xref: /aosp_15_r20/external/llvm-libc/src/__support/HashTable/table.h (revision 71db0c75aadcf003ffe3238005f61d7618a3fead)
1 //===-- Resizable Monotonic HashTable ---------------------------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 
9 #ifndef LLVM_LIBC_SRC___SUPPORT_HASHTABLE_TABLE_H
10 #define LLVM_LIBC_SRC___SUPPORT_HASHTABLE_TABLE_H
11 
12 #include "include/llvm-libc-types/ENTRY.h"
13 #include "src/__support/CPP/bit.h" // bit_ceil
14 #include "src/__support/CPP/new.h"
15 #include "src/__support/HashTable/bitmask.h"
16 #include "src/__support/hash.h"
17 #include "src/__support/macros/attributes.h"
18 #include "src/__support/macros/config.h"
19 #include "src/__support/macros/optimization.h"
20 #include "src/__support/memory_size.h"
21 #include "src/string/memset.h"
22 #include "src/string/strcmp.h"
23 #include "src/string/strlen.h"
24 #include <stddef.h>
25 #include <stdint.h>
26 
27 namespace LIBC_NAMESPACE_DECL {
28 namespace internal {
29 
secondary_hash(uint64_t hash)30 LIBC_INLINE uint8_t secondary_hash(uint64_t hash) {
31   // top 7 bits of the hash.
32   return static_cast<uint8_t>(hash >> 57);
33 }
34 
35 // Probe sequence based on triangular numbers, which is guaranteed (since our
36 // table size is a power of two) to visit every group of elements exactly once.
37 //
38 // A triangular probe has us jump by 1 more group every time. So first we
39 // jump by 1 group (meaning we just continue our linear scan), then 2 groups
40 // (skipping over 1 group), then 3 groups (skipping over 2 groups), and so on.
41 //
42 // If we set sizeof(Group) to be one unit:
43 //               T[k] = sum {1 + 2 + ... + k} = k * (k + 1) / 2
44 // It is provable that T[k] mod 2^m generates a permutation of
45 //                0, 1, 2, 3, ..., 2^m - 2, 2^m - 1
46 // Detailed proof is available at:
47 // https://fgiesen.wordpress.com/2015/02/22/triangular-numbers-mod-2n/
48 struct ProbeSequence {
49   size_t position;
50   size_t stride;
51   size_t entries_mask;
52 
nextProbeSequence53   LIBC_INLINE size_t next() {
54     position += stride;
55     position &= entries_mask;
56     stride += sizeof(Group);
57     return position;
58   }
59 };
60 
61 // The number of entries is at least group width: we do not
62 // need to do the fixup when we set the control bytes.
63 // The number of entries is at least 8: we don't have to worry
64 // about special sizes when check the fullness of the table.
capacity_to_entries(size_t cap)65 LIBC_INLINE size_t capacity_to_entries(size_t cap) {
66   if (8 >= sizeof(Group) && cap < 8)
67     return 8;
68   if (16 >= sizeof(Group) && cap < 15)
69     return 16;
70   if (cap < sizeof(Group))
71     cap = sizeof(Group);
72   // overflow is always checked in allocate()
73   return cpp::bit_ceil(cap * 8 / 7);
74 }
75 
76 // The heap memory layout for N buckets HashTable is as follows:
77 //
78 //             =======================
79 //             |   N * Entry         |
80 //             ======================= <- align boundary
81 //             |   Header            |
82 //             ======================= <- align boundary (for fast resize)
83 //             |   (N + 1) * Byte    |
84 //             =======================
85 //
86 // The trailing group part is to make sure we can always load
87 // a whole group of control bytes.
88 
89 struct HashTable {
90   HashState state;
91   size_t entries_mask;    // number of buckets - 1
92   size_t available_slots; // less than capacity
93 private:
94   // How many entries are there in the table.
num_of_entriesHashTable95   LIBC_INLINE size_t num_of_entries() const { return entries_mask + 1; }
96 
97   // How many entries can we store in the table before resizing.
