Details | Last modification | View Log | RSS feed
| Rev | Author | Line No. | Line |
|---|---|---|---|
| 14 | pmbaty | 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/Support/DataTypes.h" |
||
| 48 | #include "llvm/Support/ErrorHandling.h" |
||
| 49 | #include "llvm/Support/SwapByteOrder.h" |
||
| 50 | #include "llvm/Support/type_traits.h" |
||
| 51 | #include <algorithm> |
||
| 52 | #include <cassert> |
||
| 53 | #include <cstring> |
||
| 54 | #include <optional> |
||
| 55 | #include <string> |
||
| 56 | #include <tuple> |
||
| 57 | #include <utility> |
||
| 58 | |||
| 59 | namespace llvm { |
||
| 60 | template <typename T, typename Enable> struct DenseMapInfo; |
||
| 61 | |||
| 62 | /// An opaque object representing a hash code. |
||
| 63 | /// |
||
| 64 | /// This object represents the result of hashing some entity. It is intended to |
||
| 65 | /// be used to implement hashtables or other hashing-based data structures. |
||
| 66 | /// While it wraps and exposes a numeric value, this value should not be |
||
| 67 | /// trusted to be stable or predictable across processes or executions. |
||
| 68 | /// |
||
| 69 | /// In order to obtain the hash_code for an object 'x': |
||
| 70 | /// \code |
||
| 71 | /// using llvm::hash_value; |
||
| 72 | /// llvm::hash_code code = hash_value(x); |
||
| 73 | /// \endcode |
||
| 74 | class hash_code { |
||
| 75 | size_t value; |
||
| 76 | |||
| 77 | public: |
||
| 78 | /// Default construct a hash_code. |
||
| 79 | /// Note that this leaves the value uninitialized. |
||
| 80 | hash_code() = default; |
||
| 81 | |||
| 82 | /// Form a hash code directly from a numerical value. |
||
| 83 | hash_code(size_t value) : value(value) {} |
||
| 84 | |||
| 85 | /// Convert the hash code to its numerical value for use. |
||
| 86 | /*explicit*/ operator size_t() const { return value; } |
||
| 87 | |||
| 88 | friend bool operator==(const hash_code &lhs, const hash_code &rhs) { |
||
| 89 | return lhs.value == rhs.value; |
||
| 90 | } |
||
| 91 | friend bool operator!=(const hash_code &lhs, const hash_code &rhs) { |
||
| 92 | return lhs.value != rhs.value; |
||
| 93 | } |
||
| 94 | |||
| 95 | /// Allow a hash_code to be directly run through hash_value. |
||
| 96 | friend size_t hash_value(const hash_code &code) { return code.value; } |
||
| 97 | }; |
||
| 98 | |||
| 99 | /// Compute a hash_code for any integer value. |
||
| 100 | /// |
||
| 101 | /// Note that this function is intended to compute the same hash_code for |
||
| 102 | /// a particular value without regard to the pre-promotion type. This is in |
||
| 103 | /// contrast to hash_combine which may produce different hash_codes for |
||
| 104 | /// differing argument types even if they would implicit promote to a common |
||
| 105 | /// type without changing the value. |
||
| 106 | template <typename T> |
||
| 107 | std::enable_if_t<is_integral_or_enum<T>::value, hash_code> hash_value(T value); |
||
| 108 | |||
| 109 | /// Compute a hash_code for a pointer's address. |
||
| 110 | /// |
||
| 111 | /// N.B.: This hashes the *address*. Not the value and not the type. |
||
| 112 | template <typename T> hash_code hash_value(const T *ptr); |
||
| 113 | |||
| 114 | /// Compute a hash_code for a pair of objects. |
||
| 115 | template <typename T, typename U> |
||
| 116 | hash_code hash_value(const std::pair<T, U> &arg); |
||
| 117 | |||
| 118 | /// Compute a hash_code for a tuple. |
||
| 119 | template <typename... Ts> |
||
| 120 | hash_code hash_value(const std::tuple<Ts...> &arg); |
||
| 121 | |||
| 122 | /// Compute a hash_code for a standard string. |
||
| 123 | template <typename T> |
||
| 124 | hash_code hash_value(const std::basic_string<T> &arg); |
||
| 125 | |||
| 126 | /// Compute a hash_code for a standard string. |
||
| 127 | template <typename T> hash_code hash_value(const std::optional<T> &arg); |
||
| 128 | |||
