1 use crate::alloc::alloc::{handle_alloc_error, Layout};
2 use crate::scopeguard::guard;
3 use crate::TryReserveError;
4 #[cfg(feature = "nightly")]
5 use crate::UnavailableMutError;
6 use core::hint;
7 use core::iter::FusedIterator;
8 use core::marker::PhantomData;
9 use core::mem;
10 use core::mem::ManuallyDrop;
11 #[cfg(feature = "nightly")]
12 use core::mem::MaybeUninit;
13 use core::ptr::NonNull;
14 
15 cfg_if! {
16     // Use the SSE2 implementation if possible: it allows us to scan 16 buckets
17     // at once instead of 8. We don't bother with AVX since it would require
18     // runtime dispatch and wouldn't gain us much anyways: the probability of
19     // finding a match drops off drastically after the first few buckets.
20     //
21     // I attempted an implementation on ARM using NEON instructions, but it
22     // turns out that most NEON instructions have multi-cycle latency, which in
23     // the end outweighs any gains over the generic implementation.
24     if #[cfg(all(
25         target_feature = "sse2",
26         any(target_arch = "x86", target_arch = "x86_64"),
27         not(miri)
28     ))] {
29         mod sse2;
30         use sse2 as imp;
31     } else {
32         #[path = "generic.rs"]
33         mod generic;
34         use generic as imp;
35     }
36 }
37 
38 mod alloc;
39 pub(crate) use self::alloc::{do_alloc, Allocator, Global};
40 
41 mod bitmask;
42 
43 use self::bitmask::{BitMask, BitMaskIter};
44 use self::imp::Group;
45 
46 // Branch prediction hint. This is currently only available on nightly but it
47 // consistently improves performance by 10-15%.
48 #[cfg(feature = "nightly")]
49 use core::intrinsics::{likely, unlikely};
50 
51 // On stable we can use #[cold] to get a equivalent effect: this attributes
52 // suggests that the function is unlikely to be called
53 #[cfg(not(feature = "nightly"))]
54 #[inline]
55 #[cold]
cold()56 fn cold() {}
57 
58 #[cfg(not(feature = "nightly"))]
59 #[inline]
likely(b: bool) -> bool60 fn likely(b: bool) -> bool {
61     if !b {
62         cold()
63     }
64     b
65 }
66 #[cfg(not(feature = "nightly"))]
67 #[inline]
unlikely(b: bool) -> bool68 fn unlikely(b: bool) -> bool {
69     if b {
70         cold()
71     }
72     b
73 }
74 
75 #[cfg(feature = "nightly")]
76 #[cfg_attr(feature = "inline-more", inline)]
offset_from<T>(to: *const T, from: *const T) -> usize77 unsafe fn offset_from<T>(to: *const T, from: *const T) -> usize {
78     to.offset_from(from) as usize
79 }
80 #[cfg(not(feature = "nightly"))]
81 #[cfg_attr(feature = "inline-more", inline)]
offset_from<T>(to: *const T, from: *const T) -> usize82 unsafe fn offset_from<T>(to: *const T, from: *const T) -> usize {
83     (to as usize - from as usize) / mem::size_of::<T>()
84 }
85 
86 /// Whether memory allocation errors should return an error or abort.
87 #[derive(Copy, Clone)]
88 enum Fallibility {
89     Fallible,
90     Infallible,
91 }
92 
93 impl Fallibility {
94     /// Error to return on capacity overflow.
95     #[cfg_attr(feature = "inline-more", inline)]
capacity_overflow(self) -> TryReserveError96     fn capacity_overflow(self) -> TryReserveError {
97         match self {
98             Fallibility::Fallible => TryReserveError::CapacityOverflow,
99             Fallibility::Infallible => panic!("Hash table capacity overflow"),
100         }
101     }
102 
103     /// Error to return on allocation error.
104     #[cfg_attr(feature = "inline-more", inline)]
alloc_err(self, layout: Layout) -> TryReserveError105     fn alloc_err(self, layout: Layout) -> TryReserveError {
106         match self {
107             Fallibility::Fallible => TryReserveError::AllocError { layout },
108             Fallibility::Infallible => handle_alloc_error(layout),
109         }
110     }
111 }
112 
113 /// Control byte value for an empty bucket.
114 const EMPTY: u8 = 0b1111_1111;
115 
116 /// Control byte value for a deleted bucket.
117 const DELETED: u8 = 0b1000_0000;
118 
119 /// Checks whether a control byte represents a full bucket (top bit is clear).
120 #[inline]
is_full(ctrl: u8) -> bool121 fn is_full(ctrl: u8) -> bool {
122     ctrl & 0x80 == 0
123 }
124 
125 /// Checks whether a control byte represents a special value (top bit is set).
126 #[inline]
is_special(ctrl: u8) -> bool127 fn is_special(ctrl: u8) -> bool {
128     ctrl & 0x80 != 0
129 }
130 
131 /// Checks whether a special control value is EMPTY (just check 1 bit).
132 #[inline]
special_is_empty(ctrl: u8) -> bool133 fn special_is_empty(ctrl: u8) -> bool {
134     debug_assert!(is_special(ctrl));
135     ctrl & 0x01 != 0
136 }
137 
138 /// Primary hash function, used to select the initial bucket to probe from.
139 #[inline]
140 #[allow(clippy::cast_possible_truncation)]
h1(hash: u64) -> usize141 fn h1(hash: u64) -> usize {
142     // On 32-bit platforms we simply ignore the higher hash bits.
143     hash as usize
144 }
145 
146 /// Secondary hash function, saved in the low 7 bits of the control byte.
147 #[inline]
148 #[allow(clippy::cast_possible_truncation)]
h2(hash: u64) -> u8149 fn h2(hash: u64) -> u8 {
150     // Grab the top 7 bits of the hash. While the hash is normally a full 64-bit
151     // value, some hash functions (such as FxHash) produce a usize result
152     // instead, which means that the top 32 bits are 0 on 32-bit platforms.
153     let hash_len = usize::min(mem::size_of::<usize>(), mem::size_of::<u64>());
154     let top7 = hash >> (hash_len * 8 - 7);
155     (top7 & 0x7f) as u8 // truncation
156 }
157 
158 /// Probe sequence based on triangular numbers, which is guaranteed (since our
159 /// table size is a power of two) to visit every group of elements exactly once.
160 ///
161 /// A triangular probe has us jump by 1 more group every time. So first we
162 /// jump by 1 group (meaning we just continue our linear scan), then 2 groups
163 /// (skipping over 1 group), then 3 groups (skipping over 2 groups), and so on.
164 ///
165 /// Proof that the probe will visit every group in the table:
166 /// <https://fgiesen.wordpress.com/2015/02/22/triangular-numbers-mod-2n/>
167 struct ProbeSeq {
168     pos: usize,
169     stride: usize,
170 }
171 
172 impl ProbeSeq {
173     #[inline]
move_next(&mut self, bucket_mask: usize)174     fn move_next(&mut self, bucket_mask: usize) {
175         // We should have found an empty bucket by now and ended the probe.
176         debug_assert!(
177             self.stride <= bucket_mask,
178             "Went past end of probe sequence"
179         );
180 
181         self.stride += Group::WIDTH;
182         self.pos += self.stride;
183         self.pos &= bucket_mask;
184     }
185 }
186 
187 /// Returns the number of buckets needed to hold the given number of items,
188 /// taking the maximum load factor into account.
189 ///
190 /// Returns `None` if an overflow occurs.
191 // Workaround for emscripten bug emscripten-core/emscripten-fastcomp#258
192 #[cfg_attr(target_os = "emscripten", inline(never))]
193 #[cfg_attr(not(target_os = "emscripten"), inline)]
capacity_to_buckets(cap: usize) -> Option<usize>194 fn capacity_to_buckets(cap: usize) -> Option<usize> {
195     debug_assert_ne!(cap, 0);
196 
197     // For small tables we require at least 1 empty bucket so that lookups are
198     // guaranteed to terminate if an element doesn't exist in the table.
199     if cap < 8 {
200         // We don't bother with a table size of 2 buckets since that can only
201         // hold a single element. Instead we skip directly to a 4 bucket table
202         // which can hold 3 elements.
203         return Some(if cap < 4 { 4 } else { 8 });
204     }
205 
206     // Otherwise require 1/8 buckets to be empty (87.5% load)
207     //
208     // Be careful when modifying this, calculate_layout relies on the
209     // overflow check here.
210     let adjusted_cap = cap.checked_mul(8)? / 7;
211 
212     // Any overflows will have been caught by the checked_mul. Also, any
213     // rounding errors from the division above will be cleaned up by
214     // next_power_of_two (which can't overflow because of the previous divison).
215     Some(adjusted_cap.next_power_of_two())
216 }
217 
218 /// Returns the maximum effective capacity for the given bucket mask, taking
219 /// the maximum load factor into account.
220 #[inline]
bucket_mask_to_capacity(bucket_mask: usize) -> usize221 fn bucket_mask_to_capacity(bucket_mask: usize) -> usize {
222     if bucket_mask < 8 {
223         // For tables with 1/2/4/8 buckets, we always reserve one empty slot.
224         // Keep in mind that the bucket mask is one less than the bucket count.
225         bucket_mask
226     } else {
227         // For larger tables we reserve 12.5% of the slots as empty.
228         ((bucket_mask + 1) / 8) * 7
229     }
230 }
231 
232 /// Helper which allows the max calculation for ctrl_align to be statically computed for each T
233 /// while keeping the rest of `calculate_layout_for` independent of `T`
234 #[derive(Copy, Clone)]
235 struct TableLayout {
236     size: usize,
237     ctrl_align: usize,
238 }
239 
240 impl TableLayout {
241     #[inline]
new<T>() -> Self242     fn new<T>() -> Self {
243         let layout = Layout::new::<T>();
244         Self {
245             size: layout.size(),
246             ctrl_align: usize::max(layout.align(), Group::WIDTH),
247         }
248     }
249 
250     #[inline]
calculate_layout_for(self, buckets: usize) -> Option<(Layout, usize)>251     fn calculate_layout_for(self, buckets: usize) -> Option<(Layout, usize)> {
252         debug_assert!(buckets.is_power_of_two());
253 
254         let TableLayout { size, ctrl_align } = self;
255         // Manual layout calculation since Layout methods are not yet stable.
256         let ctrl_offset =
257             size.checked_mul(buckets)?.checked_add(ctrl_align - 1)? & !(ctrl_align - 1);
258         let len = ctrl_offset.checked_add(buckets + Group::WIDTH)?;
259 
260         Some((
261             unsafe { Layout::from_size_align_unchecked(len, ctrl_align) },
262             ctrl_offset,
263         ))
264     }
265 }
266 
267 /// Returns a Layout which describes the allocation required for a hash table,
268 /// and the offset of the control bytes in the allocation.
269 /// (the offset is also one past last element of buckets)
270 ///
271 /// Returns `None` if an overflow occurs.
272 #[cfg_attr(feature = "inline-more", inline)]
calculate_layout<T>(buckets: usize) -> Option<(Layout, usize)>273 fn calculate_layout<T>(buckets: usize) -> Option<(Layout, usize)> {
274     TableLayout::new::<T>().calculate_layout_for(buckets)
275 }
276 
277 /// A reference to a hash table bucket containing a `T`.
278 ///
279 /// This is usually just a pointer to the element itself. However if the element
280 /// is a ZST, then we instead track the index of the element in the table so
281 /// that `erase` works properly.
282 pub struct Bucket<T> {
283     // Actually it is pointer to next element than element itself
284     // this is needed to maintain pointer arithmetic invariants
285     // keeping direct pointer to element introduces difficulty.
286     // Using `NonNull` for variance and niche layout
287     ptr: NonNull<T>,
288 }
289 
290 // This Send impl is needed for rayon support. This is safe since Bucket is
291 // never exposed in a public API.
