1//===-- llvm/ADT/Hashing.h - Utilities for hashing --------------*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements the newly proposed standard C++ interfaces for hashing
10// arbitrary data and building hash functions for user-defined types. This
11// interface was originally proposed in N3333[1] and is currently under review
12// for inclusion in a future TR and/or standard.
13//
14// The primary interfaces provide are comprised of one type and three functions:
15//
16// -- 'hash_code' class is an opaque type representing the hash code for some
17// data. It is the intended product of hashing, and can be used to implement
18// hash tables, checksumming, and other common uses of hashes. It is not an
19// integer type (although it can be converted to one) because it is risky
20// to assume much about the internals of a hash_code. In particular, each
21// execution of the program has a high probability of producing a different
22// hash_code for a given input. Thus their values are not stable to save or
23// persist, and should only be used during the execution for the
24// construction of hashing datastructures.
25//
26// -- 'hash_value' is a function designed to be overloaded for each
27// user-defined type which wishes to be used within a hashing context. It
28// should be overloaded within the user-defined type's namespace and found
29// via ADL. Overloads for primitive types are provided by this library.
30//
31// -- 'hash_combine' and 'hash_combine_range' are functions designed to aid
32// programmers in easily and intuitively combining a set of data into
33// a single hash_code for their object. They should only logically be used
34// within the implementation of a 'hash_value' routine or similar context.
35//
36// Note that 'hash_combine_range' contains very special logic for hashing
37// a contiguous array of integers or pointers. This logic is *extremely* fast,
38// on a modern Intel "Gainestown" Xeon (Nehalem uarch) @2.2 GHz, these were
39// benchmarked at over 6.5 GiB/s for large keys, and <20 cycles/hash for keys
40// under 32-bytes.
41//
42//===----------------------------------------------------------------------===//
43
44#ifndef LLVM_ADT_HASHING_H
45#define LLVM_ADT_HASHING_H
46
47#include "llvm/Config/abi-breaking.h"
48#include "llvm/Support/DataTypes.h"
49#include "llvm/Support/ErrorHandling.h"
50#include "llvm/Support/SwapByteOrder.h"
51#include "llvm/Support/type_traits.h"
52#include <algorithm>
53#include <cassert>
54#include <cstring>
55#include <optional>
56#include <string>
57#include <tuple>
58#include <utility>
59
60namespace llvm {
61template <typename T, typename Enable> struct DenseMapInfo;
62
63/// An opaque object representing a hash code.
64///
65/// This object represents the result of hashing some entity. It is intended to
66/// be used to implement hashtables or other hashing-based data structures.
67/// While it wraps and exposes a numeric value, this value should not be
68/// trusted to be stable or predictable across processes or executions.
69///
70/// In order to obtain the hash_code for an object 'x':
71/// \code
72/// using llvm::hash_value;
73/// llvm::hash_code code = hash_value(x);
74/// \endcode
75class hash_code {
76 size_t value;
77
78public:
79 /// Default construct a hash_code.
80 /// Note that this leaves the value uninitialized.
81 hash_code() = default;
82
83 /// Form a hash code directly from a numerical value.
84 hash_code(size_t value) : value(value) {}
85
86 /// Convert the hash code to its numerical value for use.
87 /*explicit*/ operator size_t() const { return value; }
88
89 friend bool operator==(const hash_code &lhs, const hash_code &rhs) {
90 return lhs.value == rhs.value;
91 }
92 friend bool operator!=(const hash_code &lhs, const hash_code &rhs) {
93 return lhs.value != rhs.value;
94 }
95
96 /// Allow a hash_code to be directly run through hash_value.
97 friend size_t hash_value(const hash_code &code) { return code.value; }
98};
99
100/// Compute a hash_code for any integer value.
101///
102/// Note that this function is intended to compute the same hash_code for
103/// a particular value without regard to the pre-promotion type. This is in
104/// contrast to hash_combine which may produce different hash_codes for
105/// differing argument types even if they would implicit promote to a common
106/// type without changing the value.
107template <typename T>
108std::enable_if_t<is_integral_or_enum<T>::value, hash_code> hash_value(T value);
109
110/// Compute a hash_code for a pointer's address.