| 129 | /// Override the execution seed with a fixed value. |
||
| 130 | /// |
||
| 131 | /// This hashing library uses a per-execution seed designed to change on each |
||
| 132 | /// run with high probability in order to ensure that the hash codes are not |
||
| 133 | /// attackable and to ensure that output which is intended to be stable does |
||
| 134 | /// not rely on the particulars of the hash codes produced. |
||
| 135 | /// |
||
| 136 | /// That said, there are use cases where it is important to be able to |
||
| 137 | /// reproduce *exactly* a specific behavior. To that end, we provide a function |
||
| 138 | /// which will forcibly set the seed to a fixed value. This must be done at the |
||
| 139 | /// start of the program, before any hashes are computed. Also, it cannot be |
||
| 140 | /// undone. This makes it thread-hostile and very hard to use outside of |
||
| 141 | /// immediately on start of a simple program designed for reproducible |
||
| 142 | /// behavior. |
||
| 143 | void set_fixed_execution_hash_seed(uint64_t fixed_value); |
||
| 144 | |||
| 145 | |||
| 146 | // All of the implementation details of actually computing the various hash |
||
| 147 | // code values are held within this namespace. These routines are included in |
||
| 148 | // the header file mainly to allow inlining and constant propagation. |
||
| 149 | namespace hashing { |
||
| 150 | namespace detail { |
||
| 151 | |||
| 152 | inline uint64_t fetch64(const char *p) { |
||
| 153 | uint64_t result; |
||
| 154 | memcpy(&result, p, sizeof(result)); |
||
| 155 | if (sys::IsBigEndianHost) |
||
| 156 | sys::swapByteOrder(result); |
||
| 157 | return result; |
||
| 158 | } |
||
| 159 | |||
| 160 | inline uint32_t fetch32(const char *p) { |
||
| 161 | uint32_t result; |
||
| 162 | memcpy(&result, p, sizeof(result)); |
||
| 163 | if (sys::IsBigEndianHost) |
||
| 164 | sys::swapByteOrder(result); |
||
| 165 | return result; |
||
| 166 | } |
||
| 167 | |||
| 168 | /// Some primes between 2^63 and 2^64 for various uses. |
||
| 169 | static constexpr uint64_t k0 = 0xc3a5c85c97cb3127ULL; |
||
| 170 | static constexpr uint64_t k1 = 0xb492b66fbe98f273ULL; |
||
| 171 | static constexpr uint64_t k2 = 0x9ae16a3b2f90404fULL; |
||
| 172 | static constexpr uint64_t k3 = 0xc949d7c7509e6557ULL; |
||
| 173 | |||
| 174 | /// Bitwise right rotate. |
||
| 175 | /// Normally this will compile to a single instruction, especially if the |
||
| 176 | /// shift is a manifest constant. |
||
| 177 | inline uint64_t rotate(uint64_t val, size_t shift) { |
||
| 178 | // Avoid shifting by 64: doing so yields an undefined result. |
||
| 179 | return shift == 0 ? val : ((val >> shift) | (val << (64 - shift))); |
||
| 180 | } |
||
| 181 | |||
| 182 | inline uint64_t shift_mix(uint64_t val) { |
||
| 183 | return val ^ (val >> 47); |
||
| 184 | } |
||
| 185 | |||
| 186 | inline uint64_t hash_16_bytes(uint64_t low, uint64_t high) { |
||
| 187 | // Murmur-inspired hashing. |
||
| 188 | const uint64_t kMul = 0x9ddfea08eb382d69ULL; |
||
| 189 | uint64_t a = (low ^ high) * kMul; |
||
| 190 | a ^= (a >> 47); |
||
| 191 | uint64_t b = (high ^ a) * kMul; |
||
| 192 | b ^= (b >> 47); |
||
| 193 | b *= kMul; |
||
| 194 | return b; |
||
| 195 | } |
||
| 196 | |||
| 197 | inline uint64_t hash_1to3_bytes(const char *s, size_t len, uint64_t seed) { |
||
| 198 | uint8_t a = s[0]; |
||
| 199 | uint8_t b = s[len >> 1]; |
||
| 200 | uint8_t c = s[len - 1]; |
||
| 201 | uint32_t y = static_cast<uint32_t>(a) + (static_cast<uint32_t>(b) << 8); |
||
| 202 | uint32_t z = static_cast<uint32_t>(len) + (static_cast<uint32_t>(c) << 2); |
||
| 203 | return shift_mix(y * k2 ^ z * k3 ^ seed) * k2; |
||
| 204 | } |
||
| 205 | |||
| 206 | inline uint64_t hash_4to8_bytes(const char *s, size_t len, uint64_t seed) { |
||
| 207 | uint64_t a = fetch32(s); |
||
| 208 | return hash_16_bytes(len + (a << 3), seed ^ fetch32(s + len - 4)); |
