LLVM API Documentation

Hashing.h
Go to the documentation of this file.
00001 //===-- llvm/ADT/Hashing.h - Utilities for hashing --------------*- C++ -*-===//
00002 //
00003 //                     The LLVM Compiler Infrastructure
00004 //
00005 // This file is distributed under the University of Illinois Open Source
00006 // License. See LICENSE.TXT for details.
00007 //
00008 //===----------------------------------------------------------------------===//
00009 //
00010 // This file implements the newly proposed standard C++ interfaces for hashing
00011 // arbitrary data and building hash functions for user-defined types. This
00012 // interface was originally proposed in N3333[1] and is currently under review
00013 // for inclusion in a future TR and/or standard.
00014 //
00015 // The primary interfaces provide are comprised of one type and three functions:
00016 //
00017 //  -- 'hash_code' class is an opaque type representing the hash code for some
00018 //     data. It is the intended product of hashing, and can be used to implement
00019 //     hash tables, checksumming, and other common uses of hashes. It is not an
00020 //     integer type (although it can be converted to one) because it is risky
00021 //     to assume much about the internals of a hash_code. In particular, each
00022 //     execution of the program has a high probability of producing a different
00023 //     hash_code for a given input. Thus their values are not stable to save or
00024 //     persist, and should only be used during the execution for the
00025 //     construction of hashing datastructures.
00026 //
00027 //  -- 'hash_value' is a function designed to be overloaded for each
00028 //     user-defined type which wishes to be used within a hashing context. It
00029 //     should be overloaded within the user-defined type's namespace and found
00030 //     via ADL. Overloads for primitive types are provided by this library.
00031 //
00032 //  -- 'hash_combine' and 'hash_combine_range' are functions designed to aid
00033 //      programmers in easily and intuitively combining a set of data into
00034 //      a single hash_code for their object. They should only logically be used
00035 //      within the implementation of a 'hash_value' routine or similar context.
00036 //
00037 // Note that 'hash_combine_range' contains very special logic for hashing
00038 // a contiguous array of integers or pointers. This logic is *extremely* fast,
00039 // on a modern Intel "Gainestown" Xeon (Nehalem uarch) @2.2 GHz, these were
00040 // benchmarked at over 6.5 GiB/s for large keys, and <20 cycles/hash for keys
00041 // under 32-bytes.
00042 //
00043 //===----------------------------------------------------------------------===//
00044 
00045 #ifndef LLVM_ADT_HASHING_H
00046 #define LLVM_ADT_HASHING_H
00047 
00048 #include "llvm/Support/DataTypes.h"
00049 #include "llvm/Support/Host.h"
00050 #include "llvm/Support/SwapByteOrder.h"
00051 #include "llvm/Support/type_traits.h"
00052 #include <algorithm>
00053 #include <cassert>
00054 #include <cstring>
00055 #include <iterator>
00056 #include <utility>
00057 
00058 // Allow detecting C++11 feature availability when building with Clang without
00059 // breaking other compilers.
00060 #ifndef __has_feature
00061 # define __has_feature(x) 0
00062 #endif
00063 
00064 namespace llvm {
00065 
00066 /// \brief An opaque object representing a hash code.
00067 ///
00068 /// This object represents the result of hashing some entity. It is intended to
00069 /// be used to implement hashtables or other hashing-based data structures.
00070 /// While it wraps and exposes a numeric value, this value should not be
00071 /// trusted to be stable or predictable across processes or executions.
00072 ///
00073 /// In order to obtain the hash_code for an object 'x':
00074 /// \code
00075 ///   using llvm::hash_value;
00076 ///   llvm::hash_code code = hash_value(x);
00077 /// \endcode
00078 class hash_code {
00079   size_t value;
00080 
00081 public:
00082   /// \brief Default construct a hash_code.
00083   /// Note that this leaves the value uninitialized.
00084   hash_code() {}
00085 
00086   /// \brief Form a hash code directly from a numerical value.
00087   hash_code(size_t value) : value(value) {}
00088 
00089   /// \brief Convert the hash code to its numerical value for use.
00090   /*explicit*/ operator size_t() const { return value; }
00091 
00092   friend bool operator==(const hash_code &lhs, const hash_code &rhs) {
00093     return lhs.value == rhs.value;
00094   }
00095   friend bool operator!=(const hash_code &lhs, const hash_code &rhs) {
00096     return lhs.value != rhs.value;
00097   }
00098 
00099   /// \brief Allow a hash_code to be directly run through hash_value.
