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/BinaryFormat/Wasm.h" 15 #include "llvm/BinaryFormat/WasmTraits.h" 16 #include "llvm/Config/llvm-config.h" 17 #include "llvm/MC/MCAsmBackend.h" 18 #include "llvm/MC/MCAsmLayout.h" 19 #include "llvm/MC/MCAssembler.h" 20 #include "llvm/MC/MCContext.h" 21 #include "llvm/MC/MCExpr.h" 22 #include "llvm/MC/MCFixupKindInfo.h" 23 #include "llvm/MC/MCObjectWriter.h" 24 #include "llvm/MC/MCSectionWasm.h" 25 #include "llvm/MC/MCSymbolWasm.h" 26 #include "llvm/MC/MCValue.h" 27 #include "llvm/MC/MCWasmObjectWriter.h" 28 #include "llvm/Support/Casting.h" 29 #include "llvm/Support/Debug.h" 30 #include "llvm/Support/EndianStream.h" 31 #include "llvm/Support/ErrorHandling.h" 32 #include "llvm/Support/LEB128.h" 33 #include <vector> 34 35 using namespace llvm; 36 37 #define DEBUG_TYPE "mc" 38 39 namespace { 40 41 // When we create the indirect function table we start at 1, so that there is 42 // and empty 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 // A wasm data segment. A wasm binary contains only a single data section 59 // but that can contain many segments, each with their own virtual location 60 // in memory. Each MCSection data created by llvm is modeled as its own 61 // wasm data segment. 62 struct WasmDataSegment { 63 MCSectionWasm *Section; 64 StringRef Name; 65 uint32_t InitFlags; 66 uint64_t Offset; 67 uint32_t Alignment; 68 uint32_t LinkingFlags; 69 SmallVector<char, 4> Data; 70 }; 71 72 // A wasm function to be written into the function section. 73 struct WasmFunction { 74 uint32_t SigIndex; 75 const MCSymbolWasm *Sym; 76 }; 77 78 // A wasm global to be written into the global section. 79 struct WasmGlobal { 80 wasm::WasmGlobalType Type; 81 uint64_t InitialValue; 82 }; 83 84 // Information about a single item which is part of a COMDAT. For each data 85 // segment or function which is in the COMDAT, there is a corresponding 86 // WasmComdatEntry. 87 struct WasmComdatEntry { 88 unsigned Kind; 89 uint32_t Index; 90 }; 91 92 // Information about a single relocation. 93 struct WasmRelocationEntry { 94 uint64_t Offset; // Where is the relocation. 95 const MCSymbolWasm *Symbol; // The symbol to relocate with. 96 int64_t Addend; // A value to add to the symbol. 97 unsigned Type; // The type of the relocation. 98 const MCSectionWasm *FixupSection; // The section the relocation is targeting. 99 100 WasmRelocationEntry(uint64_t Offset, const MCSymbolWasm *Symbol, 101 int64_t Addend, unsigned Type, 102 const MCSectionWasm *FixupSection) 103 : Offset(Offset), Symbol(Symbol), Addend(Addend), Type(Type), 104 FixupSection(FixupSection) {} 105 106 bool hasAddend() const { return wasm::relocTypeHasAddend(Type); } 107 108 void print(raw_ostream &Out) const { 109 Out << wasm::relocTypetoString(Type) << " Off=" << Offset 110 << ", Sym=" << *Symbol << ", Addend=" << Addend 111 << ", FixupSection=" << FixupSection->getName(); 112 } 113 114 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 115 LLVM_DUMP_METHOD void dump() const { print(dbgs()); } 116 #endif 117 }; 118 119 static const uint32_t InvalidIndex = -1; 120 121 struct WasmCustomSection { 122 123 StringRef Name; 124 MCSectionWasm *Section; 125 126 uint32_t OutputContentsOffset = 0; 127 uint32_t OutputIndex = InvalidIndex; 128 129 WasmCustomSection(StringRef Name, MCSectionWasm *Section) 130 : Name(Name), Section(Section) {} 131 }; 132 133 #if !defined(NDEBUG) 134 raw_ostream &operator<<(raw_ostream &OS, const WasmRelocationEntry &Rel) { 135 Rel.print(OS); 136 return OS; 137 } 138 #endif 139 140 // Write X as an (unsigned) LEB value at offset Offset in Stream, padded 141 // to allow patching. 142 template <int W> 143 void writePatchableLEB(raw_pwrite_stream &Stream, uint64_t X, uint64_t Offset) { 144 uint8_t Buffer[W]; 145 unsigned SizeLen = encodeULEB128(X, Buffer, W); 146 assert(SizeLen == W); 147 Stream.pwrite((char *)Buffer, SizeLen, Offset); 148 } 149 150 // Write X as an signed LEB value at offset Offset in Stream, padded 151 // to allow patching. 152 template <int W> 153 void writePatchableSLEB(raw_pwrite_stream &Stream, int64_t X, uint64_t Offset) { 154 uint8_t Buffer[W]; 155 unsigned SizeLen = encodeSLEB128(X, Buffer, W); 156 assert(SizeLen == W); 157 Stream.pwrite((char *)Buffer, SizeLen, Offset); 158 } 159 160 // Write X as a plain integer value at offset Offset in Stream. 161 static void patchI32(raw_pwrite_stream &Stream, uint32_t X, uint64_t Offset) { 162 uint8_t Buffer[4]; 163 support::endian::write32le(Buffer, X); 164 Stream.pwrite((char *)Buffer, sizeof(Buffer), Offset); 165 } 166 167 static void patchI64(raw_pwrite_stream &Stream, uint64_t X, uint64_t Offset) { 168 uint8_t Buffer[8]; 169 support::endian::write64le(Buffer, X); 170 Stream.pwrite((char *)Buffer, sizeof(Buffer), Offset); 171 } 172 173 bool isDwoSection(const MCSection &Sec) { 174 return Sec.getName().endswith(".dwo"); 175 } 176 177 class WasmObjectWriter : public MCObjectWriter { 178 support::endian::Writer *W; 179 180 /// The target specific Wasm writer instance. 181 std::unique_ptr<MCWasmObjectTargetWriter> TargetObjectWriter; 182 183 // Relocations for fixing up references in the code section. 184 std::vector<WasmRelocationEntry> CodeRelocations; 185 // Relocations for fixing up references in the data section. 186 std::vector<WasmRelocationEntry> DataRelocations; 187 188 // Index values to use for fixing up call_indirect type indices. 189 // Maps function symbols to the index of the type of the function 190 DenseMap<const MCSymbolWasm *, uint32_t> TypeIndices; 191 // Maps function symbols to the table element index space. Used 192 // for TABLE_INDEX relocation types (i.e. address taken functions). 193 DenseMap<const MCSymbolWasm *, uint32_t> TableIndices; 194 // Maps function/global/table symbols to the 195 // function/global/table/tag/section index space. 196 DenseMap<const MCSymbolWasm *, uint32_t> WasmIndices; 197 DenseMap<const MCSymbolWasm *, uint32_t> GOTIndices; 198 // Maps data symbols to the Wasm segment and offset/size with the segment. 199 DenseMap<const MCSymbolWasm *, wasm::WasmDataReference> DataLocations; 200 201 // Stores output data (index, relocations, content offset) for custom 202 // section. 203 std::vector<WasmCustomSection> CustomSections; 204 std::unique_ptr<WasmCustomSection> ProducersSection; 205 std::unique_ptr<WasmCustomSection> TargetFeaturesSection; 206 // Relocations for fixing up references in the custom sections. 207 DenseMap<const MCSectionWasm *, std::vector<WasmRelocationEntry>> 208 CustomSectionsRelocations; 209 210 // Map from section to defining function symbol. 211 DenseMap<const MCSection *, const MCSymbol *> SectionFunctions; 212 213 DenseMap<wasm::WasmSignature, uint32_t> SignatureIndices; 214 SmallVector<wasm::WasmSignature, 4> Signatures; 215 SmallVector<WasmDataSegment, 4> DataSegments; 216 unsigned NumFunctionImports = 0; 217 unsigned NumGlobalImports = 0; 218 unsigned NumTableImports = 0; 219 unsigned NumTagImports = 0; 220 uint32_t SectionCount = 0; 221 222 enum class DwoMode { 223 AllSections, 224 NonDwoOnly, 225 DwoOnly, 226 }; 227 bool IsSplitDwarf = false; 228 raw_pwrite_stream *OS = nullptr; 229 raw_pwrite_stream *DwoOS = nullptr; 230 231 // TargetObjectWriter wranppers. 232 bool is64Bit() const { return TargetObjectWriter->is64Bit(); } 233 bool isEmscripten() const { return TargetObjectWriter->isEmscripten(); } 234 235 void startSection(SectionBookkeeping &Section, unsigned SectionId); 236 void startCustomSection(SectionBookkeeping &Section, StringRef Name); 237 void endSection(SectionBookkeeping &Section); 238 239 public: 240 WasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 241 raw_pwrite_stream &OS_) 242 : TargetObjectWriter(std::move(MOTW)), OS(&OS_) {} 243 244 WasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 245 raw_pwrite_stream &OS_, raw_pwrite_stream &DwoOS_) 246 : TargetObjectWriter(std::move(MOTW)), IsSplitDwarf(true), OS(&OS_), 247 DwoOS(&DwoOS_) {} 248 249 private: 250 void reset() override { 251 CodeRelocations.clear(); 252 DataRelocations.clear(); 253 TypeIndices.clear(); 254 WasmIndices.clear(); 255 GOTIndices.clear(); 256 TableIndices.clear(); 257 DataLocations.clear(); 258 CustomSections.clear(); 259 ProducersSection.reset(); 260 TargetFeaturesSection.reset(); 261 CustomSectionsRelocations.clear(); 262 SignatureIndices.clear(); 263 Signatures.clear(); 264 DataSegments.clear(); 265 SectionFunctions.clear(); 266 NumFunctionImports = 0; 267 NumGlobalImports = 0; 268 NumTableImports = 0; 269 MCObjectWriter::reset(); 270 } 271 272 void writeHeader(const MCAssembler &Asm); 273 274 void recordRelocation(MCAssembler &Asm, const