1 //===-- llvm/ADT/Hashing.h - Utilities for hashing --------------*- C++ -*-===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements the newly proposed standard C++ interfaces for hashing 11 // arbitrary data and building hash functions for user-defined types. This 12 // interface was originally proposed in N3333[1] and is currently under review 13 // for inclusion in a future TR and/or standard. 14 // 15 // The primary interfaces provide are comprised of one type and three functions: 16 // 17 // -- 'hash_code' class is an opaque type representing the hash code for some 18 // data. It is the intended product of hashing, and can be used to implement 19 // hash tables, checksumming, and other common uses of hashes. It is not an 20 // integer type (although it can be converted to one) because it is risky 21 // to assume much about the internals of a hash_code. In particular, each 22 // execution of the program has a high probability of producing a different 23 // hash_code for a given input. Thus their values are not stable to save or 24 // persist, and should only be used during the execution for the 25 // construction of hashing datastructures. 26 // 27 // -- 'hash_value' is a function designed to be overloaded for each 28 // user-defined type which wishes to be used within a hashing context. It 29 // should be overloaded within the user-defined type's namespace and found 30 // via ADL. Overloads for primitive types are provided by this library. 31 // 32 // -- 'hash_combine' and 'hash_combine_range' are functions designed to aid 33 // programmers in easily and intuitively combining a set of data into 34 // a single hash_code for their object. They should only logically be used 35 // within the implementation of a 'hash_value' routine or similar context. 36 // 37 // Note that 'hash_combine_range' contains very special logic for hashing 38 // a contiguous array of integers or pointers. This logic is *extremely* fast, 39 // on a modern Intel "Gainestown" Xeon (Nehalem uarch) @2.2 GHz, these were 40 // benchmarked at over 6.5 GiB/s for large keys, and <20 cycles/hash for keys 41 // under 32-bytes. 42 // 43 //===----------------------------------------------------------------------===// 44 45 #ifndef LLVM_ADT_HASHING_H 46 #define LLVM_ADT_HASHING_H 47 48 #include "llvm/ADT/STLExtras.h" 49 #include "llvm/Support/DataTypes.h" 50 #include "llvm/Support/Host.h" 51 #include "llvm/Support/SwapByteOrder.h" 52 #include "llvm/Support/type_traits.h" 53 #include <algorithm> 54 #include <cassert> 55 #include <cstring> 56 #include <iterator> 57 #include <utility> 58 59 // Allow detecting C++11 feature availability when building with Clang without 60 // breaking other compilers. 61 #ifndef __has_feature 62 # define __has_feature(x) 0 63 #endif 64 65 namespace llvm { 66 67 /// \brief An opaque object representing a hash code. 68 /// 69 /// This object represents the result of hashing some entity. It is intended to 70 /// be used to implement hashtables or other hashing-based data structures. 71 /// While it wraps and exposes a numeric value, this value should not be 72 /// trusted to be stable or predictable across processes or executions. 73 /// 74 /// In order to obtain the hash_code for an object 'x': 75 /// \code 76 /// using llvm::hash_value; 77 /// llvm::hash_code code = hash_value(x); 78 /// \endcode 79 /// 80 /// Also note that there are two numerical values which are reserved, and the 81 /// implementation ensures will never be produced for real hash_codes. These 82 /// can be used as sentinels within hashing data structures. 83 class hash_code { 84 size_t value; 85 86 public: 87 /// \brief Default construct a hash_code. 88 /// Note that this leaves the value uninitialized. 89 hash_code() {} 90 91 /// \brief Form a hash code directly from a numerical value. 