1 //===--- MultiOnDiskHashTable.h - Merged set of hash tables -----*- C++ -*-===//
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 provides a hash table data structure suitable for incremental and
11 // distributed storage across a set of files.
13 // Multiple hash tables from different files are implicitly merged to improve
14 // performance, and on reload the merged table will override those from other
17 //===----------------------------------------------------------------------===//
18 #ifndef LLVM_CLANG_LIB_SERIALIZATION_MULTIONDISKHASHTABLE_H
19 #define LLVM_CLANG_LIB_SERIALIZATION_MULTIONDISKHASHTABLE_H
21 #include "llvm/ADT/PointerUnion.h"
22 #include "llvm/Support/EndianStream.h"
23 #include "llvm/Support/OnDiskHashTable.h"
26 namespace serialization {
30 /// \brief A collection of on-disk hash tables, merged when relevant for performance.
31 template<typename Info> class MultiOnDiskHashTable {
33 /// A handle to a file, used when overriding tables.
34 typedef typename Info::file_type file_type;
35 /// A pointer to an on-disk representation of the hash table.
36 typedef const unsigned char *storage_type;
38 typedef typename Info::external_key_type external_key_type;
39 typedef typename Info::internal_key_type internal_key_type;
40 typedef typename Info::data_type data_type;
41 typedef typename Info::data_type_builder data_type_builder;
42 typedef unsigned hash_value_type;
45 /// \brief A hash table stored on disk.
47 typedef llvm::OnDiskIterableChainedHashTable<Info> HashTable;
52 OnDiskTable(file_type File, unsigned NumBuckets, unsigned NumEntries,
53 storage_type Buckets, storage_type Payload, storage_type Base,
56 Table(NumBuckets, NumEntries, Buckets, Payload, Base, InfoObj) {}
60 std::vector<file_type> Files;
61 llvm::DenseMap<internal_key_type, data_type> Data;
64 typedef llvm::PointerUnion<OnDiskTable*, MergedTable*> Table;
65 typedef llvm::TinyPtrVector<void*> TableVector;
67 /// \brief The current set of on-disk and merged tables.
68 /// We manually store the opaque value of the Table because TinyPtrVector
69 /// can't cope with holding a PointerUnion directly.
70 /// There can be at most one MergedTable in this vector, and if present,
71 /// it is the first table.
74 /// \brief Files corresponding to overridden tables that we've not yet
76 llvm::TinyPtrVector<file_type> PendingOverrides;
78 struct AsOnDiskTable {
79 typedef OnDiskTable *result_type;
80 result_type operator()(void *P) const {
81 return Table::getFromOpaqueValue(P).template get<OnDiskTable *>();
84 typedef llvm::mapped_iterator<TableVector::iterator, AsOnDiskTable>
86 typedef llvm::iterator_range<table_iterator> table_range;
88 /// \brief The current set of on-disk tables.
89 table_range tables() {
90 auto Begin = Tables.begin(), End = Tables.end();
93 return llvm::make_range(llvm::map_iterator(Begin, AsOnDiskTable()),
94 llvm::map_iterator(End, AsOnDiskTable()));
97 MergedTable *getMergedTable() const {
98 // If we already have a merged table, it's the first one.
99 return Tables.empty() ? nullptr : Table::getFromOpaqueValue(*Tables.begin())
100 .template dyn_cast<MergedTable*>();
103 /// \brief Delete all our current on-disk tables.
105 for (auto *T : tables())
107 if (auto *M = getMergedTable())
112 void removeOverriddenTables() {
113 llvm::DenseSet<file_type> Files;
114 Files.insert(PendingOverrides.begin(), PendingOverrides.end());
115 // Explicitly capture Files to work around an MSVC 2015 rejects-valid bug.
116 auto ShouldRemove = [&Files](void *T) -> bool {
117 auto *ODT = Table::getFromOpaqueValue(T).template get<OnDiskTable *>();
118 bool Remove = Files.count(ODT->File);
123 Tables.erase(std::remove_if(tables().begin().getCurrent(), Tables.end(),
126 PendingOverrides.clear();
130 MergedTable *Merged = getMergedTable();
132 Merged = new MergedTable;
134 // Read in all the tables and merge them together.
135 // FIXME: Be smarter about which tables we merge.
136 for (auto *ODT : tables()) {
137 auto &HT = ODT->Table;
138 Info &InfoObj = HT.getInfoObj();
140 for (auto I = HT.data_begin(), E = HT.data_end(); I != E; ++I) {
141 auto *LocalPtr = I.getItem();
143 // FIXME: Don't rely on the OnDiskHashTable format here.
144 auto L = InfoObj.ReadKeyDataLength(LocalPtr);
145 const internal_key_type &Key = InfoObj.ReadKey(LocalPtr, L.first);
146 data_type_builder ValueBuilder(Merged->Data[Key]);
147 InfoObj.ReadDataInto(Key, LocalPtr + L.first, L.second,
151 Merged->Files.push_back(ODT->File);
156 Tables.push_back(Table(Merged).getOpaqueValue());
159 /// The generator is permitted to read our merged table.
160 template<typename ReaderInfo, typename WriterInfo>
161 friend class MultiOnDiskHashTableGenerator;
164 MultiOnDiskHashTable() {}
165 MultiOnDiskHashTable(MultiOnDiskHashTable &&O)
166 : Tables(std::move(O.Tables)),
167 PendingOverrides(std::move(O.PendingOverrides)) {
170 MultiOnDiskHashTable &operator=(MultiOnDiskHashTable &&O) {
174 Tables = std::move(O.Tables);
176 PendingOverrides = std::move(O.PendingOverrides);
179 ~MultiOnDiskHashTable() { clear(); }
181 /// \brief Add the table \p Data loaded from file \p File.
