/src/abseil-cpp/absl/container/internal/raw_hash_set.cc
Line | Count | Source |
1 | | // Copyright 2018 The Abseil Authors. |
2 | | // |
3 | | // Licensed under the Apache License, Version 2.0 (the "License"); |
4 | | // you may not use this file except in compliance with the License. |
5 | | // You may obtain a copy of the License at |
6 | | // |
7 | | // https://www.apache.org/licenses/LICENSE-2.0 |
8 | | // |
9 | | // Unless required by applicable law or agreed to in writing, software |
10 | | // distributed under the License is distributed on an "AS IS" BASIS, |
11 | | // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
12 | | // See the License for the specific language governing permissions and |
13 | | // limitations under the License. |
14 | | |
15 | | #include "absl/container/internal/raw_hash_set.h" |
16 | | |
17 | | #include <algorithm> |
18 | | #include <atomic> |
19 | | #include <cassert> |
20 | | #include <cstddef> |
21 | | #include <cstdint> |
22 | | #include <cstring> |
23 | | #include <memory> |
24 | | #include <tuple> |
25 | | #include <type_traits> |
26 | | #include <utility> |
27 | | |
28 | | #include "absl/base/attributes.h" |
29 | | #include "absl/base/config.h" |
30 | | #include "absl/base/dynamic_annotations.h" |
31 | | #include "absl/base/internal/endian.h" |
32 | | #include "absl/base/internal/raw_logging.h" |
33 | | #include "absl/base/optimization.h" |
34 | | #include "absl/container/internal/container_memory.h" |
35 | | #include "absl/container/internal/hashtable_control_bytes.h" |
36 | | #include "absl/container/internal/hashtablez_sampler.h" |
37 | | #include "absl/container/internal/raw_hash_set_resize_impl.h" |
38 | | #include "absl/functional/function_ref.h" |
39 | | #include "absl/hash/hash.h" |
40 | | |
41 | | namespace absl { |
42 | | ABSL_NAMESPACE_BEGIN |
43 | | namespace container_internal { |
44 | | |
45 | | // Represents a control byte corresponding to a full slot with arbitrary hash. |
46 | 0 | constexpr ctrl_t ZeroCtrlT() { return static_cast<ctrl_t>(0); } |
47 | | |
48 | | // A single byte for default-constructed iterators. We leave it uninitialized |
49 | | // because reading this memory is a bug. |
50 | | ABSL_DLL char kDefaultIterSlot; |
51 | | |
52 | | // We need one full byte followed by a sentinel byte for iterator::operator++. |
53 | | ABSL_CONST_INIT ABSL_DLL const ctrl_t kSooControl[2] = {ZeroCtrlT(), |
54 | | ctrl_t::kSentinel}; |
55 | | // We need one full byte followed by a sentinel byte for iterator::operator++. |
56 | | ABSL_CONST_INIT ABSL_DLL const ctrl_t kInsertIteratorControl[2] = { |
57 | | ZeroCtrlT(), ctrl_t::kSentinel}; |
58 | | |
59 | | namespace { |
60 | | |
61 | | #ifdef ABSL_SWISSTABLE_ASSERT |
62 | | #error ABSL_SWISSTABLE_ASSERT cannot be directly set |
63 | | #else |
64 | | // We use this macro for assertions that users may see when the table is in an |
65 | | // invalid state that sanitizers may help diagnose. |
66 | | #define ABSL_SWISSTABLE_ASSERT(CONDITION) \ |
67 | 57.9M | assert((CONDITION) && "Try enabling sanitizers.") |
68 | | #endif |
69 | | |
70 | | void ValidateMaxSize([[maybe_unused]] size_t size, |
71 | | [[maybe_unused]] size_t key_size, |
72 | 0 | [[maybe_unused]] size_t slot_size) { |
73 | 0 | ABSL_SWISSTABLE_ASSERT(size <= MaxValidSize(key_size, slot_size)); |
74 | 0 | } |
75 | 0 | void ValidateMaxCapacity(size_t capacity, size_t key_size, size_t slot_size) { |
76 | 0 | if (capacity <= 1) return; |
77 | 0 | ValidateMaxSize(CapacityToGrowth(PreviousCapacity(capacity)), key_size, |
78 | 0 | slot_size); |
79 | 0 | } |
80 | | |
81 | | // Returns "random" seed. |
82 | 56.6k | inline size_t RandomSeed() { |
83 | 56.6k | constexpr size_t kIncrement = 0xad53; |
84 | 56.6k | #ifdef ABSL_HAVE_THREAD_LOCAL |
85 | 56.6k | static thread_local size_t counter = 0; |
86 | 56.6k | counter += kIncrement; |
87 | 56.6k | size_t value = counter; |
88 | | #else // ABSL_HAVE_THREAD_LOCAL |
89 | | static std::atomic<size_t> counter(0); |
90 | | size_t value = counter.fetch_add(kIncrement, std::memory_order_relaxed); |
91 | | #endif // ABSL_HAVE_THREAD_LOCAL |
92 | 56.6k | return value ^ static_cast<size_t>(reinterpret_cast<uintptr_t>(&counter)); |
93 | 56.6k | } |
94 | | |
95 | 0 | bool ShouldRehashForBugDetection(size_t capacity) { |
96 | | // Note: we can't use the abseil-random library because abseil-random |
97 | | // depends on swisstable. We want to return true with probability |
98 | | // `min(1, RehashProbabilityConstant() / capacity())`. In order to do this, |
99 | | // we probe based on a random hash and see if the offset is less than |
100 | | // RehashProbabilityConstant(). |
101 | 0 | return probe(ProbeCapacity{capacity}, absl::HashOf(RandomSeed())) |
102 | 0 | .offset() < RehashProbabilityConstant(); |
103 | 0 | } |
104 | | |
105 | | // Returns the offset of the new element after resize from capacity 1 to 3. |
106 | 56.6k | size_t Resize1To3NewOffset(size_t hash) { |
107 | | // After resize from capacity 1 to 3, we always have exactly the slot with |
108 | | // index 1 occupied, so we need to insert either at index 0 or index 2. |
109 | 56.6k | static_assert(SooSlotIndex() == 1); |
110 | 56.6k | return hash & 2; |
111 | 56.6k | } |
112 | | |
113 | | // Returns the address of the ith slot in slots where each slot occupies |
114 | | // slot_size. |
115 | 11.1M | inline void* SlotAddress(void* slot_array, size_t slot, size_t slot_size) { |
116 | 11.1M | return static_cast<void*>(static_cast<char*>(slot_array) + |
117 | 11.1M | (slot * slot_size)); |
118 | 11.1M | } |
119 | | |
120 | | // Returns the address of the slot `i` iterations after `slot` assuming each |
121 | | // slot has the specified size. |
122 | 100k | inline void* NextSlot(void* slot, size_t slot_size, size_t i = 1) { |
123 | 100k | return reinterpret_cast<void*>(reinterpret_cast<uintptr_t>(slot) + |
124 | 100k | slot_size * i); |
125 | 100k | } |
126 | | |
127 | | // Returns the address of the slot just before `slot` assuming each slot has the |
128 | | // specified size. |
129 | 0 | inline void* PrevSlot(void* slot, size_t slot_size) { |
130 | 0 | return reinterpret_cast<void*>(reinterpret_cast<uintptr_t>(slot) - slot_size); |
131 | 0 | } |
132 | | |
133 | | } // namespace |
134 | | |
135 | | // Must be defined out-of-line to avoid MSVC error C2482 on some platforms, |
136 | | // which is caused by non-constexpr initialization. |
137 | 56.6k | uint8_t NextHashTableSeed() { |
138 | 56.6k | return static_cast<uint8_t>(RandomSeed()); |
139 | 56.6k | } |
140 | | |
141 | 0 | GenerationType* EmptyGeneration() { |
142 | 0 | if (SwisstableGenerationsEnabled()) { |
143 | 0 | constexpr size_t kNumEmptyGenerations = 1024; |
144 | 0 | static constexpr GenerationType kEmptyGenerations[kNumEmptyGenerations]{}; |
145 | 0 | return const_cast<GenerationType*>( |
146 | 0 | &kEmptyGenerations[RandomSeed() % kNumEmptyGenerations]); |
147 | 0 | } |
148 | 0 | return nullptr; |
149 | 0 | } |
150 | | |
151 | | bool CommonFieldsGenerationInfoEnabled:: |
152 | 0 | should_rehash_for_bug_detection_on_insert(size_t capacity) const { |
153 | 0 | if (reserved_growth_ == kReservedGrowthJustRanOut) return true; |
154 | 0 | if (reserved_growth_ > 0) return false; |
155 | 0 | return ShouldRehashForBugDetection(capacity); |
156 | 0 | } |
157 | | |
158 | | bool CommonFieldsGenerationInfoEnabled::should_rehash_for_bug_detection_on_move( |
159 | 0 | size_t capacity) const { |
160 | 0 | return ShouldRehashForBugDetection(capacity); |
161 | 0 | } |
162 | | |
163 | | namespace { |
164 | | |
165 | | // Probes an array of control bits using a probe sequence, |
166 | | // and returns the mask corresponding to the first group with a deleted or empty |
167 | | // slot. |
168 | | inline Group::NonIterableBitMaskType probe_till_first_non_full_group( |
169 | | const ctrl_t* ctrl, probe_seq<Group::kWidth>& seq, |
170 | 117k | [[maybe_unused]] size_t capacity) { |
171 | 121k | while (true) { |
172 | 121k | GroupFullEmptyOrDeleted g{ctrl + seq.offset()}; |
173 | 121k | auto mask = g.MaskEmptyOrDeleted(); |
174 | 121k | if (mask) { |
175 | 117k | return mask; |
176 | 117k | } |
177 | 3.74k | seq.next(); |
178 | 3.74k | ABSL_SWISSTABLE_ASSERT(seq.index() <= capacity && "full table!"); |
179 | 3.74k | } |
180 | 117k | } |
181 | | |
182 | | FindInfo find_first_non_full_from_h1(const ctrl_t* ctrl, size_t h1, |
183 | 226k | HashtableCapacity capacity) { |
184 | 226k | const size_t cap = capacity.capacity(); |
185 | 226k | auto seq = probe_h1(ProbeCapacity{cap}, h1); |
186 | 226k | if (IsEmptyOrDeleted(ctrl[seq.offset()])) { |
187 | 108k | return {seq.offset(), /*probe_length=*/0}; |
188 | 108k | } |
189 | 117k | auto mask = probe_till_first_non_full_group(ctrl, seq, cap); |
190 | 117k | return {seq.offset(mask.LowestBitSet()), seq.index()}; |
191 | 226k | } |
192 | | |
193 | | // Probes an array of control bits using a probe sequence derived from `hash`, |
194 | | // and returns the offset corresponding to the first deleted or empty slot. |
195 | | // |
196 | | // Behavior when the entire table is full is undefined. |
197 | | // |
198 | | // NOTE: this function must work with tables having both empty and deleted |
199 | | // slots in the same group. Such tables appear during `erase()`. |
200 | 113k | FindInfo find_first_non_full(const CommonFields& common, size_t hash) { |
201 | 113k | return find_first_non_full_from_h1(common.control(), H1(hash), |
202 | 113k | common.capacity_impl()); |
203 | 113k | } |
204 | | |
205 | | // Same as `find_first_non_full`, but returns the mask corresponding to the |
206 | | // first group with a deleted or empty slot. |
207 | | std::pair<FindInfo, Group::NonIterableBitMaskType> find_first_non_full_group( |
208 | 0 | const CommonFields& common, size_t hash) { |
209 | 0 | auto seq = probe(common, hash); |
210 | 0 | auto mask = |
211 | 0 | probe_till_first_non_full_group(common.control(), seq, common.capacity()); |
212 | 0 | return {{seq.offset(), seq.index()}, mask}; |
213 | 0 | } |
214 | | |
215 | | // Whether a table fits in half a group. A half-group table fits entirely into a |
216 | | // probing group, i.e., has a capacity < `Group::kWidth`. |
217 | | // |
218 | | // In half-group mode we are able to use the whole capacity. The extra control |
219 | | // bytes give us at least one "empty" control byte to stop the iteration. |
220 | | // This is important to make 1 a valid capacity. |
221 | | // |
222 | | // In half-group mode only the first `capacity` control bytes after the sentinel |
223 | | // are valid. The rest contain dummy ctrl_t::kEmpty values that do not |
224 | | // represent a real slot. |
225 | 0 | constexpr bool is_half_group(size_t capacity) { |
226 | 0 | return capacity < Group::kWidth - 1; |
227 | 0 | } |
228 | | |
229 | | template <class Fn> |
230 | 0 | void IterateOverFullSlotsImpl(const CommonFields& c, size_t slot_size, Fn cb) { |
231 | 0 | const size_t cap = c.capacity(); |
232 | 0 | ABSL_ASSUME(cap > kMaxSmallCapacity); |
233 | 0 | const ctrl_t* ctrl = c.control(); |
234 | 0 | void* slot = c.slot_array(cap); |
235 | 0 | if (is_half_group(cap)) { |
236 | | // Mirrored/cloned control bytes in half-group table are also located in the |
237 | | // first group (starting from position 0). We are taking group from position |
238 | | // `capacity` in order to avoid duplicates. |
239 | | |
240 | | // Half-group tables capacity fits into portable group, where |
241 | | // GroupPortableImpl::MaskFull is more efficient for the |
242 | | // capacity <= GroupPortableImpl::kWidth. |
243 | 0 | ABSL_SWISSTABLE_ASSERT(cap <= GroupPortableImpl::kWidth && |
244 | 0 | "unexpectedly large half-group capacity"); |
245 | 0 | static_assert(Group::kWidth >= GroupPortableImpl::kWidth, |
246 | 0 | "unexpected group width"); |
247 | | // Group starts from kSentinel slot, so indices in the mask will |
248 | | // be increased by 1. |
249 | 0 | const auto mask = GroupPortableImpl(ctrl + cap).MaskFull(); |
250 | 0 | --ctrl; |
251 | 0 | slot = PrevSlot(slot, slot_size); |
252 | 0 | for (uint32_t i : mask) { |
253 | 0 | cb(ctrl + i, SlotAddress(slot, i, slot_size)); |
254 | 0 | } |
255 | 0 | return; |
256 | 0 | } |
257 | 0 | size_t remaining = c.size(); |
258 | 0 | [[maybe_unused]] const size_t original_size_for_assert = remaining; |
259 | 0 | while (remaining != 0) { |
260 | 0 | for (uint32_t i : GroupFullEmptyOrDeleted(ctrl).MaskFull()) { |
261 | 0 | ABSL_SWISSTABLE_ASSERT(IsFull(ctrl[i]) && |
262 | 0 | "hash table was modified unexpectedly"); |
263 | 0 | cb(ctrl + i, SlotAddress(slot, i, slot_size)); |
264 | 0 | --remaining; |
265 | 0 | } |
266 | 0 | ctrl += Group::kWidth; |
267 | 0 | slot = NextSlot(slot, slot_size, Group::kWidth); |
268 | 0 | ABSL_SWISSTABLE_ASSERT( |
269 | 0 | (remaining == 0 || *(ctrl - 1) != ctrl_t::kSentinel) && |
270 | 0 | "hash table was modified unexpectedly"); |
271 | 0 | } |
272 | | // NOTE: erasure of the current element is allowed in callback for |
273 | | // absl::erase_if specialization. So we use `>=`. |
274 | 0 | ABSL_SWISSTABLE_ASSERT(original_size_for_assert >= c.size() && |
275 | 0 | "hash table was modified unexpectedly"); |
276 | 0 | } Unexecuted instantiation: raw_hash_set.cc:void absl::container_internal::(anonymous namespace)::IterateOverFullSlotsImpl<absl::FunctionRef<void (absl::container_internal::ctrl_t const*, void*)> >(absl::container_internal::CommonFields const&, unsigned long, absl::FunctionRef<void (absl::container_internal::ctrl_t const*, void*)>) Unexecuted instantiation: raw_hash_set.cc:void absl::container_internal::(anonymous namespace)::IterateOverFullSlotsImpl<absl::container_internal::DestroySlots(absl::container_internal::CommonFields&, unsigned long, void (*)(void*, void*))::$_0>(absl::container_internal::CommonFields const&, unsigned long, absl::container_internal::DestroySlots(absl::container_internal::CommonFields&, unsigned long, void (*)(void*, void*))::$_0) Unexecuted instantiation: raw_hash_set.cc:void absl::container_internal::(anonymous namespace)::IterateOverFullSlotsImpl<absl::container_internal::Copy(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, absl::container_internal::CommonFields const&, absl::FunctionRef<void (void*, void const*)>)::$_0>(absl::container_internal::CommonFields const&, unsigned long, absl::container_internal::Copy(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, absl::container_internal::CommonFields const&, absl::FunctionRef<void (void*, void const*)>)::$_0) |
277 | | |
278 | | } // namespace |
279 | | |
280 | 613k | void CommonFields::AssertNotDebugCapacityImpl() const { |
281 | 613k | const HashtableCapacity cap = maybe_invalid_capacity(); |
282 | 613k | if (ABSL_PREDICT_TRUE(cap.IsValid())) { |
283 | 613k | return; |
284 | 613k | } |
285 | 613k | assert(!cap.IsReentrance() && |
286 | 0 | "Reentrant container access during element construction/destruction " |
287 | 0 | "is not allowed."); |
288 | 0 | if (cap.IsDestroyed()) { |
289 | 0 | ABSL_RAW_LOG(FATAL, "Use of destroyed hash table."); |
290 | 0 | } |
291 | 0 | if (SwisstableGenerationsEnabled() && ABSL_PREDICT_FALSE(cap.IsMovedFrom())) { |
292 | 0 | if (cap.IsSelfMovedFrom()) { |
293 | | // If this log triggers, then a hash table was move-assigned to itself |
294 | | // and then used again later without being reinitialized. |
295 | 0 | ABSL_RAW_LOG(FATAL, "Use of self-move-assigned hash table."); |
296 | 0 | } |
297 | 0 | ABSL_RAW_LOG(FATAL, "Use of moved-from hash table."); |
298 | 0 | } |
299 | 0 | } |
300 | | |
301 | 0 | void ConvertDeletedToEmptyAndFullToDeleted(ctrl_t* ctrl, size_t capacity) { |
302 | 0 | ABSL_SWISSTABLE_ASSERT(ctrl[capacity] == ctrl_t::kSentinel); |
303 | 0 | ABSL_SWISSTABLE_ASSERT(IsValidCapacity(capacity)); |
304 | 0 | for (ctrl_t* pos = ctrl; pos < ctrl + capacity; pos += Group::kWidth) { |
305 | 0 | Group{pos}.ConvertSpecialToEmptyAndFullToDeleted(pos); |
306 | 0 | } |
307 | | // Copy the cloned ctrl bytes. |
308 | 0 | std::memcpy(ctrl + capacity + 1, ctrl, NumClonedBytes()); |
309 | 0 | ctrl[capacity] = ctrl_t::kSentinel; |
310 | 0 | } |
311 | | |
312 | | void IterateOverFullSlots(const CommonFields& c, size_t slot_size, |
313 | 0 | absl::FunctionRef<void(const ctrl_t*, void*)> cb) { |
314 | 0 | IterateOverFullSlotsImpl(c, slot_size, cb); |
315 | 0 | } |
316 | | |
317 | | void CommonFields::InitGrowthLeftNoDeleted(size_t growth_left, |
318 | 502k | size_t capacity) { |
319 | 502k | if (capacity <= GrowthInfoLowerBound::kMaxGrowthLeftLowerBound) { |
320 | 463k | inline_data_.set_growth_info_lower_bound( |
321 | 463k | GrowthInfoLowerBound(static_cast<uint8_t>(growth_left))); |
322 | 463k | } else { |
323 | 38.9k | size_t lower_bound = (std::min)( |
324 | 38.9k | growth_left, |
325 | 38.9k | GrowthInfoLowerBound::kMaxGrowthLeftLowerBound); |
326 | 38.9k | inline_data_.set_growth_info_lower_bound( |
327 | 38.9k | GrowthInfoLowerBound(static_cast<uint8_t>(lower_bound))); |
328 | 38.9k | SetGrowthInfoOverflow(growth_left - lower_bound); |
329 | 38.9k | } |
330 | 502k | } |
331 | | |
332 | 177k | size_t CommonFields::GetGrowthLeftTotalSlow(size_t capacity) const { |
333 | 177k | size_t result = inline_data_.growth_info_lower_bound().GetGrowthLeft(); |
334 | 177k | if (capacity > GrowthInfoLowerBound::kMaxGrowthLeftLowerBound) { |
335 | 17.2k | result += GetOverflowGrowthLeft(); |
336 | 17.2k | } |
337 | 177k | return result; |
338 | 177k | } |
339 | | |
340 | 0 | void CommonFields::OverwriteFullAsEmpty() { |
341 | 0 | if (inline_data_.growth_info_lower_bound().GetGrowthLeft() < |
342 | 0 | GrowthInfoLowerBound::kMaxGrowthLeftLowerBound) { |
343 | 0 | inline_data_.overwrite_full_as_empty_in_lower_bound(); |
344 | 0 | } else { |
345 | 0 | SetGrowthInfoOverflow(GetOverflowGrowthLeft() + 1); |
346 | 0 | } |
347 | 0 | } |
348 | | |
349 | | GrowthInfoLowerBound CommonFields::RebalanceGrowthLeftLowerBound( |
350 | 183k | size_t capacity) { |
351 | 183k | auto growth_left_lower_bound = GetGrowthInfoLowerBound(); |
352 | 183k | if (capacity <= GrowthInfoLowerBound::kMaxGrowthLeftLowerBound || |
353 | | // For tables with deleted slots, we often call rebalance even if |
354 | | // we have growth left in the lower bound. |
355 | 159k | growth_left_lower_bound.HasDeletedAndGrowthLeft()) { |
356 | 159k | return growth_left_lower_bound; |
357 | 159k | } else { |
358 | 23.2k | return RebalanceGrowthLeftLowerBoundLargeCapacity(); |
359 | 23.2k | } |
360 | 183k | } |
361 | | |
362 | | ABSL_ATTRIBUTE_NOINLINE GrowthInfoLowerBound |
363 | 23.2k | CommonFields::RebalanceGrowthLeftLowerBoundLargeCapacity() { |
364 | 23.2k | size_t overflow_growth_left = GetOverflowGrowthLeft(); |
365 | 23.2k | size_t lower_bound_growth_left = GetGrowthLeftLowerBound(); |
366 | 23.2k | size_t overflow_to_lower_bound_size = |
367 | 23.2k | (std::min)(overflow_growth_left, |
