Adding upstream version 127.0.upstream/127.0

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

6 files changed, 3091 insertions, 580 deletions

diff --git a/third_party/rust/hashbrown/src/raw/alloc.rs b/third_party/rust/hashbrown/src/raw/alloc.rs
index ba09ea9de7..15299e7b09 100644
--- a/third_party/rust/hashbrown/src/raw/alloc.rs
+++ b/third_party/rust/hashbrown/src/raw/alloc.rs
@@ -1,5 +1,9 @@
 pub(crate) use self::inner::{do_alloc, Allocator, Global};
 
+// Nightly-case.
+// Use unstable `allocator_api` feature.
+// This is compatible with `allocator-api2` which can be enabled or not.
+// This is used when building for `std`.
 #[cfg(feature = "nightly")]
 mod inner {
     use crate::alloc::alloc::Layout;
@@ -7,28 +11,44 @@ mod inner {
     use core::ptr::NonNull;
 
     #[allow(clippy::map_err_ignore)]
-    pub fn do_alloc<A: Allocator>(alloc: &A, layout: Layout) -> Result<NonNull<u8>, ()> {
+    pub(crate) fn do_alloc<A: Allocator>(alloc: &A, layout: Layout) -> Result<NonNull<u8>, ()> {
         match alloc.allocate(layout) {
             Ok(ptr) => Ok(ptr.as_non_null_ptr()),
             Err(_) => Err(()),
         }
     }
+}
 
-    #[cfg(feature = "bumpalo")]
-    unsafe impl Allocator for crate::BumpWrapper<'_> {
-        #[inline]
-        fn allocate(&self, layout: Layout) -> Result<NonNull<[u8]>, core::alloc::AllocError> {
-            match self.0.try_alloc_layout(layout) {
-                Ok(ptr) => Ok(NonNull::slice_from_raw_parts(ptr, layout.size())),
-                Err(_) => Err(core::alloc::AllocError),
-            }
+// Basic non-nightly case.
+// This uses `allocator-api2` enabled by default.
+// If any crate enables "nightly" in `allocator-api2`,
+// this will be equivalent to the nightly case,
+// since `allocator_api2::alloc::Allocator` would be re-export of
+// `core::alloc::Allocator`.
+#[cfg(all(not(feature = "nightly"), feature = "allocator-api2"))]
+mod inner {
+    use crate::alloc::alloc::Layout;
+    pub use allocator_api2::alloc::{Allocator, Global};
+    use core::ptr::NonNull;
+
+    #[allow(clippy::map_err_ignore)]
+    pub(crate) fn do_alloc<A: Allocator>(alloc: &A, layout: Layout) -> Result<NonNull<u8>, ()> {
+        match alloc.allocate(layout) {
+            Ok(ptr) => Ok(ptr.cast()),
+            Err(_) => Err(()),
         }
-        #[inline]
-        unsafe fn deallocate(&self, _ptr: NonNull<u8>, _layout: Layout) {}
     }
 }
 
-#[cfg(not(feature = "nightly"))]
+// No-defaults case.
+// When building with default-features turned off and
+// neither `nightly` nor `allocator-api2` is enabled,
+// this will be used.
+// Making it impossible to use any custom allocator with collections defined
+// in this crate.
+// Any crate in build-tree can enable `allocator-api2`,
+// or `nightly` without disturbing users that don't want to use it.
+#[cfg(not(any(feature = "nightly", feature = "allocator-api2")))]
 mod inner {
     use crate::alloc::alloc::{alloc, dealloc, Layout};
     use core::ptr::NonNull;
@@ -41,6 +61,7 @@ mod inner {
 
     #[derive(Copy, Clone)]
     pub struct Global;
+
     unsafe impl Allocator for Global {
         #[inline]
         fn allocate(&self, layout: Layout) -> Result<NonNull<u8>, ()> {
@@ -51,6 +72,7 @@ mod inner {
             dealloc(ptr.as_ptr(), layout);
         }
     }
+
     impl Default for Global {
         #[inline]
         fn default() -> Self {
@@ -58,16 +80,7 @@ mod inner {
         }
     }
 
-    pub fn do_alloc<A: Allocator>(alloc: &A, layout: Layout) -> Result<NonNull<u8>, ()> {
+    pub(crate) fn do_alloc<A: Allocator>(alloc: &A, layout: Layout) -> Result<NonNull<u8>, ()> {
         alloc.allocate(layout)
     }
-
-    #[cfg(feature = "bumpalo")]
-    unsafe impl Allocator for crate::BumpWrapper<'_> {
-        #[allow(clippy::map_err_ignore)]
-        fn allocate(&self, layout: Layout) -> Result<NonNull<u8>, ()> {
-            self.0.try_alloc_layout(layout).map_err(|_| ())
-        }
-        unsafe fn deallocate(&self, _ptr: NonNull<u8>, _layout: Layout) {}
-    }
 }
diff --git a/third_party/rust/hashbrown/src/raw/bitmask.rs b/third_party/rust/hashbrown/src/raw/bitmask.rs
index 7d4f9fc387..6576b3c5c0 100644
--- a/third_party/rust/hashbrown/src/raw/bitmask.rs
+++ b/third_party/rust/hashbrown/src/raw/bitmask.rs
@@ -1,6 +1,6 @@
-use super::imp::{BitMaskWord, BITMASK_MASK, BITMASK_STRIDE};
-#[cfg(feature = "nightly")]
-use core::intrinsics;
+use super::imp::{
+    BitMaskWord, NonZeroBitMaskWord, BITMASK_ITER_MASK, BITMASK_MASK, BITMASK_STRIDE,
+};
 
 /// A bit mask which contains the result of a `Match` operation on a `Group` and
 /// allows iterating through them.
@@ -8,75 +8,55 @@ use core::intrinsics;
 /// The bit mask is arranged so that low-order bits represent lower memory
 /// addresses for group match results.
 ///
-/// For implementation reasons, the bits in the set may be sparsely packed, so
-/// that there is only one bit-per-byte used (the high bit, 7). If this is the
+/// For implementation reasons, the bits in the set may be sparsely packed with
+/// groups of 8 bits representing one element. If any of these bits are non-zero
+/// then this element is considered to true in the mask. If this is the
 /// case, `BITMASK_STRIDE` will be 8 to indicate a divide-by-8 should be
 /// performed on counts/indices to normalize this difference. `BITMASK_MASK` is
 /// similarly a mask of all the actually-used bits.
+///
+/// To iterate over a bit mask, it must be converted to a form where only 1 bit
+/// is set per element. This is done by applying `BITMASK_ITER_MASK` on the
+/// mask bits.
 #[derive(Copy, Clone)]
-pub struct BitMask(pub BitMaskWord);
+pub(crate) struct BitMask(pub(crate) BitMaskWord);
 
 #[allow(clippy::use_self)]
 impl BitMask {
     /// Returns a new `BitMask` with all bits inverted.
     #[inline]
     #[must_use]
-    pub fn invert(self) -> Self {
+    #[allow(dead_code)]
+    pub(crate) fn invert(self) -> Self {
         BitMask(self.0 ^ BITMASK_MASK)
     }
 
-    /// Flip the bit in the mask for the entry at the given index.
-    ///
-    /// Returns the bit's previous state.
-    #[inline]
-    #[allow(clippy::cast_ptr_alignment)]
-    #[cfg(feature = "raw")]
-    pub unsafe fn flip(&mut self, index: usize) -> bool {
-        // NOTE: The + BITMASK_STRIDE - 1 is to set the high bit.
-        let mask = 1 << (index * BITMASK_STRIDE + BITMASK_STRIDE - 1);
-        self.0 ^= mask;
-        // The bit was set if the bit is now 0.
-        self.0 & mask == 0
-    }
-
     /// Returns a new `BitMask` with the lowest bit removed.
     #[inline]
     #[must_use]
-    pub fn remove_lowest_bit(self) -> Self {
+    fn remove_lowest_bit(self) -> Self {
         BitMask(self.0 & (self.0 - 1))
     }
+
     /// Returns whether the `BitMask` has at least one set bit.
     #[inline]
-    pub fn any_bit_set(self) -> bool {
+    pub(crate) fn any_bit_set(self) -> bool {
         self.0 != 0
     }
 
     /// Returns the first set bit in the `BitMask`, if there is one.
     #[inline]
-    pub fn lowest_set_bit(self) -> Option<usize> {
-        if self.0 == 0 {
-            None
+    pub(crate) fn lowest_set_bit(self) -> Option<usize> {
+        if let Some(nonzero) = NonZeroBitMaskWord::new(self.0) {
+            Some(Self::nonzero_trailing_zeros(nonzero))
         } else {
-            Some(unsafe { self.lowest_set_bit_nonzero() })
+            None
         }
     }
 
-    /// Returns the first set bit in the `BitMask`, if there is one. The
-    /// bitmask must not be empty.
-    #[inline]
-    #[cfg(feature = "nightly")]
-    pub unsafe fn lowest_set_bit_nonzero(self) -> usize {
-        intrinsics::cttz_nonzero(self.0) as usize / BITMASK_STRIDE
-    }
-    #[inline]
-    #[cfg(not(feature = "nightly"))]
-    pub unsafe fn lowest_set_bit_nonzero(self) -> usize {
-        self.trailing_zeros()
-    }
-
     /// Returns the number of trailing zeroes in the `BitMask`.
     #[inline]
-    pub fn trailing_zeros(self) -> usize {
+    pub(crate) fn trailing_zeros(self) -> usize {
         // ARM doesn't have a trailing_zeroes instruction, and instead uses
         // reverse_bits (RBIT) + leading_zeroes (CLZ). However older ARM
         // versions (pre-ARMv7) don't have RBIT and need to emulate it
@@ -89,9 +69,21 @@ impl BitMask {
         }
     }
 
+    /// Same as above but takes a `NonZeroBitMaskWord`.
+    #[inline]
+    fn nonzero_trailing_zeros(nonzero: NonZeroBitMaskWord) -> usize {
+        if cfg!(target_arch = "arm") && BITMASK_STRIDE % 8 == 0 {
+            // SAFETY: A byte-swapped non-zero value is still non-zero.
+            let swapped = unsafe { NonZeroBitMaskWord::new_unchecked(nonzero.get().swap_bytes()) };
+            swapped.leading_zeros() as usize / BITMASK_STRIDE
+        } else {
+            nonzero.trailing_zeros() as usize / BITMASK_STRIDE
+        }
+    }
+
     /// Returns the number of leading zeroes in the `BitMask`.
     #[inline]
-    pub fn leading_zeros(self) -> usize {
+    pub(crate) fn leading_zeros(self) -> usize {
         self.0.leading_zeros() as usize / BITMASK_STRIDE
     }
 }
@@ -102,13 +94,32 @@ impl IntoIterator for BitMask {
 
     #[inline]
     fn into_iter(self) -> BitMaskIter {
-        BitMaskIter(self)
+        // A BitMask only requires each element (group of bits) to be non-zero.
+        // However for iteration we need each element to only contain 1 bit.
+        BitMaskIter(BitMask(self.0 & BITMASK_ITER_MASK))
     }
 }
 
 /// Iterator over the contents of a `BitMask`, returning the indices of set
 /// bits.
-pub struct BitMaskIter(BitMask);
+#[derive(Copy, Clone)]
+pub(crate) struct BitMaskIter(pub(crate) BitMask);
+
+impl BitMaskIter {
+    /// Flip the bit in the mask for the entry at the given index.
+    ///
+    /// Returns the bit's previous state.
+    #[inline]
+    #[allow(clippy::cast_ptr_alignment)]
+    #[cfg(feature = "raw")]
+    pub(crate) unsafe fn flip(&mut self, index: usize) -> bool {
+        // NOTE: The + BITMASK_STRIDE - 1 is to set the high bit.
+        let mask = 1 << (index * BITMASK_STRIDE + BITMASK_STRIDE - 1);
+        self.0 .0 ^= mask;
+        // The bit was set if the bit is now 0.
+        self.0 .0 & mask == 0
+    }
+}
 
 impl Iterator for BitMaskIter {
     type Item = usize;
diff --git a/third_party/rust/hashbrown/src/raw/generic.rs b/third_party/rust/hashbrown/src/raw/generic.rs
index b4d31e62c2..c668b0642a 100644
--- a/third_party/rust/hashbrown/src/raw/generic.rs
+++ b/third_party/rust/hashbrown/src/raw/generic.rs
@@ -5,26 +5,29 @@ use core::{mem, ptr};
 // Use the native word size as the group size. Using a 64-bit group size on
 // a 32-bit architecture will just end up being more expensive because
 // shifts and multiplies will need to be emulated.
-#[cfg(any(
-    target_pointer_width = "64",
-    target_arch = "aarch64",
-    target_arch = "x86_64",
-    target_arch = "wasm32",
-))]
-type GroupWord = u64;
-#[cfg(all(
-    target_pointer_width = "32",
-    not(target_arch = "aarch64"),
-    not(target_arch = "x86_64"),
-    not(target_arch = "wasm32"),
-))]
-type GroupWord = u32;
 
-pub type BitMaskWord = GroupWord;
-pub const BITMASK_STRIDE: usize = 8;
+cfg_if! {
+    if #[cfg(any(
+        target_pointer_width = "64",
+        target_arch = "aarch64",
+        target_arch = "x86_64",
+        target_arch = "wasm32",
+    ))] {
+        type GroupWord = u64;
+        type NonZeroGroupWord = core::num::NonZeroU64;
+    } else {
+        type GroupWord = u32;
+        type NonZeroGroupWord = core::num::NonZeroU32;
+    }
+}
+
+pub(crate) type BitMaskWord = GroupWord;
+pub(crate) type NonZeroBitMaskWord = NonZeroGroupWord;
+pub(crate) const BITMASK_STRIDE: usize = 8;
 // We only care about the highest bit of each byte for the mask.
 #[allow(clippy::cast_possible_truncation, clippy::unnecessary_cast)]
-pub const BITMASK_MASK: BitMaskWord = 0x8080_8080_8080_8080_u64 as GroupWord;
+pub(crate) const BITMASK_MASK: BitMaskWord = 0x8080_8080_8080_8080_u64 as GroupWord;
+pub(crate) const BITMASK_ITER_MASK: BitMaskWord = !0;
 
 /// Helper function to replicate a byte across a `GroupWord`.
 #[inline]
@@ -37,7 +40,7 @@ fn repeat(byte: u8) -> GroupWord {
 ///
 /// This implementation uses a word-sized integer.
 #[derive(Copy, Clone)]
-pub struct Group(GroupWord);
+pub(crate) struct Group(GroupWord);
 
 // We perform all operations in the native endianness, and convert to
 // little-endian just before creating a BitMask. The can potentially
@@ -46,14 +49,14 @@ pub struct Group(GroupWord);
 #[allow(clippy::use_self)]
 impl Group {
     /// Number of bytes in the group.
-    pub const WIDTH: usize = mem::size_of::<Self>();
+    pub(crate) const WIDTH: usize = mem::size_of::<Self>();
 
     /// Returns a full group of empty bytes, suitable for use as the initial
     /// value for an empty hash table.
     ///
     /// This is guaranteed to be aligned to the group size.
     #[inline]
-    pub const fn static_empty() -> &'static [u8; Group::WIDTH] {
+    pub(crate) const fn static_empty() -> &'static [u8; Group::WIDTH] {
         #[repr(C)]
         struct AlignedBytes {
             _align: [Group; 0],
@@ -69,7 +72,7 @@ impl Group {
     /// Loads a group of bytes starting at the given address.
     #[inline]
     #[allow(clippy::cast_ptr_alignment)] // unaligned load
-    pub unsafe fn load(ptr: *const u8) -> Self {
+    pub(crate) unsafe fn load(ptr: *const u8) -> Self {
         Group(ptr::read_unaligned(ptr.cast()))
     }
 
@@ -77,7 +80,7 @@ impl Group {
     /// aligned to `mem::align_of::<Group>()`.
     #[inline]
     #[allow(clippy::cast_ptr_alignment)]
-    pub unsafe fn load_aligned(ptr: *const u8) -> Self {
+    pub(crate) unsafe fn load_aligned(ptr: *const u8) -> Self {
         // FIXME: use align_offset once it stabilizes
         debug_assert_eq!(ptr as usize & (mem::align_of::<Self>() - 1), 0);
         Group(ptr::read(ptr.cast()))
@@ -87,7 +90,7 @@ impl Group {
     /// aligned to `mem::align_of::<Group>()`.
     #[inline]
     #[allow(clippy::cast_ptr_alignment)]
-    pub unsafe fn store_aligned(self, ptr: *mut u8) {
+    pub(crate) unsafe fn store_aligned(self, ptr: *mut u8) {
         // FIXME: use align_offset once it stabilizes
         debug_assert_eq!(ptr as usize & (mem::align_of::<Self>() - 1), 0);
         ptr::write(ptr.cast(), self.0);
@@ -104,7 +107,7 @@ impl Group {
     /// - This only happens if there is at least 1 true match.
     /// - The chance of this happening is very low (< 1% chance per byte).
     #[inline]
-    pub fn match_byte(self, byte: u8) -> BitMask {
+    pub(crate) fn match_byte(self, byte: u8) -> BitMask {
         // This algorithm is derived from
         // https://graphics.stanford.edu/~seander/bithacks.html##ValueInWord
         let cmp = self.0 ^ repeat(byte);
@@ -114,7 +117,7 @@ impl Group {
     /// Returns a `BitMask` indicating all bytes in the group which are
     /// `EMPTY`.
     #[inline]
-    pub fn match_empty(self) -> BitMask {
+    pub(crate) fn match_empty(self) -> BitMask {
         // If the high bit is set, then the byte must be either:
         // 1111_1111 (EMPTY) or 1000_0000 (DELETED).
         // So we can just check if the top two bits are 1 by ANDing them.
@@ -124,14 +127,14 @@ impl Group {
     /// Returns a `BitMask` indicating all bytes in the group which are
     /// `EMPTY` or `DELETED`.
     #[inline]
-    pub fn match_empty_or_deleted(self) -> BitMask {
+    pub(crate) fn match_empty_or_deleted(self) -> BitMask {
         // A byte is EMPTY or DELETED iff the high bit is set
         BitMask((self.0 & repeat(0x80)).to_le())
     }
 
     /// Returns a `BitMask` indicating all bytes in the group which are full.
     #[inline]
-    pub fn match_full(self) -> BitMask {
+    pub(crate) fn match_full(self) -> BitMask {
         self.match_empty_or_deleted().invert()
     }
 
@@ -140,7 +143,7 @@ impl Group {
     /// - `DELETED => EMPTY`
     /// - `FULL => DELETED`
     #[inline]
-    pub fn convert_special_to_empty_and_full_to_deleted(self) -> Self {
+    pub(crate) fn convert_special_to_empty_and_full_to_deleted(self) -> Self {
         // Map high_bit = 1 (EMPTY or DELETED) to 1111_1111
         // and high_bit = 0 (FULL) to 1000_0000
         //
diff --git a/third_party/rust/hashbrown/src/raw/mod.rs b/third_party/rust/hashbrown/src/raw/mod.rs
index 211b818a5f..c8e8e29122 100644
--- a/third_party/rust/hashbrown/src/raw/mod.rs
+++ b/third_party/rust/hashbrown/src/raw/mod.rs
@@ -4,7 +4,6 @@ use crate::TryReserveError;
 use core::iter::FusedIterator;
 use core::marker::PhantomData;
 use core::mem;
-use core::mem::ManuallyDrop;
 use core::mem::MaybeUninit;
 use core::ptr::NonNull;
 use core::{hint, ptr};
@@ -21,12 +20,21 @@ cfg_if! {
     if #[cfg(all(
         target_feature = "sse2",
         any(target_arch = "x86", target_arch = "x86_64"),
-        not(miri)
+        not(miri),
     ))] {
         mod sse2;
         use sse2 as imp;
+    } else if #[cfg(all(
+        target_arch = "aarch64",
+        target_feature = "neon",
+        // NEON intrinsics are currently broken on big-endian targets.
+        // See https://github.com/rust-lang/stdarch/issues/1484.
+        target_endian = "little",
+        not(miri),
+    ))] {
+        mod neon;
+        use neon as imp;
     } else {
-        #[path = "generic.rs"]
         mod generic;
         use generic as imp;
     }
@@ -37,36 +45,24 @@ pub(crate) use self::alloc::{do_alloc, Allocator, Global};
 
 mod bitmask;
 
-use self::bitmask::{BitMask, BitMaskIter};
+use self::bitmask::BitMaskIter;
 use self::imp::Group;
 
 // Branch prediction hint. This is currently only available on nightly but it
 // consistently improves performance by 10-15%.
+#[cfg(not(feature = "nightly"))]
+use core::convert::identity as likely;
+#[cfg(not(feature = "nightly"))]
+use core::convert::identity as unlikely;
 #[cfg(feature = "nightly")]
 use core::intrinsics::{likely, unlikely};
 
-// On stable we can use #[cold] to get a equivalent effect: this attributes
-// suggests that the function is unlikely to be called
-#[cfg(not(feature = "nightly"))]
-#[inline]
-#[cold]
-fn cold() {}
-
-#[cfg(not(feature = "nightly"))]
-#[inline]
-fn likely(b: bool) -> bool {
-    if !b {
-        cold();
-    }
-    b
-}
-#[cfg(not(feature = "nightly"))]
-#[inline]
-fn unlikely(b: bool) -> bool {
-    if b {
-        cold();
-    }
-    b
+// FIXME: use strict provenance functions once they are stable.
+// Implement it with a transmute for now.
+#[inline(always)]
+#[allow(clippy::useless_transmute)] // clippy is wrong, cast and transmute are different here
+fn invalid_mut<T>(addr: usize) -> *mut T {
+    unsafe { core::mem::transmute(addr) }
 }
 
 #[inline]
@@ -101,6 +97,13 @@ impl Fallibility {
     }
 }
 
+trait SizedTypeProperties: Sized {
+    const IS_ZERO_SIZED: bool = mem::size_of::<Self>() == 0;
+    const NEEDS_DROP: bool = mem::needs_drop::<Self>();
+}
+
+impl<T> SizedTypeProperties for T {}
+
 /// Control byte value for an empty bucket.
 const EMPTY: u8 = 0b1111_1111;
 
@@ -134,6 +137,13 @@ fn h1(hash: u64) -> usize {
     hash as usize
 }
 
+// Constant for h2 function that grabing the top 7 bits of the hash.
+const MIN_HASH_LEN: usize = if mem::size_of::<usize>() < mem::size_of::<u64>() {
+    mem::size_of::<usize>()
+} else {
+    mem::size_of::<u64>()
+};
+
 /// Secondary hash function, saved in the low 7 bits of the control byte.
 #[inline]
 #[allow(clippy::cast_possible_truncation)]
@@ -141,8 +151,8 @@ fn h2(hash: u64) -> u8 {
     // Grab the top 7 bits of the hash. While the hash is normally a full 64-bit
     // value, some hash functions (such as FxHash) produce a usize result
     // instead, which means that the top 32 bits are 0 on 32-bit platforms.
-    let hash_len = usize::min(mem::size_of::<usize>(), mem::size_of::<u64>());
-    let top7 = hash >> (hash_len * 8 - 7);
+    // So we use MIN_HASH_LEN constant to handle this.
+    let top7 = hash >> (MIN_HASH_LEN * 8 - 7);
     (top7 & 0x7f) as u8 // truncation
 }
 
@@ -230,11 +240,15 @@ struct TableLayout {
 
 impl TableLayout {
     #[inline]
-    fn new<T>() -> Self {
+    const fn new<T>() -> Self {
         let layout = Layout::new::<T>();
         Self {
             size: layout.size(),
-            ctrl_align: usize::max(layout.align(), Group::WIDTH),
+            ctrl_align: if layout.align() > Group::WIDTH {
+                layout.align()
+            } else {
+                Group::WIDTH
+            },
         }
     }
 
@@ -248,6 +262,12 @@ impl TableLayout {
             size.checked_mul(buckets)?.checked_add(ctrl_align - 1)? & !(ctrl_align - 1);
         let len = ctrl_offset.checked_add(buckets + Group::WIDTH)?;
 
+        // We need an additional check to ensure that the allocation doesn't
+        // exceed `isize::MAX` (https://github.com/rust-lang/rust/pull/95295).
+        if len > isize::MAX as usize - (ctrl_align - 1) {
+            return None;
+        }
+
         Some((
             unsafe { Layout::from_size_align_unchecked(len, ctrl_align) },
             ctrl_offset,
@@ -255,14 +275,9 @@ impl TableLayout {
     }
 }
 
-/// Returns a Layout which describes the allocation required for a hash table,
-/// and the offset of the control bytes in the allocation.
-/// (the offset is also one past last element of buckets)
-///
-/// Returns `None` if an overflow occurs.
-#[cfg_attr(feature = "inline-more", inline)]
-fn calculate_layout<T>(buckets: usize) -> Option<(Layout, usize)> {
-    TableLayout::new::<T>().calculate_layout_for(buckets)
+/// A reference to an empty bucket into which an can be inserted.
+pub struct InsertSlot {
+    index: usize,
 }
 
