1 //===- lib/MC/WasmObjectWriter.cpp - Wasm File Writer ---------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements Wasm object file writer information. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "llvm/ADT/STLExtras.h" 14 #include "llvm/ADT/SmallPtrSet.h" 15 #include "llvm/BinaryFormat/Wasm.h" 16 #include "llvm/BinaryFormat/WasmTraits.h" 17 #include "llvm/Config/llvm-config.h" 18 #include "llvm/MC/MCAsmBackend.h" 19 #include "llvm/MC/MCAsmLayout.h" 20 #include "llvm/MC/MCAssembler.h" 21 #include "llvm/MC/MCContext.h" 22 #include "llvm/MC/MCExpr.h" 23 #include "llvm/MC/MCFixupKindInfo.h" 24 #include "llvm/MC/MCObjectWriter.h" 25 #include "llvm/MC/MCSectionWasm.h" 26 #include "llvm/MC/MCSymbolWasm.h" 27 #include "llvm/MC/MCValue.h" 28 #include "llvm/MC/MCWasmObjectWriter.h" 29 #include "llvm/Support/Casting.h" 30 #include "llvm/Support/Debug.h" 31 #include "llvm/Support/EndianStream.h" 32 #include "llvm/Support/ErrorHandling.h" 33 #include "llvm/Support/LEB128.h" 34 #include "llvm/Support/StringSaver.h" 35 #include <vector> 36 37 using namespace llvm; 38 39 #define DEBUG_TYPE "mc" 40 41 namespace { 42 43 // When we create the indirect function table we start at 1, so that there is 44 // and empty slot at 0 and therefore calling a null function pointer will trap. 45 static const uint32_t InitialTableOffset = 1; 46 47 // For patching purposes, we need to remember where each section starts, both 48 // for patching up the section size field, and for patching up references to 49 // locations within the section. 50 struct SectionBookkeeping { 51 // Where the size of the section is written. 52 uint64_t SizeOffset; 53 // Where the section header ends (without custom section name). 54 uint64_t PayloadOffset; 55 // Where the contents of the section starts. 56 uint64_t ContentsOffset; 57 uint32_t Index; 58 }; 59 60 // A wasm data segment. A wasm binary contains only a single data section 61 // but that can contain many segments, each with their own virtual location 62 // in memory. Each MCSection data created by llvm is modeled as its own 63 // wasm data segment. 64 struct WasmDataSegment { 65 MCSectionWasm *Section; 66 StringRef Name; 67 uint32_t InitFlags; 68 uint64_t Offset; 69 uint32_t Alignment; 70 uint32_t LinkingFlags; 71 SmallVector<char, 4> Data; 72 }; 73 74 // A wasm function to be written into the function section. 75 struct WasmFunction { 76 uint32_t SigIndex; 77 const MCSymbolWasm *Sym; 78 }; 79 80 // A wasm global to be written into the global section. 81 struct WasmGlobal { 82 wasm::WasmGlobalType Type; 83 uint64_t InitialValue; 84 }; 85 86 // Information about a single item which is part of a COMDAT. For each data 87 // segment or function which is in the COMDAT, there is a corresponding 88 // WasmComdatEntry. 89 struct WasmComdatEntry { 90 unsigned Kind; 91 uint32_t Index; 92 }; 93 94 // Information about a single relocation. 95 struct WasmRelocationEntry { 96 uint64_t Offset; // Where is the relocation. 97 const MCSymbolWasm *Symbol; // The symbol to relocate with. 98 int64_t Addend; // A value to add to the symbol. 99 unsigned Type; // The type of the relocation. 100 const MCSectionWasm *FixupSection; // The section the relocation is targeting. 101 102 WasmRelocationEntry(uint64_t Offset, const MCSymbolWasm *Symbol, 103 int64_t Addend, unsigned Type, 104 const MCSectionWasm *FixupSection) 105 : Offset(Offset), Symbol(Symbol), Addend(Addend), Type(Type), 106 FixupSection(FixupSection) {} 107 108 bool hasAddend() const { return wasm::relocTypeHasAddend(Type); } 109 110 void print(raw_ostream &Out) const { 111 Out << wasm::relocTypetoString(Type) << " Off=" << Offset 112 << ", Sym=" << *Symbol << ", Addend=" << Addend 113 << ", FixupSection=" << FixupSection->getName(); 114 } 115 116 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 117 LLVM_DUMP_METHOD void dump() const { print(dbgs()); } 118 #endif 119 }; 120 121 static const uint32_t InvalidIndex = -1; 122 123 struct WasmCustomSection { 124 125 StringRef Name; 126 MCSectionWasm *Section; 127 128 uint32_t OutputContentsOffset; 129 uint32_t OutputIndex; 130 131 WasmCustomSection(StringRef Name, MCSectionWasm *Section) 132 : Name(Name), Section(Section), OutputContentsOffset(0), 133 OutputIndex(InvalidIndex) {} 134 }; 135 136 #if !defined(NDEBUG) 137 raw_ostream &operator<<(raw_ostream &OS, const WasmRelocationEntry &Rel) { 138 Rel.print(OS); 139 return OS; 140 } 141 #endif 142 143 // Write Value as an (unsigned) LEB value at offset Offset in Stream, padded 144 // to allow patching. 145 template <typename T, int W> 146 void writePatchableULEB(raw_pwrite_stream &Stream, T Value, uint64_t Offset) { 147 uint8_t Buffer[W]; 148 unsigned SizeLen = encodeULEB128(Value, Buffer, W); 149 assert(SizeLen == W); 150 Stream.pwrite((char *)Buffer, SizeLen, Offset); 151 } 152 153 // Write Value as an signed LEB value at offset Offset in Stream, padded 154 // to allow patching. 155 template <typename T, int W> 156 void writePatchableSLEB(raw_pwrite_stream &Stream, T Value, uint64_t Offset) { 157 uint8_t Buffer[W]; 158 unsigned SizeLen = encodeSLEB128(Value, Buffer, W); 159 assert(SizeLen == W); 160 Stream.pwrite((char *)Buffer, SizeLen, Offset); 161 } 162 163 static void writePatchableU32(raw_pwrite_stream &Stream, uint32_t Value, 164 uint64_t Offset) { 165 writePatchableULEB<uint32_t, 5>(Stream, Value, Offset); 166 } 167 168 static void writePatchableS32(raw_pwrite_stream &Stream, int32_t Value, 169 uint64_t Offset) { 170 writePatchableSLEB<int32_t, 5>(Stream, Value, Offset); 171 } 172 173 static void writePatchableU64(raw_pwrite_stream &Stream, uint64_t Value, 174 uint64_t Offset) { 175 writePatchableSLEB<uint64_t, 10>(Stream, Value, Offset); 176 } 177 178 static void writePatchableS64(raw_pwrite_stream &Stream, int64_t Value, 179 uint64_t Offset) { 180 writePatchableSLEB<int64_t, 10>(Stream, Value, Offset); 181 } 182 183 // Write Value as a plain integer value at offset Offset in Stream. 184 static void patchI32(raw_pwrite_stream &Stream, uint32_t Value, 185 uint64_t Offset) { 186 uint8_t Buffer[4]; 187 support::endian::write32le(Buffer, Value); 188 Stream.pwrite((char *)Buffer, sizeof(Buffer), Offset); 189 } 190 191 static void patchI64(raw_pwrite_stream &Stream, uint64_t Value, 192 uint64_t Offset) { 193 uint8_t Buffer[8]; 194 support::endian::write64le(Buffer, Value); 195 Stream.pwrite((char *)Buffer, sizeof(Buffer), Offset); 196 } 197 198 bool isDwoSection(const MCSection &Sec) { 199 return Sec.getName().endswith(".dwo"); 200 } 201 202 class WasmObjectWriter : public MCObjectWriter { 203 support::endian::Writer *W; 204 205 /// The target specific Wasm writer instance. 206 std::unique_ptr<MCWasmObjectTargetWriter> TargetObjectWriter; 207 208 // Relocations for fixing up references in the code section. 209 std::vector<WasmRelocationEntry> CodeRelocations; 210 // Relocations for fixing up references in the data section. 211 std::vector<WasmRelocationEntry> DataRelocations; 212 213 // Index values to use for fixing up call_indirect type indices. 214 // Maps function symbols to the index of the type of the function 215 DenseMap<const MCSymbolWasm *, uint32_t> TypeIndices; 216 // Maps function symbols to the table element index space. Used 217 // for TABLE_INDEX relocation types (i.e. address taken functions). 218 DenseMap<const MCSymbolWasm *, uint32_t> TableIndices; 219 // Maps function/global/table symbols to the 220 // function/global/table/tag/section index space. 221 DenseMap<const MCSymbolWasm *, uint32_t> WasmIndices; 222 DenseMap<const MCSymbolWasm *, uint32_t> GOTIndices; 223 // Maps data symbols to the Wasm segment and offset/size with the segment. 224 DenseMap<const MCSymbolWasm *, wasm::WasmDataReference> DataLocations; 225 226 // Stores output data (index, relocations, content offset) for custom 227 // section. 228 std::vector<WasmCustomSection> CustomSections; 229 std::unique_ptr<WasmCustomSection> ProducersSection; 230 std::unique_ptr<WasmCustomSection> TargetFeaturesSection; 231 // Relocations for fixing up references in the custom sections. 232 DenseMap<const MCSectionWasm *, std::vector<WasmRelocationEntry>> 233 CustomSectionsRelocations; 234 235 // Map from section to defining function symbol. 236 DenseMap<const MCSection *, const MCSymbol *> SectionFunctions; 237 238 DenseMap<wasm::WasmSignature, uint32_t> SignatureIndices; 239 SmallVector<wasm::WasmSignature, 4> Signatures; 240 SmallVector<WasmDataSegment, 4> DataSegments; 241 unsigned NumFunctionImports = 0; 242 unsigned NumGlobalImports = 0; 243 unsigned NumTableImports = 0; 244 unsigned NumTagImports = 0; 245 uint32_t SectionCount = 0; 246 247 enum class DwoMode { 248 AllSections, 249 NonDwoOnly, 250 DwoOnly, 251 }; 252 bool IsSplitDwarf = false; 253 raw_pwrite_stream *OS = nullptr; 254 raw_pwrite_stream *DwoOS = nullptr; 255 256 // TargetObjectWriter wranppers. 