368 | 23.2k | GrowthInfoLowerBound::kMaxGrowthLeftLowerBound - |
369 | 23.2k | lower_bound_growth_left); |
370 | 23.2k | SetGrowthInfoOverflow(overflow_growth_left - overflow_to_lower_bound_size); |
371 | 23.2k | inline_data_.increment_growth_info_lower_bound(overflow_to_lower_bound_size); |
372 | 23.2k | auto result = GetGrowthInfoLowerBound(); |
373 | 23.2k | ABSL_SWISSTABLE_ASSERT(result.HasNoDeleted() == |
374 | 23.2k | GetGrowthInfoLowerBound().HasNoDeleted()); |
375 | 23.2k | ABSL_SWISSTABLE_ASSERT( |
376 | 23.2k | (result.GetGrowthLeft() > 0 || |
377 | 23.2k | (GetGrowthLeftLowerBound() + GetOverflowGrowthLeft() == 0)) && |
378 | 23.2k | "rebalance may return 0 only if we have absolutely no growth left"); |
379 | 23.2k | return result; |
380 | 23.2k | } |
381 | | |
382 | 0 | HashtablezInfoHandle CommonFields::infoz_ptr() const { |
383 | | // growth_info is stored before control bytes. |
384 | 0 | ABSL_SWISSTABLE_ASSERT(has_infoz()); |
385 | 0 | HashtablezInfoHandle res; |
386 | 0 | void* src = reinterpret_cast<char*>(control()) - |
387 | 0 | MetadataBeforeControlSize(/*has_infoz=*/true, capacity()); |
388 | 0 | std::memcpy(&res, src, sizeof(HashtablezInfoHandle)); |
389 | 0 | return res; |
390 | 0 | } |
391 | | |
392 | 0 | void CommonFields::set_infoz(HashtablezInfoHandle infoz) { |
393 | 0 | ABSL_SWISSTABLE_ASSERT(has_infoz()); |
394 | 0 | void* dst = reinterpret_cast<char*>(control()) - |
395 | 0 | MetadataBeforeControlSize(/*has_infoz=*/true, capacity()); |
396 | 0 | std::memcpy(dst, &infoz, sizeof(HashtablezInfoHandle)); |
397 | 0 | } |
398 | | |
399 | | namespace { |
400 | | void DeallocBackingArrayImpl(void* alloc, size_t capacity, ctrl_t* ctrl, |
401 | | size_t slot_size, size_t slot_align, |
402 | | bool has_infoz, size_t blocked_element_count, |
403 | 267k | DeallocBackingArrayFn dealloc) { |
404 | 267k | RawHashSetLayout layout(capacity, slot_size, slot_align, has_infoz, |
405 | 267k | blocked_element_count); |
406 | 267k | void* backing_array = ctrl - layout.control_offset(); |
407 | | // Unpoison before returning the memory to the allocator. |
408 | 267k | SanitizerUnpoisonMemoryRegion(backing_array, layout.alloc_size()); |
409 | 267k | dealloc(alloc, backing_array, layout.alloc_size()); |
410 | 267k | } |
411 | | |
412 | | void DeallocBackingArrayImpl(CommonFields& c, |
413 | | const PolicyFunctions& __restrict policy, |
414 | 17.3k | void* alloc) { |
415 | 17.3k | DeallocBackingArrayImpl(alloc, c.capacity(), c.control(), policy.slot_size, |
416 | 17.3k | policy.slot_align, c.has_infoz(), |
417 | 17.3k | c.blocked_element_count(), policy.dealloc); |
418 | 17.3k | } |
419 | | |
420 | | } // namespace |
421 | | |
422 | | void UnregisterAndDeallocBackingArray(CommonFields& c, |
423 | | const DtorPolicy& __restrict policy, |
424 | | DeallocBackingArrayFn dealloc, |
425 | 41.8k | void* alloc) { |
426 | 41.8k | size_t cap = c.capacity(); // capacity is already in register, so storing it |
427 | | // in a local variable before Unregister(). |
428 | 41.8k | c.infoz().Unregister(); |
429 | 41.8k | DeallocBackingArrayImpl(alloc, cap, c.control(), policy.slot_size, |
430 | 41.8k | policy.slot_align, c.has_infoz(), |
431 | 41.8k | c.blocked_element_count(), dealloc); |
432 | 41.8k | } |
433 | | |
434 | | namespace { |
435 | | |
436 | | void ResetGrowthLeft(size_t capacity, size_t occupied_elements, |
437 | 445k | CommonFields& common) { |
438 | 445k | common.InitGrowthLeftNoDeleted(CapacityToGrowth(capacity) - occupied_elements, |
439 | 445k | capacity); |
440 | 445k | } |
441 | | |
442 | | // Finds guaranteed to exists empty slot from the given position. |
443 | | // NOTE: this function is almost never triggered inside of the |
444 | | // DropDeletesWithoutResize, so we keep it simple. |
445 | | // The table is rather sparse, so empty slot will be found very quickly. |
446 | 0 | size_t FindEmptySlot(size_t start, size_t end, const ctrl_t* ctrl) { |
447 | 0 | for (size_t i = start; i < end; ++i) { |
448 | 0 | if (IsEmpty(ctrl[i])) { |
449 | 0 | return i; |
450 | 0 | } |
451 | 0 | } |
452 | 0 | ABSL_UNREACHABLE(); |
453 | 0 | } |
454 | | |
455 | | // Finds guaranteed to exist full slot starting from the given position. |
456 | | // NOTE: this function is only triggered for rehash(0), when we need to |
457 | | // go back to SOO state, so we keep it simple. |
458 | 0 | size_t FindFirstFullSlot(size_t start, size_t end, const ctrl_t* ctrl) { |
459 | 0 | for (size_t i = start; i < end; ++i) { |
460 | 0 | if (IsFull(ctrl[i])) { |
461 | 0 | return i; |
462 | 0 | } |
463 | 0 | } |
464 | 0 | ABSL_UNREACHABLE(); |
465 | 0 | } |
466 | | |
467 | 10.9M | void PrepareInsertCommon(CommonFields& common) { |
468 | 10.9M | common.increment_size(); |
469 | 10.9M | common.maybe_increment_generation_on_insert(); |
470 | 10.9M | } |
471 | | |
472 | | // Sets sanitizer poisoning for slot corresponding to control byte being set. |
473 | | inline void DoSanitizeOnSetCtrl(const CommonFields& c, size_t i, ctrl_t h, |
474 | 10.9M | size_t slot_size) { |
475 | 10.9M | const size_t cap = c.capacity(); |
476 | 10.9M | ABSL_ASSUME(cap > kMaxSmallCapacity); |
477 | 10.9M | ABSL_SWISSTABLE_ASSERT(i < cap); |
478 | 10.9M | auto* slot_i = static_cast<const char*>(c.slot_array(cap)) + i * slot_size; |
479 | 10.9M | if (IsFull(h)) { |
480 | 10.9M | SanitizerUnpoisonMemoryRegion(slot_i, slot_size); |
481 | 10.9M | } else { |
482 | 0 | SanitizerPoisonMemoryRegion(slot_i, slot_size); |
483 | 0 | } |
484 | 10.9M | } |
485 | | |
486 | | // Sets `ctrl[i]` to `h`. |
487 | | // |
488 | | // Unlike setting it directly, this function will perform bounds checks and |
489 | | // mirror the value to the cloned tail if necessary. |
490 | 10.6M | inline void SetCtrlNoSanitizeImpl(const CommonFields& c, size_t i, ctrl_t h) { |
491 | 10.6M | ABSL_SWISSTABLE_ASSERT(i < c.capacity()); |
492 | 10.6M | ctrl_t* ctrl = c.control(); |
493 | 10.6M | const size_t cap = c.capacity(); |
494 | 10.6M | ctrl[i] = h; |
495 | 10.6M | ctrl[((i - NumClonedBytes()) & cap) + (NumClonedBytes() & cap)] = h; |
496 | 10.6M | } |
497 | | |
498 | | inline void SetCtrl(const CommonFields& c, size_t i, ctrl_t h, |
499 | 10.6M | size_t slot_size) { |
500 | 10.6M | ABSL_SWISSTABLE_ASSERT(!c.is_small()); |
501 | 10.6M | DoSanitizeOnSetCtrl(c, i, h, slot_size); |
502 | 10.6M | SetCtrlNoSanitizeImpl(c, i, h); |
503 | 10.6M | } |
504 | | // Overload for setting to an occupied `h2_t` rather than a special `ctrl_t`. |
505 | 10.6M | inline void SetCtrl(const CommonFields& c, size_t i, h2_t h, size_t slot_size) { |
506 | 10.6M | SetCtrl(c, i, static_cast<ctrl_t>(h), slot_size); |
507 | 10.6M | } |
508 | | |
509 | | // Sets `ctrl[i]` to `ctrl_t::kSentinel`. |
510 | | // |
511 | | // Unlike setting it directly, this function will perform bounds checks and |
512 | | // mirror the value to the cloned tail if necessary. |
513 | 0 | inline void BlockCtrl(const CommonFields& c, size_t i) { |
514 | 0 | ABSL_SWISSTABLE_ASSERT(!c.is_small()); |
515 | 0 | SetCtrlNoSanitizeImpl(c, i, ctrl_t::kSentinel); |
516 | 0 | } |
517 | | |
518 | | // Like SetCtrl, but in a single group table, we can save some operations when |
519 | | // setting the cloned control byte. |
520 | | inline void SetCtrlInSingleGroupTable(const CommonFields& c, size_t i, ctrl_t h, |
521 | 69.2k | size_t slot_size) { |
522 | 69.2k | const size_t cap = c.capacity(); |
523 | 69.2k | ABSL_SWISSTABLE_ASSERT(!c.is_small()); |
524 | 69.2k | ABSL_SWISSTABLE_ASSERT(is_single_group(cap)); |
525 | 69.2k | DoSanitizeOnSetCtrl(c, i, h, slot_size); |
526 | 69.2k | ctrl_t* ctrl = c.control(); |
527 | 69.2k | ctrl[i] = h; |
528 | 69.2k | ctrl[i + cap + 1] = h; |
529 | 69.2k | } |
530 | | // Overload for setting to an occupied `h2_t` rather than a special `ctrl_t`. |
531 | | inline void SetCtrlInSingleGroupTable(const CommonFields& c, size_t i, h2_t h, |
532 | 69.2k | size_t slot_size) { |
533 | 69.2k | SetCtrlInSingleGroupTable(c, i, static_cast<ctrl_t>(h), slot_size); |
534 | 69.2k | } |
535 | | |
536 | | // Like SetCtrl, but in a table with capacity >= Group::kWidth - 1, |
537 | | // we can save some operations when setting the cloned control byte. |
538 | | inline void SetCtrlInLargeTable(const CommonFields& c, size_t i, ctrl_t h, |
539 | 226k | size_t slot_size) { |
540 | 226k | ABSL_SWISSTABLE_ASSERT(c.capacity() >= Group::kWidth - 1); |
541 | 226k | DoSanitizeOnSetCtrl(c, i, h, slot_size); |
542 | 226k | ctrl_t* ctrl = c.control(); |
543 | 226k | ctrl[i] = h; |
544 | 226k | ctrl[((i - NumClonedBytes()) & c.capacity()) + NumClonedBytes()] = h; |
545 | 226k | } |
546 | | // Overload for setting to an occupied `h2_t` rather than a special `ctrl_t`. |
547 | | inline void SetCtrlInLargeTable(const CommonFields& c, size_t i, h2_t h, |
548 | 226k | size_t slot_size) { |
549 | 226k | SetCtrlInLargeTable(c, i, static_cast<ctrl_t>(h), slot_size); |
550 | 226k | } |
551 | | |
552 | 268k | void BlockControlBytes(CommonFields& common, size_t blocked_element_count) { |
553 | 268k | const size_t capacity = common.capacity(); |
554 | 268k | while (blocked_element_count > 0) { |
555 | 0 | BlockCtrl(common, capacity - blocked_element_count); |
556 | 0 | --blocked_element_count; |
557 | 0 | } |
558 | 268k | } |
559 | | |
560 | | void* DropDeletesWithoutResizeAndPrepareInsert( |
561 | | CommonFields& common, const PolicyFunctions& __restrict policy, |
562 | 0 | size_t new_hash) { |
563 | 0 | void* set = &common; |
564 | 0 | const size_t capacity = common.capacity(); |
565 | 0 | ABSL_SWISSTABLE_ASSERT(IsValidCapacity(capacity)); |
566 | 0 | ABSL_SWISSTABLE_ASSERT(!is_single_group(capacity)); |
567 | 0 | ABSL_ASSUME(capacity > kMaxSmallCapacity); |
568 | | |
569 | 0 | ctrl_t* ctrl = common.control(); |
570 | 0 | void* slot_array = common.slot_array(capacity); |
571 | | // Algorithm: |
572 | | // - mark all DELETED slots as EMPTY |
573 | | // - mark all FULL slots as DELETED |
574 | | // - for each slot marked as DELETED |
575 | | // hash = Hash(element) |
576 | | // target = find_first_non_full(hash) |
577 | | // if target is in the same group |
578 | | // mark slot as FULL |
579 | | // else if target is EMPTY |
580 | | // transfer element to target |
581 | | // mark slot as EMPTY |
582 | | // mark target as FULL |
583 | | // else if target is DELETED |
584 | | // swap current element with target element |
585 | | // mark target as FULL |
586 | | // repeat procedure for current slot with moved from element (target) |
587 | 0 | const size_t blocked_element_count = common.blocked_element_count(); |
588 | 0 | ConvertDeletedToEmptyAndFullToDeleted(ctrl, capacity); |
589 | 0 | BlockControlBytes(common, blocked_element_count); |
590 | 0 | const void* hash_fn = policy.hash_fn(common); |
591 | 0 | auto hasher = policy.hash_slot; |
592 | 0 | auto transfer_n = policy.transfer_n; |
593 | 0 | const size_t slot_size = policy.slot_size; |
594 | |
|
595 | 0 | size_t total_probe_length = 0; |
596 | 0 | void* slot_ptr = SlotAddress(slot_array, 0, slot_size); |
597 | | |
598 | | // The index of an empty slot that can be used as temporary memory for |
599 | | // the swap operation. |
600 | 0 | constexpr size_t kUnknownId = ~size_t{}; |
601 | 0 | size_t tmp_space_id = kUnknownId; |
602 | |
|
603 | 0 | for (size_t i = 0; i != capacity; |
604 | 0 | ++i, slot_ptr = NextSlot(slot_ptr, slot_size)) { |
605 | 0 | ABSL_SWISSTABLE_ASSERT(slot_ptr == SlotAddress(slot_array, i, slot_size)); |
606 | 0 | if (IsEmpty(ctrl[i])) { |
607 | 0 | tmp_space_id = i; |
608 | 0 | continue; |
609 | 0 | } |
610 | 0 | if (!IsDeleted(ctrl[i])) continue; |
611 | 0 | const size_t hash = (*hasher)(hash_fn, slot_ptr, common.seed().seed()); |
612 | 0 | const FindInfo target = find_first_non_full(common, hash); |
613 | 0 | const size_t new_i = target.offset; |
614 | 0 | total_probe_length += target.probe_length; |
615 | | |
616 | | // Verify if the old and new i fall within the same group wrt the hash. |
617 | | // If they do, we don't need to move the object as it falls already in the |
618 | | // best probe we can. |
619 | 0 | const size_t probe_offset = probe(common, hash).offset(); |
620 | 0 | const h2_t h2 = H2(hash); |
621 | 0 | const auto probe_index = [probe_offset, capacity](size_t pos) { |
622 | 0 | return ((pos - probe_offset) & capacity) / Group::kWidth; |
623 | 0 | }; |
624 | | |
625 | | // Element doesn't move. |
626 | 0 | if (ABSL_PREDICT_TRUE(probe_index(new_i) == probe_index(i))) { |
627 | 0 | SetCtrlInLargeTable(common, i, h2, slot_size); |
628 | 0 | continue; |
629 | 0 | } |
630 | | |
631 | 0 | void* new_slot_ptr = SlotAddress(slot_array, new_i, slot_size); |
632 | 0 | if (IsEmpty(ctrl[new_i])) { |
633 | | // Transfer element to the empty spot. |
634 | | // SetCtrl poisons/unpoisons the slots so we have to call it at the |
635 | | // right time. |
636 | 0 | SetCtrlInLargeTable(common, new_i, h2, slot_size); |
637 | 0 | (*transfer_n)(set, new_slot_ptr, slot_ptr, 1); |
638 | 0 | SetCtrlInLargeTable(common, i, ctrl_t::kEmpty, slot_size); |
639 | | // Initialize or change empty space id. |
640 | 0 | tmp_space_id = i; |
641 | 0 | } else { |
642 | 0 | ABSL_SWISSTABLE_ASSERT(IsDeleted(ctrl[new_i])); |
643 | 0 | SetCtrlInLargeTable(common, new_i, h2, slot_size); |
644 | | // Until we are done rehashing, DELETED marks previously FULL slots. |
645 | |
|
646 | 0 | if (tmp_space_id == kUnknownId) { |
647 | 0 | tmp_space_id = FindEmptySlot(i + 1, capacity, ctrl); |
648 | 0 | } |
649 | 0 | void* tmp_space = SlotAddress(slot_array, tmp_space_id, slot_size); |
650 | 0 | SanitizerUnpoisonMemoryRegion(tmp_space, slot_size); |
651 | | |
652 | | // Swap i and new_i elements. |
653 | 0 | (*transfer_n)(set, tmp_space, new_slot_ptr, 1); |
654 | 0 | (*transfer_n)(set, new_slot_ptr, slot_ptr, 1); |
655 | 0 | (*transfer_n)(set, slot_ptr, tmp_space, 1); |
656 | |
|
657 | 0 | SanitizerPoisonMemoryRegion(tmp_space, slot_size); |
658 | | |
659 | | // repeat the processing of the ith slot |
660 | 0 | --i; |
661 | 0 | slot_ptr = PrevSlot(slot_ptr, slot_size); |
662 | 0 | } |
663 | 0 | } |
664 | | // Prepare insert for the new element. |
665 | 0 | PrepareInsertCommon(common); |
666 | 0 | ResetGrowthLeft(capacity, common.size() + blocked_element_count, common); |
667 | 0 | FindInfo find_info = find_first_non_full(common, new_hash); |
668 | 0 | SetCtrlInLargeTable(common, find_info.offset, H2(new_hash), slot_size); |
669 | 0 | common.infoz().RecordInsertMiss(new_hash, find_info.probe_length); |
670 | 0 | common.infoz().RecordRehash(total_probe_length); |
671 | 0 | return SlotAddress(slot_array, find_info.offset, slot_size); |
672 | 0 | } |
673 | | |
674 | 0 | bool WasNeverFull(CommonFields& c, size_t index) { |
675 | 0 | if (is_single_group(c.capacity())) { |
676 | 0 | return true; |
677 | 0 | } |
678 | 0 | const size_t index_before = (index - Group::kWidth) & c.capacity(); |
679 | 0 | const auto empty_after = Group(c.control() + index).MaskEmpty(); |
680 | 0 | const auto empty_before = Group(c.control() + index_before).MaskEmpty(); |
681 | | |
682 | | // We count how many consecutive non empties we have to the right and to the |
683 | | // left of `it`. If the sum is >= kWidth then there is at least one probe |
684 | | // window that might have seen a full group. |
685 | 0 | return empty_before && empty_after && |
686 | 0 | static_cast<size_t>(empty_after.TrailingZeros()) + |
687 | 0 | empty_before.LeadingZeros() < |
688 | 0 | Group::kWidth; |
689 | 0 | } |
690 | | |
691 | | // Updates the control bytes to indicate a completely empty table such that all |
692 | | // control bytes are kEmpty except for the kSentinel bytes. |
693 | | // If the table has blocked elements, last `blocked_element_count` are set to |
694 | | // kSentinel. |
695 | | void ResetCtrl(CommonFields& common, size_t slot_size, |
696 | 268k | size_t blocked_element_count) { |
697 | 268k | ABSL_SWISSTABLE_ASSERT(IsCapacityValidForBlockedElements(common.capacity()) || |
698 | 268k | blocked_element_count == 0); |
699 | 268k | const size_t capacity = common.capacity(); |
700 | 268k | ctrl_t* ctrl = common.control(); |
701 | 268k | static constexpr size_t kTwoGroupCapacity = 2 * Group::kWidth - 1; |
702 | 268k | if (ABSL_PREDICT_TRUE(capacity <= kTwoGroupCapacity)) { |
703 | 162k | if (IsSmallCapacity(capacity)) return; |
704 | 162k | std::memset(ctrl, static_cast<int8_t>(ctrl_t::kEmpty), Group::kWidth); |
705 | 162k | std::memset(ctrl + capacity, static_cast<int8_t>(ctrl_t::kEmpty), |
706 | 162k | Group::kWidth); |
707 | 162k | if (capacity == kTwoGroupCapacity) { |
708 | 37.7k | std::memset(ctrl + Group::kWidth, static_cast<int8_t>(ctrl_t::kEmpty), |
709 | 37.7k | Group::kWidth); |
710 | 37.7k | } |
711 | 162k | } else { |
712 | 106k | std::memset(ctrl, static_cast<int8_t>(ctrl_t::kEmpty), |
713 | 106k | capacity + 1 + NumClonedBytes()); |
714 | 106k | } |
715 | 268k | ctrl[capacity] = ctrl_t::kSentinel; |
716 | 268k | SanitizerPoisonMemoryRegion(common.slot_array(capacity), |
717 | 268k | slot_size * (capacity - blocked_element_count)); |
718 | 268k | BlockControlBytes(common, blocked_element_count); |
719 | 268k | } |
720 | | |
721 | | // Initializes control bytes for growing from capacity 1 to 3. |
722 | | // `orig_h2` is placed in the position `SooSlotIndex()`. |
723 | | // `new_h2` is placed in the position `new_offset`. |
724 | | ABSL_ATTRIBUTE_ALWAYS_INLINE inline void InitializeThreeElementsControlBytes( |
725 | 56.6k | h2_t orig_h2, h2_t new_h2, size_t new_offset, ctrl_t* new_ctrl) { |
726 | 56.6k | static constexpr size_t kNewCapacity = NextCapacity(SooCapacity()); |
727 | 56.6k | static_assert(kNewCapacity == 3); |
728 | 56.6k | static_assert(is_single_group(kNewCapacity)); |
729 | 56.6k | static_assert(SooSlotIndex() == 1); |
730 | 56.6k | ABSL_SWISSTABLE_ASSERT(new_offset == 0 || new_offset == 2); |
731 | | |
732 | 56.6k | static constexpr uint64_t kEmptyXorSentinel = |
733 | 56.6k | static_cast<uint8_t>(ctrl_t::kEmpty) ^ |
734 | 56.6k | static_cast<uint8_t>(ctrl_t::kSentinel); |
735 | 56.6k | static constexpr uint64_t kEmpty64 = static_cast<uint8_t>(ctrl_t::kEmpty); |
736 | 56.6k | static constexpr size_t kMirroredSooSlotIndex = |
737 | 56.6k | SooSlotIndex() + kNewCapacity + 1; |
738 | | // The first 8 bytes, where SOO slot original and mirrored positions are |
739 | | // replaced with 0. |
740 | | // Result will look like: E0ESE0EE |
741 | 56.6k | static constexpr uint64_t kFirstCtrlBytesWithZeroes = |
742 | 56.6k | k8EmptyBytes ^ (kEmpty64 << (8 * SooSlotIndex())) ^ |
743 | 56.6k | (kEmptyXorSentinel << (8 * kNewCapacity)) ^ |
744 | 56.6k | (kEmpty64 << (8 * kMirroredSooSlotIndex)); |
745 | | |
746 | 56.6k | const uint64_t soo_h2 = static_cast<uint64_t>(orig_h2); |
747 | 56.6k | const uint64_t new_h2_xor_empty = |
748 | 56.6k | static_cast<uint64_t>(new_h2 ^ static_cast<uint8_t>(ctrl_t::kEmpty)); |
