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 LinkingFlags; 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/tag/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 NumTagImports = 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 writeStringWithAlignment(const StringRef Str, unsigned Alignment); 296 297 void writeI32(int32_t val) { 298 char Buffer[4]; 299 support::endian::write32le(Buffer, val); 300 W->OS.write(Buffer, sizeof(Buffer)); 301 } 302 303 void writeI64(int64_t val) { 304 char Buffer[8]; 305 support::endian::write64le(Buffer, val); 306 W->OS.write(Buffer, sizeof(Buffer)); 307 } 308 309 void writeValueType(wasm::ValType Ty) { W->OS << static_cast<char>(Ty); } 310 311 void writeTypeSection(ArrayRef<wasm::WasmSignature> Signatures); 312 void writeImportSection(ArrayRef<wasm::WasmImport> Imports, uint64_t DataSize, 313 uint32_t NumElements); 314 void writeFunctionSection(ArrayRef<WasmFunction> Functions); 315 void writeExportSection(ArrayRef<wasm::WasmExport> Exports); 316 void writeElemSection(const MCSymbolWasm *IndirectFunctionTable, 317 ArrayRef<uint32_t> TableElems); 318 void writeDataCountSection(); 319 uint32_t writeCodeSection(const MCAssembler &Asm, const MCAsmLayout &Layout, 320 ArrayRef<WasmFunction> Functions); 321 uint32_t writeDataSection(const MCAsmLayout &Layout); 322 void writeTagSection(ArrayRef<uint32_t> TagTypes); 323 void writeGlobalSection(ArrayRef<wasm::WasmGlobal> Globals); 324 void writeTableSection(ArrayRef<wasm::WasmTable> Tables); 325 void writeRelocSection(uint32_t SectionIndex, StringRef Name, 326 std::vector<WasmRelocationEntry> &Relocations); 327 void writeLinkingMetaDataSection( 328 ArrayRef<wasm::WasmSymbolInfo> SymbolInfos, 329 ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs, 330 const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats); 331 void writeCustomSection(WasmCustomSection &CustomSection, 332 const MCAssembler &Asm, const MCAsmLayout &Layout); 333 void writeCustomRelocSections(); 334 335 uint64_t getProvisionalValue(const WasmRelocationEntry &RelEntry, 336 const MCAsmLayout &Layout); 337 void applyRelocations(ArrayRef<WasmRelocationEntry> Relocations, 338 uint64_t ContentsOffset, const MCAsmLayout &Layout); 339 340 uint32_t getRelocationIndexValue(const WasmRelocationEntry &RelEntry); 341 uint32_t getFunctionType(const MCSymbolWasm &Symbol); 342 uint32_t getTagType(const MCSymbolWasm &Symbol); 343 void registerFunctionType(const MCSymbolWasm &Symbol); 344 void registerTagType(const MCSymbolWasm &Symbol); 345 }; 346 347 } // end anonymous namespace 348 349 // Write out a section header and a patchable section size field. 350 void WasmObjectWriter::startSection(SectionBookkeeping &Section, 351 unsigned SectionId) { 352 LLVM_DEBUG(dbgs() << "startSection " << SectionId << "\n"); 353 W->OS << char(SectionId); 354 355 Section.SizeOffset = W->OS.tell(); 356 357 // The section size. We don't know the size yet, so reserve enough space 358 // for any 32-bit value; we'll patch it later. 359 encodeULEB128(0, W->OS, 5); 360 361 // The position where the section starts, for measuring its size. 362 Section.ContentsOffset = W->OS.tell(); 363 Section.PayloadOffset = W->OS.tell(); 364 Section.Index = SectionCount++; 365 } 366 367 // Write a string with extra paddings for trailing alignment 368 // TODO: support alignment at asm and llvm level? 369 void WasmObjectWriter::writeStringWithAlignment(const StringRef Str, 370 unsigned Alignment) { 371 372 // Calculate the encoded size of str length and add pads based on it and 373 // alignment. 374 raw_null_ostream NullOS; 375 uint64_t StrSizeLength = encodeULEB128(Str.size(), NullOS); 376 uint64_t Offset = W->OS.tell() + StrSizeLength + Str.size(); 377 uint64_t Paddings = offsetToAlignment(Offset, Align(Alignment)); 378 Offset += Paddings; 379 380 // LEB128 greater than 5 bytes is invalid 381 assert((StrSizeLength + Paddings) <= 5 && "too long string to align"); 382 383 encodeSLEB128(Str.size(), W->OS, StrSizeLength + Paddings); 384 W->OS << Str; 385 386 assert(W->OS.tell() == Offset && "invalid padding"); 387 } 388 389 void WasmObjectWriter::startCustomSection(SectionBookkeeping &Section, 390 StringRef Name) { 391 LLVM_DEBUG(dbgs() << "startCustomSection " << Name << "\n"); 392 startSection(Section, wasm::WASM_SEC_CUSTOM); 393 394 // The position where the section header ends, for measuring its size. 395 Section.PayloadOffset = W->OS.tell(); 396 397 // Custom sections in wasm also have a string identifier. 398 if (Name != "__clangast") { 399 writeString(Name); 400 } else { 401 // The on-disk hashtable in clangast needs to be aligned by 4 bytes. 402 writeStringWithAlignment(Name, 4); 403 } 404 405 // The position where the custom section starts. 406 Section.ContentsOffset = W->OS.tell(); 407 } 408 409 // Now that the section is complete and we know how big it is, patch up the 410 // section size field at the start of the section. 411 void WasmObjectWriter::endSection(SectionBookkeeping &Section) { 412 uint64_t Size = W->OS.tell(); 413 // /dev/null doesn't support seek/tell and can report offset of 0. 414 // Simply skip this patching in that case. 415 if (!Size) 416 return; 417 418 Size -= Section.PayloadOffset; 419 if (uint32_t(Size) != Size) 420 report_fatal_error("section size does not fit in a uint32_t"); 421 422 LLVM_DEBUG(dbgs() << "endSection size=" << Size << "\n"); 423 424 // Write the final section size to the payload_len field, which follows 425 // the section id byte. 426 writePatchableLEB<5>(static_cast<raw_pwrite_stream &>(W->OS), Size, 427 Section.SizeOffset); 428 } 429 430 // Emit the Wasm header. 431 void WasmObjectWriter::writeHeader(const MCAssembler &Asm) { 432 W->OS.write(wasm::WasmMagic, sizeof(wasm::WasmMagic)); 433 W->write<uint32_t>(wasm::WasmVersion); 434 } 435 436 void WasmObjectWriter::executePostLayoutBinding(MCAssembler &Asm, 437 const MCAsmLayout &Layout) { 438 // Some compilation units require the indirect function table to be present 439 // but don't explicitly reference it. This is the case for call_indirect 440 // without the reference-types feature, and also function bitcasts in all 441 // cases. In those cases the __indirect_function_table has the 442 // WASM_SYMBOL_NO_STRIP attribute. Here we make sure this symbol makes it to 443 // the assembler, if needed. 444 if (auto *Sym = Asm.getContext().lookupSymbol("__indirect_function_table")) { 445 const auto *WasmSym = static_cast<const MCSymbolWasm *>(Sym); 446 if (WasmSym->isNoStrip()) 447 Asm.registerSymbol(*Sym); 448 } 449 450 // Build a map of sections to the function that defines them, for use 451 // in recordRelocation. 452 for (const MCSymbol &S : Asm.symbols()) { 453 const auto &WS = static_cast<const MCSymbolWasm &>(S); 454 if (WS.isDefined() && WS.isFunction() && !WS.isVariable()) { 455 const auto &Sec = static_cast<const MCSectionWasm &>(S.getSection()); 456 auto Pair = SectionFunctions.insert(std::make_pair(&Sec, &S)); 457 if (!Pair.second) 458 report_fatal_error("section already has a defining function: " + 459 Sec.getName()); 460 } 461 } 462 } 463 464 void WasmObjectWriter::recordRelocation(MCAssembler &Asm, 465 const MCAsmLayout &Layout, 466 const MCFragment *Fragment, 467 const MCFixup &Fixup, MCValue Target, 468 uint64_t &FixedValue) { 469 // The WebAssembly backend should never generate FKF_IsPCRel fixups 470 assert(!