00100   friend size_t hash_value(const hash_code &code) { return code.value; }
00101 };
00102 
00103 /// \brief Compute a hash_code for any integer value.
00104 ///
00105 /// Note that this function is intended to compute the same hash_code for
00106 /// a particular value without regard to the pre-promotion type. This is in
00107 /// contrast to hash_combine which may produce different hash_codes for
00108 /// differing argument types even if they would implicit promote to a common
00109 /// type without changing the value.
00110 template <typename T>
00111 typename std::enable_if<is_integral_or_enum<T>::value, hash_code>::type
00112 hash_value(T value);
00113 
00114 /// \brief Compute a hash_code for a pointer's address.
00115 ///
00116 /// N.B.: This hashes the *address*. Not the value and not the type.
00117 template <typename T> hash_code hash_value(const T *ptr);
00118 
00119 /// \brief Compute a hash_code for a pair of objects.
00120 template <typename T, typename U>
00121 hash_code hash_value(const std::pair<T, U> &arg);
00122 
00123 /// \brief Compute a hash_code for a standard string.
00124 template <typename T>
00125 hash_code hash_value(const std::basic_string<T> &arg);
00126 
00127 
00128 /// \brief Override the execution seed with a fixed value.
00129 ///
00130 /// This hashing library uses a per-execution seed designed to change on each
00131 /// run with high probability in order to ensure that the hash codes are not
00132 /// attackable and to ensure that output which is intended to be stable does
00133 /// not rely on the particulars of the hash codes produced.
00134 ///
00135 /// That said, there are use cases where it is important to be able to
00136 /// reproduce *exactly* a specific behavior. To that end, we provide a function
00137 /// which will forcibly set the seed to a fixed value. This must be done at the
00138 /// start of the program, before any hashes are computed. Also, it cannot be
00139 /// undone. This makes it thread-hostile and very hard to use outside of
00140 /// immediately on start of a simple program designed for reproducible
00141 /// behavior.
00142 void set_fixed_execution_hash_seed(size_t fixed_value);
00143 
00144 
00145 // All of the implementation details of actually computing the various hash
00146 // code values are held within this namespace. These routines are included in
00147 // the header file mainly to allow inlining and constant propagation.
00148 namespace hashing {
00149 namespace detail {
00150 
00151 inline uint64_t fetch64(const char *p) {
00152   uint64_t result;
00153   memcpy(&result, p, sizeof(result));
00154   if (sys::IsBigEndianHost)
00155     return sys::SwapByteOrder(result);
00156   return result;
00157 }
00158 
00159 inline uint32_t fetch32(const char *p) {
00160   uint32_t result;
00161   memcpy(&result, p, sizeof(result));
00162   if (sys::IsBigEndianHost)
00163     return sys::SwapByteOrder(result);
00164   return result;
00165 }
00166 
00167 /// Some primes between 2^63 and 2^64 for various uses.
00168 static const uint64_t k0 = 0xc3a5c85c97cb3127ULL;
00169 static const uint64_t k1 = 0xb492b66fbe98f273ULL;
00170 static const uint64_t k2 = 0x9ae16a3b2f90404fULL;
00171 static const uint64_t k3 = 0xc949d7c7509e6557ULL;
00172 
00173 /// \brief Bitwise right rotate.
00174 /// Normally this will compile to a single instruction, especially if the
00175 /// shift is a manifest constant.
00176 inline uint64_t rotate(uint64_t val, size_t shift) {
00177   // Avoid shifting by 64: doing so yields an undefined result.
00178   return shift == 0 ? val : ((val >> shift) | (val << (64 - shift)));
00179 }
00180 
00181 inline uint64_t shift_mix(uint64_t val) {
00182   return val ^ (val >> 47);
00183 }
00184 
00185 inline uint64_t hash_16_bytes(uint64_t low, uint64_t high) {
00186   // Murmur-inspired hashing.