 /// A reference to a hash table bucket containing a `T`.
@@ -290,11 +305,79 @@ impl<T> Clone for Bucket<T> {
 }
 
 impl<T> Bucket<T> {
+    /// Creates a [`Bucket`] that contain pointer to the data.
+    /// The pointer calculation is performed by calculating the
+    /// offset from given `base` pointer (convenience for
+    /// `base.as_ptr().sub(index)`).
+    ///
+    /// `index` is in units of `T`; e.g., an `index` of 3 represents a pointer
+    /// offset of `3 * size_of::<T>()` bytes.
+    ///
+    /// If the `T` is a ZST, then we instead track the index of the element
+    /// in the table so that `erase` works properly (return
+    /// `NonNull::new_unchecked((index + 1) as *mut T)`)
+    ///
+    /// # Safety
+    ///
+    /// If `mem::size_of::<T>() != 0`, then the safety rules are directly derived
+    /// from the safety rules for [`<*mut T>::sub`] method of `*mut T` and the safety
+    /// rules of [`NonNull::new_unchecked`] function.
+    ///
+    /// Thus, in order to uphold the safety contracts for the [`<*mut T>::sub`] method
+    /// and [`NonNull::new_unchecked`] function, as well as for the correct
+    /// logic of the work of this crate, the following rules are necessary and
+    /// sufficient:
+    ///
+    /// * the `base` pointer must not be `dangling` and must points to the
+    ///   end of the first `value element` from the `data part` of the table, i.e.
+    ///   must be the pointer that returned by [`RawTable::data_end`] or by
+    ///   [`RawTableInner::data_end<T>`];
+    ///
+    /// * `index` must not be greater than `RawTableInner.bucket_mask`, i.e.
+    ///   `index <= RawTableInner.bucket_mask` or, in other words, `(index + 1)`
+    ///   must be no greater than the number returned by the function
+    ///   [`RawTable::buckets`] or [`RawTableInner::buckets`].
+    ///
+    /// If `mem::size_of::<T>() == 0`, then the only requirement is that the
+    /// `index` must not be greater than `RawTableInner.bucket_mask`, i.e.
+    /// `index <= RawTableInner.bucket_mask` or, in other words, `(index + 1)`
+    /// must be no greater than the number returned by the function
+    /// [`RawTable::buckets`] or [`RawTableInner::buckets`].
+    ///
+    /// [`Bucket`]: crate::raw::Bucket
+    /// [`<*mut T>::sub`]: https://doc.rust-lang.org/core/primitive.pointer.html#method.sub-1
+    /// [`NonNull::new_unchecked`]: https://doc.rust-lang.org/stable/std/ptr/struct.NonNull.html#method.new_unchecked
+    /// [`RawTable::data_end`]: crate::raw::RawTable::data_end
+    /// [`RawTableInner::data_end<T>`]: RawTableInner::data_end<T>
+    /// [`RawTable::buckets`]: crate::raw::RawTable::buckets
+    /// [`RawTableInner::buckets`]: RawTableInner::buckets
     #[inline]
     unsafe fn from_base_index(base: NonNull<T>, index: usize) -> Self {
-        let ptr = if mem::size_of::<T>() == 0 {
-            // won't overflow because index must be less than length
-            (index + 1) as *mut T
+        // If mem::size_of::<T>() != 0 then return a pointer to an `element` in
+        // the data part of the table (we start counting from "0", so that
+        // in the expression T[last], the "last" index actually one less than the
+        // "buckets" number in the table, i.e. "last = RawTableInner.bucket_mask"):
+        //
+        //                   `from_base_index(base, 1).as_ptr()` returns a pointer that
+        //                   points here in the data part of the table
+        //                   (to the start of T1)
+        //                        |
+        //                        |        `base: NonNull<T>` must point here
+        //                        |         (to the end of T0 or to the start of C0)
+        //                        v         v
+        // [Padding], Tlast, ..., |T1|, T0, |C0, C1, ..., Clast
+        //                           ^
+        //                           `from_base_index(base, 1)` returns a pointer
+        //                           that points here in the data part of the table
+        //                           (to the end of T1)
+        //
+        // where: T0...Tlast - our stored data; C0...Clast - control bytes
+        // or metadata for data.
+        let ptr = if T::IS_ZERO_SIZED {
+            // won't overflow because index must be less than length (bucket_mask)
+            // and bucket_mask is guaranteed to be less than `isize::MAX`
+            // (see TableLayout::calculate_layout_for method)
+            invalid_mut(index + 1)
         } else {
             base.as_ptr().sub(index)
         };
@@ -302,27 +385,183 @@ impl<T> Bucket<T> {
             ptr: NonNull::new_unchecked(ptr),
         }
     }
+
+    /// Calculates the index of a [`Bucket`] as distance between two pointers
+    /// (convenience for `base.as_ptr().offset_from(self.ptr.as_ptr()) as usize`).
+    /// The returned value is in units of T: the distance in bytes divided by
+    /// [`core::mem::size_of::<T>()`].
+    ///
+    /// If the `T` is a ZST, then we return the index of the element in
+    /// the table so that `erase` works properly (return `self.ptr.as_ptr() as usize - 1`).
+    ///
+    /// This function is the inverse of [`from_base_index`].
+    ///
+    /// # Safety
+    ///
+    /// If `mem::size_of::<T>() != 0`, then the safety rules are directly derived
+    /// from the safety rules for [`<*const T>::offset_from`] method of `*const T`.
+    ///
+    /// Thus, in order to uphold the safety contracts for [`<*const T>::offset_from`]
+    /// method, as well as for the correct logic of the work of this crate, the
+    /// following rules are necessary and sufficient:
+    ///
+    /// * `base` contained pointer must not be `dangling` and must point to the
+    ///   end of the first `element` from the `data part` of the table, i.e.
+    ///   must be a pointer that returns by [`RawTable::data_end`] or by
+    ///   [`RawTableInner::data_end<T>`];
+    ///
+    /// * `self` also must not contain dangling pointer;
+    ///
+    /// * both `self` and `base` must be created from the same [`RawTable`]
+    ///   (or [`RawTableInner`]).
+    ///
+    /// If `mem::size_of::<T>() == 0`, this function is always safe.
+    ///
+    /// [`Bucket`]: crate::raw::Bucket
+    /// [`from_base_index`]: crate::raw::Bucket::from_base_index
+    /// [`RawTable::data_end`]: crate::raw::RawTable::data_end
+    /// [`RawTableInner::data_end<T>`]: RawTableInner::data_end<T>
+    /// [`RawTable`]: crate::raw::RawTable
+    /// [`RawTableInner`]: RawTableInner
+    /// [`<*const T>::offset_from`]: https://doc.rust-lang.org/nightly/core/primitive.pointer.html#method.offset_from
     #[inline]
     unsafe fn to_base_index(&self, base: NonNull<T>) -> usize {
-        if mem::size_of::<T>() == 0 {
+        // If mem::size_of::<T>() != 0 then return an index under which we used to store the
+        // `element` in the data part of the table (we start counting from "0", so
+        // that in the expression T[last], the "last" index actually is one less than the
+        // "buckets" number in the table, i.e. "last = RawTableInner.bucket_mask").
+        // For example for 5th element in table calculation is performed like this:
+        //
+        //                        mem::size_of::<T>()
+        //                          |
+        //                          |         `self = from_base_index(base, 5)` that returns pointer
+        //                          |         that points here in tha data part of the table
+        //                          |         (to the end of T5)
+        //                          |           |                    `base: NonNull<T>` must point here
+        //                          v           |                    (to the end of T0 or to the start of C0)
+        //                        /???\         v                      v
+        // [Padding], Tlast, ..., |T10|, ..., T5|, T4, T3, T2, T1, T0, |C0, C1, C2, C3, C4, C5, ..., C10, ..., Clast
+        //                                      \__________  __________/
+        //                                                 \/
+        //                                     `bucket.to_base_index(base)` = 5
+        //                                     (base.as_ptr() as usize - self.ptr.as_ptr() as usize) / mem::size_of::<T>()
+        //
+        // where: T0...Tlast - our stored data; C0...Clast - control bytes or metadata for data.
+        if T::IS_ZERO_SIZED {
+            // this can not be UB
             self.ptr.as_ptr() as usize - 1
         } else {
             offset_from(base.as_ptr(), self.ptr.as_ptr())
         }
     }
+
+    /// Acquires the underlying raw pointer `*mut T` to `data`.
+    ///
+    /// # Note
+    ///
+    /// If `T` is not [`Copy`], do not use `*mut T` methods that can cause calling the
+    /// destructor of `T` (for example the [`<*mut T>::drop_in_place`] method), because
+    /// for properly dropping the data we also need to clear `data` control bytes. If we
+    /// drop data, but do not clear `data control byte` it leads to double drop when
+    /// [`RawTable`] goes out of scope.
+    ///
+    /// If you modify an already initialized `value`, so [`Hash`] and [`Eq`] on the new
+    /// `T` value and its borrowed form *must* match those for the old `T` value, as the map
+    /// will not re-evaluate where the new value should go, meaning the value may become
+    /// "lost" if their location does not reflect their state.
+    ///
+    /// [`RawTable`]: crate::raw::RawTable
+    /// [`<*mut T>::drop_in_place`]: https://doc.rust-lang.org/core/primitive.pointer.html#method.drop_in_place
+    /// [`Hash`]: https://doc.rust-lang.org/core/hash/trait.Hash.html
+    /// [`Eq`]: https://doc.rust-lang.org/core/cmp/trait.Eq.html
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// # #[cfg(feature = "raw")]
+    /// # fn test() {
+    /// use core::hash::{BuildHasher, Hash};
+    /// use hashbrown::raw::{Bucket, RawTable};
+    ///
+    /// type NewHashBuilder = core::hash::BuildHasherDefault<ahash::AHasher>;
+    ///
+    /// fn make_hash<K: Hash + ?Sized, S: BuildHasher>(hash_builder: &S, key: &K) -> u64 {
+    ///     use core::hash::Hasher;
+    ///     let mut state = hash_builder.build_hasher();
+    ///     key.hash(&mut state);
+    ///     state.finish()
+    /// }
+    ///
+    /// let hash_builder = NewHashBuilder::default();
+    /// let mut table = RawTable::new();
+    ///
+    /// let value = ("a", 100);
+    /// let hash = make_hash(&hash_builder, &value.0);
+    ///
+    /// table.insert(hash, value.clone(), |val| make_hash(&hash_builder, &val.0));
+    ///
+    /// let bucket: Bucket<(&str, i32)> = table.find(hash, |(k1, _)| k1 == &value.0).unwrap();
+    ///
+    /// assert_eq!(unsafe { &*bucket.as_ptr() }, &("a", 100));
+    /// # }
+    /// # fn main() {
+    /// #     #[cfg(feature = "raw")]
+    /// #     test()
+    /// # }
+    /// ```
     #[inline]
     pub fn as_ptr(&self) -> *mut T {
-        if mem::size_of::<T>() == 0 {
+        if T::IS_ZERO_SIZED {
             // Just return an arbitrary ZST pointer which is properly aligned
-            mem::align_of::<T>() as *mut T
+            // invalid pointer is good enough for ZST
+            invalid_mut(mem::align_of::<T>())
         } else {
             unsafe { self.ptr.as_ptr().sub(1) }
         }
     }
+
+    /// Create a new [`Bucket`] that is offset from the `self` by the given
+    /// `offset`. The pointer calculation is performed by calculating the
+    /// offset from `self` pointer (convenience for `self.ptr.as_ptr().sub(offset)`).
+    /// This function is used for iterators.
+    ///
+    /// `offset` is in units of `T`; e.g., a `offset` of 3 represents a pointer
+    /// offset of `3 * size_of::<T>()` bytes.
+    ///
+    /// # Safety
+    ///
+    /// If `mem::size_of::<T>() != 0`, then the safety rules are directly derived
+    /// from the safety rules for [`<*mut T>::sub`] method of `*mut T` and safety
+    /// rules of [`NonNull::new_unchecked`] function.
+    ///
+    /// Thus, in order to uphold the safety contracts for [`<*mut T>::sub`] method
+    /// and [`NonNull::new_unchecked`] function, as well as for the correct
+    /// logic of the work of this crate, the following rules are necessary and
+    /// sufficient:
+    ///
+    /// * `self` contained pointer must not be `dangling`;
+    ///
+    /// * `self.to_base_index() + ofset` must not be greater than `RawTableInner.bucket_mask`,
+    ///   i.e. `(self.to_base_index() + ofset) <= RawTableInner.bucket_mask` or, in other
+    ///   words, `self.to_base_index() + ofset + 1` must be no greater than the number returned
+    ///   by the function [`RawTable::buckets`] or [`RawTableInner::buckets`].
+    ///
+    /// If `mem::size_of::<T>() == 0`, then the only requirement is that the
+    /// `self.to_base_index() + ofset` must not be greater than `RawTableInner.bucket_mask`,
+    /// i.e. `(self.to_base_index() + ofset) <= RawTableInner.bucket_mask` or, in other words,
+    /// `self.to_base_index() + ofset + 1` must be no greater than the number returned by the
+    /// function [`RawTable::buckets`] or [`RawTableInner::buckets`].
+    ///
+    /// [`Bucket`]: crate::raw::Bucket
+    /// [`<*mut T>::sub`]: https://doc.rust-lang.org/core/primitive.pointer.html#method.sub-1
+    /// [`NonNull::new_unchecked`]: https://doc.rust-lang.org/stable/std/ptr/struct.NonNull.html#method.new_unchecked
+    /// [`RawTable::buckets`]: crate::raw::RawTable::buckets
+    /// [`RawTableInner::buckets`]: RawTableInner::buckets
     #[inline]
     unsafe fn next_n(&self, offset: usize) -> Self {
-        let ptr = if mem::size_of::<T>() == 0 {
-            (self.ptr.as_ptr() as usize + offset) as *mut T
+        let ptr = if T::IS_ZERO_SIZED {
+            // invalid pointer is good enough for ZST
+            invalid_mut(self.ptr.as_ptr() as usize + offset)
         } else {
             self.ptr.as_ptr().sub(offset)
         };
@@ -330,26 +569,212 @@ impl<T> Bucket<T> {
             ptr: NonNull::new_unchecked(ptr),
         }
     }
+
+    /// Executes the destructor (if any) of the pointed-to `data`.
+    ///
+    /// # Safety
+    ///
+    /// See [`ptr::drop_in_place`] for safety concerns.
+    ///
+    /// You should use [`RawTable::erase`] instead of this function,
+    /// or be careful with calling this function directly, because for
+    /// properly dropping the data we need also clear `data` control bytes.
+    /// If we drop data, but do not erase `data control byte` it leads to
+    /// double drop when [`RawTable`] goes out of scope.
+    ///
+    /// [`ptr::drop_in_place`]: https://doc.rust-lang.org/core/ptr/fn.drop_in_place.html
+    /// [`RawTable`]: crate::raw::RawTable
+    /// [`RawTable::erase`]: crate::raw::RawTable::erase
     #[cfg_attr(feature = "inline-more", inline)]
-    pub unsafe fn drop(&self) {
+    pub(crate) unsafe fn drop(&self) {
         self.as_ptr().drop_in_place();
     }
+
+    /// Reads the `value` from `self` without moving it. This leaves the
+    /// memory in `self` unchanged.
+    ///
+    /// # Safety
+    ///
+    /// See [`ptr::read`] for safety concerns.
+    ///
+    /// You should use [`RawTable::remove`] instead of this function,
+    /// or be careful with calling this function directly, because compiler
+    /// calls its destructor when readed `value` goes out of scope. It
+    /// can cause double dropping when [`RawTable`] goes out of scope,
+    /// because of not erased `data control byte`.
+    ///
+    /// [`ptr::read`]: https://doc.rust-lang.org/core/ptr/fn.read.html
+    /// [`RawTable`]: crate::raw::RawTable
+    /// [`RawTable::remove`]: crate::raw::RawTable::remove
     #[inline]
-    pub unsafe fn read(&self) -> T {
+    pub(crate) unsafe fn read(&self) -> T {
         self.as_ptr().read()
     }
+
+    /// Overwrites a memory location with the given `value` without reading
+    /// or dropping the old value (like [`ptr::write`] function).
+    ///
+    /// # Safety
+    ///
+    /// See [`ptr::write`] for safety concerns.
+    ///
+    /// # Note
+    ///
+    /// [`Hash`] and [`Eq`] on the new `T` value and its borrowed form *must* match
+    /// those for the old `T` value, as the map will not re-evaluate where the new
+    /// value should go, meaning the value may become "lost" if their location
+    /// does not reflect their state.
+    ///
+    /// [`ptr::write`]: https://doc.rust-lang.org/core/ptr/fn.write.html
+    /// [`Hash`]: https://doc.rust-lang.org/core/hash/trait.Hash.html
+    /// [`Eq`]: https://doc.rust-lang.org/core/cmp/trait.Eq.html
     #[inline]
-    pub unsafe fn write(&self, val: T) {
+    pub(crate) unsafe fn write(&self, val: T) {
         self.as_ptr().write(val);
     }
+
+    /// Returns a shared immutable reference to the `value`.
+    ///
+    /// # Safety
+    ///
+    /// See [`NonNull::as_ref`] for safety concerns.
+    ///
+    /// [`NonNull::as_ref`]: https://doc.rust-lang.org/core/ptr/struct.NonNull.html#method.as_ref
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// # #[cfg(feature = "raw")]
+    /// # fn test() {
+    /// use core::hash::{BuildHasher, Hash};
+    /// use hashbrown::raw::{Bucket, RawTable};
+    ///
+    /// type NewHashBuilder = core::hash::BuildHasherDefault<ahash::AHasher>;
+    ///
+    /// fn make_hash<K: Hash + ?Sized, S: BuildHasher>(hash_builder: &S, key: &K) -> u64 {
+    ///     use core::hash::Hasher;
+    ///     let mut state = hash_builder.build_hasher();
+    ///     key.hash(&mut state);
+    ///     state.finish()
+    /// }
+    ///
+    /// let hash_builder = NewHashBuilder::default();
+    /// let mut table = RawTable::new();
+    ///
+    /// let value: (&str, String) = ("A pony", "is a small horse".to_owned());
+    /// let hash = make_hash(&hash_builder, &value.0);
+    ///
+    /// table.insert(hash, value.clone(), |val| make_hash(&hash_builder, &val.0));
+    ///
+    /// let bucket: Bucket<(&str, String)> = table.find(hash, |(k, _)| k == &value.0).unwrap();
+    ///
+    /// assert_eq!(
+    ///     unsafe { bucket.as_ref() },
+    ///     &("A pony", "is a small horse".to_owned())
+    /// );
+    /// # }
+    /// # fn main() {
+    /// #     #[cfg(feature = "raw")]
+    /// #     test()
+    /// # }
+    /// ```
     #[inline]
     pub unsafe fn as_ref<'a>(&self) -> &'a T {
         &*self.as_ptr()
     }
+
+    /// Returns a unique mutable reference to the `value`.
+    ///
+    /// # Safety
+    ///
+    /// See [`NonNull::as_mut`] for safety concerns.
+    ///
+    /// # Note
+    ///
+    /// [`Hash`] and [`Eq`] on the new `T` value and its borrowed form *must* match
+    /// those for the old `T` value, as the map will not re-evaluate where the new
+    /// value should go, meaning the value may become "lost" if their location
+    /// does not reflect their state.
+    ///
+    /// [`NonNull::as_mut`]: https://doc.rust-lang.org/core/ptr/struct.NonNull.html#method.as_mut
+    /// [`Hash`]: https://doc.rust-lang.org/core/hash/trait.Hash.html
+    /// [`Eq`]: https://doc.rust-lang.org/core/cmp/trait.Eq.html
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// # #[cfg(feature = "raw")]
+    /// # fn test() {
+    /// use core::hash::{BuildHasher, Hash};
+    /// use hashbrown::raw::{Bucket, RawTable};
+    ///
+    /// type NewHashBuilder = core::hash::BuildHasherDefault<ahash::AHasher>;
+    ///
+    /// fn make_hash<K: Hash + ?Sized, S: BuildHasher>(hash_builder: &S, key: &K) -> u64 {
+    ///     use core::hash::Hasher;
+    ///     let mut state = hash_builder.build_hasher();
+    ///     key.hash(&mut state);
+    ///     state.finish()
+    /// }
+    ///
+    /// let hash_builder = NewHashBuilder::default();
+    /// let mut table = RawTable::new();
+    ///
+    /// let value: (&str, String) = ("A pony", "is a small horse".to_owned());
+    /// let hash = make_hash(&hash_builder, &value.0);
+    ///
+    /// table.insert(hash, value.clone(), |val| make_hash(&hash_builder, &val.0));
+    ///
+    /// let bucket: Bucket<(&str, String)> = table.find(hash, |(k, _)| k == &value.0).unwrap();
+    ///
+    /// unsafe {
+    ///     bucket
+    ///         .as_mut()
+    ///         .1
+    ///         .push_str(" less than 147 cm at the withers")
+    /// };
+    /// assert_eq!(
+    ///     unsafe { bucket.as_ref() },
+    ///     &(
+    ///         "A pony",
+    ///         "is a small horse less than 147 cm at the withers".to_owned()
+    ///     )
+    /// );
+    /// # }
+    /// # fn main() {
+    /// #     #[cfg(feature = "raw")]
+    /// #     test()
+    /// # }
+    /// ```
     #[inline]
     pub unsafe fn as_mut<'a>(&self) -> &'a mut T {
         &mut *self.as_ptr()
     }
+
+    /// Copies `size_of<T>` bytes from `other` to `self`. The source
+    /// and destination may *not* overlap.
+    ///
+    /// # Safety
+    ///
+    /// See [`ptr::copy_nonoverlapping`] for safety concerns.
+    ///
+    /// Like [`read`], `copy_nonoverlapping` creates a bitwise copy of `T`, regardless of
+    /// whether `T` is [`Copy`]. If `T` is not [`Copy`], using *both* the values
+    /// in the region beginning at `*self` and the region beginning at `*other` can
+    /// [violate memory safety].
+    ///
+    /// # Note
+    ///
+    /// [`Hash`] and [`Eq`] on the new `T` value and its borrowed form *must* match
+    /// those for the old `T` value, as the map will not re-evaluate where the new
+    /// value should go, meaning the value may become "lost" if their location
+    /// does not reflect their state.
+    ///
+    /// [`ptr::copy_nonoverlapping`]: https://doc.rust-lang.org/core/ptr/fn.copy_nonoverlapping.html
+    /// [`read`]: https://doc.rust-lang.org/core/ptr/fn.read.html
+    /// [violate memory safety]: https://doc.rust-lang.org/std/ptr/fn.read.html#ownership-of-the-returned-value
+    /// [`Hash`]: https://doc.rust-lang.org/core/hash/trait.Hash.html
+    /// [`Eq`]: https://doc.rust-lang.org/core/cmp/trait.Eq.html
     #[cfg(feature = "raw")]
     #[inline]
     pub unsafe fn copy_from_nonoverlapping(&self, other: &Self) {
@@ -358,15 +783,16 @@ impl<T> Bucket<T> {
 }
 
 /// A raw hash table with an unsafe API.
-pub struct RawTable<T, A: Allocator + Clone = Global> {
-    table: RawTableInner<A>,
+pub struct RawTable<T, A: Allocator = Global> {
+    table: RawTableInner,
+    alloc: A,
     // Tell dropck that we own instances of T.
     marker: PhantomData<T>,
 }
 
 /// Non-generic part of `RawTable` which allows functions to be instantiated only once regardless
 /// of how many different key-value types are used.
-struct RawTableInner<A> {
+struct RawTableInner {
     // Mask to get an index from a hash value. The value is one less than the
     // number of buckets in the table.
     bucket_mask: usize,
@@ -380,8 +806,6 @@ struct RawTableInner<A> {
 
     // Number of elements in the table, only really used by len()
     items: usize,
-
-    alloc: A,
 }
 
 impl<T> RawTable<T, Global> {
@@ -393,7 +817,8 @@ impl<T> RawTable<T, Global> {
     #[inline]
     pub const fn new() -> Self {
         Self {
-            table: RawTableInner::new_in(Global),
+            table: RawTableInner::NEW,
+            alloc: Global,
             marker: PhantomData,
         }
     }
@@ -412,7 +837,9 @@ impl<T> RawTable<T, Global> {
     }
 }
 
-impl<T, A: Allocator + Clone> RawTable<T, A> {
+impl<T, A: Allocator> RawTable<T, A> {
+    const TABLE_LAYOUT: TableLayout = TableLayout::new::<T>();
+
     /// Creates a new empty hash table without allocating any memory, using the
     /// given allocator.
     ///
@@ -420,9 +847,10 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
     /// leave the data pointer dangling since that bucket is never written to
     /// due to our load factor forcing us to always have at least 1 free bucket.
     #[inline]
-    pub fn new_in(alloc: A) -> Self {
+    pub const fn new_in(alloc: A) -> Self {
         Self {
-            table: RawTableInner::new_in(alloc),
+            table: RawTableInner::NEW,
+            alloc,
             marker: PhantomData,
         }
     }
@@ -440,73 +868,97 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
 
         Ok(Self {
             table: RawTableInner::new_uninitialized(
-                alloc,
-                TableLayout::new::<T>(),
+                &alloc,
+                Self::TABLE_LAYOUT,
                 buckets,
                 fallibility,
             )?,
+            alloc,
             marker: PhantomData,
         })
     }
 
-    /// Attempts to allocate a new hash table with at least enough capacity
-    /// for inserting the given number of elements without reallocating.
-    fn fallible_with_capacity(
-        alloc: A,
-        capacity: usize,
-        fallibility: Fallibility,
-    ) -> Result<Self, TryReserveError> {
+    /// Attempts to allocate a new hash table using the given allocator, with at least enough
+    /// capacity for inserting the given number of elements without reallocating.
+    #[cfg(feature = "raw")]
+    pub fn try_with_capacity_in(capacity: usize, alloc: A) -> Result<Self, TryReserveError> {
         Ok(Self {
             table: RawTableInner::fallible_with_capacity(
-                alloc,
-                TableLayout::new::<T>(),
+                &alloc,
+                Self::TABLE_LAYOUT,
                 capacity,
-                fallibility,
+                Fallibility::Fallible,
             )?,
+            alloc,
             marker: PhantomData,
         })
     }
 
-    /// Attempts to allocate a new hash table using the given allocator, with at least enough
-    /// capacity for inserting the given number of elements without reallocating.
-    #[cfg(feature = "raw")]
-    pub fn try_with_capacity_in(capacity: usize, alloc: A) -> Result<Self, TryReserveError> {
-        Self::fallible_with_capacity(alloc, capacity, Fallibility::Fallible)
-    }
-
     /// Allocates a new hash table using the given allocator, with at least enough capacity for
     /// inserting the given number of elements without reallocating.
     pub fn with_capacity_in(capacity: usize, alloc: A) -> Self {
-        // Avoid `Result::unwrap_or_else` because it bloats LLVM IR.
-        match Self::fallible_with_capacity(alloc, capacity, Fallibility::Infallible) {
-            Ok(capacity) => capacity,
-            Err(_) => unsafe { hint::unreachable_unchecked() },
+        Self {
+            table: RawTableInner::with_capacity(&alloc, Self::TABLE_LAYOUT, capacity),
+            alloc,
+            marker: PhantomData,
         }
     }
 
     /// Returns a reference to the underlying allocator.
     #[inline]
     pub fn allocator(&self) -> &A {
-        &self.table.alloc
+        &self.alloc
     }
 
-    /// Deallocates the table without dropping any entries.
-    #[cfg_attr(feature = "inline-more", inline)]
-    unsafe fn free_buckets(&mut self) {
-        self.table.free_buckets(TableLayout::new::<T>());
+    /// Returns pointer to one past last `data` element in the table as viewed from
+    /// the start point of the allocation.
+    ///
+    /// The caller must ensure that the `RawTable` outlives the returned [`NonNull<T>`],
+    /// otherwise using it may result in [`undefined behavior`].
+    ///
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
+    #[inline]
+    pub fn data_end(&self) -> NonNull<T> {
+        //                        `self.table.ctrl.cast()` returns pointer that
+        //                        points here (to the end of `T0`)
+        //                          ∨
+        // [Pad], T_n, ..., T1, T0, |CT0, CT1, ..., CT_n|, CTa_0, CTa_1, ..., CTa_m
+        //                           \________  ________/
+        //                                    \/
+        //       `n = buckets - 1`, i.e. `RawTable::buckets() - 1`
+        //
+        // where: T0...T_n  - our stored data;
+        //        CT0...CT_n - control bytes or metadata for `data`.
+        //        CTa_0...CTa_m - additional control bytes, where `m = Group::WIDTH - 1` (so that the search
+        //                        with loading `Group` bytes from the heap works properly, even if the result
+        //                        of `h1(hash) & self.bucket_mask` is equal to `self.bucket_mask`). See also
+        //                        `RawTableInner::set_ctrl` function.
+        //
+        // P.S. `h1(hash) & self.bucket_mask` is the same as `hash as usize % self.buckets()` because the number
+        // of buckets is a power of two, and `self.bucket_mask = self.buckets() - 1`.
+        self.table.ctrl.cast()
     }
 
-    /// Returns pointer to one past last element of data table.
+    /// Returns pointer to start of data table.
     #[inline]
-    pub unsafe fn data_end(&self) -> NonNull<T> {
-        NonNull::new_unchecked(self.table.ctrl.as_ptr().cast())
+    #[cfg(any(feature = "raw", feature = "nightly"))]
+    pub unsafe fn data_start(&self) -> NonNull<T> {
+        NonNull::new_unchecked(self.data_end().as_ptr().wrapping_sub(self.buckets()))
     }
 
-    /// Returns pointer to start of data table.
+    /// Return the information about memory allocated by the table.
+    ///
+    /// `RawTable` allocates single memory block to store both data and metadata.
+    /// This function returns allocation size and alignment and the beginning of the area.
+    /// These are the arguments which will be passed to `dealloc` when the table is dropped.
+    ///
+    /// This function might be useful for memory profiling.
     #[inline]
-    #[cfg(feature = "nightly")]
-    pub unsafe fn data_start(&self) -> *mut T {
-        self.data_end().as_ptr().wrapping_sub(self.buckets())
+    #[cfg(feature = "raw")]
+    pub fn allocation_info(&self) -> (NonNull<u8>, Layout) {
+        // SAFETY: We use the same `table_layout` that was used to allocate
+        // this table.
+        unsafe { self.table.allocation_info_or_zero(Self::TABLE_LAYOUT) }
     }
 
