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 Value as an (unsigned) LEB value at offset Offset in Stream, padded 141 // to allow patching. 142 template <typename T, int W> 143 void writePatchableULEB(raw_pwrite_stream &Stream, T Value, uint64_t Offset) { 144 uint8_t Buffer[W]; 145 unsigned SizeLen = encodeULEB128(Value, Buffer, W); 146 assert(SizeLen == W); 147 Stream.pwrite((char *)Buffer, SizeLen, Offset); 148 } 149 150 // Write Value as an signed LEB value at offset Offset in Stream, padded 151 // to allow patching. 152 template <typename T, int W> 153 void writePatchableSLEB(raw_pwrite_stream &Stream, T Value, uint64_t Offset) { 154 uint8_t Buffer[W]; 155 unsigned SizeLen = encodeSLEB128(Value, Buffer, W); 156 assert(SizeLen == W); 157 Stream.pwrite((char *)Buffer, SizeLen, Offset); 158 } 159 160 static void writePatchableU32(raw_pwrite_stream &Stream, uint32_t Value, 161 uint64_t Offset) { 162 writePatchableULEB<uint32_t, 5>(Stream, Value, Offset); 163 } 164 165 static void writePatchableS32(raw_pwrite_stream &Stream, int32_t Value, 166 uint64_t Offset) { 167 writePatchableSLEB<int32_t, 5>(Stream, Value, Offset); 168 } 169 170 static void writePatchableU64(raw_pwrite_stream &Stream, uint64_t Value, 171 uint64_t Offset) { 172 writePatchableSLEB<uint64_t, 10>(Stream, Value, Offset); 173 } 174 175 static void writePatchableS64(raw_pwrite_stream &Stream, int64_t Value, 176 uint64_t Offset) { 177 writePatchableSLEB<int64_t, 10>(Stream, Value, Offset); 178 } 179 180 // Write Value as a plain integer value at offset Offset in Stream. 181 static void patchI32(raw_pwrite_stream &Stream, uint32_t Value, 182 uint64_t Offset) { 183 uint8_t Buffer[4]; 184 support::endian::write32le(Buffer, Value); 185 Stream.pwrite((char *)Buffer, sizeof(Buffer), Offset); 186 } 187 188 static void patchI64(raw_pwrite_stream &Stream, uint64_t Value, 189 uint64_t Offset) { 190 uint8_t Buffer[8]; 191 support::endian::write64le(Buffer, Value); 192 Stream.pwrite((char *)Buffer, sizeof(Buffer), Offset); 193 } 194 195 bool isDwoSection(const MCSection &Sec) { 196 return Sec.getName().endswith(".dwo"); 197 } 198 199 class WasmObjectWriter : public MCObjectWriter { 200 support::endian::Writer *W; 201 202 /// The target specific Wasm writer instance. 203 std::unique_ptr<MCWasmObjectTargetWriter> TargetObjectWriter; 204 205 // Relocations for fixing up references in the code section. 206 std::vector<WasmRelocationEntry> CodeRelocations; 207 // Relocations for fixing up references in the data section. 208 std::vector<WasmRelocationEntry> DataRelocations; 209 210 // Index values to use for fixing up call_indirect type indices. 211 // Maps function symbols to the index of the type of the function 212 DenseMap<const MCSymbolWasm *, uint32_t> TypeIndices; 213 // Maps function symbols to the table element index space. Used 214 // for TABLE_INDEX relocation types (i.e. address taken functions). 215 DenseMap<const MCSymbolWasm *, uint32_t> TableIndices; 216 // Maps function/global/table symbols to the 217 // function/global/table/tag/section index space. 218 DenseMap<const MCSymbolWasm *, uint32_t> WasmIndices; 219 DenseMap<const MCSymbolWasm *, uint32_t> GOTIndices; 220 // Maps data symbols to the Wasm segment and offset/size with the segment. 221 DenseMap<const MCSymbolWasm *, wasm::WasmDataReference> DataLocations; 222 223 // Stores output data (index, relocations, content offset) for custom 224 // section. 225 std::vector<WasmCustomSection> CustomSections; 226 std::unique_ptr<WasmCustomSection> ProducersSection; 227 std::unique_ptr<WasmCustomSection> TargetFeaturesSection; 228 // Relocations for fixing up references in the custom sections. 229 DenseMap<const MCSectionWasm *, std::vector<WasmRelocationEntry>> 230 CustomSectionsRelocations; 231 232 // Map from section to defining function symbol. 233 DenseMap<const MCSection *, const MCSymbol *> SectionFunctions; 234 235 DenseMap<wasm::WasmSignature, uint32_t> SignatureIndices; 236 SmallVector<wasm::WasmSignature, 4> Signatures; 237 SmallVector<WasmDataSegment, 4> DataSegments; 238 unsigned NumFunctionImports = 0; 239 unsigned NumGlobalImports = 0; 240 unsigned NumTableImports = 0; 241 unsigned NumTagImports = 0; 242 uint32_t SectionCount = 0; 243 244 enum class DwoMode { 245 AllSections, 246 NonDwoOnly, 247 DwoOnly, 248 }; 249 bool IsSplitDwarf = false; 250 raw_pwrite_stream *OS = nullptr; 251 raw_pwrite_stream *DwoOS = nullptr; 252 253 // TargetObjectWriter wranppers. 254 bool is64Bit() const { return TargetObjectWriter->is64Bit(); } 255 bool isEmscripten() const { return TargetObjectWriter->isEmscripten(); } 256 257 void startSection(SectionBookkeeping &Section, unsigned SectionId); 258 void startCustomSection(SectionBookkeeping &Section, StringRef Name); 259 void endSection(SectionBookkeeping &Section); 260 261 public: 262 WasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 263 raw_pwrite_stream &OS_) 264 : TargetObjectWriter(std::move(MOTW)), OS(&OS_) {} 265 266 WasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 267 raw_pwrite_stream &OS_, raw_pwrite_stream &DwoOS_) 268 : TargetObjectWriter(std::move(MOTW)), IsSplitDwarf(true), OS(&OS_), 269 DwoOS(&DwoOS_) {} 270 271 private: 272 void reset() override { 273 CodeRelocations.clear(); 274 DataRelocations.clear(); 275 TypeIndices.clear(); 276 WasmIndices.clear(); 277 GOTIndices.clear(); 278 TableIndices.clear(); 279 DataLocations.clear(); 280 CustomSections.clear(); 281 ProducersSection.reset(); 282 TargetFeaturesSection.reset(); 283 CustomSectionsRelocations.clear(); 284 SignatureIndices.clear(); 285 Signatures.clear(); 286 DataSegments.clear(); 287 SectionFunctions.clear(); 288 NumFunctionImports = 0; 289 NumGlobalImports = 0; 290 NumTableImports = 0; 291 MCObjectWriter::reset(); 292 } 293 294 void writeHeader(const MCAssembler &Asm); 295 296 void recordRelocation(MCAssembler &Asm, const MCAsmLayout &Layout, 297 const MCFragment *Fragment, const MCFixup &Fixup, 298 MCValue Target, uint64_t &FixedValue) override; 299 300 void executePostLayoutBinding(MCAssembler &Asm, 301 const MCAsmLayout &Layout) override; 302 void prepareImports(SmallVectorImpl<wasm::WasmImport> &Imports, 303 MCAssembler &Asm, const MCAsmLayout &Layout); 304 uint64_t writeObject(MCAssembler &Asm, const MCAsmLayout &Layout) override; 305 306 uint64_t writeOneObject(MCAssembler &Asm, const MCAsmLayout &Layout, 307 DwoMode Mode); 308 309 void writeString(const StringRef Str) { 310 encodeULEB128(Str.size(), W->OS); 311 W->OS << Str; 312 } 313 314 void writeStringWithAlignment(const StringRef Str, unsigned Alignment); 315 316 void writeI32(int32_t val) { 317 char Buffer[4]; 318 support::endian::write32le(Buffer, val); 319 W->OS.write(Buffer, sizeof(Buffer)); 320 } 321 322 void writeI64(int64_t val) { 323 char Buffer[8]; 324 support::endian::write64le(Buffer, val); 325 W->OS.write(Buffer, sizeof(Buffer)); 326 } 327 328 void writeValueType(wasm::ValType Ty) { W->OS << static_cast<char>(Ty); } 329 330 void writeTypeSection(ArrayRef<wasm::WasmSignature> Signatures); 331 void writeImportSection(ArrayRef<wasm::WasmImport> Imports, uint64_t DataSize, 332 uint32_t NumElements); 333 void writeFunctionSection(ArrayRef<WasmFunction> Functions); 334 void writeExportSection(ArrayRef<wasm::WasmExport> Exports); 335 void writeElemSection(const MCSymbolWasm *IndirectFunctionTable, 336 ArrayRef<uint32_t> TableElems); 337 void writeDataCountSection(); 338 uint32_t writeCodeSection(const MCAssembler &Asm, const MCAsmLayout &Layout, 339 ArrayRef<WasmFunction> Functions); 340 uint32_t writeDataSection(const MCAsmLayout &Layout); 341 void writeTagSection(ArrayRef<uint32_t> TagTypes); 342 void writeGlobalSection(ArrayRef<wasm::WasmGlobal> Globals); 343 void writeTableSection(ArrayRef<wasm::WasmTable> Tables); 344 void writeRelocSection(uint32_t SectionIndex, StringRef Name, 345 std::vector<WasmRelocationEntry> &Relocations); 346 void writeLinkingMetaDataSection( 347 ArrayRef<wasm::WasmSymbolInfo> SymbolInfos, 348 ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs, 349 const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats); 350 void writeCustomSection(WasmCustomSection &CustomSection, 351 const MCAssembler &Asm, const MCAsmLayout &Layout); 352 void writeCustomRelocSections(); 353 354 uint64_t getProvisionalValue(const WasmRelocationEntry &RelEntry, 355 const MCAsmLayout &Layout); 356 void applyRelocations(ArrayRef<WasmRelocationEntry> Relocations, 357 uint64_t ContentsOffset, const MCAsmLayout &Layout); 358 359 uint32_t getRelocationIndexValue(const WasmRelocationEntry &RelEntry); 360 uint32_t getFunctionType(const MCSymbolWasm &Symbol); 361 uint32_t getTagType(const MCSymbolWasm &Symbol); 362 void registerFunctionType(const MCSymbolWasm &Symbol); 363 void registerTagType(const MCSymbolWasm &Symbol); 364 }; 365 366 } // end anonymous namespace 367 368 // Write out a section header and a patchable section size field. 