111///
112/// N.B.: This hashes the *address*. Not the value and not the type.
113template <typename T> hash_code hash_value(const T *ptr);
114
115/// Compute a hash_code for a pair of objects.
116template <typename T, typename U>
117hash_code hash_value(const std::pair<T, U> &arg);
118
119/// Compute a hash_code for a tuple.
120template <typename... Ts>
121hash_code hash_value(const std::tuple<Ts...> &arg);
122
123/// Compute a hash_code for a standard string.
124template <typename T>
125hash_code hash_value(const std::basic_string<T> &arg);
126
127/// Compute a hash_code for a standard string.
128template <typename T> hash_code hash_value(const std::optional<T> &arg);
129
130// All of the implementation details of actually computing the various hash
131// code values are held within this namespace. These routines are included in
132// the header file mainly to allow inlining and constant propagation.
133namespace hashing {
134namespace detail {
135
136inline uint64_t fetch64(const char *p) {
137 uint64_t result;
138 memcpy(dest: &result, src: p, n: sizeof(result));
139 if (sys::IsBigEndianHost)
140 sys::swapByteOrder(Value&: result);
141 return result;
142}
143
144inline uint32_t fetch32(const char *p) {
145 uint32_t result;
146 memcpy(dest: &result, src: p, n: sizeof(result));
147 if (sys::IsBigEndianHost)
148 sys::swapByteOrder(Value&: result);
149 return result;
150}
151
152/// Some primes between 2^63 and 2^64 for various uses.
153static constexpr uint64_t k0 = 0xc3a5c85c97cb3127ULL;
154static constexpr uint64_t k1 = 0xb492b66fbe98f273ULL;
155static constexpr uint64_t k2 = 0x9ae16a3b2f90404fULL;
156static constexpr uint64_t k3 = 0xc949d7c7509e6557ULL;
157
158/// Bitwise right rotate.
159/// Normally this will compile to a single instruction, especially if the
160/// shift is a manifest constant.
161inline uint64_t rotate(uint64_t val, size_t shift) {
162 // Avoid shifting by 64: doing so yields an undefined result.
163 return shift == 0 ? val : ((val >> shift) | (val << (64 - shift)));
164}
165
166inline uint64_t shift_mix(uint64_t val) {
167 return val ^ (val >> 47);
168}
169
170inline uint64_t hash_16_bytes(uint64_t low, uint64_t high) {
171 // Murmur-inspired hashing.
172 const uint64_t kMul = 0x9ddfea08eb382d69ULL;
173 uint64_t a = (low ^ high) * kMul;
174 a ^= (a >> 47);
175 uint64_t b = (high ^ a) * kMul;
176 b ^= (b >> 47);
177 b *= kMul;
178 return b;
179}
180
181inline uint64_t hash_1to3_bytes(const char *s, size_t len, uint64_t seed) {
182 uint8_t a = s[0];
183 uint8_t b = s[len >> 1];
184 uint8_t c = s[len - 1];
185 uint32_t y = static_cast<uint32_t>(a) + (static_cast<uint32_t>(b) << 8);
186 uint32_t z = static_cast<uint32_t>(len) + (static_cast<uint32_t>(c) << 2);
187 return shift_mix(val: y * k2 ^ z * k3 ^ seed) * k2;
188}
189
190inline uint64_t hash_4to8_bytes(const char *s, size_t len, uint64_t seed) {
191 uint64_t a = fetch32(p: s);
192 return hash_16_bytes(low: len + (a << 3), high: seed ^ fetch32(p: s + len - 4));
193}
194
195inline uint64_t hash_9to16_bytes(const char *s, size_t len, uint64_t seed) {
196 uint64_t a = fetch64(p: s);
197 uint64_t b = fetch64(p: s + len - 8);
198 return hash_16_bytes(low: seed ^ a, high: rotate(val: b + len, shift: len)) ^ b;
199}
200
