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
23 #include "folly/Likely.h"
28 * Predicates that can be used with qfind and startsWith
30 const AsciiCaseSensitive asciiCaseSensitive = AsciiCaseSensitive();
31 const AsciiCaseInsensitive asciiCaseInsensitive = AsciiCaseInsensitive();
33 std::ostream& operator<<(std::ostream& os, const StringPiece& piece) {
34 os.write(piece.start(), piece.size());
40 size_t qfind_first_byte_of_memchr(const StringPiece& haystack,
41 const StringPiece& needles) {
42 size_t best = haystack.size();
43 for (char needle: needles) {
44 const void* ptr = memchr(haystack.data(), needle, best);
46 auto found = static_cast<const char*>(ptr) - haystack.data();
47 best = std::min<size_t>(best, found);
50 if (best == haystack.size()) {
51 return StringPiece::npos;
60 // It's okay if pages are bigger than this (as powers of two), but they should
62 constexpr size_t kMinPageSize = 4096;
63 static_assert(kMinPageSize >= 16,
64 "kMinPageSize must be at least SSE register size");
65 #define PAGE_FOR(addr) \
66 (reinterpret_cast<uintptr_t>(addr) / kMinPageSize)
68 inline size_t nextAlignedIndex(const char* arr) {
69 auto firstPossible = reinterpret_cast<uintptr_t>(arr) + 1;
70 return 1 + // add 1 because the index starts at 'arr'
71 ((firstPossible + 15) & ~0xF) // round up to next multiple of 16
75 #if FOLLY_HAVE_EMMINTRIN_H
76 // build sse4.2-optimized version even if -msse4.2 is not passed to GCC
77 size_t qfind_first_byte_of_needles16(const StringPiece& haystack,
78 const StringPiece& needles)
79 __attribute__ ((__target__("sse4.2"), noinline));
81 // helper method for case where needles.size() <= 16
82 size_t qfind_first_byte_of_needles16(const StringPiece& haystack,
83 const StringPiece& needles) {
84 DCHECK(!haystack.empty());
85 DCHECK(!needles.empty());
86 DCHECK_LE(needles.size(), 16);
87 // benchmarking shows that memchr beats out SSE for small needle-sets
88 // with large haystacks.
89 if ((needles.size() <= 2 && haystack.size() >= 256) ||
90 // must bail if we can't even SSE-load a single segment of haystack
91 (haystack.size() < 16 &&
92 PAGE_FOR(haystack.end() - 1) != PAGE_FOR(haystack.data() + 15)) ||
93 // can't load needles into SSE register if it could cross page boundary
94 PAGE_FOR(needles.end() - 1) != PAGE_FOR(needles.data() + 15)) {
95 return detail::qfind_first_byte_of_memchr(haystack, needles);
98 auto arr2 = __builtin_ia32_loaddqu(needles.data());
99 // do an unaligned load for first block of haystack
100 auto arr1 = __builtin_ia32_loaddqu(haystack.data());
101 auto index = __builtin_ia32_pcmpestri128(arr2, needles.size(),
102 arr1, haystack.size(), 0);
107 // Now, we can do aligned loads hereafter...
108 size_t i = nextAlignedIndex(haystack.data());
109 for (; i < haystack.size(); i+= 16) {
110 void* ptr1 = __builtin_assume_aligned(haystack.data() + i, 16);
111 auto arr1 = *reinterpret_cast<const __v16qi*>(ptr1);
112 auto index = __builtin_ia32_pcmpestri128(arr2, needles.size(),
113 arr1, haystack.size() - i, 0);
118 return StringPiece::npos;
120 #endif // FOLLY_HAVE_EMMINTRIN_H
122 // Aho, Hopcroft, and Ullman refer to this trick in "The Design and Analysis
123 // of Computer Algorithms" (1974), but the best description is here:
124 // http://research.swtch.com/sparse
127 FastByteSet() : size_(0) { } // no init of arrays required!
129 inline void add(uint8_t i) {
136 inline bool contains(uint8_t i) const {
137 DCHECK_LE(size_, 256);
138 return sparse_[i] < size_ && dense_[sparse_[i]] == i;
142 uint16_t size_; // can't use uint8_t because it would overflow if all
143 // possible values were inserted.
