1 | /*! |
2 | Generic crate-internal routines for the "packed pair" SIMD algorithm. |
3 | |
4 | The "packed pair" algorithm is based on the [generic SIMD] algorithm. The main |
5 | difference is that it (by default) uses a background distribution of byte |
6 | frequencies to heuristically select the pair of bytes to search for. |
7 | |
8 | [generic SIMD]: http://0x80.pl/articles/simd-strfind.html#first-and-last |
9 | */ |
10 | |
11 | use crate::{ |
12 | arch::all::{is_equal_raw, packedpair::Pair}, |
13 | ext::Pointer, |
14 | vector::{MoveMask, Vector}, |
15 | }; |
16 | |
17 | /// A generic architecture dependent "packed pair" finder. |
18 | /// |
19 | /// This finder picks two bytes that it believes have high predictive power |
20 | /// for indicating an overall match of a needle. Depending on whether |
21 | /// `Finder::find` or `Finder::find_prefilter` is used, it reports offsets |
22 | /// where the needle matches or could match. In the prefilter case, candidates |
23 | /// are reported whenever the [`Pair`] of bytes given matches. |
24 | /// |
25 | /// This is architecture dependent because it uses specific vector operations |
26 | /// to look for occurrences of the pair of bytes. |
27 | /// |
28 | /// This type is not meant to be exported and is instead meant to be used as |
29 | /// the implementation for architecture specific facades. Why? Because it's a |
30 | /// bit of a quirky API that requires `inline(always)` annotations. And pretty |
31 | /// much everything has safety obligations due (at least) to the caller needing |
32 | /// to inline calls into routines marked with |
33 | /// `#[target_feature(enable = "...")]`. |
34 | #[derive (Clone, Copy, Debug)] |
35 | pub(crate) struct Finder<V> { |
36 | pair: Pair, |
37 | v1: V, |
38 | v2: V, |
39 | min_haystack_len: usize, |
40 | } |
41 | |
42 | impl<V: Vector> Finder<V> { |
43 | /// Create a new pair searcher. The searcher returned can either report |
44 | /// exact matches of `needle` or act as a prefilter and report candidate |
45 | /// positions of `needle`. |
46 | /// |
47 | /// # Safety |
48 | /// |
49 | /// Callers must ensure that whatever vector type this routine is called |
50 | /// with is supported by the current environment. |
51 | /// |
52 | /// Callers must also ensure that `needle.len() >= 2`. |
53 | #[inline (always)] |
54 | pub(crate) unsafe fn new(needle: &[u8], pair: Pair) -> Finder<V> { |
55 | let max_index = pair.index1().max(pair.index2()); |
56 | let min_haystack_len = |
57 | core::cmp::max(needle.len(), usize::from(max_index) + V::BYTES); |
58 | let v1 = V::splat(needle[usize::from(pair.index1())]); |
59 | let v2 = V::splat(needle[usize::from(pair.index2())]); |
60 | Finder { pair, v1, v2, min_haystack_len } |
61 | } |
62 | |
63 | /// Searches the given haystack for the given needle. The needle given |
64 | /// should be the same as the needle that this finder was initialized |
65 | /// with. |
66 | /// |
67 | /// # Panics |
68 | /// |
69 | /// When `haystack.len()` is less than [`Finder::min_haystack_len`]. |
70 | /// |
71 | /// # Safety |
72 | /// |
73 | /// Since this is meant to be used with vector functions, callers need to |
74 | /// specialize this inside of a function with a `target_feature` attribute. |
75 | /// Therefore, callers must ensure that whatever target feature is being |
76 | /// used supports the vector functions that this function is specialized |
77 | /// for. (For the specific vector functions used, see the Vector trait |
78 | /// implementations.) |
79 | #[inline (always)] |
80 | pub(crate) unsafe fn find( |
81 | &self, |
82 | haystack: &[u8], |
83 | needle: &[u8], |
84 | ) -> Option<usize> { |
85 | assert!( |
86 | haystack.len() >= self.min_haystack_len, |
87 | "haystack too small, should be at least {} but got {}" , |
88 | self.min_haystack_len, |
89 | haystack.len(), |
90 | ); |
91 | |
92 | let all = V::Mask::all_zeros_except_least_significant(0); |
93 | let start = haystack.as_ptr(); |
94 | let end = start.add(haystack.len()); |
95 | let max = end.sub(self.min_haystack_len); |
96 | let mut cur = start; |
97 | |
98 | // N.B. I did experiment with unrolling the loop to deal with size(V) |
99 | // bytes at a time and 2*size(V) bytes at a time. The double unroll |
100 | // was marginally faster while the quadruple unroll was unambiguously |
101 | // slower. In the end, I decided the complexity from unrolling wasn't |