369 void WasmObjectWriter::startSection(SectionBookkeeping &Section, 370 unsigned SectionId) { 371 LLVM_DEBUG(dbgs() << "startSection " << SectionId << "\n"); 372 W->OS << char(SectionId); 373 374 Section.SizeOffset = W->OS.tell(); 375 376 // The section size. We don't know the size yet, so reserve enough space 377 // for any 32-bit value; we'll patch it later. 378 encodeULEB128(0, W->OS, 5); 379 380 // The position where the section starts, for measuring its size. 381 Section.ContentsOffset = W->OS.tell(); 382 Section.PayloadOffset = W->OS.tell(); 383 Section.Index = SectionCount++; 384 } 385 386 // Write a string with extra paddings for trailing alignment 387 // TODO: support alignment at asm and llvm level? 388 void WasmObjectWriter::writeStringWithAlignment(const StringRef Str, 389 unsigned Alignment) { 390 391 // Calculate the encoded size of str length and add pads based on it and 392 // alignment. 393 raw_null_ostream NullOS; 394 uint64_t StrSizeLength = encodeULEB128(Str.size(), NullOS); 395 uint64_t Offset = W->OS.tell() + StrSizeLength + Str.size(); 396 uint64_t Paddings = offsetToAlignment(Offset, Align(Alignment)); 397 Offset += Paddings; 398 399 // LEB128 greater than 5 bytes is invalid 400 assert((StrSizeLength + Paddings) <= 5 && "too long string to align"); 401 402 encodeSLEB128(Str.size(), W->OS, StrSizeLength + Paddings); 403 W->OS << Str; 404 405 assert(W->OS.tell() == Offset && "invalid padding"); 406 } 407 408 void WasmObjectWriter::startCustomSection(SectionBookkeeping &Section, 409 StringRef Name) { 410 LLVM_DEBUG(dbgs() << "startCustomSection " << Name << "\n"); 411 startSection(Section, wasm::WASM_SEC_CUSTOM); 412 413 // The position where the section header ends, for measuring its size. 414 Section.PayloadOffset = W->OS.tell(); 415 416 // Custom sections in wasm also have a string identifier. 417 if (Name != "__clangast") { 418 writeString(Name); 419 } else { 420 // The on-disk hashtable in clangast needs to be aligned by 4 bytes. 421 writeStringWithAlignment(Name, 4); 422 } 423 424 // The position where the custom section starts. 425 Section.ContentsOffset = W->OS.tell(); 426 } 427 428 // Now that the section is complete and we know how big it is, patch up the 429 // section size field at the start of the section. 430 void WasmObjectWriter::endSection(SectionBookkeeping &Section) { 431 uint64_t Size = W->OS.tell(); 432 // /dev/null doesn't support seek/tell and can report offset of 0. 433 // Simply skip this patching in that case. 434 if (!Size) 435 return; 436 437 Size -= Section.PayloadOffset; 438 if (uint32_t(Size) != Size) 439 report_fatal_error("section size does not fit in a uint32_t"); 440 441 LLVM_DEBUG(dbgs() << "endSection size=" << Size << "\n"); 442 443 // Write the final section size to the payload_len field, which follows 444 // the section id byte. 445 writePatchableU32(static_cast<raw_pwrite_stream &>(W->OS), Size, 446 Section.SizeOffset); 447 } 448 449 // Emit the Wasm header. 450 void WasmObjectWriter::writeHeader(const MCAssembler &Asm) { 451 W->OS.write(wasm::WasmMagic, sizeof(wasm::WasmMagic)); 452 W->write<uint32_t>(wasm::WasmVersion); 453 } 454 455 void WasmObjectWriter::executePostLayoutBinding(MCAssembler &Asm, 456 const MCAsmLayout &Layout) { 457 // Some compilation units require the indirect function table to be present 458 // but don't explicitly reference it. This is the case for call_indirect 459 // without the reference-types feature, and also function bitcasts in all 460 // cases. In those cases the __indirect_function_table has the 461 // WASM_SYMBOL_NO_STRIP attribute. Here we make sure this symbol makes it to 462 // the assembler, if needed. 463 if (auto *Sym = Asm.getContext().lookupSymbol("__indirect_function_table")) { 464 const auto *WasmSym = static_cast<const MCSymbolWasm *>(Sym); 465 if (WasmSym->isNoStrip()) 466 Asm.registerSymbol(*Sym); 467 } 468 469 // Build a map of sections to the function that defines them, for use 470 // in recordRelocation. 471 for (const MCSymbol &S : Asm.symbols()) { 472 const auto &WS = static_cast<const MCSymbolWasm &>(S); 473 if (WS.isDefined() && WS.isFunction() && !WS.isVariable()) { 474 const auto &Sec = static_cast<const MCSectionWasm &>(S.getSection()); 475 auto Pair = SectionFunctions.insert(std::make_pair(&Sec, &S)); 476 if (!Pair.second) 477 report_fatal_error("section already has a defining function: " + 478 Sec.getName()); 479 } 480 } 481 } 482 483 void WasmObjectWriter::recordRelocation(MCAssembler &Asm, 484 const MCAsmLayout &Layout, 485 const MCFragment *Fragment, 486 const MCFixup &Fixup, MCValue Target, 487 uint64_t &FixedValue) { 488 // The WebAssembly backend should never generate FKF_IsPCRel fixups 489 assert(!(Asm.getBackend().getFixupKindInfo(Fixup.getKind()).Flags & 490 MCFixupKindInfo::FKF_IsPCRel)); 491 492 const auto &FixupSection = cast<MCSectionWasm>(*Fragment->getParent()); 493 uint64_t C = Target.getConstant(); 494 uint64_t FixupOffset = Layout.getFragmentOffset(Fragment) + Fixup.getOffset(); 495 MCContext &Ctx = Asm.getContext(); 496 bool IsLocRel = false; 497 498 if (const MCSymbolRefExpr *RefB = Target.getSymB()) { 499 500 const auto &SymB = cast<MCSymbolWasm>(RefB->getSymbol()); 501 502 if (FixupSection.getKind().isText()) { 503 Ctx.reportError(Fixup.getLoc(), 504 Twine("symbol '") + SymB.getName() + 505 "' unsupported subtraction expression used in " 506 "relocation in code section."); 507 return; 508 } 509 510 if (SymB.isUndefined()) { 511 Ctx.reportError(Fixup.getLoc(), 512 Twine("symbol '") + SymB.getName() + 513 "' can not be undefined in a subtraction expression"); 514 return; 515 } 516 const MCSection &SecB = SymB.getSection(); 517 if (&SecB != &FixupSection) { 518 Ctx.reportError(Fixup.getLoc(), 519 Twine("symbol '") + SymB.getName() + 520 "' can not be placed in a different section"); 521 return; 522 } 523 IsLocRel = true; 524 C += FixupOffset - Layout.getSymbolOffset(SymB); 525 } 526 527 // We either rejected the fixup or folded B into C at this point. 528 const MCSymbolRefExpr *RefA = Target.getSymA(); 529 const auto *SymA = cast<MCSymbolWasm>(&RefA->getSymbol()); 530 531 // The .init_array isn't translated as data, so don't do relocations in it. 532 if (FixupSection.getName().startswith(".init_array")) { 533 SymA->setUsedInInitArray(); 534 return; 535 } 536 537 if (SymA->isVariable()) { 538 const MCExpr *Expr = SymA->getVariableValue(); 539 if (const auto *Inner = dyn_cast<MCSymbolRefExpr>(Expr)) 540 if (Inner->getKind() == MCSymbolRefExpr::VK_WEAKREF) 541 llvm_unreachable("weakref used in reloc not yet implemented"); 542 } 543 544 // Put any constant offset in an addend. Offsets can be negative, and 545 // LLVM expects wrapping, in contrast to wasm's immediates which can't 546 // be negative and don't wrap. 547 FixedValue = 0; 548 549 unsigned Type = 550 TargetObjectWriter->getRelocType(Target, Fixup, FixupSection, IsLocRel); 551 552 // Absolute offset within a section or a function. 553 // Currently only supported for for metadata sections. 