292 unsafe impl<T> Send for Bucket<T> {}
293 
294 impl<T> Clone for Bucket<T> {
295     #[cfg_attr(feature = "inline-more", inline)]
clone(&self) -> Self296     fn clone(&self) -> Self {
297         Self { ptr: self.ptr }
298     }
299 }
300 
301 impl<T> Bucket<T> {
302     #[cfg_attr(feature = "inline-more", inline)]
from_base_index(base: NonNull<T>, index: usize) -> Self303     unsafe fn from_base_index(base: NonNull<T>, index: usize) -> Self {
304         let ptr = if mem::size_of::<T>() == 0 {
305             // won't overflow because index must be less than length
306             (index + 1) as *mut T
307         } else {
308             base.as_ptr().sub(index)
309         };
310         Self {
311             ptr: NonNull::new_unchecked(ptr),
312         }
313     }
314     #[cfg_attr(feature = "inline-more", inline)]
to_base_index(&self, base: NonNull<T>) -> usize315     unsafe fn to_base_index(&self, base: NonNull<T>) -> usize {
316         if mem::size_of::<T>() == 0 {
317             self.ptr.as_ptr() as usize - 1
318         } else {
319             offset_from(base.as_ptr(), self.ptr.as_ptr())
320         }
321     }
322     #[cfg_attr(feature = "inline-more", inline)]
as_ptr(&self) -> *mut T323     pub fn as_ptr(&self) -> *mut T {
324         if mem::size_of::<T>() == 0 {
325             // Just return an arbitrary ZST pointer which is properly aligned
326             mem::align_of::<T>() as *mut T
327         } else {
328             unsafe { self.ptr.as_ptr().sub(1) }
329         }
330     }
331     #[cfg_attr(feature = "inline-more", inline)]
next_n(&self, offset: usize) -> Self332     unsafe fn next_n(&self, offset: usize) -> Self {
333         let ptr = if mem::size_of::<T>() == 0 {
334             (self.ptr.as_ptr() as usize + offset) as *mut T
335         } else {
336             self.ptr.as_ptr().sub(offset)
337         };
338         Self {
339             ptr: NonNull::new_unchecked(ptr),
340         }
341     }
342     #[cfg_attr(feature = "inline-more", inline)]
drop(&self)343     pub unsafe fn drop(&self) {
344         self.as_ptr().drop_in_place();
345     }
346     #[cfg_attr(feature = "inline-more", inline)]
read(&self) -> T347     pub unsafe fn read(&self) -> T {
348         self.as_ptr().read()
349     }
350     #[cfg_attr(feature = "inline-more", inline)]
write(&self, val: T)351     pub unsafe fn write(&self, val: T) {
352         self.as_ptr().write(val);
353     }
354     #[cfg_attr(feature = "inline-more", inline)]
as_ref<'a>(&self) -> &'a T355     pub unsafe fn as_ref<'a>(&self) -> &'a T {
356         &*self.as_ptr()
357     }
358     #[cfg_attr(feature = "inline-more", inline)]
as_mut<'a>(&self) -> &'a mut T359     pub unsafe fn as_mut<'a>(&self) -> &'a mut T {
360         &mut *self.as_ptr()
361     }
362     #[cfg_attr(feature = "inline-more", inline)]
copy_from_nonoverlapping(&self, other: &Self)363     pub unsafe fn copy_from_nonoverlapping(&self, other: &Self) {
364         self.as_ptr().copy_from_nonoverlapping(other.as_ptr(), 1);
365     }
366 }
367 
368 /// A raw hash table with an unsafe API.
369 pub struct RawTable<T, A: Allocator + Clone = Global> {
370     table: RawTableInner<A>,
371     // Tell dropck that we own instances of T.
372     marker: PhantomData<T>,
373 }
374 
375 /// Non-generic part of `RawTable` which allows functions to be instantiated only once regardless
376 /// of how many different key-value types are used.
377 struct RawTableInner<A> {
378     // Mask to get an index from a hash value. The value is one less than the
379     // number of buckets in the table.
380     bucket_mask: usize,
381 
382     // [Padding], T1, T2, ..., Tlast, C1, C2, ...
383     //                                ^ points here
384     ctrl: NonNull<u8>,
385 
386     // Number of elements that can be inserted before we need to grow the table
387     growth_left: usize,
388 
389     // Number of elements in the table, only really used by len()
390     items: usize,
391 
392     alloc: A,
393 }
394 
395 impl<T> RawTable<T, Global> {
396     /// Creates a new empty hash table without allocating any memory.
397     ///
398     /// In effect this returns a table with exactly 1 bucket. However we can
399     /// leave the data pointer dangling since that bucket is never written to
400     /// due to our load factor forcing us to always have at least 1 free bucket.
401     #[cfg_attr(feature = "inline-more", inline)]
new() -> Self402     pub const fn new() -> Self {
403         Self {
404             table: RawTableInner::new_in(Global),
405             marker: PhantomData,
406         }
407     }
408 
409     /// Attempts to allocate a new hash table with at least enough capacity
410     /// for inserting the given number of elements without reallocating.
411     #[cfg(feature = "raw")]
try_with_capacity(capacity: usize) -> Result<Self, TryReserveError>412     pub fn try_with_capacity(capacity: usize) -> Result<Self, TryReserveError> {
413         Self::try_with_capacity_in(capacity, Global)
414     }
415 
416     /// Allocates a new hash table with at least enough capacity for inserting
417     /// the given number of elements without reallocating.
with_capacity(capacity: usize) -> Self418     pub fn with_capacity(capacity: usize) -> Self {
419         Self::with_capacity_in(capacity, Global)
420     }
421 }
422 
423 impl<T, A: Allocator + Clone> RawTable<T, A> {
424     /// Creates a new empty hash table without allocating any memory, using the
425     /// given allocator.
426     ///
427     /// In effect this returns a table with exactly 1 bucket. However we can
428     /// leave the data pointer dangling since that bucket is never written to
429     /// due to our load factor forcing us to always have at least 1 free bucket.
430     #[cfg_attr(feature = "inline-more", inline)]
new_in(alloc: A) -> Self431     pub fn new_in(alloc: A) -> Self {
432         Self {
433             table: RawTableInner::new_in(alloc),
434             marker: PhantomData,
435         }
436     }
437 
438     /// Allocates a new hash table with the given number of buckets.
439     ///
440     /// The control bytes are left uninitialized.
441     #[cfg_attr(feature = "inline-more", inline)]
new_uninitialized( alloc: A, buckets: usize, fallibility: Fallibility, ) -> Result<Self, TryReserveError>442     unsafe fn new_uninitialized(
443         alloc: A,
444         buckets: usize,
445         fallibility: Fallibility,
446     ) -> Result<Self, TryReserveError> {
447         debug_assert!(buckets.is_power_of_two());
448 
449         Ok(Self {
450             table: RawTableInner::new_uninitialized(
451                 alloc,
452                 TableLayout::new::<T>(),
453                 buckets,
454                 fallibility,
455             )?,
456             marker: PhantomData,
457         })
458     }
459 
460     /// Attempts to allocate a new hash table with at least enough capacity
461     /// for inserting the given number of elements without reallocating.
fallible_with_capacity( alloc: A, capacity: usize, fallibility: Fallibility, ) -> Result<Self, TryReserveError>462     fn fallible_with_capacity(
463         alloc: A,
464         capacity: usize,
465         fallibility: Fallibility,
466     ) -> Result<Self, TryReserveError> {
467         Ok(Self {
468             table: RawTableInner::fallible_with_capacity(
469                 alloc,
470                 TableLayout::new::<T>(),
471                 capacity,
472                 fallibility,
473             )?,
474             marker: PhantomData,
475         })
476     }
477 
478     /// Attempts to allocate a new hash table using the given allocator, with at least enough
479     /// capacity for inserting the given number of elements without reallocating.
480     #[cfg(feature = "raw")]
try_with_capacity_in(capacity: usize, alloc: A) -> Result<Self, TryReserveError>481     pub fn try_with_capacity_in(capacity: usize, alloc: A) -> Result<Self, TryReserveError> {
482         Self::fallible_with_capacity(alloc, capacity, Fallibility::Fallible)
483     }
484 
485     /// Allocates a new hash table using the given allocator, with at least enough capacity for
486     /// inserting the given number of elements without reallocating.
with_capacity_in(capacity: usize, alloc: A) -> Self487     pub fn with_capacity_in(capacity: usize, alloc: A) -> Self {
488         // Avoid `Result::unwrap_or_else` because it bloats LLVM IR.
489         match Self::fallible_with_capacity(alloc, capacity, Fallibility::Infallible) {
490             Ok(capacity) => capacity,
491             Err(_) => unsafe { hint::unreachable_unchecked() },
492         }
493     }
494 
495     /// Deallocates the table without dropping any entries.
496     #[cfg_attr(feature = "inline-more", inline)]
free_buckets(&mut self)497     unsafe fn free_buckets(&mut self) {
498         self.table.free_buckets(TableLayout::new::<T>())
499     }
500 
501     /// Returns pointer to one past last element of data table.
502     #[cfg_attr(feature = "inline-more", inline)]
data_end(&self) -> NonNull<T>503     pub unsafe fn data_end(&self) -> NonNull<T> {
504         NonNull::new_unchecked(self.table.ctrl.as_ptr().cast())
505     }
506 
507     /// Returns pointer to start of data table.
508     #[cfg_attr(feature = "inline-more", inline)]
509     #[cfg(feature = "nightly")]
data_start(&self) -> *mut T510     pub unsafe fn data_start(&self) -> *mut T {
511         self.data_end().as_ptr().wrapping_sub(self.buckets())
512     }
513 
514     /// Returns the index of a bucket from a `Bucket`.
515     #[cfg_attr(feature = "inline-more", inline)]
bucket_index(&self, bucket: &Bucket<T>) -> usize516     pub unsafe fn bucket_index(&self, bucket: &Bucket<T>) -> usize {
517         bucket.to_base_index(self.data_end())
518     }
519 
520     /// Returns a pointer to an element in the table.
521     #[cfg_attr(feature = "inline-more", inline)]
bucket(&self, index: usize) -> Bucket<T>522     pub unsafe fn bucket(&self, index: usize) -> Bucket<T> {
523         debug_assert_ne!(self.table.bucket_mask, 0);
524         debug_assert!(index < self.buckets());
525         Bucket::from_base_index(self.data_end(), index)
526     }
527 
528     /// Erases an element from the table without dropping it.
529     #[cfg_attr(feature = "inline-more", inline)]
530     #[deprecated(since = "0.8.1", note = "use erase or remove instead")]
erase_no_drop(&mut self, item: &Bucket<T>)531     pub unsafe fn erase_no_drop(&mut self, item: &Bucket<T>) {
532         let index = self.bucket_index(item);
533         self.table.erase(index)
534     }
535 
536     /// Erases an element from the table, dropping it in place.
537     #[cfg_attr(feature = "inline-more", inline)]
538     #[allow(clippy::needless_pass_by_value)]
539     #[allow(deprecated)]
erase(&mut self, item: Bucket<T>)540     pub unsafe fn erase(&mut self, item: Bucket<T>) {
541         // Erase the element from the table first since drop might panic.
542         self.erase_no_drop(&item);
543         item.drop();
544     }
545 
546     /// Finds and erases an element from the table, dropping it in place.
547     /// Returns true if an element was found.
548     #[cfg(feature = "raw")]
549     #[cfg_attr(feature = "inline-more", inline)]
erase_entry(&mut self, hash: u64, eq: impl FnMut(&T) -> bool) -> bool550     pub fn erase_entry(&mut self, hash: u64, eq: impl FnMut(&T) -> bool) -> bool {
551         // Avoid `Option::map` because it bloats LLVM IR.
552         if let Some(bucket) = self.find(hash, eq) {
553             unsafe { self.erase(bucket) };
554             true
555         } else {
556             false
557         }
558     }
559 
560     /// Removes an element from the table, returning it.
561     #[cfg_attr(feature = "inline-more", inline)]
562     #[allow(clippy::needless_pass_by_value)]
563     #[allow(deprecated)]
remove(&mut self, item: Bucket<T>) -> T564     pub unsafe fn remove(&mut self, item: Bucket<T>) -> T {
565         self.erase_no_drop(&item);
566         item.read()
567     }
568 
569     /// Finds and removes an element from the table, returning it.
570     #[cfg_attr(feature = "inline-more", inline)]
remove_entry(&mut self, hash: u64, eq: impl FnMut(&T) -> bool) -> Option<T>571     pub fn remove_entry(&mut self, hash: u64, eq: impl FnMut(&T) -> bool) -> Option<T> {
572         // Avoid `Option::map` because it bloats LLVM IR.
