APINotes: add initial stub of APINotesWriter

[compnerd]

Summary

This adds the skeleton for serializing out the APINotes data from the
APINotes. The writer uses a private implementation pattern to reduce
the exposed surface to just the programmatic representation of the
APINotes and not expose the details of the bitcode encoding. The format
itself is not considered stable and should only be accessed through the
APINotes Reader and Writer types.

[compnerd]

clang/lib/APINotes/APINotesWriter.cpp

35

Please ignore this #if-defed portion, it is not intended for upstream, but I need to keep this working downstream. I will remove this before merging.

235

Similar

[compnerd]

Ping

[martong]

In D92797#2452497, @compnerd wrote:

Ping

I am terribly sorry for keeping you waiting always. Thanks for your patience.

clang/include/clang/APINotes/APINotesWriter.h

1

How are we going to use this class? I mean what are the use cases to call the Writer from inside Clang?
I understand that the Reader might be used to read up the serialized APINotes file.

What will be the usual process of using an APINotes file? Is it like:

1) Edit `APINotesX.yaml` by hand (by a library author)
2) Serialize `APINotesX.yaml` -> `APINotesX.serialized` (Would this require a special Clang invocation?)
3) Run Clang for the usual tasks (analysis, etc) but with feeding it with `APINotesX.serialized`

Would it be possible to feed the raw APINotesX.yaml to Clang and skip the whole serialize/deserialize steps? If yes, then maybe we should put more focus on that first (but maybe it's already too late since we added the serialization format and the reader).

[compnerd]

clang/include/clang/APINotes/APINotesWriter.h

1

As to the question of the expected usage: if a user wishes to alter the behavior of a library (the library is missing nullability annotation!? I want that checking!) they add a supplement APINotes for the library, and pass that to the compiler when using the library via flags. When clang goes through and finds that the cache is not populated, it will process the API Notes and compile it. Then it apply the API Notes to the compilation as it goes through (looking up the interfaces it encounters).

This class is going to be used by the APINotes Compiler, which will read the YAML input and then serialize it out to bitcode encoded format. This will occur implicitly by the compiler consulting a cache (similar in spirit to how modules are cached). If the APINotes Manager finds the cached copy that is used, otherwise the contents can be processed and emitted into the cache. This basically is similar in spirit to a PCH/PCM.

I dont mind switching to focus on a different part of the API if it:
a) makes it easier to understand and review, OR
b) makes it easier to test

The intent here is to chunk the pieces so that there is a way to process the changes logically and with testing. I was simply taking an approach to make progress, but as long as we can make progress, Im not really tied to a specific path through this.

In fact, I was struggling with how to create a custom compiler to serialize, deserialize the input so that it can be verified (namely a question of how to query/compare - is the best that we can do diff against a golden master?)

[compnerd]

Gentle post-holiday ping :)

[compnerd]

Ping x 2

[compnerd]

Ping x 3

[martong]

In D92797#2505254, @compnerd wrote:

Ping x 3

Sorry, just came back from holidays.
Generally, I have no major concerns.

My minor concern is that it is not easy to add a new Note. There are many different places to extend the code (i.e. adding a new DenseMap instance, adding two new write function prototypes and their definitions). I am not sure if this could be simpler though.

clang/lib/APINotes/APINotesWriter.cpp

35

Could you please remove from the patch then?

764

Some other *TableInfo classes do not have ComputeHash. E.g. GlobalVariableTableInfo. Why?
Where do we use ComputeHash?

[compnerd]

Welcome back, hope you had a good time off. Thanks for the review!

clang/lib/APINotes/APINotesWriter.cpp

35

Yeah, this is not meant to be committed with that in place, I just need to keep that around for testing locally. I will absolutely be removing them.

764

Excellent observation. The ComputeHash is not directly used, it is implemented as part of a trait used for the hashing, used in LLVM as part of llvm::OnDiskChainedHashTableGenerator which is used throughout this file. The reason that GlobalVariableTableInfo does not use it is because the versioned items derive from VersionedInfo which provides a shared implementation.

[egorzhdan]

Apologies for pinging an old review. Would it be possible to land this to facilitate future work to upstream API notes support?

[egorzhdan]

@compnerd @martong would you be OK with merging this patch?

[ronlieb]

hi, this seems to be breaking out amdgpu opemp and hip buildbots
https://lab.llvm.org/staging/#/builders/247/builds/4967

[dyung]

In D92797#4596289, @ronlieb wrote:

hi, this seems to be breaking out amdgpu opemp and hip buildbots
https://lab.llvm.org/staging/#/builders/247/builds/4967

And also some GCC based bots:
https://lab.llvm.org/buildbot/#/builders/139/builds/47730
https://lab.llvm.org/buildbot/#/builders/247/builds/7889

clang/include/clang/APINotes/APINotesWriter.h

+//===-- APINotesWriter.h - API Notes Writer ---------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_CLANG_APINOTES_WRITER_H
+#define LLVM_CLANG_APINOTES_WRITER_H
+
+#include "llvm/ADT/StringRef.h"
+
+namespace clang {
+class FileEntry;
+
+namespace api_notes {
+class APINotesWriter {
+  class Implementation;
+  std::unique_ptr<Implementation> Implementation;
+
+public:
+  APINotesWriter(llvm::StringRef ModuleName, const FileEntry *SF);
+  ~APINotesWriter();
+
+  APINotesWriter(const APINotesWriter &) = delete;
+  APINotesWriter &operator=(const APINotesWriter &) = delete;
+
+  void writeToStream(llvm::raw_ostream &OS);
+};
+} // namespace api_notes
+} // namespace clang
+
+#endif

clang/lib/APINotes/APINotesFormat.h

 /// A stored Objective-C selector.
 struct StoredObjCSelector {
   unsigned NumPieces;
   llvm::SmallVector<IdentifierID, 2> Identifiers;
 };
 } // namespace api_notes
 } // namespace clang
 
+namespace llvm {
+template <> struct DenseMapInfo<clang::api_notes::StoredObjCSelector> {
+  typedef DenseMapInfo<unsigned> UnsignedInfo;
+
+  static inline clang::api_notes::StoredObjCSelector getEmptyKey() {
+    return clang::api_notes::StoredObjCSelector{UnsignedInfo::getEmptyKey(),
+                                                {}};
+  }
+
+  static inline clang::api_notes::StoredObjCSelector getTombstoneKey() {
+    return clang::api_notes::StoredObjCSelector{UnsignedInfo::getTombstoneKey(),
+                                                {}};
+  }
+
+  static unsigned
+  getHashValue(const clang::api_notes::StoredObjCSelector &Selector) {
+    auto hash = llvm::hash_value(Selector.NumPieces);
+    hash = hash_combine(hash, Selector.Identifiers.size());
+    for (auto piece : Selector.Identifiers)
+      hash = hash_combine(hash, static_cast<unsigned>(piece));
+    // FIXME: Mix upper/lower 32-bit values together to produce
+    // unsigned rather than truncating.
+    return hash;
+  }
+
+  static bool isEqual(const clang::api_notes::StoredObjCSelector &LHS,
+                      const clang::api_notes::StoredObjCSelector &RHS) {
+    return LHS.NumPieces == RHS.NumPieces && LHS.Identifiers == RHS.Identifiers;
+  }
+};
+} // namespace llvm
+
 #endif

