[llvm] [Hashing] Replace CityHash mixers with xxh3 (PR #194567)

Fangrui Song via llvm-commits llvm-commits at lists.llvm.org
Sat May 9 16:27:54 PDT 2026


https://github.com/MaskRay updated https://github.com/llvm/llvm-project/pull/194567

>From 23bca84cab85dbba9efa43dc8ad83632df234576 Mon Sep 17 00:00:00 2001
From: Fangrui Song <i at maskray.me>
Date: Mon, 27 Apr 2026 23:34:13 -0700
Subject: [PATCH 1/2] [Hashing] Replace CityHash mixers with xxh3

Replace the CityHash-style mixer in hash_combine and (transitively)
hash_value(std::basic_string) with a flatten-and-call into xxh3_64bits,
a modern hash superior to CityHash.

hash_value(int) / hash_value(ptr) keep the existing Murmur-style
hash_16_bytes mixer; those are the dominant DenseMap key paths and a
fully-inline 16-byte mix beats inlining xxh3's larger 0..16-byte short
path.

To break dependency cycle: xxHash64, xxh3_64bits, and xxh3_128bits
ArrayRef/StringRef overloads move from llvm/Support/xxhash.h to inline
overloads in llvm/ADT/ArrayRef.h and llvm/ADT/StringRef.h, so xxhash.h
has no ADT dependencies.

A variant that inlined xxh3's 0..16-byte fast path at every
combine_bytes call site (vs. always calling out-of-line xxh3_64bits)
showed no measurable compile-time improvement on the tracker.

llvm-compile-time-tracker.com (CTMark, instructions:u)
```
  stage1-O0-g           -1.64%   (sqlite3 -3.42%)
  stage1-aarch64-O0-g   -1.41%   (sqlite3 -2.71%)
  stage1-ReleaseLTO-g   -1.14%   (tramp3d-v4 -1.64%)
  stage1-O3             -0.44%
  stage1-ReleaseThinLTO -0.43%
  stage2-O3             -0.16%
```

DenseMap-of-pointer paths (dominant at -O3) are untouched, so higher-
optimization configs see smaller wins as expected. opt's .text shrinks
~92 KB. Subsumes the StringRef-only carve-out proposed in #191115.

Aided by Claude opus 4.7
---
 llvm/include/llvm/ADT/ArrayRef.h      |  14 +
 llvm/include/llvm/ADT/FoldingSet.h    |   4 +-
 llvm/include/llvm/ADT/Hashing.h       | 389 ++++----------------------
 llvm/include/llvm/ADT/StableHashing.h |   5 +-
 llvm/include/llvm/ADT/StringRef.h     |  12 +
 llvm/include/llvm/Support/xxhash.h    |  20 +-
 llvm/lib/Support/xxhash.cpp           |  19 +-
 7 files changed, 103 insertions(+), 360 deletions(-)

diff --git a/llvm/include/llvm/ADT/ArrayRef.h b/llvm/include/llvm/ADT/ArrayRef.h
index eafc4330a1b1b..976eb6ee040de 100644
--- a/llvm/include/llvm/ADT/ArrayRef.h
+++ b/llvm/include/llvm/ADT/ArrayRef.h
@@ -13,6 +13,7 @@
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/Support/Compiler.h"
+#include "llvm/Support/xxhash.h"
 #include <algorithm>
 #include <array>
 #include <cassert>
@@ -562,6 +563,19 @@ namespace llvm {
     return hash_combine_range(S);
   }
 
+  /// Inline ArrayRef overloads of the xxhash entry points declared
+  /// out-of-line in llvm/Support/xxhash.h. They live here so xxhash.h can stay
+  /// free of ADT dependencies.
+  inline uint64_t xxHash64(ArrayRef<uint8_t> data) {
+    return xxHash64(data.data(), data.size());
+  }
+  inline uint64_t xxh3_64bits(ArrayRef<uint8_t> data) {
+    return xxh3_64bits(data.data(), data.size());
+  }
+  inline XXH128_hash_t xxh3_128bits(ArrayRef<uint8_t> data) {
+    return xxh3_128bits(data.data(), data.size());
+  }
+
   // Provide DenseMapInfo for ArrayRefs.
   template <typename T> struct DenseMapInfo<ArrayRef<T>, void> {
     static inline ArrayRef<T> getEmptyKey() {
diff --git a/llvm/include/llvm/ADT/FoldingSet.h b/llvm/include/llvm/ADT/FoldingSet.h
index 0322dc5bddfe3..590b7ca76e322 100644
--- a/llvm/include/llvm/ADT/FoldingSet.h
+++ b/llvm/include/llvm/ADT/FoldingSet.h
@@ -187,8 +187,8 @@ class FoldingSetNodeIDRef {
   // Compute a deterministic hash value across processes that is suitable for
   // on-disk serialization.
   unsigned computeStableHash() const {
-    return static_cast<unsigned>(xxh3_64bits(ArrayRef(
-        reinterpret_cast<const uint8_t *>(Data), sizeof(unsigned) * Size)));
+    return static_cast<unsigned>(xxh3_64bits(
+        reinterpret_cast<const uint8_t *>(Data), sizeof(unsigned) * Size));
   }
 
