1 //===- lib/MC/WasmObjectWriter.cpp - Wasm File Writer ---------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements Wasm object file writer information.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/ADT/STLExtras.h"
15 #include "llvm/ADT/SmallPtrSet.h"
16 #include "llvm/BinaryFormat/Wasm.h"
17 #include "llvm/MC/MCAsmBackend.h"
18 #include "llvm/MC/MCAsmInfo.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/MCObjectFileInfo.h"
25 #include "llvm/MC/MCObjectWriter.h"
26 #include "llvm/MC/MCSectionWasm.h"
27 #include "llvm/MC/MCSymbolWasm.h"
28 #include "llvm/MC/MCValue.h"
29 #include "llvm/MC/MCWasmObjectWriter.h"
30 #include "llvm/Support/Casting.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/LEB128.h"
34 #include "llvm/Support/StringSaver.h"
39 #define DEBUG_TYPE "mc"
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.
49 // Where the contents of the section starts (after the header).
50 uint64_t ContentsOffset;
53 // The signature of a wasm function, in a struct capable of being used as a
55 struct WasmFunctionType {
56 // Support empty and tombstone instances, needed by DenseMap.
57 enum { Plain, Empty, Tombstone } State;
59 // The return types of the function.
60 SmallVector<wasm::ValType, 1> Returns;
62 // The parameter types of the function.
63 SmallVector<wasm::ValType, 4> Params;
65 WasmFunctionType() : State(Plain) {}
67 bool operator==(const WasmFunctionType &Other) const {
68 return State == Other.State && Returns == Other.Returns &&
69 Params == Other.Params;
73 // Traits for using WasmFunctionType in a DenseMap.
74 struct WasmFunctionTypeDenseMapInfo {
75 static WasmFunctionType getEmptyKey() {
76 WasmFunctionType FuncTy;
77 FuncTy.State = WasmFunctionType::Empty;
80 static WasmFunctionType getTombstoneKey() {
81 WasmFunctionType FuncTy;
82 FuncTy.State = WasmFunctionType::Tombstone;
85 static unsigned getHashValue(const WasmFunctionType &FuncTy) {
86 uintptr_t Value = FuncTy.State;
87 for (wasm::ValType Ret : FuncTy.Returns)
88 Value += DenseMapInfo<int32_t>::getHashValue(int32_t(Ret));
89 for (wasm::ValType Param : FuncTy.Params)
90 Value += DenseMapInfo<int32_t>::getHashValue(int32_t(Param));
93 static bool isEqual(const WasmFunctionType &LHS,
94 const WasmFunctionType &RHS) {
99 // A wasm data segment. A wasm binary contains only a single data section
100 // but that can contain many segments, each with their own virtual location
101 // in memory. Each MCSection data created by llvm is modeled as its own
102 // wasm data segment.
103 struct WasmDataSegment {
104 MCSectionWasm *Section;
109 SmallVector<char, 4> Data;
112 // A wasm import to be written into the import section.
114 StringRef ModuleName;
120 // A wasm function to be written into the function section.
121 struct WasmFunction {
123 const MCSymbolWasm *Sym;
126 // A wasm export to be written into the export section.
133 // A wasm global to be written into the global section.
138 uint64_t InitialValue;
139 uint32_t ImportIndex;
142 // Information about a single relocation.
143 struct WasmRelocationEntry {
144 uint64_t Offset; // Where is the relocation.
145 const MCSymbolWasm *Symbol; // The symbol to relocate with.
146 int64_t Addend; // A value to add to the symbol.
147 unsigned Type; // The type of the relocation.
148 const MCSectionWasm *FixupSection;// The section the relocation is targeting.
150 WasmRelocationEntry(uint64_t Offset, const MCSymbolWasm *Symbol,
151 int64_t Addend, unsigned Type,
152 const MCSectionWasm *FixupSection)
153 : Offset(Offset), Symbol(Symbol), Addend(Addend), Type(Type),
154 FixupSection(FixupSection) {}
156 bool hasAddend() const {
158 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_LEB:
159 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_SLEB:
160 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_I32:
167 void print(raw_ostream &Out) const {
168 Out << "Off=" << Offset << ", Sym=" << *Symbol << ", Addend=" << Addend
170 << ", FixupSection=" << FixupSection->getSectionName();
173 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
174 LLVM_DUMP_METHOD void dump() const { print(dbgs()); }
179 raw_ostream &operator<<(raw_ostream &OS, const WasmRelocationEntry &Rel) {
185 class WasmObjectWriter : public MCObjectWriter {
186 /// Helper struct for containing some precomputed information on symbols.
187 struct WasmSymbolData {
188 const MCSymbolWasm *Symbol;
191 // Support lexicographic sorting.
192 bool operator<(const WasmSymbolData &RHS) const { return Name < RHS.Name; }
195 /// The target specific Wasm writer instance.
196 std::unique_ptr<MCWasmObjectTargetWriter> TargetObjectWriter;
198 // Relocations for fixing up references in the code section.
199 std::vector<WasmRelocationEntry> CodeRelocations;
201 // Relocations for fixing up references in the data section.
202 std::vector<WasmRelocationEntry> DataRelocations;
204 // Index values to use for fixing up call_indirect type indices.
205 // Maps function symbols to the index of the type of the function
206 DenseMap<const MCSymbolWasm *, uint32_t> TypeIndices;
207 // Maps function symbols to the table element index space. Used
208 // for TABLE_INDEX relocation types (i.e. address taken functions).
209 DenseMap<const MCSymbolWasm *, uint32_t> IndirectSymbolIndices;
210 // Maps function/global symbols to the function/global index space.
211 DenseMap<const MCSymbolWasm *, uint32_t> SymbolIndices;
213 DenseMap<WasmFunctionType, int32_t, WasmFunctionTypeDenseMapInfo>
215 SmallVector<WasmFunctionType, 4> FunctionTypes;
216 SmallVector<WasmGlobal, 4> Globals;
217 unsigned NumGlobalImports = 0;
219 // TargetObjectWriter wrappers.