554 // See: test/MC/WebAssembly/blockaddress.ll 555 if ((Type == wasm::R_WASM_FUNCTION_OFFSET_I32 || 556 Type == wasm::R_WASM_FUNCTION_OFFSET_I64 || 557 Type == wasm::R_WASM_SECTION_OFFSET_I32) && 558 SymA->isDefined()) { 559 // SymA can be a temp data symbol that represents a function (in which case 560 // it needs to be replaced by the section symbol), [XXX and it apparently 561 // later gets changed again to a func symbol?] or it can be a real 562 // function symbol, in which case it can be left as-is. 563 564 if (!FixupSection.getKind().isMetadata()) 565 report_fatal_error("relocations for function or section offsets are " 566 "only supported in metadata sections"); 567 568 const MCSymbol *SectionSymbol = nullptr; 569 const MCSection &SecA = SymA->getSection(); 570 if (SecA.getKind().isText()) { 571 auto SecSymIt = SectionFunctions.find(&SecA); 572 if (SecSymIt == SectionFunctions.end()) 573 report_fatal_error("section doesn\'t have defining symbol"); 574 SectionSymbol = SecSymIt->second; 575 } else { 576 SectionSymbol = SecA.getBeginSymbol(); 577 } 578 if (!SectionSymbol) 579 report_fatal_error("section symbol is required for relocation"); 580 581 C += Layout.getSymbolOffset(*SymA); 582 SymA = cast<MCSymbolWasm>(SectionSymbol); 583 } 584 585 if (Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB || 586 Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB64 || 587 Type == wasm::R_WASM_TABLE_INDEX_SLEB || 588 Type == wasm::R_WASM_TABLE_INDEX_SLEB64 || 589 Type == wasm::R_WASM_TABLE_INDEX_I32 || 590 Type == wasm::R_WASM_TABLE_INDEX_I64) { 591 // TABLE_INDEX relocs implicitly use the default indirect function table. 592 // We require the function table to have already been defined. 593 auto TableName = "__indirect_function_table"; 594 MCSymbolWasm *Sym = cast_or_null<MCSymbolWasm>(Ctx.lookupSymbol(TableName)); 595 if (!Sym) { 596 report_fatal_error("missing indirect function table symbol"); 597 } else { 598 if (!Sym->isFunctionTable()) 599 report_fatal_error("__indirect_function_table symbol has wrong type"); 600 // Ensure that __indirect_function_table reaches the output. 601 Sym->setNoStrip(); 602 Asm.registerSymbol(*Sym); 603 } 604 } 605 606 // Relocation other than R_WASM_TYPE_INDEX_LEB are required to be 607 // against a named symbol. 608 if (Type != wasm::R_WASM_TYPE_INDEX_LEB) { 609 if (SymA->getName().empty()) 610 report_fatal_error("relocations against un-named temporaries are not yet " 611 "supported by wasm"); 612 613 SymA->setUsedInReloc(); 614 } 615 616 switch (RefA->getKind()) { 617 case MCSymbolRefExpr::VK_GOT: 618 case MCSymbolRefExpr::VK_WASM_GOT_TLS: 619 SymA->setUsedInGOT(); 620 break; 621 default: 622 break; 623 } 624 625 WasmRelocationEntry Rec(FixupOffset, SymA, C, Type, &FixupSection); 626 LLVM_DEBUG(dbgs() << "WasmReloc: " << Rec << "\n"); 627 628 if (FixupSection.isWasmData()) { 629 DataRelocations.push_back(Rec); 630 } else if (FixupSection.getKind().isText()) { 631 CodeRelocations.push_back(Rec); 632 } else if (FixupSection.getKind().isMetadata()) { 633 CustomSectionsRelocations[&FixupSection].push_back(Rec); 634 } else { 635 llvm_unreachable("unexpected section type"); 636 } 637 } 638 639 // Compute a value to write into the code at the location covered 640 // by RelEntry. This value isn't used by the static linker; it just serves 641 // to make the object format more readable and more likely to be directly 642 // useable. 643 uint64_t 644 WasmObjectWriter::getProvisionalValue(const WasmRelocationEntry &RelEntry, 645 const MCAsmLayout &Layout) { 646 if ((RelEntry.Type == wasm::R_WASM_GLOBAL_INDEX_LEB || 647 RelEntry.Type == wasm::R_WASM_GLOBAL_INDEX_I32) && 648 !RelEntry.Symbol->isGlobal()) { 649 assert(GOTIndices.count(RelEntry.Symbol) > 0 && "symbol not found in GOT index space"); 650 return GOTIndices[RelEntry.Symbol]; 651 } 652 653 switch (RelEntry.Type) { 654 case wasm::R_WASM_TABLE_INDEX_REL_SLEB: 655 case wasm::R_WASM_TABLE_INDEX_REL_SLEB64: 656 case wasm::R_WASM_TABLE_INDEX_SLEB: 657 case wasm::R_WASM_TABLE_INDEX_SLEB64: 658 case wasm::R_WASM_TABLE_INDEX_I32: 659 case wasm::R_WASM_TABLE_INDEX_I64: { 660 // Provisional value is table address of the resolved symbol itself 661 const MCSymbolWasm *Base = 662 cast<MCSymbolWasm>(Layout.getBaseSymbol(*RelEntry.Symbol)); 663 assert(Base->isFunction()); 664 if (RelEntry.Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB || 665 RelEntry.Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB64) 666 return TableIndices[Base] - InitialTableOffset; 667 else 668 return TableIndices[Base]; 669 } 670 case wasm::R_WASM_TYPE_INDEX_LEB: 671 // Provisional value is same as the index 672 return getRelocationIndexValue(RelEntry); 673 case wasm::R_WASM_FUNCTION_INDEX_LEB: 674 case wasm::R_WASM_GLOBAL_INDEX_LEB: 675 case wasm::R_WASM_GLOBAL_INDEX_I32: 676 case wasm::R_WASM_TAG_INDEX_LEB: 677 case wasm::R_WASM_TABLE_NUMBER_LEB: 678 // Provisional value is function/global/tag Wasm index 679 assert(WasmIndices.count(RelEntry.Symbol) > 0 && "symbol not found in wasm index space"); 680 return WasmIndices[RelEntry.Symbol]; 681 case wasm::R_WASM_FUNCTION_OFFSET_I32: 682 case wasm::R_WASM_FUNCTION_OFFSET_I64: 683 case wasm::R_WASM_SECTION_OFFSET_I32: { 684 if (!RelEntry.Symbol->isDefined()) 685 return 0; 686 const auto &Section = 687 static_cast<const MCSectionWasm &>(RelEntry.Symbol->getSection()); 688 return Section.getSectionOffset() + RelEntry.Addend; 689 } 690 case wasm::R_WASM_MEMORY_ADDR_LEB: 691 case wasm::R_WASM_MEMORY_ADDR_LEB64: 692 case wasm::R_WASM_MEMORY_ADDR_SLEB: 693 case wasm::R_WASM_MEMORY_ADDR_SLEB64: 694 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB: 695 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB64: 696 case wasm::R_WASM_MEMORY_ADDR_I32: 697 case wasm::R_WASM_MEMORY_ADDR_I64: 698 case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB: 699 case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB64: 700 case wasm::R_WASM_MEMORY_ADDR_LOCREL_I32: { 701 // Provisional value is address of the global plus the offset 702 // For undefined symbols, use zero 703 if (!RelEntry.Symbol->isDefined()) 704 return 0; 705 const wasm::WasmDataReference &SymRef = DataLocations[RelEntry.Symbol]; 706 const WasmDataSegment &Segment = DataSegments[SymRef.Segment]; 707 // Ignore overflow. LLVM allows address arithmetic to silently wrap. 708 return Segment.Offset + SymRef.Offset + RelEntry.Addend; 709 } 710 default: 711 llvm_unreachable("invalid relocation type"); 712 } 713 } 714 715 static void addData(SmallVectorImpl<char> &DataBytes, 716 MCSectionWasm &DataSection) { 717 LLVM_DEBUG(errs() << "addData: " << DataSection.getName() << "\n"); 718 719 DataBytes.resize(alignTo(DataBytes.size(), DataSection.getAlignment())); 720 721 for (const MCFragment &Frag : DataSection) { 722 if (Frag.hasInstructions()) 723 report_fatal_error("only data supported in data sections"); 724 725 if (auto *Align = dyn_cast<MCAlignFragment>(&Frag)) { 726 if (Align->getValueSize() != 1) 727 report_fatal_error("only byte values supported for alignment"); 728 // If nops are requested, use zeros, as this is the data section. 729 uint8_t Value = Align->hasEmitNops() ? 0 : Align->getValue(); 730 uint64_t Size = 731 std::min<uint64_t>(alignTo(DataBytes.size(), Align->getAlignment()), 732 DataBytes.size() + Align->getMaxBytesToEmit()); 733 DataBytes.resize(Size, Value); 734 } else if (auto *Fill = dyn_cast<MCFillFragment>(&Frag)) { 735 int64_t NumValues; 736 if (!Fill->getNumValues().evaluateAsAbsolute(NumValues)) 737 llvm_unreachable("The fill should be an assembler constant"); 738 DataBytes.insert(DataBytes.end(), Fill->getValueSize() * NumValues, 739 Fill->getValue()); 740 } else if (auto *LEB = dyn_cast<MCLEBFragment>(&Frag)) { 741 const SmallVectorImpl<char> &Contents = LEB->getContents(); 742 llvm::append_range(DataBytes, Contents); 743 } else { 744 const auto &DataFrag = cast<MCDataFragment>(Frag); 745 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 746 llvm::append_range(DataBytes, Contents); 747 } 748 } 749 750 LLVM_DEBUG(dbgs() << "addData -> " << DataBytes.size() << "\n"); 751 } 752 753 uint32_t 754 WasmObjectWriter::getRelocationIndexValue(const WasmRelocationEntry &RelEntry) { 755 if (RelEntry.Type == wasm::R_WASM_TYPE_INDEX_LEB) { 756 if (!TypeIndices.count(RelEntry.Symbol)) 757 report_fatal_error("symbol not found in type index space: " + 758 RelEntry.Symbol->getName()); 759 return TypeIndices[RelEntry.Symbol]; 760 } 761 762 return RelEntry.Symbol->getIndex(); 763 } 764 765 // Apply the portions of the relocation records that we can handle ourselves 766 // directly. 