(Asm.getBackend().getFixupKindInfo(Fixup.getKind()).Flags & 471 MCFixupKindInfo::FKF_IsPCRel)); 472 473 const auto &FixupSection = cast<MCSectionWasm>(*Fragment->getParent()); 474 uint64_t C = Target.getConstant(); 475 uint64_t FixupOffset = Layout.getFragmentOffset(Fragment) + Fixup.getOffset(); 476 MCContext &Ctx = Asm.getContext(); 477 bool IsLocRel = false; 478 479 if (const MCSymbolRefExpr *RefB = Target.getSymB()) { 480 481 const auto &SymB = cast<MCSymbolWasm>(RefB->getSymbol()); 482 483 if (FixupSection.getKind().isText()) { 484 Ctx.reportError(Fixup.getLoc(), 485 Twine("symbol '") + SymB.getName() + 486 "' unsupported subtraction expression used in " 487 "relocation in code section."); 488 return; 489 } 490 491 if (SymB.isUndefined()) { 492 Ctx.reportError(Fixup.getLoc(), 493 Twine("symbol '") + SymB.getName() + 494 "' can not be undefined in a subtraction expression"); 495 return; 496 } 497 const MCSection &SecB = SymB.getSection(); 498 if (&SecB != &FixupSection) { 499 Ctx.reportError(Fixup.getLoc(), 500 Twine("symbol '") + SymB.getName() + 501 "' can not be placed in a different section"); 502 return; 503 } 504 IsLocRel = true; 505 C += FixupOffset - Layout.getSymbolOffset(SymB); 506 } 507 508 // We either rejected the fixup or folded B into C at this point. 509 const MCSymbolRefExpr *RefA = Target.getSymA(); 510 const auto *SymA = cast<MCSymbolWasm>(&RefA->getSymbol()); 511 512 // The .init_array isn't translated as data, so don't do relocations in it. 513 if (FixupSection.getName().startswith(".init_array")) { 514 SymA->setUsedInInitArray(); 515 return; 516 } 517 518 if (SymA->isVariable()) { 519 const MCExpr *Expr = SymA->getVariableValue(); 520 if (const auto *Inner = dyn_cast<MCSymbolRefExpr>(Expr)) 521 if (Inner->getKind() == MCSymbolRefExpr::VK_WEAKREF) 522 llvm_unreachable("weakref used in reloc not yet implemented"); 523 } 524 525 // Put any constant offset in an addend. Offsets can be negative, and 526 // LLVM expects wrapping, in contrast to wasm's immediates which can't 527 // be negative and don't wrap. 528 FixedValue = 0; 529 530 unsigned Type = 531 TargetObjectWriter->getRelocType(Target, Fixup, FixupSection, IsLocRel); 532 533 // Absolute offset within a section or a function. 534 // Currently only supported for for metadata sections. 535 // See: test/MC/WebAssembly/blockaddress.ll 536 if ((Type == wasm::R_WASM_FUNCTION_OFFSET_I32 || 537 Type == wasm::R_WASM_FUNCTION_OFFSET_I64 || 538 Type == wasm::R_WASM_SECTION_OFFSET_I32) && 539 SymA->isDefined()) { 540 // SymA can be a temp data symbol that represents a function (in which case 541 // it needs to be replaced by the section symbol), [XXX and it apparently 542 // later gets changed again to a func symbol?] or it can be a real 543 // function symbol, in which case it can be left as-is. 544 545 if (!FixupSection.getKind().isMetadata()) 546 report_fatal_error("relocations for function or section offsets are " 547 "only supported in metadata sections"); 548 549 const MCSymbol *SectionSymbol = nullptr; 550 const MCSection &SecA = SymA->getSection(); 551 if (SecA.getKind().isText()) { 552 auto SecSymIt = SectionFunctions.find(&SecA); 553 if (SecSymIt == SectionFunctions.end()) 554 report_fatal_error("section doesn\'t have defining symbol"); 555 SectionSymbol = SecSymIt->second; 556 } else { 557 SectionSymbol = SecA.getBeginSymbol(); 558 } 559 if (!SectionSymbol) 560 report_fatal_error("section symbol is required for relocation"); 561 562 C += Layout.getSymbolOffset(*SymA); 563 SymA = cast<MCSymbolWasm>(SectionSymbol); 564 } 565 566 if (Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB || 567 Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB64 || 568 Type == wasm::R_WASM_TABLE_INDEX_SLEB || 569 Type == wasm::R_WASM_TABLE_INDEX_SLEB64 || 570 Type == wasm::R_WASM_TABLE_INDEX_I32 || 571 Type == wasm::R_WASM_TABLE_INDEX_I64) { 572 // TABLE_INDEX relocs implicitly use the default indirect function table. 573 // We require the function table to have already been defined. 574 auto TableName = "__indirect_function_table"; 575 MCSymbolWasm *Sym = cast_or_null<MCSymbolWasm>(Ctx.lookupSymbol(TableName)); 576 if (!Sym) { 577 report_fatal_error("missing indirect function table symbol"); 578 } else { 579 if (!Sym->isFunctionTable()) 580 report_fatal_error("__indirect_function_table symbol has wrong type"); 581 // Ensure that __indirect_function_table reaches the output. 582 Sym->setNoStrip(); 583 Asm.registerSymbol(*Sym); 584 } 585 } 586 587 // Relocation other than R_WASM_TYPE_INDEX_LEB are required to be 588 // against a named symbol. 589 if (Type != wasm::R_WASM_TYPE_INDEX_LEB) { 590 if (SymA->getName().empty()) 591 report_fatal_error("relocations against un-named temporaries are not yet " 592 "supported by wasm"); 593 594 SymA->setUsedInReloc(); 595 } 596 597 switch (RefA->getKind()) { 598 case MCSymbolRefExpr::VK_GOT: 599 case MCSymbolRefExpr::VK_WASM_GOT_TLS: 600 SymA->setUsedInGOT(); 601 break; 602 default: 603 break; 604 } 605 606 WasmRelocationEntry Rec(FixupOffset, SymA, C, Type, &FixupSection); 607 LLVM_DEBUG(dbgs() << "WasmReloc: " << Rec << "\n"); 608 609 if (FixupSection.isWasmData()) { 610 DataRelocations.push_back(Rec); 611 } else if (FixupSection.getKind().isText()) { 612 CodeRelocations.push_back(Rec); 613 } else if (FixupSection.getKind().isMetadata()) { 614 CustomSectionsRelocations[&FixupSection].push_back(Rec); 615 } else { 616 llvm_unreachable("unexpected section type"); 617 } 618 } 619 620 // Compute a value to write into the code at the location covered 621 // by RelEntry. This value isn't used by the static linker; it just serves 622 // to make the object format more readable and more likely to be directly 623 // useable. 624 uint64_t 625 WasmObjectWriter::getProvisionalValue(const WasmRelocationEntry &RelEntry, 626 const MCAsmLayout &Layout) { 627 if ((RelEntry.Type == wasm::R_WASM_GLOBAL_INDEX_LEB || 628 RelEntry.Type == wasm::R_WASM_GLOBAL_INDEX_I32) && 629 !RelEntry.Symbol->isGlobal()) { 630 assert(GOTIndices.count(RelEntry.Symbol) > 0 && "symbol not found in GOT index space"); 631 return GOTIndices[RelEntry.Symbol]; 632 } 633 634 switch (RelEntry.Type) { 635 case wasm::R_WASM_TABLE_INDEX_REL_SLEB: 636 case wasm::R_WASM_TABLE_INDEX_REL_SLEB64: 637 case wasm::R_WASM_TABLE_INDEX_SLEB: 638 case wasm::R_WASM_TABLE_INDEX_SLEB64: 639 case wasm::R_WASM_TABLE_INDEX_I32: 640 case wasm::R_WASM_TABLE_INDEX_I64: { 641 // Provisional value is table address of the resolved symbol itself 642 const MCSymbolWasm *Base = 643 cast<MCSymbolWasm>(Layout.getBaseSymbol(*RelEntry.Symbol)); 644 assert(Base->isFunction()); 645 if (RelEntry.Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB || 646 RelEntry.Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB64) 647 return TableIndices[Base] - InitialTableOffset; 648 else 649 return TableIndices[Base]; 650 } 651 case wasm::R_WASM_TYPE_INDEX_LEB: 652 // Provisional value is same as the index 653 return getRelocationIndexValue(RelEntry); 654 case wasm::R_WASM_FUNCTION_INDEX_LEB: 655 case wasm::R_WASM_GLOBAL_INDEX_LEB: 656 case wasm::R_WASM_GLOBAL_INDEX_I32: 657 case wasm::R_WASM_TAG_INDEX_LEB: 658 case wasm::R_WASM_TABLE_NUMBER_LEB: 659 // Provisional value is function/global/tag Wasm index 660 assert(WasmIndices.count(RelEntry.Symbol) > 0 && "symbol not found in wasm index space"); 661 return WasmIndices[RelEntry.Symbol]; 662 case wasm::R_WASM_FUNCTION_OFFSET_I32: 663 case wasm::R_WASM_FUNCTION_OFFSET_I64: 664 case wasm::R_WASM_SECTION_OFFSET_I32: { 665 if (!RelEntry.Symbol->isDefined()) 666 return 0; 667 const auto &Section = 668 static_cast<const MCSectionWasm &>(RelEntry.Symbol->getSection()); 669 return Section.getSectionOffset() + RelEntry.Addend; 670 } 671 case wasm::R_WASM_MEMORY_ADDR_LEB: 672 case wasm::R_WASM_MEMORY_ADDR_LEB64: 673 case wasm::R_WASM_MEMORY_ADDR_SLEB: 674 case wasm::R_WASM_MEMORY_ADDR_SLEB64: 675 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB: 676 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB64: 677 case wasm::R_WASM_MEMORY_ADDR_I32: 678 case wasm::R_WASM_MEMORY_ADDR_I64: 679 case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB: 680 case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB64: 681 case wasm::R_WASM_MEMORY_ADDR_LOCREL_I32: { 682 // Provisional value is address of the global plus the offset 683 // For undefined symbols, use zero 684 if (!RelEntry.Symbol->isDefined()) 685 return 0; 686 const wasm::WasmDataReference &SymRef = DataLocations[RelEntry.Symbol]; 687 const WasmDataSegment &Segment = DataSegments[SymRef.Segment]; 688 // Ignore overflow. LLVM allows address arithmetic to silently wrap. 689 return Segment.Offset + SymRef.Offset + RelEntry.Addend; 690 } 691 default: 692 llvm_unreachable("invalid relocation type"); 693 } 694 } 695 696 static void addData(SmallVectorImpl<char> &DataBytes, 697 MCSectionWasm &DataSection) { 698 LLVM_DEBUG(errs() << "addData: " << DataSection.getName() << "\n"); 699 700 DataBytes.resize(alignTo(DataBytes.size(), DataSection.getAlignment())); 701 702 for (const MCFragment &Frag : DataSection) { 703 if (Frag.hasInstructions()) 704 report_fatal_error("only data supported in data sections"); 705 706 if (auto *Align = dyn_cast<MCAlignFragment>(&Frag)) { 707 if (Align->getValueSize() != 1) 708 report_fatal_error("only byte values supported for alignment"); 709 // If nops are requested, use zeros, as this is the data section. 710 uint8_t Value = Align->hasEmitNops() ? 0 : Align->getValue(); 711 uint64_t Size = 712 std::min<uint64_t>(alignTo(DataBytes.size(), Align->getAlignment()), 713 DataBytes.size() + Align->getMaxBytesToEmit()); 714 DataBytes.resize(Size, Value); 715 } else if (auto *Fill = dyn_cast<MCFillFragment>(&Frag)) { 716 int64_t NumValues; 717 if (!Fill->getNumValues().evaluateAsAbsolute(NumValues)) 718 llvm_unreachable("The fill should be an assembler constant"); 719 DataBytes.insert(DataBytes.end(), Fill->getValueSize() * NumValues, 720 Fill->getValue()); 721 } else if (auto *LEB = dyn_cast<MCLEBFragment>(&Frag)) { 722 const SmallVectorImpl<char> &Contents = LEB->getContents(); 723 llvm::append_range(DataBytes, Contents); 724 } else { 725 const auto &DataFrag = cast<MCDataFragment>(Frag); 726 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 727 llvm::append_range(DataBytes, Contents); 728 } 729 } 730 731 LLVM_DEBUG(dbgs() << "addData -> " << DataBytes.size() << "\n"); 732 } 733 734 uint32_t 735 WasmObjectWriter::getRelocationIndexValue(const WasmRelocationEntry &RelEntry) { 736 if (RelEntry.Type == wasm::R_WASM_TYPE_INDEX_LEB) { 737 if (!TypeIndices.count(RelEntry.Symbol)) 738 report_fatal_error("symbol not found in type index space: " + 739 RelEntry.Symbol->getName()); 740 return TypeIndices[RelEntry.Symbol]; 741 } 742 743 return RelEntry.Symbol->getIndex(); 744 } 745 746 // Apply the portions of the relocation records that we can handle ourselves 747 // directly. 748 void WasmObjectWriter::applyRelocations( 749 ArrayRef<WasmRelocationEntry> Relocations, uint64_t ContentsOffset, 750 const MCAsmLayout &Layout) { 751 auto &Stream = static_cast<raw_pwrite_stream &>(W->OS); 752 for (const WasmRelocationEntry &RelEntry : Relocations) { 753 uint64_t Offset = ContentsOffset + 754 RelEntry.FixupSection->getSectionOffset() + 755 RelEntry.Offset; 756 757 LLVM_DEBUG(dbgs() << "applyRelocation: " << RelEntry << "\n"); 758 auto Value = getProvisionalValue(RelEntry, Layout); 759 760 switch (RelEntry.Type) { 761 case wasm::R_WASM_FUNCTION_INDEX_LEB: 762 case wasm::R_WASM_TYPE_INDEX_LEB: 763 case wasm::R_WASM_GLOBAL_INDEX_LEB: 764 case wasm::R_WASM_MEMORY_ADDR_LEB: 765 case wasm::R_WASM_TAG_INDEX_LEB: 766 case wasm::R_WASM_TABLE_NUMBER_LEB: 767 writePatchableLEB<5>(Stream, Value, Offset); 768 break; 769 case wasm::R_WASM_MEMORY_ADDR_LEB64: 770 writePatchableLEB<10>(Stream, Value, Offset); 771 break; 772 case wasm::R_WASM_TABLE_INDEX_I32: 773 case wasm::R_WASM_MEMORY_ADDR_I32: 774 case wasm::R_WASM_FUNCTION_OFFSET_I32: 775 case wasm::R_WASM_SECTION_OFFSET_I32: 776 case wasm::R_WASM_GLOBAL_INDEX_I32: 777 case wasm::R_WASM_MEMORY_ADDR_LOCREL_I32: 778 patchI32(Stream, Value, Offset); 779 break; 780 case wasm::R_WASM_TABLE_INDEX_I64: 781 case wasm::R_WASM_MEMORY_ADDR_I64: 782 case wasm::R_WASM_FUNCTION_OFFSET_I64: 783 patchI64(Stream, Value, Offset); 784 break; 785 case wasm::R_WASM_TABLE_INDEX_SLEB: 786 case wasm::R_WASM_TABLE_INDEX_REL_SLEB: 787 case wasm::R_WASM_MEMORY_ADDR_SLEB: 788 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB: 789 case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB: 790 writePatchableSLEB<5>(Stream, Value, Offset); 791 break; 792 case wasm::R_WASM_TABLE_INDEX_SLEB64: 793 case wasm::R_WASM_TABLE_INDEX_REL_SLEB64: 794 case wasm::R_WASM_MEMORY_ADDR_SLEB64: 795 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB64: 796 case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB64: 797 writePatchableSLEB<10>(Stream, Value, Offset); 798 break; 799 default: 800 llvm_unreachable("invalid relocation type"); 801 } 802 } 803 } 804 805 void WasmObjectWriter::writeTypeSection( 806 ArrayRef<wasm::WasmSignature> Signatures) { 807 if (Signatures.empty()) 808 return; 809 810 SectionBookkeeping Section; 811 startSection(Section, wasm::WASM_SEC_TYPE); 812 813 encodeULEB128(Signatures.size(), W->OS); 814 815 for (const wasm::WasmSignature &Sig : Signatures) { 816 W->OS << char(wasm::WASM_TYPE_FUNC); 817 encodeULEB128(Sig.Params.size(), W->OS); 818 for (wasm::ValType Ty : Sig.Params) 819 writeValueType(Ty); 820 encodeULEB128(Sig.Returns.size(), W->OS); 821 for (wasm::ValType Ty : Sig.Returns) 822 writeValueType(Ty); 823 } 824 825 endSection(Section); 826 } 827 828 void WasmObjectWriter::writeImportSection(ArrayRef<wasm::WasmImport> Imports, 829 uint64_t DataSize, 830 uint32_t NumElements) { 831 if (Imports.empty()) 832 return; 833 834 uint64_t NumPages = (DataSize + wasm::WasmPageSize - 1) / wasm::WasmPageSize; 835 836 SectionBookkeeping Section; 837 startSection(Section, wasm::WASM_SEC_IMPORT); 838 839 encodeULEB128(Imports.size(), W->OS); 840 for (const wasm::WasmImport &Import : Imports) { 841 writeString(Import.Module); 842 writeString(Import.Field); 843 W->OS << char(Import.Kind); 844 845 switch (Import.Kind) { 846 case wasm::WASM_EXTERNAL_FUNCTION: 847 encodeULEB128(Import.SigIndex, W->OS); 848 break; 849 case wasm::WASM_EXTERNAL_GLOBAL: 850 W->OS << char(Import.Global.Type); 851 W->OS << char(Import.Global.Mutable ? 