MCAsmLayout &Layout, 275 const MCFragment *Fragment, const MCFixup &Fixup, 276 MCValue Target, uint64_t &FixedValue) override; 277 278 void executePostLayoutBinding(MCAssembler &Asm, 279 const MCAsmLayout &Layout) override; 280 void prepareImports(SmallVectorImpl<wasm::WasmImport> &Imports, 281 MCAssembler &Asm, const MCAsmLayout &Layout); 282 uint64_t writeObject(MCAssembler &Asm, const MCAsmLayout &Layout) override; 283 284 uint64_t writeOneObject(MCAssembler &Asm, const MCAsmLayout &Layout, 285 DwoMode Mode); 286 287 void writeString(const StringRef Str) { 288 encodeULEB128(Str.size(), W->OS); 289 W->OS << Str; 290 } 291 292 void writeStringWithAlignment(const StringRef Str, unsigned Alignment); 293 294 void writeI32(int32_t val) { 295 char Buffer[4]; 296 support::endian::write32le(Buffer, val); 297 W->OS.write(Buffer, sizeof(Buffer)); 298 } 299 300 void writeI64(int64_t val) { 301 char Buffer[8]; 302 support::endian::write64le(Buffer, val); 303 W->OS.write(Buffer, sizeof(Buffer)); 304 } 305 306 void writeValueType(wasm::ValType Ty) { W->OS << static_cast<char>(Ty); } 307 308 void writeTypeSection(ArrayRef<wasm::WasmSignature> Signatures); 309 void writeImportSection(ArrayRef<wasm::WasmImport> Imports, uint64_t DataSize, 310 uint32_t NumElements); 311 void writeFunctionSection(ArrayRef<WasmFunction> Functions); 312 void writeExportSection(ArrayRef<wasm::WasmExport> Exports); 313 void writeElemSection(const MCSymbolWasm *IndirectFunctionTable, 314 ArrayRef<uint32_t> TableElems); 315 void writeDataCountSection(); 316 uint32_t writeCodeSection(const MCAssembler &Asm, const MCAsmLayout &Layout, 317 ArrayRef<WasmFunction> Functions); 318 uint32_t writeDataSection(const MCAsmLayout &Layout); 319 void writeTagSection(ArrayRef<uint32_t> TagTypes); 320 void writeGlobalSection(ArrayRef<wasm::WasmGlobal> Globals); 321 void writeTableSection(ArrayRef<wasm::WasmTable> Tables); 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(); 331 332 uint64_t getProvisionalValue(const WasmRelocationEntry &RelEntry, 333 const MCAsmLayout &Layout); 334 void applyRelocations(ArrayRef<WasmRelocationEntry> Relocations, 335 uint64_t ContentsOffset, const MCAsmLayout &Layout); 336 337 uint32_t getRelocationIndexValue(const WasmRelocationEntry &RelEntry); 338 uint32_t getFunctionType(const MCSymbolWasm &Symbol); 339 uint32_t getTagType(const MCSymbolWasm &Symbol); 340 void registerFunctionType(const MCSymbolWasm &Symbol); 341 void registerTagType(const MCSymbolWasm &Symbol); 342 }; 343 344 } // end anonymous namespace 345 346 // Write out a section header and a patchable section size field. 347 void WasmObjectWriter::startSection(SectionBookkeeping &Section, 348 unsigned SectionId) { 349 LLVM_DEBUG(dbgs() << "startSection " << SectionId << "\n"); 350 W->OS << char(SectionId); 351 352 Section.SizeOffset = W->OS.tell(); 353 354 // The section size. We don't know the size yet, so reserve enough space 355 // for any 32-bit value; we'll patch it later. 356 encodeULEB128(0, W->OS, 5); 357 358 // The position where the section starts, for measuring its size. 359 Section.ContentsOffset = W->OS.tell(); 360 Section.PayloadOffset = W->OS.tell(); 361 Section.Index = SectionCount++; 362 } 363 364 // Write a string with extra paddings for trailing alignment 365 // TODO: support alignment at asm and llvm level? 366 void WasmObjectWriter::writeStringWithAlignment(const StringRef Str, 367 unsigned Alignment) { 368 369 // Calculate the encoded size of str length and add pads based on it and 370 // alignment. 371 raw_null_ostream NullOS; 372 uint64_t StrSizeLength = encodeULEB128(Str.size(), NullOS); 373 uint64_t Offset = W->OS.tell() + StrSizeLength + Str.size(); 374 uint64_t Paddings = offsetToAlignment(Offset, Align(Alignment)); 375 Offset += Paddings; 376 377 // LEB128 greater than 5 bytes is invalid 378 assert((StrSizeLength + Paddings) <= 5 && "too long string to align"); 379 380 encodeSLEB128(Str.size(), W->OS, StrSizeLength + Paddings); 381 W->OS << Str; 382 383 assert(W->OS.tell() == Offset && "invalid padding"); 384 } 385 386 void WasmObjectWriter::startCustomSection(SectionBookkeeping &Section, 387 StringRef Name) { 388 LLVM_DEBUG(dbgs() << "startCustomSection " << Name << "\n"); 389 startSection(Section, wasm::WASM_SEC_CUSTOM); 390 391 // The position where the section header ends, for measuring its size. 392 Section.PayloadOffset = W->OS.tell(); 393 394 // Custom sections in wasm also have a string identifier. 395 if (Name != "__clangast") { 396 writeString(Name); 397 } else { 398 // The on-disk hashtable in clangast needs to be aligned by 4 bytes. 399 writeStringWithAlignment(Name, 4); 400 } 401 402 // The position where the custom section starts. 403 Section.ContentsOffset = W->OS.tell(); 404 } 405 406 // Now that the section is complete and we know how big it is, patch up the 407 // section size field at the start of the section. 408 void WasmObjectWriter::endSection(SectionBookkeeping &Section) { 409 uint64_t Size = W->OS.tell(); 410 // /dev/null doesn't support seek/tell and can report offset of 0. 411 // Simply skip this patching in that case. 412 if (!Size) 413 return; 414 415 Size -= Section.PayloadOffset; 416 if (uint32_t(Size) != Size) 417 report_fatal_error("section size does not fit in a uint32_t"); 418 419 LLVM_DEBUG(dbgs() << "endSection size=" << Size << "\n"); 420 421 // Write the final section size to the payload_len field, which follows 422 // the section id byte. 423 writePatchableLEB<5>(static_cast<raw_pwrite_stream &>(W->OS), Size, 424 Section.SizeOffset); 425 } 426 427 // Emit the Wasm header. 428 void WasmObjectWriter::writeHeader(const MCAssembler &Asm) { 429 W->OS.write(wasm::WasmMagic, sizeof(wasm::WasmMagic)); 430 W->write<uint32_t>(wasm::WasmVersion); 431 } 432 433 void WasmObjectWriter::executePostLayoutBinding(MCAssembler &Asm, 434 const MCAsmLayout &Layout) { 435 // Some compilation units require the indirect function table to be present 436 // but don't explicitly reference it. This is the case for call_indirect 437 // without the reference-types feature, and also function bitcasts in all 438 // cases. In those cases the __indirect_function_table has the 439 // WASM_SYMBOL_NO_STRIP attribute. Here we make sure this symbol makes it to 440 // the assembler, if needed. 441 if (auto *Sym = Asm.getContext().lookupSymbol("__indirect_function_table")) { 442 const auto *WasmSym = static_cast<const MCSymbolWasm *>(Sym); 443 if (WasmSym->isNoStrip()) 444 Asm.registerSymbol(*Sym); 445 } 446 447 // Build a map of sections to the function that defines them, for use 448 // in recordRelocation. 449 for (const MCSymbol &S : Asm.symbols()) { 450 const auto &WS = static_cast<const MCSymbolWasm &>(S); 451 if (WS.isDefined() && WS.isFunction() && !WS.isVariable()) { 452 const auto &Sec = static_cast<const MCSectionWasm &>(S.getSection()); 453 auto Pair = SectionFunctions.insert(std::make_pair(&Sec, &S)); 454 if (!Pair.second) 455 report_fatal_error("section already has a defining function: " + 456 Sec.getName()); 457 } 458 } 459 } 460 461 void WasmObjectWriter::recordRelocation(MCAssembler &Asm, 462 const MCAsmLayout &Layout, 463 const MCFragment *Fragment, 464 const MCFixup &Fixup, MCValue Target, 465 uint64_t &FixedValue) { 466 // The WebAssembly backend should never generate FKF_IsPCRel fixups 467 assert(!(Asm.getBackend().getFixupKindInfo(Fixup.getKind()).Flags & 468 MCFixupKindInfo::FKF_IsPCRel)); 469 470 const auto &FixupSection = cast<MCSectionWasm>(*Fragment->getParent()); 471 uint64_t C = Target.getConstant(); 472 uint64_t FixupOffset = Layout.getFragmentOffset(Fragment) + Fixup.getOffset(); 473 MCContext &Ctx = Asm.getContext(); 474 bool IsLocRel = false; 475 476 if (const MCSymbolRefExpr *RefB = Target.getSymB()) { 477 478 const auto &SymB = cast<MCSymbolWasm>(RefB->getSymbol()); 479 480 if (FixupSection.getKind().isText()) { 481 Ctx.reportError(Fixup.getLoc(), 482 Twine("symbol '") + SymB.getName() + 483 "' unsupported subtraction expression used in " 484 "relocation in code section."); 485 return; 486 } 487 488 if (SymB.isUndefined()) { 489 Ctx.reportError(Fixup.getLoc(), 490 Twine("symbol '") + SymB.getName() + 491 "' can not be undefined in a subtraction expression"); 492 return; 493 } 494 const MCSection &SecB = SymB.getSection(); 495 if (&SecB != &FixupSection) { 496 Ctx.reportError(Fixup.getLoc(), 497 Twine("symbol '") + SymB.getName() + 498 "' can not be placed in a different section"); 499 return; 500 } 501 IsLocRel = true; 502 C += FixupOffset - Layout.getSymbolOffset(SymB); 503 } 504 505 // We either rejected the fixup or folded B into C at this point. 