92 hash_code(size_t value) : value(value) {} 93 94 /// \brief Convert the hash code to its numerical value for use. 95 /*explicit*/ operator size_t() const { return value; } 96 97 friend bool operator==(const hash_code &lhs, const hash_code &rhs) { 98 return lhs.value == rhs.value; 99 } 100 friend bool operator!=(const hash_code &lhs, const hash_code &rhs) { 101 return lhs.value != rhs.value; 102 } 103 104 /// \brief Allow a hash_code to be directly run through hash_value. 105 friend size_t hash_value(const hash_code &code) { return code.value; } 106 }; 107 108 /// \brief Compute a hash_code for any integer value. 109 /// 110 /// Note that this function is intended to compute the same hash_code for 111 /// a particular value without regard to the pre-promotion type. This is in 112 /// contrast to hash_combine which may produce different hash_codes for 113 /// differing argument types even if they would implicit promote to a common 114 /// type without changing the value. 115 template <typename T> 116 typename enable_if<is_integral_or_enum<T>, hash_code>::type hash_value(T value); 117 118 /// \brief Compute a hash_code for a pointer's address. 119 /// 120 /// N.B.: This hashes the *address*. Not the value and not the type. 121 template <typename T> hash_code hash_value(const T *ptr); 122 123 /// \brief Compute a hash_code for a pair of objects. 124 template <typename T, typename U> 125 hash_code hash_value(const std::pair<T, U> &arg); 126 127 /// \brief Compute a hash_code for a standard string. 128 template <typename T> 129 hash_code hash_value(const std::basic_string<T> &arg); 130 131 132 /// \brief Override the execution seed with a fixed value. 133 /// 134 /// This hashing library uses a per-execution seed designed to change on each 135 /// run with high probability in order to ensure that the hash codes are not 136 /// attackable and to ensure that output which is intended to be stable does 137 /// not rely on the particulars of the hash codes produced. 138 /// 139 /// That said, there are use cases where it is important to be able to 140 /// reproduce *exactly* a specific behavior. To that end, we provide a function 141 /// which will forcibly set the seed to a fixed value. This must be done at the 142 /// start of the program, before any hashes are computed. Also, it cannot be 143 /// undone. This makes it thread-hostile and very hard to use outside of 144 /// immediately on start of a simple program designed for reproducible 145 /// behavior. 146 void set_fixed_execution_hash_seed(size_t fixed_value); 147 148 149 // All of the implementation details of actually computing the various hash 150 // code values are held within this namespace. These routines are included in 151 // the header file mainly to allow inlining and constant propagation. 152 namespace hashing { 153 namespace detail { 154 155 inline uint64_t fetch64(const char *p) { 156 uint64_t result; 157 memcpy(&result, p, sizeof(result)); 158 if (sys::isBigEndianHost()) 159 return sys::SwapByteOrder(result); 160 return result; 161 } 162 163 inline uint32_t fetch32(const char *p) { 164 uint32_t result; 165 memcpy(&result, p, sizeof(result)); 166 if (sys::isBigEndianHost()) 167 return sys::SwapByteOrder(result); 168 return result; 169 } 170 171 /// Some primes between 2^63 and 2^64 for various uses. 172 static const uint64_t k0 = 0xc3a5c85c97cb3127ULL; 173 static const uint64_t k1 = 0xb492b66fbe98f273ULL; 174 static const uint64_t k2 = 0x9ae16a3b2f90404fULL; 175 static const uint64_t k3 = 0xc949d7c7509e6557ULL; 176 177 /// \brief Bitwise right rotate. 178 /// Normally this will compile to a single instruction, especially if the 179 /// shift is a manifest constant. 180 inline uint64_t rotate(uint64_t val, size_t shift) { 181 // Avoid shifting by 64: doing so yields an undefined result. 182 return shift == 0 ? val : ((val >> shift) | (val << (64 - shift))); 183 } 184 185 inline uint64_t shift_mix(uint64_t val) { 186 return val ^ (val >> 47); 187 } 188 189 inline uint64_t hash_16_bytes(uint64_t low, uint64_t high) { 190 // Murmur-inspired hashing. 