   LLVM_ABI bool operator==(FoldingSetNodeIDRef) const;
diff --git a/llvm/include/llvm/ADT/Hashing.h b/llvm/include/llvm/ADT/Hashing.h
index 0f96b857bff16..937822e783b12 100644
--- a/llvm/include/llvm/ADT/Hashing.h
+++ b/llvm/include/llvm/ADT/Hashing.h
@@ -50,9 +50,11 @@
 #include "llvm/Support/ErrorHandling.h"
 #include "llvm/Support/SwapByteOrder.h"
 #include "llvm/Support/type_traits.h"
-#include <algorithm>
+#include "llvm/Support/xxhash.h"
+#include <array>
 #include <cassert>
 #include <cstring>
+#include <memory>
 #include <optional>
 #include <string>
 #include <tuple>
@@ -136,14 +138,6 @@ template <typename T> hash_code hash_value(const std::optional<T> &arg);
 namespace hashing {
 namespace detail {
 
-inline uint64_t fetch64(const char *p) {
-  uint64_t result;
-  std::memcpy(&result, p, sizeof(result));
-  if (sys::IsBigEndianHost)
-    sys::swapByteOrder(result);
-  return result;
-}
-
 inline uint32_t fetch32(const char *p) {
   uint32_t result;
   std::memcpy(&result, p, sizeof(result));
@@ -152,22 +146,6 @@ inline uint32_t fetch32(const char *p) {
   return result;
 }
 
-/// Some primes between 2^63 and 2^64 for various uses.
-static constexpr uint64_t k0 = 0xc3a5c85c97cb3127ULL;
-static constexpr uint64_t k1 = 0xb492b66fbe98f273ULL;
-static constexpr uint64_t k2 = 0x9ae16a3b2f90404fULL;
-static constexpr uint64_t k3 = 0xc949d7c7509e6557ULL;
-
-/// Bitwise right rotate.
-/// Normally this will compile to a single instruction, especially if the
-/// shift is a manifest constant.
-constexpr uint64_t rotate(uint64_t val, size_t shift) {
-  // Avoid shifting by 64: doing so yields an undefined result.
-  return shift == 0 ? val : ((val >> shift) | (val << (64 - shift)));
-}
-
-constexpr uint64_t shift_mix(uint64_t val) { return val ^ (val >> 47); }
-
 constexpr uint64_t hash_16_bytes(uint64_t low, uint64_t high) {
   // Murmur-inspired hashing.
   const uint64_t kMul = 0x9ddfea08eb382d69ULL;
@@ -179,134 +157,6 @@ constexpr uint64_t hash_16_bytes(uint64_t low, uint64_t high) {
   return b;
 }
 