573         match self.find(hash, eq) {
574             Some(bucket) => Some(unsafe { self.remove(bucket) }),
575             None => None,
576         }
577     }
578 
579     /// Marks all table buckets as empty without dropping their contents.
580     #[cfg_attr(feature = "inline-more", inline)]
clear_no_drop(&mut self)581     pub fn clear_no_drop(&mut self) {
582         self.table.clear_no_drop()
583     }
584 
585     /// Removes all elements from the table without freeing the backing memory.
586     #[cfg_attr(feature = "inline-more", inline)]
clear(&mut self)587     pub fn clear(&mut self) {
588         // Ensure that the table is reset even if one of the drops panic
589         let mut self_ = guard(self, |self_| self_.clear_no_drop());
590         unsafe {
591             self_.drop_elements();
592         }
593     }
594 
drop_elements(&mut self)595     unsafe fn drop_elements(&mut self) {
596         if mem::needs_drop::<T>() && self.len() != 0 {
597             for item in self.iter() {
598                 item.drop();
599             }
600         }
601     }
602 
603     /// Shrinks the table to fit `max(self.len(), min_size)` elements.
604     #[cfg_attr(feature = "inline-more", inline)]
shrink_to(&mut self, min_size: usize, hasher: impl Fn(&T) -> u64)605     pub fn shrink_to(&mut self, min_size: usize, hasher: impl Fn(&T) -> u64) {
606         // Calculate the minimal number of elements that we need to reserve
607         // space for.
608         let min_size = usize::max(self.table.items, min_size);
609         if min_size == 0 {
610             *self = Self::new_in(self.table.alloc.clone());
611             return;
612         }
613 
614         // Calculate the number of buckets that we need for this number of
615         // elements. If the calculation overflows then the requested bucket
616         // count must be larger than what we have right and nothing needs to be
617         // done.
618         let min_buckets = match capacity_to_buckets(min_size) {
619             Some(buckets) => buckets,
620             None => return,
621         };
622 
623         // If we have more buckets than we need, shrink the table.
624         if min_buckets < self.buckets() {
625             // Fast path if the table is empty
626             if self.table.items == 0 {
627                 *self = Self::with_capacity_in(min_size, self.table.alloc.clone())
628             } else {
629                 // Avoid `Result::unwrap_or_else` because it bloats LLVM IR.
630                 if self
631                     .resize(min_size, hasher, Fallibility::Infallible)
632                     .is_err()
633                 {
634                     unsafe { hint::unreachable_unchecked() }
635                 }
636             }
637         }
638     }
639 
640     /// Ensures that at least `additional` items can be inserted into the table
641     /// without reallocation.
642     #[cfg_attr(feature = "inline-more", inline)]
reserve(&mut self, additional: usize, hasher: impl Fn(&T) -> u64)643     pub fn reserve(&mut self, additional: usize, hasher: impl Fn(&T) -> u64) {
644         if additional > self.table.growth_left {
645             // Avoid `Result::unwrap_or_else` because it bloats LLVM IR.
646             if self
647                 .reserve_rehash(additional, hasher, Fallibility::Infallible)
648                 .is_err()
649             {
650                 unsafe { hint::unreachable_unchecked() }
651             }
652         }
653     }
654 
655     /// Tries to ensure that at least `additional` items can be inserted into
656     /// the table without reallocation.
657     #[cfg_attr(feature = "inline-more", inline)]
try_reserve( &mut self, additional: usize, hasher: impl Fn(&T) -> u64, ) -> Result<(), TryReserveError>658     pub fn try_reserve(
659         &mut self,
660         additional: usize,
661         hasher: impl Fn(&T) -> u64,
662     ) -> Result<(), TryReserveError> {
663         if additional > self.table.growth_left {
664             self.reserve_rehash(additional, hasher, Fallibility::Fallible)
665         } else {
666             Ok(())
667         }
668     }
669 
670     /// Out-of-line slow path for `reserve` and `try_reserve`.
671     #[cold]
672     #[inline(never)]
reserve_rehash( &mut self, additional: usize, hasher: impl Fn(&T) -> u64, fallibility: Fallibility, ) -> Result<(), TryReserveError>673     fn reserve_rehash(
674         &mut self,
675         additional: usize,
676         hasher: impl Fn(&T) -> u64,
677         fallibility: Fallibility,
678     ) -> Result<(), TryReserveError> {
679         // Avoid `Option::ok_or_else` because it bloats LLVM IR.
680         let new_items = match self.table.items.checked_add(additional) {
681             Some(new_items) => new_items,
682             None => return Err(fallibility.capacity_overflow()),
683         };
684         let full_capacity = bucket_mask_to_capacity(self.table.bucket_mask);
685         if new_items <= full_capacity / 2 {
686             // Rehash in-place without re-allocating if we have plenty of spare
687             // capacity that is locked up due to DELETED entries.
688             self.rehash_in_place(hasher);
689             Ok(())
690         } else {
691             // Otherwise, conservatively resize to at least the next size up
692             // to avoid churning deletes into frequent rehashes.
693             self.resize(
694                 usize::max(new_items, full_capacity + 1),
695                 hasher,
696                 fallibility,
697             )
698         }
699     }
700 
701     /// Rehashes the contents of the table in place (i.e. without changing the
702     /// allocation).
703     ///
704     /// If `hasher` panics then some the table's contents may be lost.
rehash_in_place(&mut self, hasher: impl Fn(&T) -> u64)705     fn rehash_in_place(&mut self, hasher: impl Fn(&T) -> u64) {
706         unsafe {
707             // If the hash function panics then properly clean up any elements
708             // that we haven't rehashed yet. We unfortunately can't preserve the
709             // element since we lost their hash and have no way of recovering it
710             // without risking another panic.
711             self.table.prepare_rehash_in_place();
712 
713             let mut guard = guard(&mut self.table, move |self_| {
714                 if mem::needs_drop::<T>() {
715                     for i in 0..self_.buckets() {
716                         if *self_.ctrl(i) == DELETED {
717                             self_.set_ctrl(i, EMPTY);
718                             self_.bucket::<T>(i).drop();
719                             self_.items -= 1;
720                         }
721                     }
722                 }
723                 self_.growth_left = bucket_mask_to_capacity(self_.bucket_mask) - self_.items;
724             });
725 
726             // At this point, DELETED elements are elements that we haven't
727             // rehashed yet. Find them and re-insert them at their ideal
728             // position.
729             'outer: for i in 0..guard.buckets() {
730                 if *guard.ctrl(i) != DELETED {
731                     continue;
732                 }
733 
734                 'inner: loop {
735                     // Hash the current item
736                     let item = guard.bucket(i);
737                     let hash = hasher(item.as_ref());
738 
739                     // Search for a suitable place to put it
740                     let new_i = guard.find_insert_slot(hash);
741 
742                     // Probing works by scanning through all of the control
743                     // bytes in groups, which may not be aligned to the group
744                     // size. If both the new and old position fall within the
745                     // same unaligned group, then there is no benefit in moving
746                     // it and we can just continue to the next item.
747                     if likely(guard.is_in_same_group(i, new_i, hash)) {
748                         guard.set_ctrl_h2(i, hash);
749                         continue 'outer;
750                     }
751 
752                     // We are moving the current item to a new position. Write
753                     // our H2 to the control byte of the new position.
754                     let prev_ctrl = guard.replace_ctrl_h2(new_i, hash);
755                     if prev_ctrl == EMPTY {
756                         guard.set_ctrl(i, EMPTY);
757                         // If the target slot is empty, simply move the current
758                         // element into the new slot and clear the old control
759                         // byte.
760                         guard.bucket(new_i).copy_from_nonoverlapping(&item);
761                         continue 'outer;
762                     } else {
763                         // If the target slot is occupied, swap the two elements
764                         // and then continue processing the element that we just
765                         // swapped into the old slot.
766                         debug_assert_eq!(prev_ctrl, DELETED);
767                         mem::swap(guard.bucket(new_i).as_mut(), item.as_mut());
768                         continue 'inner;
769                     }
770                 }
771             }
772 
773             guard.growth_left = bucket_mask_to_capacity(guard.bucket_mask) - guard.items;
774             mem::forget(guard);
775         }
776     }
777 
778     /// Allocates a new table of a different size and moves the contents of the
779     /// current table into it.
resize( &mut self, capacity: usize, hasher: impl Fn(&T) -> u64, fallibility: Fallibility, ) -> Result<(), TryReserveError>780     fn resize(
781         &mut self,
782         capacity: usize,
783         hasher: impl Fn(&T) -> u64,
784         fallibility: Fallibility,
785     ) -> Result<(), TryReserveError> {
786         unsafe {
787             let mut new_table =
788                 self.table
789                     .prepare_resize(TableLayout::new::<T>(), capacity, fallibility)?;
790 
791             // Copy all elements to the new table.
792             for item in self.iter() {
793                 // This may panic.
794                 let hash = hasher(item.as_ref());
795 
796                 // We can use a simpler version of insert() here since:
797                 // - there are no DELETED entries.
798                 // - we know there is enough space in the table.
799                 // - all elements are unique.
800                 let (index, _) = new_table.prepare_insert_slot(hash);
801                 new_table.bucket(index).copy_from_nonoverlapping(&item);
802             }
803 
804             // We successfully copied all elements without panicking. Now replace
805             // self with the new table. The old table will have its memory freed but
806             // the items will not be dropped (since they have been moved into the
807             // new table).
808             mem::swap(&mut self.table, &mut new_table);
809 
810             Ok(())
811         }
812     }
813 
814     /// Inserts a new element into the table, and returns its raw bucket.
815     ///
816     /// This does not check if the given element already exists in the table.
817     #[cfg_attr(feature = "inline-more", inline)]
insert(&mut self, hash: u64, value: T, hasher: impl Fn(&T) -> u64) -> Bucket<T>818     pub fn insert(&mut self, hash: u64, value: T, hasher: impl Fn(&T) -> u64) -> Bucket<T> {
819         unsafe {
820             let mut index = self.table.find_insert_slot(hash);
821 
822             // We can avoid growing the table once we have reached our load
823             // factor if we are replacing a tombstone. This works since the
824             // number of EMPTY slots does not change in this case.
825             let old_ctrl = *self.table.ctrl(index);
826             if unlikely(self.table.growth_left == 0 && special_is_empty(old_ctrl)) {
827                 self.reserve(1, hasher);
828                 index = self.table.find_insert_slot(hash);
829             }
830 
831             self.table.record_item_insert_at(index, old_ctrl, hash);
832 
833             let bucket = self.bucket(index);
834             bucket.write(value);
835             bucket
836         }
837     }
838 
839     /// Attempts to insert a new element without growing the table and return its raw bucket.
840     ///
841     /// Returns an `Err` containing the given element if inserting it would require growing the
842     /// table.
843     ///
844     /// This does not check if the given element already exists in the table.
845     #[cfg(feature = "raw")]
846     #[cfg_attr(feature = "inline-more", inline)]
try_insert_no_grow(&mut self, hash: u64, value: T) -> Result<Bucket<T>, T>847     pub fn try_insert_no_grow(&mut self, hash: u64, value: T) -> Result<Bucket<T>, T> {
848         unsafe {
849             match self.table.prepare_insert_no_grow(hash) {
850                 Ok(index) => {
851                     let bucket = self.bucket(index);
852                     bucket.write(value);
853                     Ok(bucket)
854                 }
855                 Err(()) => Err(value),
856             }
857         }
858     }
859 
860     /// Inserts a new element into the table, and returns a mutable reference to it.
861     ///
862     /// This does not check if the given element already exists in the table.
863     #[cfg_attr(feature = "inline-more", inline)]
insert_entry(&mut self, hash: u64, value: T, hasher: impl Fn(&T) -> u64) -> &mut T864     pub fn insert_entry(&mut self, hash: u64, value: T, hasher: impl Fn(&T) -> u64) -> &mut T {
865         unsafe { self.insert(hash, value, hasher).as_mut() }
866     }
867 
868     /// Inserts a new element into the table, without growing the table.