506 const MCSymbolRefExpr *RefA = Target.getSymA(); 507 const auto *SymA = cast<MCSymbolWasm>(&RefA->getSymbol()); 508 509 // The .init_array isn't translated as data, so don't do relocations in it. 510 if (FixupSection.getName().startswith(".init_array")) { 511 SymA->setUsedInInitArray(); 512 return; 513 } 514 515 if (SymA->isVariable()) { 516 const MCExpr *Expr = SymA->getVariableValue(); 517 if (const auto *Inner = dyn_cast<MCSymbolRefExpr>(Expr)) 518 if (Inner->getKind() == MCSymbolRefExpr::VK_WEAKREF) 519 llvm_unreachable("weakref used in reloc not yet implemented"); 520 } 521 522 // Put any constant offset in an addend. Offsets can be negative, and 523 // LLVM expects wrapping, in contrast to wasm's immediates which can't 524 // be negative and don't wrap. 525 FixedValue = 0; 526 527 unsigned Type = 528 TargetObjectWriter->getRelocType(Target, Fixup, FixupSection, IsLocRel); 529 530 // Absolute offset within a section or a function. 531 // Currently only supported for for metadata sections. 532 // See: test/MC/WebAssembly/blockaddress.ll 533 if ((Type == wasm::R_WASM_FUNCTION_OFFSET_I32 || 534 Type == wasm::R_WASM_FUNCTION_OFFSET_I64 || 535 Type == wasm::R_WASM_SECTION_OFFSET_I32) && 536 SymA->isDefined()) { 537 // SymA can be a temp data symbol that represents a function (in which case 538 // it needs to be replaced by the section symbol), [XXX and it apparently 539 // later gets changed again to a func symbol?] or it can be a real 540 // function symbol, in which case it can be left as-is. 541 542 if (!FixupSection.getKind().isMetadata()) 543 report_fatal_error("relocations for function or section offsets are " 544 "only supported in metadata sections"); 545 546 const MCSymbol *SectionSymbol = nullptr; 547 const MCSection &SecA = SymA->getSection(); 548 if (SecA.getKind().isText()) { 549 auto SecSymIt = SectionFunctions.find(&SecA); 550 if (SecSymIt == SectionFunctions.end()) 551 report_fatal_error("section doesn\'t have defining symbol"); 552 SectionSymbol = SecSymIt->second; 553 } else { 554 SectionSymbol = SecA.getBeginSymbol(); 555 } 556 if (!SectionSymbol) 557 report_fatal_error("section symbol is required for relocation"); 558 559 C += Layout.getSymbolOffset(*SymA); 560 SymA = cast<MCSymbolWasm>(SectionSymbol); 561 } 562 563 if (Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB || 564 Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB64 || 565 Type == wasm::R_WASM_TABLE_INDEX_SLEB || 566 Type == wasm::R_WASM_TABLE_INDEX_SLEB64 || 567 Type == wasm::R_WASM_TABLE_INDEX_I32 || 568 Type == wasm::R_WASM_TABLE_INDEX_I64) { 569 // TABLE_INDEX relocs implicitly use the default indirect function table. 570 // We require the function table to have already been defined. 571 auto TableName = "__indirect_function_table"; 572 MCSymbolWasm *Sym = cast_or_null<MCSymbolWasm>(Ctx.lookupSymbol(TableName)); 573 if (!Sym) { 574 report_fatal_error("missing indirect function table symbol"); 575 } else { 576 if (!Sym->isFunctionTable()) 577 report_fatal_error("__indirect_function_table symbol has wrong type"); 578 // Ensure that __indirect_function_table reaches the output. 579 Sym->setNoStrip(); 580 Asm.registerSymbol(*Sym); 581 } 582 } 583 584 // Relocation other than R_WASM_TYPE_INDEX_LEB are required to be 585 // against a named symbol. 586 if (Type != wasm::R_WASM_TYPE_INDEX_LEB) { 587 if (SymA->getName().empty()) 588 report_fatal_error("relocations against un-named temporaries are not yet " 589 "supported by wasm"); 590 591 SymA->setUsedInReloc(); 592 } 593 594 switch (RefA->getKind()) { 595 case MCSymbolRefExpr::VK_GOT: 596 case MCSymbolRefExpr::VK_WASM_GOT_TLS: 597 SymA->setUsedInGOT(); 598 break; 599 default: 600 break; 601 } 602 603 WasmRelocationEntry Rec(FixupOffset, SymA, C, Type, &FixupSection); 604 LLVM_DEBUG(dbgs() << "WasmReloc: " << Rec << "\n"); 605 606 if (FixupSection.isWasmData()) { 607 DataRelocations.push_back(Rec); 608 } else if (FixupSection.getKind().isText()) { 609 CodeRelocations.push_back(Rec); 610 } else if (FixupSection.getKind().isMetadata()) { 611 CustomSectionsRelocations[&FixupSection].push_back(Rec); 612 } else { 613 llvm_unreachable("unexpected section type"); 614 } 615 } 616 617 // Compute a value to write into the code at the location covered 618 // by RelEntry. This value isn't used by the static linker; it just serves 619 // to make the object format more readable and more likely to be directly 620 // useable. 621 uint64_t 622 WasmObjectWriter::getProvisionalValue(const WasmRelocationEntry &RelEntry, 623 const MCAsmLayout &Layout) { 624 if ((RelEntry.Type == wasm::R_WASM_GLOBAL_INDEX_LEB || 625 RelEntry.Type == wasm::R_WASM_GLOBAL_INDEX_I32) && 626 !RelEntry.Symbol->isGlobal()) { 627 assert(GOTIndices.count(RelEntry.Symbol) > 0 && "symbol not found in GOT index space"); 628 return GOTIndices[RelEntry.Symbol]; 629 } 630 631 switch (RelEntry.Type) { 632 case wasm::R_WASM_TABLE_INDEX_REL_SLEB: 633 case wasm::R_WASM_TABLE_INDEX_REL_SLEB64: 634 case wasm::R_WASM_TABLE_INDEX_SLEB: 635 case wasm::R_WASM_TABLE_INDEX_SLEB64: 636 case wasm::R_WASM_TABLE_INDEX_I32: 637 case wasm::R_WASM_TABLE_INDEX_I64: { 638 // Provisional value is table address of the resolved symbol itself 639 const MCSymbolWasm *Base = 640 cast<MCSymbolWasm>(Layout.getBaseSymbol(*RelEntry.Symbol)); 641 assert(Base->isFunction()); 642 if (RelEntry.Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB || 643 RelEntry.Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB64) 644 return TableIndices[Base] - InitialTableOffset; 645 else 646 return TableIndices[Base]; 647 } 648 case wasm::R_WASM_TYPE_INDEX_LEB: 649 // Provisional value is same as the index 650 return getRelocationIndexValue(RelEntry); 651 case wasm::R_WASM_FUNCTION_INDEX_LEB: 652 case wasm::R_WASM_GLOBAL_INDEX_LEB: 653 case wasm::R_WASM_GLOBAL_INDEX_I32: 654 case wasm::R_WASM_TAG_INDEX_LEB: 655 case wasm::R_WASM_TABLE_NUMBER_LEB: 656 // Provisional value is function/global/tag Wasm index 657 assert(WasmIndices.count(RelEntry.Symbol) > 0 && "symbol not found in wasm index space"); 658 return WasmIndices[RelEntry.Symbol]; 659 case wasm::R_WASM_FUNCTION_OFFSET_I32: 660 case wasm::R_WASM_FUNCTION_OFFSET_I64: 661 case wasm::R_WASM_SECTION_OFFSET_I32: { 662 if (!RelEntry.Symbol->isDefined()) 663 return 0; 664 const auto &Section = 665 static_cast<const MCSectionWasm &>(RelEntry.Symbol->getSection()); 666 return Section.getSectionOffset() + RelEntry.Addend; 667 } 668 case wasm::R_WASM_MEMORY_ADDR_LEB: 669 case wasm::R_WASM_MEMORY_ADDR_LEB64: 670 case wasm::R_WASM_MEMORY_ADDR_SLEB: 671 case wasm::R_WASM_MEMORY_ADDR_SLEB64: 672 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB: 673 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB64: 674 case wasm::R_WASM_MEMORY_ADDR_I32: 675 case wasm::R_WASM_MEMORY_ADDR_I64: 676 case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB: 677 case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB64: 678 case wasm::R_WASM_MEMORY_ADDR_LOCREL_I32: { 679 // Provisional value is address of the global plus the offset 680 // For undefined symbols, use zero 681 if (!RelEntry.Symbol->isDefined()) 682 return 0; 683 const wasm::WasmDataReference &SymRef = DataLocations[RelEntry.Symbol]; 684 const WasmDataSegment &Segment = DataSegments[SymRef.Segment]; 685 // Ignore overflow. LLVM allows address arithmetic to silently wrap. 686 return Segment.Offset + SymRef.Offset + RelEntry.Addend; 687 } 688 default: 689 llvm_unreachable("invalid relocation type"); 690 } 691 } 692 693 static void addData(SmallVectorImpl<char> &DataBytes, 694 MCSectionWasm &DataSection) { 695 LLVM_DEBUG(errs() << "addData: " << DataSection.getName() << "\n"); 696 697 DataBytes.resize(alignTo(DataBytes.size(), DataSection.getAlignment())); 698 699 for (const MCFragment &Frag : DataSection) { 700 if (Frag.hasInstructions()) 701 report_fatal_error("only data supported in data sections"); 702 703 if (auto *Align = dyn_cast<MCAlignFragment>(&Frag)) { 704 if (Align->getValueSize() != 1) 705 report_fatal_error("only byte values supported for alignment"); 706 // If nops are requested, use zeros, as this is the data section. 707 uint8_t Value = Align->hasEmitNops() ? 