/src/abseil-cpp/absl/base/prefetch.h
Line | Count | Source (jump to first uncovered line) |
1 | | // Copyright 2023 The Abseil Authors |
2 | | // |
3 | | // Licensed under the Apache License, Version 2.0 (the "License"); |
4 | | // you may not use this file except in compliance with the License. |
5 | | // You may obtain a copy of the License at |
6 | | // |
7 | | // https://www.apache.org/licenses/LICENSE-2.0 |
8 | | // |
9 | | // Unless required by applicable law or agreed to in writing, software |
10 | | // distributed under the License is distributed on an "AS IS" BASIS, |
11 | | // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
12 | | // See the License for the specific language governing permissions and |
13 | | // limitations under the License. |
14 | | // |
15 | | // ----------------------------------------------------------------------------- |
16 | | // File: prefetch.h |
17 | | // ----------------------------------------------------------------------------- |
18 | | // |
19 | | // This header file defines prefetch functions to prefetch memory contents |
20 | | // into the first level cache (L1) for the current CPU. The prefetch logic |
21 | | // offered in this header is limited to prefetching first level cachelines |
22 | | // only, and is aimed at relatively 'simple' prefetching logic. |
23 | | // |
24 | | #ifndef ABSL_BASE_PREFETCH_H_ |
25 | | #define ABSL_BASE_PREFETCH_H_ |
26 | | |
27 | | #include "absl/base/attributes.h" |
28 | | #include "absl/base/config.h" |
29 | | |
30 | | #if defined(ABSL_INTERNAL_HAVE_SSE) |
31 | | #include <xmmintrin.h> |
32 | | #endif |
33 | | |
34 | | #if defined(_MSC_VER) && _MSC_VER >= 1900 && \ |
35 | | (defined(_M_X64) || defined(_M_IX86)) |
36 | | #include <intrin.h> |
37 | | #pragma intrinsic(_mm_prefetch) |
38 | | #endif |
39 | | |
40 | | namespace absl { |
41 | | ABSL_NAMESPACE_BEGIN |
42 | | |
43 | | // Moves data into the L1 cache before it is read, or "prefetches" it. |
44 | | // |
45 | | // The value of `addr` is the address of the memory to prefetch. If |
46 | | // the target and compiler support it, data prefetch instructions are |
47 | | // generated. If the prefetch is done some time before the memory is |
48 | | // read, it may be in the cache by the time the read occurs. |
49 | | // |
50 | | // This method prefetches data with the highest degree of temporal locality; |
51 | | // data is prefetched where possible into all levels of the cache. |
52 | | // |
53 | | // Incorrect or gratuitous use of this function can degrade performance. |
54 | | // Use this function only when representative benchmarks show an improvement. |
55 | | // |
56 | | // Example: |
57 | | // |
58 | | // // Computes incremental checksum for `data`. |
59 | | // int ComputeChecksum(int sum, absl::string_view data); |
60 | | // |
61 | | // // Computes cumulative checksum for all values in `data` |
62 | | // int ComputeChecksum(absl::Span<const std::string> data) { |
63 | | // int sum = 0; |
64 | | // auto it = data.begin(); |
65 | | // auto pit = data.begin(); |
66 | | // auto end = data.end(); |
67 | | // for (int dist = 8; dist > 0 && pit != data.end(); --dist, ++pit) { |
68 | | // absl::PrefetchToLocalCache(pit->data()); |
69 | | // } |
70 | | // for (; pit != end; ++pit, ++it) { |
71 | | // sum = ComputeChecksum(sum, *it); |
72 | | // absl::PrefetchToLocalCache(pit->data()); |
73 | | // } |
74 | | // for (; it != end; ++it) { |
75 | | // sum = ComputeChecksum(sum, *it); |
76 | | // } |
77 | | // return sum; |
78 | | // } |
79 | | // |
80 | | void PrefetchToLocalCache(const void* addr); |
81 | | |
82 | | // Moves data into the L1 cache before it is read, or "prefetches" it. |
83 | | // |
84 | | // This function is identical to `PrefetchToLocalCache()` except that it has |
85 | | // non-temporal locality: the fetched data should not be left in any of the |
86 | | // cache tiers. This is useful for cases where the data is used only once / |
87 | | // short term, for example, invoking a destructor on an object. |
88 | | // |
89 | | // Incorrect or gratuitous use of this function can degrade performance. |
90 | | // Use this function only when representative benchmarks show an improvement. |
91 | | // |
92 | | // Example: |
93 | | // |
94 | | // template <typename Iterator> |
95 | | // void DestroyPointers(Iterator begin, Iterator end) { |
96 | | // size_t distance = std::min(8U, bars.size()); |
97 | | // |
98 | | // int dist = 8; |
99 | | // auto prefetch_it = begin; |