     /// Returns the index of a bucket from a `Bucket`.
@@ -516,8 +968,55 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
     }
 
     /// Returns a pointer to an element in the table.
+    ///
+    /// The caller must ensure that the `RawTable` outlives the returned [`Bucket<T>`],
+    /// otherwise using it may result in [`undefined behavior`].
+    ///
+    /// # Safety
+    ///
+    /// If `mem::size_of::<T>() != 0`, then the caller of this function must observe the
+    /// following safety rules:
+    ///
+    /// * The table must already be allocated;
+    ///
+    /// * The `index` must not be greater than the number returned by the [`RawTable::buckets`]
+    ///   function, i.e. `(index + 1) <= self.buckets()`.
+    ///
+    /// It is safe to call this function with index of zero (`index == 0`) on a table that has
+    /// not been allocated, but using the returned [`Bucket`] results in [`undefined behavior`].
+    ///
+    /// If `mem::size_of::<T>() == 0`, then the only requirement is that the `index` must
+    /// not be greater than the number returned by the [`RawTable::buckets`] function, i.e.
+    /// `(index + 1) <= self.buckets()`.
+    ///
+    /// [`RawTable::buckets`]: RawTable::buckets
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
     #[inline]
     pub unsafe fn bucket(&self, index: usize) -> Bucket<T> {
+        // If mem::size_of::<T>() != 0 then return a pointer to the `element` in the `data part` of the table
+        // (we start counting from "0", so that in the expression T[n], the "n" index actually one less than
+        // the "buckets" number of our `RawTable`, i.e. "n = RawTable::buckets() - 1"):
+        //
+        //           `table.bucket(3).as_ptr()` returns a pointer that points here in the `data`
+        //           part of the `RawTable`, i.e. to the start of T3 (see `Bucket::as_ptr`)
+        //                  |
+        //                  |               `base = self.data_end()` points here
+        //                  |               (to the start of CT0 or to the end of T0)
+        //                  v                 v
+        // [Pad], T_n, ..., |T3|, T2, T1, T0, |CT0, CT1, CT2, CT3, ..., CT_n, CTa_0, CTa_1, ..., CTa_m
+        //                     ^                                              \__________  __________/
+        //        `table.bucket(3)` returns a pointer that points                        \/
+        //         here in the `data` part of the `RawTable` (to              additional control bytes
+        //         the end of T3)                                              `m = Group::WIDTH - 1`
+        //
+        // where: T0...T_n  - our stored data;
+        //        CT0...CT_n - control bytes or metadata for `data`;
+        //        CTa_0...CTa_m - additional control bytes (so that the search with loading `Group` bytes from
+        //                        the heap works properly, even if the result of `h1(hash) & self.table.bucket_mask`
+        //                        is equal to `self.table.bucket_mask`). See also `RawTableInner::set_ctrl` function.
+        //
+        // P.S. `h1(hash) & self.table.bucket_mask` is the same as `hash as usize % self.buckets()` because the number
+        // of buckets is a power of two, and `self.table.bucket_mask = self.buckets() - 1`.
         debug_assert_ne!(self.table.bucket_mask, 0);
         debug_assert!(index < self.buckets());
         Bucket::from_base_index(self.data_end(), index)
@@ -525,8 +1024,7 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
 
     /// Erases an element from the table without dropping it.
     #[cfg_attr(feature = "inline-more", inline)]
-    #[deprecated(since = "0.8.1", note = "use erase or remove instead")]
-    pub unsafe fn erase_no_drop(&mut self, item: &Bucket<T>) {
+    unsafe fn erase_no_drop(&mut self, item: &Bucket<T>) {
         let index = self.bucket_index(item);
         self.table.erase(index);
     }
@@ -534,7 +1032,6 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
     /// Erases an element from the table, dropping it in place.
     #[cfg_attr(feature = "inline-more", inline)]
     #[allow(clippy::needless_pass_by_value)]
-    #[allow(deprecated)]
     pub unsafe fn erase(&mut self, item: Bucket<T>) {
         // Erase the element from the table first since drop might panic.
         self.erase_no_drop(&item);
@@ -558,12 +1055,18 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
     }
 
     /// Removes an element from the table, returning it.
+    ///
+    /// This also returns an `InsertSlot` pointing to the newly free bucket.
     #[cfg_attr(feature = "inline-more", inline)]
     #[allow(clippy::needless_pass_by_value)]
-    #[allow(deprecated)]
-    pub unsafe fn remove(&mut self, item: Bucket<T>) -> T {
+    pub unsafe fn remove(&mut self, item: Bucket<T>) -> (T, InsertSlot) {
         self.erase_no_drop(&item);
-        item.read()
+        (
+            item.read(),
+            InsertSlot {
+                index: self.bucket_index(&item),
+            },
+        )
     }
 
     /// Finds and removes an element from the table, returning it.
@@ -571,7 +1074,7 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
     pub fn remove_entry(&mut self, hash: u64, eq: impl FnMut(&T) -> bool) -> Option<T> {
         // Avoid `Option::map` because it bloats LLVM IR.
         match self.find(hash, eq) {
-            Some(bucket) => Some(unsafe { self.remove(bucket) }),
+            Some(bucket) => Some(unsafe { self.remove(bucket).0 }),
             None => None,
         }
     }
@@ -585,18 +1088,17 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
     /// Removes all elements from the table without freeing the backing memory.
     #[cfg_attr(feature = "inline-more", inline)]
     pub fn clear(&mut self) {
+        if self.is_empty() {
+            // Special case empty table to avoid surprising O(capacity) time.
+            return;
+        }
         // Ensure that the table is reset even if one of the drops panic
         let mut self_ = guard(self, |self_| self_.clear_no_drop());
         unsafe {
-            self_.drop_elements();
-        }
-    }
-
-    unsafe fn drop_elements(&mut self) {
-        if mem::needs_drop::<T>() && !self.is_empty() {
-            for item in self.iter() {
-                item.drop();
-            }
+            // SAFETY: ScopeGuard sets to zero the `items` field of the table
+            // even in case of panic during the dropping of the elements so
+            // that there will be no double drop of the elements.
+            self_.table.drop_elements::<T>();
         }
     }
 
@@ -607,7 +1109,16 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
         // space for.
         let min_size = usize::max(self.table.items, min_size);
         if min_size == 0 {
-            *self = Self::new_in(self.table.alloc.clone());
+            let mut old_inner = mem::replace(&mut self.table, RawTableInner::NEW);
+            unsafe {
+                // SAFETY:
+                // 1. We call the function only once;
+                // 2. We know for sure that `alloc` and `table_layout` matches the [`Allocator`]
+                //    and [`TableLayout`] that were used to allocate this table.
+                // 3. If any elements' drop function panics, then there will only be a memory leak,
+                //    because we have replaced the inner table with a new one.
+                old_inner.drop_inner_table::<T, _>(&self.alloc, Self::TABLE_LAYOUT);
+            }
             return;
         }
 
@@ -624,14 +1135,33 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
         if min_buckets < self.buckets() {
             // Fast path if the table is empty
             if self.table.items == 0 {
-                *self = Self::with_capacity_in(min_size, self.table.alloc.clone());
+                let new_inner =
+                    RawTableInner::with_capacity(&self.alloc, Self::TABLE_LAYOUT, min_size);
+                let mut old_inner = mem::replace(&mut self.table, new_inner);
+                unsafe {
+                    // SAFETY:
+                    // 1. We call the function only once;
+                    // 2. We know for sure that `alloc` and `table_layout` matches the [`Allocator`]
+                    //    and [`TableLayout`] that were used to allocate this table.
+                    // 3. If any elements' drop function panics, then there will only be a memory leak,
+                    //    because we have replaced the inner table with a new one.
+                    old_inner.drop_inner_table::<T, _>(&self.alloc, Self::TABLE_LAYOUT);
+                }
             } else {
                 // Avoid `Result::unwrap_or_else` because it bloats LLVM IR.
-                if self
-                    .resize(min_size, hasher, Fallibility::Infallible)
-                    .is_err()
-                {
-                    unsafe { hint::unreachable_unchecked() }
+                unsafe {
+                    // SAFETY:
+                    // 1. We know for sure that `min_size >= self.table.items`.
+                    // 2. The [`RawTableInner`] must already have properly initialized control bytes since
+                    //    we will never expose RawTable::new_uninitialized in a public API.
+                    if self
+                        .resize(min_size, hasher, Fallibility::Infallible)
+                        .is_err()
+                    {
+                        // SAFETY: The result of calling the `resize` function cannot be an error
+                        // because `fallibility == Fallibility::Infallible.
+                        hint::unreachable_unchecked()
+                    }
                 }
             }
         }
@@ -641,13 +1171,18 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
     /// without reallocation.
     #[cfg_attr(feature = "inline-more", inline)]
     pub fn reserve(&mut self, additional: usize, hasher: impl Fn(&T) -> u64) {
-        if additional > self.table.growth_left {
+        if unlikely(additional > self.table.growth_left) {
             // Avoid `Result::unwrap_or_else` because it bloats LLVM IR.
-            if self
-                .reserve_rehash(additional, hasher, Fallibility::Infallible)
-                .is_err()
-            {
-                unsafe { hint::unreachable_unchecked() }
+            unsafe {
+                // SAFETY: The [`RawTableInner`] must already have properly initialized control
+                // bytes since we will never expose RawTable::new_uninitialized in a public API.
+                if self
+                    .reserve_rehash(additional, hasher, Fallibility::Infallible)
+                    .is_err()
+                {
+                    // SAFETY: All allocation errors will be caught inside `RawTableInner::reserve_rehash`.
+                    hint::unreachable_unchecked()
+                }
             }
         }
     }
@@ -661,28 +1196,45 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
         hasher: impl Fn(&T) -> u64,
     ) -> Result<(), TryReserveError> {
         if additional > self.table.growth_left {
-            self.reserve_rehash(additional, hasher, Fallibility::Fallible)
+            // SAFETY: The [`RawTableInner`] must already have properly initialized control
+            // bytes since we will never expose RawTable::new_uninitialized in a public API.
+            unsafe { self.reserve_rehash(additional, hasher, Fallibility::Fallible) }
         } else {
             Ok(())
         }
     }
 
     /// Out-of-line slow path for `reserve` and `try_reserve`.
+    ///
+    /// # Safety
+    ///
+    /// The [`RawTableInner`] must have properly initialized control bytes,
+    /// otherwise calling this function results in [`undefined behavior`]
+    ///
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
     #[cold]
     #[inline(never)]
-    fn reserve_rehash(
+    unsafe fn reserve_rehash(
         &mut self,
         additional: usize,
         hasher: impl Fn(&T) -> u64,
         fallibility: Fallibility,
     ) -> Result<(), TryReserveError> {
         unsafe {
+            // SAFETY:
+            // 1. We know for sure that `alloc` and `layout` matches the [`Allocator`] and
+            //    [`TableLayout`] that were used to allocate this table.
+            // 2. The `drop` function is the actual drop function of the elements stored in
+            //    the table.
+            // 3. The caller ensures that the control bytes of the `RawTableInner`
+            //    are already initialized.
             self.table.reserve_rehash_inner(
+                &self.alloc,
                 additional,
                 &|table, index| hasher(table.bucket::<T>(index).as_ref()),
                 fallibility,
-                TableLayout::new::<T>(),
-                if mem::needs_drop::<T>() {
+                Self::TABLE_LAYOUT,
+                if T::NEEDS_DROP {
                     Some(mem::transmute(ptr::drop_in_place::<T> as unsafe fn(*mut T)))
                 } else {
                     None
@@ -693,20 +1245,50 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
 
     /// Allocates a new table of a different size and moves the contents of the
     /// current table into it.
-    fn resize(
+    ///
+    /// # Safety
+    ///
+    /// The [`RawTableInner`] must have properly initialized control bytes,
+    /// otherwise calling this function results in [`undefined behavior`]
+    ///
+    /// The caller of this function must ensure that `capacity >= self.table.items`
+    /// otherwise:
+    ///
+    /// * If `self.table.items != 0`, calling of this function with `capacity`
+    ///   equal to 0 (`capacity == 0`) results in [`undefined behavior`].
+    ///
+    /// * If `capacity_to_buckets(capacity) < Group::WIDTH` and
+    ///   `self.table.items > capacity_to_buckets(capacity)`
+    ///   calling this function results in [`undefined behavior`].
+    ///
+    /// * If `capacity_to_buckets(capacity) >= Group::WIDTH` and
+    ///   `self.table.items > capacity_to_buckets(capacity)`
+    ///   calling this function are never return (will go into an
+    ///   infinite loop).
+    ///
+    /// See [`RawTableInner::find_insert_slot`] for more information.
+    ///
+    /// [`RawTableInner::find_insert_slot`]: RawTableInner::find_insert_slot
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
+    unsafe fn resize(
         &mut self,
         capacity: usize,
         hasher: impl Fn(&T) -> u64,
         fallibility: Fallibility,
     ) -> Result<(), TryReserveError> {
-        unsafe {
-            self.table.resize_inner(
-                capacity,
-                &|table, index| hasher(table.bucket::<T>(index).as_ref()),
-                fallibility,
-                TableLayout::new::<T>(),
-            )
-        }
+        // SAFETY:
+        // 1. The caller of this function guarantees that `capacity >= self.table.items`.
+        // 2. We know for sure that `alloc` and `layout` matches the [`Allocator`] and
+        //    [`TableLayout`] that were used to allocate this table.
+        // 3. The caller ensures that the control bytes of the `RawTableInner`
+        //    are already initialized.
+        self.table.resize_inner(
+            &self.alloc,
+            capacity,
+            &|table, index| hasher(table.bucket::<T>(index).as_ref()),
+            fallibility,
+            Self::TABLE_LAYOUT,
+        )
     }
 
     /// Inserts a new element into the table, and returns its raw bucket.
@@ -715,22 +1297,27 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
     #[cfg_attr(feature = "inline-more", inline)]
     pub fn insert(&mut self, hash: u64, value: T, hasher: impl Fn(&T) -> u64) -> Bucket<T> {
         unsafe {
-            let mut index = self.table.find_insert_slot(hash);
+            // SAFETY:
+            // 1. The [`RawTableInner`] must already have properly initialized control bytes since
+            //    we will never expose `RawTable::new_uninitialized` in a public API.
+            //
+            // 2. We reserve additional space (if necessary) right after calling this function.
+            let mut slot = self.table.find_insert_slot(hash);
 
-            // We can avoid growing the table once we have reached our load
-            // factor if we are replacing a tombstone. This works since the
-            // number of EMPTY slots does not change in this case.
-            let old_ctrl = *self.table.ctrl(index);
+            // We can avoid growing the table once we have reached our load factor if we are replacing
+            // a tombstone. This works since the number of EMPTY slots does not change in this case.
+            //
+            // SAFETY: The function is guaranteed to return [`InsertSlot`] that contains an index
+            // in the range `0..=self.buckets()`.
+            let old_ctrl = *self.table.ctrl(slot.index);
             if unlikely(self.table.growth_left == 0 && special_is_empty(old_ctrl)) {
                 self.reserve(1, hasher);
-                index = self.table.find_insert_slot(hash);
+                // SAFETY: We know for sure that `RawTableInner` has control bytes
+                // initialized and that there is extra space in the table.
+                slot = self.table.find_insert_slot(hash);
             }
 
-            self.table.record_item_insert_at(index, old_ctrl, hash);
-
-            let bucket = self.bucket(index);
-            bucket.write(value);
-            bucket
+            self.insert_in_slot(hash, slot, value)
         }
     }
 
@@ -796,9 +1383,9 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
     {
         let index = self.bucket_index(&bucket);
         let old_ctrl = *self.table.ctrl(index);
-        debug_assert!(is_full(old_ctrl));
+        debug_assert!(self.is_bucket_full(index));
         let old_growth_left = self.table.growth_left;
-        let item = self.remove(bucket);
+        let item = self.remove(bucket).0;
         if let Some(new_item) = f(item) {
             self.table.growth_left = old_growth_left;
             self.table.set_ctrl(index, old_ctrl);
@@ -810,17 +1397,78 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
         }
     }
 
+    /// Searches for an element in the table. If the element is not found,
+    /// returns `Err` with the position of a slot where an element with the
+    /// same hash could be inserted.
+    ///
+    /// This function may resize the table if additional space is required for
+    /// inserting an element.
+    #[inline]
+    pub fn find_or_find_insert_slot(
+        &mut self,
+        hash: u64,
+        mut eq: impl FnMut(&T) -> bool,
+        hasher: impl Fn(&T) -> u64,
+    ) -> Result<Bucket<T>, InsertSlot> {
+        self.reserve(1, hasher);
+
+        unsafe {
+            // SAFETY:
+            // 1. We know for sure that there is at least one empty `bucket` in the table.
+            // 2. The [`RawTableInner`] must already have properly initialized control bytes since we will
+            //    never expose `RawTable::new_uninitialized` in a public API.
+            // 3. The `find_or_find_insert_slot_inner` function returns the `index` of only the full bucket,
+            //    which is in the range `0..self.buckets()` (since there is at least one empty `bucket` in
+            //    the table), so calling `self.bucket(index)` and `Bucket::as_ref` is safe.
+            match self
+                .table
+                .find_or_find_insert_slot_inner(hash, &mut |index| eq(self.bucket(index).as_ref()))
+            {
+                // SAFETY: See explanation above.
+                Ok(index) => Ok(self.bucket(index)),
+                Err(slot) => Err(slot),
+            }
+        }
+    }
+
+    /// Inserts a new element into the table in the given slot, and returns its
+    /// raw bucket.
+    ///
+    /// # Safety
+    ///
+    /// `slot` must point to a slot previously returned by
+    /// `find_or_find_insert_slot`, and no mutation of the table must have
+    /// occurred since that call.
+    #[inline]
+    pub unsafe fn insert_in_slot(&mut self, hash: u64, slot: InsertSlot, value: T) -> Bucket<T> {
+        let old_ctrl = *self.table.ctrl(slot.index);
+        self.table.record_item_insert_at(slot.index, old_ctrl, hash);
+
+        let bucket = self.bucket(slot.index);
+        bucket.write(value);
+        bucket
+    }
+
     /// Searches for an element in the table.
     #[inline]
     pub fn find(&self, hash: u64, mut eq: impl FnMut(&T) -> bool) -> Option<Bucket<T>> {
-        let result = self.table.find_inner(hash, &mut |index| unsafe {
-            eq(self.bucket(index).as_ref())
-        });
-
-        // Avoid `Option::map` because it bloats LLVM IR.
-        match result {
-            Some(index) => Some(unsafe { self.bucket(index) }),
-            None => None,
+        unsafe {
+            // SAFETY:
+            // 1. The [`RawTableInner`] must already have properly initialized control bytes since we
+            //    will never expose `RawTable::new_uninitialized` in a public API.
+            // 1. The `find_inner` function returns the `index` of only the full bucket, which is in
+            //    the range `0..self.buckets()`, so calling `self.bucket(index)` and `Bucket::as_ref`
+            //    is safe.
+            let result = self
+                .table
+                .find_inner(hash, &mut |index| eq(self.bucket(index).as_ref()));
+
+            // Avoid `Option::map` because it bloats LLVM IR.
+            match result {
+                // SAFETY: See explanation above.
+                Some(index) => Some(self.bucket(index)),
+                None => None,
+            }
         }
     }
 
@@ -928,17 +1576,27 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
         self.table.bucket_mask + 1
     }
 
+    /// Checks whether the bucket at `index` is full.
+    ///
+    /// # Safety
+    ///
+    /// The caller must ensure `index` is less than the number of buckets.
+    #[inline]
+    pub unsafe fn is_bucket_full(&self, index: usize) -> bool {
+        self.table.is_bucket_full(index)
+    }
+
     /// Returns an iterator over every element in the table. It is up to
     /// the caller to ensure that the `RawTable` outlives the `RawIter`.
     /// Because we cannot make the `next` method unsafe on the `RawIter`
     /// struct, we have to make the `iter` method unsafe.
     #[inline]
     pub unsafe fn iter(&self) -> RawIter<T> {
-        let data = Bucket::from_base_index(self.data_end(), 0);
-        RawIter {
-            iter: RawIterRange::new(self.table.ctrl.as_ptr(), data, self.table.buckets()),
-            items: self.table.items,
-        }
+        // SAFETY:
+        // 1. The caller must uphold the safety contract for `iter` method.
+        // 2. The [`RawTableInner`] must already have properly initialized control bytes since
+        //    we will never expose RawTable::new_uninitialized in a public API.
+        self.table.iter()
     }
 
     /// Returns an iterator over occupied buckets that could match a given hash.
@@ -952,7 +1610,7 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
     /// `RawIterHash` struct, we have to make the `iter_hash` method unsafe.
     #[cfg_attr(feature = "inline-more", inline)]
     #[cfg(feature = "raw")]
-    pub unsafe fn iter_hash(&self, hash: u64) -> RawIterHash<'_, T, A> {
+    pub unsafe fn iter_hash(&self, hash: u64) -> RawIterHash<T> {
         RawIterHash::new(self, hash)
     }
 
@@ -978,8 +1636,8 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
         debug_assert_eq!(iter.len(), self.len());
         RawDrain {
             iter,
-            table: ManuallyDrop::new(mem::replace(self, Self::new_in(self.table.alloc.clone()))),
-            orig_table: NonNull::from(self),
+            table: mem::replace(&mut self.table, RawTableInner::NEW),
+            orig_table: NonNull::from(&mut self.table),
             marker: PhantomData,
         }
     }
@@ -993,31 +1651,31 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
     pub unsafe fn into_iter_from(self, iter: RawIter<T>) -> RawIntoIter<T, A> {
         debug_assert_eq!(iter.len(), self.len());
 
-        let alloc = self.table.alloc.clone();
         let allocation = self.into_allocation();
         RawIntoIter {
             iter,
             allocation,
             marker: PhantomData,
-            alloc,
         }
     }
 
     /// Converts the table into a raw allocation. The contents of the table
     /// should be dropped using a `RawIter` before freeing the allocation.
     #[cfg_attr(feature = "inline-more", inline)]
-    pub(crate) fn into_allocation(self) -> Option<(NonNull<u8>, Layout)> {
+    pub(crate) fn into_allocation(self) -> Option<(NonNull<u8>, Layout, A)> {
         let alloc = if self.table.is_empty_singleton() {
             None
         } else {
             // Avoid `Option::unwrap_or_else` because it bloats LLVM IR.
-            let (layout, ctrl_offset) = match calculate_layout::<T>(self.table.buckets()) {
-                Some(lco) => lco,
-                None => unsafe { hint::unreachable_unchecked() },
-            };
+            let (layout, ctrl_offset) =
+                match Self::TABLE_LAYOUT.calculate_layout_for(self.table.buckets()) {
+                    Some(lco) => lco,
+                    None => unsafe { hint::unreachable_unchecked() },
+                };
             Some((
                 unsafe { NonNull::new_unchecked(self.table.ctrl.as_ptr().sub(ctrl_offset)) },
                 layout,
+                unsafe { ptr::read(&self.alloc) },
             ))
         };
         mem::forget(self);
@@ -1025,41 +1683,62 @@ impl<T, A: Allocator + Clone> RawTable<T, A> {
     }
 }
 
-unsafe impl<T, A: Allocator + Clone> Send for RawTable<T, A>
+unsafe impl<T, A: Allocator> Send for RawTable<T, A>
 where
     T: Send,
     A: Send,
 {
 }
-unsafe impl<T, A: Allocator + Clone> Sync for RawTable<T, A>
+unsafe impl<T, A: Allocator> Sync for RawTable<T, A>
 where
     T: Sync,
     A: Sync,
 {
 }
 
-impl<A> RawTableInner<A> {
+impl RawTableInner {
+    const NEW: Self = RawTableInner::new();
+
+    /// Creates a new empty hash table without allocating any memory.
+    ///
+    /// In effect this returns a table with exactly 1 bucket. However we can
+    /// leave the data pointer dangling since that bucket is never accessed
+    /// due to our load factor forcing us to always have at least 1 free bucket.
     #[inline]
-    const fn new_in(alloc: A) -> Self {
+    const fn new() -> Self {
         Self {
             // Be careful to cast the entire slice to a raw pointer.
             ctrl: unsafe { NonNull::new_unchecked(Group::static_empty() as *const _ as *mut u8) },
             bucket_mask: 0,
             items: 0,
             growth_left: 0,
-            alloc,
         }
     }
 }
 
-impl<A: Allocator + Clone> RawTableInner<A> {
+impl RawTableInner {
+    /// Allocates a new [`RawTableInner`] with the given number of buckets.
+    /// The control bytes and buckets are left uninitialized.
+    ///
+    /// # Safety
+    ///
+    /// The caller of this function must ensure that the `buckets` is power of two
+    /// and also initialize all control bytes of the length `self.bucket_mask + 1 +
+    /// Group::WIDTH` with the [`EMPTY`] bytes.
+    ///
+    /// See also [`Allocator`] API for other safety concerns.
+    ///
+    /// [`Allocator`]: https://doc.rust-lang.org/alloc/alloc/trait.Allocator.html
     #[cfg_attr(feature = "inline-more", inline)]
-    unsafe fn new_uninitialized(
-        alloc: A,
+    unsafe fn new_uninitialized<A>(
+        alloc: &A,
         table_layout: TableLayout,
         buckets: usize,
         fallibility: Fallibility,
-    ) -> Result<Self, TryReserveError> {
+    ) -> Result<Self, TryReserveError>
+    where
+        A: Allocator,
+    {
         debug_assert!(buckets.is_power_of_two());
 
         // Avoid `Option::ok_or_else` because it bloats LLVM IR.
@@ -1068,45 +1747,48 @@ impl<A: Allocator + Clone> RawTableInner<A> {
             None => return Err(fallibility.capacity_overflow()),
         };
 
-        // We need an additional check to ensure that the allocation doesn't
-        // exceed `isize::MAX`. We can skip this check on 64-bit systems since
-        // such allocations will never succeed anyways.
-        //
-        // This mirrors what Vec does in the standard library.
-        if mem::size_of::<usize>() < 8 && layout.size() > isize::MAX as usize {
-            return Err(fallibility.capacity_overflow());
-        }
-
-        let ptr: NonNull<u8> = match do_alloc(&alloc, layout) {
+        let ptr: NonNull<u8> = match do_alloc(alloc, layout) {
             Ok(block) => block.cast(),
             Err(_) => return Err(fallibility.alloc_err(layout)),
         };
 
+        // SAFETY: null pointer will be caught in above check
         let ctrl = NonNull::new_unchecked(ptr.as_ptr().add(ctrl_offset));
         Ok(Self {
             ctrl,
             bucket_mask: buckets - 1,
             items: 0,
             growth_left: bucket_mask_to_capacity(buckets - 1),
-            alloc,
         })
     }
 
+    /// Attempts to allocate a new [`RawTableInner`] with at least enough
+    /// capacity for inserting the given number of elements without reallocating.
+    ///
+    /// All the control bytes are initialized with the [`EMPTY`] bytes.
     #[inline]
-    fn fallible_with_capacity(
-        alloc: A,
+    fn fallible_with_capacity<A>(
+        alloc: &A,
         table_layout: TableLayout,
         capacity: usize,
         fallibility: Fallibility,
-    ) -> Result<Self, TryReserveError> {
+    ) -> Result<Self, TryReserveError>
+    where
+        A: Allocator,
+    {
         if capacity == 0 {
-            Ok(Self::new_in(alloc))
+            Ok(Self::NEW)
         } else {
+            // SAFETY: We checked that we could successfully allocate the new table, and then
+            // initialized all control bytes with the constant `EMPTY` byte.
             unsafe {
                 let buckets =
                     capacity_to_buckets(capacity).ok_or_else(|| fallibility.capacity_overflow())?;
 
                 let result = Self::new_uninitialized(alloc, table_layout, buckets, fallibility)?;
+                // SAFETY: We checked that the table is allocated and therefore the table already has
+                // `self.bucket_mask + 1 + Group::WIDTH` number of control bytes (see TableLayout::calculate_layout_for)
+                // so writing `self.num_ctrl_bytes() == bucket_mask + 1 + Group::WIDTH` bytes is safe.
                 result.ctrl(0).write_bytes(EMPTY, result.num_ctrl_bytes());
 
