1 //===- lib/MC/WasmObjectWriter.cpp - Wasm File Writer ---------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements Wasm object file writer information. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/ADT/STLExtras.h" 15 #include "llvm/ADT/SmallPtrSet.h" 16 #include "llvm/MC/MCAsmBackend.h" 17 #include "llvm/MC/MCAsmInfo.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/MCObjectFileInfo.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/ErrorHandling.h" 32 #include "llvm/Support/LEB128.h" 33 #include "llvm/Support/StringSaver.h" 34 #include "llvm/Support/Wasm.h" 35 #include <vector> 36 37 using namespace llvm; 38 39 #undef DEBUG_TYPE 40 #define DEBUG_TYPE "reloc-info" 41 42 namespace { 43 // For patching purposes, we need to remember where each section starts, both 44 // for patching up the section size field, and for patching up references to 45 // locations within the section. 46 struct SectionBookkeeping { 47 // Where the size of the section is written. 48 uint64_t SizeOffset; 49 // Where the contents of the section starts (after the header). 50 uint64_t ContentsOffset; 51 }; 52 53 class WasmObjectWriter : public MCObjectWriter { 54 /// Helper struct for containing some precomputed information on symbols. 55 struct WasmSymbolData { 56 const MCSymbolWasm *Symbol; 57 StringRef Name; 58 59 // Support lexicographic sorting. 60 bool operator<(const WasmSymbolData &RHS) const { return Name < RHS.Name; } 61 }; 62 63 /// The target specific Wasm writer instance. 64 std::unique_ptr<MCWasmObjectTargetWriter> TargetObjectWriter; 65 66 // Relocations for fixing up references in the code section. 67 std::vector<WasmRelocationEntry> CodeRelocations; 68 69 // Relocations for fixing up references in the data section. 70 std::vector<WasmRelocationEntry> DataRelocations; 71 72 // Fixups for call_indirect type indices. 73 std::vector<WasmRelocationEntry> TypeIndexFixups; 74 75 // Index values to use for fixing up call_indirect type indices. 76 std::vector<uint32_t> TypeIndexFixupTypes; 77 78 // TargetObjectWriter wrappers. 79 bool is64Bit() const { return TargetObjectWriter->is64Bit(); } 80 unsigned getRelocType(MCContext &Ctx, const MCValue &Target, 81 const MCFixup &Fixup, bool IsPCRel) const { 82 return TargetObjectWriter->getRelocType(Ctx, Target, Fixup, IsPCRel); 83 } 84 85 void startSection(SectionBookkeeping &Section, unsigned SectionId, 86 const char *Name = nullptr); 87 void endSection(SectionBookkeeping &Section); 88 89 public: 90 WasmObjectWriter(MCWasmObjectTargetWriter *MOTW, raw_pwrite_stream &OS) 91 : MCObjectWriter(OS, /*IsLittleEndian=*/true), TargetObjectWriter(MOTW) {} 92 93 private: 94 void reset() override { 95 MCObjectWriter::reset(); 96 } 97 98 ~WasmObjectWriter() override; 99 100 void writeHeader(const MCAssembler &Asm); 101 102 void writeValueType(wasm::ValType Ty) { 103 encodeSLEB128(int32_t(Ty), getStream()); 104 } 105 106 void recordRelocation(MCAssembler &Asm, const MCAsmLayout &Layout, 107 const MCFragment *Fragment, const MCFixup &Fixup, 108 MCValue Target, bool &IsPCRel, 109 uint64_t &FixedValue) override; 110 111 void executePostLayoutBinding(MCAssembler &Asm, 112 const MCAsmLayout &Layout) override; 113 114 void writeObject(MCAssembler &Asm, const MCAsmLayout &Layout) override; 115 }; 116 } // end anonymous namespace 117 118 WasmObjectWriter::~WasmObjectWriter() {} 119 120 // Return the padding size to write a 32-bit value into a 5-byte ULEB128. 121 static unsigned PaddingFor5ByteULEB128(uint32_t X) { 122 return X == 0 ? 4 : (4u - (31u - countLeadingZeros(X)) / 7u); 123 } 124 125 // Return the padding size to write a 32-bit value into a 5-byte SLEB128. 126 static unsigned PaddingFor5ByteSLEB128(int32_t X) { 127 return 5 - getSLEB128Size(X); 128 } 129 130 // Write out a section header and a patchable section size field. 131 void WasmObjectWriter::startSection(SectionBookkeeping &Section, 132 unsigned SectionId, 133 const char *Name) { 134 assert((Name != nullptr) == (SectionId == wasm::WASM_SEC_CUSTOM) && 135 "Only custom sections can have names"); 136 137 encodeULEB128(SectionId, getStream()); 138 139 Section.SizeOffset = getStream().tell(); 140 141 // The section size. We don't know the size yet, so reserve enough space 142 // for any 32-bit value; we'll patch it later. 143 encodeULEB128(UINT32_MAX, getStream()); 144 145 // The position where the section starts, for measuring its size. 146 Section.ContentsOffset = getStream().tell(); 147 148 // Custom sections in wasm also have a string identifier. 149 if (SectionId == wasm::WASM_SEC_CUSTOM) { 150 encodeULEB128(strlen(Name), getStream()); 151 writeBytes(Name); 152 } 153 } 154 155 // Now that the section is complete and we know how big it is, patch up the 156 // section size field at the start of the section. 