767 void WasmObjectWriter::applyRelocations( 768 ArrayRef<WasmRelocationEntry> Relocations, uint64_t ContentsOffset, 769 const MCAsmLayout &Layout) { 770 auto &Stream = static_cast<raw_pwrite_stream &>(W->OS); 771 for (const WasmRelocationEntry &RelEntry : Relocations) { 772 uint64_t Offset = ContentsOffset + 773 RelEntry.FixupSection->getSectionOffset() + 774 RelEntry.Offset; 775 776 LLVM_DEBUG(dbgs() << "applyRelocation: " << RelEntry << "\n"); 777 uint64_t Value = getProvisionalValue(RelEntry, Layout); 778 779 switch (RelEntry.Type) { 780 case wasm::R_WASM_FUNCTION_INDEX_LEB: 781 case wasm::R_WASM_TYPE_INDEX_LEB: 782 case wasm::R_WASM_GLOBAL_INDEX_LEB: 783 case wasm::R_WASM_MEMORY_ADDR_LEB: 784 case wasm::R_WASM_TAG_INDEX_LEB: 785 case wasm::R_WASM_TABLE_NUMBER_LEB: 786 writePatchableU32(Stream, Value, Offset); 787 break; 788 case wasm::R_WASM_MEMORY_ADDR_LEB64: 789 writePatchableU64(Stream, Value, Offset); 790 break; 791 case wasm::R_WASM_TABLE_INDEX_I32: 792 case wasm::R_WASM_MEMORY_ADDR_I32: 793 case wasm::R_WASM_FUNCTION_OFFSET_I32: 794 case wasm::R_WASM_SECTION_OFFSET_I32: 795 case wasm::R_WASM_GLOBAL_INDEX_I32: 796 case wasm::R_WASM_MEMORY_ADDR_LOCREL_I32: 797 patchI32(Stream, Value, Offset); 798 break; 799 case wasm::R_WASM_TABLE_INDEX_I64: 800 case wasm::R_WASM_MEMORY_ADDR_I64: 801 case wasm::R_WASM_FUNCTION_OFFSET_I64: 802 patchI64(Stream, Value, Offset); 803 break; 804 case wasm::R_WASM_TABLE_INDEX_SLEB: 805 case wasm::R_WASM_TABLE_INDEX_REL_SLEB: 806 case wasm::R_WASM_MEMORY_ADDR_SLEB: 807 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB: 808 case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB: 809 writePatchableS32(Stream, Value, Offset); 810 break; 811 case wasm::R_WASM_TABLE_INDEX_SLEB64: 812 case wasm::R_WASM_TABLE_INDEX_REL_SLEB64: 813 case wasm::R_WASM_MEMORY_ADDR_SLEB64: 814 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB64: 815 case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB64: 816 writePatchableS64(Stream, Value, Offset); 817 break; 818 default: 819 llvm_unreachable("invalid relocation type"); 820 } 821 } 822 } 823 824 void WasmObjectWriter::writeTypeSection( 825 ArrayRef<wasm::WasmSignature> Signatures) { 826 if (Signatures.empty()) 827 return; 828 829 SectionBookkeeping Section; 830 startSection(Section, wasm::WASM_SEC_TYPE); 831 832 encodeULEB128(Signatures.size(), W->OS); 833 834 for (const wasm::WasmSignature &Sig : Signatures) { 835 W->OS << char(wasm::WASM_TYPE_FUNC); 836 encodeULEB128(Sig.Params.size(), W->OS); 837 for (wasm::ValType Ty : Sig.Params) 838 writeValueType(Ty); 839 encodeULEB128(Sig.Returns.size(), W->OS); 840 for (wasm::ValType Ty : Sig.Returns) 841 writeValueType(Ty); 842 } 843 844 endSection(Section); 845 } 846 847 void WasmObjectWriter::writeImportSection(ArrayRef<wasm::WasmImport> Imports, 848 uint64_t DataSize, 849 uint32_t NumElements) { 850 if (Imports.empty()) 851 return; 852 853 uint64_t NumPages = (DataSize + wasm::WasmPageSize - 1) / wasm::WasmPageSize; 854 855 SectionBookkeeping Section; 856 startSection(Section, wasm::WASM_SEC_IMPORT); 857 858 encodeULEB128(Imports.size(), W->OS); 859 for (const wasm::WasmImport &Import : Imports) { 860 writeString(Import.Module); 861 writeString(Import.Field); 862 W->OS << char(Import.Kind); 863 864 switch (Import.Kind) { 865 case wasm::WASM_EXTERNAL_FUNCTION: 866 encodeULEB128(Import.SigIndex, W->OS); 867 break; 868 case wasm::WASM_EXTERNAL_GLOBAL: 869 W->OS << char(Import.Global.Type); 870 W->OS << char(Import.Global.Mutable ? 1 : 0); 871 break; 872 case wasm::WASM_EXTERNAL_MEMORY: 873 encodeULEB128(Import.Memory.Flags, W->OS); 874 encodeULEB128(NumPages, W->OS); // initial 875 break; 876 case wasm::WASM_EXTERNAL_TABLE: 877 W->OS << char(Import.Table.ElemType); 878 encodeULEB128(0, W->OS); // flags 879 encodeULEB128(NumElements, W->OS); // initial 880 break; 881 case wasm::WASM_EXTERNAL_TAG: 882 W->OS << char(0); // Reserved 'attribute' field 883 encodeULEB128(Import.SigIndex, W->OS); 884 break; 885 default: 886 llvm_unreachable("unsupported import kind"); 887 } 888 } 889 890 endSection(Section); 891 } 892 893 void WasmObjectWriter::writeFunctionSection(ArrayRef<WasmFunction> Functions) { 894 if (Functions.empty()) 895 return; 896 897 SectionBookkeeping Section; 898 startSection(Section, wasm::WASM_SEC_FUNCTION); 899 900 encodeULEB128(Functions.size(), W->OS); 901 for (const WasmFunction &Func : Functions) 902 encodeULEB128(Func.SigIndex, W->OS); 903 904 endSection(Section); 905 } 906 907 void WasmObjectWriter::writeTagSection(ArrayRef<uint32_t> TagTypes) { 908 if (TagTypes.empty()) 909 return; 910 911 SectionBookkeeping Section; 912 startSection(Section, wasm::WASM_SEC_TAG); 913 914 encodeULEB128(TagTypes.size(), W->OS); 915 for (uint32_t Index : TagTypes) { 916 W->OS << char(0); // Reserved 'attribute' field 917 encodeULEB128(Index, W->OS); 918 } 919 920 endSection(Section); 921 } 922 923 void WasmObjectWriter::writeGlobalSection(ArrayRef<wasm::WasmGlobal> Globals) { 924 if (Globals.empty()) 925 return; 926 927 SectionBookkeeping Section; 928 startSection(Section, wasm::WASM_SEC_GLOBAL); 929 930 encodeULEB128(Globals.size(), W->OS); 931 for (const wasm::WasmGlobal &Global : Globals) { 932 encodeULEB128(Global.Type.Type, W->OS); 933 W->OS << char(Global.Type.Mutable); 934 if (Global.InitExpr.Extended) { 935 llvm_unreachable("extected init expressions not supported"); 936 } else { 937 W->OS << char(Global.InitExpr.Inst.Opcode); 938 switch (Global.Type.Type) { 939 case wasm::WASM_TYPE_I32: 940 encodeSLEB128(0, W->OS); 941 break; 942 case wasm::WASM_TYPE_I64: 943 encodeSLEB128(0, W->OS); 944 break; 945 case wasm::WASM_TYPE_F32: 946 writeI32(0); 947 break; 948 case wasm::WASM_TYPE_F64: 949 writeI64(0); 950 break; 951 case wasm::WASM_TYPE_EXTERNREF: 952 writeValueType(wasm::ValType::EXTERNREF); 953 break; 954 default: 955 llvm_unreachable("unexpected type"); 956 } 957 } 958 W->OS << char(wasm::WASM_OPCODE_END); 959 } 960 961 endSection(Section); 962 } 963 964 void WasmObjectWriter::writeTableSection(ArrayRef<wasm::WasmTable> Tables) { 965 if (Tables.empty()) 966 return; 967 968 SectionBookkeeping Section; 969 startSection(Section, wasm::WASM_SEC_TABLE); 970 971 encodeULEB128(Tables.size(), W->OS); 972 for (const wasm::WasmTable &Table : Tables) { 973 encodeULEB128(Table.Type.ElemType, W->OS); 974 encodeULEB128(Table.Type.Limits.Flags, W->OS); 975 encodeULEB128(Table.Type.Limits.Minimum, W->OS); 976 if (Table.Type.Limits.Flags & wasm::WASM_LIMITS_FLAG_HAS_MAX) 977 encodeULEB128(Table.Type.Limits.Maximum, W->OS); 978 } 979 endSection(Section); 980 } 981 982 void WasmObjectWriter::writeExportSection(ArrayRef<wasm::WasmExport> Exports) { 983 if (Exports.empty()) 984 return; 985 986 SectionBookkeeping Section; 987 startSection(Section, wasm::WASM_SEC_EXPORT); 988 989 encodeULEB128(Exports.size(), W->OS); 990 for (const wasm::WasmExport &Export : Exports) { 991 writeString(Export.Name); 992 W->OS << char(Export.Kind); 993 encodeULEB128(Export.Index, W->OS); 994 } 995 996 endSection(Section); 997 } 998 999 void WasmObjectWriter::writeElemSection( 1000 const MCSymbolWasm *IndirectFunctionTable, ArrayRef<uint32_t> TableElems) { 1001 if (TableElems.empty()) 1002 return; 1003 1004 assert(IndirectFunctionTable); 1005 1006 SectionBookkeeping Section; 1007 startSection(Section, wasm::WASM_SEC_ELEM); 1008 1009 encodeULEB128(1, W->OS); // number of "segments" 1010 1011 assert(WasmIndices.count(IndirectFunctionTable)); 1012 uint32_t TableNumber = WasmIndices.find(IndirectFunctionTable)->second; 1013 uint32_t Flags = 0; 1014 if (TableNumber) 1015 Flags |= wasm::WASM_ELEM_SEGMENT_HAS_TABLE_NUMBER; 1016 encodeULEB128(Flags, W->OS); 1017 if (Flags & wasm::WASM_ELEM_SEGMENT_HAS_TABLE_NUMBER) 1018 encodeULEB128(TableNumber, W->OS); // the table number 1019 1020 // init expr for starting offset 1021 W->OS << char(wasm::WASM_OPCODE_I32_CONST); 1022 encodeSLEB128(InitialTableOffset, W->OS); 1023 W->OS << char(wasm::WASM_OPCODE_END); 1024 1025 if (Flags & wasm::WASM_ELEM_SEGMENT_MASK_HAS_ELEM_KIND) { 1026 // We only write active function table initializers, for which the elem kind 1027 // is specified to be written as 0x00 and interpreted to mean "funcref". 1028 const uint8_t ElemKind = 0; 1029 W->OS << ElemKind; 1030 } 1031 1032 encodeULEB128(TableElems.size(), W->OS); 1033 for (uint32_t Elem : TableElems) 1034 encodeULEB128(Elem, W->OS); 1035 1036 endSection(Section); 1037 } 1038 1039 void WasmObjectWriter::writeDataCountSection() { 1040 if (DataSegments.empty()) 1041 return; 1042 1043 SectionBookkeeping Section; 1044 startSection(Section, wasm::WASM_SEC_DATACOUNT); 1045 encodeULEB128(DataSegments.size(), W->OS); 1046 endSection(Section); 1047 } 1048 1049 uint32_t WasmObjectWriter::writeCodeSection(const MCAssembler &Asm, 1050 const MCAsmLayout &Layout, 1051 ArrayRef<WasmFunction> Functions) { 1052 if (Functions.empty()) 1053 return 0; 1054 1055 SectionBookkeeping Section; 1056 startSection(Section, wasm::WASM_SEC_CODE); 1057 1058 encodeULEB128(Functions.size(), W->OS); 1059 1060 for (const WasmFunction &Func : Functions) { 1061 auto &FuncSection = static_cast<MCSectionWasm &>(Func.Sym->getSection()); 1062 1063 int64_t Size = 0; 1064 if (!Func.Sym->getSize()->evaluateAsAbsolute(Size, Layout)) 1065 report_fatal_error(".size expression must be evaluatable"); 1066 1067 encodeULEB128(Size, W->OS); 1068 FuncSection.setSectionOffset(W->OS.tell() - Section.ContentsOffset); 1069 Asm.writeSectionData(W->OS, &FuncSection, Layout); 1070 } 1071 1072 // Apply fixups. 