201inline uint64_t hash_17to32_bytes(const char *s, size_t len, uint64_t seed) {
202 uint64_t a = fetch64(p: s) * k1;
203 uint64_t b = fetch64(p: s + 8);
204 uint64_t c = fetch64(p: s + len - 8) * k2;
205 uint64_t d = fetch64(p: s + len - 16) * k0;
206 return hash_16_bytes(low: llvm::rotr<uint64_t>(V: a - b, R: 43) +
207 llvm::rotr<uint64_t>(V: c ^ seed, R: 30) + d,
208 high: a + llvm::rotr<uint64_t>(V: b ^ k3, R: 20) - c + len + seed);
209}
210
211inline uint64_t hash_33to64_bytes(const char *s, size_t len, uint64_t seed) {
212 uint64_t z = fetch64(p: s + 24);
213 uint64_t a = fetch64(p: s) + (len + fetch64(p: s + len - 16)) * k0;
214 uint64_t b = llvm::rotr<uint64_t>(V: a + z, R: 52);
215 uint64_t c = llvm::rotr<uint64_t>(V: a, R: 37);
216 a += fetch64(p: s + 8);
217 c += llvm::rotr<uint64_t>(V: a, R: 7);
218 a += fetch64(p: s + 16);
219 uint64_t vf = a + z;
220 uint64_t vs = b + llvm::rotr<uint64_t>(V: a, R: 31) + c;
221 a = fetch64(p: s + 16) + fetch64(p: s + len - 32);
222 z = fetch64(p: s + len - 8);
223 b = llvm::rotr<uint64_t>(V: a + z, R: 52);
224 c = llvm::rotr<uint64_t>(V: a, R: 37);
225 a += fetch64(p: s + len - 24);
226 c += llvm::rotr<uint64_t>(V: a, R: 7);
227 a += fetch64(p: s + len - 16);
228 uint64_t wf = a + z;
229 uint64_t ws = b + llvm::rotr<uint64_t>(V: a, R: 31) + c;
230 uint64_t r = shift_mix(val: (vf + ws) * k2 + (wf + vs) * k0);
231 return shift_mix(val: (seed ^ (r * k0)) + vs) * k2;
232}
233
234inline uint64_t hash_short(const char *s, size_t length, uint64_t seed) {
235 if (length >= 4 && length <= 8)
236 return hash_4to8_bytes(s, len: length, seed);
237 if (length > 8 && length <= 16)
238 return hash_9to16_bytes(s, len: length, seed);
239 if (length > 16 && length <= 32)
240 return hash_17to32_bytes(s, len: length, seed);
241 if (length > 32)
242 return hash_33to64_bytes(s, len: length, seed);
243 if (length != 0)
244 return hash_1to3_bytes(s, len: length, seed);
245
246 return k2 ^ seed;
247}
248
249/// The intermediate state used during hashing.
250/// Currently, the algorithm for computing hash codes is based on CityHash and
251/// keeps 56 bytes of arbitrary state.
252struct hash_state {
253 uint64_t h0 = 0, h1 = 0, h2 = 0, h3 = 0, h4 = 0, h5 = 0, h6 = 0;
254
255 /// Create a new hash_state structure and initialize it based on the
256 /// seed and the first 64-byte chunk.
257 /// This effectively performs the initial mix.
258 static hash_state create(const char *s, uint64_t seed) {
259 hash_state state = {.h0: 0,
260 .h1: seed,
261 .h2: hash_16_bytes(low: seed, high: k1),
262 .h3: llvm::rotr<uint64_t>(V: seed ^ k1, R: 49),
263 .h4: seed * k1,
264 .h5: shift_mix(val: seed),
265 .h6: 0};
266 state.h6 = hash_16_bytes(low: state.h4, high: state.h5);
267 state.mix(s);
268 return state;
269 }
270
271 /// Mix 32-bytes from the input sequence into the 16-bytes of 'a'
272 /// and 'b', including whatever is already in 'a' and 'b'.
273 static void mix_32_bytes(const char *s, uint64_t &a, uint64_t &b) {
274 a += fetch64(p: s);
275 uint64_t c = fetch64(p: s + 24);
276 b = llvm::rotr<uint64_t>(V: b + a + c, R: 21);
277 uint64_t d = a;
278 a += fetch64(p: s + 8) + fetch64(p: s + 16);
279 b += llvm::rotr<uint64_t>(V: a, R: 44) + d;
280 a += c;
281 }
282
283 /// Mix in a 64-byte buffer of data.