220 bool is64Bit() const { return TargetObjectWriter->is64Bit(); }
221 unsigned getRelocType(const MCValue &Target, const MCFixup &Fixup) const {
222 return TargetObjectWriter->getRelocType(Target, Fixup);
225 void startSection(SectionBookkeeping &Section, unsigned SectionId,
226 const char *Name = nullptr);
227 void endSection(SectionBookkeeping &Section);
230 WasmObjectWriter(MCWasmObjectTargetWriter *MOTW, raw_pwrite_stream &OS)
231 : MCObjectWriter(OS, /*IsLittleEndian=*/true), TargetObjectWriter(MOTW) {}
234 ~WasmObjectWriter() override;
236 void reset() override {
237 CodeRelocations.clear();
238 DataRelocations.clear();
240 SymbolIndices.clear();
241 IndirectSymbolIndices.clear();
242 FunctionTypeIndices.clear();
243 FunctionTypes.clear();
245 MCObjectWriter::reset();
246 NumGlobalImports = 0;
249 void writeHeader(const MCAssembler &Asm);
251 void recordRelocation(MCAssembler &Asm, const MCAsmLayout &Layout,
252 const MCFragment *Fragment, const MCFixup &Fixup,
253 MCValue Target, uint64_t &FixedValue) override;
255 void executePostLayoutBinding(MCAssembler &Asm,
256 const MCAsmLayout &Layout) override;
258 void writeObject(MCAssembler &Asm, const MCAsmLayout &Layout) override;
260 void writeString(const StringRef Str) {
261 encodeULEB128(Str.size(), getStream());
265 void writeValueType(wasm::ValType Ty) {
266 encodeSLEB128(int32_t(Ty), getStream());
269 void writeTypeSection(ArrayRef<WasmFunctionType> FunctionTypes);
270 void writeImportSection(ArrayRef<WasmImport> Imports);
271 void writeFunctionSection(ArrayRef<WasmFunction> Functions);
272 void writeTableSection(uint32_t NumElements);
273 void writeMemorySection(uint32_t DataSize);
274 void writeGlobalSection();
275 void writeExportSection(ArrayRef<WasmExport> Exports);
276 void writeElemSection(ArrayRef<uint32_t> TableElems);
277 void writeCodeSection(const MCAssembler &Asm, const MCAsmLayout &Layout,
278 ArrayRef<WasmFunction> Functions);
279 void writeDataSection(ArrayRef<WasmDataSegment> Segments);
280 void writeNameSection(ArrayRef<WasmFunction> Functions,
281 ArrayRef<WasmImport> Imports,
282 uint32_t NumFuncImports);
283 void writeCodeRelocSection();
284 void writeDataRelocSection();
285 void writeLinkingMetaDataSection(
286 ArrayRef<WasmDataSegment> Segments, uint32_t DataSize,
287 SmallVector<std::pair<StringRef, uint32_t>, 4> SymbolFlags,
288 bool HasStackPointer, uint32_t StackPointerGlobal);
290 uint32_t getProvisionalValue(const WasmRelocationEntry &RelEntry);
291 void applyRelocations(ArrayRef<WasmRelocationEntry> Relocations,
292 uint64_t ContentsOffset);
294 void writeRelocations(ArrayRef<WasmRelocationEntry> Relocations);
295 uint32_t getRelocationIndexValue(const WasmRelocationEntry &RelEntry);
296 uint32_t getFunctionType(const MCSymbolWasm& Symbol);
297 uint32_t registerFunctionType(const MCSymbolWasm& Symbol);
300 } // end anonymous namespace
302 WasmObjectWriter::~WasmObjectWriter() {}
304 // Write out a section header and a patchable section size field.
305 void WasmObjectWriter::startSection(SectionBookkeeping &Section,
308 assert((Name != nullptr) == (SectionId == wasm::WASM_SEC_CUSTOM) &&
309 "Only custom sections can have names");
311 DEBUG(dbgs() << "startSection " << SectionId << ": " << Name << "\n");
312 encodeULEB128(SectionId, getStream());
314 Section.SizeOffset = getStream().tell();
316 // The section size. We don't know the size yet, so reserve enough space
317 // for any 32-bit value; we'll patch it later.
318 encodeULEB128(UINT32_MAX, getStream());
320 // The position where the section starts, for measuring its size.
321 Section.ContentsOffset = getStream().tell();
323 // Custom sections in wasm also have a string identifier.
324 if (SectionId == wasm::WASM_SEC_CUSTOM) {
326 writeString(StringRef(Name));
330 // Now that the section is complete and we know how big it is, patch up the
331 // section size field at the start of the section.
332 void WasmObjectWriter::endSection(SectionBookkeeping &Section) {
333 uint64_t Size = getStream().tell() - Section.ContentsOffset;
334 if (uint32_t(Size) != Size)
335 report_fatal_error("section size does not fit in a uint32_t");
337 DEBUG(dbgs() << "endSection size=" << Size << "\n");
339 // Write the final section size to the payload_len field, which follows
340 // the section id byte.
342 unsigned SizeLen = encodeULEB128(Size, Buffer, 5);
343 assert(SizeLen == 5);
344 getStream().pwrite((char *)Buffer, SizeLen, Section.SizeOffset);
347 // Emit the Wasm header.