157 void WasmObjectWriter::endSection(SectionBookkeeping &Section) { 158 uint64_t Size = getStream().tell() - Section.ContentsOffset; 159 if (uint32_t(Size) != Size) 160 report_fatal_error("section size does not fit in a uint32_t"); 161 162 unsigned Padding = PaddingFor5ByteULEB128(Size); 163 164 // Write the final section size to the payload_len field, which follows 165 // the section id byte. 166 uint8_t Buffer[16]; 167 unsigned SizeLen = encodeULEB128(Size, Buffer, Padding); 168 assert(SizeLen == 5); 169 getStream().pwrite((char *)Buffer, SizeLen, Section.SizeOffset); 170 } 171 172 // Emit the Wasm header. 173 void WasmObjectWriter::writeHeader(const MCAssembler &Asm) { 174 writeBytes(StringRef(wasm::WasmMagic, sizeof(wasm::WasmMagic))); 175 writeLE32(wasm::WasmVersion); 176 } 177 178 void WasmObjectWriter::executePostLayoutBinding(MCAssembler &Asm, 179 const MCAsmLayout &Layout) { 180 } 181 182 void WasmObjectWriter::recordRelocation(MCAssembler &Asm, 183 const MCAsmLayout &Layout, 184 const MCFragment *Fragment, 185 const MCFixup &Fixup, MCValue Target, 186 bool &IsPCRel, uint64_t &FixedValue) { 187 MCSectionWasm &FixupSection = cast<MCSectionWasm>(*Fragment->getParent()); 188 uint64_t C = Target.getConstant(); 189 uint64_t FixupOffset = Layout.getFragmentOffset(Fragment) + Fixup.getOffset(); 190 MCContext &Ctx = Asm.getContext(); 191 192 if (const MCSymbolRefExpr *RefB = Target.getSymB()) { 193 assert(RefB->getKind() == MCSymbolRefExpr::VK_None && 194 "Should not have constructed this"); 195 196 // Let A, B and C being the components of Target and R be the location of 197 // the fixup. If the fixup is not pcrel, we want to compute (A - B + C). 198 // If it is pcrel, we want to compute (A - B + C - R). 199 200 // In general, Wasm has no relocations for -B. It can only represent (A + C) 201 // or (A + C - R). If B = R + K and the relocation is not pcrel, we can 202 // replace B to implement it: (A - R - K + C) 203 if (IsPCRel) { 204 Ctx.reportError( 205 Fixup.getLoc(), 206 "No relocation available to represent this relative expression"); 207 return; 208 } 209 210 const auto &SymB = cast<MCSymbolWasm>(RefB->getSymbol()); 211 212 if (SymB.isUndefined()) { 213 Ctx.reportError(Fixup.getLoc(), 214 Twine("symbol '") + SymB.getName() + 215 "' can not be undefined in a subtraction expression"); 216 return; 217 } 218 219 assert(!SymB.isAbsolute() && "Should have been folded"); 220 const MCSection &SecB = SymB.getSection(); 221 if (&SecB != &FixupSection) { 222 Ctx.reportError(Fixup.getLoc(), 223 "Cannot represent a difference across sections"); 224 return; 225 } 226 227 uint64_t SymBOffset = Layout.getSymbolOffset(SymB); 228 uint64_t K = SymBOffset - FixupOffset; 229 IsPCRel = true; 230 C -= K; 231 } 232 233 // We either rejected the fixup or folded B into C at this point. 234 const MCSymbolRefExpr *RefA = Target.getSymA(); 235 const auto *SymA = RefA ? cast<MCSymbolWasm>(&RefA->getSymbol()) : nullptr; 236 237 bool ViaWeakRef = false; 238 if (SymA && SymA->isVariable()) { 239 const MCExpr *Expr = SymA->getVariableValue(); 240 if (const auto *Inner = dyn_cast<MCSymbolRefExpr>(Expr)) { 241 if (Inner->getKind() == MCSymbolRefExpr::VK_WEAKREF) { 242 SymA = cast<MCSymbolWasm>(&Inner->getSymbol()); 243 ViaWeakRef = true; 244 } 245 } 246 } 247 248 // Put any constant offset in an addend. Offsets can be negative, and 249 // LLVM expects wrapping, in contrast to wasm's immediates which can't 250 // be negative and don't wrap. 251 FixedValue = 0; 252 253 if (SymA) { 254 if (ViaWeakRef) 255 llvm_unreachable("weakref used in reloc not yet implemented"); 256 else 257 SymA->setUsedInReloc(); 258 } 259 260 if (RefA) { 261 if (RefA->getKind() == MCSymbolRefExpr::VK_WebAssembly_TYPEINDEX) { 262 assert(C == 0); 263 WasmRelocationEntry Rec(FixupOffset, SymA, C, 264 wasm::R_WEBASSEMBLY_TYPE_INDEX_LEB, 265 &FixupSection); 266 TypeIndexFixups.push_back(Rec); 267 return; 268 } 269 } 270 271 unsigned Type = getRelocType(Ctx, Target, Fixup, IsPCRel); 272 273 WasmRelocationEntry Rec(FixupOffset, SymA, C, Type, &FixupSection); 274 275 if (FixupSection.hasInstructions()) 276 CodeRelocations.push_back(Rec); 277 else 278 DataRelocations.push_back(Rec); 279 } 280 281 namespace { 282 283 // The signature of a wasm function, in a struct capable of being used as a 284 // DenseMap key. 285 struct WasmFunctionType { 286 // Support empty and tombstone instances, needed by DenseMap. 287 enum { Plain, Empty, Tombstone } State; 288 289 // The return types of the function. 290 SmallVector<wasm::ValType, 1> Returns; 291 292 // The parameter types of the function. 