0 : Align->getValue(); 708 uint64_t Size = 709 std::min<uint64_t>(alignTo(DataBytes.size(), Align->getAlignment()), 710 DataBytes.size() + Align->getMaxBytesToEmit()); 711 DataBytes.resize(Size, Value); 712 } else if (auto *Fill = dyn_cast<MCFillFragment>(&Frag)) { 713 int64_t NumValues; 714 if (!Fill->getNumValues().evaluateAsAbsolute(NumValues)) 715 llvm_unreachable("The fill should be an assembler constant"); 716 DataBytes.insert(DataBytes.end(), Fill->getValueSize() * NumValues, 717 Fill->getValue()); 718 } else if (auto *LEB = dyn_cast<MCLEBFragment>(&Frag)) { 719 const SmallVectorImpl<char> &Contents = LEB->getContents(); 720 llvm::append_range(DataBytes, Contents); 721 } else { 722 const auto &DataFrag = cast<MCDataFragment>(Frag); 723 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 724 llvm::append_range(DataBytes, Contents); 725 } 726 } 727 728 LLVM_DEBUG(dbgs() << "addData -> " << DataBytes.size() << "\n"); 729 } 730 731 uint32_t 732 WasmObjectWriter::getRelocationIndexValue(const WasmRelocationEntry &RelEntry) { 733 if (RelEntry.Type == wasm::R_WASM_TYPE_INDEX_LEB) { 734 if (!TypeIndices.count(RelEntry.Symbol)) 735 report_fatal_error("symbol not found in type index space: " + 736 RelEntry.Symbol->getName()); 737 return TypeIndices[RelEntry.Symbol]; 738 } 739 740 return RelEntry.Symbol->getIndex(); 741 } 742 743 // Apply the portions of the relocation records that we can handle ourselves 744 // directly. 745 void WasmObjectWriter::applyRelocations( 746 ArrayRef<WasmRelocationEntry> Relocations, uint64_t ContentsOffset, 747 const MCAsmLayout &Layout) { 748 auto &Stream = static_cast<raw_pwrite_stream &>(W->OS); 749 for (const WasmRelocationEntry &RelEntry : Relocations) { 750 uint64_t Offset = ContentsOffset + 751 RelEntry.FixupSection->getSectionOffset() + 752 RelEntry.Offset; 753 754 LLVM_DEBUG(dbgs() << "applyRelocation: " << RelEntry << "\n"); 755 auto Value = getProvisionalValue(RelEntry, Layout); 756 757 switch (RelEntry.Type) { 758 case wasm::R_WASM_FUNCTION_INDEX_LEB: 759 case wasm::R_WASM_TYPE_INDEX_LEB: 760 case wasm::R_WASM_GLOBAL_INDEX_LEB: 761 case wasm::R_WASM_MEMORY_ADDR_LEB: 762 case wasm::R_WASM_TAG_INDEX_LEB: 763 case wasm::R_WASM_TABLE_NUMBER_LEB: 764 writePatchableLEB<5>(Stream, Value, Offset); 765 break; 766 case wasm::R_WASM_MEMORY_ADDR_LEB64: 767 writePatchableLEB<10>(Stream, Value, Offset); 768 break; 769 case wasm::R_WASM_TABLE_INDEX_I32: 770 case wasm::R_WASM_MEMORY_ADDR_I32: 771 case wasm::R_WASM_FUNCTION_OFFSET_I32: 772 case wasm::R_WASM_SECTION_OFFSET_I32: 773 case wasm::R_WASM_GLOBAL_INDEX_I32: 774 case wasm::R_WASM_MEMORY_ADDR_LOCREL_I32: 775 patchI32(Stream, Value, Offset); 776 break; 777 case wasm::R_WASM_TABLE_INDEX_I64: 778 case wasm::R_WASM_MEMORY_ADDR_I64: 779 case wasm::R_WASM_FUNCTION_OFFSET_I64: 780 patchI64(Stream, Value, Offset); 781 break; 782 case wasm::R_WASM_TABLE_INDEX_SLEB: 783 case wasm::R_WASM_TABLE_INDEX_REL_SLEB: 784 case wasm::R_WASM_MEMORY_ADDR_SLEB: 785 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB: 786 case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB: 787 writePatchableSLEB<5>(Stream, Value, Offset); 788 break; 789 case wasm::R_WASM_TABLE_INDEX_SLEB64: 790 case wasm::R_WASM_TABLE_INDEX_REL_SLEB64: 791 case wasm::R_WASM_MEMORY_ADDR_SLEB64: 792 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB64: 793 case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB64: 794 writePatchableSLEB<10>(Stream, Value, Offset); 795 break; 796 default: 797 llvm_unreachable("invalid relocation type"); 798 } 799 } 800 } 801 802 void WasmObjectWriter::writeTypeSection( 803 ArrayRef<wasm::WasmSignature> Signatures) { 804 if (Signatures.empty()) 805 return; 806 807 SectionBookkeeping Section; 808 startSection(Section, wasm::WASM_SEC_TYPE); 809 810 encodeULEB128(Signatures.size(), W->OS); 811 812 for (const wasm::WasmSignature &Sig : Signatures) { 813 W->OS << char(wasm::WASM_TYPE_FUNC); 814 encodeULEB128(Sig.Params.size(), W->OS); 815 for (wasm::ValType Ty : Sig.Params) 816 writeValueType(Ty); 817 encodeULEB128(Sig.Returns.size(), W->OS); 818 for (wasm::ValType Ty : Sig.Returns) 819 writeValueType(Ty); 820 } 821 822 endSection(Section); 823 } 824 825 void WasmObjectWriter::writeImportSection(ArrayRef<wasm::WasmImport> Imports, 826 uint64_t DataSize, 827 uint32_t NumElements) { 828 if (Imports.empty()) 829 return; 830 831 uint64_t NumPages = (DataSize + wasm::WasmPageSize - 1) / wasm::WasmPageSize; 832 833 SectionBookkeeping Section; 834 startSection(Section, wasm::WASM_SEC_IMPORT); 835 836 encodeULEB128(Imports.size(), W->OS); 837 for (const wasm::WasmImport &Import : Imports) { 838 writeString(Import.Module); 839 writeString(Import.Field); 840 W->OS << char(Import.Kind); 841 842 switch (Import.Kind) { 843 case wasm::WASM_EXTERNAL_FUNCTION: 844 encodeULEB128(Import.SigIndex, W->OS); 845 break; 846 case wasm::WASM_EXTERNAL_GLOBAL: 847 W->OS << char(Import.Global.Type); 848 W->OS << char(Import.Global.Mutable ? 1 : 0); 849 break; 850 case wasm::WASM_EXTERNAL_MEMORY: 851 encodeULEB128(Import.Memory.Flags, W->OS); 852 encodeULEB128(NumPages, W->OS); // initial 853 break; 854 case wasm::WASM_EXTERNAL_TABLE: 855 W->OS << char(Import.Table.ElemType); 856 encodeULEB128(0, W->OS); // flags 857 encodeULEB128(NumElements, W->OS); // initial 858 break; 859 case wasm::WASM_EXTERNAL_TAG: 860 W->OS << char(0); // Reserved 'attribute' field 861 encodeULEB128(Import.SigIndex, W->OS); 862 break; 863 default: 864 llvm_unreachable("unsupported import kind"); 865 } 866 } 867 868 endSection(Section); 869 } 870 871 void WasmObjectWriter::writeFunctionSection(ArrayRef<WasmFunction> Functions) { 872 if (Functions.empty()) 873 return; 874 875 SectionBookkeeping Section; 876 startSection(Section, wasm::WASM_SEC_FUNCTION); 877 878 encodeULEB128(Functions.size(), W->OS); 879 for (const WasmFunction &Func : Functions) 880 encodeULEB128(Func.SigIndex, W->OS); 881 882 endSection(Section); 883 } 884 885 void WasmObjectWriter::writeTagSection(ArrayRef<uint32_t> TagTypes) { 886 if (TagTypes.empty()) 887 return; 888 889 SectionBookkeeping Section; 890 startSection(Section, wasm::WASM_SEC_TAG); 891 892 encodeULEB128(TagTypes.size(), W->OS); 893 for (uint32_t Index : TagTypes) { 894 W->OS << char(0); // Reserved 'attribute' field 895 encodeULEB128(Index, W->OS); 896 } 897 898 endSection(Section); 899 } 900 901 void WasmObjectWriter::writeGlobalSection(ArrayRef<wasm::WasmGlobal> Globals) { 902 if (Globals.empty()) 903 return; 904 905 SectionBookkeeping Section; 906 startSection(Section, wasm::WASM_SEC_GLOBAL); 907 908 encodeULEB128(Globals.size(), W->OS); 909 for (const wasm::WasmGlobal &Global : Globals) { 910 encodeULEB128(Global.Type.Type, W->OS); 911 W->OS << char(Global.Type.Mutable); 912 W->OS << char(Global.InitExpr.Opcode); 913 switch (Global.Type.Type) { 914 case wasm::WASM_TYPE_I32: 915 encodeSLEB128(0, W->OS); 916 break; 917 case wasm::WASM_TYPE_I64: 918 encodeSLEB128(0, W->OS); 919 break; 920 case wasm::WASM_TYPE_F32: 921 writeI32(0); 922 break; 923 case wasm::WASM_TYPE_F64: 924 writeI64(0); 925 break; 926 case wasm::WASM_TYPE_EXTERNREF: 927 writeValueType(wasm::ValType::EXTERNREF); 928 break; 929 default: 930 llvm_unreachable("unexpected type"); 931 } 932 W->OS << char(wasm::WASM_OPCODE_END); 933 } 934 935 endSection(Section); 936 } 937 938 void WasmObjectWriter::writeTableSection(ArrayRef<wasm::WasmTable> Tables) { 939 if (Tables.empty()) 940 return; 941 942 SectionBookkeeping Section; 943 startSection(Section, wasm::WASM_SEC_TABLE); 944 945 encodeULEB128(Tables.size(), W->OS); 946 for (const wasm::WasmTable &Table : Tables) { 947 encodeULEB128(Table.Type.ElemType, W->OS); 948 encodeULEB128(Table.Type.Limits.Flags, W->OS); 949 encodeULEB128(Table.Type.Limits.Minimum, W->OS); 950 if (Table.Type.Limits.Flags & wasm::WASM_LIMITS_FLAG_HAS_MAX) 951 encodeULEB128(Table.Type.Limits.Maximum, W->OS); 952 } 953 endSection(Section); 954 } 955 956 void WasmObjectWriter::writeExportSection(ArrayRef<wasm::WasmExport> Exports) { 957 if (Exports.empty()) 958 return; 959 960 SectionBookkeeping Section; 961 startSection(Section, wasm::WASM_SEC_EXPORT); 962 963 encodeULEB128(Exports.size(), W->OS); 964 for (const wasm::WasmExport &Export : Exports) { 965 writeString(Export.Name); 966 W->OS << char(Export.Kind); 967 encodeULEB128(Export.Index, W->OS); 968 } 969 970 endSection(Section); 971 } 972 973 void WasmObjectWriter::writeElemSection( 974 const MCSymbolWasm *IndirectFunctionTable, ArrayRef<uint32_t> TableElems) { 975 if (TableElems.empty()) 976 return; 977 978 assert(IndirectFunctionTable); 979 980 SectionBookkeeping Section; 981 startSection(Section, wasm::WASM_SEC_ELEM); 982 983 encodeULEB128(1, W->OS); // number of "segments" 984 985 assert(WasmIndices.count(IndirectFunctionTable)); 986 uint32_t TableNumber = WasmIndices.find(IndirectFunctionTable)->second; 987 uint32_t Flags = 0; 988 if (TableNumber) 989 Flags |= wasm::WASM_ELEM_SEGMENT_HAS_TABLE_NUMBER; 990 encodeULEB128(Flags, W->OS); 991 if (Flags & wasm::WASM_ELEM_SEGMENT_HAS_TABLE_NUMBER) 992 encodeULEB128(TableNumber, W->OS); // the table number 993 994 // init expr for starting offset 995 W->OS << char(wasm::WASM_OPCODE_I32_CONST); 996 encodeSLEB128(InitialTableOffset, W->OS); 997 W->OS << char(wasm::WASM_OPCODE_END); 998 999 if (Flags & wasm::WASM_ELEM_SEGMENT_MASK_HAS_ELEM_KIND) { 1000 // We only write active function table initializers, for which the elem kind 1001 // is specified to be written as 0x00 and interpreted to mean "funcref". 