full_capacityHashTable98   LIBC_INLINE size_t full_capacity() const { return num_of_entries() / 8 * 7; }
99 
100   // The alignment of the whole memory area is the maximum of the alignment
101   // among the following types:
102   // - HashTable
103   // - ENTRY
104   // - Group
table_alignmentHashTable105   LIBC_INLINE constexpr static size_t table_alignment() {
106     size_t left_align = alignof(HashTable) > alignof(ENTRY) ? alignof(HashTable)
107                                                             : alignof(ENTRY);
108     return left_align > alignof(Group) ? left_align : alignof(Group);
109   }
110 
is_fullHashTable111   LIBC_INLINE bool is_full() const { return available_slots == 0; }
112 
offset_from_entriesHashTable113   LIBC_INLINE size_t offset_from_entries() const {
114     size_t entries_size = num_of_entries() * sizeof(ENTRY);
115     return entries_size +
116            SafeMemSize::offset_to(entries_size, table_alignment());
117   }
118 
offset_to_groupsHashTable119   LIBC_INLINE constexpr static size_t offset_to_groups() {
120     size_t header_size = sizeof(HashTable);
121     return header_size + SafeMemSize::offset_to(header_size, table_alignment());
122   }
123 
entryHashTable124   LIBC_INLINE ENTRY &entry(size_t i) {
125     return reinterpret_cast<ENTRY *>(this)[-i - 1];
126   }
127 
entryHashTable128   LIBC_INLINE const ENTRY &entry(size_t i) const {
129     return reinterpret_cast<const ENTRY *>(this)[-i - 1];
130   }
131 
controlHashTable132   LIBC_INLINE uint8_t &control(size_t i) {
133     uint8_t *ptr = reinterpret_cast<uint8_t *>(this) + offset_to_groups();
134     return ptr[i];
135   }
136 
controlHashTable137   LIBC_INLINE const uint8_t &control(size_t i) const {
138     const uint8_t *ptr =
139         reinterpret_cast<const uint8_t *>(this) + offset_to_groups();
140     return ptr[i];
141   }
142 
143   // We duplicate a group of control bytes to the end. Thus, it is possible that
144   // we need to set two control bytes at the same time.
set_ctrlHashTable145   LIBC_INLINE void set_ctrl(size_t index, uint8_t value) {
146     size_t index2 = ((index - sizeof(Group)) & entries_mask) + sizeof(Group);
147     control(index) = value;
148     control(index2) = value;
149   }
150 
findHashTable151   LIBC_INLINE size_t find(const char *key, uint64_t primary) {
152     uint8_t secondary = secondary_hash(primary);
153     ProbeSequence sequence{static_cast<size_t>(primary), 0, entries_mask};
154     while (true) {
155       size_t pos = sequence.next();
156       Group ctrls = Group::load(&control(pos));
157       IteratableBitMask masks = ctrls.match_byte(secondary);
158       for (size_t i : masks) {
159         size_t index = (pos + i) & entries_mask;
160         ENTRY &entry = this->entry(index);
161         if (LIBC_LIKELY(entry.key != nullptr && strcmp(entry.key, key) == 0))
162           return index;
163       }
164       BitMask available = ctrls.mask_available();
165       // Since there is no deletion, the first time we find an available slot
166       // it is also ready to be used as an insertion point. Therefore, we also
167       // return the first available slot we find. If such entry is empty, the
168       // key will be nullptr.
169       if (LIBC_LIKELY(available.any_bit_set())) {
170         size_t index =
171             (pos + available.lowest_set_bit_nonzero()) & entries_mask;
172         return index;
173       }
174     }
175   }
176 
oneshot_hashHashTable177   LIBC_INLINE uint64_t oneshot_hash(const char *key) const {
178     LIBC_NAMESPACE::internal::HashState hasher = state;
179     hasher.update(key, strlen(key));
180     return hasher.finish();
181   }
182 
183   // A fast insertion routine without checking if a key already exists.
184   // Nor does the routine check if the table is full.
185   // This is only to be used in grow() where we insert all existing entries
186   // into a new table. Hence, the requirements are naturally satisfied.
unsafe_insertHashTable187   LIBC_INLINE ENTRY *unsafe_insert(ENTRY item) {
188     uint64_t primary = oneshot_hash(item.key);
189     uint8_t secondary = secondary_hash(primary);
190     ProbeSequence sequence{static_cast<size_t>(primary), 0, entries_mask};
191     while (true) {
192       size_t pos = sequence.next();
193       Group ctrls = Group::load(&control(pos));
194       BitMask available = ctrls.mask_available();
195       if (available.any_bit_set()) {
196         size_t index =
197             (pos + available.lowest_set_bit_nonzero()) & entries_mask;
198         set_ctrl(index, secondary);
199         entry(index).key = item.key;
200         entry(index).data = item.data;
201         available_slots--;
202         return &entry(index);
203       }
204     }
205   }
206 
growHashTable207   LIBC_INLINE HashTable *grow() const {
208     size_t hint = full_capacity() + 1;
209     HashState state = this->state;
210     // migrate to a new random state
211     state.update(&hint, sizeof(hint));
212     HashTable *new_table = allocate(hint, state.finish());
213     // It is safe to call unsafe_insert() because we know that:
214     // - the new table has enough capacity to hold all the entries
215     // - there is no duplicate key in the old table
216     if (new_table != nullptr)
217       for (ENTRY e : *this)
218         new_table->unsafe_insert(e);
219     return new_table;
220   }
221 
insertHashTable222   LIBC_INLINE static ENTRY *insert(HashTable *&table, ENTRY item,
223                                    uint64_t primary) {
224     auto index = table->find(item.key, primary);
225     auto slot = &table->entry(index);
226     // SVr4 and POSIX.1-2001 specify that action is significant only for
227     // unsuccessful searches, so that an ENTER should not do anything
228     // for a successful search.