191 const uint64_t kMul = 0x9ddfea08eb382d69ULL; 192 uint64_t a = (low ^ high) * kMul; 193 a ^= (a >> 47); 194 uint64_t b = (high ^ a) * kMul; 195 b ^= (b >> 47); 196 b *= kMul; 197 return b; 198 } 199 200 inline uint64_t hash_1to3_bytes(const char *s, size_t len, uint64_t seed) { 201 uint8_t a = s[0]; 202 uint8_t b = s[len >> 1]; 203 uint8_t c = s[len - 1]; 204 uint32_t y = static_cast<uint32_t>(a) + (static_cast<uint32_t>(b) << 8); 205 uint32_t z = len + (static_cast<uint32_t>(c) << 2); 206 return shift_mix(y * k2 ^ z * k3 ^ seed) * k2; 207 } 208 209 inline uint64_t hash_4to8_bytes(const char *s, size_t len, uint64_t seed) { 210 uint64_t a = fetch32(s); 211 return hash_16_bytes(len + (a << 3), seed ^ fetch32(s + len - 4)); 212 } 213 214 inline uint64_t hash_9to16_bytes(const char *s, size_t len, uint64_t seed) { 215 uint64_t a = fetch64(s); 216 uint64_t b = fetch64(s + len - 8); 217 return hash_16_bytes(seed ^ a, rotate(b + len, len)) ^ b; 218 } 219 220 inline uint64_t hash_17to32_bytes(const char *s, size_t len, uint64_t seed) { 221 uint64_t a = fetch64(s) * k1; 222 uint64_t b = fetch64(s + 8); 223 uint64_t c = fetch64(s + len - 8) * k2; 224 uint64_t d = fetch64(s + len - 16) * k0; 225 return hash_16_bytes(rotate(a - b, 43) + rotate(c ^ seed, 30) + d, 226 a + rotate(b ^ k3, 20) - c + len + seed); 227 } 228 229 inline uint64_t hash_33to64_bytes(const char *s, size_t len, uint64_t seed) { 230 uint64_t z = fetch64(s + 24); 231 uint64_t a = fetch64(s) + (len + fetch64(s + len - 16)) * k0; 232 uint64_t b = rotate(a + z, 52); 233 uint64_t c = rotate(a, 37); 234 a += fetch64(s + 8); 235 c += rotate(a, 7); 236 a += fetch64(s + 16); 237 uint64_t vf = a + z; 238 uint64_t vs = b + rotate(a, 31) + c; 239 a = fetch64(s + 16) + fetch64(s + len - 32); 240 z = fetch64(s + len - 8); 241 b = rotate(a + z, 52); 242 c = rotate(a, 37); 243 a += fetch64(s + len - 24); 244 c += rotate(a, 7); 245 a += fetch64(s + len - 16); 246 uint64_t wf = a + z; 247 uint64_t ws = b + rotate(a, 31) + c; 248 uint64_t r = shift_mix((vf + ws) * k2 + (wf + vs) * k0); 249 return shift_mix((seed ^ (r * k0)) + vs) * k2; 250 } 251 252 inline uint64_t hash_short(const char *s, size_t length, uint64_t seed) { 253 if (length >= 4 && length <= 8) 254 return hash_4to8_bytes(s, length, seed); 255 if (length > 8 && length <= 16) 256 return hash_9to16_bytes(s, length, seed); 257 if (length > 16 && length <= 32) 258 return hash_17to32_bytes(s, length, seed); 259 if (length > 32) 260 return hash_33to64_bytes(s, length, seed); 261 if (length != 0) 262 return hash_1to3_bytes(s, length, seed); 263 264 return k2 ^ seed; 265 } 266 267 /// \brief The intermediate state used during hashing. 268 /// Currently, the algorithm for computing hash codes is based on CityHash and 269 /// keeps 56 bytes of arbitrary state. 270 struct hash_state { 271 uint64_t h0, h1, h2, h3, h4, h5, h6; 272 uint64_t seed; 273 274 /// \brief Create a new hash_state structure and initialize it based on the 275 /// seed and the first 64-byte chunk. 276 /// This effectively performs the initial mix. 277 static hash_state create(const char *s, uint64_t seed) { 278 hash_state state = { 279 0, seed, hash_16_bytes(seed, k1), rotate(seed ^ k1, 49), 280 seed * k1, shift_mix(seed), 0, seed }; 281 state.h6 = hash_16_bytes(state.h4, state.h5); 282 state.mix(s); 283 return state; 284 } 285 286 /// \brief Mix 32-bytes from the input sequence into the 16-bytes of 'a' 287 /// and 'b', including whatever is already in 'a' and 'b'. 288 static void mix_32_bytes(const char *s, uint64_t &a, uint64_t &b) { 289 a += fetch64(s); 290 uint64_t c = fetch64(s + 24); 291 b = rotate(b + a + c, 21); 292 uint64_t d = a; 293 a += fetch64(s + 8) + fetch64(s + 16); 294 b += rotate(a, 44) + d; 295 a += c; 296 } 297 298 /// \brief Mix in a 64-byte buffer of data. 299 /// We mix all 64 bytes even when the chunk length is smaller, but we 300 /// record the actual length. 