2 * Copyright 2013 Facebook, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
17 // @author Mark Rabkin (mrabkin@fb.com)
18 // @author Andrei Alexandrescu (andrei.alexandrescu@fb.com)
20 #include "folly/Range.h"
22 #include <emmintrin.h> // __v16qi
24 #include "folly/Likely.h"
29 * Predicates that can be used with qfind and startsWith
31 const AsciiCaseSensitive asciiCaseSensitive = AsciiCaseSensitive();
32 const AsciiCaseInsensitive asciiCaseInsensitive = AsciiCaseInsensitive();
34 std::ostream& operator<<(std::ostream& os, const StringPiece& piece) {
35 os.write(piece.start(), piece.size());
41 size_t qfind_first_byte_of_memchr(const StringPiece& haystack,
42 const StringPiece& needles) {
43 size_t best = haystack.size();
44 for (char needle: needles) {
45 const void* ptr = memchr(haystack.data(), needle, best);
47 auto found = static_cast<const char*>(ptr) - haystack.data();
48 best = std::min<size_t>(best, found);
51 if (best == haystack.size()) {
52 return StringPiece::npos;
61 // It's okay if pages are bigger than this (as powers of two), but they should
63 constexpr size_t kMinPageSize = 4096;
64 static_assert(kMinPageSize >= 16,
65 "kMinPageSize must be at least SSE register size");
66 #define PAGE_FOR(addr) \
67 (reinterpret_cast<uintptr_t>(addr) / kMinPageSize)
69 inline size_t nextAlignedIndex(const char* arr) {
70 auto firstPossible = reinterpret_cast<uintptr_t>(arr) + 1;
71 return 1 + // add 1 because the index starts at 'arr'
72 ((firstPossible + 15) & ~0xF) // round up to next multiple of 16
76 #if FOLLY_HAVE_EMMINTRIN_H
77 // build sse4.2-optimized version even if -msse4.2 is not passed to GCC
78 size_t qfind_first_byte_of_needles16(const StringPiece& haystack,
79 const StringPiece& needles)
80 __attribute__ ((__target__("sse4.2"), noinline));
82 // helper method for case where needles.size() <= 16
83 size_t qfind_first_byte_of_needles16(const StringPiece& haystack,
84 const StringPiece& needles) {
85 DCHECK(!haystack.empty());
86 DCHECK(!needles.empty());
87 DCHECK_LE(needles.size(), 16);
88 // benchmarking shows that memchr beats out SSE for small needle-sets
89 // with large haystacks.
90 if ((needles.size() <= 2 && haystack.size() >= 256) ||
91 // must bail if we can't even SSE-load a single segment of haystack
92 (haystack.size() < 16 &&
93 PAGE_FOR(haystack.end() - 1) != PAGE_FOR(haystack.data() + 15)) ||
94 // can't load needles into SSE register if it could cross page boundary
95 PAGE_FOR(needles.end() - 1) != PAGE_FOR(needles.data() + 15)) {
96 return detail::qfind_first_byte_of_memchr(haystack, needles);
99 auto arr2 = __builtin_ia32_loaddqu(needles.data());
100 // do an unaligned load for first block of haystack
101 auto arr1 = __builtin_ia32_loaddqu(haystack.data());
102 auto index = __builtin_ia32_pcmpestri128(arr2, needles.size(),
103 arr1, haystack.size(), 0);
108 // Now, we can do aligned loads hereafter...
109 size_t i = nextAlignedIndex(haystack.data());
110 for (; i < haystack.size(); i+= 16) {
111 void* ptr1 = __builtin_assume_aligned(haystack.data() + i, 16);
112 auto arr1 = *reinterpret_cast<const __v16qi*>(ptr1);
113 auto index = __builtin_ia32_pcmpestri128(arr2, needles.size(),
114 arr1, haystack.size() - i, 0);
119 return StringPiece::npos;
121 #endif // FOLLY_HAVE_EMMINTRIN_H
123 // Aho, Hopcroft, and Ullman refer to this trick in "The Design and Analysis
124 // of Computer Algorithms" (1974), but the best description is here:
125 // http://research.swtch.com/sparse
128 FastByteSet() : size_(0) { } // no init of arrays required!
130 inline void add(uint8_t i) {
137 inline bool contains(uint8_t i) const {
138 DCHECK_LE(size_, 256);
139 return sparse_[i] < size_ && dense_[sparse_[i]] == i;
143 uint16_t size_; // can't use uint8_t because it would overflow if all
144 // possible values were inserted.