1002 const uint8_t ElemKind = 0; 1003 W->OS << ElemKind; 1004 } 1005 1006 encodeULEB128(TableElems.size(), W->OS); 1007 for (uint32_t Elem : TableElems) 1008 encodeULEB128(Elem, W->OS); 1009 1010 endSection(Section); 1011 } 1012 1013 void WasmObjectWriter::writeDataCountSection() { 1014 if (DataSegments.empty()) 1015 return; 1016 1017 SectionBookkeeping Section; 1018 startSection(Section, wasm::WASM_SEC_DATACOUNT); 1019 encodeULEB128(DataSegments.size(), W->OS); 1020 endSection(Section); 1021 } 1022 1023 uint32_t WasmObjectWriter::writeCodeSection(const MCAssembler &Asm, 1024 const MCAsmLayout &Layout, 1025 ArrayRef<WasmFunction> Functions) { 1026 if (Functions.empty()) 1027 return 0; 1028 1029 SectionBookkeeping Section; 1030 startSection(Section, wasm::WASM_SEC_CODE); 1031 1032 encodeULEB128(Functions.size(), W->OS); 1033 1034 for (const WasmFunction &Func : Functions) { 1035 auto &FuncSection = static_cast<MCSectionWasm &>(Func.Sym->getSection()); 1036 1037 int64_t Size = 0; 1038 if (!Func.Sym->getSize()->evaluateAsAbsolute(Size, Layout)) 1039 report_fatal_error(".size expression must be evaluatable"); 1040 1041 encodeULEB128(Size, W->OS); 1042 FuncSection.setSectionOffset(W->OS.tell() - Section.ContentsOffset); 1043 Asm.writeSectionData(W->OS, &FuncSection, Layout); 1044 } 1045 1046 // Apply fixups. 1047 applyRelocations(CodeRelocations, Section.ContentsOffset, Layout); 1048 1049 endSection(Section); 1050 return Section.Index; 1051 } 1052 1053 uint32_t WasmObjectWriter::writeDataSection(const MCAsmLayout &Layout) { 1054 if (DataSegments.empty()) 1055 return 0; 1056 1057 SectionBookkeeping Section; 1058 startSection(Section, wasm::WASM_SEC_DATA); 1059 1060 encodeULEB128(DataSegments.size(), W->OS); // count 1061 1062 for (const WasmDataSegment &Segment : DataSegments) { 1063 encodeULEB128(Segment.InitFlags, W->OS); // flags 1064 if (Segment.InitFlags & wasm::WASM_DATA_SEGMENT_HAS_MEMINDEX) 1065 encodeULEB128(0, W->OS); // memory index 1066 if ((Segment.InitFlags & wasm::WASM_DATA_SEGMENT_IS_PASSIVE) == 0) { 1067 W->OS << char(is64Bit() ? wasm::WASM_OPCODE_I64_CONST 1068 : wasm::WASM_OPCODE_I32_CONST); 1069 encodeSLEB128(Segment.Offset, W->OS); // offset 1070 W->OS << char(wasm::WASM_OPCODE_END); 1071 } 1072 encodeULEB128(Segment.Data.size(), W->OS); // size 1073 Segment.Section->setSectionOffset(W->OS.tell() - Section.ContentsOffset); 1074 W->OS << Segment.Data; // data 1075 } 1076 1077 // Apply fixups. 1078 applyRelocations(DataRelocations, Section.ContentsOffset, Layout); 1079 1080 endSection(Section); 1081 return Section.Index; 1082 } 1083 1084 void WasmObjectWriter::writeRelocSection( 1085 uint32_t SectionIndex, StringRef Name, 1086 std::vector<WasmRelocationEntry> &Relocs) { 1087 // See: https://github.com/WebAssembly/tool-conventions/blob/main/Linking.md 1088 // for descriptions of the reloc sections. 1089 1090 if (Relocs.empty()) 1091 return; 1092 1093 // First, ensure the relocations are sorted in offset order. In general they 1094 // should already be sorted since `recordRelocation` is called in offset 1095 // order, but for the code section we combine many MC sections into single 1096 // wasm section, and this order is determined by the order of Asm.Symbols() 1097 // not the sections order. 1098 llvm::stable_sort( 1099 Relocs, [](const WasmRelocationEntry &A, const WasmRelocationEntry &B) { 1100 return (A.Offset + A.FixupSection->getSectionOffset()) < 1101 (B.Offset + B.FixupSection->getSectionOffset()); 1102 }); 1103 1104 SectionBookkeeping Section; 1105 startCustomSection(Section, std::string("reloc.") + Name.str()); 1106 1107 encodeULEB128(SectionIndex, W->OS); 1108 encodeULEB128(Relocs.size(), W->OS); 1109 for (const WasmRelocationEntry &RelEntry : Relocs) { 1110 uint64_t Offset = 1111 RelEntry.Offset + RelEntry.FixupSection->getSectionOffset(); 1112 uint32_t Index = getRelocationIndexValue(RelEntry); 1113 1114 W->OS << char(RelEntry.Type); 1115 encodeULEB128(Offset, W->OS); 1116 encodeULEB128(Index, W->OS); 1117 if (RelEntry.hasAddend()) 1118 encodeSLEB128(RelEntry.Addend, W->OS); 1119 } 1120 1121 endSection(Section); 1122 } 1123 1124 void WasmObjectWriter::writeCustomRelocSections() { 1125 for (const auto &Sec : CustomSections) { 1126 auto &Relocations = CustomSectionsRelocations[Sec.Section]; 1127 writeRelocSection(Sec.OutputIndex, Sec.Name, Relocations); 1128 } 1129 } 1130 1131 void WasmObjectWriter::writeLinkingMetaDataSection( 1132 ArrayRef<wasm::WasmSymbolInfo> SymbolInfos, 1133 ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs, 1134 const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats) { 1135 SectionBookkeeping Section; 1136 startCustomSection(Section, "linking"); 1137 encodeULEB128(wasm::WasmMetadataVersion, W->OS); 1138 1139 SectionBookkeeping SubSection; 1140 if (SymbolInfos.size() != 0) { 1141 startSection(SubSection, wasm::WASM_SYMBOL_TABLE); 1142 encodeULEB128(SymbolInfos.size(), W->OS); 1143 for (const wasm::WasmSymbolInfo &Sym : SymbolInfos) { 1144 encodeULEB128(Sym.Kind, W->OS); 1145 encodeULEB128(Sym.Flags, W->OS); 1146 switch (Sym.Kind) { 1147 case wasm::WASM_SYMBOL_TYPE_FUNCTION: 1148 case wasm::WASM_SYMBOL_TYPE_GLOBAL: 1149 case wasm::WASM_SYMBOL_TYPE_TAG: 1150 case wasm::WASM_SYMBOL_TYPE_TABLE: 1151 encodeULEB128(Sym.ElementIndex, W->OS); 1152 if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0 || 1153 (Sym.Flags & wasm::WASM_SYMBOL_EXPLICIT_NAME) != 0) 1154 writeString(Sym.Name); 1155 break; 1156 case wasm::WASM_SYMBOL_TYPE_DATA: 1157 writeString(Sym.Name); 1158 if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0) { 1159 encodeULEB128(Sym.DataRef.Segment, W->OS); 1160 encodeULEB128(Sym.DataRef.Offset, W->OS); 1161 encodeULEB128(Sym.DataRef.Size, W->OS); 1162 } 1163 break; 1164 case wasm::WASM_SYMBOL_TYPE_SECTION: { 1165 const uint32_t SectionIndex = 1166 CustomSections[Sym.ElementIndex].OutputIndex; 1167 encodeULEB128(SectionIndex, W->OS); 1168 break; 1169 } 1170 default: 1171 llvm_unreachable("unexpected kind"); 1172 } 1173 } 1174 endSection(SubSection); 1175 } 1176 1177 if (DataSegments.size()) { 1178 startSection(SubSection, wasm::WASM_SEGMENT_INFO); 1179 encodeULEB128(DataSegments.size(), W->OS); 1180 for (const WasmDataSegment &Segment : DataSegments) { 1181 writeString(Segment.Name); 1182 encodeULEB128(Segment.Alignment, W->OS); 1183 encodeULEB128(Segment.LinkingFlags, W->OS); 1184 } 1185 endSection(SubSection); 1186 } 1187 1188 if (!InitFuncs.empty()) { 1189 startSection(SubSection, wasm::WASM_INIT_FUNCS); 1190 encodeULEB128(InitFuncs.size(), W->OS); 1191 for (auto &StartFunc : InitFuncs) { 1192 encodeULEB128(StartFunc.first, W->OS); // priority 1193 encodeULEB128(StartFunc.second, W->OS); // function index 1194 } 1195 endSection(SubSection); 1196 } 1197 1198 if (Comdats.size()) { 1199 startSection(SubSection, wasm::WASM_COMDAT_INFO); 1200 encodeULEB128(Comdats.size(), W->OS); 1201 for (const auto &C : Comdats) { 1202 writeString(C.first); 1203 encodeULEB128(0, W->OS); // flags for future use 1204 encodeULEB128(C.second.size(), W->OS); 1205 for (const WasmComdatEntry &Entry : C.second) { 1206 encodeULEB128(Entry.Kind, W->OS); 1207 encodeULEB128(Entry.Index, W->OS); 1208 } 1209 } 1210 endSection(SubSection); 1211 } 1212 1213 endSection(Section); 1214 } 1215 1216 void WasmObjectWriter::writeCustomSection(WasmCustomSection &CustomSection, 1217 const MCAssembler &Asm, 1218 const MCAsmLayout &Layout) { 1219 SectionBookkeeping Section; 1220 auto *Sec = CustomSection.Section; 1221 startCustomSection(Section, CustomSection.Name); 1222 1223 Sec->setSectionOffset(W->OS.tell() - Section.ContentsOffset); 1224 Asm.writeSectionData(W->OS, Sec, Layout); 1225 1226 CustomSection.OutputContentsOffset = Section.ContentsOffset; 1227 CustomSection.OutputIndex = Section.Index; 1228 1229 endSection(Section); 1230 1231 // Apply fixups. 