869     ///
870     /// There must be enough space in the table to insert the new element.
871     ///
872     /// This does not check if the given element already exists in the table.
873     #[cfg_attr(feature = "inline-more", inline)]
874     #[cfg(any(feature = "raw", feature = "rustc-internal-api"))]
insert_no_grow(&mut self, hash: u64, value: T) -> Bucket<T>875     pub fn insert_no_grow(&mut self, hash: u64, value: T) -> Bucket<T> {
876         unsafe {
877             let (index, old_ctrl) = self.table.prepare_insert_slot(hash);
878             let bucket = self.table.bucket(index);
879 
880             // If we are replacing a DELETED entry then we don't need to update
881             // the load counter.
882             self.table.growth_left -= special_is_empty(old_ctrl) as usize;
883 
884             bucket.write(value);
885             self.table.items += 1;
886             bucket
887         }
888     }
889 
890     /// Temporary removes a bucket, applying the given function to the removed
891     /// element and optionally put back the returned value in the same bucket.
892     ///
893     /// Returns `true` if the bucket still contains an element
894     ///
895     /// This does not check if the given bucket is actually occupied.
896     #[cfg_attr(feature = "inline-more", inline)]
replace_bucket_with<F>(&mut self, bucket: Bucket<T>, f: F) -> bool where F: FnOnce(T) -> Option<T>,897     pub unsafe fn replace_bucket_with<F>(&mut self, bucket: Bucket<T>, f: F) -> bool
898     where
899         F: FnOnce(T) -> Option<T>,
900     {
901         let index = self.bucket_index(&bucket);
902         let old_ctrl = *self.table.ctrl(index);
903         debug_assert!(is_full(old_ctrl));
904         let old_growth_left = self.table.growth_left;
905         let item = self.remove(bucket);
906         if let Some(new_item) = f(item) {
907             self.table.growth_left = old_growth_left;
908             self.table.set_ctrl(index, old_ctrl);
909             self.table.items += 1;
910             self.bucket(index).write(new_item);
911             true
912         } else {
913             false
914         }
915     }
916 
917     /// Searches for an element in the table.
918     #[inline]
find(&self, hash: u64, mut eq: impl FnMut(&T) -> bool) -> Option<Bucket<T>>919     pub fn find(&self, hash: u64, mut eq: impl FnMut(&T) -> bool) -> Option<Bucket<T>> {
920         unsafe {
921             for bucket in self.iter_hash(hash) {
922                 let elm = bucket.as_ref();
923                 if likely(eq(elm)) {
924                     return Some(bucket);
925                 }
926             }
927             None
928         }
929     }
930 
931     /// Gets a reference to an element in the table.
932     #[inline]
get(&self, hash: u64, eq: impl FnMut(&T) -> bool) -> Option<&T>933     pub fn get(&self, hash: u64, eq: impl FnMut(&T) -> bool) -> Option<&T> {
934         // Avoid `Option::map` because it bloats LLVM IR.
935         match self.find(hash, eq) {
936             Some(bucket) => Some(unsafe { bucket.as_ref() }),
937             None => None,
938         }
939     }
940 
941     /// Gets a mutable reference to an element in the table.
942     #[inline]
get_mut(&mut self, hash: u64, eq: impl FnMut(&T) -> bool) -> Option<&mut T>943     pub fn get_mut(&mut self, hash: u64, eq: impl FnMut(&T) -> bool) -> Option<&mut T> {
944         // Avoid `Option::map` because it bloats LLVM IR.
945         match self.find(hash, eq) {
946             Some(bucket) => Some(unsafe { bucket.as_mut() }),
947             None => None,
948         }
949     }
950 
951     /// Attempts to get mutable references to `N` entries in the table at once.
952     ///
953     /// Returns an array of length `N` with the results of each query. For soundness,
954     /// at most one mutable reference will be returned to any entry. An
955     /// `Err(UnavailableMutError::Duplicate(i))` in the returned array indicates that a suitable
956     /// entry exists, but a mutable reference to it already occurs at index `i` in the returned
957     /// array.
958     ///
959     /// The `eq` argument should be a closure such that `eq(i, k)` returns true if `k` is equal to
960     /// the `i`th key to be looked up.
961     ///
962     /// This method is available only if the `nightly` feature is enabled.
963     #[cfg(feature = "nightly")]
get_each_mut<const N: usize>( &mut self, hashes: [u64; N], mut eq: impl FnMut(usize, &T) -> bool, ) -> [Result<&'_ mut T, UnavailableMutError>; N]964     pub fn get_each_mut<const N: usize>(
965         &mut self,
966         hashes: [u64; N],
967         mut eq: impl FnMut(usize, &T) -> bool,
968     ) -> [Result<&'_ mut T, UnavailableMutError>; N] {
969         // Collect the requested buckets.
970         // TODO use `MaybeUninit::uninit_array` here instead once that's stable.
971         let mut buckets: [MaybeUninit<Option<Bucket<T>>>; N] =
972             unsafe { MaybeUninit::uninit().assume_init() };
973         for i in 0..N {
974             buckets[i] = MaybeUninit::new(self.find(hashes[i], |k| eq(i, k)));
975         }
976         let buckets: [Option<Bucket<T>>; N] = unsafe { MaybeUninit::array_assume_init(buckets) };
977 
978         // Walk through the buckets, checking for duplicates and building up the output array.
979         // TODO use `MaybeUninit::uninit_array` here instead once that's stable.
980         let mut out: [MaybeUninit<Result<&'_ mut T, UnavailableMutError>>; N] =
981             unsafe { MaybeUninit::uninit().assume_init() };
982         for i in 0..N {
983             out[i] = MaybeUninit::new(
984                 #[allow(clippy::never_loop)]
985                 'outer: loop {
986                     for j in 0..i {
987                         match (&buckets[j], &buckets[i]) {
988                             // These two buckets are the same, and we can't safely return a second
989                             // mutable reference to the same entry.
990                             (Some(prev), Some(cur)) if prev.as_ptr() == cur.as_ptr() => {
991                                 break 'outer Err(UnavailableMutError::Duplicate(j));
992                             }
993                             _ => {}
994                         }
995                     }
996                     // This bucket is distinct from all previous buckets (or it doesn't exist), so
997                     // we're clear to return the result of the lookup.
998                     break match &buckets[i] {
999                         None => Err(UnavailableMutError::Absent),
1000                         Some(bkt) => unsafe { Ok(bkt.as_mut()) },
1001                     };
1002                 },
1003             )
1004         }
1005 
1006         unsafe { MaybeUninit::array_assume_init(out) }
1007     }
1008 
1009     /// Returns the number of elements the map can hold without reallocating.
1010     ///
1011     /// This number is a lower bound; the table might be able to hold
1012     /// more, but is guaranteed to be able to hold at least this many.
1013     #[cfg_attr(feature = "inline-more", inline)]
capacity(&self) -> usize1014     pub fn capacity(&self) -> usize {
1015         self.table.items + self.table.growth_left
1016     }
1017 
1018     /// Returns the number of elements in the table.
1019     #[cfg_attr(feature = "inline-more", inline)]
len(&self) -> usize1020     pub fn len(&self) -> usize {
1021         self.table.items
1022     }
1023 
1024     /// Returns the number of buckets in the table.
1025     #[cfg_attr(feature = "inline-more", inline)]
buckets(&self) -> usize1026     pub fn buckets(&self) -> usize {
1027         self.table.bucket_mask + 1
1028     }
1029 
1030     /// Returns an iterator over every element in the table. It is up to
1031     /// the caller to ensure that the `RawTable` outlives the `RawIter`.
1032     /// Because we cannot make the `next` method unsafe on the `RawIter`
1033     /// struct, we have to make the `iter` method unsafe.
1034     #[cfg_attr(feature = "inline-more", inline)]
iter(&self) -> RawIter<T>1035     pub unsafe fn iter(&self) -> RawIter<T> {
1036         let data = Bucket::from_base_index(self.data_end(), 0);
1037         RawIter {
1038             iter: RawIterRange::new(self.table.ctrl.as_ptr(), data, self.table.buckets()),
1039             items: self.table.items,
1040         }
1041     }
1042 
1043     /// Returns an iterator over occupied buckets that could match a given hash.
1044     ///
1045     /// In rare cases, the iterator may return a bucket with a different hash.
1046     ///
1047     /// It is up to the caller to ensure that the `RawTable` outlives the
1048     /// `RawIterHash`. Because we cannot make the `next` method unsafe on the
1049     /// `RawIterHash` struct, we have to make the `iter_hash` method unsafe.
1050     #[cfg_attr(feature = "inline-more", inline)]
iter_hash(&self, hash: u64) -> RawIterHash<'_, T, A>1051     pub unsafe fn iter_hash(&self, hash: u64) -> RawIterHash<'_, T, A> {
1052         RawIterHash::new(self, hash)
1053     }
1054 
1055     /// Returns an iterator which removes all elements from the table without
1056     /// freeing the memory.
1057     #[cfg_attr(feature = "inline-more", inline)]
drain(&mut self) -> RawDrain<'_, T, A>1058     pub fn drain(&mut self) -> RawDrain<'_, T, A> {
1059         unsafe {
1060             let iter = self.iter();
1061             self.drain_iter_from(iter)
1062         }
1063     }
1064 
1065     /// Returns an iterator which removes all elements from the table without
1066     /// freeing the memory.
1067     ///
1068     /// Iteration starts at the provided iterator's current location.
1069     ///
1070     /// It is up to the caller to ensure that the iterator is valid for this
1071     /// `RawTable` and covers all items that remain in the table.
1072     #[cfg_attr(feature = "inline-more", inline)]
drain_iter_from(&mut self, iter: RawIter<T>) -> RawDrain<'_, T, A>1073     pub unsafe fn drain_iter_from(&mut self, iter: RawIter<T>) -> RawDrain<'_, T, A> {
1074         debug_assert_eq!(iter.len(), self.len());
1075         RawDrain {
1076             iter,
1077             table: ManuallyDrop::new(mem::replace(self, Self::new_in(self.table.alloc.clone()))),
1078             orig_table: NonNull::from(self),
1079             marker: PhantomData,
1080         }
1081     }
1082 
1083     /// Returns an iterator which consumes all elements from the table.
1084     ///
1085     /// Iteration starts at the provided iterator's current location.
1086     ///
1087     /// It is up to the caller to ensure that the iterator is valid for this
1088     /// `RawTable` and covers all items that remain in the table.
into_iter_from(self, iter: RawIter<T>) -> RawIntoIter<T, A>1089     pub unsafe fn into_iter_from(self, iter: RawIter<T>) -> RawIntoIter<T, A> {
1090         debug_assert_eq!(iter.len(), self.len());
1091 
1092         let alloc = self.table.alloc.clone();
1093         let allocation = self.into_allocation();
1094         RawIntoIter {
1095             iter,
1096             allocation,
1097             marker: PhantomData,
1098             alloc,
1099         }
1100     }
1101 
1102     /// Converts the table into a raw allocation. The contents of the table
1103     /// should be dropped using a `RawIter` before freeing the allocation.
1104     #[cfg_attr(feature = "inline-more", inline)]
into_allocation(self) -> Option<(NonNull<u8>, Layout)>1105     pub(crate) fn into_allocation(self) -> Option<(NonNull<u8>, Layout)> {
1106         let alloc = if self.table.is_empty_singleton() {
1107             None
1108         } else {
1109             // Avoid `Option::unwrap_or_else` because it bloats LLVM IR.
1110             let (layout, ctrl_offset) = match calculate_layout::<T>(self.table.buckets()) {
1111                 Some(lco) => lco,
1112                 None => unsafe { hint::unreachable_unchecked() },
1113             };
1114             Some((
1115                 unsafe { NonNull::new_unchecked(self.table.ctrl.as_ptr().sub(ctrl_offset)) },
1116                 layout,
1117             ))
1118         };
1119         mem::forget(self);
1120         alloc
1121     }
1122 }
1123 
1124 unsafe impl<T, A: Allocator + Clone> Send for RawTable<T, A> where T: Send {}
1125 unsafe impl<T, A: Allocator + Clone> Sync for RawTable<T, A> where T: Sync {}
1126 
1127 impl<A> RawTableInner<A> {
1128     #[cfg_attr(feature = "inline-more", inline)]
new_in(alloc: A) -> Self1129     const fn new_in(alloc: A) -> Self {
1130         Self {
1131             // Be careful to cast the entire slice to a raw pointer.