                 Ok(result)
@@ -1114,66 +1796,397 @@ impl<A: Allocator + Clone> RawTableInner<A> {
         }
     }
 
-    /// Searches for an empty or deleted bucket which is suitable for inserting
-    /// a new element and sets the hash for that slot.
+    /// Allocates a new [`RawTableInner`] with at least enough capacity for inserting
+    /// the given number of elements without reallocating.
+    ///
+    /// Panics if the new capacity exceeds [`isize::MAX`] bytes and [`abort`] the program
+    /// in case of allocation error. Use [`fallible_with_capacity`] instead if you want to
+    /// handle memory allocation failure.
+    ///
+    /// All the control bytes are initialized with the [`EMPTY`] bytes.
+    ///
+    /// [`fallible_with_capacity`]: RawTableInner::fallible_with_capacity
+    /// [`abort`]: https://doc.rust-lang.org/alloc/alloc/fn.handle_alloc_error.html
+    fn with_capacity<A>(alloc: &A, table_layout: TableLayout, capacity: usize) -> Self
+    where
+        A: Allocator,
+    {
+        // Avoid `Result::unwrap_or_else` because it bloats LLVM IR.
+        match Self::fallible_with_capacity(alloc, table_layout, capacity, Fallibility::Infallible) {
+            Ok(table_inner) => table_inner,
+            // SAFETY: All allocation errors will be caught inside `RawTableInner::new_uninitialized`.
+            Err(_) => unsafe { hint::unreachable_unchecked() },
+        }
+    }
+
+    /// Fixes up an insertion slot returned by the [`RawTableInner::find_insert_slot_in_group`] method.
+    ///
+    /// In tables smaller than the group width (`self.buckets() < Group::WIDTH`), trailing control
+    /// bytes outside the range of the table are filled with [`EMPTY`] entries. These will unfortunately
+    /// trigger a match of [`RawTableInner::find_insert_slot_in_group`] function. This is because
+    /// the `Some(bit)` returned by `group.match_empty_or_deleted().lowest_set_bit()` after masking
+    /// (`(probe_seq.pos + bit) & self.bucket_mask`) may point to a full bucket that is already occupied.
+    /// We detect this situation here and perform a second scan starting at the beginning of the table.
+    /// This second scan is guaranteed to find an empty slot (due to the load factor) before hitting the
+    /// trailing control bytes (containing [`EMPTY`] bytes).
+    ///
+    /// If this function is called correctly, it is guaranteed to return [`InsertSlot`] with an
+    /// index of an empty or deleted bucket in the range `0..self.buckets()` (see `Warning` and
+    /// `Safety`).
+    ///
+    /// # Warning
+    ///
+    /// The table must have at least 1 empty or deleted `bucket`, otherwise if the table is less than
+    /// the group width (`self.buckets() < Group::WIDTH`) this function returns an index outside of the
+    /// table indices range `0..self.buckets()` (`0..=self.bucket_mask`). Attempt to write data at that
+    /// index will cause immediate [`undefined behavior`].
+    ///
+    /// # Safety
+    ///
+    /// The safety rules are directly derived from the safety rules for [`RawTableInner::ctrl`] method.
+    /// Thus, in order to uphold those safety contracts, as well as for the correct logic of the work
+    /// of this crate, the following rules are necessary and sufficient:
     ///
-    /// There must be at least 1 empty bucket in the table.
+    /// * The [`RawTableInner`] must have properly initialized control bytes otherwise calling this
+    ///   function results in [`undefined behavior`].
+    ///
+    /// * This function must only be used on insertion slots found by [`RawTableInner::find_insert_slot_in_group`]
+    ///   (after the `find_insert_slot_in_group` function, but before insertion into the table).
+    ///
+    /// * The `index` must not be greater than the `self.bucket_mask`, i.e. `(index + 1) <= self.buckets()`
+    ///   (this one is provided by the [`RawTableInner::find_insert_slot_in_group`] function).
+    ///
+    /// Calling this function with an index not provided by [`RawTableInner::find_insert_slot_in_group`]
+    /// may result in [`undefined behavior`] even if the index satisfies the safety rules of the
+    /// [`RawTableInner::ctrl`] function (`index < self.bucket_mask + 1 + Group::WIDTH`).
+    ///
+    /// [`RawTableInner::ctrl`]: RawTableInner::ctrl
+    /// [`RawTableInner::find_insert_slot_in_group`]: RawTableInner::find_insert_slot_in_group
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
     #[inline]
-    unsafe fn prepare_insert_slot(&self, hash: u64) -> (usize, u8) {
-        let index = self.find_insert_slot(hash);
+    unsafe fn fix_insert_slot(&self, mut index: usize) -> InsertSlot {
+        // SAFETY: The caller of this function ensures that `index` is in the range `0..=self.bucket_mask`.
+        if unlikely(self.is_bucket_full(index)) {
+            debug_assert!(self.bucket_mask < Group::WIDTH);
+            // SAFETY:
+            //
+            // * Since the caller of this function ensures that the control bytes are properly
+            //   initialized and `ptr = self.ctrl(0)` points to the start of the array of control
+            //   bytes, therefore: `ctrl` is valid for reads, properly aligned to `Group::WIDTH`
+            //   and points to the properly initialized control bytes (see also
+            //   `TableLayout::calculate_layout_for` and `ptr::read`);
+            //
+            // * Because the caller of this function ensures that the index was provided by the
+            //   `self.find_insert_slot_in_group()` function, so for for tables larger than the
+            //   group width (self.buckets() >= Group::WIDTH), we will never end up in the given
+            //   branch, since `(probe_seq.pos + bit) & self.bucket_mask` in `find_insert_slot_in_group`
+            //   cannot return a full bucket index. For tables smaller than the group width, calling
+            //   the `unwrap_unchecked` function is also safe, as the trailing control bytes outside
+            //   the range of the table are filled with EMPTY bytes (and we know for sure that there
+            //   is at least one FULL bucket), so this second scan either finds an empty slot (due to
+            //   the load factor) or hits the trailing control bytes (containing EMPTY).
+            index = Group::load_aligned(self.ctrl(0))
+                .match_empty_or_deleted()
+                .lowest_set_bit()
+                .unwrap_unchecked();
+        }
+        InsertSlot { index }
+    }
+
+    /// Finds the position to insert something in a group.
+    ///
+    /// **This may have false positives and must be fixed up with `fix_insert_slot`
+    /// before it's used.**
+    ///
+    /// The function is guaranteed to return the index of an empty or deleted [`Bucket`]
+    /// in the range `0..self.buckets()` (`0..=self.bucket_mask`).
+    #[inline]
+    fn find_insert_slot_in_group(&self, group: &Group, probe_seq: &ProbeSeq) -> Option<usize> {
+        let bit = group.match_empty_or_deleted().lowest_set_bit();
+
+        if likely(bit.is_some()) {
+            // This is the same as `(probe_seq.pos + bit) % self.buckets()` because the number
+            // of buckets is a power of two, and `self.bucket_mask = self.buckets() - 1`.
+            Some((probe_seq.pos + bit.unwrap()) & self.bucket_mask)
+        } else {
+            None
+        }
+    }
+
+    /// Searches for an element in the table, or a potential slot where that element could
+    /// be inserted (an empty or deleted [`Bucket`] index).
+    ///
+    /// This uses dynamic dispatch to reduce the amount of code generated, but that is
+    /// eliminated by LLVM optimizations.
+    ///
+    /// This function does not make any changes to the `data` part of the table, or any
+    /// changes to the `items` or `growth_left` field of the table.
+    ///
+    /// The table must have at least 1 empty or deleted `bucket`, otherwise, if the
+    /// `eq: &mut dyn FnMut(usize) -> bool` function does not return `true`, this function
+    /// will never return (will go into an infinite loop) for tables larger than the group
+    /// width, or return an index outside of the table indices range if the table is less
+    /// than the group width.
+    ///
+    /// This function is guaranteed to provide the `eq: &mut dyn FnMut(usize) -> bool`
+    /// function with only `FULL` buckets' indices and return the `index` of the found
+    /// element (as `Ok(index)`). If the element is not found and there is at least 1
+    /// empty or deleted [`Bucket`] in the table, the function is guaranteed to return
+    /// [InsertSlot] with an index in the range `0..self.buckets()`, but in any case,
+    /// if this function returns [`InsertSlot`], it will contain an index in the range
+    /// `0..=self.buckets()`.
+    ///
+    /// # Safety
+    ///
+    /// The [`RawTableInner`] must have properly initialized control bytes otherwise calling
+    /// this function results in [`undefined behavior`].
+    ///
+    /// Attempt to write data at the [`InsertSlot`] returned by this function when the table is
+    /// less than the group width and if there was not at least one empty or deleted bucket in
+    /// the table will cause immediate [`undefined behavior`]. This is because in this case the
+    /// function will return `self.bucket_mask + 1` as an index due to the trailing [`EMPTY]
+    /// control bytes outside the table range.
+    ///
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
+    #[inline]
+    unsafe fn find_or_find_insert_slot_inner(
+        &self,
+        hash: u64,
+        eq: &mut dyn FnMut(usize) -> bool,
+    ) -> Result<usize, InsertSlot> {
+        let mut insert_slot = None;
+
+        let h2_hash = h2(hash);
+        let mut probe_seq = self.probe_seq(hash);
+
+        loop {
+            // SAFETY:
+            // * Caller of this function ensures that the control bytes are properly initialized.
+            //
+            // * `ProbeSeq.pos` cannot be greater than `self.bucket_mask = self.buckets() - 1`
+            //   of the table due to masking with `self.bucket_mask` and also because mumber of
+            //   buckets is a power of two (see `self.probe_seq` function).
+            //
+            // * Even if `ProbeSeq.pos` returns `position == self.bucket_mask`, it is safe to
+            //   call `Group::load` due to the extended control bytes range, which is
+            //  `self.bucket_mask + 1 + Group::WIDTH` (in fact, this means that the last control
+            //   byte will never be read for the allocated table);
+            //
+            // * Also, even if `RawTableInner` is not already allocated, `ProbeSeq.pos` will
+            //   always return "0" (zero), so Group::load will read unaligned `Group::static_empty()`
+            //   bytes, which is safe (see RawTableInner::new).
+            let group = unsafe { Group::load(self.ctrl(probe_seq.pos)) };
+
+            for bit in group.match_byte(h2_hash) {
+                let index = (probe_seq.pos + bit) & self.bucket_mask;
+
+                if likely(eq(index)) {
+                    return Ok(index);
+                }
+            }
+
+            // We didn't find the element we were looking for in the group, try to get an
+            // insertion slot from the group if we don't have one yet.
+            if likely(insert_slot.is_none()) {
+                insert_slot = self.find_insert_slot_in_group(&group, &probe_seq);
+            }
+
+            // Only stop the search if the group contains at least one empty element.
+            // Otherwise, the element that we are looking for might be in a following group.
+            if likely(group.match_empty().any_bit_set()) {
+                // We must have found a insert slot by now, since the current group contains at
+                // least one. For tables smaller than the group width, there will still be an
+                // empty element in the current (and only) group due to the load factor.
+                unsafe {
+                    // SAFETY:
+                    // * Caller of this function ensures that the control bytes are properly initialized.
+                    //
+                    // * We use this function with the slot / index found by `self.find_insert_slot_in_group`
+                    return Err(self.fix_insert_slot(insert_slot.unwrap_unchecked()));
+                }
+            }
+
+            probe_seq.move_next(self.bucket_mask);
+        }
+    }
+
+    /// Searches for an empty or deleted bucket which is suitable for inserting a new
+    /// element and sets the hash for that slot. Returns an index of that slot and the
+    /// old control byte stored in the found index.
+    ///
+    /// This function does not check if the given element exists in the table. Also,
+    /// this function does not check if there is enough space in the table to insert
+    /// a new element. Caller of the funtion must make ensure that the table has at
+    /// least 1 empty or deleted `bucket`, otherwise this function will never return
+    /// (will go into an infinite loop) for tables larger than the group width, or
+    /// return an index outside of the table indices range if the table is less than
+    /// the group width.
+    ///
+    /// If there is at least 1 empty or deleted `bucket` in the table, the function is
+    /// guaranteed to return an `index` in the range `0..self.buckets()`, but in any case,
+    /// if this function returns an `index` it will be in the range `0..=self.buckets()`.
+    ///
+    /// This function does not make any changes to the `data` parts of the table,
+    /// or any changes to the `items` or `growth_left` field of the table.
+    ///
+    /// # Safety
+    ///
+    /// The safety rules are directly derived from the safety rules for the
+    /// [`RawTableInner::set_ctrl_h2`] and [`RawTableInner::find_insert_slot`] methods.
+    /// Thus, in order to uphold the safety contracts for that methods, as well as for
+    /// the correct logic of the work of this crate, you must observe the following rules
+    /// when calling this function:
+    ///
+    /// * The [`RawTableInner`] has already been allocated and has properly initialized
+    ///   control bytes otherwise calling this function results in [`undefined behavior`].
+    ///
+    /// * The caller of this function must ensure that the "data" parts of the table
+    ///   will have an entry in the returned index (matching the given hash) right
+    ///   after calling this function.
+    ///
+    /// Attempt to write data at the `index` returned by this function when the table is
+    /// less than the group width and if there was not at least one empty or deleted bucket in
+    /// the table will cause immediate [`undefined behavior`]. This is because in this case the
+    /// function will return `self.bucket_mask + 1` as an index due to the trailing [`EMPTY]
+    /// control bytes outside the table range.
+    ///
+    /// The caller must independently increase the `items` field of the table, and also,
+    /// if the old control byte was [`EMPTY`], then decrease the table's `growth_left`
+    /// field, and do not change it if the old control byte was [`DELETED`].
+    ///
+    /// See also [`Bucket::as_ptr`] method, for more information about of properly removing
+    /// or saving `element` from / into the [`RawTable`] / [`RawTableInner`].
+    ///
+    /// [`Bucket::as_ptr`]: Bucket::as_ptr
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
+    /// [`RawTableInner::ctrl`]: RawTableInner::ctrl
+    /// [`RawTableInner::set_ctrl_h2`]: RawTableInner::set_ctrl_h2
+    /// [`RawTableInner::find_insert_slot`]: RawTableInner::find_insert_slot
+    #[inline]
+    unsafe fn prepare_insert_slot(&mut self, hash: u64) -> (usize, u8) {
+        // SAFETY: Caller of this function ensures that the control bytes are properly initialized.
+        let index: usize = self.find_insert_slot(hash).index;
+        // SAFETY:
+        // 1. The `find_insert_slot` function either returns an `index` less than or
+        //    equal to `self.buckets() = self.bucket_mask + 1` of the table, or never
+        //    returns if it cannot find an empty or deleted slot.
+        // 2. The caller of this function guarantees that the table has already been
+        //    allocated
         let old_ctrl = *self.ctrl(index);
         self.set_ctrl_h2(index, hash);
         (index, old_ctrl)
     }
 
     /// Searches for an empty or deleted bucket which is suitable for inserting
-    /// a new element.
+    /// a new element, returning the `index` for the new [`Bucket`].
+    ///
+    /// This function does not make any changes to the `data` part of the table, or any
+    /// changes to the `items` or `growth_left` field of the table.
+    ///
+    /// The table must have at least 1 empty or deleted `bucket`, otherwise this function
+    /// will never return (will go into an infinite loop) for tables larger than the group
+    /// width, or return an index outside of the table indices range if the table is less
+    /// than the group width.
+    ///
+    /// If there is at least 1 empty or deleted `bucket` in the table, the function is
+    /// guaranteed to return [`InsertSlot`] with an index in the range `0..self.buckets()`,
+    /// but in any case, if this function returns [`InsertSlot`], it will contain an index
+    /// in the range `0..=self.buckets()`.
     ///
-    /// There must be at least 1 empty bucket in the table.
+    /// # Safety
+    ///
+    /// The [`RawTableInner`] must have properly initialized control bytes otherwise calling
+    /// this function results in [`undefined behavior`].
+    ///
+    /// Attempt to write data at the [`InsertSlot`] returned by this function when the table is
+    /// less than the group width and if there was not at least one empty or deleted bucket in
+    /// the table will cause immediate [`undefined behavior`]. This is because in this case the
+    /// function will return `self.bucket_mask + 1` as an index due to the trailing [`EMPTY]
+    /// control bytes outside the table range.
+    ///
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
     #[inline]
-    fn find_insert_slot(&self, hash: u64) -> usize {
+    unsafe fn find_insert_slot(&self, hash: u64) -> InsertSlot {
         let mut probe_seq = self.probe_seq(hash);
         loop {
-            unsafe {
-                let group = Group::load(self.ctrl(probe_seq.pos));
-                if let Some(bit) = group.match_empty_or_deleted().lowest_set_bit() {
-                    let result = (probe_seq.pos + bit) & self.bucket_mask;
-
-                    // In tables smaller than the group width, trailing control
-                    // bytes outside the range of the table are filled with
-                    // EMPTY entries. These will unfortunately trigger a
-                    // match, but once masked may point to a full bucket that
-                    // is already occupied. We detect this situation here and
-                    // perform a second scan starting at the beginning of the
-                    // table. This second scan is guaranteed to find an empty
-                    // slot (due to the load factor) before hitting the trailing
-                    // control bytes (containing EMPTY).
-                    if unlikely(is_full(*self.ctrl(result))) {
-                        debug_assert!(self.bucket_mask < Group::WIDTH);
-                        debug_assert_ne!(probe_seq.pos, 0);
-                        return Group::load_aligned(self.ctrl(0))
-                            .match_empty_or_deleted()
-                            .lowest_set_bit_nonzero();
-                    }
+            // SAFETY:
+            // * Caller of this function ensures that the control bytes are properly initialized.
+            //
+            // * `ProbeSeq.pos` cannot be greater than `self.bucket_mask = self.buckets() - 1`
+            //   of the table due to masking with `self.bucket_mask` and also because mumber of
+            //   buckets is a power of two (see `self.probe_seq` function).
+            //
+            // * Even if `ProbeSeq.pos` returns `position == self.bucket_mask`, it is safe to
+            //   call `Group::load` due to the extended control bytes range, which is
+            //  `self.bucket_mask + 1 + Group::WIDTH` (in fact, this means that the last control
+            //   byte will never be read for the allocated table);
+            //
+            // * Also, even if `RawTableInner` is not already allocated, `ProbeSeq.pos` will
+            //   always return "0" (zero), so Group::load will read unaligned `Group::static_empty()`
+            //   bytes, which is safe (see RawTableInner::new).
+            let group = unsafe { Group::load(self.ctrl(probe_seq.pos)) };
 
-                    return result;
+            let index = self.find_insert_slot_in_group(&group, &probe_seq);
+            if likely(index.is_some()) {
+                // SAFETY:
+                // * Caller of this function ensures that the control bytes are properly initialized.
+                //
+                // * We use this function with the slot / index found by `self.find_insert_slot_in_group`
+                unsafe {
+                    return self.fix_insert_slot(index.unwrap_unchecked());
                 }
             }
             probe_seq.move_next(self.bucket_mask);
         }
     }
 
-    /// Searches for an element in the table. This uses dynamic dispatch to reduce the amount of
-    /// code generated, but it is eliminated by LLVM optimizations.
-    #[inline]
-    fn find_inner(&self, hash: u64, eq: &mut dyn FnMut(usize) -> bool) -> Option<usize> {
+    /// Searches for an element in a table, returning the `index` of the found element.
+    /// This uses dynamic dispatch to reduce the amount of code generated, but it is
+    /// eliminated by LLVM optimizations.
+    ///
+    /// This function does not make any changes to the `data` part of the table, or any
+    /// changes to the `items` or `growth_left` field of the table.
+    ///
+    /// The table must have at least 1 empty `bucket`, otherwise, if the
+    /// `eq: &mut dyn FnMut(usize) -> bool` function does not return `true`,
+    /// this function will also never return (will go into an infinite loop).
+    ///
+    /// This function is guaranteed to provide the `eq: &mut dyn FnMut(usize) -> bool`
+    /// function with only `FULL` buckets' indices and return the `index` of the found
+    /// element as `Some(index)`, so the index will always be in the range
+    /// `0..self.buckets()`.
+    ///
+    /// # Safety
+    ///
+    /// The [`RawTableInner`] must have properly initialized control bytes otherwise calling
+    /// this function results in [`undefined behavior`].
+    ///
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
+    #[inline(always)]
+    unsafe fn find_inner(&self, hash: u64, eq: &mut dyn FnMut(usize) -> bool) -> Option<usize> {
         let h2_hash = h2(hash);
         let mut probe_seq = self.probe_seq(hash);
 
         loop {
+            // SAFETY:
+            // * Caller of this function ensures that the control bytes are properly initialized.
+            //
+            // * `ProbeSeq.pos` cannot be greater than `self.bucket_mask = self.buckets() - 1`
+            //   of the table due to masking with `self.bucket_mask`.
+            //
+            // * Even if `ProbeSeq.pos` returns `position == self.bucket_mask`, it is safe to
+            //   call `Group::load` due to the extended control bytes range, which is
+            //  `self.bucket_mask + 1 + Group::WIDTH` (in fact, this means that the last control
+            //   byte will never be read for the allocated table);
+            //
+            // * Also, even if `RawTableInner` is not already allocated, `ProbeSeq.pos` will
+            //   always return "0" (zero), so Group::load will read unaligned `Group::static_empty()`
+            //   bytes, which is safe (see RawTableInner::new_in).
             let group = unsafe { Group::load(self.ctrl(probe_seq.pos)) };
 
             for bit in group.match_byte(h2_hash) {
+                // This is the same as `(probe_seq.pos + bit) % self.buckets()` because the number
+                // of buckets is a power of two, and `self.bucket_mask = self.buckets() - 1`.
                 let index = (probe_seq.pos + bit) & self.bucket_mask;
 
                 if likely(eq(index)) {
@@ -1189,12 +2202,52 @@ impl<A: Allocator + Clone> RawTableInner<A> {
         }
     }
 
+    /// Prepares for rehashing data in place (that is, without allocating new memory).
+    /// Converts all full index `control bytes` to `DELETED` and all `DELETED` control
+    /// bytes to `EMPTY`, i.e. performs the following conversion:
+    ///
+    /// - `EMPTY` control bytes   -> `EMPTY`;
+    /// - `DELETED` control bytes -> `EMPTY`;
+    /// - `FULL` control bytes    -> `DELETED`.
+    ///
+    /// This function does not make any changes to the `data` parts of the table,
+    /// or any changes to the `items` or `growth_left` field of the table.
+    ///
+    /// # Safety
+    ///
+    /// You must observe the following safety rules when calling this function:
+    ///
+    /// * The [`RawTableInner`] has already been allocated;
+    ///
+    /// * The caller of this function must convert the `DELETED` bytes back to `FULL`
+    ///   bytes when re-inserting them into their ideal position (which was impossible
+    ///   to do during the first insert due to tombstones). If the caller does not do
+    ///   this, then calling this function may result in a memory leak.
+    ///
+    /// * The [`RawTableInner`] must have properly initialized control bytes otherwise
+    ///   calling this function results in [`undefined behavior`].
+    ///
+    /// Calling this function on a table that has not been allocated results in
+    /// [`undefined behavior`].
+    ///
+    /// See also [`Bucket::as_ptr`] method, for more information about of properly removing
+    /// or saving `data element` from / into the [`RawTable`] / [`RawTableInner`].
+    ///
+    /// [`Bucket::as_ptr`]: Bucket::as_ptr
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
     #[allow(clippy::mut_mut)]
     #[inline]
     unsafe fn prepare_rehash_in_place(&mut self) {
-        // Bulk convert all full control bytes to DELETED, and all DELETED
-        // control bytes to EMPTY. This effectively frees up all buckets
-        // containing a DELETED entry.
+        // Bulk convert all full control bytes to DELETED, and all DELETED control bytes to EMPTY.
+        // This effectively frees up all buckets containing a DELETED entry.
+        //
+        // SAFETY:
+        // 1. `i` is guaranteed to be within bounds since we are iterating from zero to `buckets - 1`;
+        // 2. Even if `i` will be `i == self.bucket_mask`, it is safe to call `Group::load_aligned`
+        //    due to the extended control bytes range, which is `self.bucket_mask + 1 + Group::WIDTH`;
+        // 3. The caller of this function guarantees that [`RawTableInner`] has already been allocated;
+        // 4. We can use `Group::load_aligned` and `Group::store_aligned` here since we start from 0
+        //    and go to the end with a step equal to `Group::WIDTH` (see TableLayout::calculate_layout_for).
         for i in (0..self.buckets()).step_by(Group::WIDTH) {
             let group = Group::load_aligned(self.ctrl(i));
             let group = group.convert_special_to_empty_and_full_to_deleted();
@@ -1203,15 +2256,245 @@ impl<A: Allocator + Clone> RawTableInner<A> {
 
         // Fix up the trailing control bytes. See the comments in set_ctrl
         // for the handling of tables smaller than the group width.
-        if self.buckets() < Group::WIDTH {
+        //
+        // SAFETY: The caller of this function guarantees that [`RawTableInner`]
+        // has already been allocated
+        if unlikely(self.buckets() < Group::WIDTH) {
+            // SAFETY: We have `self.bucket_mask + 1 + Group::WIDTH` number of control bytes,
+            // so copying `self.buckets() == self.bucket_mask + 1` bytes with offset equal to
+            // `Group::WIDTH` is safe
             self.ctrl(0)
                 .copy_to(self.ctrl(Group::WIDTH), self.buckets());
         } else {
+            // SAFETY: We have `self.bucket_mask + 1 + Group::WIDTH` number of
+            // control bytes,so copying `Group::WIDTH` bytes with offset equal
+            // to `self.buckets() == self.bucket_mask + 1` is safe
             self.ctrl(0)
                 .copy_to(self.ctrl(self.buckets()), Group::WIDTH);
         }
     }
 