00187   const uint64_t kMul = 0x9ddfea08eb382d69ULL;
00188   uint64_t a = (low ^ high) * kMul;
00189   a ^= (a >> 47);
00190   uint64_t b = (high ^ a) * kMul;
00191   b ^= (b >> 47);
00192   b *= kMul;
00193   return b;
00194 }
00195 
00196 inline uint64_t hash_1to3_bytes(const char *s, size_t len, uint64_t seed) {
00197   uint8_t a = s[0];
00198   uint8_t b = s[len >> 1];
00199   uint8_t c = s[len - 1];
00200   uint32_t y = static_cast<uint32_t>(a) + (static_cast<uint32_t>(b) << 8);
00201   uint32_t z = len + (static_cast<uint32_t>(c) << 2);
00202   return shift_mix(y * k2 ^ z * k3 ^ seed) * k2;
00203 }
00204 
00205 inline uint64_t hash_4to8_bytes(const char *s, size_t len, uint64_t seed) {
00206   uint64_t a = fetch32(s);
00207   return hash_16_bytes(len + (a << 3), seed ^ fetch32(s + len - 4));
00208 }
00209 
00210 inline uint64_t hash_9to16_bytes(const char *s, size_t len, uint64_t seed) {
00211   uint64_t a = fetch64(s);
00212   uint64_t b = fetch64(s + len - 8);
00213   return hash_16_bytes(seed ^ a, rotate(b + len, len)) ^ b;
00214 }
00215 
00216 inline uint64_t hash_17to32_bytes(const char *s, size_t len, uint64_t seed) {
00217   uint64_t a = fetch64(s) * k1;
00218   uint64_t b = fetch64(s + 8);
00219   uint64_t c = fetch64(s + len - 8) * k2;
00220   uint64_t d = fetch64(s + len - 16) * k0;
00221   return hash_16_bytes(rotate(a - b, 43) + rotate(c ^ seed, 30) + d,
00222                        a + rotate(b ^ k3, 20) - c + len + seed);
00223 }
00224 
00225 inline uint64_t hash_33to64_bytes(const char *s, size_t len, uint64_t seed) {
00226   uint64_t z = fetch64(s + 24);
00227   uint64_t a = fetch64(s) + (len + fetch64(s + len - 16)) * k0;
00228   uint64_t b = rotate(a + z, 52);
00229   uint64_t c = rotate(a, 37);
00230   a += fetch64(s + 8);
00231   c += rotate(a, 7);
00232   a += fetch64(s + 16);
00233   uint64_t vf = a + z;
00234   uint64_t vs = b + rotate(a, 31) + c;
00235   a = fetch64(s + 16) + fetch64(s + len - 32);
00236   z = fetch64(s + len - 8);
00237   b = rotate(a + z, 52);
00238   c = rotate(a, 37);
00239   a += fetch64(s + len - 24);
00240   c += rotate(a, 7);
00241   a += fetch64(s + len - 16);
00242   uint64_t wf = a + z;
00243   uint64_t ws = b + rotate(a, 31) + c;
00244   uint64_t r = shift_mix((vf + ws) * k2 + (wf + vs) * k0);
00245   return shift_mix((seed ^ (r * k0)) + vs) * k2;
00246 }
00247 
00248 inline uint64_t hash_short(const char *s, size_t length, uint64_t seed) {
00249   if (length >= 4 && length <= 8)
00250     return hash_4to8_bytes(s, length, seed);
00251   if (length > 8 && length <= 16)
00252     return hash_9to16_bytes(s, length, seed);
00253   if (length > 16 && length <= 32)
00254     return hash_17to32_bytes(s, length, seed);
00255   if (length > 32)
00256     return hash_33to64_bytes(s, length, seed);
00257   if (length != 0)
00258     return hash_1to3_bytes(s, length, seed);
00259 
00260   return k2 ^ seed;
00261 }
00262 
00263 /// \brief The intermediate state used during hashing.
00264 /// Currently, the algorithm for computing hash codes is based on CityHash and
00265 /// keeps 56 bytes of arbitrary state.
00266 struct hash_state {
00267   uint64_t h0, h1, h2, h3, h4, h5, h6;
00268   uint64_t seed;
00269 
00270   /// \brief Create a new hash_state structure and initialize it based on the
00271   /// seed and the first 64-byte chunk.
00272   /// This effectively performs the initial mix.
00273   static hash_state create(const char *s, uint64_t seed) {
00274     hash_state state = {
00275       0, seed, hash_16_bytes(seed, k1), rotate(seed ^ k1, 49),
00276       seed * k1, shift_mix(seed), 0, seed };
00277     state.h6 = hash_16_bytes(state.h4, state.h5);
00278     state.mix(s);
00279     return state;
00280   }
00281 
00282   /// \brief Mix 32-bytes from the input sequence into the 16-bytes of 'a'
00283   /// and 'b', including whatever is already in 'a' and 'b'.
00284   static void mix_32_bytes(const char *s, uint64_t &a, uint64_t &b) {
00285     a += fetch64(s);
00286     uint64_t c = fetch64(s + 24);
00287     b = rotate(b + a + c, 21);
00288     uint64_t d = a;
00289     a += fetch64(s + 8) + fetch64(s + 16);
00290     b += rotate(a, 44) + d;
00291     a += c;
00292   }
00293 
00294   /// \brief Mix in a 64-byte buffer of data.