145 uint8_t sparse_[256];
153 size_t qfind_first_byte_of_byteset(const StringPiece& haystack,
154 const StringPiece& needles) {
156 for (auto needle: needles) {
159 for (size_t index = 0; index < haystack.size(); ++index) {
160 if (s.contains(haystack[index])) {
164 return StringPiece::npos;
167 #if FOLLY_HAVE_EMMINTRIN_H
169 template <bool HAYSTACK_ALIGNED>
170 inline size_t scanHaystackBlock(const StringPiece& haystack,
171 const StringPiece& needles,
173 // inline is okay because it's only called from other sse4.2 functions
174 __attribute__ ((__target__("sse4.2")));
176 // Scans a 16-byte block of haystack (starting at blockStartIdx) to find first
177 // needle. If HAYSTACK_ALIGNED, then haystack must be 16byte aligned.
178 // If !HAYSTACK_ALIGNED, then caller must ensure that it is safe to load the
180 template <bool HAYSTACK_ALIGNED>
181 inline size_t scanHaystackBlock(const StringPiece& haystack,
182 const StringPiece& needles,
183 int64_t blockStartIdx) {
184 DCHECK_GT(needles.size(), 16); // should handled by *needles16() method
185 DCHECK(blockStartIdx + 16 <= haystack.size() ||
186 (PAGE_FOR(haystack.data() + blockStartIdx) ==
187 PAGE_FOR(haystack.data() + blockStartIdx + 15)));
190 if (HAYSTACK_ALIGNED) {
191 void* ptr1 = __builtin_assume_aligned(haystack.data() + blockStartIdx, 16);
192 arr1 = *reinterpret_cast<const __v16qi*>(ptr1);
194 arr1 = __builtin_ia32_loaddqu(haystack.data() + blockStartIdx);
197 // This load is safe because needles.size() >= 16
198 auto arr2 = __builtin_ia32_loaddqu(needles.data());
199 size_t b = __builtin_ia32_pcmpestri128(
200 arr2, 16, arr1, haystack.size() - blockStartIdx, 0);
202 size_t j = nextAlignedIndex(needles.data());
203 for (; j < needles.size(); j += 16) {
204 void* ptr2 = __builtin_assume_aligned(needles.data() + j, 16);
205 arr2 = *reinterpret_cast<const __v16qi*>(ptr2);
207 auto index = __builtin_ia32_pcmpestri128(
208 arr2, needles.size() - j, arr1, haystack.size() - blockStartIdx, 0);
209 b = std::min<size_t>(index, b);
213 return blockStartIdx + b;
215 return StringPiece::npos;
218 size_t qfind_first_byte_of_sse42(const StringPiece& haystack,
219 const StringPiece& needles)
220 __attribute__ ((__target__("sse4.2"), noinline));
222 size_t qfind_first_byte_of_sse42(const StringPiece& haystack,
223 const StringPiece& needles) {
224 if (UNLIKELY(needles.empty() || haystack.empty())) {
225 return StringPiece::npos;
226 } else if (needles.size() <= 16) {
227 // we can save some unnecessary load instructions by optimizing for
228 // the common case of needles.size() <= 16
229 return qfind_first_byte_of_needles16(haystack, needles);
232 if (haystack.size() < 16 &&
233 PAGE_FOR(haystack.end() - 1) != PAGE_FOR(haystack.data() + 16)) {
234 // We can't safely SSE-load haystack. Use a different approach.
235 if (haystack.size() <= 2) {
236 return qfind_first_of(haystack, needles, asciiCaseSensitive);
238 return qfind_first_byte_of_byteset(haystack, needles);
241 auto ret = scanHaystackBlock<false>(haystack, needles, 0);
242 if (ret != StringPiece::npos) {
246 size_t i = nextAlignedIndex(haystack.data());
247 for (; i < haystack.size(); i += 16) {
248 auto ret = scanHaystackBlock<true>(haystack, needles, i);
249 if (ret != StringPiece::npos) {
254 return StringPiece::npos;
256 #endif // FOLLY_HAVE_EMMINTRIN_H
258 size_t qfind_first_byte_of_nosse(const StringPiece& haystack,
259 const StringPiece& needles) {
260 if (UNLIKELY(needles.empty() || haystack.empty())) {
261 return StringPiece::npos;
263 // The thresholds below were empirically determined by benchmarking.
264 // This is not an exact science since it depends on the CPU, the size of
265 // needles, and the size of haystack.
266 if (haystack.size() == 1 ||
267 (haystack.size() < 4 && needles.size() <= 16)) {
268 return qfind_first_of(haystack, needles, asciiCaseSensitive);
269 } else if ((needles.size() >= 4 && haystack.size() <= 10) ||
270 (needles.size() >= 16 && haystack.size() <= 64) ||
271 needles.size() >= 32) {
272 return qfind_first_byte_of_byteset(haystack, needles);
275 return qfind_first_byte_of_memchr(haystack, needles);
278 } // namespace detail