| 1 | use super::super::{BitMask, Tag}; |
| 2 | use core::{mem, ptr}; |
| 3 | |
| 4 | // Use the native word size as the group size. Using a 64-bit group size on |
| 5 | // a 32-bit architecture will just end up being more expensive because |
| 6 | // shifts and multiplies will need to be emulated. |
| 7 | |
| 8 | cfg_if! { |
| 9 | if #[cfg(any( |
| 10 | target_pointer_width = "64" , |
| 11 | target_arch = "aarch64" , |
| 12 | target_arch = "x86_64" , |
| 13 | target_arch = "wasm32" , |
| 14 | ))] { |
| 15 | type GroupWord = u64; |
| 16 | type NonZeroGroupWord = core::num::NonZeroU64; |
| 17 | } else { |
| 18 | type GroupWord = u32; |
| 19 | type NonZeroGroupWord = core::num::NonZeroU32; |
| 20 | } |
| 21 | } |
| 22 | |
| 23 | pub(crate) type BitMaskWord = GroupWord; |
| 24 | pub(crate) type NonZeroBitMaskWord = NonZeroGroupWord; |
| 25 | pub(crate) const BITMASK_STRIDE: usize = 8; |
| 26 | // We only care about the highest bit of each tag for the mask. |
| 27 | #[allow (clippy::cast_possible_truncation, clippy::unnecessary_cast)] |
| 28 | pub(crate) const BITMASK_MASK: BitMaskWord = u64::from_ne_bytes([Tag::DELETED.0; 8]) as GroupWord; |
| 29 | pub(crate) const BITMASK_ITER_MASK: BitMaskWord = !0; |
| 30 | |
| 31 | /// Helper function to replicate a tag across a `GroupWord`. |
| 32 | #[inline ] |
| 33 | fn repeat(tag: Tag) -> GroupWord { |
| 34 | GroupWord::from_ne_bytes([tag.0; Group::WIDTH]) |
| 35 | } |
| 36 | |
| 37 | /// Abstraction over a group of control tags which can be scanned in |
| 38 | /// parallel. |
| 39 | /// |
| 40 | /// This implementation uses a word-sized integer. |
| 41 | #[derive (Copy, Clone)] |
| 42 | pub(crate) struct Group(GroupWord); |
| 43 | |
| 44 | // We perform all operations in the native endianness, and convert to |
| 45 | // little-endian just before creating a BitMask. The can potentially |
| 46 | // enable the compiler to eliminate unnecessary byte swaps if we are |
| 47 | // only checking whether a BitMask is empty. |
| 48 | #[allow (clippy::use_self)] |
| 49 | impl Group { |
| 50 | /// Number of bytes in the group. |
| 51 | pub(crate) const WIDTH: usize = mem::size_of::<Self>(); |
| 52 | |
| 53 | /// Returns a full group of empty tags, suitable for use as the initial |
| 54 | /// value for an empty hash table. |
| 55 | /// |
| 56 | /// This is guaranteed to be aligned to the group size. |
| 57 | #[inline ] |
| 58 | pub(crate) const fn static_empty() -> &'static [Tag; Group::WIDTH] { |
| 59 | #[repr (C)] |
| 60 | struct AlignedTags { |
| 61 | _align: [Group; 0], |
| 62 | tags: [Tag; Group::WIDTH], |
| 63 | } |
| 64 | const ALIGNED_TAGS: AlignedTags = AlignedTags { |
| 65 | _align: [], |
| 66 | tags: [Tag::EMPTY; Group::WIDTH], |
| 67 | }; |
| 68 | &ALIGNED_TAGS.tags |
| 69 | } |
| 70 | |
| 71 | /// Loads a group of tags starting at the given address. |
| 72 | #[inline ] |
| 73 | #[allow (clippy::cast_ptr_alignment)] // unaligned load |
| 74 | pub(crate) unsafe fn load(ptr: *const Tag) -> Self { |
| 75 | Group(ptr::read_unaligned(ptr.cast())) |
| 76 | } |
| 77 | |
| 78 | /// Loads a group of tags starting at the given address, which must be |
| 79 | /// aligned to `mem::align_of::<Group>()`. |
| 80 | #[inline ] |
| 81 | #[allow (clippy::cast_ptr_alignment)] |
| 82 | pub(crate) unsafe fn load_aligned(ptr: *const Tag) -> Self { |
