1 | use crate::alloc::alloc::{handle_alloc_error, Layout}; |
2 | use crate::control::{BitMaskIter, Group, Tag, TagSliceExt}; |
3 | use crate::scopeguard::{guard, ScopeGuard}; |
4 | use crate::util::{invalid_mut, likely, unlikely}; |
5 | use crate::TryReserveError; |
6 | use core::array; |
7 | use core::iter::FusedIterator; |
8 | use core::marker::PhantomData; |
9 | use core::mem; |
10 | use core::ptr::NonNull; |
11 | use core::slice; |
12 | use core::{hint, ptr}; |
13 | |
14 | mod alloc; |
15 | #[cfg (test)] |
16 | pub(crate) use self::alloc::AllocError; |
17 | pub(crate) use self::alloc::{do_alloc, Allocator, Global}; |
18 | |
19 | #[inline ] |
20 | unsafe fn offset_from<T>(to: *const T, from: *const T) -> usize { |
21 | to.offset_from(origin:from) as usize |
22 | } |
23 | |
24 | /// Whether memory allocation errors should return an error or abort. |
25 | #[derive (Copy, Clone)] |
26 | enum Fallibility { |
27 | Fallible, |
28 | Infallible, |
29 | } |
30 | |
31 | impl Fallibility { |
32 | /// Error to return on capacity overflow. |
33 | #[cfg_attr (feature = "inline-more" , inline)] |
34 | fn capacity_overflow(self) -> TryReserveError { |
35 | match self { |
36 | Fallibility::Fallible => TryReserveError::CapacityOverflow, |
37 | Fallibility::Infallible => panic!("Hash table capacity overflow" ), |
38 | } |
39 | } |
40 | |
41 | /// Error to return on allocation error. |
42 | #[cfg_attr (feature = "inline-more" , inline)] |
43 | fn alloc_err(self, layout: Layout) -> TryReserveError { |
44 | match self { |
45 | Fallibility::Fallible => TryReserveError::AllocError { layout }, |
46 | Fallibility::Infallible => handle_alloc_error(layout), |
47 | } |
48 | } |
49 | } |
50 | |
51 | trait SizedTypeProperties: Sized { |
52 | const IS_ZERO_SIZED: bool = mem::size_of::<Self>() == 0; |
53 | const NEEDS_DROP: bool = mem::needs_drop::<Self>(); |
54 | } |
55 | |
56 | impl<T> SizedTypeProperties for T {} |
57 | |
58 | /// Primary hash function, used to select the initial bucket to probe from. |
59 | #[inline ] |
60 | #[allow (clippy::cast_possible_truncation)] |
61 | fn h1(hash: u64) -> usize { |
62 | // On 32-bit platforms we simply ignore the higher hash bits. |
63 | hash as usize |
64 | } |
65 | |
66 | /// Probe sequence based on triangular numbers, which is guaranteed (since our |
67 | /// table size is a power of two) to visit every group of elements exactly once. |
68 | /// |
69 | /// A triangular probe has us jump by 1 more group every time. So first we |
70 | /// jump by 1 group (meaning we just continue our linear scan), then 2 groups |
71 | /// (skipping over 1 group), then 3 groups (skipping over 2 groups), and so on. |
72 | /// |
73 | /// Proof that the probe will visit every group in the table: |
74 | /// <https://fgiesen.wordpress.com/2015/02/22/triangular-numbers-mod-2n/> |
75 | #[derive (Clone)] |
76 | struct ProbeSeq { |
77 | pos: usize, |
78 | stride: usize, |
79 | } |
80 | |
81 | impl ProbeSeq { |
82 | #[inline ] |
83 | fn move_next(&mut self, bucket_mask: usize) { |
84 | // We should have found an empty bucket by now and ended the probe. |
85 | debug_assert!( |
86 | self.stride <= bucket_mask, |
87 | "Went past end of probe sequence" |
88 | ); |
89 | |
90 | self.stride += Group::WIDTH; |
91 | self.pos += self.stride; |
92 | self.pos &= bucket_mask; |
93 | } |
94 | } |
95 | |
96 | /// Returns the number of buckets needed to hold the given number of items, |
97 | /// taking the maximum load factor into account. |
98 | /// |
99 | /// Returns `None` if an overflow occurs. |
100 | // Workaround for emscripten bug emscripten-core/emscripten-fastcomp#258 |
101 | #[cfg_attr (target_os = "emscripten" , inline(never))] |
102 | #[cfg_attr (not(target_os = "emscripten" ), inline)] |
103 | fn capacity_to_buckets(cap: usize, table_layout: TableLayout) -> Option<usize> { |
104 | debug_assert_ne!(cap, 0); |
105 | |
106 | // For small tables we require at least 1 empty bucket so that lookups are |
107 | // guaranteed to terminate if an element doesn't exist in the table. |
108 | if cap < 15 { |
109 | // Consider a small TableLayout like { size: 1, ctrl_align: 16 } on a |
110 | // platform with Group::WIDTH of 16 (like x86_64 with SSE2). For small |
111 | // bucket sizes, this ends up wasting quite a few bytes just to pad to |
112 | // the relatively larger ctrl_align: |
113 | // |
114 | // | capacity | buckets | bytes allocated | bytes per item | |
115 | // | -------- | ------- | --------------- | -------------- | |
116 | // | 3 | 4 | 36 | (Yikes!) 12.0 | |
117 | // | 7 | 8 | 40 | (Poor) 5.7 | |
118 | // | 14 | 16 | 48 | 3.4 | |
119 | // | 28 | 32 | 80 | 3.3 | |
120 | // |
121 | // In general, buckets * table_layout.size >= table_layout.ctrl_align |
122 | // must be true to avoid these edges. This is implemented by adjusting |
123 | // the minimum capacity upwards for small items. This code only needs |
124 | // to handle ctrl_align which are less than or equal to Group::WIDTH, |
125 | // because valid layout sizes are always a multiple of the alignment, |
126 | // so anything with alignment over the Group::WIDTH won't hit this edge |
127 | // case. |
128 | |
129 | // This is brittle, e.g. if we ever add 32 byte groups, it will select |
130 | // 3 regardless of the table_layout.size. |
131 | let min_cap = match (Group::WIDTH, table_layout.size) { |
132 | (16, 0..=1) => 14, |
133 | (16, 2..=3) => 7, |
134 | (8, 0..=1) => 7, |
135 | _ => 3, |
136 | }; |
137 | let cap = min_cap.max(cap); |
138 | // We don't bother with a table size of 2 buckets since that can only |
139 | // hold a single element. Instead, we skip directly to a 4 bucket table |
140 | // which can hold 3 elements. |
141 | return Some(if cap < 4 { |
142 | 4 |
143 | } else if cap < 8 { |
144 | 8 |
145 | } else { |
146 | 16 |
147 | }); |
148 | } |
149 | |
150 | // Otherwise require 1/8 buckets to be empty (87.5% load) |
151 | // |
152 | // Be careful when modifying this, calculate_layout relies on the |
153 | // overflow check here. |
154 | let adjusted_cap = cap.checked_mul(8)? / 7; |
155 | |
156 | // Any overflows will have been caught by the checked_mul. Also, any |
157 | // rounding errors from the division above will be cleaned up by |
158 | // next_power_of_two (which can't overflow because of the previous division). |
159 | Some(adjusted_cap.next_power_of_two()) |
160 | } |
161 | |
162 | /// Returns the maximum effective capacity for the given bucket mask, taking |
163 | /// the maximum load factor into account. |
164 | #[inline ] |
165 | fn bucket_mask_to_capacity(bucket_mask: usize) -> usize { |
166 | if bucket_mask < 8 { |
167 | // For tables with 1/2/4/8 buckets, we always reserve one empty slot. |
168 | // Keep in mind that the bucket mask is one less than the bucket count. |
169 | bucket_mask |
170 | } else { |
171 | // For larger tables we reserve 12.5% of the slots as empty. |
172 | ((bucket_mask + 1) / 8) * 7 |
173 | } |
174 | } |
175 | |
176 | /// Helper which allows the max calculation for `ctrl_align` to be statically computed for each `T` |
177 | /// while keeping the rest of `calculate_layout_for` independent of `T` |
178 | #[derive (Copy, Clone)] |
179 | struct TableLayout { |
180 | size: usize, |
181 | ctrl_align: usize, |
182 | } |
183 | |
184 | impl TableLayout { |
185 | #[inline ] |
186 | const fn new<T>() -> Self { |
187 | let layout = Layout::new::<T>(); |
188 | Self { |
189 | size: layout.size(), |
190 | ctrl_align: if layout.align() > Group::WIDTH { |
191 | layout.align() |
192 | } else { |
193 | Group::WIDTH |
194 | }, |
195 | } |
196 | } |
197 | |
198 | #[inline ] |
199 | fn calculate_layout_for(self, buckets: usize) -> Option<(Layout, usize)> { |
200 | debug_assert!(buckets.is_power_of_two()); |
201 | |
202 | let TableLayout { size, ctrl_align } = self; |
203 | // Manual layout calculation since Layout methods are not yet stable. |
204 | let ctrl_offset = |
205 | size.checked_mul(buckets)?.checked_add(ctrl_align - 1)? & !(ctrl_align - 1); |
206 | let len = ctrl_offset.checked_add(buckets + Group::WIDTH)?; |
207 | |
208 | // We need an additional check to ensure that the allocation doesn't |
209 | // exceed `isize::MAX` (https://github.com/rust-lang/rust/pull/95295). |
210 | if len > isize::MAX as usize - (ctrl_align - 1) { |
211 | return None; |
212 | } |
213 | |
214 | Some(( |
215 | unsafe { Layout::from_size_align_unchecked(len, ctrl_align) }, |
216 | ctrl_offset, |
217 | )) |
218 | } |
219 | } |
220 | |
221 | /// A reference to an empty bucket into which an can be inserted. |
222 | pub struct InsertSlot { |
223 | index: usize, |
224 | } |
225 | |
226 | /// A reference to a hash table bucket containing a `T`. |
227 | /// |
228 | /// This is usually just a pointer to the element itself. However if the element |
229 | /// is a ZST, then we instead track the index of the element in the table so |
230 | /// that `erase` works properly. |
231 | pub struct Bucket<T> { |
232 | // Actually it is pointer to next element than element itself |
233 | // this is needed to maintain pointer arithmetic invariants |
234 | // keeping direct pointer to element introduces difficulty. |
235 | // Using `NonNull` for variance and niche layout |
236 | ptr: NonNull<T>, |
237 | } |
238 | |
239 | // This Send impl is needed for rayon support. This is safe since Bucket is |
240 | // never exposed in a public API. |
241 | unsafe impl<T> Send for Bucket<T> {} |
242 | |
243 | impl<T> Clone for Bucket<T> { |
244 | #[inline ] |
245 | fn clone(&self) -> Self { |
246 | Self { ptr: self.ptr } |
247 | } |
248 | } |
249 | |
250 | impl<T> Bucket<T> { |
251 | /// Creates a [`Bucket`] that contain pointer to the data. |
252 | /// The pointer calculation is performed by calculating the |
253 | /// offset from given `base` pointer (convenience for |
254 | /// `base.as_ptr().sub(index)`). |
255 | /// |
256 | /// `index` is in units of `T`; e.g., an `index` of 3 represents a pointer |
257 | /// offset of `3 * size_of::<T>()` bytes. |
258 | /// |
259 | /// If the `T` is a ZST, then we instead track the index of the element |
260 | /// in the table so that `erase` works properly (return |
261 | /// `NonNull::new_unchecked((index + 1) as *mut T)`) |
262 | /// |
263 | /// # Safety |
264 | /// |
265 | /// If `mem::size_of::<T>() != 0`, then the safety rules are directly derived |
266 | /// from the safety rules for [`<*mut T>::sub`] method of `*mut T` and the safety |
267 | /// rules of [`NonNull::new_unchecked`] function. |
268 | /// |
269 | /// Thus, in order to uphold the safety contracts for the [`<*mut T>::sub`] method |
270 | /// and [`NonNull::new_unchecked`] function, as well as for the correct |
271 | /// logic of the work of this crate, the following rules are necessary and |
272 | /// sufficient: |
273 | /// |
274 | /// * the `base` pointer must not be `dangling` and must points to the |
275 | /// end of the first `value element` from the `data part` of the table, i.e. |
276 | /// must be the pointer that returned by [`RawTable::data_end`] or by |
277 | /// [`RawTableInner::data_end<T>`]; |
278 | /// |
279 | /// * `index` must not be greater than `RawTableInner.bucket_mask`, i.e. |
280 | /// `index <= RawTableInner.bucket_mask` or, in other words, `(index + 1)` |
281 | /// must be no greater than the number returned by the function |
282 | /// [`RawTable::buckets`] or [`RawTableInner::buckets`]. |
283 | /// |
284 | /// If `mem::size_of::<T>() == 0`, then the only requirement is that the |
285 | /// `index` must not be greater than `RawTableInner.bucket_mask`, i.e. |
286 | /// `index <= RawTableInner.bucket_mask` or, in other words, `(index + 1)` |
287 | /// must be no greater than the number returned by the function |
288 | /// [`RawTable::buckets`] or [`RawTableInner::buckets`]. |
289 | /// |
290 | /// [`Bucket`]: crate::raw::Bucket |
291 | /// [`<*mut T>::sub`]: https://doc.rust-lang.org/core/primitive.pointer.html#method.sub-1 |
292 | /// [`NonNull::new_unchecked`]: https://doc.rust-lang.org/stable/std/ptr/struct.NonNull.html#method.new_unchecked |
293 | /// [`RawTable::data_end`]: crate::raw::RawTable::data_end |
294 | /// [`RawTableInner::data_end<T>`]: RawTableInner::data_end<T> |
295 | /// [`RawTable::buckets`]: crate::raw::RawTable::buckets |
296 | /// [`RawTableInner::buckets`]: RawTableInner::buckets |
297 | #[inline ] |
298 | unsafe fn from_base_index(base: NonNull<T>, index: usize) -> Self { |
299 | // If mem::size_of::<T>() != 0 then return a pointer to an `element` in |
300 | // the data part of the table (we start counting from "0", so that |
301 | // in the expression T[last], the "last" index actually one less than the |
302 | // "buckets" number in the table, i.e. "last = RawTableInner.bucket_mask"): |
303 | // |
304 | // `from_base_index(base, 1).as_ptr()` returns a pointer that |
305 | // points here in the data part of the table |
306 | // (to the start of T1) |
307 | // | |
308 | // | `base: NonNull<T>` must point here |
309 | // | (to the end of T0 or to the start of C0) |
310 | // v v |
311 | // [Padding], Tlast, ..., |T1|, T0, |C0, C1, ..., Clast |
312 | // ^ |
313 | // `from_base_index(base, 1)` returns a pointer |
314 | // that points here in the data part of the table |
315 | // (to the end of T1) |
316 | // |
317 | // where: T0...Tlast - our stored data; C0...Clast - control bytes |
318 | // or metadata for data. |
319 | let ptr = if T::IS_ZERO_SIZED { |
320 | // won't overflow because index must be less than length (bucket_mask) |
321 | // and bucket_mask is guaranteed to be less than `isize::MAX` |
322 | // (see TableLayout::calculate_layout_for method) |
323 | invalid_mut(index + 1) |
324 | } else { |
325 | base.as_ptr().sub(index) |
326 | }; |
327 | Self { |
328 | ptr: NonNull::new_unchecked(ptr), |
329 | } |
330 | } |
331 | |
332 | /// Calculates the index of a [`Bucket`] as distance between two pointers |
333 | /// (convenience for `base.as_ptr().offset_from(self.ptr.as_ptr()) as usize`). |
334 | /// The returned value is in units of T: the distance in bytes divided by |
335 | /// [`core::mem::size_of::<T>()`]. |
336 | /// |
337 | /// If the `T` is a ZST, then we return the index of the element in |
338 | /// the table so that `erase` works properly (return `self.ptr.as_ptr() as usize - 1`). |
339 | /// |
340 | /// This function is the inverse of [`from_base_index`]. |
341 | /// |
342 | /// # Safety |
343 | /// |
344 | /// If `mem::size_of::<T>() != 0`, then the safety rules are directly derived |
345 | /// from the safety rules for [`<*const T>::offset_from`] method of `*const T`. |
346 | /// |
347 | /// Thus, in order to uphold the safety contracts for [`<*const T>::offset_from`] |
348 | /// method, as well as for the correct logic of the work of this crate, the |
349 | /// following rules are necessary and sufficient: |
350 | /// |
351 | /// * `base` contained pointer must not be `dangling` and must point to the |
352 | /// end of the first `element` from the `data part` of the table, i.e. |
353 | /// must be a pointer that returns by [`RawTable::data_end`] or by |
354 | /// [`RawTableInner::data_end<T>`]; |
355 | /// |
356 | /// * `self` also must not contain dangling pointer; |
357 | /// |
358 | /// * both `self` and `base` must be created from the same [`RawTable`] |
359 | /// (or [`RawTableInner`]). |
360 | /// |
361 | /// If `mem::size_of::<T>() == 0`, this function is always safe. |
362 | /// |
363 | /// [`Bucket`]: crate::raw::Bucket |
364 | /// [`from_base_index`]: crate::raw::Bucket::from_base_index |
365 | /// [`RawTable::data_end`]: crate::raw::RawTable::data_end |
366 | /// [`RawTableInner::data_end<T>`]: RawTableInner::data_end<T> |
367 | /// [`RawTable`]: crate::raw::RawTable |
368 | /// [`RawTableInner`]: RawTableInner |
369 | /// [`<*const T>::offset_from`]: https://doc.rust-lang.org/nightly/core/primitive.pointer.html#method.offset_from |
370 | #[inline ] |
371 | unsafe fn to_base_index(&self, base: NonNull<T>) -> usize { |
372 | // If mem::size_of::<T>() != 0 then return an index under which we used to store the |
373 | // `element` in the data part of the table (we start counting from "0", so |
374 | // that in the expression T[last], the "last" index actually is one less than the |
375 | // "buckets" number in the table, i.e. "last = RawTableInner.bucket_mask"). |
376 | // For example for 5th element in table calculation is performed like this: |
377 | // |
378 | // mem::size_of::<T>() |
379 | // | |
380 | // | `self = from_base_index(base, 5)` that returns pointer |
381 | // | that points here in the data part of the table |
382 | // | (to the end of T5) |
383 | // | | `base: NonNull<T>` must point here |
384 | // v | (to the end of T0 or to the start of C0) |
385 | // /???\ v v |
386 | // [Padding], Tlast, ..., |T10|, ..., T5|, T4, T3, T2, T1, T0, |C0, C1, C2, C3, C4, C5, ..., C10, ..., Clast |
387 | // \__________ __________/ |
388 | // \/ |
389 | // `bucket.to_base_index(base)` = 5 |
390 | // (base.as_ptr() as usize - self.ptr.as_ptr() as usize) / mem::size_of::<T>() |
391 | // |
392 | // where: T0...Tlast - our stored data; C0...Clast - control bytes or metadata for data. |
393 | if T::IS_ZERO_SIZED { |
394 | // this can not be UB |
395 | self.ptr.as_ptr() as usize - 1 |
396 | } else { |
397 | offset_from(base.as_ptr(), self.ptr.as_ptr()) |
398 | } |
399 | } |
400 | |
401 | /// Acquires the underlying raw pointer `*mut T` to `data`. |
402 | /// |
403 | /// # Note |
404 | /// |
405 | /// If `T` is not [`Copy`], do not use `*mut T` methods that can cause calling the |
406 | /// destructor of `T` (for example the [`<*mut T>::drop_in_place`] method), because |
407 | /// for properly dropping the data we also need to clear `data` control bytes. If we |
408 | /// drop data, but do not clear `data control byte` it leads to double drop when |
409 | /// [`RawTable`] goes out of scope. |
410 | /// |
411 | /// If you modify an already initialized `value`, so [`Hash`] and [`Eq`] on the new |
412 | /// `T` value and its borrowed form *must* match those for the old `T` value, as the map |
413 | /// will not re-evaluate where the new value should go, meaning the value may become |
414 | /// "lost" if their location does not reflect their state. |
415 | /// |
416 | /// [`RawTable`]: crate::raw::RawTable |
417 | /// [`<*mut T>::drop_in_place`]: https://doc.rust-lang.org/core/primitive.pointer.html#method.drop_in_place |
418 | /// [`Hash`]: https://doc.rust-lang.org/core/hash/trait.Hash.html |
419 | /// [`Eq`]: https://doc.rust-lang.org/core/cmp/trait.Eq.html |
420 | #[inline ] |
421 | pub fn as_ptr(&self) -> *mut T { |
422 | if T::IS_ZERO_SIZED { |
423 | // Just return an arbitrary ZST pointer which is properly aligned |
424 | // invalid pointer is good enough for ZST |
425 | invalid_mut(mem::align_of::<T>()) |
426 | } else { |
427 | unsafe { self.ptr.as_ptr().sub(1) } |
428 | } |
429 | } |
430 | |
431 | /// Acquires the underlying non-null pointer `*mut T` to `data`. |
432 | #[inline ] |
433 | fn as_non_null(&self) -> NonNull<T> { |
434 | // SAFETY: `self.ptr` is already a `NonNull` |
435 | unsafe { NonNull::new_unchecked(self.as_ptr()) } |
436 | } |
437 | |
438 | /// Create a new [`Bucket`] that is offset from the `self` by the given |
439 | /// `offset`. The pointer calculation is performed by calculating the |
440 | /// offset from `self` pointer (convenience for `self.ptr.as_ptr().sub(offset)`). |
441 | /// This function is used for iterators. |
442 | /// |
443 | /// `offset` is in units of `T`; e.g., a `offset` of 3 represents a pointer |
444 | /// offset of `3 * size_of::<T>()` bytes. |
445 | /// |
446 | /// # Safety |
447 | /// |
448 | /// If `mem::size_of::<T>() != 0`, then the safety rules are directly derived |
449 | /// from the safety rules for [`<*mut T>::sub`] method of `*mut T` and safety |
450 | /// rules of [`NonNull::new_unchecked`] function. |
451 | /// |
452 | /// Thus, in order to uphold the safety contracts for [`<*mut T>::sub`] method |
453 | /// and [`NonNull::new_unchecked`] function, as well as for the correct |
454 | /// logic of the work of this crate, the following rules are necessary and |
455 | /// sufficient: |
456 | /// |
457 | /// * `self` contained pointer must not be `dangling`; |
458 | /// |
459 | /// * `self.to_base_index() + offset` must not be greater than `RawTableInner.bucket_mask`, |
460 | /// i.e. `(self.to_base_index() + offset) <= RawTableInner.bucket_mask` or, in other |
461 | /// words, `self.to_base_index() + offset + 1` must be no greater than the number returned |
462 | /// by the function [`RawTable::buckets`] or [`RawTableInner::buckets`]. |
463 | /// |
464 | /// If `mem::size_of::<T>() == 0`, then the only requirement is that the |
465 | /// `self.to_base_index() + offset` must not be greater than `RawTableInner.bucket_mask`, |
466 | /// i.e. `(self.to_base_index() + offset) <= RawTableInner.bucket_mask` or, in other words, |
467 | /// `self.to_base_index() + offset + 1` must be no greater than the number returned by the |
468 | /// function [`RawTable::buckets`] or [`RawTableInner::buckets`]. |
469 | /// |
470 | /// [`Bucket`]: crate::raw::Bucket |
471 | /// [`<*mut T>::sub`]: https://doc.rust-lang.org/core/primitive.pointer.html#method.sub-1 |
472 | /// [`NonNull::new_unchecked`]: https://doc.rust-lang.org/stable/std/ptr/struct.NonNull.html#method.new_unchecked |
473 | /// [`RawTable::buckets`]: crate::raw::RawTable::buckets |
474 | /// [`RawTableInner::buckets`]: RawTableInner::buckets |
475 | #[inline ] |
476 | unsafe fn next_n(&self, offset: usize) -> Self { |
477 | let ptr = if T::IS_ZERO_SIZED { |
478 | // invalid pointer is good enough for ZST |
479 | invalid_mut(self.ptr.as_ptr() as usize + offset) |
480 | } else { |
481 | self.ptr.as_ptr().sub(offset) |
482 | }; |
483 | Self { |
484 | ptr: NonNull::new_unchecked(ptr), |
485 | } |
486 | } |
487 | |
488 | /// Executes the destructor (if any) of the pointed-to `data`. |
489 | /// |
490 | /// # Safety |
491 | /// |
492 | /// See [`ptr::drop_in_place`] for safety concerns. |
493 | /// |
494 | /// You should use [`RawTable::erase`] instead of this function, |
495 | /// or be careful with calling this function directly, because for |
496 | /// properly dropping the data we need also clear `data` control bytes. |
497 | /// If we drop data, but do not erase `data control byte` it leads to |
498 | /// double drop when [`RawTable`] goes out of scope. |
499 | /// |
500 | /// [`ptr::drop_in_place`]: https://doc.rust-lang.org/core/ptr/fn.drop_in_place.html |
501 | /// [`RawTable`]: crate::raw::RawTable |
502 | /// [`RawTable::erase`]: crate::raw::RawTable::erase |
503 | #[cfg_attr (feature = "inline-more" , inline)] |
504 | pub(crate) unsafe fn drop(&self) { |
505 | self.as_ptr().drop_in_place(); |
506 | } |
507 | |
508 | /// Reads the `value` from `self` without moving it. This leaves the |
509 | /// memory in `self` unchanged. |
510 | /// |
511 | /// # Safety |
512 | /// |
513 | /// See [`ptr::read`] for safety concerns. |
514 | /// |
515 | /// You should use [`RawTable::remove`] instead of this function, |
516 | /// or be careful with calling this function directly, because compiler |
517 | /// calls its destructor when the read `value` goes out of scope. It |
518 | /// can cause double dropping when [`RawTable`] goes out of scope, |
519 | /// because of not erased `data control byte`. |
520 | /// |
521 | /// [`ptr::read`]: https://doc.rust-lang.org/core/ptr/fn.read.html |
522 | /// [`RawTable`]: crate::raw::RawTable |
523 | /// [`RawTable::remove`]: crate::raw::RawTable::remove |
524 | #[inline ] |
525 | pub(crate) unsafe fn read(&self) -> T { |
526 | self.as_ptr().read() |
527 | } |
528 | |
529 | /// Overwrites a memory location with the given `value` without reading |
530 | /// or dropping the old value (like [`ptr::write`] function). |
531 | /// |
532 | /// # Safety |
533 | /// |
534 | /// See [`ptr::write`] for safety concerns. |
535 | /// |
536 | /// # Note |
537 | /// |
538 | /// [`Hash`] and [`Eq`] on the new `T` value and its borrowed form *must* match |
539 | /// those for the old `T` value, as the map will not re-evaluate where the new |
540 | /// value should go, meaning the value may become "lost" if their location |
541 | /// does not reflect their state. |
542 | /// |
543 | /// [`ptr::write`]: https://doc.rust-lang.org/core/ptr/fn.write.html |
544 | /// [`Hash`]: https://doc.rust-lang.org/core/hash/trait.Hash.html |
545 | /// [`Eq`]: https://doc.rust-lang.org/core/cmp/trait.Eq.html |
546 | #[inline ] |
