/src/abseil-cpp/absl/hash/internal/hash.h
Line | Count | Source (jump to first uncovered line) |
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 | | // ----------------------------------------------------------------------------- |
16 | | // File: hash.h |
17 | | // ----------------------------------------------------------------------------- |
18 | | // |
19 | | #ifndef ABSL_HASH_INTERNAL_HASH_H_ |
20 | | #define ABSL_HASH_INTERNAL_HASH_H_ |
21 | | |
22 | | #ifdef __APPLE__ |
23 | | #include <Availability.h> |
24 | | #include <TargetConditionals.h> |
25 | | #endif |
26 | | |
27 | | // We include config.h here to make sure that ABSL_INTERNAL_CPLUSPLUS_LANG is |
28 | | // defined. |
29 | | #include "absl/base/config.h" |
30 | | |
31 | | // GCC15 warns that <ciso646> is deprecated in C++17 and suggests using |
32 | | // <version> instead, even though <version> is not available in C++17 mode prior |
33 | | // to GCC9. |
34 | | #if defined(__has_include) |
35 | | #if __has_include(<version>) |
36 | | #define ABSL_INTERNAL_VERSION_HEADER_AVAILABLE 1 |
37 | | #endif |
38 | | #endif |
39 | | |
40 | | // For feature testing and determining which headers can be included. |
41 | | #if ABSL_INTERNAL_CPLUSPLUS_LANG >= 202002L || \ |
42 | | defined(ABSL_INTERNAL_VERSION_HEADER_AVAILABLE) |
43 | | #include <version> |
44 | | #else |
45 | | #include <ciso646> |
46 | | #endif |
47 | | |
48 | | #undef ABSL_INTERNAL_VERSION_HEADER_AVAILABLE |
49 | | |
50 | | #include <algorithm> |
51 | | #include <array> |
52 | | #include <bitset> |
53 | | #include <cassert> |
54 | | #include <cmath> |
55 | | #include <cstddef> |
56 | | #include <cstdint> |
57 | | #include <cstring> |
58 | | #include <deque> |
59 | | #include <forward_list> |
60 | | #include <functional> |
61 | | #include <iterator> |
62 | | #include <limits> |
63 | | #include <list> |
64 | | #include <map> |
65 | | #include <memory> |
66 | | #include <set> |
67 | | #include <string> |
68 | | #include <string_view> |
69 | | #include <tuple> |
70 | | #include <type_traits> |
71 | | #include <unordered_map> |
72 | | #include <unordered_set> |
73 | | #include <utility> |
74 | | #include <vector> |
75 | | |
76 | | #include "absl/base/attributes.h" |
77 | | #include "absl/base/internal/unaligned_access.h" |
78 | | #include "absl/base/optimization.h" |
79 | | #include "absl/base/port.h" |
80 | | #include "absl/container/fixed_array.h" |
81 | | #include "absl/hash/internal/city.h" |
82 | | #include "absl/hash/internal/weakly_mixed_integer.h" |
83 | | #include "absl/meta/type_traits.h" |
84 | | #include "absl/numeric/bits.h" |
85 | | #include "absl/numeric/int128.h" |
86 | | #include "absl/strings/string_view.h" |
87 | | #include "absl/types/optional.h" |
88 | | #include "absl/types/variant.h" |
89 | | #include "absl/utility/utility.h" |
90 | | |
91 | | #if defined(__cpp_lib_filesystem) && __cpp_lib_filesystem >= 201703L |
92 | | #include <filesystem> // NOLINT |
93 | | #endif |
94 | | |
95 | | namespace absl { |
96 | | ABSL_NAMESPACE_BEGIN |
97 | | |
98 | | class HashState; |
99 | | |
100 | | namespace hash_internal { |
101 | | |
102 | | // Internal detail: Large buffers are hashed in smaller chunks. This function |
103 | | // returns the size of these chunks. |
104 | 7.19M | constexpr size_t PiecewiseChunkSize() { return 1024; } |
105 | | |
106 | | // PiecewiseCombiner is an internal-only helper class for hashing a piecewise |
107 | | // buffer of `char` or `unsigned char` as though it were contiguous. This class |
108 | | // provides two methods: |
109 | | // |
110 | | // H add_buffer(state, data, size) |
111 | | // H finalize(state) |
112 | | // |
113 | | // `add_buffer` can be called zero or more times, followed by a single call to |
114 | | // `finalize`. This will produce the same hash expansion as concatenating each |
115 | | // buffer piece into a single contiguous buffer, and passing this to |
116 | | // `H::combine_contiguous`. |
117 | | // |
118 | | // Example usage: |
119 | | // PiecewiseCombiner combiner; |
120 | | // for (const auto& piece : pieces) { |
121 | | // state = combiner.add_buffer(std::move(state), piece.data, piece.size); |
122 | | // } |
123 | | // return combiner.finalize(std::move(state)); |
124 | | class PiecewiseCombiner { |
125 | | public: |
126 | | PiecewiseCombiner() = default; |
127 | | PiecewiseCombiner(const PiecewiseCombiner&) = delete; |
128 | | PiecewiseCombiner& operator=(const PiecewiseCombiner&) = delete; |
129 | | |
130 | | // Appends the given range of bytes to the sequence to be hashed, which may |
131 | | // modify the provided hash state. |
132 | | template <typename H> |
133 | | H add_buffer(H state, const unsigned char* data, size_t size); |
134 | | template <typename H> |
135 | 0 | H add_buffer(H state, const char* data, size_t size) { |
136 | 0 | return add_buffer(std::move(state), |
137 | 0 | reinterpret_cast<const unsigned char*>(data), size); |
138 | 0 | } |
139 | | |
140 | | // Finishes combining the hash sequence, which may may modify the provided |
141 | | // hash state. |
142 | | // |
143 | | // Once finalize() is called, add_buffer() may no longer be called. The |
144 | | // resulting hash state will be the same as if the pieces passed to |
145 | | // add_buffer() were concatenated into a single flat buffer, and then provided |
146 | | // to H::combine_contiguous(). |
147 | | template <typename H> |
148 | | H finalize(H state); |
149 | | |
150 | | private: |
151 | | unsigned char buf_[PiecewiseChunkSize()]; |
152 | | size_t position_ = 0; |
153 | | bool added_something_ = false; |
154 | | }; |
155 | | |
156 | | // Trait class which returns true if T is hashable by the absl::Hash framework. |
157 | | // Used for the AbslHashValue implementations for composite types below. |
158 | | template <typename T> |
159 | | struct is_hashable; |
160 | | |
161 | | // HashStateBase is an internal implementation detail that contains common |
162 | | // implementation details for all of the "hash state objects" objects generated |
163 | | // by Abseil. This is not a public API; users should not create classes that |
164 | | // inherit from this. |
165 | | // |
166 | | // A hash state object is the template argument `H` passed to `AbslHashValue`. |
167 | | // It represents an intermediate state in the computation of an unspecified hash |
168 | | // algorithm. `HashStateBase` provides a CRTP style base class for hash state |
169 | | // implementations. Developers adding type support for `absl::Hash` should not |
170 | | // rely on any parts of the state object other than the following member |
171 | | // functions: |
172 | | // |
173 | | // * HashStateBase::combine() |
174 | | // * HashStateBase::combine_contiguous() |
175 | | // * HashStateBase::combine_unordered() |
176 | | // |
177 | | // A derived hash state class of type `H` must provide a public member function |
178 | | // with a signature similar to the following: |
179 | | // |
180 | | // `static H combine_contiguous(H state, const unsigned char*, size_t)`. |
181 | | // |
182 | | // It must also provide a private template method named RunCombineUnordered. |
183 | | // |
184 | | // A "consumer" is a 1-arg functor returning void. Its argument is a reference |
185 | | // to an inner hash state object, and it may be called multiple times. When |
186 | | // called, the functor consumes the entropy from the provided state object, |
187 | | // and resets that object to its empty state. |
188 | | // |
189 | | // A "combiner" is a stateless 2-arg functor returning void. Its arguments are |
190 | | // an inner hash state object and an ElementStateConsumer functor. A combiner |
191 | | // uses the provided inner hash state object to hash each element of the |
192 | | // container, passing the inner hash state object to the consumer after hashing |
193 | | // each element. |
194 | | // |
195 | | // Given these definitions, a derived hash state class of type H |
196 | | // must provide a private template method with a signature similar to the |
197 | | // following: |
198 | | // |
199 | | // `template <typename CombinerT>` |
200 | | // `static H RunCombineUnordered(H outer_state, CombinerT combiner)` |
201 | | // |
202 | | // This function is responsible for constructing the inner state object and |
203 | | // providing a consumer to the combiner. It uses side effects of the consumer |
204 | | // and combiner to mix the state of each element in an order-independent manner, |
205 | | // and uses this to return an updated value of `outer_state`. |
206 | | // |
207 | | // This inside-out approach generates efficient object code in the normal case, |
208 | | // but allows us to use stack storage to implement the absl::HashState type |
209 | | // erasure mechanism (avoiding heap allocations while hashing). |
210 | | // |
211 | | // `HashStateBase` will provide a complete implementation for a hash state |
212 | | // object in terms of these two methods. |
213 | | // |
214 | | // Example: |
215 | | // |
216 | | // // Use CRTP to define your derived class. |
217 | | // struct MyHashState : HashStateBase<MyHashState> { |
218 | | // static H combine_contiguous(H state, const unsigned char*, size_t); |
219 | | // using MyHashState::HashStateBase::combine; |
220 | | // using MyHashState::HashStateBase::combine_contiguous; |
221 | | // using MyHashState::HashStateBase::combine_unordered; |
222 | | // private: |
223 | | // template <typename CombinerT> |
224 | | // static H RunCombineUnordered(H state, CombinerT combiner); |
225 | | // }; |
226 | | template <typename H> |
227 | | class HashStateBase { |
228 | | public: |
229 | | // Combines an arbitrary number of values into a hash state, returning the |
230 | | // updated state. |
231 | | // |
232 | | // Each of the value types `T` must be separately hashable by the Abseil |
233 | | // hashing framework. |
234 | | // |
235 | | // NOTE: |
236 | | // |