-constexpr uint64_t hash_1to3_bytes(const char *s, size_t len, uint64_t seed) {
-  uint8_t a = s[0];
-  uint8_t b = s[len >> 1];
-  uint8_t c = s[len - 1];
-  uint32_t y = static_cast<uint32_t>(a) + (static_cast<uint32_t>(b) << 8);
-  uint32_t z = static_cast<uint32_t>(len) + (static_cast<uint32_t>(c) << 2);
-  return shift_mix(y * k2 ^ z * k3 ^ seed) * k2;
-}
-
-inline uint64_t hash_4to8_bytes(const char *s, size_t len, uint64_t seed) {
-  uint64_t a = fetch32(s);
-  return hash_16_bytes(len + (a << 3), seed ^ fetch32(s + len - 4));
-}
-
-inline uint64_t hash_9to16_bytes(const char *s, size_t len, uint64_t seed) {
-  uint64_t a = fetch64(s);
-  uint64_t b = fetch64(s + len - 8);
-  return hash_16_bytes(seed ^ a, rotate(b + len, len)) ^ b;
-}
-
-inline uint64_t hash_17to32_bytes(const char *s, size_t len, uint64_t seed) {
-  uint64_t a = fetch64(s) * k1;
-  uint64_t b = fetch64(s + 8);
-  uint64_t c = fetch64(s + len - 8) * k2;
-  uint64_t d = fetch64(s + len - 16) * k0;
-  return hash_16_bytes(llvm::rotr<uint64_t>(a - b, 43) +
-                           llvm::rotr<uint64_t>(c ^ seed, 30) + d,
-                       a + llvm::rotr<uint64_t>(b ^ k3, 20) - c + len + seed);
-}
-
-inline uint64_t hash_33to64_bytes(const char *s, size_t len, uint64_t seed) {
-  uint64_t z = fetch64(s + 24);
-  uint64_t a = fetch64(s) + (len + fetch64(s + len - 16)) * k0;
-  uint64_t b = llvm::rotr<uint64_t>(a + z, 52);
-  uint64_t c = llvm::rotr<uint64_t>(a, 37);
-  a += fetch64(s + 8);
-  c += llvm::rotr<uint64_t>(a, 7);
-  a += fetch64(s + 16);
-  uint64_t vf = a + z;
-  uint64_t vs = b + llvm::rotr<uint64_t>(a, 31) + c;
-  a = fetch64(s + 16) + fetch64(s + len - 32);
-  z = fetch64(s + len - 8);
-  b = llvm::rotr<uint64_t>(a + z, 52);
-  c = llvm::rotr<uint64_t>(a, 37);
-  a += fetch64(s + len - 24);
-  c += llvm::rotr<uint64_t>(a, 7);
-  a += fetch64(s + len - 16);
-  uint64_t wf = a + z;
-  uint64_t ws = b + llvm::rotr<uint64_t>(a, 31) + c;
-  uint64_t r = shift_mix((vf + ws) * k2 + (wf + vs) * k0);
-  return shift_mix((seed ^ (r * k0)) + vs) * k2;
-}
-
-inline uint64_t hash_short(const char *s, size_t length, uint64_t seed) {
-  if (length >= 4 && length <= 8)
-    return hash_4to8_bytes(s, length, seed);
-  if (length > 8 && length <= 16)
-    return hash_9to16_bytes(s, length, seed);
-  if (length > 16 && length <= 32)
-    return hash_17to32_bytes(s, length, seed);
-  if (length > 32)
-    return hash_33to64_bytes(s, length, seed);
-  if (length != 0)
-    return hash_1to3_bytes(s, length, seed);
-
-  return k2 ^ seed;
-}
-
-/// The intermediate state used during hashing.
-/// Currently, the algorithm for computing hash codes is based on CityHash and
-/// keeps 56 bytes of arbitrary state.
-struct hash_state {
-  uint64_t h0 = 0, h1 = 0, h2 = 0, h3 = 0, h4 = 0, h5 = 0, h6 = 0;
-
-  /// Create a new hash_state structure and initialize it based on the
-  /// seed and the first 64-byte chunk.
-  /// This effectively performs the initial mix.
-  static hash_state create(const char *s, uint64_t seed) {
-    hash_state state = {0,
-                        seed,
-                        hash_16_bytes(seed, k1),
-                        llvm::rotr<uint64_t>(seed ^ k1, 49),
-                        seed * k1,
-                        shift_mix(seed),
-                        0};
-    state.h6 = hash_16_bytes(state.h4, state.h5);
-    state.mix(s);
-    return state;
-  }
-
-  /// Mix 32-bytes from the input sequence into the 16-bytes of 'a'
-  /// and 'b', including whatever is already in 'a' and 'b'.
-  static void mix_32_bytes(const char *s, uint64_t &a, uint64_t &b) {
-    a += fetch64(s);
-    uint64_t c = fetch64(s + 24);
-    b = llvm::rotr<uint64_t>(b + a + c, 21);
-    uint64_t d = a;
-    a += fetch64(s + 8) + fetch64(s + 16);
-    b += llvm::rotr<uint64_t>(a, 44) + d;
-    a += c;
-  }
-
-  /// Mix in a 64-byte buffer of data.
-  /// We mix all 64 bytes even when the chunk length is smaller, but we
-  /// record the actual length.
-  void mix(const char *s) {
-    h0 = llvm::rotr<uint64_t>(h0 + h1 + h3 + fetch64(s + 8), 37) * k1;
-    h1 = llvm::rotr<uint64_t>(h1 + h4 + fetch64(s + 48), 42) * k1;
-    h0 ^= h6;
-    h1 += h3 + fetch64(s + 40);
-    h2 = llvm::rotr<uint64_t>(h2 + h5, 33) * k1;
-    h3 = h4 * k1;
-    h4 = h0 + h5;
-    mix_32_bytes(s, h3, h4);
-    h5 = h2 + h6;
-    h6 = h1 + fetch64(s + 16);
-    mix_32_bytes(s + 32, h5, h6);
-    std::swap(h2, h0);
-  }
-
-  /// Compute the final 64-bit hash code value based on the current
-  /// state and the length of bytes hashed.
-  constexpr uint64_t finalize(size_t length) {
-    return hash_16_bytes(hash_16_bytes(h3, h5) + shift_mix(h1) * k1 + h2,
-                         hash_16_bytes(h4, h6) + shift_mix(length) * k1 + h0);
-  }
-};
-
 /// In LLVM_ENABLE_ABI_BREAKING_CHECKS builds, the seed is non-deterministic
 /// per process (address of a function in LLVMSupport) to prevent having users
 /// depend on the particular hash values. On platforms without ASLR, this is
@@ -320,6 +170,13 @@ inline uint64_t get_execution_seed() {
 #endif
 }
 