301 void mix(const char *s) { 302 h0 = rotate(h0 + h1 + h3 + fetch64(s + 8), 37) * k1; 303 h1 = rotate(h1 + h4 + fetch64(s + 48), 42) * k1; 304 h0 ^= h6; 305 h1 += h3 + fetch64(s + 40); 306 h2 = rotate(h2 + h5, 33) * k1; 307 h3 = h4 * k1; 308 h4 = h0 + h5; 309 mix_32_bytes(s, h3, h4); 310 h5 = h2 + h6; 311 h6 = h1 + fetch64(s + 16); 312 mix_32_bytes(s + 32, h5, h6); 313 std::swap(h2, h0); 314 } 315 316 /// \brief Compute the final 64-bit hash code value based on the current 317 /// state and the length of bytes hashed. 318 uint64_t finalize(size_t length) { 319 return hash_16_bytes(hash_16_bytes(h3, h5) + shift_mix(h1) * k1 + h2, 320 hash_16_bytes(h4, h6) + shift_mix(length) * k1 + h0); 321 } 322 }; 323 324 325 /// \brief A global, fixed seed-override variable. 326 /// 327 /// This variable can be set using the \see llvm::set_fixed_execution_seed 328 /// function. See that function for details. Do not, under any circumstances, 329 /// set or read this variable. 330 extern size_t fixed_seed_override; 331 332 inline size_t get_execution_seed() { 333 // FIXME: This needs to be a per-execution seed. This is just a placeholder 334 // implementation. Switching to a per-execution seed is likely to flush out 335 // instability bugs and so will happen as its own commit. 336 // 337 // However, if there is a fixed seed override set the first time this is 338 // called, return that instead of the per-execution seed. 339 const uint64_t seed_prime = 0xff51afd7ed558ccdULL; 340 static size_t seed = fixed_seed_override ? fixed_seed_override 341 : (size_t)seed_prime; 342 return seed; 343 } 344 345 346 /// \brief Trait to indicate whether a type's bits can be hashed directly. 347 /// 348 /// A type trait which is true if we want to combine values for hashing by 349 /// reading the underlying data. It is false if values of this type must 350 /// first be passed to hash_value, and the resulting hash_codes combined. 351 // 352 // FIXME: We want to replace is_integral_or_enum and is_pointer here with 353 // a predicate which asserts that comparing the underlying storage of two 354 // values of the type for equality is equivalent to comparing the two values 355 // for equality. For all the platforms we care about, this holds for integers 356 // and pointers, but there are platforms where it doesn't and we would like to 357 // support user-defined types which happen to satisfy this property. 358 template <typename T> struct is_hashable_data 359 : integral_constant<bool, ((is_integral_or_enum<T>::value || 360 is_pointer<T>::value) && 361 64 % sizeof(T) == 0)> {}; 362 363 // Special case std::pair to detect when both types are viable and when there 364 // is no alignment-derived padding in the pair. This is a bit of a lie because 365 // std::pair isn't truly POD, but it's close enough in all reasonable 366 // implementations for our use case of hashing the underlying data. 367 template <typename T, typename U> struct is_hashable_data<std::pair<T, U> > 368 : integral_constant<bool, (is_hashable_data<T>::value && 369 is_hashable_data<U>::value && 370 (sizeof(T) + sizeof(U)) == 371 sizeof(std::pair<T, U>))> {}; 372 373 /// \brief Helper to get the hashable data representation for a type. 374 /// This variant is enabled when the type itself can be used. 375 template <typename T> 376 typename enable_if<is_hashable_data<T>, T>::type 377 get_hashable_data(const T &value) { 378 return value; 379 } 380 /// \brief Helper to get the hashable data representation for a type. 381 /// This variant is enabled when we must first call hash_value and use the 382 /// result as our data. 383 template <typename T> 384 typename enable_if_c<!is_hashable_data<T>::value, size_t>::type 385 get_hashable_data(const T &value) { 386 using ::llvm::hash_value; 387 return hash_value(value); 388 } 389 390 /// \brief Helper to store data from a value into a buffer and advance the 391 /// pointer into that buffer. 