clang/lib/APINotes/APINotesWriter.cpp

+//===-- APINotesWriter.cpp - API Notes Writer -------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "clang/APINotes/APINotesWriter.h"
+#include "APINotesFormat.h"
+#include "clang/APINotes/Types.h"
+#include "clang/Basic/FileManager.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/StringMap.h"
+#include "llvm/Bitstream/BitstreamWriter.h"
+#include "llvm/Support/DJB.h"
+#include "llvm/Support/OnDiskHashTable.h"
+#include "llvm/Support/VersionTuple.h"
+
+namespace clang {
+namespace api_notes {
+class APINotesWriter::Implementation {
+  template <typename T>
+  using VersionedSmallVector =
+      llvm::SmallVector<std::pair<llvm::VersionTuple, T>, 1>;
+
+  std::string ModuleName;
+  const FileEntry *SourceFile;
+
+  /// Scratch space for bitstream writing.
+  llvm::SmallVector<uint64_t, 64> Scratch;
+
+#if defined(__APPLE__) && defined(SWIFT_DOWNSTREAM)
+  bool SwiftImportAsMember = false;
+#endif
+
+  /// Mapping from strings to identifier IDs.
+  llvm::StringMap<IdentifierID> IdentifierIDs;
+
+  /// Information about Objective-C contexts (classes or protocols).
+  ///
+  /// Indexed by the identifier ID and a bit indication whether we're looking
+  /// for a class (0) or protocol (1) and provides both the context ID and
+  /// information describing the context within that module.
+  llvm::DenseMap<std::pair<unsigned, char>,
+                 std::pair<unsigned, VersionedSmallVector<ObjCContextInfo>>>
+      ObjCContexts;
+
+  /// Information about Objective-C properties.
+  ///
+  /// Indexed by the context ID, property name, and whether this is an
+  /// instance property.
+  llvm::DenseMap<
+      std::tuple<unsigned, unsigned, char>,
+      llvm::SmallVector<std::pair<VersionTuple, ObjCPropertyInfo>, 1>>
+      ObjCProperties;
+
+  /// Information about Objective-C methods.
+  ///
+  /// Indexed by the context ID, selector ID, and Boolean (stored as a char)
+  /// indicating whether this is a class or instance method.
+  llvm::DenseMap<std::tuple<unsigned, unsigned, char>,
+                 llvm::SmallVector<std::pair<VersionTuple, ObjCMethodInfo>, 1>>
+      ObjCMethods;
+
+  /// Mapping from selectors to selector ID.
+  llvm::DenseMap<StoredObjCSelector, SelectorID> SelectorIDs;
+
+  /// Information about global variables.
+  ///
+  /// Indexed by the identifier ID.
+  llvm::DenseMap<unsigned, llvm::SmallVector<
+                               std::pair<VersionTuple, GlobalVariableInfo>, 1>>
+      GlobalVariables;
+
+  /// Information about global functions.
+  ///
+  /// Indexed by the identifier ID.
+  llvm::DenseMap<unsigned, llvm::SmallVector<
+                               std::pair<VersionTuple, GlobalFunctionInfo>, 1>>
+      GlobalFunctions;
+
+  /// Information about enumerators.
+  ///
+  /// Indexed by the identifier ID.
+  llvm::DenseMap<
+      unsigned, llvm::SmallVector<std::pair<VersionTuple, EnumConstantInfo>, 1>>
+      EnumConstants;
+
+  /// Information about tags.
+  ///
+  /// Indexed by the identifier ID.
+  llvm::DenseMap<unsigned,
+                 llvm::SmallVector<std::pair<VersionTuple, TagInfo>, 1>>
+      Tags;
+
+  /// Information about typedefs.
+  ///
+  /// Indexed by the identifier ID.
+  llvm::DenseMap<unsigned,
+                 llvm::SmallVector<std::pair<VersionTuple, TypedefInfo>, 1>>
+      Typedefs;
+
+  void writeBlockInfoBlock(llvm::BitstreamWriter &Stream);
+  void writeControlBlock(llvm::BitstreamWriter &Stream);
+  void writeIdentifierBlock(llvm::BitstreamWriter &Stream);
+  void writeObjCContextBlock(llvm::BitstreamWriter &Stream);
+  void writeObjCPropertyBlock(llvm::BitstreamWriter &Stream);
+  void writeObjCMethodBlock(llvm::BitstreamWriter &Stream);
+  void writeObjCSelectorBlock(llvm::BitstreamWriter &Stream);
+  void writeGlobalVariableBlock(llvm::BitstreamWriter &Stream);
+  void writeGlobalFunctionBlock(llvm::BitstreamWriter &Stream);
+  void writeEnumConstantBlock(llvm::BitstreamWriter &Stream);
+  void writeTagBlock(llvm::BitstreamWriter &Stream);
+  void writeTypedefBlock(llvm::BitstreamWriter &Stream);
+
+public:
+  Implementation(llvm::StringRef ModuleName, const FileEntry *SF)
+      : ModuleName(std::string(ModuleName)), SourceFile(SF) {}
+
+  void writeToStream(llvm::raw_ostream &OS);
+};
+
+void APINotesWriter::Implementation::writeToStream(llvm::raw_ostream &OS) {
+  llvm::SmallVector<char, 0> Buffer;
+
+  {
+    llvm::BitstreamWriter Stream(Buffer);
+
+    // Emit the signature.
+    for (unsigned char Byte : API_NOTES_SIGNATURE)
+      Stream.Emit(Byte, 8);
+
+    // Emit the blocks.
+    writeBlockInfoBlock(Stream);
+    writeControlBlock(Stream);
+    writeIdentifierBlock(Stream);
+    writeObjCContextBlock(Stream);
+    writeObjCPropertyBlock(Stream);
+    writeObjCMethodBlock(Stream);
+    writeObjCSelectorBlock(Stream);
+    writeGlobalVariableBlock(Stream);
+    writeGlobalFunctionBlock(Stream);
+    writeEnumConstantBlock(Stream);
+    writeTagBlock(Stream);
+    writeTypedefBlock(Stream);
+  }
+
+  OS.write(Buffer.data(), Buffer.size());
+  OS.flush();
+}
+
+namespace {
+/// Record the name of a block.
+void emitBlockID(llvm::BitstreamWriter &Stream, unsigned ID,
+                 llvm::StringRef Name) {
+  Stream.EmitRecord(llvm::bitc::BLOCKINFO_CODE_SETBID,
+                    llvm::ArrayRef<unsigned>{ID});
+
+  // Emit the block name if present.
+  if (Name.empty())
+    return;
+  Stream.EmitRecord(
+      llvm::bitc::BLOCKINFO_CODE_BLOCKNAME,
+      llvm::ArrayRef<unsigned char>(