237 | | // state = H::combine(std::move(state), value1, value2, value3); |
238 | | // |
239 | | // is guaranteed to produce the same hash expansion as: |
240 | | // |
241 | | // state = H::combine(std::move(state), value1); |
242 | | // state = H::combine(std::move(state), value2); |
243 | | // state = H::combine(std::move(state), value3); |
244 | | template <typename T, typename... Ts> |
245 | | static H combine(H state, const T& value, const Ts&... values); |
246 | 3.18M | static H combine(H state) { return state; } |
247 | | |
248 | | // Combines a contiguous array of `size` elements into a hash state, returning |
249 | | // the updated state. |
250 | | // |
251 | | // NOTE: |
252 | | // |
253 | | // state = H::combine_contiguous(std::move(state), data, size); |
254 | | // |
255 | | // is NOT guaranteed to produce the same hash expansion as a for-loop (it may |
256 | | // perform internal optimizations). If you need this guarantee, use the |
257 | | // for-loop instead. |
258 | | template <typename T> |
259 | | static H combine_contiguous(H state, const T* data, size_t size); |
260 | | |
261 | | template <typename I> |
262 | | static H combine_unordered(H state, I begin, I end); |
263 | | |
264 | | using AbslInternalPiecewiseCombiner = PiecewiseCombiner; |
265 | | |
266 | | template <typename T> |
267 | | using is_hashable = absl::hash_internal::is_hashable<T>; |
268 | | |
269 | | private: |
270 | | // Common implementation of the iteration step of a "combiner", as described |
271 | | // above. |
272 | | template <typename I> |
273 | | struct CombineUnorderedCallback { |
274 | | I begin; |
275 | | I end; |
276 | | |
277 | | template <typename InnerH, typename ElementStateConsumer> |
278 | | void operator()(InnerH inner_state, ElementStateConsumer cb) { |
279 | | for (; begin != end; ++begin) { |
280 | | inner_state = H::combine(std::move(inner_state), *begin); |
281 | | cb(inner_state); |
282 | | } |
283 | | } |
284 | | }; |
285 | | }; |
286 | | |
287 | | // `is_uniquely_represented<T>` is a trait class that indicates whether `T` |
288 | | // is uniquely represented. |
289 | | // |
290 | | // A type is "uniquely represented" if two equal values of that type are |
291 | | // guaranteed to have the same bytes in their underlying storage. In other |
292 | | // words, if `a == b`, then `memcmp(&a, &b, sizeof(T))` is guaranteed to be |
293 | | // zero. This property cannot be detected automatically, so this trait is false |
294 | | // by default, but can be specialized by types that wish to assert that they are |
295 | | // uniquely represented. This makes them eligible for certain optimizations. |
296 | | // |
297 | | // If you have any doubt whatsoever, do not specialize this template. |
298 | | // The default is completely safe, and merely disables some optimizations |
299 | | // that will not matter for most types. Specializing this template, |
300 | | // on the other hand, can be very hazardous. |
301 | | // |
302 | | // To be uniquely represented, a type must not have multiple ways of |
303 | | // representing the same value; for example, float and double are not |
304 | | // uniquely represented, because they have distinct representations for |
305 | | // +0 and -0. Furthermore, the type's byte representation must consist |
306 | | // solely of user-controlled data, with no padding bits and no compiler- |
307 | | // controlled data such as vptrs or sanitizer metadata. This is usually |
308 | | // very difficult to guarantee, because in most cases the compiler can |
309 | | // insert data and padding bits at its own discretion. |
310 | | // |
311 | | // If you specialize this template for a type `T`, you must do so in the file |
312 | | // that defines that type (or in this file). If you define that specialization |
313 | | // anywhere else, `is_uniquely_represented<T>` could have different meanings |
314 | | // in different places. |
315 | | // |
316 | | // The Enable parameter is meaningless; it is provided as a convenience, |
317 | | // to support certain SFINAE techniques when defining specializations. |
318 | | template <typename T, typename Enable = void> |
319 | | struct is_uniquely_represented : std::false_type {}; |
320 | | |
321 | | // unsigned char is a synonym for "byte", so it is guaranteed to be |
322 | | // uniquely represented. |
323 | | template <> |
324 | | struct is_uniquely_represented<unsigned char> : std::true_type {}; |
325 | | |
326 | | // is_uniquely_represented for non-standard integral types |
327 | | // |
328 | | // Integral types other than bool should be uniquely represented on any |
329 | | // platform that this will plausibly be ported to. |
330 | | template <typename Integral> |
331 | | struct is_uniquely_represented< |
332 | | Integral, typename std::enable_if<std::is_integral<Integral>::value>::type> |
333 | | : std::true_type {}; |
334 | | |
335 | | template <> |
336 | | struct is_uniquely_represented<bool> : std::false_type {}; |
337 | | |
338 | | #ifdef ABSL_HAVE_INTRINSIC_INT128 |
339 | | // Specialize the trait for GNU extension types. |
340 | | template <> |
341 | | struct is_uniquely_represented<__int128> : std::true_type {}; |
342 | | template <> |
343 | | struct is_uniquely_represented<unsigned __int128> : std::true_type {}; |
344 | | #endif // ABSL_HAVE_INTRINSIC_INT128 |
345 | | |
346 | | template <typename T> |
347 | | struct FitsIn64Bits : std::integral_constant<bool, sizeof(T) <= 8> {}; |
348 | | |
349 | | struct CombineRaw { |
350 | | template <typename H> |
351 | 1.59M | H operator()(H state, uint64_t value) const { |
352 | 1.59M | return H::combine_raw(std::move(state), value); |
353 | 1.59M | } |
354 | | }; |
355 | | |
356 | | // For use in `raw_hash_set` to pass a seed to the hash function. |
357 | | struct HashWithSeed { |
358 | | template <typename Hasher, typename T> |
359 | 0 | size_t hash(const Hasher& hasher, const T& value, size_t seed) const { |
360 | 0 | // NOLINTNEXTLINE(clang-diagnostic-sign-conversion) |
361 | 0 | return hasher.hash_with_seed(value, seed); |
362 | 0 | } Unexecuted instantiation: unsigned long absl::hash_internal::HashWithSeed::hash<absl::hash_internal::Hash<std::__1::basic_string_view<char, std::__1::char_traits<char> > >, std::__1::basic_string_view<char, std::__1::char_traits<char> > >(absl::hash_internal::Hash<std::__1::basic_string_view<char, std::__1::char_traits<char> > > const&, std::__1::basic_string_view<char, std::__1::char_traits<char> > const&, unsigned long) const Unexecuted instantiation: unsigned long absl::hash_internal::HashWithSeed::hash<absl::hash_internal::Hash<absl::Cord>, absl::Cord>(absl::hash_internal::Hash<absl::Cord> const&, absl::Cord const&, unsigned long) const |
363 | | }; |
364 | | |
365 | | // Convenience function that combines `hash_state` with the byte representation |
366 | | // of `value`. |
367 | | template <typename H, typename T, |
368 | | absl::enable_if_t<FitsIn64Bits<T>::value, int> = 0> |
369 | 640k | H hash_bytes(H hash_state, const T& value) { |
370 | 640k | const unsigned char* start = reinterpret_cast<const unsigned char*>(&value); |
371 | 640k | uint64_t v; |
372 | | if constexpr (sizeof(T) == 1) { |
373 | | v = *start; |
374 | | } else if constexpr (sizeof(T) == 2) { |
375 | | v = absl::base_internal::UnalignedLoad16(start); |
376 | 376k | } else if constexpr (sizeof(T) == 4) { |
377 | 376k | v = absl::base_internal::UnalignedLoad32(start); |
378 | 376k | } else { |
379 | 263k | static_assert(sizeof(T) == 8); |
380 | 263k | v = absl::base_internal::UnalignedLoad64(start); |
381 | 263k | } |
382 | 640k | return CombineRaw()(std::move(hash_state), v); |
383 | 640k | } _ZN4absl13hash_internal10hash_bytesINS0_15MixingHashStateEiTnNSt3__19enable_ifIXsr12FitsIn64BitsIT0_EE5valueEiE4typeELi0EEET_S8_RKS5_ Line | Count | Source | 369 | 376k | H hash_bytes(H hash_state, const T& value) { | 370 | 376k | const unsigned char* start = reinterpret_cast<const unsigned char*>(&value); | 371 | 376k | uint64_t v; | 372 | | if constexpr (sizeof(T) == 1) { | 373 | | v = *start; | 374 | | } else if constexpr (sizeof(T) == 2) { | 375 | | v = absl::base_internal::UnalignedLoad16(start); | 376 | 376k | } else if constexpr (sizeof(T) == 4) { | 377 | 376k | v = absl::base_internal::UnalignedLoad32(start); | 378 | | } else { | 379 | | static_assert(sizeof(T) == 8); | 380 | | v = absl::base_internal::UnalignedLoad64(start); | 381 | | } | 382 | 376k | return CombineRaw()(std::move(hash_state), v); | 383 | 376k | } |
_ZN4absl13hash_internal10hash_bytesINS0_15MixingHashStateEmTnNSt3__19enable_ifIXsr12FitsIn64BitsIT0_EE5valueEiE4typeELi0EEET_S8_RKS5_ Line | Count | Source | 369 | 263k | H hash_bytes(H hash_state, const T& value) { | 370 | 263k | const unsigned char* start = reinterpret_cast<const unsigned char*>(&value); | 371 | 263k | uint64_t v; | 372 | | if constexpr (sizeof(T) == 1) { | 373 | | v = *start; | 374 | | } else if constexpr (sizeof(T) == 2) { | 375 | | v = absl::base_internal::UnalignedLoad16(start); | 376 | | } else if constexpr (sizeof(T) == 4) { | 377 | | v = absl::base_internal::UnalignedLoad32(start); | 378 | 263k | } else { | 379 | 263k | static_assert(sizeof(T) == 8); | 380 | 263k | v = absl::base_internal::UnalignedLoad64(start); | 381 | 263k | } | 382 | 263k | return CombineRaw()(std::move(hash_state), v); | 383 | 263k | } |
|
384 | | template <typename H, typename T, |
385 | | absl::enable_if_t<!FitsIn64Bits<T>::value, int> = 0> |
386 | | H hash_bytes(H hash_state, const T& value) { |
387 | | const unsigned char* start = reinterpret_cast<const unsigned char*>(&value); |
388 | | return H::combine_contiguous(std::move(hash_state), start, sizeof(value)); |
389 | | } |
390 | | |
391 | | template <typename H> |
392 | | H hash_weakly_mixed_integer(H hash_state, WeaklyMixedInteger value) { |
393 | | return H::combine_weakly_mixed_integer(std::move(hash_state), value); |
394 | | } |
395 | | |