+/// Hash a contiguous byte buffer to a hash_code. The execution seed is XORed
+/// into the result (not propagated through the avalanche), so a given byte
+/// stream produces the same xxh3 output modulo the per-process seed.
+inline hash_code combine_bytes(const char *data, size_t len) {
+  return xxh3_64bits(reinterpret_cast<const uint8_t *>(data), len) ^
+         get_execution_seed();
+}
 
 /// Trait to indicate whether a type's bits can be hashed directly.
 ///
@@ -362,63 +219,39 @@ template <typename T> auto get_hashable_data(const T &value) {
   }
 }
 
-/// Helper to store data from a value into a buffer and advance the
-/// pointer into that buffer.
-///
-/// This routine first checks whether there is enough space in the provided
-/// buffer, and if not immediately returns false. If there is space, it
-/// copies the underlying bytes of value into the buffer, advances the
-/// buffer_ptr past the copied bytes, and returns true.
-template <typename T>
-bool store_and_advance(char *&buffer_ptr, char *buffer_end, const T& value,
-                       size_t offset = 0) {
-  size_t store_size = sizeof(value) - offset;
-  if (buffer_ptr + store_size > buffer_end)
-    return false;
-  const char *value_data = reinterpret_cast<const char *>(&value);
-  std::memcpy(buffer_ptr, value_data + offset, store_size);
-  buffer_ptr += store_size;
-  return true;
-}
-
 /// Implement the combining of integral values into a hash_code.
 ///
 /// This overload is selected when the value type of the iterator is
 /// integral. Rather than computing a hash_code for each object and then
 /// combining them, this (as an optimization) directly combines the integers.
+///
+/// xxh3 has no streaming entry point in libLLVMSupport, so the byte stream is
+/// flattened to a buffer and hashed in one shot. A 64-byte on-stack buffer
+/// covers the common cases; longer non-contiguous ranges (the prior chunked
+/// CityHash impl was streaming and never allocated) fall back to the heap.
 template <typename InputIteratorT>
 hash_code hash_combine_range_impl(InputIteratorT first, InputIteratorT last) {
-  const uint64_t seed = get_execution_seed();
-  char buffer[64], *buffer_ptr = buffer;
-  char *const buffer_end = std::end(buffer);
-  while (first != last && store_and_advance(buffer_ptr, buffer_end,
-                                            get_hashable_data(*first)))
-    ++first;
-  if (first == last)
-    return hash_short(buffer, buffer_ptr - buffer, seed);
-  assert(buffer_ptr == buffer_end);
-
-  hash_state state = state.create(buffer, seed);
-  size_t length = 64;
-  while (first != last) {
-    // Fill up the buffer. We don't clear it, which re-mixes the last round
-    // when only a partial 64-byte chunk is left.
-    buffer_ptr = buffer;
-    while (first != last && store_and_advance(buffer_ptr, buffer_end,
-                                              get_hashable_data(*first)))
-      ++first;
-
-    // Rotate the buffer if we did a partial fill in order to simulate doing
-    // a mix of the last 64-bytes. That is how the algorithm works when we
-    // have a contiguous byte sequence, and we want to emulate that here.
-    std::rotate(buffer, buffer_ptr, buffer_end);
-
-    // Mix this chunk into the current state.
-    state.mix(buffer);
-    length += buffer_ptr - buffer;
-  };
-
-  return state.finalize(length);
+  alignas(uint64_t) char stack_buf[64];
+  std::unique_ptr<char[]> heap_buf;
+  char *buf = stack_buf;
+  size_t cap = sizeof(stack_buf);
+  size_t len = 0;
+  for (; first != last; ++first) {
+    auto data = get_hashable_data(*first);
+    if (len + sizeof(data) > cap) {
+      size_t new_cap = cap * 2;
+      while (new_cap < len + sizeof(data))
+        new_cap *= 2;
+      std::unique_ptr<char[]> new_buf(new char[new_cap]);
+      std::memcpy(new_buf.get(), buf, len);
+      heap_buf = std::move(new_buf);
+      buf = heap_buf.get();
+      cap = new_cap;
+    }
+    std::memcpy(buf + len, &data, sizeof(data));
+    len += sizeof(data);
+  }
+  return combine_bytes(buf, len);
 }
 