1 : 0); 852 break; 853 case wasm::WASM_EXTERNAL_MEMORY: 854 encodeULEB128(Import.Memory.Flags, W->OS); 855 encodeULEB128(NumPages, W->OS); // initial 856 break; 857 case wasm::WASM_EXTERNAL_TABLE: 858 W->OS << char(Import.Table.ElemType); 859 encodeULEB128(0, W->OS); // flags 860 encodeULEB128(NumElements, W->OS); // initial 861 break; 862 case wasm::WASM_EXTERNAL_TAG: 863 W->OS << char(0); // Reserved 'attribute' field 864 encodeULEB128(Import.SigIndex, W->OS); 865 break; 866 default: 867 llvm_unreachable("unsupported import kind"); 868 } 869 } 870 871 endSection(Section); 872 } 873 874 void WasmObjectWriter::writeFunctionSection(ArrayRef<WasmFunction> Functions) { 875 if (Functions.empty()) 876 return; 877 878 SectionBookkeeping Section; 879 startSection(Section, wasm::WASM_SEC_FUNCTION); 880 881 encodeULEB128(Functions.size(), W->OS); 882 for (const WasmFunction &Func : Functions) 883 encodeULEB128(Func.SigIndex, W->OS); 884 885 endSection(Section); 886 } 887 888 void WasmObjectWriter::writeTagSection(ArrayRef<uint32_t> TagTypes) { 889 if (TagTypes.empty()) 890 return; 891 892 SectionBookkeeping Section; 893 startSection(Section, wasm::WASM_SEC_TAG); 894 895 encodeULEB128(TagTypes.size(), W->OS); 896 for (uint32_t Index : TagTypes) { 897 W->OS << char(0); // Reserved 'attribute' field 898 encodeULEB128(Index, W->OS); 899 } 900 901 endSection(Section); 902 } 903 904 void WasmObjectWriter::writeGlobalSection(ArrayRef<wasm::WasmGlobal> Globals) { 905 if (Globals.empty()) 906 return; 907 908 SectionBookkeeping Section; 909 startSection(Section, wasm::WASM_SEC_GLOBAL); 910 911 encodeULEB128(Globals.size(), W->OS); 912 for (const wasm::WasmGlobal &Global : Globals) { 913 encodeULEB128(Global.Type.Type, W->OS); 914 W->OS << char(Global.Type.Mutable); 915 W->OS << char(Global.InitExpr.Opcode); 916 switch (Global.Type.Type) { 917 case wasm::WASM_TYPE_I32: 918 encodeSLEB128(0, W->OS); 919 break; 920 case wasm::WASM_TYPE_I64: 921 encodeSLEB128(0, W->OS); 922 break; 923 case wasm::WASM_TYPE_F32: 924 writeI32(0); 925 break; 926 case wasm::WASM_TYPE_F64: 927 writeI64(0); 928 break; 929 case wasm::WASM_TYPE_EXTERNREF: 930 writeValueType(wasm::ValType::EXTERNREF); 931 break; 932 default: 933 llvm_unreachable("unexpected type"); 934 } 935 W->OS << char(wasm::WASM_OPCODE_END); 936 } 937 938 endSection(Section); 939 } 940 941 void WasmObjectWriter::writeTableSection(ArrayRef<wasm::WasmTable> Tables) { 942 if (Tables.empty()) 943 return; 944 945 SectionBookkeeping Section; 946 startSection(Section, wasm::WASM_SEC_TABLE); 947 948 encodeULEB128(Tables.size(), W->OS); 949 for (const wasm::WasmTable &Table : Tables) { 950 encodeULEB128(Table.Type.ElemType, W->OS); 951 encodeULEB128(Table.Type.Limits.Flags, W->OS); 952 encodeULEB128(Table.Type.Limits.Minimum, W->OS); 953 if (Table.Type.Limits.Flags & wasm::WASM_LIMITS_FLAG_HAS_MAX) 954 encodeULEB128(Table.Type.Limits.Maximum, W->OS); 955 } 956 endSection(Section); 957 } 958 959 void WasmObjectWriter::writeExportSection(ArrayRef<wasm::WasmExport> Exports) { 960 if (Exports.empty()) 961 return; 962 963 SectionBookkeeping Section; 964 startSection(Section, wasm::WASM_SEC_EXPORT); 965 966 encodeULEB128(Exports.size(), W->OS); 967 for (const wasm::WasmExport &Export : Exports) { 968 writeString(Export.Name); 969 W->OS << char(Export.Kind); 970 encodeULEB128(Export.Index, W->OS); 971 } 972 973 endSection(Section); 974 } 975 976 void WasmObjectWriter::writeElemSection( 977 const MCSymbolWasm *IndirectFunctionTable, ArrayRef<uint32_t> TableElems) { 978 if (TableElems.empty()) 979 return; 980 981 assert(IndirectFunctionTable); 982 983 SectionBookkeeping Section; 984 startSection(Section, wasm::WASM_SEC_ELEM); 985 986 encodeULEB128(1, W->OS); // number of "segments" 987 988 assert(WasmIndices.count(IndirectFunctionTable)); 989 uint32_t TableNumber = WasmIndices.find(IndirectFunctionTable)->second; 990 uint32_t Flags = 0; 991 if (TableNumber) 992 Flags |= wasm::WASM_ELEM_SEGMENT_HAS_TABLE_NUMBER; 993 encodeULEB128(Flags, W->OS); 994 if (Flags & wasm::WASM_ELEM_SEGMENT_HAS_TABLE_NUMBER) 995 encodeULEB128(TableNumber, W->OS); // the table number 996 997 // init expr for starting offset 998 W->OS << char(wasm::WASM_OPCODE_I32_CONST); 999 encodeSLEB128(InitialTableOffset, W->OS); 1000 W->OS << char(wasm::WASM_OPCODE_END); 1001 1002 if (Flags & wasm::WASM_ELEM_SEGMENT_MASK_HAS_ELEM_KIND) { 1003 // We only write active function table initializers, for which the elem kind 1004 // is specified to be written as 0x00 and interpreted to mean "funcref". 1005 const uint8_t ElemKind = 0; 1006 W->OS << ElemKind; 1007 } 1008 1009 encodeULEB128(TableElems.size(), W->OS); 1010 for (uint32_t Elem : TableElems) 1011 encodeULEB128(Elem, W->OS); 1012 1013 endSection(Section); 1014 } 1015 1016 void WasmObjectWriter::writeDataCountSection() { 1017 if (DataSegments.empty()) 1018 return; 1019 1020 SectionBookkeeping Section; 1021 startSection(Section, wasm::WASM_SEC_DATACOUNT); 1022 encodeULEB128(DataSegments.size(), W->OS); 1023 endSection(Section); 1024 } 1025 1026 uint32_t WasmObjectWriter::writeCodeSection(const MCAssembler &Asm, 1027 const MCAsmLayout &Layout, 1028 ArrayRef<WasmFunction> Functions) { 1029 if (Functions.empty()) 1030 return 0; 1031 1032 SectionBookkeeping Section; 1033 startSection(Section, wasm::WASM_SEC_CODE); 1034 1035 encodeULEB128(Functions.size(), W->OS); 1036 1037 for (const WasmFunction &Func : Functions) { 1038 auto &FuncSection = static_cast<MCSectionWasm &>(Func.Sym->getSection()); 1039 1040 int64_t Size = 0; 1041 if (!Func.Sym->getSize()->evaluateAsAbsolute(Size, Layout)) 1042 report_fatal_error(".size expression must be evaluatable"); 1043 1044 encodeULEB128(Size, W->OS); 1045 FuncSection.setSectionOffset(W->OS.tell() - Section.ContentsOffset); 1046 Asm.writeSectionData(W->OS, &FuncSection, Layout); 1047 } 1048 1049 // Apply fixups. 1050 applyRelocations(CodeRelocations, Section.ContentsOffset, Layout); 1051 1052 endSection(Section); 1053 return Section.Index; 1054 } 1055 1056 uint32_t WasmObjectWriter::writeDataSection(const MCAsmLayout &Layout) { 1057 if (DataSegments.empty()) 1058 return 0; 1059 1060 SectionBookkeeping Section; 1061 startSection(Section, wasm::WASM_SEC_DATA); 1062 1063 encodeULEB128(DataSegments.size(), W->OS); // count 1064 1065 for (const WasmDataSegment &Segment : DataSegments) { 1066 encodeULEB128(Segment.InitFlags, W->OS); // flags 1067 if (Segment.InitFlags & wasm::WASM_DATA_SEGMENT_HAS_MEMINDEX) 1068 encodeULEB128(0, W->OS); // memory index 1069 if ((Segment.InitFlags & wasm::WASM_DATA_SEGMENT_IS_PASSIVE) == 0) { 1070 W->OS << char(is64Bit() ? wasm::WASM_OPCODE_I64_CONST 1071 : wasm::WASM_OPCODE_I32_CONST); 1072 encodeSLEB128(Segment.Offset, W->OS); // offset 1073 W->OS << char(wasm::WASM_OPCODE_END); 1074 } 1075 encodeULEB128(Segment.Data.size(), W->OS); // size 1076 Segment.Section->setSectionOffset(W->OS.tell() - Section.ContentsOffset); 1077 W->OS << Segment.Data; // data 1078 } 1079 1080 // Apply fixups. 1081 applyRelocations(DataRelocations, Section.ContentsOffset, Layout); 1082 1083 endSection(Section); 1084 return Section.Index; 1085 } 1086 1087 void WasmObjectWriter::writeRelocSection( 1088 uint32_t SectionIndex, StringRef Name, 1089 std::vector<WasmRelocationEntry> &Relocs) { 1090 // See: https://github.com/WebAssembly/tool-conventions/blob/main/Linking.md 1091 // for descriptions of the reloc sections. 1092 1093 if (Relocs.empty()) 1094 return; 1095 1096 // First, ensure the relocations are sorted in offset order. In general they 1097 // should already be sorted since `recordRelocation` is called in offset 1098 // order, but for the code section we combine many MC sections into single 1099 // wasm section, and this order is determined by the order of Asm.Symbols() 1100 // not the sections order. 