+          const_cast<unsigned char *>(
+              reinterpret_cast<const unsigned char *>(Name.data())),
+          Name.size()));
+}
+
+/// Record the name of a record within a block.
+void emitRecordID(llvm::BitstreamWriter &Stream, unsigned ID,
+                  llvm::StringRef Name) {
+  assert(ID < 256 && "can't fit record ID in next to name");
+
+  llvm::SmallVector<unsigned char, 64> Buffer;
+  Buffer.resize(Name.size() + 1);
+  Buffer[0] = ID;
+  memcpy(Buffer.data() + 1, Name.data(), Name.size());
+
+  Stream.EmitRecord(llvm::bitc::BLOCKINFO_CODE_SETRECORDNAME, Buffer);
+}
+} // namespace
+
+void APINotesWriter::Implementation::writeBlockInfoBlock(
+    llvm::BitstreamWriter &Stream) {
+  llvm::BCBlockRAII Scope(Stream, llvm::bitc::BLOCKINFO_BLOCK_ID, 2);
+
+#define BLOCK(Block) emitBlockID(Stream, Block##_ID, #Block)
+#define BLOCK_RECORD(NameSpace, Block)                                         \
+  emitRecordID(Stream, NameSpace::Block, #Block)
+  BLOCK(CONTROL_BLOCK);
+  BLOCK_RECORD(control_block, METADATA);
+  BLOCK_RECORD(control_block, MODULE_NAME);
+
+  BLOCK(IDENTIFIER_BLOCK);
+  BLOCK_RECORD(identifier_block, IDENTIFIER_DATA);
+
+  BLOCK(OBJC_CONTEXT_BLOCK);
+  BLOCK_RECORD(objc_context_block, OBJC_CONTEXT_ID_DATA);
+
+  BLOCK(OBJC_PROPERTY_BLOCK);
+  BLOCK_RECORD(objc_property_block, OBJC_PROPERTY_DATA);
+
+  BLOCK(OBJC_METHOD_BLOCK);
+  BLOCK_RECORD(objc_method_block, OBJC_METHOD_DATA);
+
+  BLOCK(OBJC_SELECTOR_BLOCK);
+  BLOCK_RECORD(objc_selector_block, OBJC_SELECTOR_DATA);
+
+  BLOCK(GLOBAL_VARIABLE_BLOCK);
+  BLOCK_RECORD(global_variable_block, GLOBAL_VARIABLE_DATA);
+
+  BLOCK(GLOBAL_FUNCTION_BLOCK);
+  BLOCK_RECORD(global_function_block, GLOBAL_FUNCTION_DATA);
+#undef BLOCK_RECORD
+#undef BLOCK
+}
+
+void APINotesWriter::Implementation::writeControlBlock(
+    llvm::BitstreamWriter &Stream) {
+  llvm::BCBlockRAII Scope(Stream, CONTROL_BLOCK_ID, 3);
+
+  control_block::MetadataLayout Metadata(Stream);
+  Metadata.emit(Scratch, VERSION_MAJOR, VERSION_MINOR);
+
+  control_block::ModuleNameLayout ModuleName(Stream);
+  ModuleName.emit(Scratch, this->ModuleName);
+
+#if defined(__APPLE__) && defined(SWIFT_DOWNSTREAM)
+  if (SwiftInferImportAsMember) {
+    control_block::ModuleOptionsLayout ModuleOptions(Stream);
+    ModuleOptions.emit(Scratch, SwiftInferImportAsMember);
+  }
+#endif
+
+  if (SourceFile) {
+    control_block::SourceFileLayout SourceFile(Stream);
+    SourceFile.emit(Scratch, this->SourceFile->getSize(),
+                    this->SourceFile->getModificationTime());
+  }
+}
+
+namespace {
+/// Used to serialize the on-disk identifier table.
+class IdentifierTableInfo {
+public:
+  using key_type = StringRef;
+  using key_type_ref = key_type;
+  using data_type = IdentifierID;
+  using data_type_ref = const data_type &;
+  using hash_value_type = uint32_t;
+  using offset_type = unsigned;
+
+  hash_value_type ComputeHash(key_type_ref Key) { return llvm::djbHash(Key); }
+
+  std::pair<unsigned, unsigned>
+  EmitKeyDataLength(raw_ostream &OS, key_type_ref Key, data_type_ref) {
+    uint32_t KeyLength = Key.size();
+    uint32_t DataLength = sizeof(uint32_t);
+
+    llvm::support::endian::Writer writer(OS, llvm::support::little);
+    writer.write<uint16_t>(KeyLength);
+    writer.write<uint16_t>(DataLength);
+    return {KeyLength, DataLength};
+  }
+
+  void EmitKey(raw_ostream &OS, key_type_ref Key, unsigned) { OS << Key; }
+
+  void EmitData(raw_ostream &OS, key_type_ref, data_type_ref Data, unsigned) {
+    llvm::support::endian::Writer writer(OS, llvm::support::little);
+    writer.write<uint32_t>(Data);
+  }
+};
+} // namespace
+
+void APINotesWriter::Implementation::writeIdentifierBlock(
+    llvm::BitstreamWriter &Stream) {
+  llvm::BCBlockRAII restoreBlock(Stream, IDENTIFIER_BLOCK_ID, 3);
+
+  if (IdentifierIDs.empty())
+    return;
+
+  llvm::SmallString<4096> HashTableBlob;
+  uint32_t Offset;
+  {
+    llvm::OnDiskChainedHashTableGenerator<IdentifierTableInfo> Generator;
+    for (auto &II : IdentifierIDs)
+      Generator.insert(II.first(), II.second);
+
+    llvm::raw_svector_ostream BlobStream(HashTableBlob);
+    // Make sure that no bucket is at offset 0
+    llvm::support::endian::write<uint32_t>(BlobStream, 0,
+                                           llvm::support::little);
+    Offset = Generator.Emit(BlobStream);
+  }
+
+  identifier_block::IdentifierDataLayout IdentifierData(Stream);
+  IdentifierData.emit(Scratch, Offset, HashTableBlob);
+}
+
+namespace {
+/// Used to serialize the on-disk Objective-C context table.
+class ObjCContextIDTableInfo {
+public:
+  using key_type = std::pair<unsigned, char>; // identifier ID, is-protocol
+  using key_type_ref = key_type;
+  using data_type = unsigned;
+  using data_type_ref = const data_type &;
+  using hash_value_type = size_t;
+  using offset_type = unsigned;
+
+  hash_value_type ComputeHash(key_type_ref Key) {
+    return static_cast<size_t>(llvm::hash_value(Key));
+  }
+
+  std::pair<unsigned, unsigned> EmitKeyDataLength(raw_ostream &OS, key_type_ref,
+                                                  data_type_ref) {
+    uint32_t KeyLength = sizeof(uint32_t) + 1;
+    uint32_t DataLength = sizeof(uint32_t);
+
+    llvm::support::endian::Writer writer(OS, llvm::support::little);
+    writer.write<uint16_t>(KeyLength);
+    writer.write<uint16_t>(DataLength);
+    return {KeyLength, DataLength};
+  }
+
+  void EmitKey(raw_ostream &OS, key_type_ref Key, unsigned) {
+    llvm::support::endian::Writer writer(OS, llvm::support::little);