1073 applyRelocations(CodeRelocations, Section.ContentsOffset, Layout); 1074 1075 endSection(Section); 1076 return Section.Index; 1077 } 1078 1079 uint32_t WasmObjectWriter::writeDataSection(const MCAsmLayout &Layout) { 1080 if (DataSegments.empty()) 1081 return 0; 1082 1083 SectionBookkeeping Section; 1084 startSection(Section, wasm::WASM_SEC_DATA); 1085 1086 encodeULEB128(DataSegments.size(), W->OS); // count 1087 1088 for (const WasmDataSegment &Segment : DataSegments) { 1089 encodeULEB128(Segment.InitFlags, W->OS); // flags 1090 if (Segment.InitFlags & wasm::WASM_DATA_SEGMENT_HAS_MEMINDEX) 1091 encodeULEB128(0, W->OS); // memory index 1092 if ((Segment.InitFlags & wasm::WASM_DATA_SEGMENT_IS_PASSIVE) == 0) { 1093 W->OS << char(is64Bit() ? wasm::WASM_OPCODE_I64_CONST 1094 : wasm::WASM_OPCODE_I32_CONST); 1095 encodeSLEB128(Segment.Offset, W->OS); // offset 1096 W->OS << char(wasm::WASM_OPCODE_END); 1097 } 1098 encodeULEB128(Segment.Data.size(), W->OS); // size 1099 Segment.Section->setSectionOffset(W->OS.tell() - Section.ContentsOffset); 1100 W->OS << Segment.Data; // data 1101 } 1102 1103 // Apply fixups. 1104 applyRelocations(DataRelocations, Section.ContentsOffset, Layout); 1105 1106 endSection(Section); 1107 return Section.Index; 1108 } 1109 1110 void WasmObjectWriter::writeRelocSection( 1111 uint32_t SectionIndex, StringRef Name, 1112 std::vector<WasmRelocationEntry> &Relocs) { 1113 // See: https://github.com/WebAssembly/tool-conventions/blob/main/Linking.md 1114 // for descriptions of the reloc sections. 1115 1116 if (Relocs.empty()) 1117 return; 1118 1119 // First, ensure the relocations are sorted in offset order. In general they 1120 // should already be sorted since `recordRelocation` is called in offset 1121 // order, but for the code section we combine many MC sections into single 1122 // wasm section, and this order is determined by the order of Asm.Symbols() 1123 // not the sections order. 1124 llvm::stable_sort( 1125 Relocs, [](const WasmRelocationEntry &A, const WasmRelocationEntry &B) { 1126 return (A.Offset + A.FixupSection->getSectionOffset()) < 1127 (B.Offset + B.FixupSection->getSectionOffset()); 1128 }); 1129 1130 SectionBookkeeping Section; 1131 startCustomSection(Section, std::string("reloc.") + Name.str()); 1132 1133 encodeULEB128(SectionIndex, W->OS); 1134 encodeULEB128(Relocs.size(), W->OS); 1135 for (const WasmRelocationEntry &RelEntry : Relocs) { 1136 uint64_t Offset = 1137 RelEntry.Offset + RelEntry.FixupSection->getSectionOffset(); 1138 uint32_t Index = getRelocationIndexValue(RelEntry); 1139 1140 W->OS << char(RelEntry.Type); 1141 encodeULEB128(Offset, W->OS); 1142 encodeULEB128(Index, W->OS); 1143 if (RelEntry.hasAddend()) 1144 encodeSLEB128(RelEntry.Addend, W->OS); 1145 } 1146 1147 endSection(Section); 1148 } 1149 1150 void WasmObjectWriter::writeCustomRelocSections() { 1151 for (const auto &Sec : CustomSections) { 1152 auto &Relocations = CustomSectionsRelocations[Sec.Section]; 1153 writeRelocSection(Sec.OutputIndex, Sec.Name, Relocations); 1154 } 1155 } 1156 1157 void WasmObjectWriter::writeLinkingMetaDataSection( 1158 ArrayRef<wasm::WasmSymbolInfo> SymbolInfos, 1159 ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs, 1160 const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats) { 1161 SectionBookkeeping Section; 1162 startCustomSection(Section, "linking"); 1163 encodeULEB128(wasm::WasmMetadataVersion, W->OS); 1164 1165 SectionBookkeeping SubSection; 1166 if (SymbolInfos.size() != 0) { 1167 startSection(SubSection, wasm::WASM_SYMBOL_TABLE); 1168 encodeULEB128(SymbolInfos.size(), W->OS); 1169 for (const wasm::WasmSymbolInfo &Sym : SymbolInfos) { 1170 encodeULEB128(Sym.Kind, W->OS); 1171 encodeULEB128(Sym.Flags, W->OS); 1172 switch (Sym.Kind) { 1173 case wasm::WASM_SYMBOL_TYPE_FUNCTION: 1174 case wasm::WASM_SYMBOL_TYPE_GLOBAL: 1175 case wasm::WASM_SYMBOL_TYPE_TAG: 1176 case wasm::WASM_SYMBOL_TYPE_TABLE: 1177 encodeULEB128(Sym.ElementIndex, W->OS); 1178 if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0 || 1179 (Sym.Flags & wasm::WASM_SYMBOL_EXPLICIT_NAME) != 0) 1180 writeString(Sym.Name); 1181 break; 1182 case wasm::WASM_SYMBOL_TYPE_DATA: 1183 writeString(Sym.Name); 1184 if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0) { 1185 encodeULEB128(Sym.DataRef.Segment, W->OS); 1186 encodeULEB128(Sym.DataRef.Offset, W->OS); 1187 encodeULEB128(Sym.DataRef.Size, W->OS); 1188 } 1189 break; 1190 case wasm::WASM_SYMBOL_TYPE_SECTION: { 1191 const uint32_t SectionIndex = 1192 CustomSections[Sym.ElementIndex].OutputIndex; 1193 encodeULEB128(SectionIndex, W->OS); 1194 break; 1195 } 1196 default: 1197 llvm_unreachable("unexpected kind"); 1198 } 1199 } 1200 endSection(SubSection); 1201 } 1202 1203 if (DataSegments.size()) { 1204 startSection(SubSection, wasm::WASM_SEGMENT_INFO); 1205 encodeULEB128(DataSegments.size(), W->OS); 1206 for (const WasmDataSegment &Segment : DataSegments) { 1207 writeString(Segment.Name); 1208 encodeULEB128(Segment.Alignment, W->OS); 1209 encodeULEB128(Segment.LinkingFlags, W->OS); 1210 } 1211 endSection(SubSection); 1212 } 1213 1214 if (!InitFuncs.empty()) { 1215 startSection(SubSection, wasm::WASM_INIT_FUNCS); 1216 encodeULEB128(InitFuncs.size(), W->OS); 1217 for (auto &StartFunc : InitFuncs) { 1218 encodeULEB128(StartFunc.first, W->OS); // priority 1219 encodeULEB128(StartFunc.second, W->OS); // function index 1220 } 1221 endSection(SubSection); 1222 } 1223 1224 if (Comdats.size()) { 1225 startSection(SubSection, wasm::WASM_COMDAT_INFO); 1226 encodeULEB128(Comdats.size(), W->OS); 1227 for (const auto &C : Comdats) { 1228 writeString(C.first); 1229 encodeULEB128(0, W->OS); // flags for future use 1230 encodeULEB128(C.second.size(), W->OS); 1231 for (const WasmComdatEntry &Entry : C.second) { 1232 encodeULEB128(Entry.Kind, W->OS); 1233 encodeULEB128(Entry.Index, W->OS); 1234 } 1235 } 1236 endSection(SubSection); 1237 } 1238 1239 endSection(Section); 1240 } 1241 1242 void WasmObjectWriter::writeCustomSection(WasmCustomSection &CustomSection, 1243 const MCAssembler &Asm, 1244 const MCAsmLayout &Layout) { 1245 SectionBookkeeping Section; 1246 auto *Sec = CustomSection.Section; 1247 startCustomSection(Section, CustomSection.Name); 1248 1249 Sec->setSectionOffset(W->OS.tell() - Section.ContentsOffset); 1250 Asm.writeSectionData(W->OS, Sec, Layout); 1251 1252 CustomSection.OutputContentsOffset = Section.ContentsOffset; 1253 CustomSection.OutputIndex = Section.Index; 1254 1255 endSection(Section); 1256 1257 // Apply fixups. 1258 auto &Relocations = CustomSectionsRelocations[CustomSection.Section]; 1259 applyRelocations(Relocations, CustomSection.OutputContentsOffset, Layout); 1260 } 1261 1262 uint32_t WasmObjectWriter::getFunctionType(const MCSymbolWasm &Symbol) { 1263 assert(Symbol.isFunction()); 1264 assert(TypeIndices.count(&Symbol)); 1265 return TypeIndices[&Symbol]; 1266 } 1267 1268 uint32_t WasmObjectWriter::getTagType(const MCSymbolWasm &Symbol) { 1269 assert(Symbol.isTag()); 1270 assert(TypeIndices.count(&Symbol)); 1271 return TypeIndices[&Symbol]; 1272 } 1273 1274 void WasmObjectWriter::registerFunctionType(const MCSymbolWasm &Symbol) { 1275 assert(Symbol.isFunction()); 1276 1277 wasm::WasmSignature S; 1278 1279 if (auto *Sig = Symbol.getSignature()) { 1280 S.Returns = Sig->Returns; 1281 S.Params = Sig->Params; 1282 } 1283 1284 auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size())); 1285 if (Pair.second) 1286 Signatures.push_back(S); 1287 TypeIndices[&Symbol] = Pair.first->second; 1288 1289 LLVM_DEBUG(dbgs() << "registerFunctionType: " << Symbol 1290 << " new:" << Pair.second << "\n"); 1291 LLVM_DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n"); 1292 } 1293 1294 void WasmObjectWriter::registerTagType(const MCSymbolWasm &Symbol) { 1295 assert(Symbol.isTag()); 1296 1297 // TODO Currently we don't generate imported exceptions, but if we do, we 1298 // should have a way of infering types of imported exceptions. 