547 | pub(crate) unsafe fn write(&self, val: T) { |
548 | self.as_ptr().write(val); |
549 | } |
550 | |
551 | /// Returns a shared immutable reference to the `value`. |
552 | /// |
553 | /// # Safety |
554 | /// |
555 | /// See [`NonNull::as_ref`] for safety concerns. |
556 | /// |
557 | /// [`NonNull::as_ref`]: https://doc.rust-lang.org/core/ptr/struct.NonNull.html#method.as_ref |
558 | #[inline ] |
559 | pub unsafe fn as_ref<'a>(&self) -> &'a T { |
560 | &*self.as_ptr() |
561 | } |
562 | |
563 | /// Returns a unique mutable reference to the `value`. |
564 | /// |
565 | /// # Safety |
566 | /// |
567 | /// See [`NonNull::as_mut`] for safety concerns. |
568 | /// |
569 | /// # Note |
570 | /// |
571 | /// [`Hash`] and [`Eq`] on the new `T` value and its borrowed form *must* match |
572 | /// those for the old `T` value, as the map will not re-evaluate where the new |
573 | /// value should go, meaning the value may become "lost" if their location |
574 | /// does not reflect their state. |
575 | /// |
576 | /// [`NonNull::as_mut`]: https://doc.rust-lang.org/core/ptr/struct.NonNull.html#method.as_mut |
577 | /// [`Hash`]: https://doc.rust-lang.org/core/hash/trait.Hash.html |
578 | /// [`Eq`]: https://doc.rust-lang.org/core/cmp/trait.Eq.html |
579 | #[inline ] |
580 | pub unsafe fn as_mut<'a>(&self) -> &'a mut T { |
581 | &mut *self.as_ptr() |
582 | } |
583 | } |
584 | |
585 | /// A raw hash table with an unsafe API. |
586 | pub struct RawTable<T, A: Allocator = Global> { |
587 | table: RawTableInner, |
588 | alloc: A, |
589 | // Tell dropck that we own instances of T. |
590 | marker: PhantomData<T>, |
591 | } |
592 | |
593 | /// Non-generic part of `RawTable` which allows functions to be instantiated only once regardless |
594 | /// of how many different key-value types are used. |
595 | struct RawTableInner { |
596 | // Mask to get an index from a hash value. The value is one less than the |
597 | // number of buckets in the table. |
598 | bucket_mask: usize, |
599 | |
600 | // [Padding], T_n, ..., T1, T0, C0, C1, ... |
601 | // ^ points here |
602 | ctrl: NonNull<u8>, |
603 | |
604 | // Number of elements that can be inserted before we need to grow the table |
605 | growth_left: usize, |
606 | |
607 | // Number of elements in the table, only really used by len() |
608 | items: usize, |
609 | } |
610 | |
611 | impl<T> RawTable<T, Global> { |
612 | /// Creates a new empty hash table without allocating any memory. |
613 | /// |
614 | /// In effect this returns a table with exactly 1 bucket. However we can |
615 | /// leave the data pointer dangling since that bucket is never written to |
616 | /// due to our load factor forcing us to always have at least 1 free bucket. |
617 | #[inline ] |
618 | #[cfg_attr (feature = "rustc-dep-of-std" , rustc_const_stable_indirect)] |
619 | pub const fn new() -> Self { |
620 | Self { |
621 | table: RawTableInner::NEW, |
622 | alloc: Global, |
623 | marker: PhantomData, |
624 | } |
625 | } |
626 | |
627 | /// Allocates a new hash table with at least enough capacity for inserting |
628 | /// the given number of elements without reallocating. |
629 | pub fn with_capacity(capacity: usize) -> Self { |
630 | Self::with_capacity_in(capacity, alloc:Global) |
631 | } |
632 | } |
633 | |
634 | impl<T, A: Allocator> RawTable<T, A> { |
635 | const TABLE_LAYOUT: TableLayout = TableLayout::new::<T>(); |
636 | |
637 | /// Creates a new empty hash table without allocating any memory, using the |
638 | /// given allocator. |
639 | /// |
640 | /// In effect this returns a table with exactly 1 bucket. However we can |
641 | /// leave the data pointer dangling since that bucket is never written to |
642 | /// due to our load factor forcing us to always have at least 1 free bucket. |
643 | #[inline ] |
644 | #[cfg_attr (feature = "rustc-dep-of-std" , rustc_const_stable_indirect)] |
645 | pub const fn new_in(alloc: A) -> Self { |
646 | Self { |
647 | table: RawTableInner::NEW, |
648 | alloc, |
649 | marker: PhantomData, |
650 | } |
651 | } |
652 | |
653 | /// Allocates a new hash table with the given number of buckets. |
654 | /// |
655 | /// The control bytes are left uninitialized. |
656 | #[cfg_attr (feature = "inline-more" , inline)] |
657 | unsafe fn new_uninitialized( |
658 | alloc: A, |
659 | buckets: usize, |
660 | fallibility: Fallibility, |
661 | ) -> Result<Self, TryReserveError> { |
662 | debug_assert!(buckets.is_power_of_two()); |
663 | |
664 | Ok(Self { |
665 | table: RawTableInner::new_uninitialized( |
666 | &alloc, |
667 | Self::TABLE_LAYOUT, |
668 | buckets, |
669 | fallibility, |
670 | )?, |
671 | alloc, |
672 | marker: PhantomData, |
673 | }) |
674 | } |
675 | |
676 | /// Allocates a new hash table using the given allocator, with at least enough capacity for |
677 | /// inserting the given number of elements without reallocating. |
678 | pub fn with_capacity_in(capacity: usize, alloc: A) -> Self { |
679 | Self { |
680 | table: RawTableInner::with_capacity(&alloc, Self::TABLE_LAYOUT, capacity), |
681 | alloc, |
682 | marker: PhantomData, |
683 | } |
684 | } |
685 | |
686 | /// Returns a reference to the underlying allocator. |
687 | #[inline ] |
688 | pub fn allocator(&self) -> &A { |
689 | &self.alloc |
690 | } |
691 | |
692 | /// Returns pointer to one past last `data` element in the table as viewed from |
693 | /// the start point of the allocation. |
694 | /// |
695 | /// The caller must ensure that the `RawTable` outlives the returned [`NonNull<T>`], |
696 | /// otherwise using it may result in [`undefined behavior`]. |
697 | /// |
698 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
699 | #[inline ] |
700 | pub fn data_end(&self) -> NonNull<T> { |
701 | // `self.table.ctrl.cast()` returns pointer that |
702 | // points here (to the end of `T0`) |
703 | // ∨ |
704 | // [Pad], T_n, ..., T1, T0, |CT0, CT1, ..., CT_n|, CTa_0, CTa_1, ..., CTa_m |
705 | // \________ ________/ |
706 | // \/ |
707 | // `n = buckets - 1`, i.e. `RawTable::buckets() - 1` |
708 | // |
709 | // where: T0...T_n - our stored data; |
710 | // CT0...CT_n - control bytes or metadata for `data`. |
711 | // CTa_0...CTa_m - additional control bytes, where `m = Group::WIDTH - 1` (so that the search |
712 | // with loading `Group` bytes from the heap works properly, even if the result |
713 | // of `h1(hash) & self.bucket_mask` is equal to `self.bucket_mask`). See also |
714 | // `RawTableInner::set_ctrl` function. |
715 | // |
716 | // P.S. `h1(hash) & self.bucket_mask` is the same as `hash as usize % self.buckets()` because the number |
717 | // of buckets is a power of two, and `self.bucket_mask = self.buckets() - 1`. |
718 | self.table.ctrl.cast() |
719 | } |
720 | |
721 | /// Returns pointer to start of data table. |
722 | #[inline ] |
723 | #[cfg (feature = "nightly" )] |
724 | pub unsafe fn data_start(&self) -> NonNull<T> { |
725 | NonNull::new_unchecked(self.data_end().as_ptr().wrapping_sub(self.buckets())) |
726 | } |
727 | |
728 | /// Returns the total amount of memory allocated internally by the hash |
729 | /// table, in bytes. |
730 | /// |
731 | /// The returned number is informational only. It is intended to be |
732 | /// primarily used for memory profiling. |
733 | #[inline ] |
734 | pub fn allocation_size(&self) -> usize { |
735 | // SAFETY: We use the same `table_layout` that was used to allocate |
736 | // this table. |
737 | unsafe { self.table.allocation_size_or_zero(Self::TABLE_LAYOUT) } |
738 | } |
739 | |
740 | /// Returns the index of a bucket from a `Bucket`. |
741 | #[inline ] |
742 | pub unsafe fn bucket_index(&self, bucket: &Bucket<T>) -> usize { |
743 | bucket.to_base_index(self.data_end()) |
744 | } |
745 | |
746 | /// Returns a pointer to an element in the table. |
747 | /// |
748 | /// The caller must ensure that the `RawTable` outlives the returned [`Bucket<T>`], |
749 | /// otherwise using it may result in [`undefined behavior`]. |
750 | /// |
751 | /// # Safety |
752 | /// |
753 | /// If `mem::size_of::<T>() != 0`, then the caller of this function must observe the |
754 | /// following safety rules: |
755 | /// |
756 | /// * The table must already be allocated; |
757 | /// |
758 | /// * The `index` must not be greater than the number returned by the [`RawTable::buckets`] |
759 | /// function, i.e. `(index + 1) <= self.buckets()`. |
760 | /// |
761 | /// It is safe to call this function with index of zero (`index == 0`) on a table that has |
762 | /// not been allocated, but using the returned [`Bucket`] results in [`undefined behavior`]. |
763 | /// |
764 | /// If `mem::size_of::<T>() == 0`, then the only requirement is that the `index` must |
765 | /// not be greater than the number returned by the [`RawTable::buckets`] function, i.e. |
766 | /// `(index + 1) <= self.buckets()`. |
767 | /// |
768 | /// [`RawTable::buckets`]: RawTable::buckets |
769 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
770 | #[inline ] |
771 | pub unsafe fn bucket(&self, index: usize) -> Bucket<T> { |
772 | // If mem::size_of::<T>() != 0 then return a pointer to the `element` in the `data part` of the table |
773 | // (we start counting from "0", so that in the expression T[n], the "n" index actually one less than |
774 | // the "buckets" number of our `RawTable`, i.e. "n = RawTable::buckets() - 1"): |
775 | // |
776 | // `table.bucket(3).as_ptr()` returns a pointer that points here in the `data` |
777 | // part of the `RawTable`, i.e. to the start of T3 (see `Bucket::as_ptr`) |
778 | // | |
779 | // | `base = self.data_end()` points here |
780 | // | (to the start of CT0 or to the end of T0) |
781 | // v v |
782 | // [Pad], T_n, ..., |T3|, T2, T1, T0, |CT0, CT1, CT2, CT3, ..., CT_n, CTa_0, CTa_1, ..., CTa_m |
783 | // ^ \__________ __________/ |
784 | // `table.bucket(3)` returns a pointer that points \/ |
785 | // here in the `data` part of the `RawTable` (to additional control bytes |
786 | // the end of T3) `m = Group::WIDTH - 1` |
787 | // |
788 | // where: T0...T_n - our stored data; |
789 | // CT0...CT_n - control bytes or metadata for `data`; |
790 | // CTa_0...CTa_m - additional control bytes (so that the search with loading `Group` bytes from |
791 | // the heap works properly, even if the result of `h1(hash) & self.table.bucket_mask` |
792 | // is equal to `self.table.bucket_mask`). See also `RawTableInner::set_ctrl` function. |
793 | // |
794 | // P.S. `h1(hash) & self.table.bucket_mask` is the same as `hash as usize % self.buckets()` because the number |
795 | // of buckets is a power of two, and `self.table.bucket_mask = self.buckets() - 1`. |
796 | debug_assert_ne!(self.table.bucket_mask, 0); |
797 | debug_assert!(index < self.buckets()); |
798 | Bucket::from_base_index(self.data_end(), index) |
799 | } |
800 | |
801 | /// Erases an element from the table without dropping it. |
802 | #[cfg_attr (feature = "inline-more" , inline)] |
803 | unsafe fn erase_no_drop(&mut self, item: &Bucket<T>) { |
804 | let index = self.bucket_index(item); |
805 | self.table.erase(index); |
806 | } |
807 | |
808 | /// Erases an element from the table, dropping it in place. |
809 | #[cfg_attr (feature = "inline-more" , inline)] |
810 | #[allow (clippy::needless_pass_by_value)] |
811 | pub unsafe fn erase(&mut self, item: Bucket<T>) { |
812 | // Erase the element from the table first since drop might panic. |
813 | self.erase_no_drop(&item); |
814 | item.drop(); |
815 | } |
816 | |
817 | /// Removes an element from the table, returning it. |
818 | /// |
819 | /// This also returns an `InsertSlot` pointing to the newly free bucket. |
820 | #[cfg_attr (feature = "inline-more" , inline)] |
821 | #[allow (clippy::needless_pass_by_value)] |
822 | pub unsafe fn remove(&mut self, item: Bucket<T>) -> (T, InsertSlot) { |
823 | self.erase_no_drop(&item); |
824 | ( |
825 | item.read(), |
826 | InsertSlot { |
827 | index: self.bucket_index(&item), |
828 | }, |
829 | ) |
830 | } |
831 | |
832 | /// Finds and removes an element from the table, returning it. |
833 | #[cfg_attr (feature = "inline-more" , inline)] |
834 | pub fn remove_entry(&mut self, hash: u64, eq: impl FnMut(&T) -> bool) -> Option<T> { |
835 | // Avoid `Option::map` because it bloats LLVM IR. |
836 | match self.find(hash, eq) { |
837 | Some(bucket) => Some(unsafe { self.remove(bucket).0 }), |
838 | None => None, |
839 | } |
840 | } |
841 | |
842 | /// Marks all table buckets as empty without dropping their contents. |
843 | #[cfg_attr (feature = "inline-more" , inline)] |
844 | pub fn clear_no_drop(&mut self) { |
845 | self.table.clear_no_drop(); |
846 | } |
847 | |
848 | /// Removes all elements from the table without freeing the backing memory. |
849 | #[cfg_attr (feature = "inline-more" , inline)] |
850 | pub fn clear(&mut self) { |
851 | if self.is_empty() { |
852 | // Special case empty table to avoid surprising O(capacity) time. |
853 | return; |
854 | } |
855 | // Ensure that the table is reset even if one of the drops panic |
856 | let mut self_ = guard(self, |self_| self_.clear_no_drop()); |
857 | unsafe { |
858 | // SAFETY: ScopeGuard sets to zero the `items` field of the table |
859 | // even in case of panic during the dropping of the elements so |
860 | // that there will be no double drop of the elements. |
861 | self_.table.drop_elements::<T>(); |
862 | } |
863 | } |
864 | |
865 | /// Shrinks the table to fit `max(self.len(), min_size)` elements. |
866 | #[cfg_attr (feature = "inline-more" , inline)] |
867 | pub fn shrink_to(&mut self, min_size: usize, hasher: impl Fn(&T) -> u64) { |
868 | // Calculate the minimal number of elements that we need to reserve |
869 | // space for. |
870 | let min_size = usize::max(self.table.items, min_size); |
871 | if min_size == 0 { |
872 | let mut old_inner = mem::replace(&mut self.table, RawTableInner::NEW); |
873 | unsafe { |
874 | // SAFETY: |
875 | // 1. We call the function only once; |
876 | // 2. We know for sure that `alloc` and `table_layout` matches the [`Allocator`] |
877 | // and [`TableLayout`] that were used to allocate this table. |
878 | // 3. If any elements' drop function panics, then there will only be a memory leak, |
879 | // because we have replaced the inner table with a new one. |
880 | old_inner.drop_inner_table::<T, _>(&self.alloc, Self::TABLE_LAYOUT); |
881 | } |
882 | return; |
883 | } |
884 | |
885 | // Calculate the number of buckets that we need for this number of |
886 | // elements. If the calculation overflows then the requested bucket |
887 | // count must be larger than what we have right and nothing needs to be |
888 | // done. |
889 | let min_buckets = match capacity_to_buckets(min_size, Self::TABLE_LAYOUT) { |
890 | Some(buckets) => buckets, |
891 | None => return, |
892 | }; |
893 | |
894 | // If we have more buckets than we need, shrink the table. |
895 | if min_buckets < self.buckets() { |
896 | // Fast path if the table is empty |
897 | if self.table.items == 0 { |
898 | let new_inner = |
899 | RawTableInner::with_capacity(&self.alloc, Self::TABLE_LAYOUT, min_size); |
900 | let mut old_inner = mem::replace(&mut self.table, new_inner); |
901 | unsafe { |
902 | // SAFETY: |
903 | // 1. We call the function only once; |
904 | // 2. We know for sure that `alloc` and `table_layout` matches the [`Allocator`] |
905 | // and [`TableLayout`] that were used to allocate this table. |
906 | // 3. If any elements' drop function panics, then there will only be a memory leak, |
907 | // because we have replaced the inner table with a new one. |
908 | old_inner.drop_inner_table::<T, _>(&self.alloc, Self::TABLE_LAYOUT); |
909 | } |
910 | } else { |
911 | // Avoid `Result::unwrap_or_else` because it bloats LLVM IR. |
912 | unsafe { |
913 | // SAFETY: |
914 | // 1. We know for sure that `min_size >= self.table.items`. |
915 | // 2. The [`RawTableInner`] must already have properly initialized control bytes since |
916 | // we will never expose RawTable::new_uninitialized in a public API. |
917 | if self |
918 | .resize(min_size, hasher, Fallibility::Infallible) |
919 | .is_err() |
920 | { |
921 | // SAFETY: The result of calling the `resize` function cannot be an error |
922 | // because `fallibility == Fallibility::Infallible. |
923 | hint::unreachable_unchecked() |
924 | } |
925 | } |
926 | } |
927 | } |
928 | } |
929 | |
930 | /// Ensures that at least `additional` items can be inserted into the table |
931 | /// without reallocation. |
932 | #[cfg_attr (feature = "inline-more" , inline)] |
933 | pub fn reserve(&mut self, additional: usize, hasher: impl Fn(&T) -> u64) { |
934 | if unlikely(additional > self.table.growth_left) { |
935 | // Avoid `Result::unwrap_or_else` because it bloats LLVM IR. |
936 | unsafe { |
937 | // SAFETY: The [`RawTableInner`] must already have properly initialized control |
938 | // bytes since we will never expose RawTable::new_uninitialized in a public API. |
939 | if self |
940 | .reserve_rehash(additional, hasher, Fallibility::Infallible) |
941 | .is_err() |
942 | { |
943 | // SAFETY: All allocation errors will be caught inside `RawTableInner::reserve_rehash`. |
944 | hint::unreachable_unchecked() |
945 | } |
946 | } |
947 | } |
948 | } |
949 | |
950 | /// Tries to ensure that at least `additional` items can be inserted into |
951 | /// the table without reallocation. |
952 | #[cfg_attr (feature = "inline-more" , inline)] |
953 | pub fn try_reserve( |
954 | &mut self, |
955 | additional: usize, |
956 | hasher: impl Fn(&T) -> u64, |
957 | ) -> Result<(), TryReserveError> { |
958 | if additional > self.table.growth_left { |
959 | // SAFETY: The [`RawTableInner`] must already have properly initialized control |
960 | // bytes since we will never expose RawTable::new_uninitialized in a public API. |
961 | unsafe { self.reserve_rehash(additional, hasher, Fallibility::Fallible) } |
962 | } else { |
963 | Ok(()) |
964 | } |
965 | } |
966 | |
967 | /// Out-of-line slow path for `reserve` and `try_reserve`. |
968 | /// |
969 | /// # Safety |
970 | /// |
971 | /// The [`RawTableInner`] must have properly initialized control bytes, |
972 | /// otherwise calling this function results in [`undefined behavior`] |
973 | /// |
974 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
975 | #[cold ] |
976 | #[inline (never)] |
977 | unsafe fn reserve_rehash( |
978 | &mut self, |
979 | additional: usize, |
980 | hasher: impl Fn(&T) -> u64, |
981 | fallibility: Fallibility, |
982 | ) -> Result<(), TryReserveError> { |
983 | unsafe { |
984 | // SAFETY: |
985 | // 1. We know for sure that `alloc` and `layout` matches the [`Allocator`] and |
986 | // [`TableLayout`] that were used to allocate this table. |
987 | // 2. The `drop` function is the actual drop function of the elements stored in |
988 | // the table. |
989 | // 3. The caller ensures that the control bytes of the `RawTableInner` |
990 | // are already initialized. |
991 | self.table.reserve_rehash_inner( |
992 | &self.alloc, |
993 | additional, |
994 | &|table, index| hasher(table.bucket::<T>(index).as_ref()), |
995 | fallibility, |
996 | Self::TABLE_LAYOUT, |
997 | if T::NEEDS_DROP { |
998 | Some(|ptr| ptr::drop_in_place(ptr as *mut T)) |
999 | } else { |
1000 | None |
1001 | }, |
1002 | ) |
1003 | } |
1004 | } |
1005 | |
1006 | /// Allocates a new table of a different size and moves the contents of the |
1007 | /// current table into it. |
1008 | /// |
1009 | /// # Safety |
1010 | /// |
1011 | /// The [`RawTableInner`] must have properly initialized control bytes, |
1012 | /// otherwise calling this function results in [`undefined behavior`] |
1013 | /// |
1014 | /// The caller of this function must ensure that `capacity >= self.table.items` |
1015 | /// otherwise: |
1016 | /// |
1017 | /// * If `self.table.items != 0`, calling of this function with `capacity` |
1018 | /// equal to 0 (`capacity == 0`) results in [`undefined behavior`]. |
1019 | /// |
1020 | /// * If `self.table.items > capacity_to_buckets(capacity, Self::TABLE_LAYOUT)` |
1021 | /// calling this function are never return (will loop infinitely). |
1022 | /// |
1023 | /// See [`RawTableInner::find_insert_slot`] for more information. |
1024 | /// |
1025 | /// [`RawTableInner::find_insert_slot`]: RawTableInner::find_insert_slot |
1026 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
1027 | unsafe fn resize( |
1028 | &mut self, |
1029 | capacity: usize, |
1030 | hasher: impl Fn(&T) -> u64, |
1031 | fallibility: Fallibility, |
1032 | ) -> Result<(), TryReserveError> { |
1033 | // SAFETY: |
1034 | // 1. The caller of this function guarantees that `capacity >= self.table.items`. |
1035 | // 2. We know for sure that `alloc` and `layout` matches the [`Allocator`] and |
1036 | // [`TableLayout`] that were used to allocate this table. |
1037 | // 3. The caller ensures that the control bytes of the `RawTableInner` |
1038 | // are already initialized. |
1039 | self.table.resize_inner( |
1040 | &self.alloc, |
1041 | capacity, |
1042 | &|table, index| hasher(table.bucket::<T>(index).as_ref()), |
1043 | fallibility, |
1044 | Self::TABLE_LAYOUT, |
1045 | ) |
1046 | } |
1047 | |
1048 | /// Inserts a new element into the table, and returns its raw bucket. |
1049 | /// |
1050 | /// This does not check if the given element already exists in the table. |
1051 | #[cfg_attr (feature = "inline-more" , inline)] |
1052 | pub fn insert(&mut self, hash: u64, value: T, hasher: impl Fn(&T) -> u64) -> Bucket<T> { |
1053 | unsafe { |
1054 | // SAFETY: |
1055 | // 1. The [`RawTableInner`] must already have properly initialized control bytes since |
1056 | // we will never expose `RawTable::new_uninitialized` in a public API. |
1057 | // |
1058 | // 2. We reserve additional space (if necessary) right after calling this function. |
1059 | let mut slot = self.table.find_insert_slot(hash); |
1060 | |
1061 | // We can avoid growing the table once we have reached our load factor if we are replacing |
1062 | // a tombstone. This works since the number of EMPTY slots does not change in this case. |
1063 | // |
1064 | // SAFETY: The function is guaranteed to return [`InsertSlot`] that contains an index |
1065 | // in the range `0..=self.buckets()`. |
1066 | let old_ctrl = *self.table.ctrl(slot.index); |
1067 | if unlikely(self.table.growth_left == 0 && old_ctrl.special_is_empty()) { |
1068 | self.reserve(1, hasher); |
1069 | // SAFETY: We know for sure that `RawTableInner` has control bytes |
1070 | // initialized and that there is extra space in the table. |
1071 | slot = self.table.find_insert_slot(hash); |
1072 | } |
1073 | |
1074 | self.insert_in_slot(hash, slot, value) |
1075 | } |
1076 | } |
1077 | |
1078 | /// Inserts a new element into the table, and returns a mutable reference to it. |
1079 | /// |
1080 | /// This does not check if the given element already exists in the table. |
1081 | #[cfg_attr (feature = "inline-more" , inline)] |
1082 | pub fn insert_entry(&mut self, hash: u64, value: T, hasher: impl Fn(&T) -> u64) -> &mut T { |
1083 | unsafe { self.insert(hash, value, hasher).as_mut() } |
1084 | } |
1085 | |
1086 | /// Inserts a new element into the table, without growing the table. |
1087 | /// |
1088 | /// There must be enough space in the table to insert the new element. |
1089 | /// |
1090 | /// This does not check if the given element already exists in the table. |
1091 | #[cfg_attr (feature = "inline-more" , inline)] |
1092 | #[cfg (feature = "rustc-internal-api" )] |
1093 | pub unsafe fn insert_no_grow(&mut self, hash: u64, value: T) -> Bucket<T> { |
1094 | let (index, old_ctrl) = self.table.prepare_insert_slot(hash); |
1095 | let bucket = self.table.bucket(index); |
1096 | |
1097 | // If we are replacing a DELETED entry then we don't need to update |
1098 | // the load counter. |
1099 | self.table.growth_left -= old_ctrl.special_is_empty() as usize; |
1100 | |
1101 | bucket.write(value); |
1102 | self.table.items += 1; |
1103 | bucket |
1104 | } |
1105 | |
1106 | /// Temporary removes a bucket, applying the given function to the removed |
1107 | /// element and optionally put back the returned value in the same bucket. |
1108 | /// |
1109 | /// Returns `true` if the bucket still contains an element |
1110 | /// |
1111 | /// This does not check if the given bucket is actually occupied. |
1112 | #[cfg_attr (feature = "inline-more" , inline)] |
1113 | pub unsafe fn replace_bucket_with<F>(&mut self, bucket: Bucket<T>, f: F) -> bool |
1114 | where |
1115 | F: FnOnce(T) -> Option<T>, |
1116 | { |
1117 | let index = self.bucket_index(&bucket); |