749 | | // Fill the original and mirrored bytes for SOO slot. |
750 | | // Result will look like: |
751 | | // EHESEHEE |
752 | | // Where H = soo_h2, E = kEmpty, S = kSentinel. |
753 | 56.6k | uint64_t first_ctrl_bytes = |
754 | 56.6k | ((soo_h2 << (8 * SooSlotIndex())) | kFirstCtrlBytesWithZeroes) | |
755 | 56.6k | (soo_h2 << (8 * kMirroredSooSlotIndex)); |
756 | | // Replace original and mirrored empty bytes for the new position. |
757 | | // Result for new_offset 0 will look like: |
758 | | // NHESNHEE |
759 | | // Where H = soo_h2, N = H2(new_hash), E = kEmpty, S = kSentinel. |
760 | | // Result for new_offset 2 will look like: |
761 | | // EHNSEHNE |
762 | 56.6k | first_ctrl_bytes ^= (new_h2_xor_empty << (8 * new_offset)); |
763 | 56.6k | size_t new_mirrored_offset = new_offset + kNewCapacity + 1; |
764 | 56.6k | first_ctrl_bytes ^= (new_h2_xor_empty << (8 * new_mirrored_offset)); |
765 | | |
766 | | // Fill last bytes with kEmpty. |
767 | 56.6k | std::memset(new_ctrl + kNewCapacity, static_cast<int8_t>(ctrl_t::kEmpty), |
768 | 56.6k | Group::kWidth); |
769 | | // Overwrite the first 8 bytes with first_ctrl_bytes. |
770 | 56.6k | absl::little_endian::Store64(new_ctrl, first_ctrl_bytes); |
771 | | |
772 | | // Example for group size 16: |
773 | | // new_ctrl after 1st memset = ???EEEEEEEEEEEEEEEE |
774 | | // new_offset 0: |
775 | | // new_ctrl after 2nd store = NHESNHEEEEEEEEEEEEE |
776 | | // new_offset 2: |
777 | | // new_ctrl after 2nd store = EHNSEHNEEEEEEEEEEEE |
778 | | |
779 | | // Example for group size 8: |
780 | | // new_ctrl after 1st memset = ???EEEEEEEE |
781 | | // new_offset 0: |
782 | | // new_ctrl after 2nd store = NHESNHEEEEE |
783 | | // new_offset 2: |
784 | | // new_ctrl after 2nd store = EHNSEHNEEEE |
785 | 56.6k | } |
786 | | |
787 | | // ClearBackingArrayNoReuse clears the backing array and sets the common |
788 | | // fields to the default values for empty non-allocated tables. |
789 | | // REQUIRES: c.capacity > policy.soo_capacity. |
790 | | void ClearBackingArrayNoReuse(CommonFields& c, |
791 | | const PolicyFunctions& __restrict policy, |
792 | 17.3k | void* alloc) { |
793 | 17.3k | ABSL_SWISSTABLE_ASSERT(c.capacity() > policy.soo_capacity()); |
794 | | // We need to record infoz before calling dealloc, which will unregister |
795 | | // infoz. |
796 | 17.3k | c.infoz().RecordClearedReservation(); |
797 | 17.3k | c.infoz().RecordStorageChanged(0, policy.soo_capacity()); |
798 | 17.3k | c.infoz().Unregister(); |
799 | 17.3k | DeallocBackingArrayImpl(c, policy, alloc); |
800 | 17.3k | c = policy.soo_enabled ? CommonFields{soo_tag_t{}} |
801 | 17.3k | : CommonFields{non_soo_tag_t{}}; |
802 | 17.3k | } |
803 | | |
804 | | template <bool kSooEnabled> |
805 | 135k | void* SingleSlotAddress(CommonFields& c) { |
806 | 135k | return kSooEnabled ? c.soo_data() : c.slot_array(/*capacity=*/1); |
807 | 135k | } raw_hash_set.cc:void* absl::container_internal::(anonymous namespace)::SingleSlotAddress<false>(absl::container_internal::CommonFields&) Line | Count | Source | 805 | 135k | void* SingleSlotAddress(CommonFields& c) { | 806 | 135k | return kSooEnabled ? c.soo_data() : c.slot_array(/*capacity=*/1); | 807 | 135k | } |
Unexecuted instantiation: raw_hash_set.cc:void* absl::container_internal::(anonymous namespace)::SingleSlotAddress<true>(absl::container_internal::CommonFields&) |
808 | | |
809 | | template <bool kSooEnabled> |
810 | 148k | void DecrementSmallSize(CommonFields& c) { |
811 | 148k | if constexpr (kSooEnabled) { |
812 | 13.1k | c.set_empty_soo(); |
813 | 135k | } else { |
814 | 135k | c.decrement_size(); |
815 | 135k | } |
816 | 148k | } raw_hash_set.cc:void absl::container_internal::(anonymous namespace)::DecrementSmallSize<true>(absl::container_internal::CommonFields&) Line | Count | Source | 810 | 13.1k | void DecrementSmallSize(CommonFields& c) { | 811 | 13.1k | if constexpr (kSooEnabled) { | 812 | 13.1k | c.set_empty_soo(); | 813 | | } else { | 814 | | c.decrement_size(); | 815 | | } | 816 | 13.1k | } |
raw_hash_set.cc:void absl::container_internal::(anonymous namespace)::DecrementSmallSize<false>(absl::container_internal::CommonFields&) Line | Count | Source | 810 | 135k | void DecrementSmallSize(CommonFields& c) { | 811 | | if constexpr (kSooEnabled) { | 812 | | c.set_empty_soo(); | 813 | 135k | } else { | 814 | 135k | c.decrement_size(); | 815 | 135k | } | 816 | 135k | } |
|
817 | | |
818 | | void DestructSoo(CommonFields& c, const DtorPolicy& __restrict policy, |
819 | 24.9k | DeallocBackingArrayFn dealloc, void* alloc) { |
820 | 24.9k | ABSL_SWISSTABLE_ASSERT(!c.is_small() || !c.empty()); |
821 | 24.9k | if (c.is_small()) { |
822 | 0 | ABSL_SWISSTABLE_ASSERT(policy.destroy_slot != nullptr); |
823 | 0 | policy.destroy_slot(&c, c.soo_data()); |
824 | 0 | return; |
825 | 0 | } |
826 | 24.9k | if (policy.destroy_slot != nullptr) { |
827 | 0 | DestroySlots(c, policy.slot_size, policy.destroy_slot); |
828 | 0 | } |
829 | 24.9k | UnregisterAndDeallocBackingArray(c, policy, dealloc, alloc); |
830 | 24.9k | } |
831 | | |
832 | | void DestructNonSoo(CommonFields& c, const DtorPolicy& __restrict policy, |
833 | 16.9k | DeallocBackingArrayFn dealloc, void* alloc) { |
834 | 16.9k | ABSL_SWISSTABLE_ASSERT(c.capacity() > 0); |
835 | 16.9k | if (policy.destroy_slot != nullptr) { |
836 | 0 | if (c.is_small()) { |
837 | 0 | if (!c.empty()) { |
838 | 0 | static_assert(kMaxSmallCapacity == 1); |
839 | 0 | policy.destroy_slot(&c, c.slot_array(/*capacity=*/1)); |
840 | 0 | } |
841 | 0 | } else { |
842 | 0 | DestroySlots(c, policy.slot_size, policy.destroy_slot); |
843 | 0 | } |
844 | 0 | } |
845 | 16.9k | UnregisterAndDeallocBackingArray(c, policy, dealloc, alloc); |
846 | 16.9k | } |
847 | | |
848 | | } // namespace |
849 | | |
850 | 0 | void EraseMetaOnlySmall(CommonFields& c, bool soo_enabled, size_t slot_size) { |
851 | 0 | ABSL_SWISSTABLE_ASSERT(c.is_small()); |
852 | 0 | if (soo_enabled) { |
853 | 0 | c.set_empty_soo(); |
854 | 0 | return; |
855 | 0 | } |
856 | 0 | c.decrement_size(); |
857 | 0 | c.infoz().RecordErase(); |
858 | 0 | SanitizerPoisonMemoryRegion(SingleSlotAddress</*kSooEnabled=*/false>(c), |
859 | 0 | slot_size); |
860 | 0 | } |
861 | | |
862 | 0 | void EraseMetaOnlyLarge(CommonFields& c, size_t index, size_t slot_size) { |
863 | 0 | ABSL_SWISSTABLE_ASSERT(!c.is_small()); |
864 | 0 | ABSL_SWISSTABLE_ASSERT(IsFull(c.control()[index]) && |
865 | 0 | "erasing a dangling iterator"); |
866 | 0 | c.decrement_size(); |
867 | 0 | c.infoz().RecordErase(); |
868 | |
|
869 | 0 | if (WasNeverFull(c, index)) { |
870 | 0 | SetCtrl(c, index, ctrl_t::kEmpty, slot_size); |
871 | 0 | c.OverwriteFullAsEmpty(); |
872 | 0 | return; |
873 | 0 | } |
874 | | |
875 | 0 | c.OverwriteFullAsDeleted(); |
876 | 0 | SetCtrlInLargeTable(c, index, ctrl_t::kDeleted, slot_size); |
877 | 0 | } |
878 | | |
879 | | void ClearBackingArray(CommonFields& c, |
880 | | const PolicyFunctions& __restrict policy, void* alloc, |
881 | 285k | bool reuse) { |
882 | 285k | ABSL_SWISSTABLE_ASSERT(c.capacity() > kMaxSmallCapacity); |
883 | 285k | if (reuse) { |
884 | 268k | const size_t blocked_element_count = c.blocked_element_count(); |
885 | 268k | c.set_size_to_zero(); |
886 | 268k | ABSL_SWISSTABLE_ASSERT(c.capacity() > policy.soo_capacity()); |
887 | 268k | ResetCtrl(c, policy.slot_size, blocked_element_count); |
888 | 268k | ResetGrowthLeft(c.capacity(), blocked_element_count, c); |
889 | 268k | ABSL_SWISSTABLE_ASSERT(c.blocked_element_count() == blocked_element_count); |
890 | 268k | c.infoz().RecordStorageChanged(0, c.capacity()); |
891 | 268k | } else { |
892 | 17.3k | ClearBackingArrayNoReuse(c, policy, alloc); |
893 | 17.3k | } |
894 | 285k | } |
895 | | |
896 | | void DestroySlots(CommonFields& c, size_t slot_size, |
897 | 0 | DestroySlotFn destroy_slot) { |
898 | 0 | ABSL_SWISSTABLE_ASSERT(!c.is_small()); |
899 | 0 | ABSL_SWISSTABLE_ASSERT(destroy_slot != nullptr); |
900 | 0 | auto destroy_slot_wrapper = [&](const ctrl_t*, void* slot) { |
901 | 0 | destroy_slot(&c, slot); |
902 | 0 | }; |
903 | 0 | if constexpr (SwisstableGenerationsOrDebugEnabled()) { |
904 | 0 | CommonFields common_copy(non_soo_tag_t{}, c); |
905 | 0 | c.set_capacity(HashtableCapacity::CreateDestroyed()); |
906 | 0 | IterateOverFullSlotsImpl(common_copy, slot_size, destroy_slot_wrapper); |
907 | 0 | c.set_capacity(common_copy.capacity()); |
908 | | } else { |
909 | | IterateOverFullSlotsImpl(c, slot_size, destroy_slot_wrapper); |
910 | | } |
911 | 0 | } |
912 | | |
913 | | template <bool kSooEnabled> |
914 | | void Clear(CommonFields& c, const PolicyFunctions& __restrict policy, |
915 | 613k | DestroySlotFn destroy_slot, void* alloc) { |
916 | 613k | if (SwisstableGenerationsEnabled() && |
917 | 0 | c.maybe_invalid_capacity().IsMovedFrom()) { |
918 | 0 | c.set_capacity(policy.soo_capacity()); |
919 | 0 | } |
920 | 613k | c.AssertNotDebugCapacity(); |
921 | 613k | const size_t cap = c.capacity(); |
922 | 613k | if constexpr (kSooEnabled) { |
923 | 37.6k | ABSL_ASSUME(cap > 0); |
924 | 37.6k | } |
925 | 613k | if (c.is_small()) { |
926 | 327k | if (!c.empty()) { |
927 | 148k | if (destroy_slot != nullptr) { |
928 | 0 | destroy_slot(&c, SingleSlotAddress<kSooEnabled>(c)); |
929 | 0 | } |
930 | 148k | DecrementSmallSize<kSooEnabled>(c); |
931 | 148k | c.infoz().RecordStorageChanged(0, cap); |
932 | 148k | } |
933 | 327k | } else { |
934 | 285k | if (destroy_slot != nullptr) { |
935 | 0 | DestroySlots(c, policy.slot_size, destroy_slot); |
936 | 0 | } |
937 | | // Iterating over this container is O(bucket_count()). When bucket_count() |
938 | | // is much greater than size(), iteration becomes prohibitively expensive. |
939 | | // For clear() it is more important to reuse the allocated array when the |
940 | | // container is small because allocation takes comparatively long time |
941 | | // compared to destruction of the elements of the container. So we pick the |
942 | | // largest bucket_count() threshold for which iteration is still fast and |
943 | | // past that we simply deallocate the array. |
944 | 285k | ClearBackingArray(c, policy, alloc, /*reuse=*/cap < 128); |
945 | 285k | } |
946 | 613k | c.set_reserved_growth(0); |
947 | 613k | c.set_reservation_size(0); |
948 | 613k | } void absl::container_internal::Clear<true>(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, void (*)(void*, void*), void*) Line | Count | Source | 915 | 37.6k | DestroySlotFn destroy_slot, void* alloc) { | 916 | 37.6k | if (SwisstableGenerationsEnabled() && | 917 | 0 | c.maybe_invalid_capacity().IsMovedFrom()) { | 918 | 0 | c.set_capacity(policy.soo_capacity()); | 919 | 0 | } | 920 | 37.6k | c.AssertNotDebugCapacity(); | 921 | 37.6k | const size_t cap = c.capacity(); | 922 | 37.6k | if constexpr (kSooEnabled) { | 923 | 37.6k | ABSL_ASSUME(cap > 0); | 924 | 37.6k | } | 925 | 37.6k | if (c.is_small()) { | 926 | 13.7k | if (!c.empty()) { | 927 | 13.1k | if (destroy_slot != nullptr) { | 928 | 0 | destroy_slot(&c, SingleSlotAddress<kSooEnabled>(c)); | 929 | 0 | } | 930 | 13.1k | DecrementSmallSize<kSooEnabled>(c); | 931 | 13.1k | c.infoz().RecordStorageChanged(0, cap); | 932 | 13.1k | } | 933 | 23.8k | } else { | 934 | 23.8k | if (destroy_slot != nullptr) { | 935 | 0 | DestroySlots(c, policy.slot_size, destroy_slot); | 936 | 0 | } | 937 | | // Iterating over this container is O(bucket_count()). When bucket_count() | 938 | | // is much greater than size(), iteration becomes prohibitively expensive. | 939 | | // For clear() it is more important to reuse the allocated array when the | 940 | | // container is small because allocation takes comparatively long time | 941 | | // compared to destruction of the elements of the container. So we pick the | 942 | | // largest bucket_count() threshold for which iteration is still fast and | 943 | | // past that we simply deallocate the array. | 944 | 23.8k | ClearBackingArray(c, policy, alloc, /*reuse=*/cap < 128); | 945 | 23.8k | } | 946 | 37.6k | c.set_reserved_growth(0); | 947 | 37.6k | c.set_reservation_size(0); | 948 | 37.6k | } |
void absl::container_internal::Clear<false>(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, void (*)(void*, void*), void*) Line | Count | Source | 915 | 576k | DestroySlotFn destroy_slot, void* alloc) { | 916 | 576k | if (SwisstableGenerationsEnabled() && | 917 | 0 | c.maybe_invalid_capacity().IsMovedFrom()) { | 918 | 0 | c.set_capacity(policy.soo_capacity()); | 919 | 0 | } | 920 | 576k | c.AssertNotDebugCapacity(); | 921 | 576k | const size_t cap = c.capacity(); | 922 | | if constexpr (kSooEnabled) { | 923 | | ABSL_ASSUME(cap > 0); | 924 | | } | 925 | 576k | if (c.is_small()) { | 926 | 313k | if (!c.empty()) { | 927 | 135k | if (destroy_slot != nullptr) { | 928 | 0 | destroy_slot(&c, SingleSlotAddress<kSooEnabled>(c)); | 929 | 0 | } | 930 | 135k | DecrementSmallSize<kSooEnabled>(c); | 931 | 135k | c.infoz().RecordStorageChanged(0, cap); | 932 | 135k | } | 933 | 313k | } else { | 934 | 262k | if (destroy_slot != nullptr) { | 935 | 0 | DestroySlots(c, policy.slot_size, destroy_slot); | 936 | 0 | } | 937 | | // Iterating over this container is O(bucket_count()). When bucket_count() | 938 | | // is much greater than size(), iteration becomes prohibitively expensive. | 939 | | // For clear() it is more important to reuse the allocated array when the | 940 | | // container is small because allocation takes comparatively long time | 941 | | // compared to destruction of the elements of the container. So we pick the | 942 | | // largest bucket_count() threshold for which iteration is still fast and | 943 | | // past that we simply deallocate the array. | 944 | 262k | ClearBackingArray(c, policy, alloc, /*reuse=*/cap < 128); | 945 | 262k | } | 946 | 576k | c.set_reserved_growth(0); | 947 | 576k | c.set_reservation_size(0); | 948 | 576k | } |
|
949 | | |
950 | | template <bool kSooEnabled> |
951 | | void Destruct(CommonFields& c, const DtorPolicy& __restrict policy, |
952 | 41.8k | DeallocBackingArrayFn dealloc, void* alloc) { |
953 | 41.8k | if constexpr (kSooEnabled) { |
954 | 24.9k | DestructSoo(c, policy, dealloc, alloc); |
955 | 24.9k | } else { |
956 | 16.9k | DestructNonSoo(c, policy, dealloc, alloc); |
957 | 16.9k | } |
958 | 41.8k | } void absl::container_internal::Destruct<true>(absl::container_internal::CommonFields&, absl::container_internal::DtorPolicy const&, void (*)(void*, void*, unsigned long), void*) Line | Count | Source | 952 | 24.9k | DeallocBackingArrayFn dealloc, void* alloc) { | 953 | 24.9k | if constexpr (kSooEnabled) { | 954 | 24.9k | DestructSoo(c, policy, dealloc, alloc); | 955 | | } else { | 956 | | DestructNonSoo(c, policy, dealloc, alloc); | 957 | | } | 958 | 24.9k | } |
void absl::container_internal::Destruct<false>(absl::container_internal::CommonFields&, absl::container_internal::DtorPolicy const&, void (*)(void*, void*, unsigned long), void*) Line | Count | Source | 952 | 16.9k | DeallocBackingArrayFn dealloc, void* alloc) { | 953 | | if constexpr (kSooEnabled) { | 954 | | DestructSoo(c, policy, dealloc, alloc); | 955 | 16.9k | } else { | 956 | 16.9k | DestructNonSoo(c, policy, dealloc, alloc); | 957 | 16.9k | } | 958 | 16.9k | } |
|
959 | | template <bool kSooEnabled> |
960 | | void Destruct(CommonFields& c, const DtorPolicy& __restrict policy, |
961 | 41.8k | DeallocBackingArrayFn dealloc) { |
962 | 41.8k | Destruct<kSooEnabled>(c, policy, dealloc, /*alloc=*/&c); |
963 | 41.8k | } void absl::container_internal::Destruct<true>(absl::container_internal::CommonFields&, absl::container_internal::DtorPolicy const&, void (*)(void*, void*, unsigned long)) Line | Count | Source | 961 | 24.9k | DeallocBackingArrayFn dealloc) { | 962 | 24.9k | Destruct<kSooEnabled>(c, policy, dealloc, /*alloc=*/&c); | 963 | 24.9k | } |
void absl::container_internal::Destruct<false>(absl::container_internal::CommonFields&, absl::container_internal::DtorPolicy const&, void (*)(void*, void*, unsigned long)) Line | Count | Source | 961 | 16.9k | DeallocBackingArrayFn dealloc) { | 962 | 16.9k | Destruct<kSooEnabled>(c, policy, dealloc, /*alloc=*/&c); | 963 | 16.9k | } |
|
964 | | template <bool kSooEnabled> |
965 | 41.8k | void Destruct(CommonFields& c, const DtorPolicy& __restrict policy) { |
966 | 41.8k | Destruct<kSooEnabled>(c, policy, kStandardDeallocBackingArrayFn); |
967 | 41.8k | } void absl::container_internal::Destruct<true>(absl::container_internal::CommonFields&, absl::container_internal::DtorPolicy const&) Line | Count | Source | 965 | 24.9k | void Destruct(CommonFields& c, const DtorPolicy& __restrict policy) { | 966 | 24.9k | Destruct<kSooEnabled>(c, policy, kStandardDeallocBackingArrayFn); | 967 | 24.9k | } |
void absl::container_internal::Destruct<false>(absl::container_internal::CommonFields&, absl::container_internal::DtorPolicy const&) Line | Count | Source | 965 | 16.9k | void Destruct(CommonFields& c, const DtorPolicy& __restrict policy) { | 966 | 16.9k | Destruct<kSooEnabled>(c, policy, kStandardDeallocBackingArrayFn); | 967 | 16.9k | } |
|
968 | | |
969 | | namespace { |
970 | | |
971 | | // Iterates over full slots in old table, finds new positions for them and |
972 | | // transfers the slots. |
973 | | // This function is used for reserving or rehashing non-empty tables. |
974 | | // This use case is rare so the function is type erased. |
975 | | // Returns the total probe length. |
976 | | size_t FindNewPositionsAndTransferSlots( |
977 | | CommonFields& common, const PolicyFunctions& __restrict policy, |
978 | 0 | ctrl_t* old_ctrl, void* old_slots, size_t old_capacity) { |
979 | 0 | void* new_slots = common.slot_array(common.capacity()); |
980 | 0 | const void* hash_fn = policy.hash_fn(common); |
981 | 0 | const size_t slot_size = policy.slot_size; |
982 | 0 | const size_t seed = common.seed().seed(); |
983 | |
|
984 | 0 | const auto insert_slot = [&](void* slot) { |
985 | 0 | size_t hash = policy.hash_slot(hash_fn, slot, seed); |
986 | 0 | FindInfo target; |
987 | 0 | if (common.is_small()) { |
988 | 0 | target = FindInfo{0, 0}; |
989 | 0 | } else { |
990 | 0 | target = find_first_non_full(common, hash); |
991 | 0 | SetCtrl(common, target.offset, H2(hash), slot_size); |
992 | 0 | } |
993 | 0 | policy.transfer_n(&common, SlotAddress(new_slots, target.offset, slot_size), |
994 | 0 | slot, 1); |
995 | 0 | return target.probe_length; |
996 | 0 | }; |
997 | 0 | if (IsSmallCapacity(old_capacity)) { |
998 | 0 | if (common.size() == 1) insert_slot(old_slots); |
999 | 0 | return 0; |
1000 | 0 | } |
1001 | 0 | size_t total_probe_length = 0; |
1002 | 0 | for (size_t i = 0; i < old_capacity; ++i) { |
1003 | 0 | if (IsFull(old_ctrl[i])) { |
1004 | 0 | total_probe_length += insert_slot(old_slots); |
1005 | 0 | } |
1006 | 0 | old_slots = NextSlot(old_slots, slot_size); |
1007 | 0 | } |
1008 | 0 | return total_probe_length; |
1009 | 0 | } |