||
| 209 | } |
||
| 210 | |||
| 211 | inline uint64_t hash_9to16_bytes(const char *s, size_t len, uint64_t seed) { |
||
| 212 | uint64_t a = fetch64(s); |
||
| 213 | uint64_t b = fetch64(s + len - 8); |
||
| 214 | return hash_16_bytes(seed ^ a, rotate(b + len, len)) ^ b; |
||
| 215 | } |
||
| 216 | |||
| 217 | inline uint64_t hash_17to32_bytes(const char *s, size_t len, uint64_t seed) { |
||
| 218 | uint64_t a = fetch64(s) * k1; |
||
| 219 | uint64_t b = fetch64(s + 8); |
||
| 220 | uint64_t c = fetch64(s + len - 8) * k2; |
||
| 221 | uint64_t d = fetch64(s + len - 16) * k0; |
||
| 222 | return hash_16_bytes(rotate(a - b, 43) + rotate(c ^ seed, 30) + d, |
||
| 223 | a + rotate(b ^ k3, 20) - c + len + seed); |
||
| 224 | } |
||
| 225 | |||
| 226 | inline uint64_t hash_33to64_bytes(const char *s, size_t len, uint64_t seed) { |
||
| 227 | uint64_t z = fetch64(s + 24); |
||
| 228 | uint64_t a = fetch64(s) + (len + fetch64(s + len - 16)) * k0; |
||
| 229 | uint64_t b = rotate(a + z, 52); |
||
| 230 | uint64_t c = rotate(a, 37); |
||
| 231 | a += fetch64(s + 8); |
||
| 232 | c += rotate(a, 7); |
||
| 233 | a += fetch64(s + 16); |
||
| 234 | uint64_t vf = a + z; |
||
| 235 | uint64_t vs = b + rotate(a, 31) + c; |
||
| 236 | a = fetch64(s + 16) + fetch64(s + len - 32); |
||
| 237 | z = fetch64(s + len - 8); |
||
| 238 | b = rotate(a + z, 52); |
||
| 239 | c = rotate(a, 37); |
||
| 240 | a += fetch64(s + len - 24); |
||
| 241 | c += rotate(a, 7); |
||
| 242 | a += fetch64(s + len - 16); |
||
| 243 | uint64_t wf = a + z; |
||
| 244 | uint64_t ws = b + rotate(a, 31) + c; |
||
| 245 | uint64_t r = shift_mix((vf + ws) * k2 + (wf + vs) * k0); |
||
| 246 | return shift_mix((seed ^ (r * k0)) + vs) * k2; |
||
| 247 | } |
||
| 248 | |||
| 249 | inline uint64_t hash_short(const char *s, size_t length, uint64_t seed) { |
||
| 250 | if (length >= 4 && length <= 8) |
||
| 251 | return hash_4to8_bytes(s, length, seed); |
||
| 252 | if (length > 8 && length <= 16) |
||
| 253 | return hash_9to16_bytes(s, length, seed); |
||
| 254 | if (length > 16 && length <= 32) |
||
| 255 | return hash_17to32_bytes(s, length, seed); |
||
| 256 | if (length > 32) |
||
| 257 | return hash_33to64_bytes(s, length, seed); |
||
| 258 | if (length != 0) |
||
| 259 | return hash_1to3_bytes(s, length, seed); |
||
| 260 | |||
| 261 | return k2 ^ seed; |
||
| 262 | } |
||
| 263 | |||
| 264 | /// The intermediate state used during hashing. |
||
| 265 | /// Currently, the algorithm for computing hash codes is based on CityHash and |
||
| 266 | /// keeps 56 bytes of arbitrary state. |
||
| 267 | struct hash_state { |
||
| 268 | uint64_t h0 = 0, h1 = 0, h2 = 0, h3 = 0, h4 = 0, h5 = 0, h6 = 0; |
||
| 269 | |||
| 270 | /// Create a new hash_state structure and initialize it based on the |
||
| 271 | /// seed and the first 64-byte chunk. |
||
| 272 | /// This effectively performs the initial mix. |
||
| 273 | static hash_state create(const char *s, uint64_t seed) { |
||
| 274 | hash_state state = { |
||
| 275 | 0, seed, hash_16_bytes(seed, k1), rotate(seed ^ k1, 49), |
||
| 276 | seed * k1, shift_mix(seed), 0 }; |
||
| 277 | state.h6 = hash_16_bytes(state.h4, state.h5); |
||
| 278 | state.mix(s); |
||
| 279 | return state; |
||
| 280 | } |
||
| 281 | |||
| 282 | /// Mix 32-bytes from the input sequence into the 16-bytes of 'a' |
||
| 283 | /// and 'b', including whatever is already in 'a' and 'b'. |
||
| 284 | static void mix_32_bytes(const char *s, uint64_t &a, uint64_t &b) { |
||
| 285 | a += fetch64(s); |
||
| 286 | uint64_t c = fetch64(s + 24); |
||
| 287 | b = rotate(b + a + c, 21); |
||
| 288 | uint64_t d = a; |
||
| 289 | a += fetch64(s + 8) + fetch64(s + 16); |
||
| 290 | b += rotate(a, 44) + d; |
||
| 291 | a += c; |
||
| 292 | } |
||
| 293 | |||
| 294 | /// Mix in a 64-byte buffer of data. |
||
| 295 | /// We mix all 64 bytes even when the chunk length is smaller, but we |