00295   /// We mix all 64 bytes even when the chunk length is smaller, but we
00296   /// record the actual length.
00297   void mix(const char *s) {
00298     h0 = rotate(h0 + h1 + h3 + fetch64(s + 8), 37) * k1;
00299     h1 = rotate(h1 + h4 + fetch64(s + 48), 42) * k1;
00300     h0 ^= h6;
00301     h1 += h3 + fetch64(s + 40);
00302     h2 = rotate(h2 + h5, 33) * k1;
00303     h3 = h4 * k1;
00304     h4 = h0 + h5;
00305     mix_32_bytes(s, h3, h4);
00306     h5 = h2 + h6;
00307     h6 = h1 + fetch64(s + 16);
00308     mix_32_bytes(s + 32, h5, h6);
00309     std::swap(h2, h0);
00310   }
00311 
00312   /// \brief Compute the final 64-bit hash code value based on the current
00313   /// state and the length of bytes hashed.
00314   uint64_t finalize(size_t length) {
00315     return hash_16_bytes(hash_16_bytes(h3, h5) + shift_mix(h1) * k1 + h2,
00316                          hash_16_bytes(h4, h6) + shift_mix(length) * k1 + h0);
00317   }
00318 };
00319 
00320 
00321 /// \brief A global, fixed seed-override variable.
00322 ///
00323 /// This variable can be set using the \see llvm::set_fixed_execution_seed
00324 /// function. See that function for details. Do not, under any circumstances,
00325 /// set or read this variable.
00326 extern size_t fixed_seed_override;
00327 
00328 inline size_t get_execution_seed() {
00329   // FIXME: This needs to be a per-execution seed. This is just a placeholder
00330   // implementation. Switching to a per-execution seed is likely to flush out
00331   // instability bugs and so will happen as its own commit.
00332   //
00333   // However, if there is a fixed seed override set the first time this is
00334   // called, return that instead of the per-execution seed.
00335   const uint64_t seed_prime = 0xff51afd7ed558ccdULL;
00336   static size_t seed = fixed_seed_override ? fixed_seed_override
00337                                            : (size_t)seed_prime;
00338   return seed;
00339 }
00340 
00341 
00342 /// \brief Trait to indicate whether a type's bits can be hashed directly.
00343 ///
00344 /// A type trait which is true if we want to combine values for hashing by
00345 /// reading the underlying data. It is false if values of this type must
00346 /// first be passed to hash_value, and the resulting hash_codes combined.
00347 //
00348 // FIXME: We want to replace is_integral_or_enum and is_pointer here with
00349 // a predicate which asserts that comparing the underlying storage of two
00350 // values of the type for equality is equivalent to comparing the two values
00351 // for equality. For all the platforms we care about, this holds for integers
00352 // and pointers, but there are platforms where it doesn't and we would like to
00353 // support user-defined types which happen to satisfy this property.
00354 template <typename T> struct is_hashable_data
00355   : std::integral_constant<bool, ((is_integral_or_enum<T>::value ||
00356                                    std::is_pointer<T>::value) &&
00357                                   64 % sizeof(T) == 0)> {};
00358 
00359 // Special case std::pair to detect when both types are viable and when there
00360 // is no alignment-derived padding in the pair. This is a bit of a lie because
00361 // std::pair isn't truly POD, but it's close enough in all reasonable
00362 // implementations for our use case of hashing the underlying data.
00363 template <typename T, typename U> struct is_hashable_data<std::pair<T, U> >
00364   : std::integral_constant<bool, (is_hashable_data<T>::value &&
00365                                   is_hashable_data<U>::value &&
00366                                   (sizeof(T) + sizeof(U)) ==
00367                                    sizeof(std::pair<T, U>))> {};
00368 
00369 /// \brief Helper to get the hashable data representation for a type.
00370 /// This variant is enabled when the type itself can be used.
00371 template <typename T>
00372 typename std::enable_if<is_hashable_data<T>::value, T>::type
00373 get_hashable_data(const T &value) {
00374   return value;
00375 }
00376 /// \brief Helper to get the hashable data representation for a type.
00377 /// This variant is enabled when we must first call hash_value and use the
00378 /// result as our data.