+    writer.write<uint32_t>(Key.first);
+    writer.write<uint8_t>(Key.second);
+  }
+
+  void EmitData(raw_ostream &OS, key_type_ref, data_type_ref Data, unsigned) {
+    llvm::support::endian::Writer writer(OS, llvm::support::little);
+    writer.write<uint32_t>(Data);
+  }
+};
+
+/// Localized helper to make a type dependent, thwarting template argument
+/// deduction.
+template <typename T> struct MakeDependent { typedef T Type; };
+
+/// Retrieve the serialized size of the given VersionTuple, for use in
+/// on-disk hash tables.
+unsigned getVersionTupleSize(const VersionTuple &VT) {
+  unsigned size = sizeof(uint8_t) + /*major*/ sizeof(uint32_t);
+  if (VT.getMinor())
+    size += sizeof(uint32_t);
+  if (VT.getSubminor())
+    size += sizeof(uint32_t);
+  if (VT.getBuild())
+    size += sizeof(uint32_t);
+  return size;
+}
+
+/// Determine the size of an array of versioned information,
+template <typename T>
+unsigned getVersionedInfoSize(
+    const llvm::SmallVectorImpl<std::pair<llvm::VersionTuple, T>> &VI,
+    llvm::function_ref<unsigned(const typename MakeDependent<T>::Type &)>
+        getInfoSize) {
+  unsigned result = sizeof(uint16_t); // # of elements
+  for (const auto &E : VI) {
+    result += getVersionTupleSize(E.first);
+    result += getInfoSize(E.second);
+  }
+  return result;
+}
+
+/// Emit a serialized representation of a version tuple.
+void emitVersionTuple(raw_ostream &OS, const VersionTuple &VT) {
+  llvm::support::endian::Writer writer(OS, llvm::support::little);
+
+  // First byte contains the number of components beyond the 'major' component.
+  uint8_t descriptor;
+  if (VT.getBuild())
+    descriptor = 3;
+  else if (VT.getSubminor())
+    descriptor = 2;
+  else if (VT.getMinor())
+    descriptor = 1;
+  else
+    descriptor = 0;
+  writer.write<uint8_t>(descriptor);
+
+  // Write the components.
+  writer.write<uint32_t>(VT.getMajor());
+  if (auto minor = VT.getMinor())
+    writer.write<uint32_t>(*minor);
+  if (auto subminor = VT.getSubminor())
+    writer.write<uint32_t>(*subminor);
+  if (auto build = VT.getBuild())
+    writer.write<uint32_t>(*build);
+}
+
+/// Emit versioned information.
+template <typename T>
+void emitVersionedInfo(
+    raw_ostream &OS, llvm::SmallVectorImpl<std::pair<VersionTuple, T>> &VI,
+    llvm::function_ref<void(raw_ostream &,
+                            const typename MakeDependent<T>::Type &)>
+        emitInfo) {
+  std::sort(VI.begin(), VI.end(),
+            [](const std::pair<VersionTuple, T> &LHS,
+               const std::pair<VersionTuple, T> &RHS) -> bool {
+              assert(LHS.first != RHS.first &&
+                     "two entries for the same version");
+              return LHS.first < RHS.first;
+            });
+
+  llvm::support::endian::Writer writer(OS, llvm::support::little);
+  writer.write<uint16_t>(VI.size());
+  for (const auto &E : VI) {
+    emitVersionTuple(OS, E.first);
+    emitInfo(OS, E.second);
+  }
+}
+
+/// On-disk hash table info key base for handling versioned data.
+template <typename Derived, typename KeyType, typename UnversionedDataType>
+class VersionedTableInfo {
+  Derived &asDerived() { return *static_cast<Derived *>(this); }
+
+  const Derived &asDerived() const {
+    return *static_cast<const Derived *>(this);
+  }
+
+public:
+  using key_type = KeyType;
+  using key_type_ref = key_type;
+  using data_type =
+      llvm::SmallVector<std::pair<llvm::VersionTuple, UnversionedDataType>, 1>;
+  using data_type_ref = data_type &;
+  using hash_value_type = size_t;
+  using offset_type = unsigned;
+
+  hash_value_type ComputeHash(key_type_ref Key) {
+    return llvm::hash_value(Key);
+  }
+
+  std::pair<unsigned, unsigned>
+  EmitKeyDataLength(raw_ostream &OS, key_type_ref Key, data_type_ref Data) {
+    uint32_t KeyLength = asDerived().getKeyLength(Key);
+    uint32_t DataLength =
+        getVersionedInfoSize(Data, [this](const UnversionedDataType &UI) {
+          return asDerived().getUnversionedInfoSize(UI);
+        });
+
+    llvm::support::endian::Writer writer(OS, llvm::support::little);
+    writer.write<uint16_t>(KeyLength);
+    writer.write<uint16_t>(DataLength);
+    return {KeyLength, DataLength};
+  }
+
+  void EmitData(raw_ostream &OS, key_type_ref, data_type_ref Data, unsigned) {
+    emitVersionedInfo(
+        OS, Data, [this](llvm::raw_ostream &OS, const UnversionedDataType &UI) {
+          asDerived().emitUnversionedInfo(OS, UI);
+        });
+  }
+};
+
+/// Emit a serialized representation of the common entity information.
+void emitCommonEntityInfo(raw_ostream &OS, const CommonEntityInfo &CEI) {
+  llvm::support::endian::Writer writer(OS, llvm::support::little);
+
+  uint8_t payload = 0;
+  if (auto swiftPrivate = CEI.isSwiftPrivate()) {
+    payload |= 0x01;
+    if (*swiftPrivate)
+      payload |= 0x02;
+  }
+  payload <<= 1;
+  payload |= CEI.Unavailable;
+  payload <<= 1;
+  payload |= CEI.UnavailableInSwift;
+
+  writer.write<uint8_t>(payload);
+
+  writer.write<uint16_t>(CEI.UnavailableMsg.size());
+  OS.write(CEI.UnavailableMsg.c_str(), CEI.UnavailableMsg.size());
+
+  writer.write<uint16_t>(CEI.SwiftName.size());
+  OS.write(CEI.SwiftName.c_str(), CEI.SwiftName.size());
+}
+
+/// Retrieve the serialized size of the given CommonEntityInfo, for use in
+/// on-disk hash tables.
+unsigned getCommonEntityInfoSize(const CommonEntityInfo &CEI) {
+  return 5 + CEI.UnavailableMsg.size() + CEI.SwiftName.size();
+}
+
+// Retrieve the serialized size of the given CommonTypeInfo, for use