1132             ctrl: unsafe { NonNull::new_unchecked(Group::static_empty() as *const _ as *mut u8) },
1133             bucket_mask: 0,
1134             items: 0,
1135             growth_left: 0,
1136             alloc,
1137         }
1138     }
1139 }
1140 
1141 impl<A: Allocator + Clone> RawTableInner<A> {
1142     #[cfg_attr(feature = "inline-more", inline)]
new_uninitialized( alloc: A, table_layout: TableLayout, buckets: usize, fallibility: Fallibility, ) -> Result<Self, TryReserveError>1143     unsafe fn new_uninitialized(
1144         alloc: A,
1145         table_layout: TableLayout,
1146         buckets: usize,
1147         fallibility: Fallibility,
1148     ) -> Result<Self, TryReserveError> {
1149         debug_assert!(buckets.is_power_of_two());
1150 
1151         // Avoid `Option::ok_or_else` because it bloats LLVM IR.
1152         let (layout, ctrl_offset) = match table_layout.calculate_layout_for(buckets) {
1153             Some(lco) => lco,
1154             None => return Err(fallibility.capacity_overflow()),
1155         };
1156 
1157         let ptr: NonNull<u8> = match do_alloc(&alloc, layout) {
1158             Ok(block) => block.cast(),
1159             Err(_) => return Err(fallibility.alloc_err(layout)),
1160         };
1161 
1162         let ctrl = NonNull::new_unchecked(ptr.as_ptr().add(ctrl_offset));
1163         Ok(Self {
1164             ctrl,
1165             bucket_mask: buckets - 1,
1166             items: 0,
1167             growth_left: bucket_mask_to_capacity(buckets - 1),
1168             alloc,
1169         })
1170     }
1171 
1172     #[inline]
fallible_with_capacity( alloc: A, table_layout: TableLayout, capacity: usize, fallibility: Fallibility, ) -> Result<Self, TryReserveError>1173     fn fallible_with_capacity(
1174         alloc: A,
1175         table_layout: TableLayout,
1176         capacity: usize,
1177         fallibility: Fallibility,
1178     ) -> Result<Self, TryReserveError> {
1179         if capacity == 0 {
1180             Ok(Self::new_in(alloc))
1181         } else {
1182             unsafe {
1183                 let buckets =
1184                     capacity_to_buckets(capacity).ok_or_else(|| fallibility.capacity_overflow())?;
1185 
1186                 let result = Self::new_uninitialized(alloc, table_layout, buckets, fallibility)?;
1187                 result.ctrl(0).write_bytes(EMPTY, result.num_ctrl_bytes());
1188 
1189                 Ok(result)
1190             }
1191         }
1192     }
1193 
1194     /// Searches for an empty or deleted bucket which is suitable for inserting
1195     /// a new element and sets the hash for that slot.
1196     ///
1197     /// There must be at least 1 empty bucket in the table.
1198     #[inline]
prepare_insert_slot(&self, hash: u64) -> (usize, u8)1199     unsafe fn prepare_insert_slot(&self, hash: u64) -> (usize, u8) {
1200         let index = self.find_insert_slot(hash);
1201         let old_ctrl = *self.ctrl(index);
1202         self.set_ctrl_h2(index, hash);
1203         (index, old_ctrl)
1204     }
1205 
1206     /// Searches for an empty or deleted bucket which is suitable for inserting
1207     /// a new element.
1208     ///
1209     /// There must be at least 1 empty bucket in the table.
1210     #[inline]
find_insert_slot(&self, hash: u64) -> usize1211     fn find_insert_slot(&self, hash: u64) -> usize {
1212         let mut probe_seq = self.probe_seq(hash);
1213         loop {
1214             unsafe {
1215                 let group = Group::load(self.ctrl(probe_seq.pos));
1216                 if let Some(bit) = group.match_empty_or_deleted().lowest_set_bit() {
1217                     let result = (probe_seq.pos + bit) & self.bucket_mask;
1218 
1219                     // In tables smaller than the group width, trailing control
1220                     // bytes outside the range of the table are filled with
1221                     // EMPTY entries. These will unfortunately trigger a
1222                     // match, but once masked may point to a full bucket that
1223                     // is already occupied. We detect this situation here and
1224                     // perform a second scan starting at the begining of the
1225                     // table. This second scan is guaranteed to find an empty
1226                     // slot (due to the load factor) before hitting the trailing
1227                     // control bytes (containing EMPTY).
1228                     if unlikely(is_full(*self.ctrl(result))) {
1229                         debug_assert!(self.bucket_mask < Group::WIDTH);
1230                         debug_assert_ne!(probe_seq.pos, 0);
1231                         return Group::load_aligned(self.ctrl(0))
1232                             .match_empty_or_deleted()
1233                             .lowest_set_bit_nonzero();
1234                     }
1235 
1236                     return result;
1237                 }
1238             }
1239             probe_seq.move_next(self.bucket_mask);
1240         }
1241     }
1242 
1243     #[allow(clippy::mut_mut)]
1244     #[inline]
prepare_rehash_in_place(&mut self)1245     unsafe fn prepare_rehash_in_place(&mut self) {
1246         // Bulk convert all full control bytes to DELETED, and all DELETED
1247         // control bytes to EMPTY. This effectively frees up all buckets
1248         // containing a DELETED entry.
1249         for i in (0..self.buckets()).step_by(Group::WIDTH) {
1250             let group = Group::load_aligned(self.ctrl(i));
1251             let group = group.convert_special_to_empty_and_full_to_deleted();
1252             group.store_aligned(self.ctrl(i));
1253         }
1254 
1255         // Fix up the trailing control bytes. See the comments in set_ctrl
1256         // for the handling of tables smaller than the group width.
1257         if self.buckets() < Group::WIDTH {
1258             self.ctrl(0)
1259                 .copy_to(self.ctrl(Group::WIDTH), self.buckets());
1260         } else {
1261             self.ctrl(0)
1262                 .copy_to(self.ctrl(self.buckets()), Group::WIDTH);
1263         }
1264     }
1265 
1266     #[cfg_attr(feature = "inline-more", inline)]
bucket<T>(&self, index: usize) -> Bucket<T>1267     unsafe fn bucket<T>(&self, index: usize) -> Bucket<T> {
1268         debug_assert_ne!(self.bucket_mask, 0);
1269         debug_assert!(index < self.buckets());
1270         Bucket::from_base_index(self.data_end(), index)
1271     }
1272 
1273     #[cfg_attr(feature = "inline-more", inline)]
data_end<T>(&self) -> NonNull<T>1274     unsafe fn data_end<T>(&self) -> NonNull<T> {
1275         NonNull::new_unchecked(self.ctrl.as_ptr().cast())
1276     }
1277 
1278     /// Returns an iterator-like object for a probe sequence on the table.
1279     ///
1280     /// This iterator never terminates, but is guaranteed to visit each bucket
1281     /// group exactly once. The loop using `probe_seq` must terminate upon
1282     /// reaching a group containing an empty bucket.
1283     #[inline]
probe_seq(&self, hash: u64) -> ProbeSeq1284     fn probe_seq(&self, hash: u64) -> ProbeSeq {
1285         ProbeSeq {
1286             pos: h1(hash) & self.bucket_mask,
1287             stride: 0,
1288         }
1289     }
1290 
1291     /// Returns the index of a bucket for which a value must be inserted if there is enough rooom
1292     /// in the table, otherwise returns error
1293     #[cfg(feature = "raw")]
1294     #[inline]
prepare_insert_no_grow(&mut self, hash: u64) -> Result<usize, ()>1295     unsafe fn prepare_insert_no_grow(&mut self, hash: u64) -> Result<usize, ()> {
1296         let index = self.find_insert_slot(hash);
1297         let old_ctrl = *self.ctrl(index);
1298         if unlikely(self.growth_left == 0 && special_is_empty(old_ctrl)) {
1299             Err(())
1300         } else {
1301             self.record_item_insert_at(index, old_ctrl, hash);
1302             Ok(index)
1303         }
1304     }
1305 
1306     #[inline]
record_item_insert_at(&mut self, index: usize, old_ctrl: u8, hash: u64)1307     unsafe fn record_item_insert_at(&mut self, index: usize, old_ctrl: u8, hash: u64) {
1308         self.growth_left -= special_is_empty(old_ctrl) as usize;
1309         self.set_ctrl_h2(index, hash);
1310         self.items += 1;
1311     }
1312 
1313     #[inline]
is_in_same_group(&self, i: usize, new_i: usize, hash: u64) -> bool1314     fn is_in_same_group(&self, i: usize, new_i: usize, hash: u64) -> bool {
1315         let probe_seq_pos = self.probe_seq(hash).pos;
1316         let probe_index =
1317             |pos: usize| (pos.wrapping_sub(probe_seq_pos) & self.bucket_mask) / Group::WIDTH;
1318         probe_index(i) == probe_index(new_i)
1319     }
1320 
1321     /// Sets a control byte to the hash, and possibly also the replicated control byte at
1322     /// the end of the array.
1323     #[inline]
set_ctrl_h2(&self, index: usize, hash: u64)1324     unsafe fn set_ctrl_h2(&self, index: usize, hash: u64) {
1325         self.set_ctrl(index, h2(hash))
1326     }
1327 
1328     #[inline]
replace_ctrl_h2(&self, index: usize, hash: u64) -> u81329     unsafe fn replace_ctrl_h2(&self, index: usize, hash: u64) -> u8 {
1330         let prev_ctrl = *self.ctrl(index);
1331         self.set_ctrl_h2(index, hash);
1332         prev_ctrl
1333     }
1334 
1335     /// Sets a control byte, and possibly also the replicated control byte at
1336     /// the end of the array.
1337     #[inline]
set_ctrl(&self, index: usize, ctrl: u8)1338     unsafe fn set_ctrl(&self, index: usize, ctrl: u8) {
1339         // Replicate the first Group::WIDTH control bytes at the end of
1340         // the array without using a branch:
1341         // - If index >= Group::WIDTH then index == index2.
1342         // - Otherwise index2 == self.bucket_mask + 1 + index.
1343         //
1344         // The very last replicated control byte is never actually read because
1345         // we mask the initial index for unaligned loads, but we write it
1346         // anyways because it makes the set_ctrl implementation simpler.
1347         //
1348         // If there are fewer buckets than Group::WIDTH then this code will
1349         // replicate the buckets at the end of the trailing group. For example
1350         // with 2 buckets and a group size of 4, the control bytes will look
1351         // like this:
1352         //
1353         //     Real    |             Replicated
1354         // ---------------------------------------------
1355         // | [A] | [B] | [EMPTY] | [EMPTY] | [A] | [B] |
1356         // ---------------------------------------------
1357         let index2 = ((index.wrapping_sub(Group::WIDTH)) & self.bucket_mask) + Group::WIDTH;
1358 
1359         *self.ctrl(index) = ctrl;
1360         *self.ctrl(index2) = ctrl;
1361     }
1362 
1363     /// Returns a pointer to a control byte.
1364     #[inline]
ctrl(&self, index: usize) -> *mut u81365     unsafe fn ctrl(&self, index: usize) -> *mut u8 {
1366         debug_assert!(index < self.num_ctrl_bytes());
1367         self.ctrl.as_ptr().add(index)
1368     }
1369 
1370     #[inline]
buckets(&self) -> usize1371     fn buckets(&self) -> usize {
1372         self.bucket_mask + 1
1373     }
1374 
1375     #[inline]
num_ctrl_bytes(&self) -> usize1376     fn num_ctrl_bytes(&self) -> usize {
1377         self.bucket_mask + 1 + Group::WIDTH
1378     }
1379 
1380     #[inline]
is_empty_singleton(&self) -> bool1381     fn is_empty_singleton(&self) -> bool {
1382         self.bucket_mask == 0
1383     }
1384 
1385     #[allow(clippy::mut_mut)]
1386     #[inline]
prepare_resize( &self, table_layout: TableLayout, capacity: usize, fallibility: Fallibility, ) -> Result<crate::scopeguard::ScopeGuard<Self, impl FnMut(&mut Self)>, TryReserveError>1387     unsafe fn prepare_resize(
1388         &self,
1389         table_layout: TableLayout,
1390         capacity: usize,
1391         fallibility: Fallibility,
1392     ) -> Result<crate::scopeguard::ScopeGuard<Self, impl FnMut(&mut Self)>, TryReserveError> {
1393         debug_assert!(self.items <= capacity);
1394 
1395         // Allocate and initialize the new table.