||
| 296 | /// record the actual length. |
||
| 297 | void mix(const char *s) { |
||
| 298 | h0 = rotate(h0 + h1 + h3 + fetch64(s + 8), 37) * k1; |
||
| 299 | h1 = rotate(h1 + h4 + fetch64(s + 48), 42) * k1; |
||
| 300 | h0 ^= h6; |
||
| 301 | h1 += h3 + fetch64(s + 40); |
||
| 302 | h2 = rotate(h2 + h5, 33) * k1; |
||
| 303 | h3 = h4 * k1; |
||
| 304 | h4 = h0 + h5; |
||
| 305 | mix_32_bytes(s, h3, h4); |
||
| 306 | h5 = h2 + h6; |
||
| 307 | h6 = h1 + fetch64(s + 16); |
||
| 308 | mix_32_bytes(s + 32, h5, h6); |
||
| 309 | std::swap(h2, h0); |
||
| 310 | } |
||
| 311 | |||
| 312 | /// Compute the final 64-bit hash code value based on the current |
||
| 313 | /// state and the length of bytes hashed. |
||
| 314 | uint64_t finalize(size_t length) { |
||
| 315 | return hash_16_bytes(hash_16_bytes(h3, h5) + shift_mix(h1) * k1 + h2, |
||
| 316 | hash_16_bytes(h4, h6) + shift_mix(length) * k1 + h0); |
||
| 317 | } |
||
| 318 | }; |
||
| 319 | |||
| 320 | |||
| 321 | /// A global, fixed seed-override variable. |
||
| 322 | /// |
||
| 323 | /// This variable can be set using the \see llvm::set_fixed_execution_seed |
||
| 324 | /// function. See that function for details. Do not, under any circumstances, |
||
| 325 | /// set or read this variable. |
||
| 326 | extern uint64_t fixed_seed_override; |
||
| 327 | |||
| 328 | inline uint64_t get_execution_seed() { |
||
| 329 | // FIXME: This needs to be a per-execution seed. This is just a placeholder |
||
| 330 | // implementation. Switching to a per-execution seed is likely to flush out |
||
| 331 | // instability bugs and so will happen as its own commit. |
||
| 332 | // |
||
| 333 | // However, if there is a fixed seed override set the first time this is |
||
| 334 | // called, return that instead of the per-execution seed. |
||
| 335 | const uint64_t seed_prime = 0xff51afd7ed558ccdULL; |
||
| 336 | static uint64_t seed = fixed_seed_override ? fixed_seed_override : seed_prime; |
||
| 337 | return seed; |
||
| 338 | } |
||
| 339 | |||
| 340 | |||
| 341 | /// Trait to indicate whether a type's bits can be hashed directly. |
||
| 342 | /// |
||
| 343 | /// A type trait which is true if we want to combine values for hashing by |
||
| 344 | /// reading the underlying data. It is false if values of this type must |
||
| 345 | /// first be passed to hash_value, and the resulting hash_codes combined. |
||
| 346 | // |
||
| 347 | // FIXME: We want to replace is_integral_or_enum and is_pointer here with |
||
| 348 | // a predicate which asserts that comparing the underlying storage of two |
||
| 349 | // values of the type for equality is equivalent to comparing the two values |
||
| 350 | // for equality. For all the platforms we care about, this holds for integers |
||
| 351 | // and pointers, but there are platforms where it doesn't and we would like to |
||
| 352 | // support user-defined types which happen to satisfy this property. |
||
| 353 | template <typename T> struct is_hashable_data |
||
| 354 | : std::integral_constant<bool, ((is_integral_or_enum<T>::value || |
||
| 355 | std::is_pointer<T>::value) && |
||
| 356 | 64 % sizeof(T) == 0)> {}; |
||
| 357 | |||
| 358 | // Special case std::pair to detect when both types are viable and when there |
||
| 359 | // is no alignment-derived padding in the pair. This is a bit of a lie because |
||
| 360 | // std::pair isn't truly POD, but it's close enough in all reasonable |
||
| 361 | // implementations for our use case of hashing the underlying data. |
||
| 362 | template <typename T, typename U> struct is_hashable_data<std::pair<T, U> > |
||
| 363 | : std::integral_constant<bool, (is_hashable_data<T>::value && |
||
| 364 | is_hashable_data<U>::value && |
||
| 365 | (sizeof(T) + sizeof(U)) == |
||
| 366 | sizeof(std::pair<T, U>))> {}; |
||
| 367 | |||
| 368 | /// Helper to get the hashable data representation for a type. |
