2 * Copyright 2014 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 #if FOLLY_HAVE_EMMINTRIN_H
23 #include <emmintrin.h> // __v16qi
30 * Predicates that can be used with qfind and startsWith
32 const AsciiCaseSensitive asciiCaseSensitive = AsciiCaseSensitive();
33 const AsciiCaseInsensitive asciiCaseInsensitive = AsciiCaseInsensitive();
35 std::ostream& operator<<(std::ostream& os, const StringPiece& piece) {
36 os.write(piece.start(), piece.size());
40 std::ostream& operator<<(std::ostream& os, const MutableStringPiece& piece) {
41 os.write(piece.start(), piece.size());
47 size_t qfind_first_byte_of_memchr(const StringPiece& haystack,
48 const StringPiece& needles) {
49 size_t best = haystack.size();
50 for (char needle: needles) {
51 const void* ptr = memchr(haystack.data(), needle, best);
53 auto found = static_cast<const char*>(ptr) - haystack.data();
54 best = std::min<size_t>(best, found);
57 if (best == haystack.size()) {
58 return StringPiece::npos;
67 // It's okay if pages are bigger than this (as powers of two), but they should
69 constexpr size_t kMinPageSize = 4096;
70 static_assert(kMinPageSize >= 16,
71 "kMinPageSize must be at least SSE register size");
72 #define PAGE_FOR(addr) \
73 (reinterpret_cast<uintptr_t>(addr) / kMinPageSize)
76 // Earlier versions of GCC (for example, Clang on Mac OS X, which is based on
77 // GCC 4.2) do not have a full compliment of SSE builtins.
78 #if FOLLY_HAVE_EMMINTRIN_H && __GNUC_PREREQ(4, 6)
79 inline size_t nextAlignedIndex(const char* arr) {
80 auto firstPossible = reinterpret_cast<uintptr_t>(arr) + 1;
81 return 1 + // add 1 because the index starts at 'arr'
82 ((firstPossible + 15) & ~0xF) // round up to next multiple of 16
86 // build sse4.2-optimized version even if -msse4.2 is not passed to GCC
87 size_t qfind_first_byte_of_needles16(const StringPiece& haystack,
88 const StringPiece& needles)
89 __attribute__ ((__target__("sse4.2"), noinline))
90 FOLLY_DISABLE_ADDRESS_SANITIZER;
92 // helper method for case where needles.size() <= 16
93 size_t qfind_first_byte_of_needles16(const StringPiece& haystack,
94 const StringPiece& needles) {
95 DCHECK(!haystack.empty());
96 DCHECK(!needles.empty());
97 DCHECK_LE(needles.size(), 16);
98 // benchmarking shows that memchr beats out SSE for small needle-sets
99 // with large haystacks.
100 if ((needles.size() <= 2 && haystack.size() >= 256) ||
101 // must bail if we can't even SSE-load a single segment of haystack
102 (haystack.size() < 16 &&
103 PAGE_FOR(haystack.end() - 1) != PAGE_FOR(haystack.data() + 15)) ||
104 // can't load needles into SSE register if it could cross page boundary
105 PAGE_FOR(needles.end() - 1) != PAGE_FOR(needles.data() + 15)) {
106 return detail::qfind_first_byte_of_memchr(haystack, needles);
109 auto arr2 = __builtin_ia32_loaddqu(needles.data());
110 // do an unaligned load for first block of haystack
111 auto arr1 = __builtin_ia32_loaddqu(haystack.data());
112 auto index = __builtin_ia32_pcmpestri128(arr2, needles.size(),
113 arr1, haystack.size(), 0);
118 // Now, we can do aligned loads hereafter...
119 size_t i = nextAlignedIndex(haystack.data());
120 for (; i < haystack.size(); i+= 16) {
121 void* ptr1 = __builtin_assume_aligned(haystack.data() + i, 16);
122 auto arr1 = *reinterpret_cast<const __v16qi*>(ptr1);
123 auto index = __builtin_ia32_pcmpestri128(arr2, needles.size(),
124 arr1, haystack.size() - i, 0);
129 return StringPiece::npos;
131 #endif // FOLLY_HAVE_EMMINTRIN_H && GCC 4.6+
133 // Aho, Hopcroft, and Ullman refer to this trick in "The Design and Analysis
134 // of Computer Algorithms" (1974), but the best description is here:
135 // http://research.swtch.com/sparse
138 FastByteSet() : size_(0) { } // no init of arrays required!
140 inline void add(uint8_t i) {
147 inline bool contains(uint8_t i) const {
148 DCHECK_LE(size_, 256);
149 return sparse_[i] < size_ && dense_[sparse_[i]] == i;
153 uint16_t size_; // can't use uint8_t because it would overflow if all
154 // possible values were inserted.