1299 wasm::WasmSignature S; 1300 if (auto *Sig = Symbol.getSignature()) { 1301 S.Returns = Sig->Returns; 1302 S.Params = Sig->Params; 1303 } 1304 1305 auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size())); 1306 if (Pair.second) 1307 Signatures.push_back(S); 1308 TypeIndices[&Symbol] = Pair.first->second; 1309 1310 LLVM_DEBUG(dbgs() << "registerTagType: " << Symbol << " new:" << Pair.second 1311 << "\n"); 1312 LLVM_DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n"); 1313 } 1314 1315 static bool isInSymtab(const MCSymbolWasm &Sym) { 1316 if (Sym.isUsedInReloc() || Sym.isUsedInInitArray()) 1317 return true; 1318 1319 if (Sym.isComdat() && !Sym.isDefined()) 1320 return false; 1321 1322 if (Sym.isTemporary()) 1323 return false; 1324 1325 if (Sym.isSection()) 1326 return false; 1327 1328 if (Sym.omitFromLinkingSection()) 1329 return false; 1330 1331 return true; 1332 } 1333 1334 void WasmObjectWriter::prepareImports( 1335 SmallVectorImpl<wasm::WasmImport> &Imports, MCAssembler &Asm, 1336 const MCAsmLayout &Layout) { 1337 // For now, always emit the memory import, since loads and stores are not 1338 // valid without it. In the future, we could perhaps be more clever and omit 1339 // it if there are no loads or stores. 1340 wasm::WasmImport MemImport; 1341 MemImport.Module = "env"; 1342 MemImport.Field = "__linear_memory"; 1343 MemImport.Kind = wasm::WASM_EXTERNAL_MEMORY; 1344 MemImport.Memory.Flags = is64Bit() ? wasm::WASM_LIMITS_FLAG_IS_64 1345 : wasm::WASM_LIMITS_FLAG_NONE; 1346 Imports.push_back(MemImport); 1347 1348 // Populate SignatureIndices, and Imports and WasmIndices for undefined 1349 // symbols. This must be done before populating WasmIndices for defined 1350 // symbols. 1351 for (const MCSymbol &S : Asm.symbols()) { 1352 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1353 1354 // Register types for all functions, including those with private linkage 1355 // (because wasm always needs a type signature). 1356 if (WS.isFunction()) { 1357 const auto *BS = Layout.getBaseSymbol(S); 1358 if (!BS) 1359 report_fatal_error(Twine(S.getName()) + 1360 ": absolute addressing not supported!"); 1361 registerFunctionType(*cast<MCSymbolWasm>(BS)); 1362 } 1363 1364 if (WS.isTag()) 1365 registerTagType(WS); 1366 1367 if (WS.isTemporary()) 1368 continue; 1369 1370 // If the symbol is not defined in this translation unit, import it. 1371 if (!WS.isDefined() && !WS.isComdat()) { 1372 if (WS.isFunction()) { 1373 wasm::WasmImport Import; 1374 Import.Module = WS.getImportModule(); 1375 Import.Field = WS.getImportName(); 1376 Import.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1377 Import.SigIndex = getFunctionType(WS); 1378 Imports.push_back(Import); 1379 assert(WasmIndices.count(&WS) == 0); 1380 WasmIndices[&WS] = NumFunctionImports++; 1381 } else if (WS.isGlobal()) { 1382 if (WS.isWeak()) 1383 report_fatal_error("undefined global symbol cannot be weak"); 1384 1385 wasm::WasmImport Import; 1386 Import.Field = WS.getImportName(); 1387 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1388 Import.Module = WS.getImportModule(); 1389 Import.Global = WS.getGlobalType(); 1390 Imports.push_back(Import); 1391 assert(WasmIndices.count(&WS) == 0); 1392 WasmIndices[&WS] = NumGlobalImports++; 1393 } else if (WS.isTag()) { 1394 if (WS.isWeak()) 1395 report_fatal_error("undefined tag symbol cannot be weak"); 1396 1397 wasm::WasmImport Import; 1398 Import.Module = WS.getImportModule(); 1399 Import.Field = WS.getImportName(); 1400 Import.Kind = wasm::WASM_EXTERNAL_TAG; 1401 Import.SigIndex = getTagType(WS); 1402 Imports.push_back(Import); 1403 assert(WasmIndices.count(&WS) == 0); 1404 WasmIndices[&WS] = NumTagImports++; 1405 } else if (WS.isTable()) { 1406 if (WS.isWeak()) 1407 report_fatal_error("undefined table symbol cannot be weak"); 1408 1409 wasm::WasmImport Import; 1410 Import.Module = WS.getImportModule(); 1411 Import.Field = WS.getImportName(); 1412 Import.Kind = wasm::WASM_EXTERNAL_TABLE; 1413 Import.Table = WS.getTableType(); 1414 Imports.push_back(Import); 1415 assert(WasmIndices.count(&WS) == 0); 1416 WasmIndices[&WS] = NumTableImports++; 1417 } 1418 } 1419 } 1420 1421 // Add imports for GOT globals 1422 for (const MCSymbol &S : Asm.symbols()) { 1423 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1424 if (WS.isUsedInGOT()) { 1425 wasm::WasmImport Import; 1426 if (WS.isFunction()) 1427 Import.Module = "GOT.func"; 1428 else 1429 Import.Module = "GOT.mem"; 1430 Import.Field = WS.getName(); 1431 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1432 Import.Global = {wasm::WASM_TYPE_I32, true}; 1433 Imports.push_back(Import); 1434 assert(GOTIndices.count(&WS) == 0); 1435 GOTIndices[&WS] = NumGlobalImports++; 1436 } 1437 } 1438 } 1439 1440 uint64_t WasmObjectWriter::writeObject(MCAssembler &Asm, 1441 const MCAsmLayout &Layout) { 1442 support::endian::Writer MainWriter(*OS, support::little); 1443 W = &MainWriter; 1444 if (IsSplitDwarf) { 1445 uint64_t TotalSize = writeOneObject(Asm, Layout, DwoMode::NonDwoOnly); 1446 assert(DwoOS); 1447 support::endian::Writer DwoWriter(*DwoOS, support::little); 1448 W = &DwoWriter; 1449 return TotalSize + writeOneObject(Asm, Layout, DwoMode::DwoOnly); 1450 } else { 1451 return writeOneObject(Asm, Layout, DwoMode::AllSections); 1452 } 1453 } 1454 1455 uint64_t WasmObjectWriter::writeOneObject(MCAssembler &Asm, 1456 const MCAsmLayout &Layout, 1457 DwoMode Mode) { 1458 uint64_t StartOffset = W->OS.tell(); 1459 SectionCount = 0; 1460 CustomSections.clear(); 1461 1462 LLVM_DEBUG(dbgs() << "WasmObjectWriter::writeObject\n"); 1463 1464 // Collect information from the available symbols. 1465 SmallVector<WasmFunction, 4> Functions; 1466 SmallVector<uint32_t, 4> TableElems; 1467 SmallVector<wasm::WasmImport, 4> Imports; 1468 SmallVector<wasm::WasmExport, 4> Exports; 1469 SmallVector<uint32_t, 2> TagTypes; 1470 SmallVector<wasm::WasmGlobal, 1> Globals; 1471 SmallVector<wasm::WasmTable, 1> Tables; 1472 SmallVector<wasm::WasmSymbolInfo, 4> SymbolInfos; 1473 SmallVector<std::pair<uint16_t, uint32_t>, 2> InitFuncs; 1474 std::map<StringRef, std::vector<WasmComdatEntry>> Comdats; 1475 uint64_t DataSize = 0; 1476 if (Mode != DwoMode::DwoOnly) { 1477 prepareImports(Imports, Asm, Layout); 1478 } 1479 1480 // Populate DataSegments and CustomSections, which must be done before 1481 // populating DataLocations. 1482 for (MCSection &Sec : Asm) { 1483 auto &Section = static_cast<MCSectionWasm &>(Sec); 1484 StringRef SectionName = Section.getName(); 1485 1486 if (Mode == DwoMode::NonDwoOnly && isDwoSection(Sec)) 1487 continue; 1488 if (Mode == DwoMode::DwoOnly && !isDwoSection(Sec)) 1489 continue; 1490 1491 LLVM_DEBUG(dbgs() << "Processing Section " << SectionName << " group " 1492 << Section.getGroup() << "\n";); 1493 1494 // .init_array sections are handled specially elsewhere. 1495 if (SectionName.startswith(".init_array")) 1496 continue; 1497 1498 // Code is handled separately 1499 if (Section.getKind().isText()) 1500 continue; 1501 1502 if (Section.isWasmData()) { 1503 uint32_t SegmentIndex = DataSegments.size(); 1504 DataSize = alignTo(DataSize, Section.getAlignment()); 1505 DataSegments.emplace_back(); 1506 WasmDataSegment &Segment = DataSegments.back(); 1507 Segment.Name = SectionName; 1508 Segment.InitFlags = Section.getPassive() 1509 ? (uint32_t)wasm::WASM_DATA_SEGMENT_IS_PASSIVE 1510 : 0; 1511 Segment.Offset = DataSize; 1512 Segment.Section = &Section; 1513 addData(Segment.Data, Section); 1514 Segment.Alignment = Log2_32(Section.getAlignment()); 1515 Segment.LinkingFlags = Section.getSegmentFlags(); 1516 DataSize += Segment.Data.size(); 1517 Section.setSegmentIndex(SegmentIndex); 1518 1519 if (const MCSymbolWasm *C = Section.getGroup()) { 1520 Comdats[C->getName()].emplace_back( 1521 WasmComdatEntry{wasm::WASM_COMDAT_DATA, SegmentIndex}); 1522 } 1523 } else { 1524 // Create custom sections 1525 assert(Sec.getKind().isMetadata()); 1526 1527 StringRef Name = SectionName; 1528 1529 // For user-defined custom sections, strip the prefix 1530 if (Name.startswith(".custom_section.")) 