1010 | | |
1011 | | void ReportGrowthToInfozImpl(CommonFields& common, HashtablezInfoHandle infoz, |
1012 | | size_t hash, size_t total_probe_length, |
1013 | 0 | size_t distance_from_desired) { |
1014 | 0 | ABSL_SWISSTABLE_ASSERT(infoz.IsSampled()); |
1015 | 0 | infoz.RecordStorageChanged(common.size() - 1, common.capacity()); |
1016 | 0 | infoz.RecordRehash(total_probe_length); |
1017 | 0 | infoz.RecordInsertMiss(hash, distance_from_desired); |
1018 | 0 | common.set_has_infoz(); |
1019 | 0 | common.set_infoz(infoz); |
1020 | 0 | } |
1021 | | |
1022 | | // Specialization to avoid passing two 0s from hot function. |
1023 | | ABSL_ATTRIBUTE_NOINLINE void ReportSingleGroupTableGrowthToInfoz( |
1024 | 0 | CommonFields& common, HashtablezInfoHandle infoz, size_t hash) { |
1025 | 0 | ReportGrowthToInfozImpl(common, infoz, hash, /*total_probe_length=*/0, |
1026 | 0 | /*distance_from_desired=*/0); |
1027 | 0 | } |
1028 | | |
1029 | | ABSL_ATTRIBUTE_NOINLINE void ReportGrowthToInfoz(CommonFields& common, |
1030 | | HashtablezInfoHandle infoz, |
1031 | | size_t hash, |
1032 | | size_t total_probe_length, |
1033 | 0 | size_t distance_from_desired) { |
1034 | 0 | ReportGrowthToInfozImpl(common, infoz, hash, total_probe_length, |
1035 | 0 | distance_from_desired); |
1036 | 0 | } |
1037 | | |
1038 | | ABSL_ATTRIBUTE_NOINLINE void ReportResizeToInfoz(CommonFields& common, |
1039 | | HashtablezInfoHandle infoz, |
1040 | 0 | size_t total_probe_length) { |
1041 | 0 | ABSL_SWISSTABLE_ASSERT(infoz.IsSampled()); |
1042 | 0 | infoz.RecordStorageChanged(common.size(), common.capacity()); |
1043 | 0 | infoz.RecordRehash(total_probe_length); |
1044 | 0 | common.set_has_infoz(); |
1045 | 0 | common.set_infoz(infoz); |
1046 | 0 | } |
1047 | | |
1048 | | struct BackingArrayPtrs { |
1049 | | ctrl_t* ctrl; |
1050 | | void* slots; |
1051 | | }; |
1052 | | |
1053 | | BackingArrayPtrs AllocBackingArray(CommonFields& common, |
1054 | | const PolicyFunctions& __restrict policy, |
1055 | | size_t new_capacity, bool has_infoz, |
1056 | 267k | void* alloc, size_t blocked_element_count) { |
1057 | 267k | RawHashSetLayout layout(new_capacity, policy.slot_size, policy.slot_align, |
1058 | 267k | has_infoz, blocked_element_count); |
1059 | | // Perform a direct call in the common case to allow for profile-guided |
1060 | | // heap optimization (PGHO) to understand which allocation function is used. |
1061 | 267k | constexpr size_t kDefaultAlignment = BackingArrayAlignment(alignof(size_t)); |
1062 | 267k | char* mem = static_cast<char*>( |
1063 | 267k | ABSL_PREDICT_TRUE( |
1064 | 267k | policy.alloc == |
1065 | 267k | (&AllocateBackingArray<kDefaultAlignment, std::allocator<char>>)) |
1066 | 267k | ? AllocateBackingArray<kDefaultAlignment, std::allocator<char>>( |
1067 | 267k | alloc, layout.alloc_size()) |
1068 | 267k | : policy.alloc(alloc, layout.alloc_size())); |
1069 | 267k | const GenerationType old_generation = common.generation(); |
1070 | 267k | common.set_generation_ptr( |
1071 | 267k | reinterpret_cast<GenerationType*>(mem + layout.generation_offset())); |
1072 | 267k | common.set_generation(NextGeneration(old_generation)); |
1073 | | |
1074 | 267k | return {reinterpret_cast<ctrl_t*>(mem + layout.control_offset()), |
1075 | 267k | mem + layout.slot_offset()}; |
1076 | 267k | } |
1077 | | |
1078 | | void ResizeEmptyNonAllocatedTableImpl(CommonFields& common, |
1079 | | const PolicyFunctions& __restrict policy, |
1080 | | size_t new_capacity, |
1081 | | size_t blocked_element_count, |
1082 | 0 | bool force_infoz) { |
1083 | 0 | ABSL_SWISSTABLE_ASSERT(IsValidCapacity(new_capacity)); |
1084 | 0 | ABSL_SWISSTABLE_ASSERT(new_capacity > policy.soo_capacity()); |
1085 | 0 | ABSL_SWISSTABLE_ASSERT(!force_infoz || policy.soo_enabled); |
1086 | 0 | ABSL_SWISSTABLE_ASSERT(common.capacity() == policy.soo_capacity()); |
1087 | 0 | ABSL_SWISSTABLE_ASSERT(common.empty()); |
1088 | 0 | const size_t slot_size = policy.slot_size; |
1089 | 0 | HashtablezInfoHandle infoz; |
1090 | 0 | const bool should_sample = |
1091 | 0 | policy.is_hashtablez_eligible && (force_infoz || ShouldSampleNextTable()); |
1092 | 0 | if (ABSL_PREDICT_FALSE(should_sample)) { |
1093 | 0 | infoz = ForcedTrySample(slot_size, policy.key_size, policy.value_size, |
1094 | 0 | policy.soo_capacity()); |
1095 | 0 | } |
1096 | 0 | const bool has_infoz = infoz.IsSampled(); |
1097 | 0 | void* alloc = policy.get_char_alloc(common); |
1098 | |
|
1099 | 0 | common.set_capacity(new_capacity); |
1100 | 0 | common.init_blocked_element_count(blocked_element_count); |
1101 | 0 | const auto [new_ctrl, new_slots] = AllocBackingArray( |
1102 | 0 | common, policy, new_capacity, has_infoz, alloc, blocked_element_count); |
1103 | 0 | common.set_control(new_ctrl); |
1104 | 0 | common.generate_new_seed(has_infoz); |
1105 | |
|
1106 | 0 | ResetCtrl(common, slot_size, blocked_element_count); |
1107 | 0 | ResetGrowthLeft(new_capacity, blocked_element_count, common); |
1108 | |
|
1109 | 0 | if (ABSL_PREDICT_FALSE(has_infoz)) { |
1110 | 0 | ReportResizeToInfoz(common, infoz, 0); |
1111 | 0 | } |
1112 | 0 | } |
1113 | | |
1114 | | // If the table was SOO, initializes new control bytes and transfers slot. |
1115 | | // After transferring the slot, sets control and slots in CommonFields. |
1116 | | // It is rare to resize an SOO table with one element to a large size. |
1117 | | // Requires: `c` contains SOO data. |
1118 | | void InsertOldSooSlotAndInitializeControlBytes( |
1119 | | CommonFields& c, const PolicyFunctions& __restrict policy, ctrl_t* new_ctrl, |
1120 | 0 | void* new_slots, bool has_infoz) { |
1121 | 0 | ABSL_SWISSTABLE_ASSERT(c.size() == policy.soo_capacity()); |
1122 | 0 | ABSL_SWISSTABLE_ASSERT(policy.soo_enabled); |
1123 | 0 | const size_t new_capacity = c.capacity(); |
1124 | |
|
1125 | 0 | c.generate_new_seed(has_infoz); |
1126 | |
|
1127 | 0 | const size_t soo_slot_hash = |
1128 | 0 | policy.hash_slot(policy.hash_fn(c), c.soo_data(), c.seed().seed()); |
1129 | 0 | size_t offset = probe(ProbeCapacity{new_capacity}, soo_slot_hash).offset(); |
1130 | 0 | offset = offset == new_capacity ? 0 : offset; |
1131 | 0 | SanitizerPoisonMemoryRegion(new_slots, policy.slot_size * new_capacity); |
1132 | 0 | void* target_slot = SlotAddress(new_slots, offset, policy.slot_size); |
1133 | 0 | SanitizerUnpoisonMemoryRegion(target_slot, policy.slot_size); |
1134 | 0 | policy.transfer_n(&c, target_slot, c.soo_data(), 1); |
1135 | 0 | c.set_control(new_ctrl); |
1136 | 0 | ResetCtrl(c, policy.slot_size, /*blocked_element_count=*/0); |
1137 | 0 | SetCtrl(c, offset, H2(soo_slot_hash), policy.slot_size); |
1138 | 0 | } |
1139 | | |
1140 | | enum class ResizeFullSooTableSamplingMode { |
1141 | | kNoSampling, |
1142 | | // Force sampling. If the table was still not sampled, do not resize. |
1143 | | kForceSampleNoResizeIfUnsampled, |
1144 | | }; |
1145 | | |
1146 | | void AssertSoo([[maybe_unused]] CommonFields& common, |
1147 | 25.2k | [[maybe_unused]] const PolicyFunctions& __restrict policy) { |
1148 | 25.2k | ABSL_SWISSTABLE_ASSERT(policy.soo_enabled); |
1149 | 25.2k | ABSL_SWISSTABLE_ASSERT(common.capacity() == policy.soo_capacity()); |
1150 | 25.2k | } |
1151 | | void AssertFullSoo([[maybe_unused]] CommonFields& common, |
1152 | 0 | [[maybe_unused]] const PolicyFunctions& __restrict policy) { |
1153 | 0 | AssertSoo(common, policy); |
1154 | 0 | ABSL_SWISSTABLE_ASSERT(common.size() == policy.soo_capacity()); |
1155 | 0 | } |
1156 | | |
1157 | | void ResizeFullSooTable(CommonFields& common, |
1158 | | const PolicyFunctions& __restrict policy, |
1159 | | size_t new_capacity, |
1160 | 0 | ResizeFullSooTableSamplingMode sampling_mode) { |
1161 | 0 | AssertFullSoo(common, policy); |
1162 | 0 | const size_t slot_size = policy.slot_size; |
1163 | 0 | void* alloc = policy.get_char_alloc(common); |
1164 | 0 | constexpr size_t kTableSize = 1; |
1165 | |
|
1166 | 0 | HashtablezInfoHandle infoz; |
1167 | 0 | bool has_infoz = false; |
1168 | 0 | if (sampling_mode == |
1169 | 0 | ResizeFullSooTableSamplingMode::kForceSampleNoResizeIfUnsampled) { |
1170 | 0 | if (ABSL_PREDICT_FALSE(policy.is_hashtablez_eligible)) { |
1171 | 0 | infoz = ForcedTrySample(slot_size, policy.key_size, policy.value_size, |
1172 | 0 | policy.soo_capacity()); |
1173 | 0 | } |
1174 | |
|
1175 | 0 | if (!infoz.IsSampled()) return; |
1176 | 0 | has_infoz = true; |
1177 | 0 | } |
1178 | | |
1179 | 0 | common.set_capacity(new_capacity); |
1180 | | |
1181 | | // We do not set control and slots in CommonFields yet to avoid overriding |
1182 | | // SOO data. |
1183 | 0 | const auto [new_ctrl, new_slots] = |
1184 | 0 | AllocBackingArray(common, policy, new_capacity, has_infoz, alloc, |
1185 | 0 | /*blocked_element_count=*/0); |
1186 | |
|
1187 | 0 | InsertOldSooSlotAndInitializeControlBytes(common, policy, new_ctrl, new_slots, |
1188 | 0 | has_infoz); |
1189 | 0 | ResetGrowthLeft(new_capacity, kTableSize, common); |
1190 | 0 | if (has_infoz) { |
1191 | 0 | common.set_has_infoz(); |
1192 | 0 | common.set_infoz(infoz); |
1193 | 0 | infoz.RecordStorageChanged(kTableSize, new_capacity); |
1194 | 0 | } |
1195 | 0 | } |
1196 | | |
1197 | | void GrowIntoSingleGroupShuffleControlBytes(ctrl_t* __restrict old_ctrl, |
1198 | | size_t old_capacity, |
1199 | | size_t old_blocked_element_count, |
1200 | | ctrl_t* __restrict new_ctrl, |
1201 | 69.2k | size_t new_capacity) { |
1202 | 69.2k | ABSL_SWISSTABLE_ASSERT(is_single_group(new_capacity)); |
1203 | 69.2k | constexpr size_t kHalfWidth = Group::kWidth / 2; |
1204 | 69.2k | ABSL_ASSUME(old_capacity < kHalfWidth); |
1205 | 69.2k | ABSL_ASSUME(old_capacity > 0); |
1206 | 69.2k | static_assert(Group::kWidth == 8 || Group::kWidth == 16, |
1207 | 69.2k | "Group size is not supported."); |
1208 | | |
1209 | | // NOTE: operations are done with compile time known size = 8. |
1210 | | // Compiler optimizes that into single ASM operation. |
1211 | | |
1212 | | // Load the bytes from old_capacity. This contains |
1213 | | // - the sentinel byte |
1214 | | // - all the old control bytes |
1215 | | // - the rest is filled with kEmpty bytes |
1216 | | // Example: |
1217 | | // old_ctrl = 012S012EEEEEEEEE... |
1218 | | // copied_bytes = S012EEEE |
1219 | | // Example with blocked elements: |
1220 | | // old_ctrl = 01SS01SEEEEEEEEE... |
1221 | | // copied_bytes = S01SEEEE |
1222 | 69.2k | uint64_t copied_bytes = absl::little_endian::Load64(old_ctrl + old_capacity); |
1223 | | |
1224 | | // We change the sentinel byte to kEmpty before storing to both the start of |
1225 | | // the new_ctrl, and past the end of the new_ctrl later for the new cloned |
1226 | | // bytes. Note that this is faster than setting the sentinel byte to kEmpty |
1227 | | // after the copy directly in new_ctrl because we are limited on store |
1228 | | // bandwidth. |
1229 | 69.2k | static constexpr uint64_t kEmptyXorSentinel = |
1230 | 69.2k | static_cast<uint8_t>(ctrl_t::kEmpty) ^ |
1231 | 69.2k | static_cast<uint8_t>(ctrl_t::kSentinel); |
1232 | | |
1233 | | // Replace the first byte kSentinel with kEmpty. |
1234 | | // Resulting bytes will be shifted by one byte old control blocks. |
1235 | | // Example: |
1236 | | // old_ctrl = 012S012EEEEEEEEE... |
1237 | | // before = S012EEEE |
1238 | | // after = E012EEEE |
1239 | 69.2k | copied_bytes ^= kEmptyXorSentinel; |
1240 | | |
1241 | 69.2k | if (ABSL_PREDICT_FALSE(old_blocked_element_count > 0)) { |
1242 | | // Replacing blocked sentinel elements with kEmpty. |
1243 | 0 | static constexpr uint64_t kAllBytesEmptyXorSentinel = |
1244 | 0 | kEmptyXorSentinel * uint64_t{0x0101010101010101}; |
1245 | 0 | uint64_t blocked_mask = kAllBytesEmptyXorSentinel; |
1246 | | // Keep old_blocked_element_count bytes in the mask. |
1247 | 0 | blocked_mask >>= 64 - old_blocked_element_count * 8; |
1248 | | // Shift the mask to the start of the blocked elements bytes. |
1249 | 0 | blocked_mask <<= (old_capacity - old_blocked_element_count + 1) * 8; |
1250 | | // Example with blocked elements: |
1251 | | // old_ctrl = 0SSS0SSEEEEEEEEE... |
1252 | | // before = E0SSEEEE |
1253 | | // after = E0EEEEEE |
1254 | 0 | copied_bytes ^= blocked_mask; |
1255 | 0 | } |
1256 | | |
1257 | 69.2k | if (Group::kWidth == 8) { |
1258 | | // With group size 8, we can grow with two write operations. |
1259 | 0 | ABSL_SWISSTABLE_ASSERT(old_capacity < 8 && |
1260 | 0 | "old_capacity is too large for group size 8"); |
1261 | 0 | absl::little_endian::Store64(new_ctrl, copied_bytes); |
1262 | |
|
1263 | 0 | static constexpr uint64_t kSentinal64 = |
1264 | 0 | static_cast<uint8_t>(ctrl_t::kSentinel); |
1265 | | |
1266 | | // Prepend kSentinel byte to the beginning of copied_bytes. |
1267 | | // We have maximum 3 non-empty bytes at the beginning of copied_bytes for |
1268 | | // group size 8. |
1269 | | // Example: |
1270 | | // old_ctrl = 012S012EEEE |
1271 | | // before = E012EEEE |
1272 | | // after = SE012EEE |
1273 | 0 | copied_bytes = (copied_bytes << 8) ^ kSentinal64; |
1274 | 0 | absl::little_endian::Store64(new_ctrl + new_capacity, copied_bytes); |
1275 | | // Example for capacity 3: |
1276 | | // old_ctrl = 012S012EEEE |
1277 | | // After the first store: |
1278 | | // >! |
1279 | | // new_ctrl = E012EEEE??????? |
1280 | | // After the second store: |
1281 | | // >! |
1282 | | // new_ctrl = E012EEESE012EEE |
1283 | 0 | return; |
1284 | 0 | } |
1285 | | |
1286 | 69.2k | ABSL_SWISSTABLE_ASSERT(Group::kWidth == 16); // NOLINT(misc-static-assert) |
1287 | | |
1288 | | // Fill the second half of the main control bytes with kEmpty. |
1289 | | // For small capacity that may write into mirrored control bytes. |
1290 | | // It is fine as we will overwrite all the bytes later. |
1291 | 69.2k | std::memset(new_ctrl + kHalfWidth, static_cast<int8_t>(ctrl_t::kEmpty), |
1292 | 69.2k | kHalfWidth); |
1293 | | // Fill the second half of the mirrored control bytes with kEmpty. |
1294 | 69.2k | std::memset(new_ctrl + new_capacity + kHalfWidth, |
1295 | 69.2k | static_cast<int8_t>(ctrl_t::kEmpty), kHalfWidth); |
1296 | | // Copy the first half of the non-mirrored control bytes. |
1297 | 69.2k | absl::little_endian::Store64(new_ctrl, copied_bytes); |
1298 | 69.2k | new_ctrl[new_capacity] = ctrl_t::kSentinel; |
1299 | | // Copy the first half of the mirrored control bytes. |
1300 | 69.2k | absl::little_endian::Store64(new_ctrl + new_capacity + 1, copied_bytes); |
1301 | | |
1302 | | // Example for growth capacity 1->3: |
1303 | | // old_ctrl = 0S0EEEEEEEEEEEEEE |
1304 | | // new_ctrl at the end = E0ESE0EEEEEEEEEEEEE |
1305 | | // >! |
1306 | | // new_ctrl after 1st memset = ????????EEEEEEEE??? |
1307 | | // >! |
1308 | | // new_ctrl after 2nd memset = ????????EEEEEEEEEEE |
1309 | | // >! |
1310 | | // new_ctrl after 1st store = E0EEEEEEEEEEEEEEEEE |
1311 | | // new_ctrl after kSentinel = E0ESEEEEEEEEEEEEEEE |
1312 | | // >! |
1313 | | // new_ctrl after 2nd store = E0ESE0EEEEEEEEEEEEE |
1314 | | |
1315 | | // Example for growth capacity 3->7: |
1316 | | // old_ctrl = 012S012EEEEEEEEEEEE |
1317 | | // new_ctrl at the end = E012EEESE012EEEEEEEEEEE |
1318 | | // >! |
1319 | | // new_ctrl after 1st memset = ????????EEEEEEEE??????? |
1320 | | // >! |
1321 | | // new_ctrl after 2nd memset = ????????EEEEEEEEEEEEEEE |
1322 | | // >! |
1323 | | // new_ctrl after 1st store = E012EEEEEEEEEEEEEEEEEEE |
1324 | | // new_ctrl after kSentinel = E012EEESEEEEEEEEEEEEEEE |
1325 | | // >! |
1326 | | // new_ctrl after 2nd store = E012EEESE012EEEEEEEEEEE |
1327 | | |
1328 | | // Example for growth capacity 7->15: |
1329 | | // old_ctrl = 0123456S0123456EEEEEEEE |
1330 | | // new_ctrl at the end = E0123456EEEEEEESE0123456EEEEEEE |
1331 | | // >! |
1332 | | // new_ctrl after 1st memset = ????????EEEEEEEE??????????????? |
1333 | | // >! |
1334 | | // new_ctrl after 2nd memset = ????????EEEEEEEE???????EEEEEEEE |
1335 | | // >! |
1336 | | // new_ctrl after 1st store = E0123456EEEEEEEE???????EEEEEEEE |
1337 | | // new_ctrl after kSentinel = E0123456EEEEEEES???????EEEEEEEE |
1338 | | // >! |
1339 | | // new_ctrl after 2nd store = E0123456EEEEEEESE0123456EEEEEEE |
1340 | 69.2k | } |
1341 | | |
1342 | | // Size of the buffer we allocate on stack for storing probed elements in |
1343 | | // GrowToNextCapacity algorithm. |
1344 | | constexpr size_t kProbedElementsBufferSize = 512; |
1345 | | |
1346 | | // Decodes information about probed elements from contiguous memory. |
1347 | | // Finds new position for each element and transfers it to the new slots. |
1348 | | // Returns the total probe length. |
1349 | | template <typename ProbedItem> |
1350 | | ABSL_ATTRIBUTE_NOINLINE size_t DecodeAndInsertImpl( |
1351 | | CommonFields& c, const PolicyFunctions& __restrict policy, |
1352 | 48.1k | const ProbedItem* start, const ProbedItem* end, void* old_slots) { |
1353 | 48.1k | const HashtableCapacity new_capacity = c.capacity_impl(); |
1354 | | |
1355 | 48.1k | void* new_slots = c.slot_array(new_capacity.capacity()); |
1356 | 48.1k | ctrl_t* new_ctrl = c.control(); |
1357 | 48.1k | size_t total_probe_length = 0; |
1358 | | |
1359 | 48.1k | const size_t slot_size = policy.slot_size; |
1360 | 48.1k | auto transfer_n = policy.transfer_n; |
1361 | | |
1362 | 160k | for (; start < end; ++start) { |
1363 | 112k | const FindInfo target = find_first_non_full_from_h1( |
1364 | 112k | new_ctrl, static_cast<size_t>(start->h1), new_capacity); |
1365 | 112k | total_probe_length += target.probe_length; |
1366 | 112k | const size_t old_index = static_cast<size_t>(start->source_offset); |
1367 | 112k | const size_t new_i = target.offset; |
1368 | 112k | ABSL_SWISSTABLE_ASSERT(old_index < new_capacity.capacity() / 2); |
1369 | 112k | ABSL_SWISSTABLE_ASSERT(new_i < new_capacity.capacity()); |
1370 | 112k | ABSL_SWISSTABLE_ASSERT(IsEmpty(new_ctrl[new_i])); |
1371 | 112k | void* src_slot = SlotAddress(old_slots, old_index, slot_size); |
1372 | 112k | void* dst_slot = SlotAddress(new_slots, new_i, slot_size); |
1373 | 112k | SanitizerUnpoisonMemoryRegion(dst_slot, slot_size); |
1374 | 112k | transfer_n(&c, dst_slot, src_slot, 1); |
1375 | 112k | SetCtrlInLargeTable(c, new_i, static_cast<h2_t>(start->h2), slot_size); |
1376 | 112k | } |
1377 | 48.1k | return total_probe_length; |