1101 llvm::stable_sort( 1102 Relocs, [](const WasmRelocationEntry &A, const WasmRelocationEntry &B) { 1103 return (A.Offset + A.FixupSection->getSectionOffset()) < 1104 (B.Offset + B.FixupSection->getSectionOffset()); 1105 }); 1106 1107 SectionBookkeeping Section; 1108 startCustomSection(Section, std::string("reloc.") + Name.str()); 1109 1110 encodeULEB128(SectionIndex, W->OS); 1111 encodeULEB128(Relocs.size(), W->OS); 1112 for (const WasmRelocationEntry &RelEntry : Relocs) { 1113 uint64_t Offset = 1114 RelEntry.Offset + RelEntry.FixupSection->getSectionOffset(); 1115 uint32_t Index = getRelocationIndexValue(RelEntry); 1116 1117 W->OS << char(RelEntry.Type); 1118 encodeULEB128(Offset, W->OS); 1119 encodeULEB128(Index, W->OS); 1120 if (RelEntry.hasAddend()) 1121 encodeSLEB128(RelEntry.Addend, W->OS); 1122 } 1123 1124 endSection(Section); 1125 } 1126 1127 void WasmObjectWriter::writeCustomRelocSections() { 1128 for (const auto &Sec : CustomSections) { 1129 auto &Relocations = CustomSectionsRelocations[Sec.Section]; 1130 writeRelocSection(Sec.OutputIndex, Sec.Name, Relocations); 1131 } 1132 } 1133 1134 void WasmObjectWriter::writeLinkingMetaDataSection( 1135 ArrayRef<wasm::WasmSymbolInfo> SymbolInfos, 1136 ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs, 1137 const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats) { 1138 SectionBookkeeping Section; 1139 startCustomSection(Section, "linking"); 1140 encodeULEB128(wasm::WasmMetadataVersion, W->OS); 1141 1142 SectionBookkeeping SubSection; 1143 if (SymbolInfos.size() != 0) { 1144 startSection(SubSection, wasm::WASM_SYMBOL_TABLE); 1145 encodeULEB128(SymbolInfos.size(), W->OS); 1146 for (const wasm::WasmSymbolInfo &Sym : SymbolInfos) { 1147 encodeULEB128(Sym.Kind, W->OS); 1148 encodeULEB128(Sym.Flags, W->OS); 1149 switch (Sym.Kind) { 1150 case wasm::WASM_SYMBOL_TYPE_FUNCTION: 1151 case wasm::WASM_SYMBOL_TYPE_GLOBAL: 1152 case wasm::WASM_SYMBOL_TYPE_TAG: 1153 case wasm::WASM_SYMBOL_TYPE_TABLE: 1154 encodeULEB128(Sym.ElementIndex, W->OS); 1155 if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0 || 1156 (Sym.Flags & wasm::WASM_SYMBOL_EXPLICIT_NAME) != 0) 1157 writeString(Sym.Name); 1158 break; 1159 case wasm::WASM_SYMBOL_TYPE_DATA: 1160 writeString(Sym.Name); 1161 if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0) { 1162 encodeULEB128(Sym.DataRef.Segment, W->OS); 1163 encodeULEB128(Sym.DataRef.Offset, W->OS); 1164 encodeULEB128(Sym.DataRef.Size, W->OS); 1165 } 1166 break; 1167 case wasm::WASM_SYMBOL_TYPE_SECTION: { 1168 const uint32_t SectionIndex = 1169 CustomSections[Sym.ElementIndex].OutputIndex; 1170 encodeULEB128(SectionIndex, W->OS); 1171 break; 1172 } 1173 default: 1174 llvm_unreachable("unexpected kind"); 1175 } 1176 } 1177 endSection(SubSection); 1178 } 1179 1180 if (DataSegments.size()) { 1181 startSection(SubSection, wasm::WASM_SEGMENT_INFO); 1182 encodeULEB128(DataSegments.size(), W->OS); 1183 for (const WasmDataSegment &Segment : DataSegments) { 1184 writeString(Segment.Name); 1185 encodeULEB128(Segment.Alignment, W->OS); 1186 encodeULEB128(Segment.LinkingFlags, W->OS); 1187 } 1188 endSection(SubSection); 1189 } 1190 1191 if (!InitFuncs.empty()) { 1192 startSection(SubSection, wasm::WASM_INIT_FUNCS); 1193 encodeULEB128(InitFuncs.size(), W->OS); 1194 for (auto &StartFunc : InitFuncs) { 1195 encodeULEB128(StartFunc.first, W->OS); // priority 1196 encodeULEB128(StartFunc.second, W->OS); // function index 1197 } 1198 endSection(SubSection); 1199 } 1200 1201 if (Comdats.size()) { 1202 startSection(SubSection, wasm::WASM_COMDAT_INFO); 1203 encodeULEB128(Comdats.size(), W->OS); 1204 for (const auto &C : Comdats) { 1205 writeString(C.first); 1206 encodeULEB128(0, W->OS); // flags for future use 1207 encodeULEB128(C.second.size(), W->OS); 1208 for (const WasmComdatEntry &Entry : C.second) { 1209 encodeULEB128(Entry.Kind, W->OS); 1210 encodeULEB128(Entry.Index, W->OS); 1211 } 1212 } 1213 endSection(SubSection); 1214 } 1215 1216 endSection(Section); 1217 } 1218 1219 void WasmObjectWriter::writeCustomSection(WasmCustomSection &CustomSection, 1220 const MCAssembler &Asm, 1221 const MCAsmLayout &Layout) { 1222 SectionBookkeeping Section; 1223 auto *Sec = CustomSection.Section; 1224 startCustomSection(Section, CustomSection.Name); 1225 1226 Sec->setSectionOffset(W->OS.tell() - Section.ContentsOffset); 1227 Asm.writeSectionData(W->OS, Sec, Layout); 1228 1229 CustomSection.OutputContentsOffset = Section.ContentsOffset; 1230 CustomSection.OutputIndex = Section.Index; 1231 1232 endSection(Section); 1233 1234 // Apply fixups. 1235 auto &Relocations = CustomSectionsRelocations[CustomSection.Section]; 1236 applyRelocations(Relocations, CustomSection.OutputContentsOffset, Layout); 1237 } 1238 1239 uint32_t WasmObjectWriter::getFunctionType(const MCSymbolWasm &Symbol) { 1240 assert(Symbol.isFunction()); 1241 assert(TypeIndices.count(&Symbol)); 1242 return TypeIndices[&Symbol]; 1243 } 1244 1245 uint32_t WasmObjectWriter::getTagType(const MCSymbolWasm &Symbol) { 1246 assert(Symbol.isTag()); 1247 assert(TypeIndices.count(&Symbol)); 1248 return TypeIndices[&Symbol]; 1249 } 1250 1251 void WasmObjectWriter::registerFunctionType(const MCSymbolWasm &Symbol) { 1252 assert(Symbol.isFunction()); 1253 1254 wasm::WasmSignature S; 1255 1256 if (auto *Sig = Symbol.getSignature()) { 1257 S.Returns = Sig->Returns; 1258 S.Params = Sig->Params; 1259 } 1260 1261 auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size())); 1262 if (Pair.second) 1263 Signatures.push_back(S); 1264 TypeIndices[&Symbol] = Pair.first->second; 1265 1266 LLVM_DEBUG(dbgs() << "registerFunctionType: " << Symbol 1267 << " new:" << Pair.second << "\n"); 1268 LLVM_DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n"); 1269 } 1270 1271 void WasmObjectWriter::registerTagType(const MCSymbolWasm &Symbol) { 1272 assert(Symbol.isTag()); 1273 1274 // TODO Currently we don't generate imported exceptions, but if we do, we 1275 // should have a way of infering types of imported exceptions. 1276 wasm::WasmSignature S; 1277 if (auto *Sig = Symbol.getSignature()) { 1278 S.Returns = Sig->Returns; 1279 S.Params = Sig->Params; 1280 } 1281 1282 auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size())); 1283 if (Pair.second) 1284 Signatures.push_back(S); 1285 TypeIndices[&Symbol] = Pair.first->second; 1286 1287 LLVM_DEBUG(dbgs() << "registerTagType: " << Symbol << " new:" << Pair.second 1288 << "\n"); 1289 LLVM_DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n"); 1290 } 1291 1292 static bool isInSymtab(const MCSymbolWasm &Sym) { 1293 if (Sym.isUsedInReloc() || Sym.isUsedInInitArray()) 1294 return true; 1295 1296 if (Sym.isComdat() && !Sym.isDefined()) 1297 return false; 1298 1299 if (Sym.isTemporary()) 1300 return false; 1301 1302 if (Sym.isSection()) 1303 return false; 1304 1305 if (Sym.omitFromLinkingSection()) 1306 return false; 1307 1308 return true; 1309 } 1310 1311 void WasmObjectWriter::prepareImports( 1312 SmallVectorImpl<wasm::WasmImport> &Imports, MCAssembler &Asm, 1313 const MCAsmLayout &Layout) { 1314 // For now, always emit the memory import, since loads and stores are not 1315 // valid without it. In the future, we could perhaps be more clever and omit 1316 // it if there are no loads or stores. 1317 wasm::WasmImport MemImport; 1318 MemImport.Module = "env"; 1319 MemImport.Field = "__linear_memory"; 1320 MemImport.Kind = wasm::WASM_EXTERNAL_MEMORY; 1321 MemImport.Memory.Flags = is64Bit() ? wasm::WASM_LIMITS_FLAG_IS_64 1322 : wasm::WASM_LIMITS_FLAG_NONE; 1323 Imports.push_back(MemImport); 1324 1325 // Populate SignatureIndices, and Imports and WasmIndices for undefined 1326 // symbols. This must be done before populating WasmIndices for defined 1327 // symbols. 1328 for (const MCSymbol &S : Asm.symbols()) { 1329 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1330 1331 // Register types for all functions, including those with private linkage 1332 // (because wasm always needs a type signature). 