257 bool is64Bit() const { return TargetObjectWriter->is64Bit(); } 258 bool isEmscripten() const { return TargetObjectWriter->isEmscripten(); } 259 260 void startSection(SectionBookkeeping &Section, unsigned SectionId); 261 void startCustomSection(SectionBookkeeping &Section, StringRef Name); 262 void endSection(SectionBookkeeping &Section); 263 264 public: 265 WasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 266 raw_pwrite_stream &OS_) 267 : TargetObjectWriter(std::move(MOTW)), OS(&OS_) {} 268 269 WasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 270 raw_pwrite_stream &OS_, raw_pwrite_stream &DwoOS_) 271 : TargetObjectWriter(std::move(MOTW)), IsSplitDwarf(true), OS(&OS_), 272 DwoOS(&DwoOS_) {} 273 274 private: 275 void reset() override { 276 CodeRelocations.clear(); 277 DataRelocations.clear(); 278 TypeIndices.clear(); 279 WasmIndices.clear(); 280 GOTIndices.clear(); 281 TableIndices.clear(); 282 DataLocations.clear(); 283 CustomSections.clear(); 284 ProducersSection.reset(); 285 TargetFeaturesSection.reset(); 286 CustomSectionsRelocations.clear(); 287 SignatureIndices.clear(); 288 Signatures.clear(); 289 DataSegments.clear(); 290 SectionFunctions.clear(); 291 NumFunctionImports = 0; 292 NumGlobalImports = 0; 293 NumTableImports = 0; 294 MCObjectWriter::reset(); 295 } 296 297 void writeHeader(const MCAssembler &Asm); 298 299 void recordRelocation(MCAssembler &Asm, const MCAsmLayout &Layout, 300 const MCFragment *Fragment, const MCFixup &Fixup, 301 MCValue Target, uint64_t &FixedValue) override; 302 303 void executePostLayoutBinding(MCAssembler &Asm, 304 const MCAsmLayout &Layout) override; 305 void prepareImports(SmallVectorImpl<wasm::WasmImport> &Imports, 306 MCAssembler &Asm, const MCAsmLayout &Layout); 307 uint64_t writeObject(MCAssembler &Asm, const MCAsmLayout &Layout) override; 308 309 uint64_t writeOneObject(MCAssembler &Asm, const MCAsmLayout &Layout, 310 DwoMode Mode); 311 312 void writeString(const StringRef Str) { 313 encodeULEB128(Str.size(), W->OS); 314 W->OS << Str; 315 } 316 317 void writeStringWithAlignment(const StringRef Str, unsigned Alignment); 318 319 void writeI32(int32_t val) { 320 char Buffer[4]; 321 support::endian::write32le(Buffer, val); 322 W->OS.write(Buffer, sizeof(Buffer)); 323 } 324 325 void writeI64(int64_t val) { 326 char Buffer[8]; 327 support::endian::write64le(Buffer, val); 328 W->OS.write(Buffer, sizeof(Buffer)); 329 } 330 331 void writeValueType(wasm::ValType Ty) { W->OS << static_cast<char>(Ty); } 332 333 void writeTypeSection(ArrayRef<wasm::WasmSignature> Signatures); 334 void writeImportSection(ArrayRef<wasm::WasmImport> Imports, uint64_t DataSize, 335 uint32_t NumElements); 336 void writeFunctionSection(ArrayRef<WasmFunction> Functions); 337 void writeExportSection(ArrayRef<wasm::WasmExport> Exports); 338 void writeElemSection(const MCSymbolWasm *IndirectFunctionTable, 339 ArrayRef<uint32_t> TableElems); 340 void writeDataCountSection(); 341 uint32_t writeCodeSection(const MCAssembler &Asm, const MCAsmLayout &Layout, 342 ArrayRef<WasmFunction> Functions); 343 uint32_t writeDataSection(const MCAsmLayout &Layout); 344 void writeTagSection(ArrayRef<uint32_t> TagTypes); 345 void writeGlobalSection(ArrayRef<wasm::WasmGlobal> Globals); 346 void writeTableSection(ArrayRef<wasm::WasmTable> Tables); 347 void writeRelocSection(uint32_t SectionIndex, StringRef Name, 348 std::vector<WasmRelocationEntry> &Relocations); 349 void writeLinkingMetaDataSection( 350 ArrayRef<wasm::WasmSymbolInfo> SymbolInfos, 351 ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs, 352 const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats); 353 void writeCustomSection(WasmCustomSection &CustomSection, 354 const MCAssembler &Asm, const MCAsmLayout &Layout); 355 void writeCustomRelocSections(); 356 357 uint64_t getProvisionalValue(const WasmRelocationEntry &RelEntry, 358 const MCAsmLayout &Layout); 359 void applyRelocations(ArrayRef<WasmRelocationEntry> Relocations, 360 uint64_t ContentsOffset, const MCAsmLayout &Layout); 361 362 uint32_t getRelocationIndexValue(const WasmRelocationEntry &RelEntry); 363 uint32_t getFunctionType(const MCSymbolWasm &Symbol); 364 uint32_t getTagType(const MCSymbolWasm &Symbol); 365 void registerFunctionType(const MCSymbolWasm &Symbol); 366 void registerTagType(const MCSymbolWasm &Symbol); 367 }; 368 369 } // end anonymous namespace 370 371 // Write out a section header and a patchable section size field. 372 void WasmObjectWriter::startSection(SectionBookkeeping &Section, 373 unsigned SectionId) { 374 LLVM_DEBUG(dbgs() << "startSection " << SectionId << "\n"); 375 W->OS << char(SectionId); 376 377 Section.SizeOffset = W->OS.tell(); 378 379 // The section size. We don't know the size yet, so reserve enough space 380 // for any 32-bit value; we'll patch it later. 381 encodeULEB128(0, W->OS, 5); 382 383 // The position where the section starts, for measuring its size. 384 Section.ContentsOffset = W->OS.tell(); 385 Section.PayloadOffset = W->OS.tell(); 386 Section.Index = SectionCount++; 387 } 388 389 // Write a string with extra paddings for trailing alignment 390 // TODO: support alignment at asm and llvm level? 391 void WasmObjectWriter::writeStringWithAlignment(const StringRef Str, 392 unsigned Alignment) { 393 394 // Calculate the encoded size of str length and add pads based on it and 395 // alignment. 396 raw_null_ostream NullOS; 397 uint64_t StrSizeLength = encodeULEB128(Str.size(), NullOS); 398 uint64_t Offset = W->OS.tell() + StrSizeLength + Str.size(); 399 uint64_t Paddings = offsetToAlignment(Offset, Align(Alignment)); 400 Offset += Paddings; 401 402 // LEB128 greater than 5 bytes is invalid 403 assert((StrSizeLength + Paddings) <= 5 && "too long string to align"); 404 405 encodeSLEB128(Str.size(), W->OS, StrSizeLength + Paddings); 406 W->OS << Str; 407 408 assert(W->OS.tell() == Offset && "invalid padding"); 409 } 410 411 void WasmObjectWriter::startCustomSection(SectionBookkeeping &Section, 412 StringRef Name) { 413 LLVM_DEBUG(dbgs() << "startCustomSection " << Name << "\n"); 414 startSection(Section, wasm::WASM_SEC_CUSTOM); 415 416 // The position where the section header ends, for measuring its size. 417 Section.PayloadOffset = W->OS.tell(); 418 419 // Custom sections in wasm also have a string identifier. 420 if (Name != "__clangast") { 421 writeString(Name); 422 } else { 423 // The on-disk hashtable in clangast needs to be aligned by 4 bytes. 424 writeStringWithAlignment(Name, 4); 425 } 426 427 // The position where the custom section starts. 428 Section.ContentsOffset = W->OS.tell(); 429 } 430 431 // Now that the section is complete and we know how big it is, patch up the 432 // section size field at the start of the section. 433 void WasmObjectWriter::endSection(SectionBookkeeping &Section) { 434 uint64_t Size = W->OS.tell(); 435 // /dev/null doesn't support seek/tell and can report offset of 0. 436 // Simply skip this patching in that case. 437 if (!Size) 438 return; 439 440 Size -= Section.PayloadOffset; 441 if (uint32_t(Size) != Size) 442 report_fatal_error("section size does not fit in a uint32_t"); 443 444 LLVM_DEBUG(dbgs() << "endSection size=" << Size << "\n"); 445 446 // Write the final section size to the payload_len field, which follows 447 // the section id byte. 448 writePatchableU32(static_cast<raw_pwrite_stream &>(W->OS), Size, 449 Section.SizeOffset); 450 } 451 452 // Emit the Wasm header. 453 void WasmObjectWriter::writeHeader(const MCAssembler &Asm) { 454 W->OS.write(wasm::WasmMagic, sizeof(wasm::WasmMagic)); 455 W->write<uint32_t>(wasm::WasmVersion); 456 } 457 458 void WasmObjectWriter::executePostLayoutBinding(MCAssembler &Asm, 459 const MCAsmLayout &Layout) { 460 // Some compilation units require the indirect function table to be present 461 // but don't explicitly reference it. This is the case for call_indirect 462 // without the reference-types feature, and also function bitcasts in all 463 // cases. In those cases the __indirect_function_table has the 464 // WASM_SYMBOL_NO_STRIP attribute. Here we make sure this symbol makes it to 465 // the assembler, if needed. 466 if (auto *Sym = Asm.getContext().lookupSymbol("__indirect_function_table")) { 467 const auto *WasmSym = static_cast<const MCSymbolWasm *>(Sym); 468 if (WasmSym->isNoStrip()) 469 Asm.registerSymbol(*Sym); 470 } 471 472 // Build a map of sections to the function that defines them, for use 473 // in recordRelocation. 474 for (const MCSymbol &S : Asm.symbols()) { 475 const auto &WS = static_cast<const MCSymbolWasm &>(S); 476 if (WS.isDefined() && WS.isFunction() && !WS.isVariable()) { 477 const auto &Sec = static_cast<const MCSectionWasm &>(S.getSection()); 478 auto Pair = SectionFunctions.insert(std::make_pair(&Sec, &S)); 479 if (!Pair.second) 480 report_fatal_error("section already has a defining function: " + 481 Sec.getName()); 482 } 483 } 484 } 485 486 void WasmObjectWriter::recordRelocation(MCAssembler &Asm, 487 const MCAsmLayout &Layout, 488 const MCFragment *Fragment, 489 const MCFixup &Fixup, MCValue Target, 490 uint64_t &FixedValue) { 491 // The WebAssembly backend should never generate FKF_IsPCRel fixups 492 assert(!