1396         let mut new_table = RawTableInner::fallible_with_capacity(
1397             self.alloc.clone(),
1398             table_layout,
1399             capacity,
1400             fallibility,
1401         )?;
1402         new_table.growth_left -= self.items;
1403         new_table.items = self.items;
1404 
1405         // The hash function may panic, in which case we simply free the new
1406         // table without dropping any elements that may have been copied into
1407         // it.
1408         //
1409         // This guard is also used to free the old table on success, see
1410         // the comment at the bottom of this function.
1411         Ok(guard(new_table, move |self_| {
1412             if !self_.is_empty_singleton() {
1413                 self_.free_buckets(table_layout);
1414             }
1415         }))
1416     }
1417 
1418     #[inline]
free_buckets(&mut self, table_layout: TableLayout)1419     unsafe fn free_buckets(&mut self, table_layout: TableLayout) {
1420         // Avoid `Option::unwrap_or_else` because it bloats LLVM IR.
1421         let (layout, ctrl_offset) = match table_layout.calculate_layout_for(self.buckets()) {
1422             Some(lco) => lco,
1423             None => hint::unreachable_unchecked(),
1424         };
1425         self.alloc.deallocate(
1426             NonNull::new_unchecked(self.ctrl.as_ptr().sub(ctrl_offset)),
1427             layout,
1428         );
1429     }
1430 
1431     /// Marks all table buckets as empty without dropping their contents.
1432     #[inline]
clear_no_drop(&mut self)1433     fn clear_no_drop(&mut self) {
1434         if !self.is_empty_singleton() {
1435             unsafe {
1436                 self.ctrl(0).write_bytes(EMPTY, self.num_ctrl_bytes());
1437             }
1438         }
1439         self.items = 0;
1440         self.growth_left = bucket_mask_to_capacity(self.bucket_mask);
1441     }
1442 
1443     #[inline]
erase(&mut self, index: usize)1444     unsafe fn erase(&mut self, index: usize) {
1445         debug_assert!(is_full(*self.ctrl(index)));
1446         let index_before = index.wrapping_sub(Group::WIDTH) & self.bucket_mask;
1447         let empty_before = Group::load(self.ctrl(index_before)).match_empty();
1448         let empty_after = Group::load(self.ctrl(index)).match_empty();
1449 
1450         // If we are inside a continuous block of Group::WIDTH full or deleted
1451         // cells then a probe window may have seen a full block when trying to
1452         // insert. We therefore need to keep that block non-empty so that
1453         // lookups will continue searching to the next probe window.
1454         //
1455         // Note that in this context `leading_zeros` refers to the bytes at the
1456         // end of a group, while `trailing_zeros` refers to the bytes at the
1457         // begining of a group.
1458         let ctrl = if empty_before.leading_zeros() + empty_after.trailing_zeros() >= Group::WIDTH {
1459             DELETED
1460         } else {
1461             self.growth_left += 1;
1462             EMPTY
1463         };
1464         self.set_ctrl(index, ctrl);
1465         self.items -= 1;
1466     }
1467 }
1468 
1469 impl<T: Clone, A: Allocator + Clone> Clone for RawTable<T, A> {
clone(&self) -> Self1470     fn clone(&self) -> Self {
1471         if self.table.is_empty_singleton() {
1472             Self::new_in(self.table.alloc.clone())
1473         } else {
1474             unsafe {
1475                 let mut new_table = ManuallyDrop::new(
1476                     // Avoid `Result::ok_or_else` because it bloats LLVM IR.
1477                     match Self::new_uninitialized(
1478                         self.table.alloc.clone(),
1479                         self.table.buckets(),
1480                         Fallibility::Infallible,
1481                     ) {
1482                         Ok(table) => table,
1483                         Err(_) => hint::unreachable_unchecked(),
1484                     },
1485                 );
1486 
1487                 new_table.clone_from_spec(self, |new_table| {
1488                     // We need to free the memory allocated for the new table.
1489                     new_table.free_buckets();
1490                 });
1491 
1492                 // Return the newly created table.
1493                 ManuallyDrop::into_inner(new_table)
1494             }
1495         }
1496     }
1497 
clone_from(&mut self, source: &Self)1498     fn clone_from(&mut self, source: &Self) {
1499         if source.table.is_empty_singleton() {
1500             *self = Self::new_in(self.table.alloc.clone());
1501         } else {
1502             unsafe {
1503                 // First, drop all our elements without clearing the control bytes.
1504                 self.drop_elements();
1505 
1506                 // If necessary, resize our table to match the source.
1507                 if self.buckets() != source.buckets() {
1508                     // Skip our drop by using ptr::write.
1509                     if !self.table.is_empty_singleton() {
1510                         self.free_buckets();
1511                     }
1512                     (self as *mut Self).write(
1513                         // Avoid `Result::unwrap_or_else` because it bloats LLVM IR.
1514                         match Self::new_uninitialized(
1515                             self.table.alloc.clone(),
1516                             source.buckets(),
1517                             Fallibility::Infallible,
1518                         ) {
1519                             Ok(table) => table,
1520                             Err(_) => hint::unreachable_unchecked(),
1521                         },
1522                     );
1523                 }
1524 
1525                 self.clone_from_spec(source, |self_| {
1526                     // We need to leave the table in an empty state.
1527                     self_.clear_no_drop()
1528                 });
1529             }
1530         }
1531     }
1532 }
1533 
1534 /// Specialization of `clone_from` for `Copy` types
1535 trait RawTableClone {
clone_from_spec(&mut self, source: &Self, on_panic: impl FnMut(&mut Self))1536     unsafe fn clone_from_spec(&mut self, source: &Self, on_panic: impl FnMut(&mut Self));
1537 }
1538 impl<T: Clone, A: Allocator + Clone> RawTableClone for RawTable<T, A> {
1539     #[cfg_attr(feature = "inline-more", inline)]
1540     default_fn! {
1541         unsafe fn clone_from_spec(&mut self, source: &Self, on_panic: impl FnMut(&mut Self)) {
1542             self.clone_from_impl(source, on_panic);
1543         }
1544     }
1545 }
1546 #[cfg(feature = "nightly")]
1547 impl<T: Copy, A: Allocator + Clone> RawTableClone for RawTable<T, A> {
1548     #[cfg_attr(feature = "inline-more", inline)]
clone_from_spec(&mut self, source: &Self, _on_panic: impl FnMut(&mut Self))1549     unsafe fn clone_from_spec(&mut self, source: &Self, _on_panic: impl FnMut(&mut Self)) {
1550         source
1551             .table
1552             .ctrl(0)
1553             .copy_to_nonoverlapping(self.table.ctrl(0), self.table.num_ctrl_bytes());
1554         source
1555             .data_start()
1556             .copy_to_nonoverlapping(self.data_start(), self.table.buckets());
1557 
1558         self.table.items = source.table.items;
1559         self.table.growth_left = source.table.growth_left;
1560     }
1561 }
1562 
1563 impl<T: Clone, A: Allocator + Clone> RawTable<T, A> {
1564     /// Common code for clone and clone_from. Assumes `self.buckets() == source.buckets()`.
1565     #[cfg_attr(feature = "inline-more", inline)]
clone_from_impl(&mut self, source: &Self, mut on_panic: impl FnMut(&mut Self))1566     unsafe fn clone_from_impl(&mut self, source: &Self, mut on_panic: impl FnMut(&mut Self)) {
1567         // Copy the control bytes unchanged. We do this in a single pass
1568         source
1569             .table
1570             .ctrl(0)
1571             .copy_to_nonoverlapping(self.table.ctrl(0), self.table.num_ctrl_bytes());
1572 
1573         // The cloning of elements may panic, in which case we need
1574         // to make sure we drop only the elements that have been
1575         // cloned so far.
1576         let mut guard = guard((0, &mut *self), |(index, self_)| {
1577             if mem::needs_drop::<T>() && self_.len() != 0 {
1578                 for i in 0..=*index {
1579                     if is_full(*self_.table.ctrl(i)) {
1580                         self_.bucket(i).drop();
1581                     }
1582                 }
1583             }
1584 
1585             // Depending on whether we were called from clone or clone_from, we
1586             // either need to free the memory for the destination table or just
1587             // clear the control bytes.
1588             on_panic(self_);
1589         });
1590 
1591         for from in source.iter() {
1592             let index = source.bucket_index(&from);
1593             let to = guard.1.bucket(index);
1594             to.write(from.as_ref().clone());
1595 
1596             // Update the index in case we need to unwind.
1597             guard.0 = index;
1598         }
1599 
1600         // Successfully cloned all items, no need to clean up.
1601         mem::forget(guard);
1602 
1603         self.table.items = source.table.items;
1604         self.table.growth_left = source.table.growth_left;
1605     }
1606 
1607     /// Variant of `clone_from` to use when a hasher is available.
1608     #[cfg(feature = "raw")]
clone_from_with_hasher(&mut self, source: &Self, hasher: impl Fn(&T) -> u64)1609     pub fn clone_from_with_hasher(&mut self, source: &Self, hasher: impl Fn(&T) -> u64) {
1610         // If we have enough capacity in the table, just clear it and insert
1611         // elements one by one. We don't do this if we have the same number of
1612         // buckets as the source since we can just copy the contents directly
1613         // in that case.
1614         if self.table.buckets() != source.table.buckets()
1615             && bucket_mask_to_capacity(self.table.bucket_mask) >= source.len()
1616         {
1617             self.clear();
1618 
1619             let guard_self = guard(&mut *self, |self_| {
1620                 // Clear the partially copied table if a panic occurs, otherwise
1621                 // items and growth_left will be out of sync with the contents
1622                 // of the table.
1623                 self_.clear();
1624             });
1625 
1626             unsafe {
1627                 for item in source.iter() {
1628                     // This may panic.
1629                     let item = item.as_ref().clone();
1630                     let hash = hasher(&item);
1631 
1632                     // We can use a simpler version of insert() here since:
1633                     // - there are no DELETED entries.
1634                     // - we know there is enough space in the table.
1635                     // - all elements are unique.
1636                     let (index, _) = guard_self.table.prepare_insert_slot(hash);
1637                     guard_self.bucket(index).write(item);
1638                 }
1639             }
1640 
1641             // Successfully cloned all items, no need to clean up.
1642             mem::forget(guard_self);
1643 
1644             self.table.items = source.table.items;
1645             self.table.growth_left -= source.table.items;
1646         } else {
1647             self.clone_from(source);
1648         }
1649     }
1650 }
1651 
1652 impl<T, A: Allocator + Clone + Default> Default for RawTable<T, A> {
1653     #[cfg_attr(feature = "inline-more", inline)]
default() -> Self1654     fn default() -> Self {
1655         Self::new_in(Default::default())
1656     }
1657 }
1658 
1659 #[cfg(feature = "nightly")]
1660 unsafe impl<#[may_dangle] T, A: Allocator + Clone> Drop for RawTable<T, A> {
1661     #[cfg_attr(feature = "inline-more", inline)]
drop(&mut self)1662     fn drop(&mut self) {
1663         if !self.table.is_empty_singleton() {
1664             unsafe {
1665                 self.drop_elements();
1666                 self.free_buckets();
1667             }
1668         }
1669     }
1670 }
1671 #[cfg(not(feature = "nightly"))]
1672 impl<T, A: Allocator + Clone> Drop for RawTable<T, A> {
1673     #[cfg_attr(feature = "inline-more", inline)]
drop(&mut self)1674     fn drop(&mut self) {
1675         if !self.table.is_empty_singleton() {
1676             unsafe {
1677                 self.drop_elements();
1678                 self.free_buckets();
1679             }
1680         }
1681     }
1682 }
1683 
1684 impl<T, A: Allocator + Clone> IntoIterator for RawTable<T, A> {
1685     type Item = T;
1686     type IntoIter = RawIntoIter<T, A>;
1687 
1688     #[cfg_attr(feature = "inline-more", inline)]
into_iter(self) -> RawIntoIter<T, A>1689     fn into_iter(self) -> RawIntoIter<T, A> {
1690         unsafe {
1691             let iter = self.iter();
1692             self.into_iter_from(iter)
1693         }
1694     }
1695 }
1696 
1697 /// Iterator over a sub-range of a table. Unlike `RawIter` this iterator does
1698 /// not track an item count.