1333 if (WS.isFunction()) { 1334 const auto *BS = Layout.getBaseSymbol(S); 1335 if (!BS) 1336 report_fatal_error(Twine(S.getName()) + 1337 ": absolute addressing not supported!"); 1338 registerFunctionType(*cast<MCSymbolWasm>(BS)); 1339 } 1340 1341 if (WS.isTag()) 1342 registerTagType(WS); 1343 1344 if (WS.isTemporary()) 1345 continue; 1346 1347 // If the symbol is not defined in this translation unit, import it. 1348 if (!WS.isDefined() && !WS.isComdat()) { 1349 if (WS.isFunction()) { 1350 wasm::WasmImport Import; 1351 Import.Module = WS.getImportModule(); 1352 Import.Field = WS.getImportName(); 1353 Import.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1354 Import.SigIndex = getFunctionType(WS); 1355 Imports.push_back(Import); 1356 assert(WasmIndices.count(&WS) == 0); 1357 WasmIndices[&WS] = NumFunctionImports++; 1358 } else if (WS.isGlobal()) { 1359 if (WS.isWeak()) 1360 report_fatal_error("undefined global symbol cannot be weak"); 1361 1362 wasm::WasmImport Import; 1363 Import.Field = WS.getImportName(); 1364 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1365 Import.Module = WS.getImportModule(); 1366 Import.Global = WS.getGlobalType(); 1367 Imports.push_back(Import); 1368 assert(WasmIndices.count(&WS) == 0); 1369 WasmIndices[&WS] = NumGlobalImports++; 1370 } else if (WS.isTag()) { 1371 if (WS.isWeak()) 1372 report_fatal_error("undefined tag symbol cannot be weak"); 1373 1374 wasm::WasmImport Import; 1375 Import.Module = WS.getImportModule(); 1376 Import.Field = WS.getImportName(); 1377 Import.Kind = wasm::WASM_EXTERNAL_TAG; 1378 Import.SigIndex = getTagType(WS); 1379 Imports.push_back(Import); 1380 assert(WasmIndices.count(&WS) == 0); 1381 WasmIndices[&WS] = NumTagImports++; 1382 } else if (WS.isTable()) { 1383 if (WS.isWeak()) 1384 report_fatal_error("undefined table symbol cannot be weak"); 1385 1386 wasm::WasmImport Import; 1387 Import.Module = WS.getImportModule(); 1388 Import.Field = WS.getImportName(); 1389 Import.Kind = wasm::WASM_EXTERNAL_TABLE; 1390 Import.Table = WS.getTableType(); 1391 Imports.push_back(Import); 1392 assert(WasmIndices.count(&WS) == 0); 1393 WasmIndices[&WS] = NumTableImports++; 1394 } 1395 } 1396 } 1397 1398 // Add imports for GOT globals 1399 for (const MCSymbol &S : Asm.symbols()) { 1400 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1401 if (WS.isUsedInGOT()) { 1402 wasm::WasmImport Import; 1403 if (WS.isFunction()) 1404 Import.Module = "GOT.func"; 1405 else 1406 Import.Module = "GOT.mem"; 1407 Import.Field = WS.getName(); 1408 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1409 Import.Global = {wasm::WASM_TYPE_I32, true}; 1410 Imports.push_back(Import); 1411 assert(GOTIndices.count(&WS) == 0); 1412 GOTIndices[&WS] = NumGlobalImports++; 1413 } 1414 } 1415 } 1416 1417 uint64_t WasmObjectWriter::writeObject(MCAssembler &Asm, 1418 const MCAsmLayout &Layout) { 1419 support::endian::Writer MainWriter(*OS, support::little); 1420 W = &MainWriter; 1421 if (IsSplitDwarf) { 1422 uint64_t TotalSize = writeOneObject(Asm, Layout, DwoMode::NonDwoOnly); 1423 assert(DwoOS); 1424 support::endian::Writer DwoWriter(*DwoOS, support::little); 1425 W = &DwoWriter; 1426 return TotalSize + writeOneObject(Asm, Layout, DwoMode::DwoOnly); 1427 } else { 1428 return writeOneObject(Asm, Layout, DwoMode::AllSections); 1429 } 1430 } 1431 1432 uint64_t WasmObjectWriter::writeOneObject(MCAssembler &Asm, 1433 const MCAsmLayout &Layout, 1434 DwoMode Mode) { 1435 uint64_t StartOffset = W->OS.tell(); 1436 SectionCount = 0; 1437 CustomSections.clear(); 1438 1439 LLVM_DEBUG(dbgs() << "WasmObjectWriter::writeObject\n"); 1440 1441 // Collect information from the available symbols. 1442 SmallVector<WasmFunction, 4> Functions; 1443 SmallVector<uint32_t, 4> TableElems; 1444 SmallVector<wasm::WasmImport, 4> Imports; 1445 SmallVector<wasm::WasmExport, 4> Exports; 1446 SmallVector<uint32_t, 2> TagTypes; 1447 SmallVector<wasm::WasmGlobal, 1> Globals; 1448 SmallVector<wasm::WasmTable, 1> Tables; 1449 SmallVector<wasm::WasmSymbolInfo, 4> SymbolInfos; 1450 SmallVector<std::pair<uint16_t, uint32_t>, 2> InitFuncs; 1451 std::map<StringRef, std::vector<WasmComdatEntry>> Comdats; 1452 uint64_t DataSize = 0; 1453 if (Mode != DwoMode::DwoOnly) { 1454 prepareImports(Imports, Asm, Layout); 1455 } 1456 1457 // Populate DataSegments and CustomSections, which must be done before 1458 // populating DataLocations. 1459 for (MCSection &Sec : Asm) { 1460 auto &Section = static_cast<MCSectionWasm &>(Sec); 1461 StringRef SectionName = Section.getName(); 1462 1463 if (Mode == DwoMode::NonDwoOnly && isDwoSection(Sec)) 1464 continue; 1465 if (Mode == DwoMode::DwoOnly && !isDwoSection(Sec)) 1466 continue; 1467 1468 LLVM_DEBUG(dbgs() << "Processing Section " << SectionName << " group " 1469 << Section.getGroup() << "\n";); 1470 1471 // .init_array sections are handled specially elsewhere. 1472 if (SectionName.startswith(".init_array")) 1473 continue; 1474 1475 // Code is handled separately 1476 if (Section.getKind().isText()) 1477 continue; 1478 1479 if (Section.isWasmData()) { 1480 uint32_t SegmentIndex = DataSegments.size(); 1481 DataSize = alignTo(DataSize, Section.getAlignment()); 1482 DataSegments.emplace_back(); 1483 WasmDataSegment &Segment = DataSegments.back(); 1484 Segment.Name = SectionName; 1485 Segment.InitFlags = Section.getPassive() 1486 ? (uint32_t)wasm::WASM_DATA_SEGMENT_IS_PASSIVE 1487 : 0; 1488 Segment.Offset = DataSize; 1489 Segment.Section = &Section; 1490 addData(Segment.Data, Section); 1491 Segment.Alignment = Log2_32(Section.getAlignment()); 1492 Segment.LinkingFlags = Section.getSegmentFlags(); 1493 DataSize += Segment.Data.size(); 1494 Section.setSegmentIndex(SegmentIndex); 1495 1496 if (const MCSymbolWasm *C = Section.getGroup()) { 1497 Comdats[C->getName()].emplace_back( 1498 WasmComdatEntry{wasm::WASM_COMDAT_DATA, SegmentIndex}); 1499 } 1500 } else { 1501 // Create custom sections 1502 assert(Sec.getKind().isMetadata()); 1503 1504 StringRef Name = SectionName; 1505 1506 // For user-defined custom sections, strip the prefix 1507 if (Name.startswith(".custom_section.")) 1508 Name = Name.substr(strlen(".custom_section.")); 1509 1510 MCSymbol *Begin = Sec.getBeginSymbol(); 1511 if (Begin) { 1512 assert(WasmIndices.count(cast<MCSymbolWasm>(Begin)) == 0); 1513 WasmIndices[cast<MCSymbolWasm>(Begin)] = CustomSections.size(); 1514 } 1515 1516 // Separate out the producers and target features sections 1517 if (Name == "producers") { 1518 ProducersSection = std::make_unique<WasmCustomSection>(Name, &Section); 1519 continue; 1520 } 1521 if (Name == "target_features") { 1522 TargetFeaturesSection = 1523 std::make_unique<WasmCustomSection>(Name, &Section); 1524 continue; 1525 } 1526 1527 // Custom sections can also belong to COMDAT groups. In this case the 1528 // decriptor's "index" field is the section index (in the final object 1529 // file), but that is not known until after layout, so it must be fixed up 1530 // later 1531 if (const MCSymbolWasm *C = Section.getGroup()) { 1532 Comdats[C->getName()].emplace_back( 1533 WasmComdatEntry{wasm::WASM_COMDAT_SECTION, 1534 static_cast<uint32_t>(CustomSections.size())}); 1535 } 1536 1537 CustomSections.emplace_back(Name, &Section); 1538 } 1539 } 1540 1541 if (Mode != DwoMode::DwoOnly) { 1542 // Populate WasmIndices and DataLocations for defined symbols. 1543 for (const MCSymbol &S : Asm.symbols()) { 1544 // Ignore unnamed temporary symbols, which aren't ever exported, imported, 1545 // or used in relocations. 