1378 | 48.1k | } raw_hash_set.cc:unsigned long absl::container_internal::(anonymous namespace)::DecodeAndInsertImpl<absl::container_internal::ProbedItemImpl<unsigned int, 32ul> >(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, absl::container_internal::ProbedItemImpl<unsigned int, 32ul> const*, absl::container_internal::ProbedItemImpl<unsigned int, 32ul> const*, void*) Line | Count | Source | 1352 | 48.1k | const ProbedItem* start, const ProbedItem* end, void* old_slots) { | 1353 | 48.1k | const HashtableCapacity new_capacity = c.capacity_impl(); | 1354 | | | 1355 | 48.1k | void* new_slots = c.slot_array(new_capacity.capacity()); | 1356 | 48.1k | ctrl_t* new_ctrl = c.control(); | 1357 | 48.1k | size_t total_probe_length = 0; | 1358 | | | 1359 | 48.1k | const size_t slot_size = policy.slot_size; | 1360 | 48.1k | auto transfer_n = policy.transfer_n; | 1361 | | | 1362 | 160k | for (; start < end; ++start) { | 1363 | 112k | const FindInfo target = find_first_non_full_from_h1( | 1364 | 112k | new_ctrl, static_cast<size_t>(start->h1), new_capacity); | 1365 | 112k | total_probe_length += target.probe_length; | 1366 | 112k | const size_t old_index = static_cast<size_t>(start->source_offset); | 1367 | 112k | const size_t new_i = target.offset; | 1368 | 112k | ABSL_SWISSTABLE_ASSERT(old_index < new_capacity.capacity() / 2); | 1369 | 112k | ABSL_SWISSTABLE_ASSERT(new_i < new_capacity.capacity()); | 1370 | 112k | ABSL_SWISSTABLE_ASSERT(IsEmpty(new_ctrl[new_i])); | 1371 | 112k | void* src_slot = SlotAddress(old_slots, old_index, slot_size); | 1372 | 112k | void* dst_slot = SlotAddress(new_slots, new_i, slot_size); | 1373 | 112k | SanitizerUnpoisonMemoryRegion(dst_slot, slot_size); | 1374 | 112k | transfer_n(&c, dst_slot, src_slot, 1); | 1375 | 112k | SetCtrlInLargeTable(c, new_i, static_cast<h2_t>(start->h2), slot_size); | 1376 | 112k | } | 1377 | 48.1k | return total_probe_length; | 1378 | 48.1k | } |
Unexecuted instantiation: raw_hash_set.cc:unsigned long absl::container_internal::(anonymous namespace)::DecodeAndInsertImpl<absl::container_internal::ProbedItemImpl<unsigned long, 64ul> >(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, absl::container_internal::ProbedItemImpl<unsigned long, 64ul> const*, absl::container_internal::ProbedItemImpl<unsigned long, 64ul> const*, void*) Unexecuted instantiation: raw_hash_set.cc:unsigned long absl::container_internal::(anonymous namespace)::DecodeAndInsertImpl<absl::container_internal::ProbedItemImpl<unsigned long, 122ul> >(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, absl::container_internal::ProbedItemImpl<unsigned long, 122ul> const*, absl::container_internal::ProbedItemImpl<unsigned long, 122ul> const*, void*) |
1379 | | |
1380 | | // Sentinel value for the start of marked elements. |
1381 | | // Signals that there are no marked elements. |
1382 | | constexpr size_t kNoMarkedElementsSentinel = ~size_t{}; |
1383 | | |
1384 | | // Process probed elements that did not fit into available buffers. |
1385 | | // We marked them in control bytes as kMarkedForSlowTransfer. |
1386 | | // Hash recomputation and full probing is done here. |
1387 | | // This use case should be extremely rare. |
1388 | | ABSL_ATTRIBUTE_NOINLINE size_t ProcessProbedMarkedElements( |
1389 | | CommonFields& c, const PolicyFunctions& __restrict policy, ctrl_t* old_ctrl, |
1390 | 0 | void* old_slots, size_t start) { |
1391 | 0 | size_t old_capacity = PreviousCapacity(c.capacity()); |
1392 | 0 | const size_t slot_size = policy.slot_size; |
1393 | 0 | void* new_slots = c.slot_array(c.capacity()); |
1394 | 0 | size_t total_probe_length = 0; |
1395 | 0 | const void* hash_fn = policy.hash_fn(c); |
1396 | 0 | auto hash_slot = policy.hash_slot; |
1397 | 0 | auto transfer_n = policy.transfer_n; |
1398 | 0 | const size_t seed = c.seed().seed(); |
1399 | 0 | for (size_t old_index = start; old_index < old_capacity; ++old_index) { |
1400 | 0 | if (old_ctrl[old_index] != ctrl_t::kMarkedForSlowTransfer) { |
1401 | 0 | continue; |
1402 | 0 | } |
1403 | 0 | void* src_slot = SlotAddress(old_slots, old_index, slot_size); |
1404 | 0 | const size_t hash = hash_slot(hash_fn, src_slot, seed); |
1405 | 0 | const FindInfo target = find_first_non_full(c, hash); |
1406 | 0 | total_probe_length += target.probe_length; |
1407 | 0 | const size_t new_i = target.offset; |
1408 | 0 | void* dst_slot = SlotAddress(new_slots, new_i, slot_size); |
1409 | 0 | SetCtrlInLargeTable(c, new_i, H2(hash), slot_size); |
1410 | 0 | transfer_n(&c, dst_slot, src_slot, 1); |
1411 | 0 | } |
1412 | 0 | return total_probe_length; |
1413 | 0 | } |
1414 | | |
1415 | | // The largest old capacity for which it is guaranteed that all probed elements |
1416 | | // fit in ProbedItemEncoder's local buffer. |
1417 | | // For such tables, `encode_probed_element` is trivial. |
1418 | | constexpr size_t kMaxLocalBufferOldCapacity = |
1419 | | kProbedElementsBufferSize / sizeof(ProbedItem4Bytes) - 1; |
1420 | | static_assert(IsValidCapacity(kMaxLocalBufferOldCapacity)); |
1421 | | constexpr size_t kMaxLocalBufferNewCapacity = |
1422 | | NextCapacity(kMaxLocalBufferOldCapacity); |
1423 | | static_assert(kMaxLocalBufferNewCapacity <= ProbedItem4Bytes::kMaxNewCapacity); |
1424 | | static_assert(NextCapacity(kMaxLocalBufferNewCapacity) <= |
1425 | | ProbedItem4Bytes::kMaxNewCapacity); |
1426 | | |
1427 | | // Initializes mirrored control bytes after |
1428 | | // transfer_unprobed_elements_to_next_capacity. |
1429 | 107k | void InitializeMirroredControlBytes(ctrl_t* new_ctrl, size_t new_capacity) { |
1430 | 107k | std::memcpy(new_ctrl + new_capacity + 1, new_ctrl, Group::kWidth - 1); |
1431 | 107k | new_ctrl[new_capacity] = ctrl_t::kSentinel; |
1432 | 107k | } |
1433 | | |
1434 | | // Encodes probed elements into available memory. |
1435 | | // At first, a local (on stack) buffer is used. The size of the buffer is |
1436 | | // kProbedElementsBufferSize bytes. |
1437 | | // When the local buffer is full, we switch to `control_` buffer. We are allowed |
1438 | | // to overwrite `control_` buffer till the `source_offset` byte. In case we have |
1439 | | // no space in `control_` buffer, we fallback to a naive algorithm for all the |
1440 | | // rest of the probed elements. We mark elements as kSentinel in control bytes |
1441 | | // and later process them fully. See ProcessMarkedElements for details. It |
1442 | | // should be extremely rare. |
1443 | | template <typename ProbedItemType, |
1444 | | // If true, we only use the local buffer and never switch to the |
1445 | | // control buffer. |
1446 | | bool kGuaranteedFitToBuffer = false> |
1447 | | class ProbedItemEncoder { |
1448 | | public: |
1449 | | using ProbedItem = ProbedItemType; |
1450 | 107k | explicit ProbedItemEncoder(ctrl_t* control) : control_(control) {}raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned int, 32ul>, true>::ProbedItemEncoder(absl::container_internal::ctrl_t*) Line | Count | Source | 1450 | 90.6k | explicit ProbedItemEncoder(ctrl_t* control) : control_(control) {} |
raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned int, 32ul>, false>::ProbedItemEncoder(absl::container_internal::ctrl_t*) Line | Count | Source | 1450 | 17.2k | explicit ProbedItemEncoder(ctrl_t* control) : control_(control) {} |
Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned long, 64ul>, false>::ProbedItemEncoder(absl::container_internal::ctrl_t*) Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned long, 122ul>, false>::ProbedItemEncoder(absl::container_internal::ctrl_t*) |
1451 | | |
1452 | | // Encode item into the best available location. |
1453 | 112k | void EncodeItem(ProbedItem item) { |
1454 | 112k | if (ABSL_PREDICT_FALSE(!kGuaranteedFitToBuffer && pos_ >= end_)) { |
1455 | 0 | return ProcessEncodeWithOverflow(item); |
1456 | 0 | } |
1457 | 112k | ABSL_SWISSTABLE_ASSERT(pos_ < end_); |
1458 | 112k | *pos_ = item; |
1459 | 112k | ++pos_; |
1460 | 112k | } raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned int, 32ul>, true>::EncodeItem(absl::container_internal::ProbedItemImpl<unsigned int, 32ul>) Line | Count | Source | 1453 | 88.7k | void EncodeItem(ProbedItem item) { | 1454 | 88.7k | if (ABSL_PREDICT_FALSE(!kGuaranteedFitToBuffer && pos_ >= end_)) { | 1455 | 0 | return ProcessEncodeWithOverflow(item); | 1456 | 0 | } | 1457 | 88.7k | ABSL_SWISSTABLE_ASSERT(pos_ < end_); | 1458 | 88.7k | *pos_ = item; | 1459 | 88.7k | ++pos_; | 1460 | 88.7k | } |
raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned int, 32ul>, false>::EncodeItem(absl::container_internal::ProbedItemImpl<unsigned int, 32ul>) Line | Count | Source | 1453 | 23.4k | void EncodeItem(ProbedItem item) { | 1454 | 23.4k | if (ABSL_PREDICT_FALSE(!kGuaranteedFitToBuffer && pos_ >= end_)) { | 1455 | 0 | return ProcessEncodeWithOverflow(item); | 1456 | 0 | } | 1457 | 23.4k | ABSL_SWISSTABLE_ASSERT(pos_ < end_); | 1458 | 23.4k | *pos_ = item; | 1459 | 23.4k | ++pos_; | 1460 | 23.4k | } |
Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned long, 64ul>, false>::EncodeItem(absl::container_internal::ProbedItemImpl<unsigned long, 64ul>) Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned long, 122ul>, false>::EncodeItem(absl::container_internal::ProbedItemImpl<unsigned long, 122ul>) |
1461 | | |
1462 | | // Decodes information about probed elements from all available sources. |
1463 | | // Finds new position for each element and transfers it to the new slots. |
1464 | | // Returns the total probe length. |
1465 | | size_t DecodeAndInsertToTable(CommonFields& common, |
1466 | | const PolicyFunctions& __restrict policy, |
1467 | 107k | void* old_slots) const { |
1468 | 107k | if (pos_ == buffer_) { |
1469 | 59.7k | return 0; |
1470 | 59.7k | } |
1471 | 48.1k | if constexpr (kGuaranteedFitToBuffer) { |
1472 | 39.7k | return DecodeAndInsertImpl(common, policy, buffer_, pos_, old_slots); |
1473 | 39.7k | } |
1474 | 0 | size_t total_probe_length = DecodeAndInsertImpl( |
1475 | 48.1k | common, policy, buffer_, |
1476 | 48.1k | local_buffer_full_ ? buffer_ + kBufferSize : pos_, old_slots); |
1477 | 48.1k | if (!local_buffer_full_) { |
1478 | 8.37k | return total_probe_length; |
1479 | 8.37k | } |
1480 | 39.7k | total_probe_length += |
1481 | 39.7k | DecodeAndInsertToTableOverflow(common, policy, old_slots); |
1482 | 39.7k | return total_probe_length; |
1483 | 48.1k | } raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned int, 32ul>, true>::DecodeAndInsertToTable(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, void*) const Line | Count | Source | 1467 | 90.6k | void* old_slots) const { | 1468 | 90.6k | if (pos_ == buffer_) { | 1469 | 50.8k | return 0; | 1470 | 50.8k | } | 1471 | 39.7k | if constexpr (kGuaranteedFitToBuffer) { | 1472 | 39.7k | return DecodeAndInsertImpl(common, policy, buffer_, pos_, old_slots); | 1473 | 39.7k | } | 1474 | 0 | size_t total_probe_length = DecodeAndInsertImpl( | 1475 | 39.7k | common, policy, buffer_, | 1476 | 39.7k | local_buffer_full_ ? buffer_ + kBufferSize : pos_, old_slots); | 1477 | 39.7k | if (!local_buffer_full_) { | 1478 | 0 | return total_probe_length; | 1479 | 0 | } | 1480 | 39.7k | total_probe_length += | 1481 | 39.7k | DecodeAndInsertToTableOverflow(common, policy, old_slots); | 1482 | 39.7k | return total_probe_length; | 1483 | 39.7k | } |
raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned int, 32ul>, false>::DecodeAndInsertToTable(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, void*) const Line | Count | Source | 1467 | 17.2k | void* old_slots) const { | 1468 | 17.2k | if (pos_ == buffer_) { | 1469 | 8.89k | return 0; | 1470 | 8.89k | } | 1471 | | if constexpr (kGuaranteedFitToBuffer) { | 1472 | | return DecodeAndInsertImpl(common, policy, buffer_, pos_, old_slots); | 1473 | | } | 1474 | 8.37k | size_t total_probe_length = DecodeAndInsertImpl( | 1475 | 8.37k | common, policy, buffer_, | 1476 | 8.37k | local_buffer_full_ ? buffer_ + kBufferSize : pos_, old_slots); | 1477 | 8.37k | if (!local_buffer_full_) { | 1478 | 8.37k | return total_probe_length; | 1479 | 8.37k | } | 1480 | 0 | total_probe_length += | 1481 | 0 | DecodeAndInsertToTableOverflow(common, policy, old_slots); | 1482 | 0 | return total_probe_length; | 1483 | 8.37k | } |
Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned long, 64ul>, false>::DecodeAndInsertToTable(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, void*) const Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned long, 122ul>, false>::DecodeAndInsertToTable(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, void*) const |
1484 | | |
1485 | | private: |
1486 | 0 | static ProbedItem* AlignToNextItem(void* ptr) { |
1487 | 0 | return reinterpret_cast<ProbedItem*>(AlignUpTo( |
1488 | 0 | reinterpret_cast<uintptr_t>(ptr), alignof(ProbedItem))); |
1489 | 0 | } Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned int, 32ul>, false>::AlignToNextItem(void*) Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned long, 64ul>, false>::AlignToNextItem(void*) Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned long, 122ul>, false>::AlignToNextItem(void*) Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned int, 32ul>, true>::AlignToNextItem(void*) |
1490 | | |
1491 | 0 | ProbedItem* OverflowBufferStart() const { |
1492 | 0 | ABSL_SWISSTABLE_ASSERT(!kGuaranteedFitToBuffer && |
1493 | 0 | "OverflowBufferStart should not be called when " |
1494 | 0 | "kGuaranteedFitToBuffer is true."); |
1495 | | // We reuse GrowthInfo memory as well. |
1496 | 0 | return AlignToNextItem( |
1497 | 0 | control_ - MetadataBeforeControlSize(/*has_infoz=*/false, |
1498 | 0 | NextCapacity(kMaxLocalBufferOldCapacity))); |
1499 | 0 | } Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned int, 32ul>, false>::OverflowBufferStart() const Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned long, 64ul>, false>::OverflowBufferStart() const Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned long, 122ul>, false>::OverflowBufferStart() const Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned int, 32ul>, true>::OverflowBufferStart() const |
1500 | | |
1501 | | // Encodes item when previously allocated buffer is full. |
1502 | | // At first that happens when local buffer is full. |
1503 | | // We switch from the local buffer to the control buffer. |
1504 | | // Every time this function is called, the available buffer is extended till |
1505 | | // `item.source_offset` byte in the control buffer. |
1506 | | // After the buffer is extended, this function wouldn't be called till the |
1507 | | // buffer is exhausted. |
1508 | | // |
1509 | | // If there's no space in the control buffer, we fallback to naive algorithm |
1510 | | // and mark probed elements as kMarkedForSlowTransfer in the control buffer. |
1511 | | // In this case, we will call this function for every subsequent probed |
1512 | | // element. |
1513 | 0 | ABSL_ATTRIBUTE_NOINLINE void ProcessEncodeWithOverflow(ProbedItem item) { |
1514 | 0 | if (!local_buffer_full_) { |
1515 | 0 | local_buffer_full_ = true; |
1516 | 0 | pos_ = OverflowBufferStart(); |
1517 | 0 | } |
1518 | 0 | const size_t source_offset = static_cast<size_t>(item.source_offset); |
1519 | | // We are in fallback mode so we can't reuse control buffer anymore. |
1520 | | // Probed elements are marked as kMarkedForSlowTransfer in the control |
1521 | | // buffer. |
1522 | 0 | if (ABSL_PREDICT_FALSE(marked_elements_starting_position_ != |
1523 | 0 | kNoMarkedElementsSentinel)) { |
1524 | 0 | control_[source_offset] = ctrl_t::kMarkedForSlowTransfer; |
1525 | 0 | return; |
1526 | 0 | } |
1527 | | // Refresh the end pointer to the new available position. |
1528 | | // Invariant: if pos < end, then we have at least sizeof(ProbedItem) bytes |
1529 | | // to write. |
1530 | 0 | end_ = control_ + source_offset + 1 - sizeof(ProbedItem); |
1531 | 0 | if (ABSL_PREDICT_TRUE(pos_ < end_)) { |
1532 | 0 | *pos_ = item; |
1533 | 0 | ++pos_; |
1534 | 0 | return; |
1535 | 0 | } |
1536 | 0 | control_[source_offset] = ctrl_t::kMarkedForSlowTransfer; |
1537 | 0 | marked_elements_starting_position_ = source_offset; |
1538 | | // Now we will always fall down to `ProcessEncodeWithOverflow`. |
1539 | 0 | ABSL_SWISSTABLE_ASSERT(pos_ >= end_); |
1540 | 0 | } Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned int, 32ul>, false>::ProcessEncodeWithOverflow(absl::container_internal::ProbedItemImpl<unsigned int, 32ul>) Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned long, 64ul>, false>::ProcessEncodeWithOverflow(absl::container_internal::ProbedItemImpl<unsigned long, 64ul>) Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned long, 122ul>, false>::ProcessEncodeWithOverflow(absl::container_internal::ProbedItemImpl<unsigned long, 122ul>) Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned int, 32ul>, true>::ProcessEncodeWithOverflow(absl::container_internal::ProbedItemImpl<unsigned int, 32ul>) |
1541 | | |
1542 | | // Decodes information about probed elements from control buffer and processes |
1543 | | // marked elements. |
1544 | | // Finds new position for each element and transfers it to the new slots. |
1545 | | // Returns the total probe length. |
1546 | | ABSL_ATTRIBUTE_NOINLINE size_t DecodeAndInsertToTableOverflow( |
1547 | | CommonFields& common, const PolicyFunctions& __restrict policy, |
1548 | 0 | void* old_slots) const { |
1549 | 0 | ABSL_SWISSTABLE_ASSERT(local_buffer_full_ && |
1550 | 0 | "must not be called when local buffer is not full"); |
1551 | 0 | size_t total_probe_length = DecodeAndInsertImpl( |
1552 | 0 | common, policy, OverflowBufferStart(), pos_, old_slots); |
1553 | 0 | if (ABSL_PREDICT_TRUE(marked_elements_starting_position_ == |
1554 | 0 | kNoMarkedElementsSentinel)) { |
1555 | 0 | return total_probe_length; |
1556 | 0 | } |
1557 | 0 | total_probe_length += |
1558 | 0 | ProcessProbedMarkedElements(common, policy, control_, old_slots, |
1559 | 0 | marked_elements_starting_position_); |
1560 | 0 | return total_probe_length; |
1561 | 0 | } Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned int, 32ul>, false>::DecodeAndInsertToTableOverflow(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, void*) const Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned long, 64ul>, false>::DecodeAndInsertToTableOverflow(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, void*) const Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned long, 122ul>, false>::DecodeAndInsertToTableOverflow(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, void*) const Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned int, 32ul>, true>::DecodeAndInsertToTableOverflow(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, void*) const |
1562 | | |
1563 | | static constexpr size_t kBufferSize = |
1564 | | kProbedElementsBufferSize / sizeof(ProbedItem); |
1565 | | ProbedItem buffer_[kBufferSize]; |
1566 | | // If local_buffer_full_ is false, then pos_/end_ are in the local buffer, |
1567 | | // otherwise, they're in the overflow buffer. |
1568 | | ProbedItem* pos_ = buffer_; |
1569 | | const void* end_ = buffer_ + kBufferSize; |
1570 | | ctrl_t* const control_; |
1571 | | size_t marked_elements_starting_position_ = kNoMarkedElementsSentinel; |
1572 | | bool local_buffer_full_ = false; |
1573 | | }; |
1574 | | |
1575 | | // Grows to next capacity with specified encoder type. |
1576 | | // Encoder is used to store probed elements that are processed later. |