102 | // worth it. I used the memmem/krate/prebuilt/huge-en/ benchmarks to |
103 | // compare. |
104 | while cur <= max { |
105 | if let Some(chunki) = self.find_in_chunk(needle, cur, end, all) { |
106 | return Some(matched(start, cur, chunki)); |
107 | } |
108 | cur = cur.add(V::BYTES); |
109 | } |
110 | if cur < end { |
111 | let remaining = end.distance(cur); |
112 | debug_assert!( |
113 | remaining < self.min_haystack_len, |
114 | "remaining bytes should be smaller than the minimum haystack \ |
115 | length of {}, but there are {} bytes remaining" , |
116 | self.min_haystack_len, |
117 | remaining, |
118 | ); |
119 | if remaining < needle.len() { |
120 | return None; |
121 | } |
122 | debug_assert!( |
123 | max < cur, |
124 | "after main loop, cur should have exceeded max" , |
125 | ); |
126 | let overlap = cur.distance(max); |
127 | debug_assert!( |
128 | overlap > 0, |
129 | "overlap ( {}) must always be non-zero" , |
130 | overlap, |
131 | ); |
132 | debug_assert!( |
133 | overlap < V::BYTES, |
134 | "overlap ( {}) cannot possibly be >= than a vector ( {})" , |
135 | overlap, |
136 | V::BYTES, |
137 | ); |
138 | // The mask has all of its bits set except for the first N least |
139 | // significant bits, where N=overlap. This way, any matches that |
140 | // occur in find_in_chunk within the overlap are automatically |
141 | // ignored. |
142 | let mask = V::Mask::all_zeros_except_least_significant(overlap); |
143 | cur = max; |
144 | let m = self.find_in_chunk(needle, cur, end, mask); |
145 | if let Some(chunki) = m { |
146 | return Some(matched(start, cur, chunki)); |
147 | } |
148 | } |
149 | None |
150 | } |
151 | |
152 | /// Searches the given haystack for offsets that represent candidate |
153 | /// matches of the `needle` given to this finder's constructor. The offsets |
154 | /// returned, if they are a match, correspond to the starting offset of |
155 | /// `needle` in the given `haystack`. |
156 | /// |
157 | /// # Panics |
158 | /// |
159 | /// When `haystack.len()` is less than [`Finder::min_haystack_len`]. |
160 | /// |
161 | /// # Safety |
162 | /// |
163 | /// Since this is meant to be used with vector functions, callers need to |
164 | /// specialize this inside of a function with a `target_feature` attribute. |
165 | /// Therefore, callers must ensure that whatever target feature is being |
166 | /// used supports the vector functions that this function is specialized |
167 | /// for. (For the specific vector functions used, see the Vector trait |
168 | /// implementations.) |
169 | #[inline (always)] |
170 | pub(crate) unsafe fn find_prefilter( |
171 | &self, |
172 | haystack: &[u8], |
173 | ) -> Option<usize> { |
174 | assert!( |
175 | haystack.len() >= self.min_haystack_len, |
176 | "haystack too small, should be at least {} but got {}" , |
177 | self.min_haystack_len, |
178 | haystack.len(), |
179 | ); |
180 | |
181 | let start = haystack.as_ptr(); |
182 | let end = start.add(haystack.len()); |
183 | let max = end.sub(self.min_haystack_len); |
184 | let mut cur = start; |
185 | |
186 | // N.B. I did experiment with unrolling the loop to deal with size(V) |
187 | // bytes at a time and 2*size(V) bytes at a time. The double unroll |
188 | // was marginally faster while the quadruple unroll was unambiguously |
189 | // slower. In the end, I decided the complexity from unrolling wasn't |
190 | // worth it. I used the memmem/krate/prebuilt/huge-en/ benchmarks to |
191 | // compare. |
192 | while cur <= max { |
193 | if let Some(chunki) = self.find_prefilter_in_chunk(cur) { |
194 | return Some(matched(start, cur, chunki)); |
195 | } |
196 | cur = cur.add(V::BYTES); |
197 | } |
198 | if cur < end { |
199 | // This routine immediately quits if a candidate match is found. |
200 | // That means that if we're here, no candidate matches have been |
201 | // found at or before 'ptr'. Thus, we don't need to mask anything |
202 | // out even though we might technically search part of the haystack |
203 | // that we've already searched (because we know it can't match). |
204 | cur = max; |
205 | if let Some(chunki) = self.find_prefilter_in_chunk(cur) { |
206 | return Some(matched(start, cur, chunki)); |
207 | } |
208 | } |
209 | None |
210 | } |
211 | |
212 | /// Search for an occurrence of our byte pair from the needle in the chunk |