396 | | // ----------------------------------------------------------------------------- |
397 | | // AbslHashValue for Basic Types |
398 | | // ----------------------------------------------------------------------------- |
399 | | |
400 | | // Note: Default `AbslHashValue` implementations live in `hash_internal`. This |
401 | | // allows us to block lexical scope lookup when doing an unqualified call to |
402 | | // `AbslHashValue` below. User-defined implementations of `AbslHashValue` can |
403 | | // only be found via ADL. |
404 | | |
405 | | // AbslHashValue() for hashing bool values |
406 | | // |
407 | | // We use SFINAE to ensure that this overload only accepts bool, not types that |
408 | | // are convertible to bool. |
409 | | template <typename H, typename B> |
410 | | typename std::enable_if<std::is_same<B, bool>::value, H>::type AbslHashValue( |
411 | | H hash_state, B value) { |
412 | | // We use ~size_t{} instead of 1 so that all bits are different between |
413 | | // true/false instead of only 1. |
414 | | return H::combine(std::move(hash_state), |
415 | | static_cast<size_t>(value ? ~size_t{} : 0)); |
416 | | } |
417 | | |
418 | | // AbslHashValue() for hashing enum values |
419 | | template <typename H, typename Enum> |
420 | | typename std::enable_if<std::is_enum<Enum>::value, H>::type AbslHashValue( |
421 | | H hash_state, Enum e) { |
422 | | // In practice, we could almost certainly just invoke hash_bytes directly, |
423 | | // but it's possible that a sanitizer might one day want to |
424 | | // store data in the unused bits of an enum. To avoid that risk, we |
425 | | // convert to the underlying type before hashing. Hopefully this will get |
426 | | // optimized away; if not, we can reopen discussion with c-toolchain-team. |
427 | | return H::combine(std::move(hash_state), |
428 | | static_cast<typename std::underlying_type<Enum>::type>(e)); |
429 | | } |
430 | | // AbslHashValue() for hashing floating-point values |
431 | | template <typename H, typename Float> |
432 | | typename std::enable_if<std::is_same<Float, float>::value || |
433 | | std::is_same<Float, double>::value, |
434 | | H>::type |
435 | | AbslHashValue(H hash_state, Float value) { |
436 | | return hash_internal::hash_bytes(std::move(hash_state), |
437 | | value == 0 ? 0 : value); |
438 | | } |
439 | | |
440 | | // Long double has the property that it might have extra unused bytes in it. |
441 | | // For example, in x86 sizeof(long double)==16 but it only really uses 80-bits |
442 | | // of it. This means we can't use hash_bytes on a long double and have to |
443 | | // convert it to something else first. |
444 | | template <typename H, typename LongDouble> |
445 | | typename std::enable_if<std::is_same<LongDouble, long double>::value, H>::type |
446 | | AbslHashValue(H hash_state, LongDouble value) { |
447 | | const int category = std::fpclassify(value); |
448 | | switch (category) { |
449 | | case FP_INFINITE: |
450 | | // Add the sign bit to differentiate between +Inf and -Inf |
451 | | hash_state = H::combine(std::move(hash_state), std::signbit(value)); |
452 | | break; |
453 | | |
454 | | case FP_NAN: |
455 | | case FP_ZERO: |
456 | | default: |
457 | | // Category is enough for these. |
458 | | break; |
459 | | |
460 | | case FP_NORMAL: |
461 | | case FP_SUBNORMAL: |
462 | | // We can't convert `value` directly to double because this would have |
463 | | // undefined behavior if the value is out of range. |
464 | | // std::frexp gives us a value in the range (-1, -.5] or [.5, 1) that is |
465 | | // guaranteed to be in range for `double`. The truncation is |
466 | | // implementation defined, but that works as long as it is deterministic. |
467 | | int exp; |
468 | | auto mantissa = static_cast<double>(std::frexp(value, &exp)); |
469 | | hash_state = H::combine(std::move(hash_state), mantissa, exp); |
470 | | } |
471 | | |
472 | | return H::combine(std::move(hash_state), category); |
473 | | } |
474 | | |
475 | | // Without this overload, an array decays to a pointer and we hash that, which |
476 | | // is not likely to be what the caller intended. |
477 | | template <typename H, typename T, size_t N> |
478 | | H AbslHashValue(H hash_state, T (&)[N]) { |
479 | | static_assert( |
480 | | sizeof(T) == -1, |
481 | | "Hashing C arrays is not allowed. For string literals, wrap the literal " |
482 | | "in absl::string_view(). To hash the array contents, use " |
483 | | "absl::MakeSpan() or make the array an std::array. To hash the array " |
484 | | "address, use &array[0]."); |
485 | | return hash_state; |
486 | | } |
487 | | |
488 | | // AbslHashValue() for hashing pointers |
489 | | template <typename H, typename T> |
490 | | std::enable_if_t<std::is_pointer<T>::value, H> AbslHashValue(H hash_state, |
491 | | T ptr) { |
492 | | auto v = reinterpret_cast<uintptr_t>(ptr); |
493 | | // Due to alignment, pointers tend to have low bits as zero, and the next few |
494 | | // bits follow a pattern since they are also multiples of some base value. |
495 | | // Mix pointers twice to ensure we have good entropy in low bits. |
496 | | return H::combine(std::move(hash_state), v, v); |
497 | | } |
498 | | |
499 | | // AbslHashValue() for hashing nullptr_t |
500 | | template <typename H> |
501 | | H AbslHashValue(H hash_state, std::nullptr_t) { |
502 | | return H::combine(std::move(hash_state), static_cast<void*>(nullptr)); |
503 | | } |
504 | | |
505 | | // AbslHashValue() for hashing pointers-to-member |
506 | | template <typename H, typename T, typename C> |
507 | | H AbslHashValue(H hash_state, T C::*ptr) { |
508 | | auto salient_ptm_size = [](std::size_t n) -> std::size_t { |
509 | | #if defined(_MSC_VER) |
510 | | // Pointers-to-member-function on MSVC consist of one pointer plus 0, 1, 2, |
511 | | // or 3 ints. In 64-bit mode, they are 8-byte aligned and thus can contain |
512 | | // padding (namely when they have 1 or 3 ints). The value below is a lower |
513 | | // bound on the number of salient, non-padding bytes that we use for |
514 | | // hashing. |
515 | | if constexpr (alignof(T C::*) == alignof(int)) { |
516 | | // No padding when all subobjects have the same size as the total |
517 | | // alignment. This happens in 32-bit mode. |
518 | | return n; |
519 | | } else { |
520 | | // Padding for 1 int (size 16) or 3 ints (size 24). |
521 | | // With 2 ints, the size is 16 with no padding, which we pessimize. |
522 | | return n == 24 ? 20 : n == 16 ? 12 : n; |
523 | | } |
524 | | #else |
525 | | // On other platforms, we assume that pointers-to-members do not have |
526 | | // padding. |
527 | | #ifdef __cpp_lib_has_unique_object_representations |
528 | | static_assert(std::has_unique_object_representations<T C::*>::value); |
529 | | #endif // __cpp_lib_has_unique_object_representations |
530 | | return n; |
531 | | #endif |
532 | | }; |
533 | | return H::combine_contiguous(std::move(hash_state), |
534 | | reinterpret_cast<unsigned char*>(&ptr), |
535 | | salient_ptm_size(sizeof ptr)); |
536 | | } |
537 | | |
538 | | // ----------------------------------------------------------------------------- |
539 | | // AbslHashValue for Composite Types |
540 | | // ----------------------------------------------------------------------------- |
541 | | |
542 | | // AbslHashValue() for hashing pairs |
543 | | template <typename H, typename T1, typename T2> |
544 | | typename std::enable_if<is_hashable<T1>::value && is_hashable<T2>::value, |
545 | | H>::type |
546 | | AbslHashValue(H hash_state, const std::pair<T1, T2>& p) { |
547 | | return H::combine(std::move(hash_state), p.first, p.second); |
548 | | } |
549 | | |
550 | | // Helper function for hashing a tuple. The third argument should |
551 | | // be an index_sequence running from 0 to tuple_size<Tuple> - 1. |
552 | | template <typename H, typename Tuple, size_t... Is> |
553 | 0 | H hash_tuple(H hash_state, const Tuple& t, absl::index_sequence<Is...>) { |
554 | 0 | return H::combine(std::move(hash_state), std::get<Is>(t)...); |
555 | 0 | } Unexecuted instantiation: absl::hash_internal::MixingHashState absl::hash_internal::hash_tuple<absl::hash_internal::MixingHashState, std::__1::tuple<std::__1::basic_string_view<char, std::__1::char_traits<char> > const&, int const&>, 0ul, 1ul>(absl::hash_internal::MixingHashState, std::__1::tuple<std::__1::basic_string_view<char, std::__1::char_traits<char> > const&, int const&> const&, std::__1::integer_sequence<unsigned long, 0ul, 1ul>) Unexecuted instantiation: absl::hash_internal::MixingHashState absl::hash_internal::hash_tuple<absl::hash_internal::MixingHashState, std::__1::tuple<unsigned long const&>, 0ul>(absl::hash_internal::MixingHashState, std::__1::tuple<unsigned long const&> const&, std::__1::integer_sequence<unsigned long, 0ul>) |
556 | | |
557 | | // AbslHashValue for hashing tuples |
558 | | template <typename H, typename... Ts> |
559 | | #if defined(_MSC_VER) |
560 | | // This SFINAE gets MSVC confused under some conditions. Let's just disable it |
561 | | // for now. |
562 | | H |
563 | | #else // _MSC_VER |
564 | | typename std::enable_if<absl::conjunction<is_hashable<Ts>...>::value, H>::type |
565 | | #endif // _MSC_VER |
566 | 0 | AbslHashValue(H hash_state, const std::tuple<Ts...>& t) { |
567 | 0 | return hash_internal::hash_tuple(std::move(hash_state), t, |
568 | 0 | absl::make_index_sequence<sizeof...(Ts)>()); |
569 | 0 | } Unexecuted instantiation: _ZN4absl13hash_internal13AbslHashValueINS0_15MixingHashStateEJRKNSt3__117basic_string_viewIcNS3_11char_traitsIcEEEERKiEEENS3_9enable_ifIXsr4absl11conjunctionIDpNS0_11is_hashableIT0_EEEE5valueET_E4typeESH_RKNS3_5tupleIJDpSE_EEE Unexecuted instantiation: _ZN4absl13hash_internal13AbslHashValueINS0_15MixingHashStateEJRKmEEENSt3__19enable_ifIXsr4absl11conjunctionIDpNS0_11is_hashableIT0_EEEE5valueET_E4typeESB_RKNS5_5tupleIJDpS8_EEE |
570 | | |
571 | | // ----------------------------------------------------------------------------- |