1577 | | // Different encoder is used depending on the capacity of the table. |
1578 | | // Returns total probe length. |
1579 | | template <typename Encoder> |
1580 | | size_t GrowToNextCapacity(CommonFields& common, |
1581 | | const PolicyFunctions& __restrict policy, |
1582 | 107k | ctrl_t* old_ctrl, void* old_slots) { |
1583 | 107k | using ProbedItem = typename Encoder::ProbedItem; |
1584 | 107k | ABSL_SWISSTABLE_ASSERT(common.capacity() <= ProbedItem::kMaxNewCapacity); |
1585 | 107k | Encoder encoder(old_ctrl); |
1586 | 107k | policy.transfer_unprobed_elements_to_next_capacity( |
1587 | 107k | common, old_ctrl, old_slots, &encoder, |
1588 | 112k | [](void* probed_storage, h2_t h2, size_t source_offset, size_t h1) { |
1589 | 112k | auto encoder_ptr = static_cast<Encoder*>(probed_storage); |
1590 | 112k | encoder_ptr->EncodeItem(ProbedItem(h2, source_offset, h1)); |
1591 | 112k | }); raw_hash_set.cc:absl::container_internal::(anonymous namespace)::GrowToNextCapacity<absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned int, 32ul>, true> >(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, absl::container_internal::ctrl_t*, void*)::{lambda(void*, unsigned char, unsigned long, unsigned long)#1}::operator()(void*, unsigned char, unsigned long, unsigned long) constLine | Count | Source | 1588 | 88.7k | [](void* probed_storage, h2_t h2, size_t source_offset, size_t h1) { | 1589 | 88.7k | auto encoder_ptr = static_cast<Encoder*>(probed_storage); | 1590 | 88.7k | encoder_ptr->EncodeItem(ProbedItem(h2, source_offset, h1)); | 1591 | 88.7k | }); |
raw_hash_set.cc:absl::container_internal::(anonymous namespace)::GrowToNextCapacity<absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned int, 32ul>, false> >(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, absl::container_internal::ctrl_t*, void*)::{lambda(void*, unsigned char, unsigned long, unsigned long)#1}::operator()(void*, unsigned char, unsigned long, unsigned long) constLine | Count | Source | 1588 | 23.4k | [](void* probed_storage, h2_t h2, size_t source_offset, size_t h1) { | 1589 | 23.4k | auto encoder_ptr = static_cast<Encoder*>(probed_storage); | 1590 | 23.4k | encoder_ptr->EncodeItem(ProbedItem(h2, source_offset, h1)); | 1591 | 23.4k | }); |
Unexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::GrowToNextCapacity<absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned long, 64ul>, false> >(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, absl::container_internal::ctrl_t*, void*)::{lambda(void*, unsigned char, unsigned long, unsigned long)#1}::operator()(void*, unsigned char, unsigned long, unsigned long) constUnexecuted instantiation: raw_hash_set.cc:absl::container_internal::(anonymous namespace)::GrowToNextCapacity<absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned long, 122ul>, false> >(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, absl::container_internal::ctrl_t*, void*)::{lambda(void*, unsigned char, unsigned long, unsigned long)#1}::operator()(void*, unsigned char, unsigned long, unsigned long) const |
1592 | 107k | InitializeMirroredControlBytes(common.control(), common.capacity()); |
1593 | 107k | return encoder.DecodeAndInsertToTable(common, policy, old_slots); |
1594 | 107k | } raw_hash_set.cc:unsigned long absl::container_internal::(anonymous namespace)::GrowToNextCapacity<absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned int, 32ul>, true> >(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, absl::container_internal::ctrl_t*, void*) Line | Count | Source | 1582 | 90.6k | ctrl_t* old_ctrl, void* old_slots) { | 1583 | 90.6k | using ProbedItem = typename Encoder::ProbedItem; | 1584 | 90.6k | ABSL_SWISSTABLE_ASSERT(common.capacity() <= ProbedItem::kMaxNewCapacity); | 1585 | 90.6k | Encoder encoder(old_ctrl); | 1586 | 90.6k | policy.transfer_unprobed_elements_to_next_capacity( | 1587 | 90.6k | common, old_ctrl, old_slots, &encoder, | 1588 | 90.6k | [](void* probed_storage, h2_t h2, size_t source_offset, size_t h1) { | 1589 | 90.6k | auto encoder_ptr = static_cast<Encoder*>(probed_storage); | 1590 | 90.6k | encoder_ptr->EncodeItem(ProbedItem(h2, source_offset, h1)); | 1591 | 90.6k | }); | 1592 | 90.6k | InitializeMirroredControlBytes(common.control(), common.capacity()); | 1593 | 90.6k | return encoder.DecodeAndInsertToTable(common, policy, old_slots); | 1594 | 90.6k | } |
raw_hash_set.cc:unsigned long absl::container_internal::(anonymous namespace)::GrowToNextCapacity<absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned int, 32ul>, false> >(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, absl::container_internal::ctrl_t*, void*) Line | Count | Source | 1582 | 17.2k | ctrl_t* old_ctrl, void* old_slots) { | 1583 | 17.2k | using ProbedItem = typename Encoder::ProbedItem; | 1584 | 17.2k | ABSL_SWISSTABLE_ASSERT(common.capacity() <= ProbedItem::kMaxNewCapacity); | 1585 | 17.2k | Encoder encoder(old_ctrl); | 1586 | 17.2k | policy.transfer_unprobed_elements_to_next_capacity( | 1587 | 17.2k | common, old_ctrl, old_slots, &encoder, | 1588 | 17.2k | [](void* probed_storage, h2_t h2, size_t source_offset, size_t h1) { | 1589 | 17.2k | auto encoder_ptr = static_cast<Encoder*>(probed_storage); | 1590 | 17.2k | encoder_ptr->EncodeItem(ProbedItem(h2, source_offset, h1)); | 1591 | 17.2k | }); | 1592 | 17.2k | InitializeMirroredControlBytes(common.control(), common.capacity()); | 1593 | 17.2k | return encoder.DecodeAndInsertToTable(common, policy, old_slots); | 1594 | 17.2k | } |
Unexecuted instantiation: raw_hash_set.cc:unsigned long absl::container_internal::(anonymous namespace)::GrowToNextCapacity<absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned long, 64ul>, false> >(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, absl::container_internal::ctrl_t*, void*) Unexecuted instantiation: raw_hash_set.cc:unsigned long absl::container_internal::(anonymous namespace)::GrowToNextCapacity<absl::container_internal::(anonymous namespace)::ProbedItemEncoder<absl::container_internal::ProbedItemImpl<unsigned long, 122ul>, false> >(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, absl::container_internal::ctrl_t*, void*) |
1595 | | |
1596 | | // Grows to next capacity for relatively small tables so that even if all |
1597 | | // elements are probed, we don't need to overflow the local buffer. |
1598 | | // Returns total probe length. |
1599 | | size_t GrowToNextCapacityThatFitsInLocalBuffer( |
1600 | | CommonFields& common, const PolicyFunctions& __restrict policy, |
1601 | 90.6k | ctrl_t* old_ctrl, void* old_slots) { |
1602 | 90.6k | ABSL_SWISSTABLE_ASSERT(common.capacity() <= kMaxLocalBufferNewCapacity); |
1603 | 90.6k | return GrowToNextCapacity< |
1604 | 90.6k | ProbedItemEncoder<ProbedItem4Bytes, /*kGuaranteedFitToBuffer=*/true>>( |
1605 | 90.6k | common, policy, old_ctrl, old_slots); |
1606 | 90.6k | } |
1607 | | |
1608 | | // Grows to next capacity with different encodings. Returns total probe length. |
1609 | | // These functions are useful to simplify profile analysis. |
1610 | | size_t GrowToNextCapacity4BytesEncoder(CommonFields& common, |
1611 | | const PolicyFunctions& __restrict policy, |
1612 | 17.2k | ctrl_t* old_ctrl, void* old_slots) { |
1613 | 17.2k | return GrowToNextCapacity<ProbedItemEncoder<ProbedItem4Bytes>>( |
1614 | 17.2k | common, policy, old_ctrl, old_slots); |
1615 | 17.2k | } |
1616 | | size_t GrowToNextCapacity8BytesEncoder(CommonFields& common, |
1617 | | const PolicyFunctions& __restrict policy, |
1618 | 0 | ctrl_t* old_ctrl, void* old_slots) { |
1619 | 0 | return GrowToNextCapacity<ProbedItemEncoder<ProbedItem8Bytes>>( |
1620 | 0 | common, policy, old_ctrl, old_slots); |
1621 | 0 | } |
1622 | | size_t GrowToNextCapacity16BytesEncoder( |
1623 | | CommonFields& common, const PolicyFunctions& __restrict policy, |
1624 | 0 | ctrl_t* old_ctrl, void* old_slots) { |
1625 | 0 | return GrowToNextCapacity<ProbedItemEncoder<ProbedItem16Bytes>>( |
1626 | 0 | common, policy, old_ctrl, old_slots); |
1627 | 0 | } |
1628 | | |
1629 | | // Grows to next capacity for tables with relatively large capacity so that we |
1630 | | // can't guarantee that all probed elements fit in the local buffer. Returns |
1631 | | // total probe length. |
1632 | | size_t GrowToNextCapacityOverflowLocalBuffer( |
1633 | | CommonFields& common, const PolicyFunctions& __restrict policy, |
1634 | 17.2k | ctrl_t* old_ctrl, void* old_slots) { |
1635 | 17.2k | const size_t new_capacity = common.capacity(); |
1636 | 17.2k | if (ABSL_PREDICT_TRUE(new_capacity <= ProbedItem4Bytes::kMaxNewCapacity)) { |
1637 | 17.2k | return GrowToNextCapacity4BytesEncoder(common, policy, old_ctrl, old_slots); |
1638 | 17.2k | } |
1639 | 0 | if (ABSL_PREDICT_TRUE(new_capacity <= ProbedItem8Bytes::kMaxNewCapacity)) { |
1640 | 0 | return GrowToNextCapacity8BytesEncoder(common, policy, old_ctrl, old_slots); |
1641 | 0 | } |
1642 | | // 16 bytes encoding supports the maximum swisstable capacity. |
1643 | 0 | return GrowToNextCapacity16BytesEncoder(common, policy, old_ctrl, old_slots); |
1644 | 0 | } |
1645 | | |
1646 | | // Dispatches to the appropriate `GrowToNextCapacity*` function based on the |
1647 | | // capacity of the table. Returns total probe length. |
1648 | | ABSL_ATTRIBUTE_NOINLINE |
1649 | | size_t GrowToNextCapacityDispatch(CommonFields& common, |
1650 | | const PolicyFunctions& __restrict policy, |
1651 | 107k | ctrl_t* old_ctrl, void* old_slots) { |
1652 | 107k | const size_t new_capacity = common.capacity(); |
1653 | 107k | if (ABSL_PREDICT_TRUE(new_capacity <= kMaxLocalBufferNewCapacity)) { |
1654 | 90.6k | return GrowToNextCapacityThatFitsInLocalBuffer(common, policy, old_ctrl, |
1655 | 90.6k | old_slots); |
1656 | 90.6k | } else { |
1657 | 17.2k | return GrowToNextCapacityOverflowLocalBuffer(common, policy, old_ctrl, |
1658 | 17.2k | old_slots); |
1659 | 17.2k | } |
1660 | 107k | } |
1661 | | |
1662 | | void IncrementSmallSizeNonSoo(CommonFields& common, |
1663 | 135k | const PolicyFunctions& __restrict policy) { |
1664 | 135k | ABSL_SWISSTABLE_ASSERT(common.is_small()); |
1665 | 135k | common.increment_size(); |
1666 | 135k | SanitizerUnpoisonMemoryRegion( |
1667 | 135k | SingleSlotAddress</*kSooEnabled=*/false>(common), policy.slot_size); |
1668 | 135k | } |
1669 | | |
1670 | | void IncrementSmallSize(CommonFields& common, |
1671 | 0 | const PolicyFunctions& __restrict policy) { |
1672 | 0 | ABSL_SWISSTABLE_ASSERT(common.is_small()); |
1673 | 0 | if (policy.soo_enabled) { |
1674 | 0 | common.set_full_soo(); |
1675 | 0 | } else { |
1676 | 0 | IncrementSmallSizeNonSoo(common, policy); |
1677 | 0 | } |
1678 | 0 | } |
1679 | | |
1680 | | void* Grow1To3AndPrepareInsert(CommonFields& common, |
1681 | | const PolicyFunctions& __restrict policy, |
1682 | 31.3k | absl::FunctionRef<size_t(size_t)> get_hash) { |
1683 | | // TODO(b/413062340): Refactor to reuse more code with |
1684 | | // GrowSooTableToNextCapacityAndPrepareInsert. |
1685 | 31.3k | ABSL_SWISSTABLE_ASSERT(common.capacity() == 1); |
1686 | 31.3k | ABSL_SWISSTABLE_ASSERT(!common.empty()); |
1687 | 31.3k | ABSL_SWISSTABLE_ASSERT(!policy.soo_enabled); |
1688 | | // 1-element tables can't have any blocked elements. |
1689 | 31.3k | ABSL_SWISSTABLE_ASSERT(common.blocked_element_count() == 0); |
1690 | 31.3k | constexpr size_t kOldCapacity = 1; |
1691 | 31.3k | constexpr size_t kNewCapacity = NextCapacity(kOldCapacity); |
1692 | 31.3k | void* old_slots = common.slot_array(kOldCapacity); |
1693 | | |
1694 | 31.3k | const size_t slot_size = policy.slot_size; |
1695 | 31.3k | const size_t slot_align = policy.slot_align; |
1696 | 31.3k | void* alloc = policy.get_char_alloc(common); |
1697 | 31.3k | HashtablezInfoHandle infoz = common.infoz(); |
1698 | 31.3k | const bool has_infoz = infoz.IsSampled(); |
1699 | 31.3k | common.set_capacity(kNewCapacity); |
1700 | | |
1701 | 31.3k | const auto [new_ctrl, new_slots] = |
1702 | 31.3k | AllocBackingArray(common, policy, kNewCapacity, has_infoz, alloc, |
1703 | 31.3k | /*blocked_element_count=*/0); |
1704 | 31.3k | common.set_control(new_ctrl); |
1705 | 31.3k | SanitizerPoisonMemoryRegion(new_slots, kNewCapacity * slot_size); |
1706 | | |
1707 | 31.3k | if (ABSL_PREDICT_TRUE(!has_infoz)) { |
1708 | | // When we're sampled, we already have a seed. |
1709 | 31.3k | common.generate_new_seed(/*has_infoz=*/false); |
1710 | 31.3k | } |
1711 | 31.3k | const size_t new_hash = get_hash(common.seed().seed()); |
1712 | 31.3k | h2_t new_h2 = H2(new_hash); |
1713 | 31.3k | size_t orig_hash = |
1714 | 31.3k | policy.hash_slot(policy.hash_fn(common), old_slots, common.seed().seed()); |
1715 | 31.3k | size_t offset = Resize1To3NewOffset(new_hash); |
1716 | 31.3k | InitializeThreeElementsControlBytes(H2(orig_hash), new_h2, offset, new_ctrl); |
1717 | | |
1718 | 31.3k | void* old_element_target = NextSlot(new_slots, slot_size); |
1719 | 31.3k | SanitizerUnpoisonMemoryRegion(old_element_target, slot_size); |
1720 | 31.3k | policy.transfer_n(&common, old_element_target, old_slots, 1); |
1721 | | |
1722 | 31.3k | void* new_element_target_slot = SlotAddress(new_slots, offset, slot_size); |
1723 | 31.3k | SanitizerUnpoisonMemoryRegion(new_element_target_slot, slot_size); |
1724 | | |
1725 | 31.3k | DeallocBackingArrayImpl(alloc, kOldCapacity, |
1726 | | // old_slots == old_ctrl in case of capacity == 1. |
1727 | 31.3k | static_cast<ctrl_t*>(old_slots), slot_size, |
1728 | 31.3k | slot_align, has_infoz, |
1729 | 31.3k | /*blocked_element_count=*/0, policy.dealloc); |
1730 | 31.3k | PrepareInsertCommon(common); |
1731 | 31.3k | ABSL_SWISSTABLE_ASSERT(common.size() == 2); |
1732 | 31.3k | common.InitGrowthLeftNoDeleted(kNewCapacity - 2, kNewCapacity); |
1733 | | |
1734 | 31.3k | if (ABSL_PREDICT_FALSE(has_infoz)) { |
1735 | 0 | ReportSingleGroupTableGrowthToInfoz(common, infoz, new_hash); |
1736 | 0 | } |
1737 | 31.3k | return new_element_target_slot; |
1738 | 31.3k | } |
1739 | | |
1740 | | // Grows to next capacity and prepares insert for the given new_hash. |
1741 | | // Returns the offset of the new element. |
1742 | | void* GrowToNextCapacityAndPrepareInsert( |
1743 | | CommonFields& common, const PolicyFunctions& __restrict policy, |
1744 | 177k | size_t new_hash) { |
1745 | 177k | const size_t old_capacity = common.capacity(); |
1746 | 177k | ABSL_SWISSTABLE_ASSERT(common.GetGrowthLeftTotalSlow(old_capacity) == 0); |
1747 | 177k | ABSL_SWISSTABLE_ASSERT(old_capacity > policy.soo_capacity()); |
1748 | 177k | ABSL_SWISSTABLE_ASSERT(!IsSmallCapacity(old_capacity)); |
1749 | 177k | ABSL_ASSUME(old_capacity > kMaxSmallCapacity); |
1750 | | |
1751 | 177k | const size_t new_capacity = NextCapacity(old_capacity); |
1752 | 177k | ctrl_t* old_ctrl = common.control(); |
1753 | 177k | void* old_slots = common.slot_array(old_capacity); |
1754 | 177k | size_t old_blocked_element_count = common.blocked_element_count(); |
1755 | | |
1756 | 177k | HashtablezInfoHandle infoz = common.infoz(); |
1757 | 177k | const bool has_infoz = infoz.IsSampled(); |
1758 | 177k | common.set_capacity(new_capacity); |
1759 | 177k | common.set_blocked_element_count_to_zero(); |
1760 | 177k | const size_t slot_size = policy.slot_size; |
1761 | 177k | const size_t slot_align = policy.slot_align; |
1762 | 177k | void* alloc = policy.get_char_alloc(common); |
1763 | | |
1764 | 177k | const auto [new_ctrl, new_slots] = |
1765 | 177k | AllocBackingArray(common, policy, new_capacity, has_infoz, alloc, |
1766 | 177k | /*blocked_element_count=*/0); |
1767 | 177k | common.set_control(new_ctrl); |
1768 | 177k | SanitizerPoisonMemoryRegion(new_slots, new_capacity * slot_size); |
1769 | | |
1770 | 177k | h2_t new_h2 = H2(new_hash); |
1771 | 177k | size_t total_probe_length = 0; |
1772 | 177k | FindInfo find_info; |
1773 | 177k | if (ABSL_PREDICT_TRUE(is_single_group(new_capacity))) { |
1774 | 69.2k | size_t offset; |
1775 | 69.2k | const size_t old_size = common.size(); |
1776 | 69.2k | GrowIntoSingleGroupShuffleControlBytes(old_ctrl, old_capacity, |
1777 | 69.2k | old_blocked_element_count, new_ctrl, |
1778 | 69.2k | new_capacity); |
1779 | | // We put the new element either at the beginning or at the end of the |
1780 | | // table with approximately equal probability. |
1781 | 69.2k | offset = new_hash & 1 ? 0 : new_capacity - 1; |
1782 | | |
1783 | 69.2k | ABSL_SWISSTABLE_ASSERT(IsEmpty(new_ctrl[offset])); |
1784 | 69.2k | SetCtrlInSingleGroupTable(common, offset, new_h2, policy.slot_size); |
1785 | 69.2k | find_info = FindInfo{offset, 0}; |
1786 | | // Single group tables have all slots full on resize. So we can transfer |
1787 | | // all slots without checking the control bytes. |
1788 | 69.2k | ABSL_SWISSTABLE_ASSERT(common.size() + old_blocked_element_count == |
1789 | 69.2k | old_capacity); |
1790 | 69.2k | void* target = NextSlot(new_slots, slot_size); |
1791 | 69.2k | SanitizerUnpoisonMemoryRegion(target, old_size * slot_size); |
1792 | 69.2k | policy.transfer_n(&common, target, old_slots, old_size); |
1793 | 107k | } else { |
1794 | 107k | total_probe_length = |
1795 | 107k | GrowToNextCapacityDispatch(common, policy, old_ctrl, old_slots); |
1796 | 107k | find_info = find_first_non_full(common, new_hash); |
1797 | 107k | SetCtrlInLargeTable(common, find_info.offset, new_h2, policy.slot_size); |
1798 | 107k | } |
1799 | 177k | ABSL_SWISSTABLE_ASSERT(old_capacity > policy.soo_capacity()); |
1800 | 177k | DeallocBackingArrayImpl(alloc, old_capacity, old_ctrl, slot_size, slot_align, |
1801 | 177k | has_infoz, old_blocked_element_count, policy.dealloc); |
1802 | 177k | PrepareInsertCommon(common); |
1803 | 177k | ResetGrowthLeft(new_capacity, common.size(), common); |
1804 | | |
1805 | 177k | if (ABSL_PREDICT_FALSE(has_infoz)) { |
1806 | 0 | ReportGrowthToInfoz(common, infoz, new_hash, total_probe_length, |
1807 | 0 | find_info.probe_length); |
1808 | 0 | } |
1809 | 177k | return SlotAddress(new_slots, find_info.offset, policy.slot_size); |
1810 | 177k | } |
1811 | | |
1812 | | } // namespace |
1813 | | |
1814 | | void* PrepareInsertSmallNonSoo(CommonFields& common, |
1815 | | const PolicyFunctions& __restrict policy, |
1816 | 200k | absl::FunctionRef<size_t(size_t)> get_hash) { |
1817 | 200k | ABSL_SWISSTABLE_ASSERT(common.is_small()); |
1818 | 200k | ABSL_SWISSTABLE_ASSERT(!policy.soo_enabled); |
1819 | 200k | if (common.capacity() == 1) { |
1820 | 166k | if (common.empty()) { |
1821 | 135k | IncrementSmallSizeNonSoo(common, policy); |
1822 | 135k | if (common.has_infoz()) { |
1823 | 0 | common.infoz().RecordInsertMiss(get_hash(common.seed().seed()), |
1824 | 0 | /*distance_from_desired=*/0); |
1825 | 0 | } |
1826 | 135k | return common.slot_array(/*capacity=*/1); |
1827 | 135k | } else { |
1828 | 31.3k | return Grow1To3AndPrepareInsert(common, policy, get_hash); |
1829 | 31.3k | } |
1830 | 166k | } |
1831 | | |