+// in on-disk hash tables.
+unsigned getCommonTypeInfoSize(const CommonTypeInfo &CTI) {
+  return 2 + (CTI.getSwiftBridge() ? CTI.getSwiftBridge()->size() : 0) + 2 +
+         (CTI.getNSErrorDomain() ? CTI.getNSErrorDomain()->size() : 0) +
+         getCommonEntityInfoSize(CTI);
+}
+
+/// Emit a serialized representation of the common type information.
+void emitCommonTypeInfo(raw_ostream &OS, const CommonTypeInfo &CTI) {
+  emitCommonEntityInfo(OS, CTI);
+
+  llvm::support::endian::Writer writer(OS, llvm::support::little);
+  if (auto swiftBridge = CTI.getSwiftBridge()) {
+    writer.write<uint16_t>(swiftBridge->size() + 1);
+    OS.write(swiftBridge->c_str(), swiftBridge->size());
+  } else {
+    writer.write<uint16_t>(0);
+  }
+  if (auto nsErrorDomain = CTI.getNSErrorDomain()) {
+    writer.write<uint16_t>(nsErrorDomain->size() + 1);
+    OS.write(nsErrorDomain->c_str(), CTI.getNSErrorDomain()->size());
+  } else {
+    writer.write<uint16_t>(0);
+  }
+}
+
+/// Used to serialize the on-disk Objective-C property table.
+class ObjCContextInfoTableInfo
+    : public VersionedTableInfo<ObjCContextInfoTableInfo, unsigned,
+                                ObjCContextInfo> {
+public:
+  unsigned getKeyLength(key_type_ref) { return sizeof(uint32_t); }
+
+  void EmitKey(raw_ostream &OS, key_type_ref Key, unsigned) {
+    llvm::support::endian::Writer writer(OS, llvm::support::little);
+    writer.write<uint32_t>(Key);
+  }
+
+  unsigned getUnversionedInfoSize(const ObjCContextInfo &OCI) {
+    return getCommonTypeInfoSize(OCI) + 1;
+  }
+
+  void emitUnversionedInfo(raw_ostream &OS, const ObjCContextInfo &OCI) {
+    emitCommonTypeInfo(OS, OCI);
+
+    uint8_t payload = 0;
+    if (auto swiftImportAsNonGeneric = OCI.getSwiftImportAsNonGeneric())
+      payload |= (0x01 << 1) | swiftImportAsNonGeneric.getValue();
+    payload <<= 2;
+    if (auto swiftObjCMembers = OCI.getSwiftObjCMembers())
+      payload |= (0x01 << 1) | swiftObjCMembers.getValue();
+    payload <<= 3;
+    if (auto nullable = OCI.getDefaultNullability())
+      payload |= (0x01 << 2) | static_cast<uint8_t>(*nullable);
+    payload = (payload << 1) | (OCI.hasDesignatedInits() ? 1 : 0);
+
+    OS << payload;
+  }
+};
+} // namespace
+
+void APINotesWriter::Implementation::writeObjCContextBlock(
+    llvm::BitstreamWriter &Stream) {
+  llvm::BCBlockRAII restoreBlock(Stream, OBJC_CONTEXT_BLOCK_ID, 3);
+
+  if (ObjCContexts.empty())
+    return;
+
+  {
+    llvm::SmallString<4096> HashTableBlob;
+    uint32_t Offset;
+    {
+      llvm::OnDiskChainedHashTableGenerator<ObjCContextIDTableInfo> Generator;
+      for (auto &OC : ObjCContexts)
+        Generator.insert(OC.first, OC.second.first);
+
+      llvm::raw_svector_ostream BlobStream(HashTableBlob);
+      // Make sure that no bucket is at offset 0
+      llvm::support::endian::write<uint32_t>(BlobStream, 0,
+                                             llvm::support::little);
+      Offset = Generator.Emit(BlobStream);
+    }
+
+    objc_context_block::ObjCContextIDLayout ObjCContextID(Stream);
+    ObjCContextID.emit(Scratch, Offset, HashTableBlob);
+  }
+
+  {
+    llvm::SmallString<4096> HashTableBlob;
+    uint32_t Offset;
+    {
+      llvm::OnDiskChainedHashTableGenerator<ObjCContextInfoTableInfo> Generator;
+      for (auto &OC : ObjCContexts)
+        Generator.insert(OC.second.first, OC.second.second);
+
+      llvm::raw_svector_ostream BlobStream(HashTableBlob);
+      // Make sure that no bucket is at offset 0
+      llvm::support::endian::write<uint32_t>(BlobStream, 0,
+                                             llvm::support::little);
+      Offset = Generator.Emit(BlobStream);
+    }
+
+    objc_context_block::ObjCContextInfoLayout ObjCContextInfo(Stream);
+    ObjCContextInfo.emit(Scratch, Offset, HashTableBlob);
+  }
+}
+
+namespace {
+/// Retrieve the serialized size of the given VariableInfo, for use in
+/// on-disk hash tables.
+unsigned getVariableInfoSize(const VariableInfo &VI) {
+  return 2 + getCommonEntityInfoSize(VI) + 2 + VI.getType().size();
+}
+
+/// Emit a serialized representation of the variable information.
+void emitVariableInfo(raw_ostream &OS, const VariableInfo &VI) {
+  emitCommonEntityInfo(OS, VI);
+
+  uint8_t bytes[2] = {0, 0};
+  if (auto nullable = VI.getNullability()) {
+    bytes[0] = 1;
+    bytes[1] = static_cast<uint8_t>(*nullable);
+  } else {
+    // Nothing to do.
+  }
+
+  OS.write(reinterpret_cast<const char *>(bytes), 2);
+
+  llvm::support::endian::Writer writer(OS, llvm::support::little);
+  writer.write<uint16_t>(VI.getType().size());
+  OS.write(VI.getType().data(), VI.getType().size());
+}
+
+/// Used to serialize the on-disk Objective-C property table.
+class ObjCPropertyTableInfo
+    : public VersionedTableInfo<ObjCPropertyTableInfo,
+                                std::tuple<unsigned, unsigned, char>,
+                                ObjCPropertyInfo> {
+public:
+  unsigned getKeyLength(key_type_ref) {
+    return sizeof(uint32_t) + sizeof(uint32_t) + sizeof(uint8_t);
+  }
+
+  void EmitKey(raw_ostream &OS, key_type_ref Key, unsigned) {
+    llvm::support::endian::Writer writer(OS, llvm::support::little);
+    writer.write<uint32_t>(std::get<0>(Key));
+    writer.write<uint32_t>(std::get<1>(Key));
+    writer.write<uint8_t>(std::get<2>(Key));
+  }
+
+  unsigned getUnversionedInfoSize(const ObjCPropertyInfo &OPI) {
+    return getVariableInfoSize(OPI) + 1;
+  }
+
+  void emitUnversionedInfo(raw_ostream &OS, const ObjCPropertyInfo &OPI) {
+    emitVariableInfo(OS, OPI);