284 /// We mix all 64 bytes even when the chunk length is smaller, but we
285 /// record the actual length.
286 void mix(const char *s) {
287 h0 = llvm::rotr<uint64_t>(V: h0 + h1 + h3 + fetch64(p: s + 8), R: 37) * k1;
288 h1 = llvm::rotr<uint64_t>(V: h1 + h4 + fetch64(p: s + 48), R: 42) * k1;
289 h0 ^= h6;
290 h1 += h3 + fetch64(p: s + 40);
291 h2 = llvm::rotr<uint64_t>(V: h2 + h5, R: 33) * k1;
292 h3 = h4 * k1;
293 h4 = h0 + h5;
294 mix_32_bytes(s, a&: h3, b&: h4);
295 h5 = h2 + h6;
296 h6 = h1 + fetch64(p: s + 16);
297 mix_32_bytes(s: s + 32, a&: h5, b&: h6);
298 std::swap(a&: h2, b&: h0);
299 }
300
301 /// Compute the final 64-bit hash code value based on the current
302 /// state and the length of bytes hashed.
303 uint64_t finalize(size_t length) {
304 return hash_16_bytes(low: hash_16_bytes(low: h3, high: h5) + shift_mix(val: h1) * k1 + h2,
305 high: hash_16_bytes(low: h4, high: h6) + shift_mix(val: length) * k1 + h0);
306 }
307};
308
309/// In LLVM_ENABLE_ABI_BREAKING_CHECKS builds, the seed is non-deterministic
310/// per process (address of a function in LLVMSupport) to prevent having users
311/// depend on the particular hash values. On platforms without ASLR, this is
312/// still likely non-deterministic per build.
313inline uint64_t get_execution_seed() {
314#if LLVM_ENABLE_ABI_BREAKING_CHECKS
315 return static_cast<uint64_t>(
316 reinterpret_cast<uintptr_t>(&install_fatal_error_handler));
317#else
318 return 0xff51afd7ed558ccdULL;
319#endif
320}
321
322
323/// Trait to indicate whether a type's bits can be hashed directly.
324///
325/// A type trait which is true if we want to combine values for hashing by
326/// reading the underlying data. It is false if values of this type must
327/// first be passed to hash_value, and the resulting hash_codes combined.
328//
329// FIXME: We want to replace is_integral_or_enum and is_pointer here with
330// a predicate which asserts that comparing the underlying storage of two
331// values of the type for equality is equivalent to comparing the two values
332// for equality. For all the platforms we care about, this holds for integers
333// and pointers, but there are platforms where it doesn't and we would like to
334// support user-defined types which happen to satisfy this property.
335template <typename T> struct is_hashable_data
336 : std::integral_constant<bool, ((is_integral_or_enum<T>::value ||
337 std::is_pointer<T>::value) &&
338 64 % sizeof(T) == 0)> {};
339
340// Special case std::pair to detect when both types are viable and when there
341// is no alignment-derived padding in the pair. This is a bit of a lie because
342// std::pair isn't truly POD, but it's close enough in all reasonable
343// implementations for our use case of hashing the underlying data.
344template <typename T, typename U> struct is_hashable_data<std::pair<T, U> >
345 : std::integral_constant<bool, (is_hashable_data<T>::value &&
346 is_hashable_data<U>::value &&
347 (sizeof(T) + sizeof(U)) ==
348 sizeof(std::pair<T, U>))> {};
349
350/// Helper to get the hashable data representation for a type.
351/// This variant is enabled when the type itself can be used.
352template <typename T>
353std::enable_if_t<is_hashable_data<T>::value, T>
354get_hashable_data(const T &value) {
355 return value;
356}
357/// Helper to get the hashable data representation for a type.
358/// This variant is enabled when we must first call hash_value and use the
359/// result as our data.
360template <typename T>
361std::enable_if_t<!is_hashable_data<T>::value, size_t>
362get_hashable_data(const T &value) {
363 using ::llvm::hash_value;
364 return hash_value(value);
365}
366
367/// Helper to store data from a value into a buffer and advance the
368/// pointer into that buffer.