1531 Name = Name.substr(strlen(".custom_section.")); 1532 1533 MCSymbol *Begin = Sec.getBeginSymbol(); 1534 if (Begin) { 1535 assert(WasmIndices.count(cast<MCSymbolWasm>(Begin)) == 0); 1536 WasmIndices[cast<MCSymbolWasm>(Begin)] = CustomSections.size(); 1537 } 1538 1539 // Separate out the producers and target features sections 1540 if (Name == "producers") { 1541 ProducersSection = std::make_unique<WasmCustomSection>(Name, &Section); 1542 continue; 1543 } 1544 if (Name == "target_features") { 1545 TargetFeaturesSection = 1546 std::make_unique<WasmCustomSection>(Name, &Section); 1547 continue; 1548 } 1549 1550 // Custom sections can also belong to COMDAT groups. In this case the 1551 // decriptor's "index" field is the section index (in the final object 1552 // file), but that is not known until after layout, so it must be fixed up 1553 // later 1554 if (const MCSymbolWasm *C = Section.getGroup()) { 1555 Comdats[C->getName()].emplace_back( 1556 WasmComdatEntry{wasm::WASM_COMDAT_SECTION, 1557 static_cast<uint32_t>(CustomSections.size())}); 1558 } 1559 1560 CustomSections.emplace_back(Name, &Section); 1561 } 1562 } 1563 1564 if (Mode != DwoMode::DwoOnly) { 1565 // Populate WasmIndices and DataLocations for defined symbols. 1566 for (const MCSymbol &S : Asm.symbols()) { 1567 // Ignore unnamed temporary symbols, which aren't ever exported, imported, 1568 // or used in relocations. 1569 if (S.isTemporary() && S.getName().empty()) 1570 continue; 1571 1572 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1573 LLVM_DEBUG( 1574 dbgs() << "MCSymbol: " 1575 << toString(WS.getType().value_or(wasm::WASM_SYMBOL_TYPE_DATA)) 1576 << " '" << S << "'" 1577 << " isDefined=" << S.isDefined() << " isExternal=" 1578 << S.isExternal() << " isTemporary=" << S.isTemporary() 1579 << " isWeak=" << WS.isWeak() << " isHidden=" << WS.isHidden() 1580 << " isVariable=" << WS.isVariable() << "\n"); 1581 1582 if (WS.isVariable()) 1583 continue; 1584 if (WS.isComdat() && !WS.isDefined()) 1585 continue; 1586 1587 if (WS.isFunction()) { 1588 unsigned Index; 1589 if (WS.isDefined()) { 1590 if (WS.getOffset() != 0) 1591 report_fatal_error( 1592 "function sections must contain one function each"); 1593 1594 if (WS.getSize() == nullptr) 1595 report_fatal_error( 1596 "function symbols must have a size set with .size"); 1597 1598 // A definition. Write out the function body. 1599 Index = NumFunctionImports + Functions.size(); 1600 WasmFunction Func; 1601 Func.SigIndex = getFunctionType(WS); 1602 Func.Sym = &WS; 1603 assert(WasmIndices.count(&WS) == 0); 1604 WasmIndices[&WS] = Index; 1605 Functions.push_back(Func); 1606 1607 auto &Section = static_cast<MCSectionWasm &>(WS.getSection()); 1608 if (const MCSymbolWasm *C = Section.getGroup()) { 1609 Comdats[C->getName()].emplace_back( 1610 WasmComdatEntry{wasm::WASM_COMDAT_FUNCTION, Index}); 1611 } 1612 1613 if (WS.hasExportName()) { 1614 wasm::WasmExport Export; 1615 Export.Name = WS.getExportName(); 1616 Export.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1617 Export.Index = Index; 1618 Exports.push_back(Export); 1619 } 1620 } else { 1621 // An import; the index was assigned above. 1622 Index = WasmIndices.find(&WS)->second; 1623 } 1624 1625 LLVM_DEBUG(dbgs() << " -> function index: " << Index << "\n"); 1626 1627 } else if (WS.isData()) { 1628 if (!isInSymtab(WS)) 1629 continue; 1630 1631 if (!WS.isDefined()) { 1632 LLVM_DEBUG(dbgs() << " -> segment index: -1" 1633 << "\n"); 1634 continue; 1635 } 1636 1637 if (!WS.getSize()) 1638 report_fatal_error("data symbols must have a size set with .size: " + 1639 WS.getName()); 1640 1641 int64_t Size = 0; 1642 if (!WS.getSize()->evaluateAsAbsolute(Size, Layout)) 1643 report_fatal_error(".size expression must be evaluatable"); 1644 1645 auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection()); 1646 if (!DataSection.isWasmData()) 1647 report_fatal_error("data symbols must live in a data section: " + 1648 WS.getName()); 1649 1650 // For each data symbol, export it in the symtab as a reference to the 1651 // corresponding Wasm data segment. 1652 wasm::WasmDataReference Ref = wasm::WasmDataReference{ 1653 DataSection.getSegmentIndex(), Layout.getSymbolOffset(WS), 1654 static_cast<uint64_t>(Size)}; 1655 assert(DataLocations.count(&WS) == 0); 1656 DataLocations[&WS] = Ref; 1657 LLVM_DEBUG(dbgs() << " -> segment index: " << Ref.Segment << "\n"); 1658 1659 } else if (WS.isGlobal()) { 1660 // A "true" Wasm global (currently just __stack_pointer) 1661 if (WS.isDefined()) { 1662 wasm::WasmGlobal Global; 1663 Global.Type = WS.getGlobalType(); 1664 Global.Index = NumGlobalImports + Globals.size(); 1665 Global.InitExpr.Extended = false; 1666 switch (Global.Type.Type) { 1667 case wasm::WASM_TYPE_I32: 1668 Global.InitExpr.Inst.Opcode = wasm::WASM_OPCODE_I32_CONST; 1669 break; 1670 case wasm::WASM_TYPE_I64: 1671 Global.InitExpr.Inst.Opcode = wasm::WASM_OPCODE_I64_CONST; 1672 break; 1673 case wasm::WASM_TYPE_F32: 1674 Global.InitExpr.Inst.Opcode = wasm::WASM_OPCODE_F32_CONST; 1675 break; 1676 case wasm::WASM_TYPE_F64: 1677 Global.InitExpr.Inst.Opcode = wasm::WASM_OPCODE_F64_CONST; 1678 break; 1679 case wasm::WASM_TYPE_EXTERNREF: 1680 Global.InitExpr.Inst.Opcode = wasm::WASM_OPCODE_REF_NULL; 1681 break; 1682 default: 1683 llvm_unreachable("unexpected type"); 1684 } 1685 assert(WasmIndices.count(&WS) == 0); 1686 WasmIndices[&WS] = Global.Index; 1687 Globals.push_back(Global); 1688 } else { 1689 // An import; the index was assigned above 1690 LLVM_DEBUG(dbgs() << " -> global index: " 1691 << WasmIndices.find(&WS)->second << "\n"); 1692 } 1693 } else if (WS.isTable()) { 1694 if (WS.isDefined()) { 1695 wasm::WasmTable Table; 1696 Table.Index = NumTableImports + Tables.size(); 1697 Table.Type = WS.getTableType(); 1698 assert(WasmIndices.count(&WS) == 0); 1699 WasmIndices[&WS] = Table.Index; 1700 Tables.push_back(Table); 1701 } 1702 LLVM_DEBUG(dbgs() << " -> table index: " 1703 << WasmIndices.find(&WS)->second << "\n"); 1704 } else if (WS.isTag()) { 1705 // C++ exception symbol (__cpp_exception) or longjmp symbol 1706 // (__c_longjmp) 1707 unsigned Index; 1708 if (WS.isDefined()) { 1709 Index = NumTagImports + TagTypes.size(); 1710 uint32_t SigIndex = getTagType(WS); 1711 assert(WasmIndices.count(&WS) == 0); 1712 WasmIndices[&WS] = Index; 1713 TagTypes.push_back(SigIndex); 1714 } else { 1715 // An import; the index was assigned above. 1716 assert(WasmIndices.count(&WS) > 0); 1717 } 1718 LLVM_DEBUG(dbgs() << " -> tag index: " << WasmIndices.find(&WS)->second 1719 << "\n"); 1720 1721 } else { 1722 assert(WS.isSection()); 1723 } 1724 } 1725 1726 // Populate WasmIndices and DataLocations for aliased symbols. We need to 1727 // process these in a separate pass because we need to have processed the 1728 // target of the alias before the alias itself and the symbols are not 1729 // necessarily ordered in this way. 1730 for (const MCSymbol &S : Asm.symbols()) { 1731 if (!S.isVariable()) 1732 continue; 1733 1734 assert(S.isDefined()); 1735 1736 const auto *BS = Layout.getBaseSymbol(S); 1737 if (!BS) 1738 report_fatal_error(Twine(S.getName()) + 1739 ": absolute addressing not supported!"); 1740 const MCSymbolWasm *Base = cast<MCSymbolWasm>(BS); 1741 1742 // Find the target symbol of this weak alias and export that index 1743 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1744 LLVM_DEBUG(dbgs() << WS.getName() << ": weak alias of '" << *Base 1745 << "'\n"); 1746 1747 if (Base->isFunction()) { 1748 assert(WasmIndices.count(Base) > 0); 1749 uint32_t WasmIndex = WasmIndices.find(Base)->second; 1750 assert(WasmIndices.count(&WS) == 0); 1751 WasmIndices[&WS] = WasmIndex; 1752 LLVM_DEBUG(dbgs() << " -> index:" << WasmIndex << "\n"); 1753 } else if (Base->isData()) { 1754 auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection()); 1755 uint64_t Offset = Layout.getSymbolOffset(S); 1756 int64_t Size = 0; 1757 // For data symbol alias we use the size of the base symbol as the 1758 // size of the alias. When an offset from the base is involved this 1759 // can result in a offset + size goes past the end of the data section 1760 // which out object format doesn't support. So we must clamp it. 1761 if (!Base->getSize()->evaluateAsAbsolute(Size, Layout)) 1762 report_fatal_error(".size expression must be evaluatable"); 1763 const WasmDataSegment &Segment = 1764 DataSegments[DataSection.getSegmentIndex()]; 1765 Size = 1766 std::min(static_cast<uint64_t>(Size), Segment.Data.size() - Offset); 1767 wasm::WasmDataReference Ref = wasm::WasmDataReference{ 1768 DataSection.getSegmentIndex(), 1769 static_cast<uint32_t>(Layout.getSymbolOffset(S)), 1770 static_cast<uint32_t>(Size)}; 1771 DataLocations[&WS] = Ref; 1772 LLVM_DEBUG(dbgs() << " -> index:" << Ref.Segment << "\n"); 1773 } else { 1774 report_fatal_error("don't yet support global/tag aliases"); 1775 } 1776 } 1777 } 1778 1779 // Finally, populate the symbol table itself, in its "natural" order. 