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());
42 size_t qfind_first_byte_of_memchr(const StringPiece& haystack,
43 const StringPiece& needles) {
44 size_t best = haystack.size();
45 for (char needle: needles) {
46 const void* ptr = memchr(haystack.data(), needle, best);
48 auto found = static_cast<const char*>(ptr) - haystack.data();
49 best = std::min<size_t>(best, found);
52 if (best == haystack.size()) {
53 return StringPiece::npos;
62 // It's okay if pages are bigger than this (as powers of two), but they should
64 constexpr size_t kMinPageSize = 4096;
65 static_assert(kMinPageSize >= 16,
66 "kMinPageSize must be at least SSE register size");
67 #define PAGE_FOR(addr) \
68 (reinterpret_cast<uintptr_t>(addr) / kMinPageSize)
71 // Earlier versions of GCC (for example, Clang on Mac OS X, which is based on
72 // GCC 4.2) do not have a full compliment of SSE builtins.
73 #if FOLLY_HAVE_EMMINTRIN_H && __GNUC_PREREQ(4, 6)
74 inline size_t nextAlignedIndex(const char* arr) {
75 auto firstPossible = reinterpret_cast<uintptr_t>(arr) + 1;
76 return 1 + // add 1 because the index starts at 'arr'
77 ((firstPossible + 15) & ~0xF) // round up to next multiple of 16
81 // build sse4.2-optimized version even if -msse4.2 is not passed to GCC
82 size_t qfind_first_byte_of_needles16(const StringPiece& haystack,
83 const StringPiece& needles)
84 __attribute__ ((__target__("sse4.2"), noinline))
85 FOLLY_DISABLE_ADDRESS_SANITIZER;
87 // helper method for case where needles.size() <= 16
88 size_t qfind_first_byte_of_needles16(const StringPiece& haystack,
89 const StringPiece& needles) {
90 DCHECK(!haystack.empty());
91 DCHECK(!needles.empty());
92 DCHECK_LE(needles.size(), 16);
93 // benchmarking shows that memchr beats out SSE for small needle-sets
94 // with large haystacks.
95 if ((needles.size() <= 2 && haystack.size() >= 256) ||
96 // must bail if we can't even SSE-load a single segment of haystack
97 (haystack.size() < 16 &&
98 PAGE_FOR(haystack.end() - 1) != PAGE_FOR(haystack.data() + 15)) ||
99 // can't load needles into SSE register if it could cross page boundary
100 PAGE_FOR(needles.end() - 1) != PAGE_FOR(needles.data() + 15)) {
101 return detail::qfind_first_byte_of_memchr(haystack, needles);
104 auto arr2 = __builtin_ia32_loaddqu(needles.data());
105 // do an unaligned load for first block of haystack
106 auto arr1 = __builtin_ia32_loaddqu(haystack.data());
107 auto index = __builtin_ia32_pcmpestri128(arr2, needles.size(),
108 arr1, haystack.size(), 0);
113 // Now, we can do aligned loads hereafter...
114 size_t i = nextAlignedIndex(haystack.data());
115 for (; i < haystack.size(); i+= 16) {
116 void* ptr1 = __builtin_assume_aligned(haystack.data() + i, 16);
117 auto arr1 = *reinterpret_cast<const __v16qi*>(ptr1);
118 auto index = __builtin_ia32_pcmpestri128(arr2, needles.size(),
119 arr1, haystack.size() - i, 0);
124 return StringPiece::npos;
126 #endif // FOLLY_HAVE_EMMINTRIN_H && GCC 4.6+
128 // Aho, Hopcroft, and Ullman refer to this trick in "The Design and Analysis
129 // of Computer Algorithms" (1974), but the best description is here:
130 // http://research.swtch.com/sparse
133 FastByteSet() : size_(0) { } // no init of arrays required!
135 inline void add(uint8_t i) {
142 inline bool contains(uint8_t i) const {
143 DCHECK_LE(size_, 256);
144 return sparse_[i] < size_ && dense_[sparse_[i]] == i;
148 uint16_t size_; // can't use uint8_t because it would overflow if all
149 // possible values were inserted.
150 uint8_t sparse_[256];
158 size_t qfind_first_byte_of_byteset(const StringPiece& haystack,
159 const StringPiece& needles) {
161 for (auto needle: needles) {
164 for (size_t index = 0; index < haystack.size(); ++index) {
165 if (s.contains(haystack[index])) {
169 return StringPiece::npos;
172 #if FOLLY_HAVE_EMMINTRIN_H && __GNUC_PREREQ(4, 6)
174 template <bool HAYSTACK_ALIGNED>
175 size_t scanHaystackBlock(const StringPiece& haystack,
176 const StringPiece& needles,
178 // inline is okay because it's only called from other sse4.2 functions
179 __attribute__ ((__target__("sse4.2")))
180 // Turn off ASAN because the "arr2 = ..." assignment in the loop below reads
181 // up to 15 bytes beyond end of the buffer in #needles#. That is ok because
182 // ptr2 is always 16-byte aligned, so the read can never span a page boundary.