1699 pub(crate) struct RawIterRange<T> {
1700     // Mask of full buckets in the current group. Bits are cleared from this
1701     // mask as each element is processed.
1702     current_group: BitMask,
1703 
1704     // Pointer to the buckets for the current group.
1705     data: Bucket<T>,
1706 
1707     // Pointer to the next group of control bytes,
1708     // Must be aligned to the group size.
1709     next_ctrl: *const u8,
1710 
1711     // Pointer one past the last control byte of this range.
1712     end: *const u8,
1713 }
1714 
1715 impl<T> RawIterRange<T> {
1716     /// Returns a `RawIterRange` covering a subset of a table.
1717     ///
1718     /// The control byte address must be aligned to the group size.
1719     #[cfg_attr(feature = "inline-more", inline)]
new(ctrl: *const u8, data: Bucket<T>, len: usize) -> Self1720     unsafe fn new(ctrl: *const u8, data: Bucket<T>, len: usize) -> Self {
1721         debug_assert_ne!(len, 0);
1722         debug_assert_eq!(ctrl as usize % Group::WIDTH, 0);
1723         let end = ctrl.add(len);
1724 
1725         // Load the first group and advance ctrl to point to the next group
1726         let current_group = Group::load_aligned(ctrl).match_full();
1727         let next_ctrl = ctrl.add(Group::WIDTH);
1728 
1729         Self {
1730             current_group,
1731             data,
1732             next_ctrl,
1733             end,
1734         }
1735     }
1736 
1737     /// Splits a `RawIterRange` into two halves.
1738     ///
1739     /// Returns `None` if the remaining range is smaller than or equal to the
1740     /// group width.
1741     #[cfg_attr(feature = "inline-more", inline)]
1742     #[cfg(feature = "rayon")]
split(mut self) -> (Self, Option<RawIterRange<T>>)1743     pub(crate) fn split(mut self) -> (Self, Option<RawIterRange<T>>) {
1744         unsafe {
1745             if self.end <= self.next_ctrl {
1746                 // Nothing to split if the group that we are current processing
1747                 // is the last one.
1748                 (self, None)
1749             } else {
1750                 // len is the remaining number of elements after the group that
1751                 // we are currently processing. It must be a multiple of the
1752                 // group size (small tables are caught by the check above).
1753                 let len = offset_from(self.end, self.next_ctrl);
1754                 debug_assert_eq!(len % Group::WIDTH, 0);
1755 
1756                 // Split the remaining elements into two halves, but round the
1757                 // midpoint down in case there is an odd number of groups
1758                 // remaining. This ensures that:
1759                 // - The tail is at least 1 group long.
1760                 // - The split is roughly even considering we still have the
1761                 //   current group to process.
1762                 let mid = (len / 2) & !(Group::WIDTH - 1);
1763 
1764                 let tail = Self::new(
1765                     self.next_ctrl.add(mid),
1766                     self.data.next_n(Group::WIDTH).next_n(mid),
1767                     len - mid,
1768                 );
1769                 debug_assert_eq!(
1770                     self.data.next_n(Group::WIDTH).next_n(mid).ptr,
1771                     tail.data.ptr
1772                 );
1773                 debug_assert_eq!(self.end, tail.end);
1774                 self.end = self.next_ctrl.add(mid);
1775                 debug_assert_eq!(self.end.add(Group::WIDTH), tail.next_ctrl);
1776                 (self, Some(tail))
1777             }
1778         }
1779     }
1780 }
1781 
1782 // We make raw iterators unconditionally Send and Sync, and let the PhantomData
1783 // in the actual iterator implementations determine the real Send/Sync bounds.
1784 unsafe impl<T> Send for RawIterRange<T> {}
1785 unsafe impl<T> Sync for RawIterRange<T> {}
1786 
1787 impl<T> Clone for RawIterRange<T> {
1788     #[cfg_attr(feature = "inline-more", inline)]
clone(&self) -> Self1789     fn clone(&self) -> Self {
1790         Self {
1791             data: self.data.clone(),
1792             next_ctrl: self.next_ctrl,
1793             current_group: self.current_group,
1794             end: self.end,
1795         }
1796     }
1797 }
1798 
1799 impl<T> Iterator for RawIterRange<T> {
1800     type Item = Bucket<T>;
1801 
1802     #[cfg_attr(feature = "inline-more", inline)]
next(&mut self) -> Option<Bucket<T>>1803     fn next(&mut self) -> Option<Bucket<T>> {
1804         unsafe {
1805             loop {
1806                 if let Some(index) = self.current_group.lowest_set_bit() {
1807                     self.current_group = self.current_group.remove_lowest_bit();
1808                     return Some(self.data.next_n(index));
1809                 }
1810 
1811                 if self.next_ctrl >= self.end {
1812                     return None;
1813                 }
1814 
1815                 // We might read past self.end up to the next group boundary,
1816                 // but this is fine because it only occurs on tables smaller
1817                 // than the group size where the trailing control bytes are all
1818                 // EMPTY. On larger tables self.end is guaranteed to be aligned
1819                 // to the group size (since tables are power-of-two sized).
1820                 self.current_group = Group::load_aligned(self.next_ctrl).match_full();
1821                 self.data = self.data.next_n(Group::WIDTH);
1822                 self.next_ctrl = self.next_ctrl.add(Group::WIDTH);
1823             }
1824         }
1825     }
1826 
1827     #[cfg_attr(feature = "inline-more", inline)]
size_hint(&self) -> (usize, Option<usize>)1828     fn size_hint(&self) -> (usize, Option<usize>) {
1829         // We don't have an item count, so just guess based on the range size.
1830         (
1831             0,
1832             Some(unsafe { offset_from(self.end, self.next_ctrl) + Group::WIDTH }),
1833         )
1834     }
1835 }
1836 
1837 impl<T> FusedIterator for RawIterRange<T> {}
1838 
1839 /// Iterator which returns a raw pointer to every full bucket in the table.
1840 ///
1841 /// For maximum flexibility this iterator is not bound by a lifetime, but you
1842 /// must observe several rules when using it:
1843 /// - You must not free the hash table while iterating (including via growing/shrinking).
1844 /// - It is fine to erase a bucket that has been yielded by the iterator.
1845 /// - Erasing a bucket that has not yet been yielded by the iterator may still
1846 ///   result in the iterator yielding that bucket (unless `reflect_remove` is called).
1847 /// - It is unspecified whether an element inserted after the iterator was
1848 ///   created will be yielded by that iterator (unless `reflect_insert` is called).
1849 /// - The order in which the iterator yields bucket is unspecified and may
1850 ///   change in the future.
1851 pub struct RawIter<T> {
1852     pub(crate) iter: RawIterRange<T>,
1853     items: usize,
1854 }
1855 
1856 impl<T> RawIter<T> {
1857     /// Refresh the iterator so that it reflects a removal from the given bucket.
1858     ///
1859     /// For the iterator to remain valid, this method must be called once
1860     /// for each removed bucket before `next` is called again.
1861     ///
1862     /// This method should be called _before_ the removal is made. It is not necessary to call this
1863     /// method if you are removing an item that this iterator yielded in the past.
1864     #[cfg(feature = "raw")]
reflect_remove(&mut self, b: &Bucket<T>)1865     pub fn reflect_remove(&mut self, b: &Bucket<T>) {
1866         self.reflect_toggle_full(b, false);
1867     }
1868 
1869     /// Refresh the iterator so that it reflects an insertion into the given bucket.
1870     ///
1871     /// For the iterator to remain valid, this method must be called once
1872     /// for each insert before `next` is called again.
1873     ///
1874     /// This method does not guarantee that an insertion of a bucket witha greater
1875     /// index than the last one yielded will be reflected in the iterator.
1876     ///
1877     /// This method should be called _after_ the given insert is made.
1878     #[cfg(feature = "raw")]
reflect_insert(&mut self, b: &Bucket<T>)1879     pub fn reflect_insert(&mut self, b: &Bucket<T>) {
1880         self.reflect_toggle_full(b, true);
1881     }
1882 
1883     /// Refresh the iterator so that it reflects a change to the state of the given bucket.
1884     #[cfg(feature = "raw")]
reflect_toggle_full(&mut self, b: &Bucket<T>, is_insert: bool)1885     fn reflect_toggle_full(&mut self, b: &Bucket<T>, is_insert: bool) {
1886         unsafe {
1887             if b.as_ptr() > self.iter.data.as_ptr() {
1888                 // The iterator has already passed the bucket's group.
1889                 // So the toggle isn't relevant to this iterator.
1890                 return;
1891             }
1892 
1893             if self.iter.next_ctrl < self.iter.end
1894                 && b.as_ptr() <= self.iter.data.next_n(Group::WIDTH).as_ptr()
1895             {
1896                 // The iterator has not yet reached the bucket's group.
1897                 // We don't need to reload anything, but we do need to adjust the item count.
1898 
1899                 if cfg!(debug_assertions) {
1900                     // Double-check that the user isn't lying to us by checking the bucket state.
1901                     // To do that, we need to find its control byte. We know that self.iter.data is
1902                     // at self.iter.next_ctrl - Group::WIDTH, so we work from there:
1903                     let offset = offset_from(self.iter.data.as_ptr(), b.as_ptr());
1904                     let ctrl = self.iter.next_ctrl.sub(Group::WIDTH).add(offset);
1905                     // This method should be called _before_ a removal, or _after_ an insert,
1906                     // so in both cases the ctrl byte should indicate that the bucket is full.
1907                     assert!(is_full(*ctrl));
1908                 }
1909 
1910                 if is_insert {
1911                     self.items += 1;
1912                 } else {
1913                     self.items -= 1;
1914                 }
1915 
1916                 return;
1917             }
1918 
1919             // The iterator is at the bucket group that the toggled bucket is in.
1920             // We need to do two things:
1921             //
1922             //  - Determine if the iterator already yielded the toggled bucket.
1923             //    If it did, we're done.
1924             //  - Otherwise, update the iterator cached group so that it won't
1925             //    yield a to-be-removed bucket, or _will_ yield a to-be-added bucket.
1926             //    We'll also need ot update the item count accordingly.
1927             if let Some(index) = self.iter.current_group.lowest_set_bit() {
1928                 let next_bucket = self.iter.data.next_n(index);
1929                 if b.as_ptr() > next_bucket.as_ptr() {
1930                     // The toggled bucket is "before" the bucket the iterator would yield next. We
1931                     // therefore don't need to do anything --- the iterator has already passed the
1932                     // bucket in question.
1933                     //
1934                     // The item count must already be correct, since a removal or insert "prior" to
1935                     // the iterator's position wouldn't affect the item count.
1936                 } else {
1937                     // The removed bucket is an upcoming bucket. We need to make sure it does _not_
1938                     // get yielded, and also that it's no longer included in the item count.
1939                     //
1940                     // NOTE: We can't just reload the group here, both since that might reflect
1941                     // inserts we've already passed, and because that might inadvertently unset the
1942                     // bits for _other_ removals. If we do that, we'd have to also decrement the
1943                     // item count for those other bits that we unset. But the presumably subsequent
1944                     // call to reflect for those buckets might _also_ decrement the item count.
1945                     // Instead, we _just_ flip the bit for the particular bucket the caller asked
1946                     // us to reflect.