1546 if (S.isTemporary() && S.getName().empty()) 1547 continue; 1548 1549 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1550 LLVM_DEBUG(dbgs() 1551 << "MCSymbol: " 1552 << toString(WS.getType().getValueOr(wasm::WASM_SYMBOL_TYPE_DATA)) 1553 << " '" << S << "'" 1554 << " isDefined=" << S.isDefined() << " isExternal=" 1555 << S.isExternal() << " isTemporary=" << S.isTemporary() 1556 << " isWeak=" << WS.isWeak() << " isHidden=" << WS.isHidden() 1557 << " isVariable=" << WS.isVariable() << "\n"); 1558 1559 if (WS.isVariable()) 1560 continue; 1561 if (WS.isComdat() && !WS.isDefined()) 1562 continue; 1563 1564 if (WS.isFunction()) { 1565 unsigned Index; 1566 if (WS.isDefined()) { 1567 if (WS.getOffset() != 0) 1568 report_fatal_error( 1569 "function sections must contain one function each"); 1570 1571 if (WS.getSize() == nullptr) 1572 report_fatal_error( 1573 "function symbols must have a size set with .size"); 1574 1575 // A definition. Write out the function body. 1576 Index = NumFunctionImports + Functions.size(); 1577 WasmFunction Func; 1578 Func.SigIndex = getFunctionType(WS); 1579 Func.Sym = &WS; 1580 assert(WasmIndices.count(&WS) == 0); 1581 WasmIndices[&WS] = Index; 1582 Functions.push_back(Func); 1583 1584 auto &Section = static_cast<MCSectionWasm &>(WS.getSection()); 1585 if (const MCSymbolWasm *C = Section.getGroup()) { 1586 Comdats[C->getName()].emplace_back( 1587 WasmComdatEntry{wasm::WASM_COMDAT_FUNCTION, Index}); 1588 } 1589 1590 if (WS.hasExportName()) { 1591 wasm::WasmExport Export; 1592 Export.Name = WS.getExportName(); 1593 Export.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1594 Export.Index = Index; 1595 Exports.push_back(Export); 1596 } 1597 } else { 1598 // An import; the index was assigned above. 1599 Index = WasmIndices.find(&WS)->second; 1600 } 1601 1602 LLVM_DEBUG(dbgs() << " -> function index: " << Index << "\n"); 1603 1604 } else if (WS.isData()) { 1605 if (!isInSymtab(WS)) 1606 continue; 1607 1608 if (!WS.isDefined()) { 1609 LLVM_DEBUG(dbgs() << " -> segment index: -1" 1610 << "\n"); 1611 continue; 1612 } 1613 1614 if (!WS.getSize()) 1615 report_fatal_error("data symbols must have a size set with .size: " + 1616 WS.getName()); 1617 1618 int64_t Size = 0; 1619 if (!WS.getSize()->evaluateAsAbsolute(Size, Layout)) 1620 report_fatal_error(".size expression must be evaluatable"); 1621 1622 auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection()); 1623 if (!DataSection.isWasmData()) 1624 report_fatal_error("data symbols must live in a data section: " + 1625 WS.getName()); 1626 1627 // For each data symbol, export it in the symtab as a reference to the 1628 // corresponding Wasm data segment. 1629 wasm::WasmDataReference Ref = wasm::WasmDataReference{ 1630 DataSection.getSegmentIndex(), Layout.getSymbolOffset(WS), 1631 static_cast<uint64_t>(Size)}; 1632 assert(DataLocations.count(&WS) == 0); 1633 DataLocations[&WS] = Ref; 1634 LLVM_DEBUG(dbgs() << " -> segment index: " << Ref.Segment << "\n"); 1635 1636 } else if (WS.isGlobal()) { 1637 // A "true" Wasm global (currently just __stack_pointer) 1638 if (WS.isDefined()) { 1639 wasm::WasmGlobal Global; 1640 Global.Type = WS.getGlobalType(); 1641 Global.Index = NumGlobalImports + Globals.size(); 1642 switch (Global.Type.Type) { 1643 case wasm::WASM_TYPE_I32: 1644 Global.InitExpr.Opcode = wasm::WASM_OPCODE_I32_CONST; 1645 break; 1646 case wasm::WASM_TYPE_I64: 1647 Global.InitExpr.Opcode = wasm::WASM_OPCODE_I64_CONST; 1648 break; 1649 case wasm::WASM_TYPE_F32: 1650 Global.InitExpr.Opcode = wasm::WASM_OPCODE_F32_CONST; 1651 break; 1652 case wasm::WASM_TYPE_F64: 1653 Global.InitExpr.Opcode = wasm::WASM_OPCODE_F64_CONST; 1654 break; 1655 case wasm::WASM_TYPE_EXTERNREF: 1656 Global.InitExpr.Opcode = wasm::WASM_OPCODE_REF_NULL; 1657 break; 1658 default: 1659 llvm_unreachable("unexpected type"); 1660 } 1661 assert(WasmIndices.count(&WS) == 0); 1662 WasmIndices[&WS] = Global.Index; 1663 Globals.push_back(Global); 1664 } else { 1665 // An import; the index was assigned above 1666 LLVM_DEBUG(dbgs() << " -> global index: " 1667 << WasmIndices.find(&WS)->second << "\n"); 1668 } 1669 } else if (WS.isTable()) { 1670 if (WS.isDefined()) { 1671 wasm::WasmTable Table; 1672 Table.Index = NumTableImports + Tables.size(); 1673 Table.Type = WS.getTableType(); 1674 assert(WasmIndices.count(&WS) == 0); 1675 WasmIndices[&WS] = Table.Index; 1676 Tables.push_back(Table); 1677 } 1678 LLVM_DEBUG(dbgs() << " -> table index: " 1679 << WasmIndices.find(&WS)->second << "\n"); 1680 } else if (WS.isTag()) { 1681 // C++ exception symbol (__cpp_exception) or longjmp symbol 1682 // (__c_longjmp) 1683 unsigned Index; 1684 if (WS.isDefined()) { 1685 Index = NumTagImports + TagTypes.size(); 1686 uint32_t SigIndex = getTagType(WS); 1687 assert(WasmIndices.count(&WS) == 0); 1688 WasmIndices[&WS] = Index; 1689 TagTypes.push_back(SigIndex); 1690 } else { 1691 // An import; the index was assigned above. 1692 assert(WasmIndices.count(&WS) > 0); 1693 } 1694 LLVM_DEBUG(dbgs() << " -> tag index: " << WasmIndices.find(&WS)->second 1695 << "\n"); 1696 1697 } else { 1698 assert(WS.isSection()); 1699 } 1700 } 1701 1702 // Populate WasmIndices and DataLocations for aliased symbols. We need to 1703 // process these in a separate pass because we need to have processed the 1704 // target of the alias before the alias itself and the symbols are not 1705 // necessarily ordered in this way. 1706 for (const MCSymbol &S : Asm.symbols()) { 1707 if (!S.isVariable()) 1708 continue; 1709 1710 assert(S.isDefined()); 1711 1712 const auto *BS = Layout.getBaseSymbol(S); 1713 if (!BS) 1714 report_fatal_error(Twine(S.getName()) + 1715 ": absolute addressing not supported!"); 1716 const MCSymbolWasm *Base = cast<MCSymbolWasm>(BS); 1717 1718 // Find the target symbol of this weak alias and export that index 1719 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1720 LLVM_DEBUG(dbgs() << WS.getName() << ": weak alias of '" << *Base 1721 << "'\n"); 1722 1723 if (Base->isFunction()) { 1724 assert(WasmIndices.count(Base) > 0); 1725 uint32_t WasmIndex = WasmIndices.find(Base)->second; 1726 assert(WasmIndices.count(&WS) == 0); 1727 WasmIndices[&WS] = WasmIndex; 1728 LLVM_DEBUG(dbgs() << " -> index:" << WasmIndex << "\n"); 1729 } else if (Base->isData()) { 1730 auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection()); 1731 uint64_t Offset = Layout.getSymbolOffset(S); 1732 int64_t Size = 0; 1733 // For data symbol alias we use the size of the base symbol as the 1734 // size of the alias. When an offset from the base is involved this 1735 // can result in a offset + size goes past the end of the data section 1736 // which out object format doesn't support. So we must clamp it. 1737 if (!Base->getSize()->evaluateAsAbsolute(Size, Layout)) 1738 report_fatal_error(".size expression must be evaluatable"); 1739 const WasmDataSegment &Segment = 1740 DataSegments[DataSection.getSegmentIndex()]; 1741 Size = 1742 std::min(static_cast<uint64_t>(Size), Segment.Data.size() - Offset); 1743 wasm::WasmDataReference Ref = wasm::WasmDataReference{ 1744 DataSection.getSegmentIndex(), 1745 static_cast<uint32_t>(Layout.getSymbolOffset(S)), 1746 static_cast<uint32_t>(Size)}; 1747 DataLocations[&WS] = Ref; 1748 LLVM_DEBUG(dbgs() << " -> index:" << Ref.Segment << "\n"); 1749 } else { 1750 report_fatal_error("don't yet support global/tag aliases"); 1751 } 1752 } 1753 } 1754 1755 // Finally, populate the symbol table itself, in its "natural" order. 