(Asm.getBackend().getFixupKindInfo(Fixup.getKind()).Flags & 493 MCFixupKindInfo::FKF_IsPCRel)); 494 495 const auto &FixupSection = cast<MCSectionWasm>(*Fragment->getParent()); 496 uint64_t C = Target.getConstant(); 497 uint64_t FixupOffset = Layout.getFragmentOffset(Fragment) + Fixup.getOffset(); 498 MCContext &Ctx = Asm.getContext(); 499 bool IsLocRel = false; 500 501 if (const MCSymbolRefExpr *RefB = Target.getSymB()) { 502 503 const auto &SymB = cast<MCSymbolWasm>(RefB->getSymbol()); 504 505 if (FixupSection.getKind().isText()) { 506 Ctx.reportError(Fixup.getLoc(), 507 Twine("symbol '") + SymB.getName() + 508 "' unsupported subtraction expression used in " 509 "relocation in code section."); 510 return; 511 } 512 513 if (SymB.isUndefined()) { 514 Ctx.reportError(Fixup.getLoc(), 515 Twine("symbol '") + SymB.getName() + 516 "' can not be undefined in a subtraction expression"); 517 return; 518 } 519 const MCSection &SecB = SymB.getSection(); 520 if (&SecB != &FixupSection) { 521 Ctx.reportError(Fixup.getLoc(), 522 Twine("symbol '") + SymB.getName() + 523 "' can not be placed in a different section"); 524 return; 525 } 526 IsLocRel = true; 527 C += FixupOffset - Layout.getSymbolOffset(SymB); 528 } 529 530 // We either rejected the fixup or folded B into C at this point. 531 const MCSymbolRefExpr *RefA = Target.getSymA(); 532 const auto *SymA = cast<MCSymbolWasm>(&RefA->getSymbol()); 533 534 // The .init_array isn't translated as data, so don't do relocations in it. 535 if (FixupSection.getName().startswith(".init_array")) { 536 SymA->setUsedInInitArray(); 537 return; 538 } 539 540 if (SymA->isVariable()) { 541 const MCExpr *Expr = SymA->getVariableValue(); 542 if (const auto *Inner = dyn_cast<MCSymbolRefExpr>(Expr)) 543 if (Inner->getKind() == MCSymbolRefExpr::VK_WEAKREF) 544 llvm_unreachable("weakref used in reloc not yet implemented"); 545 } 546 547 // Put any constant offset in an addend. Offsets can be negative, and 548 // LLVM expects wrapping, in contrast to wasm's immediates which can't 549 // be negative and don't wrap. 550 FixedValue = 0; 551 552 unsigned Type = 553 TargetObjectWriter->getRelocType(Target, Fixup, FixupSection, IsLocRel); 554 555 // Absolute offset within a section or a function. 556 // Currently only supported for for metadata sections. 557 // See: test/MC/WebAssembly/blockaddress.ll 558 if ((Type == wasm::R_WASM_FUNCTION_OFFSET_I32 || 559 Type == wasm::R_WASM_FUNCTION_OFFSET_I64 || 560 Type == wasm::R_WASM_SECTION_OFFSET_I32) && 561 SymA->isDefined()) { 562 // SymA can be a temp data symbol that represents a function (in which case 563 // it needs to be replaced by the section symbol), [XXX and it apparently 564 // later gets changed again to a func symbol?] or it can be a real 565 // function symbol, in which case it can be left as-is. 566 567 if (!FixupSection.getKind().isMetadata()) 568 report_fatal_error("relocations for function or section offsets are " 569 "only supported in metadata sections"); 570 571 const MCSymbol *SectionSymbol = nullptr; 572 const MCSection &SecA = SymA->getSection(); 573 if (SecA.getKind().isText()) { 574 auto SecSymIt = SectionFunctions.find(&SecA); 575 if (SecSymIt == SectionFunctions.end()) 576 report_fatal_error("section doesn\'t have defining symbol"); 577 SectionSymbol = SecSymIt->second; 578 } else { 579 SectionSymbol = SecA.getBeginSymbol(); 580 } 581 if (!SectionSymbol) 582 report_fatal_error("section symbol is required for relocation"); 583 584 C += Layout.getSymbolOffset(*SymA); 585 SymA = cast<MCSymbolWasm>(SectionSymbol); 586 } 587 588 if (Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB || 589 Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB64 || 590 Type == wasm::R_WASM_TABLE_INDEX_SLEB || 591 Type == wasm::R_WASM_TABLE_INDEX_SLEB64 || 592 Type == wasm::R_WASM_TABLE_INDEX_I32 || 593 Type == wasm::R_WASM_TABLE_INDEX_I64) { 594 // TABLE_INDEX relocs implicitly use the default indirect function table. 595 // We require the function table to have already been defined. 596 auto TableName = "__indirect_function_table"; 597 MCSymbolWasm *Sym = cast_or_null<MCSymbolWasm>(Ctx.lookupSymbol(TableName)); 598 if (!Sym) { 599 report_fatal_error("missing indirect function table symbol"); 600 } else { 601 if (!Sym->isFunctionTable()) 602 report_fatal_error("__indirect_function_table symbol has wrong type"); 603 // Ensure that __indirect_function_table reaches the output. 604 Sym->setNoStrip(); 605 Asm.registerSymbol(*Sym); 606 } 607 } 608 609 // Relocation other than R_WASM_TYPE_INDEX_LEB are required to be 610 // against a named symbol. 611 if (Type != wasm::R_WASM_TYPE_INDEX_LEB) { 612 if (SymA->getName().empty()) 613 report_fatal_error("relocations against un-named temporaries are not yet " 614 "supported by wasm"); 615 616 SymA->setUsedInReloc(); 617 } 618 619 switch (RefA->getKind()) { 620 case MCSymbolRefExpr::VK_GOT: 621 case MCSymbolRefExpr::VK_WASM_GOT_TLS: 622 SymA->setUsedInGOT(); 623 break; 624 default: 625 break; 626 } 627 628 WasmRelocationEntry Rec(FixupOffset, SymA, C, Type, &FixupSection); 629 LLVM_DEBUG(dbgs() << "WasmReloc: " << Rec << "\n"); 630 631 if (FixupSection.isWasmData()) { 632 DataRelocations.push_back(Rec); 633 } else if (FixupSection.getKind().isText()) { 634 CodeRelocations.push_back(Rec); 635 } else if (FixupSection.getKind().isMetadata()) { 636 CustomSectionsRelocations[&FixupSection].push_back(Rec); 637 } else { 638 llvm_unreachable("unexpected section type"); 639 } 640 } 641 642 // Compute a value to write into the code at the location covered 643 // by RelEntry. This value isn't used by the static linker; it just serves 644 // to make the object format more readable and more likely to be directly 645 // useable. 646 uint64_t 647 WasmObjectWriter::getProvisionalValue(const WasmRelocationEntry &RelEntry, 648 const MCAsmLayout &Layout) { 649 if ((RelEntry.Type == wasm::R_WASM_GLOBAL_INDEX_LEB || 650 RelEntry.Type == wasm::R_WASM_GLOBAL_INDEX_I32) && 651 !RelEntry.Symbol->isGlobal()) { 652 assert(GOTIndices.count(RelEntry.Symbol) > 0 && "symbol not found in GOT index space"); 653 return GOTIndices[RelEntry.Symbol]; 654 } 655 656 switch (RelEntry.Type) { 657 case wasm::R_WASM_TABLE_INDEX_REL_SLEB: 658 case wasm::R_WASM_TABLE_INDEX_REL_SLEB64: 659 case wasm::R_WASM_TABLE_INDEX_SLEB: 660 case wasm::R_WASM_TABLE_INDEX_SLEB64: 661 case wasm::R_WASM_TABLE_INDEX_I32: 662 case wasm::R_WASM_TABLE_INDEX_I64: { 663 // Provisional value is table address of the resolved symbol itself 664 const MCSymbolWasm *Base = 665 cast<MCSymbolWasm>(Layout.getBaseSymbol(*RelEntry.Symbol)); 666 assert(Base->isFunction()); 667 if (RelEntry.Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB || 668 RelEntry.Type == wasm::R_WASM_TABLE_INDEX_REL_SLEB64) 669 return TableIndices[Base] - InitialTableOffset; 670 else 671 return TableIndices[Base]; 672 } 673 case wasm::R_WASM_TYPE_INDEX_LEB: 674 // Provisional value is same as the index 675 return getRelocationIndexValue(RelEntry); 676 case wasm::R_WASM_FUNCTION_INDEX_LEB: 677 case wasm::R_WASM_GLOBAL_INDEX_LEB: 678 case wasm::R_WASM_GLOBAL_INDEX_I32: 679 case wasm::R_WASM_TAG_INDEX_LEB: 680 case wasm::R_WASM_TABLE_NUMBER_LEB: 681 // Provisional value is function/global/tag Wasm index 682 assert(WasmIndices.count(RelEntry.Symbol) > 0 && "symbol not found in wasm index space"); 683 return WasmIndices[RelEntry.Symbol]; 684 case wasm::R_WASM_FUNCTION_OFFSET_I32: 685 case wasm::R_WASM_FUNCTION_OFFSET_I64: 686 case wasm::R_WASM_SECTION_OFFSET_I32: { 687 if (!RelEntry.Symbol->isDefined()) 688 return 0; 689 const auto &Section = 690 static_cast<const MCSectionWasm &>(RelEntry.Symbol->getSection()); 691 return Section.getSectionOffset() + RelEntry.Addend; 692 } 693 case wasm::R_WASM_MEMORY_ADDR_LEB: 694 case wasm::R_WASM_MEMORY_ADDR_LEB64: 695 case wasm::R_WASM_MEMORY_ADDR_SLEB: 696 case wasm::R_WASM_MEMORY_ADDR_SLEB64: 697 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB: 698 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB64: 699 case wasm::R_WASM_MEMORY_ADDR_I32: 700 case wasm::R_WASM_MEMORY_ADDR_I64: 701 case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB: 702 case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB64: 703 case wasm::R_WASM_MEMORY_ADDR_LOCREL_I32: { 704 // Provisional value is address of the global plus the offset 705 // For undefined symbols, use zero 706 if (!RelEntry.Symbol->isDefined()) 707 return 0; 708 const wasm::WasmDataReference &SymRef = DataLocations[RelEntry.Symbol]; 709 const WasmDataSegment &Segment = DataSegments[SymRef.Segment]; 710 // Ignore overflow. LLVM allows address arithmetic to silently wrap. 711 return Segment.Offset + SymRef.Offset + RelEntry.Addend; 712 } 713 default: 714 llvm_unreachable("invalid relocation type"); 715 } 716 } 717 718 static void addData(SmallVectorImpl<char> &DataBytes, 719 MCSectionWasm &DataSection) { 720 LLVM_DEBUG(errs() << "addData: " << DataSection.getName() << "\n"); 721 722 DataBytes.resize(alignTo(DataBytes.size(), DataSection.getAlignment())); 723 724 for (const MCFragment &Frag : DataSection) { 725 if (Frag.hasInstructions()) 726 report_fatal_error("only data supported in data sections"); 727 728 if (auto *Align = dyn_cast<MCAlignFragment>(&Frag)) { 729 if (Align->getValueSize() != 1) 730 report_fatal_error("only byte values supported for alignment"); 731 // If nops are requested, use zeros, as this is the data section. 