||
| 369 | /// This variant is enabled when the type itself can be used. |
||
| 370 | template <typename T> |
||
| 371 | std::enable_if_t<is_hashable_data<T>::value, T> |
||
| 372 | get_hashable_data(const T &value) { |
||
| 373 | return value; |
||
| 374 | } |
||
| 375 | /// Helper to get the hashable data representation for a type. |
||
| 376 | /// This variant is enabled when we must first call hash_value and use the |
||
| 377 | /// result as our data. |
||
| 378 | template <typename T> |
||
| 379 | std::enable_if_t<!is_hashable_data<T>::value, size_t> |
||
| 380 | get_hashable_data(const T &value) { |
||
| 381 | using ::llvm::hash_value; |
||
| 382 | return hash_value(value); |
||
| 383 | } |
||
| 384 | |||
| 385 | /// Helper to store data from a value into a buffer and advance the |
||
| 386 | /// pointer into that buffer. |
||
| 387 | /// |
||
| 388 | /// This routine first checks whether there is enough space in the provided |
||
| 389 | /// buffer, and if not immediately returns false. If there is space, it |
||
| 390 | /// copies the underlying bytes of value into the buffer, advances the |
||
| 391 | /// buffer_ptr past the copied bytes, and returns true. |
||
| 392 | template <typename T> |
||
| 393 | bool store_and_advance(char *&buffer_ptr, char *buffer_end, const T& value, |
||
| 394 | size_t offset = 0) { |
||
| 395 | size_t store_size = sizeof(value) - offset; |
||
| 396 | if (buffer_ptr + store_size > buffer_end) |
||
| 397 | return false; |
||
| 398 | const char *value_data = reinterpret_cast<const char *>(&value); |
||
| 399 | memcpy(buffer_ptr, value_data + offset, store_size); |
||
| 400 | buffer_ptr += store_size; |
||
| 401 | return true; |
||
| 402 | } |
||
| 403 | |||
| 404 | /// Implement the combining of integral values into a hash_code. |
||
| 405 | /// |
||
| 406 | /// This overload is selected when the value type of the iterator is |
||
| 407 | /// integral. Rather than computing a hash_code for each object and then |
||
| 408 | /// combining them, this (as an optimization) directly combines the integers. |
||
| 409 | template <typename InputIteratorT> |
||
| 410 | hash_code hash_combine_range_impl(InputIteratorT first, InputIteratorT last) { |
||
| 411 | const uint64_t seed = get_execution_seed(); |
||
| 412 | char buffer[64], *buffer_ptr = buffer; |
||
| 413 | char *const buffer_end = std::end(buffer); |
||
| 414 | while (first != last && store_and_advance(buffer_ptr, buffer_end, |
||
| 415 | get_hashable_data(*first))) |
||
| 416 | ++first; |
||
| 417 | if (first == last) |
||
| 418 | return hash_short(buffer, buffer_ptr - buffer, seed); |
||
| 419 | assert(buffer_ptr == buffer_end); |
||
| 420 | |||
| 421 | hash_state state = state.create(buffer, seed); |
||
| 422 | size_t length = 64; |
||
| 423 | while (first != last) { |
||
| 424 | // Fill up the buffer. We don't clear it, which re-mixes the last round |
||
| 425 | // when only a partial 64-byte chunk is left. |
||
| 426 | buffer_ptr = buffer; |
||
| 427 | while (first != last && store_and_advance(buffer_ptr, buffer_end, |
||
| 428 | get_hashable_data(*first))) |
||
| 429 | ++first; |
||
| 430 | |||
| 431 | // Rotate the buffer if we did a partial fill in order to simulate doing |
||
| 432 | // a mix of the last 64-bytes. That is how the algorithm works when we |
||
| 433 | // have a contiguous byte sequence, and we want to emulate that here. |
||
| 434 | std::rotate(buffer, buffer_ptr, buffer_end); |
||
| 435 | |||
| 436 | // Mix this chunk into the current state. |
||
| 437 | state.mix(buffer); |
||
| 438 | length += buffer_ptr - buffer; |
||
| 439 | }; |
||
| 440 | |||
| 441 | return state.finalize(length); |
||
| 442 | } |
||
| 443 | |||
| 444 | /// Implement the combining of integral values into a hash_code. |
||
| 445 | /// |
||
| 446 | /// This overload is selected when the value type of the iterator is integral |
||