392 /// 393 /// This routine first checks whether there is enough space in the provided 394 /// buffer, and if not immediately returns false. If there is space, it 395 /// copies the underlying bytes of value into the buffer, advances the 396 /// buffer_ptr past the copied bytes, and returns true. 397 template <typename T> 398 bool store_and_advance(char *&buffer_ptr, char *buffer_end, const T& value, 399 size_t offset = 0) { 400 size_t store_size = sizeof(value) - offset; 401 if (buffer_ptr + store_size > buffer_end) 402 return false; 403 const char *value_data = reinterpret_cast<const char *>(&value); 404 memcpy(buffer_ptr, value_data + offset, store_size); 405 buffer_ptr += store_size; 406 return true; 407 } 408 409 /// \brief Implement the combining of integral values into a hash_code. 410 /// 411 /// This overload is selected when the value type of the iterator is 412 /// integral. Rather than computing a hash_code for each object and then 413 /// combining them, this (as an optimization) directly combines the integers. 414 template <typename InputIteratorT> 415 hash_code hash_combine_range_impl(InputIteratorT first, InputIteratorT last) { 416 typedef typename std::iterator_traits<InputIteratorT>::value_type ValueT; 417 const size_t seed = get_execution_seed(); 418 char buffer[64], *buffer_ptr = buffer; 419 char *const buffer_end = buffer_ptr + array_lengthof(buffer); 420 while (first != last && store_and_advance(buffer_ptr, buffer_end, 421 get_hashable_data(*first))) 422 ++first; 423 if (first == last) 424 return hash_short(buffer, buffer_ptr - buffer, seed); 425 assert(buffer_ptr == buffer_end); 426 427 hash_state state = state.create(buffer, seed); 428 size_t length = 64; 429 while (first != last) { 430 // Fill up the buffer. We don't clear it, which re-mixes the last round 431 // when only a partial 64-byte chunk is left. 432 buffer_ptr = buffer; 433 while (first != last && store_and_advance(buffer_ptr, buffer_end, 434 get_hashable_data(*first))) 435 ++first; 436 437 // Rotate the buffer if we did a partial fill in order to simulate doing 438 // a mix of the last 64-bytes. That is how the algorithm works when we 439 // have a contiguous byte sequence, and we want to emulate that here. 440 std::rotate(buffer, buffer_ptr, buffer_end); 441 442 // Mix this chunk into the current state. 443 state.mix(buffer); 444 length += buffer_ptr - buffer; 445 }; 446 447 return state.finalize(length); 448 } 449 450 /// \brief Implement the combining of integral values into a hash_code. 451 /// 452 /// This overload is selected when the value type of the iterator is integral 453 /// and when the input iterator is actually a pointer. Rather than computing 454 /// a hash_code for each object and then combining them, this (as an 455 /// optimization) directly combines the integers. Also, because the integers 456 /// are stored in contiguous memory, this routine avoids copying each value 457 /// and directly reads from the underlying memory. 458 template <typename ValueT> 459 typename enable_if<is_hashable_data<ValueT>, hash_code>::type 460 hash_combine_range_impl(ValueT *first, ValueT *last) { 461 const size_t seed = get_execution_seed(); 462 const char *s_begin = reinterpret_cast<const char *>(first); 463 const char *s_end = reinterpret_cast<const char *>(last); 464 const size_t length = std::distance(s_begin, s_end); 465 if (length <= 64) 466 return hash_short(s_begin, length, seed); 467 468 const char *s_aligned_end = s_begin + (length & ~63); 469 hash_state state = state.create(s_begin, seed); 470 s_begin += 64; 471 while (s_begin != s_aligned_end) { 472 state.mix(s_begin); 473 s_begin += 64; 474 } 475 if (length & 63) 476 state.mix(s_end - 64); 477 478 return state.finalize(length); 479 } 480 481 } // namespace detail 482 } // namespace hashing 483 484 485 /// \brief Compute a hash_code for a sequence of values. 486 /// 487 /// This hashes a sequence of values. It produces the same hash_code as 488 /// 'hash_combine(a, b, c, ...)', but can run over arbitrary sized sequences 489 /// and is significantly faster given pointers and types which can be hashed as 490 /// a sequence of bytes. 