1780 for (const MCSymbol &S : Asm.symbols()) { 1781 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1782 if (!isInSymtab(WS)) { 1783 WS.setIndex(InvalidIndex); 1784 continue; 1785 } 1786 LLVM_DEBUG(dbgs() << "adding to symtab: " << WS << "\n"); 1787 1788 uint32_t Flags = 0; 1789 if (WS.isWeak()) 1790 Flags |= wasm::WASM_SYMBOL_BINDING_WEAK; 1791 if (WS.isHidden()) 1792 Flags |= wasm::WASM_SYMBOL_VISIBILITY_HIDDEN; 1793 if (!WS.isExternal() && WS.isDefined()) 1794 Flags |= wasm::WASM_SYMBOL_BINDING_LOCAL; 1795 if (WS.isUndefined()) 1796 Flags |= wasm::WASM_SYMBOL_UNDEFINED; 1797 if (WS.isNoStrip()) { 1798 Flags |= wasm::WASM_SYMBOL_NO_STRIP; 1799 if (isEmscripten()) { 1800 Flags |= wasm::WASM_SYMBOL_EXPORTED; 1801 } 1802 } 1803 if (WS.hasImportName()) 1804 Flags |= wasm::WASM_SYMBOL_EXPLICIT_NAME; 1805 if (WS.hasExportName()) 1806 Flags |= wasm::WASM_SYMBOL_EXPORTED; 1807 if (WS.isTLS()) 1808 Flags |= wasm::WASM_SYMBOL_TLS; 1809 1810 wasm::WasmSymbolInfo Info; 1811 Info.Name = WS.getName(); 1812 Info.Kind = WS.getType().value_or(wasm::WASM_SYMBOL_TYPE_DATA); 1813 Info.Flags = Flags; 1814 if (!WS.isData()) { 1815 assert(WasmIndices.count(&WS) > 0); 1816 Info.ElementIndex = WasmIndices.find(&WS)->second; 1817 } else if (WS.isDefined()) { 1818 assert(DataLocations.count(&WS) > 0); 1819 Info.DataRef = DataLocations.find(&WS)->second; 1820 } 1821 WS.setIndex(SymbolInfos.size()); 1822 SymbolInfos.emplace_back(Info); 1823 } 1824 1825 { 1826 auto HandleReloc = [&](const WasmRelocationEntry &Rel) { 1827 // Functions referenced by a relocation need to put in the table. This is 1828 // purely to make the object file's provisional values readable, and is 1829 // ignored by the linker, which re-calculates the relocations itself. 1830 if (Rel.Type != wasm::R_WASM_TABLE_INDEX_I32 && 1831 Rel.Type != wasm::R_WASM_TABLE_INDEX_I64 && 1832 Rel.Type != wasm::R_WASM_TABLE_INDEX_SLEB && 1833 Rel.Type != wasm::R_WASM_TABLE_INDEX_SLEB64 && 1834 Rel.Type != wasm::R_WASM_TABLE_INDEX_REL_SLEB && 1835 Rel.Type != wasm::R_WASM_TABLE_INDEX_REL_SLEB64) 1836 return; 1837 assert(Rel.Symbol->isFunction()); 1838 const MCSymbolWasm *Base = 1839 cast<MCSymbolWasm>(Layout.getBaseSymbol(*Rel.Symbol)); 1840 uint32_t FunctionIndex = WasmIndices.find(Base)->second; 1841 uint32_t TableIndex = TableElems.size() + InitialTableOffset; 1842 if (TableIndices.try_emplace(Base, TableIndex).second) { 1843 LLVM_DEBUG(dbgs() << " -> adding " << Base->getName() 1844 << " to table: " << TableIndex << "\n"); 1845 TableElems.push_back(FunctionIndex); 1846 registerFunctionType(*Base); 1847 } 1848 }; 1849 1850 for (const WasmRelocationEntry &RelEntry : CodeRelocations) 1851 HandleReloc(RelEntry); 1852 for (const WasmRelocationEntry &RelEntry : DataRelocations) 1853 HandleReloc(RelEntry); 1854 } 1855 1856 // Translate .init_array section contents into start functions. 1857 for (const MCSection &S : Asm) { 1858 const auto &WS = static_cast<const MCSectionWasm &>(S); 1859 if (WS.getName().startswith(".fini_array")) 1860 report_fatal_error(".fini_array sections are unsupported"); 1861 if (!WS.getName().startswith(".init_array")) 1862 continue; 1863 if (WS.getFragmentList().empty()) 1864 continue; 1865 1866 // init_array is expected to contain a single non-empty data fragment 1867 if (WS.getFragmentList().size() != 3) 1868 report_fatal_error("only one .init_array section fragment supported"); 1869 1870 auto IT = WS.begin(); 1871 const MCFragment &EmptyFrag = *IT; 1872 if (EmptyFrag.getKind() != MCFragment::FT_Data) 1873 report_fatal_error(".init_array section should be aligned"); 1874 1875 IT = std::next(IT); 1876 const MCFragment &AlignFrag = *IT; 1877 if (AlignFrag.getKind() != MCFragment::FT_Align) 1878 report_fatal_error(".init_array section should be aligned"); 1879 if (cast<MCAlignFragment>(AlignFrag).getAlignment() != 1880 Align(is64Bit() ? 8 : 4)) 1881 report_fatal_error(".init_array section should be aligned for pointers"); 1882 1883 const MCFragment &Frag = *std::next(IT); 1884 if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data) 1885 report_fatal_error("only data supported in .init_array section"); 1886 1887 uint16_t Priority = UINT16_MAX; 1888 unsigned PrefixLength = strlen(".init_array"); 1889 if (WS.getName().size() > PrefixLength) { 1890 if (WS.getName()[PrefixLength] != '.') 1891 report_fatal_error( 1892 ".init_array section priority should start with '.'"); 1893 if (WS.getName().substr(PrefixLength + 1).getAsInteger(10, Priority)) 1894 report_fatal_error("invalid .init_array section priority"); 1895 } 1896 const auto &DataFrag = cast<MCDataFragment>(Frag); 1897 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 1898 for (const uint8_t * 1899 P = (const uint8_t *)Contents.data(), 1900 *End = (const uint8_t *)Contents.data() + Contents.size(); 1901 P != End; ++P) { 1902 if (*P != 0) 1903 report_fatal_error("non-symbolic data in .init_array section"); 1904 } 1905 for (const MCFixup &Fixup : DataFrag.getFixups()) { 1906 assert(Fixup.getKind() == 1907 MCFixup::getKindForSize(is64Bit() ? 8 : 4, false)); 1908 const MCExpr *Expr = Fixup.getValue(); 1909 auto *SymRef = dyn_cast<MCSymbolRefExpr>(Expr); 1910 if (!SymRef) 1911 report_fatal_error("fixups in .init_array should be symbol references"); 1912 const auto &TargetSym = cast<const MCSymbolWasm>(SymRef->getSymbol()); 1913 if (TargetSym.getIndex() == InvalidIndex) 1914 report_fatal_error("symbols in .init_array should exist in symtab"); 1915 if (!TargetSym.isFunction()) 1916 report_fatal_error("symbols in .init_array should be for functions"); 1917 InitFuncs.push_back( 1918 std::make_pair(Priority, TargetSym.getIndex())); 1919 } 1920 } 1921 1922 // Write out the Wasm header. 1923 writeHeader(Asm); 1924 1925 uint32_t CodeSectionIndex, DataSectionIndex; 1926 if (Mode != DwoMode::DwoOnly) { 1927 writeTypeSection(Signatures); 1928 writeImportSection(Imports, DataSize, TableElems.size()); 1929 writeFunctionSection(Functions); 1930 writeTableSection(Tables); 1931 // Skip the "memory" section; we import the memory instead. 1932 writeTagSection(TagTypes); 1933 writeGlobalSection(Globals); 1934 writeExportSection(Exports); 1935 const MCSymbol *IndirectFunctionTable = 1936 Asm.getContext().lookupSymbol("__indirect_function_table"); 1937 writeElemSection(cast_or_null<const MCSymbolWasm>(IndirectFunctionTable), 1938 TableElems); 1939 writeDataCountSection(); 1940 1941 CodeSectionIndex = writeCodeSection(Asm, Layout, Functions); 1942 DataSectionIndex = writeDataSection(Layout); 1943 } 1944 1945 // The Sections in the COMDAT list have placeholder indices (their index among 1946 // custom sections, rather than among all sections). Fix them up here. 1947 for (auto &Group : Comdats) { 1948 for (auto &Entry : Group.second) { 1949 if (Entry.Kind == wasm::WASM_COMDAT_SECTION) { 1950 Entry.Index += SectionCount; 1951 } 1952 } 1953 } 1954 for (auto &CustomSection : CustomSections) 1955 writeCustomSection(CustomSection, Asm, Layout); 1956 1957 if (Mode != DwoMode::DwoOnly) { 1958 writeLinkingMetaDataSection(SymbolInfos, InitFuncs, Comdats); 1959 1960 writeRelocSection(CodeSectionIndex, "CODE", CodeRelocations); 1961 writeRelocSection(DataSectionIndex, "DATA", DataRelocations); 1962 } 1963 writeCustomRelocSections(); 1964 if (ProducersSection) 1965 writeCustomSection(*ProducersSection, Asm, Layout); 1966 if (TargetFeaturesSection) 1967 writeCustomSection(*TargetFeaturesSection, Asm, Layout); 1968 1969 // TODO: Translate the .comment section to the output. 1970 return W->OS.tell() - StartOffset; 1971 } 1972 1973 std::unique_ptr<MCObjectWriter> 1974 llvm::createWasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 1975 raw_pwrite_stream &OS) { 1976 return std::make_unique<WasmObjectWriter>(std::move(MOTW), OS); 1977 } 1978 1979 std::unique_ptr<MCObjectWriter> 1980 llvm::createWasmDwoObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 1981 raw_pwrite_stream &OS, 1982 raw_pwrite_stream &DwoOS) { 1983 return std::make_unique<WasmObjectWriter>(std::move(MOTW), OS, DwoOS); 1984 } 1985