348 void WasmObjectWriter::writeHeader(const MCAssembler &Asm) {
349 writeBytes(StringRef(wasm::WasmMagic, sizeof(wasm::WasmMagic)));
350 writeLE32(wasm::WasmVersion);
353 void WasmObjectWriter::executePostLayoutBinding(MCAssembler &Asm,
354 const MCAsmLayout &Layout) {
357 void WasmObjectWriter::recordRelocation(MCAssembler &Asm,
358 const MCAsmLayout &Layout,
359 const MCFragment *Fragment,
360 const MCFixup &Fixup, MCValue Target,
361 uint64_t &FixedValue) {
362 MCAsmBackend &Backend = Asm.getBackend();
363 bool IsPCRel = Backend.getFixupKindInfo(Fixup.getKind()).Flags &
364 MCFixupKindInfo::FKF_IsPCRel;
365 const auto &FixupSection = cast<MCSectionWasm>(*Fragment->getParent());
366 uint64_t C = Target.getConstant();
367 uint64_t FixupOffset = Layout.getFragmentOffset(Fragment) + Fixup.getOffset();
368 MCContext &Ctx = Asm.getContext();
370 if (const MCSymbolRefExpr *RefB = Target.getSymB()) {
371 assert(RefB->getKind() == MCSymbolRefExpr::VK_None &&
372 "Should not have constructed this");
374 // Let A, B and C being the components of Target and R be the location of
375 // the fixup. If the fixup is not pcrel, we want to compute (A - B + C).
376 // If it is pcrel, we want to compute (A - B + C - R).
378 // In general, Wasm has no relocations for -B. It can only represent (A + C)
379 // or (A + C - R). If B = R + K and the relocation is not pcrel, we can
380 // replace B to implement it: (A - R - K + C)
384 "No relocation available to represent this relative expression");
388 const auto &SymB = cast<MCSymbolWasm>(RefB->getSymbol());
390 if (SymB.isUndefined()) {
391 Ctx.reportError(Fixup.getLoc(),
392 Twine("symbol '") + SymB.getName() +
393 "' can not be undefined in a subtraction expression");
397 assert(!SymB.isAbsolute() && "Should have been folded");
398 const MCSection &SecB = SymB.getSection();
399 if (&SecB != &FixupSection) {
400 Ctx.reportError(Fixup.getLoc(),
401 "Cannot represent a difference across sections");
405 uint64_t SymBOffset = Layout.getSymbolOffset(SymB);
406 uint64_t K = SymBOffset - FixupOffset;
411 // We either rejected the fixup or folded B into C at this point.
412 const MCSymbolRefExpr *RefA = Target.getSymA();
413 const auto *SymA = RefA ? cast<MCSymbolWasm>(&RefA->getSymbol()) : nullptr;
415 if (SymA && SymA->isVariable()) {
416 const MCExpr *Expr = SymA->getVariableValue();
417 const auto *Inner = cast<MCSymbolRefExpr>(Expr);
418 if (Inner->getKind() == MCSymbolRefExpr::VK_WEAKREF)
419 llvm_unreachable("weakref used in reloc not yet implemented");
422 // Put any constant offset in an addend. Offsets can be negative, and
423 // LLVM expects wrapping, in contrast to wasm's immediates which can't
424 // be negative and don't wrap.
428 SymA->setUsedInReloc();
433 unsigned Type = getRelocType(Target, Fixup);
435 WasmRelocationEntry Rec(FixupOffset, SymA, C, Type, &FixupSection);
436 DEBUG(dbgs() << "WasmReloc: " << Rec << "\n");
438 if (FixupSection.hasInstructions())
439 CodeRelocations.push_back(Rec);
441 DataRelocations.push_back(Rec);
444 // Write X as an (unsigned) LEB value at offset Offset in Stream, padded
445 // to allow patching.
447 WritePatchableLEB(raw_pwrite_stream &Stream, uint32_t X, uint64_t Offset) {
449 unsigned SizeLen = encodeULEB128(X, Buffer, 5);
450 assert(SizeLen == 5);
451 Stream.pwrite((char *)Buffer, SizeLen, Offset);
454 // Write X as an signed LEB value at offset Offset in Stream, padded
455 // to allow patching.
457 WritePatchableSLEB(raw_pwrite_stream &Stream, int32_t X, uint64_t Offset) {
459 unsigned SizeLen = encodeSLEB128(X, Buffer, 5);
460 assert(SizeLen == 5);
461 Stream.pwrite((char *)Buffer, SizeLen, Offset);
464 // Write X as a plain integer value at offset Offset in Stream.
465 static void WriteI32(raw_pwrite_stream &Stream, uint32_t X, uint64_t Offset) {
467 support::endian::write32le(Buffer, X);
468 Stream.pwrite((char *)Buffer, sizeof(Buffer), Offset);
471 static const MCSymbolWasm* ResolveSymbol(const MCSymbolWasm& Symbol) {
472 if (Symbol.isVariable()) {
473 const MCExpr *Expr = Symbol.getVariableValue();
474 auto *Inner = cast<MCSymbolRefExpr>(Expr);
475 return cast<MCSymbolWasm>(&Inner->getSymbol());
480 // Compute a value to write into the code at the location covered
481 // by RelEntry. This value isn't used by the static linker, since
482 // we have addends; it just serves to make the code more readable
483 // and to make standalone wasm modules directly usable.
485 WasmObjectWriter::getProvisionalValue(const WasmRelocationEntry &RelEntry) {
486 const MCSymbolWasm *Sym = ResolveSymbol(*RelEntry.Symbol);
488 // For undefined symbols, use a hopefully invalid value.
489 if (!Sym->isDefined(/*SetUsed=*/false))
492 uint32_t GlobalIndex = SymbolIndices[Sym];
493 const WasmGlobal& Global = Globals[GlobalIndex - NumGlobalImports];
494 uint64_t Address = Global.InitialValue + RelEntry.Addend;
496 // Ignore overflow. LLVM allows address arithmetic to silently wrap.
497 uint32_t Value = Address;
502 static void addData(SmallVectorImpl<char> &DataBytes,
503 MCSectionWasm &DataSection) {
504 DEBUG(errs() << "addData: " << DataSection.getSectionName() << "\n");
506 DataBytes.resize(alignTo(DataBytes.size(), DataSection.getAlignment()));
508 for (const MCFragment &Frag : DataSection) {
509 if (Frag.hasInstructions())
510 report_fatal_error("only data supported in data sections");
512 if (auto *Align = dyn_cast<MCAlignFragment>(&Frag)) {
513 if (Align->getValueSize() != 1)
514 report_fatal_error("only byte values supported for alignment");
515 // If nops are requested, use zeros, as this is the data section.
