(); 1568 P != End; ++P) { 1569 if (*P != 0) 1570 report_fatal_error("non-symbolic data in .init_array section"); 1571 } 1572 for (const MCFixup &Fixup : DataFrag.getFixups()) { 1573 assert(Fixup.getKind() == 1574 MCFixup::getKindForSize(is64Bit() ? 8 : 4, false)); 1575 const MCExpr *Expr = Fixup.getValue(); 1576 auto *SymRef = dyn_cast<MCSymbolRefExpr>(Expr); 1577 if (!SymRef) 1578 report_fatal_error("fixups in .init_array should be symbol references"); 1579 const auto &TargetSym = cast<const MCSymbolWasm>(SymRef->getSymbol()); 1580 if (TargetSym.getIndex() == InvalidIndex) 1581 report_fatal_error("symbols in .init_array should exist in symbtab"); 1582 if (!TargetSym.isFunction()) 1583 report_fatal_error("symbols in .init_array should be for functions"); 1584 InitFuncs.push_back( 1585 std::make_pair(Priority, TargetSym.getIndex())); 1586 } 1587 } 1588 1589 // Write out the Wasm header. 1590 writeHeader(Asm); 1591 1592 writeTypeSection(Signatures); 1593 writeImportSection(Imports, DataSize, TableElems.size()); 1594 writeFunctionSection(Functions); 1595 // Skip the "table" section; we import the table instead. 1596 // Skip the "memory" section; we import the memory instead. 1597 writeEventSection(Events); 1598 writeExportSection(Exports); 1599 writeElemSection(TableElems); 1600 writeDataCountSection(); 1601 writeCodeSection(Asm, Layout, Functions); 1602 writeDataSection(); 1603 for (auto &CustomSection : CustomSections) 1604 writeCustomSection(CustomSection, Asm, Layout); 1605 writeLinkingMetaDataSection(SymbolInfos, InitFuncs, Comdats); 1606 writeRelocSection(CodeSectionIndex, "CODE", CodeRelocations); 1607 writeRelocSection(DataSectionIndex, "DATA", DataRelocations); 1608 writeCustomRelocSections(); 1609 if (ProducersSection) 1610 writeCustomSection(*ProducersSection, Asm, Layout); 1611 if (TargetFeaturesSection) 1612 writeCustomSection(*TargetFeaturesSection, Asm, Layout); 1613 1614 // TODO: Translate the .comment section to the output. 1615 return W.OS.tell() - StartOffset; 1616 } 1617 1618 std::unique_ptr<MCObjectWriter> 1619 llvm::createWasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 1620 raw_pwrite_stream &OS) { 1621 return llvm::make_unique<WasmObjectWriter>(std::move(MOTW), OS); 1622 } 1623