 /// Implement the combining of integral values into a hash_code.
@@ -432,24 +265,22 @@ hash_code hash_combine_range_impl(InputIteratorT first, InputIteratorT last) {
 template <typename ValueT>
 std::enable_if_t<is_hashable_data<ValueT>::value, hash_code>
 hash_combine_range_impl(ValueT *first, ValueT *last) {
-  const uint64_t seed = get_execution_seed();
-  const char *s_begin = reinterpret_cast<const char *>(first);
-  const char *s_end = reinterpret_cast<const char *>(last);
-  const size_t length = std::distance(s_begin, s_end);
-  if (length <= 64)
-    return hash_short(s_begin, length, seed);
-
-  const char *s_aligned_end = s_begin + (length & ~63);
-  hash_state state = state.create(s_begin, seed);
-  s_begin += 64;
-  while (s_begin != s_aligned_end) {
-    state.mix(s_begin);
-    s_begin += 64;
-  }
-  if (length & 63)
-    state.mix(s_end - 64);
+  return combine_bytes(reinterpret_cast<const char *>(first),
+                       size_t(last - first) * sizeof(ValueT));
+}
+
+/// Sum of `sizeof(get_hashable_data(arg))` across a parameter pack.
+template <typename... Ts> constexpr size_t total_hashable_size() {
+  return (size_t(0) + ... +
+          sizeof(decltype(get_hashable_data(std::declval<Ts>()))));
+}
 
-  return state.finalize(length);
+/// Copy `get_hashable_data(arg)` into `buf` at offset `off`, advancing `off`.
+template <typename T>
+inline void store_hashable_data(char *buf, size_t &off, const T &arg) {
+  auto data = get_hashable_data(arg);
+  std::memcpy(buf + off, &data, sizeof(data));
+  off += sizeof(data);
 }
 
 } // namespace detail
@@ -472,112 +303,6 @@ template <typename RangeT> hash_code hash_combine_range(RangeT &&R) {
   return hash_combine_range(adl_begin(R), adl_end(R));
 }
 