| 447 | /// and when the input iterator is actually a pointer. Rather than computing |
||
| 448 | /// a hash_code for each object and then combining them, this (as an |
||
| 449 | /// optimization) directly combines the integers. Also, because the integers |
||
| 450 | /// are stored in contiguous memory, this routine avoids copying each value |
||
| 451 | /// and directly reads from the underlying memory. |
||
| 452 | template <typename ValueT> |
||
| 453 | std::enable_if_t<is_hashable_data<ValueT>::value, hash_code> |
||
| 454 | hash_combine_range_impl(ValueT *first, ValueT *last) { |
||
| 455 | const uint64_t seed = get_execution_seed(); |
||
| 456 | const char *s_begin = reinterpret_cast<const char *>(first); |
||
| 457 | const char *s_end = reinterpret_cast<const char *>(last); |
||
| 458 | const size_t length = std::distance(s_begin, s_end); |
||
| 459 | if (length <= 64) |
||
| 460 | return hash_short(s_begin, length, seed); |
||
| 461 | |||
| 462 | const char *s_aligned_end = s_begin + (length & ~63); |
||
| 463 | hash_state state = state.create(s_begin, seed); |
||
| 464 | s_begin += 64; |
||
| 465 | while (s_begin != s_aligned_end) { |
||
| 466 | state.mix(s_begin); |
||
| 467 | s_begin += 64; |
||
| 468 | } |
||
| 469 | if (length & 63) |
||
| 470 | state.mix(s_end - 64); |
||
| 471 | |||
| 472 | return state.finalize(length); |
||
| 473 | } |
||
| 474 | |||
| 475 | } // namespace detail |
||
| 476 | } // namespace hashing |
||
| 477 | |||
| 478 | |||
| 479 | /// Compute a hash_code for a sequence of values. |
||
| 480 | /// |
||
| 481 | /// This hashes a sequence of values. It produces the same hash_code as |
||
| 482 | /// 'hash_combine(a, b, c, ...)', but can run over arbitrary sized sequences |
||
| 483 | /// and is significantly faster given pointers and types which can be hashed as |
||
| 484 | /// a sequence of bytes. |
||
| 485 | template <typename InputIteratorT> |
||
| 486 | hash_code hash_combine_range(InputIteratorT first, InputIteratorT last) { |
||
| 487 | return ::llvm::hashing::detail::hash_combine_range_impl(first, last); |
||
| 488 | } |
||
| 489 | |||
| 490 | |||
| 491 | // Implementation details for hash_combine. |
||
| 492 | namespace hashing { |
||
| 493 | namespace detail { |
||
| 494 | |||
| 495 | /// Helper class to manage the recursive combining of hash_combine |
||
| 496 | /// arguments. |
||
| 497 | /// |
||
| 498 | /// This class exists to manage the state and various calls involved in the |
||
| 499 | /// recursive combining of arguments used in hash_combine. It is particularly |
||
| 500 | /// useful at minimizing the code in the recursive calls to ease the pain |
||
| 501 | /// caused by a lack of variadic functions. |
||
| 502 | struct hash_combine_recursive_helper { |
||
| 503 | char buffer[64] = {}; |
||
| 504 | hash_state state; |
||
| 505 | const uint64_t seed; |
||
| 506 | |||
| 507 | public: |
||
| 508 | /// Construct a recursive hash combining helper. |
||
| 509 | /// |
||
| 510 | /// This sets up the state for a recursive hash combine, including getting |
||
| 511 | /// the seed and buffer setup. |
||
| 512 | hash_combine_recursive_helper() |
||
| 513 | : seed(get_execution_seed()) {} |
||
| 514 | |||
| 515 | /// Combine one chunk of data into the current in-flight hash. |
||
| 516 | /// |
||
| 517 | /// This merges one chunk of data into the hash. First it tries to buffer |
||
| 518 | /// the data. If the buffer is full, it hashes the buffer into its |
||
| 519 | /// hash_state, empties it, and then merges the new chunk in. This also |
||
| 520 | /// handles cases where the data straddles the end of the buffer. |
||
| 521 | template <typename T> |
||
| 522 | char *combine_data(size_t &length, char *buffer_ptr, char *buffer_end, T data) { |
||
| 523 | if (!store_and_advance(buffer_ptr, buffer_end, data)) { |
||
| 524 | // Check for skew which prevents the buffer from being packed, and do |