+    /// Returns an iterator over every element in the table.
+    ///
+    /// # Safety
+    ///
+    /// If any of the following conditions are violated, the result
+    /// is [`undefined behavior`]:
+    ///
+    /// * The caller has to ensure that the `RawTableInner` outlives the
+    ///   `RawIter`. Because we cannot make the `next` method unsafe on
+    ///   the `RawIter` struct, we have to make the `iter` method unsafe.
+    ///
+    /// * The [`RawTableInner`] must have properly initialized control bytes.
+    ///
+    /// The type `T` must be the actual type of the elements stored in the table,
+    /// otherwise using the returned [`RawIter`] results in [`undefined behavior`].
+    ///
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
+    #[inline]
+    unsafe fn iter<T>(&self) -> RawIter<T> {
+        // SAFETY:
+        // 1. Since the caller of this function ensures that the control bytes
+        //    are properly initialized and `self.data_end()` points to the start
+        //    of the array of control bytes, therefore: `ctrl` is valid for reads,
+        //    properly aligned to `Group::WIDTH` and points to the properly initialized
+        //    control bytes.
+        // 2. `data` bucket index in the table is equal to the `ctrl` index (i.e.
+        //    equal to zero).
+        // 3. We pass the exact value of buckets of the table to the function.
+        //
+        //                         `ctrl` points here (to the start
+        //                         of the first control byte `CT0`)
+        //                          ∨
+        // [Pad], T_n, ..., T1, T0, |CT0, CT1, ..., CT_n|, CTa_0, CTa_1, ..., CTa_m
+        //                           \________  ________/
+        //                                    \/
+        //       `n = buckets - 1`, i.e. `RawTableInner::buckets() - 1`
+        //
+        // where: T0...T_n  - our stored data;
+        //        CT0...CT_n - control bytes or metadata for `data`.
+        //        CTa_0...CTa_m - additional control bytes, where `m = Group::WIDTH - 1` (so that the search
+        //                        with loading `Group` bytes from the heap works properly, even if the result
+        //                        of `h1(hash) & self.bucket_mask` is equal to `self.bucket_mask`). See also
+        //                        `RawTableInner::set_ctrl` function.
+        //
+        // P.S. `h1(hash) & self.bucket_mask` is the same as `hash as usize % self.buckets()` because the number
+        // of buckets is a power of two, and `self.bucket_mask = self.buckets() - 1`.
+        let data = Bucket::from_base_index(self.data_end(), 0);
+        RawIter {
+            // SAFETY: See explanation above
+            iter: RawIterRange::new(self.ctrl.as_ptr(), data, self.buckets()),
+            items: self.items,
+        }
+    }
+
+    /// Executes the destructors (if any) of the values stored in the table.
+    ///
+    /// # Note
+    ///
+    /// This function does not erase the control bytes of the table and does
+    /// not make any changes to the `items` or `growth_left` fields of the
+    /// table. If necessary, the caller of this function must manually set
+    /// up these table fields, for example using the [`clear_no_drop`] function.
+    ///
+    /// Be careful during calling this function, because drop function of
+    /// the elements can panic, and this can leave table in an inconsistent
+    /// state.
+    ///
+    /// # Safety
+    ///
+    /// The type `T` must be the actual type of the elements stored in the table,
+    /// otherwise calling this function may result in [`undefined behavior`].
+    ///
+    /// If `T` is a type that should be dropped and **the table is not empty**,
+    /// calling this function more than once results in [`undefined behavior`].
+    ///
+    /// If `T` is not [`Copy`], attempting to use values stored in the table after
+    /// calling this function may result in [`undefined behavior`].
+    ///
+    /// It is safe to call this function on a table that has not been allocated,
+    /// on a table with uninitialized control bytes, and on a table with no actual
+    /// data but with `Full` control bytes if `self.items == 0`.
+    ///
+    /// See also [`Bucket::drop`] / [`Bucket::as_ptr`] methods, for more information
+    /// about of properly removing or saving `element` from / into the [`RawTable`] /
+    /// [`RawTableInner`].
+    ///
+    /// [`Bucket::drop`]: Bucket::drop
+    /// [`Bucket::as_ptr`]: Bucket::as_ptr
+    /// [`clear_no_drop`]: RawTableInner::clear_no_drop
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
+    unsafe fn drop_elements<T>(&mut self) {
+        // Check that `self.items != 0`. Protects against the possibility
+        // of creating an iterator on an table with uninitialized control bytes.
+        if T::NEEDS_DROP && self.items != 0 {
+            // SAFETY: We know for sure that RawTableInner will outlive the
+            // returned `RawIter` iterator, and the caller of this function
+            // must uphold the safety contract for `drop_elements` method.
+            for item in self.iter::<T>() {
+                // SAFETY: The caller must uphold the safety contract for
+                // `drop_elements` method.
+                item.drop();
+            }
+        }
+    }
+
+    /// Executes the destructors (if any) of the values stored in the table and than
+    /// deallocates the table.
+    ///
+    /// # Note
+    ///
+    /// Calling this function automatically makes invalid (dangling) all instances of
+    /// buckets ([`Bucket`]) and makes invalid (dangling) the `ctrl` field of the table.
+    ///
+    /// This function does not make any changes to the `bucket_mask`, `items` or `growth_left`
+    /// fields of the table. If necessary, the caller of this function must manually set
+    /// up these table fields.
+    ///
+    /// # Safety
+    ///
+    /// If any of the following conditions are violated, the result is [`undefined behavior`]:
+    ///
+    /// * Calling this function more than once;
+    ///
+    /// * The type `T` must be the actual type of the elements stored in the table.
+    ///
+    /// * The `alloc` must be the same [`Allocator`] as the `Allocator` that was used
+    ///   to allocate this table.
+    ///
+    /// * The `table_layout` must be the same [`TableLayout`] as the `TableLayout` that
+    ///   was used to allocate this table.
+    ///
+    /// The caller of this function should pay attention to the possibility of the
+    /// elements' drop function panicking, because this:
+    ///
+    ///    * May leave the table in an inconsistent state;
+    ///
+    ///    * Memory is never deallocated, so a memory leak may occur.
+    ///
+    /// Attempt to use the `ctrl` field of the table (dereference) after calling this
+    /// function results in [`undefined behavior`].
+    ///
+    /// It is safe to call this function on a table that has not been allocated,
+    /// on a table with uninitialized control bytes, and on a table with no actual
+    /// data but with `Full` control bytes if `self.items == 0`.
+    ///
+    /// See also [`RawTableInner::drop_elements`] or [`RawTableInner::free_buckets`]
+    /// for more  information.
+    ///
+    /// [`RawTableInner::drop_elements`]: RawTableInner::drop_elements
+    /// [`RawTableInner::free_buckets`]: RawTableInner::free_buckets
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
+    unsafe fn drop_inner_table<T, A: Allocator>(&mut self, alloc: &A, table_layout: TableLayout) {
+        if !self.is_empty_singleton() {
+            unsafe {
+                // SAFETY: The caller must uphold the safety contract for `drop_inner_table` method.
+                self.drop_elements::<T>();
+                // SAFETY:
+                // 1. We have checked that our table is allocated.
+                // 2. The caller must uphold the safety contract for `drop_inner_table` method.
+                self.free_buckets(alloc, table_layout);
+            }
+        }
+    }
+
+    /// Returns a pointer to an element in the table (convenience for
+    /// `Bucket::from_base_index(self.data_end::<T>(), index)`).
+    ///
+    /// The caller must ensure that the `RawTableInner` outlives the returned [`Bucket<T>`],
+    /// otherwise using it may result in [`undefined behavior`].
+    ///
+    /// # Safety
+    ///
+    /// If `mem::size_of::<T>() != 0`, then the safety rules are directly derived from the
+    /// safety rules of the [`Bucket::from_base_index`] function. Therefore, when calling
+    /// this function, the following safety rules must be observed:
+    ///
+    /// * The table must already be allocated;
+    ///
+    /// * The `index` must not be greater than the number returned by the [`RawTableInner::buckets`]
+    ///   function, i.e. `(index + 1) <= self.buckets()`.
+    ///
+    /// * The type `T` must be the actual type of the elements stored in the table, otherwise
+    ///   using the returned [`Bucket`] may result in [`undefined behavior`].
+    ///
+    /// It is safe to call this function with index of zero (`index == 0`) on a table that has
+    /// not been allocated, but using the returned [`Bucket`] results in [`undefined behavior`].
+    ///
+    /// If `mem::size_of::<T>() == 0`, then the only requirement is that the `index` must
+    /// not be greater than the number returned by the [`RawTable::buckets`] function, i.e.
+    /// `(index + 1) <= self.buckets()`.
+    ///
+    /// ```none
+    /// If mem::size_of::<T>() != 0 then return a pointer to the `element` in the `data part` of the table
+    /// (we start counting from "0", so that in the expression T[n], the "n" index actually one less than
+    /// the "buckets" number of our `RawTableInner`, i.e. "n = RawTableInner::buckets() - 1"):
+    ///
+    ///           `table.bucket(3).as_ptr()` returns a pointer that points here in the `data`
+    ///           part of the `RawTableInner`, i.e. to the start of T3 (see [`Bucket::as_ptr`])
+    ///                  |
+    ///                  |               `base = table.data_end::<T>()` points here
+    ///                  |               (to the start of CT0 or to the end of T0)
+    ///                  v                 v
+    /// [Pad], T_n, ..., |T3|, T2, T1, T0, |CT0, CT1, CT2, CT3, ..., CT_n, CTa_0, CTa_1, ..., CTa_m
+    ///                     ^                                              \__________  __________/
+    ///        `table.bucket(3)` returns a pointer that points                        \/
+    ///         here in the `data` part of the `RawTableInner`             additional control bytes
+    ///         (to the end of T3)                                          `m = Group::WIDTH - 1`
+    ///
+    /// where: T0...T_n  - our stored data;
+    ///        CT0...CT_n - control bytes or metadata for `data`;
+    ///        CTa_0...CTa_m - additional control bytes (so that the search with loading `Group` bytes from
+    ///                        the heap works properly, even if the result of `h1(hash) & self.bucket_mask`
+    ///                        is equal to `self.bucket_mask`). See also `RawTableInner::set_ctrl` function.
+    ///
+    /// P.S. `h1(hash) & self.bucket_mask` is the same as `hash as usize % self.buckets()` because the number
+    /// of buckets is a power of two, and `self.bucket_mask = self.buckets() - 1`.
+    /// ```
+    ///
+    /// [`Bucket::from_base_index`]: Bucket::from_base_index
+    /// [`RawTableInner::buckets`]: RawTableInner::buckets
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
     #[inline]
     unsafe fn bucket<T>(&self, index: usize) -> Bucket<T> {
         debug_assert_ne!(self.bucket_mask, 0);
@@ -1219,6 +2502,52 @@ impl<A: Allocator + Clone> RawTableInner<A> {
         Bucket::from_base_index(self.data_end(), index)
     }
 
+    /// Returns a raw `*mut u8` pointer to the start of the `data` element in the table
+    /// (convenience for `self.data_end::<u8>().as_ptr().sub((index + 1) * size_of)`).
+    ///
+    /// The caller must ensure that the `RawTableInner` outlives the returned `*mut u8`,
+    /// otherwise using it may result in [`undefined behavior`].
+    ///
+    /// # Safety
+    ///
+    /// If any of the following conditions are violated, the result is [`undefined behavior`]:
+    ///
+    /// * The table must already be allocated;
+    ///
+    /// * The `index` must not be greater than the number returned by the [`RawTableInner::buckets`]
+    ///   function, i.e. `(index + 1) <= self.buckets()`;
+    ///
+    /// * The `size_of` must be equal to the size of the elements stored in the table;
+    ///
+    /// ```none
+    /// If mem::size_of::<T>() != 0 then return a pointer to the `element` in the `data part` of the table
+    /// (we start counting from "0", so that in the expression T[n], the "n" index actually one less than
+    /// the "buckets" number of our `RawTableInner`, i.e. "n = RawTableInner::buckets() - 1"):
+    ///
+    ///           `table.bucket_ptr(3, mem::size_of::<T>())` returns a pointer that points here in the
+    ///           `data` part of the `RawTableInner`, i.e. to the start of T3
+    ///                  |
+    ///                  |               `base = table.data_end::<u8>()` points here
+    ///                  |               (to the start of CT0 or to the end of T0)
+    ///                  v                 v
+    /// [Pad], T_n, ..., |T3|, T2, T1, T0, |CT0, CT1, CT2, CT3, ..., CT_n, CTa_0, CTa_1, ..., CTa_m
+    ///                                                                    \__________  __________/
+    ///                                                                               \/
+    ///                                                                    additional control bytes
+    ///                                                                     `m = Group::WIDTH - 1`
+    ///
+    /// where: T0...T_n  - our stored data;
+    ///        CT0...CT_n - control bytes or metadata for `data`;
+    ///        CTa_0...CTa_m - additional control bytes (so that the search with loading `Group` bytes from
+    ///                        the heap works properly, even if the result of `h1(hash) & self.bucket_mask`
+    ///                        is equal to `self.bucket_mask`). See also `RawTableInner::set_ctrl` function.
+    ///
+    /// P.S. `h1(hash) & self.bucket_mask` is the same as `hash as usize % self.buckets()` because the number
+    /// of buckets is a power of two, and `self.bucket_mask = self.buckets() - 1`.
+    /// ```
+    ///
+    /// [`RawTableInner::buckets`]: RawTableInner::buckets
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
     #[inline]
     unsafe fn bucket_ptr(&self, index: usize, size_of: usize) -> *mut u8 {
         debug_assert_ne!(self.bucket_mask, 0);
@@ -1227,9 +2556,44 @@ impl<A: Allocator + Clone> RawTableInner<A> {
         base.sub((index + 1) * size_of)
     }
 
+    /// Returns pointer to one past last `data` element in the table as viewed from
+    /// the start point of the allocation (convenience for `self.ctrl.cast()`).
+    ///
+    /// This function actually returns a pointer to the end of the `data element` at
+    /// index "0" (zero).
+    ///
+    /// The caller must ensure that the `RawTableInner` outlives the returned [`NonNull<T>`],
+    /// otherwise using it may result in [`undefined behavior`].
+    ///
+    /// # Note
+    ///
+    /// The type `T` must be the actual type of the elements stored in the table, otherwise
+    /// using the returned [`NonNull<T>`] may result in [`undefined behavior`].
+    ///
+    /// ```none
+    ///                        `table.data_end::<T>()` returns pointer that points here
+    ///                        (to the end of `T0`)
+    ///                          ∨
+    /// [Pad], T_n, ..., T1, T0, |CT0, CT1, ..., CT_n|, CTa_0, CTa_1, ..., CTa_m
+    ///                           \________  ________/
+    ///                                    \/
+    ///       `n = buckets - 1`, i.e. `RawTableInner::buckets() - 1`
+    ///
+    /// where: T0...T_n  - our stored data;
+    ///        CT0...CT_n - control bytes or metadata for `data`.
+    ///        CTa_0...CTa_m - additional control bytes, where `m = Group::WIDTH - 1` (so that the search
+    ///                        with loading `Group` bytes from the heap works properly, even if the result
+    ///                        of `h1(hash) & self.bucket_mask` is equal to `self.bucket_mask`). See also
+    ///                        `RawTableInner::set_ctrl` function.
+    ///
+    /// P.S. `h1(hash) & self.bucket_mask` is the same as `hash as usize % self.buckets()` because the number
+    /// of buckets is a power of two, and `self.bucket_mask = self.buckets() - 1`.
+    /// ```
+    ///
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
     #[inline]
-    unsafe fn data_end<T>(&self) -> NonNull<T> {
-        NonNull::new_unchecked(self.ctrl.as_ptr().cast())
+    fn data_end<T>(&self) -> NonNull<T> {
+        self.ctrl.cast()
     }
 
     /// Returns an iterator-like object for a probe sequence on the table.
@@ -1240,6 +2604,8 @@ impl<A: Allocator + Clone> RawTableInner<A> {
     #[inline]
     fn probe_seq(&self, hash: u64) -> ProbeSeq {
         ProbeSeq {
+            // This is the same as `hash as usize % self.buckets()` because the number
+            // of buckets is a power of two, and `self.bucket_mask = self.buckets() - 1`.
             pos: h1(hash) & self.bucket_mask,
             stride: 0,
         }
@@ -1250,7 +2616,7 @@ impl<A: Allocator + Clone> RawTableInner<A> {
     #[cfg(feature = "raw")]
     #[inline]
     unsafe fn prepare_insert_no_grow(&mut self, hash: u64) -> Result<usize, ()> {
-        let index = self.find_insert_slot(hash);
+        let index = self.find_insert_slot(hash).index;
         let old_ctrl = *self.ctrl(index);
         if unlikely(self.growth_left == 0 && special_is_empty(old_ctrl)) {
             Err(())
@@ -1277,13 +2643,68 @@ impl<A: Allocator + Clone> RawTableInner<A> {
 
     /// Sets a control byte to the hash, and possibly also the replicated control byte at
     /// the end of the array.
+    ///
+    /// This function does not make any changes to the `data` parts of the table,
+    /// or any changes to the `items` or `growth_left` field of the table.
+    ///
+    /// # Safety
+    ///
+    /// The safety rules are directly derived from the safety rules for [`RawTableInner::set_ctrl`]
+    /// method. Thus, in order to uphold the safety contracts for the method, you must observe the
+    /// following rules when calling this function:
+    ///
+    /// * The [`RawTableInner`] has already been allocated;
+    ///
+    /// * The `index` must not be greater than the `RawTableInner.bucket_mask`, i.e.
+    ///   `index <= RawTableInner.bucket_mask` or, in other words, `(index + 1)` must
+    ///   be no greater than the number returned by the function [`RawTableInner::buckets`].
+    ///
+    /// Calling this function on a table that has not been allocated results in [`undefined behavior`].
+    ///
+    /// See also [`Bucket::as_ptr`] method, for more information about of properly removing
+    /// or saving `data element` from / into the [`RawTable`] / [`RawTableInner`].
+    ///
+    /// [`RawTableInner::set_ctrl`]: RawTableInner::set_ctrl
+    /// [`RawTableInner::buckets`]: RawTableInner::buckets
+    /// [`Bucket::as_ptr`]: Bucket::as_ptr
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
     #[inline]
-    unsafe fn set_ctrl_h2(&self, index: usize, hash: u64) {
+    unsafe fn set_ctrl_h2(&mut self, index: usize, hash: u64) {
+        // SAFETY: The caller must uphold the safety rules for the [`RawTableInner::set_ctrl_h2`]
         self.set_ctrl(index, h2(hash));
     }
 
+    /// Replaces the hash in the control byte at the given index with the provided one,
+    /// and possibly also replicates the new control byte at the end of the array of control
+    /// bytes, returning the old control byte.
+    ///
+    /// This function does not make any changes to the `data` parts of the table,
+    /// or any changes to the `items` or `growth_left` field of the table.
+    ///
+    /// # Safety
+    ///
+    /// The safety rules are directly derived from the safety rules for [`RawTableInner::set_ctrl_h2`]
+    /// and [`RawTableInner::ctrl`] methods. Thus, in order to uphold the safety contracts for both
+    /// methods, you must observe the following rules when calling this function:
+    ///
+    /// * The [`RawTableInner`] has already been allocated;
+    ///
+    /// * The `index` must not be greater than the `RawTableInner.bucket_mask`, i.e.
+    ///   `index <= RawTableInner.bucket_mask` or, in other words, `(index + 1)` must
+    ///   be no greater than the number returned by the function [`RawTableInner::buckets`].
+    ///
+    /// Calling this function on a table that has not been allocated results in [`undefined behavior`].
+    ///
+    /// See also [`Bucket::as_ptr`] method, for more information about of properly removing
+    /// or saving `data element` from / into the [`RawTable`] / [`RawTableInner`].
+    ///
+    /// [`RawTableInner::set_ctrl_h2`]: RawTableInner::set_ctrl_h2
+    /// [`RawTableInner::buckets`]: RawTableInner::buckets
+    /// [`Bucket::as_ptr`]: Bucket::as_ptr
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
     #[inline]
-    unsafe fn replace_ctrl_h2(&self, index: usize, hash: u64) -> u8 {
+    unsafe fn replace_ctrl_h2(&mut self, index: usize, hash: u64) -> u8 {
+        // SAFETY: The caller must uphold the safety rules for the [`RawTableInner::replace_ctrl_h2`]
         let prev_ctrl = *self.ctrl(index);
         self.set_ctrl_h2(index, hash);
         prev_ctrl
@@ -1291,10 +2712,35 @@ impl<A: Allocator + Clone> RawTableInner<A> {
 
     /// Sets a control byte, and possibly also the replicated control byte at
     /// the end of the array.
+    ///
+    /// This function does not make any changes to the `data` parts of the table,
+    /// or any changes to the `items` or `growth_left` field of the table.
+    ///
+    /// # Safety
+    ///
+    /// You must observe the following safety rules when calling this function:
+    ///
+    /// * The [`RawTableInner`] has already been allocated;
+    ///
+    /// * The `index` must not be greater than the `RawTableInner.bucket_mask`, i.e.
+    ///   `index <= RawTableInner.bucket_mask` or, in other words, `(index + 1)` must
+    ///   be no greater than the number returned by the function [`RawTableInner::buckets`].
+    ///
+    /// Calling this function on a table that has not been allocated results in [`undefined behavior`].
+    ///
+    /// See also [`Bucket::as_ptr`] method, for more information about of properly removing
+    /// or saving `data element` from / into the [`RawTable`] / [`RawTableInner`].
+    ///
+    /// [`RawTableInner::buckets`]: RawTableInner::buckets
+    /// [`Bucket::as_ptr`]: Bucket::as_ptr
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
     #[inline]
-    unsafe fn set_ctrl(&self, index: usize, ctrl: u8) {
+    unsafe fn set_ctrl(&mut self, index: usize, ctrl: u8) {
         // Replicate the first Group::WIDTH control bytes at the end of
-        // the array without using a branch:
+        // the array without using a branch. If the tables smaller than
+        // the group width (self.buckets() < Group::WIDTH),
+        // `index2 = Group::WIDTH + index`, otherwise `index2` is:
+        //
         // - If index >= Group::WIDTH then index == index2.
         // - Otherwise index2 == self.bucket_mask + 1 + index.
         //
@@ -1311,16 +2757,43 @@ impl<A: Allocator + Clone> RawTableInner<A> {
         // ---------------------------------------------
         // | [A] | [B] | [EMPTY] | [EMPTY] | [A] | [B] |
         // ---------------------------------------------
+
+        // This is the same as `(index.wrapping_sub(Group::WIDTH)) % self.buckets() + Group::WIDTH`
+        // because the number of buckets is a power of two, and `self.bucket_mask = self.buckets() - 1`.
         let index2 = ((index.wrapping_sub(Group::WIDTH)) & self.bucket_mask) + Group::WIDTH;
 
+        // SAFETY: The caller must uphold the safety rules for the [`RawTableInner::set_ctrl`]
         *self.ctrl(index) = ctrl;
         *self.ctrl(index2) = ctrl;
     }
 
     /// Returns a pointer to a control byte.
+    ///
+    /// # Safety
+    ///
+    /// For the allocated [`RawTableInner`], the result is [`Undefined Behavior`],
+    /// if the `index` is greater than the `self.bucket_mask + 1 + Group::WIDTH`.
+    /// In that case, calling this function with `index == self.bucket_mask + 1 + Group::WIDTH`
+    /// will return a pointer to the end of the allocated table and it is useless on its own.
+    ///
+    /// Calling this function with `index >= self.bucket_mask + 1 + Group::WIDTH` on a
+    /// table that has not been allocated results in [`Undefined Behavior`].
+    ///
+    /// So to satisfy both requirements you should always follow the rule that
+    /// `index < self.bucket_mask + 1 + Group::WIDTH`
+    ///
+    /// Calling this function on [`RawTableInner`] that are not already allocated is safe
+    /// for read-only purpose.
+    ///
+    /// See also [`Bucket::as_ptr()`] method, for more information about of properly removing
+    /// or saving `data element` from / into the [`RawTable`] / [`RawTableInner`].
+    ///
+    /// [`Bucket::as_ptr()`]: Bucket::as_ptr()
+    /// [`Undefined Behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
     #[inline]
     unsafe fn ctrl(&self, index: usize) -> *mut u8 {
         debug_assert!(index < self.num_ctrl_bytes());
+        // SAFETY: The caller must uphold the safety rules for the [`RawTableInner::ctrl`]
         self.ctrl.as_ptr().add(index)
     }
 
@@ -1329,6 +2802,17 @@ impl<A: Allocator + Clone> RawTableInner<A> {
         self.bucket_mask + 1
     }
 
+    /// Checks whether the bucket at `index` is full.
+    ///
+    /// # Safety
+    ///
+    /// The caller must ensure `index` is less than the number of buckets.
+    #[inline]
+    unsafe fn is_bucket_full(&self, index: usize) -> bool {
+        debug_assert!(index < self.buckets());
+        is_full(*self.ctrl(index))
+    }
+
     #[inline]
     fn num_ctrl_bytes(&self) -> usize {
         self.bucket_mask + 1 + Group::WIDTH
@@ -1339,25 +2823,45 @@ impl<A: Allocator + Clone> RawTableInner<A> {
         self.bucket_mask == 0
     }
 
+    /// Attempts to allocate a new hash table with at least enough capacity
+    /// for inserting the given number of elements without reallocating,
+    /// and return it inside ScopeGuard to protect against panic in the hash
+    /// function.
+    ///
+    /// # Note
+    ///
+    /// It is recommended (but not required):
+    ///
+    /// * That the new table's `capacity` be greater than or equal to `self.items`.
+    ///
+    /// * The `alloc` is the same [`Allocator`] as the `Allocator` used
+    ///   to allocate this table.
+    ///
+    /// * The `table_layout` is the same [`TableLayout`] as the `TableLayout` used
+    ///   to allocate this table.
+    ///
+    /// If `table_layout` does not match the `TableLayout` that was used to allocate
+    /// this table, then using `mem::swap` with the `self` and the new table returned
+    /// by this function results in [`undefined behavior`].
+    ///
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
     #[allow(clippy::mut_mut)]
     #[inline]
-    unsafe fn prepare_resize(
+    fn prepare_resize<'a, A>(
         &self,
+        alloc: &'a A,
         table_layout: TableLayout,
         capacity: usize,
         fallibility: Fallibility,
-    ) -> Result<crate::scopeguard::ScopeGuard<Self, impl FnMut(&mut Self)>, TryReserveError> {
+    ) -> Result<crate::scopeguard::ScopeGuard<Self, impl FnMut(&mut Self) + 'a>, TryReserveError>
+    where
+        A: Allocator,
+    {
         debug_assert!(self.items <= capacity);
 
         // Allocate and initialize the new table.
-        let mut new_table = RawTableInner::fallible_with_capacity(
-            self.alloc.clone(),
-            table_layout,
-            capacity,
-            fallibility,
-        )?;
-        new_table.growth_left -= self.items;
-        new_table.items = self.items;
+        let new_table =
+            RawTableInner::fallible_with_capacity(alloc, table_layout, capacity, fallibility)?;
 
         // The hash function may panic, in which case we simply free the new
         // table without dropping any elements that may have been copied into
@@ -1367,7 +2871,11 @@ impl<A: Allocator + Clone> RawTableInner<A> {
         // the comment at the bottom of this function.
         Ok(guard(new_table, move |self_| {
             if !self_.is_empty_singleton() {
-                self_.free_buckets(table_layout);
+                // SAFETY:
+                // 1. We have checked that our table is allocated.
+                // 2. We know for sure that the `alloc` and `table_layout` matches the
+                //    [`Allocator`] and [`TableLayout`] used to allocate this table.
+                unsafe { self_.free_buckets(alloc, table_layout) };
             }
         }))
     }
@@ -1376,16 +2884,38 @@ impl<A: Allocator + Clone> RawTableInner<A> {
     ///
     /// This uses dynamic dispatch to reduce the amount of
     /// code generated, but it is eliminated by LLVM optimizations when inlined.
+    ///
+    /// # Safety
+    ///
+    /// If any of the following conditions are violated, the result is
+    /// [`undefined behavior`]:
+    ///
+    /// * The `alloc` must be the same [`Allocator`] as the `Allocator` used
+    ///   to allocate this table.
+    ///
+    /// * The `layout` must be the same [`TableLayout`] as the `TableLayout`
+    ///   used to allocate this table.
+    ///
+    /// * The `drop` function (`fn(*mut u8)`) must be the actual drop function of
+    ///   the elements stored in the table.
+    ///
+    /// * The [`RawTableInner`] must have properly initialized control bytes.
+    ///
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
     #[allow(clippy::inline_always)]
     #[inline(always)]
-    unsafe fn reserve_rehash_inner(
+    unsafe fn reserve_rehash_inner<A>(
         &mut self,
+        alloc: &A,
         additional: usize,
         hasher: &dyn Fn(&mut Self, usize) -> u64,
         fallibility: Fallibility,
         layout: TableLayout,
         drop: Option<fn(*mut u8)>,
-    ) -> Result<(), TryReserveError> {
+    ) -> Result<(), TryReserveError>
+    where
+        A: Allocator,
+    {
         // Avoid `Option::ok_or_else` because it bloats LLVM IR.
         let new_items = match self.items.checked_add(additional) {
             Some(new_items) => new_items,
@@ -1395,12 +2925,30 @@ impl<A: Allocator + Clone> RawTableInner<A> {
         if new_items <= full_capacity / 2 {
             // Rehash in-place without re-allocating if we have plenty of spare
             // capacity that is locked up due to DELETED entries.
+
+            // SAFETY:
+            // 1. We know for sure that `[`RawTableInner`]` has already been allocated
+            //    (since new_items <= full_capacity / 2);
+            // 2. The caller ensures that `drop` function is the actual drop function of
+            //    the elements stored in the table.
+            // 3. The caller ensures that `layout` matches the [`TableLayout`] that was
+            //    used to allocate this table.
+            // 4. The caller ensures that the control bytes of the `RawTableInner`
+            //    are already initialized.
             self.rehash_in_place(hasher, layout.size, drop);
             Ok(())
         } else {
             // Otherwise, conservatively resize to at least the next size up
             // to avoid churning deletes into frequent rehashes.
+            //
+            // SAFETY:
+            // 1. We know for sure that `capacity >= self.items`.
+            // 2. The caller ensures that `alloc` and `layout` matches the [`Allocator`] and
+            //    [`TableLayout`] that were used to allocate this table.
+            // 3. The caller ensures that the control bytes of the `RawTableInner`
+            //    are already initialized.
             self.resize_inner(
+                alloc,
                 usize::max(new_items, full_capacity + 1),
                 hasher,
                 fallibility,
@@ -1409,48 +2957,160 @@ impl<A: Allocator + Clone> RawTableInner<A> {
         }
     }
 