+
+    uint8_t flags = 0;
+    if (Optional<bool> value = OPI.getSwiftImportAsAccessors()) {
+      flags |= 1 << 0;
+      flags |= value.getValue() << 1;
+    }
+    OS << flags;
+  }
+};
+} // namespace
+
+void APINotesWriter::Implementation::writeObjCPropertyBlock(
+    llvm::BitstreamWriter &Stream) {
+  llvm::BCBlockRAII Scope(Stream, OBJC_PROPERTY_BLOCK_ID, 3);
+
+  if (ObjCProperties.empty())
+    return;
+
+  {
+    llvm::SmallString<4096> HashTableBlob;
+    uint32_t Offset;
+    {
+      llvm::OnDiskChainedHashTableGenerator<ObjCPropertyTableInfo> Generator;
+      for (auto &OP : ObjCProperties)
+        Generator.insert(OP.first, OP.second);
+
+      llvm::raw_svector_ostream BlobStream(HashTableBlob);
+      // Make sure that no bucket is at offset 0
+      llvm::support::endian::write<uint32_t>(BlobStream, 0,
+                                             llvm::support::little);
+      Offset = Generator.Emit(BlobStream);
+    }
+
+    objc_property_block::ObjCPropertyDataLayout ObjCPropertyData(Stream);
+    ObjCPropertyData.emit(Scratch, Offset, HashTableBlob);
+  }
+}
+
+namespace {
+unsigned getFunctionInfoSize(const FunctionInfo &);
+void emitFunctionInfo(llvm::raw_ostream &, const FunctionInfo &);
+
+/// Used to serialize the on-disk Objective-C method table.
+class ObjCMethodTableInfo
+    : public VersionedTableInfo<ObjCMethodTableInfo,
+                                std::tuple<unsigned, unsigned, char>,
+                                ObjCMethodInfo> {
+public:
+  unsigned getKeyLength(key_type_ref) {
+    return sizeof(uint32_t) + sizeof(uint32_t) + sizeof(uint8_t);
+  }
+
+  void EmitKey(raw_ostream &OS, key_type_ref Key, unsigned) {
+    llvm::support::endian::Writer writer(OS, llvm::support::little);
+    writer.write<uint32_t>(std::get<0>(Key));
+    writer.write<uint32_t>(std::get<1>(Key));
+    writer.write<uint8_t>(std::get<2>(Key));
+  }
+
+  unsigned getUnversionedInfoSize(const ObjCMethodInfo &OMI) {
+    return getFunctionInfoSize(OMI) + 1;
+  }
+
+  void emitUnversionedInfo(raw_ostream &OS, const ObjCMethodInfo &OMI) {
+    uint8_t flags = 0;
+    llvm::support::endian::Writer writer(OS, llvm::support::little);
+    flags = (flags << 1) | OMI.DesignatedInit;
+    flags = (flags << 1) | OMI.RequiredInit;
+    writer.write<uint8_t>(flags);
+
+    emitFunctionInfo(OS, OMI);
+  }
+};
+} // namespace
+
+void APINotesWriter::Implementation::writeObjCMethodBlock(
+    llvm::BitstreamWriter &Stream) {
+  llvm::BCBlockRAII Scope(Stream, OBJC_METHOD_BLOCK_ID, 3);
+
+  if (ObjCMethods.empty())
+    return;
+
+  {
+    llvm::SmallString<4096> HashTableBlob;
+    uint32_t Offset;
+    {
+      llvm::OnDiskChainedHashTableGenerator<ObjCMethodTableInfo> Generator;
+      for (auto &OM : ObjCMethods)
+        Generator.insert(OM.first, OM.second);
+
+      llvm::raw_svector_ostream BlobStream(HashTableBlob);
+      // Make sure that no bucket is at offset 0
+      llvm::support::endian::write<uint32_t>(BlobStream, 0,
+                                             llvm::support::little);
+      Offset = Generator.Emit(BlobStream);
+    }
+
+    objc_method_block::ObjCMethodDataLayout ObjCMethodData(Stream);
+    ObjCMethodData.emit(Scratch, Offset, HashTableBlob);
+  }
+}
+
+namespace {
+/// Used to serialize the on-disk Objective-C selector table.
+class ObjCSelectorTableInfo {
+public:
+  using key_type = StoredObjCSelector;
+  using key_type_ref = const key_type &;
+  using data_type = SelectorID;
+  using data_type_ref = data_type;
+  using hash_value_type = unsigned;
+  using offset_type = unsigned;
+
+  hash_value_type ComputeHash(key_type_ref Key) {
+    return llvm::DenseMapInfo<StoredObjCSelector>::getHashValue(Key);
+  }
+
+  std::pair<unsigned, unsigned>
+  EmitKeyDataLength(raw_ostream &OS, key_type_ref Key, data_type_ref) {
+    uint32_t KeyLength =
+        sizeof(uint16_t) + sizeof(uint32_t) * Key.Identifiers.size();
+    uint32_t DataLength = sizeof(uint32_t);
+
+    llvm::support::endian::Writer writer(OS, llvm::support::little);
+    writer.write<uint16_t>(KeyLength);
+    writer.write<uint16_t>(DataLength);
+    return {KeyLength, DataLength};
+  }
+
+  void EmitKey(raw_ostream &OS, key_type_ref Key, unsigned) {
+    llvm::support::endian::Writer writer(OS, llvm::support::little);
+    writer.write<uint16_t>(Key.NumPieces);
+    for (auto Identifier : Key.Identifiers)
+      writer.write<uint32_t>(Identifier);
+  }
+
+  void EmitData(raw_ostream &OS, key_type_ref, data_type_ref Data, unsigned) {
+    llvm::support::endian::Writer writer(OS, llvm::support::little);
+    writer.write<uint32_t>(Data);
+  }
+};
+} // namespace
+
+void APINotesWriter::Implementation::writeObjCSelectorBlock(
+    llvm::BitstreamWriter &Stream) {
+  llvm::BCBlockRAII Scope(Stream, OBJC_SELECTOR_BLOCK_ID, 3);
+
+  if (SelectorIDs.empty())
+    return;
+
+  {
+    llvm::SmallString<4096> HashTableBlob;
+    uint32_t Offset;
+    {
+      llvm::OnDiskChainedHashTableGenerator<ObjCSelectorTableInfo> Generator;
+      for (auto &S : SelectorIDs)
+        Generator.insert(S.first, S.second);
+
+      llvm::raw_svector_ostream BlobStream(HashTableBlob);
+      // Make sure that no bucket is at offset 0
+      llvm::support::endian::write<uint32_t>(BlobStream, 0,
+                                             llvm::support::little);
+      Offset = Generator.Emit(BlobStream);
+    }
+
+    objc_selector_block::ObjCSelectorDataLayout ObjCSelectorData(Stream);
+    ObjCSelectorData.emit(Scratch, Offset, HashTableBlob);