293 SmallVector<wasm::ValType, 4> Params; 294 295 WasmFunctionType() : State(Plain) {} 296 297 bool operator==(const WasmFunctionType &Other) const { 298 return State == Other.State && Returns == Other.Returns && 299 Params == Other.Params; 300 } 301 }; 302 303 // Traits for using WasmFunctionType in a DenseMap. 304 struct WasmFunctionTypeDenseMapInfo { 305 static WasmFunctionType getEmptyKey() { 306 WasmFunctionType FuncTy; 307 FuncTy.State = WasmFunctionType::Empty; 308 return FuncTy; 309 } 310 static WasmFunctionType getTombstoneKey() { 311 WasmFunctionType FuncTy; 312 FuncTy.State = WasmFunctionType::Tombstone; 313 return FuncTy; 314 } 315 static unsigned getHashValue(const WasmFunctionType &FuncTy) { 316 uintptr_t Value = FuncTy.State; 317 for (wasm::ValType Ret : FuncTy.Returns) 318 Value += DenseMapInfo<int32_t>::getHashValue(int32_t(Ret)); 319 for (wasm::ValType Param : FuncTy.Params) 320 Value += DenseMapInfo<int32_t>::getHashValue(int32_t(Param)); 321 return Value; 322 } 323 static bool isEqual(const WasmFunctionType &LHS, 324 const WasmFunctionType &RHS) { 325 return LHS == RHS; 326 } 327 }; 328 329 // A wasm import to be written into the import section. 330 struct WasmImport { 331 StringRef ModuleName; 332 StringRef FieldName; 333 unsigned Kind; 334 int32_t Type; 335 }; 336 337 // A wasm function to be written into the function section. 338 struct WasmFunction { 339 int32_t Type; 340 const MCSymbolWasm *Sym; 341 }; 342 343 // A wasm export to be written into the export section. 344 struct WasmExport { 345 StringRef FieldName; 346 unsigned Kind; 347 uint32_t Index; 348 }; 349 350 // A wasm global to be written into the global section. 351 struct WasmGlobal { 352 wasm::ValType Type; 353 bool IsMutable; 354 bool HasImport; 355 uint64_t InitialValue; 356 uint32_t ImportIndex; 357 }; 358 359 } // end anonymous namespace 360 361 // Write X as an (unsigned) LEB value at offset Offset in Stream, padded 362 // to allow patching. 363 static void 364 WritePatchableLEB(raw_pwrite_stream &Stream, uint32_t X, uint64_t Offset) { 365 uint8_t Buffer[5]; 366 unsigned Padding = PaddingFor5ByteULEB128(X); 367 unsigned SizeLen = encodeULEB128(X, Buffer, Padding); 368 assert(SizeLen == 5); 369 Stream.pwrite((char *)Buffer, SizeLen, Offset); 370 } 371 372 // Write X as an signed LEB value at offset Offset in Stream, padded 373 // to allow patching. 374 static void 375 WritePatchableSLEB(raw_pwrite_stream &Stream, int32_t X, uint64_t Offset) { 376 uint8_t Buffer[5]; 377 unsigned Padding = PaddingFor5ByteSLEB128(X); 378 unsigned SizeLen = encodeSLEB128(X, Buffer, Padding); 379 assert(SizeLen == 5); 380 Stream.pwrite((char *)Buffer, SizeLen, Offset); 381 } 382 383 // Write X as a plain integer value at offset Offset in Stream. 384 static void WriteI32(raw_pwrite_stream &Stream, uint32_t X, uint64_t Offset) { 385 uint8_t Buffer[4]; 386 support::endian::write32le(Buffer, X); 387 Stream.pwrite((char *)Buffer, sizeof(Buffer), Offset); 388 } 389 390 // Compute a value to write into the code at the location covered 391 // by RelEntry. This value isn't used by the static linker, since 392 // we have addends; it just serves to make the code more readable 393 // and to make standalone wasm modules directly usable. 394 static uint32_t ProvisionalValue(const WasmRelocationEntry &RelEntry) { 395 const MCSymbolWasm *Sym = RelEntry.Symbol; 396 397 // For undefined symbols, use a hopefully invalid value. 398 if (!Sym->isDefined(false)) 399 return UINT32_MAX; 400 401 MCSectionWasm &Section = 402 cast<MCSectionWasm>(RelEntry.Symbol->getSection(false)); 403 uint64_t Address = Section.getSectionOffset() + RelEntry.Addend; 404 405 // Ignore overflow. LLVM allows address arithmetic to silently wrap. 406 uint32_t Value = Address; 407 408 return Value; 409 } 410 411 // Apply the portions of the relocation records that we can handle ourselves 412 // directly. 413 static void ApplyRelocations( 414 ArrayRef<WasmRelocationEntry> Relocations, 415 raw_pwrite_stream &Stream, 416 DenseMap<const MCSymbolWasm *, uint32_t> &SymbolIndices, 417 uint64_t ContentsOffset) 418 { 419 for (const WasmRelocationEntry &RelEntry : Relocations) { 420 uint64_t Offset = ContentsOffset + 421 RelEntry.FixupSection->getSectionOffset() + 422 RelEntry.Offset; 423 switch (RelEntry.Type) { 424 case wasm::R_WEBASSEMBLY_FUNCTION_INDEX_LEB: { 425 uint32_t Index = SymbolIndices[RelEntry.Symbol]; 426 assert(RelEntry.Addend == 0); 427 428 WritePatchableLEB(Stream, Index, Offset); 429 break; 430 } 431 case wasm::R_WEBASSEMBLY_TABLE_INDEX_SLEB: { 432 uint32_t Index = SymbolIndices[RelEntry.Symbol]; 433 assert(RelEntry.Addend == 0); 434 435 WritePatchableSLEB(Stream, Index, Offset); 436 break; 437 } 438 case wasm::R_WEBASSEMBLY_GLOBAL_ADDR_SLEB: { 439 uint32_t Value = ProvisionalValue(RelEntry); 440 441 WritePatchableSLEB(Stream, Value, Offset); 442 break; 443 } 444 case wasm::R_WEBASSEMBLY_GLOBAL_ADDR_LEB: { 445 uint32_t Value = ProvisionalValue(RelEntry); 446 447 WritePatchableLEB(Stream, Value, Offset); 448 break; 449 } 450 case wasm::R_WEBASSEMBLY_TABLE_INDEX_I32: { 451 uint32_t Index = SymbolIndices[RelEntry.Symbol]; 452 assert(RelEntry.Addend == 0); 453 454 WriteI32(Stream, Index, Offset); 455 break; 456 } 457 case wasm::R_WEBASSEMBLY_GLOBAL_ADDR_I32: { 458 uint32_t Value = ProvisionalValue(RelEntry); 459 460 WriteI32(Stream, Value, Offset); 461 break; 462 } 463 default: 464 break; 465 } 466 } 467 } 468 469 // Write out the portions of the relocation records that the linker will 470 // need to handle. 