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 #include "absl/base/config.h"
28
29 // For feature testing and determining which headers can be included.
30 #if ABSL_INTERNAL_CPLUSPLUS_LANG >= 202002L
31 #include <version>
32 #else
33 #include <ciso646>
34 #endif
35
36 #include <algorithm>
37 #include <array>
38 #include <bitset>
39 #include <cassert>
40 #include <cmath>
41 #include <cstddef>
42 #include <cstdint>
43 #include <cstring>
44 #include <deque>
45 #include <forward_list>
46 #include <functional>
47 #include <iterator>
48 #include <limits>
49 #include <list>
50 #include <map>
51 #include <memory>
52 #include <set>
53 #include <string>
54 #include <tuple>
55 #include <type_traits>
56 #include <unordered_map>
57 #include <unordered_set>
58 #include <utility>
59 #include <vector>
60
61 #include "absl/base/attributes.h"
62 #include "absl/base/internal/endian.h"
63 #include "absl/base/internal/unaligned_access.h"
64 #include "absl/base/optimization.h"
65 #include "absl/base/port.h"
66 #include "absl/container/fixed_array.h"
67 #include "absl/hash/internal/city.h"
68 #include "absl/hash/internal/low_level_hash.h"
69 #include "absl/meta/type_traits.h"
70 #include "absl/numeric/bits.h"
71 #include "absl/numeric/int128.h"
72 #include "absl/strings/string_view.h"
73 #include "absl/types/optional.h"
74 #include "absl/types/variant.h"
75 #include "absl/utility/utility.h"
76
77 #if defined(__cpp_lib_filesystem) && __cpp_lib_filesystem >= 201703L
78 #include <filesystem> // NOLINT
79 #endif
80
81 #ifdef ABSL_HAVE_STD_STRING_VIEW
82 #include <string_view>
83 #endif
84
85 #ifdef __ARM_ACLE
86 #include <arm_acle.h>
87 #endif
88
89 namespace absl {
90 ABSL_NAMESPACE_BEGIN
91
92 class HashState;
93
94 namespace hash_internal {
95
96 // Internal detail: Large buffers are hashed in smaller chunks. This function
97 // returns the size of these chunks.
PiecewiseChunkSize()98 constexpr size_t PiecewiseChunkSize() { return 1024; }
99
100 // PiecewiseCombiner
101 //
102 // PiecewiseCombiner is an internal-only helper class for hashing a piecewise
103 // buffer of `char` or `unsigned char` as though it were contiguous. This class
104 // provides two methods:
105 //
106 // H add_buffer(state, data, size)
107 // H finalize(state)
108 //
109 // `add_buffer` can be called zero or more times, followed by a single call to
110 // `finalize`. This will produce the same hash expansion as concatenating each
111 // buffer piece into a single contiguous buffer, and passing this to
112 // `H::combine_contiguous`.
113 //
114 // Example usage:
115 // PiecewiseCombiner combiner;
116 // for (const auto& piece : pieces) {
117 // state = combiner.add_buffer(std::move(state), piece.data, piece.size);
118 // }
119 // return combiner.finalize(std::move(state));
120 class PiecewiseCombiner {
121 public:
PiecewiseCombiner()122 PiecewiseCombiner() : position_(0) {}
123 PiecewiseCombiner(const PiecewiseCombiner&) = delete;
124 PiecewiseCombiner& operator=(const PiecewiseCombiner&) = delete;
125
126 // PiecewiseCombiner::add_buffer()
127 //
128 // Appends the given range of bytes to the sequence to be hashed, which may
129 // modify the provided hash state.
130 template <typename H>
131 H add_buffer(H state, const unsigned char* data, size_t size);
132 template <typename H>
add_buffer(H state,const char * data,size_t size)133 H add_buffer(H state, const char* data, size_t size) {
134 return add_buffer(std::move(state),
135 reinterpret_cast<const unsigned char*>(data), size);
136 }
137
138 // PiecewiseCombiner::finalize()
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_;
153 };
154
155 // is_hashable()
156 //
157 // Trait class which returns true if T is hashable by the absl::Hash framework.
158 // Used for the AbslHashValue implementations for composite types below.
159 template <typename T>
160 struct is_hashable;
161
162 // HashStateBase
163 //
164 // An internal implementation detail that contains common implementation details
165 // for all of the "hash state objects" objects generated by Abseil. This is not
166 // a public API; users should not create classes that inherit from this.
167 //
168 // A hash state object is the template argument `H` passed to `AbslHashValue`.
169 // It represents an intermediate state in the computation of an unspecified hash
170 // algorithm. `HashStateBase` provides a CRTP style base class for hash state
171 // implementations. Developers adding type support for `absl::Hash` should not
172 // rely on any parts of the state object other than the following member
173 // functions:
174 //
175 // * HashStateBase::combine()
176 // * HashStateBase::combine_contiguous()
177 // * HashStateBase::combine_unordered()
178 //
179 // A derived hash state class of type `H` must provide a public member function
180 // with a signature similar to the following:
181 //
182 // `static H combine_contiguous(H state, const unsigned char*, size_t)`.
183 //
184 // It must also provide a private template method named RunCombineUnordered.
185 //
186 // A "consumer" is a 1-arg functor returning void. Its argument is a reference
187 // to an inner hash state object, and it may be called multiple times. When
188 // called, the functor consumes the entropy from the provided state object,
189 // and resets that object to its empty state.
190 //
191 // A "combiner" is a stateless 2-arg functor returning void. Its arguments are
192 // an inner hash state object and an ElementStateConsumer functor. A combiner
193 // uses the provided inner hash state object to hash each element of the
194 // container, passing the inner hash state object to the consumer after hashing
195 // each element.
196 //
197 // Given these definitions, a derived hash state class of type H
198 // must provide a private template method with a signature similar to the
199 // following:
200 //
201 // `template <typename CombinerT>`
202 // `static H RunCombineUnordered(H outer_state, CombinerT combiner)`
203 //
204 // This function is responsible for constructing the inner state object and
205 // providing a consumer to the combiner. It uses side effects of the consumer
206 // and combiner to mix the state of each element in an order-independent manner,
207 // and uses this to return an updated value of `outer_state`.
208 //
209 // This inside-out approach generates efficient object code in the normal case,
210 // but allows us to use stack storage to implement the absl::HashState type
211 // erasure mechanism (avoiding heap allocations while hashing).
212 //
213 // `HashStateBase` will provide a complete implementation for a hash state
214 // object in terms of these two methods.
215 //
216 // Example:
217 //
218 // // Use CRTP to define your derived class.
219 // struct MyHashState : HashStateBase<MyHashState> {
220 // static H combine_contiguous(H state, const unsigned char*, size_t);
221 // using MyHashState::HashStateBase::combine;
222 // using MyHashState::HashStateBase::combine_contiguous;
223 // using MyHashState::HashStateBase::combine_unordered;
224 // private:
225 // template <typename CombinerT>
226 // static H RunCombineUnordered(H state, CombinerT combiner);
227 // };
228 template <typename H>
229 class HashStateBase {
230 public:
231 // HashStateBase::combine()
232 //
233 // Combines an arbitrary number of values into a hash state, returning the
234 // updated state.
235 //
236 // Each of the value types `T` must be separately hashable by the Abseil
237 // hashing framework.