572 | | // AbslHashValue for Pointers |
573 | | // ----------------------------------------------------------------------------- |
574 | | |
575 | | // AbslHashValue for hashing unique_ptr |
576 | | template <typename H, typename T, typename D> |
577 | | H AbslHashValue(H hash_state, const std::unique_ptr<T, D>& ptr) { |
578 | | return H::combine(std::move(hash_state), ptr.get()); |
579 | | } |
580 | | |
581 | | // AbslHashValue for hashing shared_ptr |
582 | | template <typename H, typename T> |
583 | | H AbslHashValue(H hash_state, const std::shared_ptr<T>& ptr) { |
584 | | return H::combine(std::move(hash_state), ptr.get()); |
585 | | } |
586 | | |
587 | | // ----------------------------------------------------------------------------- |
588 | | // AbslHashValue for String-Like Types |
589 | | // ----------------------------------------------------------------------------- |
590 | | |
591 | | // AbslHashValue for hashing strings |
592 | | // |
593 | | // All the string-like types supported here provide the same hash expansion for |
594 | | // the same character sequence. These types are: |
595 | | // |
596 | | // - `absl::Cord` |
597 | | // - `std::string` (and std::basic_string<T, std::char_traits<T>, A> for |
598 | | // any allocator A and any T in {char, wchar_t, char16_t, char32_t}) |
599 | | // - `absl::string_view`, `std::string_view`, `std::wstring_view`, |
600 | | // `std::u16string_view`, and `std::u32_string_view`. |
601 | | // |
602 | | // For simplicity, we currently support only strings built on `char`, `wchar_t`, |
603 | | // `char16_t`, or `char32_t`. This support may be broadened, if necessary, but |
604 | | // with some caution - this overload would misbehave in cases where the traits' |
605 | | // `eq()` member isn't equivalent to `==` on the underlying character type. |
606 | | template <typename H> |
607 | 289k | H AbslHashValue(H hash_state, absl::string_view str) { |
608 | 289k | return H::combine_contiguous(std::move(hash_state), str.data(), str.size()); |
609 | 289k | } |
610 | | |
611 | | // Support std::wstring, std::u16string and std::u32string. |
612 | | template <typename Char, typename Alloc, typename H, |
613 | | typename = absl::enable_if_t<std::is_same<Char, wchar_t>::value || |
614 | | std::is_same<Char, char16_t>::value || |
615 | | std::is_same<Char, char32_t>::value>> |
616 | | H AbslHashValue( |
617 | | H hash_state, |
618 | | const std::basic_string<Char, std::char_traits<Char>, Alloc>& str) { |
619 | | return H::combine_contiguous(std::move(hash_state), str.data(), str.size()); |
620 | | } |
621 | | |
622 | | // Support std::wstring_view, std::u16string_view and std::u32string_view. |
623 | | template <typename Char, typename H, |
624 | | typename = absl::enable_if_t<std::is_same<Char, wchar_t>::value || |
625 | | std::is_same<Char, char16_t>::value || |
626 | | std::is_same<Char, char32_t>::value>> |
627 | | H AbslHashValue(H hash_state, std::basic_string_view<Char> str) { |
628 | | return H::combine_contiguous(std::move(hash_state), str.data(), str.size()); |
629 | | } |
630 | | |
631 | | #if defined(__cpp_lib_filesystem) && __cpp_lib_filesystem >= 201703L && \ |
632 | | (!defined(__ENVIRONMENT_IPHONE_OS_VERSION_MIN_REQUIRED__) || \ |
633 | | __ENVIRONMENT_IPHONE_OS_VERSION_MIN_REQUIRED__ >= 130000) && \ |
634 | | (!defined(__ENVIRONMENT_MAC_OS_X_VERSION_MIN_REQUIRED__) || \ |
635 | | __ENVIRONMENT_MAC_OS_X_VERSION_MIN_REQUIRED__ >= 101500) |
636 | | |
637 | | #define ABSL_INTERNAL_STD_FILESYSTEM_PATH_HASH_AVAILABLE 1 |
638 | | |
639 | | // Support std::filesystem::path. The SFINAE is required because some string |
640 | | // types are implicitly convertible to std::filesystem::path. |
641 | | template <typename Path, typename H, |
642 | | typename = absl::enable_if_t< |
643 | | std::is_same_v<Path, std::filesystem::path>>> |
644 | | H AbslHashValue(H hash_state, const Path& path) { |
645 | | // This is implemented by deferring to the standard library to compute the |
646 | | // hash. The standard library requires that for two paths, `p1 == p2`, then |
647 | | // `hash_value(p1) == hash_value(p2)`. `AbslHashValue` has the same |
648 | | // requirement. Since `operator==` does platform specific matching, deferring |
649 | | // to the standard library is the simplest approach. |
650 | | return H::combine(std::move(hash_state), std::filesystem::hash_value(path)); |
651 | | } |
652 | | |
653 | | #endif // ABSL_INTERNAL_STD_FILESYSTEM_PATH_HASH_AVAILABLE |
654 | | |
655 | | // ----------------------------------------------------------------------------- |
656 | | // AbslHashValue for Sequence Containers |
657 | | // ----------------------------------------------------------------------------- |
658 | | |
659 | | // AbslHashValue for hashing std::array |
660 | | template <typename H, typename T, size_t N> |
661 | | typename std::enable_if<is_hashable<T>::value, H>::type AbslHashValue( |
662 | | H hash_state, const std::array<T, N>& array) { |
663 | | return H::combine_contiguous(std::move(hash_state), array.data(), |
664 | | array.size()); |
665 | | } |
666 | | |
667 | | // AbslHashValue for hashing std::deque |
668 | | template <typename H, typename T, typename Allocator> |
669 | | typename std::enable_if<is_hashable<T>::value, H>::type AbslHashValue( |
670 | | H hash_state, const std::deque<T, Allocator>& deque) { |
671 | | // TODO(gromer): investigate a more efficient implementation taking |
672 | | // advantage of the chunk structure. |
673 | | for (const auto& t : deque) { |
674 | | hash_state = H::combine(std::move(hash_state), t); |
675 | | } |
676 | | return H::combine(std::move(hash_state), WeaklyMixedInteger{deque.size()}); |
677 | | } |
678 | | |
679 | | // AbslHashValue for hashing std::forward_list |
680 | | template <typename H, typename T, typename Allocator> |
681 | | typename std::enable_if<is_hashable<T>::value, H>::type AbslHashValue( |
682 | | H hash_state, const std::forward_list<T, Allocator>& list) { |
683 | | size_t size = 0; |
684 | | for (const T& t : list) { |
685 | | hash_state = H::combine(std::move(hash_state), t); |
686 | | ++size; |
687 | | } |
688 | | return H::combine(std::move(hash_state), WeaklyMixedInteger{size}); |
689 | | } |
690 | | |
691 | | // AbslHashValue for hashing std::list |
692 | | template <typename H, typename T, typename Allocator> |
693 | | typename std::enable_if<is_hashable<T>::value, H>::type AbslHashValue( |
694 | | H hash_state, const std::list<T, Allocator>& list) { |
695 | | for (const auto& t : list) { |
696 | | hash_state = H::combine(std::move(hash_state), t); |
697 | | } |
698 | | return H::combine(std::move(hash_state), WeaklyMixedInteger{list.size()}); |
699 | | } |
700 | | |
701 | | // AbslHashValue for hashing std::vector |
702 | | // |
703 | | // Do not use this for vector<bool> on platforms that have a working |
704 | | // implementation of std::hash. It does not have a .data(), and a fallback for |
705 | | // std::hash<> is most likely faster. |
706 | | template <typename H, typename T, typename Allocator> |
707 | | typename std::enable_if<is_hashable<T>::value && !std::is_same<T, bool>::value, |
708 | | H>::type |
709 | | AbslHashValue(H hash_state, const std::vector<T, Allocator>& vector) { |
710 | | return H::combine_contiguous(std::move(hash_state), vector.data(), |
711 | | vector.size()); |
712 | | } |
713 | | |
714 | | // AbslHashValue special cases for hashing std::vector<bool> |
715 | | |
716 | | #if defined(ABSL_IS_BIG_ENDIAN) && \ |
717 | | (defined(__GLIBCXX__) || defined(__GLIBCPP__)) |
718 | | |
719 | | // std::hash in libstdc++ does not work correctly with vector<bool> on Big |
720 | | // Endian platforms therefore we need to implement a custom AbslHashValue for |
721 | | // it. More details on the bug: |
722 | | // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102531 |
723 | | template <typename H, typename T, typename Allocator> |
724 | | typename std::enable_if<is_hashable<T>::value && std::is_same<T, bool>::value, |
725 | | H>::type |
726 | | AbslHashValue(H hash_state, const std::vector<T, Allocator>& vector) { |
727 | | typename H::AbslInternalPiecewiseCombiner combiner; |
728 | | for (const auto& i : vector) { |
729 | | unsigned char c = static_cast<unsigned char>(i); |
730 | | hash_state = combiner.add_buffer(std::move(hash_state), &c, sizeof(c)); |
731 | | } |
732 | | return H::combine(combiner.finalize(std::move(hash_state)), |
733 | | WeaklyMixedInteger{vector.size()}); |
734 | | } |
735 | | #else |
736 | | // When not working around the libstdc++ bug above, we still have to contend |
737 | | // with the fact that std::hash<vector<bool>> is often poor quality, hashing |
738 | | // directly on the internal words and on no other state. On these platforms, |
739 | | // vector<bool>{1, 1} and vector<bool>{1, 1, 0} hash to the same value. |
740 | | // |
741 | | // Mixing in the size (as we do in our other vector<> implementations) on top |
742 | | // of the library-provided hash implementation avoids this QOI issue. |
743 | | template <typename H, typename T, typename Allocator> |
744 | | typename std::enable_if<is_hashable<T>::value && std::is_same<T, bool>::value, |
745 | | H>::type |
746 | | AbslHashValue(H hash_state, const std::vector<T, Allocator>& vector) { |
747 | | return H::combine(std::move(hash_state), |
748 | | std::hash<std::vector<T, Allocator>>{}(vector), |
749 | | WeaklyMixedInteger{vector.size()}); |
750 | | } |
751 | | #endif |
752 | | |
753 | | // ----------------------------------------------------------------------------- |
754 | | // AbslHashValue for Ordered Associative Containers |
755 | | // ----------------------------------------------------------------------------- |
756 | | |
757 | | // AbslHashValue for hashing std::map |
758 | | template <typename H, typename Key, typename T, typename Compare, |
759 | | typename Allocator> |