471 static void 472 WriteRelocations(ArrayRef<WasmRelocationEntry> Relocations, 473 raw_pwrite_stream &Stream, 474 DenseMap<const MCSymbolWasm *, uint32_t> &SymbolIndices, 475 uint64_t HeaderSize) { 476 for (const WasmRelocationEntry RelEntry : Relocations) { 477 encodeULEB128(RelEntry.Type, Stream); 478 479 uint64_t Offset = RelEntry.Offset + 480 RelEntry.FixupSection->getSectionOffset() + HeaderSize; 481 uint32_t Index = SymbolIndices[RelEntry.Symbol]; 482 int64_t Addend = RelEntry.Addend; 483 484 switch (RelEntry.Type) { 485 case wasm::R_WEBASSEMBLY_FUNCTION_INDEX_LEB: 486 case wasm::R_WEBASSEMBLY_TABLE_INDEX_SLEB: 487 case wasm::R_WEBASSEMBLY_TABLE_INDEX_I32: 488 encodeULEB128(Offset, Stream); 489 encodeULEB128(Index, Stream); 490 assert(Addend == 0 && "addends not supported for functions"); 491 break; 492 case wasm::R_WEBASSEMBLY_GLOBAL_ADDR_LEB: 493 case wasm::R_WEBASSEMBLY_GLOBAL_ADDR_SLEB: 494 case wasm::R_WEBASSEMBLY_GLOBAL_ADDR_I32: 495 encodeULEB128(Offset, Stream); 496 encodeULEB128(Index, Stream); 497 encodeSLEB128(Addend, Stream); 498 break; 499 default: 500 llvm_unreachable("unsupported relocation type"); 501 } 502 } 503 } 504 505 // Write out the the type relocation records that the linker will 506 // need to handle. 507 static void WriteTypeRelocations( 508 ArrayRef<WasmRelocationEntry> TypeIndexFixups, 509 ArrayRef<uint32_t> TypeIndexFixupTypes, 510 raw_pwrite_stream &Stream) 511 { 512 for (size_t i = 0, e = TypeIndexFixups.size(); i < e; ++i) { 513 const WasmRelocationEntry &Fixup = TypeIndexFixups[i]; 514 uint32_t Type = TypeIndexFixupTypes[i]; 515 516 assert(Fixup.Type == wasm::R_WEBASSEMBLY_TYPE_INDEX_LEB); 517 assert(Fixup.Addend == 0); 518 519 uint64_t Offset = Fixup.Offset + 520 Fixup.FixupSection->getSectionOffset(); 521 522 encodeULEB128(Fixup.Type, Stream); 523 encodeULEB128(Offset, Stream); 524 encodeULEB128(Type, Stream); 525 } 526 } 527 528 void WasmObjectWriter::writeObject(MCAssembler &Asm, 529 const MCAsmLayout &Layout) { 530 MCContext &Ctx = Asm.getContext(); 531 wasm::ValType PtrType = is64Bit() ? wasm::ValType::I64 : wasm::ValType::I32; 532 533 // Collect information from the available symbols. 534 DenseMap<WasmFunctionType, int32_t, WasmFunctionTypeDenseMapInfo> 535 FunctionTypeIndices; 536 SmallVector<WasmFunctionType, 4> FunctionTypes; 537 SmallVector<WasmFunction, 4> Functions; 538 SmallVector<uint32_t, 4> TableElems; 539 SmallVector<WasmGlobal, 4> Globals; 540 SmallVector<WasmImport, 4> Imports; 541 SmallVector<WasmExport, 4> Exports; 542 DenseMap<const MCSymbolWasm *, uint32_t> SymbolIndices; 543 SmallPtrSet<const MCSymbolWasm *, 4> IsAddressTaken; 544 unsigned NumFuncImports = 0; 545 unsigned NumGlobalImports = 0; 546 SmallVector<char, 0> DataBytes; 547 uint32_t StackPointerGlobal = 0; 548 bool HasStackPointer = false; 549 550 // Populate the IsAddressTaken set. 551 for (WasmRelocationEntry RelEntry : CodeRelocations) { 552 switch (RelEntry.Type) { 553 case wasm::R_WEBASSEMBLY_TABLE_INDEX_SLEB: 554 case wasm::R_WEBASSEMBLY_GLOBAL_ADDR_SLEB: 555 IsAddressTaken.insert(RelEntry.Symbol); 556 break; 557 default: 558 break; 559 } 560 } 561 for (WasmRelocationEntry RelEntry : DataRelocations) { 562 switch (RelEntry.Type) { 563 case wasm::R_WEBASSEMBLY_TABLE_INDEX_I32: 564 case wasm::R_WEBASSEMBLY_GLOBAL_ADDR_I32: 565 IsAddressTaken.insert(RelEntry.Symbol); 566 break; 567 default: 568 break; 569 } 570 } 571 572 // Populate the Imports set. 573 for (const MCSymbol &S : Asm.symbols()) { 574 const auto &WS = static_cast<const MCSymbolWasm &>(S); 575 int32_t Type; 576 577 if (WS.isFunction()) { 578 // Prepare the function's type, if we haven't seen it yet. 579 WasmFunctionType F; 580 F.Returns = WS.getReturns(); 581 F.Params = WS.getParams(); 582 auto Pair = 583 FunctionTypeIndices.insert(std::make_pair(F, FunctionTypes.size())); 584 if (Pair.second) 585 FunctionTypes.push_back(F); 586 587 Type = Pair.first->second; 588 } else { 589 Type = int32_t(PtrType); 590 } 591 592 // If the symbol is not defined in this translation unit, import it. 593 if (!WS.isTemporary() && !WS.isDefined(/*SetUsed=*/false)) { 