-// Implementation details for hash_combine.
-namespace hashing {
-namespace detail {
-
-/// Helper class to manage the recursive combining of hash_combine
-/// arguments.
-///
-/// This class exists to manage the state and various calls involved in the
-/// recursive combining of arguments used in hash_combine. It is particularly
-/// useful at minimizing the code in the recursive calls to ease the pain
-/// caused by a lack of variadic functions.
-struct hash_combine_recursive_helper {
-  char buffer[64] = {};
-  hash_state state;
-  const uint64_t seed;
-
-public:
-  /// Construct a recursive hash combining helper.
-  ///
-  /// This sets up the state for a recursive hash combine, including getting
-  /// the seed and buffer setup.
-  hash_combine_recursive_helper()
-    : seed(get_execution_seed()) {}
-
-  /// Combine one chunk of data into the current in-flight hash.
-  ///
-  /// This merges one chunk of data into the hash. First it tries to buffer
-  /// the data. If the buffer is full, it hashes the buffer into its
-  /// hash_state, empties it, and then merges the new chunk in. This also
-  /// handles cases where the data straddles the end of the buffer.
-  template <typename T>
-  char *combine_data(size_t &length, char *buffer_ptr, char *buffer_end, T data) {
-    if (!store_and_advance(buffer_ptr, buffer_end, data)) {
-      // Check for skew which prevents the buffer from being packed, and do
-      // a partial store into the buffer to fill it. This is only a concern
-      // with the variadic combine because that formation can have varying
-      // argument types.
-      size_t partial_store_size = buffer_end - buffer_ptr;
-      std::memcpy(buffer_ptr, &data, partial_store_size);
-
-      // If the store fails, our buffer is full and ready to hash. We have to
-      // either initialize the hash state (on the first full buffer) or mix
-      // this buffer into the existing hash state. Length tracks the *hashed*
-      // length, not the buffered length.
-      if (length == 0) {
-        state = state.create(buffer, seed);
-        length = 64;
-      } else {
-        // Mix this chunk into the current state and bump length up by 64.
-        state.mix(buffer);
-        length += 64;
-      }
-      // Reset the buffer_ptr to the head of the buffer for the next chunk of
-      // data.
-      buffer_ptr = buffer;
-
-      // Try again to store into the buffer -- this cannot fail as we only
-      // store types smaller than the buffer.
-      if (!store_and_advance(buffer_ptr, buffer_end, data,
-                             partial_store_size))
-        llvm_unreachable("buffer smaller than stored type");
-    }
-    return buffer_ptr;
-  }
-
-  /// Recursive, variadic combining method.
-  ///
-  /// This function recurses through each argument, combining that argument
-  /// into a single hash.
-  template <typename T, typename ...Ts>
-  hash_code combine(size_t length, char *buffer_ptr, char *buffer_end,
-                    const T &arg, const Ts &...args) {
-    buffer_ptr = combine_data(length, buffer_ptr, buffer_end, get_hashable_data(arg));
-
-    // Recurse to the next argument.
-    return combine(length, buffer_ptr, buffer_end, args...);
-  }
-
-  /// Base case for recursive, variadic combining.
-  ///
-  /// The base case when combining arguments recursively is reached when all
-  /// arguments have been handled. It flushes the remaining buffer and
-  /// constructs a hash_code.
-  hash_code combine(size_t length, char *buffer_ptr, char *buffer_end) {
-    // Check whether the entire set of values fit in the buffer. If so, we'll
-    // use the optimized short hashing routine and skip state entirely.
-    if (length == 0)
-      return hash_short(buffer, buffer_ptr - buffer, seed);
-
-    // Mix the final buffer, rotating it if we did a partial fill in order to
-    // simulate doing a mix of the last 64-bytes. That is how the algorithm
-    // works when we have a contiguous byte sequence, and we want to emulate
-    // that here.
-    std::rotate(buffer, buffer_ptr, buffer_end);
-
-    // Mix this chunk into the current state.
-    state.mix(buffer);
-    length += buffer_ptr - buffer;
-
-    return state.finalize(length);
-  }
-};
-
-} // namespace detail
-} // namespace hashing
-
 /// Combine values into a single hash_code.
 ///
 /// This routine accepts a varying number of arguments of any type. It will
@@ -589,10 +314,14 @@ struct hash_combine_recursive_helper {
 /// The result is suitable for returning from a user's hash_value
 /// *implementation* for their user-defined type. Consumers of a type should
 /// *not* call this routine, they should instead call 'hash_value'.
-template <typename ...Ts> hash_code hash_combine(const Ts &...args) {
-  // Recursively hash each argument using a helper class.
-  ::llvm::hashing::detail::hash_combine_recursive_helper helper;
-  return helper.combine(0, helper.buffer, helper.buffer + 64, args...);
+template <typename... Ts> hash_code hash_combine(const Ts &...args) {
+  constexpr size_t Total = hashing::detail::total_hashable_size<Ts...>();
+  // Round up so `data()` is non-null when Total == 0; combine_bytes won't
+  // read the buffer in that case (len=0 short-circuits in xxh3_64bits).
+  std::array<char, Total == 0 ? 1 : Total> buf;
+  size_t off = 0;
+  (hashing::detail::store_hashable_data(buf.data(), off, args), ...);
+  return hashing::detail::combine_bytes(buf.data(), Total);
 }
 
 // Implementation details for implementations of hash_value overloads provided
diff --git a/llvm/include/llvm/ADT/StableHashing.h b/llvm/include/llvm/ADT/StableHashing.h
index 0dd83be639424..1beb5b85c9967 100644
--- a/llvm/include/llvm/ADT/StableHashing.h
+++ b/llvm/include/llvm/ADT/StableHashing.h
@@ -30,9 +30,8 @@ namespace llvm {
 using stable_hash = uint64_t;
 
 inline stable_hash stable_hash_combine(ArrayRef<stable_hash> Buffer) {
-  const uint8_t *Ptr = reinterpret_cast<const uint8_t *>(Buffer.data());
-  size_t Size = Buffer.size() * sizeof(stable_hash);
-  return xxh3_64bits(ArrayRef<uint8_t>(Ptr, Size));
+  return xxh3_64bits(reinterpret_cast<const uint8_t *>(Buffer.data()),
+                     Buffer.size() * sizeof(stable_hash));
 }
 
 inline stable_hash stable_hash_combine(stable_hash A, stable_hash B) {
diff --git a/llvm/include/llvm/ADT/StringRef.h b/llvm/include/llvm/ADT/StringRef.h
index 1a2f2799fe960..c77608d4a2432 100644
--- a/llvm/include/llvm/ADT/StringRef.h
+++ b/llvm/include/llvm/ADT/StringRef.h
@@ -13,6 +13,7 @@
 #include "llvm/ADT/STLFunctionalExtras.h"
 #include "llvm/ADT/iterator_range.h"
 #include "llvm/Support/Compiler.h"
+#include "llvm/Support/xxhash.h"
 #include <algorithm>
 #include <cassert>
 #include <cstddef>
@@ -941,6 +942,17 @@ inline std::string &operator+=(std::string &buffer, StringRef string) {
 /// Compute a hash_code for a StringRef.
 [[nodiscard]] LLVM_ABI hash_code hash_value(StringRef S);
 