491 template <typename InputIteratorT> 492 hash_code hash_combine_range(InputIteratorT first, InputIteratorT last) { 493 return ::llvm::hashing::detail::hash_combine_range_impl(first, last); 494 } 495 496 497 // Implementation details for hash_combine. 498 namespace hashing { 499 namespace detail { 500 501 /// \brief Helper class to manage the recursive combining of hash_combine 502 /// arguments. 503 /// 504 /// This class exists to manage the state and various calls involved in the 505 /// recursive combining of arguments used in hash_combine. It is particularly 506 /// useful at minimizing the code in the recursive calls to ease the pain 507 /// caused by a lack of variadic functions. 508 struct hash_combine_recursive_helper { 509 char buffer[64]; 510 hash_state state; 511 const size_t seed; 512 513 public: 514 /// \brief Construct a recursive hash combining helper. 515 /// 516 /// This sets up the state for a recursive hash combine, including getting 517 /// the seed and buffer setup. 518 hash_combine_recursive_helper() 519 : seed(get_execution_seed()) {} 520 521 /// \brief Combine one chunk of data into the current in-flight hash. 522 /// 523 /// This merges one chunk of data into the hash. First it tries to buffer 524 /// the data. If the buffer is full, it hashes the buffer into its 525 /// hash_state, empties it, and then merges the new chunk in. This also 526 /// handles cases where the data straddles the end of the buffer. 527 template <typename T> 528 char *combine_data(size_t &length, char *buffer_ptr, char *buffer_end, T data) { 529 if (!store_and_advance(buffer_ptr, buffer_end, data)) { 530 // Check for skew which prevents the buffer from being packed, and do 531 // a partial store into the buffer to fill it. This is only a concern 532 // with the variadic combine because that formation can have varying 533 // argument types. 534 size_t partial_store_size = buffer_end - buffer_ptr; 535 memcpy(buffer_ptr, &data, partial_store_size); 536 537 // If the store fails, our buffer is full and ready to hash. We have to 538 // either initialize the hash state (on the first full buffer) or mix 539 // this buffer into the existing hash state. Length tracks the *hashed* 540 // length, not the buffered length. 541 if (length == 0) { 542 state = state.create(buffer, seed); 543 length = 64; 544 } else { 545 // Mix this chunk into the current state and bump length up by 64. 546 state.mix(buffer); 547 length += 64; 548 } 549 // Reset the buffer_ptr to the head of the buffer for the next chunk of 550 // data. 551 buffer_ptr = buffer; 552 553 // Try again to store into the buffer -- this cannot fail as we only 554 // store types smaller than the buffer. 555 if (!store_and_advance(buffer_ptr, buffer_end, data, 556 partial_store_size)) 557 abort(); 558 } 559 return buffer_ptr; 560 } 561 562 #if defined(__has_feature) && __has_feature(__cxx_variadic_templates__) 563 564 /// \brief Recursive, variadic combining method. 565 /// 566 /// This function recurses through each argument, combining that argument 567 /// into a single hash. 568 template <typename T, typename ...Ts> 569 hash_code combine(size_t length, char *buffer_ptr, char *buffer_end, 570 const T &arg, const Ts &...args) { 571 buffer_ptr = combine_data(length, buffer_ptr, buffer_end, get_hashable_data(arg)); 572 573 // Recurse to the next argument. 574 return combine(length, buffer_ptr, buffer_end, args...); 575 } 576 577 #else 578 // Manually expanded recursive combining methods. See variadic above for 579 // documentation. 