213 | /// pointed to by cur, with the end of the haystack pointed to by end. |
214 | /// When an occurrence is found, memcmp is run to check if a match occurs |
215 | /// at the corresponding position. |
216 | /// |
217 | /// `mask` should have bits set corresponding the positions in the chunk |
218 | /// in which matches are considered. This is only used for the last vector |
219 | /// load where the beginning of the vector might have overlapped with the |
220 | /// last load in the main loop. The mask lets us avoid visiting positions |
221 | /// that have already been discarded as matches. |
222 | /// |
223 | /// # Safety |
224 | /// |
225 | /// It must be safe to do an unaligned read of size(V) bytes starting at |
226 | /// both (cur + self.index1) and (cur + self.index2). It must also be safe |
227 | /// to do unaligned loads on cur up to (end - needle.len()). |
228 | #[inline (always)] |
229 | unsafe fn find_in_chunk( |
230 | &self, |
231 | needle: &[u8], |
232 | cur: *const u8, |
233 | end: *const u8, |
234 | mask: V::Mask, |
235 | ) -> Option<usize> { |
236 | let index1 = usize::from(self.pair.index1()); |
237 | let index2 = usize::from(self.pair.index2()); |
238 | let chunk1 = V::load_unaligned(cur.add(index1)); |
239 | let chunk2 = V::load_unaligned(cur.add(index2)); |
240 | let eq1 = chunk1.cmpeq(self.v1); |
241 | let eq2 = chunk2.cmpeq(self.v2); |
242 | |
243 | let mut offsets = eq1.and(eq2).movemask().and(mask); |
244 | while offsets.has_non_zero() { |
245 | let offset = offsets.first_offset(); |
246 | let cur = cur.add(offset); |
247 | if end.sub(needle.len()) < cur { |
248 | return None; |
249 | } |
250 | if is_equal_raw(needle.as_ptr(), cur, needle.len()) { |
251 | return Some(offset); |
252 | } |
253 | offsets = offsets.clear_least_significant_bit(); |
254 | } |
255 | None |
256 | } |
257 | |
258 | /// Search for an occurrence of our byte pair from the needle in the chunk |
259 | /// pointed to by cur, with the end of the haystack pointed to by end. |
260 | /// When an occurrence is found, memcmp is run to check if a match occurs |
261 | /// at the corresponding position. |
262 | /// |
263 | /// # Safety |
264 | /// |
265 | /// It must be safe to do an unaligned read of size(V) bytes starting at |
266 | /// both (cur + self.index1) and (cur + self.index2). It must also be safe |
267 | /// to do unaligned reads on cur up to (end - needle.len()). |
268 | #[inline (always)] |
269 | unsafe fn find_prefilter_in_chunk(&self, cur: *const u8) -> Option<usize> { |
270 | let index1 = usize::from(self.pair.index1()); |
271 | let index2 = usize::from(self.pair.index2()); |
272 | let chunk1 = V::load_unaligned(cur.add(index1)); |
273 | let chunk2 = V::load_unaligned(cur.add(index2)); |
274 | let eq1 = chunk1.cmpeq(self.v1); |
275 | let eq2 = chunk2.cmpeq(self.v2); |
276 | |
277 | let offsets = eq1.and(eq2).movemask(); |
278 | if !offsets.has_non_zero() { |
279 | return None; |
280 | } |
281 | Some(offsets.first_offset()) |
282 | } |
283 | |
284 | /// Returns the pair of offsets (into the needle) used to check as a |
285 | /// predicate before confirming whether a needle exists at a particular |
286 | /// position. |
287 | #[inline ] |
288 | pub(crate) fn pair(&self) -> &Pair { |
289 | &self.pair |
290 | } |
291 | |
292 | /// Returns the minimum haystack length that this `Finder` can search. |
293 | /// |
294 | /// Providing a haystack to this `Finder` shorter than this length is |
295 | /// guaranteed to result in a panic. |
296 | #[inline (always)] |
297 | pub(crate) fn min_haystack_len(&self) -> usize { |
298 | self.min_haystack_len |
299 | } |
300 | } |
301 | |
302 | /// Accepts a chunk-relative offset and returns a haystack relative offset. |
303 | /// |
304 | /// This used to be marked `#[cold]` and `#[inline(never)]`, but I couldn't |
305 | /// observe a consistent measureable difference between that and just inlining |
306 | /// it. So we go with inlining it. |
307 | /// |
308 | /// # Safety |
309 | /// |
310 | /// Same at `ptr::offset_from` in addition to `cur >= start`. |
311 | #[inline (always)] |
312 | unsafe fn matched(start: *const u8, cur: *const u8, chunki: usize) -> usize { |
313 | cur.distance(origin:start) + chunki |
314 | } |
315 | |
316 | // If you're looking for tests, those are run for each instantiation of the |
317 | // above code. So for example, see arch::x86_64::sse2::packedpair. |
318 | |