1232 auto &Relocations = CustomSectionsRelocations[CustomSection.Section]; 1233 applyRelocations(Relocations, CustomSection.OutputContentsOffset, Layout); 1234 } 1235 1236 uint32_t WasmObjectWriter::getFunctionType(const MCSymbolWasm &Symbol) { 1237 assert(Symbol.isFunction()); 1238 assert(TypeIndices.count(&Symbol)); 1239 return TypeIndices[&Symbol]; 1240 } 1241 1242 uint32_t WasmObjectWriter::getTagType(const MCSymbolWasm &Symbol) { 1243 assert(Symbol.isTag()); 1244 assert(TypeIndices.count(&Symbol)); 1245 return TypeIndices[&Symbol]; 1246 } 1247 1248 void WasmObjectWriter::registerFunctionType(const MCSymbolWasm &Symbol) { 1249 assert(Symbol.isFunction()); 1250 1251 wasm::WasmSignature S; 1252 1253 if (auto *Sig = Symbol.getSignature()) { 1254 S.Returns = Sig->Returns; 1255 S.Params = Sig->Params; 1256 } 1257 1258 auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size())); 1259 if (Pair.second) 1260 Signatures.push_back(S); 1261 TypeIndices[&Symbol] = Pair.first->second; 1262 1263 LLVM_DEBUG(dbgs() << "registerFunctionType: " << Symbol 1264 << " new:" << Pair.second << "\n"); 1265 LLVM_DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n"); 1266 } 1267 1268 void WasmObjectWriter::registerTagType(const MCSymbolWasm &Symbol) { 1269 assert(Symbol.isTag()); 1270 1271 // TODO Currently we don't generate imported exceptions, but if we do, we 1272 // should have a way of infering types of imported exceptions. 1273 wasm::WasmSignature S; 1274 if (auto *Sig = Symbol.getSignature()) { 1275 S.Returns = Sig->Returns; 1276 S.Params = Sig->Params; 1277 } 1278 1279 auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size())); 1280 if (Pair.second) 1281 Signatures.push_back(S); 1282 TypeIndices[&Symbol] = Pair.first->second; 1283 1284 LLVM_DEBUG(dbgs() << "registerTagType: " << Symbol << " new:" << Pair.second 1285 << "\n"); 1286 LLVM_DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n"); 1287 } 1288 1289 static bool isInSymtab(const MCSymbolWasm &Sym) { 1290 if (Sym.isUsedInReloc() || Sym.isUsedInInitArray()) 1291 return true; 1292 1293 if (Sym.isComdat() && !Sym.isDefined()) 1294 return false; 1295 1296 if (Sym.isTemporary()) 1297 return false; 1298 1299 if (Sym.isSection()) 1300 return false; 1301 1302 if (Sym.omitFromLinkingSection()) 1303 return false; 1304 1305 return true; 1306 } 1307 1308 void WasmObjectWriter::prepareImports( 1309 SmallVectorImpl<wasm::WasmImport> &Imports, MCAssembler &Asm, 1310 const MCAsmLayout &Layout) { 1311 // For now, always emit the memory import, since loads and stores are not 1312 // valid without it. In the future, we could perhaps be more clever and omit 1313 // it if there are no loads or stores. 1314 wasm::WasmImport MemImport; 1315 MemImport.Module = "env"; 1316 MemImport.Field = "__linear_memory"; 1317 MemImport.Kind = wasm::WASM_EXTERNAL_MEMORY; 1318 MemImport.Memory.Flags = is64Bit() ? wasm::WASM_LIMITS_FLAG_IS_64 1319 : wasm::WASM_LIMITS_FLAG_NONE; 1320 Imports.push_back(MemImport); 1321 1322 // Populate SignatureIndices, and Imports and WasmIndices for undefined 1323 // symbols. This must be done before populating WasmIndices for defined 1324 // symbols. 1325 for (const MCSymbol &S : Asm.symbols()) { 1326 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1327 1328 // Register types for all functions, including those with private linkage 1329 // (because wasm always needs a type signature). 1330 if (WS.isFunction()) { 1331 const auto *BS = Layout.getBaseSymbol(S); 1332 if (!BS) 1333 report_fatal_error(Twine(S.getName()) + 1334 ": absolute addressing not supported!"); 1335 registerFunctionType(*cast<MCSymbolWasm>(BS)); 1336 } 1337 1338 if (WS.isTag()) 1339 registerTagType(WS); 1340 1341 if (WS.isTemporary()) 1342 continue; 1343 1344 // If the symbol is not defined in this translation unit, import it. 1345 if (!WS.isDefined() && !WS.isComdat()) { 1346 if (WS.isFunction()) { 1347 wasm::WasmImport Import; 1348 Import.Module = WS.getImportModule(); 1349 Import.Field = WS.getImportName(); 1350 Import.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1351 Import.SigIndex = getFunctionType(WS); 1352 Imports.push_back(Import); 1353 assert(WasmIndices.count(&WS) == 0); 1354 WasmIndices[&WS] = NumFunctionImports++; 1355 } else if (WS.isGlobal()) { 1356 if (WS.isWeak()) 1357 report_fatal_error("undefined global symbol cannot be weak"); 1358 1359 wasm::WasmImport Import; 1360 Import.Field = WS.getImportName(); 1361 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1362 Import.Module = WS.getImportModule(); 1363 Import.Global = WS.getGlobalType(); 1364 Imports.push_back(Import); 1365 assert(WasmIndices.count(&WS) == 0); 1366 WasmIndices[&WS] = NumGlobalImports++; 1367 } else if (WS.isTag()) { 1368 if (WS.isWeak()) 1369 report_fatal_error("undefined tag symbol cannot be weak"); 1370 1371 wasm::WasmImport Import; 1372 Import.Module = WS.getImportModule(); 1373 Import.Field = WS.getImportName(); 1374 Import.Kind = wasm::WASM_EXTERNAL_TAG; 1375 Import.SigIndex = getTagType(WS); 1376 Imports.push_back(Import); 1377 assert(WasmIndices.count(&WS) == 0); 1378 WasmIndices[&WS] = NumTagImports++; 1379 } else if (WS.isTable()) { 1380 if (WS.isWeak()) 1381 report_fatal_error("undefined table symbol cannot be weak"); 1382 1383 wasm::WasmImport Import; 1384 Import.Module = WS.getImportModule(); 1385 Import.Field = WS.getImportName(); 1386 Import.Kind = wasm::WASM_EXTERNAL_TABLE; 1387 Import.Table = WS.getTableType(); 1388 Imports.push_back(Import); 1389 assert(WasmIndices.count(&WS) == 0); 1390 WasmIndices[&WS] = NumTableImports++; 1391 } 1392 } 1393 } 1394 1395 // Add imports for GOT globals 1396 for (const MCSymbol &S : Asm.symbols()) { 1397 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1398 if (WS.isUsedInGOT()) { 1399 wasm::WasmImport Import; 1400 if (WS.isFunction()) 1401 Import.Module = "GOT.func"; 1402 else 1403 Import.Module = "GOT.mem"; 1404 Import.Field = WS.getName(); 1405 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1406 Import.Global = {wasm::WASM_TYPE_I32, true}; 1407 Imports.push_back(Import); 1408 assert(GOTIndices.count(&WS) == 0); 1409 GOTIndices[&WS] = NumGlobalImports++; 1410 } 1411 } 1412 } 1413 1414 uint64_t WasmObjectWriter::writeObject(MCAssembler &Asm, 1415 const MCAsmLayout &Layout) { 1416 support::endian::Writer MainWriter(*OS, support::little); 1417 W = &MainWriter; 1418 if (IsSplitDwarf) { 1419 uint64_t TotalSize = writeOneObject(Asm, Layout, DwoMode::NonDwoOnly); 1420 assert(DwoOS); 1421 support::endian::Writer DwoWriter(*DwoOS, support::little); 1422 W = &DwoWriter; 1423 return TotalSize + writeOneObject(Asm, Layout, DwoMode::DwoOnly); 1424 } else { 1425 return writeOneObject(Asm, Layout, DwoMode::AllSections); 1426 } 1427 } 1428 1429 uint64_t WasmObjectWriter::writeOneObject(MCAssembler &Asm, 1430 const MCAsmLayout &Layout, 1431 DwoMode Mode) { 1432 uint64_t StartOffset = W->OS.tell(); 1433 SectionCount = 0; 1434 CustomSections.clear(); 1435 1436 LLVM_DEBUG(dbgs() << "WasmObjectWriter::writeObject\n"); 1437 1438 // Collect information from the available symbols. 1439 SmallVector<WasmFunction, 4> Functions; 1440 SmallVector<uint32_t, 4> TableElems; 1441 SmallVector<wasm::WasmImport, 4> Imports; 1442 SmallVector<wasm::WasmExport, 4> Exports; 1443 SmallVector<uint32_t, 2> TagTypes; 1444 SmallVector<wasm::WasmGlobal, 1> Globals; 1445 SmallVector<wasm::WasmTable, 1> Tables; 1446 SmallVector<wasm::WasmSymbolInfo, 4> SymbolInfos; 1447 SmallVector<std::pair<uint16_t, uint32_t>, 2> InitFuncs; 1448 std::map<StringRef, std::vector<WasmComdatEntry>> Comdats; 1449 uint64_t DataSize = 0; 1450 if (Mode != DwoMode::DwoOnly) { 1451 prepareImports(Imports, Asm, Layout); 1452 } 1453 1454 // Populate DataSegments and CustomSections, which must be done before 1455 // populating DataLocations. 1456 for (MCSection &Sec : Asm) { 1457 auto &Section = static_cast<MCSectionWasm &>(Sec); 1458 StringRef SectionName = Section.getName(); 1459 1460 if (Mode == DwoMode::NonDwoOnly && isDwoSection(Sec)) 1461 continue; 1462 if (Mode == DwoMode::DwoOnly && !isDwoSection(Sec)) 1463 continue; 1464 1465 LLVM_DEBUG(dbgs() << "Processing Section " << SectionName << " group " 1466 << Section.getGroup() << "\n";); 1467 1468 // .init_array sections are handled specially elsewhere. 1469 if (SectionName.startswith(".init_array")) 1470 continue; 1471 1472 // Code is handled separately 1473 if (Section.getKind().isText()) 1474 continue; 1475 1476 if (Section.isWasmData()) { 1477 uint32_t SegmentIndex = DataSegments.size(); 1478 DataSize = alignTo(DataSize, Section.getAlignment()); 1479 DataSegments.emplace_back(); 1480 WasmDataSegment &Segment = DataSegments.back(); 1481 Segment.Name = SectionName; 1482 Segment.InitFlags = Section.getPassive() 1483 ? (uint32_t)wasm::WASM_DATA_SEGMENT_IS_PASSIVE 1484 : 0; 1485 Segment.Offset = DataSize; 1486 Segment.Section = &Section; 1487 addData(Segment.Data, Section); 1488 Segment.Alignment = Log2_32(Section.getAlignment()); 1489 Segment.LinkingFlags = Section.getSegmentFlags(); 1490 DataSize += Segment.Data.size(); 1491 Section.setSegmentIndex(SegmentIndex); 1492 1493 if (const MCSymbolWasm *C = Section.getGroup()) { 1494 Comdats[C->getName()].emplace_back( 1495 WasmComdatEntry{wasm::WASM_COMDAT_DATA, SegmentIndex}); 1496 } 1497 } else { 1498 // Create custom sections 1499 assert(Sec.getKind().isMetadata()); 1500 1501 StringRef Name = SectionName; 1502 1503 // For user-defined custom sections, strip the prefix 1504 if (Name.startswith(".custom_section.")) 