1118 | let old_ctrl = *self.table.ctrl(index); |
1119 | debug_assert!(self.is_bucket_full(index)); |
1120 | let old_growth_left = self.table.growth_left; |
1121 | let item = self.remove(bucket).0; |
1122 | if let Some(new_item) = f(item) { |
1123 | self.table.growth_left = old_growth_left; |
1124 | self.table.set_ctrl(index, old_ctrl); |
1125 | self.table.items += 1; |
1126 | self.bucket(index).write(new_item); |
1127 | true |
1128 | } else { |
1129 | false |
1130 | } |
1131 | } |
1132 | |
1133 | /// Searches for an element in the table. If the element is not found, |
1134 | /// returns `Err` with the position of a slot where an element with the |
1135 | /// same hash could be inserted. |
1136 | /// |
1137 | /// This function may resize the table if additional space is required for |
1138 | /// inserting an element. |
1139 | #[inline ] |
1140 | pub fn find_or_find_insert_slot( |
1141 | &mut self, |
1142 | hash: u64, |
1143 | mut eq: impl FnMut(&T) -> bool, |
1144 | hasher: impl Fn(&T) -> u64, |
1145 | ) -> Result<Bucket<T>, InsertSlot> { |
1146 | self.reserve(1, hasher); |
1147 | |
1148 | unsafe { |
1149 | // SAFETY: |
1150 | // 1. We know for sure that there is at least one empty `bucket` in the table. |
1151 | // 2. The [`RawTableInner`] must already have properly initialized control bytes since we will |
1152 | // never expose `RawTable::new_uninitialized` in a public API. |
1153 | // 3. The `find_or_find_insert_slot_inner` function returns the `index` of only the full bucket, |
1154 | // which is in the range `0..self.buckets()` (since there is at least one empty `bucket` in |
1155 | // the table), so calling `self.bucket(index)` and `Bucket::as_ref` is safe. |
1156 | match self |
1157 | .table |
1158 | .find_or_find_insert_slot_inner(hash, &mut |index| eq(self.bucket(index).as_ref())) |
1159 | { |
1160 | // SAFETY: See explanation above. |
1161 | Ok(index) => Ok(self.bucket(index)), |
1162 | Err(slot) => Err(slot), |
1163 | } |
1164 | } |
1165 | } |
1166 | |
1167 | /// Inserts a new element into the table in the given slot, and returns its |
1168 | /// raw bucket. |
1169 | /// |
1170 | /// # Safety |
1171 | /// |
1172 | /// `slot` must point to a slot previously returned by |
1173 | /// `find_or_find_insert_slot`, and no mutation of the table must have |
1174 | /// occurred since that call. |
1175 | #[inline ] |
1176 | pub unsafe fn insert_in_slot(&mut self, hash: u64, slot: InsertSlot, value: T) -> Bucket<T> { |
1177 | let old_ctrl = *self.table.ctrl(slot.index); |
1178 | self.table.record_item_insert_at(slot.index, old_ctrl, hash); |
1179 | |
1180 | let bucket = self.bucket(slot.index); |
1181 | bucket.write(value); |
1182 | bucket |
1183 | } |
1184 | |
1185 | /// Searches for an element in the table. |
1186 | #[inline ] |
1187 | pub fn find(&self, hash: u64, mut eq: impl FnMut(&T) -> bool) -> Option<Bucket<T>> { |
1188 | unsafe { |
1189 | // SAFETY: |
1190 | // 1. The [`RawTableInner`] must already have properly initialized control bytes since we |
1191 | // will never expose `RawTable::new_uninitialized` in a public API. |
1192 | // 1. The `find_inner` function returns the `index` of only the full bucket, which is in |
1193 | // the range `0..self.buckets()`, so calling `self.bucket(index)` and `Bucket::as_ref` |
1194 | // is safe. |
1195 | let result = self |
1196 | .table |
1197 | .find_inner(hash, &mut |index| eq(self.bucket(index).as_ref())); |
1198 | |
1199 | // Avoid `Option::map` because it bloats LLVM IR. |
1200 | match result { |
1201 | // SAFETY: See explanation above. |
1202 | Some(index) => Some(self.bucket(index)), |
1203 | None => None, |
1204 | } |
1205 | } |
1206 | } |
1207 | |
1208 | /// Gets a reference to an element in the table. |
1209 | #[inline ] |
1210 | pub fn get(&self, hash: u64, eq: impl FnMut(&T) -> bool) -> Option<&T> { |
1211 | // Avoid `Option::map` because it bloats LLVM IR. |
1212 | match self.find(hash, eq) { |
1213 | Some(bucket) => Some(unsafe { bucket.as_ref() }), |
1214 | None => None, |
1215 | } |
1216 | } |
1217 | |
1218 | /// Gets a mutable reference to an element in the table. |
1219 | #[inline ] |
1220 | pub fn get_mut(&mut self, hash: u64, eq: impl FnMut(&T) -> bool) -> Option<&mut T> { |
1221 | // Avoid `Option::map` because it bloats LLVM IR. |
1222 | match self.find(hash, eq) { |
1223 | Some(bucket) => Some(unsafe { bucket.as_mut() }), |
1224 | None => None, |
1225 | } |
1226 | } |
1227 | |
1228 | /// Attempts to get mutable references to `N` entries in the table at once. |
1229 | /// |
1230 | /// Returns an array of length `N` with the results of each query. |
1231 | /// |
1232 | /// At most one mutable reference will be returned to any entry. `None` will be returned if any |
1233 | /// of the hashes are duplicates. `None` will be returned if the hash is not found. |
1234 | /// |
1235 | /// The `eq` argument should be a closure such that `eq(i, k)` returns true if `k` is equal to |
1236 | /// the `i`th key to be looked up. |
1237 | pub fn get_many_mut<const N: usize>( |
1238 | &mut self, |
1239 | hashes: [u64; N], |
1240 | eq: impl FnMut(usize, &T) -> bool, |
1241 | ) -> [Option<&'_ mut T>; N] { |
1242 | unsafe { |
1243 | let ptrs = self.get_many_mut_pointers(hashes, eq); |
1244 | |
1245 | for (i, cur) in ptrs.iter().enumerate() { |
1246 | if cur.is_some() && ptrs[..i].contains(cur) { |
1247 | panic!("duplicate keys found" ); |
1248 | } |
1249 | } |
1250 | // All bucket are distinct from all previous buckets so we're clear to return the result |
1251 | // of the lookup. |
1252 | |
1253 | ptrs.map(|ptr| ptr.map(|mut ptr| ptr.as_mut())) |
1254 | } |
1255 | } |
1256 | |
1257 | pub unsafe fn get_many_unchecked_mut<const N: usize>( |
1258 | &mut self, |
1259 | hashes: [u64; N], |
1260 | eq: impl FnMut(usize, &T) -> bool, |
1261 | ) -> [Option<&'_ mut T>; N] { |
1262 | let ptrs = self.get_many_mut_pointers(hashes, eq); |
1263 | ptrs.map(|ptr| ptr.map(|mut ptr| ptr.as_mut())) |
1264 | } |
1265 | |
1266 | unsafe fn get_many_mut_pointers<const N: usize>( |
1267 | &mut self, |
1268 | hashes: [u64; N], |
1269 | mut eq: impl FnMut(usize, &T) -> bool, |
1270 | ) -> [Option<NonNull<T>>; N] { |
1271 | array::from_fn(|i| { |
1272 | self.find(hashes[i], |k| eq(i, k)) |
1273 | .map(|cur| cur.as_non_null()) |
1274 | }) |
1275 | } |
1276 | |
1277 | /// Returns the number of elements the map can hold without reallocating. |
1278 | /// |
1279 | /// This number is a lower bound; the table might be able to hold |
1280 | /// more, but is guaranteed to be able to hold at least this many. |
1281 | #[inline ] |
1282 | pub fn capacity(&self) -> usize { |
1283 | self.table.items + self.table.growth_left |
1284 | } |
1285 | |
1286 | /// Returns the number of elements in the table. |
1287 | #[inline ] |
1288 | pub fn len(&self) -> usize { |
1289 | self.table.items |
1290 | } |
1291 | |
1292 | /// Returns `true` if the table contains no elements. |
1293 | #[inline ] |
1294 | pub fn is_empty(&self) -> bool { |
1295 | self.len() == 0 |
1296 | } |
1297 | |
1298 | /// Returns the number of buckets in the table. |
1299 | #[inline ] |
1300 | pub fn buckets(&self) -> usize { |
1301 | self.table.bucket_mask + 1 |
1302 | } |
1303 | |
1304 | /// Checks whether the bucket at `index` is full. |
1305 | /// |
1306 | /// # Safety |
1307 | /// |
1308 | /// The caller must ensure `index` is less than the number of buckets. |
1309 | #[inline ] |
1310 | pub unsafe fn is_bucket_full(&self, index: usize) -> bool { |
1311 | self.table.is_bucket_full(index) |
1312 | } |
1313 | |
1314 | /// Returns an iterator over every element in the table. It is up to |
1315 | /// the caller to ensure that the `RawTable` outlives the `RawIter`. |
1316 | /// Because we cannot make the `next` method unsafe on the `RawIter` |
1317 | /// struct, we have to make the `iter` method unsafe. |
1318 | #[inline ] |
1319 | pub unsafe fn iter(&self) -> RawIter<T> { |
1320 | // SAFETY: |
1321 | // 1. The caller must uphold the safety contract for `iter` method. |
1322 | // 2. The [`RawTableInner`] must already have properly initialized control bytes since |
1323 | // we will never expose RawTable::new_uninitialized in a public API. |
1324 | self.table.iter() |
1325 | } |
1326 | |
1327 | /// Returns an iterator over occupied buckets that could match a given hash. |
1328 | /// |
1329 | /// `RawTable` only stores 7 bits of the hash value, so this iterator may |
1330 | /// return items that have a hash value different than the one provided. You |
1331 | /// should always validate the returned values before using them. |
1332 | /// |
1333 | /// It is up to the caller to ensure that the `RawTable` outlives the |
1334 | /// `RawIterHash`. Because we cannot make the `next` method unsafe on the |
1335 | /// `RawIterHash` struct, we have to make the `iter_hash` method unsafe. |
1336 | #[cfg_attr (feature = "inline-more" , inline)] |
1337 | pub unsafe fn iter_hash(&self, hash: u64) -> RawIterHash<T> { |
1338 | RawIterHash::new(self, hash) |
1339 | } |
1340 | |
1341 | /// Returns an iterator which removes all elements from the table without |
1342 | /// freeing the memory. |
1343 | #[cfg_attr (feature = "inline-more" , inline)] |
1344 | pub fn drain(&mut self) -> RawDrain<'_, T, A> { |
1345 | unsafe { |
1346 | let iter = self.iter(); |
1347 | self.drain_iter_from(iter) |
1348 | } |
1349 | } |
1350 | |
1351 | /// Returns an iterator which removes all elements from the table without |
1352 | /// freeing the memory. |
1353 | /// |
1354 | /// Iteration starts at the provided iterator's current location. |
1355 | /// |
1356 | /// It is up to the caller to ensure that the iterator is valid for this |
1357 | /// `RawTable` and covers all items that remain in the table. |
1358 | #[cfg_attr (feature = "inline-more" , inline)] |
1359 | pub unsafe fn drain_iter_from(&mut self, iter: RawIter<T>) -> RawDrain<'_, T, A> { |
1360 | debug_assert_eq!(iter.len(), self.len()); |
1361 | RawDrain { |
1362 | iter, |
1363 | table: mem::replace(&mut self.table, RawTableInner::NEW), |
1364 | orig_table: NonNull::from(&mut self.table), |
1365 | marker: PhantomData, |
1366 | } |
1367 | } |
1368 | |
1369 | /// Returns an iterator which consumes all elements from the table. |
1370 | /// |
1371 | /// Iteration starts at the provided iterator's current location. |
1372 | /// |
1373 | /// It is up to the caller to ensure that the iterator is valid for this |
1374 | /// `RawTable` and covers all items that remain in the table. |
1375 | pub unsafe fn into_iter_from(self, iter: RawIter<T>) -> RawIntoIter<T, A> { |
1376 | debug_assert_eq!(iter.len(), self.len()); |
1377 | |
1378 | let allocation = self.into_allocation(); |
1379 | RawIntoIter { |
1380 | iter, |
1381 | allocation, |
1382 | marker: PhantomData, |
1383 | } |
1384 | } |
1385 | |
1386 | /// Converts the table into a raw allocation. The contents of the table |
1387 | /// should be dropped using a `RawIter` before freeing the allocation. |
1388 | #[cfg_attr (feature = "inline-more" , inline)] |
1389 | pub(crate) fn into_allocation(self) -> Option<(NonNull<u8>, Layout, A)> { |
1390 | let alloc = if self.table.is_empty_singleton() { |
1391 | None |
1392 | } else { |
1393 | // Avoid `Option::unwrap_or_else` because it bloats LLVM IR. |
1394 | let (layout, ctrl_offset) = |
1395 | match Self::TABLE_LAYOUT.calculate_layout_for(self.table.buckets()) { |
1396 | Some(lco) => lco, |
1397 | None => unsafe { hint::unreachable_unchecked() }, |
1398 | }; |
1399 | Some(( |
1400 | unsafe { NonNull::new_unchecked(self.table.ctrl.as_ptr().sub(ctrl_offset).cast()) }, |
1401 | layout, |
1402 | unsafe { ptr::read(&self.alloc) }, |
1403 | )) |
1404 | }; |
1405 | mem::forget(self); |
1406 | alloc |
1407 | } |
1408 | } |
1409 | |
1410 | unsafe impl<T, A: Allocator> Send for RawTable<T, A> |
1411 | where |
1412 | T: Send, |
1413 | A: Send, |
1414 | { |
1415 | } |
1416 | unsafe impl<T, A: Allocator> Sync for RawTable<T, A> |
1417 | where |
1418 | T: Sync, |
1419 | A: Sync, |
1420 | { |
1421 | } |
1422 | |
1423 | impl RawTableInner { |
1424 | const NEW: Self = RawTableInner::new(); |
1425 | |
1426 | /// Creates a new empty hash table without allocating any memory. |
1427 | /// |
1428 | /// In effect this returns a table with exactly 1 bucket. However we can |
1429 | /// leave the data pointer dangling since that bucket is never accessed |
1430 | /// due to our load factor forcing us to always have at least 1 free bucket. |
1431 | #[inline ] |
1432 | const fn new() -> Self { |
1433 | Self { |
1434 | // Be careful to cast the entire slice to a raw pointer. |
1435 | ctrl: unsafe { |
1436 | NonNull::new_unchecked(ptr:Group::static_empty().as_ptr().cast_mut().cast()) |
1437 | }, |
1438 | bucket_mask: 0, |
1439 | items: 0, |
1440 | growth_left: 0, |
1441 | } |
1442 | } |
1443 | } |
1444 | |
1445 | impl RawTableInner { |
1446 | /// Allocates a new [`RawTableInner`] with the given number of buckets. |
1447 | /// The control bytes and buckets are left uninitialized. |
1448 | /// |
1449 | /// # Safety |
1450 | /// |
1451 | /// The caller of this function must ensure that the `buckets` is power of two |
1452 | /// and also initialize all control bytes of the length `self.bucket_mask + 1 + |
1453 | /// Group::WIDTH` with the [`Tag::EMPTY`] bytes. |
1454 | /// |
1455 | /// See also [`Allocator`] API for other safety concerns. |
1456 | /// |
1457 | /// [`Allocator`]: https://doc.rust-lang.org/alloc/alloc/trait.Allocator.html |
1458 | #[cfg_attr (feature = "inline-more" , inline)] |
1459 | unsafe fn new_uninitialized<A>( |
1460 | alloc: &A, |
1461 | table_layout: TableLayout, |
1462 | buckets: usize, |
1463 | fallibility: Fallibility, |
1464 | ) -> Result<Self, TryReserveError> |
1465 | where |
1466 | A: Allocator, |
1467 | { |
1468 | debug_assert!(buckets.is_power_of_two()); |
1469 | |
1470 | // Avoid `Option::ok_or_else` because it bloats LLVM IR. |
1471 | let (layout, ctrl_offset) = match table_layout.calculate_layout_for(buckets) { |
1472 | Some(lco) => lco, |
1473 | None => return Err(fallibility.capacity_overflow()), |
1474 | }; |
1475 | |
1476 | let ptr: NonNull<u8> = match do_alloc(alloc, layout) { |
1477 | Ok(block) => block.cast(), |
1478 | Err(_) => return Err(fallibility.alloc_err(layout)), |
1479 | }; |
1480 | |
1481 | // SAFETY: null pointer will be caught in above check |
1482 | let ctrl = NonNull::new_unchecked(ptr.as_ptr().add(ctrl_offset)); |
1483 | Ok(Self { |
1484 | ctrl, |
1485 | bucket_mask: buckets - 1, |
1486 | items: 0, |
1487 | growth_left: bucket_mask_to_capacity(buckets - 1), |
1488 | }) |
1489 | } |
1490 | |
1491 | /// Attempts to allocate a new [`RawTableInner`] with at least enough |
1492 | /// capacity for inserting the given number of elements without reallocating. |
1493 | /// |
1494 | /// All the control bytes are initialized with the [`Tag::EMPTY`] bytes. |
1495 | #[inline ] |
1496 | fn fallible_with_capacity<A>( |
1497 | alloc: &A, |
1498 | table_layout: TableLayout, |
1499 | capacity: usize, |
1500 | fallibility: Fallibility, |
1501 | ) -> Result<Self, TryReserveError> |
1502 | where |
1503 | A: Allocator, |
1504 | { |
1505 | if capacity == 0 { |
1506 | Ok(Self::NEW) |
1507 | } else { |
1508 | // SAFETY: We checked that we could successfully allocate the new table, and then |
1509 | // initialized all control bytes with the constant `Tag::EMPTY` byte. |
1510 | unsafe { |
1511 | let buckets = capacity_to_buckets(capacity, table_layout) |
1512 | .ok_or_else(|| fallibility.capacity_overflow())?; |
1513 | |
1514 | let mut result = |
1515 | Self::new_uninitialized(alloc, table_layout, buckets, fallibility)?; |
1516 | // SAFETY: We checked that the table is allocated and therefore the table already has |
1517 | // `self.bucket_mask + 1 + Group::WIDTH` number of control bytes (see TableLayout::calculate_layout_for) |
1518 | // so writing `self.num_ctrl_bytes() == bucket_mask + 1 + Group::WIDTH` bytes is safe. |
1519 | result.ctrl_slice().fill_empty(); |
1520 | |
1521 | Ok(result) |
1522 | } |
1523 | } |
1524 | } |
1525 | |
1526 | /// Allocates a new [`RawTableInner`] with at least enough capacity for inserting |
1527 | /// the given number of elements without reallocating. |
1528 | /// |
1529 | /// Panics if the new capacity exceeds [`isize::MAX`] bytes and [`abort`] the program |
1530 | /// in case of allocation error. Use [`fallible_with_capacity`] instead if you want to |
1531 | /// handle memory allocation failure. |
1532 | /// |
1533 | /// All the control bytes are initialized with the [`Tag::EMPTY`] bytes. |
1534 | /// |
1535 | /// [`fallible_with_capacity`]: RawTableInner::fallible_with_capacity |
1536 | /// [`abort`]: https://doc.rust-lang.org/alloc/alloc/fn.handle_alloc_error.html |
1537 | fn with_capacity<A>(alloc: &A, table_layout: TableLayout, capacity: usize) -> Self |
1538 | where |
1539 | A: Allocator, |
1540 | { |
1541 | // Avoid `Result::unwrap_or_else` because it bloats LLVM IR. |
1542 | match Self::fallible_with_capacity(alloc, table_layout, capacity, Fallibility::Infallible) { |
1543 | Ok(table_inner) => table_inner, |
1544 | // SAFETY: All allocation errors will be caught inside `RawTableInner::new_uninitialized`. |
1545 | Err(_) => unsafe { hint::unreachable_unchecked() }, |
1546 | } |
1547 | } |
1548 | |
1549 | /// Fixes up an insertion slot returned by the [`RawTableInner::find_insert_slot_in_group`] method. |
1550 | /// |
1551 | /// In tables smaller than the group width (`self.buckets() < Group::WIDTH`), trailing control |
1552 | /// bytes outside the range of the table are filled with [`Tag::EMPTY`] entries. These will unfortunately |
1553 | /// trigger a match of [`RawTableInner::find_insert_slot_in_group`] function. This is because |
1554 | /// the `Some(bit)` returned by `group.match_empty_or_deleted().lowest_set_bit()` after masking |
1555 | /// (`(probe_seq.pos + bit) & self.bucket_mask`) may point to a full bucket that is already occupied. |
1556 | /// We detect this situation here and perform a second scan starting at the beginning of the table. |
1557 | /// This second scan is guaranteed to find an empty slot (due to the load factor) before hitting the |
1558 | /// trailing control bytes (containing [`Tag::EMPTY`] bytes). |
1559 | /// |
1560 | /// If this function is called correctly, it is guaranteed to return [`InsertSlot`] with an |
1561 | /// index of an empty or deleted bucket in the range `0..self.buckets()` (see `Warning` and |
1562 | /// `Safety`). |
1563 | /// |
1564 | /// # Warning |
1565 | /// |
1566 | /// The table must have at least 1 empty or deleted `bucket`, otherwise if the table is less than |
1567 | /// the group width (`self.buckets() < Group::WIDTH`) this function returns an index outside of the |
1568 | /// table indices range `0..self.buckets()` (`0..=self.bucket_mask`). Attempt to write data at that |
1569 | /// index will cause immediate [`undefined behavior`]. |
1570 | /// |
1571 | /// # Safety |
1572 | /// |
1573 | /// The safety rules are directly derived from the safety rules for [`RawTableInner::ctrl`] method. |
1574 | /// Thus, in order to uphold those safety contracts, as well as for the correct logic of the work |
1575 | /// of this crate, the following rules are necessary and sufficient: |
1576 | /// |
1577 | /// * The [`RawTableInner`] must have properly initialized control bytes otherwise calling this |
1578 | /// function results in [`undefined behavior`]. |
1579 | /// |
1580 | /// * This function must only be used on insertion slots found by [`RawTableInner::find_insert_slot_in_group`] |
1581 | /// (after the `find_insert_slot_in_group` function, but before insertion into the table). |
1582 | /// |
1583 | /// * The `index` must not be greater than the `self.bucket_mask`, i.e. `(index + 1) <= self.buckets()` |
1584 | /// (this one is provided by the [`RawTableInner::find_insert_slot_in_group`] function). |
1585 | /// |
1586 | /// Calling this function with an index not provided by [`RawTableInner::find_insert_slot_in_group`] |
1587 | /// may result in [`undefined behavior`] even if the index satisfies the safety rules of the |
1588 | /// [`RawTableInner::ctrl`] function (`index < self.bucket_mask + 1 + Group::WIDTH`). |
1589 | /// |
1590 | /// [`RawTableInner::ctrl`]: RawTableInner::ctrl |
1591 | /// [`RawTableInner::find_insert_slot_in_group`]: RawTableInner::find_insert_slot_in_group |
1592 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
1593 | #[inline ] |
1594 | unsafe fn fix_insert_slot(&self, mut index: usize) -> InsertSlot { |
1595 | // SAFETY: The caller of this function ensures that `index` is in the range `0..=self.bucket_mask`. |
1596 | if unlikely(self.is_bucket_full(index)) { |
1597 | debug_assert!(self.bucket_mask < Group::WIDTH); |
1598 | // SAFETY: |
1599 | // |
1600 | // * Since the caller of this function ensures that the control bytes are properly |
1601 | // initialized and `ptr = self.ctrl(0)` points to the start of the array of control |
1602 | // bytes, therefore: `ctrl` is valid for reads, properly aligned to `Group::WIDTH` |
1603 | // and points to the properly initialized control bytes (see also |
1604 | // `TableLayout::calculate_layout_for` and `ptr::read`); |
1605 | // |
1606 | // * Because the caller of this function ensures that the index was provided by the |
1607 | // `self.find_insert_slot_in_group()` function, so for for tables larger than the |
1608 | // group width (self.buckets() >= Group::WIDTH), we will never end up in the given |
1609 | // branch, since `(probe_seq.pos + bit) & self.bucket_mask` in `find_insert_slot_in_group` |
1610 | // cannot return a full bucket index. For tables smaller than the group width, calling |
1611 | // the `unwrap_unchecked` function is also safe, as the trailing control bytes outside |
1612 | // the range of the table are filled with EMPTY bytes (and we know for sure that there |
1613 | // is at least one FULL bucket), so this second scan either finds an empty slot (due to |
1614 | // the load factor) or hits the trailing control bytes (containing EMPTY). |
1615 | index = Group::load_aligned(self.ctrl(0)) |
1616 | .match_empty_or_deleted() |
1617 | .lowest_set_bit() |
1618 | .unwrap_unchecked(); |
1619 | } |
1620 | InsertSlot { index } |
1621 | } |
1622 | |
1623 | /// Finds the position to insert something in a group. |
1624 | /// |
1625 | /// **This may have false positives and must be fixed up with `fix_insert_slot` |
1626 | /// before it's used.** |
1627 | /// |
1628 | /// The function is guaranteed to return the index of an empty or deleted [`Bucket`] |
1629 | /// in the range `0..self.buckets()` (`0..=self.bucket_mask`). |
1630 | #[inline ] |
1631 | fn find_insert_slot_in_group(&self, group: &Group, probe_seq: &ProbeSeq) -> Option<usize> { |
1632 | let bit = group.match_empty_or_deleted().lowest_set_bit(); |
1633 | |
1634 | if likely(bit.is_some()) { |
1635 | // This is the same as `(probe_seq.pos + bit) % self.buckets()` because the number |
1636 | // of buckets is a power of two, and `self.bucket_mask = self.buckets() - 1`. |
1637 | Some((probe_seq.pos + bit.unwrap()) & self.bucket_mask) |
1638 | } else { |
1639 | None |
1640 | } |
1641 | } |
1642 | |
1643 | /// Searches for an element in the table, or a potential slot where that element could |
1644 | /// be inserted (an empty or deleted [`Bucket`] index). |
1645 | /// |
1646 | /// This uses dynamic dispatch to reduce the amount of code generated, but that is |
1647 | /// eliminated by LLVM optimizations. |
1648 | /// |
1649 | /// This function does not make any changes to the `data` part of the table, or any |
1650 | /// changes to the `items` or `growth_left` field of the table. |
1651 | /// |
1652 | /// The table must have at least 1 empty or deleted `bucket`, otherwise, if the |
1653 | /// `eq: &mut dyn FnMut(usize) -> bool` function does not return `true`, this function |
1654 | /// will never return (will go into an infinite loop) for tables larger than the group |
1655 | /// width, or return an index outside of the table indices range if the table is less |
1656 | /// than the group width. |
1657 | /// |
1658 | /// This function is guaranteed to provide the `eq: &mut dyn FnMut(usize) -> bool` |
1659 | /// function with only `FULL` buckets' indices and return the `index` of the found |
1660 | /// element (as `Ok(index)`). If the element is not found and there is at least 1 |
1661 | /// empty or deleted [`Bucket`] in the table, the function is guaranteed to return |