144 uint8_t sparse_[256];
152 size_t qfind_first_byte_of_byteset(const StringPiece& haystack,
153 const StringPiece& needles) {
155 for (auto needle: needles) {
158 for (size_t index = 0; index < haystack.size(); ++index) {
159 if (s.contains(haystack[index])) {
163 return StringPiece::npos;
166 #if FOLLY_HAVE_EMMINTRIN_H
168 template <bool HAYSTACK_ALIGNED>
169 inline size_t scanHaystackBlock(const StringPiece& haystack,
170 const StringPiece& needles,
172 // inline is okay because it's only called from other sse4.2 functions
173 __attribute__ ((__target__("sse4.2")));
175 // Scans a 16-byte block of haystack (starting at blockStartIdx) to find first
176 // needle. If HAYSTACK_ALIGNED, then haystack must be 16byte aligned.
177 // If !HAYSTACK_ALIGNED, then caller must ensure that it is safe to load the
179 template <bool HAYSTACK_ALIGNED>
180 inline size_t scanHaystackBlock(const StringPiece& haystack,
181 const StringPiece& needles,
182 int64_t blockStartIdx) {
183 DCHECK_GT(needles.size(), 16); // should handled by *needles16() method
184 DCHECK(blockStartIdx + 16 <= haystack.size() ||
185 (PAGE_FOR(haystack.data() + blockStartIdx) ==
186 PAGE_FOR(haystack.data() + blockStartIdx + 15)));
189 if (HAYSTACK_ALIGNED) {
190 void* ptr1 = __builtin_assume_aligned(haystack.data() + blockStartIdx, 16);
191 arr1 = *reinterpret_cast<const __v16qi*>(ptr1);
193 arr1 = __builtin_ia32_loaddqu(haystack.data() + blockStartIdx);
196 // This load is safe because needles.size() >= 16
197 auto arr2 = __builtin_ia32_loaddqu(needles.data());
198 size_t b = __builtin_ia32_pcmpestri128(
199 arr2, 16, arr1, haystack.size() - blockStartIdx, 0);
201 size_t j = nextAlignedIndex(needles.data());
202 for (; j < needles.size(); j += 16) {
203 void* ptr2 = __builtin_assume_aligned(needles.data() + j, 16);
204 arr2 = *reinterpret_cast<const __v16qi*>(ptr2);
206 auto index = __builtin_ia32_pcmpestri128(
207 arr2, needles.size() - j, arr1, haystack.size() - blockStartIdx, 0);
208 b = std::min<size_t>(index, b);
212 return blockStartIdx + b;
214 return StringPiece::npos;
217 size_t qfind_first_byte_of_sse42(const StringPiece& haystack,
218 const StringPiece& needles)
219 __attribute__ ((__target__("sse4.2"), noinline));
221 size_t qfind_first_byte_of_sse42(const StringPiece& haystack,
222 const StringPiece& needles) {
223 if (UNLIKELY(needles.empty() || haystack.empty())) {
224 return StringPiece::npos;
225 } else if (needles.size() <= 16) {
226 // we can save some unnecessary load instructions by optimizing for
227 // the common case of needles.size() <= 16
228 return qfind_first_byte_of_needles16(haystack, needles);
231 if (haystack.size() < 16 &&
232 PAGE_FOR(haystack.end() - 1) != PAGE_FOR(haystack.data() + 16)) {
233 // We can't safely SSE-load haystack. Use a different approach.
234 if (haystack.size() <= 2) {
235 return qfind_first_of(haystack, needles, asciiCaseSensitive);
237 return qfind_first_byte_of_byteset(haystack, needles);
240 auto ret = scanHaystackBlock<false>(haystack, needles, 0);
241 if (ret != StringPiece::npos) {
245 size_t i = nextAlignedIndex(haystack.data());
246 for (; i < haystack.size(); i += 16) {
247 auto ret = scanHaystackBlock<true>(haystack, needles, i);
248 if (ret != StringPiece::npos) {
253 return StringPiece::npos;
255 #endif // FOLLY_HAVE_EMMINTRIN_H
257 size_t qfind_first_byte_of_nosse(const StringPiece& haystack,
258 const StringPiece& needles) {
259 if (UNLIKELY(needles.empty() || haystack.empty())) {
260 return StringPiece::npos;
262 // The thresholds below were empirically determined by benchmarking.
263 // This is not an exact science since it depends on the CPU, the size of
264 // needles, and the size of haystack.
265 if (haystack.size() == 1 ||
266 (haystack.size() < 4 && needles.size() <= 16)) {
267 return qfind_first_of(haystack, needles, asciiCaseSensitive);
268 } else if ((needles.size() >= 4 && haystack.size() <= 10) ||
269 (needles.size() >= 16 && haystack.size() <= 64) ||
270 needles.size() >= 32) {
271 return qfind_first_byte_of_byteset(haystack, needles);
274 return qfind_first_byte_of_memchr(haystack, needles);
277 } // namespace detail