+    /// Returns an iterator over full buckets indices in the table.
+    ///
+    /// # Safety
+    ///
+    /// Behavior is undefined if any of the following conditions are violated:
+    ///
+    /// * The caller has to ensure that the `RawTableInner` outlives the
+    ///   `FullBucketsIndices`. Because we cannot make the `next` method
+    ///   unsafe on the `FullBucketsIndices` struct, we have to make the
+    ///   `full_buckets_indices` method unsafe.
+    ///
+    /// * The [`RawTableInner`] must have properly initialized control bytes.
+    #[inline(always)]
+    unsafe fn full_buckets_indices(&self) -> FullBucketsIndices {
+        // SAFETY:
+        // 1. Since the caller of this function ensures that the control bytes
+        //    are properly initialized and `self.ctrl(0)` points to the start
+        //    of the array of control bytes, therefore: `ctrl` is valid for reads,
+        //    properly aligned to `Group::WIDTH` and points to the properly initialized
+        //    control bytes.
+        // 2. The value of `items` is equal to the amount of data (values) added
+        //    to the table.
+        //
+        //                         `ctrl` points here (to the start
+        //                         of the first control byte `CT0`)
+        //                          ∨
+        // [Pad], T_n, ..., T1, T0, |CT0, CT1, ..., CT_n|, Group::WIDTH
+        //                           \________  ________/
+        //                                    \/
+        //       `n = buckets - 1`, i.e. `RawTableInner::buckets() - 1`
+        //
+        // where: T0...T_n  - our stored data;
+        //        CT0...CT_n - control bytes or metadata for `data`.
+        let ctrl = NonNull::new_unchecked(self.ctrl(0));
+
+        FullBucketsIndices {
+            // Load the first group
+            // SAFETY: See explanation above.
+            current_group: Group::load_aligned(ctrl.as_ptr()).match_full().into_iter(),
+            group_first_index: 0,
+            ctrl,
+            items: self.items,
+        }
+    }
+
     /// Allocates a new table of a different size and moves the contents of the
     /// current table into it.
     ///
     /// This uses dynamic dispatch to reduce the amount of
     /// code generated, but it is eliminated by LLVM optimizations when inlined.
+    ///
+    /// # Safety
+    ///
+    /// If any of the following conditions are violated, the result is
+    /// [`undefined behavior`]:
+    ///
+    /// * The `alloc` must be the same [`Allocator`] as the `Allocator` used
+    ///   to allocate this table;
+    ///
+    /// * The `layout` must be the same [`TableLayout`] as the `TableLayout`
+    ///   used to allocate this table;
+    ///
+    /// * The [`RawTableInner`] must have properly initialized control bytes.
+    ///
+    /// The caller of this function must ensure that `capacity >= self.items`
+    /// otherwise:
+    ///
+    /// * If `self.items != 0`, calling of this function with `capacity == 0`
+    ///   results in [`undefined behavior`].
+    ///
+    /// * If `capacity_to_buckets(capacity) < Group::WIDTH` and
+    ///   `self.items > capacity_to_buckets(capacity)` calling this function
+    ///   results in [`undefined behavior`].
+    ///
+    /// * If `capacity_to_buckets(capacity) >= Group::WIDTH` and
+    ///   `self.items > capacity_to_buckets(capacity)` calling this function
+    ///   are never return (will go into an infinite loop).
+    ///
+    /// Note: It is recommended (but not required) that the new table's `capacity`
+    /// be greater than or equal to `self.items`. In case if `capacity <= self.items`
+    /// this function can never return. See [`RawTableInner::find_insert_slot`] for
+    /// more information.
+    ///
+    /// [`RawTableInner::find_insert_slot`]: RawTableInner::find_insert_slot
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
     #[allow(clippy::inline_always)]
     #[inline(always)]
-    unsafe fn resize_inner(
+    unsafe fn resize_inner<A>(
         &mut self,
+        alloc: &A,
         capacity: usize,
         hasher: &dyn Fn(&mut Self, usize) -> u64,
         fallibility: Fallibility,
         layout: TableLayout,
-    ) -> Result<(), TryReserveError> {
-        let mut new_table = self.prepare_resize(layout, capacity, fallibility)?;
-
-        // Copy all elements to the new table.
-        for i in 0..self.buckets() {
-            if !is_full(*self.ctrl(i)) {
-                continue;
-            }
-
+    ) -> Result<(), TryReserveError>
+    where
+        A: Allocator,
+    {
+        // SAFETY: We know for sure that `alloc` and `layout` matches the [`Allocator`] and [`TableLayout`]
+        // that were used to allocate this table.
+        let mut new_table = self.prepare_resize(alloc, layout, capacity, fallibility)?;
+
+        // SAFETY: We know for sure that RawTableInner will outlive the
+        // returned `FullBucketsIndices` iterator, and the caller of this
+        // function ensures that the control bytes are properly initialized.
+        for full_byte_index in self.full_buckets_indices() {
             // This may panic.
-            let hash = hasher(self, i);
+            let hash = hasher(self, full_byte_index);
 
+            // SAFETY:
             // We can use a simpler version of insert() here since:
-            // - there are no DELETED entries.
-            // - we know there is enough space in the table.
-            // - all elements are unique.
-            let (index, _) = new_table.prepare_insert_slot(hash);
-
+            // 1. There are no DELETED entries.
+            // 2. We know there is enough space in the table.
+            // 3. All elements are unique.
+            // 4. The caller of this function guarantees that `capacity > 0`
+            //    so `new_table` must already have some allocated memory.
+            // 5. We set `growth_left` and `items` fields of the new table
+            //    after the loop.
+            // 6. We insert into the table, at the returned index, the data
+            //    matching the given hash immediately after calling this function.
+            let (new_index, _) = new_table.prepare_insert_slot(hash);
+
+            // SAFETY:
+            //
+            // * `src` is valid for reads of `layout.size` bytes, since the
+            //   table is alive and the `full_byte_index` is guaranteed to be
+            //   within bounds (see `FullBucketsIndices::next_impl`);
+            //
+            // * `dst` is valid for writes of `layout.size` bytes, since the
+            //   caller ensures that `table_layout` matches the [`TableLayout`]
+            //   that was used to allocate old table and we have the `new_index`
+            //   returned by `prepare_insert_slot`.
+            //
+            // * Both `src` and `dst` are properly aligned.
+            //
+            // * Both `src` and `dst` point to different region of memory.
             ptr::copy_nonoverlapping(
-                self.bucket_ptr(i, layout.size),
-                new_table.bucket_ptr(index, layout.size),
+                self.bucket_ptr(full_byte_index, layout.size),
+                new_table.bucket_ptr(new_index, layout.size),
                 layout.size,
             );
         }
 
+        // The hash function didn't panic, so we can safely set the
+        // `growth_left` and `items` fields of the new table.
+        new_table.growth_left -= self.items;
+        new_table.items = self.items;
+
         // We successfully copied all elements without panicking. Now replace
         // self with the new table. The old table will have its memory freed but
         // the items will not be dropped (since they have been moved into the
         // new table).
+        // SAFETY: The caller ensures that `table_layout` matches the [`TableLayout`]
+        // that was used to allocate this table.
         mem::swap(self, &mut new_table);
 
         Ok(())
@@ -1463,6 +3123,21 @@ impl<A: Allocator + Clone> RawTableInner<A> {
     ///
     /// This uses dynamic dispatch to reduce the amount of
     /// code generated, but it is eliminated by LLVM optimizations when inlined.
+    ///
+    /// # Safety
+    ///
+    /// If any of the following conditions are violated, the result is [`undefined behavior`]:
+    ///
+    /// * The `size_of` must be equal to the size of the elements stored in the table;
+    ///
+    /// * The `drop` function (`fn(*mut u8)`) must be the actual drop function of
+    ///   the elements stored in the table.
+    ///
+    /// * The [`RawTableInner`] has already been allocated;
+    ///
+    /// * The [`RawTableInner`] must have properly initialized control bytes.
+    ///
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
     #[allow(clippy::inline_always)]
     #[cfg_attr(feature = "inline-more", inline(always))]
     #[cfg_attr(not(feature = "inline-more"), inline)]
@@ -1506,8 +3181,10 @@ impl<A: Allocator + Clone> RawTableInner<A> {
                 let hash = hasher(*guard, i);
 
                 // Search for a suitable place to put it
-                let new_i = guard.find_insert_slot(hash);
-                let new_i_p = guard.bucket_ptr(new_i, size_of);
+                //
+                // SAFETY: Caller of this function ensures that the control bytes
+                // are properly initialized.
+                let new_i = guard.find_insert_slot(hash).index;
 
                 // Probing works by scanning through all of the control
                 // bytes in groups, which may not be aligned to the group
@@ -1519,6 +3196,8 @@ impl<A: Allocator + Clone> RawTableInner<A> {
                     continue 'outer;
                 }
 
+                let new_i_p = guard.bucket_ptr(new_i, size_of);
+
                 // We are moving the current item to a new position. Write
                 // our H2 to the control byte of the new position.
                 let prev_ctrl = guard.replace_ctrl_h2(new_i, hash);
@@ -1545,17 +3224,107 @@ impl<A: Allocator + Clone> RawTableInner<A> {
         mem::forget(guard);
     }
 
+    /// Deallocates the table without dropping any entries.
+    ///
+    /// # Note
+    ///
+    /// This function must be called only after [`drop_elements`](RawTableInner::drop_elements),
+    /// else it can lead to leaking of memory. Also calling this function automatically
+    /// makes invalid (dangling) all instances of buckets ([`Bucket`]) and makes invalid
+    /// (dangling) the `ctrl` field of the table.
+    ///
+    /// # Safety
+    ///
+    /// If any of the following conditions are violated, the result is [`Undefined Behavior`]:
+    ///
+    /// * The [`RawTableInner`] has already been allocated;
+    ///
+    /// * The `alloc` must be the same [`Allocator`] as the `Allocator` that was used
+    ///   to allocate this table.
+    ///
+    /// * The `table_layout` must be the same [`TableLayout`] as the `TableLayout` that was used
+    ///   to allocate this table.
+    ///
+    /// See also [`GlobalAlloc::dealloc`] or [`Allocator::deallocate`] for more  information.
+    ///
+    /// [`Undefined Behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
+    /// [`GlobalAlloc::dealloc`]: https://doc.rust-lang.org/alloc/alloc/trait.GlobalAlloc.html#tymethod.dealloc
+    /// [`Allocator::deallocate`]: https://doc.rust-lang.org/alloc/alloc/trait.Allocator.html#tymethod.deallocate
+    #[inline]
+    unsafe fn free_buckets<A>(&mut self, alloc: &A, table_layout: TableLayout)
+    where
+        A: Allocator,
+    {
+        // SAFETY: The caller must uphold the safety contract for `free_buckets`
+        // method.
+        let (ptr, layout) = self.allocation_info(table_layout);
+        alloc.deallocate(ptr, layout);
+    }
+
+    /// Returns a pointer to the allocated memory and the layout that was used to
+    /// allocate the table.
+    ///
+    /// # Safety
+    ///
+    /// Caller of this function must observe the following safety rules:
+    ///
+    /// * The [`RawTableInner`] has already been allocated, otherwise
+    ///   calling this function results in [`undefined behavior`]
+    ///
+    /// * The `table_layout` must be the same [`TableLayout`] as the `TableLayout`
+    ///   that was used to allocate this table. Failure to comply with this condition
+    ///   may result in [`undefined behavior`].
+    ///
+    /// See also [`GlobalAlloc::dealloc`] or [`Allocator::deallocate`] for more  information.
+    ///
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
+    /// [`GlobalAlloc::dealloc`]: https://doc.rust-lang.org/alloc/alloc/trait.GlobalAlloc.html#tymethod.dealloc
+    /// [`Allocator::deallocate`]: https://doc.rust-lang.org/alloc/alloc/trait.Allocator.html#tymethod.deallocate
     #[inline]
-    unsafe fn free_buckets(&mut self, table_layout: TableLayout) {
+    unsafe fn allocation_info(&self, table_layout: TableLayout) -> (NonNull<u8>, Layout) {
+        debug_assert!(
+            !self.is_empty_singleton(),
+            "this function can only be called on non-empty tables"
+        );
+
         // Avoid `Option::unwrap_or_else` because it bloats LLVM IR.
         let (layout, ctrl_offset) = match table_layout.calculate_layout_for(self.buckets()) {
             Some(lco) => lco,
-            None => hint::unreachable_unchecked(),
+            None => unsafe { hint::unreachable_unchecked() },
         };
-        self.alloc.deallocate(
-            NonNull::new_unchecked(self.ctrl.as_ptr().sub(ctrl_offset)),
+        (
+            // SAFETY: The caller must uphold the safety contract for `allocation_info` method.
+            unsafe { NonNull::new_unchecked(self.ctrl.as_ptr().sub(ctrl_offset)) },
             layout,
-        );
+        )
+    }
+
+    /// Returns a pointer to the allocated memory and the layout that was used to
+    /// allocate the table. If [`RawTableInner`] has not been allocated, this
+    /// function return `dangling` pointer and `()` (unit) layout.
+    ///
+    /// # Safety
+    ///
+    /// The `table_layout` must be the same [`TableLayout`] as the `TableLayout`
+    /// that was used to allocate this table. Failure to comply with this condition
+    /// may result in [`undefined behavior`].
+    ///
+    /// See also [`GlobalAlloc::dealloc`] or [`Allocator::deallocate`] for more  information.
+    ///
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
+    /// [`GlobalAlloc::dealloc`]: https://doc.rust-lang.org/alloc/alloc/trait.GlobalAlloc.html#tymethod.dealloc
+    /// [`Allocator::deallocate`]: https://doc.rust-lang.org/alloc/alloc/trait.Allocator.html#tymethod.deallocate
+    #[cfg(feature = "raw")]
+    unsafe fn allocation_info_or_zero(&self, table_layout: TableLayout) -> (NonNull<u8>, Layout) {
+        if self.is_empty_singleton() {
+            (NonNull::dangling(), Layout::new::<()>())
+        } else {
+            // SAFETY:
+            // 1. We have checked that our table is allocated.
+            // 2. The caller ensures that `table_layout` matches the [`TableLayout`]
+            // that was used to allocate this table.
+            unsafe { self.allocation_info(table_layout) }
+        }
     }
 
     /// Marks all table buckets as empty without dropping their contents.
@@ -1570,27 +3339,95 @@ impl<A: Allocator + Clone> RawTableInner<A> {
         self.growth_left = bucket_mask_to_capacity(self.bucket_mask);
     }
 
+    /// Erases the [`Bucket`]'s control byte at the given index so that it does not
+    /// triggered as full, decreases the `items` of the table and, if it can be done,
+    /// increases `self.growth_left`.
+    ///
+    /// This function does not actually erase / drop the [`Bucket`] itself, i.e. it
+    /// does not make any changes to the `data` parts of the table. The caller of this
+    /// function must take care to properly drop the `data`, otherwise calling this
+    /// function may result in a memory leak.
+    ///
+    /// # Safety
+    ///
+    /// You must observe the following safety rules when calling this function:
+    ///
+    /// * The [`RawTableInner`] has already been allocated;
+    ///
+    /// * It must be the full control byte at the given position;
+    ///
+    /// * The `index` must not be greater than the `RawTableInner.bucket_mask`, i.e.
+    ///   `index <= RawTableInner.bucket_mask` or, in other words, `(index + 1)` must
+    ///   be no greater than the number returned by the function [`RawTableInner::buckets`].
+    ///
+    /// Calling this function on a table that has not been allocated results in [`undefined behavior`].
+    ///
+    /// Calling this function on a table with no elements is unspecified, but calling subsequent
+    /// functions is likely to result in [`undefined behavior`] due to overflow subtraction
+    /// (`self.items -= 1 cause overflow when self.items == 0`).
+    ///
+    /// See also [`Bucket::as_ptr`] method, for more information about of properly removing
+    /// or saving `data element` from / into the [`RawTable`] / [`RawTableInner`].
+    ///
+    /// [`RawTableInner::buckets`]: RawTableInner::buckets
+    /// [`Bucket::as_ptr`]: Bucket::as_ptr
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
     #[inline]
     unsafe fn erase(&mut self, index: usize) {
-        debug_assert!(is_full(*self.ctrl(index)));
+        debug_assert!(self.is_bucket_full(index));
+
+        // This is the same as `index.wrapping_sub(Group::WIDTH) % self.buckets()` because
+        // the number of buckets is a power of two, and `self.bucket_mask = self.buckets() - 1`.
         let index_before = index.wrapping_sub(Group::WIDTH) & self.bucket_mask;
+        // SAFETY:
+        // - The caller must uphold the safety contract for `erase` method;
+        // - `index_before` is guaranteed to be in range due to masking with `self.bucket_mask`
         let empty_before = Group::load(self.ctrl(index_before)).match_empty();
         let empty_after = Group::load(self.ctrl(index)).match_empty();
 
-        // If we are inside a continuous block of Group::WIDTH full or deleted
-        // cells then a probe window may have seen a full block when trying to
-        // insert. We therefore need to keep that block non-empty so that
-        // lookups will continue searching to the next probe window.
+        // Inserting and searching in the map is performed by two key functions:
+        //
+        // - The `find_insert_slot` function that looks up the index of any `EMPTY` or `DELETED`
+        //   slot in a group to be able to insert. If it doesn't find an `EMPTY` or `DELETED`
+        //   slot immediately in the first group, it jumps to the next `Group` looking for it,
+        //   and so on until it has gone through all the groups in the control bytes.
+        //
+        // - The `find_inner` function that looks for the index of the desired element by looking
+        //   at all the `FULL` bytes in the group. If it did not find the element right away, and
+        //   there is no `EMPTY` byte in the group, then this means that the `find_insert_slot`
+        //   function may have found a suitable slot in the next group. Therefore, `find_inner`
+        //   jumps further, and if it does not find the desired element and again there is no `EMPTY`
+        //   byte, then it jumps further, and so on. The search stops only if `find_inner` function
+        //   finds the desired element or hits an `EMPTY` slot/byte.
+        //
+        // Accordingly, this leads to two consequences:
         //
-        // Note that in this context `leading_zeros` refers to the bytes at the
-        // end of a group, while `trailing_zeros` refers to the bytes at the
-        // beginning of a group.
+        // - The map must have `EMPTY` slots (bytes);
+        //
+        // - You can't just mark the byte to be erased as `EMPTY`, because otherwise the `find_inner`
+        //   function may stumble upon an `EMPTY` byte before finding the desired element and stop
+        //   searching.
+        //
+        // Thus it is necessary to check all bytes after and before the erased element. If we are in
+        // a contiguous `Group` of `FULL` or `DELETED` bytes (the number of `FULL` or `DELETED` bytes
+        // before and after is greater than or equal to `Group::WIDTH`), then we must mark our byte as
+        // `DELETED` in order for the `find_inner` function to go further. On the other hand, if there
+        // is at least one `EMPTY` slot in the `Group`, then the `find_inner` function will still stumble
+        // upon an `EMPTY` byte, so we can safely mark our erased byte as `EMPTY` as well.
+        //
+        // Finally, since `index_before == (index.wrapping_sub(Group::WIDTH) & self.bucket_mask) == index`
+        // and given all of the above, tables smaller than the group width (self.buckets() < Group::WIDTH)
+        // cannot have `DELETED` bytes.
+        //
+        // Note that in this context `leading_zeros` refers to the bytes at the end of a group, while
+        // `trailing_zeros` refers to the bytes at the beginning of a group.
         let ctrl = if empty_before.leading_zeros() + empty_after.trailing_zeros() >= Group::WIDTH {
             DELETED
         } else {
             self.growth_left += 1;
             EMPTY
         };
+        // SAFETY: the caller must uphold the safety contract for `erase` method.
         self.set_ctrl(index, ctrl);
         self.items -= 1;
     }
@@ -1599,12 +3436,16 @@ impl<A: Allocator + Clone> RawTableInner<A> {
 impl<T: Clone, A: Allocator + Clone> Clone for RawTable<T, A> {
     fn clone(&self) -> Self {
         if self.table.is_empty_singleton() {
-            Self::new_in(self.table.alloc.clone())
+            Self::new_in(self.alloc.clone())
         } else {
             unsafe {
                 // Avoid `Result::ok_or_else` because it bloats LLVM IR.
-                let new_table = match Self::new_uninitialized(
-                    self.table.alloc.clone(),
+                //
+                // SAFETY: This is safe as we are taking the size of an already allocated table
+                // and therefore сapacity overflow cannot occur, `self.table.buckets()` is power
+                // of two and all allocator errors will be caught inside `RawTableInner::new_uninitialized`.
+                let mut new_table = match Self::new_uninitialized(
+                    self.alloc.clone(),
                     self.table.buckets(),
                     Fallibility::Infallible,
                 ) {
@@ -1612,24 +3453,32 @@ impl<T: Clone, A: Allocator + Clone> Clone for RawTable<T, A> {
                     Err(_) => hint::unreachable_unchecked(),
                 };
 
-                // If cloning fails then we need to free the allocation for the
-                // new table. However we don't run its drop since its control
-                // bytes are not initialized yet.
-                let mut guard = guard(ManuallyDrop::new(new_table), |new_table| {
-                    new_table.free_buckets();
-                });
-
-                guard.clone_from_spec(self);
-
-                // Disarm the scope guard and return the newly created table.
-                ManuallyDrop::into_inner(ScopeGuard::into_inner(guard))
+                // Cloning elements may fail (the clone function may panic). But we don't
+                // need to worry about uninitialized control bits, since:
+                // 1. The number of items (elements) in the table is zero, which means that
+                //    the control bits will not be readed by Drop function.
+                // 2. The `clone_from_spec` method will first copy all control bits from
+                //    `self` (thus initializing them). But this will not affect the `Drop`
+                //    function, since the `clone_from_spec` function sets `items` only after
+                //    successfully clonning all elements.
+                new_table.clone_from_spec(self);
+                new_table
             }
         }
     }
 
     fn clone_from(&mut self, source: &Self) {
         if source.table.is_empty_singleton() {
-            *self = Self::new_in(self.table.alloc.clone());
+            let mut old_inner = mem::replace(&mut self.table, RawTableInner::NEW);
+            unsafe {
+                // SAFETY:
+                // 1. We call the function only once;
+                // 2. We know for sure that `alloc` and `table_layout` matches the [`Allocator`]
+                //    and [`TableLayout`] that were used to allocate this table.
+                // 3. If any elements' drop function panics, then there will only be a memory leak,
+                //    because we have replaced the inner table with a new one.
+                old_inner.drop_inner_table::<T, _>(&self.alloc, Self::TABLE_LAYOUT);
+            }
         } else {
             unsafe {
                 // Make sure that if any panics occurs, we clear the table and
@@ -1644,27 +3493,38 @@ impl<T: Clone, A: Allocator + Clone> Clone for RawTable<T, A> {
                 //
                 // This leak is unavoidable: we can't try dropping more elements
                 // since this could lead to another panic and abort the process.
-                self_.drop_elements();
+                //
+                // SAFETY: If something gets wrong we clear our table right after
+                // dropping the elements, so there is no double drop, since `items`
+                // will be equal to zero.
+                self_.table.drop_elements::<T>();
 
                 // If necessary, resize our table to match the source.
                 if self_.buckets() != source.buckets() {
-                    // Skip our drop by using ptr::write.
-                    if !self_.table.is_empty_singleton() {
-                        self_.free_buckets();
+                    let new_inner = match RawTableInner::new_uninitialized(
+                        &self_.alloc,
+                        Self::TABLE_LAYOUT,
+                        source.buckets(),
+                        Fallibility::Infallible,
+                    ) {
+                        Ok(table) => table,
+                        Err(_) => hint::unreachable_unchecked(),
+                    };
+                    // Replace the old inner with new uninitialized one. It's ok, since if something gets
+                    // wrong `ScopeGuard` will initialize all control bytes and leave empty table.
+                    let mut old_inner = mem::replace(&mut self_.table, new_inner);
+                    if !old_inner.is_empty_singleton() {
+                        // SAFETY:
+                        // 1. We have checked that our table is allocated.
+                        // 2. We know for sure that `alloc` and `table_layout` matches
+                        // the [`Allocator`] and [`TableLayout`] that were used to allocate this table.
+                        old_inner.free_buckets(&self_.alloc, Self::TABLE_LAYOUT);
                     }
-                    (&mut **self_ as *mut Self).write(
-                        // Avoid `Result::unwrap_or_else` because it bloats LLVM IR.
-                        match Self::new_uninitialized(
-                            self_.table.alloc.clone(),
-                            source.buckets(),
-                            Fallibility::Infallible,
-                        ) {
-                            Ok(table) => table,
-                            Err(_) => hint::unreachable_unchecked(),
-                        },
-                    );
                 }
 
+                // Cloning elements may fail (the clone function may panic), but the `ScopeGuard`
+                // inside the `clone_from_impl` function will take care of that, dropping all
+                // cloned elements if necessary. Our `ScopeGuard` will clear the table.
                 self_.clone_from_spec(source);
 
                 // Disarm the scope guard if cloning was successful.
@@ -1696,7 +3556,8 @@ impl<T: Copy, A: Allocator + Clone> RawTableClone for RawTable<T, A> {
             .copy_to_nonoverlapping(self.table.ctrl(0), self.table.num_ctrl_bytes());
         source
             .data_start()
-            .copy_to_nonoverlapping(self.data_start(), self.table.buckets());
+            .as_ptr()
+            .copy_to_nonoverlapping(self.data_start().as_ptr(), self.table.buckets());
 
         self.table.items = source.table.items;
         self.table.growth_left = source.table.growth_left;
@@ -1720,9 +3581,9 @@ impl<T: Clone, A: Allocator + Clone> RawTable<T, A> {
         // to make sure we drop only the elements that have been
         // cloned so far.
         let mut guard = guard((0, &mut *self), |(index, self_)| {
-            if mem::needs_drop::<T>() && !self_.is_empty() {
-                for i in 0..=*index {
-                    if is_full(*self_.table.ctrl(i)) {
+            if T::NEEDS_DROP {
+                for i in 0..*index {
+                    if self_.is_bucket_full(i) {
                         self_.bucket(i).drop();
                     }
                 }
@@ -1735,7 +3596,7 @@ impl<T: Clone, A: Allocator + Clone> RawTable<T, A> {
             to.write(from.as_ref().clone());
 