760 | | typename std::enable_if<is_hashable<Key>::value && is_hashable<T>::value, |
761 | | H>::type |
762 | | AbslHashValue(H hash_state, const std::map<Key, T, Compare, Allocator>& map) { |
763 | | for (const auto& t : map) { |
764 | | hash_state = H::combine(std::move(hash_state), t); |
765 | | } |
766 | | return H::combine(std::move(hash_state), WeaklyMixedInteger{map.size()}); |
767 | | } |
768 | | |
769 | | // AbslHashValue for hashing std::multimap |
770 | | template <typename H, typename Key, typename T, typename Compare, |
771 | | typename Allocator> |
772 | | typename std::enable_if<is_hashable<Key>::value && is_hashable<T>::value, |
773 | | H>::type |
774 | | AbslHashValue(H hash_state, |
775 | | const std::multimap<Key, T, Compare, Allocator>& map) { |
776 | | for (const auto& t : map) { |
777 | | hash_state = H::combine(std::move(hash_state), t); |
778 | | } |
779 | | return H::combine(std::move(hash_state), WeaklyMixedInteger{map.size()}); |
780 | | } |
781 | | |
782 | | // AbslHashValue for hashing std::set |
783 | | template <typename H, typename Key, typename Compare, typename Allocator> |
784 | | typename std::enable_if<is_hashable<Key>::value, H>::type AbslHashValue( |
785 | | H hash_state, const std::set<Key, Compare, Allocator>& set) { |
786 | | for (const auto& t : set) { |
787 | | hash_state = H::combine(std::move(hash_state), t); |
788 | | } |
789 | | return H::combine(std::move(hash_state), WeaklyMixedInteger{set.size()}); |
790 | | } |
791 | | |
792 | | // AbslHashValue for hashing std::multiset |
793 | | template <typename H, typename Key, typename Compare, typename Allocator> |
794 | | typename std::enable_if<is_hashable<Key>::value, H>::type AbslHashValue( |
795 | | H hash_state, const std::multiset<Key, Compare, Allocator>& set) { |
796 | | for (const auto& t : set) { |
797 | | hash_state = H::combine(std::move(hash_state), t); |
798 | | } |
799 | | return H::combine(std::move(hash_state), WeaklyMixedInteger{set.size()}); |
800 | | } |
801 | | |
802 | | // ----------------------------------------------------------------------------- |
803 | | // AbslHashValue for Unordered Associative Containers |
804 | | // ----------------------------------------------------------------------------- |
805 | | |
806 | | // AbslHashValue for hashing std::unordered_set |
807 | | template <typename H, typename Key, typename Hash, typename KeyEqual, |
808 | | typename Alloc> |
809 | | typename std::enable_if<is_hashable<Key>::value, H>::type AbslHashValue( |
810 | | H hash_state, const std::unordered_set<Key, Hash, KeyEqual, Alloc>& s) { |
811 | | return H::combine( |
812 | | H::combine_unordered(std::move(hash_state), s.begin(), s.end()), |
813 | | WeaklyMixedInteger{s.size()}); |
814 | | } |
815 | | |
816 | | // AbslHashValue for hashing std::unordered_multiset |
817 | | template <typename H, typename Key, typename Hash, typename KeyEqual, |
818 | | typename Alloc> |
819 | | typename std::enable_if<is_hashable<Key>::value, H>::type AbslHashValue( |
820 | | H hash_state, |
821 | | const std::unordered_multiset<Key, Hash, KeyEqual, Alloc>& s) { |
822 | | return H::combine( |
823 | | H::combine_unordered(std::move(hash_state), s.begin(), s.end()), |
824 | | WeaklyMixedInteger{s.size()}); |
825 | | } |
826 | | |
827 | | // AbslHashValue for hashing std::unordered_set |
828 | | template <typename H, typename Key, typename T, typename Hash, |
829 | | typename KeyEqual, typename Alloc> |
830 | | typename std::enable_if<is_hashable<Key>::value && is_hashable<T>::value, |
831 | | H>::type |
832 | | AbslHashValue(H hash_state, |
833 | | const std::unordered_map<Key, T, Hash, KeyEqual, Alloc>& s) { |
834 | | return H::combine( |
835 | | H::combine_unordered(std::move(hash_state), s.begin(), s.end()), |
836 | | WeaklyMixedInteger{s.size()}); |
837 | | } |
838 | | |
839 | | // AbslHashValue for hashing std::unordered_multiset |
840 | | template <typename H, typename Key, typename T, typename Hash, |
841 | | typename KeyEqual, typename Alloc> |
842 | | typename std::enable_if<is_hashable<Key>::value && is_hashable<T>::value, |
843 | | H>::type |
844 | | AbslHashValue(H hash_state, |
845 | | const std::unordered_multimap<Key, T, Hash, KeyEqual, Alloc>& s) { |
846 | | return H::combine( |
847 | | H::combine_unordered(std::move(hash_state), s.begin(), s.end()), |
848 | | WeaklyMixedInteger{s.size()}); |
849 | | } |
850 | | |
851 | | // ----------------------------------------------------------------------------- |
852 | | // AbslHashValue for Wrapper Types |
853 | | // ----------------------------------------------------------------------------- |
854 | | |
855 | | // AbslHashValue for hashing std::reference_wrapper |
856 | | template <typename H, typename T> |
857 | | typename std::enable_if<is_hashable<T>::value, H>::type AbslHashValue( |
858 | | H hash_state, std::reference_wrapper<T> opt) { |
859 | | return H::combine(std::move(hash_state), opt.get()); |
860 | | } |
861 | | |
862 | | // AbslHashValue for hashing absl::optional |
863 | | template <typename H, typename T> |
864 | | typename std::enable_if<is_hashable<T>::value, H>::type AbslHashValue( |
865 | | H hash_state, const absl::optional<T>& opt) { |
866 | | if (opt) hash_state = H::combine(std::move(hash_state), *opt); |
867 | | return H::combine(std::move(hash_state), opt.has_value()); |
868 | | } |
869 | | |
870 | | template <typename H> |
871 | | struct VariantVisitor { |
872 | | H&& hash_state; |
873 | | template <typename T> |
874 | | H operator()(const T& t) const { |
875 | | return H::combine(std::move(hash_state), t); |
876 | | } |
877 | | }; |
878 | | |
879 | | // AbslHashValue for hashing absl::variant |
880 | | template <typename H, typename... T> |
881 | | typename std::enable_if<conjunction<is_hashable<T>...>::value, H>::type |
882 | | AbslHashValue(H hash_state, const absl::variant<T...>& v) { |
883 | | if (!v.valueless_by_exception()) { |
884 | | hash_state = absl::visit(VariantVisitor<H>{std::move(hash_state)}, v); |
885 | | } |
886 | | return H::combine(std::move(hash_state), v.index()); |
887 | | } |
888 | | |
889 | | // ----------------------------------------------------------------------------- |
890 | | // AbslHashValue for Other Types |
891 | | // ----------------------------------------------------------------------------- |
892 | | |
893 | | // AbslHashValue for hashing std::bitset is not defined on Little Endian |
894 | | // platforms, for the same reason as for vector<bool> (see std::vector above): |
895 | | // It does not expose the raw bytes, and a fallback to std::hash<> is most |
896 | | // likely faster. |
897 | | |
898 | | #if defined(ABSL_IS_BIG_ENDIAN) && \ |
899 | | (defined(__GLIBCXX__) || defined(__GLIBCPP__)) |
900 | | // AbslHashValue for hashing std::bitset |
901 | | // |
902 | | // std::hash in libstdc++ does not work correctly with std::bitset on Big Endian |
903 | | // platforms therefore we need to implement a custom AbslHashValue for it. More |
904 | | // details on the bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102531 |
905 | | template <typename H, size_t N> |
906 | | H AbslHashValue(H hash_state, const std::bitset<N>& set) { |
907 | | typename H::AbslInternalPiecewiseCombiner combiner; |
908 | | for (size_t i = 0; i < N; i++) { |
909 | | unsigned char c = static_cast<unsigned char>(set[i]); |
910 | | hash_state = combiner.add_buffer(std::move(hash_state), &c, sizeof(c)); |
911 | | } |
912 | | return H::combine(combiner.finalize(std::move(hash_state)), N); |
913 | | } |
914 | | #endif |
915 | | |
916 | | // ----------------------------------------------------------------------------- |
917 | | |
918 | | // Mixes all values in the range [data, data+size) into the hash state. |
919 | | // This overload accepts only uniquely-represented types, and hashes them by |
920 | | // hashing the entire range of bytes. |
921 | | template <typename H, typename T> |
922 | | typename std::enable_if<is_uniquely_represented<T>::value, H>::type |
923 | 289k | hash_range_or_bytes(H hash_state, const T* data, size_t size) { |
924 | 289k | const auto* bytes = reinterpret_cast<const unsigned char*>(data); |
925 | 289k | return H::combine_contiguous(std::move(hash_state), bytes, sizeof(T) * size); |
926 | 289k | } |
927 | | |
928 | | template <typename H, typename T> |
929 | | typename std::enable_if<!is_uniquely_represented<T>::value, H>::type |
930 | | hash_range_or_bytes(H hash_state, const T* data, size_t size) { |
931 | | for (const auto end = data + size; data < end; ++data) { |
932 | | hash_state = H::combine(std::move(hash_state), *data); |
933 | | } |
934 | | return H::combine(std::move(hash_state), |
935 | | hash_internal::WeaklyMixedInteger{size}); |
936 | | } |
937 | | |
938 | | inline constexpr uint64_t kMul = uint64_t{0x79d5f9e0de1e8cf5}; |
939 | | |
940 | | // Random data taken from the hexadecimal digits of Pi's fractional component. |
941 | | // https://en.wikipedia.org/wiki/Nothing-up-my-sleeve_number |
942 | | ABSL_CACHELINE_ALIGNED inline constexpr uint64_t kStaticRandomData[] = { |
943 | | 0x243f'6a88'85a3'08d3, 0x1319'8a2e'0370'7344, 0xa409'3822'299f'31d0, |
944 | | 0x082e'fa98'ec4e'6c89, 0x4528'21e6'38d0'1377, |
945 | | }; |
946 | | |
947 | | // Extremely weak mixture of length that is mixed into the state before |
948 | | // combining the data. It is used only for small strings. This also ensures that |
949 | | // we have high entropy in all bits of the state. |
950 | | inline uint64_t PrecombineLengthMix(uint64_t state, size_t len) { |
951 | | ABSL_ASSUME(len + sizeof(uint64_t) <= sizeof(kStaticRandomData)); |
952 | | uint64_t data = absl::base_internal::UnalignedLoad64( |
953 | | reinterpret_cast<const unsigned char*>(&kStaticRandomData[0]) + len); |
954 | | return state ^ data; |
955 | | } |
956 | | |
957 | 176M | ABSL_ATTRIBUTE_ALWAYS_INLINE inline uint64_t Mix(uint64_t lhs, uint64_t rhs) { |