238 //
239 // NOTE:
240 //
241 // state = H::combine(std::move(state), value1, value2, value3);
242 //
243 // is guaranteed to produce the same hash expansion as:
244 //
245 // state = H::combine(std::move(state), value1);
246 // state = H::combine(std::move(state), value2);
247 // state = H::combine(std::move(state), value3);
248 template <typename T, typename... Ts>
249 static H combine(H state, const T& value, const Ts&... values);
combine(H state)250 static H combine(H state) { return state; }
251
252 // HashStateBase::combine_contiguous()
253 //
254 // Combines a contiguous array of `size` elements into a hash state, returning
255 // the updated state.
256 //
257 // NOTE:
258 //
259 // state = H::combine_contiguous(std::move(state), data, size);
260 //
261 // is NOT guaranteed to produce the same hash expansion as a for-loop (it may
262 // perform internal optimizations). If you need this guarantee, use the
263 // for-loop instead.
264 template <typename T>
265 static H combine_contiguous(H state, const T* data, size_t size);
266
267 template <typename I>
268 static H combine_unordered(H state, I begin, I end);
269
270 using AbslInternalPiecewiseCombiner = PiecewiseCombiner;
271
272 template <typename T>
273 using is_hashable = absl::hash_internal::is_hashable<T>;
274
275 private:
276 // Common implementation of the iteration step of a "combiner", as described
277 // above.
278 template <typename I>
279 struct CombineUnorderedCallback {
280 I begin;
281 I end;
282
283 template <typename InnerH, typename ElementStateConsumer>
operatorCombineUnorderedCallback284 void operator()(InnerH inner_state, ElementStateConsumer cb) {
285 for (; begin != end; ++begin) {
286 inner_state = H::combine(std::move(inner_state), *begin);
287 cb(inner_state);
288 }
289 }
290 };
291 };
292
293 // is_uniquely_represented
294 //
295 // `is_uniquely_represented<T>` is a trait class that indicates whether `T`
296 // is uniquely represented.
297 //
298 // A type is "uniquely represented" if two equal values of that type are
299 // guaranteed to have the same bytes in their underlying storage. In other
300 // words, if `a == b`, then `memcmp(&a, &b, sizeof(T))` is guaranteed to be
301 // zero. This property cannot be detected automatically, so this trait is false
302 // by default, but can be specialized by types that wish to assert that they are
303 // uniquely represented. This makes them eligible for certain optimizations.
304 //
305 // If you have any doubt whatsoever, do not specialize this template.
306 // The default is completely safe, and merely disables some optimizations
307 // that will not matter for most types. Specializing this template,
308 // on the other hand, can be very hazardous.
309 //
310 // To be uniquely represented, a type must not have multiple ways of
311 // representing the same value; for example, float and double are not
312 // uniquely represented, because they have distinct representations for
313 // +0 and -0. Furthermore, the type's byte representation must consist
314 // solely of user-controlled data, with no padding bits and no compiler-
315 // controlled data such as vptrs or sanitizer metadata. This is usually
316 // very difficult to guarantee, because in most cases the compiler can
317 // insert data and padding bits at its own discretion.
318 //
319 // If you specialize this template for a type `T`, you must do so in the file
320 // that defines that type (or in this file). If you define that specialization
321 // anywhere else, `is_uniquely_represented<T>` could have different meanings
322 // in different places.
323 //
324 // The Enable parameter is meaningless; it is provided as a convenience,
325 // to support certain SFINAE techniques when defining specializations.
326 template <typename T, typename Enable = void>
327 struct is_uniquely_represented : std::false_type {};
328
329 // is_uniquely_represented<unsigned char>
330 //
331 // unsigned char is a synonym for "byte", so it is guaranteed to be
332 // uniquely represented.
333 template <>
334 struct is_uniquely_represented<unsigned char> : std::true_type {};
335
336 // is_uniquely_represented for non-standard integral types
337 //
338 // Integral types other than bool should be uniquely represented on any
339 // platform that this will plausibly be ported to.
340 template <typename Integral>
341 struct is_uniquely_represented<
342 Integral, typename std::enable_if<std::is_integral<Integral>::value>::type>
343 : std::true_type {};
344
345 // is_uniquely_represented<bool>
346 //
347 //
348 template <>
349 struct is_uniquely_represented<bool> : std::false_type {};
350
351 #ifdef ABSL_HAVE_INTRINSIC_INT128
352 // Specialize the trait for GNU extension types.
353 template <>
354 struct is_uniquely_represented<__int128> : std::true_type {};
355 template <>
356 struct is_uniquely_represented<unsigned __int128> : std::true_type {};
357 #endif // ABSL_HAVE_INTRINSIC_INT128
358
359 template <typename T>
360 struct FitsIn64Bits : std::integral_constant<bool, sizeof(T) <= 8> {};
361
362 struct CombineRaw {
363 template <typename H>
364 H operator()(H state, uint64_t value) const {
365 return H::combine_raw(std::move(state), value);
366 }
367 };
368
369 // hash_bytes()
370 //
371 // Convenience function that combines `hash_state` with the byte representation
372 // of `value`.
373 template <typename H, typename T,
374 absl::enable_if_t<FitsIn64Bits<T>::value, int> = 0>
375 H hash_bytes(H hash_state, const T& value) {
376 const unsigned char* start = reinterpret_cast<const unsigned char*>(&value);
377 uint64_t v;
378 if (sizeof(T) == 1) {
379 v = *start;
380 } else if (sizeof(T) == 2) {
381 v = absl::base_internal::UnalignedLoad16(start);
382 } else if (sizeof(T) == 4) {
383 v = absl::base_internal::UnalignedLoad32(start);
384 } else {
385 assert(sizeof(T) == 8);
386 v = absl::base_internal::UnalignedLoad64(start);
387 }
388 return CombineRaw()(std::move(hash_state), v);
389 }
390 template <typename H, typename T,
391 absl::enable_if_t<!FitsIn64Bits<T>::value, int> = 0>
392 H hash_bytes(H hash_state, const T& value) {
393 const unsigned char* start = reinterpret_cast<const unsigned char*>(&value);
394 return H::combine_contiguous(std::move(hash_state), start, sizeof(value));
395 }
396
397 // -----------------------------------------------------------------------------
398 // AbslHashValue for Basic Types
399 // -----------------------------------------------------------------------------
400
401 // Note: Default `AbslHashValue` implementations live in `hash_internal`. This
402 // allows us to block lexical scope lookup when doing an unqualified call to
403 // `AbslHashValue` below. User-defined implementations of `AbslHashValue` can
404 // only be found via ADL.
405
406 // AbslHashValue() for hashing bool values
407 //
408 // We use SFINAE to ensure that this overload only accepts bool, not types that
409 // are convertible to bool.
410 template <typename H, typename B>
411 typename std::enable_if<std::is_same<B, bool>::value, H>::type AbslHashValue(
412 H hash_state, B value) {
413 return H::combine(std::move(hash_state),
414 static_cast<unsigned char>(value ? 1 : 0));
415 }
416
417 // AbslHashValue() for hashing enum values
418 template <typename H, typename Enum>
419 typename std::enable_if<std::is_enum<Enum>::value, H>::type AbslHashValue(
420 H hash_state, Enum e) {
421 // In practice, we could almost certainly just invoke hash_bytes directly,
422 // but it's possible that a sanitizer might one day want to
423 // store data in the unused bits of an enum. To avoid that risk, we
424 // convert to the underlying type before hashing. Hopefully this will get
425 // optimized away; if not, we can reopen discussion with c-toolchain-team.