580 581 template <typename T1, typename T2, typename T3, typename T4, typename T5, 582 typename T6> 583 hash_code combine(size_t length, char *buffer_ptr, char *buffer_end, 584 const T1 &arg1, const T2 &arg2, const T3 &arg3, 585 const T4 &arg4, const T5 &arg5, const T6 &arg6) { 586 buffer_ptr = combine_data(length, buffer_ptr, buffer_end, get_hashable_data(arg1)); 587 return combine(length, buffer_ptr, buffer_end, arg2, arg3, arg4, arg5, arg6); 588 } 589 template <typename T1, typename T2, typename T3, typename T4, typename T5> 590 hash_code combine(size_t length, char *buffer_ptr, char *buffer_end, 591 const T1 &arg1, const T2 &arg2, const T3 &arg3, 592 const T4 &arg4, const T5 &arg5) { 593 buffer_ptr = combine_data(length, buffer_ptr, buffer_end, get_hashable_data(arg1)); 594 return combine(length, buffer_ptr, buffer_end, arg2, arg3, arg4, arg5); 595 } 596 template <typename T1, typename T2, typename T3, typename T4> 597 hash_code combine(size_t length, char *buffer_ptr, char *buffer_end, 598 const T1 &arg1, const T2 &arg2, const T3 &arg3, 599 const T4 &arg4) { 600 buffer_ptr = combine_data(length, buffer_ptr, buffer_end, get_hashable_data(arg1)); 601 return combine(length, buffer_ptr, buffer_end, arg2, arg3, arg4); 602 } 603 template <typename T1, typename T2, typename T3> 604 hash_code combine(size_t length, char *buffer_ptr, char *buffer_end, 605 const T1 &arg1, const T2 &arg2, const T3 &arg3) { 606 buffer_ptr = combine_data(length, buffer_ptr, buffer_end, get_hashable_data(arg1)); 607 return combine(length, buffer_ptr, buffer_end, arg2, arg3); 608 } 609 template <typename T1, typename T2> 610 hash_code combine(size_t length, char *buffer_ptr, char *buffer_end, 611 const T1 &arg1, const T2 &arg2) { 612 buffer_ptr = combine_data(length, buffer_ptr, buffer_end, get_hashable_data(arg1)); 613 return combine(length, buffer_ptr, buffer_end, arg2); 614 } 615 template <typename T1> 616 hash_code combine(size_t length, char *buffer_ptr, char *buffer_end, 617 const T1 &arg1) { 618 buffer_ptr = combine_data(length, buffer_ptr, buffer_end, get_hashable_data(arg1)); 619 return combine(length, buffer_ptr, buffer_end); 620 } 621 622 #endif 623 624 /// \brief Base case for recursive, variadic combining. 625 /// 626 /// The base case when combining arguments recursively is reached when all 627 /// arguments have been handled. It flushes the remaining buffer and 628 /// constructs a hash_code. 629 hash_code combine(size_t length, char *buffer_ptr, char *buffer_end) { 630 // Check whether the entire set of values fit in the buffer. If so, we'll 631 // use the optimized short hashing routine and skip state entirely. 632 if (length == 0) 633 return hash_short(buffer, buffer_ptr - buffer, seed); 634 635 // Mix the final buffer, rotating it if we did a partial fill in order to 636 // simulate doing a mix of the last 64-bytes. That is how the algorithm 637 // works when we have a contiguous byte sequence, and we want to emulate 638 // that here. 639 std::rotate(buffer, buffer_ptr, buffer_end); 640 641 // Mix this chunk into the current state. 642 state.mix(buffer); 643 length += buffer_ptr - buffer; 644 645 return state.finalize(length); 646 } 647 }; 648 649 } // namespace detail 650 } // namespace hashing 651 652 653 #if __has_feature(__cxx_variadic_templates__) 654 655 /// \brief Combine values into a single hash_code. 656 /// 657 /// This routine accepts a varying number of arguments of any type. It will 658 /// attempt to combine them into a single hash_code. For user-defined types it 659 /// attempts to call a \see hash_value overload (via ADL) for the type. For 660 /// integer and pointer types it directly combines their data into the 661 /// resulting hash_code. 662 /// 663 /// The result is suitable for returning from a user's hash_value 664 /// *implementation* for their user-defined type. Consumers of a type should 665 /// *not* call this routine, they should instead call 'hash_value'. 666 template <typename ...Ts> hash_code hash_combine(const Ts &...args) { 667 // Recursively hash each argument using a helper class. 668 ::llvm::hashing::detail::hash_combine_recursive_helper helper; 669 return helper.combine(0, helper.buffer, helper.buffer + 64, args...); 670 } 671 672 #else 673 674 // What follows are manually exploded overloads for each argument width. See 675 // the above variadic definition for documentation and specification. 