||
| 525 | // a partial store into the buffer to fill it. This is only a concern |
||
| 526 | // with the variadic combine because that formation can have varying |
||
| 527 | // argument types. |
||
| 528 | size_t partial_store_size = buffer_end - buffer_ptr; |
||
| 529 | memcpy(buffer_ptr, &data, partial_store_size); |
||
| 530 | |||
| 531 | // If the store fails, our buffer is full and ready to hash. We have to |
||
| 532 | // either initialize the hash state (on the first full buffer) or mix |
||
| 533 | // this buffer into the existing hash state. Length tracks the *hashed* |
||
| 534 | // length, not the buffered length. |
||
| 535 | if (length == 0) { |
||
| 536 | state = state.create(buffer, seed); |
||
| 537 | length = 64; |
||
| 538 | } else { |
||
| 539 | // Mix this chunk into the current state and bump length up by 64. |
||
| 540 | state.mix(buffer); |
||
| 541 | length += 64; |
||
| 542 | } |
||
| 543 | // Reset the buffer_ptr to the head of the buffer for the next chunk of |
||
| 544 | // data. |
||
| 545 | buffer_ptr = buffer; |
||
| 546 | |||
| 547 | // Try again to store into the buffer -- this cannot fail as we only |
||
| 548 | // store types smaller than the buffer. |
||
| 549 | if (!store_and_advance(buffer_ptr, buffer_end, data, |
||
| 550 | partial_store_size)) |
||
| 551 | llvm_unreachable("buffer smaller than stored type"); |
||
| 552 | } |
||
| 553 | return buffer_ptr; |
||
| 554 | } |
||
| 555 | |||
| 556 | /// Recursive, variadic combining method. |
||
| 557 | /// |
||
| 558 | /// This function recurses through each argument, combining that argument |
||
| 559 | /// into a single hash. |
||
| 560 | template <typename T, typename ...Ts> |
||
| 561 | hash_code combine(size_t length, char *buffer_ptr, char *buffer_end, |
||
| 562 | const T &arg, const Ts &...args) { |
||
| 563 | buffer_ptr = combine_data(length, buffer_ptr, buffer_end, get_hashable_data(arg)); |
||
| 564 | |||
| 565 | // Recurse to the next argument. |
||
| 566 | return combine(length, buffer_ptr, buffer_end, args...); |
||
| 567 | } |
||
| 568 | |||
| 569 | /// Base case for recursive, variadic combining. |
||
| 570 | /// |
||
| 571 | /// The base case when combining arguments recursively is reached when all |
||
| 572 | /// arguments have been handled. It flushes the remaining buffer and |
||
| 573 | /// constructs a hash_code. |
||
| 574 | hash_code combine(size_t length, char *buffer_ptr, char *buffer_end) { |
||
| 575 | // Check whether the entire set of values fit in the buffer. If so, we'll |
||
| 576 | // use the optimized short hashing routine and skip state entirely. |
||
| 577 | if (length == 0) |
||
| 578 | return hash_short(buffer, buffer_ptr - buffer, seed); |
||
| 579 | |||
| 580 | // Mix the final buffer, rotating it if we did a partial fill in order to |
||
| 581 | // simulate doing a mix of the last 64-bytes. That is how the algorithm |
||
| 582 | // works when we have a contiguous byte sequence, and we want to emulate |
||
| 583 | // that here. |
||
| 584 | std::rotate(buffer, buffer_ptr, buffer_end); |
||
| 585 | |||
| 586 | // Mix this chunk into the current state. |
||
| 587 | state.mix(buffer); |
||
| 588 | length += buffer_ptr - buffer; |
||
| 589 | |||
| 590 | return state.finalize(length); |
||
| 591 | } |
||
| 592 | }; |
||
| 593 | |||
| 594 | } // namespace detail |
||
| 595 | } // namespace hashing |
||
| 596 | |||
| 597 | /// Combine values into a single hash_code. |
||
| 598 | /// |
||
| 599 | /// This routine accepts a varying number of arguments of any type. It will |
||
| 600 | /// attempt to combine them into a single hash_code. For user-defined types it |
||
| 601 | /// attempts to call a \see hash_value overload (via ADL) for the type. For |
||
| 602 | /// integer and pointer types it directly combines their data into the |
||
| 603 | /// resulting hash_code. |
||
| 604 | /// |