426 return H::combine(std::move(hash_state),
427 static_cast<typename std::underlying_type<Enum>::type>(e));
428 }
429 // AbslHashValue() for hashing floating-point values
430 template <typename H, typename Float>
431 typename std::enable_if<std::is_same<Float, float>::value ||
432 std::is_same<Float, double>::value,
433 H>::type
434 AbslHashValue(H hash_state, Float value) {
435 return hash_internal::hash_bytes(std::move(hash_state),
436 value == 0 ? 0 : value);
437 }
438
439 // Long double has the property that it might have extra unused bytes in it.
440 // For example, in x86 sizeof(long double)==16 but it only really uses 80-bits
441 // of it. This means we can't use hash_bytes on a long double and have to
442 // convert it to something else first.
443 template <typename H, typename LongDouble>
444 typename std::enable_if<std::is_same<LongDouble, long double>::value, H>::type
445 AbslHashValue(H hash_state, LongDouble value) {
446 const int category = std::fpclassify(value);
447 switch (category) {
448 case FP_INFINITE:
449 // Add the sign bit to differentiate between +Inf and -Inf
450 hash_state = H::combine(std::move(hash_state), std::signbit(value));
451 break;
452
453 case FP_NAN:
454 case FP_ZERO:
455 default:
456 // Category is enough for these.
457 break;
458
459 case FP_NORMAL:
460 case FP_SUBNORMAL:
461 // We can't convert `value` directly to double because this would have
462 // undefined behavior if the value is out of range.
463 // std::frexp gives us a value in the range (-1, -.5] or [.5, 1) that is
464 // guaranteed to be in range for `double`. The truncation is
465 // implementation defined, but that works as long as it is deterministic.
466 int exp;
467 auto mantissa = static_cast<double>(std::frexp(value, &exp));
468 hash_state = H::combine(std::move(hash_state), mantissa, exp);
469 }
470
471 return H::combine(std::move(hash_state), category);
472 }
473
474 // Without this overload, an array decays to a pointer and we hash that, which
475 // is not likely to be what the caller intended.
476 template <typename H, typename T, size_t N>
477 H AbslHashValue(H hash_state, T (&)[N]) {
478 static_assert(
479 sizeof(T) == -1,
480 "Hashing C arrays is not allowed. For string literals, wrap the literal "
481 "in absl::string_view(). To hash the array contents, use "
482 "absl::MakeSpan() or make the array an std::array. To hash the array "
483 "address, use &array[0].");
484 return hash_state;
485 }
486
487 // AbslHashValue() for hashing pointers
488 template <typename H, typename T>
489 std::enable_if_t<std::is_pointer<T>::value, H> AbslHashValue(H hash_state,
490 T ptr) {
491 auto v = reinterpret_cast<uintptr_t>(ptr);
492 // Due to alignment, pointers tend to have low bits as zero, and the next few
493 // bits follow a pattern since they are also multiples of some base value. The
494 // byte swap in WeakMix helps ensure we still have good entropy in low bits.
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 (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 // hash_tuple()
551 //
552 // Helper function for hashing a tuple. The third argument should
553 // be an index_sequence running from 0 to tuple_size<Tuple> - 1.
554 template <typename H, typename Tuple, size_t... Is>
555 H hash_tuple(H hash_state, const Tuple& t, absl::index_sequence<Is...>) {
556 return H::combine(std::move(hash_state), std::get<Is>(t)...);
557 }
558
559 // AbslHashValue for hashing tuples
560 template <typename H, typename... Ts>
561 #if defined(_MSC_VER)
562 // This SFINAE gets MSVC confused under some conditions. Let's just disable it
563 // for now.
564 H
565 #else // _MSC_VER
566 typename std::enable_if<absl::conjunction<is_hashable<Ts>...>::value, H>::type
567 #endif // _MSC_VER
568 AbslHashValue(H hash_state, const std::tuple<Ts...>& t) {
569 return hash_internal::hash_tuple(std::move(hash_state), t,
570 absl::make_index_sequence<sizeof...(Ts)>());
571 }
572
573 // -----------------------------------------------------------------------------
574 // AbslHashValue for Pointers
575 // -----------------------------------------------------------------------------
576
577 // AbslHashValue for hashing unique_ptr
578 template <typename H, typename T, typename D>
579 H AbslHashValue(H hash_state, const std::unique_ptr<T, D>& ptr) {
580 return H::combine(std::move(hash_state), ptr.get());
581 }
582
583 // AbslHashValue for hashing shared_ptr
584 template <typename H, typename T>
585 H AbslHashValue(H hash_state, const std::shared_ptr<T>& ptr) {
586 return H::combine(std::move(hash_state), ptr.get());
587 }
588
589 // -----------------------------------------------------------------------------
590 // AbslHashValue for String-Like Types
591 // -----------------------------------------------------------------------------
592
593 // AbslHashValue for hashing strings
594 //
595 // All the string-like types supported here provide the same hash expansion for
596 // the same character sequence. These types are:
597 //
598 // - `absl::Cord`
599 // - `std::string` (and std::basic_string<T, std::char_traits<T>, A> for
600 // any allocator A and any T in {char, wchar_t, char16_t, char32_t})
601 // - `absl::string_view`, `std::string_view`, `std::wstring_view`,
602 // `std::u16string_view`, and `std::u32_string_view`.
603 //
604 // For simplicity, we currently support only strings built on `char`, `wchar_t`,
605 // `char16_t`, or `char32_t`. This support may be broadened, if necessary, but
606 // with some caution - this overload would misbehave in cases where the traits'
607 // `eq()` member isn't equivalent to `==` on the underlying character type.
608 template <typename H>
609 H AbslHashValue(H hash_state, absl::string_view str) {
610 return H::combine(
611 H::combine_contiguous(std::move(hash_state), str.data(), str.size()),
612 str.size());
613 }
614
615 // Support std::wstring, std::u16string and std::u32string.
616 template <typename Char, typename Alloc, typename H,
617 typename = absl::enable_if_t<std::is_same<Char, wchar_t>::value ||
618 std::is_same<Char, char16_t>::value ||
619 std::is_same<Char, char32_t>::value>>
620 H AbslHashValue(
621 H hash_state,
622 const std::basic_string<Char, std::char_traits<Char>, Alloc>& str) {
623 return H::combine(
624 H::combine_contiguous(std::move(hash_state), str.data(), str.size()),
625 str.size());
626 }
627
628 #ifdef ABSL_HAVE_STD_STRING_VIEW
629
630 // Support std::wstring_view, std::u16string_view and std::u32string_view.
631 template <typename Char, typename H,
632 typename = absl::enable_if_t<std::is_same<Char, wchar_t>::value ||
633 std::is_same<Char, char16_t>::value ||
634 std::is_same<Char, char32_t>::value>>
635 H AbslHashValue(H hash_state, std::basic_string_view<Char> str) {
636 return H::combine(
637 H::combine_contiguous(std::move(hash_state), str.data(), str.size()),
638 str.size());
639 }
640
641 #endif // ABSL_HAVE_STD_STRING_VIEW
642
643 #if defined(__cpp_lib_filesystem) && __cpp_lib_filesystem >= 201703L && \
644 (!defined(__ENVIRONMENT_IPHONE_OS_VERSION_MIN_REQUIRED__) || \
645 __ENVIRONMENT_IPHONE_OS_VERSION_MIN_REQUIRED__ >= 130000) && \
646 (!defined(__ENVIRONMENT_MAC_OS_X_VERSION_MIN_REQUIRED__) || \
647 __ENVIRONMENT_MAC_OS_X_VERSION_MIN_REQUIRED__ >= 101500)
648
649 #define ABSL_INTERNAL_STD_FILESYSTEM_PATH_HASH_AVAILABLE 1
650
651 // Support std::filesystem::path. The SFINAE is required because some string
652 // types are implicitly convertible to std::filesystem::path.