732 uint8_t Value = Align->hasEmitNops() ? 0 : Align->getValue(); 733 uint64_t Size = 734 std::min<uint64_t>(alignTo(DataBytes.size(), Align->getAlignment()), 735 DataBytes.size() + Align->getMaxBytesToEmit()); 736 DataBytes.resize(Size, Value); 737 } else if (auto *Fill = dyn_cast<MCFillFragment>(&Frag)) { 738 int64_t NumValues; 739 if (!Fill->getNumValues().evaluateAsAbsolute(NumValues)) 740 llvm_unreachable("The fill should be an assembler constant"); 741 DataBytes.insert(DataBytes.end(), Fill->getValueSize() * NumValues, 742 Fill->getValue()); 743 } else if (auto *LEB = dyn_cast<MCLEBFragment>(&Frag)) { 744 const SmallVectorImpl<char> &Contents = LEB->getContents(); 745 llvm::append_range(DataBytes, Contents); 746 } else { 747 const auto &DataFrag = cast<MCDataFragment>(Frag); 748 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 749 llvm::append_range(DataBytes, Contents); 750 } 751 } 752 753 LLVM_DEBUG(dbgs() << "addData -> " << DataBytes.size() << "\n"); 754 } 755 756 uint32_t 757 WasmObjectWriter::getRelocationIndexValue(const WasmRelocationEntry &RelEntry) { 758 if (RelEntry.Type == wasm::R_WASM_TYPE_INDEX_LEB) { 759 if (!TypeIndices.count(RelEntry.Symbol)) 760 report_fatal_error("symbol not found in type index space: " + 761 RelEntry.Symbol->getName()); 762 return TypeIndices[RelEntry.Symbol]; 763 } 764 765 return RelEntry.Symbol->getIndex(); 766 } 767 768 // Apply the portions of the relocation records that we can handle ourselves 769 // directly. 770 void WasmObjectWriter::applyRelocations( 771 ArrayRef<WasmRelocationEntry> Relocations, uint64_t ContentsOffset, 772 const MCAsmLayout &Layout) { 773 auto &Stream = static_cast<raw_pwrite_stream &>(W->OS); 774 for (const WasmRelocationEntry &RelEntry : Relocations) { 775 uint64_t Offset = ContentsOffset + 776 RelEntry.FixupSection->getSectionOffset() + 777 RelEntry.Offset; 778 779 LLVM_DEBUG(dbgs() << "applyRelocation: " << RelEntry << "\n"); 780 uint64_t Value = getProvisionalValue(RelEntry, Layout); 781 782 switch (RelEntry.Type) { 783 case wasm::R_WASM_FUNCTION_INDEX_LEB: 784 case wasm::R_WASM_TYPE_INDEX_LEB: 785 case wasm::R_WASM_GLOBAL_INDEX_LEB: 786 case wasm::R_WASM_MEMORY_ADDR_LEB: 787 case wasm::R_WASM_TAG_INDEX_LEB: 788 case wasm::R_WASM_TABLE_NUMBER_LEB: 789 writePatchableU32(Stream, Value, Offset); 790 break; 791 case wasm::R_WASM_MEMORY_ADDR_LEB64: 792 writePatchableU64(Stream, Value, Offset); 793 break; 794 case wasm::R_WASM_TABLE_INDEX_I32: 795 case wasm::R_WASM_MEMORY_ADDR_I32: 796 case wasm::R_WASM_FUNCTION_OFFSET_I32: 797 case wasm::R_WASM_SECTION_OFFSET_I32: 798 case wasm::R_WASM_GLOBAL_INDEX_I32: 799 case wasm::R_WASM_MEMORY_ADDR_LOCREL_I32: 800 patchI32(Stream, Value, Offset); 801 break; 802 case wasm::R_WASM_TABLE_INDEX_I64: 803 case wasm::R_WASM_MEMORY_ADDR_I64: 804 case wasm::R_WASM_FUNCTION_OFFSET_I64: 805 patchI64(Stream, Value, Offset); 806 break; 807 case wasm::R_WASM_TABLE_INDEX_SLEB: 808 case wasm::R_WASM_TABLE_INDEX_REL_SLEB: 809 case wasm::R_WASM_MEMORY_ADDR_SLEB: 810 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB: 811 case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB: 812 writePatchableS32(Stream, Value, Offset); 813 break; 814 case wasm::R_WASM_TABLE_INDEX_SLEB64: 815 case wasm::R_WASM_TABLE_INDEX_REL_SLEB64: 816 case wasm::R_WASM_MEMORY_ADDR_SLEB64: 817 case wasm::R_WASM_MEMORY_ADDR_REL_SLEB64: 818 case wasm::R_WASM_MEMORY_ADDR_TLS_SLEB64: 819 writePatchableS64(Stream, Value, Offset); 820 break; 821 default: 822 llvm_unreachable("invalid relocation type"); 823 } 824 } 825 } 826 827 void WasmObjectWriter::writeTypeSection( 828 ArrayRef<wasm::WasmSignature> Signatures) { 829 if (Signatures.empty()) 830 return; 831 832 SectionBookkeeping Section; 833 startSection(Section, wasm::WASM_SEC_TYPE); 834 835 encodeULEB128(Signatures.size(), W->OS); 836 837 for (const wasm::WasmSignature &Sig : Signatures) { 838 W->OS << char(wasm::WASM_TYPE_FUNC); 839 encodeULEB128(Sig.Params.size(), W->OS); 840 for (wasm::ValType Ty : Sig.Params) 841 writeValueType(Ty); 842 encodeULEB128(Sig.Returns.size(), W->OS); 843 for (wasm::ValType Ty : Sig.Returns) 844 writeValueType(Ty); 845 } 846 847 endSection(Section); 848 } 849 850 void WasmObjectWriter::writeImportSection(ArrayRef<wasm::WasmImport> Imports, 851 uint64_t DataSize, 852 uint32_t NumElements) { 853 if (Imports.empty()) 854 return; 855 856 uint64_t NumPages = (DataSize + wasm::WasmPageSize - 1) / wasm::WasmPageSize; 857 858 SectionBookkeeping Section; 859 startSection(Section, wasm::WASM_SEC_IMPORT); 860 861 encodeULEB128(Imports.size(), W->OS); 862 for (const wasm::WasmImport &Import : Imports) { 863 writeString(Import.Module); 864 writeString(Import.Field); 865 W->OS << char(Import.Kind); 866 867 switch (Import.Kind) { 868 case wasm::WASM_EXTERNAL_FUNCTION: 869 encodeULEB128(Import.SigIndex, W->OS); 870 break; 871 case wasm::WASM_EXTERNAL_GLOBAL: 872 W->OS << char(Import.Global.Type); 873 W->OS << char(Import.Global.Mutable ? 1 : 0); 874 break; 875 case wasm::WASM_EXTERNAL_MEMORY: 876 encodeULEB128(Import.Memory.Flags, W->OS); 877 encodeULEB128(NumPages, W->OS); // initial 878 break; 879 case wasm::WASM_EXTERNAL_TABLE: 880 W->OS << char(Import.Table.ElemType); 881 encodeULEB128(0, W->OS); // flags 882 encodeULEB128(NumElements, W->OS); // initial 883 break; 884 case wasm::WASM_EXTERNAL_TAG: 885 W->OS << char(0); // Reserved 'attribute' field 886 encodeULEB128(Import.SigIndex, W->OS); 887 break; 888 default: 889 llvm_unreachable("unsupported import kind"); 890 } 891 } 892 893 endSection(Section); 894 } 895 896 void WasmObjectWriter::writeFunctionSection(ArrayRef<WasmFunction> Functions) { 897 if (Functions.empty()) 898 return; 899 900 SectionBookkeeping Section; 901 startSection(Section, wasm::WASM_SEC_FUNCTION); 902 903 encodeULEB128(Functions.size(), W->OS); 904 for (const WasmFunction &Func : Functions) 905 encodeULEB128(Func.SigIndex, W->OS); 906 907 endSection(Section); 908 } 909 910 void WasmObjectWriter::writeTagSection(ArrayRef<uint32_t> TagTypes) { 911 if (TagTypes.empty()) 912 return; 913 914 SectionBookkeeping Section; 915 startSection(Section, wasm::WASM_SEC_TAG); 916 917 encodeULEB128(TagTypes.size(), W->OS); 918 for (uint32_t Index : TagTypes) { 919 W->OS << char(0); // Reserved 'attribute' field 920 encodeULEB128(Index, W->OS); 921 } 922 923 endSection(Section); 924 } 925 926 void WasmObjectWriter::writeGlobalSection(ArrayRef<wasm::WasmGlobal> Globals) { 927 if (Globals.empty()) 928 return; 929 930 SectionBookkeeping Section; 931 startSection(Section, wasm::WASM_SEC_GLOBAL); 932 933 encodeULEB128(Globals.size(), W->OS); 934 for (const wasm::WasmGlobal &Global : Globals) { 935 encodeULEB128(Global.Type.Type, W->OS); 936 W->OS << char(Global.Type.Mutable); 937 W->OS << char(Global.InitExpr.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 W->OS << char(wasm::WASM_OPCODE_END); 958 } 959 960 endSection(Section); 961 } 962 963 void WasmObjectWriter::writeTableSection(ArrayRef<wasm::WasmTable> Tables) { 964 if (Tables.empty()) 965 return; 966 967 SectionBookkeeping Section; 968 startSection(Section, wasm::WASM_SEC_TABLE); 969 970 encodeULEB128(Tables.size(), W->OS); 971 for (const wasm::WasmTable &Table : Tables) { 972 encodeULEB128(Table.Type.ElemType, W->OS); 973 encodeULEB128(Table.Type.Limits.Flags, W->OS); 974 encodeULEB128(Table.Type.Limits.Minimum, W->OS); 975 if (Table.Type.Limits.Flags & wasm::WASM_LIMITS_FLAG_HAS_MAX) 976 encodeULEB128(Table.Type.Limits.Maximum, W->OS); 977 } 978 endSection(Section); 979 } 980 981 void WasmObjectWriter::writeExportSection(ArrayRef<wasm::WasmExport> Exports) { 982 if (Exports.empty()) 983 return; 984 985 SectionBookkeeping Section; 986 startSection(Section, wasm::WASM_SEC_EXPORT); 987 988 encodeULEB128(Exports.size(), W->OS); 989 for (const wasm::WasmExport &Export : Exports) { 990 writeString(Export.Name); 991 W->OS << char(Export.Kind); 992 encodeULEB128(Export.Index, W->OS); 993 } 994 995 endSection(Section); 996 } 997 998 void WasmObjectWriter::writeElemSection( 999 const MCSymbolWasm *IndirectFunctionTable, ArrayRef<uint32_t> TableElems) { 1000 if (TableElems.empty()) 1001 return; 1002 1003 assert(IndirectFunctionTable); 1004 1005 SectionBookkeeping Section; 1006 startSection(Section, wasm::WASM_SEC_ELEM); 1007 1008 encodeULEB128(1, W->OS); // number of "segments" 1009 1010 assert(WasmIndices.count(IndirectFunctionTable)); 1011 uint32_t TableNumber = WasmIndices.find(IndirectFunctionTable)->second; 1012 uint32_t Flags = 0; 1013 if (TableNumber) 1014 Flags |= wasm::WASM_ELEM_SEGMENT_HAS_TABLE_NUMBER; 1015 encodeULEB128(Flags, W->OS); 1016 if (Flags & wasm::WASM_ELEM_SEGMENT_HAS_TABLE_NUMBER) 1017 encodeULEB128(TableNumber, W->OS); // the table number 1018 1019 // init expr for starting offset 1020 W->OS << char(wasm::WASM_OPCODE_I32_CONST); 1021 encodeSLEB128(InitialTableOffset, W->OS); 1022 W->OS << char(wasm::WASM_OPCODE_END); 1023 1024 if (Flags & wasm::WASM_ELEM_SEGMENT_MASK_HAS_ELEM_KIND) { 1025 // We only write active function table initializers, for which the elem kind 1026 // is specified to be written as 0x00 and interpreted to mean "funcref". 