             // Update the index in case we need to unwind.
-            guard.0 = index;
+            guard.0 = index + 1;
         }
 
         // Successfully cloned all items, no need to clean up.
@@ -1757,7 +3618,7 @@ impl<T: Clone, A: Allocator + Clone> RawTable<T, A> {
         {
             self.clear();
 
-            let guard_self = guard(&mut *self, |self_| {
+            let mut guard_self = guard(&mut *self, |self_| {
                 // Clear the partially copied table if a panic occurs, otherwise
                 // items and growth_left will be out of sync with the contents
                 // of the table.
@@ -1790,7 +3651,7 @@ impl<T: Clone, A: Allocator + Clone> RawTable<T, A> {
     }
 }
 
-impl<T, A: Allocator + Clone + Default> Default for RawTable<T, A> {
+impl<T, A: Allocator + Default> Default for RawTable<T, A> {
     #[inline]
     fn default() -> Self {
         Self::new_in(Default::default())
@@ -1798,31 +3659,41 @@ impl<T, A: Allocator + Clone + Default> Default for RawTable<T, A> {
 }
 
 #[cfg(feature = "nightly")]
-unsafe impl<#[may_dangle] T, A: Allocator + Clone> Drop for RawTable<T, A> {
+unsafe impl<#[may_dangle] T, A: Allocator> Drop for RawTable<T, A> {
     #[cfg_attr(feature = "inline-more", inline)]
     fn drop(&mut self) {
-        if !self.table.is_empty_singleton() {
-            unsafe {
-                self.drop_elements();
-                self.free_buckets();
-            }
+        unsafe {
+            // SAFETY:
+            // 1. We call the function only once;
+            // 2. We know for sure that `alloc` and `table_layout` matches the [`Allocator`]
+            //    and [`TableLayout`] that were used to allocate this table.
+            // 3. If the drop function of any elements fails, then only a memory leak will occur,
+            //    and we don't care because we are inside the `Drop` function of the `RawTable`,
+            //    so there won't be any table left in an inconsistent state.
+            self.table
+                .drop_inner_table::<T, _>(&self.alloc, Self::TABLE_LAYOUT);
         }
     }
 }
 #[cfg(not(feature = "nightly"))]
-impl<T, A: Allocator + Clone> Drop for RawTable<T, A> {
+impl<T, A: Allocator> Drop for RawTable<T, A> {
     #[cfg_attr(feature = "inline-more", inline)]
     fn drop(&mut self) {
-        if !self.table.is_empty_singleton() {
-            unsafe {
-                self.drop_elements();
-                self.free_buckets();
-            }
+        unsafe {
+            // SAFETY:
+            // 1. We call the function only once;
+            // 2. We know for sure that `alloc` and `table_layout` matches the [`Allocator`]
+            //    and [`TableLayout`] that were used to allocate this table.
+            // 3. If the drop function of any elements fails, then only a memory leak will occur,
+            //    and we don't care because we are inside the `Drop` function of the `RawTable`,
+            //    so there won't be any table left in an inconsistent state.
+            self.table
+                .drop_inner_table::<T, _>(&self.alloc, Self::TABLE_LAYOUT);
         }
     }
 }
 
-impl<T, A: Allocator + Clone> IntoIterator for RawTable<T, A> {
+impl<T, A: Allocator> IntoIterator for RawTable<T, A> {
     type Item = T;
     type IntoIter = RawIntoIter<T, A>;
 
@@ -1840,7 +3711,7 @@ impl<T, A: Allocator + Clone> IntoIterator for RawTable<T, A> {
 pub(crate) struct RawIterRange<T> {
     // Mask of full buckets in the current group. Bits are cleared from this
     // mask as each element is processed.
-    current_group: BitMask,
+    current_group: BitMaskIter,
 
     // Pointer to the buckets for the current group.
     data: Bucket<T>,
@@ -1856,19 +3727,44 @@ pub(crate) struct RawIterRange<T> {
 impl<T> RawIterRange<T> {
     /// Returns a `RawIterRange` covering a subset of a table.
     ///
-    /// The control byte address must be aligned to the group size.
+    /// # Safety
+    ///
+    /// If any of the following conditions are violated, the result is
+    /// [`undefined behavior`]:
+    ///
+    /// * `ctrl` must be [valid] for reads, i.e. table outlives the `RawIterRange`;
+    ///
+    /// * `ctrl` must be properly aligned to the group size (Group::WIDTH);
+    ///
+    /// * `ctrl` must point to the array of properly initialized control bytes;
+    ///
+    /// * `data` must be the [`Bucket`] at the `ctrl` index in the table;
+    ///
+    /// * the value of `len` must be less than or equal to the number of table buckets,
+    ///   and the returned value of `ctrl.as_ptr().add(len).offset_from(ctrl.as_ptr())`
+    ///   must be positive.
+    ///
+    /// * The `ctrl.add(len)` pointer must be either in bounds or one
+    ///   byte past the end of the same [allocated table].
+    ///
+    /// * The `len` must be a power of two.
+    ///
+    /// [valid]: https://doc.rust-lang.org/std/ptr/index.html#safety
+    /// [`undefined behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
     #[cfg_attr(feature = "inline-more", inline)]
     unsafe fn new(ctrl: *const u8, data: Bucket<T>, len: usize) -> Self {
         debug_assert_ne!(len, 0);
         debug_assert_eq!(ctrl as usize % Group::WIDTH, 0);
+        // SAFETY: The caller must uphold the safety rules for the [`RawIterRange::new`]
         let end = ctrl.add(len);
 
         // Load the first group and advance ctrl to point to the next group
+        // SAFETY: The caller must uphold the safety rules for the [`RawIterRange::new`]
         let current_group = Group::load_aligned(ctrl).match_full();
         let next_ctrl = ctrl.add(Group::WIDTH);
 
         Self {
-            current_group,
+            current_group: current_group.into_iter(),
             data,
             next_ctrl,
             end,
@@ -1925,8 +3821,7 @@ impl<T> RawIterRange<T> {
     #[cfg_attr(feature = "inline-more", inline)]
     unsafe fn next_impl<const DO_CHECK_PTR_RANGE: bool>(&mut self) -> Option<Bucket<T>> {
         loop {
-            if let Some(index) = self.current_group.lowest_set_bit() {
-                self.current_group = self.current_group.remove_lowest_bit();
+            if let Some(index) = self.current_group.next() {
                 return Some(self.data.next_n(index));
             }
 
@@ -1939,7 +3834,86 @@ impl<T> RawIterRange<T> {
             // than the group size where the trailing control bytes are all
             // EMPTY. On larger tables self.end is guaranteed to be aligned
             // to the group size (since tables are power-of-two sized).
-            self.current_group = Group::load_aligned(self.next_ctrl).match_full();
+            self.current_group = Group::load_aligned(self.next_ctrl).match_full().into_iter();
+            self.data = self.data.next_n(Group::WIDTH);
+            self.next_ctrl = self.next_ctrl.add(Group::WIDTH);
+        }
+    }
+
+    /// Folds every element into an accumulator by applying an operation,
+    /// returning the final result.
+    ///
+    /// `fold_impl()` takes three arguments: the number of items remaining in
+    /// the iterator, an initial value, and a closure with two arguments: an
+    /// 'accumulator', and an element. The closure returns the value that the
+    /// accumulator should have for the next iteration.
+    ///
+    /// The initial value is the value the accumulator will have on the first call.
+    ///
+    /// After applying this closure to every element of the iterator, `fold_impl()`
+    /// returns the accumulator.
+    ///
+    /// # Safety
+    ///
+    /// If any of the following conditions are violated, the result is
+    /// [`Undefined Behavior`]:
+    ///
+    /// * The [`RawTableInner`] / [`RawTable`] must be alive and not moved,
+    ///   i.e. table outlives the `RawIterRange`;
+    ///
+    /// * The provided `n` value must match the actual number of items
+    ///   in the table.
+    ///
+    /// [`Undefined Behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
+    #[allow(clippy::while_let_on_iterator)]
+    #[cfg_attr(feature = "inline-more", inline)]
+    unsafe fn fold_impl<F, B>(mut self, mut n: usize, mut acc: B, mut f: F) -> B
+    where
+        F: FnMut(B, Bucket<T>) -> B,
+    {
+        loop {
+            while let Some(index) = self.current_group.next() {
+                // The returned `index` will always be in the range `0..Group::WIDTH`,
+                // so that calling `self.data.next_n(index)` is safe (see detailed explanation below).
+                debug_assert!(n != 0);
+                let bucket = self.data.next_n(index);
+                acc = f(acc, bucket);
+                n -= 1;
+            }
+
+            if n == 0 {
+                return acc;
+            }
+
+            // SAFETY: The caller of this function ensures that:
+            //
+            // 1. The provided `n` value matches the actual number of items in the table;
+            // 2. The table is alive and did not moved.
+            //
+            // Taking the above into account, we always stay within the bounds, because:
+            //
+            // 1. For tables smaller than the group width (self.buckets() <= Group::WIDTH),
+            //    we will never end up in the given branch, since we should have already
+            //    yielded all the elements of the table.
+            //
+            // 2. For tables larger than the group width. The number of buckets is a
+            //    power of two (2 ^ n), Group::WIDTH is also power of two (2 ^ k). Since
+            //    `(2 ^ n) > (2 ^ k)`, than `(2 ^ n) % (2 ^ k) = 0`. As we start from the
+            //    start of the array of control bytes, and never try to iterate after
+            //    getting all the elements, the last `self.current_group` will read bytes
+            //    from the `self.buckets() - Group::WIDTH` index.  We know also that
+            //    `self.current_group.next()` will always retun indices within the range
+            //    `0..Group::WIDTH`.
+            //
+            //    Knowing all of the above and taking into account that we are synchronizing
+            //    the `self.data` index with the index we used to read the `self.current_group`,
+            //    the subsequent `self.data.next_n(index)` will always return a bucket with
+            //    an index number less than `self.buckets()`.
+            //
+            //    The last `self.next_ctrl`, whose index would be `self.buckets()`, will never
+            //    actually be read, since we should have already yielded all the elements of
+            //    the table.
+            self.current_group = Group::load_aligned(self.next_ctrl).match_full().into_iter();
             self.data = self.data.next_n(Group::WIDTH);
             self.next_ctrl = self.next_ctrl.add(Group::WIDTH);
         }
@@ -2016,7 +3990,7 @@ impl<T> RawIter<T> {
     /// This method should be called _before_ the removal is made. It is not necessary to call this
     /// method if you are removing an item that this iterator yielded in the past.
     #[cfg(feature = "raw")]
-    pub fn reflect_remove(&mut self, b: &Bucket<T>) {
+    pub unsafe fn reflect_remove(&mut self, b: &Bucket<T>) {
         self.reflect_toggle_full(b, false);
     }
 
@@ -2030,36 +4004,76 @@ impl<T> RawIter<T> {
     ///
     /// This method should be called _after_ the given insert is made.
     #[cfg(feature = "raw")]
-    pub fn reflect_insert(&mut self, b: &Bucket<T>) {
+    pub unsafe fn reflect_insert(&mut self, b: &Bucket<T>) {
         self.reflect_toggle_full(b, true);
     }
 
     /// Refresh the iterator so that it reflects a change to the state of the given bucket.
     #[cfg(feature = "raw")]
-    fn reflect_toggle_full(&mut self, b: &Bucket<T>, is_insert: bool) {
-        unsafe {
-            if b.as_ptr() > self.iter.data.as_ptr() {
-                // The iterator has already passed the bucket's group.
-                // So the toggle isn't relevant to this iterator.
-                return;
+    unsafe fn reflect_toggle_full(&mut self, b: &Bucket<T>, is_insert: bool) {
+        if b.as_ptr() > self.iter.data.as_ptr() {
+            // The iterator has already passed the bucket's group.
+            // So the toggle isn't relevant to this iterator.
+            return;
+        }
+
+        if self.iter.next_ctrl < self.iter.end
+            && b.as_ptr() <= self.iter.data.next_n(Group::WIDTH).as_ptr()
+        {
+            // The iterator has not yet reached the bucket's group.
+            // We don't need to reload anything, but we do need to adjust the item count.
+
+            if cfg!(debug_assertions) {
+                // Double-check that the user isn't lying to us by checking the bucket state.
+                // To do that, we need to find its control byte. We know that self.iter.data is
+                // at self.iter.next_ctrl - Group::WIDTH, so we work from there:
+                let offset = offset_from(self.iter.data.as_ptr(), b.as_ptr());
+                let ctrl = self.iter.next_ctrl.sub(Group::WIDTH).add(offset);
+                // This method should be called _before_ a removal, or _after_ an insert,
+                // so in both cases the ctrl byte should indicate that the bucket is full.
+                assert!(is_full(*ctrl));
             }
 
-            if self.iter.next_ctrl < self.iter.end
-                && b.as_ptr() <= self.iter.data.next_n(Group::WIDTH).as_ptr()
-            {
-                // The iterator has not yet reached the bucket's group.
-                // We don't need to reload anything, but we do need to adjust the item count.
+            if is_insert {
+                self.items += 1;
+            } else {
+                self.items -= 1;
+            }
 
-                if cfg!(debug_assertions) {
-                    // Double-check that the user isn't lying to us by checking the bucket state.
-                    // To do that, we need to find its control byte. We know that self.iter.data is
-                    // at self.iter.next_ctrl - Group::WIDTH, so we work from there:
-                    let offset = offset_from(self.iter.data.as_ptr(), b.as_ptr());
-                    let ctrl = self.iter.next_ctrl.sub(Group::WIDTH).add(offset);
-                    // This method should be called _before_ a removal, or _after_ an insert,
-                    // so in both cases the ctrl byte should indicate that the bucket is full.
-                    assert!(is_full(*ctrl));
-                }
+            return;
+        }
+
+        // The iterator is at the bucket group that the toggled bucket is in.
+        // We need to do two things:
+        //
+        //  - Determine if the iterator already yielded the toggled bucket.
+        //    If it did, we're done.
+        //  - Otherwise, update the iterator cached group so that it won't
+        //    yield a to-be-removed bucket, or _will_ yield a to-be-added bucket.
+        //    We'll also need to update the item count accordingly.
+        if let Some(index) = self.iter.current_group.0.lowest_set_bit() {
+            let next_bucket = self.iter.data.next_n(index);
+            if b.as_ptr() > next_bucket.as_ptr() {
+                // The toggled bucket is "before" the bucket the iterator would yield next. We
+                // therefore don't need to do anything --- the iterator has already passed the
+                // bucket in question.
+                //
+                // The item count must already be correct, since a removal or insert "prior" to
+                // the iterator's position wouldn't affect the item count.
+            } else {
+                // The removed bucket is an upcoming bucket. We need to make sure it does _not_
+                // get yielded, and also that it's no longer included in the item count.
+                //
+                // NOTE: We can't just reload the group here, both since that might reflect
+                // inserts we've already passed, and because that might inadvertently unset the
+                // bits for _other_ removals. If we do that, we'd have to also decrement the
+                // item count for those other bits that we unset. But the presumably subsequent
+                // call to reflect for those buckets might _also_ decrement the item count.
+                // Instead, we _just_ flip the bit for the particular bucket the caller asked
+                // us to reflect.
+                let our_bit = offset_from(self.iter.data.as_ptr(), b.as_ptr());
+                let was_full = self.iter.current_group.flip(our_bit);
+                debug_assert_ne!(was_full, is_insert);
 
                 if is_insert {
                     self.items += 1;
@@ -2067,65 +4081,23 @@ impl<T> RawIter<T> {
                     self.items -= 1;
                 }
 
-                return;
-            }
-
-            // The iterator is at the bucket group that the toggled bucket is in.
-            // We need to do two things:
-            //
-            //  - Determine if the iterator already yielded the toggled bucket.
-            //    If it did, we're done.
-            //  - Otherwise, update the iterator cached group so that it won't
-            //    yield a to-be-removed bucket, or _will_ yield a to-be-added bucket.
-            //    We'll also need to update the item count accordingly.
-            if let Some(index) = self.iter.current_group.lowest_set_bit() {
-                let next_bucket = self.iter.data.next_n(index);
-                if b.as_ptr() > next_bucket.as_ptr() {
-                    // The toggled bucket is "before" the bucket the iterator would yield next. We
-                    // therefore don't need to do anything --- the iterator has already passed the
-                    // bucket in question.
-                    //
-                    // The item count must already be correct, since a removal or insert "prior" to
-                    // the iterator's position wouldn't affect the item count.
-                } else {
-                    // The removed bucket is an upcoming bucket. We need to make sure it does _not_
-                    // get yielded, and also that it's no longer included in the item count.
-                    //
-                    // NOTE: We can't just reload the group here, both since that might reflect
-                    // inserts we've already passed, and because that might inadvertently unset the
-                    // bits for _other_ removals. If we do that, we'd have to also decrement the
-                    // item count for those other bits that we unset. But the presumably subsequent
-                    // call to reflect for those buckets might _also_ decrement the item count.
-                    // Instead, we _just_ flip the bit for the particular bucket the caller asked
-                    // us to reflect.
-                    let our_bit = offset_from(self.iter.data.as_ptr(), b.as_ptr());
-                    let was_full = self.iter.current_group.flip(our_bit);
-                    debug_assert_ne!(was_full, is_insert);
-
-                    if is_insert {
-                        self.items += 1;
+                if cfg!(debug_assertions) {
+                    if b.as_ptr() == next_bucket.as_ptr() {
+                        // The removed bucket should no longer be next
+                        debug_assert_ne!(self.iter.current_group.0.lowest_set_bit(), Some(index));
                     } else {
-                        self.items -= 1;
-                    }
-
-                    if cfg!(debug_assertions) {
-                        if b.as_ptr() == next_bucket.as_ptr() {
-                            // The removed bucket should no longer be next
-                            debug_assert_ne!(self.iter.current_group.lowest_set_bit(), Some(index));
-                        } else {
-                            // We should not have changed what bucket comes next.
-                            debug_assert_eq!(self.iter.current_group.lowest_set_bit(), Some(index));
-                        }
+                        // We should not have changed what bucket comes next.
+                        debug_assert_eq!(self.iter.current_group.0.lowest_set_bit(), Some(index));
                     }
                 }
-            } else {
-                // We must have already iterated past the removed item.
             }
+        } else {
+            // We must have already iterated past the removed item.
         }
     }
 
     unsafe fn drop_elements(&mut self) {
-        if mem::needs_drop::<T>() && self.len() != 0 {
+        if T::NEEDS_DROP && self.items != 0 {
             for item in self {
                 item.drop();
             }
@@ -2159,9 +4131,8 @@ impl<T> Iterator for RawIter<T> {
             self.iter.next_impl::<false>()
         };
 
-        if nxt.is_some() {
-            self.items -= 1;
-        }
+        debug_assert!(nxt.is_some());
+        self.items -= 1;
 
         nxt
     }
@@ -2170,33 +4141,160 @@ impl<T> Iterator for RawIter<T> {
     fn size_hint(&self) -> (usize, Option<usize>) {
         (self.items, Some(self.items))
     }
+
+    #[inline]
+    fn fold<B, F>(self, init: B, f: F) -> B
+    where
+        Self: Sized,
+        F: FnMut(B, Self::Item) -> B,
+    {
+        unsafe { self.iter.fold_impl(self.items, init, f) }
+    }
 }
 
 impl<T> ExactSizeIterator for RawIter<T> {}
 impl<T> FusedIterator for RawIter<T> {}
 
+/// Iterator which returns an index of every full bucket in the table.
+///
+/// For maximum flexibility this iterator is not bound by a lifetime, but you
+/// must observe several rules when using it:
+/// - You must not free the hash table while iterating (including via growing/shrinking).
+/// - It is fine to erase a bucket that has been yielded by the iterator.
+/// - Erasing a bucket that has not yet been yielded by the iterator may still
+///   result in the iterator yielding index of that bucket.
+/// - It is unspecified whether an element inserted after the iterator was
+///   created will be yielded by that iterator.
+/// - The order in which the iterator yields indices of the buckets is unspecified
+///   and may change in the future.
+pub(crate) struct FullBucketsIndices {
+    // Mask of full buckets in the current group. Bits are cleared from this
+    // mask as each element is processed.
+    current_group: BitMaskIter,
+
+    // Initial value of the bytes' indices of the current group (relative
+    // to the start of the control bytes).
+    group_first_index: usize,
+
+    // Pointer to the current group of control bytes,
+    // Must be aligned to the group size (Group::WIDTH).
+    ctrl: NonNull<u8>,
+
+    // Number of elements in the table.
+    items: usize,
+}
+
+impl FullBucketsIndices {
+    /// Advances the iterator and returns the next value.
+    ///
+    /// # Safety
+    ///
+    /// If any of the following conditions are violated, the result is
+    /// [`Undefined Behavior`]:
+    ///
+    /// * The [`RawTableInner`] / [`RawTable`] must be alive and not moved,
+    ///   i.e. table outlives the `FullBucketsIndices`;
+    ///
+    /// * It never tries to iterate after getting all elements.
+    ///
+    /// [`Undefined Behavior`]: https://doc.rust-lang.org/reference/behavior-considered-undefined.html
+    #[inline(always)]
+    unsafe fn next_impl(&mut self) -> Option<usize> {
+        loop {
+            if let Some(index) = self.current_group.next() {
+                // The returned `self.group_first_index + index` will always
+                // be in the range `0..self.buckets()`. See explanation below.
+                return Some(self.group_first_index + index);
+            }
+
+            // SAFETY: The caller of this function ensures that:
+            //
+            // 1. It never tries to iterate after getting all the elements;
+            // 2. The table is alive and did not moved;
+            // 3. The first `self.ctrl` pointed to the start of the array of control bytes.
+            //
+            // Taking the above into account, we always stay within the bounds, because:
+            //
+            // 1. For tables smaller than the group width (self.buckets() <= Group::WIDTH),
+            //    we will never end up in the given branch, since we should have already
+            //    yielded all the elements of the table.
+            //
+            // 2. For tables larger than the group width. The number of buckets is a
+            //    power of two (2 ^ n), Group::WIDTH is also power of two (2 ^ k). Since
+            //    `(2 ^ n) > (2 ^ k)`, than `(2 ^ n) % (2 ^ k) = 0`. As we start from the
+            //    the start of the array of control bytes, and never try to iterate after
+            //    getting all the elements, the last `self.ctrl` will be equal to
+            //    the `self.buckets() - Group::WIDTH`, so `self.current_group.next()`
+            //    will always contains indices within the range `0..Group::WIDTH`,
+            //    and subsequent `self.group_first_index + index` will always return a
+            //    number less than `self.buckets()`.
+            self.ctrl = NonNull::new_unchecked(self.ctrl.as_ptr().add(Group::WIDTH));
+
+            // SAFETY: See explanation above.
+            self.current_group = Group::load_aligned(self.ctrl.as_ptr())
+                .match_full()
+                .into_iter();
+            self.group_first_index += Group::WIDTH;
+        }
+    }
+}
+
+impl Iterator for FullBucketsIndices {
+    type Item = usize;
+
+    /// Advances the iterator and returns the next value. It is up to
+    /// the caller to ensure that the `RawTable` outlives the `FullBucketsIndices`,
+    /// because we cannot make the `next` method unsafe.
+    #[inline(always)]
+    fn next(&mut self) -> Option<usize> {
+        // Return if we already yielded all items.
+        if self.items == 0 {
+            return None;
+        }
+
+        let nxt = unsafe {
+            // SAFETY:
+            // 1. We check number of items to yield using `items` field.
+            // 2. The caller ensures that the table is alive and has not moved.
+            self.next_impl()
+        };
+
+        debug_assert!(nxt.is_some());
+        self.items -= 1;
+
+        nxt
+    }
+
+    #[inline(always)]
+    fn size_hint(&self) -> (usize, Option<usize>) {
+        (self.items, Some(self.items))
+    }
+}
+
+impl ExactSizeIterator for FullBucketsIndices {}
+impl FusedIterator for FullBucketsIndices {}
+
 /// Iterator which consumes a table and returns elements.
-pub struct RawIntoIter<T, A: Allocator + Clone = Global> {
+pub struct RawIntoIter<T, A: Allocator = Global> {
     iter: RawIter<T>,
-    allocation: Option<(NonNull<u8>, Layout)>,
+    allocation: Option<(NonNull<u8>, Layout, A)>,
     marker: PhantomData<T>,
-    alloc: A,
 }
 
-impl<T, A: Allocator + Clone> RawIntoIter<T, A> {
+impl<T, A: Allocator> RawIntoIter<T, A> {
     #[cfg_attr(feature = "inline-more", inline)]
     pub fn iter(&self) -> RawIter<T> {
         self.iter.clone()
     }
 }
 
-unsafe impl<T, A: Allocator + Clone> Send for RawIntoIter<T, A>
+unsafe impl<T, A: Allocator> Send for RawIntoIter<T, A>
 where
     T: Send,
     A: Send,
 {
 }
-unsafe impl<T, A: Allocator + Clone> Sync for RawIntoIter<T, A>
+unsafe impl<T, A: Allocator> Sync for RawIntoIter<T, A>
 where
     T: Sync,
     A: Sync,
@@ -2204,7 +4302,7 @@ where
 }
 
 #[cfg(feature = "nightly")]
-unsafe impl<#[may_dangle] T, A: Allocator + Clone> Drop for RawIntoIter<T, A> {
+unsafe impl<#[may_dangle] T, A: Allocator> Drop for RawIntoIter<T, A> {
     #[cfg_attr(feature = "inline-more", inline)]
     fn drop(&mut self) {
         unsafe {
@@ -2212,14 +4310,14 @@ unsafe impl<#[may_dangle] T, A: Allocator + Clone> Drop for RawIntoIter<T, A> {
             self.iter.drop_elements();
 
             // Free the table
-            if let Some((ptr, layout)) = self.allocation {
-                self.alloc.deallocate(ptr, layout);
+            if let Some((ptr, layout, ref alloc)) = self.allocation {
+                alloc.deallocate(ptr, layout);
             }
         }
     }
 }
 #[cfg(not(feature = "nightly"))]
-impl<T, A: Allocator + Clone> Drop for RawIntoIter<T, A> {
+impl<T, A: Allocator> Drop for RawIntoIter<T, A> {
     #[cfg_attr(feature = "inline-more", inline)]
     fn drop(&mut self) {
         unsafe {
@@ -2227,14 +4325,14 @@ impl<T, A: Allocator + Clone> Drop for RawIntoIter<T, A> {
             self.iter.drop_elements();
 
             // Free the table
-            if let Some((ptr, layout)) = self.allocation {
-                self.alloc.deallocate(ptr, layout);
+            if let Some((ptr, layout, ref alloc)) = self.allocation {
+                alloc.deallocate(ptr, layout);
             }
         }
     }
 }
 
-impl<T, A: Allocator + Clone> Iterator for RawIntoIter<T, A> {
+impl<T, A: Allocator> Iterator for RawIntoIter<T, A> {
     type Item = T;
 
     #[cfg_attr(feature = "inline-more", inline)]
@@ -2248,45 +4346,45 @@ impl<T, A: Allocator + Clone> Iterator for RawIntoIter<T, A> {
     }
 }
 
-impl<T, A: Allocator + Clone> ExactSizeIterator for RawIntoIter<T, A> {}
-impl<T, A: Allocator + Clone> FusedIterator for RawIntoIter<T, A> {}
+impl<T, A: Allocator> ExactSizeIterator for RawIntoIter<T, A> {}
+impl<T, A: Allocator> FusedIterator for RawIntoIter<T, A> {}
 