1662 | /// [`InsertSlot`] with an index in the range `0..self.buckets()`, but in any case, |
1663 | /// if this function returns [`InsertSlot`], it will contain an index in the range |
1664 | /// `0..=self.buckets()`. |
1665 | /// |
1666 | /// # Safety |
1667 | /// |
1668 | /// The [`RawTableInner`] must have properly initialized control bytes otherwise calling |
1669 | /// this function results in [`undefined behavior`]. |
1670 | /// |
1671 | /// Attempt to write data at the [`InsertSlot`] returned by this function when the table is |
1672 | /// less than the group width and if there was not at least one empty or deleted bucket in |
1673 | /// the table will cause immediate [`undefined behavior`]. This is because in this case the |
1674 | /// function will return `self.bucket_mask + 1` as an index due to the trailing [`Tag::EMPTY`] |
1675 | /// control bytes outside the table range. |
1676 | /// |
1677 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
1678 | #[inline ] |
1679 | unsafe fn find_or_find_insert_slot_inner( |
1680 | &self, |
1681 | hash: u64, |
1682 | eq: &mut dyn FnMut(usize) -> bool, |
1683 | ) -> Result<usize, InsertSlot> { |
1684 | let mut insert_slot = None; |
1685 | |
1686 | let tag_hash = Tag::full(hash); |
1687 | let mut probe_seq = self.probe_seq(hash); |
1688 | |
1689 | loop { |
1690 | // SAFETY: |
1691 | // * Caller of this function ensures that the control bytes are properly initialized. |
1692 | // |
1693 | // * `ProbeSeq.pos` cannot be greater than `self.bucket_mask = self.buckets() - 1` |
1694 | // of the table due to masking with `self.bucket_mask` and also because the number |
1695 | // of buckets is a power of two (see `self.probe_seq` function). |
1696 | // |
1697 | // * Even if `ProbeSeq.pos` returns `position == self.bucket_mask`, it is safe to |
1698 | // call `Group::load` due to the extended control bytes range, which is |
1699 | // `self.bucket_mask + 1 + Group::WIDTH` (in fact, this means that the last control |
1700 | // byte will never be read for the allocated table); |
1701 | // |
1702 | // * Also, even if `RawTableInner` is not already allocated, `ProbeSeq.pos` will |
1703 | // always return "0" (zero), so Group::load will read unaligned `Group::static_empty()` |
1704 | // bytes, which is safe (see RawTableInner::new). |
1705 | let group = unsafe { Group::load(self.ctrl(probe_seq.pos)) }; |
1706 | |
1707 | for bit in group.match_tag(tag_hash) { |
1708 | let index = (probe_seq.pos + bit) & self.bucket_mask; |
1709 | |
1710 | if likely(eq(index)) { |
1711 | return Ok(index); |
1712 | } |
1713 | } |
1714 | |
1715 | // We didn't find the element we were looking for in the group, try to get an |
1716 | // insertion slot from the group if we don't have one yet. |
1717 | if likely(insert_slot.is_none()) { |
1718 | insert_slot = self.find_insert_slot_in_group(&group, &probe_seq); |
1719 | } |
1720 | |
1721 | if let Some(insert_slot) = insert_slot { |
1722 | // Only stop the search if the group contains at least one empty element. |
1723 | // Otherwise, the element that we are looking for might be in a following group. |
1724 | if likely(group.match_empty().any_bit_set()) { |
1725 | // We must have found a insert slot by now, since the current group contains at |
1726 | // least one. For tables smaller than the group width, there will still be an |
1727 | // empty element in the current (and only) group due to the load factor. |
1728 | unsafe { |
1729 | // SAFETY: |
1730 | // * Caller of this function ensures that the control bytes are properly initialized. |
1731 | // |
1732 | // * We use this function with the slot / index found by `self.find_insert_slot_in_group` |
1733 | return Err(self.fix_insert_slot(insert_slot)); |
1734 | } |
1735 | } |
1736 | } |
1737 | |
1738 | probe_seq.move_next(self.bucket_mask); |
1739 | } |
1740 | } |
1741 | |
1742 | /// Searches for an empty or deleted bucket which is suitable for inserting a new |
1743 | /// element and sets the hash for that slot. Returns an index of that slot and the |
1744 | /// old control byte stored in the found index. |
1745 | /// |
1746 | /// This function does not check if the given element exists in the table. Also, |
1747 | /// this function does not check if there is enough space in the table to insert |
1748 | /// a new element. The caller of the function must make sure that the table has at |
1749 | /// least 1 empty or deleted `bucket`, otherwise this function will never return |
1750 | /// (will go into an infinite loop) for tables larger than the group width, or |
1751 | /// return an index outside of the table indices range if the table is less than |
1752 | /// the group width. |
1753 | /// |
1754 | /// If there is at least 1 empty or deleted `bucket` in the table, the function is |
1755 | /// guaranteed to return an `index` in the range `0..self.buckets()`, but in any case, |
1756 | /// if this function returns an `index` it will be in the range `0..=self.buckets()`. |
1757 | /// |
1758 | /// This function does not make any changes to the `data` parts of the table, |
1759 | /// or any changes to the `items` or `growth_left` field of the table. |
1760 | /// |
1761 | /// # Safety |
1762 | /// |
1763 | /// The safety rules are directly derived from the safety rules for the |
1764 | /// [`RawTableInner::set_ctrl_hash`] and [`RawTableInner::find_insert_slot`] methods. |
1765 | /// Thus, in order to uphold the safety contracts for that methods, as well as for |
1766 | /// the correct logic of the work of this crate, you must observe the following rules |
1767 | /// when calling this function: |
1768 | /// |
1769 | /// * The [`RawTableInner`] has already been allocated and has properly initialized |
1770 | /// control bytes otherwise calling this function results in [`undefined behavior`]. |
1771 | /// |
1772 | /// * The caller of this function must ensure that the "data" parts of the table |
1773 | /// will have an entry in the returned index (matching the given hash) right |
1774 | /// after calling this function. |
1775 | /// |
1776 | /// Attempt to write data at the `index` returned by this function when the table is |
1777 | /// less than the group width and if there was not at least one empty or deleted bucket in |
1778 | /// the table will cause immediate [`undefined behavior`]. This is because in this case the |
1779 | /// function will return `self.bucket_mask + 1` as an index due to the trailing [`Tag::EMPTY`] |
1780 | /// control bytes outside the table range. |
1781 | /// |
1782 | /// The caller must independently increase the `items` field of the table, and also, |
1783 | /// if the old control byte was [`Tag::EMPTY`], then decrease the table's `growth_left` |
1784 | /// field, and do not change it if the old control byte was [`Tag::DELETED`]. |
1785 | /// |
1786 | /// See also [`Bucket::as_ptr`] method, for more information about of properly removing |
1787 | /// or saving `element` from / into the [`RawTable`] / [`RawTableInner`]. |
1788 | /// |
1789 | /// [`Bucket::as_ptr`]: Bucket::as_ptr |
1790 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
1791 | /// [`RawTableInner::ctrl`]: RawTableInner::ctrl |
1792 | /// [`RawTableInner::set_ctrl_hash`]: RawTableInner::set_ctrl_hash |
1793 | /// [`RawTableInner::find_insert_slot`]: RawTableInner::find_insert_slot |
1794 | #[inline ] |
1795 | unsafe fn prepare_insert_slot(&mut self, hash: u64) -> (usize, Tag) { |
1796 | // SAFETY: Caller of this function ensures that the control bytes are properly initialized. |
1797 | let index: usize = self.find_insert_slot(hash).index; |
1798 | // SAFETY: |
1799 | // 1. The `find_insert_slot` function either returns an `index` less than or |
1800 | // equal to `self.buckets() = self.bucket_mask + 1` of the table, or never |
1801 | // returns if it cannot find an empty or deleted slot. |
1802 | // 2. The caller of this function guarantees that the table has already been |
1803 | // allocated |
1804 | let old_ctrl = *self.ctrl(index); |
1805 | self.set_ctrl_hash(index, hash); |
1806 | (index, old_ctrl) |
1807 | } |
1808 | |
1809 | /// Searches for an empty or deleted bucket which is suitable for inserting |
1810 | /// a new element, returning the `index` for the new [`Bucket`]. |
1811 | /// |
1812 | /// This function does not make any changes to the `data` part of the table, or any |
1813 | /// changes to the `items` or `growth_left` field of the table. |
1814 | /// |
1815 | /// The table must have at least 1 empty or deleted `bucket`, otherwise this function |
1816 | /// will never return (will go into an infinite loop) for tables larger than the group |
1817 | /// width, or return an index outside of the table indices range if the table is less |
1818 | /// than the group width. |
1819 | /// |
1820 | /// If there is at least 1 empty or deleted `bucket` in the table, the function is |
1821 | /// guaranteed to return [`InsertSlot`] with an index in the range `0..self.buckets()`, |
1822 | /// but in any case, if this function returns [`InsertSlot`], it will contain an index |
1823 | /// in the range `0..=self.buckets()`. |
1824 | /// |
1825 | /// # Safety |
1826 | /// |
1827 | /// The [`RawTableInner`] must have properly initialized control bytes otherwise calling |
1828 | /// this function results in [`undefined behavior`]. |
1829 | /// |
1830 | /// Attempt to write data at the [`InsertSlot`] returned by this function when the table is |
1831 | /// less than the group width and if there was not at least one empty or deleted bucket in |
1832 | /// the table will cause immediate [`undefined behavior`]. This is because in this case the |
1833 | /// function will return `self.bucket_mask + 1` as an index due to the trailing [`Tag::EMPTY`] |
1834 | /// control bytes outside the table range. |
1835 | /// |
1836 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
1837 | #[inline ] |
1838 | unsafe fn find_insert_slot(&self, hash: u64) -> InsertSlot { |
1839 | let mut probe_seq = self.probe_seq(hash); |
1840 | loop { |
1841 | // SAFETY: |
1842 | // * Caller of this function ensures that the control bytes are properly initialized. |
1843 | // |
1844 | // * `ProbeSeq.pos` cannot be greater than `self.bucket_mask = self.buckets() - 1` |
1845 | // of the table due to masking with `self.bucket_mask` and also because the number |
1846 | // of buckets is a power of two (see `self.probe_seq` function). |
1847 | // |
1848 | // * Even if `ProbeSeq.pos` returns `position == self.bucket_mask`, it is safe to |
1849 | // call `Group::load` due to the extended control bytes range, which is |
1850 | // `self.bucket_mask + 1 + Group::WIDTH` (in fact, this means that the last control |
1851 | // byte will never be read for the allocated table); |
1852 | // |
1853 | // * Also, even if `RawTableInner` is not already allocated, `ProbeSeq.pos` will |
1854 | // always return "0" (zero), so Group::load will read unaligned `Group::static_empty()` |
1855 | // bytes, which is safe (see RawTableInner::new). |
1856 | let group = unsafe { Group::load(self.ctrl(probe_seq.pos)) }; |
1857 | |
1858 | let index = self.find_insert_slot_in_group(&group, &probe_seq); |
1859 | if likely(index.is_some()) { |
1860 | // SAFETY: |
1861 | // * Caller of this function ensures that the control bytes are properly initialized. |
1862 | // |
1863 | // * We use this function with the slot / index found by `self.find_insert_slot_in_group` |
1864 | unsafe { |
1865 | return self.fix_insert_slot(index.unwrap_unchecked()); |
1866 | } |
1867 | } |
1868 | probe_seq.move_next(self.bucket_mask); |
1869 | } |
1870 | } |
1871 | |
1872 | /// Searches for an element in a table, returning the `index` of the found element. |
1873 | /// This uses dynamic dispatch to reduce the amount of code generated, but it is |
1874 | /// eliminated by LLVM optimizations. |
1875 | /// |
1876 | /// This function does not make any changes to the `data` part of the table, or any |
1877 | /// changes to the `items` or `growth_left` field of the table. |
1878 | /// |
1879 | /// The table must have at least 1 empty `bucket`, otherwise, if the |
1880 | /// `eq: &mut dyn FnMut(usize) -> bool` function does not return `true`, |
1881 | /// this function will also never return (will go into an infinite loop). |
1882 | /// |
1883 | /// This function is guaranteed to provide the `eq: &mut dyn FnMut(usize) -> bool` |
1884 | /// function with only `FULL` buckets' indices and return the `index` of the found |
1885 | /// element as `Some(index)`, so the index will always be in the range |
1886 | /// `0..self.buckets()`. |
1887 | /// |
1888 | /// # Safety |
1889 | /// |
1890 | /// The [`RawTableInner`] must have properly initialized control bytes otherwise calling |
1891 | /// this function results in [`undefined behavior`]. |
1892 | /// |
1893 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
1894 | #[inline (always)] |
1895 | unsafe fn find_inner(&self, hash: u64, eq: &mut dyn FnMut(usize) -> bool) -> Option<usize> { |
1896 | let tag_hash = Tag::full(hash); |
1897 | let mut probe_seq = self.probe_seq(hash); |
1898 | |
1899 | loop { |
1900 | // SAFETY: |
1901 | // * Caller of this function ensures that the control bytes are properly initialized. |
1902 | // |
1903 | // * `ProbeSeq.pos` cannot be greater than `self.bucket_mask = self.buckets() - 1` |
1904 | // of the table due to masking with `self.bucket_mask`. |
1905 | // |
1906 | // * Even if `ProbeSeq.pos` returns `position == self.bucket_mask`, it is safe to |
1907 | // call `Group::load` due to the extended control bytes range, which is |
1908 | // `self.bucket_mask + 1 + Group::WIDTH` (in fact, this means that the last control |
1909 | // byte will never be read for the allocated table); |
1910 | // |
1911 | // * Also, even if `RawTableInner` is not already allocated, `ProbeSeq.pos` will |
1912 | // always return "0" (zero), so Group::load will read unaligned `Group::static_empty()` |
1913 | // bytes, which is safe (see RawTableInner::new_in). |
1914 | let group = unsafe { Group::load(self.ctrl(probe_seq.pos)) }; |
1915 | |
1916 | for bit in group.match_tag(tag_hash) { |
1917 | // This is the same as `(probe_seq.pos + bit) % self.buckets()` because the number |
1918 | // of buckets is a power of two, and `self.bucket_mask = self.buckets() - 1`. |
1919 | let index = (probe_seq.pos + bit) & self.bucket_mask; |
1920 | |
1921 | if likely(eq(index)) { |
1922 | return Some(index); |
1923 | } |
1924 | } |
1925 | |
1926 | if likely(group.match_empty().any_bit_set()) { |
1927 | return None; |
1928 | } |
1929 | |
1930 | probe_seq.move_next(self.bucket_mask); |
1931 | } |
1932 | } |
1933 | |
1934 | /// Prepares for rehashing data in place (that is, without allocating new memory). |
1935 | /// Converts all full index `control bytes` to `Tag::DELETED` and all `Tag::DELETED` control |
1936 | /// bytes to `Tag::EMPTY`, i.e. performs the following conversion: |
1937 | /// |
1938 | /// - `Tag::EMPTY` control bytes -> `Tag::EMPTY`; |
1939 | /// - `Tag::DELETED` control bytes -> `Tag::EMPTY`; |
1940 | /// - `FULL` control bytes -> `Tag::DELETED`. |
1941 | /// |
1942 | /// This function does not make any changes to the `data` parts of the table, |
1943 | /// or any changes to the `items` or `growth_left` field of the table. |
1944 | /// |
1945 | /// # Safety |
1946 | /// |
1947 | /// You must observe the following safety rules when calling this function: |
1948 | /// |
1949 | /// * The [`RawTableInner`] has already been allocated; |
1950 | /// |
1951 | /// * The caller of this function must convert the `Tag::DELETED` bytes back to `FULL` |
1952 | /// bytes when re-inserting them into their ideal position (which was impossible |
1953 | /// to do during the first insert due to tombstones). If the caller does not do |
1954 | /// this, then calling this function may result in a memory leak. |
1955 | /// |
1956 | /// * The [`RawTableInner`] must have properly initialized control bytes otherwise |
1957 | /// calling this function results in [`undefined behavior`]. |
1958 | /// |
1959 | /// Calling this function on a table that has not been allocated results in |
1960 | /// [`undefined behavior`]. |
1961 | /// |
1962 | /// See also [`Bucket::as_ptr`] method, for more information about of properly removing |
1963 | /// or saving `data element` from / into the [`RawTable`] / [`RawTableInner`]. |
1964 | /// |
1965 | /// [`Bucket::as_ptr`]: Bucket::as_ptr |
1966 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
1967 | #[allow (clippy::mut_mut)] |
1968 | #[inline ] |
1969 | unsafe fn prepare_rehash_in_place(&mut self) { |
1970 | // Bulk convert all full control bytes to DELETED, and all DELETED control bytes to EMPTY. |
1971 | // This effectively frees up all buckets containing a DELETED entry. |
1972 | // |
1973 | // SAFETY: |
1974 | // 1. `i` is guaranteed to be within bounds since we are iterating from zero to `buckets - 1`; |
1975 | // 2. Even if `i` will be `i == self.bucket_mask`, it is safe to call `Group::load_aligned` |
1976 | // due to the extended control bytes range, which is `self.bucket_mask + 1 + Group::WIDTH`; |
1977 | // 3. The caller of this function guarantees that [`RawTableInner`] has already been allocated; |
1978 | // 4. We can use `Group::load_aligned` and `Group::store_aligned` here since we start from 0 |
1979 | // and go to the end with a step equal to `Group::WIDTH` (see TableLayout::calculate_layout_for). |
1980 | for i in (0..self.buckets()).step_by(Group::WIDTH) { |
1981 | let group = Group::load_aligned(self.ctrl(i)); |
1982 | let group = group.convert_special_to_empty_and_full_to_deleted(); |
1983 | group.store_aligned(self.ctrl(i)); |
1984 | } |
1985 | |
1986 | // Fix up the trailing control bytes. See the comments in set_ctrl |
1987 | // for the handling of tables smaller than the group width. |
1988 | // |
1989 | // SAFETY: The caller of this function guarantees that [`RawTableInner`] |
1990 | // has already been allocated |
1991 | if unlikely(self.buckets() < Group::WIDTH) { |
1992 | // SAFETY: We have `self.bucket_mask + 1 + Group::WIDTH` number of control bytes, |
1993 | // so copying `self.buckets() == self.bucket_mask + 1` bytes with offset equal to |
1994 | // `Group::WIDTH` is safe |
1995 | self.ctrl(0) |
1996 | .copy_to(self.ctrl(Group::WIDTH), self.buckets()); |
1997 | } else { |
1998 | // SAFETY: We have `self.bucket_mask + 1 + Group::WIDTH` number of |
1999 | // control bytes,so copying `Group::WIDTH` bytes with offset equal |
2000 | // to `self.buckets() == self.bucket_mask + 1` is safe |
2001 | self.ctrl(0) |
2002 | .copy_to(self.ctrl(self.buckets()), Group::WIDTH); |
2003 | } |
2004 | } |
2005 | |
2006 | /// Returns an iterator over every element in the table. |
2007 | /// |
2008 | /// # Safety |
2009 | /// |
2010 | /// If any of the following conditions are violated, the result |
2011 | /// is [`undefined behavior`]: |
2012 | /// |
2013 | /// * The caller has to ensure that the `RawTableInner` outlives the |
2014 | /// `RawIter`. Because we cannot make the `next` method unsafe on |
2015 | /// the `RawIter` struct, we have to make the `iter` method unsafe. |
2016 | /// |
2017 | /// * The [`RawTableInner`] must have properly initialized control bytes. |
2018 | /// |
2019 | /// The type `T` must be the actual type of the elements stored in the table, |
2020 | /// otherwise using the returned [`RawIter`] results in [`undefined behavior`]. |
2021 | /// |
2022 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
2023 | #[inline ] |
2024 | unsafe fn iter<T>(&self) -> RawIter<T> { |
2025 | // SAFETY: |
2026 | // 1. Since the caller of this function ensures that the control bytes |
2027 | // are properly initialized and `self.data_end()` points to the start |
2028 | // of the array of control bytes, therefore: `ctrl` is valid for reads, |
2029 | // properly aligned to `Group::WIDTH` and points to the properly initialized |
2030 | // control bytes. |
2031 | // 2. `data` bucket index in the table is equal to the `ctrl` index (i.e. |
2032 | // equal to zero). |
2033 | // 3. We pass the exact value of buckets of the table to the function. |
2034 | // |
2035 | // `ctrl` points here (to the start |
2036 | // of the first control byte `CT0`) |
2037 | // ∨ |
2038 | // [Pad], T_n, ..., T1, T0, |CT0, CT1, ..., CT_n|, CTa_0, CTa_1, ..., CTa_m |
2039 | // \________ ________/ |
2040 | // \/ |
2041 | // `n = buckets - 1`, i.e. `RawTableInner::buckets() - 1` |
2042 | // |
2043 | // where: T0...T_n - our stored data; |
2044 | // CT0...CT_n - control bytes or metadata for `data`. |
2045 | // CTa_0...CTa_m - additional control bytes, where `m = Group::WIDTH - 1` (so that the search |
2046 | // with loading `Group` bytes from the heap works properly, even if the result |
2047 | // of `h1(hash) & self.bucket_mask` is equal to `self.bucket_mask`). See also |
2048 | // `RawTableInner::set_ctrl` function. |
2049 | // |
2050 | // P.S. `h1(hash) & self.bucket_mask` is the same as `hash as usize % self.buckets()` because the number |
2051 | // of buckets is a power of two, and `self.bucket_mask = self.buckets() - 1`. |
2052 | let data = Bucket::from_base_index(self.data_end(), 0); |
2053 | RawIter { |
2054 | // SAFETY: See explanation above |
2055 | iter: RawIterRange::new(self.ctrl.as_ptr(), data, self.buckets()), |
2056 | items: self.items, |
2057 | } |
2058 | } |
2059 | |
2060 | /// Executes the destructors (if any) of the values stored in the table. |
2061 | /// |
2062 | /// # Note |
2063 | /// |
2064 | /// This function does not erase the control bytes of the table and does |
2065 | /// not make any changes to the `items` or `growth_left` fields of the |
2066 | /// table. If necessary, the caller of this function must manually set |
2067 | /// up these table fields, for example using the [`clear_no_drop`] function. |
2068 | /// |
2069 | /// Be careful during calling this function, because drop function of |
2070 | /// the elements can panic, and this can leave table in an inconsistent |
2071 | /// state. |
2072 | /// |
2073 | /// # Safety |
2074 | /// |
2075 | /// The type `T` must be the actual type of the elements stored in the table, |
2076 | /// otherwise calling this function may result in [`undefined behavior`]. |
2077 | /// |
2078 | /// If `T` is a type that should be dropped and **the table is not empty**, |
2079 | /// calling this function more than once results in [`undefined behavior`]. |
2080 | /// |
2081 | /// If `T` is not [`Copy`], attempting to use values stored in the table after |
2082 | /// calling this function may result in [`undefined behavior`]. |
2083 | /// |
2084 | /// It is safe to call this function on a table that has not been allocated, |
2085 | /// on a table with uninitialized control bytes, and on a table with no actual |
2086 | /// data but with `Full` control bytes if `self.items == 0`. |
2087 | /// |
2088 | /// See also [`Bucket::drop`] / [`Bucket::as_ptr`] methods, for more information |
2089 | /// about of properly removing or saving `element` from / into the [`RawTable`] / |
2090 | /// [`RawTableInner`]. |
2091 | /// |
2092 | /// [`Bucket::drop`]: Bucket::drop |
2093 | /// [`Bucket::as_ptr`]: Bucket::as_ptr |
2094 | /// [`clear_no_drop`]: RawTableInner::clear_no_drop |
2095 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
2096 | unsafe fn drop_elements<T>(&mut self) { |
2097 | // Check that `self.items != 0`. Protects against the possibility |
2098 | // of creating an iterator on an table with uninitialized control bytes. |
2099 | if T::NEEDS_DROP && self.items != 0 { |
2100 | // SAFETY: We know for sure that RawTableInner will outlive the |
2101 | // returned `RawIter` iterator, and the caller of this function |
2102 | // must uphold the safety contract for `drop_elements` method. |
2103 | for item in self.iter::<T>() { |
2104 | // SAFETY: The caller must uphold the safety contract for |
2105 | // `drop_elements` method. |
2106 | item.drop(); |
2107 | } |
2108 | } |
2109 | } |
2110 | |
2111 | /// Executes the destructors (if any) of the values stored in the table and than |
2112 | /// deallocates the table. |
2113 | /// |
2114 | /// # Note |
2115 | /// |
2116 | /// Calling this function automatically makes invalid (dangling) all instances of |
2117 | /// buckets ([`Bucket`]) and makes invalid (dangling) the `ctrl` field of the table. |
2118 | /// |
2119 | /// This function does not make any changes to the `bucket_mask`, `items` or `growth_left` |
2120 | /// fields of the table. If necessary, the caller of this function must manually set |