00379 template <typename T>
00380 typename std::enable_if<!is_hashable_data<T>::value, size_t>::type
00381 get_hashable_data(const T &value) {
00382   using ::llvm::hash_value;
00383   return hash_value(value);
00384 }
00385 
00386 /// \brief Helper to store data from a value into a buffer and advance the
00387 /// pointer into that buffer.
00388 ///
00389 /// This routine first checks whether there is enough space in the provided
00390 /// buffer, and if not immediately returns false. If there is space, it
00391 /// copies the underlying bytes of value into the buffer, advances the
00392 /// buffer_ptr past the copied bytes, and returns true.
00393 template <typename T>
00394 bool store_and_advance(char *&buffer_ptr, char *buffer_end, const T& value,
00395                        size_t offset = 0) {
00396   size_t store_size = sizeof(value) - offset;
00397   if (buffer_ptr + store_size > buffer_end)
00398     return false;
00399   const char *value_data = reinterpret_cast<const char *>(&value);
00400   memcpy(buffer_ptr, value_data + offset, store_size);
00401   buffer_ptr += store_size;
00402   return true;
00403 }
00404 
00405 /// \brief Implement the combining of integral values into a hash_code.
00406 ///
00407 /// This overload is selected when the value type of the iterator is
00408 /// integral. Rather than computing a hash_code for each object and then
00409 /// combining them, this (as an optimization) directly combines the integers.
00410 template <typename InputIteratorT>
00411 hash_code hash_combine_range_impl(InputIteratorT first, InputIteratorT last) {
00412   const size_t seed = get_execution_seed();
00413   char buffer[64], *buffer_ptr = buffer;
00414   char *const buffer_end = std::end(buffer);
00415   while (first != last && store_and_advance(buffer_ptr, buffer_end,
00416                                             get_hashable_data(*first)))
00417     ++first;
00418   if (first == last)
00419     return hash_short(buffer, buffer_ptr - buffer, seed);
00420   assert(buffer_ptr == buffer_end);
00421 
00422   hash_state state = state.create(buffer, seed);
00423   size_t length = 64;
00424   while (first != last) {
00425     // Fill up the buffer. We don't clear it, which re-mixes the last round
00426     // when only a partial 64-byte chunk is left.
00427     buffer_ptr = buffer;
00428     while (first != last && store_and_advance(buffer_ptr, buffer_end,
00429                                               get_hashable_data(*first)))
00430       ++first;
00431 
00432     // Rotate the buffer if we did a partial fill in order to simulate doing
00433     // a mix of the last 64-bytes. That is how the algorithm works when we
00434     // have a contiguous byte sequence, and we want to emulate that here.
00435     std::rotate(buffer, buffer_ptr, buffer_end);
00436 
00437     // Mix this chunk into the current state.
00438     state.mix(buffer);
00439     length += buffer_ptr - buffer;
00440   };
00441 
00442   return state.finalize(length);
00443 }
00444 
00445 /// \brief Implement the combining of integral values into a hash_code.
00446 ///
00447 /// This overload is selected when the value type of the iterator is integral
00448 /// and when the input iterator is actually a pointer. Rather than computing
00449 /// a hash_code for each object and then combining them, this (as an
00450 /// optimization) directly combines the integers. Also, because the integers
00451 /// are stored in contiguous memory, this routine avoids copying each value
00452 /// and directly reads from the underlying memory.
00453 template <typename ValueT>
00454 typename std::enable_if<is_hashable_data<ValueT>::value, hash_code>::type
00455 hash_combine_range_impl(ValueT *first, ValueT *last) {
00456   const size_t seed = get_execution_seed();
00457   const char *s_begin = reinterpret_cast<const char *>(first);
00458   const char *s_end = reinterpret_cast<const char *>(last);
00459   const size_t length = std::distance(s_begin, s_end);
00460   if (length <= 64)
00461     return hash_short(s_begin, length, seed);
00462 
00463   const char *s_aligned_end = s_begin + (length & ~63);
00464   hash_state state = state.create(s_begin, seed);
00465   s_begin += 64;
00466   while (s_begin != s_aligned_end) {
00467     state.mix(s_begin);
00468     s_begin += 64;
00469   }
00470   if (length & 63)
00471     state.mix(s_end - 64);
00472 
00473   return state.finalize(length);
00474 }
00475 
00476 } // namespace detail
00477 } // namespace hashing
00478 
00479 
00480 /// \brief Compute a hash_code for a sequence of values.
00481 ///
00482 /// This hashes a sequence of values. It produces the same hash_code as
00483 /// 'hash_combine(a, b, c, ...)', but can run over arbitrary sized sequences
00484 /// and is significantly faster given pointers and types which can be hashed as
00485 /// a sequence of bytes.