183 // Also, the extra data that may be read is never actually used.
184 FOLLY_DISABLE_ADDRESS_SANITIZER;
186 // Scans a 16-byte block of haystack (starting at blockStartIdx) to find first
187 // needle. If HAYSTACK_ALIGNED, then haystack must be 16byte aligned.
188 // If !HAYSTACK_ALIGNED, then caller must ensure that it is safe to load the
190 template <bool HAYSTACK_ALIGNED>
191 size_t scanHaystackBlock(const StringPiece& haystack,
192 const StringPiece& needles,
193 int64_t blockStartIdx) {
194 DCHECK_GT(needles.size(), 16); // should handled by *needles16() method
195 DCHECK(blockStartIdx + 16 <= haystack.size() ||
196 (PAGE_FOR(haystack.data() + blockStartIdx) ==
197 PAGE_FOR(haystack.data() + blockStartIdx + 15)));
200 if (HAYSTACK_ALIGNED) {
201 void* ptr1 = __builtin_assume_aligned(haystack.data() + blockStartIdx, 16);
202 arr1 = *reinterpret_cast<const __v16qi*>(ptr1);
204 arr1 = __builtin_ia32_loaddqu(haystack.data() + blockStartIdx);
207 // This load is safe because needles.size() >= 16
208 auto arr2 = __builtin_ia32_loaddqu(needles.data());
209 size_t b = __builtin_ia32_pcmpestri128(
210 arr2, 16, arr1, haystack.size() - blockStartIdx, 0);
212 size_t j = nextAlignedIndex(needles.data());
213 for (; j < needles.size(); j += 16) {
214 void* ptr2 = __builtin_assume_aligned(needles.data() + j, 16);
215 arr2 = *reinterpret_cast<const __v16qi*>(ptr2);
217 auto index = __builtin_ia32_pcmpestri128(
218 arr2, needles.size() - j, arr1, haystack.size() - blockStartIdx, 0);
219 b = std::min<size_t>(index, b);
223 return blockStartIdx + b;
225 return StringPiece::npos;
228 size_t qfind_first_byte_of_sse42(const StringPiece& haystack,
229 const StringPiece& needles)
230 __attribute__ ((__target__("sse4.2"), noinline));
232 size_t qfind_first_byte_of_sse42(const StringPiece& haystack,
233 const StringPiece& needles) {
234 if (UNLIKELY(needles.empty() || haystack.empty())) {
235 return StringPiece::npos;
236 } else if (needles.size() <= 16) {
237 // we can save some unnecessary load instructions by optimizing for
238 // the common case of needles.size() <= 16
239 return qfind_first_byte_of_needles16(haystack, needles);
242 if (haystack.size() < 16 &&
243 PAGE_FOR(haystack.end() - 1) != PAGE_FOR(haystack.data() + 16)) {
244 // We can't safely SSE-load haystack. Use a different approach.
245 if (haystack.size() <= 2) {
246 return qfind_first_of(haystack, needles, asciiCaseSensitive);
248 return qfind_first_byte_of_byteset(haystack, needles);
251 auto ret = scanHaystackBlock<false>(haystack, needles, 0);
252 if (ret != StringPiece::npos) {
256 size_t i = nextAlignedIndex(haystack.data());
257 for (; i < haystack.size(); i += 16) {
258 auto ret = scanHaystackBlock<true>(haystack, needles, i);
259 if (ret != StringPiece::npos) {
264 return StringPiece::npos;
266 #endif // FOLLY_HAVE_EMMINTRIN_H && GCC 4.6+
268 size_t qfind_first_byte_of_nosse(const StringPiece& haystack,
269 const StringPiece& needles) {
270 if (UNLIKELY(needles.empty() || haystack.empty())) {
271 return StringPiece::npos;
273 // The thresholds below were empirically determined by benchmarking.
274 // This is not an exact science since it depends on the CPU, the size of
275 // needles, and the size of haystack.
276 if (haystack.size() == 1 ||
277 (haystack.size() < 4 && needles.size() <= 16)) {
278 return qfind_first_of(haystack, needles, asciiCaseSensitive);
279 } else if ((needles.size() >= 4 && haystack.size() <= 10) ||
280 (needles.size() >= 16 && haystack.size() <= 64) ||
281 needles.size() >= 32) {
282 return qfind_first_byte_of_byteset(haystack, needles);
285 return qfind_first_byte_of_memchr(haystack, needles);
288 } // namespace detail