2121 | /// up these table fields. |
2122 | /// |
2123 | /// # Safety |
2124 | /// |
2125 | /// If any of the following conditions are violated, the result is [`undefined behavior`]: |
2126 | /// |
2127 | /// * Calling this function more than once; |
2128 | /// |
2129 | /// * The type `T` must be the actual type of the elements stored in the table. |
2130 | /// |
2131 | /// * The `alloc` must be the same [`Allocator`] as the `Allocator` that was used |
2132 | /// to allocate this table. |
2133 | /// |
2134 | /// * The `table_layout` must be the same [`TableLayout`] as the `TableLayout` that |
2135 | /// was used to allocate this table. |
2136 | /// |
2137 | /// The caller of this function should pay attention to the possibility of the |
2138 | /// elements' drop function panicking, because this: |
2139 | /// |
2140 | /// * May leave the table in an inconsistent state; |
2141 | /// |
2142 | /// * Memory is never deallocated, so a memory leak may occur. |
2143 | /// |
2144 | /// Attempt to use the `ctrl` field of the table (dereference) after calling this |
2145 | /// function results in [`undefined behavior`]. |
2146 | /// |
2147 | /// It is safe to call this function on a table that has not been allocated, |
2148 | /// on a table with uninitialized control bytes, and on a table with no actual |
2149 | /// data but with `Full` control bytes if `self.items == 0`. |
2150 | /// |
2151 | /// See also [`RawTableInner::drop_elements`] or [`RawTableInner::free_buckets`] |
2152 | /// for more information. |
2153 | /// |
2154 | /// [`RawTableInner::drop_elements`]: RawTableInner::drop_elements |
2155 | /// [`RawTableInner::free_buckets`]: RawTableInner::free_buckets |
2156 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
2157 | unsafe fn drop_inner_table<T, A: Allocator>(&mut self, alloc: &A, table_layout: TableLayout) { |
2158 | if !self.is_empty_singleton() { |
2159 | unsafe { |
2160 | // SAFETY: The caller must uphold the safety contract for `drop_inner_table` method. |
2161 | self.drop_elements::<T>(); |
2162 | // SAFETY: |
2163 | // 1. We have checked that our table is allocated. |
2164 | // 2. The caller must uphold the safety contract for `drop_inner_table` method. |
2165 | self.free_buckets(alloc, table_layout); |
2166 | } |
2167 | } |
2168 | } |
2169 | |
2170 | /// Returns a pointer to an element in the table (convenience for |
2171 | /// `Bucket::from_base_index(self.data_end::<T>(), index)`). |
2172 | /// |
2173 | /// The caller must ensure that the `RawTableInner` outlives the returned [`Bucket<T>`], |
2174 | /// otherwise using it may result in [`undefined behavior`]. |
2175 | /// |
2176 | /// # Safety |
2177 | /// |
2178 | /// If `mem::size_of::<T>() != 0`, then the safety rules are directly derived from the |
2179 | /// safety rules of the [`Bucket::from_base_index`] function. Therefore, when calling |
2180 | /// this function, the following safety rules must be observed: |
2181 | /// |
2182 | /// * The table must already be allocated; |
2183 | /// |
2184 | /// * The `index` must not be greater than the number returned by the [`RawTableInner::buckets`] |
2185 | /// function, i.e. `(index + 1) <= self.buckets()`. |
2186 | /// |
2187 | /// * The type `T` must be the actual type of the elements stored in the table, otherwise |
2188 | /// using the returned [`Bucket`] may result in [`undefined behavior`]. |
2189 | /// |
2190 | /// It is safe to call this function with index of zero (`index == 0`) on a table that has |
2191 | /// not been allocated, but using the returned [`Bucket`] results in [`undefined behavior`]. |
2192 | /// |
2193 | /// If `mem::size_of::<T>() == 0`, then the only requirement is that the `index` must |
2194 | /// not be greater than the number returned by the [`RawTable::buckets`] function, i.e. |
2195 | /// `(index + 1) <= self.buckets()`. |
2196 | /// |
2197 | /// ```none |
2198 | /// If mem::size_of::<T>() != 0 then return a pointer to the `element` in the `data part` of the table |
2199 | /// (we start counting from "0", so that in the expression T[n], the "n" index actually one less than |
2200 | /// the "buckets" number of our `RawTableInner`, i.e. "n = RawTableInner::buckets() - 1"): |
2201 | /// |
2202 | /// `table.bucket(3).as_ptr()` returns a pointer that points here in the `data` |
2203 | /// part of the `RawTableInner`, i.e. to the start of T3 (see [`Bucket::as_ptr`]) |
2204 | /// | |
2205 | /// | `base = table.data_end::<T>()` points here |
2206 | /// | (to the start of CT0 or to the end of T0) |
2207 | /// v v |
2208 | /// [Pad], T_n, ..., |T3|, T2, T1, T0, |CT0, CT1, CT2, CT3, ..., CT_n, CTa_0, CTa_1, ..., CTa_m |
2209 | /// ^ \__________ __________/ |
2210 | /// `table.bucket(3)` returns a pointer that points \/ |
2211 | /// here in the `data` part of the `RawTableInner` additional control bytes |
2212 | /// (to the end of T3) `m = Group::WIDTH - 1` |
2213 | /// |
2214 | /// where: T0...T_n - our stored data; |
2215 | /// CT0...CT_n - control bytes or metadata for `data`; |
2216 | /// CTa_0...CTa_m - additional control bytes (so that the search with loading `Group` bytes from |
2217 | /// the heap works properly, even if the result of `h1(hash) & self.bucket_mask` |
2218 | /// is equal to `self.bucket_mask`). See also `RawTableInner::set_ctrl` function. |
2219 | /// |
2220 | /// P.S. `h1(hash) & self.bucket_mask` is the same as `hash as usize % self.buckets()` because the number |
2221 | /// of buckets is a power of two, and `self.bucket_mask = self.buckets() - 1`. |
2222 | /// ``` |
2223 | /// |
2224 | /// [`Bucket::from_base_index`]: Bucket::from_base_index |
2225 | /// [`RawTableInner::buckets`]: RawTableInner::buckets |
2226 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
2227 | #[inline ] |
2228 | unsafe fn bucket<T>(&self, index: usize) -> Bucket<T> { |
2229 | debug_assert_ne!(self.bucket_mask, 0); |
2230 | debug_assert!(index < self.buckets()); |
2231 | Bucket::from_base_index(self.data_end(), index) |
2232 | } |
2233 | |
2234 | /// Returns a raw `*mut u8` pointer to the start of the `data` element in the table |
2235 | /// (convenience for `self.data_end::<u8>().as_ptr().sub((index + 1) * size_of)`). |
2236 | /// |
2237 | /// The caller must ensure that the `RawTableInner` outlives the returned `*mut u8`, |
2238 | /// otherwise using it may result in [`undefined behavior`]. |
2239 | /// |
2240 | /// # Safety |
2241 | /// |
2242 | /// If any of the following conditions are violated, the result is [`undefined behavior`]: |
2243 | /// |
2244 | /// * The table must already be allocated; |
2245 | /// |
2246 | /// * The `index` must not be greater than the number returned by the [`RawTableInner::buckets`] |
2247 | /// function, i.e. `(index + 1) <= self.buckets()`; |
2248 | /// |
2249 | /// * The `size_of` must be equal to the size of the elements stored in the table; |
2250 | /// |
2251 | /// ```none |
2252 | /// If mem::size_of::<T>() != 0 then return a pointer to the `element` in the `data part` of the table |
2253 | /// (we start counting from "0", so that in the expression T[n], the "n" index actually one less than |
2254 | /// the "buckets" number of our `RawTableInner`, i.e. "n = RawTableInner::buckets() - 1"): |
2255 | /// |
2256 | /// `table.bucket_ptr(3, mem::size_of::<T>())` returns a pointer that points here in the |
2257 | /// `data` part of the `RawTableInner`, i.e. to the start of T3 |
2258 | /// | |
2259 | /// | `base = table.data_end::<u8>()` points here |
2260 | /// | (to the start of CT0 or to the end of T0) |
2261 | /// v v |
2262 | /// [Pad], T_n, ..., |T3|, T2, T1, T0, |CT0, CT1, CT2, CT3, ..., CT_n, CTa_0, CTa_1, ..., CTa_m |
2263 | /// \__________ __________/ |
2264 | /// \/ |
2265 | /// additional control bytes |
2266 | /// `m = Group::WIDTH - 1` |
2267 | /// |
2268 | /// where: T0...T_n - our stored data; |
2269 | /// CT0...CT_n - control bytes or metadata for `data`; |
2270 | /// CTa_0...CTa_m - additional control bytes (so that the search with loading `Group` bytes from |
2271 | /// the heap works properly, even if the result of `h1(hash) & self.bucket_mask` |
2272 | /// is equal to `self.bucket_mask`). See also `RawTableInner::set_ctrl` function. |
2273 | /// |
2274 | /// P.S. `h1(hash) & self.bucket_mask` is the same as `hash as usize % self.buckets()` because the number |
2275 | /// of buckets is a power of two, and `self.bucket_mask = self.buckets() - 1`. |
2276 | /// ``` |
2277 | /// |
2278 | /// [`RawTableInner::buckets`]: RawTableInner::buckets |
2279 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
2280 | #[inline ] |
2281 | unsafe fn bucket_ptr(&self, index: usize, size_of: usize) -> *mut u8 { |
2282 | debug_assert_ne!(self.bucket_mask, 0); |
2283 | debug_assert!(index < self.buckets()); |
2284 | let base: *mut u8 = self.data_end().as_ptr(); |
2285 | base.sub((index + 1) * size_of) |
2286 | } |
2287 | |
2288 | /// Returns pointer to one past last `data` element in the table as viewed from |
2289 | /// the start point of the allocation (convenience for `self.ctrl.cast()`). |
2290 | /// |
2291 | /// This function actually returns a pointer to the end of the `data element` at |
2292 | /// index "0" (zero). |
2293 | /// |
2294 | /// The caller must ensure that the `RawTableInner` outlives the returned [`NonNull<T>`], |
2295 | /// otherwise using it may result in [`undefined behavior`]. |
2296 | /// |
2297 | /// # Note |
2298 | /// |
2299 | /// The type `T` must be the actual type of the elements stored in the table, otherwise |
2300 | /// using the returned [`NonNull<T>`] may result in [`undefined behavior`]. |
2301 | /// |
2302 | /// ```none |
2303 | /// `table.data_end::<T>()` returns pointer that points here |
2304 | /// (to the end of `T0`) |
2305 | /// ∨ |
2306 | /// [Pad], T_n, ..., T1, T0, |CT0, CT1, ..., CT_n|, CTa_0, CTa_1, ..., CTa_m |
2307 | /// \________ ________/ |
2308 | /// \/ |
2309 | /// `n = buckets - 1`, i.e. `RawTableInner::buckets() - 1` |
2310 | /// |
2311 | /// where: T0...T_n - our stored data; |
2312 | /// CT0...CT_n - control bytes or metadata for `data`. |
2313 | /// CTa_0...CTa_m - additional control bytes, where `m = Group::WIDTH - 1` (so that the search |
2314 | /// with loading `Group` bytes from the heap works properly, even if the result |
2315 | /// of `h1(hash) & self.bucket_mask` is equal to `self.bucket_mask`). See also |
2316 | /// `RawTableInner::set_ctrl` function. |
2317 | /// |
2318 | /// P.S. `h1(hash) & self.bucket_mask` is the same as `hash as usize % self.buckets()` because the number |
2319 | /// of buckets is a power of two, and `self.bucket_mask = self.buckets() - 1`. |
2320 | /// ``` |
2321 | /// |
2322 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
2323 | #[inline ] |
2324 | fn data_end<T>(&self) -> NonNull<T> { |
2325 | self.ctrl.cast() |
2326 | } |
2327 | |
2328 | /// Returns an iterator-like object for a probe sequence on the table. |
2329 | /// |
2330 | /// This iterator never terminates, but is guaranteed to visit each bucket |
2331 | /// group exactly once. The loop using `probe_seq` must terminate upon |
2332 | /// reaching a group containing an empty bucket. |
2333 | #[inline ] |
2334 | fn probe_seq(&self, hash: u64) -> ProbeSeq { |
2335 | ProbeSeq { |
2336 | // This is the same as `hash as usize % self.buckets()` because the number |
2337 | // of buckets is a power of two, and `self.bucket_mask = self.buckets() - 1`. |
2338 | pos: h1(hash) & self.bucket_mask, |
2339 | stride: 0, |
2340 | } |
2341 | } |
2342 | |
2343 | #[inline ] |
2344 | unsafe fn record_item_insert_at(&mut self, index: usize, old_ctrl: Tag, hash: u64) { |
2345 | self.growth_left -= usize::from(old_ctrl.special_is_empty()); |
2346 | self.set_ctrl_hash(index, hash); |
2347 | self.items += 1; |
2348 | } |
2349 | |
2350 | #[inline ] |
2351 | fn is_in_same_group(&self, i: usize, new_i: usize, hash: u64) -> bool { |
2352 | let probe_seq_pos = self.probe_seq(hash).pos; |
2353 | let probe_index = |
2354 | |pos: usize| (pos.wrapping_sub(probe_seq_pos) & self.bucket_mask) / Group::WIDTH; |
2355 | probe_index(i) == probe_index(new_i) |
2356 | } |
2357 | |
2358 | /// Sets a control byte to the hash, and possibly also the replicated control byte at |
2359 | /// the end of the array. |
2360 | /// |
2361 | /// This function does not make any changes to the `data` parts of the table, |
2362 | /// or any changes to the `items` or `growth_left` field of the table. |
2363 | /// |
2364 | /// # Safety |
2365 | /// |
2366 | /// The safety rules are directly derived from the safety rules for [`RawTableInner::set_ctrl`] |
2367 | /// method. Thus, in order to uphold the safety contracts for the method, you must observe the |
2368 | /// following rules when calling this function: |
2369 | /// |
2370 | /// * The [`RawTableInner`] has already been allocated; |
2371 | /// |
2372 | /// * The `index` must not be greater than the `RawTableInner.bucket_mask`, i.e. |
2373 | /// `index <= RawTableInner.bucket_mask` or, in other words, `(index + 1)` must |
2374 | /// be no greater than the number returned by the function [`RawTableInner::buckets`]. |
2375 | /// |
2376 | /// Calling this function on a table that has not been allocated results in [`undefined behavior`]. |
2377 | /// |
2378 | /// See also [`Bucket::as_ptr`] method, for more information about of properly removing |
2379 | /// or saving `data element` from / into the [`RawTable`] / [`RawTableInner`]. |
2380 | /// |
2381 | /// [`RawTableInner::set_ctrl`]: RawTableInner::set_ctrl |
2382 | /// [`RawTableInner::buckets`]: RawTableInner::buckets |
2383 | /// [`Bucket::as_ptr`]: Bucket::as_ptr |
2384 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
2385 | #[inline ] |
2386 | unsafe fn set_ctrl_hash(&mut self, index: usize, hash: u64) { |
2387 | // SAFETY: The caller must uphold the safety rules for the [`RawTableInner::set_ctrl_hash`] |
2388 | self.set_ctrl(index, Tag::full(hash)); |
2389 | } |
2390 | |
2391 | /// Replaces the hash in the control byte at the given index with the provided one, |
2392 | /// and possibly also replicates the new control byte at the end of the array of control |
2393 | /// bytes, returning the old control byte. |
2394 | /// |
2395 | /// This function does not make any changes to the `data` parts of the table, |
2396 | /// or any changes to the `items` or `growth_left` field of the table. |
2397 | /// |
2398 | /// # Safety |
2399 | /// |
2400 | /// The safety rules are directly derived from the safety rules for [`RawTableInner::set_ctrl_hash`] |
2401 | /// and [`RawTableInner::ctrl`] methods. Thus, in order to uphold the safety contracts for both |
2402 | /// methods, you must observe the following rules when calling this function: |
2403 | /// |
2404 | /// * The [`RawTableInner`] has already been allocated; |
2405 | /// |
2406 | /// * The `index` must not be greater than the `RawTableInner.bucket_mask`, i.e. |
2407 | /// `index <= RawTableInner.bucket_mask` or, in other words, `(index + 1)` must |
2408 | /// be no greater than the number returned by the function [`RawTableInner::buckets`]. |
2409 | /// |
2410 | /// Calling this function on a table that has not been allocated results in [`undefined behavior`]. |
2411 | /// |
2412 | /// See also [`Bucket::as_ptr`] method, for more information about of properly removing |
2413 | /// or saving `data element` from / into the [`RawTable`] / [`RawTableInner`]. |
2414 | /// |
2415 | /// [`RawTableInner::set_ctrl_hash`]: RawTableInner::set_ctrl_hash |
2416 | /// [`RawTableInner::buckets`]: RawTableInner::buckets |
2417 | /// [`Bucket::as_ptr`]: Bucket::as_ptr |
2418 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
2419 | #[inline ] |
2420 | unsafe fn replace_ctrl_hash(&mut self, index: usize, hash: u64) -> Tag { |
2421 | // SAFETY: The caller must uphold the safety rules for the [`RawTableInner::replace_ctrl_hash`] |
2422 | let prev_ctrl = *self.ctrl(index); |
2423 | self.set_ctrl_hash(index, hash); |
2424 | prev_ctrl |
2425 | } |
2426 | |
2427 | /// Sets a control byte, and possibly also the replicated control byte at |
2428 | /// the end of the array. |
2429 | /// |
2430 | /// This function does not make any changes to the `data` parts of the table, |
2431 | /// or any changes to the `items` or `growth_left` field of the table. |
2432 | /// |
2433 | /// # Safety |
2434 | /// |
2435 | /// You must observe the following safety rules when calling this function: |
2436 | /// |
2437 | /// * The [`RawTableInner`] has already been allocated; |
2438 | /// |
2439 | /// * The `index` must not be greater than the `RawTableInner.bucket_mask`, i.e. |
2440 | /// `index <= RawTableInner.bucket_mask` or, in other words, `(index + 1)` must |
2441 | /// be no greater than the number returned by the function [`RawTableInner::buckets`]. |
2442 | /// |
2443 | /// Calling this function on a table that has not been allocated results in [`undefined behavior`]. |
2444 | /// |
2445 | /// See also [`Bucket::as_ptr`] method, for more information about of properly removing |
2446 | /// or saving `data element` from / into the [`RawTable`] / [`RawTableInner`]. |
2447 | /// |
2448 | /// [`RawTableInner::buckets`]: RawTableInner::buckets |
2449 | /// [`Bucket::as_ptr`]: Bucket::as_ptr |
2450 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
2451 | #[inline ] |
2452 | unsafe fn set_ctrl(&mut self, index: usize, ctrl: Tag) { |
2453 | // Replicate the first Group::WIDTH control bytes at the end of |
2454 | // the array without using a branch. If the tables smaller than |
2455 | // the group width (self.buckets() < Group::WIDTH), |
2456 | // `index2 = Group::WIDTH + index`, otherwise `index2` is: |
2457 | // |
2458 | // - If index >= Group::WIDTH then index == index2. |
2459 | // - Otherwise index2 == self.bucket_mask + 1 + index. |
2460 | // |
2461 | // The very last replicated control byte is never actually read because |
2462 | // we mask the initial index for unaligned loads, but we write it |
2463 | // anyways because it makes the set_ctrl implementation simpler. |
2464 | // |
2465 | // If there are fewer buckets than Group::WIDTH then this code will |
2466 | // replicate the buckets at the end of the trailing group. For example |
2467 | // with 2 buckets and a group size of 4, the control bytes will look |
2468 | // like this: |
2469 | // |
2470 | // Real | Replicated |
2471 | // --------------------------------------------- |
2472 | // | [A] | [B] | [Tag::EMPTY] | [EMPTY] | [A] | [B] | |
2473 | // --------------------------------------------- |
2474 | |
2475 | // This is the same as `(index.wrapping_sub(Group::WIDTH)) % self.buckets() + Group::WIDTH` |
2476 | // because the number of buckets is a power of two, and `self.bucket_mask = self.buckets() - 1`. |
2477 | let index2 = ((index.wrapping_sub(Group::WIDTH)) & self.bucket_mask) + Group::WIDTH; |
2478 | |
2479 | // SAFETY: The caller must uphold the safety rules for the [`RawTableInner::set_ctrl`] |
2480 | *self.ctrl(index) = ctrl; |
2481 | *self.ctrl(index2) = ctrl; |
2482 | } |
2483 | |
2484 | /// Returns a pointer to a control byte. |
2485 | /// |
2486 | /// # Safety |
2487 | /// |
2488 | /// For the allocated [`RawTableInner`], the result is [`Undefined Behavior`], |
2489 | /// if the `index` is greater than the `self.bucket_mask + 1 + Group::WIDTH`. |
2490 | /// In that case, calling this function with `index == self.bucket_mask + 1 + Group::WIDTH` |
2491 | /// will return a pointer to the end of the allocated table and it is useless on its own. |
2492 | /// |
2493 | /// Calling this function with `index >= self.bucket_mask + 1 + Group::WIDTH` on a |
2494 | /// table that has not been allocated results in [`Undefined Behavior`]. |
2495 | /// |
2496 | /// So to satisfy both requirements you should always follow the rule that |
2497 | /// `index < self.bucket_mask + 1 + Group::WIDTH` |
2498 | /// |
2499 | /// Calling this function on [`RawTableInner`] that are not already allocated is safe |
2500 | /// for read-only purpose. |
2501 | /// |
2502 | /// See also [`Bucket::as_ptr()`] method, for more information about of properly removing |
2503 | /// or saving `data element` from / into the [`RawTable`] / [`RawTableInner`]. |
2504 | /// |
2505 | /// [`Bucket::as_ptr()`]: Bucket::as_ptr() |
2506 | /// [`Undefined Behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
2507 | #[inline ] |
2508 | unsafe fn ctrl(&self, index: usize) -> *mut Tag { |
2509 | debug_assert!(index < self.num_ctrl_bytes()); |
2510 | // SAFETY: The caller must uphold the safety rules for the [`RawTableInner::ctrl`] |
2511 | self.ctrl.as_ptr().add(index).cast() |
2512 | } |
2513 | |
2514 | /// Gets the slice of all control bytes. |
2515 | fn ctrl_slice(&mut self) -> &mut [Tag] { |
2516 | // SAFETY: We've intiailized all control bytes, and have the correct number. |
2517 | unsafe { slice::from_raw_parts_mut(self.ctrl.as_ptr().cast(), self.num_ctrl_bytes()) } |
2518 | } |
2519 | |
2520 | #[inline ] |
2521 | fn buckets(&self) -> usize { |
2522 | self.bucket_mask + 1 |
2523 | } |
2524 | |
2525 | /// Checks whether the bucket at `index` is full. |
2526 | /// |
2527 | /// # Safety |
2528 | /// |
2529 | /// The caller must ensure `index` is less than the number of buckets. |
2530 | #[inline ] |
2531 | unsafe fn is_bucket_full(&self, index: usize) -> bool { |
2532 | debug_assert!(index < self.buckets()); |
2533 | (*self.ctrl(index)).is_full() |
2534 | } |
2535 | |
2536 | #[inline ] |
2537 | fn num_ctrl_bytes(&self) -> usize { |
2538 | self.bucket_mask + 1 + Group::WIDTH |
2539 | } |
2540 | |
2541 | #[inline ] |
2542 | fn is_empty_singleton(&self) -> bool { |
2543 | self.bucket_mask == 0 |
2544 | } |
2545 | |
2546 | /// Attempts to allocate a new hash table with at least enough capacity |
2547 | /// for inserting the given number of elements without reallocating, |
2548 | /// and return it inside `ScopeGuard` to protect against panic in the hash |
2549 | /// function. |
2550 | /// |
2551 | /// # Note |
2552 | /// |
2553 | /// It is recommended (but not required): |
2554 | /// |
2555 | /// * That the new table's `capacity` be greater than or equal to `self.items`. |
2556 | /// |
2557 | /// * The `alloc` is the same [`Allocator`] as the `Allocator` used |
2558 | /// to allocate this table. |
2559 | /// |
2560 | /// * The `table_layout` is the same [`TableLayout`] as the `TableLayout` used |
2561 | /// to allocate this table. |
2562 | /// |
2563 | /// If `table_layout` does not match the `TableLayout` that was used to allocate |
2564 | /// this table, then using `mem::swap` with the `self` and the new table returned |
2565 | /// by this function results in [`undefined behavior`]. |
2566 | /// |
2567 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
2568 | #[allow (clippy::mut_mut)] |
2569 | #[inline ] |
2570 | fn prepare_resize<'a, A>( |
2571 | &self, |
2572 | alloc: &'a A, |
2573 | table_layout: TableLayout, |
2574 | capacity: usize, |
2575 | fallibility: Fallibility, |
2576 | ) -> Result<crate::scopeguard::ScopeGuard<Self, impl FnMut(&mut Self) + 'a>, TryReserveError> |
2577 | where |
2578 | A: Allocator, |
2579 | { |
2580 | debug_assert!(self.items <= capacity); |
2581 | |
2582 | // Allocate and initialize the new table. |
2583 | let new_table = |
2584 | RawTableInner::fallible_with_capacity(alloc, table_layout, capacity, fallibility)?; |
2585 | |
2586 | // The hash function may panic, in which case we simply free the new |
2587 | // table without dropping any elements that may have been copied into |
2588 | // it. |
2589 | // |
2590 | // This guard is also used to free the old table on success, see |
2591 | // the comment at the bottom of this function. |
2592 | Ok(guard(new_table, move |self_| { |
2593 | if !self_.is_empty_singleton() { |
2594 | // SAFETY: |
2595 | // 1. We have checked that our table is allocated. |
2596 | // 2. We know for sure that the `alloc` and `table_layout` matches the |
2597 | // [`Allocator`] and [`TableLayout`] used to allocate this table. |
2598 | unsafe { self_.free_buckets(alloc, table_layout) }; |
2599 | } |
2600 | })) |
2601 | } |
2602 | |
2603 | /// Reserves or rehashes to make room for `additional` more elements. |
2604 | /// |
2605 | /// This uses dynamic dispatch to reduce the amount of |
2606 | /// code generated, but it is eliminated by LLVM optimizations when inlined. |
2607 | /// |
2608 | /// # Safety |
2609 | /// |
2610 | /// If any of the following conditions are violated, the result is |
2611 | /// [`undefined behavior`]: |
2612 | /// |