516 uint8_t Value = Align->hasEmitNops() ? 0 : Align->getValue();
517 uint64_t Size = std::min<uint64_t>(alignTo(DataBytes.size(),
518 Align->getAlignment()),
520 Align->getMaxBytesToEmit());
521 DataBytes.resize(Size, Value);
522 } else if (auto *Fill = dyn_cast<MCFillFragment>(&Frag)) {
523 DataBytes.insert(DataBytes.end(), Fill->getSize(), Fill->getValue());
525 const auto &DataFrag = cast<MCDataFragment>(Frag);
526 const SmallVectorImpl<char> &Contents = DataFrag.getContents();
528 DataBytes.insert(DataBytes.end(), Contents.begin(), Contents.end());
532 DEBUG(dbgs() << "addData -> " << DataBytes.size() << "\n");
535 uint32_t WasmObjectWriter::getRelocationIndexValue(
536 const WasmRelocationEntry &RelEntry) {
537 switch (RelEntry.Type) {
538 case wasm::R_WEBASSEMBLY_TABLE_INDEX_SLEB:
539 case wasm::R_WEBASSEMBLY_TABLE_INDEX_I32:
540 if (!IndirectSymbolIndices.count(RelEntry.Symbol))
541 report_fatal_error("symbol not found table index space: " +
542 RelEntry.Symbol->getName());
543 return IndirectSymbolIndices[RelEntry.Symbol];
544 case wasm::R_WEBASSEMBLY_FUNCTION_INDEX_LEB:
545 case wasm::R_WEBASSEMBLY_GLOBAL_INDEX_LEB:
546 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_LEB:
547 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_SLEB:
548 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_I32:
549 if (!SymbolIndices.count(RelEntry.Symbol))
550 report_fatal_error("symbol not found function/global index space: " +
551 RelEntry.Symbol->getName());
552 return SymbolIndices[RelEntry.Symbol];
553 case wasm::R_WEBASSEMBLY_TYPE_INDEX_LEB:
554 if (!TypeIndices.count(RelEntry.Symbol))
555 report_fatal_error("symbol not found in type index space: " +
556 RelEntry.Symbol->getName());
557 return TypeIndices[RelEntry.Symbol];
559 llvm_unreachable("invalid relocation type");
563 // Apply the portions of the relocation records that we can handle ourselves
565 void WasmObjectWriter::applyRelocations(
566 ArrayRef<WasmRelocationEntry> Relocations, uint64_t ContentsOffset) {
567 raw_pwrite_stream &Stream = getStream();
568 for (const WasmRelocationEntry &RelEntry : Relocations) {
569 uint64_t Offset = ContentsOffset +
570 RelEntry.FixupSection->getSectionOffset() +
573 DEBUG(dbgs() << "applyRelocation: " << RelEntry << "\n");
574 switch (RelEntry.Type) {
575 case wasm::R_WEBASSEMBLY_TABLE_INDEX_SLEB:
576 case wasm::R_WEBASSEMBLY_FUNCTION_INDEX_LEB:
577 case wasm::R_WEBASSEMBLY_TYPE_INDEX_LEB:
578 case wasm::R_WEBASSEMBLY_GLOBAL_INDEX_LEB: {
579 uint32_t Index = getRelocationIndexValue(RelEntry);
580 WritePatchableSLEB(Stream, Index, Offset);
583 case wasm::R_WEBASSEMBLY_TABLE_INDEX_I32: {
584 uint32_t Index = getRelocationIndexValue(RelEntry);
585 WriteI32(Stream, Index, Offset);
588 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_SLEB: {
589 uint32_t Value = getProvisionalValue(RelEntry);
590 WritePatchableSLEB(Stream, Value, Offset);
593 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_LEB: {
594 uint32_t Value = getProvisionalValue(RelEntry);
595 WritePatchableLEB(Stream, Value, Offset);
598 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_I32: {
599 uint32_t Value = getProvisionalValue(RelEntry);
600 WriteI32(Stream, Value, Offset);
604 llvm_unreachable("invalid relocation type");
609 // Write out the portions of the relocation records that the linker will
611 void WasmObjectWriter::writeRelocations(
612 ArrayRef<WasmRelocationEntry> Relocations) {
613 raw_pwrite_stream &Stream = getStream();
614 for (const WasmRelocationEntry& RelEntry : Relocations) {
616 uint64_t Offset = RelEntry.Offset +
617 RelEntry.FixupSection->getSectionOffset();
618 uint32_t Index = getRelocationIndexValue(RelEntry);
620 encodeULEB128(RelEntry.Type, Stream);
621 encodeULEB128(Offset, Stream);
622 encodeULEB128(Index, Stream);
623 if (RelEntry.hasAddend())
624 encodeSLEB128(RelEntry.Addend, Stream);
628 void WasmObjectWriter::writeTypeSection(
629 ArrayRef<WasmFunctionType> FunctionTypes) {
630 if (FunctionTypes.empty())
633 SectionBookkeeping Section;
634 startSection(Section, wasm::WASM_SEC_TYPE);
636 encodeULEB128(FunctionTypes.size(), getStream());
638 for (const WasmFunctionType &FuncTy : FunctionTypes) {
639 encodeSLEB128(wasm::WASM_TYPE_FUNC, getStream());
640 encodeULEB128(FuncTy.Params.size(), getStream());
641 for (wasm::ValType Ty : FuncTy.Params)
643 encodeULEB128(FuncTy.Returns.size(), getStream());
644 for (wasm::ValType Ty : FuncTy.Returns)
651 void WasmObjectWriter::writeImportSection(ArrayRef<WasmImport> Imports) {
655 SectionBookkeeping Section;
656 startSection(Section, wasm::WASM_SEC_IMPORT);
658 encodeULEB128(Imports.size(), getStream());
659 for (const WasmImport &Import : Imports) {
660 writeString(Import.ModuleName);
661 writeString(Import.FieldName);
663 encodeULEB128(Import.Kind, getStream());
665 switch (Import.Kind) {
666 case wasm::WASM_EXTERNAL_FUNCTION:
667 encodeULEB128(Import.Type, getStream());
669 case wasm::WASM_EXTERNAL_GLOBAL:
670 encodeSLEB128(int32_t(Import.Type), getStream());
671 encodeULEB128(0, getStream()); // mutability
674 llvm_unreachable("unsupported import kind");
681 void WasmObjectWriter::writeFunctionSection(ArrayRef<WasmFunction> Functions) {
682 if (Functions.empty())
685 SectionBookkeeping Section;
686 startSection(Section, wasm::WASM_SEC_FUNCTION);
688 encodeULEB128(Functions.size(), getStream());
689 for (const WasmFunction &Func : Functions)
690 encodeULEB128(Func.Type, getStream());
695 void WasmObjectWriter::writeTableSection(uint32_t NumElements) {
696 // For now, always emit the table section, since indirect calls are not
697 // valid without it. In the future, we could perhaps be more clever and omit
698 // it if there are no indirect calls.