1027 const uint8_t ElemKind = 0; 1028 W->OS << ElemKind; 1029 } 1030 1031 encodeULEB128(TableElems.size(), W->OS); 1032 for (uint32_t Elem : TableElems) 1033 encodeULEB128(Elem, W->OS); 1034 1035 endSection(Section); 1036 } 1037 1038 void WasmObjectWriter::writeDataCountSection() { 1039 if (DataSegments.empty()) 1040 return; 1041 1042 SectionBookkeeping Section; 1043 startSection(Section, wasm::WASM_SEC_DATACOUNT); 1044 encodeULEB128(DataSegments.size(), W->OS); 1045 endSection(Section); 1046 } 1047 1048 uint32_t WasmObjectWriter::writeCodeSection(const MCAssembler &Asm, 1049 const MCAsmLayout &Layout, 1050 ArrayRef<WasmFunction> Functions) { 1051 if (Functions.empty()) 1052 return 0; 1053 1054 SectionBookkeeping Section; 1055 startSection(Section, wasm::WASM_SEC_CODE); 1056 1057 encodeULEB128(Functions.size(), W->OS); 1058 1059 for (const WasmFunction &Func : Functions) { 1060 auto &FuncSection = static_cast<MCSectionWasm &>(Func.Sym->getSection()); 1061 1062 int64_t Size = 0; 1063 if (!Func.Sym->getSize()->evaluateAsAbsolute(Size, Layout)) 1064 report_fatal_error(".size expression must be evaluatable"); 1065 1066 encodeULEB128(Size, W->OS); 1067 FuncSection.setSectionOffset(W->OS.tell() - Section.ContentsOffset); 1068 Asm.writeSectionData(W->OS, &FuncSection, Layout); 1069 } 1070 1071 // Apply fixups. 1072 applyRelocations(CodeRelocations, Section.ContentsOffset, Layout); 1073 1074 endSection(Section); 1075 return Section.Index; 1076 } 1077 1078 uint32_t WasmObjectWriter::writeDataSection(const MCAsmLayout &Layout) { 1079 if (DataSegments.empty()) 1080 return 0; 1081 1082 SectionBookkeeping Section; 1083 startSection(Section, wasm::WASM_SEC_DATA); 1084 1085 encodeULEB128(DataSegments.size(), W->OS); // count 1086 1087 for (const WasmDataSegment &Segment : DataSegments) { 1088 encodeULEB128(Segment.InitFlags, W->OS); // flags 1089 if (Segment.InitFlags & wasm::WASM_DATA_SEGMENT_HAS_MEMINDEX) 1090 encodeULEB128(0, W->OS); // memory index 1091 if ((Segment.InitFlags & wasm::WASM_DATA_SEGMENT_IS_PASSIVE) == 0) { 1092 W->OS << char(is64Bit() ? wasm::WASM_OPCODE_I64_CONST 1093 : wasm::WASM_OPCODE_I32_CONST); 1094 encodeSLEB128(Segment.Offset, W->OS); // offset 1095 W->OS << char(wasm::WASM_OPCODE_END); 1096 } 1097 encodeULEB128(Segment.Data.size(), W->OS); // size 1098 Segment.Section->setSectionOffset(W->OS.tell() - Section.ContentsOffset); 1099 W->OS << Segment.Data; // data 1100 } 1101 1102 // Apply fixups. 1103 applyRelocations(DataRelocations, Section.ContentsOffset, Layout); 1104 1105 endSection(Section); 1106 return Section.Index; 1107 } 1108 1109 void WasmObjectWriter::writeRelocSection( 1110 uint32_t SectionIndex, StringRef Name, 1111 std::vector<WasmRelocationEntry> &Relocs) { 1112 // See: https://github.com/WebAssembly/tool-conventions/blob/main/Linking.md 1113 // for descriptions of the reloc sections. 1114 1115 if (Relocs.empty()) 1116 return; 1117 1118 // First, ensure the relocations are sorted in offset order. In general they 1119 // should already be sorted since `recordRelocation` is called in offset 1120 // order, but for the code section we combine many MC sections into single 1121 // wasm section, and this order is determined by the order of Asm.Symbols() 1122 // not the sections order. 1123 llvm::stable_sort( 1124 Relocs, [](const WasmRelocationEntry &A, const WasmRelocationEntry &B) { 1125 return (A.Offset + A.FixupSection->getSectionOffset()) < 1126 (B.Offset + B.FixupSection->getSectionOffset()); 1127 }); 1128 1129 SectionBookkeeping Section; 1130 startCustomSection(Section, std::string("reloc.") + Name.str()); 1131 1132 encodeULEB128(SectionIndex, W->OS); 1133 encodeULEB128(Relocs.size(), W->OS); 1134 for (const WasmRelocationEntry &RelEntry : Relocs) { 1135 uint64_t Offset = 1136 RelEntry.Offset + RelEntry.FixupSection->getSectionOffset(); 1137 uint32_t Index = getRelocationIndexValue(RelEntry); 1138 1139 W->OS << char(RelEntry.Type); 1140 encodeULEB128(Offset, W->OS); 1141 encodeULEB128(Index, W->OS); 1142 if (RelEntry.hasAddend()) 1143 encodeSLEB128(RelEntry.Addend, W->OS); 1144 } 1145 1146 endSection(Section); 1147 } 1148 1149 void WasmObjectWriter::writeCustomRelocSections() { 1150 for (const auto &Sec : CustomSections) { 1151 auto &Relocations = CustomSectionsRelocations[Sec.Section]; 1152 writeRelocSection(Sec.OutputIndex, Sec.Name, Relocations); 1153 } 1154 } 1155 1156 void WasmObjectWriter::writeLinkingMetaDataSection( 1157 ArrayRef<wasm::WasmSymbolInfo> SymbolInfos, 1158 ArrayRef<std::pair<uint16_t, uint32_t>> InitFuncs, 1159 const std::map<StringRef, std::vector<WasmComdatEntry>> &Comdats) { 1160 SectionBookkeeping Section; 1161 startCustomSection(Section, "linking"); 1162 encodeULEB128(wasm::WasmMetadataVersion, W->OS); 1163 1164 SectionBookkeeping SubSection; 1165 if (SymbolInfos.size() != 0) { 1166 startSection(SubSection, wasm::WASM_SYMBOL_TABLE); 1167 encodeULEB128(SymbolInfos.size(), W->OS); 1168 for (const wasm::WasmSymbolInfo &Sym : SymbolInfos) { 1169 encodeULEB128(Sym.Kind, W->OS); 1170 encodeULEB128(Sym.Flags, W->OS); 1171 switch (Sym.Kind) { 1172 case wasm::WASM_SYMBOL_TYPE_FUNCTION: 1173 case wasm::WASM_SYMBOL_TYPE_GLOBAL: 1174 case wasm::WASM_SYMBOL_TYPE_TAG: 1175 case wasm::WASM_SYMBOL_TYPE_TABLE: 1176 encodeULEB128(Sym.ElementIndex, W->OS); 1177 if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0 || 1178 (Sym.Flags & wasm::WASM_SYMBOL_EXPLICIT_NAME) != 0) 1179 writeString(Sym.Name); 1180 break; 1181 case wasm::WASM_SYMBOL_TYPE_DATA: 1182 writeString(Sym.Name); 1183 if ((Sym.Flags & wasm::WASM_SYMBOL_UNDEFINED) == 0) { 1184 encodeULEB128(Sym.DataRef.Segment, W->OS); 1185 encodeULEB128(Sym.DataRef.Offset, W->OS); 1186 encodeULEB128(Sym.DataRef.Size, W->OS); 1187 } 1188 break; 1189 case wasm::WASM_SYMBOL_TYPE_SECTION: { 1190 const uint32_t SectionIndex = 1191 CustomSections[Sym.ElementIndex].OutputIndex; 1192 encodeULEB128(SectionIndex, W->OS); 1193 break; 1194 } 1195 default: 1196 llvm_unreachable("unexpected kind"); 1197 } 1198 } 1199 endSection(SubSection); 1200 } 1201 1202 if (DataSegments.size()) { 1203 startSection(SubSection, wasm::WASM_SEGMENT_INFO); 1204 encodeULEB128(DataSegments.size(), W->OS); 1205 for (const WasmDataSegment &Segment : DataSegments) { 1206 writeString(Segment.Name); 1207 encodeULEB128(Segment.Alignment, W->OS); 1208 encodeULEB128(Segment.LinkingFlags, W->OS); 1209 } 1210 endSection(SubSection); 1211 } 1212 1213 if (!InitFuncs.empty()) { 1214 startSection(SubSection, wasm::WASM_INIT_FUNCS); 1215 encodeULEB128(InitFuncs.size(), W->OS); 1216 for (auto &StartFunc : InitFuncs) { 1217 encodeULEB128(StartFunc.first, W->OS); // priority 1218 encodeULEB128(StartFunc.second, W->OS); // function index 1219 } 1220 endSection(SubSection); 1221 } 1222 1223 if (Comdats.size()) { 1224 startSection(SubSection, wasm::WASM_COMDAT_INFO); 1225 encodeULEB128(Comdats.size(), W->OS); 1226 for (const auto &C : Comdats) { 1227 writeString(C.first); 1228 encodeULEB128(0, W->OS); // flags for future use 1229 encodeULEB128(C.second.size(), W->OS); 1230 for (const WasmComdatEntry &Entry : C.second) { 1231 encodeULEB128(Entry.Kind, W->OS); 1232 encodeULEB128(Entry.Index, W->OS); 1233 } 1234 } 1235 endSection(SubSection); 1236 } 1237 1238 endSection(Section); 1239 } 1240 1241 void WasmObjectWriter::writeCustomSection(WasmCustomSection &CustomSection, 1242 const MCAssembler &Asm, 1243 const MCAsmLayout &Layout) { 1244 SectionBookkeeping Section; 1245 auto *Sec = CustomSection.Section; 1246 startCustomSection(Section, CustomSection.Name); 1247 1248 Sec->setSectionOffset(W->OS.tell() - Section.ContentsOffset); 1249 Asm.writeSectionData(W->OS, Sec, Layout); 1250 1251 CustomSection.OutputContentsOffset = Section.ContentsOffset; 1252 CustomSection.OutputIndex = Section.Index; 1253 1254 endSection(Section); 1255 1256 // Apply fixups. 