1947                     let our_bit = offset_from(self.iter.data.as_ptr(), b.as_ptr());
1948                     let was_full = self.iter.current_group.flip(our_bit);
1949                     debug_assert_ne!(was_full, is_insert);
1950 
1951                     if is_insert {
1952                         self.items += 1;
1953                     } else {
1954                         self.items -= 1;
1955                     }
1956 
1957                     if cfg!(debug_assertions) {
1958                         if b.as_ptr() == next_bucket.as_ptr() {
1959                             // The removed bucket should no longer be next
1960                             debug_assert_ne!(self.iter.current_group.lowest_set_bit(), Some(index));
1961                         } else {
1962                             // We should not have changed what bucket comes next.
1963                             debug_assert_eq!(self.iter.current_group.lowest_set_bit(), Some(index));
1964                         }
1965                     }
1966                 }
1967             } else {
1968                 // We must have already iterated past the removed item.
1969             }
1970         }
1971     }
1972 
drop_elements(&mut self)1973     unsafe fn drop_elements(&mut self) {
1974         if mem::needs_drop::<T>() && self.len() != 0 {
1975             for item in self {
1976                 item.drop();
1977             }
1978         }
1979     }
1980 }
1981 
1982 impl<T> Clone for RawIter<T> {
1983     #[cfg_attr(feature = "inline-more", inline)]
clone(&self) -> Self1984     fn clone(&self) -> Self {
1985         Self {
1986             iter: self.iter.clone(),
1987             items: self.items,
1988         }
1989     }
1990 }
1991 
1992 impl<T> Iterator for RawIter<T> {
1993     type Item = Bucket<T>;
1994 
1995     #[cfg_attr(feature = "inline-more", inline)]
next(&mut self) -> Option<Bucket<T>>1996     fn next(&mut self) -> Option<Bucket<T>> {
1997         if let Some(b) = self.iter.next() {
1998             self.items -= 1;
1999             Some(b)
2000         } else {
2001             // We don't check against items == 0 here to allow the
2002             // compiler to optimize away the item count entirely if the
2003             // iterator length is never queried.
2004             debug_assert_eq!(self.items, 0);
2005             None
2006         }
2007     }
2008 
2009     #[cfg_attr(feature = "inline-more", inline)]
size_hint(&self) -> (usize, Option<usize>)2010     fn size_hint(&self) -> (usize, Option<usize>) {
2011         (self.items, Some(self.items))
2012     }
2013 }
2014 
2015 impl<T> ExactSizeIterator for RawIter<T> {}
2016 impl<T> FusedIterator for RawIter<T> {}
2017 
2018 /// Iterator which consumes a table and returns elements.
2019 pub struct RawIntoIter<T, A: Allocator + Clone = Global> {
2020     iter: RawIter<T>,
2021     allocation: Option<(NonNull<u8>, Layout)>,
2022     marker: PhantomData<T>,
2023     alloc: A,
2024 }
2025 
2026 impl<T, A: Allocator + Clone> RawIntoIter<T, A> {
2027     #[cfg_attr(feature = "inline-more", inline)]
iter(&self) -> RawIter<T>2028     pub fn iter(&self) -> RawIter<T> {
2029         self.iter.clone()
2030     }
2031 }
2032 
2033 unsafe impl<T, A: Allocator + Clone> Send for RawIntoIter<T, A> where T: Send {}
2034 unsafe impl<T, A: Allocator + Clone> Sync for RawIntoIter<T, A> where T: Sync {}
2035 
2036 #[cfg(feature = "nightly")]
2037 unsafe impl<#[may_dangle] T, A: Allocator + Clone> Drop for RawIntoIter<T, A> {
2038     #[cfg_attr(feature = "inline-more", inline)]
drop(&mut self)2039     fn drop(&mut self) {
2040         unsafe {
2041             // Drop all remaining elements
2042             self.iter.drop_elements();
2043 
2044             // Free the table
2045             if let Some((ptr, layout)) = self.allocation {
2046                 self.alloc.deallocate(ptr, layout);
2047             }
2048         }
2049     }
2050 }
2051 #[cfg(not(feature = "nightly"))]
2052 impl<T, A: Allocator + Clone> Drop for RawIntoIter<T, A> {
2053     #[cfg_attr(feature = "inline-more", inline)]
drop(&mut self)2054     fn drop(&mut self) {
2055         unsafe {
2056             // Drop all remaining elements
2057             self.iter.drop_elements();
2058 
2059             // Free the table
2060             if let Some((ptr, layout)) = self.allocation {
2061                 self.alloc.deallocate(ptr, layout);
2062             }
2063         }
2064     }
2065 }
2066 
2067 impl<T, A: Allocator + Clone> Iterator for RawIntoIter<T, A> {
2068     type Item = T;
2069 
2070     #[cfg_attr(feature = "inline-more", inline)]
next(&mut self) -> Option<T>2071     fn next(&mut self) -> Option<T> {
2072         unsafe { Some(self.iter.next()?.read()) }
2073     }
2074 
2075     #[cfg_attr(feature = "inline-more", inline)]
size_hint(&self) -> (usize, Option<usize>)2076     fn size_hint(&self) -> (usize, Option<usize>) {
2077         self.iter.size_hint()
2078     }
2079 }
2080 
2081 impl<T, A: Allocator + Clone> ExactSizeIterator for RawIntoIter<T, A> {}
2082 impl<T, A: Allocator + Clone> FusedIterator for RawIntoIter<T, A> {}
2083 
2084 /// Iterator which consumes elements without freeing the table storage.
2085 pub struct RawDrain<'a, T, A: Allocator + Clone = Global> {
2086     iter: RawIter<T>,
2087 
2088     // The table is moved into the iterator for the duration of the drain. This
2089     // ensures that an empty table is left if the drain iterator is leaked
2090     // without dropping.
2091     table: ManuallyDrop<RawTable<T, A>>,
2092     orig_table: NonNull<RawTable<T, A>>,
2093 
2094     // We don't use a &'a mut RawTable<T> because we want RawDrain to be
2095     // covariant over T.
2096     marker: PhantomData<&'a RawTable<T, A>>,
2097 }
2098 
2099 impl<T, A: Allocator + Clone> RawDrain<'_, T, A> {
2100     #[cfg_attr(feature = "inline-more", inline)]
iter(&self) -> RawIter<T>2101     pub fn iter(&self) -> RawIter<T> {
2102         self.iter.clone()
2103     }
2104 }
2105 
2106 unsafe impl<T, A: Allocator + Copy> Send for RawDrain<'_, T, A> where T: Send {}
2107 unsafe impl<T, A: Allocator + Copy> Sync for RawDrain<'_, T, A> where T: Sync {}
2108 
2109 impl<T, A: Allocator + Clone> Drop for RawDrain<'_, T, A> {
2110     #[cfg_attr(feature = "inline-more", inline)]
drop(&mut self)2111     fn drop(&mut self) {
2112         unsafe {
2113             // Drop all remaining elements. Note that this may panic.
2114             self.iter.drop_elements();
2115 
2116             // Reset the contents of the table now that all elements have been
2117             // dropped.
2118             self.table.clear_no_drop();
2119 
2120             // Move the now empty table back to its original location.
2121             self.orig_table
2122                 .as_ptr()
2123                 .copy_from_nonoverlapping(&*self.table, 1);
2124         }
2125     }
2126 }
2127 
2128 impl<T, A: Allocator + Clone> Iterator for RawDrain<'_, T, A> {
2129     type Item = T;
2130 
2131     #[cfg_attr(feature = "inline-more", inline)]
next(&mut self) -> Option<T>2132     fn next(&mut self) -> Option<T> {
2133         unsafe {
2134             let item = self.iter.next()?;
2135             Some(item.read())
2136         }
2137     }
2138 
2139     #[cfg_attr(feature = "inline-more", inline)]
size_hint(&self) -> (usize, Option<usize>)2140     fn size_hint(&self) -> (usize, Option<usize>) {
2141         self.iter.size_hint()
2142     }
2143 }
2144 
2145 impl<T, A: Allocator + Clone> ExactSizeIterator for RawDrain<'_, T, A> {}
2146 impl<T, A: Allocator + Clone> FusedIterator for RawDrain<'_, T, A> {}
2147 
2148 /// Iterator over occupied buckets that could match a given hash.
2149 ///
2150 /// In rare cases, the iterator may return a bucket with a different hash.
2151 pub struct RawIterHash<'a, T, A: Allocator + Clone = Global> {
2152     inner: RawIterHashInner<'a, A>,
2153     _marker: PhantomData<T>,
2154 }
2155 
2156 struct RawIterHashInner<'a, A: Allocator + Clone> {
2157     table: &'a RawTableInner<A>,
2158 
2159     // The top 7 bits of the hash.
2160     h2_hash: u8,
2161 
2162     // The sequence of groups to probe in the search.
2163     probe_seq: ProbeSeq,
2164 
2165     group: Group,
2166 
2167     // The elements within the group with a matching h2-hash.
2168     bitmask: BitMaskIter,
2169 }
2170 
2171 impl<'a, T, A: Allocator + Clone> RawIterHash<'a, T, A> {
2172     #[cfg_attr(feature = "inline-more", inline)]
new(table: &'a RawTable<T, A>, hash: u64) -> Self2173     fn new(table: &'a RawTable<T, A>, hash: u64) -> Self {
2174         RawIterHash {
2175             inner: RawIterHashInner::new(&table.table, hash),
2176             _marker: PhantomData,
2177         }
2178     }
2179 }
2180 impl<'a, A: Allocator + Clone> RawIterHashInner<'a, A> {
2181     #[cfg_attr(feature = "inline-more", inline)]
new(table: &'a RawTableInner<A>, hash: u64) -> Self2182     fn new(table: &'a RawTableInner<A>, hash: u64) -> Self {
2183         unsafe {
2184             let h2_hash = h2(hash);
2185             let probe_seq = table.probe_seq(hash);
2186             let group = Group::load(table.ctrl(probe_seq.pos));
2187             let bitmask = group.match_byte(h2_hash).into_iter();
2188 
2189             RawIterHashInner {
2190                 table,
2191                 h2_hash,
2192                 probe_seq,
2193                 group,
2194                 bitmask,
2195             }
2196         }
2197     }
2198 }
2199 
2200 impl<'a, T, A: Allocator + Clone> Iterator for RawIterHash<'a, T, A> {
2201     type Item = Bucket<T>;
2202 
next(&mut self) -> Option<Bucket<T>>2203     fn next(&mut self) -> Option<Bucket<T>> {
2204         unsafe {
2205             match self.inner.next() {
2206                 Some(index) => Some(self.inner.table.bucket(index)),
2207                 None => None,
2208             }
2209         }
2210     }
2211 }
2212 
2213 impl<'a, A: Allocator + Clone> Iterator for RawIterHashInner<'a, A> {
2214     type Item = usize;
2215 
next(&mut self) -> Option<Self::Item>2216     fn next(&mut self) -> Option<Self::Item> {
2217         unsafe {
2218             loop {
2219                 if let Some(bit) = self.bitmask.next() {
2220                     let index = (self.probe_seq.pos + bit) & self.table.bucket_mask;
2221                     return Some(index);
2222                 }
2223                 if likely(self.group.match_empty().any_bit_set()) {
2224                     return None;
2225                 }
2226                 self.probe_seq.move_next(self.table.bucket_mask);
2227                 self.group = Group::load(self.table.ctrl(self.probe_seq.pos));
2228                 self.bitmask = self.group.match_byte(self.h2_hash).into_iter();
2229             }
2230         }
2231     }
2232 }
2233 
2234 #[cfg(test)]
2235 mod test_map {
2236     use super::*;
2237 
2238     #[test]
rehash()2239     fn rehash() {
2240         let mut table = RawTable::new();
2241         let hasher = |i: &u64| *i;
2242         for i in 0..100 {
2243             table.insert(i, i, hasher);
2244         }
2245 
2246         for i in 0..100 {
2247             unsafe {
2248                 assert_eq!(table.find(i, |x| *x == i).map(|b| b.read()), Some(i));
2249             }
2250             assert!(table.find(i + 100, |x| *x == i + 100).is_none());
2251         }
2252 
2253         table.rehash_in_place(hasher);
2254 
2255         for i in 0..100 {
2256             unsafe {
2257                 assert_eq!(table.find(i, |x| *x == i).map(|b| b.read()), Some(i));
2258             }
2259             assert!(table.find(i + 100, |x| *x == i + 100).is_none());
2260         }
2261     }
2262 }
2263