369///
370/// This routine first checks whether there is enough space in the provided
371/// buffer, and if not immediately returns false. If there is space, it
372/// copies the underlying bytes of value into the buffer, advances the
373/// buffer_ptr past the copied bytes, and returns true.
374template <typename T>
375bool store_and_advance(char *&buffer_ptr, char *buffer_end, const T& value,
376 size_t offset = 0) {
377 size_t store_size = sizeof(value) - offset;
378 if (buffer_ptr + store_size > buffer_end)
379 return false;
380 const char *value_data = reinterpret_cast<const char *>(&value);
381 memcpy(dest: buffer_ptr, src: value_data + offset, n: store_size);
382 buffer_ptr += store_size;
383 return true;
384}
385
386/// Implement the combining of integral values into a hash_code.
387///
388/// This overload is selected when the value type of the iterator is
389/// integral. Rather than computing a hash_code for each object and then
390/// combining them, this (as an optimization) directly combines the integers.
391template <typename InputIteratorT>
392hash_code hash_combine_range_impl(InputIteratorT first, InputIteratorT last) {
393 const uint64_t seed = get_execution_seed();
394 char buffer[64], *buffer_ptr = buffer;
395 char *const buffer_end = std::end(arr&: buffer);
396 while (first != last && store_and_advance(buffer_ptr, buffer_end,
397 get_hashable_data(*first)))
398 ++first;
399 if (first == last)
400 return hash_short(s: buffer, length: buffer_ptr - buffer, seed);
401 assert(buffer_ptr == buffer_end);
402
403 hash_state state = state.create(s: buffer, seed);
404 size_t length = 64;
405 while (first != last) {
406 // Fill up the buffer. We don't clear it, which re-mixes the last round
407 // when only a partial 64-byte chunk is left.
408 buffer_ptr = buffer;
409 while (first != last && store_and_advance(buffer_ptr, buffer_end,
410 get_hashable_data(*first)))
411 ++first;
412
413 // Rotate the buffer if we did a partial fill in order to simulate doing
414 // a mix of the last 64-bytes. That is how the algorithm works when we
415 // have a contiguous byte sequence, and we want to emulate that here.
416 std::rotate(first: buffer, middle: buffer_ptr, last: buffer_end);
417
418 // Mix this chunk into the current state.
419 state.mix(s: buffer);
420 length += buffer_ptr - buffer;
421 };
422
423 return state.finalize(length);
424}
425
426/// Implement the combining of integral values into a hash_code.
427///
428/// This overload is selected when the value type of the iterator is integral
429/// and when the input iterator is actually a pointer. Rather than computing
430/// a hash_code for each object and then combining them, this (as an
431/// optimization) directly combines the integers. Also, because the integers
432/// are stored in contiguous memory, this routine avoids copying each value
433/// and directly reads from the underlying memory.
434template <typename ValueT>
435std::enable_if_t<is_hashable_data<ValueT>::value, hash_code>
436hash_combine_range_impl(ValueT *first, ValueT *last) {
437 const uint64_t seed = get_execution_seed();
438 const char *s_begin = reinterpret_cast<const char *>(first);
439 const char *s_end = reinterpret_cast<const char *>(last);
440 const size_t length = std::distance(first: s_begin, last: s_end);
441 if (length <= 64)
442 return hash_short(s: s_begin, length, seed);
443
444 const char *s_aligned_end = s_begin + (length & ~63);
445 hash_state state = state.create(s: s_begin, seed);
446 s_begin += 64;
447 while (s_begin != s_aligned_end) {
448 state.mix(s: s_begin);
449 s_begin += 64;
450 }
451 if (length & 63)
452 state.mix(s: s_end - 64);
453
454 return state.finalize(length);
455}
456
457} // namespace detail
458} // namespace hashing
459
460
461/// Compute a hash_code for a sequence of values.
462///
463/// This hashes a sequence of values. It produces the same hash_code as
464/// 'hash_combine(a, b, c, ...)', but can run over arbitrary sized sequences
465/// and is significantly faster given pointers and types which can be hashed as
466/// a sequence of bytes.