1257 auto &Relocations = CustomSectionsRelocations[CustomSection.Section]; 1258 applyRelocations(Relocations, CustomSection.OutputContentsOffset, Layout); 1259 } 1260 1261 uint32_t WasmObjectWriter::getFunctionType(const MCSymbolWasm &Symbol) { 1262 assert(Symbol.isFunction()); 1263 assert(TypeIndices.count(&Symbol)); 1264 return TypeIndices[&Symbol]; 1265 } 1266 1267 uint32_t WasmObjectWriter::getTagType(const MCSymbolWasm &Symbol) { 1268 assert(Symbol.isTag()); 1269 assert(TypeIndices.count(&Symbol)); 1270 return TypeIndices[&Symbol]; 1271 } 1272 1273 void WasmObjectWriter::registerFunctionType(const MCSymbolWasm &Symbol) { 1274 assert(Symbol.isFunction()); 1275 1276 wasm::WasmSignature S; 1277 1278 if (auto *Sig = Symbol.getSignature()) { 1279 S.Returns = Sig->Returns; 1280 S.Params = Sig->Params; 1281 } 1282 1283 auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size())); 1284 if (Pair.second) 1285 Signatures.push_back(S); 1286 TypeIndices[&Symbol] = Pair.first->second; 1287 1288 LLVM_DEBUG(dbgs() << "registerFunctionType: " << Symbol 1289 << " new:" << Pair.second << "\n"); 1290 LLVM_DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n"); 1291 } 1292 1293 void WasmObjectWriter::registerTagType(const MCSymbolWasm &Symbol) { 1294 assert(Symbol.isTag()); 1295 1296 // TODO Currently we don't generate imported exceptions, but if we do, we 1297 // should have a way of infering types of imported exceptions. 1298 wasm::WasmSignature S; 1299 if (auto *Sig = Symbol.getSignature()) { 1300 S.Returns = Sig->Returns; 1301 S.Params = Sig->Params; 1302 } 1303 1304 auto Pair = SignatureIndices.insert(std::make_pair(S, Signatures.size())); 1305 if (Pair.second) 1306 Signatures.push_back(S); 1307 TypeIndices[&Symbol] = Pair.first->second; 1308 1309 LLVM_DEBUG(dbgs() << "registerTagType: " << Symbol << " new:" << Pair.second 1310 << "\n"); 1311 LLVM_DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n"); 1312 } 1313 1314 static bool isInSymtab(const MCSymbolWasm &Sym) { 1315 if (Sym.isUsedInReloc() || Sym.isUsedInInitArray()) 1316 return true; 1317 1318 if (Sym.isComdat() && !Sym.isDefined()) 1319 return false; 1320 1321 if (Sym.isTemporary()) 1322 return false; 1323 1324 if (Sym.isSection()) 1325 return false; 1326 1327 if (Sym.omitFromLinkingSection()) 1328 return false; 1329 1330 return true; 1331 } 1332 1333 void WasmObjectWriter::prepareImports( 1334 SmallVectorImpl<wasm::WasmImport> &Imports, MCAssembler &Asm, 1335 const MCAsmLayout &Layout) { 1336 // For now, always emit the memory import, since loads and stores are not 1337 // valid without it. In the future, we could perhaps be more clever and omit 1338 // it if there are no loads or stores. 1339 wasm::WasmImport MemImport; 1340 MemImport.Module = "env"; 1341 MemImport.Field = "__linear_memory"; 1342 MemImport.Kind = wasm::WASM_EXTERNAL_MEMORY; 1343 MemImport.Memory.Flags = is64Bit() ? wasm::WASM_LIMITS_FLAG_IS_64 1344 : wasm::WASM_LIMITS_FLAG_NONE; 1345 Imports.push_back(MemImport); 1346 1347 // Populate SignatureIndices, and Imports and WasmIndices for undefined 1348 // symbols. This must be done before populating WasmIndices for defined 1349 // symbols. 1350 for (const MCSymbol &S : Asm.symbols()) { 1351 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1352 1353 // Register types for all functions, including those with private linkage 1354 // (because wasm always needs a type signature). 1355 if (WS.isFunction()) { 1356 const auto *BS = Layout.getBaseSymbol(S); 1357 if (!BS) 1358 report_fatal_error(Twine(S.getName()) + 1359 ": absolute addressing not supported!"); 1360 registerFunctionType(*cast<MCSymbolWasm>(BS)); 1361 } 1362 1363 if (WS.isTag()) 1364 registerTagType(WS); 1365 1366 if (WS.isTemporary()) 1367 continue; 1368 1369 // If the symbol is not defined in this translation unit, import it. 1370 if (!WS.isDefined() && !WS.isComdat()) { 1371 if (WS.isFunction()) { 1372 wasm::WasmImport Import; 1373 Import.Module = WS.getImportModule(); 1374 Import.Field = WS.getImportName(); 1375 Import.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1376 Import.SigIndex = getFunctionType(WS); 1377 Imports.push_back(Import); 1378 assert(WasmIndices.count(&WS) == 0); 1379 WasmIndices[&WS] = NumFunctionImports++; 1380 } else if (WS.isGlobal()) { 1381 if (WS.isWeak()) 1382 report_fatal_error("undefined global symbol cannot be weak"); 1383 1384 wasm::WasmImport Import; 1385 Import.Field = WS.getImportName(); 1386 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1387 Import.Module = WS.getImportModule(); 1388 Import.Global = WS.getGlobalType(); 1389 Imports.push_back(Import); 1390 assert(WasmIndices.count(&WS) == 0); 1391 WasmIndices[&WS] = NumGlobalImports++; 1392 } else if (WS.isTag()) { 1393 if (WS.isWeak()) 1394 report_fatal_error("undefined tag symbol cannot be weak"); 1395 1396 wasm::WasmImport Import; 1397 Import.Module = WS.getImportModule(); 1398 Import.Field = WS.getImportName(); 1399 Import.Kind = wasm::WASM_EXTERNAL_TAG; 1400 Import.SigIndex = getTagType(WS); 1401 Imports.push_back(Import); 1402 assert(WasmIndices.count(&WS) == 0); 1403 WasmIndices[&WS] = NumTagImports++; 1404 } else if (WS.isTable()) { 1405 if (WS.isWeak()) 1406 report_fatal_error("undefined table symbol cannot be weak"); 1407 1408 wasm::WasmImport Import; 1409 Import.Module = WS.getImportModule(); 1410 Import.Field = WS.getImportName(); 1411 Import.Kind = wasm::WASM_EXTERNAL_TABLE; 1412 Import.Table = WS.getTableType(); 1413 Imports.push_back(Import); 1414 assert(WasmIndices.count(&WS) == 0); 1415 WasmIndices[&WS] = NumTableImports++; 1416 } 1417 } 1418 } 1419 1420 // Add imports for GOT globals 1421 for (const MCSymbol &S : Asm.symbols()) { 1422 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1423 if (WS.isUsedInGOT()) { 1424 wasm::WasmImport Import; 1425 if (WS.isFunction()) 1426 Import.Module = "GOT.func"; 1427 else 1428 Import.Module = "GOT.mem"; 1429 Import.Field = WS.getName(); 1430 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 1431 Import.Global = {wasm::WASM_TYPE_I32, true}; 1432 Imports.push_back(Import); 1433 assert(GOTIndices.count(&WS) == 0); 1434 GOTIndices[&WS] = NumGlobalImports++; 1435 } 1436 } 1437 } 1438 1439 uint64_t WasmObjectWriter::writeObject(MCAssembler &Asm, 1440 const MCAsmLayout &Layout) { 1441 support::endian::Writer MainWriter(*OS, support::little); 1442 W = &MainWriter; 1443 if (IsSplitDwarf) { 1444 uint64_t TotalSize = writeOneObject(Asm, Layout, DwoMode::NonDwoOnly); 1445 assert(DwoOS); 1446 support::endian::Writer DwoWriter(*DwoOS, support::little); 1447 W = &DwoWriter; 1448 return TotalSize + writeOneObject(Asm, Layout, DwoMode::DwoOnly); 1449 } else { 1450 return writeOneObject(Asm, Layout, DwoMode::AllSections); 1451 } 1452 } 1453 1454 uint64_t WasmObjectWriter::writeOneObject(MCAssembler &Asm, 1455 const MCAsmLayout &Layout, 1456 DwoMode Mode) { 1457 uint64_t StartOffset = W->OS.tell(); 1458 SectionCount = 0; 1459 CustomSections.clear(); 1460 1461 LLVM_DEBUG(dbgs() << "WasmObjectWriter::writeObject\n"); 1462 1463 // Collect information from the available symbols. 1464 SmallVector<WasmFunction, 4> Functions; 1465 SmallVector<uint32_t, 4> TableElems; 1466 SmallVector<wasm::WasmImport, 4> Imports; 1467 SmallVector<wasm::WasmExport, 4> Exports; 1468 SmallVector<uint32_t, 2> TagTypes; 1469 SmallVector<wasm::WasmGlobal, 1> Globals; 1470 SmallVector<wasm::WasmTable, 1> Tables; 1471 SmallVector<wasm::WasmSymbolInfo, 4> SymbolInfos; 1472 SmallVector<std::pair<uint16_t, uint32_t>, 2> InitFuncs; 1473 std::map<StringRef, std::vector<WasmComdatEntry>> Comdats; 1474 uint64_t DataSize = 0; 1475 if (Mode != DwoMode::DwoOnly) { 1476 prepareImports(Imports, Asm, Layout); 1477 } 1478 1479 // Populate DataSegments and CustomSections, which must be done before 1480 // populating DataLocations. 1481 for (MCSection &Sec : Asm) { 1482 auto &Section = static_cast<MCSectionWasm &>(Sec); 1483 StringRef SectionName = Section.getName(); 1484 1485 if (Mode == DwoMode::NonDwoOnly && isDwoSection(Sec)) 1486 continue; 1487 if (Mode == DwoMode::DwoOnly && !isDwoSection(Sec)) 1488 continue; 1489 1490 LLVM_DEBUG(dbgs() << "Processing Section " << SectionName << " group " 1491 << Section.getGroup() << "\n";); 1492 1493 // .init_array sections are handled specially elsewhere. 