700 SectionBookkeeping Section;
701 startSection(Section, wasm::WASM_SEC_TABLE);
703 encodeULEB128(1, getStream()); // The number of tables.
704 // Fixed to 1 for now.
705 encodeSLEB128(wasm::WASM_TYPE_ANYFUNC, getStream()); // Type of table
706 encodeULEB128(0, getStream()); // flags
707 encodeULEB128(NumElements, getStream()); // initial
712 void WasmObjectWriter::writeMemorySection(uint32_t DataSize) {
713 // For now, always emit the memory section, since loads and stores are not
714 // valid without it. In the future, we could perhaps be more clever and omit
715 // it if there are no loads or stores.
716 SectionBookkeeping Section;
717 uint32_t NumPages = (DataSize + wasm::WasmPageSize - 1) / wasm::WasmPageSize;
719 startSection(Section, wasm::WASM_SEC_MEMORY);
720 encodeULEB128(1, getStream()); // number of memory spaces
722 encodeULEB128(0, getStream()); // flags
723 encodeULEB128(NumPages, getStream()); // initial
728 void WasmObjectWriter::writeGlobalSection() {
732 SectionBookkeeping Section;
733 startSection(Section, wasm::WASM_SEC_GLOBAL);
735 encodeULEB128(Globals.size(), getStream());
736 for (const WasmGlobal &Global : Globals) {
737 writeValueType(Global.Type);
738 write8(Global.IsMutable);
740 if (Global.HasImport) {
741 assert(Global.InitialValue == 0);
742 write8(wasm::WASM_OPCODE_GET_GLOBAL);
743 encodeULEB128(Global.ImportIndex, getStream());
745 assert(Global.ImportIndex == 0);
746 write8(wasm::WASM_OPCODE_I32_CONST);
747 encodeSLEB128(Global.InitialValue, getStream()); // offset
749 write8(wasm::WASM_OPCODE_END);
755 void WasmObjectWriter::writeExportSection(ArrayRef<WasmExport> Exports) {
759 SectionBookkeeping Section;
760 startSection(Section, wasm::WASM_SEC_EXPORT);
762 encodeULEB128(Exports.size(), getStream());
763 for (const WasmExport &Export : Exports) {
764 writeString(Export.FieldName);
765 encodeSLEB128(Export.Kind, getStream());
766 encodeULEB128(Export.Index, getStream());
772 void WasmObjectWriter::writeElemSection(ArrayRef<uint32_t> TableElems) {
773 if (TableElems.empty())
776 SectionBookkeeping Section;
777 startSection(Section, wasm::WASM_SEC_ELEM);
779 encodeULEB128(1, getStream()); // number of "segments"
780 encodeULEB128(0, getStream()); // the table index
782 // init expr for starting offset
783 write8(wasm::WASM_OPCODE_I32_CONST);
784 encodeSLEB128(0, getStream());
785 write8(wasm::WASM_OPCODE_END);
787 encodeULEB128(TableElems.size(), getStream());
788 for (uint32_t Elem : TableElems)
789 encodeULEB128(Elem, getStream());
794 void WasmObjectWriter::writeCodeSection(const MCAssembler &Asm,
795 const MCAsmLayout &Layout,
796 ArrayRef<WasmFunction> Functions) {
797 if (Functions.empty())
800 SectionBookkeeping Section;
801 startSection(Section, wasm::WASM_SEC_CODE);
803 encodeULEB128(Functions.size(), getStream());
805 for (const WasmFunction &Func : Functions) {
806 auto &FuncSection = static_cast<MCSectionWasm &>(Func.Sym->getSection());
809 if (!Func.Sym->getSize()->evaluateAsAbsolute(Size, Layout))
810 report_fatal_error(".size expression must be evaluatable");
812 encodeULEB128(Size, getStream());
813 FuncSection.setSectionOffset(getStream().tell() - Section.ContentsOffset);
814 Asm.writeSectionData(&FuncSection, Layout);
818 applyRelocations(CodeRelocations, Section.ContentsOffset);
823 void WasmObjectWriter::writeDataSection(ArrayRef<WasmDataSegment> Segments) {
824 if (Segments.empty())
827 SectionBookkeeping Section;
828 startSection(Section, wasm::WASM_SEC_DATA);
830 encodeULEB128(Segments.size(), getStream()); // count
832 for (const WasmDataSegment & Segment : Segments) {
833 encodeULEB128(0, getStream()); // memory index
834 write8(wasm::WASM_OPCODE_I32_CONST);
835 encodeSLEB128(Segment.Offset, getStream()); // offset
836 write8(wasm::WASM_OPCODE_END);
837 encodeULEB128(Segment.Data.size(), getStream()); // size
838 Segment.Section->setSectionOffset(getStream().tell() - Section.ContentsOffset);
839 writeBytes(Segment.Data); // data
843 applyRelocations(DataRelocations, Section.ContentsOffset);
848 void WasmObjectWriter::writeNameSection(
849 ArrayRef<WasmFunction> Functions,
850 ArrayRef<WasmImport> Imports,
851 unsigned NumFuncImports) {
852 uint32_t TotalFunctions = NumFuncImports + Functions.size();
853 if (TotalFunctions == 0)
856 SectionBookkeeping Section;
857 startSection(Section, wasm::WASM_SEC_CUSTOM, "name");
858 SectionBookkeeping SubSection;
859 startSection(SubSection, wasm::WASM_NAMES_FUNCTION);
861 encodeULEB128(TotalFunctions, getStream());
863 for (const WasmImport &Import : Imports) {
864 if (Import.Kind == wasm::WASM_EXTERNAL_FUNCTION) {
865 encodeULEB128(Index, getStream());
866 writeString(Import.FieldName);
870 for (const WasmFunction &Func : Functions) {
871 encodeULEB128(Index, getStream());
872 writeString(Func.Sym->getName());
876 endSection(SubSection);
880 void WasmObjectWriter::writeCodeRelocSection() {
881 // See: https://github.com/WebAssembly/tool-conventions/blob/master/Linking.md
882 // for descriptions of the reloc sections.
