Skip to content

Commit 8e67632

Browse files
Optimize HashBytes for 64-bit and remove dead code (#129640)
Optimize `HashBytes` (`src/coreclr/inc/utilcode.h`) to use xxHash32 primitives (`XXHash32_QueueRound`/`XXHash32_MixFinal`) processing 4 bytes at a time, replacing the original byte-at-a-time DJB2 hash. `HashBytes` is used by the string literal interning path (`EEUnicodeStringLiteralHashTableHelper::Hash`) which hashes every string literal resolved during JIT compilation (`ldstr` token resolution). This makes it part of the steady-state cost for all managed applications. The xxHash32 primitives are extracted from `typehashingalgorithms.h` into a new shared header `src/coreclr/inc/dn_xxhash.h` so they can be reused by both the type hashing and `HashBytes` paths. Also removes unused dead code: `EEUnicodeHashTableHelper`, `EEUnicodeStringHashTable`, `EEStringData::IsOnlyLowChars`, `HashStringN`, `HashiStringA`, `HashiStringN`, and `CaseInsensitiveStringCompareHash`. ### Measurement A microbenchmark calling `string.IsInterned()` 200M times against a pool of 1000 non-interned strings was used to isolate the hashing cost. Profiled on macOS ARM64 (Apple M5) with a full Release build (`clr+libs -c Release`). Each row is the median of 5 runs. | Version | Median | vs Baseline | |---------|--------|-------------| | Upstream baseline (byte-at-a-time DJB2) | 8.53s | — | | xxHash32 primitives (QueueRound loop) | 7.65s | **-10.3%** | | xxHash32 class (uint32_t) | 8.30s | -2.7% | | xxHash32 class (uint64_t) | 9.06s | +6.2% | The direct primitives approach was chosen because the `xxHash` class queue machinery (position tracking, branching, accumulator state) adds overhead that erases the gain when used for bulk byte-stream hashing. The primitives inline cleanly into a tight loop. --------- Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
1 parent f4430a4 commit 8e67632