1832 | | // Growing from 0 to 1 capacity. |
1833 | 34.0k | ABSL_SWISSTABLE_ASSERT(common.capacity() == 0); |
1834 | 34.0k | constexpr size_t kNewCapacity = 1; |
1835 | | |
1836 | 34.0k | common.set_capacity(kNewCapacity); |
1837 | 34.0k | HashtablezInfoHandle infoz; |
1838 | 34.0k | const bool should_sample = |
1839 | 34.0k | policy.is_hashtablez_eligible && ShouldSampleNextTable(); |
1840 | 34.0k | if (ABSL_PREDICT_FALSE(should_sample)) { |
1841 | 0 | infoz = ForcedTrySample(policy.slot_size, policy.key_size, |
1842 | 0 | policy.value_size, policy.soo_capacity()); |
1843 | 0 | } |
1844 | 34.0k | const bool has_infoz = infoz.IsSampled(); |
1845 | 34.0k | void* alloc = policy.get_char_alloc(common); |
1846 | | |
1847 | 34.0k | const auto [new_ctrl, new_slots] = |
1848 | 34.0k | AllocBackingArray(common, policy, kNewCapacity, has_infoz, alloc, |
1849 | 34.0k | /*blocked_element_count=*/0); |
1850 | 34.0k | common.set_control(new_ctrl); |
1851 | | |
1852 | 34.0k | static_assert(NextCapacity(0) == 1); |
1853 | 34.0k | PrepareInsertCommon(common); |
1854 | | |
1855 | 34.0k | if (ABSL_PREDICT_FALSE(has_infoz)) { |
1856 | 0 | common.generate_new_seed(/*has_infoz=*/true); |
1857 | 0 | ReportSingleGroupTableGrowthToInfoz(common, infoz, |
1858 | 0 | get_hash(common.seed().seed())); |
1859 | 0 | } |
1860 | 34.0k | return new_slots; |
1861 | 34.0k | } |
1862 | | |
1863 | | namespace { |
1864 | | |
1865 | | // Called whenever the table needs to vacate empty slots either by removing |
1866 | | // tombstones via rehash or growth to next capacity. |
1867 | | ABSL_ATTRIBUTE_NOINLINE |
1868 | | void* RehashOrGrowToNextCapacityAndPrepareInsert( |
1869 | | CommonFields& common, const PolicyFunctions& __restrict policy, |
1870 | 0 | size_t new_hash) { |
1871 | 0 | ABSL_SWISSTABLE_ASSERT(!common.GetGrowthInfoLowerBound().HasNoDeleted()); |
1872 | 0 | const size_t cap = common.capacity(); |
1873 | 0 | ABSL_ASSUME(cap > 0); |
1874 | | // Do these calculations in 64-bit to avoid overflow. |
1875 | 0 | if (common.size() * uint64_t{32} <= |
1876 | 0 | (cap - kMaxBlockedElementsForLargeTables) * uint64_t{25}) { |
1877 | | // Squash DELETED without growing if there is enough capacity. |
1878 | | // |
1879 | | // Rehash in place if the current size is <= 25/32 of capacity. |
1880 | | // Rationale for such a high factor: 1) DropDeletesWithoutResize() is |
1881 | | // faster than resize, and 2) it takes quite a bit of work to add |
1882 | | // tombstones. In the worst case, seems to take approximately 4 |
1883 | | // insert/erase pairs to create a single tombstone and so if we are |
1884 | | // rehashing because of tombstones, we can afford to rehash-in-place as |
1885 | | // long as we are reclaiming at least 1/8 the capacity without doing more |
1886 | | // than 2X the work. (Where "work" is defined to be size() for rehashing |
1887 | | // or rehashing in place, and 1 for an insert or erase.) But rehashing in |
1888 | | // place is faster per operation than inserting or even doubling the size |
1889 | | // of the table, so we actually afford to reclaim even less space from a |
1890 | | // resize-in-place. The decision is to rehash in place if we can reclaim |
1891 | | // at about 1/8th of the usable capacity (specifically 3/28 of the |
1892 | | // capacity) which means that the total cost of rehashing will be a small |
1893 | | // fraction of the total work. |
1894 | | // |
1895 | | // Here is output of an experiment using the BM_CacheInSteadyState |
1896 | | // benchmark running the old case (where we rehash-in-place only if we can |
1897 | | // reclaim at least 7/16*capacity) vs. this code (which rehashes in place |
1898 | | // if we can recover 3/32*capacity). |
1899 | | // |
1900 | | // Note that although in the worst-case number of rehashes jumped up from |
1901 | | // 15 to 190, but the number of operations per second is almost the same. |
1902 | | // |
1903 | | // Abridged output of running BM_CacheInSteadyState benchmark from |
1904 | | // raw_hash_set_benchmark. N is the number of insert/erase operations. |
1905 | | // |
1906 | | // | OLD (recover >= 7/16 | NEW (recover >= 3/32) |
1907 | | // size | N/s LoadFactor NRehashes | N/s LoadFactor NRehashes |
1908 | | // 448 | 145284 0.44 18 | 140118 0.44 19 |
1909 | | // 493 | 152546 0.24 11 | 151417 0.48 28 |
1910 | | // 538 | 151439 0.26 11 | 151152 0.53 38 |
1911 | | // 583 | 151765 0.28 11 | 150572 0.57 50 |
1912 | | // 628 | 150241 0.31 11 | 150853 0.61 66 |
1913 | | // 672 | 149602 0.33 12 | 150110 0.66 90 |
1914 | | // 717 | 149998 0.35 12 | 149531 0.70 129 |
1915 | | // 762 | 149836 0.37 13 | 148559 0.74 190 |
1916 | | // 807 | 149736 0.39 14 | 151107 0.39 14 |
1917 | | // 852 | 150204 0.42 15 | 151019 0.42 15 |
1918 | 0 | return DropDeletesWithoutResizeAndPrepareInsert(common, policy, new_hash); |
1919 | 0 | } else { |
1920 | | // Otherwise grow the container. |
1921 | 0 | return GrowToNextCapacityAndPrepareInsert(common, policy, new_hash); |
1922 | 0 | } |
1923 | 0 | } |
1924 | | |
1925 | | // Slow path for PrepareInsertLarge that is called when the table has deleted |
1926 | | // slots or need to be resized or rehashed. |
1927 | | ABSL_ATTRIBUTE_NOINLINE |
1928 | | void* PrepareInsertLargeSlow(CommonFields& common, |
1929 | | const PolicyFunctions& __restrict policy, |
1930 | 183k | size_t hash) { |
1931 | 183k | const size_t cap = common.capacity(); |
1932 | 183k | ABSL_ASSUME(cap > kMaxSmallCapacity); |
1933 | 183k | GrowthInfoLowerBound growth_info_lower_bound = |
1934 | 183k | common.RebalanceGrowthLeftLowerBound(cap); |
1935 | 183k | if (ABSL_PREDICT_TRUE( |
1936 | 183k | growth_info_lower_bound.HasNoGrowthLeftAndNoDeleted())) { |
1937 | | // Table without deleted slots (>95% cases) that needs to be resized. |
1938 | 177k | return GrowToNextCapacityAndPrepareInsert(common, policy, hash); |
1939 | 177k | } |
1940 | 6.00k | if (ABSL_PREDICT_FALSE( |
1941 | 6.00k | growth_info_lower_bound.HasNoGrowthLeftAndHaveDeleted())) { |
1942 | | // Table with deleted slots that needs to be rehashed or resized. |
1943 | 0 | return RehashOrGrowToNextCapacityAndPrepareInsert(common, policy, hash); |
1944 | 0 | } |
1945 | | // Covers two cases: |
1946 | | // 1. Table with deleted slots that has space for the inserting element. |
1947 | | // 2. Table without deleted slots that has space and GrowthInfoView was |
1948 | | // rebalanced. |
1949 | 6.00k | FindInfo target = find_first_non_full(common, hash); |
1950 | 6.00k | PrepareInsertCommon(common); |
1951 | 6.00k | common.OverwriteControlAsFull(common.control()[target.offset]); |
1952 | 6.00k | SetCtrlInLargeTable(common, target.offset, H2(hash), policy.slot_size); |
1953 | 6.00k | common.infoz().RecordInsertMiss(hash, target.probe_length); |
1954 | 6.00k | return SlotAddress(common.slot_array(cap), target.offset, policy.slot_size); |
1955 | 6.00k | } |
1956 | | |
1957 | | // Resizes empty non-allocated SOO table to NextCapacity(SooCapacity()), |
1958 | | // forces the table to be sampled and prepares the insert. |
1959 | | // SOO tables need to switch from SOO to heap in order to store the infoz. |
1960 | | // Requires: |
1961 | | // 1. `c.capacity() == SooCapacity()`. |
1962 | | // 2. `c.empty()`. |
1963 | | ABSL_ATTRIBUTE_NOINLINE void* |
1964 | | GrowEmptySooTableToNextCapacityForceSamplingAndPrepareInsert( |
1965 | | CommonFields& common, const PolicyFunctions& __restrict policy, |
1966 | 0 | absl::FunctionRef<size_t(size_t)> get_hash) { |
1967 | 0 | const size_t kNewCapacity = NextCapacity(SooCapacity()); |
1968 | 0 | ResizeEmptyNonAllocatedTableImpl(common, policy, kNewCapacity, |
1969 | 0 | /*blocked_element_count=*/0, |
1970 | 0 | /*force_infoz=*/true); |
1971 | 0 | PrepareInsertCommon(common); |
1972 | 0 | common.OverwriteEmptyAsFull(); |
1973 | 0 | const size_t new_hash = get_hash(common.seed().seed()); |
1974 | 0 | SetCtrlInSingleGroupTable(common, SooSlotIndex(), H2(new_hash), |
1975 | 0 | policy.slot_size); |
1976 | 0 | common.infoz().RecordInsertMiss(new_hash, /*distance_from_desired=*/0); |
1977 | 0 | return SlotAddress(common.slot_array(kNewCapacity), SooSlotIndex(), |
1978 | 0 | policy.slot_size); |
1979 | 0 | } |
1980 | | |
1981 | | // Returns the number of elements to block for the given capacity and reserved |
1982 | | // size. |
1983 | | size_t BlockedElementCountForReservedTable(size_t capacity, |
1984 | 0 | size_t reserved_size) { |
1985 | 0 | if (!IsCapacityValidForBlockedElements(capacity)) { |
1986 | 0 | return 0; |
1987 | 0 | } |
1988 | 0 | const size_t blocked_elements = CapacityToGrowth(capacity) - reserved_size; |
1989 | 0 | if (is_single_group(capacity)) { |
1990 | | // Single group tables never probes, so we can block all the slots. |
1991 | 0 | return blocked_elements; |
1992 | 0 | } |
1993 | 0 | return (std::min)(blocked_elements, kMaxBlockedElementsForLargeTables); |
1994 | 0 | } |
1995 | | |
1996 | | // Resizes empty non-allocated table to the capacity to fit new_size elements. |
1997 | | // Requires: |
1998 | | // 1. `c.capacity() == policy.soo_capacity()`. |
1999 | | // 2. `c.empty()`. |
2000 | | // 3. `new_size > policy.soo_capacity()`. |
2001 | | // The table will be attempted to be sampled. |
2002 | | void ReserveEmptyNonAllocatedTableToFitNewSize( |
2003 | | CommonFields& common, const PolicyFunctions& __restrict policy, |
2004 | 0 | size_t new_size) { |
2005 | 0 | ValidateMaxSize(new_size, policy.key_size, policy.slot_size); |
2006 | 0 | ABSL_ASSUME(new_size > 0); |
2007 | 0 | const size_t new_capacity = SizeToCapacity(new_size); |
2008 | 0 | ResizeEmptyNonAllocatedTableImpl( |
2009 | 0 | common, policy, new_capacity, |
2010 | 0 | BlockedElementCountForReservedTable(new_capacity, new_size), |
2011 | 0 | /*force_infoz=*/false); |
2012 | | // This is after resize, to ensure that we have completed the allocation |
2013 | | // and have potentially sampled the hashtable. |
2014 | 0 | common.infoz().RecordReservation(new_size); |
2015 | 0 | } |
2016 | | |
2017 | | // Type erased version of raw_hash_set::reserve for tables that have an |
2018 | | // allocated backing array. |
2019 | | // |
2020 | | // Requires: |
2021 | | // 1. `c.capacity() > policy.soo_capacity()` OR `!c.empty()`. |
2022 | | // Reserving already allocated tables is considered to be a rare case. |
2023 | | ABSL_ATTRIBUTE_NOINLINE void ReserveAllocatedTable( |
2024 | | CommonFields& common, const PolicyFunctions& __restrict policy, |
2025 | 0 | size_t new_size) { |
2026 | 0 | const size_t cap = common.capacity(); |
2027 | 0 | ValidateMaxSize(new_size, policy.key_size, policy.slot_size); |
2028 | 0 | ABSL_ASSUME(new_size > 0); |
2029 | 0 | const size_t new_capacity = SizeToCapacity(new_size); |
2030 | 0 | if (cap == policy.soo_capacity()) { |
2031 | 0 | ABSL_SWISSTABLE_ASSERT(!common.empty()); |
2032 | 0 | ResizeFullSooTable(common, policy, new_capacity, |
2033 | 0 | ResizeFullSooTableSamplingMode::kNoSampling); |
2034 | 0 | } else { |
2035 | 0 | ABSL_SWISSTABLE_ASSERT(cap > policy.soo_capacity()); |
2036 | | // TODO(b/382423690): consider using GrowToNextCapacity, when applicable. |
2037 | 0 | ResizeAllocatedTableWithSeedChange(common, policy, new_capacity); |
2038 | 0 | } |
2039 | 0 | common.infoz().RecordReservation(new_size); |
2040 | 0 | } |
2041 | | |
2042 | | // As `ResizeFullSooTableToNextCapacity`, except that we also force the SOO |
2043 | | // table to be sampled. SOO tables need to switch from SOO to heap in order to |
2044 | | // store the infoz. No-op if sampling is disabled or not possible. |
2045 | | void GrowFullSooTableToNextCapacityForceSampling( |
2046 | 0 | CommonFields& common, const PolicyFunctions& __restrict policy) { |
2047 | 0 | AssertFullSoo(common, policy); |
2048 | 0 | ResizeFullSooTable( |
2049 | 0 | common, policy, NextCapacity(SooCapacity()), |
2050 | 0 | ResizeFullSooTableSamplingMode::kForceSampleNoResizeIfUnsampled); |
2051 | 0 | } |
2052 | | |
2053 | | } // namespace |
2054 | | |
2055 | 324k | void* GetRefForEmptyClass(CommonFields& common) { |
2056 | | // Empty base optimization typically make the empty base class address to be |
2057 | | // the same as the first address of the derived class object. |
2058 | | // But we generally assume that for empty classes we can return any valid |
2059 | | // pointer. |
2060 | 324k | return &common; |
2061 | 324k | } |
2062 | | |
2063 | | void ResizeAllocatedTableWithSeedChange( |
2064 | | CommonFields& common, const PolicyFunctions& __restrict policy, |
2065 | 0 | size_t new_capacity) { |
2066 | 0 | ABSL_SWISSTABLE_ASSERT(IsValidCapacity(new_capacity)); |
2067 | 0 | ABSL_SWISSTABLE_ASSERT(new_capacity > policy.soo_capacity()); |
2068 | | |
2069 | 0 | const size_t old_capacity = common.capacity(); |
2070 | 0 | ctrl_t* const old_ctrl = common.control(); |
2071 | 0 | void* const old_slots = common.slot_array(old_capacity); |
2072 | 0 | const size_t old_blocked_element_count = common.blocked_element_count(); |
2073 | |
|
2074 | 0 | const size_t slot_size = policy.slot_size; |
2075 | 0 | const size_t slot_align = policy.slot_align; |
2076 | 0 | HashtablezInfoHandle infoz = common.infoz(); |
2077 | 0 | const bool has_infoz = infoz.IsSampled(); |
2078 | 0 | void* alloc = policy.get_char_alloc(common); |
2079 | |
|
2080 | 0 | common.set_capacity(new_capacity); |
2081 | 0 | common.set_blocked_element_count_to_zero(); |
2082 | 0 | const auto [new_ctrl, new_slots] = |
2083 | 0 | AllocBackingArray(common, policy, new_capacity, has_infoz, alloc, |
2084 | 0 | /*blocked_element_count=*/0); |
2085 | 0 | common.set_control(new_ctrl); |
2086 | 0 | common.generate_new_seed(has_infoz); |
2087 | |
|
2088 | 0 | size_t total_probe_length = 0; |
2089 | 0 | ResetCtrl(common, slot_size, /*blocked_element_count=*/0); |
2090 | 0 | ABSL_SWISSTABLE_ASSERT(old_capacity > 0); |
2091 | 0 | total_probe_length = FindNewPositionsAndTransferSlots( |
2092 | 0 | common, policy, old_ctrl, old_slots, old_capacity); |
2093 | 0 | DeallocBackingArrayImpl(alloc, old_capacity, old_ctrl, slot_size, slot_align, |
2094 | 0 | has_infoz, old_blocked_element_count, policy.dealloc); |
2095 | 0 | ResetGrowthLeft(new_capacity, common.size(), common); |
2096 | |
|
2097 | 0 | if (ABSL_PREDICT_FALSE(has_infoz)) { |
2098 | 0 | ReportResizeToInfoz(common, infoz, total_probe_length); |
2099 | 0 | } |
2100 | 0 | } |
2101 | | |
2102 | | // Resizes a full SOO table to the NextCapacity(SooCapacity()). |
2103 | | template <size_t SooSlotMemcpySize, bool TransferUsesMemcpy> |
2104 | | void* GrowSooTableToNextCapacityAndPrepareInsert( |
2105 | | CommonFields& common, const PolicyFunctions& __restrict policy, |
2106 | 25.2k | absl::FunctionRef<size_t(size_t)> get_hash, bool force_sampling) { |
2107 | 25.2k | AssertSoo(common, policy); |
2108 | 25.2k | if (ABSL_PREDICT_FALSE(force_sampling)) { |
2109 | | // The table is empty, it is only used for forced sampling of SOO tables. |
2110 | 0 | return GrowEmptySooTableToNextCapacityForceSamplingAndPrepareInsert( |
2111 | 0 | common, policy, get_hash); |
2112 | 0 | } |
2113 | 25.2k | ABSL_SWISSTABLE_ASSERT(common.size() == policy.soo_capacity()); |
2114 | 25.2k | static constexpr size_t kNewCapacity = NextCapacity(SooCapacity()); |
2115 | 25.2k | const size_t slot_size = policy.slot_size; |
2116 | 25.2k | void* alloc = policy.get_char_alloc(common); |
2117 | 25.2k | common.set_capacity(kNewCapacity); |
2118 | | |
2119 | | // Since the table is not empty, it will not be sampled. |
2120 | | // The decision to sample was already made during the first insertion. |
2121 | | // |
2122 | | // We do not set control and slots in CommonFields yet to avoid overriding |
2123 | | // SOO data. |
2124 | 25.2k | const auto [new_ctrl, new_slots] = AllocBackingArray( |
2125 | 25.2k | common, policy, kNewCapacity, /*has_infoz=*/false, alloc, |
2126 | 25.2k | /*blocked_element_count=*/0); |
2127 | | |
2128 | 25.2k | PrepareInsertCommon(common); |
2129 | 25.2k | ABSL_SWISSTABLE_ASSERT(common.size() == 2); |
2130 | 25.2k | common.InitGrowthLeftNoDeleted(kNewCapacity - 2, kNewCapacity); |
2131 | 25.2k | common.generate_new_seed(/*has_infoz=*/false); |
2132 | 25.2k | const h2_t soo_slot_h2 = H2(policy.hash_slot( |
2133 | 25.2k | policy.hash_fn(common), common.soo_data(), common.seed().seed())); |
2134 | 25.2k | const size_t new_hash = get_hash(common.seed().seed()); |
2135 | | |
2136 | 25.2k | const size_t offset = Resize1To3NewOffset(new_hash); |
2137 | 25.2k | InitializeThreeElementsControlBytes(soo_slot_h2, H2(new_hash), offset, |
2138 | 25.2k | new_ctrl); |
2139 | | |
2140 | 25.2k | SanitizerPoisonMemoryRegion(new_slots, slot_size * kNewCapacity); |
2141 | 25.2k | void* target_slot = SlotAddress(new_slots, SooSlotIndex(), slot_size); |
2142 | 25.2k | SanitizerUnpoisonMemoryRegion(target_slot, slot_size); |
2143 | 25.2k | if constexpr (TransferUsesMemcpy) { |
2144 | | // Target slot is placed at index 1, but capacity is at |
2145 | | // minimum 3. So we are allowed to copy at least twice as much |
2146 | | // memory. |
2147 | 25.2k | static_assert(SooSlotIndex() == 1); |
2148 | 25.2k | static_assert(SooSlotMemcpySize > 0); |
2149 | 25.2k | static_assert(SooSlotMemcpySize <= MaxSooSlotSize()); |
2150 | 25.2k | ABSL_SWISSTABLE_ASSERT(SooSlotMemcpySize <= 2 * slot_size); |
2151 | 25.2k | ABSL_SWISSTABLE_ASSERT(SooSlotMemcpySize >= slot_size); |
2152 | 25.2k | void* next_slot = SlotAddress(target_slot, 1, slot_size); |
2153 | 25.2k | SanitizerUnpoisonMemoryRegion(next_slot, SooSlotMemcpySize - slot_size); |
2154 | 25.2k | std::memcpy(target_slot, common.soo_data(), SooSlotMemcpySize); |
2155 | 25.2k | SanitizerPoisonMemoryRegion(next_slot, SooSlotMemcpySize - slot_size); |
2156 | 25.2k | } else { |
2157 | 0 | static_assert(SooSlotMemcpySize == 0); |
2158 | 0 | policy.transfer_n(&common, target_slot, common.soo_data(), 1); |
2159 | 0 | } |
2160 | 0 | common.set_control(new_ctrl); |
2161 | | |
2162 | | // Full SOO table couldn't be sampled. If SOO table is sampled, it would |
2163 | | // have been resized to the next capacity. |
2164 | 25.2k | ABSL_SWISSTABLE_ASSERT(!common.infoz().IsSampled()); |
2165 | 25.2k | void* new_slot = SlotAddress(new_slots, offset, slot_size); |
2166 | 25.2k | SanitizerUnpoisonMemoryRegion(new_slot, slot_size); |
2167 | 25.2k | return new_slot; |