467template <typename InputIteratorT>
468hash_code hash_combine_range(InputIteratorT first, InputIteratorT last) {
469 return ::llvm::hashing::detail::hash_combine_range_impl(first, last);
470}
471
472
473// Implementation details for hash_combine.
474namespace hashing {
475namespace detail {
476
477/// Helper class to manage the recursive combining of hash_combine
478/// arguments.
479///
480/// This class exists to manage the state and various calls involved in the
481/// recursive combining of arguments used in hash_combine. It is particularly
482/// useful at minimizing the code in the recursive calls to ease the pain
483/// caused by a lack of variadic functions.
484struct hash_combine_recursive_helper {
485 char buffer[64] = {};
486 hash_state state;
487 const uint64_t seed;
488
489public:
490 /// Construct a recursive hash combining helper.
491 ///
492 /// This sets up the state for a recursive hash combine, including getting
493 /// the seed and buffer setup.
494 hash_combine_recursive_helper()
495 : seed(get_execution_seed()) {}
496
497 /// Combine one chunk of data into the current in-flight hash.
498 ///
499 /// This merges one chunk of data into the hash. First it tries to buffer
500 /// the data. If the buffer is full, it hashes the buffer into its
501 /// hash_state, empties it, and then merges the new chunk in. This also
502 /// handles cases where the data straddles the end of the buffer.
503 template <typename T>
504 char *combine_data(size_t &length, char *buffer_ptr, char *buffer_end, T data) {
505 if (!store_and_advance(buffer_ptr, buffer_end, data)) {
506 // Check for skew which prevents the buffer from being packed, and do
507 // a partial store into the buffer to fill it. This is only a concern
508 // with the variadic combine because that formation can have varying
509 // argument types.
510 size_t partial_store_size = buffer_end - buffer_ptr;
511 memcpy(buffer_ptr, &data, partial_store_size);
512
513 // If the store fails, our buffer is full and ready to hash. We have to
514 // either initialize the hash state (on the first full buffer) or mix
515 // this buffer into the existing hash state. Length tracks the *hashed*
516 // length, not the buffered length.
517 if (length == 0) {
518 state = state.create(s: buffer, seed);
519 length = 64;
520 } else {
521 // Mix this chunk into the current state and bump length up by 64.
522 state.mix(s: buffer);
523 length += 64;
524 }
525 // Reset the buffer_ptr to the head of the buffer for the next chunk of
526 // data.
527 buffer_ptr = buffer;
528
529 // Try again to store into the buffer -- this cannot fail as we only
530 // store types smaller than the buffer.
531 if (!store_and_advance(buffer_ptr, buffer_end, data,
532 partial_store_size))
533 llvm_unreachable("buffer smaller than stored type");
534 }
535 return buffer_ptr;
536 }
537
538 /// Recursive, variadic combining method.
539 ///
540 /// This function recurses through each argument, combining that argument
541 /// into a single hash.
542 template <typename T, typename ...Ts>
543 hash_code combine(size_t length, char *buffer_ptr, char *buffer_end,
544 const T &arg, const Ts &...args) {
545 buffer_ptr = combine_data(length, buffer_ptr, buffer_end, get_hashable_data(arg));
546
547 // Recurse to the next argument.
548 return combine(length, buffer_ptr, buffer_end, args...);
549 }
550
551 /// Base case for recursive, variadic combining.
552 ///
553 /// The base case when combining arguments recursively is reached when all
554 /// arguments have been handled. It flushes the remaining buffer and
555 /// constructs a hash_code.
556 hash_code combine(size_t length, char *buffer_ptr, char *buffer_end) {
557 // Check whether the entire set of values fit in the buffer. If so, we'll
558 // use the optimized short hashing routine and skip state entirely.
559 if (length == 0)
560 return hash_short(s: buffer, length: buffer_ptr - buffer, seed);
561
562 // Mix the final buffer, rotating it if we did a partial fill in order to
563 // simulate doing a mix of the last 64-bytes. That is how the algorithm
564 // works when we have a contiguous byte sequence, and we want to emulate
565 // that here.
566 std::rotate(first: buffer, middle: buffer_ptr, last: buffer_end);
567
568 // Mix this chunk into the current state.