00486 template <typename InputIteratorT>
00487 hash_code hash_combine_range(InputIteratorT first, InputIteratorT last) {
00488   return ::llvm::hashing::detail::hash_combine_range_impl(first, last);
00489 }
00490 
00491 
00492 // Implementation details for hash_combine.
00493 namespace hashing {
00494 namespace detail {
00495 
00496 /// \brief Helper class to manage the recursive combining of hash_combine
00497 /// arguments.
00498 ///
00499 /// This class exists to manage the state and various calls involved in the
00500 /// recursive combining of arguments used in hash_combine. It is particularly
00501 /// useful at minimizing the code in the recursive calls to ease the pain
00502 /// caused by a lack of variadic functions.
00503 struct hash_combine_recursive_helper {
00504   char buffer[64];
00505   hash_state state;
00506   const size_t seed;
00507 
00508 public:
00509   /// \brief Construct a recursive hash combining helper.
00510   ///
00511   /// This sets up the state for a recursive hash combine, including getting
00512   /// the seed and buffer setup.
00513   hash_combine_recursive_helper()
00514     : seed(get_execution_seed()) {}
00515 
00516   /// \brief Combine one chunk of data into the current in-flight hash.
00517   ///
00518   /// This merges one chunk of data into the hash. First it tries to buffer
00519   /// the data. If the buffer is full, it hashes the buffer into its
00520   /// hash_state, empties it, and then merges the new chunk in. This also
00521   /// handles cases where the data straddles the end of the buffer.
00522   template <typename T>
00523   char *combine_data(size_t &length, char *buffer_ptr, char *buffer_end, T data) {
00524     if (!store_and_advance(buffer_ptr, buffer_end, data)) {
00525       // Check for skew which prevents the buffer from being packed, and do
00526       // a partial store into the buffer to fill it. This is only a concern
00527       // with the variadic combine because that formation can have varying
00528       // argument types.
00529       size_t partial_store_size = buffer_end - buffer_ptr;
00530       memcpy(buffer_ptr, &data, partial_store_size);
00531 
00532       // If the store fails, our buffer is full and ready to hash. We have to
00533       // either initialize the hash state (on the first full buffer) or mix
00534       // this buffer into the existing hash state. Length tracks the *hashed*
00535       // length, not the buffered length.
00536       if (length == 0) {
00537         state = state.create(buffer, seed);
00538         length = 64;
00539       } else {
00540         // Mix this chunk into the current state and bump length up by 64.
00541         state.mix(buffer);
00542         length += 64;
00543       }
00544       // Reset the buffer_ptr to the head of the buffer for the next chunk of
00545       // data.
00546       buffer_ptr = buffer;
00547 
00548       // Try again to store into the buffer -- this cannot fail as we only
00549       // store types smaller than the buffer.
00550       if (!store_and_advance(buffer_ptr, buffer_end, data,
00551                              partial_store_size))
00552         abort();
00553     }
00554     return buffer_ptr;
00555   }
00556 
00557 #if defined(__has_feature) && __has_feature(__cxx_variadic_templates__)
00558 
00559   /// \brief Recursive, variadic combining method.
00560   ///
00561   /// This function recurses through each argument, combining that argument
00562   /// into a single hash.
00563   template <typename T, typename ...Ts>
00564   hash_code combine(size_t length, char *buffer_ptr, char *buffer_end,
00565                     const T &arg, const Ts &...args) {
00566     buffer_ptr = combine_data(length, buffer_ptr, buffer_end, get_hashable_data(arg));
00567 
00568     // Recurse to the next argument.
00569     return combine(length, buffer_ptr, buffer_end, args...);
00570   }
00571 
00572 #else
00573   // Manually expanded recursive combining methods. See variadic above for
00574   // documentation.