| 83 | debug_assert_eq!(ptr.align_offset(mem::align_of::<Self>()), 0); |
| 84 | Group(ptr::read(ptr.cast())) |
| 85 | } |
| 86 | |
| 87 | /// Stores the group of tags to the given address, which must be |
| 88 | /// aligned to `mem::align_of::<Group>()`. |
| 89 | #[inline ] |
| 90 | #[allow (clippy::cast_ptr_alignment)] |
| 91 | pub(crate) unsafe fn store_aligned(self, ptr: *mut Tag) { |
| 92 | debug_assert_eq!(ptr.align_offset(mem::align_of::<Self>()), 0); |
| 93 | ptr::write(ptr.cast(), self.0); |
| 94 | } |
| 95 | |
| 96 | /// Returns a `BitMask` indicating all tags in the group which *may* |
| 97 | /// have the given value. |
| 98 | /// |
| 99 | /// This function may return a false positive in certain cases where |
| 100 | /// the tag in the group differs from the searched value only in its |
| 101 | /// lowest bit. This is fine because: |
| 102 | /// - This never happens for `EMPTY` and `DELETED`, only full entries. |
| 103 | /// - The check for key equality will catch these. |
| 104 | /// - This only happens if there is at least 1 true match. |
| 105 | /// - The chance of this happening is very low (< 1% chance per byte). |
| 106 | #[inline ] |
| 107 | pub(crate) fn match_tag(self, tag: Tag) -> BitMask { |
| 108 | // This algorithm is derived from |
| 109 | // https://graphics.stanford.edu/~seander/bithacks.html##ValueInWord |
| 110 | let cmp = self.0 ^ repeat(tag); |
| 111 | BitMask((cmp.wrapping_sub(repeat(Tag(0x01))) & !cmp & repeat(Tag::DELETED)).to_le()) |
| 112 | } |
| 113 | |
| 114 | /// Returns a `BitMask` indicating all tags in the group which are |
| 115 | /// `EMPTY`. |
| 116 | #[inline ] |
| 117 | pub(crate) fn match_empty(self) -> BitMask { |
| 118 | // If the high bit is set, then the tag must be either: |
| 119 | // 1111_1111 (EMPTY) or 1000_0000 (DELETED). |
| 120 | // So we can just check if the top two bits are 1 by ANDing them. |
| 121 | BitMask((self.0 & (self.0 << 1) & repeat(Tag::DELETED)).to_le()) |
| 122 | } |
| 123 | |
| 124 | /// Returns a `BitMask` indicating all tags in the group which are |
| 125 | /// `EMPTY` or `DELETED`. |
| 126 | #[inline ] |
| 127 | pub(crate) fn match_empty_or_deleted(self) -> BitMask { |
| 128 | // A tag is EMPTY or DELETED iff the high bit is set |
| 129 | BitMask((self.0 & repeat(Tag::DELETED)).to_le()) |
| 130 | } |
| 131 | |
| 132 | /// Returns a `BitMask` indicating all tags in the group which are full. |
| 133 | #[inline ] |
| 134 | pub(crate) fn match_full(self) -> BitMask { |
| 135 | self.match_empty_or_deleted().invert() |
| 136 | } |
| 137 | |
| 138 | /// Performs the following transformation on all tags in the group: |
| 139 | /// - `EMPTY => EMPTY` |
| 140 | /// - `DELETED => EMPTY` |
| 141 | /// - `FULL => DELETED` |
| 142 | #[inline ] |
| 143 | pub(crate) fn convert_special_to_empty_and_full_to_deleted(self) -> Self { |
| 144 | // Map high_bit = 1 (EMPTY or DELETED) to 1111_1111 |
| 145 | // and high_bit = 0 (FULL) to 1000_0000 |
| 146 | // |
| 147 | // Here's this logic expanded to concrete values: |
| 148 | // let full = 1000_0000 (true) or 0000_0000 (false) |
| 149 | // !1000_0000 + 1 = 0111_1111 + 1 = 1000_0000 (no carry) |
| 150 | // !0000_0000 + 0 = 1111_1111 + 0 = 1111_1111 (no carry) |
| 151 | let full = !self.0 & repeat(Tag::DELETED); |
| 152 | Group(!full + (full >> 7)) |
| 153 | } |
| 154 | } |
| 155 | |