884 if (CodeRelocations.empty())
887 SectionBookkeeping Section;
888 startSection(Section, wasm::WASM_SEC_CUSTOM, "reloc.CODE");
890 encodeULEB128(wasm::WASM_SEC_CODE, getStream());
891 encodeULEB128(CodeRelocations.size(), getStream());
893 writeRelocations(CodeRelocations);
898 void WasmObjectWriter::writeDataRelocSection() {
899 // See: https://github.com/WebAssembly/tool-conventions/blob/master/Linking.md
900 // for descriptions of the reloc sections.
902 if (DataRelocations.empty())
905 SectionBookkeeping Section;
906 startSection(Section, wasm::WASM_SEC_CUSTOM, "reloc.DATA");
908 encodeULEB128(wasm::WASM_SEC_DATA, getStream());
909 encodeULEB128(DataRelocations.size(), getStream());
911 writeRelocations(DataRelocations);
916 void WasmObjectWriter::writeLinkingMetaDataSection(
917 ArrayRef<WasmDataSegment> Segments, uint32_t DataSize,
918 SmallVector<std::pair<StringRef, uint32_t>, 4> SymbolFlags,
919 bool HasStackPointer, uint32_t StackPointerGlobal) {
920 SectionBookkeeping Section;
921 startSection(Section, wasm::WASM_SEC_CUSTOM, "linking");
922 SectionBookkeeping SubSection;
924 if (HasStackPointer) {
925 startSection(SubSection, wasm::WASM_STACK_POINTER);
926 encodeULEB128(StackPointerGlobal, getStream()); // id
927 endSection(SubSection);
930 if (SymbolFlags.size() != 0) {
931 startSection(SubSection, wasm::WASM_SYMBOL_INFO);
932 encodeULEB128(SymbolFlags.size(), getStream());
933 for (auto Pair: SymbolFlags) {
934 writeString(Pair.first);
935 encodeULEB128(Pair.second, getStream());
937 endSection(SubSection);
941 startSection(SubSection, wasm::WASM_DATA_SIZE);
942 encodeULEB128(DataSize, getStream());
943 endSection(SubSection);
946 if (Segments.size()) {
947 startSection(SubSection, wasm::WASM_SEGMENT_INFO);
948 encodeULEB128(Segments.size(), getStream());
949 for (const WasmDataSegment &Segment : Segments) {
950 writeString(Segment.Name);
951 encodeULEB128(Segment.Alignment, getStream());
952 encodeULEB128(Segment.Flags, getStream());
954 endSection(SubSection);
960 uint32_t WasmObjectWriter::getFunctionType(const MCSymbolWasm& Symbol) {
961 assert(Symbol.isFunction());
962 assert(TypeIndices.count(&Symbol));
963 return TypeIndices[&Symbol];
966 uint32_t WasmObjectWriter::registerFunctionType(const MCSymbolWasm& Symbol) {
967 assert(Symbol.isFunction());
970 const MCSymbolWasm* ResolvedSym = ResolveSymbol(Symbol);
971 F.Returns = ResolvedSym->getReturns();
972 F.Params = ResolvedSym->getParams();
975 FunctionTypeIndices.insert(std::make_pair(F, FunctionTypes.size()));
977 FunctionTypes.push_back(F);
978 TypeIndices[&Symbol] = Pair.first->second;
980 DEBUG(dbgs() << "registerFunctionType: " << Symbol << " new:" << Pair.second << "\n");
981 DEBUG(dbgs() << " -> type index: " << Pair.first->second << "\n");
982 return Pair.first->second;
985 void WasmObjectWriter::writeObject(MCAssembler &Asm,
986 const MCAsmLayout &Layout) {
987 DEBUG(dbgs() << "WasmObjectWriter::writeObject\n");
988 MCContext &Ctx = Asm.getContext();
989 wasm::ValType PtrType = is64Bit() ? wasm::ValType::I64 : wasm::ValType::I32;
991 // Collect information from the available symbols.
992 SmallVector<WasmFunction, 4> Functions;
993 SmallVector<uint32_t, 4> TableElems;
994 SmallVector<WasmImport, 4> Imports;
995 SmallVector<WasmExport, 4> Exports;
996 SmallVector<std::pair<StringRef, uint32_t>, 4> SymbolFlags;
997 SmallPtrSet<const MCSymbolWasm *, 4> IsAddressTaken;
998 unsigned NumFuncImports = 0;
999 SmallVector<WasmDataSegment, 4> DataSegments;
1000 uint32_t StackPointerGlobal = 0;
1001 uint32_t DataSize = 0;
1002 bool HasStackPointer = false;
1004 // Populate the IsAddressTaken set.