+  }
+}
+
+namespace {
+/// Used to serialize the on-disk global variable table.
+class GlobalVariableTableInfo
+    : public VersionedTableInfo<GlobalVariableTableInfo, unsigned,
+                                GlobalVariableInfo> {
+public:
+  unsigned getKeyLength(key_type_ref) { return sizeof(uint32_t); }
+
+  void EmitKey(raw_ostream &OS, key_type_ref Key, unsigned) {
+    llvm::support::endian::Writer writer(OS, llvm::support::little);
+    writer.write<uint32_t>(Key);
+  }
+
+  unsigned getUnversionedInfoSize(const GlobalVariableInfo &GVI) {
+    return getVariableInfoSize(GVI);
+  }
+
+  void emitUnversionedInfo(raw_ostream &OS, const GlobalVariableInfo &GVI) {
+    emitVariableInfo(OS, GVI);
+  }
+};
+} // namespace
+
+void APINotesWriter::Implementation::writeGlobalVariableBlock(
+    llvm::BitstreamWriter &Stream) {
+  llvm::BCBlockRAII Scope(Stream, GLOBAL_VARIABLE_BLOCK_ID, 3);
+
+  if (GlobalVariables.empty())
+    return;
+
+  {
+    llvm::SmallString<4096> HashTableBlob;
+    uint32_t Offset;
+    {
+      llvm::OnDiskChainedHashTableGenerator<GlobalVariableTableInfo> Generator;
+      for (auto &GV : GlobalVariables)
+        Generator.insert(GV.first, GV.second);
+
+      llvm::raw_svector_ostream BlobStream(HashTableBlob);
+      // Make sure that no bucket is at offset 0
+      llvm::support::endian::write<uint32_t>(BlobStream, 0,
+                                             llvm::support::little);
+      Offset = Generator.Emit(BlobStream);
+    }
+
+    global_variable_block::GlobalVariableDataLayout GlobalVariableData(Stream);
+    GlobalVariableData.emit(Scratch, Offset, HashTableBlob);
+  }
+}
+
+namespace {
+unsigned getParamInfoSize(const ParamInfo &PI) {
+  return getVariableInfoSize(PI) + 1;
+}
+
+void emitParamInfo(raw_ostream &OS, const ParamInfo &PI) {
+  emitVariableInfo(OS, PI);
+
+  uint8_t flags = 0;
+  if (auto noescape = PI.isNoEscape()) {
+    flags |= 0x01;
+    if (*noescape)
+      flags |= 0x02;
+  }
+  flags <<= 3;
+  if (auto RCC = PI.getRetainCountConvention())
+    flags |= static_cast<uint8_t>(RCC.getValue()) + 1;
+
+  llvm::support::endian::Writer writer(OS, llvm::support::little);
+  writer.write<uint8_t>(flags);
+}
+
+/// Retrieve the serialized size of the given FunctionInfo, for use in on-disk
+/// hash tables.
+unsigned getFunctionInfoSize(const FunctionInfo &FI) {
+  unsigned size = getCommonEntityInfoSize(FI) + 2 + sizeof(uint64_t);
+  size += sizeof(uint16_t);
+  for (const auto &P : FI.Params)
+    size += getParamInfoSize(P);
+  size += sizeof(uint16_t) + FI.ResultType.size();
+  return size;
+}
+
+/// Emit a serialized representation of the function information.
+static void emitFunctionInfo(raw_ostream &OS, const FunctionInfo &FI) {
+  emitCommonEntityInfo(OS, FI);
+
+  uint8_t flags = 0;
+  flags |= FI.NullabilityAudited;
+  flags <<= 3;
+  if (auto RCC = FI.getRetainCountConvention())
+    flags |= static_cast<uint8_t>(RCC.getValue()) + 1;
+
+  llvm::support::endian::Writer writer(OS, llvm::support::little);
+
+  writer.write<uint8_t>(flags);
+  writer.write<uint8_t>(FI.NumAdjustedNullable);
+  writer.write<uint64_t>(FI.NullabilityPayload);
+
+  writer.write<uint16_t>(FI.Params.size());
+  for (const auto &PI : FI.Params)
+    emitParamInfo(OS, PI);
+
+  writer.write<uint16_t>(FI.ResultType.size());
+  writer.write(ArrayRef<char>{FI.ResultType.data(), FI.ResultType.size()});
+}
+
+/// Used to serialize the on-disk global function table.
+class GlobalFunctionTableInfo
+    : public VersionedTableInfo<GlobalFunctionTableInfo, unsigned,
+                                GlobalFunctionInfo> {
+public:
+  unsigned getKeyLength(key_type_ref) { return sizeof(uint32_t); }
+
+  void EmitKey(raw_ostream &OS, key_type_ref Key, unsigned) {
+    llvm::support::endian::Writer writer(OS, llvm::support::little);
+    writer.write<uint32_t>(Key);
+  }
+
+  unsigned getUnversionedInfoSize(const GlobalFunctionInfo &GFI) {
+    return getFunctionInfoSize(GFI);
+  }
+
+  void emitUnversionedInfo(raw_ostream &OS, const GlobalFunctionInfo &GFI) {
+    emitFunctionInfo(OS, GFI);
+  }
+};
+} // namespace
+
+void APINotesWriter::Implementation::writeGlobalFunctionBlock(
+    llvm::BitstreamWriter &Stream) {
+  llvm::BCBlockRAII Scope(Stream, GLOBAL_FUNCTION_BLOCK_ID, 3);
+
+  if (GlobalFunctions.empty())
+    return;
+
+  {
+    llvm::SmallString<4096> HashTableBlob;
+    uint32_t Offset;
+    {
+      llvm::OnDiskChainedHashTableGenerator<GlobalFunctionTableInfo> Generator;
+      for (auto &F : GlobalFunctions)
+        Generator.insert(F.first, F.second);
+
+      llvm::raw_svector_ostream BlobStream(HashTableBlob);
+      // Make sure that no bucket is at offset 0
+      llvm::support::endian::write<uint32_t>(BlobStream, 0,
+                                             llvm::support::little);
+      Offset = Generator.Emit(BlobStream);
+    }
+
+    global_function_block::GlobalFunctionDataLayout GlobalFunctionData(Stream);
+    GlobalFunctionData.emit(Scratch, Offset, HashTableBlob);
+  }
+}
+
+namespace {
+/// Used to serialize the on-disk global enum constant.
+class EnumConstantTableInfo
+    : public VersionedTableInfo<EnumConstantTableInfo, unsigned,
+                                EnumConstantInfo> {
+public:
+  unsigned getKeyLength(key_type_ref) { return sizeof(uint32_t); }
+
+  void EmitKey(raw_ostream &OS, key_type_ref Key, unsigned) {
+    llvm::support::endian::Writer writer(OS, llvm::support::little);
+    writer.write<uint32_t>(Key);
+  }
+
+  unsigned getUnversionedInfoSize(const EnumConstantInfo &ECI) {