100 | | // while (prefetch_it != end && --dist;) { |
101 | | // absl::PrefetchToLocalCacheNta(*prefetch_it++); |
102 | | // } |
103 | | // while (prefetch_it != end) { |
104 | | // delete *begin++; |
105 | | // absl::PrefetchToLocalCacheNta(*prefetch_it++); |
106 | | // } |
107 | | // while (begin != end) { |
108 | | // delete *begin++; |
109 | | // } |
110 | | // } |
111 | | // |
112 | | void PrefetchToLocalCacheNta(const void* addr); |
113 | | |
114 | | // Moves data into the L1 cache with the intent to modify it. |
115 | | // |
116 | | // This function is similar to `PrefetchToLocalCache()` except that it |
117 | | // prefetches cachelines with an 'intent to modify' This typically includes |
118 | | // invalidating cache entries for this address in all other cache tiers, and an |
119 | | // exclusive access intent. |
120 | | // |
121 | | // Incorrect or gratuitous use of this function can degrade performance. As this |
122 | | // function can invalidate cached cachelines on other caches and computer cores, |
123 | | // incorrect usage of this function can have an even greater negative impact |
124 | | // than incorrect regular prefetches. |
125 | | // Use this function only when representative benchmarks show an improvement. |
126 | | // |
127 | | // Example: |
128 | | // |
129 | | // void* Arena::Allocate(size_t size) { |
130 | | // void* ptr = AllocateBlock(size); |
131 | | // absl::PrefetchToLocalCacheForWrite(p); |
132 | | // return ptr; |
133 | | // } |
134 | | // |
135 | | void PrefetchToLocalCacheForWrite(const void* addr); |
136 | | |
137 | | #if ABSL_HAVE_BUILTIN(__builtin_prefetch) || defined(__GNUC__) |
138 | | |
139 | | #define ABSL_HAVE_PREFETCH 1 |
140 | | |
141 | | // See __builtin_prefetch: |
142 | | // https://gcc.gnu.org/onlinedocs/gcc/Other-Builtins.html. |
143 | | // |
144 | | ABSL_ATTRIBUTE_ALWAYS_INLINE inline void PrefetchToLocalCache( |
145 | 56 | const void* addr) { |
146 | 56 | __builtin_prefetch(addr, 0, 3); |
147 | 56 | } |
148 | | |
149 | | ABSL_ATTRIBUTE_ALWAYS_INLINE inline void PrefetchToLocalCacheNta( |
150 | 0 | const void* addr) { |
151 | 0 | __builtin_prefetch(addr, 0, 0); |
152 | 0 | } |
153 | | |
154 | | ABSL_ATTRIBUTE_ALWAYS_INLINE inline void PrefetchToLocalCacheForWrite( |
155 | 0 | const void* addr) { |
156 | 0 | // [x86] gcc/clang don't generate PREFETCHW for __builtin_prefetch(.., 1) |
157 | 0 | // unless -march=broadwell or newer; this is not generally the default, so we |
158 | 0 | // manually emit prefetchw. PREFETCHW is recognized as a no-op on older Intel |
159 | 0 | // processors and has been present on AMD processors since the K6-2. |
160 | 0 | #if defined(__x86_64__) |
161 | 0 | asm("prefetchw (%0)" : : "r"(addr)); |
162 | 0 | #else |
163 | 0 | __builtin_prefetch(addr, 1, 3); |
164 | 0 | #endif |
165 | 0 | } |
166 | | |
167 | | #elif defined(ABSL_INTERNAL_HAVE_SSE) |
168 | | |
169 | | #define ABSL_HAVE_PREFETCH 1 |
170 | | |
171 | | ABSL_ATTRIBUTE_ALWAYS_INLINE inline void PrefetchToLocalCache( |
172 | | const void* addr) { |
173 | | _mm_prefetch(reinterpret_cast<const char*>(addr), _MM_HINT_T0); |
174 | | } |
175 | | |
176 | | ABSL_ATTRIBUTE_ALWAYS_INLINE inline void PrefetchToLocalCacheNta( |
177 | | const void* addr) { |
178 | | _mm_prefetch(reinterpret_cast<const char*>(addr), _MM_HINT_NTA); |
179 | | } |
180 | | |
181 | | ABSL_ATTRIBUTE_ALWAYS_INLINE inline void PrefetchToLocalCacheForWrite( |
182 | | const void* addr) { |
183 | | #if defined(_MM_HINT_ET0) |
184 | | _mm_prefetch(reinterpret_cast<const char*>(addr), _MM_HINT_ET0); |
185 | | #elif !defined(_MSC_VER) && defined(__x86_64__) |
186 | | // _MM_HINT_ET0 is not universally supported. As we commented further |
187 | | // up, PREFETCHW is recognized as a no-op on older Intel processors |
188 | | // and has been present on AMD processors since the K6-2. We have this |
189 | | // disabled for MSVC compilers as this miscompiles on older MSVC compilers. |
190 | | asm("prefetchw (%0)" : : "r"(addr)); |
191 | | #endif |
192 | | } |
193 | | |
194 | | #else |
195 | | |
196 | | ABSL_ATTRIBUTE_ALWAYS_INLINE inline void PrefetchToLocalCache( |
197 | | const void* addr) {} |
198 | | ABSL_ATTRIBUTE_ALWAYS_INLINE inline void PrefetchToLocalCacheNta( |
199 | | const void* addr) {} |
200 | | ABSL_ATTRIBUTE_ALWAYS_INLINE inline void PrefetchToLocalCacheForWrite( |
201 | | const void* addr) {} |
202 | | |
203 | | #endif |
204 | | |
205 | | ABSL_NAMESPACE_END |
206 | | } // namespace absl |
207 | | |
208 | | #endif // ABSL_BASE_PREFETCH_H_ |