1005 for (const WasmRelocationEntry &RelEntry : CodeRelocations) {
1006 switch (RelEntry.Type) {
1007 case wasm::R_WEBASSEMBLY_TABLE_INDEX_SLEB:
1008 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_SLEB:
1009 IsAddressTaken.insert(RelEntry.Symbol);
1015 for (const WasmRelocationEntry &RelEntry : DataRelocations) {
1016 switch (RelEntry.Type) {
1017 case wasm::R_WEBASSEMBLY_TABLE_INDEX_I32:
1018 case wasm::R_WEBASSEMBLY_MEMORY_ADDR_I32:
1019 IsAddressTaken.insert(RelEntry.Symbol);
1026 // Populate FunctionTypeIndices and Imports.
1027 for (const MCSymbol &S : Asm.symbols()) {
1028 const auto &WS = static_cast<const MCSymbolWasm &>(S);
1030 if (WS.isTemporary())
1033 if (WS.isFunction())
1034 registerFunctionType(WS);
1036 // If the symbol is not defined in this translation unit, import it.
1037 if (!WS.isDefined(/*SetUsed=*/false)) {
1039 Import.ModuleName = WS.getModuleName();
1040 Import.FieldName = WS.getName();
1042 if (WS.isFunction()) {
1043 Import.Kind = wasm::WASM_EXTERNAL_FUNCTION;
1044 Import.Type = getFunctionType(WS);
1045 SymbolIndices[&WS] = NumFuncImports;
1048 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL;
1049 Import.Type = int32_t(PtrType);
1050 SymbolIndices[&WS] = NumGlobalImports;
1054 Imports.push_back(Import);
1058 // In the special .global_variables section, we've encoded global
1059 // variables used by the function. Translate them into the Globals
1061 MCSectionWasm *GlobalVars = Ctx.getWasmSection(".global_variables", wasm::WASM_SEC_DATA);
1062 if (!GlobalVars->getFragmentList().empty()) {
1063 if (GlobalVars->getFragmentList().size() != 1)
1064 report_fatal_error("only one .global_variables fragment supported");
1065 const MCFragment &Frag = *GlobalVars->begin();
1066 if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data)
1067 report_fatal_error("only data supported in .global_variables");
1068 const auto &DataFrag = cast<MCDataFragment>(Frag);
1069 if (!DataFrag.getFixups().empty())
1070 report_fatal_error("fixups not supported in .global_variables");
1071 const SmallVectorImpl<char> &Contents = DataFrag.getContents();
1072 for (const uint8_t *p = (const uint8_t *)Contents.data(),
1073 *end = (const uint8_t *)Contents.data() + Contents.size();
1077 report_fatal_error("truncated global variable encoding");
1078 G.Type = wasm::ValType(int8_t(*p++));
1079 G.IsMutable = bool(*p++);
1080 G.HasImport = bool(*p++);
1085 Import.ModuleName = (const char *)p;
1086 const uint8_t *nul = (const uint8_t *)memchr(p, '\0', end - p);
1088 report_fatal_error("global module name must be nul-terminated");
1090 nul = (const uint8_t *)memchr(p, '\0', end - p);
1092 report_fatal_error("global base name must be nul-terminated");
1093 Import.FieldName = (const char *)p;
1096 Import.Kind = wasm::WASM_EXTERNAL_GLOBAL;
1097 Import.Type = int32_t(G.Type);
1099 G.ImportIndex = NumGlobalImports;
1102 Imports.push_back(Import);
1105 G.InitialValue = decodeSLEB128(p, &n);
1107 if ((ptrdiff_t)n > end - p)
1108 report_fatal_error("global initial value must be valid SLEB128");
1111 Globals.push_back(G);
1115 // In the special .stack_pointer section, we've encoded the stack pointer
1117 MCSectionWasm *StackPtr = Ctx.getWasmSection(".stack_pointer", wasm::WASM_SEC_DATA);
1118 if (!StackPtr->getFragmentList().empty()) {
1119 if (StackPtr->getFragmentList().size() != 1)
1120 report_fatal_error("only one .stack_pointer fragment supported");
1121 const MCFragment &Frag = *StackPtr->begin();
1122 if (Frag.hasInstructions() || Frag.getKind() != MCFragment::FT_Data)
1123 report_fatal_error("only data supported in .stack_pointer");
1124 const auto &DataFrag = cast<MCDataFragment>(Frag);
1125 if (!DataFrag.getFixups().empty())
1126 report_fatal_error("fixups not supported in .stack_pointer");
1127 const SmallVectorImpl<char> &Contents = DataFrag.getContents();
1128 if (Contents.size() != 4)
1129 report_fatal_error("only one entry supported in .stack_pointer");
1130 HasStackPointer = true;
1131 StackPointerGlobal = NumGlobalImports + *(const int32_t *)Contents.data();
1134 for (MCSection &Sec : Asm) {
1135 auto &Section = static_cast<MCSectionWasm &>(Sec);
1136 if (Section.getType() != wasm::WASM_SEC_DATA)
1139 DataSize = alignTo(DataSize, Section.getAlignment());
1140 DataSegments.emplace_back();
1141 WasmDataSegment &Segment = DataSegments.back();
1142 Segment.Name = Section.getSectionName();
1143 Segment.Offset = DataSize;
1144 Segment.Section = &Section;
1145 addData(Segment.Data, Section);
1146 Segment.Alignment = Section.getAlignment();
1148 DataSize += Segment.Data.size();
1149 Section.setMemoryOffset(Segment.Offset);
1152 // Handle regular defined and undefined symbols.
1153 for (const MCSymbol &S : Asm.symbols()) {
1154 // Ignore unnamed temporary symbols, which aren't ever exported, imported,
1155 // or used in relocations.