+    return getCommonEntityInfoSize(ECI);
+  }
+
+  void emitUnversionedInfo(raw_ostream &OS, const EnumConstantInfo &ECI) {
+    emitCommonEntityInfo(OS, ECI);
+  }
+};
+} // namespace
+
+void APINotesWriter::Implementation::writeEnumConstantBlock(
+    llvm::BitstreamWriter &Stream) {
+  llvm::BCBlockRAII Scope(Stream, ENUM_CONSTANT_BLOCK_ID, 3);
+
+  if (EnumConstants.empty())
+    return;
+
+  {
+    llvm::SmallString<4096> HashTableBlob;
+    uint32_t Offset;
+    {
+      llvm::OnDiskChainedHashTableGenerator<EnumConstantTableInfo> Generator;
+      for (auto &EC : EnumConstants)
+        Generator.insert(EC.first, EC.second);
+
+      llvm::raw_svector_ostream BlobStream(HashTableBlob);
+      // Make sure that no bucket is at offset 0
+      llvm::support::endian::write<uint32_t>(BlobStream, 0,
+                                             llvm::support::little);
+      Offset = Generator.Emit(BlobStream);
+    }
+
+    enum_constant_block::EnumConstantDataLayout EnumConstantData(Stream);
+    EnumConstantData.emit(Scratch, Offset, HashTableBlob);
+  }
+}
+
+namespace {
+template <typename Derived, typename UnversionedDataType>
+class CommonTypeTableInfo
+    : public VersionedTableInfo<Derived, unsigned, UnversionedDataType> {
+public:
+  using key_type_ref = typename CommonTypeTableInfo::key_type_ref;
+
+  unsigned getKeyLength(key_type_ref) { return sizeof(IdentifierID); }
+
+  void EmitKey(raw_ostream &OS, key_type_ref Key, unsigned) {
+    llvm::support::endian::Writer writer(OS, llvm::support::little);
+    writer.write<IdentifierID>(Key);
+  }
+
+  unsigned getUnversionedInfoSize(const UnversionedDataType &UDT) {
+    return getCommonTypeInfoSize(UDT);
+  }
+
+  void emitUnversionedInfo(raw_ostream &OS, const UnversionedDataType &UDT) {
+    emitCommonTypeInfo(OS, UDT);
+  }
+};
+
+/// Used to serialize the on-disk tag table.
+class TagTableInfo : public CommonTypeTableInfo<TagTableInfo, TagInfo> {
+public:
+  unsigned getUnversionedInfoSize(const TagInfo &TI) {
+    return 1 + getCommonTypeInfoSize(TI);
+  }
+
+  void emitUnversionedInfo(raw_ostream &OS, const TagInfo &TI) {
+    llvm::support::endian::Writer writer(OS, llvm::support::little);
+
+    uint8_t Flags = 0;
+    if (auto extensibility = TI.EnumExtensibility) {
+      Flags |= static_cast<uint8_t>(extensibility.getValue()) + 1;
+      assert((Flags < (1 << 2)) && "must fit in two bits");
+    }
+
+    Flags <<= 2;
+    if (llvm::Optional<bool> Value = TI.isFlagEnum())
+      Flags |= (Value.getValue() << 1 | 1 << 0);
+
+    writer.write<uint8_t>(Flags);
+
+    emitCommonTypeInfo(OS, TI);
+  }
+};
+} // namespace
+
+void APINotesWriter::Implementation::writeTagBlock(
+    llvm::BitstreamWriter &Stream) {
+  llvm::BCBlockRAII Scope(Stream, TAG_BLOCK_ID, 3);
+
+  if (Tags.empty())
+    return;
+
+  {
+    llvm::SmallString<4096> HashTableBlob;
+    uint32_t Offset;
+    {
+      llvm::OnDiskChainedHashTableGenerator<TagTableInfo> Generator;
+      for (auto &T : Tags)
+        Generator.insert(T.first, T.second);
+
+      llvm::raw_svector_ostream BlobStream(HashTableBlob);
+      // Make sure that no bucket is at offset 0
+      llvm::support::endian::write<uint32_t>(BlobStream, 0,
+                                             llvm::support::little);
+      Offset = Generator.Emit(BlobStream);
+    }
+
+    tag_block::TagDataLayout TagData(Stream);
+    TagData.emit(Scratch, Offset, HashTableBlob);
+  }
+}
+
+namespace {
+/// Used to serialize the on-disk typedef table.
+class TypedefTableInfo
+    : public CommonTypeTableInfo<TypedefTableInfo, TypedefInfo> {
+public:
+  unsigned getUnversionedInfoSize(const TypedefInfo &TI) {
+    return 1 + getCommonTypeInfoSize(TI);
+  }
+
+  void emitUnversionedInfo(raw_ostream &OS, const TypedefInfo &TI) {
+    llvm::support::endian::Writer writer(OS, llvm::support::little);
+
+    uint8_t Flags = 0;
+    if (auto swiftWrapper = TI.SwiftWrapper)
+      Flags |= static_cast<uint8_t>(*swiftWrapper) + 1;
+
+    writer.write<uint8_t>(Flags);
+
+    emitCommonTypeInfo(OS, TI);
+  }
+};
+} // namespace
+
+void APINotesWriter::Implementation::writeTypedefBlock(
+    llvm::BitstreamWriter &Stream) {
+  llvm::BCBlockRAII Scope(Stream, TYPEDEF_BLOCK_ID, 3);
+
+  if (Typedefs.empty())
+    return;
+
+  {
+    llvm::SmallString<4096> HashTableBlob;
+    uint32_t Offset;
+    {
+      llvm::OnDiskChainedHashTableGenerator<TypedefTableInfo> Generator;
+      for (auto &T : Typedefs)
+        Generator.insert(T.first, T.second);
+
+      llvm::raw_svector_ostream BlobStream(HashTableBlob);
+      // Make sure that no bucket is at offset 0
+      llvm::support::endian::write<uint32_t>(BlobStream, 0,
+                                             llvm::support::little);
+      Offset = Generator.Emit(BlobStream);
+    }
+
+    typedef_block::TypedefDataLayout TypedefData(Stream);
+    TypedefData.emit(Scratch, Offset, HashTableBlob);
+  }
+}
+
+// APINotesWriter
+
+APINotesWriter::APINotesWriter(llvm::StringRef ModuleName, const FileEntry *SF)
+    : Implementation(new class Implementation(ModuleName, SF)) {}
+
+APINotesWriter::~APINotesWriter() = default;
+
+void APINotesWriter::writeToStream(llvm::raw_ostream &OS) {
+  Implementation->writeToStream(OS);
+}
+} // namespace api_notes
+} // namespace clang

clang/lib/APINotes/CMakeLists.txt

 set(LLVM_LINK_COMPONENTS
   Support)
 add_clang_library(clangAPINotes
   APINotesTypes.cpp
+  APINotesWriter.cpp
   APINotesYAMLCompiler.cpp
   LINK_LIBS
     clangBasic)