594 WasmImport Import; 595 Import.ModuleName = WS.getModuleName(); 596 Import.FieldName = WS.getName(); 597 598 if (WS.isFunction()) { 599 Import.Kind = wasm::WASM_EXTERNAL_FUNCTION; 600 Import.Type = Type; 601 SymbolIndices[&WS] = NumFuncImports; 602 ++NumFuncImports; 603 } else { 604 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 605 Import.Type = Type; 606 SymbolIndices[&WS] = NumGlobalImports; 607 ++NumGlobalImports; 608 } 609 610 Imports.push_back(Import); 611 } 612 } 613 614 // In the special .global_variables section, we've encoded global 615 // variables used by the function. Translate them into the Globals 616 // list. 617 MCSectionWasm *GlobalVars = Ctx.getWasmSection(".global_variables", 0, 0); 618 if (!GlobalVars->getFragmentList().empty()) { 619 if (GlobalVars->getFragmentList().size() != 1) 620 report_fatal_error("only one .global_variables fragment supported"); 621 const MCFragment &Frag = *GlobalVars->begin(); 622 if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data) 623 report_fatal_error("only data supported in .global_variables"); 624 const MCDataFragment &DataFrag = cast<MCDataFragment>(Frag); 625 if (!DataFrag.getFixups().empty()) 626 report_fatal_error("fixups not supported in .global_variables"); 627 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 628 for (const uint8_t *p = (const uint8_t *)Contents.data(), 629 *end = (const uint8_t *)Contents.data() + Contents.size(); 630 p != end; ) { 631 WasmGlobal G; 632 if (end - p < 3) 633 report_fatal_error("truncated global variable encoding"); 634 G.Type = wasm::ValType(int8_t(*p++)); 635 G.IsMutable = bool(*p++); 636 G.HasImport = bool(*p++); 637 if (G.HasImport) { 638 G.InitialValue = 0; 639 640 WasmImport Import; 641 Import.ModuleName = (const char *)p; 642 const uint8_t *nul = (const uint8_t *)memchr(p, '\0', end - p); 643 if (!nul) 644 report_fatal_error("global module name must be nul-terminated"); 645 p = nul + 1; 646 nul = (const uint8_t *)memchr(p, '\0', end - p); 647 if (!nul) 648 report_fatal_error("global base name must be nul-terminated"); 649 Import.FieldName = (const char *)p; 650 p = nul + 1; 651 652 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL; 653 Import.Type = int32_t(G.Type); 654 655 G.ImportIndex = NumGlobalImports; 656 ++NumGlobalImports; 657 658 Imports.push_back(Import); 659 } else { 660 unsigned n; 661 G.InitialValue = decodeSLEB128(p, &n); 662 G.ImportIndex = 0; 663 if ((ptrdiff_t)n > end - p) 664 report_fatal_error("global initial value must be valid SLEB128"); 665 p += n; 666 } 667 Globals.push_back(G); 668 } 669 } 670 671 // In the special .stack_pointer section, we've encoded the stack pointer 672 // index. 673 MCSectionWasm *StackPtr = Ctx.getWasmSection(".stack_pointer", 0, 0); 674 if (!StackPtr->getFragmentList().empty()) { 675 if (StackPtr->getFragmentList().size() != 1) 676 report_fatal_error("only one .stack_pointer fragment supported"); 677 const MCFragment &Frag = *StackPtr->begin(); 678 if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data) 679 report_fatal_error("only data supported in .stack_pointer"); 680 const MCDataFragment &DataFrag = cast<MCDataFragment>(Frag); 681 if (!DataFrag.getFixups().empty()) 682 report_fatal_error("fixups not supported in .stack_pointer"); 683 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 684 if (Contents.size() != 4) 685 report_fatal_error("only one entry supported in .stack_pointer"); 686 HasStackPointer = true; 687 StackPointerGlobal = NumGlobalImports + *(const int32_t *)Contents.data(); 688 } 689 690 // Handle defined symbols. 691 for (const MCSymbol &S : Asm.symbols()) { 692 // Ignore unnamed temporary symbols, which aren't ever exported, imported, 693 // or used in relocations. 694 if (S.isTemporary() && S.getName().empty()) 695 continue; 696 const auto &WS = static_cast<const MCSymbolWasm &>(S); 697 unsigned Index; 698 if (WS.isFunction()) { 699 // Prepare the function's type, if we haven't seen it yet. 700 WasmFunctionType F; 701 F.Returns = WS.getReturns(); 702 F.Params = WS.getParams(); 703 auto Pair = 704 FunctionTypeIndices.insert(std::make_pair(F, FunctionTypes.size())); 705 if (Pair.second) 706 FunctionTypes.push_back(F); 707 708 int32_t Type = Pair.first->second; 709 710 if (WS.isDefined(/*SetUsed=*/false)) { 711 // A definition. Take the next available index. 712 Index = NumFuncImports + Functions.size(); 713 714 // Prepare the function. 715 WasmFunction Func; 716 Func.Type = Type; 717 Func.Sym = &WS; 718 SymbolIndices[&WS] = Index; 719 Functions.push_back(Func); 720 } else { 721 // An import; the index was assigned above. 722 Index = SymbolIndices.find(&WS)->second; 723 } 724 725 // If needed, prepare the function to be called indirectly. 