229     if (slot->key != nullptr)
230       return slot;
231 
232     // if table of full, we try to grow the table
233     if (table->is_full()) {
234       HashTable *new_table = table->grow();
235       // allocation failed, return nullptr to indicate failure
236       if (new_table == nullptr)
237         return nullptr;
238       // resized sccuessfully: clean up the old table and use the new one
239       deallocate(table);
240       table = new_table;
241       // it is still valid to use the fastpath insertion.
242       return table->unsafe_insert(item);
243     }
244 
245     table->set_ctrl(index, secondary_hash(primary));
246     slot->key = item.key;
247     slot->data = item.data;
248     table->available_slots--;
249     return slot;
250   }
251 
252 public:
deallocateHashTable253   LIBC_INLINE static void deallocate(HashTable *table) {
254     if (table) {
255       void *ptr =
256           reinterpret_cast<uint8_t *>(table) - table->offset_from_entries();
257       operator delete(ptr, std::align_val_t{table_alignment()});
258     }
259   }
260 
allocateHashTable261   LIBC_INLINE static HashTable *allocate(size_t capacity, uint64_t randomness) {
262     // check if capacity_to_entries overflows MAX_MEM_SIZE
263     if (capacity > size_t{1} << (8 * sizeof(size_t) - 1 - 3))
264       return nullptr;
265     SafeMemSize entries{capacity_to_entries(capacity)};
266     SafeMemSize entries_size = entries * SafeMemSize{sizeof(ENTRY)};
267     SafeMemSize align_boundary = entries_size.align_up(table_alignment());
268     SafeMemSize ctrl_sizes = entries + SafeMemSize{sizeof(Group)};
269     SafeMemSize header_size{offset_to_groups()};
270     SafeMemSize total_size =
271         (align_boundary + header_size + ctrl_sizes).align_up(table_alignment());
272     if (!total_size.valid())
273       return nullptr;
274     AllocChecker ac;
275 
276     void *mem = operator new(total_size, std::align_val_t{table_alignment()},
277                              ac);
278 
279     HashTable *table = reinterpret_cast<HashTable *>(
280         static_cast<uint8_t *>(mem) + align_boundary);
281     if (ac) {
282       table->entries_mask = entries - 1u;
283       table->available_slots = entries / 8 * 7;
284       table->state = HashState{randomness};
285       memset(&table->control(0), 0x80, ctrl_sizes);
286       memset(mem, 0, table->offset_from_entries());
287     }
288     return table;
289   }
290 
291   struct FullTableIterator {
292     size_t current_offset;
293     size_t remaining;
294     IteratableBitMask current_mask;
295     const HashTable &table;
296 
297     // It is fine to use remaining to represent the iterator:
298     // - this comparison only happens with the same table
299     // - hashtable will not be mutated during the iteration
300     LIBC_INLINE bool operator==(const FullTableIterator &other) const {
301       return remaining == other.remaining;
302     }
303     LIBC_INLINE bool operator!=(const FullTableIterator &other) const {
304       return remaining != other.remaining;
305     }
306 
307     LIBC_INLINE FullTableIterator &operator++() {
308       this->ensure_valid_group();
309       current_mask.remove_lowest_bit();
310       remaining--;
311       return *this;
312     }
313     LIBC_INLINE const ENTRY &operator*() {
314       this->ensure_valid_group();
315       return table.entry(
316           (current_offset + current_mask.lowest_set_bit_nonzero()) &
317           table.entries_mask);
318     }
319 
320   private:
ensure_valid_groupHashTable::FullTableIterator321     LIBC_INLINE void ensure_valid_group() {
322       while (!current_mask.any_bit_set()) {
323         current_offset += sizeof(Group);
324         // It is ensured that the load will only happen at aligned boundaries.
325         current_mask =
326             Group::load_aligned(&table.control(current_offset)).occupied();
327       }
328     }
329   };
330 
331   using value_type = ENTRY;
332   using iterator = FullTableIterator;
beginHashTable333   iterator begin() const {
334     return {0, full_capacity() - available_slots,
335             Group::load_aligned(&control(0)).occupied(), *this};
336   }
endHashTable337   iterator end() const { return {0, 0, {BitMask{0}}, *this}; }
338 
findHashTable339   LIBC_INLINE ENTRY *find(const char *key) {
340     uint64_t primary = oneshot_hash(key);
341     ENTRY &entry = this->entry(find(key, primary));
342     if (entry.key == nullptr)
343       return nullptr;
344     return &entry;
345   }
346 
insertHashTable347   LIBC_INLINE static ENTRY *insert(HashTable *&table, ENTRY item) {
348     uint64_t primary = table->oneshot_hash(item.key);
349     return insert(table, item, primary);
350   }
351 };
352 } // namespace internal
353 } // namespace LIBC_NAMESPACE_DECL
354 
355 #endif // LLVM_LIBC_SRC___SUPPORT_HASHTABLE_TABLE_H
356