2613 | /// * The `alloc` must be the same [`Allocator`] as the `Allocator` used |
2614 | /// to allocate this table. |
2615 | /// |
2616 | /// * The `layout` must be the same [`TableLayout`] as the `TableLayout` |
2617 | /// used to allocate this table. |
2618 | /// |
2619 | /// * The `drop` function (`fn(*mut u8)`) must be the actual drop function of |
2620 | /// the elements stored in the table. |
2621 | /// |
2622 | /// * The [`RawTableInner`] must have properly initialized control bytes. |
2623 | /// |
2624 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
2625 | #[allow (clippy::inline_always)] |
2626 | #[inline (always)] |
2627 | unsafe fn reserve_rehash_inner<A>( |
2628 | &mut self, |
2629 | alloc: &A, |
2630 | additional: usize, |
2631 | hasher: &dyn Fn(&mut Self, usize) -> u64, |
2632 | fallibility: Fallibility, |
2633 | layout: TableLayout, |
2634 | drop: Option<unsafe fn(*mut u8)>, |
2635 | ) -> Result<(), TryReserveError> |
2636 | where |
2637 | A: Allocator, |
2638 | { |
2639 | // Avoid `Option::ok_or_else` because it bloats LLVM IR. |
2640 | let new_items = match self.items.checked_add(additional) { |
2641 | Some(new_items) => new_items, |
2642 | None => return Err(fallibility.capacity_overflow()), |
2643 | }; |
2644 | let full_capacity = bucket_mask_to_capacity(self.bucket_mask); |
2645 | if new_items <= full_capacity / 2 { |
2646 | // Rehash in-place without re-allocating if we have plenty of spare |
2647 | // capacity that is locked up due to DELETED entries. |
2648 | |
2649 | // SAFETY: |
2650 | // 1. We know for sure that `[`RawTableInner`]` has already been allocated |
2651 | // (since new_items <= full_capacity / 2); |
2652 | // 2. The caller ensures that `drop` function is the actual drop function of |
2653 | // the elements stored in the table. |
2654 | // 3. The caller ensures that `layout` matches the [`TableLayout`] that was |
2655 | // used to allocate this table. |
2656 | // 4. The caller ensures that the control bytes of the `RawTableInner` |
2657 | // are already initialized. |
2658 | self.rehash_in_place(hasher, layout.size, drop); |
2659 | Ok(()) |
2660 | } else { |
2661 | // Otherwise, conservatively resize to at least the next size up |
2662 | // to avoid churning deletes into frequent rehashes. |
2663 | // |
2664 | // SAFETY: |
2665 | // 1. We know for sure that `capacity >= self.items`. |
2666 | // 2. The caller ensures that `alloc` and `layout` matches the [`Allocator`] and |
2667 | // [`TableLayout`] that were used to allocate this table. |
2668 | // 3. The caller ensures that the control bytes of the `RawTableInner` |
2669 | // are already initialized. |
2670 | self.resize_inner( |
2671 | alloc, |
2672 | usize::max(new_items, full_capacity + 1), |
2673 | hasher, |
2674 | fallibility, |
2675 | layout, |
2676 | ) |
2677 | } |
2678 | } |
2679 | |
2680 | /// Returns an iterator over full buckets indices in the table. |
2681 | /// |
2682 | /// # Safety |
2683 | /// |
2684 | /// Behavior is undefined if any of the following conditions are violated: |
2685 | /// |
2686 | /// * The caller has to ensure that the `RawTableInner` outlives the |
2687 | /// `FullBucketsIndices`. Because we cannot make the `next` method |
2688 | /// unsafe on the `FullBucketsIndices` struct, we have to make the |
2689 | /// `full_buckets_indices` method unsafe. |
2690 | /// |
2691 | /// * The [`RawTableInner`] must have properly initialized control bytes. |
2692 | #[inline (always)] |
2693 | unsafe fn full_buckets_indices(&self) -> FullBucketsIndices { |
2694 | // SAFETY: |
2695 | // 1. Since the caller of this function ensures that the control bytes |
2696 | // are properly initialized and `self.ctrl(0)` points to the start |
2697 | // of the array of control bytes, therefore: `ctrl` is valid for reads, |
2698 | // properly aligned to `Group::WIDTH` and points to the properly initialized |
2699 | // control bytes. |
2700 | // 2. The value of `items` is equal to the amount of data (values) added |
2701 | // to the table. |
2702 | // |
2703 | // `ctrl` points here (to the start |
2704 | // of the first control byte `CT0`) |
2705 | // ∨ |
2706 | // [Pad], T_n, ..., T1, T0, |CT0, CT1, ..., CT_n|, Group::WIDTH |
2707 | // \________ ________/ |
2708 | // \/ |
2709 | // `n = buckets - 1`, i.e. `RawTableInner::buckets() - 1` |
2710 | // |
2711 | // where: T0...T_n - our stored data; |
2712 | // CT0...CT_n - control bytes or metadata for `data`. |
2713 | let ctrl = NonNull::new_unchecked(self.ctrl(0).cast::<u8>()); |
2714 | |
2715 | FullBucketsIndices { |
2716 | // Load the first group |
2717 | // SAFETY: See explanation above. |
2718 | current_group: Group::load_aligned(ctrl.as_ptr().cast()) |
2719 | .match_full() |
2720 | .into_iter(), |
2721 | group_first_index: 0, |
2722 | ctrl, |
2723 | items: self.items, |
2724 | } |
2725 | } |
2726 | |
2727 | /// Allocates a new table of a different size and moves the contents of the |
2728 | /// current table into it. |
2729 | /// |
2730 | /// This uses dynamic dispatch to reduce the amount of |
2731 | /// code generated, but it is eliminated by LLVM optimizations when inlined. |
2732 | /// |
2733 | /// # Safety |
2734 | /// |
2735 | /// If any of the following conditions are violated, the result is |
2736 | /// [`undefined behavior`]: |
2737 | /// |
2738 | /// * The `alloc` must be the same [`Allocator`] as the `Allocator` used |
2739 | /// to allocate this table; |
2740 | /// |
2741 | /// * The `layout` must be the same [`TableLayout`] as the `TableLayout` |
2742 | /// used to allocate this table; |
2743 | /// |
2744 | /// * The [`RawTableInner`] must have properly initialized control bytes. |
2745 | /// |
2746 | /// The caller of this function must ensure that `capacity >= self.items` |
2747 | /// otherwise: |
2748 | /// |
2749 | /// * If `self.items != 0`, calling of this function with `capacity == 0` |
2750 | /// results in [`undefined behavior`]. |
2751 | /// |
2752 | /// * If `capacity_to_buckets(capacity) < Group::WIDTH` and |
2753 | /// `self.items > capacity_to_buckets(capacity)` calling this function |
2754 | /// results in [`undefined behavior`]. |
2755 | /// |
2756 | /// * If `capacity_to_buckets(capacity) >= Group::WIDTH` and |
2757 | /// `self.items > capacity_to_buckets(capacity)` calling this function |
2758 | /// are never return (will go into an infinite loop). |
2759 | /// |
2760 | /// Note: It is recommended (but not required) that the new table's `capacity` |
2761 | /// be greater than or equal to `self.items`. In case if `capacity <= self.items` |
2762 | /// this function can never return. See [`RawTableInner::find_insert_slot`] for |
2763 | /// more information. |
2764 | /// |
2765 | /// [`RawTableInner::find_insert_slot`]: RawTableInner::find_insert_slot |
2766 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
2767 | #[allow (clippy::inline_always)] |
2768 | #[inline (always)] |
2769 | unsafe fn resize_inner<A>( |
2770 | &mut self, |
2771 | alloc: &A, |
2772 | capacity: usize, |
2773 | hasher: &dyn Fn(&mut Self, usize) -> u64, |
2774 | fallibility: Fallibility, |
2775 | layout: TableLayout, |
2776 | ) -> Result<(), TryReserveError> |
2777 | where |
2778 | A: Allocator, |
2779 | { |
2780 | // SAFETY: We know for sure that `alloc` and `layout` matches the [`Allocator`] and [`TableLayout`] |
2781 | // that were used to allocate this table. |
2782 | let mut new_table = self.prepare_resize(alloc, layout, capacity, fallibility)?; |
2783 | |
2784 | // SAFETY: We know for sure that RawTableInner will outlive the |
2785 | // returned `FullBucketsIndices` iterator, and the caller of this |
2786 | // function ensures that the control bytes are properly initialized. |
2787 | for full_byte_index in self.full_buckets_indices() { |
2788 | // This may panic. |
2789 | let hash = hasher(self, full_byte_index); |
2790 | |
2791 | // SAFETY: |
2792 | // We can use a simpler version of insert() here since: |
2793 | // 1. There are no DELETED entries. |
2794 | // 2. We know there is enough space in the table. |
2795 | // 3. All elements are unique. |
2796 | // 4. The caller of this function guarantees that `capacity > 0` |
2797 | // so `new_table` must already have some allocated memory. |
2798 | // 5. We set `growth_left` and `items` fields of the new table |
2799 | // after the loop. |
2800 | // 6. We insert into the table, at the returned index, the data |
2801 | // matching the given hash immediately after calling this function. |
2802 | let (new_index, _) = new_table.prepare_insert_slot(hash); |
2803 | |
2804 | // SAFETY: |
2805 | // |
2806 | // * `src` is valid for reads of `layout.size` bytes, since the |
2807 | // table is alive and the `full_byte_index` is guaranteed to be |
2808 | // within bounds (see `FullBucketsIndices::next_impl`); |
2809 | // |
2810 | // * `dst` is valid for writes of `layout.size` bytes, since the |
2811 | // caller ensures that `table_layout` matches the [`TableLayout`] |
2812 | // that was used to allocate old table and we have the `new_index` |
2813 | // returned by `prepare_insert_slot`. |
2814 | // |
2815 | // * Both `src` and `dst` are properly aligned. |
2816 | // |
2817 | // * Both `src` and `dst` point to different region of memory. |
2818 | ptr::copy_nonoverlapping( |
2819 | self.bucket_ptr(full_byte_index, layout.size), |
2820 | new_table.bucket_ptr(new_index, layout.size), |
2821 | layout.size, |
2822 | ); |
2823 | } |
2824 | |
2825 | // The hash function didn't panic, so we can safely set the |
2826 | // `growth_left` and `items` fields of the new table. |
2827 | new_table.growth_left -= self.items; |
2828 | new_table.items = self.items; |
2829 | |
2830 | // We successfully copied all elements without panicking. Now replace |
2831 | // self with the new table. The old table will have its memory freed but |
2832 | // the items will not be dropped (since they have been moved into the |
2833 | // new table). |
2834 | // SAFETY: The caller ensures that `table_layout` matches the [`TableLayout`] |
2835 | // that was used to allocate this table. |
2836 | mem::swap(self, &mut new_table); |
2837 | |
2838 | Ok(()) |
2839 | } |
2840 | |
2841 | /// Rehashes the contents of the table in place (i.e. without changing the |
2842 | /// allocation). |
2843 | /// |
2844 | /// If `hasher` panics then some the table's contents may be lost. |
2845 | /// |
2846 | /// This uses dynamic dispatch to reduce the amount of |
2847 | /// code generated, but it is eliminated by LLVM optimizations when inlined. |
2848 | /// |
2849 | /// # Safety |
2850 | /// |
2851 | /// If any of the following conditions are violated, the result is [`undefined behavior`]: |
2852 | /// |
2853 | /// * The `size_of` must be equal to the size of the elements stored in the table; |
2854 | /// |
2855 | /// * The `drop` function (`fn(*mut u8)`) must be the actual drop function of |
2856 | /// the elements stored in the table. |
2857 | /// |
2858 | /// * The [`RawTableInner`] has already been allocated; |
2859 | /// |
2860 | /// * The [`RawTableInner`] must have properly initialized control bytes. |
2861 | /// |
2862 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
2863 | #[allow (clippy::inline_always)] |
2864 | #[cfg_attr (feature = "inline-more" , inline(always))] |
2865 | #[cfg_attr (not(feature = "inline-more" ), inline)] |
2866 | unsafe fn rehash_in_place( |
2867 | &mut self, |
2868 | hasher: &dyn Fn(&mut Self, usize) -> u64, |
2869 | size_of: usize, |
2870 | drop: Option<unsafe fn(*mut u8)>, |
2871 | ) { |
2872 | // If the hash function panics then properly clean up any elements |
2873 | // that we haven't rehashed yet. We unfortunately can't preserve the |
2874 | // element since we lost their hash and have no way of recovering it |
2875 | // without risking another panic. |
2876 | self.prepare_rehash_in_place(); |
2877 | |
2878 | let mut guard = guard(self, move |self_| { |
2879 | if let Some(drop) = drop { |
2880 | for i in 0..self_.buckets() { |
2881 | if *self_.ctrl(i) == Tag::DELETED { |
2882 | self_.set_ctrl(i, Tag::EMPTY); |
2883 | drop(self_.bucket_ptr(i, size_of)); |
2884 | self_.items -= 1; |
2885 | } |
2886 | } |
2887 | } |
2888 | self_.growth_left = bucket_mask_to_capacity(self_.bucket_mask) - self_.items; |
2889 | }); |
2890 | |
2891 | // At this point, DELETED elements are elements that we haven't |
2892 | // rehashed yet. Find them and re-insert them at their ideal |
2893 | // position. |
2894 | 'outer: for i in 0..guard.buckets() { |
2895 | if *guard.ctrl(i) != Tag::DELETED { |
2896 | continue; |
2897 | } |
2898 | |
2899 | let i_p = guard.bucket_ptr(i, size_of); |
2900 | |
2901 | 'inner: loop { |
2902 | // Hash the current item |
2903 | let hash = hasher(*guard, i); |
2904 | |
2905 | // Search for a suitable place to put it |
2906 | // |
2907 | // SAFETY: Caller of this function ensures that the control bytes |
2908 | // are properly initialized. |
2909 | let new_i = guard.find_insert_slot(hash).index; |
2910 | |
2911 | // Probing works by scanning through all of the control |
2912 | // bytes in groups, which may not be aligned to the group |
2913 | // size. If both the new and old position fall within the |
2914 | // same unaligned group, then there is no benefit in moving |
2915 | // it and we can just continue to the next item. |
2916 | if likely(guard.is_in_same_group(i, new_i, hash)) { |
2917 | guard.set_ctrl_hash(i, hash); |
2918 | continue 'outer; |
2919 | } |
2920 | |
2921 | let new_i_p = guard.bucket_ptr(new_i, size_of); |
2922 | |
2923 | // We are moving the current item to a new position. Write |
2924 | // our H2 to the control byte of the new position. |
2925 | let prev_ctrl = guard.replace_ctrl_hash(new_i, hash); |
2926 | if prev_ctrl == Tag::EMPTY { |
2927 | guard.set_ctrl(i, Tag::EMPTY); |
2928 | // If the target slot is empty, simply move the current |
2929 | // element into the new slot and clear the old control |
2930 | // byte. |
2931 | ptr::copy_nonoverlapping(i_p, new_i_p, size_of); |
2932 | continue 'outer; |
2933 | } else { |
2934 | // If the target slot is occupied, swap the two elements |
2935 | // and then continue processing the element that we just |
2936 | // swapped into the old slot. |
2937 | debug_assert_eq!(prev_ctrl, Tag::DELETED); |
2938 | ptr::swap_nonoverlapping(i_p, new_i_p, size_of); |
2939 | continue 'inner; |
2940 | } |
2941 | } |
2942 | } |
2943 | |
2944 | guard.growth_left = bucket_mask_to_capacity(guard.bucket_mask) - guard.items; |
2945 | |
2946 | mem::forget(guard); |
2947 | } |
2948 | |
2949 | /// Deallocates the table without dropping any entries. |
2950 | /// |
2951 | /// # Note |
2952 | /// |
2953 | /// This function must be called only after [`drop_elements`](RawTableInner::drop_elements), |
2954 | /// else it can lead to leaking of memory. Also calling this function automatically |
2955 | /// makes invalid (dangling) all instances of buckets ([`Bucket`]) and makes invalid |
2956 | /// (dangling) the `ctrl` field of the table. |
2957 | /// |
2958 | /// # Safety |
2959 | /// |
2960 | /// If any of the following conditions are violated, the result is [`Undefined Behavior`]: |
2961 | /// |
2962 | /// * The [`RawTableInner`] has already been allocated; |
2963 | /// |
2964 | /// * The `alloc` must be the same [`Allocator`] as the `Allocator` that was used |
2965 | /// to allocate this table. |
2966 | /// |
2967 | /// * The `table_layout` must be the same [`TableLayout`] as the `TableLayout` that was used |
2968 | /// to allocate this table. |
2969 | /// |
2970 | /// See also [`GlobalAlloc::dealloc`] or [`Allocator::deallocate`] for more information. |
2971 | /// |
2972 | /// [`Undefined Behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
2973 | /// [`GlobalAlloc::dealloc`]: https://doc.rust-lang.org/alloc/alloc/trait.GlobalAlloc.html#tymethod.dealloc |
2974 | /// [`Allocator::deallocate`]: https://doc.rust-lang.org/alloc/alloc/trait.Allocator.html#tymethod.deallocate |
2975 | #[inline ] |
2976 | unsafe fn free_buckets<A>(&mut self, alloc: &A, table_layout: TableLayout) |
2977 | where |
2978 | A: Allocator, |
2979 | { |
2980 | // SAFETY: The caller must uphold the safety contract for `free_buckets` |
2981 | // method. |
2982 | let (ptr, layout) = self.allocation_info(table_layout); |
2983 | alloc.deallocate(ptr, layout); |
2984 | } |
2985 | |
2986 | /// Returns a pointer to the allocated memory and the layout that was used to |
2987 | /// allocate the table. |
2988 | /// |
2989 | /// # Safety |
2990 | /// |
2991 | /// Caller of this function must observe the following safety rules: |
2992 | /// |
2993 | /// * The [`RawTableInner`] has already been allocated, otherwise |
2994 | /// calling this function results in [`undefined behavior`] |
2995 | /// |
2996 | /// * The `table_layout` must be the same [`TableLayout`] as the `TableLayout` |
2997 | /// that was used to allocate this table. Failure to comply with this condition |
2998 | /// may result in [`undefined behavior`]. |
2999 | /// |
3000 | /// See also [`GlobalAlloc::dealloc`] or [`Allocator::deallocate`] for more information. |
3001 | /// |
3002 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
3003 | /// [`GlobalAlloc::dealloc`]: https://doc.rust-lang.org/alloc/alloc/trait.GlobalAlloc.html#tymethod.dealloc |
3004 | /// [`Allocator::deallocate`]: https://doc.rust-lang.org/alloc/alloc/trait.Allocator.html#tymethod.deallocate |
3005 | #[inline ] |
3006 | unsafe fn allocation_info(&self, table_layout: TableLayout) -> (NonNull<u8>, Layout) { |
3007 | debug_assert!( |
3008 | !self.is_empty_singleton(), |
3009 | "this function can only be called on non-empty tables" |
3010 | ); |
3011 | |
3012 | // Avoid `Option::unwrap_or_else` because it bloats LLVM IR. |
3013 | let (layout, ctrl_offset) = match table_layout.calculate_layout_for(self.buckets()) { |
3014 | Some(lco) => lco, |
3015 | None => unsafe { hint::unreachable_unchecked() }, |
3016 | }; |
3017 | ( |
3018 | // SAFETY: The caller must uphold the safety contract for `allocation_info` method. |
3019 | unsafe { NonNull::new_unchecked(self.ctrl.as_ptr().sub(ctrl_offset)) }, |
3020 | layout, |
3021 | ) |
3022 | } |
3023 | |
3024 | /// Returns the total amount of memory allocated internally by the hash |
3025 | /// table, in bytes. |
3026 | /// |
3027 | /// The returned number is informational only. It is intended to be |
3028 | /// primarily used for memory profiling. |
3029 | /// |
3030 | /// # Safety |
3031 | /// |
3032 | /// The `table_layout` must be the same [`TableLayout`] as the `TableLayout` |
3033 | /// that was used to allocate this table. Failure to comply with this condition |
3034 | /// may result in [`undefined behavior`]. |
3035 | /// |
3036 | /// |
3037 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
3038 | #[inline ] |
3039 | unsafe fn allocation_size_or_zero(&self, table_layout: TableLayout) -> usize { |
3040 | if self.is_empty_singleton() { |
3041 | 0 |
3042 | } else { |
3043 | // SAFETY: |
3044 | // 1. We have checked that our table is allocated. |
3045 | // 2. The caller ensures that `table_layout` matches the [`TableLayout`] |
3046 | // that was used to allocate this table. |
3047 | unsafe { self.allocation_info(table_layout).1.size() } |
3048 | } |
3049 | } |
3050 | |
3051 | /// Marks all table buckets as empty without dropping their contents. |
3052 | #[inline ] |
3053 | fn clear_no_drop(&mut self) { |
3054 | if !self.is_empty_singleton() { |
3055 | self.ctrl_slice().fill_empty(); |
3056 | } |
3057 | self.items = 0; |
3058 | self.growth_left = bucket_mask_to_capacity(self.bucket_mask); |
3059 | } |
3060 | |
3061 | /// Erases the [`Bucket`]'s control byte at the given index so that it does not |
3062 | /// triggered as full, decreases the `items` of the table and, if it can be done, |
3063 | /// increases `self.growth_left`. |
3064 | /// |
3065 | /// This function does not actually erase / drop the [`Bucket`] itself, i.e. it |
3066 | /// does not make any changes to the `data` parts of the table. The caller of this |
3067 | /// function must take care to properly drop the `data`, otherwise calling this |
3068 | /// function may result in a memory leak. |
3069 | /// |
3070 | /// # Safety |
3071 | /// |
3072 | /// You must observe the following safety rules when calling this function: |
3073 | /// |
3074 | /// * The [`RawTableInner`] has already been allocated; |
3075 | /// |
3076 | /// * It must be the full control byte at the given position; |
3077 | /// |
3078 | /// * The `index` must not be greater than the `RawTableInner.bucket_mask`, i.e. |
3079 | /// `index <= RawTableInner.bucket_mask` or, in other words, `(index + 1)` must |
3080 | /// be no greater than the number returned by the function [`RawTableInner::buckets`]. |
3081 | /// |
3082 | /// Calling this function on a table that has not been allocated results in [`undefined behavior`]. |
3083 | /// |
3084 | /// Calling this function on a table with no elements is unspecified, but calling subsequent |
3085 | /// functions is likely to result in [`undefined behavior`] due to overflow subtraction |
3086 | /// (`self.items -= 1 cause overflow when self.items == 0`). |
3087 | /// |
3088 | /// See also [`Bucket::as_ptr`] method, for more information about of properly removing |
3089 | /// or saving `data element` from / into the [`RawTable`] / [`RawTableInner`]. |
3090 | /// |
3091 | /// [`RawTableInner::buckets`]: RawTableInner::buckets |
3092 | /// [`Bucket::as_ptr`]: Bucket::as_ptr |
3093 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
3094 | #[inline ] |
3095 | unsafe fn erase(&mut self, index: usize) { |
3096 | debug_assert!(self.is_bucket_full(index)); |
3097 | |
3098 | // This is the same as `index.wrapping_sub(Group::WIDTH) % self.buckets()` because |
3099 | // the number of buckets is a power of two, and `self.bucket_mask = self.buckets() - 1`. |
3100 | let index_before = index.wrapping_sub(Group::WIDTH) & self.bucket_mask; |
3101 | // SAFETY: |
3102 | // - The caller must uphold the safety contract for `erase` method; |
3103 | // - `index_before` is guaranteed to be in range due to masking with `self.bucket_mask` |
3104 | let empty_before = Group::load(self.ctrl(index_before)).match_empty(); |
3105 | let empty_after = Group::load(self.ctrl(index)).match_empty(); |
3106 | |
3107 | // Inserting and searching in the map is performed by two key functions: |
3108 | // |
3109 | // - The `find_insert_slot` function that looks up the index of any `Tag::EMPTY` or `Tag::DELETED` |
3110 | // slot in a group to be able to insert. If it doesn't find an `Tag::EMPTY` or `Tag::DELETED` |
3111 | // slot immediately in the first group, it jumps to the next `Group` looking for it, |
3112 | // and so on until it has gone through all the groups in the control bytes. |
3113 | // |
3114 | // - The `find_inner` function that looks for the index of the desired element by looking |
3115 | // at all the `FULL` bytes in the group. If it did not find the element right away, and |
3116 | // there is no `Tag::EMPTY` byte in the group, then this means that the `find_insert_slot` |
3117 | // function may have found a suitable slot in the next group. Therefore, `find_inner` |
3118 | // jumps further, and if it does not find the desired element and again there is no `Tag::EMPTY` |
3119 | // byte, then it jumps further, and so on. The search stops only if `find_inner` function |
3120 | // finds the desired element or hits an `Tag::EMPTY` slot/byte. |
3121 | // |
3122 | // Accordingly, this leads to two consequences: |
3123 | // |
3124 | // - The map must have `Tag::EMPTY` slots (bytes); |
3125 | // |
3126 | // - You can't just mark the byte to be erased as `Tag::EMPTY`, because otherwise the `find_inner` |
3127 | // function may stumble upon an `Tag::EMPTY` byte before finding the desired element and stop |
3128 | // searching. |
3129 | // |
3130 | // Thus it is necessary to check all bytes after and before the erased element. If we are in |
3131 | // a contiguous `Group` of `FULL` or `Tag::DELETED` bytes (the number of `FULL` or `Tag::DELETED` bytes |
3132 | // before and after is greater than or equal to `Group::WIDTH`), then we must mark our byte as |
3133 | // `Tag::DELETED` in order for the `find_inner` function to go further. On the other hand, if there |
3134 | // is at least one `Tag::EMPTY` slot in the `Group`, then the `find_inner` function will still stumble |
3135 | // upon an `Tag::EMPTY` byte, so we can safely mark our erased byte as `Tag::EMPTY` as well. |
3136 | // |