676 677 template <typename T1, typename T2, typename T3, typename T4, typename T5, 678 typename T6> 679 hash_code hash_combine(const T1 &arg1, const T2 &arg2, const T3 &arg3, 680 const T4 &arg4, const T5 &arg5, const T6 &arg6) { 681 ::llvm::hashing::detail::hash_combine_recursive_helper helper; 682 return helper.combine(0, helper.buffer, helper.buffer + 64, 683 arg1, arg2, arg3, arg4, arg5, arg6); 684 } 685 template <typename T1, typename T2, typename T3, typename T4, typename T5> 686 hash_code hash_combine(const T1 &arg1, const T2 &arg2, const T3 &arg3, 687 const T4 &arg4, const T5 &arg5) { 688 ::llvm::hashing::detail::hash_combine_recursive_helper helper; 689 return helper.combine(0, helper.buffer, helper.buffer + 64, 690 arg1, arg2, arg3, arg4, arg5); 691 } 692 template <typename T1, typename T2, typename T3, typename T4> 693 hash_code hash_combine(const T1 &arg1, const T2 &arg2, const T3 &arg3, 694 const T4 &arg4) { 695 ::llvm::hashing::detail::hash_combine_recursive_helper helper; 696 return helper.combine(0, helper.buffer, helper.buffer + 64, 697 arg1, arg2, arg3, arg4); 698 } 699 template <typename T1, typename T2, typename T3> 700 hash_code hash_combine(const T1 &arg1, const T2 &arg2, const T3 &arg3) { 701 ::llvm::hashing::detail::hash_combine_recursive_helper helper; 702 return helper.combine(0, helper.buffer, helper.buffer + 64, arg1, arg2, arg3); 703 } 704 template <typename T1, typename T2> 705 hash_code hash_combine(const T1 &arg1, const T2 &arg2) { 706 ::llvm::hashing::detail::hash_combine_recursive_helper helper; 707 return helper.combine(0, helper.buffer, helper.buffer + 64, arg1, arg2); 708 } 709 template <typename T1> 710 hash_code hash_combine(const T1 &arg1) { 711 ::llvm::hashing::detail::hash_combine_recursive_helper helper; 712 return helper.combine(0, helper.buffer, helper.buffer + 64, arg1); 713 } 714 715 #endif 716 717 718 // Implementation details for implementatinos of hash_value overloads provided 719 // here. 720 namespace hashing { 721 namespace detail { 722 723 /// \brief Helper to hash the value of a single integer. 724 /// 725 /// Overloads for smaller integer types are not provided to ensure consistent 726 /// behavior in the presence of integral promotions. Essentially, 727 /// "hash_value('4')" and "hash_value('0' + 4)" should be the same. 728 inline hash_code hash_integer_value(uint64_t value) { 729 // Similar to hash_4to8_bytes but using a seed instead of length. 730 const uint64_t seed = get_execution_seed(); 731 const char *s = reinterpret_cast<const char *>(&value); 732 const uint64_t a = fetch32(s); 733 return hash_16_bytes(seed + (a << 3), fetch32(s + 4)); 734 } 735 736 } // namespace detail 737 } // namespace hashing 738 739 // Declared and documented above, but defined here so that any of the hashing 740 // infrastructure is available. 741 template <typename T> 742 typename enable_if<is_integral_or_enum<T>, hash_code>::type 743 hash_value(T value) { 744 return ::llvm::hashing::detail::hash_integer_value(value); 745 } 746 747 // Declared and documented above, but defined here so that any of the hashing 748 // infrastructure is available. 749 template <typename T> hash_code hash_value(const T *ptr) { 750 return ::llvm::hashing::detail::hash_integer_value( 751 reinterpret_cast<uintptr_t>(ptr)); 752 } 753 754 // Declared and documented above, but defined here so that any of the hashing 755 // infrastructure is available. 756 template <typename T, typename U> 757 hash_code hash_value(const std::pair<T, U> &arg) { 758 return hash_combine(arg.first, arg.second); 759 } 760 761 // Declared and documented above, but defined here so that any of the hashing 762 // infrastructure is available. 763 template <typename T> 764 hash_code hash_value(const std::basic_string<T> &arg) { 765 return hash_combine_range(arg.begin(), arg.end()); 766 } 767 768 } // namespace llvm 769 770 #endif 771