14 files changed

Lines changed: 306 additions & 463 deletions

‎src/coreclr/inc/dn_xxhash.h‎

Lines changed: 253 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,253 @@
1+
// Licensed to the .NET Foundation under one or more agreements.
2+
// The .NET Foundation licenses this file to you under the MIT license.
3+
4+
// ---------------------------------------------------------------------------
5+
// xxHash32 primitives and helpers
6+
//
7+
// Based on the xxHash32 logic implemented in System.HashCode, which is in
8+
// turn based on the code published by Yann Collet:
9+
// https://raw.githubusercontent.com/Cyan4973/xxHash/5c174cfa4e45a42f94082dc0d4539b39696afea1/xxhash.c
10+
//
11+
// xxHash - Fast Hash algorithm
12+
// Copyright (C) 2012-2016, Yann Collet
13+
//
14+
// BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
15+
//
16+
// Redistribution and use in source and binary forms, with or without
17+
// modification, are permitted provided that the following conditions are
18+
// met:
19+
//
20+
// * Redistributions of source code must retain the above copyright
21+
// notice, this list of conditions and the following disclaimer.
22+
// * Redistributions in binary form must reproduce the above
23+
// copyright notice, this list of conditions and the following disclaimer
24+
// in the documentation and/or other materials provided with the
25+
// distribution.
26+
//
27+
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
28+
// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
29+
// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
30+
// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
31+
// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
32+
// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
33+
// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
34+
// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
35+
// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
36+
// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
37+
// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
38+
//
39+
// You can contact the author at :
40+
// - xxHash homepage: http://www.xxhash.com
41+
// - xxHash source repository : https://github.com/Cyan4973/xxHash
42+
// ---------------------------------------------------------------------------
43+
44+
#pragma once
45+
#include <stdlib.h>
46+
#include <type_traits>
47+
#include "clrtypes.h"
48+
49+
struct xxHashDefaultTraits
50+
{
51+
static uint32_t GenerateGlobalSeed();
52+
};
53+
54+
// This is a port of the System.HashCode logic for computing a hashcode using the xxHash algorithm.
55+
// The traits type T must provide a static GenerateGlobalSeed() method that returns the seed value.
56+
template <typename T>
57+
class xxHash
58+
{
59+
static_assert(std::is_same<decltype(T::GenerateGlobalSeed()), uint32_t>::value,
60+
"T must provide a static uint32_t GenerateGlobalSeed() method");
61+
62+
static constexpr uint32_t Prime1 = 2654435761U;
63+
static constexpr uint32_t Prime2 = 2246822519U;
64+
static constexpr uint32_t Prime3 = 3266489917U;
65+
static constexpr uint32_t Prime4 = 668265263U;
66+
static constexpr uint32_t Prime5 = 374761393U;
67+
68+
public: // static
69+
static uint32_t MixEmptyState()
70+
{
71+
return T::GenerateGlobalSeed() + Prime5;
72+
}
73+
74+
static uint32_t MixState(uint32_t v1, uint32_t v2, uint32_t v3, uint32_t v4)
75+
{
76+
return (uint32_t)_rotl(v1, 1) + (uint32_t)_rotl(v2, 7) + (uint32_t)_rotl(v3, 12) + (uint32_t)_rotl(v4, 18);
77+
}
78+
79+
static uint32_t QueueRound(uint32_t hash, uint32_t queuedValue)
80+
{
81+
return ((uint32_t)_rotl((int)(hash + queuedValue * Prime3), 17)) * Prime4;
82+
}
83+
84+
static uint32_t Round(uint32_t hash, uint32_t input)
85+
{
86+
return ((uint32_t)_rotl((int)(hash + input * Prime2), 13)) * Prime1;
87+
}
88+
89+
static uint32_t MixFinal(uint32_t hash)
90+
{
91+
hash ^= hash >> 15;
92+
hash *= Prime2;
93+
hash ^= hash >> 13;
94+
hash *= Prime3;
95+
hash ^= hash >> 16;
96+
return hash;
97+
}
98+
99+
private:
100+
static constexpr uint32_t MixEmptyState(uint32_t seed)
101+
{
102+
return seed + Prime5;
103+
}
104+
105+
const uint32_t seed;
106+
uint32_t _v1;
107+
uint32_t _v2;
108+
uint32_t _v3;
109+
uint32_t _v4;
110+
uint32_t _queue1 = 0;
111+
uint32_t _queue2 = 0;
112+
uint32_t _queue3 = 0;
113+
uint32_t _length = 0;
114+
115+
public:
116+
xxHash()
117+
: seed(T::GenerateGlobalSeed())
118+
, _v1{seed + Prime1 + Prime2}
119+
, _v2{seed + Prime2}
120+
, _v3{seed}
121+
, _v4{seed - Prime1}
122+
{
123+
}
124+
125+
void Add(uint32_t val)
126+
{
127+
// The original xxHash works as follows:
128+
// 0. Initialize immediately. We can't do this in a struct (no
129+
// default ctor).
130+
// 1. Accumulate blocks of length 16 (4 uints) into 4 accumulators.
131+
// 2. Accumulate remaining blocks of length 4 (1 uint) into the
132+
// hash.
133+
// 3. Accumulate remaining blocks of length 1 into the hash.
134+
135+
// There is no need for #3 as this type only accepts ints. _queue1,
136+
// _queue2 and _queue3 are basically a buffer so that when
137+
// ToHashCode is called we can execute #2 correctly.
138+
139+
// Storing the value of _length locally shaves of quite a few bytes
140+
// in the resulting machine code.
141+
uint32_t previousLength = _length++;
142+
uint32_t position = previousLength % 4;
143+
144+
// Switch can't be inlined.
145+
146+
if (position == 0)
147+
_queue1 = val;
148+
else if (position == 1)
149+
_queue2 = val;
150+
else if (position == 2)
151+
_queue3 = val;
152+
else // position == 3
153+
{
154+
_v1 = Round(_v1, _queue1);
155+
_v2 = Round(_v2, _queue2);
156+
_v3 = Round(_v3, _queue3);
157+
_v4 = Round(_v4, val);
158+
}
159+
}
160+
161+
void AddPointer(void* ptr)
162+
{
163+
#ifdef HOST_64BIT
164+
Add((uint32_t)(UINT_PTR)ptr);
165+
Add((uint32_t)(((UINT_PTR)ptr) >> 32));
166+
#else
167+
Add((uint32_t)(UINT_PTR)ptr);
168+
#endif
169+
}
170+
171+
uint32_t ToHashCode()
172+
{
173+
// Storing the value of _length locally shaves of quite a few bytes
174+
// in the resulting machine code.
175+
uint32_t length = _length;
176+
177+
// position refers to the *next* queue position in this method, so
178+
// position == 1 means that _queue1 is populated; _queue2 would have
179+
// been populated on the next call to Add.
180+
uint32_t position = length % 4;
181+
182+
// If the length is less than 4, _v1 to _v4 don't contain anything
183+
// yet. xxHash32 treats this differently.
184+
185+
uint32_t hash = length < 4 ? MixEmptyState(seed) : MixState(_v1, _v2, _v3, _v4);
186+
187+
// _length is incremented once per Add(Int32) and is therefore 4
188+
// times too small (xxHash length is in bytes, not ints).
189+
190+
hash += length * 4;
191+
192+
// Mix what remains in the queue
193+
194+
// Switch can't be inlined right now, so use as few branches as
195+
// possible by manually excluding impossible scenarios (position > 1
196+
// is always false if position is not > 0).
197+
if (position > 0)
198+
{
199+
hash = QueueRound(hash, _queue1);
200+
if (position > 1)
201+
{
202+
hash = QueueRound(hash, _queue2);
203+
if (position > 2)
204+
hash = QueueRound(hash, _queue3);
205+
}
206+
}
207+
208+
hash = MixFinal(hash);
209+
return hash;
210+
}
211+
};
212+
213+
template <typename T = xxHashDefaultTraits>
214+
inline static uint32_t MixOneValueIntoHash(uint32_t value1)
215+
{
216+
// This matches the behavior of System.HashCode.Combine(value1) as of the time of authoring
217+
218+
// Provide a way of diffusing bits from something with a limited
219+
// input hash space. For example, many enums only have a few
220+
// possible hashes, only using the bottom few bits of the code. Some
221+
// collections are built on the assumption that hashes are spread
222+
// over a larger space, so diffusing the bits may help the
223+
// collection work more efficiently.
224+
225+
uint32_t hash = xxHash<T>::MixEmptyState();
226+
hash += 4;
227+
hash = xxHash<T>::QueueRound(hash, value1);
228+
hash = xxHash<T>::MixFinal(hash);
229+
return hash;
230+
}
231+
232+
template <typename T = xxHashDefaultTraits>
233+
inline static uint32_t CombineTwoValuesIntoHash(uint32_t value1, uint32_t value2)
234+
{
235+
// This matches the behavior of System.HashCode.Combine(value1, value2) as of the time of authoring
236+
uint32_t hash = xxHash<T>::MixEmptyState();
237+
hash += 8;
238+
hash = xxHash<T>::QueueRound(hash, value1);
239+
hash = xxHash<T>::QueueRound(hash, value2);
240+
hash = xxHash<T>::MixFinal(hash);
241+
return hash;
242+
}
243+
244+
template <typename T = xxHashDefaultTraits>
245+
inline static uint32_t MixPointerIntoHash(void* ptr)
246+
{
247+
// This matches the behavior of System.HashCode.Combine(ptr) as of the time of authoring
248+
#ifdef HOST_64BIT
249+
return CombineTwoValuesIntoHash<T>((uint32_t)(UINT_PTR)ptr, (uint32_t)(((UINT64)(UINT_PTR)ptr) >> 32));
250+
#else
251+
return MixOneValueIntoHash<T>((uint32_t)ptr);
252+
#endif
253+
}