653 template <typename Path, typename H,
654 typename = absl::enable_if_t<
655 std::is_same_v<Path, std::filesystem::path>>>
656 H AbslHashValue(H hash_state, const Path& path) {
657 // This is implemented by deferring to the standard library to compute the
658 // hash. The standard library requires that for two paths, `p1 == p2`, then
659 // `hash_value(p1) == hash_value(p2)`. `AbslHashValue` has the same
660 // requirement. Since `operator==` does platform specific matching, deferring
661 // to the standard library is the simplest approach.
662 return H::combine(std::move(hash_state), std::filesystem::hash_value(path));
663 }
664
665 #endif // ABSL_INTERNAL_STD_FILESYSTEM_PATH_HASH_AVAILABLE
666
667 // -----------------------------------------------------------------------------
668 // AbslHashValue for Sequence Containers
669 // -----------------------------------------------------------------------------
670
671 // AbslHashValue for hashing std::array
672 template <typename H, typename T, size_t N>
673 typename std::enable_if<is_hashable<T>::value, H>::type AbslHashValue(
674 H hash_state, const std::array<T, N>& array) {
675 return H::combine_contiguous(std::move(hash_state), array.data(),
676 array.size());
677 }
678
679 // AbslHashValue for hashing std::deque
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::deque<T, Allocator>& deque) {
683 // TODO(gromer): investigate a more efficient implementation taking
684 // advantage of the chunk structure.
685 for (const auto& t : deque) {
686 hash_state = H::combine(std::move(hash_state), t);
687 }
688 return H::combine(std::move(hash_state), deque.size());
689 }
690
691 // AbslHashValue for hashing std::forward_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::forward_list<T, Allocator>& list) {
695 size_t size = 0;
696 for (const T& t : list) {
697 hash_state = H::combine(std::move(hash_state), t);
698 ++size;
699 }
700 return H::combine(std::move(hash_state), size);
701 }
702
703 // AbslHashValue for hashing std::list
704 template <typename H, typename T, typename Allocator>
705 typename std::enable_if<is_hashable<T>::value, H>::type AbslHashValue(
706 H hash_state, const std::list<T, Allocator>& list) {
707 for (const auto& t : list) {
708 hash_state = H::combine(std::move(hash_state), t);
709 }
710 return H::combine(std::move(hash_state), list.size());
711 }
712
713 // AbslHashValue for hashing std::vector
714 //
715 // Do not use this for vector<bool> on platforms that have a working
716 // implementation of std::hash. It does not have a .data(), and a fallback for
717 // std::hash<> is most likely faster.
718 template <typename H, typename T, typename Allocator>
719 typename std::enable_if<is_hashable<T>::value && !std::is_same<T, bool>::value,
720 H>::type
721 AbslHashValue(H hash_state, const std::vector<T, Allocator>& vector) {
722 return H::combine(H::combine_contiguous(std::move(hash_state), vector.data(),
723 vector.size()),
724 vector.size());
725 }
726
727 // AbslHashValue special cases for hashing std::vector<bool>
728
729 #if defined(ABSL_IS_BIG_ENDIAN) && \
730 (defined(__GLIBCXX__) || defined(__GLIBCPP__))
731
732 // std::hash in libstdc++ does not work correctly with vector<bool> on Big
733 // Endian platforms therefore we need to implement a custom AbslHashValue for
734 // it. More details on the bug:
735 // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102531
736 template <typename H, typename T, typename Allocator>
737 typename std::enable_if<is_hashable<T>::value && std::is_same<T, bool>::value,
738 H>::type
739 AbslHashValue(H hash_state, const std::vector<T, Allocator>& vector) {
740 typename H::AbslInternalPiecewiseCombiner combiner;
741 for (const auto& i : vector) {
742 unsigned char c = static_cast<unsigned char>(i);
743 hash_state = combiner.add_buffer(std::move(hash_state), &c, sizeof(c));
744 }
745 return H::combine(combiner.finalize(std::move(hash_state)), vector.size());
746 }
747 #else
748 // When not working around the libstdc++ bug above, we still have to contend
749 // with the fact that std::hash<vector<bool>> is often poor quality, hashing
750 // directly on the internal words and on no other state. On these platforms,
751 // vector<bool>{1, 1} and vector<bool>{1, 1, 0} hash to the same value.
752 //
753 // Mixing in the size (as we do in our other vector<> implementations) on top
754 // of the library-provided hash implementation avoids this QOI issue.
755 template <typename H, typename T, typename Allocator>
756 typename std::enable_if<is_hashable<T>::value && std::is_same<T, bool>::value,
757 H>::type
758 AbslHashValue(H hash_state, const std::vector<T, Allocator>& vector) {
759 return H::combine(std::move(hash_state),
760 std::hash<std::vector<T, Allocator>>{}(vector),
761 vector.size());
762 }
763 #endif
764
765 // -----------------------------------------------------------------------------
766 // AbslHashValue for Ordered Associative Containers
767 // -----------------------------------------------------------------------------
768
769 // AbslHashValue for hashing std::map
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, const std::map<Key, T, Compare, Allocator>& map) {
775 for (const auto& t : map) {
776 hash_state = H::combine(std::move(hash_state), t);
777 }
778 return H::combine(std::move(hash_state), map.size());
779 }
780
781 // AbslHashValue for hashing std::multimap
782 template <typename H, typename Key, typename T, typename Compare,
783 typename Allocator>
784 typename std::enable_if<is_hashable<Key>::value && is_hashable<T>::value,
785 H>::type
786 AbslHashValue(H hash_state,
787 const std::multimap<Key, T, Compare, Allocator>& map) {
788 for (const auto& t : map) {
789 hash_state = H::combine(std::move(hash_state), t);
790 }
791 return H::combine(std::move(hash_state), map.size());
792 }
793
794 // AbslHashValue for hashing std::set
795 template <typename H, typename Key, typename Compare, typename Allocator>
796 typename std::enable_if<is_hashable<Key>::value, H>::type AbslHashValue(
797 H hash_state, const std::set<Key, Compare, Allocator>& set) {
798 for (const auto& t : set) {
799 hash_state = H::combine(std::move(hash_state), t);
800 }
801 return H::combine(std::move(hash_state), set.size());
802 }
803
804 // AbslHashValue for hashing std::multiset
805 template <typename H, typename Key, typename Compare, typename Allocator>
806 typename std::enable_if<is_hashable<Key>::value, H>::type AbslHashValue(