00575 
00576   template <typename T1, typename T2, typename T3, typename T4, typename T5,
00577             typename T6>
00578   hash_code combine(size_t length, char *buffer_ptr, char *buffer_end,
00579                     const T1 &arg1, const T2 &arg2, const T3 &arg3,
00580                     const T4 &arg4, const T5 &arg5, const T6 &arg6) {
00581     buffer_ptr = combine_data(length, buffer_ptr, buffer_end, get_hashable_data(arg1));
00582     return combine(length, buffer_ptr, buffer_end, arg2, arg3, arg4, arg5, arg6);
00583   }
00584   template <typename T1, typename T2, typename T3, typename T4, typename T5>
00585   hash_code combine(size_t length, char *buffer_ptr, char *buffer_end,
00586                     const T1 &arg1, const T2 &arg2, const T3 &arg3,
00587                     const T4 &arg4, const T5 &arg5) {
00588     buffer_ptr = combine_data(length, buffer_ptr, buffer_end, get_hashable_data(arg1));
00589     return combine(length, buffer_ptr, buffer_end, arg2, arg3, arg4, arg5);
00590   }
00591   template <typename T1, typename T2, typename T3, typename T4>
00592   hash_code combine(size_t length, char *buffer_ptr, char *buffer_end,
00593                     const T1 &arg1, const T2 &arg2, const T3 &arg3,
00594                     const T4 &arg4) {
00595     buffer_ptr = combine_data(length, buffer_ptr, buffer_end, get_hashable_data(arg1));
00596     return combine(length, buffer_ptr, buffer_end, arg2, arg3, arg4);
00597   }
00598   template <typename T1, typename T2, typename T3>
00599   hash_code combine(size_t length, char *buffer_ptr, char *buffer_end,
00600                     const T1 &arg1, const T2 &arg2, const T3 &arg3) {
00601     buffer_ptr = combine_data(length, buffer_ptr, buffer_end, get_hashable_data(arg1));
00602     return combine(length, buffer_ptr, buffer_end, arg2, arg3);
00603   }
00604   template <typename T1, typename T2>
00605   hash_code combine(size_t length, char *buffer_ptr, char *buffer_end,
00606                     const T1 &arg1, const T2 &arg2) {
00607     buffer_ptr = combine_data(length, buffer_ptr, buffer_end, get_hashable_data(arg1));
00608     return combine(length, buffer_ptr, buffer_end, arg2);
00609   }
00610   template <typename T1>
00611   hash_code combine(size_t length, char *buffer_ptr, char *buffer_end,
00612                     const T1 &arg1) {
00613     buffer_ptr = combine_data(length, buffer_ptr, buffer_end, get_hashable_data(arg1));
00614     return combine(length, buffer_ptr, buffer_end);
00615   }
00616 
00617 #endif
00618 
00619   /// \brief Base case for recursive, variadic combining.
00620   ///
00621   /// The base case when combining arguments recursively is reached when all
00622   /// arguments have been handled. It flushes the remaining buffer and
00623   /// constructs a hash_code.
00624   hash_code combine(size_t length, char *buffer_ptr, char *buffer_end) {
00625     // Check whether the entire set of values fit in the buffer. If so, we'll
00626     // use the optimized short hashing routine and skip state entirely.
00627     if (length == 0)
00628       return hash_short(buffer, buffer_ptr - buffer, seed);
00629 
00630     // Mix the final buffer, rotating it if we did a partial fill in order to
00631     // simulate doing a mix of the last 64-bytes. That is how the algorithm
00632     // works when we have a contiguous byte sequence, and we want to emulate
00633     // that here.
00634     std::rotate(buffer, buffer_ptr, buffer_end);
00635 
00636     // Mix this chunk into the current state.
00637     state.mix(buffer);
00638     length += buffer_ptr - buffer;
00639 
00640     return state.finalize(length);
00641   }
00642 };
00643 
00644 } // namespace detail
00645 } // namespace hashing
00646 
00647 
00648 #if __has_feature(__cxx_variadic_templates__)
00649 
00650 /// \brief Combine values into a single hash_code.
00651 ///
00652 /// This routine accepts a varying number of arguments of any type. It will
00653 /// attempt to combine them into a single hash_code. For user-defined types it
00654 /// attempts to call a \see hash_value overload (via ADL) for the type. For
00655 /// integer and pointer types it directly combines their data into the
00656 /// resulting hash_code.
00657 ///
00658 /// The result is suitable for returning from a user's hash_value
00659 /// *implementation* for their user-defined type. Consumers of a type should
00660 /// *not* call this routine, they should instead call 'hash_value'.
00661 template <typename ...Ts> hash_code hash_combine(const Ts &...args) {
00662   // Recursively hash each argument using a helper class.
00663   ::llvm::hashing::detail::hash_combine_recursive_helper helper;
00664   return helper.combine(0, helper.buffer, helper.buffer + 64, args...);
00665 }
00666 
00667 #else
00668 
00669 // What follows are manually exploded overloads for each argument width. See
00670 // the above variadic definition for documentation and specification.