569 state.mix(s: buffer);
570 length += buffer_ptr - buffer;
571
572 return state.finalize(length);
573 }
574};
575
576} // namespace detail
577} // namespace hashing
578
579/// Combine values into a single hash_code.
580///
581/// This routine accepts a varying number of arguments of any type. It will
582/// attempt to combine them into a single hash_code. For user-defined types it
583/// attempts to call a \see hash_value overload (via ADL) for the type. For
584/// integer and pointer types it directly combines their data into the
585/// resulting hash_code.
586///
587/// The result is suitable for returning from a user's hash_value
588/// *implementation* for their user-defined type. Consumers of a type should
589/// *not* call this routine, they should instead call 'hash_value'.
590template <typename ...Ts> hash_code hash_combine(const Ts &...args) {
591 // Recursively hash each argument using a helper class.
592 ::llvm::hashing::detail::hash_combine_recursive_helper helper;
593 return helper.combine(0, helper.buffer, helper.buffer + 64, args...);
594}
595
596// Implementation details for implementations of hash_value overloads provided
597// here.
598namespace hashing {
599namespace detail {
600
601/// Helper to hash the value of a single integer.
602///
603/// Overloads for smaller integer types are not provided to ensure consistent
604/// behavior in the presence of integral promotions. Essentially,
605/// "hash_value('4')" and "hash_value('0' + 4)" should be the same.
606inline hash_code hash_integer_value(uint64_t value) {
607 // Similar to hash_4to8_bytes but using a seed instead of length.
608 const uint64_t seed = get_execution_seed();
609 const char *s = reinterpret_cast<const char *>(&value);
610 const uint64_t a = fetch32(p: s);
611 return hash_16_bytes(low: seed + (a << 3), high: fetch32(p: s + 4));
612}
613
614} // namespace detail
615} // namespace hashing
616
617// Declared and documented above, but defined here so that any of the hashing
618// infrastructure is available.
619template <typename T>
620std::enable_if_t<is_integral_or_enum<T>::value, hash_code> hash_value(T value) {
621 return ::llvm::hashing::detail::hash_integer_value(
622 value: static_cast<uint64_t>(value));
623}
624
625// Declared and documented above, but defined here so that any of the hashing
626// infrastructure is available.
627template <typename T> hash_code hash_value(const T *ptr) {
628 return ::llvm::hashing::detail::hash_integer_value(
629 value: reinterpret_cast<uintptr_t>(ptr));
630}
631
632// Declared and documented above, but defined here so that any of the hashing
633// infrastructure is available.
634template <typename T, typename U>
635hash_code hash_value(const std::pair<T, U> &arg) {
636 return hash_combine(arg.first, arg.second);
637}
638
639template <typename... Ts> hash_code hash_value(const std::tuple<Ts...> &arg) {
640 return std::apply([](const auto &...xs) { return hash_combine(xs...); }, arg);
641}
642
643// Declared and documented above, but defined here so that any of the hashing
644// infrastructure is available.
645template <typename T>
646hash_code hash_value(const std::basic_string<T> &arg) {
647 return hash_combine_range(arg.begin(), arg.end());
648}
649
650template <typename T> hash_code hash_value(const std::optional<T> &arg) {
651 return arg ? hash_combine(true, *arg) : hash_value(value: false);
652}
653
654template <> struct DenseMapInfo<hash_code, void> {
655 static inline hash_code getEmptyKey() { return hash_code(-1); }
656 static inline hash_code getTombstoneKey() { return hash_code(-2); }
657 static unsigned getHashValue(hash_code val) {
658 return static_cast<unsigned>(size_t(val));
659 }
660 static bool isEqual(hash_code LHS, hash_code RHS) { return LHS == RHS; }
661};
662
663} // namespace llvm
664
665/// Implement std::hash so that hash_code can be used in STL containers.
666namespace std {
667
668template<>
669struct hash<llvm::hash_code> {
670 size_t operator()(llvm::hash_code const& Val) const {
671 return Val;
672 }
673};
674
675} // namespace std;
676
677#endif
678

source code of include/llvm-20/llvm/ADT/Hashing.h