958 | | // Though the 128-bit product needs multiple instructions on non-x86-64 |
959 | | // platforms, it is still a good balance between speed and hash quality. |
960 | 176M | absl::uint128 m = lhs; |
961 | 176M | m *= rhs; |
962 | 176M | return Uint128High64(m) ^ Uint128Low64(m); |
963 | 176M | } |
964 | | |
965 | | // Reads 8 bytes from p. |
966 | 399k | inline uint64_t Read8(const unsigned char* p) { |
967 | | // Suppress erroneous array bounds errors on GCC. |
968 | | #if defined(__GNUC__) && !defined(__clang__) |
969 | | #pragma GCC diagnostic push |
970 | | #pragma GCC diagnostic ignored "-Warray-bounds" |
971 | | #endif |
972 | 399k | return absl::base_internal::UnalignedLoad64(p); |
973 | | #if defined(__GNUC__) && !defined(__clang__) |
974 | | #pragma GCC diagnostic pop |
975 | | #endif |
976 | 399k | } |
977 | | |
978 | | // Reads 9 to 16 bytes from p. |
979 | | // The first 8 bytes are in .first, and the rest of the bytes are in .second |
980 | | // along with duplicated bytes from .first if len<16. |
981 | | inline std::pair<uint64_t, uint64_t> Read9To16(const unsigned char* p, |
982 | 82.4k | size_t len) { |
983 | 82.4k | return {Read8(p), Read8(p + len - 8)}; |
984 | 82.4k | } |
985 | | |
986 | | // Reads 4 to 8 bytes from p. |
987 | | // Bytes are permuted and some input bytes may be duplicated in output. |
988 | 250k | inline uint64_t Read4To8(const unsigned char* p, size_t len) { |
989 | | // If `len < 8`, we duplicate bytes. We always put low memory at the end. |
990 | | // E.g., on little endian platforms: |
991 | | // `ABCD` will be read as `ABCDABCD`. |
992 | | // `ABCDE` will be read as `BCDEABCD`. |
993 | | // `ABCDEF` will be read as `CDEFABCD`. |
994 | | // `ABCDEFG` will be read as `DEFGABCD`. |
995 | | // `ABCDEFGH` will be read as `EFGHABCD`. |
996 | | // We also do not care about endianness. On big-endian platforms, bytes will |
997 | | // be permuted differently. We always shift low memory by 32, because that |
998 | | // can be pipelined earlier. Reading high memory requires computing |
999 | | // `p + len - 4`. |
1000 | 250k | uint64_t most_significant = |
1001 | 250k | static_cast<uint64_t>(absl::base_internal::UnalignedLoad32(p)) << 32; |
1002 | 250k | uint64_t least_significant = |
1003 | 250k | absl::base_internal::UnalignedLoad32(p + len - 4); |
1004 | 250k | return most_significant | least_significant; |
1005 | 250k | } |
1006 | | |
1007 | | // Reads 1 to 3 bytes from p. Some input bytes may be duplicated in output. |
1008 | 85.5k | inline uint32_t Read1To3(const unsigned char* p, size_t len) { |
1009 | | // The trick used by this implementation is to avoid branches. |
1010 | | // We always read three bytes by duplicating. |
1011 | | // E.g., |
1012 | | // `A` is read as `AAA`. |
1013 | | // `AB` is read as `ABB`. |
1014 | | // `ABC` is read as `ABC`. |
1015 | | // We always shift `p[0]` so that it can be pipelined better. |
1016 | | // Other bytes require extra computation to find indices. |
1017 | 85.5k | uint32_t mem0 = (static_cast<uint32_t>(p[0]) << 16) | p[len - 1]; |
1018 | 85.5k | uint32_t mem1 = static_cast<uint32_t>(p[len / 2]) << 8; |
1019 | 85.5k | return mem0 | mem1; |
1020 | 85.5k | } |
1021 | | |
1022 | | ABSL_ATTRIBUTE_ALWAYS_INLINE inline uint64_t CombineRawImpl(uint64_t state, |
1023 | 63.5M | uint64_t value) { |
1024 | 63.5M | return Mix(state ^ value, kMul); |
1025 | 63.5M | } |
1026 | | |
1027 | | // Slow dispatch path for calls to CombineContiguousImpl with a size argument |
1028 | | // larger than inlined size. Has the same effect as calling |
1029 | | // CombineContiguousImpl() repeatedly with the chunk stride size. |
1030 | | uint64_t CombineLargeContiguousImplOn32BitLengthGt8(const unsigned char* first, |
1031 | | size_t len, uint64_t state); |
1032 | | uint64_t CombineLargeContiguousImplOn64BitLengthGt32(const unsigned char* first, |
1033 | | size_t len, |
1034 | | uint64_t state); |
1035 | | |
1036 | | ABSL_ATTRIBUTE_ALWAYS_INLINE inline uint64_t CombineSmallContiguousImpl( |
1037 | | uint64_t state, const unsigned char* first, size_t len) { |
1038 | | ABSL_ASSUME(len <= 8); |
1039 | | uint64_t v; |
1040 | | if (len >= 4) { |
1041 | | v = Read4To8(first, len); |
1042 | | } else if (len > 0) { |
1043 | | v = Read1To3(first, len); |
1044 | | } else { |
1045 | | // Empty string must modify the state. |
1046 | | v = 0x57; |
1047 | | } |
1048 | | return CombineRawImpl(state, v); |
1049 | | } |
1050 | | |
1051 | | ABSL_ATTRIBUTE_ALWAYS_INLINE inline uint64_t CombineContiguousImpl9to16( |
1052 | | uint64_t state, const unsigned char* first, size_t len) { |
1053 | | ABSL_ASSUME(len >= 9); |
1054 | | ABSL_ASSUME(len <= 16); |
1055 | | // Note: any time one half of the mix function becomes zero it will fail to |
1056 | | // incorporate any bits from the other half. However, there is exactly 1 in |
1057 | | // 2^64 values for each side that achieve this, and only when the size is |
1058 | | // exactly 16 -- for smaller sizes there is an overlapping byte that makes |
1059 | | // this impossible unless the seed is *also* incredibly unlucky. |
1060 | | auto p = Read9To16(first, len); |
1061 | | return Mix(state ^ p.first, kMul ^ p.second); |
1062 | | } |
1063 | | |
1064 | | ABSL_ATTRIBUTE_ALWAYS_INLINE inline uint64_t CombineContiguousImpl17to32( |
1065 | | uint64_t state, const unsigned char* first, size_t len) { |
1066 | | ABSL_ASSUME(len >= 17); |
1067 | | ABSL_ASSUME(len <= 32); |
1068 | | // Do two mixes of overlapping 16-byte ranges in parallel to minimize |
1069 | | // latency. |
1070 | | const uint64_t m0 = |
1071 | | Mix(Read8(first) ^ kStaticRandomData[1], Read8(first + 8) ^ state); |
1072 | | |
1073 | | const unsigned char* tail_16b_ptr = first + (len - 16); |
1074 | | const uint64_t m1 = Mix(Read8(tail_16b_ptr) ^ kStaticRandomData[3], |
1075 | | Read8(tail_16b_ptr + 8) ^ state); |
1076 | | return m0 ^ m1; |
1077 | | } |
1078 | | |
1079 | | // Implementation of the base case for combine_contiguous where we actually |
1080 | | // mix the bytes into the state. |
1081 | | // Dispatch to different implementations of combine_contiguous depending |
1082 | | // on the value of `sizeof(size_t)`. |
1083 | | inline uint64_t CombineContiguousImpl( |
1084 | | uint64_t state, const unsigned char* first, size_t len, |
1085 | 0 | std::integral_constant<int, 4> /* sizeof_size_t */) { |
1086 | 0 | // For large values we use CityHash, for small ones we use custom low latency |
1087 | 0 | // hash. |
1088 | 0 | if (len <= 8) { |
1089 | 0 | return CombineSmallContiguousImpl(PrecombineLengthMix(state, len), first, |
1090 | 0 | len); |
1091 | 0 | } |
1092 | 0 | return CombineLargeContiguousImplOn32BitLengthGt8(first, len, state); |
1093 | 0 | } |
1094 | | |
1095 | | inline uint64_t CombineContiguousImpl( |
1096 | | uint64_t state, const unsigned char* first, size_t len, |
1097 | 1.05M | std::integral_constant<int, 8> /* sizeof_size_t */) { |
1098 | | // For large values we use LowLevelHash or CityHash depending on the platform, |
1099 | | // for small ones we use custom low latency hash. |
1100 | 1.05M | if (len <= 8) { |
1101 | 392k | return CombineSmallContiguousImpl(PrecombineLengthMix(state, len), first, |
1102 | 392k | len); |
1103 | 392k | } |
1104 | 661k | if (len <= 16) { |
1105 | 82.4k | return CombineContiguousImpl9to16(PrecombineLengthMix(state, len), first, |
1106 | 82.4k | len); |
1107 | 82.4k | } |
1108 | 578k | if (len <= 32) { |
1109 | 58.5k | return CombineContiguousImpl17to32(PrecombineLengthMix(state, len), first, |
1110 | 58.5k | len); |
1111 | 58.5k | } |
1112 | | // We must not mix length into the state here because calling |
1113 | | // CombineContiguousImpl twice with PiecewiseChunkSize() must be equivalent |
1114 | | // to calling CombineLargeContiguousImpl once with 2 * PiecewiseChunkSize(). |
1115 | 520k | return CombineLargeContiguousImplOn64BitLengthGt32(first, len, state); |
1116 | 578k | } |
1117 | | |
1118 | | #if defined(ABSL_INTERNAL_LEGACY_HASH_NAMESPACE) && \ |
1119 | | ABSL_META_INTERNAL_STD_HASH_SFINAE_FRIENDLY_ |
1120 | | #define ABSL_HASH_INTERNAL_SUPPORT_LEGACY_HASH_ 1 |
1121 | | #else |
1122 | | #define ABSL_HASH_INTERNAL_SUPPORT_LEGACY_HASH_ 0 |
1123 | | #endif |
1124 | | |
1125 | | // Type trait to select the appropriate hash implementation to use. |
1126 | | // HashSelect::type<T> will give the proper hash implementation, to be invoked |
1127 | | // as: |
1128 | | // HashSelect::type<T>::Invoke(state, value) |
1129 | | // Also, HashSelect::type<T>::value is a boolean equal to `true` if there is a |
1130 | | // valid `Invoke` function. Types that are not hashable will have a ::value of |
1131 | | // `false`. |
1132 | | struct HashSelect { |
1133 | | private: |
1134 | | struct WeaklyMixedIntegerProbe { |
1135 | | template <typename H> |
1136 | | static H Invoke(H state, WeaklyMixedInteger value) { |
1137 | | return hash_internal::hash_weakly_mixed_integer(std::move(state), value); |
1138 | | } |
1139 | | }; |
1140 | | |
1141 | | struct State : HashStateBase<State> { |
1142 | | static State combine_contiguous(State hash_state, const unsigned char*, |
1143 | | size_t); |
1144 | | using State::HashStateBase::combine_contiguous; |
1145 | | static State combine_raw(State state, uint64_t value); |
1146 | | static State combine_weakly_mixed_integer(State hash_state, |
1147 | | WeaklyMixedInteger value); |
1148 | | }; |
1149 | | |
1150 | | struct UniquelyRepresentedProbe { |
1151 | | template <typename H, typename T> |
1152 | | static auto Invoke(H state, const T& value) |
1153 | 640k | -> absl::enable_if_t<is_uniquely_represented<T>::value, H> { |
1154 | 640k | return hash_internal::hash_bytes(std::move(state), value); |
1155 | 640k | } _ZN4absl13hash_internal10HashSelect24UniquelyRepresentedProbe6InvokeINS0_15MixingHashStateEiEENSt3__19enable_ifIXsr23is_uniquely_representedIT0_EE5valueET_E4typeES8_RKS7_ Line | Count | Source | 1153 | 376k | -> absl::enable_if_t<is_uniquely_represented<T>::value, H> { | 1154 | 376k | return hash_internal::hash_bytes(std::move(state), value); | 1155 | 376k | } |