807 H hash_state, const std::multiset<Key, Compare, Allocator>& set) {
808 for (const auto& t : set) {
809 hash_state = H::combine(std::move(hash_state), t);
810 }
811 return H::combine(std::move(hash_state), set.size());
812 }
813
814 // -----------------------------------------------------------------------------
815 // AbslHashValue for Unordered Associative Containers
816 // -----------------------------------------------------------------------------
817
818 // AbslHashValue for hashing std::unordered_set
819 template <typename H, typename Key, typename Hash, typename KeyEqual,
820 typename Alloc>
821 typename std::enable_if<is_hashable<Key>::value, H>::type AbslHashValue(
822 H hash_state, const std::unordered_set<Key, Hash, KeyEqual, Alloc>& s) {
823 return H::combine(
824 H::combine_unordered(std::move(hash_state), s.begin(), s.end()),
825 s.size());
826 }
827
828 // AbslHashValue for hashing std::unordered_multiset
829 template <typename H, typename Key, typename Hash, typename KeyEqual,
830 typename Alloc>
831 typename std::enable_if<is_hashable<Key>::value, H>::type AbslHashValue(
832 H hash_state,
833 const std::unordered_multiset<Key, Hash, KeyEqual, Alloc>& s) {
834 return H::combine(
835 H::combine_unordered(std::move(hash_state), s.begin(), s.end()),
836 s.size());
837 }
838
839 // AbslHashValue for hashing std::unordered_set
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_map<Key, T, Hash, KeyEqual, Alloc>& s) {
846 return H::combine(
847 H::combine_unordered(std::move(hash_state), s.begin(), s.end()),
848 s.size());
849 }
850
851 // AbslHashValue for hashing std::unordered_multiset
852 template <typename H, typename Key, typename T, typename Hash,
853 typename KeyEqual, typename Alloc>
854 typename std::enable_if<is_hashable<Key>::value && is_hashable<T>::value,
855 H>::type
856 AbslHashValue(H hash_state,
857 const std::unordered_multimap<Key, T, Hash, KeyEqual, Alloc>& s) {
858 return H::combine(
859 H::combine_unordered(std::move(hash_state), s.begin(), s.end()),
860 s.size());
861 }
862
863 // -----------------------------------------------------------------------------
864 // AbslHashValue for Wrapper Types
865 // -----------------------------------------------------------------------------
866
867 // AbslHashValue for hashing std::reference_wrapper
868 template <typename H, typename T>
869 typename std::enable_if<is_hashable<T>::value, H>::type AbslHashValue(
870 H hash_state, std::reference_wrapper<T> opt) {
871 return H::combine(std::move(hash_state), opt.get());
872 }
873
874 // AbslHashValue for hashing absl::optional
875 template <typename H, typename T>
876 typename std::enable_if<is_hashable<T>::value, H>::type AbslHashValue(
877 H hash_state, const absl::optional<T>& opt) {
878 if (opt) hash_state = H::combine(std::move(hash_state), *opt);
879 return H::combine(std::move(hash_state), opt.has_value());
880 }
881
882 // VariantVisitor
883 template <typename H>
884 struct VariantVisitor {
885 H&& hash_state;
886 template <typename T>
887 H operator()(const T& t) const {
888 return H::combine(std::move(hash_state), t);
889 }
890 };
891
892 // AbslHashValue for hashing absl::variant
893 template <typename H, typename... T>
894 typename std::enable_if<conjunction<is_hashable<T>...>::value, H>::type
895 AbslHashValue(H hash_state, const absl::variant<T...>& v) {
896 if (!v.valueless_by_exception()) {
897 hash_state = absl::visit(VariantVisitor<H>{std::move(hash_state)}, v);
898 }
899 return H::combine(std::move(hash_state), v.index());
900 }
901
902 // -----------------------------------------------------------------------------
903 // AbslHashValue for Other Types
904 // -----------------------------------------------------------------------------
905
906 // AbslHashValue for hashing std::bitset is not defined on Little Endian
907 // platforms, for the same reason as for vector<bool> (see std::vector above):
908 // It does not expose the raw bytes, and a fallback to std::hash<> is most
909 // likely faster.
910
911 #if defined(ABSL_IS_BIG_ENDIAN) && \
912 (defined(__GLIBCXX__) || defined(__GLIBCPP__))
913 // AbslHashValue for hashing std::bitset
914 //
915 // std::hash in libstdc++ does not work correctly with std::bitset on Big Endian
916 // platforms therefore we need to implement a custom AbslHashValue for it. More
917 // details on the bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102531
918 template <typename H, size_t N>
919 H AbslHashValue(H hash_state, const std::bitset<N>& set) {
920 typename H::AbslInternalPiecewiseCombiner combiner;
921 for (size_t i = 0; i < N; i++) {
922 unsigned char c = static_cast<unsigned char>(set[i]);
923 hash_state = combiner.add_buffer(std::move(hash_state), &c, sizeof(c));
924 }
925 return H::combine(combiner.finalize(std::move(hash_state)), N);
926 }
927 #endif
928
929 // -----------------------------------------------------------------------------
930
931 // hash_range_or_bytes()
932 //
933 // Mixes all values in the range [data, data+size) into the hash state.
934 // This overload accepts only uniquely-represented types, and hashes them by
935 // hashing the entire range of bytes.
936 template <typename H, typename T>
937 typename std::enable_if<is_uniquely_represented<T>::value, H>::type
938 hash_range_or_bytes(H hash_state, const T* data, size_t size) {
939 const auto* bytes = reinterpret_cast<const unsigned char*>(data);
940 return H::combine_contiguous(std::move(hash_state), bytes, sizeof(T) * size);
941 }
942
943 // hash_range_or_bytes()
944 template <typename H, typename T>
945 typename std::enable_if<!is_uniquely_represented<T>::value, H>::type
946 hash_range_or_bytes(H hash_state, const T* data, size_t size) {
947 for (const auto end = data + size; data < end; ++data) {
948 hash_state = H::combine(std::move(hash_state), *data);
949 }
950 return hash_state;
951 }
952
953 #if defined(ABSL_INTERNAL_LEGACY_HASH_NAMESPACE) && \
954 ABSL_META_INTERNAL_STD_HASH_SFINAE_FRIENDLY_
955 #define ABSL_HASH_INTERNAL_SUPPORT_LEGACY_HASH_ 1
956 #else
957 #define ABSL_HASH_INTERNAL_SUPPORT_LEGACY_HASH_ 0
958 #endif
959
960 // HashSelect
961 //
962 // Type trait to select the appropriate hash implementation to use.
963 // HashSelect::type<T> will give the proper hash implementation, to be invoked
964 // as:
965 // HashSelect::type<T>::Invoke(state, value)
966 // Also, HashSelect::type<T>::value is a boolean equal to `true` if there is a
967 // valid `Invoke` function. Types that are not hashable will have a ::value of
968 // `false`.