3137 | // Finally, since `index_before == (index.wrapping_sub(Group::WIDTH) & self.bucket_mask) == index` |
3138 | // and given all of the above, tables smaller than the group width (self.buckets() < Group::WIDTH) |
3139 | // cannot have `Tag::DELETED` bytes. |
3140 | // |
3141 | // Note that in this context `leading_zeros` refers to the bytes at the end of a group, while |
3142 | // `trailing_zeros` refers to the bytes at the beginning of a group. |
3143 | let ctrl = if empty_before.leading_zeros() + empty_after.trailing_zeros() >= Group::WIDTH { |
3144 | Tag::DELETED |
3145 | } else { |
3146 | self.growth_left += 1; |
3147 | Tag::EMPTY |
3148 | }; |
3149 | // SAFETY: the caller must uphold the safety contract for `erase` method. |
3150 | self.set_ctrl(index, ctrl); |
3151 | self.items -= 1; |
3152 | } |
3153 | } |
3154 | |
3155 | impl<T: Clone, A: Allocator + Clone> Clone for RawTable<T, A> { |
3156 | fn clone(&self) -> Self { |
3157 | if self.table.is_empty_singleton() { |
3158 | Self::new_in(self.alloc.clone()) |
3159 | } else { |
3160 | unsafe { |
3161 | // Avoid `Result::ok_or_else` because it bloats LLVM IR. |
3162 | // |
3163 | // SAFETY: This is safe as we are taking the size of an already allocated table |
3164 | // and therefore capacity overflow cannot occur, `self.table.buckets()` is power |
3165 | // of two and all allocator errors will be caught inside `RawTableInner::new_uninitialized`. |
3166 | let mut new_table = match Self::new_uninitialized( |
3167 | self.alloc.clone(), |
3168 | self.table.buckets(), |
3169 | Fallibility::Infallible, |
3170 | ) { |
3171 | Ok(table) => table, |
3172 | Err(_) => hint::unreachable_unchecked(), |
3173 | }; |
3174 | |
3175 | // Cloning elements may fail (the clone function may panic). But we don't |
3176 | // need to worry about uninitialized control bits, since: |
3177 | // 1. The number of items (elements) in the table is zero, which means that |
3178 | // the control bits will not be read by Drop function. |
3179 | // 2. The `clone_from_spec` method will first copy all control bits from |
3180 | // `self` (thus initializing them). But this will not affect the `Drop` |
3181 | // function, since the `clone_from_spec` function sets `items` only after |
3182 | // successfully cloning all elements. |
3183 | new_table.clone_from_spec(self); |
3184 | new_table |
3185 | } |
3186 | } |
3187 | } |
3188 | |
3189 | fn clone_from(&mut self, source: &Self) { |
3190 | if source.table.is_empty_singleton() { |
3191 | let mut old_inner = mem::replace(&mut self.table, RawTableInner::NEW); |
3192 | unsafe { |
3193 | // SAFETY: |
3194 | // 1. We call the function only once; |
3195 | // 2. We know for sure that `alloc` and `table_layout` matches the [`Allocator`] |
3196 | // and [`TableLayout`] that were used to allocate this table. |
3197 | // 3. If any elements' drop function panics, then there will only be a memory leak, |
3198 | // because we have replaced the inner table with a new one. |
3199 | old_inner.drop_inner_table::<T, _>(&self.alloc, Self::TABLE_LAYOUT); |
3200 | } |
3201 | } else { |
3202 | unsafe { |
3203 | // Make sure that if any panics occurs, we clear the table and |
3204 | // leave it in an empty state. |
3205 | let mut self_ = guard(self, |self_| { |
3206 | self_.clear_no_drop(); |
3207 | }); |
3208 | |
3209 | // First, drop all our elements without clearing the control |
3210 | // bytes. If this panics then the scope guard will clear the |
3211 | // table, leaking any elements that were not dropped yet. |
3212 | // |
3213 | // This leak is unavoidable: we can't try dropping more elements |
3214 | // since this could lead to another panic and abort the process. |
3215 | // |
3216 | // SAFETY: If something gets wrong we clear our table right after |
3217 | // dropping the elements, so there is no double drop, since `items` |
3218 | // will be equal to zero. |
3219 | self_.table.drop_elements::<T>(); |
3220 | |
3221 | // If necessary, resize our table to match the source. |
3222 | if self_.buckets() != source.buckets() { |
3223 | let new_inner = match RawTableInner::new_uninitialized( |
3224 | &self_.alloc, |
3225 | Self::TABLE_LAYOUT, |
3226 | source.buckets(), |
3227 | Fallibility::Infallible, |
3228 | ) { |
3229 | Ok(table) => table, |
3230 | Err(_) => hint::unreachable_unchecked(), |
3231 | }; |
3232 | // Replace the old inner with new uninitialized one. It's ok, since if something gets |
3233 | // wrong `ScopeGuard` will initialize all control bytes and leave empty table. |
3234 | let mut old_inner = mem::replace(&mut self_.table, new_inner); |
3235 | if !old_inner.is_empty_singleton() { |
3236 | // SAFETY: |
3237 | // 1. We have checked that our table is allocated. |
3238 | // 2. We know for sure that `alloc` and `table_layout` matches |
3239 | // the [`Allocator`] and [`TableLayout`] that were used to allocate this table. |
3240 | old_inner.free_buckets(&self_.alloc, Self::TABLE_LAYOUT); |
3241 | } |
3242 | } |
3243 | |
3244 | // Cloning elements may fail (the clone function may panic), but the `ScopeGuard` |
3245 | // inside the `clone_from_impl` function will take care of that, dropping all |
3246 | // cloned elements if necessary. Our `ScopeGuard` will clear the table. |
3247 | self_.clone_from_spec(source); |
3248 | |
3249 | // Disarm the scope guard if cloning was successful. |
3250 | ScopeGuard::into_inner(self_); |
3251 | } |
3252 | } |
3253 | } |
3254 | } |
3255 | |
3256 | /// Specialization of `clone_from` for `Copy` types |
3257 | trait RawTableClone { |
3258 | unsafe fn clone_from_spec(&mut self, source: &Self); |
3259 | } |
3260 | impl<T: Clone, A: Allocator + Clone> RawTableClone for RawTable<T, A> { |
3261 | default_fn! { |
3262 | #[cfg_attr (feature = "inline-more" , inline)] |
3263 | unsafe fn clone_from_spec(&mut self, source: &Self) { |
3264 | self.clone_from_impl(source); |
3265 | } |
3266 | } |
3267 | } |
3268 | #[cfg (feature = "nightly" )] |
3269 | impl<T: Copy, A: Allocator + Clone> RawTableClone for RawTable<T, A> { |
3270 | #[cfg_attr (feature = "inline-more" , inline)] |
3271 | unsafe fn clone_from_spec(&mut self, source: &Self) { |
3272 | source |
3273 | .table |
3274 | .ctrl(0) |
3275 | .copy_to_nonoverlapping(self.table.ctrl(0), self.table.num_ctrl_bytes()); |
3276 | source |
3277 | .data_start() |
3278 | .as_ptr() |
3279 | .copy_to_nonoverlapping(self.data_start().as_ptr(), self.table.buckets()); |
3280 | |
3281 | self.table.items = source.table.items; |
3282 | self.table.growth_left = source.table.growth_left; |
3283 | } |
3284 | } |
3285 | |
3286 | impl<T: Clone, A: Allocator + Clone> RawTable<T, A> { |
3287 | /// Common code for `clone` and `clone_from`. Assumes: |
3288 | /// - `self.buckets() == source.buckets()`. |
3289 | /// - Any existing elements have been dropped. |
3290 | /// - The control bytes are not initialized yet. |
3291 | #[cfg_attr (feature = "inline-more" , inline)] |
3292 | unsafe fn clone_from_impl(&mut self, source: &Self) { |
3293 | // Copy the control bytes unchanged. We do this in a single pass |
3294 | source |
3295 | .table |
3296 | .ctrl(0) |
3297 | .copy_to_nonoverlapping(self.table.ctrl(0), self.table.num_ctrl_bytes()); |
3298 | |
3299 | // The cloning of elements may panic, in which case we need |
3300 | // to make sure we drop only the elements that have been |
3301 | // cloned so far. |
3302 | let mut guard = guard((0, &mut *self), |(index, self_)| { |
3303 | if T::NEEDS_DROP { |
3304 | for i in 0..*index { |
3305 | if self_.is_bucket_full(i) { |
3306 | self_.bucket(i).drop(); |
3307 | } |
3308 | } |
3309 | } |
3310 | }); |
3311 | |
3312 | for from in source.iter() { |
3313 | let index = source.bucket_index(&from); |
3314 | let to = guard.1.bucket(index); |
3315 | to.write(from.as_ref().clone()); |
3316 | |
3317 | // Update the index in case we need to unwind. |
3318 | guard.0 = index + 1; |
3319 | } |
3320 | |
3321 | // Successfully cloned all items, no need to clean up. |
3322 | mem::forget(guard); |
3323 | |
3324 | self.table.items = source.table.items; |
3325 | self.table.growth_left = source.table.growth_left; |
3326 | } |
3327 | } |
3328 | |
3329 | impl<T, A: Allocator + Default> Default for RawTable<T, A> { |
3330 | #[inline ] |
3331 | fn default() -> Self { |
3332 | Self::new_in(alloc:Default::default()) |
3333 | } |
3334 | } |
3335 | |
3336 | #[cfg (feature = "nightly" )] |
3337 | unsafe impl<#[may_dangle ] T, A: Allocator> Drop for RawTable<T, A> { |
3338 | #[cfg_attr (feature = "inline-more" , inline)] |
3339 | fn drop(&mut self) { |
3340 | unsafe { |
3341 | // SAFETY: |
3342 | // 1. We call the function only once; |
3343 | // 2. We know for sure that `alloc` and `table_layout` matches the [`Allocator`] |
3344 | // and [`TableLayout`] that were used to allocate this table. |
3345 | // 3. If the drop function of any elements fails, then only a memory leak will occur, |
3346 | // and we don't care because we are inside the `Drop` function of the `RawTable`, |
3347 | // so there won't be any table left in an inconsistent state. |
3348 | self.table |
3349 | .drop_inner_table::<T, _>(&self.alloc, Self::TABLE_LAYOUT); |
3350 | } |
3351 | } |
3352 | } |
3353 | #[cfg (not(feature = "nightly" ))] |
3354 | impl<T, A: Allocator> Drop for RawTable<T, A> { |
3355 | #[cfg_attr (feature = "inline-more" , inline)] |
3356 | fn drop(&mut self) { |
3357 | unsafe { |
3358 | // SAFETY: |
3359 | // 1. We call the function only once; |
3360 | // 2. We know for sure that `alloc` and `table_layout` matches the [`Allocator`] |
3361 | // and [`TableLayout`] that were used to allocate this table. |
3362 | // 3. If the drop function of any elements fails, then only a memory leak will occur, |
3363 | // and we don't care because we are inside the `Drop` function of the `RawTable`, |
3364 | // so there won't be any table left in an inconsistent state. |
3365 | self.table |
3366 | .drop_inner_table::<T, _>(&self.alloc, Self::TABLE_LAYOUT); |
3367 | } |
3368 | } |
3369 | } |
3370 | |
3371 | impl<T, A: Allocator> IntoIterator for RawTable<T, A> { |
3372 | type Item = T; |
3373 | type IntoIter = RawIntoIter<T, A>; |
3374 | |
3375 | #[cfg_attr (feature = "inline-more" , inline)] |
3376 | fn into_iter(self) -> RawIntoIter<T, A> { |
3377 | unsafe { |
3378 | let iter: RawIter = self.iter(); |
3379 | self.into_iter_from(iter) |
3380 | } |
3381 | } |
3382 | } |
3383 | |
3384 | /// Iterator over a sub-range of a table. Unlike `RawIter` this iterator does |
3385 | /// not track an item count. |
3386 | pub(crate) struct RawIterRange<T> { |
3387 | // Mask of full buckets in the current group. Bits are cleared from this |
3388 | // mask as each element is processed. |
3389 | current_group: BitMaskIter, |
3390 | |
3391 | // Pointer to the buckets for the current group. |
3392 | data: Bucket<T>, |
3393 | |
3394 | // Pointer to the next group of control bytes, |
3395 | // Must be aligned to the group size. |
3396 | next_ctrl: *const u8, |
3397 | |
3398 | // Pointer one past the last control byte of this range. |
3399 | end: *const u8, |
3400 | } |
3401 | |
3402 | impl<T> RawIterRange<T> { |
3403 | /// Returns a `RawIterRange` covering a subset of a table. |
3404 | /// |
3405 | /// # Safety |
3406 | /// |
3407 | /// If any of the following conditions are violated, the result is |
3408 | /// [`undefined behavior`]: |
3409 | /// |
3410 | /// * `ctrl` must be [valid] for reads, i.e. table outlives the `RawIterRange`; |
3411 | /// |
3412 | /// * `ctrl` must be properly aligned to the group size (`Group::WIDTH`); |
3413 | /// |
3414 | /// * `ctrl` must point to the array of properly initialized control bytes; |
3415 | /// |
3416 | /// * `data` must be the [`Bucket`] at the `ctrl` index in the table; |
3417 | /// |
3418 | /// * the value of `len` must be less than or equal to the number of table buckets, |
3419 | /// and the returned value of `ctrl.as_ptr().add(len).offset_from(ctrl.as_ptr())` |
3420 | /// must be positive. |
3421 | /// |
3422 | /// * The `ctrl.add(len)` pointer must be either in bounds or one |
3423 | /// byte past the end of the same [allocated table]. |
3424 | /// |
3425 | /// * The `len` must be a power of two. |
3426 | /// |
3427 | /// [valid]: https://doc.rust-lang.org/std/ptr/index.html#safety |
3428 | /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
3429 | #[cfg_attr (feature = "inline-more" , inline)] |
3430 | unsafe fn new(ctrl: *const u8, data: Bucket<T>, len: usize) -> Self { |
3431 | debug_assert_ne!(len, 0); |
3432 | debug_assert_eq!(ctrl as usize % Group::WIDTH, 0); |
3433 | // SAFETY: The caller must uphold the safety rules for the [`RawIterRange::new`] |
3434 | let end = ctrl.add(len); |
3435 | |
3436 | // Load the first group and advance ctrl to point to the next group |
3437 | // SAFETY: The caller must uphold the safety rules for the [`RawIterRange::new`] |
3438 | let current_group = Group::load_aligned(ctrl.cast()).match_full(); |
3439 | let next_ctrl = ctrl.add(Group::WIDTH); |
3440 | |
3441 | Self { |
3442 | current_group: current_group.into_iter(), |
3443 | data, |
3444 | next_ctrl, |
3445 | end, |
3446 | } |
3447 | } |
3448 | |
3449 | /// Splits a `RawIterRange` into two halves. |
3450 | /// |
3451 | /// Returns `None` if the remaining range is smaller than or equal to the |
3452 | /// group width. |
3453 | #[cfg_attr (feature = "inline-more" , inline)] |
3454 | #[cfg (feature = "rayon" )] |
3455 | pub(crate) fn split(mut self) -> (Self, Option<RawIterRange<T>>) { |
3456 | unsafe { |
3457 | if self.end <= self.next_ctrl { |
3458 | // Nothing to split if the group that we are current processing |
3459 | // is the last one. |
3460 | (self, None) |
3461 | } else { |
3462 | // len is the remaining number of elements after the group that |
3463 | // we are currently processing. It must be a multiple of the |
3464 | // group size (small tables are caught by the check above). |
3465 | let len = offset_from(self.end, self.next_ctrl); |
3466 | debug_assert_eq!(len % Group::WIDTH, 0); |
3467 | |
3468 | // Split the remaining elements into two halves, but round the |
3469 | // midpoint down in case there is an odd number of groups |
3470 | // remaining. This ensures that: |
3471 | // - The tail is at least 1 group long. |
3472 | // - The split is roughly even considering we still have the |
3473 | // current group to process. |
3474 | let mid = (len / 2) & !(Group::WIDTH - 1); |
3475 | |
3476 | let tail = Self::new( |
3477 | self.next_ctrl.add(mid), |
3478 | self.data.next_n(Group::WIDTH).next_n(mid), |
3479 | len - mid, |
3480 | ); |
3481 | debug_assert_eq!( |
3482 | self.data.next_n(Group::WIDTH).next_n(mid).ptr, |
3483 | tail.data.ptr |
3484 | ); |
3485 | debug_assert_eq!(self.end, tail.end); |
3486 | self.end = self.next_ctrl.add(mid); |
3487 | debug_assert_eq!(self.end.add(Group::WIDTH), tail.next_ctrl); |
3488 | (self, Some(tail)) |
3489 | } |
3490 | } |
3491 | } |
3492 | |
3493 | /// # Safety |
3494 | /// If `DO_CHECK_PTR_RANGE` is false, caller must ensure that we never try to iterate |
3495 | /// after yielding all elements. |
3496 | #[cfg_attr (feature = "inline-more" , inline)] |
3497 | unsafe fn next_impl<const DO_CHECK_PTR_RANGE: bool>(&mut self) -> Option<Bucket<T>> { |
3498 | loop { |
3499 | if let Some(index) = self.current_group.next() { |
3500 | return Some(self.data.next_n(index)); |
3501 | } |
3502 | |
3503 | if DO_CHECK_PTR_RANGE && self.next_ctrl >= self.end { |
3504 | return None; |
3505 | } |
3506 | |
3507 | // We might read past self.end up to the next group boundary, |
3508 | // but this is fine because it only occurs on tables smaller |
3509 | // than the group size where the trailing control bytes are all |
3510 | // EMPTY. On larger tables self.end is guaranteed to be aligned |
3511 | // to the group size (since tables are power-of-two sized). |
3512 | self.current_group = Group::load_aligned(self.next_ctrl.cast()) |
3513 | .match_full() |
3514 | .into_iter(); |
3515 | self.data = self.data.next_n(Group::WIDTH); |
3516 | self.next_ctrl = self.next_ctrl.add(Group::WIDTH); |
3517 | } |
3518 | } |
3519 | |
3520 | /// Folds every element into an accumulator by applying an operation, |
3521 | /// returning the final result. |
3522 | /// |
3523 | /// `fold_impl()` takes three arguments: the number of items remaining in |
3524 | /// the iterator, an initial value, and a closure with two arguments: an |
3525 | /// 'accumulator', and an element. The closure returns the value that the |
3526 | /// accumulator should have for the next iteration. |
3527 | /// |
3528 | /// The initial value is the value the accumulator will have on the first call. |
3529 | /// |
3530 | /// After applying this closure to every element of the iterator, `fold_impl()` |
3531 | /// returns the accumulator. |
3532 | /// |
3533 | /// # Safety |
3534 | /// |
3535 | /// If any of the following conditions are violated, the result is |
3536 | /// [`Undefined Behavior`]: |
3537 | /// |
3538 | /// * The [`RawTableInner`] / [`RawTable`] must be alive and not moved, |
3539 | /// i.e. table outlives the `RawIterRange`; |
3540 | /// |
3541 | /// * The provided `n` value must match the actual number of items |
3542 | /// in the table. |
3543 | /// |
3544 | /// [`Undefined Behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
3545 | #[allow (clippy::while_let_on_iterator)] |
3546 | #[cfg_attr (feature = "inline-more" , inline)] |
3547 | unsafe fn fold_impl<F, B>(mut self, mut n: usize, mut acc: B, mut f: F) -> B |
3548 | where |
3549 | F: FnMut(B, Bucket<T>) -> B, |
3550 | { |
3551 | loop { |
3552 | while let Some(index) = self.current_group.next() { |
3553 | // The returned `index` will always be in the range `0..Group::WIDTH`, |
3554 | // so that calling `self.data.next_n(index)` is safe (see detailed explanation below). |
3555 | debug_assert!(n != 0); |
3556 | let bucket = self.data.next_n(index); |
3557 | acc = f(acc, bucket); |
3558 | n -= 1; |
3559 | } |
3560 | |
3561 | if n == 0 { |
3562 | return acc; |
3563 | } |
3564 | |
3565 | // SAFETY: The caller of this function ensures that: |
3566 | // |
3567 | // 1. The provided `n` value matches the actual number of items in the table; |
3568 | // 2. The table is alive and did not moved. |
3569 | // |
3570 | // Taking the above into account, we always stay within the bounds, because: |
3571 | // |
3572 | // 1. For tables smaller than the group width (self.buckets() <= Group::WIDTH), |
3573 | // we will never end up in the given branch, since we should have already |
3574 | // yielded all the elements of the table. |
3575 | // |
3576 | // 2. For tables larger than the group width. The number of buckets is a |
3577 | // power of two (2 ^ n), Group::WIDTH is also power of two (2 ^ k). Since |
3578 | // `(2 ^ n) > (2 ^ k)`, than `(2 ^ n) % (2 ^ k) = 0`. As we start from the |
3579 | // start of the array of control bytes, and never try to iterate after |
3580 | // getting all the elements, the last `self.current_group` will read bytes |
3581 | // from the `self.buckets() - Group::WIDTH` index. We know also that |
3582 | // `self.current_group.next()` will always return indices within the range |
3583 | // `0..Group::WIDTH`. |
3584 | // |
3585 | // Knowing all of the above and taking into account that we are synchronizing |
3586 | // the `self.data` index with the index we used to read the `self.current_group`, |
3587 | // the subsequent `self.data.next_n(index)` will always return a bucket with |
3588 | // an index number less than `self.buckets()`. |
3589 | // |
3590 | // The last `self.next_ctrl`, whose index would be `self.buckets()`, will never |
3591 | // actually be read, since we should have already yielded all the elements of |
3592 | // the table. |
3593 | self.current_group = Group::load_aligned(self.next_ctrl.cast()) |
3594 | .match_full() |
3595 | .into_iter(); |
3596 | self.data = self.data.next_n(Group::WIDTH); |
3597 | self.next_ctrl = self.next_ctrl.add(Group::WIDTH); |
3598 | } |
3599 | } |
3600 | } |
3601 | |
3602 | // We make raw iterators unconditionally Send and Sync, and let the PhantomData |
3603 | // in the actual iterator implementations determine the real Send/Sync bounds. |
3604 | unsafe impl<T> Send for RawIterRange<T> {} |
3605 | unsafe impl<T> Sync for RawIterRange<T> {} |
3606 | |
3607 | impl<T> Clone for RawIterRange<T> { |
3608 | #[cfg_attr (feature = "inline-more" , inline)] |
3609 | fn clone(&self) -> Self { |
3610 | Self { |
3611 | data: self.data.clone(), |
3612 | next_ctrl: self.next_ctrl, |
3613 | current_group: self.current_group.clone(), |
3614 | end: self.end, |
3615 | } |
3616 | } |
3617 | } |
3618 | |
3619 | impl<T> Iterator for RawIterRange<T> { |
3620 | type Item = Bucket<T>; |
3621 | |
3622 | #[cfg_attr (feature = "inline-more" , inline)] |
3623 | fn next(&mut self) -> Option<Bucket<T>> { |
3624 | unsafe { |
3625 | // SAFETY: We set checker flag to true. |
3626 | self.next_impl::<true>() |
3627 | } |
3628 | } |
3629 | |
3630 | #[inline ] |
3631 | fn size_hint(&self) -> (usize, Option<usize>) { |
3632 | // We don't have an item count, so just guess based on the range size. |
3633 | let remaining_buckets: usize = if self.end > self.next_ctrl { |
3634 | unsafe { offset_from(self.end, self.next_ctrl) } |
3635 | } else { |
3636 | 0 |
3637 | }; |
3638 | |
3639 | // Add a group width to include the group we are currently processing. |
3640 | (0, Some(Group::WIDTH + remaining_buckets)) |
3641 | } |
3642 | } |
3643 | |
3644 | impl<T> FusedIterator for RawIterRange<T> {} |
3645 | |
3646 | /// Iterator which returns a raw pointer to every full bucket in the table. |
3647 | /// |
3648 | /// For maximum flexibility this iterator is not bound by a lifetime, but you |
3649 | /// must observe several rules when using it: |
3650 | /// - You must not free the hash table while iterating (including via growing/shrinking). |
3651 | /// - It is fine to erase a bucket that has been yielded by the iterator. |
3652 | /// - Erasing a bucket that has not yet been yielded by the iterator may still |
3653 | /// result in the iterator yielding that bucket (unless `reflect_remove` is called). |
3654 | /// - It is unspecified whether an element inserted after the iterator was |
3655 | /// created will be yielded by that iterator (unless `reflect_insert` is called). |
3656 | /// - The order in which the iterator yields bucket is unspecified and may |
3657 | /// change in the future. |
3658 | pub struct RawIter<T> { |
3659 | pub(crate) iter: RawIterRange<T>, |
3660 | items: usize, |
3661 | } |
3662 | |
3663 | impl<T> RawIter<T> { |
3664 | unsafe fn drop_elements(&mut self) { |
3665 | if T::NEEDS_DROP && self.items != 0 { |
3666 | for item: Bucket in self { |
3667 | item.drop(); |
3668 | } |
3669 | } |
3670 | } |
3671 | } |
3672 | |
3673 | impl<T> Clone for RawIter<T> { |
3674 | #[cfg_attr (feature = "inline-more" , inline)] |
3675 | fn clone(&self) -> Self { |
3676 | Self { |
3677 | iter: self.iter.clone(), |
3678 | items: self.items, |
3679 | } |
3680 | } |
3681 | } |
3682 | impl<T> Default for RawIter<T> { |
3683 | #[cfg_attr (feature = "inline-more" , inline)] |
3684 | fn default() -> Self { |
3685 | // SAFETY: Because the table is static, it always outlives the iter. |
3686 | unsafe { RawTableInner::NEW.iter() } |
3687 | } |
3688 | } |
3689 | |
3690 | impl<T> Iterator for RawIter<T> { |
3691 | type Item = Bucket<T>; |
3692 | |
3693 | #[cfg_attr (feature = "inline-more" , inline)] |
3694 | fn next(&mut self) -> Option<Bucket<T>> { |
3695 | // Inner iterator iterates over buckets |
3696 | // so it can do unnecessary work if we already yielded all items. |
3697 | if self.items == 0 { |
3698 | return None; |
3699 | } |
3700 | |
3701 | let nxt = unsafe { |
3702 | // SAFETY: We check number of items to yield using `items` field. |
3703 | self.iter.next_impl::<false>() |
3704 | }; |
3705 | |
3706 | debug_assert!(nxt.is_some()); |
3707 | self.items -= 1; |
3708 | |
3709 | nxt |
3710 | } |
3711 | |
3712 | #[inline ] |
3713 | fn size_hint(&self) -> (usize, Option<usize>) { |
3714 | (self.items, Some(self.items)) |
3715 | } |
3716 | |
3717 | #[inline ] |
3718 | fn fold<B, F>(self, init: B, f: F) -> B |
3719 | where |
3720 | Self: Sized, |
3721 | F: FnMut(B, Self::Item) -> B, |
3722 | { |
3723 | unsafe { self.iter.fold_impl(self.items, init, f) } |
3724 | } |
3725 | } |
3726 | |
3727 | impl<T> ExactSizeIterator for RawIter<T> {} |
3728 | impl<T> FusedIterator for RawIter<T> {} |
3729 | |
3730 | /// Iterator which returns an index of every full bucket in the table. |
3731 | /// |