_ZN4absl13hash_internal10HashSelect24UniquelyRepresentedProbe6InvokeINS0_15MixingHashStateEmEENSt3__19enable_ifIXsr23is_uniquely_representedIT0_EE5valueET_E4typeES8_RKS7_ Line | Count | Source | 1153 | 263k | -> absl::enable_if_t<is_uniquely_represented<T>::value, H> { | 1154 | 263k | return hash_internal::hash_bytes(std::move(state), value); | 1155 | 263k | } |
|
1156 | | }; |
1157 | | |
1158 | | struct HashValueProbe { |
1159 | | template <typename H, typename T> |
1160 | | static auto Invoke(H state, const T& value) -> absl::enable_if_t< |
1161 | | std::is_same<H, |
1162 | | decltype(AbslHashValue(std::move(state), value))>::value, |
1163 | 0 | H> { |
1164 | 0 | return AbslHashValue(std::move(state), value); |
1165 | 0 | } Unexecuted instantiation: _ZN4absl13hash_internal10HashSelect14HashValueProbe6InvokeINS0_15MixingHashStateENSt3__15tupleIJRKNS5_17basic_string_viewIcNS5_11char_traitsIcEEEERKiEEEEENS5_9enable_ifIXsr3std7is_sameIT_DTcl13AbslHashValueclsr3stdE4movefp_Efp0_EEEE5valueESH_E4typeESH_RKT0_ Unexecuted instantiation: _ZN4absl13hash_internal10HashSelect14HashValueProbe6InvokeINS0_15MixingHashStateENSt3__117basic_string_viewIcNS5_11char_traitsIcEEEEEENS5_9enable_ifIXsr3std7is_sameIT_DTcl13AbslHashValueclsr3stdE4movefp_Efp0_EEEE5valueESB_E4typeESB_RKT0_ Unexecuted instantiation: _ZN4absl13hash_internal10HashSelect14HashValueProbe6InvokeINS0_15MixingHashStateENSt3__15tupleIJRKmEEEEENS5_9enable_ifIXsr3std7is_sameIT_DTcl13AbslHashValueclsr3stdE4movefp_Efp0_EEEE5valueESB_E4typeESB_RKT0_ Unexecuted instantiation: _ZN4absl13hash_internal10HashSelect14HashValueProbe6InvokeINS0_15MixingHashStateENS_4CordEEENSt3__19enable_ifIXsr3std7is_sameIT_DTcl13AbslHashValueclsr3stdE4movefp_Efp0_EEEE5valueES8_E4typeES8_RKT0_ |
1166 | | }; |
1167 | | |
1168 | | struct LegacyHashProbe { |
1169 | | #if ABSL_HASH_INTERNAL_SUPPORT_LEGACY_HASH_ |
1170 | | template <typename H, typename T> |
1171 | | static auto Invoke(H state, const T& value) -> absl::enable_if_t< |
1172 | | std::is_convertible< |
1173 | | decltype(ABSL_INTERNAL_LEGACY_HASH_NAMESPACE::hash<T>()(value)), |
1174 | | size_t>::value, |
1175 | | H> { |
1176 | | return hash_internal::hash_bytes( |
1177 | | std::move(state), |
1178 | | ABSL_INTERNAL_LEGACY_HASH_NAMESPACE::hash<T>{}(value)); |
1179 | | } |
1180 | | #endif // ABSL_HASH_INTERNAL_SUPPORT_LEGACY_HASH_ |
1181 | | }; |
1182 | | |
1183 | | struct StdHashProbe { |
1184 | | template <typename H, typename T> |
1185 | | static auto Invoke(H state, const T& value) |
1186 | | -> absl::enable_if_t<type_traits_internal::IsHashable<T>::value, H> { |
1187 | | return hash_internal::hash_bytes(std::move(state), std::hash<T>{}(value)); |
1188 | | } |
1189 | | }; |
1190 | | |
1191 | | template <typename Hash, typename T> |
1192 | | struct Probe : Hash { |
1193 | | private: |
1194 | | template <typename H, typename = decltype(H::Invoke( |
1195 | | std::declval<State>(), std::declval<const T&>()))> |
1196 | | static std::true_type Test(int); |
1197 | | template <typename U> |
1198 | | static std::false_type Test(char); |
1199 | | |
1200 | | public: |
1201 | | static constexpr bool value = decltype(Test<Hash>(0))::value; |
1202 | | }; |
1203 | | |
1204 | | public: |
1205 | | // Probe each implementation in order. |
1206 | | // disjunction provides short circuiting wrt instantiation. |
1207 | | template <typename T> |
1208 | | using Apply = absl::disjunction< // |
1209 | | Probe<WeaklyMixedIntegerProbe, T>, // |
1210 | | Probe<UniquelyRepresentedProbe, T>, // |
1211 | | Probe<HashValueProbe, T>, // |
1212 | | Probe<LegacyHashProbe, T>, // |
1213 | | Probe<StdHashProbe, T>, // |
1214 | | std::false_type>; |
1215 | | }; |
1216 | | |
1217 | | template <typename T> |
1218 | | struct is_hashable |
1219 | | : std::integral_constant<bool, HashSelect::template Apply<T>::value> {}; |
1220 | | |
1221 | | class ABSL_DLL MixingHashState : public HashStateBase<MixingHashState> { |
1222 | | template <typename T> |
1223 | | using IntegralFastPath = |
1224 | | conjunction<std::is_integral<T>, is_uniquely_represented<T>, |
1225 | | FitsIn64Bits<T>>; |
1226 | | |
1227 | | public: |
1228 | | // Move only |
1229 | | MixingHashState(MixingHashState&&) = default; |
1230 | | MixingHashState& operator=(MixingHashState&&) = default; |
1231 | | |
1232 | | // Fundamental base case for hash recursion: mixes the given range of bytes |
1233 | | // into the hash state. |
1234 | | static MixingHashState combine_contiguous(MixingHashState hash_state, |
1235 | | const unsigned char* first, |
1236 | 902k | size_t size) { |
1237 | 902k | return MixingHashState( |
1238 | 902k | CombineContiguousImpl(hash_state.state_, first, size, |
1239 | 902k | std::integral_constant<int, sizeof(size_t)>{})); |
1240 | 902k | } |
1241 | | using MixingHashState::HashStateBase::combine_contiguous; |
1242 | | |
1243 | | template <typename T> |
1244 | 0 | static size_t hash(const T& value) { |
1245 | 0 | return hash_with_seed(value, Seed()); |
1246 | 0 | } Unexecuted instantiation: unsigned long absl::hash_internal::MixingHashState::hash<std::__1::tuple<std::__1::basic_string_view<char, std::__1::char_traits<char> > const&, int const&> >(std::__1::tuple<std::__1::basic_string_view<char, std::__1::char_traits<char> > const&, int const&> const&) Unexecuted instantiation: unsigned long absl::hash_internal::MixingHashState::hash<std::__1::tuple<unsigned long const&> >(std::__1::tuple<unsigned long const&> const&) Unexecuted instantiation: unsigned long absl::hash_internal::MixingHashState::hash<std::__1::basic_string_view<char, std::__1::char_traits<char> > >(std::__1::basic_string_view<char, std::__1::char_traits<char> > const&) Unexecuted instantiation: unsigned long absl::hash_internal::MixingHashState::hash<absl::Cord>(absl::Cord const&) |
1247 | | |
1248 | | // For performance reasons in non-opt mode, we specialize this for |
1249 | | // integral types. |
1250 | | // Otherwise we would be instantiating and calling dozens of functions for |
1251 | | // something that is just one multiplication and a couple xor's. |
1252 | | // The result should be the same as running the whole algorithm, but faster. |
1253 | | template <typename T, absl::enable_if_t<IntegralFastPath<T>::value, int> = 0> |
1254 | | static size_t hash_with_seed(T value, size_t seed) { |
1255 | | return static_cast<size_t>( |
1256 | | CombineRawImpl(seed, static_cast<std::make_unsigned_t<T>>(value))); |
1257 | | } |
1258 | | |
1259 | | template <typename T, absl::enable_if_t<!IntegralFastPath<T>::value, int> = 0> |
1260 | 0 | static size_t hash_with_seed(const T& value, size_t seed) { |
1261 | 0 | return static_cast<size_t>(combine(MixingHashState{seed}, value).state_); |
1262 | 0 | } Unexecuted instantiation: _ZN4absl13hash_internal15MixingHashState14hash_with_seedINSt3__15tupleIJRKNS3_17basic_string_viewIcNS3_11char_traitsIcEEEERKiEEETnNS3_9enable_ifIXntsr16IntegralFastPathIT_EE5valueEiE4typeELi0EEEmRKSF_m Unexecuted instantiation: _ZN4absl13hash_internal15MixingHashState14hash_with_seedINSt3__15tupleIJRKmEEETnNS3_9enable_ifIXntsr16IntegralFastPathIT_EE5valueEiE4typeELi0EEEmRKS9_m Unexecuted instantiation: _ZN4absl13hash_internal15MixingHashState14hash_with_seedINSt3__117basic_string_viewIcNS3_11char_traitsIcEEEETnNS3_9enable_ifIXntsr16IntegralFastPathIT_EE5valueEiE4typeELi0EEEmRKS9_m Unexecuted instantiation: _ZN4absl13hash_internal15MixingHashState14hash_with_seedINS_4CordETnNSt3__19enable_ifIXntsr16IntegralFastPathIT_EE5valueEiE4typeELi0EEEmRKS6_m |
1263 | | |
1264 | | private: |
1265 | | friend class MixingHashState::HashStateBase; |
1266 | | template <typename H> |
1267 | | friend H absl::hash_internal::hash_weakly_mixed_integer(H, |
1268 | | WeaklyMixedInteger); |
1269 | | // Allow the HashState type-erasure implementation to invoke |
1270 | | // RunCombinedUnordered() directly. |
1271 | | friend class absl::HashState; |
1272 | | friend struct CombineRaw; |
1273 | | |
1274 | | // For use in Seed(). |
1275 | | static const void* const kSeed; |
1276 | | |
1277 | | // Invoked only once for a given argument; that plus the fact that this is |
1278 | | // move-only ensures that there is only one non-moved-from object. |
1279 | 0 | MixingHashState() : state_(Seed()) {} |
1280 | | |
1281 | | // Workaround for MSVC bug. |
1282 | | // We make the type copyable to fix the calling convention, even though we |
1283 | | // never actually copy it. Keep it private to not affect the public API of the |
1284 | | // type. |
1285 | | MixingHashState(const MixingHashState&) = default; |
1286 | | |
1287 | 3.56M | explicit MixingHashState(uint64_t state) : state_(state) {} |
1288 | | |
1289 | | // Combines a raw value from e.g. integrals/floats/pointers/etc. This allows |
1290 | | // us to be consistent with IntegralFastPath when combining raw types, but |
1291 | | // optimize Read1To3 and Read4To8 differently for the string case. |
1292 | | static MixingHashState combine_raw(MixingHashState hash_state, |
1293 | 1.59M | uint64_t value) { |
1294 | 1.59M | return MixingHashState(CombineRawImpl(hash_state.state_, value)); |
1295 | 1.59M | } |
1296 | | |
1297 | | static MixingHashState combine_weakly_mixed_integer( |
1298 | 0 | MixingHashState hash_state, WeaklyMixedInteger value) { |
1299 | 0 | // Some transformation for the value is needed to make an empty |
1300 | 0 | // string/container change the mixing hash state. |
1301 | 0 | // We use constant smaller than 8 bits to make compiler use |
1302 | 0 | // `add` with an immediate operand with 1 byte value. |
1303 | 0 | return MixingHashState{hash_state.state_ + (0x57 + value.value)}; |
1304 | 0 | } |
1305 | | |
1306 | | template <typename CombinerT> |
1307 | | static MixingHashState RunCombineUnordered(MixingHashState state, |