1494 if (SectionName.startswith(".init_array")) 1495 continue; 1496 1497 // Code is handled separately 1498 if (Section.getKind().isText()) 1499 continue; 1500 1501 if (Section.isWasmData()) { 1502 uint32_t SegmentIndex = DataSegments.size(); 1503 DataSize = alignTo(DataSize, Section.getAlignment()); 1504 DataSegments.emplace_back(); 1505 WasmDataSegment &Segment = DataSegments.back(); 1506 Segment.Name = SectionName; 1507 Segment.InitFlags = Section.getPassive() 1508 ? (uint32_t)wasm::WASM_DATA_SEGMENT_IS_PASSIVE 1509 : 0; 1510 Segment.Offset = DataSize; 1511 Segment.Section = &Section; 1512 addData(Segment.Data, Section); 1513 Segment.Alignment = Log2_32(Section.getAlignment()); 1514 Segment.LinkingFlags = Section.getSegmentFlags(); 1515 DataSize += Segment.Data.size(); 1516 Section.setSegmentIndex(SegmentIndex); 1517 1518 if (const MCSymbolWasm *C = Section.getGroup()) { 1519 Comdats[C->getName()].emplace_back( 1520 WasmComdatEntry{wasm::WASM_COMDAT_DATA, SegmentIndex}); 1521 } 1522 } else { 1523 // Create custom sections 1524 assert(Sec.getKind().isMetadata()); 1525 1526 StringRef Name = SectionName; 1527 1528 // For user-defined custom sections, strip the prefix 1529 if (Name.startswith(".custom_section.")) 1530 Name = Name.substr(strlen(".custom_section.")); 1531 1532 MCSymbol *Begin = Sec.getBeginSymbol(); 1533 if (Begin) { 1534 assert(WasmIndices.count(cast<MCSymbolWasm>(Begin)) == 0); 1535 WasmIndices[cast<MCSymbolWasm>(Begin)] = CustomSections.size(); 1536 } 1537 1538 // Separate out the producers and target features sections 1539 if (Name == "producers") { 1540 ProducersSection = std::make_unique<WasmCustomSection>(Name, &Section); 1541 continue; 1542 } 1543 if (Name == "target_features") { 1544 TargetFeaturesSection = 1545 std::make_unique<WasmCustomSection>(Name, &Section); 1546 continue; 1547 } 1548 1549 // Custom sections can also belong to COMDAT groups. In this case the 1550 // decriptor's "index" field is the section index (in the final object 1551 // file), but that is not known until after layout, so it must be fixed up 1552 // later 1553 if (const MCSymbolWasm *C = Section.getGroup()) { 1554 Comdats[C->getName()].emplace_back( 1555 WasmComdatEntry{wasm::WASM_COMDAT_SECTION, 1556 static_cast<uint32_t>(CustomSections.size())}); 1557 } 1558 1559 CustomSections.emplace_back(Name, &Section); 1560 } 1561 } 1562 1563 if (Mode != DwoMode::DwoOnly) { 1564 // Populate WasmIndices and DataLocations for defined symbols. 1565 for (const MCSymbol &S : Asm.symbols()) { 1566 // Ignore unnamed temporary symbols, which aren't ever exported, imported, 1567 // or used in relocations. 1568 if (S.isTemporary() && S.getName().empty()) 1569 continue; 1570 1571 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1572 LLVM_DEBUG(dbgs() 1573 << "MCSymbol: " 1574 << toString(WS.getType().getValueOr(wasm::WASM_SYMBOL_TYPE_DATA)) 1575 << " '" << S << "'" 1576 << " isDefined=" << S.isDefined() << " isExternal=" 1577 << S.isExternal() << " isTemporary=" << S.isTemporary() 1578 << " isWeak=" << WS.isWeak() << " isHidden=" << WS.isHidden() 1579 << " isVariable=" << WS.isVariable() << "\n"); 1580 1581 if (WS.isVariable()) 1582 continue; 1583 if (WS.isComdat() && !WS.isDefined()) 1584 continue; 1585 1586 if (WS.isFunction()) { 1587 unsigned Index; 1588 if (WS.isDefined()) { 1589 if (WS.getOffset() != 0) 1590 report_fatal_error( 1591 "function sections must contain one function each"); 1592 1593 if (WS.getSize() == nullptr) 1594 report_fatal_error( 1595 "function symbols must have a size set with .size"); 1596 1597 // A definition. Write out the function body. 1598 Index = NumFunctionImports + Functions.size(); 1599 WasmFunction Func; 1600 Func.SigIndex = getFunctionType(WS); 1601 Func.Sym = &WS; 1602 assert(WasmIndices.count(&WS) == 0); 1603 WasmIndices[&WS] = Index; 1604 Functions.push_back(Func); 1605 1606 auto &Section = static_cast<MCSectionWasm &>(WS.getSection()); 1607 if (const MCSymbolWasm *C = Section.getGroup()) { 1608 Comdats[C->getName()].emplace_back( 1609 WasmComdatEntry{wasm::WASM_COMDAT_FUNCTION, Index}); 1610 } 1611 1612 if (WS.hasExportName()) { 1613 wasm::WasmExport Export; 1614 Export.Name = WS.getExportName(); 1615 Export.Kind = wasm::WASM_EXTERNAL_FUNCTION; 1616 Export.Index = Index; 1617 Exports.push_back(Export); 1618 } 1619 } else { 1620 // An import; the index was assigned above. 1621 Index = WasmIndices.find(&WS)->second; 1622 } 1623 1624 LLVM_DEBUG(dbgs() << " -> function index: " << Index << "\n"); 1625 1626 } else if (WS.isData()) { 1627 if (!isInSymtab(WS)) 1628 continue; 1629 1630 if (!WS.isDefined()) { 1631 LLVM_DEBUG(dbgs() << " -> segment index: -1" 1632 << "\n"); 1633 continue; 1634 } 1635 1636 if (!WS.getSize()) 1637 report_fatal_error("data symbols must have a size set with .size: " + 1638 WS.getName()); 1639 1640 int64_t Size = 0; 1641 if (!WS.getSize()->evaluateAsAbsolute(Size, Layout)) 1642 report_fatal_error(".size expression must be evaluatable"); 1643 1644 auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection()); 1645 if (!DataSection.isWasmData()) 1646 report_fatal_error("data symbols must live in a data section: " + 1647 WS.getName()); 1648 1649 // For each data symbol, export it in the symtab as a reference to the 1650 // corresponding Wasm data segment. 1651 wasm::WasmDataReference Ref = wasm::WasmDataReference{ 1652 DataSection.getSegmentIndex(), Layout.getSymbolOffset(WS), 1653 static_cast<uint64_t>(Size)}; 1654 assert(DataLocations.count(&WS) == 0); 1655 DataLocations[&WS] = Ref; 1656 LLVM_DEBUG(dbgs() << " -> segment index: " << Ref.Segment << "\n"); 1657 1658 } else if (WS.isGlobal()) { 1659 // A "true" Wasm global (currently just __stack_pointer) 1660 if (WS.isDefined()) { 1661 wasm::WasmGlobal Global; 1662 Global.Type = WS.getGlobalType(); 1663 Global.Index = NumGlobalImports + Globals.size(); 1664 switch (Global.Type.Type) { 1665 case wasm::WASM_TYPE_I32: 1666 Global.InitExpr.Opcode = wasm::WASM_OPCODE_I32_CONST; 1667 break; 1668 case wasm::WASM_TYPE_I64: 1669 Global.InitExpr.Opcode = wasm::WASM_OPCODE_I64_CONST; 1670 break; 1671 case wasm::WASM_TYPE_F32: 1672 Global.InitExpr.Opcode = wasm::WASM_OPCODE_F32_CONST; 1673 break; 1674 case wasm::WASM_TYPE_F64: 1675 Global.InitExpr.Opcode = wasm::WASM_OPCODE_F64_CONST; 1676 break; 1677 case wasm::WASM_TYPE_EXTERNREF: 1678 Global.InitExpr.Opcode = wasm::WASM_OPCODE_REF_NULL; 1679 break; 1680 default: 1681 llvm_unreachable("unexpected type"); 1682 } 1683 assert(WasmIndices.count(&WS) == 0); 1684 WasmIndices[&WS] = Global.Index; 1685 Globals.push_back(Global); 1686 } else { 1687 // An import; the index was assigned above 1688 LLVM_DEBUG(dbgs() << " -> global index: " 1689 << WasmIndices.find(&WS)->second << "\n"); 1690 } 1691 } else if (WS.isTable()) { 1692 if (WS.isDefined()) { 1693 wasm::WasmTable Table; 1694 Table.Index = NumTableImports + Tables.size(); 1695 Table.Type = WS.getTableType(); 1696 assert(WasmIndices.count(&WS) == 0); 1697 WasmIndices[&WS] = Table.Index; 1698 Tables.push_back(Table); 1699 } 1700 LLVM_DEBUG(dbgs() << " -> table index: " 1701 << WasmIndices.find(&WS)->second << "\n"); 1702 } else if (WS.isTag()) { 1703 // C++ exception symbol (__cpp_exception) or longjmp symbol 1704 // (__c_longjmp) 1705 unsigned Index; 1706 if (WS.isDefined()) { 1707 Index = NumTagImports + TagTypes.size(); 1708 uint32_t SigIndex = getTagType(WS); 1709 assert(WasmIndices.count(&WS) == 0); 1710 WasmIndices[&WS] = Index; 1711 TagTypes.push_back(SigIndex); 1712 } else { 1713 // An import; the index was assigned above. 1714 assert(WasmIndices.count(&WS) > 0); 1715 } 1716 LLVM_DEBUG(dbgs() << " -> tag index: " << WasmIndices.find(&WS)->second 1717 << "\n"); 1718 1719 } else { 1720 assert(WS.isSection()); 1721 } 1722 } 1723 1724 // Populate WasmIndices and DataLocations for aliased symbols. We need to 1725 // process these in a separate pass because we need to have processed the 1726 // target of the alias before the alias itself and the symbols are not 1727 // necessarily ordered in this way. 1728 for (const MCSymbol &S : Asm.symbols()) { 1729 if (!S.isVariable()) 1730 continue; 1731 1732 assert(S.isDefined()); 1733 1734 const auto *BS = Layout.getBaseSymbol(S); 1735 if (!BS) 1736 report_fatal_error(Twine(S.getName()) + 1737 ": absolute addressing not supported!"); 1738 const MCSymbolWasm *Base = cast<MCSymbolWasm>(BS); 1739 1740 // Find the target symbol of this weak alias and export that index 1741 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1742 LLVM_DEBUG(dbgs() << WS.getName() << ": weak alias of '" << *Base 1743 << "'\n"); 1744 1745 if (Base->isFunction()) { 1746 assert(WasmIndices.count(Base) > 0); 1747 uint32_t WasmIndex = WasmIndices.find(Base)->second; 1748 assert(WasmIndices.count(&WS) == 0); 1749 WasmIndices[&WS] = WasmIndex; 1750 LLVM_DEBUG(dbgs() << " -> index:" << WasmIndex << "\n"); 1751 } else if (Base->isData()) { 1752 auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection()); 1753 uint64_t Offset = Layout.getSymbolOffset(S); 1754 int64_t Size = 0; 1755 // For data symbol alias we use the size of the base symbol as the 1756 // size of the alias. When an offset from the base is involved this 1757 // can result in a offset + size goes past the end of the data section 1758 // which out object format doesn't support. So we must clamp it. 1759 if (!Base->getSize()->evaluateAsAbsolute(Size, Layout)) 1760 report_fatal_error(".size expression must be evaluatable"); 1761 const WasmDataSegment &Segment = 1762 DataSegments[DataSection.getSegmentIndex()]; 1763 Size = 1764 std::min(static_cast<uint64_t>(Size), Segment.Data.size() - Offset); 1765 wasm::WasmDataReference Ref = wasm::WasmDataReference{ 1766 DataSection.getSegmentIndex(), 1767 static_cast<uint32_t>(Layout.getSymbolOffset(S)), 1768 static_cast<uint32_t>(Size)}; 1769 DataLocations[&WS] = Ref; 1770 LLVM_DEBUG(dbgs() << " -> index:" << Ref.Segment << "\n"); 1771 } else { 1772 report_fatal_error("don't yet support global/tag aliases"); 1773 } 1774 } 1775 } 1776 1777 // Finally, populate the symbol table itself, in its "natural" order. 