1 | //! Code for efficiently counting the number of `char`s in a UTF-8 encoded |
2 | //! string. |
3 | //! |
4 | //! Broadly, UTF-8 encodes `char`s as a "leading" byte which begins the `char`, |
5 | //! followed by some number (possibly 0) of continuation bytes. |
6 | //! |
7 | //! The leading byte can have a number of bit-patterns (with the specific |
8 | //! pattern indicating how many continuation bytes follow), but the continuation |
9 | //! bytes are always in the format `0b10XX_XXXX` (where the `X`s can take any |
10 | //! value). That is, the most significant bit is set, and the second most |
11 | //! significant bit is unset. |
12 | //! |
13 | //! To count the number of characters, we can just count the number of bytes in |
14 | //! the string which are not continuation bytes, which can be done many bytes at |
15 | //! a time fairly easily. |
16 | //! |
17 | //! Note: Because the term "leading byte" can sometimes be ambiguous (for |
18 | //! example, it could also refer to the first byte of a slice), we'll often use |
19 | //! the term "non-continuation byte" to refer to these bytes in the code. |
20 | use core::intrinsics::unlikely; |
21 | |
22 | const USIZE_SIZE: usize = core::mem::size_of::<usize>(); |
23 | const UNROLL_INNER: usize = 4; |
24 | |
25 | #[inline ] |
26 | pub(super) fn count_chars(s: &str) -> usize { |
27 | if s.len() < USIZE_SIZE * UNROLL_INNER { |
28 | // Avoid entering the optimized implementation for strings where the |
29 | // difference is not likely to matter, or where it might even be slower. |
30 | // That said, a ton of thought was not spent on the particular threshold |
31 | // here, beyond "this value seems to make sense". |
32 | char_count_general_case(s.as_bytes()) |
33 | } else { |
34 | do_count_chars(s) |
35 | } |
36 | } |
37 | |
38 | fn do_count_chars(s: &str) -> usize { |
39 | // For correctness, `CHUNK_SIZE` must be: |
40 | // |
41 | // - Less than or equal to 255, otherwise we'll overflow bytes in `counts`. |
42 | // - A multiple of `UNROLL_INNER`, otherwise our `break` inside the |
43 | // `body.chunks(CHUNK_SIZE)` loop is incorrect. |
44 | // |
45 | // For performance, `CHUNK_SIZE` should be: |
46 | // - Relatively cheap to `/` against (so some simple sum of powers of two). |
47 | // - Large enough to avoid paying for the cost of the `sum_bytes_in_usize` |
48 | // too often. |
49 | const CHUNK_SIZE: usize = 192; |
50 | |
51 | // Check the properties of `CHUNK_SIZE` and `UNROLL_INNER` that are required |
52 | // for correctness. |
53 | const _: () = assert!(CHUNK_SIZE < 256); |
54 | const _: () = assert!(CHUNK_SIZE % UNROLL_INNER == 0); |
55 | |
56 | // SAFETY: transmuting `[u8]` to `[usize]` is safe except for size |
57 | // differences which are handled by `align_to`. |
58 | let (head, body, tail) = unsafe { s.as_bytes().align_to::<usize>() }; |
59 | |
60 | // This should be quite rare, and basically exists to handle the degenerate |
61 | // cases where align_to fails (as well as miri under symbolic alignment |
62 | // mode). |
63 | // |
64 | // The `unlikely` helps discourage LLVM from inlining the body, which is |
65 | // nice, as we would rather not mark the `char_count_general_case` function |
66 | // as cold. |
67 | if unlikely(body.is_empty() || head.len() > USIZE_SIZE || tail.len() > USIZE_SIZE) { |
68 | return char_count_general_case(s.as_bytes()); |
69 | } |
70 | |
71 | let mut total = char_count_general_case(head) + char_count_general_case(tail); |
72 | // Split `body` into `CHUNK_SIZE` chunks to reduce the frequency with which |
73 | // we call `sum_bytes_in_usize`. |
74 | for chunk in body.chunks(CHUNK_SIZE) { |
75 | // We accumulate intermediate sums in `counts`, where each byte contains |
76 | // a subset of the sum of this chunk, like a `[u8; size_of::<usize>()]`. |
77 | let mut counts = 0; |
78 | |
79 | let (unrolled_chunks, remainder) = chunk.as_chunks::<UNROLL_INNER>(); |
80 | for unrolled in unrolled_chunks { |
81 | for &word in unrolled { |
82 | // Because `CHUNK_SIZE` is < 256, this addition can't cause the |
83 | // count in any of the bytes to overflow into a subsequent byte. |
84 | counts += contains_non_continuation_byte(word); |
85 | } |
86 | } |
87 | |
88 | // Sum the values in `counts` (which, again, is conceptually a `[u8; |
89 | // size_of::<usize>()]`), and accumulate the result into `total`. |
90 | total += sum_bytes_in_usize(counts); |
91 | |
92 | // If there's any data in `remainder`, then handle it. This will only |
93 | // happen for the last `chunk` in `body.chunks()` (because `CHUNK_SIZE` |
94 | // is divisible by `UNROLL_INNER`), so we explicitly break at the end |
95 | // (which seems to help LLVM out). |
96 | if !remainder.is_empty() { |
97 | // Accumulate all the data in the remainder. |
98 | let mut counts = 0; |
99 | for &word in remainder { |
100 | counts += contains_non_continuation_byte(word); |
101 | } |
102 | total += sum_bytes_in_usize(counts); |
103 | break; |
104 | } |
105 | } |
106 | total |
107 | } |
108 | |
109 | // Checks each byte of `w` to see if it contains the first byte in a UTF-8 |
110 | // sequence. Bytes in `w` which are continuation bytes are left as `0x00` (e.g. |
111 | // false), and bytes which are non-continuation bytes are left as `0x01` (e.g. |
112 | // true) |
113 | #[inline ] |
114 | fn contains_non_continuation_byte(w: usize) -> usize { |
115 | const LSB: usize = usize::repeat_u8(0x01); |
116 | ((!w >> 7) | (w >> 6)) & LSB |
117 | } |
118 | |
119 | // Morally equivalent to `values.to_ne_bytes().into_iter().sum::<usize>()`, but |
120 | // more efficient. |
121 | #[inline ] |
122 | fn sum_bytes_in_usize(values: usize) -> usize { |
123 | const LSB_SHORTS: usize = usize::repeat_u16(0x0001); |
124 | const SKIP_BYTES: usize = usize::repeat_u16(0x00ff); |
125 | |
126 | let pair_sum: usize = (values & SKIP_BYTES) + ((values >> 8) & SKIP_BYTES); |
127 | pair_sum.wrapping_mul(LSB_SHORTS) >> ((USIZE_SIZE - 2) * 8) |
128 | } |
129 | |
130 | // This is the most direct implementation of the concept of "count the number of |
131 | // bytes in the string which are not continuation bytes", and is used for the |
132 | // head and tail of the input string (the first and last item in the tuple |
133 | // returned by `slice::align_to`). |
134 | fn char_count_general_case(s: &[u8]) -> usize { |
135 | s.iter().filter(|&&byte: u8| !super::validations::utf8_is_cont_byte(byte)).count() |
136 | } |
137 | |