969 struct HashSelect {
970 private:
971 struct State : HashStateBase<State> {
972 static State combine_contiguous(State hash_state, const unsigned char*,
973 size_t);
974 using State::HashStateBase::combine_contiguous;
975 static State combine_raw(State state, uint64_t value);
976 };
977
978 struct UniquelyRepresentedProbe {
979 template <typename H, typename T>
980 static auto Invoke(H state, const T& value)
981 -> absl::enable_if_t<is_uniquely_represented<T>::value, H> {
982 return hash_internal::hash_bytes(std::move(state), value);
983 }
984 };
985
986 struct HashValueProbe {
987 template <typename H, typename T>
988 static auto Invoke(H state, const T& value) -> absl::enable_if_t<
989 std::is_same<H,
990 decltype(AbslHashValue(std::move(state), value))>::value,
991 H> {
992 return AbslHashValue(std::move(state), value);
993 }
994 };
995
996 struct LegacyHashProbe {
997 #if ABSL_HASH_INTERNAL_SUPPORT_LEGACY_HASH_
998 template <typename H, typename T>
999 static auto Invoke(H state, const T& value) -> absl::enable_if_t<
1000 std::is_convertible<
1001 decltype(ABSL_INTERNAL_LEGACY_HASH_NAMESPACE::hash<T>()(value)),
1002 size_t>::value,
1003 H> {
1004 return hash_internal::hash_bytes(
1005 std::move(state),
1006 ABSL_INTERNAL_LEGACY_HASH_NAMESPACE::hash<T>{}(value));
1007 }
1008 #endif // ABSL_HASH_INTERNAL_SUPPORT_LEGACY_HASH_
1009 };
1010
1011 struct StdHashProbe {
1012 template <typename H, typename T>
1013 static auto Invoke(H state, const T& value)
1014 -> absl::enable_if_t<type_traits_internal::IsHashable<T>::value, H> {
1015 return hash_internal::hash_bytes(std::move(state), std::hash<T>{}(value));
1016 }
1017 };
1018
1019 template <typename Hash, typename T>
1020 struct Probe : Hash {
1021 private:
1022 template <typename H, typename = decltype(H::Invoke(
1023 std::declval<State>(), std::declval<const T&>()))>
1024 static std::true_type Test(int);
1025 template <typename U>
1026 static std::false_type Test(char);
1027
1028 public:
1029 static constexpr bool value = decltype(Test<Hash>(0))::value;
1030 };
1031
1032 public:
1033 // Probe each implementation in order.
1034 // disjunction provides short circuiting wrt instantiation.
1035 template <typename T>
1036 using Apply = absl::disjunction< //
1037 Probe<UniquelyRepresentedProbe, T>, //
1038 Probe<HashValueProbe, T>, //
1039 Probe<LegacyHashProbe, T>, //
1040 Probe<StdHashProbe, T>, //
1041 std::false_type>;
1042 };
1043
1044 template <typename T>
1045 struct is_hashable
1046 : std::integral_constant<bool, HashSelect::template Apply<T>::value> {};
1047
1048 // MixingHashState
1049 class ABSL_DLL MixingHashState : public HashStateBase<MixingHashState> {
1050 // absl::uint128 is not an alias or a thin wrapper around the intrinsic.
1051 // We use the intrinsic when available to improve performance.
1052 #ifdef ABSL_HAVE_INTRINSIC_INT128
1053 using uint128 = __uint128_t;
1054 #else // ABSL_HAVE_INTRINSIC_INT128
1055 using uint128 = absl::uint128;
1056 #endif // ABSL_HAVE_INTRINSIC_INT128
1057
1058 // Random data taken from the hexadecimal digits of Pi's fractional component.
1059 // https://en.wikipedia.org/wiki/Nothing-up-my-sleeve_number
1060 ABSL_CACHELINE_ALIGNED static constexpr uint64_t kStaticRandomData[] = {
1061 0x243f'6a88'85a3'08d3, 0x1319'8a2e'0370'7344, 0xa409'3822'299f'31d0,
1062 0x082e'fa98'ec4e'6c89, 0x4528'21e6'38d0'1377,
1063 };
1064
1065 static constexpr uint64_t kMul =
1066 sizeof(size_t) == 4 ? uint64_t{0xcc9e2d51}
1067 : uint64_t{0xdcb22ca68cb134ed};
1068
1069 template <typename T>
1070 using IntegralFastPath =
1071 conjunction<std::is_integral<T>, is_uniquely_represented<T>,
1072 FitsIn64Bits<T>>;
1073
1074 public:
1075 // Move only
1076 MixingHashState(MixingHashState&&) = default;
1077 MixingHashState& operator=(MixingHashState&&) = default;
1078
1079 // MixingHashState::combine_contiguous()
1080 //
1081 // Fundamental base case for hash recursion: mixes the given range of bytes
1082 // into the hash state.
1083 static MixingHashState combine_contiguous(MixingHashState hash_state,
1084 const unsigned char* first,
1085 size_t size) {
1086 return MixingHashState(
1087 CombineContiguousImpl(hash_state.state_, first, size,
1088 std::integral_constant<int, sizeof(size_t)>{}));
1089 }
1090 using MixingHashState::HashStateBase::combine_contiguous;
1091
1092 // MixingHashState::hash()
1093 //
1094 // For performance reasons in non-opt mode, we specialize this for
1095 // integral types.
1096 // Otherwise we would be instantiating and calling dozens of functions for
1097 // something that is just one multiplication and a couple xor's.
1098 // The result should be the same as running the whole algorithm, but faster.
1099 template <typename T, absl::enable_if_t<IntegralFastPath<T>::value, int> = 0>
1100 static size_t hash(T value) {
1101 return static_cast<size_t>(
1102 WeakMix(Seed() ^ static_cast<std::make_unsigned_t<T>>(value)));
1103 }
1104
1105 // Overload of MixingHashState::hash()
1106 template <typename T, absl::enable_if_t<!IntegralFastPath<T>::value, int> = 0>
1107 static size_t hash(const T& value) {
1108 return static_cast<size_t>(combine(MixingHashState{}, value).state_);
1109 }
1110
1111 private:
1112 // Invoked only once for a given argument; that plus the fact that this is
1113 // move-only ensures that there is only one non-moved-from object.
1114 MixingHashState() : state_(Seed()) {}
1115
1116 friend class MixingHashState::HashStateBase;
1117
1118 template <typename CombinerT>
1119 static MixingHashState RunCombineUnordered(MixingHashState state,
1120 CombinerT combiner) {
1121 uint64_t unordered_state = 0;
1122 combiner(MixingHashState{}, [&](MixingHashState& inner_state) {
1123 // Add the hash state of the element to the running total, but mix the
1124 // carry bit back into the low bit. This in intended to avoid losing
1125 // entropy to overflow, especially when unordered_multisets contain
1126 // multiple copies of the same value.
1127 auto element_state = inner_state.state_;
1128 unordered_state += element_state;
1129 if (unordered_state < element_state) {
1130 ++unordered_state;
1131 }
1132 inner_state = MixingHashState{};
1133 });
1134 return MixingHashState::combine(std::move(state), unordered_state);
1135 }
1136
1137 // Allow the HashState type-erasure implementation to invoke
1138 // RunCombinedUnordered() directly.
1139 friend class absl::HashState;
1140 friend struct CombineRaw;
1141
1142 // Workaround for MSVC bug.
1143 // We make the type copyable to fix the calling convention, even though we
1144 // never actually copy it. Keep it private to not affect the public API of the
1145 // type.
1146 MixingHashState(const MixingHashState&) = default;
1147
1148 explicit MixingHashState(uint64_t state) : state_(state) {}
1149
1150 // Combines a raw value from e.g. integrals/floats/pointers/etc. This allows
1151 // us to be consistent with IntegralFastPath when combining raw types, but
1152 // optimize Read1To3 and Read4To8 differently for the string case.
1153 static MixingHashState combine_raw(MixingHashState hash_state,
1154 uint64_t value) {
1155 return MixingHashState(WeakMix(hash_state.state_ ^ value));
1156 }
1157
1158 // Implementation of the base case for combine_contiguous where we actually
1159 // mix the bytes into the state.
1160 // Dispatch to different implementations of the combine_contiguous depending
1161 // on the value of `sizeof(size_t)`.