726 if (IsAddressTaken.count(&WS)) 727 TableElems.push_back(Index); 728 } else { 729 // For now, ignore temporary non-function symbols. 730 if (S.isTemporary()) 731 continue; 732 733 if (WS.getOffset() != 0) 734 report_fatal_error("data sections must contain one variable each"); 735 if (!WS.getSize()) 736 report_fatal_error("data symbols must have a size set with .size"); 737 738 int64_t Size = 0; 739 if (!WS.getSize()->evaluateAsAbsolute(Size, Layout)) 740 report_fatal_error(".size expression must be evaluatable"); 741 742 if (WS.isDefined(false)) { 743 MCSectionWasm &DataSection = 744 static_cast<MCSectionWasm &>(WS.getSection()); 745 746 if (uint64_t(Size) != Layout.getSectionFileSize(&DataSection)) 747 report_fatal_error("data sections must contain at most one variable"); 748 749 DataBytes.resize(alignTo(DataBytes.size(), DataSection.getAlignment())); 750 751 DataSection.setSectionOffset(DataBytes.size()); 752 753 for (MCSection::iterator I = DataSection.begin(), E = DataSection.end(); 754 I != E; ++I) { 755 const MCFragment &Frag = *I; 756 if (Frag.hasInstructions()) 757 report_fatal_error("only data supported in data sections"); 758 759 if (const MCAlignFragment *Align = dyn_cast<MCAlignFragment>(&Frag)) { 760 if (Align->getValueSize() != 1) 761 report_fatal_error("only byte values supported for alignment"); 762 // If nops are requested, use zeros, as this is the data section. 763 uint8_t Value = Align->hasEmitNops() ? 0 : Align->getValue(); 764 uint64_t Size = std::min<uint64_t>(alignTo(DataBytes.size(), 765 Align->getAlignment()), 766 DataBytes.size() + 767 Align->getMaxBytesToEmit()); 768 DataBytes.resize(Size, Value); 769 } else if (const MCFillFragment *Fill = 770 dyn_cast<MCFillFragment>(&Frag)) { 771 DataBytes.insert(DataBytes.end(), Size, Fill->getValue()); 772 } else { 773 const MCDataFragment &DataFrag = cast<MCDataFragment>(Frag); 774 const SmallVectorImpl<char> &Contents = DataFrag.getContents(); 775 776 DataBytes.insert(DataBytes.end(), Contents.begin(), Contents.end()); 777 } 778 } 779 780 // For each external global, prepare a corresponding wasm global 781 // holding its address. 782 if (WS.isExternal()) { 783 Index = NumGlobalImports + Globals.size(); 784 785 WasmGlobal Global; 786 Global.Type = PtrType; 787 Global.IsMutable = false; 788 Global.HasImport = false; 789 Global.InitialValue = DataSection.getSectionOffset(); 790 Global.ImportIndex = 0; 791 SymbolIndices[&WS] = Index; 792 Globals.push_back(Global); 793 } 794 } 795 } 796 797 // If the symbol is visible outside this translation unit, export it. 798 if (WS.isExternal()) { 799 assert(WS.isDefined(false)); 800 WasmExport Export; 801 Export.FieldName = WS.getName(); 802 Export.Index = Index; 803 804 if (WS.isFunction()) 805 Export.Kind = wasm::WASM_EXTERNAL_FUNCTION; 806 else 807 Export.Kind = wasm::WASM_EXTERNAL_GLOBAL; 808 809 Exports.push_back(Export); 810 } 811 } 812 813 // Add types for indirect function calls. 814 for (const WasmRelocationEntry &Fixup : TypeIndexFixups) { 815 assert(Fixup.Addend == 0); 816 assert(Fixup.Type == wasm::R_WEBASSEMBLY_TYPE_INDEX_LEB); 817 818 WasmFunctionType F; 819 F.Returns = Fixup.Symbol->getReturns(); 820 F.Params = Fixup.Symbol->getParams(); 821 auto Pair = 822 FunctionTypeIndices.insert(std::make_pair(F, FunctionTypes.size())); 823 if (Pair.second) 824 FunctionTypes.push_back(F); 825 826 TypeIndexFixupTypes.push_back(Pair.first->second); 827 } 828 829 // Write out the Wasm header. 830 writeHeader(Asm); 831 832 SectionBookkeeping Section; 833 834 // === Type Section ========================================================= 835 if (!FunctionTypes.empty()) { 836 startSection(Section, wasm::WASM_SEC_TYPE); 837 838 encodeULEB128(FunctionTypes.size(), getStream()); 839 840 for (WasmFunctionType &FuncTy : FunctionTypes) { 841 encodeSLEB128(wasm::WASM_TYPE_FUNC, getStream()); 842 encodeULEB128(FuncTy.Params.size(), getStream()); 843 for (wasm::ValType Ty : FuncTy.Params) 844 writeValueType(Ty); 845 encodeULEB128(FuncTy.Returns.size(), getStream()); 846 for (wasm::ValType Ty : FuncTy.Returns) 847 writeValueType(Ty); 848 } 849 850 endSection(Section); 851 } 852 853 // === Import Section ======================================================== 854 if (!Imports.empty()) { 855 startSection(Section, wasm::WASM_SEC_IMPORT); 856 857 encodeULEB128(Imports.size(), getStream()); 858 for (const WasmImport &Import : Imports) { 859 StringRef ModuleName = Import.ModuleName; 860 encodeULEB128(ModuleName.size(), getStream()); 861 writeBytes(ModuleName); 862 863 StringRef FieldName = Import.FieldName; 864 encodeULEB128(FieldName.size(), getStream()); 865 writeBytes(FieldName); 866 867 encodeULEB128(Import.Kind, getStream()); 868 869 switch (Import.Kind) { 870 case wasm::WASM_EXTERNAL_FUNCTION: 871 