1505 Name = Name.substr(strlen(".custom_section.")); 1506 1507 MCSymbol *Begin = Sec.getBeginSymbol(); 1508 if (Begin) { 1509 assert(WasmIndices.count(cast<MCSymbolWasm>(Begin)) == 0); 1510 WasmIndices[cast<MCSymbolWasm>(Begin)] = CustomSections.size(); 1511 } 1512 1513 // Separate out the producers and target features sections 1514 if (Name == "producers") { 1515 ProducersSection = std::make_unique<WasmCustomSection>(Name, &Section); 1516 continue; 1517 } 1518 if (Name == "target_features") { 1519 TargetFeaturesSection = 1520 std::make_unique<WasmCustomSection>(Name, &Section); 1521 continue; 1522 } 1523 1524 // Custom sections can also belong to COMDAT groups. In this case the 1525 // decriptor's "index" field is the section index (in the final object 1526 // file), but that is not known until after layout, so it must be fixed up 1527 // later 1528 if (const MCSymbolWasm *C = Section.getGroup()) { 1529 Comdats[C->getName()].emplace_back( 1530 WasmComdatEntry{wasm::WASM_COMDAT_SECTION, 1531 static_cast<uint32_t>(CustomSections.size())}); 1532 } 1533 1534 CustomSections.emplace_back(Name, &Section); 1535 } 1536 } 1537 1538 if (Mode != DwoMode::DwoOnly) { 1539 // Populate WasmIndices and DataLocations for defined symbols. 1540 for (const MCSymbol &S : Asm.symbols()) { 1541 // Ignore unnamed temporary symbols, which aren't ever exported, imported, 1542 // or used in relocations. 1543 if (S.isTemporary() && S.getName().empty()) 1544 continue; 1545 1546 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1547 LLVM_DEBUG(dbgs() 1548 << "MCSymbol: " 1549 << toString(WS.getType().getValueOr(wasm::WASM_SYMBOL_TYPE_DATA)) 1550 << " '" << S << "'" 1551 << " isDefined=" << S.isDefined() << " isExternal=" 1552 << S.isExternal() << " isTemporary=" << S.isTemporary() 1553 << " isWeak=" << WS.isWeak() << " isHidden=" << WS.isHidden() 1554 << " isVariable=" << WS.isVariable() << "\n"); 1555 1556 if (WS.isVariable()) 1557 continue; 1558 if (WS.isComdat() && !WS.isDefined()) 1559 continue; 1560 1561 if (WS.isFunction()) { 1562 unsigned Index; 1563 if (WS.isDefined()) { 1564 if (WS.getOffset() != 0) 1565 report_fatal_error( 1566 "function sections must contain one function each"); 1567 1568 if (WS.getSize() == nullptr) 1569 report_fatal_error( 1570 "function symbols must have a size set with .size"); 1571 1572 // A definition. Write out the function body. 1573 Index = NumFunctionImports + Functions.size(); 1574 WasmFunction Func; 1575 Func.SigIndex = getFunctionType(WS); 1576 Func.Sym = &WS; 1577 assert(WasmIndices.count(&WS) == 0); 1578 WasmIndices[&WS] = Index; 1579 Functions.push_back(Func); 1580 1581 auto &Section = static_cast<MCSectionWasm &>(WS.getSection()); 1582 if (const MCSymbolWasm *C = Section.getGroup()) { 1583 Comdats[C->getName()].emplace_back( 1584 WasmComdatEntry{wasm::WASM_COMDAT_FUNCTION, Index}); 1585 } 1586 1587 if (WS.hasExportName()) { 1588 wasm::WasmExport Export; 1589 Export.Name = WS.getExportName(); 1590 Export.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1591 Export.Index = Index; 1592 Exports.push_back(Export); 1593 } 1594 } else { 1595 // An import; the index was assigned above. 1596 Index = WasmIndices.find(&WS)->second; 1597 } 1598 1599 LLVM_DEBUG(dbgs() << " -> function index: " << Index << "\n"); 1600 1601 } else if (WS.isData()) { 1602 if (!isInSymtab(WS)) 1603 continue; 1604 1605 if (!WS.isDefined()) { 1606 LLVM_DEBUG(dbgs() << " -> segment index: -1" 1607 << "\n"); 1608 continue; 1609 } 1610 1611 if (!WS.getSize()) 1612 report_fatal_error("data symbols must have a size set with .size: " + 1613 WS.getName()); 1614 1615 int64_t Size = 0; 1616 if (!WS.getSize()->evaluateAsAbsolute(Size, Layout)) 1617 report_fatal_error(".size expression must be evaluatable"); 1618 1619 auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection()); 1620 if (!DataSection.isWasmData()) 1621 report_fatal_error("data symbols must live in a data section: " + 1622 WS.getName()); 1623 1624 // For each data symbol, export it in the symtab as a reference to the 1625 // corresponding Wasm data segment. 1626 wasm::WasmDataReference Ref = wasm::WasmDataReference{ 1627 DataSection.getSegmentIndex(), Layout.getSymbolOffset(WS), 1628 static_cast<uint64_t>(Size)}; 1629 assert(DataLocations.count(&WS) == 0); 1630 DataLocations[&WS] = Ref; 1631 LLVM_DEBUG(dbgs() << " -> segment index: " << Ref.Segment << "\n"); 1632 1633 } else if (WS.isGlobal()) { 1634 // A "true" Wasm global (currently just __stack_pointer) 1635 if (WS.isDefined()) { 1636 wasm::WasmGlobal Global; 1637 Global.Type = WS.getGlobalType(); 1638 Global.Index = NumGlobalImports + Globals.size(); 1639 switch (Global.Type.Type) { 1640 case wasm::WASM_TYPE_I32: 1641 Global.InitExpr.Opcode = wasm::WASM_OPCODE_I32_CONST; 1642 break; 1643 case wasm::WASM_TYPE_I64: 1644 Global.InitExpr.Opcode = wasm::WASM_OPCODE_I64_CONST; 1645 break; 1646 case wasm::WASM_TYPE_F32: 1647 Global.InitExpr.Opcode = wasm::WASM_OPCODE_F32_CONST; 1648 break; 1649 case wasm::WASM_TYPE_F64: 1650 Global.InitExpr.Opcode = wasm::WASM_OPCODE_F64_CONST; 1651 break; 1652 case wasm::WASM_TYPE_EXTERNREF: 1653 Global.InitExpr.Opcode = wasm::WASM_OPCODE_REF_NULL; 1654 break; 1655 default: 1656 llvm_unreachable("unexpected type"); 1657 } 1658 assert(WasmIndices.count(&WS) == 0); 1659 WasmIndices[&WS] = Global.Index; 1660 Globals.push_back(Global); 1661 } else { 1662 // An import; the index was assigned above 1663 LLVM_DEBUG(dbgs() << " -> global index: " 1664 << WasmIndices.find(&WS)->second << "\n"); 1665 } 1666 } else if (WS.isTable()) { 1667 if (WS.isDefined()) { 1668 wasm::WasmTable Table; 1669 Table.Index = NumTableImports + Tables.size(); 1670 Table.Type = WS.getTableType(); 1671 assert(WasmIndices.count(&WS) == 0); 1672 WasmIndices[&WS] = Table.Index; 1673 Tables.push_back(Table); 1674 } 1675 LLVM_DEBUG(dbgs() << " -> table index: " 1676 << WasmIndices.find(&WS)->second << "\n"); 1677 } else if (WS.isTag()) { 1678 // C++ exception symbol (__cpp_exception) or longjmp symbol 1679 // (__c_longjmp) 1680 unsigned Index; 1681 if (WS.isDefined()) { 1682 Index = NumTagImports + TagTypes.size(); 1683 uint32_t SigIndex = getTagType(WS); 1684 assert(WasmIndices.count(&WS) == 0); 1685 WasmIndices[&WS] = Index; 1686 TagTypes.push_back(SigIndex); 1687 } else { 1688 // An import; the index was assigned above. 1689 assert(WasmIndices.count(&WS) > 0); 1690 } 1691 LLVM_DEBUG(dbgs() << " -> tag index: " << WasmIndices.find(&WS)->second 1692 << "\n"); 1693 1694 } else { 1695 assert(WS.isSection()); 1696 } 1697 } 1698 1699 // Populate WasmIndices and DataLocations for aliased symbols. We need to 1700 // process these in a separate pass because we need to have processed the 1701 // target of the alias before the alias itself and the symbols are not 1702 // necessarily ordered in this way. 1703 for (const MCSymbol &S : Asm.symbols()) { 1704 if (!S.isVariable()) 1705 continue; 1706 1707 assert(S.isDefined()); 1708 1709 const auto *BS = Layout.getBaseSymbol(S); 1710 if (!BS) 1711 report_fatal_error(Twine(S.getName()) + 1712 ": absolute addressing not supported!"); 1713 const MCSymbolWasm *Base = cast<MCSymbolWasm>(BS); 1714 1715 // Find the target symbol of this weak alias and export that index 1716 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1717 LLVM_DEBUG(dbgs() << WS.getName() << ": weak alias of '" << *Base 1718 << "'\n"); 1719 1720 if (Base->isFunction()) { 1721 assert(WasmIndices.count(Base) > 0); 1722 uint32_t WasmIndex = WasmIndices.find(Base)->second; 1723 assert(WasmIndices.count(&WS) == 0); 1724 WasmIndices[&WS] = WasmIndex; 1725 LLVM_DEBUG(dbgs() << " -> index:" << WasmIndex << "\n"); 1726 } else if (Base->isData()) { 1727 auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection()); 1728 uint64_t Offset = Layout.getSymbolOffset(S); 1729 int64_t Size = 0; 1730 // For data symbol alias we use the size of the base symbol as the 1731 // size of the alias. When an offset from the base is involved this 1732 // can result in a offset + size goes past the end of the data section 1733 // which out object format doesn't support. So we must clamp it. 1734 if (!Base->getSize()->evaluateAsAbsolute(Size, Layout)) 1735 report_fatal_error(".size expression must be evaluatable"); 1736 const WasmDataSegment &Segment = 1737 DataSegments[DataSection.getSegmentIndex()]; 1738 Size = 1739 std::min(static_cast<uint64_t>(Size), Segment.Data.size() - Offset); 1740 wasm::WasmDataReference Ref = wasm::WasmDataReference{ 1741 DataSection.getSegmentIndex(), 1742 static_cast<uint32_t>(Layout.getSymbolOffset(S)), 1743 static_cast<uint32_t>(Size)}; 1744 DataLocations[&WS] = Ref; 1745 LLVM_DEBUG(dbgs() << " -> index:" << Ref.Segment << "\n"); 1746 } else { 1747 report_fatal_error("don't yet support global/tag aliases"); 1748 } 1749 } 1750 } 1751 1752 // Finally, populate the symbol table itself, in its "natural" order. 