155 uint8_t sparse_[256];
163 size_t qfind_first_byte_of_byteset(const StringPiece& haystack,
164 const StringPiece& needles) {
166 for (auto needle: needles) {
169 for (size_t index = 0; index < haystack.size(); ++index) {
170 if (s.contains(haystack[index])) {
174 return StringPiece::npos;
177 #if FOLLY_HAVE_EMMINTRIN_H && __GNUC_PREREQ(4, 6)
179 template <bool HAYSTACK_ALIGNED>
180 size_t scanHaystackBlock(const StringPiece& haystack,
181 const StringPiece& needles,
183 // inline is okay because it's only called from other sse4.2 functions
184 __attribute__ ((__target__("sse4.2")))
185 // Turn off ASAN because the "arr2 = ..." assignment in the loop below reads
186 // up to 15 bytes beyond end of the buffer in #needles#. That is ok because
187 // ptr2 is always 16-byte aligned, so the read can never span a page boundary.
188 // Also, the extra data that may be read is never actually used.
189 FOLLY_DISABLE_ADDRESS_SANITIZER;
191 // Scans a 16-byte block of haystack (starting at blockStartIdx) to find first
192 // needle. If HAYSTACK_ALIGNED, then haystack must be 16byte aligned.
193 // If !HAYSTACK_ALIGNED, then caller must ensure that it is safe to load the
195 template <bool HAYSTACK_ALIGNED>
196 size_t scanHaystackBlock(const StringPiece& haystack,
197 const StringPiece& needles,
198 int64_t blockStartIdx) {
199 DCHECK_GT(needles.size(), 16); // should handled by *needles16() method
200 DCHECK(blockStartIdx + 16 <= haystack.size() ||
201 (PAGE_FOR(haystack.data() + blockStartIdx) ==
202 PAGE_FOR(haystack.data() + blockStartIdx + 15)));
205 if (HAYSTACK_ALIGNED) {
206 void* ptr1 = __builtin_assume_aligned(haystack.data() + blockStartIdx, 16);
207 arr1 = *reinterpret_cast<const __v16qi*>(ptr1);
209 arr1 = __builtin_ia32_loaddqu(haystack.data() + blockStartIdx);
212 // This load is safe because needles.size() >= 16
213 auto arr2 = __builtin_ia32_loaddqu(needles.data());
214 size_t b = __builtin_ia32_pcmpestri128(
215 arr2, 16, arr1, haystack.size() - blockStartIdx, 0);
217 size_t j = nextAlignedIndex(needles.data());
218 for (; j < needles.size(); j += 16) {
219 void* ptr2 = __builtin_assume_aligned(needles.data() + j, 16);
220 arr2 = *reinterpret_cast<const __v16qi*>(ptr2);
222 auto index = __builtin_ia32_pcmpestri128(
223 arr2, needles.size() - j, arr1, haystack.size() - blockStartIdx, 0);
224 b = std::min<size_t>(index, b);
228 return blockStartIdx + b;
230 return StringPiece::npos;
233 size_t qfind_first_byte_of_sse42(const StringPiece& haystack,
234 const StringPiece& needles)
235 __attribute__ ((__target__("sse4.2"), noinline));
237 size_t qfind_first_byte_of_sse42(const StringPiece& haystack,
238 const StringPiece& needles) {
239 if (UNLIKELY(needles.empty() || haystack.empty())) {
240 return StringPiece::npos;
241 } else if (needles.size() <= 16) {
242 // we can save some unnecessary load instructions by optimizing for
243 // the common case of needles.size() <= 16
244 return qfind_first_byte_of_needles16(haystack, needles);
247 if (haystack.size() < 16 &&
248 PAGE_FOR(haystack.end() - 1) != PAGE_FOR(haystack.data() + 16)) {
249 // We can't safely SSE-load haystack. Use a different approach.
250 if (haystack.size() <= 2) {
251 return qfind_first_of(haystack, needles, asciiCaseSensitive);
253 return qfind_first_byte_of_byteset(haystack, needles);
256 auto ret = scanHaystackBlock<false>(haystack, needles, 0);
257 if (ret != StringPiece::npos) {
261 size_t i = nextAlignedIndex(haystack.data());
262 for (; i < haystack.size(); i += 16) {
263 auto ret = scanHaystackBlock<true>(haystack, needles, i);
264 if (ret != StringPiece::npos) {
269 return StringPiece::npos;
271 #endif // FOLLY_HAVE_EMMINTRIN_H && GCC 4.6+
273 size_t qfind_first_byte_of_nosse(const StringPiece& haystack,
274 const StringPiece& needles) {
275 if (UNLIKELY(needles.empty() || haystack.empty())) {
276 return StringPiece::npos;
278 // The thresholds below were empirically determined by benchmarking.
279 // This is not an exact science since it depends on the CPU, the size of
280 // needles, and the size of haystack.
281 if (haystack.size() == 1 ||
282 (haystack.size() < 4 && needles.size() <= 16)) {
283 return qfind_first_of(haystack, needles, asciiCaseSensitive);
284 } else if ((needles.size() >= 4 && haystack.size() <= 10) ||
285 (needles.size() >= 16 && haystack.size() <= 64) ||
286 needles.size() >= 32) {
287 return qfind_first_byte_of_byteset(haystack, needles);
290 return qfind_first_byte_of_memchr(haystack, needles);
293 } // namespace detail