 /// Iterator which consumes elements without freeing the table storage.
-pub struct RawDrain<'a, T, A: Allocator + Clone = Global> {
+pub struct RawDrain<'a, T, A: Allocator = Global> {
     iter: RawIter<T>,
 
     // The table is moved into the iterator for the duration of the drain. This
     // ensures that an empty table is left if the drain iterator is leaked
     // without dropping.
-    table: ManuallyDrop<RawTable<T, A>>,
-    orig_table: NonNull<RawTable<T, A>>,
+    table: RawTableInner,
+    orig_table: NonNull<RawTableInner>,
 
     // We don't use a &'a mut RawTable<T> because we want RawDrain to be
     // covariant over T.
     marker: PhantomData<&'a RawTable<T, A>>,
 }
 
-impl<T, A: Allocator + Clone> RawDrain<'_, T, A> {
+impl<T, A: Allocator> RawDrain<'_, T, A> {
     #[cfg_attr(feature = "inline-more", inline)]
     pub fn iter(&self) -> RawIter<T> {
         self.iter.clone()
     }
 }
 
-unsafe impl<T, A: Allocator + Copy> Send for RawDrain<'_, T, A>
+unsafe impl<T, A: Allocator> Send for RawDrain<'_, T, A>
 where
     T: Send,
     A: Send,
 {
 }
-unsafe impl<T, A: Allocator + Copy> Sync for RawDrain<'_, T, A>
+unsafe impl<T, A: Allocator> Sync for RawDrain<'_, T, A>
 where
     T: Sync,
     A: Sync,
 {
 }
 
-impl<T, A: Allocator + Clone> Drop for RawDrain<'_, T, A> {
+impl<T, A: Allocator> Drop for RawDrain<'_, T, A> {
     #[cfg_attr(feature = "inline-more", inline)]
     fn drop(&mut self) {
         unsafe {
@@ -2300,12 +4398,12 @@ impl<T, A: Allocator + Clone> Drop for RawDrain<'_, T, A> {
             // Move the now empty table back to its original location.
             self.orig_table
                 .as_ptr()
-                .copy_from_nonoverlapping(&*self.table, 1);
+                .copy_from_nonoverlapping(&self.table, 1);
         }
     }
 }
 
-impl<T, A: Allocator + Clone> Iterator for RawDrain<'_, T, A> {
+impl<T, A: Allocator> Iterator for RawDrain<'_, T, A> {
     type Item = T;
 
     #[cfg_attr(feature = "inline-more", inline)]
@@ -2322,21 +4420,36 @@ impl<T, A: Allocator + Clone> Iterator for RawDrain<'_, T, A> {
     }
 }
 
-impl<T, A: Allocator + Clone> ExactSizeIterator for RawDrain<'_, T, A> {}
-impl<T, A: Allocator + Clone> FusedIterator for RawDrain<'_, T, A> {}
+impl<T, A: Allocator> ExactSizeIterator for RawDrain<'_, T, A> {}
+impl<T, A: Allocator> FusedIterator for RawDrain<'_, T, A> {}
 
 /// Iterator over occupied buckets that could match a given hash.
 ///
 /// `RawTable` only stores 7 bits of the hash value, so this iterator may return
 /// items that have a hash value different than the one provided. You should
 /// always validate the returned values before using them.
-pub struct RawIterHash<'a, T, A: Allocator + Clone = Global> {
-    inner: RawIterHashInner<'a, A>,
+///
+/// For maximum flexibility this iterator is not bound by a lifetime, but you
+/// must observe several rules when using it:
+/// - You must not free the hash table while iterating (including via growing/shrinking).
+/// - It is fine to erase a bucket that has been yielded by the iterator.
+/// - Erasing a bucket that has not yet been yielded by the iterator may still
+///   result in the iterator yielding that bucket.
+/// - It is unspecified whether an element inserted after the iterator was
+///   created will be yielded by that iterator.
+/// - The order in which the iterator yields buckets is unspecified and may
+///   change in the future.
+pub struct RawIterHash<T> {
+    inner: RawIterHashInner,
     _marker: PhantomData<T>,
 }
 
-struct RawIterHashInner<'a, A: Allocator + Clone> {
-    table: &'a RawTableInner<A>,
+struct RawIterHashInner {
+    // See `RawTableInner`'s corresponding fields for details.
+    // We can't store a `*const RawTableInner` as it would get
+    // invalidated by the user calling `&mut` methods on `RawTable`.
+    bucket_mask: usize,
+    ctrl: NonNull<u8>,
 
     // The top 7 bits of the hash.
     h2_hash: u8,
@@ -2350,71 +4463,105 @@ struct RawIterHashInner<'a, A: Allocator + Clone> {
     bitmask: BitMaskIter,
 }
 
-impl<'a, T, A: Allocator + Clone> RawIterHash<'a, T, A> {
+impl<T> RawIterHash<T> {
     #[cfg_attr(feature = "inline-more", inline)]
     #[cfg(feature = "raw")]
-    fn new(table: &'a RawTable<T, A>, hash: u64) -> Self {
+    unsafe fn new<A: Allocator>(table: &RawTable<T, A>, hash: u64) -> Self {
         RawIterHash {
             inner: RawIterHashInner::new(&table.table, hash),
             _marker: PhantomData,
         }
     }
 }
-impl<'a, A: Allocator + Clone> RawIterHashInner<'a, A> {
+impl RawIterHashInner {
     #[cfg_attr(feature = "inline-more", inline)]
     #[cfg(feature = "raw")]
-    fn new(table: &'a RawTableInner<A>, hash: u64) -> Self {
-        unsafe {
-            let h2_hash = h2(hash);
-            let probe_seq = table.probe_seq(hash);
-            let group = Group::load(table.ctrl(probe_seq.pos));
-            let bitmask = group.match_byte(h2_hash).into_iter();
-
-            RawIterHashInner {
-                table,
-                h2_hash,
-                probe_seq,
-                group,
-                bitmask,
-            }
+    unsafe fn new(table: &RawTableInner, hash: u64) -> Self {
+        let h2_hash = h2(hash);
+        let probe_seq = table.probe_seq(hash);
+        let group = Group::load(table.ctrl(probe_seq.pos));
+        let bitmask = group.match_byte(h2_hash).into_iter();
+
+        RawIterHashInner {
+            bucket_mask: table.bucket_mask,
+            ctrl: table.ctrl,
+            h2_hash,
+            probe_seq,
+            group,
+            bitmask,
         }
     }
 }
 
-impl<'a, T, A: Allocator + Clone> Iterator for RawIterHash<'a, T, A> {
+impl<T> Iterator for RawIterHash<T> {
     type Item = Bucket<T>;
 
     fn next(&mut self) -> Option<Bucket<T>> {
         unsafe {
             match self.inner.next() {
-                Some(index) => Some(self.inner.table.bucket(index)),
+                Some(index) => {
+                    // Can't use `RawTable::bucket` here as we don't have
+                    // an actual `RawTable` reference to use.
+                    debug_assert!(index <= self.inner.bucket_mask);
+                    let bucket = Bucket::from_base_index(self.inner.ctrl.cast(), index);
+                    Some(bucket)
+                }
                 None => None,
             }
         }
     }
 }
 
-impl<'a, A: Allocator + Clone> Iterator for RawIterHashInner<'a, A> {
+impl Iterator for RawIterHashInner {
     type Item = usize;
 
     fn next(&mut self) -> Option<Self::Item> {
         unsafe {
             loop {
                 if let Some(bit) = self.bitmask.next() {
-                    let index = (self.probe_seq.pos + bit) & self.table.bucket_mask;
+                    let index = (self.probe_seq.pos + bit) & self.bucket_mask;
                     return Some(index);
                 }
                 if likely(self.group.match_empty().any_bit_set()) {
                     return None;
                 }
-                self.probe_seq.move_next(self.table.bucket_mask);
-                self.group = Group::load(self.table.ctrl(self.probe_seq.pos));
+                self.probe_seq.move_next(self.bucket_mask);
+
+                // Can't use `RawTableInner::ctrl` here as we don't have
+                // an actual `RawTableInner` reference to use.
+                let index = self.probe_seq.pos;
+                debug_assert!(index < self.bucket_mask + 1 + Group::WIDTH);
+                let group_ctrl = self.ctrl.as_ptr().add(index);
+
+                self.group = Group::load(group_ctrl);
                 self.bitmask = self.group.match_byte(self.h2_hash).into_iter();
             }
         }
     }
 }
 
+pub(crate) struct RawExtractIf<'a, T, A: Allocator> {
+    pub iter: RawIter<T>,
+    pub table: &'a mut RawTable<T, A>,
+}
+
+impl<T, A: Allocator> RawExtractIf<'_, T, A> {
+    #[cfg_attr(feature = "inline-more", inline)]
+    pub(crate) fn next<F>(&mut self, mut f: F) -> Option<T>
+    where
+        F: FnMut(&mut T) -> bool,
+    {
+        unsafe {
+            for item in &mut self.iter {
+                if f(item.as_mut()) {
+                    return Some(self.table.remove(item).0);
+                }
+            }
+        }
+        None
+    }
+}
+
 #[cfg(test)]
 mod test_map {
     use super::*;
@@ -2457,4 +4604,214 @@ mod test_map {
             assert!(table.find(i + 100, |x| *x == i + 100).is_none());
         }
     }
+
+    /// CHECKING THAT WE ARE NOT TRYING TO READ THE MEMORY OF
+    /// AN UNINITIALIZED TABLE DURING THE DROP
+    #[test]
+    fn test_drop_uninitialized() {
+        use ::alloc::vec::Vec;
+
+        let table = unsafe {
+            // SAFETY: The `buckets` is power of two and we're not
+            // trying to actually use the returned RawTable.
+            RawTable::<(u64, Vec<i32>)>::new_uninitialized(Global, 8, Fallibility::Infallible)
+                .unwrap()
+        };
+        drop(table);
+    }
+
+    /// CHECKING THAT WE DON'T TRY TO DROP DATA IF THE `ITEMS`
+    /// ARE ZERO, EVEN IF WE HAVE `FULL` CONTROL BYTES.
+    #[test]
+    fn test_drop_zero_items() {
+        use ::alloc::vec::Vec;
+        unsafe {
+            // SAFETY: The `buckets` is power of two and we're not
+            // trying to actually use the returned RawTable.
+            let table =
+                RawTable::<(u64, Vec<i32>)>::new_uninitialized(Global, 8, Fallibility::Infallible)
+                    .unwrap();
+
+            // WE SIMULATE, AS IT WERE, A FULL TABLE.
+
+            // SAFETY: We checked that the table is allocated and therefore the table already has
+            // `self.bucket_mask + 1 + Group::WIDTH` number of control bytes (see TableLayout::calculate_layout_for)
+            // so writing `table.table.num_ctrl_bytes() == bucket_mask + 1 + Group::WIDTH` bytes is safe.
+            table
+                .table
+                .ctrl(0)
+                .write_bytes(EMPTY, table.table.num_ctrl_bytes());
+
+            // SAFETY: table.capacity() is guaranteed to be smaller than table.buckets()
+            table.table.ctrl(0).write_bytes(0, table.capacity());
+
+            // Fix up the trailing control bytes. See the comments in set_ctrl
+            // for the handling of tables smaller than the group width.
+            if table.buckets() < Group::WIDTH {
+                // SAFETY: We have `self.bucket_mask + 1 + Group::WIDTH` number of control bytes,
+                // so copying `self.buckets() == self.bucket_mask + 1` bytes with offset equal to
+                // `Group::WIDTH` is safe
+                table
+                    .table
+                    .ctrl(0)
+                    .copy_to(table.table.ctrl(Group::WIDTH), table.table.buckets());
+            } else {
+                // SAFETY: We have `self.bucket_mask + 1 + Group::WIDTH` number of
+                // control bytes,so copying `Group::WIDTH` bytes with offset equal
+                // to `self.buckets() == self.bucket_mask + 1` is safe
+                table
+                    .table
+                    .ctrl(0)
+                    .copy_to(table.table.ctrl(table.table.buckets()), Group::WIDTH);
+            }
+            drop(table);
+        }
+    }
+
+    /// CHECKING THAT WE DON'T TRY TO DROP DATA IF THE `ITEMS`
+    /// ARE ZERO, EVEN IF WE HAVE `FULL` CONTROL BYTES.
+    #[test]
+    fn test_catch_panic_clone_from() {
+        use ::alloc::sync::Arc;
+        use ::alloc::vec::Vec;
+        use allocator_api2::alloc::{AllocError, Allocator, Global};
+        use core::sync::atomic::{AtomicI8, Ordering};
+        use std::thread;
+
+        struct MyAllocInner {
+            drop_count: Arc<AtomicI8>,
+        }
+
+        #[derive(Clone)]
+        struct MyAlloc {
+            _inner: Arc<MyAllocInner>,
+        }
+
+        impl Drop for MyAllocInner {
+            fn drop(&mut self) {
+                println!("MyAlloc freed.");
+                self.drop_count.fetch_sub(1, Ordering::SeqCst);
+            }
+        }
+
+        unsafe impl Allocator for MyAlloc {
+            fn allocate(&self, layout: Layout) -> std::result::Result<NonNull<[u8]>, AllocError> {
+                let g = Global;
+                g.allocate(layout)
+            }
+
+            unsafe fn deallocate(&self, ptr: NonNull<u8>, layout: Layout) {
+                let g = Global;
+                g.deallocate(ptr, layout)
+            }
+        }
+
+        const DISARMED: bool = false;
+        const ARMED: bool = true;
+
+        struct CheckedCloneDrop {
+            panic_in_clone: bool,
+            dropped: bool,
+            need_drop: Vec<u64>,
+        }
+
+        impl Clone for CheckedCloneDrop {
+            fn clone(&self) -> Self {
+                if self.panic_in_clone {
+                    panic!("panic in clone")
+                }
+                Self {
+                    panic_in_clone: self.panic_in_clone,
+                    dropped: self.dropped,
+                    need_drop: self.need_drop.clone(),
+                }
+            }
+        }
+
+        impl Drop for CheckedCloneDrop {
+            fn drop(&mut self) {
+                if self.dropped {
+                    panic!("double drop");
+                }
+                self.dropped = true;
+            }
+        }
+
+        let dropped: Arc<AtomicI8> = Arc::new(AtomicI8::new(2));
+
+        let mut table = RawTable::new_in(MyAlloc {
+            _inner: Arc::new(MyAllocInner {
+                drop_count: dropped.clone(),
+            }),
+        });
+
+        for (idx, panic_in_clone) in core::iter::repeat(DISARMED).take(7).enumerate() {
+            let idx = idx as u64;
+            table.insert(
+                idx,
+                (
+                    idx,
+                    CheckedCloneDrop {
+                        panic_in_clone,
+                        dropped: false,
+                        need_drop: vec![idx],
+                    },
+                ),
+                |(k, _)| *k,
+            );
+        }
+
+        assert_eq!(table.len(), 7);
+
+        thread::scope(|s| {
+            let result = s.spawn(|| {
+                let armed_flags = [
+                    DISARMED, DISARMED, ARMED, DISARMED, DISARMED, DISARMED, DISARMED,
+                ];
+                let mut scope_table = RawTable::new_in(MyAlloc {
+                    _inner: Arc::new(MyAllocInner {
+                        drop_count: dropped.clone(),
+                    }),
+                });
+                for (idx, &panic_in_clone) in armed_flags.iter().enumerate() {
+                    let idx = idx as u64;
+                    scope_table.insert(
+                        idx,
+                        (
+                            idx,
+                            CheckedCloneDrop {
+                                panic_in_clone,
+                                dropped: false,
+                                need_drop: vec![idx + 100],
+                            },
+                        ),
+                        |(k, _)| *k,
+                    );
+                }
+                table.clone_from(&scope_table);
+            });
+            assert!(result.join().is_err());
+        });
+
+        // Let's check that all iterators work fine and do not return elements
+        // (especially `RawIterRange`, which does not depend on the number of
+        // elements in the table, but looks directly at the control bytes)
+        //
+        // SAFETY: We know for sure that `RawTable` will outlive
+        // the returned `RawIter / RawIterRange` iterator.
+        assert_eq!(table.len(), 0);
+        assert_eq!(unsafe { table.iter().count() }, 0);
+        assert_eq!(unsafe { table.iter().iter.count() }, 0);
+
+        for idx in 0..table.buckets() {
+            let idx = idx as u64;
+            assert!(
+                table.find(idx, |(k, _)| *k == idx).is_none(),
+                "Index: {idx}"
+            );
+        }
+
+        // All allocator clones should already be dropped.
+        assert_eq!(dropped.load(Ordering::SeqCst), 1);
+    }
 }
diff --git a/third_party/rust/hashbrown/src/raw/neon.rs b/third_party/rust/hashbrown/src/raw/neon.rs
new file mode 100644
index 0000000000..44e82d57d5
--- /dev/null
+++ b/third_party/rust/hashbrown/src/raw/neon.rs
@@ -0,0 +1,124 @@
+use super::bitmask::BitMask;
+use super::EMPTY;
+use core::arch::aarch64 as neon;
+use core::mem;
+use core::num::NonZeroU64;
+
+pub(crate) type BitMaskWord = u64;
+pub(crate) type NonZeroBitMaskWord = NonZeroU64;
+pub(crate) const BITMASK_STRIDE: usize = 8;
+pub(crate) const BITMASK_MASK: BitMaskWord = !0;
+pub(crate) const BITMASK_ITER_MASK: BitMaskWord = 0x8080_8080_8080_8080;
+
+/// Abstraction over a group of control bytes which can be scanned in
+/// parallel.
+///
+/// This implementation uses a 64-bit NEON value.
+#[derive(Copy, Clone)]
+pub(crate) struct Group(neon::uint8x8_t);
+
+#[allow(clippy::use_self)]
+impl Group {
+    /// Number of bytes in the group.
+    pub(crate) const WIDTH: usize = mem::size_of::<Self>();
+
+    /// Returns a full group of empty bytes, suitable for use as the initial
+    /// value for an empty hash table.
+    ///
+    /// This is guaranteed to be aligned to the group size.
+    #[inline]
+    pub(crate) const fn static_empty() -> &'static [u8; Group::WIDTH] {
+        #[repr(C)]
+        struct AlignedBytes {
+            _align: [Group; 0],
+            bytes: [u8; Group::WIDTH],
+        }
+        const ALIGNED_BYTES: AlignedBytes = AlignedBytes {
+            _align: [],
+            bytes: [EMPTY; Group::WIDTH],
+        };
+        &ALIGNED_BYTES.bytes
+    }
+
+    /// Loads a group of bytes starting at the given address.
+    #[inline]
+    #[allow(clippy::cast_ptr_alignment)] // unaligned load
+    pub(crate) unsafe fn load(ptr: *const u8) -> Self {
+        Group(neon::vld1_u8(ptr))
+    }
+
+    /// Loads a group of bytes starting at the given address, which must be
+    /// aligned to `mem::align_of::<Group>()`.
+    #[inline]
+    #[allow(clippy::cast_ptr_alignment)]
+    pub(crate) unsafe fn load_aligned(ptr: *const u8) -> Self {
+        // FIXME: use align_offset once it stabilizes
+        debug_assert_eq!(ptr as usize & (mem::align_of::<Self>() - 1), 0);
+        Group(neon::vld1_u8(ptr))
+    }
+
+    /// Stores the group of bytes to the given address, which must be
+    /// aligned to `mem::align_of::<Group>()`.
+    #[inline]
+    #[allow(clippy::cast_ptr_alignment)]
+    pub(crate) unsafe fn store_aligned(self, ptr: *mut u8) {
+        // FIXME: use align_offset once it stabilizes
+        debug_assert_eq!(ptr as usize & (mem::align_of::<Self>() - 1), 0);
+        neon::vst1_u8(ptr, self.0);
+    }
+
+    /// Returns a `BitMask` indicating all bytes in the group which *may*
+    /// have the given value.
+    #[inline]
+    pub(crate) fn match_byte(self, byte: u8) -> BitMask {
+        unsafe {
+            let cmp = neon::vceq_u8(self.0, neon::vdup_n_u8(byte));
+            BitMask(neon::vget_lane_u64(neon::vreinterpret_u64_u8(cmp), 0))
+        }
+    }
+
+    /// Returns a `BitMask` indicating all bytes in the group which are
+    /// `EMPTY`.
+    #[inline]
+    pub(crate) fn match_empty(self) -> BitMask {
+        self.match_byte(EMPTY)
+    }
+
+    /// Returns a `BitMask` indicating all bytes in the group which are
+    /// `EMPTY` or `DELETED`.
+    #[inline]
+    pub(crate) fn match_empty_or_deleted(self) -> BitMask {
+        unsafe {
+            let cmp = neon::vcltz_s8(neon::vreinterpret_s8_u8(self.0));
+            BitMask(neon::vget_lane_u64(neon::vreinterpret_u64_u8(cmp), 0))
+        }
+    }
+
+    /// Returns a `BitMask` indicating all bytes in the group which are full.
+    #[inline]
+    pub(crate) fn match_full(self) -> BitMask {
+        unsafe {
+            let cmp = neon::vcgez_s8(neon::vreinterpret_s8_u8(self.0));
+            BitMask(neon::vget_lane_u64(neon::vreinterpret_u64_u8(cmp), 0))
+        }
+    }
+
+    /// Performs the following transformation on all bytes in the group:
+    /// - `EMPTY => EMPTY`
+    /// - `DELETED => EMPTY`
+    /// - `FULL => DELETED`
+    #[inline]
+    pub(crate) fn convert_special_to_empty_and_full_to_deleted(self) -> Self {
+        // Map high_bit = 1 (EMPTY or DELETED) to 1111_1111
+        // and high_bit = 0 (FULL) to 1000_0000
+        //
+        // Here's this logic expanded to concrete values:
+        //   let special = 0 > byte = 1111_1111 (true) or 0000_0000 (false)
+        //   1111_1111 | 1000_0000 = 1111_1111
+        //   0000_0000 | 1000_0000 = 1000_0000
+        unsafe {
+            let special = neon::vcltz_s8(neon::vreinterpret_s8_u8(self.0));
+            Group(neon::vorr_u8(special, neon::vdup_n_u8(0x80)))
+        }
+    }
+}
diff --git a/third_party/rust/hashbrown/src/raw/sse2.rs b/third_party/rust/hashbrown/src/raw/sse2.rs
index a0bf6da804..956ba5d265 100644
--- a/third_party/rust/hashbrown/src/raw/sse2.rs
+++ b/third_party/rust/hashbrown/src/raw/sse2.rs
@@ -1,28 +1,31 @@
 use super::bitmask::BitMask;
 use super::EMPTY;
 use core::mem;
+use core::num::NonZeroU16;
 
 #[cfg(target_arch = "x86")]
 use core::arch::x86;
 #[cfg(target_arch = "x86_64")]
 use core::arch::x86_64 as x86;
 
-pub type BitMaskWord = u16;
-pub const BITMASK_STRIDE: usize = 1;
-pub const BITMASK_MASK: BitMaskWord = 0xffff;
+pub(crate) type BitMaskWord = u16;
+pub(crate) type NonZeroBitMaskWord = NonZeroU16;
+pub(crate) const BITMASK_STRIDE: usize = 1;
+pub(crate) const BITMASK_MASK: BitMaskWord = 0xffff;
+pub(crate) const BITMASK_ITER_MASK: BitMaskWord = !0;
 
 /// Abstraction over a group of control bytes which can be scanned in
 /// parallel.
 ///
 /// This implementation uses a 128-bit SSE value.
 #[derive(Copy, Clone)]
-pub struct Group(x86::__m128i);
+pub(crate) struct Group(x86::__m128i);
 
 // FIXME: https://github.com/rust-lang/rust-clippy/issues/3859
 #[allow(clippy::use_self)]
 impl Group {
     /// Number of bytes in the group.
-    pub const WIDTH: usize = mem::size_of::<Self>();
+    pub(crate) const WIDTH: usize = mem::size_of::<Self>();
 
     /// Returns a full group of empty bytes, suitable for use as the initial
     /// value for an empty hash table.
@@ -30,7 +33,7 @@ impl Group {
     /// This is guaranteed to be aligned to the group size.
     #[inline]
     #[allow(clippy::items_after_statements)]
-    pub const fn static_empty() -> &'static [u8; Group::WIDTH] {
+    pub(crate) const fn static_empty() -> &'static [u8; Group::WIDTH] {
         #[repr(C)]
         struct AlignedBytes {
             _align: [Group; 0],
@@ -46,7 +49,7 @@ impl Group {
     /// Loads a group of bytes starting at the given address.
     #[inline]
     #[allow(clippy::cast_ptr_alignment)] // unaligned load
-    pub unsafe fn load(ptr: *const u8) -> Self {
+    pub(crate) unsafe fn load(ptr: *const u8) -> Self {
         Group(x86::_mm_loadu_si128(ptr.cast()))
     }
 
@@ -54,7 +57,7 @@ impl Group {
     /// aligned to `mem::align_of::<Group>()`.
     #[inline]
     #[allow(clippy::cast_ptr_alignment)]
-    pub unsafe fn load_aligned(ptr: *const u8) -> Self {
+    pub(crate) unsafe fn load_aligned(ptr: *const u8) -> Self {
         // FIXME: use align_offset once it stabilizes
         debug_assert_eq!(ptr as usize & (mem::align_of::<Self>() - 1), 0);
         Group(x86::_mm_load_si128(ptr.cast()))
@@ -64,7 +67,7 @@ impl Group {
     /// aligned to `mem::align_of::<Group>()`.
     #[inline]
     #[allow(clippy::cast_ptr_alignment)]
-    pub unsafe fn store_aligned(self, ptr: *mut u8) {
+    pub(crate) unsafe fn store_aligned(self, ptr: *mut u8) {
         // FIXME: use align_offset once it stabilizes
         debug_assert_eq!(ptr as usize & (mem::align_of::<Self>() - 1), 0);
         x86::_mm_store_si128(ptr.cast(), self.0);
@@ -73,7 +76,7 @@ impl Group {
     /// Returns a `BitMask` indicating all bytes in the group which have
     /// the given value.
     #[inline]
-    pub fn match_byte(self, byte: u8) -> BitMask {
+    pub(crate) fn match_byte(self, byte: u8) -> BitMask {
         #[allow(
             clippy::cast_possible_wrap, // byte: u8 as i8
             // byte: i32 as u16
@@ -91,14 +94,14 @@ impl Group {
     /// Returns a `BitMask` indicating all bytes in the group which are
     /// `EMPTY`.
     #[inline]
-    pub fn match_empty(self) -> BitMask {
+    pub(crate) fn match_empty(self) -> BitMask {
         self.match_byte(EMPTY)
     }
 
     /// Returns a `BitMask` indicating all bytes in the group which are
     /// `EMPTY` or `DELETED`.
     #[inline]
-    pub fn match_empty_or_deleted(self) -> BitMask {
+    pub(crate) fn match_empty_or_deleted(self) -> BitMask {
         #[allow(
             // byte: i32 as u16
             //   note: _mm_movemask_epi8 returns a 16-bit mask in a i32, the
@@ -114,7 +117,7 @@ impl Group {
 
     /// Returns a `BitMask` indicating all bytes in the group which are full.
     #[inline]
-    pub fn match_full(&self) -> BitMask {
+    pub(crate) fn match_full(&self) -> BitMask {
         self.match_empty_or_deleted().invert()
     }
 
@@ -123,7 +126,7 @@ impl Group {
     /// - `DELETED => EMPTY`
     /// - `FULL => DELETED`
     #[inline]
-    pub fn convert_special_to_empty_and_full_to_deleted(self) -> Self {
+    pub(crate) fn convert_special_to_empty_and_full_to_deleted(self) -> Self {
         // Map high_bit = 1 (EMPTY or DELETED) to 1111_1111
         // and high_bit = 0 (FULL) to 1000_0000
         //