encodeULEB128(Import.Type, getStream()); 872 break; 873 case wasm::WASM_EXTERNAL_GLOBAL: 874 encodeSLEB128(int32_t(Import.Type), getStream()); 875 encodeULEB128(0, getStream()); // mutability 876 break; 877 default: 878 llvm_unreachable("unsupported import kind"); 879 } 880 } 881 882 endSection(Section); 883 } 884 885 // === Function Section ====================================================== 886 if (!Functions.empty()) { 887 startSection(Section, wasm::WASM_SEC_FUNCTION); 888 889 encodeULEB128(Functions.size(), getStream()); 890 for (const WasmFunction &Func : Functions) 891 encodeULEB128(Func.Type, getStream()); 892 893 endSection(Section); 894 } 895 896 // === Table Section ========================================================= 897 // For now, always emit the table section, since indirect calls are not 898 // valid without it. In the future, we could perhaps be more clever and omit 899 // it if there are no indirect calls. 900 startSection(Section, wasm::WASM_SEC_TABLE); 901 902 // The number of tables, fixed to 1 for now. 903 encodeULEB128(1, getStream()); 904 905 encodeSLEB128(wasm::WASM_TYPE_ANYFUNC, getStream()); 906 907 encodeULEB128(0, getStream()); // flags 908 encodeULEB128(TableElems.size(), getStream()); // initial 909 910 endSection(Section); 911 912 // === Memory Section ======================================================== 913 // For now, always emit the memory section, since loads and stores are not 914 // valid without it. In the future, we could perhaps be more clever and omit 915 // it if there are no loads or stores. 916 uint32_t NumPages = 917 (DataBytes.size() + wasm::WasmPageSize - 1) / wasm::WasmPageSize; 918 919 startSection(Section, wasm::WASM_SEC_MEMORY); 920 encodeULEB128(1, getStream()); // number of memory spaces 921 922 encodeULEB128(0, getStream()); // flags 923 encodeULEB128(NumPages, getStream()); // initial 924 925 endSection(Section); 926 927 // === Global Section ======================================================== 928 if (!Globals.empty()) { 929 startSection(Section, wasm::WASM_SEC_GLOBAL); 930 931 encodeULEB128(Globals.size(), getStream()); 932 for (const WasmGlobal &Global : Globals) { 933 writeValueType(Global.Type); 934 write8(Global.IsMutable); 935 936 if (Global.HasImport) { 937 assert(Global.InitialValue == 0); 938 write8(wasm::WASM_OPCODE_GET_GLOBAL); 939 encodeULEB128(Global.ImportIndex, getStream()); 940 } else { 941 assert(Global.ImportIndex == 0); 942 write8(wasm::WASM_OPCODE_I32_CONST); 943 encodeSLEB128(Global.InitialValue, getStream()); // offset 944 } 945 write8(wasm::WASM_OPCODE_END); 946 } 947 948 endSection(Section); 949 } 950 951 // === Export Section ======================================================== 952 if (!Exports.empty()) { 953 startSection(Section, wasm::WASM_SEC_EXPORT); 954 955 encodeULEB128(Exports.size(), getStream()); 956 for (const WasmExport &Export : Exports) { 957 encodeULEB128(Export.FieldName.size(), getStream()); 958 writeBytes(Export.FieldName); 959 960 encodeSLEB128(Export.Kind, getStream()); 961 962 encodeULEB128(Export.Index, getStream()); 963 } 964 965 endSection(Section); 966 } 967 968 #if 0 // TODO: Start Section 969 if (HaveStartFunction) { 970 // === Start Section ========================================================= 971 startSection(Section, wasm::WASM_SEC_START); 972 973 encodeSLEB128(StartFunction, getStream()); 974 975 endSection(Section); 976 } 977 #endif 978 979 // === Elem Section ========================================================== 980 if (!TableElems.empty()) { 981 startSection(Section, wasm::WASM_SEC_ELEM); 982 983 encodeULEB128(1, getStream()); // number of "segments" 984 encodeULEB128(0, getStream()); // the table index 985 986 // init expr for starting offset 987 write8(wasm::WASM_OPCODE_I32_CONST); 988 encodeSLEB128(0, getStream()); 989 write8(wasm::WASM_OPCODE_END); 990 991 encodeULEB128(TableElems.size(), getStream()); 992 for (uint32_t Elem : TableElems) 993 encodeULEB128(Elem, getStream()); 994 995 endSection(Section); 996 } 997 998 // === Code Section ========================================================== 999 if (!Functions.empty()) { 1000 startSection(Section, wasm::WASM_SEC_CODE); 1001 1002 encodeULEB128(Functions.size(), getStream()); 1003 1004 for (const WasmFunction &Func : Functions) { 1005 MCSectionWasm &FuncSection = 1006 static_cast<MCSectionWasm &>(Func.Sym->getSection()); 1007 1008 if (Func.Sym->isVariable()) 1009 report_fatal_error("weak symbols not supported yet"); 1010 1011 if (Func.Sym->getOffset() != 0) 1012 report_fatal_error("function sections must contain one function each"); 1013 1014 if (!Func.Sym->getSize()) 1015 report_fatal_error("function symbols must have a size set with .size"); 1016 1017 int64_t Size = 0; 1018 if (!Func.Sym->getSize()->evaluateAsAbsolute(Size, Layout)) 1019 report_fatal_error(".size expression must be evaluatable"); 1020 1021 encodeULEB128(Size, getStream()); 1022 1023 FuncSection.setSectionOffset(getStream().tell() - 1024 Section.ContentsOffset); 1025 1026 Asm.writeSectionData(&FuncSection, Layout); 1027 } 1028 1029 // Apply the type index fixups for call_indirect etc. instructions. 