||
| 605 | /// The result is suitable for returning from a user's hash_value |
||
| 606 | /// *implementation* for their user-defined type. Consumers of a type should |
||
| 607 | /// *not* call this routine, they should instead call 'hash_value'. |
||
| 608 | template <typename ...Ts> hash_code hash_combine(const Ts &...args) { |
||
| 609 | // Recursively hash each argument using a helper class. |
||
| 610 | ::llvm::hashing::detail::hash_combine_recursive_helper helper; |
||
| 611 | return helper.combine(0, helper.buffer, helper.buffer + 64, args...); |
||
| 612 | } |
||
| 613 | |||
| 614 | // Implementation details for implementations of hash_value overloads provided |
||
| 615 | // here. |
||
| 616 | namespace hashing { |
||
| 617 | namespace detail { |
||
| 618 | |||
| 619 | /// Helper to hash the value of a single integer. |
||
| 620 | /// |
||
| 621 | /// Overloads for smaller integer types are not provided to ensure consistent |
||
| 622 | /// behavior in the presence of integral promotions. Essentially, |
||
| 623 | /// "hash_value('4')" and "hash_value('0' + 4)" should be the same. |
||
| 624 | inline hash_code hash_integer_value(uint64_t value) { |
||
| 625 | // Similar to hash_4to8_bytes but using a seed instead of length. |
||
| 626 | const uint64_t seed = get_execution_seed(); |
||
| 627 | const char *s = reinterpret_cast<const char *>(&value); |
||
| 628 | const uint64_t a = fetch32(s); |
||
| 629 | return hash_16_bytes(seed + (a << 3), fetch32(s + 4)); |
||
| 630 | } |
||
| 631 | |||
| 632 | } // namespace detail |
||
| 633 | } // namespace hashing |
||
| 634 | |||
| 635 | // Declared and documented above, but defined here so that any of the hashing |
||
| 636 | // infrastructure is available. |
||
| 637 | template <typename T> |
||
| 638 | std::enable_if_t<is_integral_or_enum<T>::value, hash_code> hash_value(T value) { |
||
| 639 | return ::llvm::hashing::detail::hash_integer_value( |
||
| 640 | static_cast<uint64_t>(value)); |
||
| 641 | } |
||
| 642 | |||
| 643 | // Declared and documented above, but defined here so that any of the hashing |
||
| 644 | // infrastructure is available. |
||
| 645 | template <typename T> hash_code hash_value(const T *ptr) { |
||
| 646 | return ::llvm::hashing::detail::hash_integer_value( |
||
| 647 | reinterpret_cast<uintptr_t>(ptr)); |
||
| 648 | } |
||
| 649 | |||
| 650 | // Declared and documented above, but defined here so that any of the hashing |
||
| 651 | // infrastructure is available. |
||
| 652 | template <typename T, typename U> |
||
| 653 | hash_code hash_value(const std::pair<T, U> &arg) { |
||
| 654 | return hash_combine(arg.first, arg.second); |
||
| 655 | } |
||
| 656 | |||
| 657 | template <typename... Ts> hash_code hash_value(const std::tuple<Ts...> &arg) { |
||
| 658 | return std::apply([](const auto &...xs) { return hash_combine(xs...); }, arg); |
||
| 659 | } |
||
| 660 | |||
| 661 | // Declared and documented above, but defined here so that any of the hashing |
||
| 662 | // infrastructure is available. |
||
| 663 | template <typename T> |
||
| 664 | hash_code hash_value(const std::basic_string<T> &arg) { |
||
| 665 | return hash_combine_range(arg.begin(), arg.end()); |
||
| 666 | } |
||
| 667 | |||
| 668 | template <typename T> hash_code hash_value(const std::optional<T> &arg) { |
||
| 669 | return arg ? hash_combine(true, *arg) : hash_value(false); |
||
| 670 | } |
||
| 671 | |||
| 672 | template <> struct DenseMapInfo<hash_code, void> { |
||
| 673 | static inline hash_code getEmptyKey() { return hash_code(-1); } |
||
| 674 | static inline hash_code getTombstoneKey() { return hash_code(-2); } |
||
| 675 | static unsigned getHashValue(hash_code val) { return val; } |
||
| 676 | static bool isEqual(hash_code LHS, hash_code RHS) { return LHS == RHS; } |
||
| 677 | }; |
||
| 678 | |||
| 679 | } // namespace llvm |
||
| 680 | |||
| 681 | #endif |