‎src/coreclr/inc/shash.h‎

Lines changed: 0 additions & 45 deletions
Original file line numberDiff line numberDiff line change
@@ -729,51 +729,6 @@ struct CaseSensitiveStringCompareHash
729729
}
730730
};
731731

732-
// Provides case-insensitive comparison and hashing functionality through static
733-
// and functor object methods. Can be instantiated with CHAR or WCHAR.
734-
template <typename CharT>
735-
struct CaseInsensitiveStringCompareHash
736-
{
737-
private:
738-
typedef CharT const * str_t;
739-
740-
static size_t _strcmp(str_t left, str_t right)
741-
{
742-
return ::SString::_tstricmp(left, right);
743-
}
744-
745-
static size_t _hash(CHAR const *str)
746-
{
747-
return HashiStringA(str);
748-
}
749-
750-
static size_t _hash(WCHAR const *str)
751-
{
752-
return HashiString(str);
753-
}
754-
755-
public:
756-
static size_t compare(str_t left, str_t right)
757-
{
758-
return _strcmp(left, right);
759-
}
760-
761-
size_t operator()(str_t left, str_t right)
762-
{
763-
return compare(left, right);
764-
}
765-
766-
static size_t hash(str_t str)
767-
{
768-
return _hash(str);
769-
}
770-
771-
size_t operator()(str_t str)
772-
{
773-
return hash(str);
774-
}
775-
};
776-
777732
// StringSHashTraits is a traits class useful for string-keyed
778733
// pointer hash tables.
779734

0 commit comments

Comments
 (0)