1162 static uint64_t CombineContiguousImpl(uint64_t state,
1163 const unsigned char* first, size_t len,
1164 std::integral_constant<int, 4>
1165 /* sizeof_size_t */);
1166 static uint64_t CombineContiguousImpl(uint64_t state,
1167 const unsigned char* first, size_t len,
1168 std::integral_constant<int, 8>
1169 /* sizeof_size_t */);
1170
1171 ABSL_ATTRIBUTE_ALWAYS_INLINE static uint64_t CombineSmallContiguousImpl(
1172 uint64_t state, const unsigned char* first, size_t len) {
1173 ABSL_ASSUME(len <= 8);
1174 uint64_t v;
1175 if (len >= 4) {
1176 v = Read4To8(first, len);
1177 } else if (len > 0) {
1178 v = Read1To3(first, len);
1179 } else {
1180 // Empty ranges have no effect.
1181 return state;
1182 }
1183 return WeakMix(state ^ v);
1184 }
1185
1186 ABSL_ATTRIBUTE_ALWAYS_INLINE static uint64_t CombineContiguousImpl9to16(
1187 uint64_t state, const unsigned char* first, size_t len) {
1188 ABSL_ASSUME(len >= 9);
1189 ABSL_ASSUME(len <= 16);
1190 // Note: any time one half of the mix function becomes zero it will fail to
1191 // incorporate any bits from the other half. However, there is exactly 1 in
1192 // 2^64 values for each side that achieve this, and only when the size is
1193 // exactly 16 -- for smaller sizes there is an overlapping byte that makes
1194 // this impossible unless the seed is *also* incredibly unlucky.
1195 auto p = Read9To16(first, len);
1196 return Mix(state ^ p.first, kMul ^ p.second);
1197 }
1198
1199 ABSL_ATTRIBUTE_ALWAYS_INLINE static uint64_t CombineContiguousImpl17to32(
1200 uint64_t state, const unsigned char* first, size_t len) {
1201 ABSL_ASSUME(len >= 17);
1202 ABSL_ASSUME(len <= 32);
1203 // Do two mixes of overlapping 16-byte ranges in parallel to minimize
1204 // latency.
1205 const uint64_t m0 =
1206 Mix(Read8(first) ^ kStaticRandomData[1], Read8(first + 8) ^ state);
1207
1208 const unsigned char* tail_16b_ptr = first + (len - 16);
1209 const uint64_t m1 = Mix(Read8(tail_16b_ptr) ^ kStaticRandomData[3],
1210 Read8(tail_16b_ptr + 8) ^ state);
1211 return m0 ^ m1;
1212 }
1213
1214 // Slow dispatch path for calls to CombineContiguousImpl with a size argument
1215 // larger than PiecewiseChunkSize(). Has the same effect as calling
1216 // CombineContiguousImpl() repeatedly with the chunk stride size.
1217 static uint64_t CombineLargeContiguousImpl32(uint64_t state,
1218 const unsigned char* first,
1219 size_t len);
1220 static uint64_t CombineLargeContiguousImpl64(uint64_t state,
1221 const unsigned char* first,
1222 size_t len);
1223
1224 // Reads 9 to 16 bytes from p.
1225 // The least significant 8 bytes are in .first, the rest (zero padded) bytes
1226 // are in .second.
1227 static std::pair<uint64_t, uint64_t> Read9To16(const unsigned char* p,
1228 size_t len) {
1229 uint64_t low_mem = Read8(p);
1230 uint64_t high_mem = Read8(p + len - 8);
1231 #ifdef ABSL_IS_LITTLE_ENDIAN
1232 uint64_t most_significant = high_mem;
1233 uint64_t least_significant = low_mem;
1234 #else
1235 uint64_t most_significant = low_mem;
1236 uint64_t least_significant = high_mem;
1237 #endif
1238 return {least_significant, most_significant};
1239 }
1240
1241 // Reads 8 bytes from p.
1242 static uint64_t Read8(const unsigned char* p) {
1243 // Suppress erroneous array bounds errors on GCC.
1244 #if defined(__GNUC__) && !defined(__clang__)
1245 #pragma GCC diagnostic push
1246 #pragma GCC diagnostic ignored "-Warray-bounds"
1247 #endif
1248 return absl::base_internal::UnalignedLoad64(p);
1249 #if defined(__GNUC__) && !defined(__clang__)
1250 #pragma GCC diagnostic pop
1251 #endif
1252 }
1253
1254 // Reads 4 to 8 bytes from p. Zero pads to fill uint64_t.
1255 // TODO(b/384509507): consider optimizing this by not requiring the output to
1256 // be equivalent to an integer load for 4/8 bytes. Currently, we rely on this
1257 // behavior for the HashConsistentAcrossIntTypes test case. Ditto for
1258 // Read1To3.
1259 static uint64_t Read4To8(const unsigned char* p, size_t len) {
1260 // If `len < 8`, we duplicate bytes in the middle.
1261 // E.g.:
1262 // `ABCD` will be read as `ABCDABCD`.
1263 // `ABCDE` will be read as `ABCDBCDE`.
1264 // `ABCDEF` will be read as `ABCDCDEF`.
1265 // `ABCDEFG` will be read as `ABCDDEFG`.
1266 // We also do not care about endianness. On big-endian platforms, bytes will
1267 // be shuffled (it's fine). We always shift low memory by 32, because that
1268 // can be pipelined earlier. Reading high memory requires computing
1269 // `p + len - 4`.
1270 uint64_t most_significant =
1271 static_cast<uint64_t>(absl::base_internal::UnalignedLoad32(p)) << 32;
1272 uint64_t least_significant =
1273 absl::base_internal::UnalignedLoad32(p + len - 4);
1274 return most_significant | least_significant;
1275 }
1276
1277 // Reads 1 to 3 bytes from p. Zero pads to fill uint32_t.
1278 static uint32_t Read1To3(const unsigned char* p, size_t len) {
1279 // The trick used by this implementation is to avoid branches.
1280 // We always read three bytes by duplicating.
1281 // E.g.,
1282 // `A` is read as `AAA`.
1283 // `AB` is read as `ABB`.
1284 // `ABC` is read as `ABC`.
1285 // We always shift `p[0]` so that it can be pipelined better.
1286 // Other bytes require extra computation to find indices.
1287 uint32_t mem0 = (static_cast<uint32_t>(p[0]) << 16) | p[len - 1];
1288 uint32_t mem1 = static_cast<uint32_t>(p[len / 2]) << 8;
1289 return mem0 | mem1;
1290 }
1291
1292 ABSL_ATTRIBUTE_ALWAYS_INLINE static uint64_t Mix(uint64_t lhs, uint64_t rhs) {
1293 // Though the 128-bit product on AArch64 needs two instructions, it is
1294 // still a good balance between speed and hash quality.
1295 using MultType =
1296 absl::conditional_t<sizeof(size_t) == 4, uint64_t, uint128>;
1297 MultType m = lhs;
1298 m *= rhs;
1299 return static_cast<uint64_t>(m ^ (m >> (sizeof(m) * 8 / 2)));
1300 }
1301
1302 // Slightly lower latency than Mix, but with lower quality. The byte swap
1303 // helps ensure that low bits still have high quality.
1304 ABSL_ATTRIBUTE_ALWAYS_INLINE static uint64_t WeakMix(uint64_t n) {
1305 // WeakMix doesn't work well on 32-bit platforms so just use Mix.
1306 if (sizeof(size_t) < 8) return Mix(n, kMul);
1307 #ifdef __ARM_ACLE
1308 // gbswap_64 compiles to `rev` on ARM, but `rbit` is better because it
1309 // reverses bits rather than reversing bytes.