1308 | | CombinerT combiner) { |
1309 | | uint64_t unordered_state = 0; |
1310 | | combiner(MixingHashState{}, [&](MixingHashState& inner_state) { |
1311 | | // Add the hash state of the element to the running total, but mix the |
1312 | | // carry bit back into the low bit. This in intended to avoid losing |
1313 | | // entropy to overflow, especially when unordered_multisets contain |
1314 | | // multiple copies of the same value. |
1315 | | auto element_state = inner_state.state_; |
1316 | | unordered_state += element_state; |
1317 | | if (unordered_state < element_state) { |
1318 | | ++unordered_state; |
1319 | | } |
1320 | | inner_state = MixingHashState{}; |
1321 | | }); |
1322 | | return MixingHashState::combine(std::move(state), unordered_state); |
1323 | | } |
1324 | | |
1325 | | // A non-deterministic seed. |
1326 | | // |
1327 | | // The current purpose of this seed is to generate non-deterministic results |
1328 | | // and prevent having users depend on the particular hash values. |
1329 | | // It is not meant as a security feature right now, but it leaves the door |
1330 | | // open to upgrade it to a true per-process random seed. A true random seed |
1331 | | // costs more and we don't need to pay for that right now. |
1332 | | // |
1333 | | // On platforms with ASLR, we take advantage of it to make a per-process |
1334 | | // random value. |
1335 | | // See https://en.wikipedia.org/wiki/Address_space_layout_randomization |
1336 | | // |
1337 | | // On other platforms this is still going to be non-deterministic but most |
1338 | | // probably per-build and not per-process. |
1339 | 1.06M | ABSL_ATTRIBUTE_ALWAYS_INLINE static size_t Seed() { |
1340 | 1.06M | #if (!defined(__clang__) || __clang_major__ > 11) && \ |
1341 | 1.06M | (!defined(__apple_build_version__) || \ |
1342 | 1.06M | __apple_build_version__ >= 19558921) // Xcode 12 |
1343 | 1.06M | return static_cast<size_t>(reinterpret_cast<uintptr_t>(&kSeed)); |
1344 | | #else |
1345 | | // Workaround the absence of |
1346 | | // https://github.com/llvm/llvm-project/commit/bc15bf66dcca76cc06fe71fca35b74dc4d521021. |
1347 | | return static_cast<size_t>(reinterpret_cast<uintptr_t>(kSeed)); |
1348 | | #endif |
1349 | 1.06M | } |
1350 | | |
1351 | | uint64_t state_; |
1352 | | }; |
1353 | | |
1354 | | struct AggregateBarrier {}; |
1355 | | |
1356 | | // Add a private base class to make sure this type is not an aggregate. |
1357 | | // Aggregates can be aggregate initialized even if the default constructor is |
1358 | | // deleted. |
1359 | | struct PoisonedHash : private AggregateBarrier { |
1360 | | PoisonedHash() = delete; |
1361 | | PoisonedHash(const PoisonedHash&) = delete; |
1362 | | PoisonedHash& operator=(const PoisonedHash&) = delete; |
1363 | | }; |
1364 | | |
1365 | | template <typename T> |
1366 | | struct HashImpl { |
1367 | 0 | size_t operator()(const T& value) const { |
1368 | 0 | return MixingHashState::hash(value); |
1369 | 0 | } Unexecuted instantiation: absl::hash_internal::HashImpl<std::__1::tuple<std::__1::basic_string_view<char, std::__1::char_traits<char> > const&, int const&> >::operator()(std::__1::tuple<std::__1::basic_string_view<char, std::__1::char_traits<char> > const&, int const&> const&) const Unexecuted instantiation: absl::hash_internal::HashImpl<std::__1::tuple<unsigned long const&> >::operator()(std::__1::tuple<unsigned long const&> const&) const Unexecuted instantiation: absl::hash_internal::HashImpl<std::__1::basic_string_view<char, std::__1::char_traits<char> > >::operator()(std::__1::basic_string_view<char, std::__1::char_traits<char> > const&) const Unexecuted instantiation: absl::hash_internal::HashImpl<absl::Cord>::operator()(absl::Cord const&) const |
1370 | | |
1371 | | private: |
1372 | | friend struct HashWithSeed; |
1373 | | |
1374 | 0 | size_t hash_with_seed(const T& value, size_t seed) const { |
1375 | 0 | return MixingHashState::hash_with_seed(value, seed); |
1376 | 0 | } Unexecuted instantiation: absl::hash_internal::HashImpl<std::__1::basic_string_view<char, std::__1::char_traits<char> > >::hash_with_seed(std::__1::basic_string_view<char, std::__1::char_traits<char> > const&, unsigned long) const Unexecuted instantiation: absl::hash_internal::HashImpl<absl::Cord>::hash_with_seed(absl::Cord const&, unsigned long) const |
1377 | | }; |
1378 | | |
1379 | | template <typename T> |
1380 | | struct Hash |
1381 | | : absl::conditional_t<is_hashable<T>::value, HashImpl<T>, PoisonedHash> {}; |
1382 | | |
1383 | | template <typename H> |
1384 | | template <typename T, typename... Ts> |
1385 | 536k | H HashStateBase<H>::combine(H state, const T& value, const Ts&... values) { |
1386 | 536k | return H::combine(hash_internal::HashSelect::template Apply<T>::Invoke( |
1387 | 536k | std::move(state), value), |
1388 | 536k | values...); |
1389 | 536k | } Unexecuted instantiation: absl::hash_internal::MixingHashState absl::hash_internal::HashStateBase<absl::hash_internal::MixingHashState>::combine<std::__1::tuple<std::__1::basic_string_view<char, std::__1::char_traits<char> > const&, int const&>>(absl::hash_internal::MixingHashState, std::__1::tuple<std::__1::basic_string_view<char, std::__1::char_traits<char> > const&, int const&> const&) Unexecuted instantiation: absl::hash_internal::MixingHashState absl::hash_internal::HashStateBase<absl::hash_internal::MixingHashState>::combine<std::__1::basic_string_view<char, std::__1::char_traits<char> >, int>(absl::hash_internal::MixingHashState, std::__1::basic_string_view<char, std::__1::char_traits<char> > const&, int const&) absl::hash_internal::MixingHashState absl::hash_internal::HashStateBase<absl::hash_internal::MixingHashState>::combine<int>(absl::hash_internal::MixingHashState, int const&) Line | Count | Source | 1385 | 376k | H HashStateBase<H>::combine(H state, const T& value, const Ts&... values) { | 1386 | 376k | return H::combine(hash_internal::HashSelect::template Apply<T>::Invoke( | 1387 | 376k | std::move(state), value), | 1388 | 376k | values...); | 1389 | 376k | } |
Unexecuted instantiation: absl::hash_internal::MixingHashState absl::hash_internal::HashStateBase<absl::hash_internal::MixingHashState>::combine<std::__1::tuple<unsigned long const&>>(absl::hash_internal::MixingHashState, std::__1::tuple<unsigned long const&> const&) absl::hash_internal::MixingHashState absl::hash_internal::HashStateBase<absl::hash_internal::MixingHashState>::combine<unsigned long>(absl::hash_internal::MixingHashState, unsigned long const&) Line | Count | Source | 1385 | 159k | H HashStateBase<H>::combine(H state, const T& value, const Ts&... values) { | 1386 | 159k | return H::combine(hash_internal::HashSelect::template Apply<T>::Invoke( | 1387 | 159k | std::move(state), value), | 1388 | 159k | values...); | 1389 | 159k | } |
Unexecuted instantiation: absl::hash_internal::MixingHashState absl::hash_internal::HashStateBase<absl::hash_internal::MixingHashState>::combine<std::__1::basic_string_view<char, std::__1::char_traits<char> >>(absl::hash_internal::MixingHashState, std::__1::basic_string_view<char, std::__1::char_traits<char> > const&) Unexecuted instantiation: absl::hash_internal::MixingHashState absl::hash_internal::HashStateBase<absl::hash_internal::MixingHashState>::combine<absl::Cord>(absl::hash_internal::MixingHashState, absl::Cord const&) |
1390 | | |
1391 | | template <typename H> |
1392 | | template <typename T> |
1393 | 289k | H HashStateBase<H>::combine_contiguous(H state, const T* data, size_t size) { |
1394 | 289k | return hash_internal::hash_range_or_bytes(std::move(state), data, size); |
1395 | 289k | } |
1396 | | |
1397 | | template <typename H> |
1398 | | template <typename I> |
1399 | | H HashStateBase<H>::combine_unordered(H state, I begin, I end) { |
1400 | | return H::RunCombineUnordered(std::move(state), |
1401 | | CombineUnorderedCallback<I>{begin, end}); |
1402 | | } |
1403 | | |
1404 | | template <typename H> |
1405 | | H PiecewiseCombiner::add_buffer(H state, const unsigned char* data, |
1406 | 0 | size_t size) { |
1407 | 0 | if (position_ + size < PiecewiseChunkSize()) { |
1408 | 0 | // This partial chunk does not fill our existing buffer |
1409 | 0 | memcpy(buf_ + position_, data, size); |
1410 | 0 | position_ += size; |
1411 | 0 | return state; |
1412 | 0 | } |
1413 | 0 | added_something_ = true; |
1414 | 0 | // If the buffer is partially filled we need to complete the buffer |
1415 | 0 | // and hash it. |
1416 | 0 | if (position_ != 0) { |
1417 | 0 | const size_t bytes_needed = PiecewiseChunkSize() - position_; |
1418 | 0 | memcpy(buf_ + position_, data, bytes_needed); |
1419 | 0 | state = H::combine_contiguous(std::move(state), buf_, PiecewiseChunkSize()); |
1420 | 0 | data += bytes_needed; |
1421 | 0 | size -= bytes_needed; |
1422 | 0 | } |
1423 | 0 |
|
1424 | 0 | // Hash whatever chunks we can without copying |
1425 | 0 | while (size >= PiecewiseChunkSize()) { |
1426 | 0 | state = H::combine_contiguous(std::move(state), data, PiecewiseChunkSize()); |
1427 | 0 | data += PiecewiseChunkSize(); |
1428 | 0 | size -= PiecewiseChunkSize(); |
1429 | 0 | } |
1430 | 0 | // Fill the buffer with the remainder |
1431 | 0 | memcpy(buf_, data, size); |
1432 | 0 | position_ = size; |
1433 | 0 | return state; |
1434 | 0 | } |
1435 | | |
1436 | | template <typename H> |
1437 | 0 | H PiecewiseCombiner::finalize(H state) { |
1438 | 0 | // Do not call combine_contiguous with empty remainder since it is modifying |
1439 | 0 | // state. |
1440 | 0 | if (added_something_ && position_ == 0) { |
1441 | 0 | return state; |
1442 | 0 | } |
1443 | 0 | // We still call combine_contiguous for the entirely empty buffer. |
1444 | 0 | return H::combine_contiguous(std::move(state), buf_, position_); |
1445 | 0 | } |
1446 | | |
1447 | | } // namespace hash_internal |
1448 | | ABSL_NAMESPACE_END |
1449 | | } // namespace absl |
1450 | | |
1451 | | #endif // ABSL_HASH_INTERNAL_HASH_H_ |