00671 
00672 template <typename T1, typename T2, typename T3, typename T4, typename T5,
00673           typename T6>
00674 hash_code hash_combine(const T1 &arg1, const T2 &arg2, const T3 &arg3,
00675                        const T4 &arg4, const T5 &arg5, const T6 &arg6) {
00676   ::llvm::hashing::detail::hash_combine_recursive_helper helper;
00677   return helper.combine(0, helper.buffer, helper.buffer + 64,
00678                         arg1, arg2, arg3, arg4, arg5, arg6);
00679 }
00680 template <typename T1, typename T2, typename T3, typename T4, typename T5>
00681 hash_code hash_combine(const T1 &arg1, const T2 &arg2, const T3 &arg3,
00682                        const T4 &arg4, const T5 &arg5) {
00683   ::llvm::hashing::detail::hash_combine_recursive_helper helper;
00684   return helper.combine(0, helper.buffer, helper.buffer + 64,
00685                         arg1, arg2, arg3, arg4, arg5);
00686 }
00687 template <typename T1, typename T2, typename T3, typename T4>
00688 hash_code hash_combine(const T1 &arg1, const T2 &arg2, const T3 &arg3,
00689                        const T4 &arg4) {
00690   ::llvm::hashing::detail::hash_combine_recursive_helper helper;
00691   return helper.combine(0, helper.buffer, helper.buffer + 64,
00692                         arg1, arg2, arg3, arg4);
00693 }
00694 template <typename T1, typename T2, typename T3>
00695 hash_code hash_combine(const T1 &arg1, const T2 &arg2, const T3 &arg3) {
00696   ::llvm::hashing::detail::hash_combine_recursive_helper helper;
00697   return helper.combine(0, helper.buffer, helper.buffer + 64, arg1, arg2, arg3);
00698 }
00699 template <typename T1, typename T2>
00700 hash_code hash_combine(const T1 &arg1, const T2 &arg2) {
00701   ::llvm::hashing::detail::hash_combine_recursive_helper helper;
00702   return helper.combine(0, helper.buffer, helper.buffer + 64, arg1, arg2);
00703 }
00704 template <typename T1>
00705 hash_code hash_combine(const T1 &arg1) {
00706   ::llvm::hashing::detail::hash_combine_recursive_helper helper;
00707   return helper.combine(0, helper.buffer, helper.buffer + 64, arg1);
00708 }
00709 
00710 #endif
00711 
00712 
00713 // Implementation details for implementations of hash_value overloads provided
00714 // here.
00715 namespace hashing {
00716 namespace detail {
00717 
00718 /// \brief Helper to hash the value of a single integer.
00719 ///
00720 /// Overloads for smaller integer types are not provided to ensure consistent
00721 /// behavior in the presence of integral promotions. Essentially,
00722 /// "hash_value('4')" and "hash_value('0' + 4)" should be the same.
00723 inline hash_code hash_integer_value(uint64_t value) {
00724   // Similar to hash_4to8_bytes but using a seed instead of length.
00725   const uint64_t seed = get_execution_seed();
00726   const char *s = reinterpret_cast<const char *>(&value);
00727   const uint64_t a = fetch32(s);
00728   return hash_16_bytes(seed + (a << 3), fetch32(s + 4));
00729 }
00730 
00731 } // namespace detail
00732 } // namespace hashing
00733 
00734 // Declared and documented above, but defined here so that any of the hashing
00735 // infrastructure is available.
00736 template <typename T>
00737 typename std::enable_if<is_integral_or_enum<T>::value, hash_code>::type
00738 hash_value(T value) {
00739   return ::llvm::hashing::detail::hash_integer_value(value);
00740 }
00741 
00742 // Declared and documented above, but defined here so that any of the hashing
00743 // infrastructure is available.
00744 template <typename T> hash_code hash_value(const T *ptr) {
00745   return ::llvm::hashing::detail::hash_integer_value(
00746     reinterpret_cast<uintptr_t>(ptr));
00747 }
00748 
00749 // Declared and documented above, but defined here so that any of the hashing
00750 // infrastructure is available.
00751 template <typename T, typename U>
00752 hash_code hash_value(const std::pair<T, U> &arg) {
00753   return hash_combine(arg.first, arg.second);
00754 }
00755 
00756 // Declared and documented above, but defined here so that any of the hashing
00757 // infrastructure is available.
00758 template <typename T>
00759 hash_code hash_value(const std::basic_string<T> &arg) {
00760   return hash_combine_range(arg.begin(), arg.end());
00761 }
00762 
00763 } // namespace llvm
00764 
00765 #endif