1156 if (S.isTemporary() && S.getName().empty())
1159 const auto &WS = static_cast<const MCSymbolWasm &>(S);
1160 DEBUG(dbgs() << "MCSymbol: '" << S << "'"
1161 << " isDefined=" << S.isDefined() << " isExternal="
1162 << S.isExternal() << " isTemporary=" << S.isTemporary()
1163 << " isFunction=" << WS.isFunction()
1164 << " isWeak=" << WS.isWeak()
1165 << " isVariable=" << WS.isVariable() << "\n");
1168 SymbolFlags.emplace_back(WS.getName(), wasm::WASM_SYMBOL_BINDING_WEAK);
1170 if (WS.isVariable())
1175 if (WS.isFunction()) {
1176 if (WS.isDefined(/*SetUsed=*/false)) {
1177 if (WS.getOffset() != 0)
1179 "function sections must contain one function each");
1181 if (WS.getSize() == 0)
1183 "function symbols must have a size set with .size");
1185 // A definition. Take the next available index.
1186 Index = NumFuncImports + Functions.size();
1188 // Prepare the function.
1190 Func.Type = getFunctionType(WS);
1192 SymbolIndices[&WS] = Index;
1193 Functions.push_back(Func);
1195 // An import; the index was assigned above.
1196 Index = SymbolIndices.find(&WS)->second;
1199 DEBUG(dbgs() << " -> function index: " << Index << "\n");
1201 // If needed, prepare the function to be called indirectly.
1202 if (IsAddressTaken.count(&WS) != 0) {
1203 IndirectSymbolIndices[&WS] = TableElems.size();
1204 DEBUG(dbgs() << " -> adding to table: " << TableElems.size() << "\n");
1205 TableElems.push_back(Index);
1208 if (WS.isTemporary() && !WS.getSize())
1211 if (!WS.isDefined(/*SetUsed=*/false))
1215 report_fatal_error("data symbols must have a size set with .size: " +
1219 if (!WS.getSize()->evaluateAsAbsolute(Size, Layout))
1220 report_fatal_error(".size expression must be evaluatable");
1222 // For each global, prepare a corresponding wasm global holding its
1223 // address. For externals these will also be named exports.
1224 Index = NumGlobalImports + Globals.size();
1225 auto &DataSection = static_cast<MCSectionWasm &>(WS.getSection());
1228 Global.Type = PtrType;
1229 Global.IsMutable = false;
1230 Global.HasImport = false;
1231 Global.InitialValue = DataSection.getMemoryOffset() + Layout.getSymbolOffset(WS);
1232 Global.ImportIndex = 0;
1233 SymbolIndices[&WS] = Index;
1234 DEBUG(dbgs() << " -> global index: " << Index << "\n");
1235 Globals.push_back(Global);
1238 // If the symbol is visible outside this translation unit, export it.
1239 if (WS.isDefined(/*SetUsed=*/false)) {
1241 Export.FieldName = WS.getName();
1242 Export.Index = Index;
1243 if (WS.isFunction())
1244 Export.Kind = wasm::WASM_EXTERNAL_FUNCTION;
1246 Export.Kind = wasm::WASM_EXTERNAL_GLOBAL;
1247 DEBUG(dbgs() << " -> export " << Exports.size() << "\n");
1248 Exports.push_back(Export);
1249 if (!WS.isExternal())
1250 SymbolFlags.emplace_back(WS.getName(), wasm::WASM_SYMBOL_BINDING_LOCAL);
1254 // Handle weak aliases. We need to process these in a separate pass because
1255 // we need to have processed the target of the alias before the alias itself
1256 // and the symbols are not necessarily ordered in this way.
1257 for (const MCSymbol &S : Asm.symbols()) {
1258 if (!S.isVariable())
1261 assert(S.isDefined(/*SetUsed=*/false));
1263 // Find the target symbol of this weak alias and export that index
1264 const auto &WS = static_cast<const MCSymbolWasm &>(S);
1265 const MCSymbolWasm *ResolvedSym = ResolveSymbol(WS);
1266 DEBUG(dbgs() << WS.getName() << ": weak alias of '" << *ResolvedSym << "'\n");
1267 assert(SymbolIndices.count(ResolvedSym) > 0);
1268 uint32_t Index = SymbolIndices.find(ResolvedSym)->second;
1269 DEBUG(dbgs() << " -> index:" << Index << "\n");
1271 SymbolIndices[&WS] = Index;
1273 Export.FieldName = WS.getName();
1274 Export.Index = Index;
1275 if (WS.isFunction())
1276 Export.Kind = wasm::WASM_EXTERNAL_FUNCTION;
1278 Export.Kind = wasm::WASM_EXTERNAL_GLOBAL;
1279 DEBUG(dbgs() << " -> export " << Exports.size() << "\n");
1280 Exports.push_back(Export);
1282 if (!WS.isExternal())
1283 SymbolFlags.emplace_back(WS.getName(), wasm::WASM_SYMBOL_BINDING_LOCAL);
1286 // Add types for indirect function calls.
1287 for (const WasmRelocationEntry &Fixup : CodeRelocations) {
1288 if (Fixup.Type != wasm::R_WEBASSEMBLY_TYPE_INDEX_LEB)
1291 registerFunctionType(*Fixup.Symbol);
1294 // Write out the Wasm header.
1297 writeTypeSection(FunctionTypes);
1298 writeImportSection(Imports);
1299 writeFunctionSection(Functions);
1300 writeTableSection(TableElems.size());
1301 writeMemorySection(DataSize);
1302 writeGlobalSection();
1303 writeExportSection(Exports);
1304 // TODO: Start Section
1305 writeElemSection(TableElems);
1306 writeCodeSection(Asm, Layout, Functions);
1307 writeDataSection(DataSegments);
1308 writeNameSection(Functions, Imports, NumFuncImports);
1309 writeCodeRelocSection();
1310 writeDataRelocSection();
1311 writeLinkingMetaDataSection(DataSegments, DataSize, SymbolFlags,
1312 HasStackPointer, StackPointerGlobal);
1314 // TODO: Translate the .comment section to the output.
1315 // TODO: Translate debug sections to the output.
1318 MCObjectWriter *llvm::createWasmObjectWriter(MCWasmObjectTargetWriter *MOTW,
1319 raw_pwrite_stream &OS) {
1320 return new WasmObjectWriter(MOTW, OS);