+/// Inline StringRef overloads of the xxhash entry points declared out-of-line
+/// in llvm/Support/xxhash.h. They live here so xxhash.h can stay free of ADT
+/// dependencies.
+inline uint64_t xxHash64(StringRef data) {
+  return xxHash64(reinterpret_cast<const uint8_t *>(data.data()), data.size());
+}
+inline uint64_t xxh3_64bits(StringRef data) {
+  return xxh3_64bits(reinterpret_cast<const uint8_t *>(data.data()),
+                     data.size());
+}
+
 // Provide DenseMapInfo for StringRefs.
 template <> struct DenseMapInfo<StringRef, void> {
   static inline StringRef getEmptyKey() {
diff --git a/llvm/include/llvm/Support/xxhash.h b/llvm/include/llvm/Support/xxhash.h
index b521adbef3456..15c4f1bfd4563 100644
--- a/llvm/include/llvm/Support/xxhash.h
+++ b/llvm/include/llvm/Support/xxhash.h
@@ -38,19 +38,18 @@
 #ifndef LLVM_SUPPORT_XXHASH_H
 #define LLVM_SUPPORT_XXHASH_H
 
-#include "llvm/ADT/ArrayRef.h"
-#include "llvm/ADT/StringRef.h"
 #include "llvm/Support/Compiler.h"
+#include <cstddef>
+#include <cstdint>
 
 namespace llvm {
 
-LLVM_ABI uint64_t xxHash64(llvm::StringRef Data);
-LLVM_ABI uint64_t xxHash64(llvm::ArrayRef<uint8_t> Data);
+// Deprecated pre-xxh3 64-bit hash.
+LLVM_ABI uint64_t xxHash64(const uint8_t *data, size_t len);
 
-LLVM_ABI uint64_t xxh3_64bits(ArrayRef<uint8_t> data);
-inline uint64_t xxh3_64bits(StringRef data) {
-  return xxh3_64bits(ArrayRef(data.bytes_begin(), data.size()));
-}
+/// XXH3's 64-bit variant. Inline ArrayRef and StringRef overloads live in
+/// llvm/ADT/ArrayRef.h and llvm/ADT/StringRef.h.
+LLVM_ABI uint64_t xxh3_64bits(const uint8_t *data, size_t len);
 
 /*-**********************************************************************
  *  XXH3 128-bit variant
@@ -72,8 +71,9 @@ struct XXH128_hash_t {
   }
 };
 
-/// XXH3's 128-bit variant.
-LLVM_ABI XXH128_hash_t xxh3_128bits(ArrayRef<uint8_t> data);
+/// XXH3's 128-bit variant. Inline ArrayRef overload lives in
+/// llvm/ADT/ArrayRef.h.
+LLVM_ABI XXH128_hash_t xxh3_128bits(const uint8_t *data, size_t len);
 
 } // namespace llvm
 
diff --git a/llvm/lib/Support/xxhash.cpp b/llvm/lib/Support/xxhash.cpp
index cdb76d57e2c1d..6997fed7e8336 100644
--- a/llvm/lib/Support/xxhash.cpp
+++ b/llvm/lib/Support/xxhash.cpp
@@ -100,11 +100,9 @@ static uint64_t XXH64_avalanche(uint64_t hash) {
   return hash;
 }
 
-uint64_t llvm::xxHash64(StringRef Data) {
-  size_t Len = Data.size();
+uint64_t llvm::xxHash64(const uint8_t *P, size_t Len) {
   uint64_t Seed = 0;
-  const unsigned char *P = Data.bytes_begin();
-  const unsigned char *const BEnd = Data.bytes_end();
+  const uint8_t *const BEnd = P + Len;
   uint64_t H64;
 
   if (Len >= 32) {
@@ -160,10 +158,6 @@ uint64_t llvm::xxHash64(StringRef Data) {
   return XXH64_avalanche(H64);
 }
 
-uint64_t llvm::xxHash64(ArrayRef<uint8_t> Data) {
-  return xxHash64({(const char *)Data.data(), Data.size()});
-}
-
 constexpr size_t XXH3_SECRETSIZE_MIN = 136;
 constexpr size_t XXH_SECRET_DEFAULT_SIZE = 192;
 