1756 for (const MCSymbol &S : Asm.symbols()) { 1757 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1758 if (!isInSymtab(WS)) { 1759 WS.setIndex(InvalidIndex); 1760 continue; 1761 } 1762 LLVM_DEBUG(dbgs() << "adding to symtab: " << WS << "\n"); 1763 1764 uint32_t Flags = 0; 1765 if (WS.isWeak()) 1766 Flags |= wasm::WASM_SYMBOL_BINDING_WEAK; 1767 if (WS.isHidden()) 1768 Flags |= wasm::WASM_SYMBOL_VISIBILITY_HIDDEN; 1769 if (!WS.isExternal() && WS.isDefined()) 1770 Flags |= wasm::WASM_SYMBOL_BINDING_LOCAL; 1771 if (WS.isUndefined()) 1772 Flags |= wasm::WASM_SYMBOL_UNDEFINED; 1773 if (WS.isNoStrip()) { 1774 Flags |= wasm::WASM_SYMBOL_NO_STRIP; 1775 if (isEmscripten()) { 1776 Flags |= wasm::WASM_SYMBOL_EXPORTED; 1777 } 1778 } 1779 if (WS.hasImportName()) 1780 Flags |= wasm::WASM_SYMBOL_EXPLICIT_NAME; 1781 if (WS.hasExportName()) 1782 Flags |= wasm::WASM_SYMBOL_EXPORTED; 1783 if (WS.isTLS()) 1784 Flags |= wasm::WASM_SYMBOL_TLS; 1785 1786 wasm::WasmSymbolInfo Info; 1787 Info.Name = WS.getName(); 1788 Info.Kind = WS.getType().getValueOr(wasm::WASM_SYMBOL_TYPE_DATA); 1789 Info.Flags = Flags; 1790 if (!WS.isData()) { 1791 assert(WasmIndices.count(&WS) > 0); 1792 Info.ElementIndex = WasmIndices.find(&WS)->second; 1793 } else if (WS.isDefined()) { 1794 assert(DataLocations.count(&WS) > 0); 1795 Info.DataRef = DataLocations.find(&WS)->second; 1796 } 1797 WS.setIndex(SymbolInfos.size()); 1798 SymbolInfos.emplace_back(Info); 1799 } 1800 1801 { 1802 auto HandleReloc = [&](const WasmRelocationEntry &Rel) { 1803 // Functions referenced by a relocation need to put in the table. This is 1804 // purely to make the object file's provisional values readable, and is 1805 // ignored by the linker, which re-calculates the relocations itself. 1806 if (Rel.Type != wasm::R_WASM_TABLE_INDEX_I32 && 1807 Rel.Type != wasm::R_WASM_TABLE_INDEX_I64 && 1808 Rel.Type != wasm::R_WASM_TABLE_INDEX_SLEB && 1809 Rel.Type != wasm::R_WASM_TABLE_INDEX_SLEB64 && 1810 Rel.Type != wasm::R_WASM_TABLE_INDEX_REL_SLEB && 1811 Rel.Type != wasm::R_WASM_TABLE_INDEX_REL_SLEB64) 1812 return; 1813 assert(Rel.Symbol->isFunction()); 1814 const MCSymbolWasm *Base = 1815 cast<MCSymbolWasm>(Layout.getBaseSymbol(*Rel.Symbol)); 1816 uint32_t FunctionIndex = WasmIndices.find(Base)->second; 1817 uint32_t TableIndex = TableElems.size() + InitialTableOffset; 1818 if (TableIndices.try_emplace(Base, TableIndex).second) { 1819 LLVM_DEBUG(dbgs() << " -> adding " << Base->getName() 1820 << " to table: " << TableIndex << "\n"); 1821 TableElems.push_back(FunctionIndex); 1822 registerFunctionType(*Base); 1823 } 1824 }; 1825 1826 for (const WasmRelocationEntry &RelEntry : CodeRelocations) 1827 HandleReloc(RelEntry); 1828 for (const WasmRelocationEntry &RelEntry : DataRelocations) 1829 HandleReloc(RelEntry); 1830 } 1831 1832 // Translate .init_array section contents into start functions. 1833 for (const MCSection &S : Asm) { 1834 const auto &WS = static_cast<const MCSectionWasm &>(S); 1835 if (WS.getName().startswith(".fini_array")) 1836 report_fatal_error(".fini_array sections are unsupported"); 1837 if (!WS.getName().startswith(".init_array")) 1838 continue; 1839 if (WS.getFragmentList().empty()) 1840 continue; 1841 1842 // init_array is expected to contain a single non-empty data fragment 1843 if (WS.getFragmentList().size() != 3) 1844 report_fatal_error("only one .init_array section fragment supported"); 1845 1846 auto IT = WS.begin(); 1847 const MCFragment &EmptyFrag = *IT; 1848 if (EmptyFrag.getKind() != MCFragment::FT_Data) 1849 report_fatal_error(".init_array section should be aligned"); 1850 1851 IT = std::next(IT); 1852 const MCFragment &AlignFrag = *IT; 1853 if (AlignFrag.getKind() != MCFragment::FT_Align) 1854 report_fatal_error(".init_array section should be aligned"); 1855 if (cast<MCAlignFragment>(AlignFrag).getAlignment() != (is64Bit() ? 8 : 4)) 1856 report_fatal_error(".init_array section should be aligned for pointers"); 1857 1858 const MCFragment &Frag = *std::next(IT); 1859 if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data) 1860 report_fatal_error("only data supported in .init_array section"); 1861 1862 uint16_t Priority = UINT16_MAX; 1863 unsigned PrefixLength = strlen(".init_array"); 1864 if (WS.getName().size() > PrefixLength) { 1865 if (WS.getName()[PrefixLength] != '.') 1866 report_fatal_error( 1867 ".init_array section priority should start with '.'"); 1868 if (WS.getName().substr(PrefixLength + 1).getAsInteger(10, Priority)) 1869 report_fatal_error("invalid .init_array section priority"); 1870 } 1871 const auto &DataFrag = cast<MCDataFragment>(Frag); 1872 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 1873 for (const uint8_t * 1874 P = (const uint8_t *)Contents.data(), 1875 *End = (const uint8_t *)Contents.data() + Contents.size(); 1876 P != End; ++P) { 1877 if (*P != 0) 1878 report_fatal_error("non-symbolic data in .init_array section"); 1879 } 1880 for (const MCFixup &Fixup : DataFrag.getFixups()) { 1881 assert(Fixup.getKind() == 1882 MCFixup::getKindForSize(is64Bit() ? 8 : 4, false)); 1883 const MCExpr *Expr = Fixup.getValue(); 1884 auto *SymRef = dyn_cast<MCSymbolRefExpr>(Expr); 1885 if (!SymRef) 1886 report_fatal_error("fixups in .init_array should be symbol references"); 1887 const auto &TargetSym = cast<const MCSymbolWasm>(SymRef->getSymbol()); 1888 if (TargetSym.getIndex() == InvalidIndex) 1889 report_fatal_error("symbols in .init_array should exist in symtab"); 1890 if (!TargetSym.isFunction()) 1891 report_fatal_error("symbols in .init_array should be for functions"); 1892 InitFuncs.push_back( 1893 std::make_pair(Priority, TargetSym.getIndex())); 1894 } 1895 } 1896 1897 // Write out the Wasm header. 1898 writeHeader(Asm); 1899 1900 uint32_t CodeSectionIndex, DataSectionIndex; 1901 if (Mode != DwoMode::DwoOnly) { 1902 writeTypeSection(Signatures); 1903 writeImportSection(Imports, DataSize, TableElems.size()); 1904 writeFunctionSection(Functions); 1905 writeTableSection(Tables); 1906 // Skip the "memory" section; we import the memory instead. 1907 writeTagSection(TagTypes); 1908 writeGlobalSection(Globals); 1909 writeExportSection(Exports); 1910 const MCSymbol *IndirectFunctionTable = 1911 Asm.getContext().lookupSymbol("__indirect_function_table"); 1912 writeElemSection(cast_or_null<const MCSymbolWasm>(IndirectFunctionTable), 1913 TableElems); 1914 writeDataCountSection(); 1915 1916 CodeSectionIndex = writeCodeSection(Asm, Layout, Functions); 1917 DataSectionIndex = writeDataSection(Layout); 1918 } 1919 1920 // The Sections in the COMDAT list have placeholder indices (their index among 1921 // custom sections, rather than among all sections). Fix them up here. 1922 for (auto &Group : Comdats) { 1923 for (auto &Entry : Group.second) { 1924 if (Entry.Kind == wasm::WASM_COMDAT_SECTION) { 1925 Entry.Index += SectionCount; 1926 } 1927 } 1928 } 1929 for (auto &CustomSection : CustomSections) 1930 writeCustomSection(CustomSection, Asm, Layout); 1931 1932 if (Mode != DwoMode::DwoOnly) { 1933 writeLinkingMetaDataSection(SymbolInfos, InitFuncs, Comdats); 1934 1935 writeRelocSection(CodeSectionIndex, "CODE", CodeRelocations); 1936 writeRelocSection(DataSectionIndex, "DATA", DataRelocations); 1937 } 1938 writeCustomRelocSections(); 1939 if (ProducersSection) 1940 writeCustomSection(*ProducersSection, Asm, Layout); 1941 if (TargetFeaturesSection) 1942 writeCustomSection(*TargetFeaturesSection, Asm, Layout); 1943 1944 // TODO: Translate the .comment section to the output. 1945 return W->OS.tell() - StartOffset; 1946 } 1947 1948 std::unique_ptr<MCObjectWriter> 1949 llvm::createWasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 1950 raw_pwrite_stream &OS) { 1951 return std::make_unique<WasmObjectWriter>(std::move(MOTW), OS); 1952 } 1953 1954 std::unique_ptr<MCObjectWriter> 1955 llvm::createWasmDwoObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 1956 raw_pwrite_stream &OS, 1957 raw_pwrite_stream &DwoOS) { 1958 return std::make_unique<WasmObjectWriter>(std::move(MOTW), OS, DwoOS); 1959 } 1960