1030 for (size_t i = 0, e = TypeIndexFixups.size(); i < e; ++i) { 1031 uint32_t Type = TypeIndexFixupTypes[i]; 1032 unsigned Padding = PaddingFor5ByteULEB128(Type); 1033 1034 const WasmRelocationEntry &Fixup = TypeIndexFixups[i]; 1035 assert(Fixup.Addend == 0); 1036 assert(Fixup.Type == wasm::R_WEBASSEMBLY_TYPE_INDEX_LEB); 1037 uint64_t Offset = Fixup.Offset + 1038 Fixup.FixupSection->getSectionOffset(); 1039 1040 uint8_t Buffer[16]; 1041 unsigned SizeLen = encodeULEB128(Type, Buffer, Padding); 1042 assert(SizeLen == 5); 1043 getStream().pwrite((char *)Buffer, SizeLen, 1044 Section.ContentsOffset + Offset); 1045 } 1046 1047 // Apply fixups. 1048 ApplyRelocations(CodeRelocations, getStream(), SymbolIndices, 1049 Section.ContentsOffset); 1050 1051 endSection(Section); 1052 } 1053 1054 // === Data Section ========================================================== 1055 uint32_t DataSectionHeaderSize = 0; 1056 if (!DataBytes.empty()) { 1057 startSection(Section, wasm::WASM_SEC_DATA); 1058 1059 encodeULEB128(1, getStream()); // count 1060 encodeULEB128(0, getStream()); // memory index 1061 write8(wasm::WASM_OPCODE_I32_CONST); 1062 encodeSLEB128(0, getStream()); // offset 1063 write8(wasm::WASM_OPCODE_END); 1064 encodeULEB128(DataBytes.size(), getStream()); // size 1065 DataSectionHeaderSize = getStream().tell() - Section.ContentsOffset; 1066 writeBytes(DataBytes); // data 1067 1068 // Apply fixups. 1069 ApplyRelocations(DataRelocations, getStream(), SymbolIndices, 1070 Section.ContentsOffset + DataSectionHeaderSize); 1071 1072 endSection(Section); 1073 } 1074 1075 // === Name Section ========================================================== 1076 uint32_t TotalFunctions = NumFuncImports + Functions.size(); 1077 if (TotalFunctions != 0) { 1078 startSection(Section, wasm::WASM_SEC_CUSTOM, "name"); 1079 SectionBookkeeping SubSection; 1080 startSection(SubSection, wasm::WASM_NAMES_FUNCTION); 1081 1082 encodeULEB128(TotalFunctions, getStream()); 1083 uint32_t Index = 0; 1084 for (const WasmImport &Import : Imports) { 1085 if (Import.Kind == wasm::WASM_EXTERNAL_FUNCTION) { 1086 encodeULEB128(Index, getStream()); 1087 encodeULEB128(Import.FieldName.size(), getStream()); 1088 writeBytes(Import.FieldName); 1089 ++Index; 1090 } 1091 } 1092 for (const WasmFunction &Func : Functions) { 1093 encodeULEB128(Index, getStream()); 1094 encodeULEB128(Func.Sym->getName().size(), getStream()); 1095 writeBytes(Func.Sym->getName()); 1096 ++Index; 1097 } 1098 1099 endSection(SubSection); 1100 endSection(Section); 1101 } 1102 1103 // See: https://github.com/WebAssembly/tool-conventions/blob/master/Linking.md 1104 // for descriptions of the reloc sections. 1105 1106 // === Code Reloc Section ==================================================== 1107 if (!CodeRelocations.empty()) { 1108 startSection(Section, wasm::WASM_SEC_CUSTOM, "reloc.CODE"); 1109 1110 encodeULEB128(wasm::WASM_SEC_CODE, getStream()); 1111 1112 encodeULEB128(CodeRelocations.size() + TypeIndexFixups.size(), getStream()); 1113 1114 WriteRelocations(CodeRelocations, getStream(), SymbolIndices, 0); 1115 WriteTypeRelocations(TypeIndexFixups, TypeIndexFixupTypes, getStream()); 1116 1117 endSection(Section); 1118 } 1119 1120 // === Data Reloc Section ==================================================== 1121 if (!DataRelocations.empty()) { 1122 startSection(Section, wasm::WASM_SEC_CUSTOM, "reloc.DATA"); 1123 1124 encodeULEB128(wasm::WASM_SEC_DATA, getStream()); 1125 1126 encodeULEB128(DataRelocations.size(), getStream()); 1127 1128 WriteRelocations(DataRelocations, getStream(), SymbolIndices, 1129 DataSectionHeaderSize); 1130 1131 endSection(Section); 1132 } 1133 1134 // === Linking Metadata Section ============================================== 1135 if (HasStackPointer) { 1136 startSection(Section, wasm::WASM_SEC_CUSTOM, "linking"); 1137 1138 encodeULEB128(1, getStream()); // count 1139 1140 encodeULEB128(wasm::WASM_STACK_POINTER, getStream()); // type 1141 encodeULEB128(StackPointerGlobal, getStream()); // id 1142 1143 endSection(Section); 1144 } 1145 1146 // TODO: Translate the .comment section to the output. 1147 1148 // TODO: Translate debug sections to the output. 1149 } 1150 1151 MCObjectWriter *llvm::createWasmObjectWriter(MCWasmObjectTargetWriter *MOTW, 1152 raw_pwrite_stream &OS) { 1153 return new WasmObjectWriter(MOTW, OS); 1154 } 1155