1753 for (const MCSymbol &S : Asm.symbols()) { 1754 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1755 if (!isInSymtab(WS)) { 1756 WS.setIndex(InvalidIndex); 1757 continue; 1758 } 1759 LLVM_DEBUG(dbgs() << "adding to symtab: " << WS << "\n"); 1760 1761 uint32_t Flags = 0; 1762 if (WS.isWeak()) 1763 Flags |= wasm::WASM_SYMBOL_BINDING_WEAK; 1764 if (WS.isHidden()) 1765 Flags |= wasm::WASM_SYMBOL_VISIBILITY_HIDDEN; 1766 if (!WS.isExternal() && WS.isDefined()) 1767 Flags |= wasm::WASM_SYMBOL_BINDING_LOCAL; 1768 if (WS.isUndefined()) 1769 Flags |= wasm::WASM_SYMBOL_UNDEFINED; 1770 if (WS.isNoStrip()) { 1771 Flags |= wasm::WASM_SYMBOL_NO_STRIP; 1772 if (isEmscripten()) { 1773 Flags |= wasm::WASM_SYMBOL_EXPORTED; 1774 } 1775 } 1776 if (WS.hasImportName()) 1777 Flags |= wasm::WASM_SYMBOL_EXPLICIT_NAME; 1778 if (WS.hasExportName()) 1779 Flags |= wasm::WASM_SYMBOL_EXPORTED; 1780 if (WS.isTLS()) 1781 Flags |= wasm::WASM_SYMBOL_TLS; 1782 1783 wasm::WasmSymbolInfo Info; 1784 Info.Name = WS.getName(); 1785 Info.Kind = WS.getType().getValueOr(wasm::WASM_SYMBOL_TYPE_DATA); 1786 Info.Flags = Flags; 1787 if (!WS.isData()) { 1788 assert(WasmIndices.count(&WS) > 0); 1789 Info.ElementIndex = WasmIndices.find(&WS)->second; 1790 } else if (WS.isDefined()) { 1791 assert(DataLocations.count(&WS) > 0); 1792 Info.DataRef = DataLocations.find(&WS)->second; 1793 } 1794 WS.setIndex(SymbolInfos.size()); 1795 SymbolInfos.emplace_back(Info); 1796 } 1797 1798 { 1799 auto HandleReloc = [&](const WasmRelocationEntry &Rel) { 1800 // Functions referenced by a relocation need to put in the table. This is 1801 // purely to make the object file's provisional values readable, and is 1802 // ignored by the linker, which re-calculates the relocations itself. 1803 if (Rel.Type != wasm::R_WASM_TABLE_INDEX_I32 && 1804 Rel.Type != wasm::R_WASM_TABLE_INDEX_I64 && 1805 Rel.Type != wasm::R_WASM_TABLE_INDEX_SLEB && 1806 Rel.Type != wasm::R_WASM_TABLE_INDEX_SLEB64 && 1807 Rel.Type != wasm::R_WASM_TABLE_INDEX_REL_SLEB && 1808 Rel.Type != wasm::R_WASM_TABLE_INDEX_REL_SLEB64) 1809 return; 1810 assert(Rel.Symbol->isFunction()); 1811 const MCSymbolWasm *Base = 1812 cast<MCSymbolWasm>(Layout.getBaseSymbol(*Rel.Symbol)); 1813 uint32_t FunctionIndex = WasmIndices.find(Base)->second; 1814 uint32_t TableIndex = TableElems.size() + InitialTableOffset; 1815 if (TableIndices.try_emplace(Base, TableIndex).second) { 1816 LLVM_DEBUG(dbgs() << " -> adding " << Base->getName() 1817 << " to table: " << TableIndex << "\n"); 1818 TableElems.push_back(FunctionIndex); 1819 registerFunctionType(*Base); 1820 } 1821 }; 1822 1823 for (const WasmRelocationEntry &RelEntry : CodeRelocations) 1824 HandleReloc(RelEntry); 1825 for (const WasmRelocationEntry &RelEntry : DataRelocations) 1826 HandleReloc(RelEntry); 1827 } 1828 1829 // Translate .init_array section contents into start functions. 1830 for (const MCSection &S : Asm) { 1831 const auto &WS = static_cast<const MCSectionWasm &>(S); 1832 if (WS.getName().startswith(".fini_array")) 1833 report_fatal_error(".fini_array sections are unsupported"); 1834 if (!WS.getName().startswith(".init_array")) 1835 continue; 1836 if (WS.getFragmentList().empty()) 1837 continue; 1838 1839 // init_array is expected to contain a single non-empty data fragment 1840 if (WS.getFragmentList().size() != 3) 1841 report_fatal_error("only one .init_array section fragment supported"); 1842 1843 auto IT = WS.begin(); 1844 const MCFragment &EmptyFrag = *IT; 1845 if (EmptyFrag.getKind() != MCFragment::FT_Data) 1846 report_fatal_error(".init_array section should be aligned"); 1847 1848 IT = std::next(IT); 1849 const MCFragment &AlignFrag = *IT; 1850 if (AlignFrag.getKind() != MCFragment::FT_Align) 1851 report_fatal_error(".init_array section should be aligned"); 1852 if (cast<MCAlignFragment>(AlignFrag).getAlignment() != (is64Bit() ? 8 : 4)) 1853 report_fatal_error(".init_array section should be aligned for pointers"); 1854 1855 const MCFragment &Frag = *std::next(IT); 1856 if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data) 1857 report_fatal_error("only data supported in .init_array section"); 1858 1859 uint16_t Priority = UINT16_MAX; 1860 unsigned PrefixLength = strlen(".init_array"); 1861 if (WS.getName().size() > PrefixLength) { 1862 if (WS.getName()[PrefixLength] != '.') 1863 report_fatal_error( 1864 ".init_array section priority should start with '.'"); 1865 if (WS.getName().substr(PrefixLength + 1).getAsInteger(10, Priority)) 1866 report_fatal_error("invalid .init_array section priority"); 1867 } 1868 const auto &DataFrag = cast<MCDataFragment>(Frag); 1869 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 1870 for (const uint8_t * 1871 P = (const uint8_t *)Contents.data(), 1872 *End = (const uint8_t *)Contents.data() + Contents.size(); 1873 P != End; ++P) { 1874 if (*P != 0) 1875 report_fatal_error("non-symbolic data in .init_array section"); 1876 } 1877 for (const MCFixup &Fixup : DataFrag.getFixups()) { 1878 assert(Fixup.getKind() == 1879 MCFixup::getKindForSize(is64Bit() ? 8 : 4, false)); 1880 const MCExpr *Expr = Fixup.getValue(); 1881 auto *SymRef = dyn_cast<MCSymbolRefExpr>(Expr); 1882 if (!SymRef) 1883 report_fatal_error("fixups in .init_array should be symbol references"); 1884 const auto &TargetSym = cast<const MCSymbolWasm>(SymRef->getSymbol()); 1885 if (TargetSym.getIndex() == InvalidIndex) 1886 report_fatal_error("symbols in .init_array should exist in symtab"); 1887 if (!TargetSym.isFunction()) 1888 report_fatal_error("symbols in .init_array should be for functions"); 1889 InitFuncs.push_back( 1890 std::make_pair(Priority, TargetSym.getIndex())); 1891 } 1892 } 1893 1894 // Write out the Wasm header. 1895 writeHeader(Asm); 1896 1897 uint32_t CodeSectionIndex, DataSectionIndex; 1898 if (Mode != DwoMode::DwoOnly) { 1899 writeTypeSection(Signatures); 1900 writeImportSection(Imports, DataSize, TableElems.size()); 1901 writeFunctionSection(Functions); 1902 writeTableSection(Tables); 1903 // Skip the "memory" section; we import the memory instead. 1904 writeTagSection(TagTypes); 1905 writeGlobalSection(Globals); 1906 writeExportSection(Exports); 1907 const MCSymbol *IndirectFunctionTable = 1908 Asm.getContext().lookupSymbol("__indirect_function_table"); 1909 writeElemSection(cast_or_null<const MCSymbolWasm>(IndirectFunctionTable), 1910 TableElems); 1911 writeDataCountSection(); 1912 1913 CodeSectionIndex = writeCodeSection(Asm, Layout, Functions); 1914 DataSectionIndex = writeDataSection(Layout); 1915 } 1916 1917 // The Sections in the COMDAT list have placeholder indices (their index among 1918 // custom sections, rather than among all sections). Fix them up here. 1919 for (auto &Group : Comdats) { 1920 for (auto &Entry : Group.second) { 1921 if (Entry.Kind == wasm::WASM_COMDAT_SECTION) { 1922 Entry.Index += SectionCount; 1923 } 1924 } 1925 } 1926 for (auto &CustomSection : CustomSections) 1927 writeCustomSection(CustomSection, Asm, Layout); 1928 1929 if (Mode != DwoMode::DwoOnly) { 1930 writeLinkingMetaDataSection(SymbolInfos, InitFuncs, Comdats); 1931 1932 writeRelocSection(CodeSectionIndex, "CODE", CodeRelocations); 1933 writeRelocSection(DataSectionIndex, "DATA", DataRelocations); 1934 } 1935 writeCustomRelocSections(); 1936 if (ProducersSection) 1937 writeCustomSection(*ProducersSection, Asm, Layout); 1938 if (TargetFeaturesSection) 1939 writeCustomSection(*TargetFeaturesSection, Asm, Layout); 1940 1941 // TODO: Translate the .comment section to the output. 1942 return W->OS.tell() - StartOffset; 1943 } 1944 1945 std::unique_ptr<MCObjectWriter> 1946 llvm::createWasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 1947 raw_pwrite_stream &OS) { 1948 return std::make_unique<WasmObjectWriter>(std::move(MOTW), OS); 1949 } 1950 1951 std::unique_ptr<MCObjectWriter> 1952 llvm::createWasmDwoObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 1953 raw_pwrite_stream &OS, 1954 raw_pwrite_stream &DwoOS) { 1955 return std::make_unique<WasmObjectWriter>(std::move(MOTW), OS, DwoOS); 1956 } 1957