1778 for (const MCSymbol &S : Asm.symbols()) { 1779 const auto &WS = static_cast<const MCSymbolWasm &>(S); 1780 if (!isInSymtab(WS)) { 1781 WS.setIndex(InvalidIndex); 1782 continue; 1783 } 1784 LLVM_DEBUG(dbgs() << "adding to symtab: " << WS << "\n"); 1785 1786 uint32_t Flags = 0; 1787 if (WS.isWeak()) 1788 Flags |= wasm::WASM_SYMBOL_BINDING_WEAK; 1789 if (WS.isHidden()) 1790 Flags |= wasm::WASM_SYMBOL_VISIBILITY_HIDDEN; 1791 if (!WS.isExternal() && WS.isDefined()) 1792 Flags |= wasm::WASM_SYMBOL_BINDING_LOCAL; 1793 if (WS.isUndefined()) 1794 Flags |= wasm::WASM_SYMBOL_UNDEFINED; 1795 if (WS.isNoStrip()) { 1796 Flags |= wasm::WASM_SYMBOL_NO_STRIP; 1797 if (isEmscripten()) { 1798 Flags |= wasm::WASM_SYMBOL_EXPORTED; 1799 } 1800 } 1801 if (WS.hasImportName()) 1802 Flags |= wasm::WASM_SYMBOL_EXPLICIT_NAME; 1803 if (WS.hasExportName()) 1804 Flags |= wasm::WASM_SYMBOL_EXPORTED; 1805 if (WS.isTLS()) 1806 Flags |= wasm::WASM_SYMBOL_TLS; 1807 1808 wasm::WasmSymbolInfo Info; 1809 Info.Name = WS.getName(); 1810 Info.Kind = WS.getType().getValueOr(wasm::WASM_SYMBOL_TYPE_DATA); 1811 Info.Flags = Flags; 1812 if (!WS.isData()) { 1813 assert(WasmIndices.count(&WS) > 0); 1814 Info.ElementIndex = WasmIndices.find(&WS)->second; 1815 } else if (WS.isDefined()) { 1816 assert(DataLocations.count(&WS) > 0); 1817 Info.DataRef = DataLocations.find(&WS)->second; 1818 } 1819 WS.setIndex(SymbolInfos.size()); 1820 SymbolInfos.emplace_back(Info); 1821 } 1822 1823 { 1824 auto HandleReloc = [&](const WasmRelocationEntry &Rel) { 1825 // Functions referenced by a relocation need to put in the table. This is 1826 // purely to make the object file's provisional values readable, and is 1827 // ignored by the linker, which re-calculates the relocations itself. 1828 if (Rel.Type != wasm::R_WASM_TABLE_INDEX_I32 && 1829 Rel.Type != wasm::R_WASM_TABLE_INDEX_I64 && 1830 Rel.Type != wasm::R_WASM_TABLE_INDEX_SLEB && 1831 Rel.Type != wasm::R_WASM_TABLE_INDEX_SLEB64 && 1832 Rel.Type != wasm::R_WASM_TABLE_INDEX_REL_SLEB && 1833 Rel.Type != wasm::R_WASM_TABLE_INDEX_REL_SLEB64) 1834 return; 1835 assert(Rel.Symbol->isFunction()); 1836 const MCSymbolWasm *Base = 1837 cast<MCSymbolWasm>(Layout.getBaseSymbol(*Rel.Symbol)); 1838 uint32_t FunctionIndex = WasmIndices.find(Base)->second; 1839 uint32_t TableIndex = TableElems.size() + InitialTableOffset; 1840 if (TableIndices.try_emplace(Base, TableIndex).second) { 1841 LLVM_DEBUG(dbgs() << " -> adding " << Base->getName() 1842 << " to table: " << TableIndex << "\n"); 1843 TableElems.push_back(FunctionIndex); 1844 registerFunctionType(*Base); 1845 } 1846 }; 1847 1848 for (const WasmRelocationEntry &RelEntry : CodeRelocations) 1849 HandleReloc(RelEntry); 1850 for (const WasmRelocationEntry &RelEntry : DataRelocations) 1851 HandleReloc(RelEntry); 1852 } 1853 1854 // Translate .init_array section contents into start functions. 1855 for (const MCSection &S : Asm) { 1856 const auto &WS = static_cast<const MCSectionWasm &>(S); 1857 if (WS.getName().startswith(".fini_array")) 1858 report_fatal_error(".fini_array sections are unsupported"); 1859 if (!WS.getName().startswith(".init_array")) 1860 continue; 1861 if (WS.getFragmentList().empty()) 1862 continue; 1863 1864 // init_array is expected to contain a single non-empty data fragment 1865 if (WS.getFragmentList().size() != 3) 1866 report_fatal_error("only one .init_array section fragment supported"); 1867 1868 auto IT = WS.begin(); 1869 const MCFragment &EmptyFrag = *IT; 1870 if (EmptyFrag.getKind() != MCFragment::FT_Data) 1871 report_fatal_error(".init_array section should be aligned"); 1872 1873 IT = std::next(IT); 1874 const MCFragment &AlignFrag = *IT; 1875 if (AlignFrag.getKind() != MCFragment::FT_Align) 1876 report_fatal_error(".init_array section should be aligned"); 1877 if (cast<MCAlignFragment>(AlignFrag).getAlignment() != (is64Bit() ? 8 : 4)) 1878 report_fatal_error(".init_array section should be aligned for pointers"); 1879 1880 const MCFragment &Frag = *std::next(IT); 1881 if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data) 1882 report_fatal_error("only data supported in .init_array section"); 1883 1884 uint16_t Priority = UINT16_MAX; 1885 unsigned PrefixLength = strlen(".init_array"); 1886 if (WS.getName().size() > PrefixLength) { 1887 if (WS.getName()[PrefixLength] != '.') 1888 report_fatal_error( 1889 ".init_array section priority should start with '.'"); 1890 if (WS.getName().substr(PrefixLength + 1).getAsInteger(10, Priority)) 1891 report_fatal_error("invalid .init_array section priority"); 1892 } 1893 const auto &DataFrag = cast<MCDataFragment>(Frag); 1894 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 1895 for (const uint8_t * 1896 P = (const uint8_t *)Contents.data(), 1897 *End = (const uint8_t *)Contents.data() + Contents.size(); 1898 P != End; ++P) { 1899 if (*P != 0) 1900 report_fatal_error("non-symbolic data in .init_array section"); 1901 } 1902 for (const MCFixup &Fixup : DataFrag.getFixups()) { 1903 assert(Fixup.getKind() == 1904 MCFixup::getKindForSize(is64Bit() ? 8 : 4, false)); 1905 const MCExpr *Expr = Fixup.getValue(); 1906 auto *SymRef = dyn_cast<MCSymbolRefExpr>(Expr); 1907 if (!SymRef) 1908 report_fatal_error("fixups in .init_array should be symbol references"); 1909 const auto &TargetSym = cast<const MCSymbolWasm>(SymRef->getSymbol()); 1910 if (TargetSym.getIndex() == InvalidIndex) 1911 report_fatal_error("symbols in .init_array should exist in symtab"); 1912 if (!TargetSym.isFunction()) 1913 report_fatal_error("symbols in .init_array should be for functions"); 1914 InitFuncs.push_back( 1915 std::make_pair(Priority, TargetSym.getIndex())); 1916 } 1917 } 1918 1919 // Write out the Wasm header. 1920 writeHeader(Asm); 1921 1922 uint32_t CodeSectionIndex, DataSectionIndex; 1923 if (Mode != DwoMode::DwoOnly) { 1924 writeTypeSection(Signatures); 1925 writeImportSection(Imports, DataSize, TableElems.size()); 1926 writeFunctionSection(Functions); 1927 writeTableSection(Tables); 1928 // Skip the "memory" section; we import the memory instead. 1929 writeTagSection(TagTypes); 1930 writeGlobalSection(Globals); 1931 writeExportSection(Exports); 1932 const MCSymbol *IndirectFunctionTable = 1933 Asm.getContext().lookupSymbol("__indirect_function_table"); 1934 writeElemSection(cast_or_null<const MCSymbolWasm>(IndirectFunctionTable), 1935 TableElems); 1936 writeDataCountSection(); 1937 1938 CodeSectionIndex = writeCodeSection(Asm, Layout, Functions); 1939 DataSectionIndex = writeDataSection(Layout); 1940 } 1941 1942 // The Sections in the COMDAT list have placeholder indices (their index among 1943 // custom sections, rather than among all sections). Fix them up here. 1944 for (auto &Group : Comdats) { 1945 for (auto &Entry : Group.second) { 1946 if (Entry.Kind == wasm::WASM_COMDAT_SECTION) { 1947 Entry.Index += SectionCount; 1948 } 1949 } 1950 } 1951 for (auto &CustomSection : CustomSections) 1952 writeCustomSection(CustomSection, Asm, Layout); 1953 1954 if (Mode != DwoMode::DwoOnly) { 1955 writeLinkingMetaDataSection(SymbolInfos, InitFuncs, Comdats); 1956 1957 writeRelocSection(CodeSectionIndex, "CODE", CodeRelocations); 1958 writeRelocSection(DataSectionIndex, "DATA", DataRelocations); 1959 } 1960 writeCustomRelocSections(); 1961 if (ProducersSection) 1962 writeCustomSection(*ProducersSection, Asm, Layout); 1963 if (TargetFeaturesSection) 1964 writeCustomSection(*TargetFeaturesSection, Asm, Layout); 1965 1966 // TODO: Translate the .comment section to the output. 1967 return W->OS.tell() - StartOffset; 1968 } 1969 1970 std::unique_ptr<MCObjectWriter> 1971 llvm::createWasmObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 1972 raw_pwrite_stream &OS) { 1973 return std::make_unique<WasmObjectWriter>(std::move(MOTW), OS); 1974 } 1975 1976 std::unique_ptr<MCObjectWriter> 1977 llvm::createWasmDwoObjectWriter(std::unique_ptr<MCWasmObjectTargetWriter> MOTW, 1978 raw_pwrite_stream &OS, 1979 raw_pwrite_stream &DwoOS) { 1980 return std::make_unique<WasmObjectWriter>(std::move(MOTW), OS, DwoOS); 1981 } 1982