@@ -550,9 +544,7 @@ static uint64_t XXH3_hashLong_64b(const uint8_t *input, size_t len,
                         (uint64_t)len * PRIME64_1);
 }
 
-uint64_t llvm::xxh3_64bits(ArrayRef<uint8_t> data) {
-  auto *in = data.data();
-  size_t len = data.size();
+uint64_t llvm::xxh3_64bits(const uint8_t *in, size_t len) {
   if (len <= 16)
     return XXH3_len_0to16_64b(in, len, kSecret, 0);
   if (len <= 128)
@@ -1020,10 +1012,7 @@ XXH3_hashLong_128b(const uint8_t *input, size_t len, const uint8_t *secret,
   return h128;
 }
 
-llvm::XXH128_hash_t llvm::xxh3_128bits(ArrayRef<uint8_t> data) {
-  size_t len = data.size();
-  const uint8_t *input = data.data();
-
+llvm::XXH128_hash_t llvm::xxh3_128bits(const uint8_t *input, size_t len) {
   /*
    * If an action is to be taken if `secret` conditions are not respected,
    * it should be done here.

>From d6660fe5732405048c7a3810b898b4d8d7c5ccd5 Mon Sep 17 00:00:00 2001
From: Fangrui Song <i at maskray.me>
Date: Sat, 9 May 2026 16:27:44 -0700
Subject: [PATCH 2/2] benchmark; comment

---
 llvm/include/llvm/ADT/Hashing.h    |  9 ++++-----
 llvm/unittests/ADT/HashingTest.cpp | 19 +++++++++++++++++++
 2 files changed, 23 insertions(+), 5 deletions(-)

diff --git a/llvm/include/llvm/ADT/Hashing.h b/llvm/include/llvm/ADT/Hashing.h
index 2c49e04a3d148..c3cc37683c79a 100644
--- a/llvm/include/llvm/ADT/Hashing.h
+++ b/llvm/include/llvm/ADT/Hashing.h
@@ -33,11 +33,10 @@
 //      a single hash_code for their object. They should only logically be used
 //      within the implementation of a 'hash_value' routine or similar context.
 //
-// Note that 'hash_combine_range' contains very special logic for hashing
-// a contiguous array of integers or pointers. This logic is *extremely* fast,
-// on a modern Intel "Gainestown" Xeon (Nehalem uarch) @2.2 GHz, these were
-// benchmarked at over 6.5 GiB/s for large keys, and <20 cycles/hash for keys
-// under 32-bytes.
+// 'hash_combine_range' hashes the byte stream of the range via xxh3. The
+// contiguous-array overload hashes the range in place; the iterator overload
+// materializes the byte stream into a 256-byte on-stack buffer, falling back
+// to the heap for ranges that exceed it.
 //
 //===----------------------------------------------------------------------===//
 
diff --git a/llvm/unittests/ADT/HashingTest.cpp b/llvm/unittests/ADT/HashingTest.cpp
index e116ee934a0a4..b6daf8f4695f8 100644
--- a/llvm/unittests/ADT/HashingTest.cpp
+++ b/llvm/unittests/ADT/HashingTest.cpp
@@ -202,6 +202,25 @@ TEST(HashingTest, HashCombineRangeBasicTest) {
   EXPECT_EQ(arr5_hash, d_arr5_hash);
 }
 
+TEST(HashingTest, HashCombineRangeIteratorOverInlineBuffer) {
+  // Drive the non-pointer iterator overload past the inline stack buffer
+  // (>256 bytes of hashable data) into the heap-grow path.
+  constexpr size_t N = 100; // 100 * sizeof(size_t) = 800 bytes
+  std::vector<size_t> v(N);
+  for (size_t i = 0; i < N; ++i)
+    v[i] = i * 0x9E3779B97F4A7C15ULL;
+  std::list<size_t> l(v.begin(), v.end());
+
+  // Iterator and pointer paths see the same byte stream and agree past the
+  // inline buffer threshold.
+  EXPECT_EQ(hash_combine_range(v), hash_combine_range(l));
+  EXPECT_EQ(hash_combine_range(l), hash_combine_range(l));
+
+  std::list<size_t> l2 = l;
+  l2.push_back(0xDEADBEEFu);
+  EXPECT_NE(hash_combine_range(l), hash_combine_range(l2));
+}
+
 TEST(HashingTest, HashCombineRangeLengthDiff) {
   // Test that as only the length varies, we compute different hash codes for
   // sequences.



More information about the llvm-commits mailing list