2168 | 25.2k | } Unexecuted instantiation: void* absl::container_internal::GrowSooTableToNextCapacityAndPrepareInsert<0ul, false>(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, absl::FunctionRef<unsigned long (unsigned long)>, bool) Unexecuted instantiation: void* absl::container_internal::GrowSooTableToNextCapacityAndPrepareInsert<1ul, true>(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, absl::FunctionRef<unsigned long (unsigned long)>, bool) Unexecuted instantiation: void* absl::container_internal::GrowSooTableToNextCapacityAndPrepareInsert<4ul, true>(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, absl::FunctionRef<unsigned long (unsigned long)>, bool) void* absl::container_internal::GrowSooTableToNextCapacityAndPrepareInsert<8ul, true>(absl::container_internal::CommonFields&, absl::container_internal::PolicyFunctions const&, absl::FunctionRef<unsigned long (unsigned long)>, bool) Line | Count | Source | 2106 | 25.2k | absl::FunctionRef<size_t(size_t)> get_hash, bool force_sampling) { | 2107 | 25.2k | AssertSoo(common, policy); | 2108 | 25.2k | if (ABSL_PREDICT_FALSE(force_sampling)) { | 2109 | | // The table is empty, it is only used for forced sampling of SOO tables. | 2110 | 0 | return GrowEmptySooTableToNextCapacityForceSamplingAndPrepareInsert( | 2111 | 0 | common, policy, get_hash); | 2112 | 0 | } | 2113 | 25.2k | ABSL_SWISSTABLE_ASSERT(common.size() == policy.soo_capacity()); | 2114 | 25.2k | static constexpr size_t kNewCapacity = NextCapacity(SooCapacity()); | 2115 | 25.2k | const size_t slot_size = policy.slot_size; | 2116 | 25.2k | void* alloc = policy.get_char_alloc(common); | 2117 | 25.2k | common.set_capacity(kNewCapacity); | 2118 | | | 2119 | | // Since the table is not empty, it will not be sampled. | 2120 | | // The decision to sample was already made during the first insertion. | 2121 | | // | 2122 | | // We do not set control and slots in CommonFields yet to avoid overriding | 2123 | | // SOO data. | 2124 | 25.2k | const auto [new_ctrl, new_slots] = AllocBackingArray( | 2125 | 25.2k | common, policy, kNewCapacity, /*has_infoz=*/false, alloc, | 2126 | 25.2k | /*blocked_element_count=*/0); | 2127 | | | 2128 | 25.2k | PrepareInsertCommon(common); | 2129 | 25.2k | ABSL_SWISSTABLE_ASSERT(common.size() == 2); | 2130 | 25.2k | common.InitGrowthLeftNoDeleted(kNewCapacity - 2, kNewCapacity); | 2131 | 25.2k | common.generate_new_seed(/*has_infoz=*/false); | 2132 | 25.2k | const h2_t soo_slot_h2 = H2(policy.hash_slot( | 2133 | 25.2k | policy.hash_fn(common), common.soo_data(), common.seed().seed())); | 2134 | 25.2k | const size_t new_hash = get_hash(common.seed().seed()); | 2135 | | | 2136 | 25.2k | const size_t offset = Resize1To3NewOffset(new_hash); | 2137 | 25.2k | InitializeThreeElementsControlBytes(soo_slot_h2, H2(new_hash), offset, | 2138 | 25.2k | new_ctrl); | 2139 | | | 2140 | 25.2k | SanitizerPoisonMemoryRegion(new_slots, slot_size * kNewCapacity); | 2141 | 25.2k | void* target_slot = SlotAddress(new_slots, SooSlotIndex(), slot_size); | 2142 | 25.2k | SanitizerUnpoisonMemoryRegion(target_slot, slot_size); | 2143 | 25.2k | if constexpr (TransferUsesMemcpy) { | 2144 | | // Target slot is placed at index 1, but capacity is at | 2145 | | // minimum 3. So we are allowed to copy at least twice as much | 2146 | | // memory. | 2147 | 25.2k | static_assert(SooSlotIndex() == 1); | 2148 | 25.2k | static_assert(SooSlotMemcpySize > 0); | 2149 | 25.2k | static_assert(SooSlotMemcpySize <= MaxSooSlotSize()); | 2150 | 25.2k | ABSL_SWISSTABLE_ASSERT(SooSlotMemcpySize <= 2 * slot_size); | 2151 | 25.2k | ABSL_SWISSTABLE_ASSERT(SooSlotMemcpySize >= slot_size); | 2152 | 25.2k | void* next_slot = SlotAddress(target_slot, 1, slot_size); | 2153 | 25.2k | SanitizerUnpoisonMemoryRegion(next_slot, SooSlotMemcpySize - slot_size); | 2154 | 25.2k | std::memcpy(target_slot, common.soo_data(), SooSlotMemcpySize); | 2155 | 25.2k | SanitizerPoisonMemoryRegion(next_slot, SooSlotMemcpySize - slot_size); | 2156 | | } else { | 2157 | | static_assert(SooSlotMemcpySize == 0); | 2158 | | policy.transfer_n(&common, target_slot, common.soo_data(), 1); | 2159 | | } | 2160 | 0 | common.set_control(new_ctrl); | 2161 | | | 2162 | | // Full SOO table couldn't be sampled. If SOO table is sampled, it would | 2163 | | // have been resized to the next capacity. | 2164 | 25.2k | ABSL_SWISSTABLE_ASSERT(!common.infoz().IsSampled()); | 2165 | 25.2k | void* new_slot = SlotAddress(new_slots, offset, slot_size); | 2166 | 25.2k | SanitizerUnpoisonMemoryRegion(new_slot, slot_size); | 2167 | 25.2k | return new_slot; | 2168 | 25.2k | } |
|
2169 | | |
2170 | | void Rehash(CommonFields& common, const PolicyFunctions& __restrict policy, |
2171 | 0 | size_t n) { |
2172 | 0 | const size_t cap = common.capacity(); |
2173 | |
|
2174 | 0 | auto clear_backing_array = [&]() { |
2175 | 0 | ClearBackingArrayNoReuse(common, policy, policy.get_char_alloc(common)); |
2176 | 0 | }; |
2177 | |
|
2178 | 0 | const size_t slot_size = policy.slot_size; |
2179 | |
|
2180 | 0 | if (n == 0) { |
2181 | 0 | if (cap <= policy.soo_capacity()) return; |
2182 | 0 | if (common.empty()) { |
2183 | 0 | clear_backing_array(); |
2184 | 0 | return; |
2185 | 0 | } |
2186 | 0 | if (common.size() <= policy.soo_capacity()) { |
2187 | | // When the table is already sampled, we keep it sampled. |
2188 | 0 | if (common.infoz().IsSampled()) { |
2189 | 0 | static constexpr size_t kInitialSampledCapacity = |
2190 | 0 | NextCapacity(SooCapacity()); |
2191 | 0 | if (cap > kInitialSampledCapacity) { |
2192 | 0 | ResizeAllocatedTableWithSeedChange(common, policy, |
2193 | 0 | kInitialSampledCapacity); |
2194 | 0 | } |
2195 | | // This asserts that we didn't lose sampling coverage in `resize`. |
2196 | 0 | ABSL_SWISSTABLE_ASSERT(common.infoz().IsSampled()); |
2197 | 0 | return; |
2198 | 0 | } |
2199 | 0 | ABSL_SWISSTABLE_ASSERT(slot_size <= sizeof(HeapOrSoo)); |
2200 | 0 | ABSL_SWISSTABLE_ASSERT(policy.slot_align <= alignof(HeapOrSoo)); |
2201 | 0 | HeapOrSoo tmp_slot; |
2202 | 0 | size_t begin_offset = FindFirstFullSlot(0, cap, common.control()); |
2203 | 0 | policy.transfer_n( |
2204 | 0 | &common, &tmp_slot, |
2205 | 0 | SlotAddress(common.slot_array(cap), begin_offset, slot_size), 1); |
2206 | 0 | clear_backing_array(); |
2207 | 0 | policy.transfer_n(&common, common.soo_data(), &tmp_slot, 1); |
2208 | 0 | common.set_full_soo(); |
2209 | 0 | return; |
2210 | 0 | } |
2211 | 0 | } |
2212 | | |
2213 | | // bitor is a faster way of doing `max` here. We will round up to the next |
2214 | | // power-of-2-minus-1, so bitor is good enough. |
2215 | 0 | const size_t new_capacity = |
2216 | 0 | NormalizeCapacity(n | SizeToCapacity(common.size())); |
2217 | 0 | ValidateMaxCapacity(new_capacity, policy.key_size, policy.slot_size); |
2218 | | // n == 0 unconditionally rehashes as per the standard. |
2219 | 0 | if (n == 0 || new_capacity > cap) { |
2220 | 0 | if (cap == policy.soo_capacity()) { |
2221 | 0 | if (common.empty()) { |
2222 | 0 | ResizeEmptyNonAllocatedTableImpl(common, policy, new_capacity, |
2223 | 0 | /*blocked_element_count=*/0, |
2224 | 0 | /*force_infoz=*/false); |
2225 | 0 | } else { |
2226 | 0 | ResizeFullSooTable(common, policy, new_capacity, |
2227 | 0 | ResizeFullSooTableSamplingMode::kNoSampling); |
2228 | 0 | } |
2229 | 0 | } else { |
2230 | 0 | ResizeAllocatedTableWithSeedChange(common, policy, new_capacity); |
2231 | 0 | } |
2232 | | // This is after resize, to ensure that we have completed the allocation |
2233 | | // and have potentially sampled the hashtable. |
2234 | 0 | common.infoz().RecordReservation(n); |
2235 | 0 | } |
2236 | 0 | } |
2237 | | |
2238 | | void Copy(CommonFields& common, const PolicyFunctions& __restrict policy, |
2239 | | const CommonFields& other, |
2240 | 0 | absl::FunctionRef<void(void*, const void*)> copy_fn) { |
2241 | 0 | const size_t size = other.size(); |
2242 | 0 | ABSL_SWISSTABLE_ASSERT(size > 0); |
2243 | 0 | const size_t soo_capacity = policy.soo_capacity(); |
2244 | 0 | const size_t slot_size = policy.slot_size; |
2245 | 0 | const bool soo_enabled = policy.soo_enabled; |
2246 | 0 | if (size == 1) { |
2247 | 0 | if (!soo_enabled) { |
2248 | 0 | ReserveEmptyNonAllocatedTableToFitNewSize(common, policy, 1); |
2249 | 0 | common.infoz().RecordStorageChanged(1, 1); |
2250 | 0 | } |
2251 | 0 | IncrementSmallSize(common, policy); |
2252 | 0 | const size_t other_capacity = other.capacity(); |
2253 | 0 | const void* other_slot = |
2254 | 0 | other_capacity <= soo_capacity ? other.soo_data() |
2255 | 0 | : IsSmallCapacity(other_capacity) |
2256 | 0 | ? other.slot_array(other_capacity) |
2257 | 0 | : SlotAddress(other.slot_array(other_capacity), |
2258 | 0 | FindFirstFullSlot(0, other_capacity, other.control()), |
2259 | 0 | slot_size); |
2260 | 0 | copy_fn(soo_enabled ? common.soo_data() |
2261 | 0 | : SingleSlotAddress</*kSooEnabled=*/false>(common), |
2262 | 0 | other_slot); |
2263 | |
|
2264 | 0 | if (soo_enabled && policy.is_hashtablez_eligible && |
2265 | 0 | ShouldSampleNextTable()) { |
2266 | 0 | GrowFullSooTableToNextCapacityForceSampling(common, policy); |
2267 | 0 | } |
2268 | 0 | return; |
2269 | 0 | } |
2270 | | |
2271 | 0 | ReserveTableToFitNewSize(common, policy, size); |
2272 | 0 | const size_t blocked_element_count = common.blocked_element_count(); |
2273 | 0 | auto infoz = common.infoz(); |
2274 | 0 | ABSL_SWISSTABLE_ASSERT(other.capacity() > soo_capacity); |
2275 | 0 | const size_t cap = common.capacity(); |
2276 | 0 | ABSL_SWISSTABLE_ASSERT(cap > soo_capacity); |
2277 | 0 | ABSL_ASSUME(cap > kMaxSmallCapacity); |
2278 | 0 | size_t offset = cap; |
2279 | 0 | const void* hash_fn = policy.hash_fn(common); |
2280 | 0 | auto hasher = policy.hash_slot; |
2281 | 0 | const size_t seed = common.seed().seed(); |
2282 | 0 | void* target_slot_array = common.slot_array(cap); |
2283 | 0 | IterateOverFullSlotsImpl( |
2284 | 0 | other, slot_size, [&](const ctrl_t*, void* that_slot) { |
2285 | | // The table is guaranteed to be empty, so we can do faster than |
2286 | | // a full `insert`. |
2287 | 0 | const size_t hash = (*hasher)(hash_fn, that_slot, seed); |
2288 | 0 | FindInfo target = find_first_non_full(common, hash); |
2289 | 0 | infoz.RecordInsertMiss(hash, target.probe_length); |
2290 | 0 | offset = target.offset; |
2291 | 0 | SetCtrl(common, offset, H2(hash), slot_size); |
2292 | 0 | copy_fn(SlotAddress(target_slot_array, offset, slot_size), that_slot); |
2293 | 0 | common.maybe_increment_generation_on_insert(); |
2294 | 0 | }); |
2295 | 0 | common.increment_size(size); |
2296 | 0 | ResetGrowthLeft(cap, size + blocked_element_count, common); |
2297 | 0 | } |
2298 | | |
2299 | | void ReserveTableToFitNewSize(CommonFields& common, |
2300 | | const PolicyFunctions& __restrict policy, |
2301 | 0 | size_t new_size) { |
2302 | 0 | new_size = |
2303 | 0 | std::min(new_size, MaxValidSize(policy.key_size, policy.slot_size)); |
2304 | 0 | common.reset_reserved_growth(new_size); |
2305 | 0 | common.set_reservation_size(new_size); |
2306 | 0 | ABSL_SWISSTABLE_ASSERT(new_size > policy.soo_capacity()); |
2307 | 0 | const size_t cap = common.capacity(); |
2308 | 0 | if (ABSL_PREDICT_TRUE(common.empty() && cap <= policy.soo_capacity())) { |
2309 | 0 | return ReserveEmptyNonAllocatedTableToFitNewSize(common, policy, new_size); |
2310 | 0 | } |
2311 | | |
2312 | 0 | ABSL_SWISSTABLE_ASSERT(!common.empty() || cap > policy.soo_capacity()); |
2313 | 0 | ABSL_SWISSTABLE_ASSERT(cap > 0); |
2314 | 0 | const size_t max_size_before_growth = |
2315 | 0 | IsSmallCapacity(cap) ? cap |
2316 | 0 | : common.size() + common.GetGrowthLeftTotalSlow(cap); |
2317 | 0 | if (new_size <= max_size_before_growth) { |
2318 | 0 | return; |
2319 | 0 | } |
2320 | 0 | ReserveAllocatedTable(common, policy, new_size); |
2321 | 0 | } |
2322 | | |
2323 | | namespace { |
2324 | | void* PrepareInsertLargeImpl(CommonFields& common, |
2325 | | const PolicyFunctions& __restrict policy, |
2326 | | size_t hash, |
2327 | | Group::NonIterableBitMaskType mask_empty, |
2328 | 10.8M | FindInfo target_group) { |
2329 | 10.8M | ABSL_SWISSTABLE_ASSERT(!common.is_small()); |
2330 | | // When there are no deleted slots in the table |
2331 | | // and growth_left is positive, we can insert at the first |
2332 | | // empty slot in the probe sequence (target). |
2333 | 10.8M | if (ABSL_PREDICT_FALSE( |
2334 | 10.8M | !common.GetGrowthInfoLowerBound().HasNoDeletedAndGrowthLeft())) { |
2335 | 183k | return PrepareInsertLargeSlow(common, policy, hash); |
2336 | 183k | } |
2337 | 10.6M | PrepareInsertCommon(common); |
2338 | 10.6M | common.OverwriteEmptyAsFull(); |
2339 | 10.6M | const size_t cap = common.capacity(); |
2340 | 10.6M | ABSL_ASSUME(cap > kMaxSmallCapacity); |
2341 | 10.6M | target_group.offset += mask_empty.LowestBitSet(); |
2342 | 10.6M | target_group.offset &= cap; |
2343 | 10.6M | SetCtrl(common, target_group.offset, H2(hash), policy.slot_size); |
2344 | 10.6M | common.infoz().RecordInsertMiss(hash, target_group.probe_length); |
2345 | 10.6M | return SlotAddress(common.slot_array(cap), target_group.offset, |
2346 | 10.6M | policy.slot_size); |
2347 | 10.6M | } |
2348 | | } // namespace |
2349 | | |
2350 | | void* PrepareInsertLarge(CommonFields& common, |
2351 | | const PolicyFunctions& __restrict policy, size_t hash, |
2352 | | Group::NonIterableBitMaskType mask_empty, |
2353 | 10.8M | FindInfo target_group) { |
2354 | | // NOLINTNEXTLINE(misc-static-assert) |
2355 | 10.8M | ABSL_SWISSTABLE_ASSERT(!SwisstableGenerationsEnabled()); |
2356 | 10.8M | return PrepareInsertLargeImpl(common, policy, hash, mask_empty, target_group); |
2357 | 10.8M | } |
2358 | | |
2359 | | void* PrepareInsertLargeGenerationsEnabled( |
2360 | | CommonFields& common, const PolicyFunctions& __restrict policy, size_t hash, |
2361 | | Group::NonIterableBitMaskType mask_empty, FindInfo target_group, |
2362 | 0 | absl::FunctionRef<size_t(size_t)> recompute_hash) { |
2363 | | // NOLINTNEXTLINE(misc-static-assert) |
2364 | 0 | ABSL_SWISSTABLE_ASSERT(SwisstableGenerationsEnabled()); |
2365 | 0 | const size_t cap = common.capacity(); |
2366 | 0 | const size_t growth_left = common.GetGrowthLeftTotalSlow(cap); |
2367 | | // As an optimization, we avoid calling ShouldRehashForBugDetection if we |
2368 | | // will end up rehashing anyways. |
2369 | 0 | if (growth_left > 0 && common.should_rehash_for_bug_detection_on_insert()) { |
2370 | | // Move to a different heap allocation in order to detect bugs. |
2371 | 0 | ResizeAllocatedTableWithSeedChange(common, policy, cap); |
2372 | 0 | hash = recompute_hash(common.seed().seed()); |
2373 | 0 | std::tie(target_group, mask_empty) = |
2374 | 0 | find_first_non_full_group(common, hash); |
2375 | 0 | } |
2376 | 0 | return PrepareInsertLargeImpl(common, policy, hash, mask_empty, target_group); |
2377 | 0 | } |
2378 | | |
2379 | | namespace { |
2380 | | // Returns true if the following is true |
2381 | | // 1. OptimalMemcpySizeForSooSlotTransfer(left) > |
2382 | | // OptimalMemcpySizeForSooSlotTransfer(left - 1) |
2383 | | // 2. OptimalMemcpySizeForSooSlotTransfer(left) are equal for all i in [left, |
2384 | | // right]. |
2385 | | // This function is used to verify that we have all the possible template |
2386 | | // instantiations for GrowFullSooTableToNextCapacity. |
2387 | | // With this verification the problem may be detected at compile time instead of |
2388 | | // link time. |
2389 | | constexpr bool VerifyOptimalMemcpySizeForSooSlotTransferRange(size_t left, |
2390 | 0 | size_t right) { |
2391 | 0 | size_t optimal_size_for_range = OptimalMemcpySizeForSooSlotTransfer(left); |
2392 | 0 | if (optimal_size_for_range <= OptimalMemcpySizeForSooSlotTransfer(left - 1)) { |
2393 | 0 | return false; |
2394 | 0 | } |
2395 | 0 | for (size_t i = left + 1; i <= right; ++i) { |
2396 | 0 | if (OptimalMemcpySizeForSooSlotTransfer(i) != optimal_size_for_range) { |
2397 | 0 | return false; |
2398 | 0 | } |
2399 | 0 | } |
2400 | 0 | return true; |
2401 | 0 | } |
2402 | | } // namespace |
2403 | | |
2404 | | // Extern template instantiation for inline function. |
2405 | | template size_t TryFindNewIndexWithoutProbing(size_t h1, size_t old_index, |
2406 | | size_t old_capacity, |
2407 | | ctrl_t* new_ctrl, |
2408 | | size_t new_capacity); |
2409 | | |
2410 | | // We need to instantiate ALL possible template combinations because we define |
2411 | | // the function in the cc file. |
2412 | | template void* GrowSooTableToNextCapacityAndPrepareInsert<0, false>( |
2413 | | CommonFields&, const PolicyFunctions&, absl::FunctionRef<size_t(size_t)>, |
2414 | | bool); |
2415 | | template void* GrowSooTableToNextCapacityAndPrepareInsert< |
2416 | | OptimalMemcpySizeForSooSlotTransfer(1), true>( |
2417 | | CommonFields&, const PolicyFunctions&, absl::FunctionRef<size_t(size_t)>, |
2418 | | bool); |
2419 | | |
2420 | | static_assert(VerifyOptimalMemcpySizeForSooSlotTransferRange(2, 3)); |
2421 | | template void* GrowSooTableToNextCapacityAndPrepareInsert< |
2422 | | OptimalMemcpySizeForSooSlotTransfer(3), true>( |
2423 | | CommonFields&, const PolicyFunctions&, absl::FunctionRef<size_t(size_t)>, |
2424 | | bool); |
2425 | | |
2426 | | #if UINTPTR_MAX == UINT32_MAX |
2427 | | static_assert(MaxSooSlotSize() == 4); |
2428 | | static_assert(VerifyOptimalMemcpySizeForSooSlotTransferRange(2, 4)); |
2429 | | #else |
2430 | | static_assert(VerifyOptimalMemcpySizeForSooSlotTransferRange(4, 8)); |
2431 | | template void* GrowSooTableToNextCapacityAndPrepareInsert< |
2432 | | OptimalMemcpySizeForSooSlotTransfer(8), true>( |
2433 | | CommonFields&, const PolicyFunctions&, absl::FunctionRef<size_t(size_t)>, |
2434 | | bool); |
2435 | | static_assert(MaxSooSlotSize() == 8); |
2436 | | #endif |
2437 | | |
2438 | | template void* AllocateBackingArray<kStandardBackingArrayAlignment, |
2439 | | std::allocator<char>>(void* alloc, |
2440 | | size_t n); |
2441 | | template void DeallocateBackingArray<kStandardBackingArrayAlignment, |
2442 | | std::allocator<char>>(void* alloc, |
2443 | | void* backing_array, |
2444 | | size_t n); |
2445 | | |
2446 | | template void Clear</*kSooEnabled=*/true>(CommonFields& c, |
2447 | | const PolicyFunctions& policy, |
2448 | | DestroySlotFn destroy_slot, |
2449 | | void* alloc); |
2450 | | template void Clear</*kSooEnabled=*/false>(CommonFields& c, |
2451 | | const PolicyFunctions& policy, |
2452 | | DestroySlotFn destroy_slot, |
2453 | | void* alloc); |
2454 | | |
2455 | | template void Destruct</*kSooEnabled=*/true>(CommonFields& c, |
2456 | | const DtorPolicy& policy, |
2457 | | DeallocBackingArrayFn dealloc, |
2458 | | void* alloc); |
2459 | | template void Destruct</*kSooEnabled=*/true>(CommonFields& c, |
2460 | | const DtorPolicy& policy, |
2461 | | DeallocBackingArrayFn dealloc); |
2462 | | template void Destruct</*kSooEnabled=*/true>(CommonFields& c, |
2463 | | const DtorPolicy& policy); |
2464 | | template void Destruct</*kSooEnabled=*/false>(CommonFields& c, |
2465 | | const DtorPolicy& policy, |
2466 | | DeallocBackingArrayFn dealloc, |
2467 | | void* alloc); |
2468 | | template void Destruct</*kSooEnabled=*/false>(CommonFields& c, |
2469 | | const DtorPolicy& policy, |
2470 | | DeallocBackingArrayFn dealloc); |
2471 | | template void Destruct</*kSooEnabled=*/false>(CommonFields& c, |
2472 | | const DtorPolicy& policy); |
2473 | | |
2474 | | } // namespace container_internal |
2475 | | ABSL_NAMESPACE_END |
2476 | | } // namespace absl |