182 void add(file_type File, storage_type Data, Info InfoObj = Info()) {
183 using namespace llvm::support;
184 storage_type Ptr = Data;
186 uint32_t BucketOffset = endian::readNext<uint32_t, little, unaligned>(Ptr);
188 // Read the list of overridden files.
189 uint32_t NumFiles = endian::readNext<uint32_t, little, unaligned>(Ptr);
190 // FIXME: Add a reserve() to TinyPtrVector so that we don't need to make
191 // an additional copy.
192 llvm::SmallVector<file_type, 16> OverriddenFiles;
193 OverriddenFiles.reserve(NumFiles);
194 for (/**/; NumFiles != 0; --NumFiles)
195 OverriddenFiles.push_back(InfoObj.ReadFileRef(Ptr));
196 PendingOverrides.insert(PendingOverrides.end(), OverriddenFiles.begin(),
197 OverriddenFiles.end());
199 // Read the OnDiskChainedHashTable header.
200 storage_type Buckets = Data + BucketOffset;
201 auto NumBucketsAndEntries =
202 OnDiskTable::HashTable::readNumBucketsAndEntries(Buckets);
204 // Register the table.
205 Table NewTable = new OnDiskTable(File, NumBucketsAndEntries.first,
206 NumBucketsAndEntries.second,
207 Buckets, Ptr, Data, std::move(InfoObj));
208 Tables.push_back(NewTable.getOpaqueValue());
211 /// \brief Find and read the lookup results for \p EKey.
212 data_type find(const external_key_type &EKey) {
215 if (!PendingOverrides.empty())
216 removeOverriddenTables();
218 if (Tables.size() > static_cast<unsigned>(Info::MaxTables))
221 internal_key_type Key = Info::GetInternalKey(EKey);
222 auto KeyHash = Info::ComputeHash(Key);
224 if (MergedTable *M = getMergedTable()) {
225 auto It = M->Data.find(Key);
226 if (It != M->Data.end())
230 data_type_builder ResultBuilder(Result);
232 for (auto *ODT : tables()) {
233 auto &HT = ODT->Table;
234 auto It = HT.find_hashed(Key, KeyHash);
236 HT.getInfoObj().ReadDataInto(Key, It.getDataPtr(), It.getDataLen(),
243 /// \brief Read all the lookup results into a single value. This only makes
244 /// sense if merging values across keys is meaningful.
245 data_type findAll() {
247 data_type_builder ResultBuilder(Result);
249 if (!PendingOverrides.empty())
250 removeOverriddenTables();
252 if (MergedTable *M = getMergedTable()) {
253 for (auto &KV : M->Data)
254 Info::MergeDataInto(KV.second, ResultBuilder);
257 for (auto *ODT : tables()) {
258 auto &HT = ODT->Table;
259 Info &InfoObj = HT.getInfoObj();
260 for (auto I = HT.data_begin(), E = HT.data_end(); I != E; ++I) {
261 auto *LocalPtr = I.getItem();
263 // FIXME: Don't rely on the OnDiskHashTable format here.
264 auto L = InfoObj.ReadKeyDataLength(LocalPtr);
265 const internal_key_type &Key = InfoObj.ReadKey(LocalPtr, L.first);
266 InfoObj.ReadDataInto(Key, LocalPtr + L.first, L.second, ResultBuilder);
274 /// \brief Writer for the on-disk hash table.
275 template<typename ReaderInfo, typename WriterInfo>
276 class MultiOnDiskHashTableGenerator {
277 typedef MultiOnDiskHashTable<ReaderInfo> BaseTable;
278 typedef llvm::OnDiskChainedHashTableGenerator<WriterInfo> Generator;
283 MultiOnDiskHashTableGenerator() : Gen() {}
285 void insert(typename WriterInfo::key_type_ref Key,
286 typename WriterInfo::data_type_ref Data, WriterInfo &Info) {
287 Gen.insert(Key, Data, Info);
290 void emit(llvm::SmallVectorImpl<char> &Out, WriterInfo &Info,
291 const BaseTable *Base) {
292 using namespace llvm::support;
293 llvm::raw_svector_ostream OutStream(Out);
295 // Write our header information.
297 endian::Writer<little> Writer(OutStream);
299 // Reserve four bytes for the bucket offset.
300 Writer.write<uint32_t>(0);
302 if (auto *Merged = Base ? Base->getMergedTable() : nullptr) {
303 // Write list of overridden files.
304 Writer.write<uint32_t>(Merged->Files.size());
305 for (const auto &F : Merged->Files)
306 Info.EmitFileRef(OutStream, F);
308 // Add all merged entries from Base to the generator.
309 for (auto &KV : Merged->Data) {
310 if (!Gen.contains(KV.first, Info))
311 Gen.insert(KV.first, Info.ImportData(KV.second), Info);
314 Writer.write<uint32_t>(0);
318 // Write the table itself.
319 uint32_t BucketOffset = Gen.Emit(OutStream, Info);
321 // Replace the first four bytes with the bucket offset.
322 endian::write32le(Out.data(), BucketOffset);
326 } // end namespace clang::serialization
327 } // end namespace clang