3732 | /// For maximum flexibility this iterator is not bound by a lifetime, but you |
3733 | /// must observe several rules when using it: |
3734 | /// - You must not free the hash table while iterating (including via growing/shrinking). |
3735 | /// - It is fine to erase a bucket that has been yielded by the iterator. |
3736 | /// - Erasing a bucket that has not yet been yielded by the iterator may still |
3737 | /// result in the iterator yielding index of that bucket. |
3738 | /// - It is unspecified whether an element inserted after the iterator was |
3739 | /// created will be yielded by that iterator. |
3740 | /// - The order in which the iterator yields indices of the buckets is unspecified |
3741 | /// and may change in the future. |
3742 | pub(crate) struct FullBucketsIndices { |
3743 | // Mask of full buckets in the current group. Bits are cleared from this |
3744 | // mask as each element is processed. |
3745 | current_group: BitMaskIter, |
3746 | |
3747 | // Initial value of the bytes' indices of the current group (relative |
3748 | // to the start of the control bytes). |
3749 | group_first_index: usize, |
3750 | |
3751 | // Pointer to the current group of control bytes, |
3752 | // Must be aligned to the group size (Group::WIDTH). |
3753 | ctrl: NonNull<u8>, |
3754 | |
3755 | // Number of elements in the table. |
3756 | items: usize, |
3757 | } |
3758 | |
3759 | impl FullBucketsIndices { |
3760 | /// Advances the iterator and returns the next value. |
3761 | /// |
3762 | /// # Safety |
3763 | /// |
3764 | /// If any of the following conditions are violated, the result is |
3765 | /// [`Undefined Behavior`]: |
3766 | /// |
3767 | /// * The [`RawTableInner`] / [`RawTable`] must be alive and not moved, |
3768 | /// i.e. table outlives the `FullBucketsIndices`; |
3769 | /// |
3770 | /// * It never tries to iterate after getting all elements. |
3771 | /// |
3772 | /// [`Undefined Behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html |
3773 | #[inline (always)] |
3774 | unsafe fn next_impl(&mut self) -> Option<usize> { |
3775 | loop { |
3776 | if let Some(index) = self.current_group.next() { |
3777 | // The returned `self.group_first_index + index` will always |
3778 | // be in the range `0..self.buckets()`. See explanation below. |
3779 | return Some(self.group_first_index + index); |
3780 | } |
3781 | |
3782 | // SAFETY: The caller of this function ensures that: |
3783 | // |
3784 | // 1. It never tries to iterate after getting all the elements; |
3785 | // 2. The table is alive and did not moved; |
3786 | // 3. The first `self.ctrl` pointed to the start of the array of control bytes. |
3787 | // |
3788 | // Taking the above into account, we always stay within the bounds, because: |
3789 | // |
3790 | // 1. For tables smaller than the group width (self.buckets() <= Group::WIDTH), |
3791 | // we will never end up in the given branch, since we should have already |
3792 | // yielded all the elements of the table. |
3793 | // |
3794 | // 2. For tables larger than the group width. The number of buckets is a |
3795 | // power of two (2 ^ n), Group::WIDTH is also power of two (2 ^ k). Since |
3796 | // `(2 ^ n) > (2 ^ k)`, than `(2 ^ n) % (2 ^ k) = 0`. As we start from the |
3797 | // the start of the array of control bytes, and never try to iterate after |
3798 | // getting all the elements, the last `self.ctrl` will be equal to |
3799 | // the `self.buckets() - Group::WIDTH`, so `self.current_group.next()` |
3800 | // will always contains indices within the range `0..Group::WIDTH`, |
3801 | // and subsequent `self.group_first_index + index` will always return a |
3802 | // number less than `self.buckets()`. |
3803 | self.ctrl = NonNull::new_unchecked(self.ctrl.as_ptr().add(Group::WIDTH)); |
3804 | |
3805 | // SAFETY: See explanation above. |
3806 | self.current_group = Group::load_aligned(self.ctrl.as_ptr().cast()) |
3807 | .match_full() |
3808 | .into_iter(); |
3809 | self.group_first_index += Group::WIDTH; |
3810 | } |
3811 | } |
3812 | } |
3813 | |
3814 | impl Iterator for FullBucketsIndices { |
3815 | type Item = usize; |
3816 | |
3817 | /// Advances the iterator and returns the next value. It is up to |
3818 | /// the caller to ensure that the `RawTable` outlives the `FullBucketsIndices`, |
3819 | /// because we cannot make the `next` method unsafe. |
3820 | #[inline (always)] |
3821 | fn next(&mut self) -> Option<usize> { |
3822 | // Return if we already yielded all items. |
3823 | if self.items == 0 { |
3824 | return None; |
3825 | } |
3826 | |
3827 | let nxt = unsafe { |
3828 | // SAFETY: |
3829 | // 1. We check number of items to yield using `items` field. |
3830 | // 2. The caller ensures that the table is alive and has not moved. |
3831 | self.next_impl() |
3832 | }; |
3833 | |
3834 | debug_assert!(nxt.is_some()); |
3835 | self.items -= 1; |
3836 | |
3837 | nxt |
3838 | } |
3839 | |
3840 | #[inline (always)] |
3841 | fn size_hint(&self) -> (usize, Option<usize>) { |
3842 | (self.items, Some(self.items)) |
3843 | } |
3844 | } |
3845 | |
3846 | impl ExactSizeIterator for FullBucketsIndices {} |
3847 | impl FusedIterator for FullBucketsIndices {} |
3848 | |
3849 | /// Iterator which consumes a table and returns elements. |
3850 | pub struct RawIntoIter<T, A: Allocator = Global> { |
3851 | iter: RawIter<T>, |
3852 | allocation: Option<(NonNull<u8>, Layout, A)>, |
3853 | marker: PhantomData<T>, |
3854 | } |
3855 | |
3856 | impl<T, A: Allocator> RawIntoIter<T, A> { |
3857 | #[cfg_attr (feature = "inline-more" , inline)] |
3858 | pub fn iter(&self) -> RawIter<T> { |
3859 | self.iter.clone() |
3860 | } |
3861 | } |
3862 | |
3863 | unsafe impl<T, A: Allocator> Send for RawIntoIter<T, A> |
3864 | where |
3865 | T: Send, |
3866 | A: Send, |
3867 | { |
3868 | } |
3869 | unsafe impl<T, A: Allocator> Sync for RawIntoIter<T, A> |
3870 | where |
3871 | T: Sync, |
3872 | A: Sync, |
3873 | { |
3874 | } |
3875 | |
3876 | #[cfg (feature = "nightly" )] |
3877 | unsafe impl<#[may_dangle ] T, A: Allocator> Drop for RawIntoIter<T, A> { |
3878 | #[cfg_attr (feature = "inline-more" , inline)] |
3879 | fn drop(&mut self) { |
3880 | unsafe { |
3881 | // Drop all remaining elements |
3882 | self.iter.drop_elements(); |
3883 | |
3884 | // Free the table |
3885 | if let Some((ptr: NonNull, layout: Layout, ref alloc: &A)) = self.allocation { |
3886 | alloc.deallocate(ptr, layout); |
3887 | } |
3888 | } |
3889 | } |
3890 | } |
3891 | #[cfg (not(feature = "nightly" ))] |
3892 | impl<T, A: Allocator> Drop for RawIntoIter<T, A> { |
3893 | #[cfg_attr (feature = "inline-more" , inline)] |
3894 | fn drop(&mut self) { |
3895 | unsafe { |
3896 | // Drop all remaining elements |
3897 | self.iter.drop_elements(); |
3898 | |
3899 | // Free the table |
3900 | if let Some((ptr, layout, ref alloc)) = self.allocation { |
3901 | alloc.deallocate(ptr, layout); |
3902 | } |
3903 | } |
3904 | } |
3905 | } |
3906 | |
3907 | impl<T, A: Allocator> Default for RawIntoIter<T, A> { |
3908 | fn default() -> Self { |
3909 | Self { |
3910 | iter: Default::default(), |
3911 | allocation: None, |
3912 | marker: PhantomData, |
3913 | } |
3914 | } |
3915 | } |
3916 | impl<T, A: Allocator> Iterator for RawIntoIter<T, A> { |
3917 | type Item = T; |
3918 | |
3919 | #[cfg_attr (feature = "inline-more" , inline)] |
3920 | fn next(&mut self) -> Option<T> { |
3921 | unsafe { Some(self.iter.next()?.read()) } |
3922 | } |
3923 | |
3924 | #[inline ] |
3925 | fn size_hint(&self) -> (usize, Option<usize>) { |
3926 | self.iter.size_hint() |
3927 | } |
3928 | } |
3929 | |
3930 | impl<T, A: Allocator> ExactSizeIterator for RawIntoIter<T, A> {} |
3931 | impl<T, A: Allocator> FusedIterator for RawIntoIter<T, A> {} |
3932 | |
3933 | /// Iterator which consumes elements without freeing the table storage. |
3934 | pub struct RawDrain<'a, T, A: Allocator = Global> { |
3935 | iter: RawIter<T>, |
3936 | |
3937 | // The table is moved into the iterator for the duration of the drain. This |
3938 | // ensures that an empty table is left if the drain iterator is leaked |
3939 | // without dropping. |
3940 | table: RawTableInner, |
3941 | orig_table: NonNull<RawTableInner>, |
3942 | |
3943 | // We don't use a &'a mut RawTable<T> because we want RawDrain to be |
3944 | // covariant over T. |
3945 | marker: PhantomData<&'a RawTable<T, A>>, |
3946 | } |
3947 | |
3948 | impl<T, A: Allocator> RawDrain<'_, T, A> { |
3949 | #[cfg_attr (feature = "inline-more" , inline)] |
3950 | pub fn iter(&self) -> RawIter<T> { |
3951 | self.iter.clone() |
3952 | } |
3953 | } |
3954 | |
3955 | unsafe impl<T, A: Allocator> Send for RawDrain<'_, T, A> |
3956 | where |
3957 | T: Send, |
3958 | A: Send, |
3959 | { |
3960 | } |
3961 | unsafe impl<T, A: Allocator> Sync for RawDrain<'_, T, A> |
3962 | where |
3963 | T: Sync, |
3964 | A: Sync, |
3965 | { |
3966 | } |
3967 | |
3968 | impl<T, A: Allocator> Drop for RawDrain<'_, T, A> { |
3969 | #[cfg_attr (feature = "inline-more" , inline)] |
3970 | fn drop(&mut self) { |
3971 | unsafe { |
3972 | // Drop all remaining elements. Note that this may panic. |
3973 | self.iter.drop_elements(); |
3974 | |
3975 | // Reset the contents of the table now that all elements have been |
3976 | // dropped. |
3977 | self.table.clear_no_drop(); |
3978 | |
3979 | // Move the now empty table back to its original location. |
3980 | self.orig_table |
3981 | .as_ptr() |
3982 | .copy_from_nonoverlapping(&self.table, count:1); |
3983 | } |
3984 | } |
3985 | } |
3986 | |
3987 | impl<T, A: Allocator> Iterator for RawDrain<'_, T, A> { |
3988 | type Item = T; |
3989 | |
3990 | #[cfg_attr (feature = "inline-more" , inline)] |
3991 | fn next(&mut self) -> Option<T> { |
3992 | unsafe { |
3993 | let item: Bucket = self.iter.next()?; |
3994 | Some(item.read()) |
3995 | } |
3996 | } |
3997 | |
3998 | #[inline ] |
3999 | fn size_hint(&self) -> (usize, Option<usize>) { |
4000 | self.iter.size_hint() |
4001 | } |
4002 | } |
4003 | |
4004 | impl<T, A: Allocator> ExactSizeIterator for RawDrain<'_, T, A> {} |
4005 | impl<T, A: Allocator> FusedIterator for RawDrain<'_, T, A> {} |
4006 | |
4007 | /// Iterator over occupied buckets that could match a given hash. |
4008 | /// |
4009 | /// `RawTable` only stores 7 bits of the hash value, so this iterator may return |
4010 | /// items that have a hash value different than the one provided. You should |
4011 | /// always validate the returned values before using them. |
4012 | /// |
4013 | /// For maximum flexibility this iterator is not bound by a lifetime, but you |
4014 | /// must observe several rules when using it: |
4015 | /// - You must not free the hash table while iterating (including via growing/shrinking). |
4016 | /// - It is fine to erase a bucket that has been yielded by the iterator. |
4017 | /// - Erasing a bucket that has not yet been yielded by the iterator may still |
4018 | /// result in the iterator yielding that bucket. |
4019 | /// - It is unspecified whether an element inserted after the iterator was |
4020 | /// created will be yielded by that iterator. |
4021 | /// - The order in which the iterator yields buckets is unspecified and may |
4022 | /// change in the future. |
4023 | pub struct RawIterHash<T> { |
4024 | inner: RawIterHashInner, |
4025 | _marker: PhantomData<T>, |
4026 | } |
4027 | |
4028 | #[derive (Clone)] |
4029 | struct RawIterHashInner { |
4030 | // See `RawTableInner`'s corresponding fields for details. |
4031 | // We can't store a `*const RawTableInner` as it would get |
4032 | // invalidated by the user calling `&mut` methods on `RawTable`. |
4033 | bucket_mask: usize, |
4034 | ctrl: NonNull<u8>, |
4035 | |
4036 | // The top 7 bits of the hash. |
4037 | tag_hash: Tag, |
4038 | |
4039 | // The sequence of groups to probe in the search. |
4040 | probe_seq: ProbeSeq, |
4041 | |
4042 | group: Group, |
4043 | |
4044 | // The elements within the group with a matching tag-hash. |
4045 | bitmask: BitMaskIter, |
4046 | } |
4047 | |
4048 | impl<T> RawIterHash<T> { |
4049 | #[cfg_attr (feature = "inline-more" , inline)] |
4050 | unsafe fn new<A: Allocator>(table: &RawTable<T, A>, hash: u64) -> Self { |
4051 | RawIterHash { |
4052 | inner: RawIterHashInner::new(&table.table, hash), |
4053 | _marker: PhantomData, |
4054 | } |
4055 | } |
4056 | } |
4057 | |
4058 | impl<T> Clone for RawIterHash<T> { |
4059 | #[cfg_attr (feature = "inline-more" , inline)] |
4060 | fn clone(&self) -> Self { |
4061 | Self { |
4062 | inner: self.inner.clone(), |
4063 | _marker: PhantomData, |
4064 | } |
4065 | } |
4066 | } |
4067 | |
4068 | impl<T> Default for RawIterHash<T> { |
4069 | #[cfg_attr (feature = "inline-more" , inline)] |
4070 | fn default() -> Self { |
4071 | Self { |
4072 | // SAFETY: Because the table is static, it always outlives the iter. |
4073 | inner: unsafe { RawIterHashInner::new(&RawTableInner::NEW, hash:0) }, |
4074 | _marker: PhantomData, |
4075 | } |
4076 | } |
4077 | } |
4078 | |
4079 | impl RawIterHashInner { |
4080 | #[cfg_attr (feature = "inline-more" , inline)] |
4081 | unsafe fn new(table: &RawTableInner, hash: u64) -> Self { |
4082 | let tag_hash: Tag = Tag::full(hash); |
4083 | let probe_seq: ProbeSeq = table.probe_seq(hash); |
4084 | let group: Group = Group::load(ptr:table.ctrl(index:probe_seq.pos)); |
4085 | let bitmask: BitMaskIter = group.match_tag(tag_hash).into_iter(); |
4086 | |
4087 | RawIterHashInner { |
4088 | bucket_mask: table.bucket_mask, |
4089 | ctrl: table.ctrl, |
4090 | tag_hash, |
4091 | probe_seq, |
4092 | group, |
4093 | bitmask, |
4094 | } |
4095 | } |
4096 | } |
4097 | |
4098 | impl<T> Iterator for RawIterHash<T> { |
4099 | type Item = Bucket<T>; |
4100 | |
4101 | fn next(&mut self) -> Option<Bucket<T>> { |
4102 | unsafe { |
4103 | match self.inner.next() { |
4104 | Some(index: usize) => { |
4105 | // Can't use `RawTable::bucket` here as we don't have |
4106 | // an actual `RawTable` reference to use. |
4107 | debug_assert!(index <= self.inner.bucket_mask); |
4108 | let bucket: Bucket = Bucket::from_base_index(self.inner.ctrl.cast(), index); |
4109 | Some(bucket) |
4110 | } |
4111 | None => None, |
4112 | } |
4113 | } |
4114 | } |
4115 | } |
4116 | |
4117 | impl Iterator for RawIterHashInner { |
4118 | type Item = usize; |
4119 | |
4120 | fn next(&mut self) -> Option<Self::Item> { |
4121 | unsafe { |
4122 | loop { |
4123 | if let Some(bit) = self.bitmask.next() { |
4124 | let index = (self.probe_seq.pos + bit) & self.bucket_mask; |
4125 | return Some(index); |
4126 | } |
4127 | if likely(self.group.match_empty().any_bit_set()) { |
4128 | return None; |
4129 | } |
4130 | self.probe_seq.move_next(self.bucket_mask); |
4131 | |
4132 | // Can't use `RawTableInner::ctrl` here as we don't have |
4133 | // an actual `RawTableInner` reference to use. |
4134 | let index = self.probe_seq.pos; |
4135 | debug_assert!(index < self.bucket_mask + 1 + Group::WIDTH); |
4136 | let group_ctrl = self.ctrl.as_ptr().add(index).cast(); |
4137 | |
4138 | self.group = Group::load(group_ctrl); |
4139 | self.bitmask = self.group.match_tag(self.tag_hash).into_iter(); |
4140 | } |
4141 | } |
4142 | } |
4143 | } |
4144 | |
4145 | pub(crate) struct RawExtractIf<'a, T, A: Allocator> { |
4146 | pub iter: RawIter<T>, |
4147 | pub table: &'a mut RawTable<T, A>, |
4148 | } |
4149 | |
4150 | impl<T, A: Allocator> RawExtractIf<'_, T, A> { |
4151 | #[cfg_attr (feature = "inline-more" , inline)] |
4152 | pub(crate) fn next<F>(&mut self, mut f: F) -> Option<T> |
4153 | where |
4154 | F: FnMut(&mut T) -> bool, |
4155 | { |
4156 | unsafe { |
4157 | for item: Bucket in &mut self.iter { |
4158 | if f(item.as_mut()) { |
4159 | return Some(self.table.remove(item).0); |
4160 | } |
4161 | } |
4162 | } |
4163 | None |
4164 | } |
4165 | } |
4166 | |
4167 | #[cfg (test)] |
4168 | mod test_map { |
4169 | use super::*; |
4170 | |
4171 | #[test ] |
4172 | fn test_minimum_capacity_for_small_types() { |
4173 | #[track_caller ] |
4174 | fn test_t<T>() { |
4175 | let raw_table: RawTable<T> = RawTable::with_capacity(1); |
4176 | let actual_buckets = raw_table.buckets(); |
4177 | let min_buckets = Group::WIDTH / core::mem::size_of::<T>(); |
4178 | assert!( |
4179 | actual_buckets >= min_buckets, |
4180 | "expected at least {min_buckets} buckets, got {actual_buckets} buckets" |
4181 | ); |
4182 | } |
4183 | |
4184 | test_t::<u8>(); |
4185 | |
4186 | // This is only "small" for some platforms, like x86_64 with SSE2, but |
4187 | // there's no harm in running it on other platforms. |
4188 | test_t::<u16>(); |
4189 | } |
4190 | |
4191 | fn rehash_in_place<T>(table: &mut RawTable<T>, hasher: impl Fn(&T) -> u64) { |
4192 | unsafe { |
4193 | table.table.rehash_in_place( |
4194 | &|table, index| hasher(table.bucket::<T>(index).as_ref()), |
4195 | mem::size_of::<T>(), |
4196 | if mem::needs_drop::<T>() { |
4197 | Some(|ptr| ptr::drop_in_place(ptr as *mut T)) |
4198 | } else { |
4199 | None |
4200 | }, |
4201 | ); |
4202 | } |
4203 | } |
4204 | |
4205 | #[test ] |
4206 | fn rehash() { |
4207 | let mut table = RawTable::new(); |
4208 | let hasher = |i: &u64| *i; |
4209 | for i in 0..100 { |
4210 | table.insert(i, i, hasher); |
4211 | } |
4212 | |
4213 | for i in 0..100 { |
4214 | unsafe { |
4215 | assert_eq!(table.find(i, |x| *x == i).map(|b| b.read()), Some(i)); |
4216 | } |
4217 | assert!(table.find(i + 100, |x| *x == i + 100).is_none()); |
4218 | } |
4219 | |
4220 | rehash_in_place(&mut table, hasher); |
4221 | |
4222 | for i in 0..100 { |
4223 | unsafe { |
4224 | assert_eq!(table.find(i, |x| *x == i).map(|b| b.read()), Some(i)); |
4225 | } |
4226 | assert!(table.find(i + 100, |x| *x == i + 100).is_none()); |
4227 | } |
4228 | } |
4229 | |
4230 | /// CHECKING THAT WE ARE NOT TRYING TO READ THE MEMORY OF |
4231 | /// AN UNINITIALIZED TABLE DURING THE DROP |
4232 | #[test ] |
4233 | fn test_drop_uninitialized() { |
4234 | use ::alloc::vec::Vec; |
4235 | |
4236 | let table = unsafe { |
4237 | // SAFETY: The `buckets` is power of two and we're not |
4238 | // trying to actually use the returned RawTable. |
4239 | RawTable::<(u64, Vec<i32>)>::new_uninitialized(Global, 8, Fallibility::Infallible) |
4240 | .unwrap() |
4241 | }; |
4242 | drop(table); |
4243 | } |
4244 | |
4245 | /// CHECKING THAT WE DON'T TRY TO DROP DATA IF THE `ITEMS` |
4246 | /// ARE ZERO, EVEN IF WE HAVE `FULL` CONTROL BYTES. |
4247 | #[test ] |
4248 | fn test_drop_zero_items() { |
4249 | use ::alloc::vec::Vec; |
4250 | unsafe { |
4251 | // SAFETY: The `buckets` is power of two and we're not |
4252 | // trying to actually use the returned RawTable. |
4253 | let mut table = |
4254 | RawTable::<(u64, Vec<i32>)>::new_uninitialized(Global, 8, Fallibility::Infallible) |
4255 | .unwrap(); |
4256 | |
4257 | // WE SIMULATE, AS IT WERE, A FULL TABLE. |
4258 | |
4259 | // SAFETY: We checked that the table is allocated and therefore the table already has |
4260 | // `self.bucket_mask + 1 + Group::WIDTH` number of control bytes (see TableLayout::calculate_layout_for) |
4261 | // so writing `table.table.num_ctrl_bytes() == bucket_mask + 1 + Group::WIDTH` bytes is safe. |
4262 | table.table.ctrl_slice().fill_empty(); |
4263 | |
4264 | // SAFETY: table.capacity() is guaranteed to be smaller than table.buckets() |
4265 | table.table.ctrl(0).write_bytes(0, table.capacity()); |
4266 | |
4267 | // Fix up the trailing control bytes. See the comments in set_ctrl |
4268 | // for the handling of tables smaller than the group width. |
4269 | if table.buckets() < Group::WIDTH { |
4270 | // SAFETY: We have `self.bucket_mask + 1 + Group::WIDTH` number of control bytes, |
4271 | // so copying `self.buckets() == self.bucket_mask + 1` bytes with offset equal to |
4272 | // `Group::WIDTH` is safe |
4273 | table |
4274 | .table |
4275 | .ctrl(0) |
4276 | .copy_to(table.table.ctrl(Group::WIDTH), table.table.buckets()); |
4277 | } else { |
4278 | // SAFETY: We have `self.bucket_mask + 1 + Group::WIDTH` number of |
4279 | // control bytes,so copying `Group::WIDTH` bytes with offset equal |
4280 | // to `self.buckets() == self.bucket_mask + 1` is safe |
4281 | table |
4282 | .table |
4283 | .ctrl(0) |
4284 | .copy_to(table.table.ctrl(table.table.buckets()), Group::WIDTH); |
4285 | } |
4286 | drop(table); |
4287 | } |
4288 | } |
4289 | |
4290 | /// CHECKING THAT WE DON'T TRY TO DROP DATA IF THE `ITEMS` |
4291 | /// ARE ZERO, EVEN IF WE HAVE `FULL` CONTROL BYTES. |
4292 | #[test ] |
4293 | fn test_catch_panic_clone_from() { |
4294 | use super::{AllocError, Allocator, Global}; |
4295 | use ::alloc::sync::Arc; |
4296 | use ::alloc::vec::Vec; |
4297 | use core::sync::atomic::{AtomicI8, Ordering}; |
4298 | use std::thread; |
4299 | |
4300 | struct MyAllocInner { |
4301 | drop_count: Arc<AtomicI8>, |
4302 | } |
4303 | |
4304 | #[derive (Clone)] |
4305 | struct MyAlloc { |
4306 | _inner: Arc<MyAllocInner>, |
4307 | } |
4308 | |
4309 | impl Drop for MyAllocInner { |
4310 | fn drop(&mut self) { |
4311 | println!("MyAlloc freed." ); |
4312 | self.drop_count.fetch_sub(1, Ordering::SeqCst); |
4313 | } |
4314 | } |
4315 | |
4316 | unsafe impl Allocator for MyAlloc { |
4317 | fn allocate(&self, layout: Layout) -> std::result::Result<NonNull<[u8]>, AllocError> { |
4318 | let g = Global; |
4319 | g.allocate(layout) |
4320 | } |
4321 | |
4322 | unsafe fn deallocate(&self, ptr: NonNull<u8>, layout: Layout) { |
4323 | let g = Global; |
4324 | g.deallocate(ptr, layout) |
4325 | } |
4326 | } |
4327 | |
4328 | const DISARMED: bool = false; |
4329 | const ARMED: bool = true; |
4330 | |
4331 | struct CheckedCloneDrop { |
4332 | panic_in_clone: bool, |
4333 | dropped: bool, |
4334 | need_drop: Vec<u64>, |
4335 | } |
4336 | |
4337 | impl Clone for CheckedCloneDrop { |
4338 | fn clone(&self) -> Self { |
4339 | if self.panic_in_clone { |
4340 | panic!("panic in clone" ) |
4341 | } |
4342 | Self { |
4343 | panic_in_clone: self.panic_in_clone, |
4344 | dropped: self.dropped, |
4345 | need_drop: self.need_drop.clone(), |
4346 | } |
4347 | } |
4348 | } |
4349 | |
4350 | impl Drop for CheckedCloneDrop { |
4351 | fn drop(&mut self) { |
4352 | if self.dropped { |
4353 | panic!("double drop" ); |
4354 | } |
4355 | self.dropped = true; |
4356 | } |
4357 | } |
4358 | |
4359 | let dropped: Arc<AtomicI8> = Arc::new(AtomicI8::new(2)); |
4360 | |
4361 | let mut table = RawTable::new_in(MyAlloc { |
4362 | _inner: Arc::new(MyAllocInner { |
4363 | drop_count: dropped.clone(), |
4364 | }), |
4365 | }); |
4366 | |
4367 | for (idx, panic_in_clone) in core::iter::repeat(DISARMED).take(7).enumerate() { |
4368 | let idx = idx as u64; |
4369 | table.insert( |
4370 | idx, |
4371 | ( |
4372 | idx, |
4373 | CheckedCloneDrop { |
4374 | panic_in_clone, |
4375 | dropped: false, |
4376 | need_drop: vec![idx], |
4377 | }, |
4378 | ), |
4379 | |(k, _)| *k, |
4380 | ); |
4381 | } |
4382 | |
4383 | assert_eq!(table.len(), 7); |
4384 | |
4385 | thread::scope(|s| { |
4386 | let result = s.spawn(|| { |
4387 | let armed_flags = [ |
4388 | DISARMED, DISARMED, ARMED, DISARMED, DISARMED, DISARMED, DISARMED, |
4389 | ]; |
4390 | let mut scope_table = RawTable::new_in(MyAlloc { |
4391 | _inner: Arc::new(MyAllocInner { |
4392 | drop_count: dropped.clone(), |
4393 | }), |
4394 | }); |
4395 | for (idx, &panic_in_clone) in armed_flags.iter().enumerate() { |
4396 | let idx = idx as u64; |
4397 | scope_table.insert( |
4398 | idx, |
4399 | ( |
4400 | idx, |
4401 | CheckedCloneDrop { |
4402 | panic_in_clone, |
4403 | dropped: false, |
4404 | need_drop: vec![idx + 100], |
4405 | }, |
4406 | ), |
4407 | |(k, _)| *k, |
4408 | ); |
4409 | } |
4410 | table.clone_from(&scope_table); |
4411 | }); |
4412 | assert!(result.join().is_err()); |
4413 | }); |
4414 | |
4415 | // Let's check that all iterators work fine and do not return elements |
4416 | // (especially `RawIterRange`, which does not depend on the number of |
4417 | // elements in the table, but looks directly at the control bytes) |
4418 | // |
4419 | // SAFETY: We know for sure that `RawTable` will outlive |
4420 | // the returned `RawIter / RawIterRange` iterator. |
4421 | assert_eq!(table.len(), 0); |
4422 | assert_eq!(unsafe { table.iter().count() }, 0); |
4423 | assert_eq!(unsafe { table.iter().iter.count() }, 0); |
4424 | |
4425 | for idx in 0..table.buckets() { |
4426 | let idx = idx as u64; |
4427 | assert!( |
4428 | table.find(idx, |(k, _)| *k == idx).is_none(), |
4429 | "Index: {idx}" |
4430 | ); |
4431 | } |
4432 | |
4433 | // All allocator clones should already be dropped. |
4434 | assert_eq!(dropped.load(Ordering::SeqCst), 1); |
4435 | } |
4436 | } |
4437 | |