1310 return __rbitll(n * kMul);
1311 #else
1312 return absl::gbswap_64(n * kMul);
1313 #endif
1314 }
1315
1316 // An extern to avoid bloat on a direct call to LowLevelHash() with fixed
1317 // values for both the seed and salt parameters.
1318 static uint64_t LowLevelHashImpl(const unsigned char* data, size_t len);
1319
1320 ABSL_ATTRIBUTE_ALWAYS_INLINE static uint64_t Hash64(const unsigned char* data,
1321 size_t len) {
1322 #ifdef ABSL_HAVE_INTRINSIC_INT128
1323 return LowLevelHashImpl(data, len);
1324 #else
1325 return hash_internal::CityHash64(reinterpret_cast<const char*>(data), len);
1326 #endif
1327 }
1328
1329 // Seed()
1330 //
1331 // A non-deterministic seed.
1332 //
1333 // The current purpose of this seed is to generate non-deterministic results
1334 // and prevent having users depend on the particular hash values.
1335 // It is not meant as a security feature right now, but it leaves the door
1336 // open to upgrade it to a true per-process random seed. A true random seed
1337 // costs more and we don't need to pay for that right now.
1338 //
1339 // On platforms with ASLR, we take advantage of it to make a per-process
1340 // random value.
1341 // See https://en.wikipedia.org/wiki/Address_space_layout_randomization
1342 //
1343 // On other platforms this is still going to be non-deterministic but most
1344 // probably per-build and not per-process.
1345 ABSL_ATTRIBUTE_ALWAYS_INLINE static uint64_t Seed() {
1346 #if (!defined(__clang__) || __clang_major__ > 11) && \
1347 (!defined(__apple_build_version__) || \
1348 __apple_build_version__ >= 19558921) // Xcode 12
1349 return static_cast<uint64_t>(reinterpret_cast<uintptr_t>(&kSeed));
1350 #else
1351 // Workaround the absence of
1352 // https://github.com/llvm/llvm-project/commit/bc15bf66dcca76cc06fe71fca35b74dc4d521021.
1353 return static_cast<uint64_t>(reinterpret_cast<uintptr_t>(kSeed));
1354 #endif
1355 }
1356 static const void* const kSeed;
1357
1358 uint64_t state_;
1359 };
1360
1361 // MixingHashState::CombineContiguousImpl()
1362 inline uint64_t MixingHashState::CombineContiguousImpl(
1363 uint64_t state, const unsigned char* first, size_t len,
1364 std::integral_constant<int, 4> /* sizeof_size_t */) {
1365 // For large values we use CityHash, for small ones we just use a
1366 // multiplicative hash.
1367 if (len <= 8) {
1368 return CombineSmallContiguousImpl(state, first, len);
1369 }
1370 if (ABSL_PREDICT_TRUE(len <= PiecewiseChunkSize())) {
1371 return Mix(state ^ hash_internal::CityHash32(
1372 reinterpret_cast<const char*>(first), len),
1373 kMul);
1374 }
1375 return CombineLargeContiguousImpl32(state, first, len);
1376 }
1377
1378 // Overload of MixingHashState::CombineContiguousImpl()
1379 inline uint64_t MixingHashState::CombineContiguousImpl(
1380 uint64_t state, const unsigned char* first, size_t len,
1381 std::integral_constant<int, 8> /* sizeof_size_t */) {
1382 // For large values we use LowLevelHash or CityHash depending on the platform,
1383 // for small ones we just use a multiplicative hash.
1384 if (len <= 8) {
1385 return CombineSmallContiguousImpl(state, first, len);
1386 }
1387 if (len <= 16) {
1388 return CombineContiguousImpl9to16(state, first, len);
1389 }
1390 if (len <= 32) {
1391 return CombineContiguousImpl17to32(state, first, len);
1392 }
1393 if (ABSL_PREDICT_TRUE(len <= PiecewiseChunkSize())) {
1394 return Mix(state ^ Hash64(first, len), kMul);
1395 }
1396 return CombineLargeContiguousImpl64(state, first, len);
1397 }
1398
1399 struct AggregateBarrier {};
1400
1401 // HashImpl
1402
1403 // Add a private base class to make sure this type is not an aggregate.
1404 // Aggregates can be aggregate initialized even if the default constructor is
1405 // deleted.
1406 struct PoisonedHash : private AggregateBarrier {
1407 PoisonedHash() = delete;
1408 PoisonedHash(const PoisonedHash&) = delete;
1409 PoisonedHash& operator=(const PoisonedHash&) = delete;
1410 };
1411
1412 template <typename T>
1413 struct HashImpl {
1414 size_t operator()(const T& value) const {
1415 return MixingHashState::hash(value);
1416 }
1417 };
1418
1419 template <typename T>
1420 struct Hash
1421 : absl::conditional_t<is_hashable<T>::value, HashImpl<T>, PoisonedHash> {};
1422
1423 template <typename H>
1424 template <typename T, typename... Ts>
1425 H HashStateBase<H>::combine(H state, const T& value, const Ts&... values) {
1426 return H::combine(hash_internal::HashSelect::template Apply<T>::Invoke(
1427 std::move(state), value),
1428 values...);
1429 }
1430
1431 // HashStateBase::combine_contiguous()
1432 template <typename H>
1433 template <typename T>
1434 H HashStateBase<H>::combine_contiguous(H state, const T* data, size_t size) {
1435 return hash_internal::hash_range_or_bytes(std::move(state), data, size);
1436 }
1437
1438 // HashStateBase::combine_unordered()
1439 template <typename H>
1440 template <typename I>
1441 H HashStateBase<H>::combine_unordered(H state, I begin, I end) {
1442 return H::RunCombineUnordered(std::move(state),
1443 CombineUnorderedCallback<I>{begin, end});
1444 }
1445
1446 // HashStateBase::PiecewiseCombiner::add_buffer()
1447 template <typename H>
1448 H PiecewiseCombiner::add_buffer(H state, const unsigned char* data,
1449 size_t size) {
1450 if (position_ + size < PiecewiseChunkSize()) {
1451 // This partial chunk does not fill our existing buffer
1452 memcpy(buf_ + position_, data, size);
1453 position_ += size;
1454 return state;
1455 }
1456
1457 // If the buffer is partially filled we need to complete the buffer
1458 // and hash it.
1459 if (position_ != 0) {
1460 const size_t bytes_needed = PiecewiseChunkSize() - position_;
1461 memcpy(buf_ + position_, data, bytes_needed);
1462 state = H::combine_contiguous(std::move(state), buf_, PiecewiseChunkSize());
1463 data += bytes_needed;
1464 size -= bytes_needed;
1465 }
1466
1467 // Hash whatever chunks we can without copying
1468 while (size >= PiecewiseChunkSize()) {
1469 state = H::combine_contiguous(std::move(state), data, PiecewiseChunkSize());
1470 data += PiecewiseChunkSize();
1471 size -= PiecewiseChunkSize();
1472 }
1473 // Fill the buffer with the remainder
1474 memcpy(buf_, data, size);
1475 position_ = size;
1476 return state;
1477 }
1478
1479 // HashStateBase::PiecewiseCombiner::finalize()
1480 template <typename H>
1481 H PiecewiseCombiner::finalize(H state) {
1482 // Hash the remainder left in the buffer, which may be empty
1483 return H::combine_contiguous(std::move(state), buf_, position_);
1484 }
1485
1486 } // namespace hash_internal
1487 ABSL_NAMESPACE_END
1488 } // namespace absl
1489
1490 #endif // ABSL_HASH_INTERNAL_HASH_H_
1491