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
21use core::intrinsics::unlikely;
22
23const USIZE_SIZE: usize = size_of::<usize>();
24const UNROLL_INNER: usize = 4;
25
26#[inline]
27pub(super) fn count_chars(s: &str) -> usize {
28 if cfg!(feature = "optimize_for_size") || s.len() < USIZE_SIZE * UNROLL_INNER {
29 // Avoid entering the optimized implementation for strings where the
30 // difference is not likely to matter, or where it might even be slower.
31 // That said, a ton of thought was not spent on the particular threshold
32 // here, beyond "this value seems to make sense".
33 char_count_general_case(s.as_bytes())
34 } else {
35 do_count_chars(s)
36 }
37}
38
39fn do_count_chars(s: &str) -> usize {
40 // For correctness, `CHUNK_SIZE` must be:
41 //
42 // - Less than or equal to 255, otherwise we'll overflow bytes in `counts`.
43 // - A multiple of `UNROLL_INNER`, otherwise our `break` inside the
44 // `body.chunks(CHUNK_SIZE)` loop is incorrect.
45 //
46 // For performance, `CHUNK_SIZE` should be:
47 // - Relatively cheap to `/` against (so some simple sum of powers of two).
48 // - Large enough to avoid paying for the cost of the `sum_bytes_in_usize`
49 // too often.
50 const CHUNK_SIZE: usize = 192;
51
52 // Check the properties of `CHUNK_SIZE` and `UNROLL_INNER` that are required
53 // for correctness.
54 const _: () = assert!(CHUNK_SIZE < 256);
55 const _: () = assert!(CHUNK_SIZE % UNROLL_INNER == 0);
56
57 // SAFETY: transmuting `[u8]` to `[usize]` is safe except for size
58 // differences which are handled by `align_to`.
59 let (head, body, tail) = unsafe { s.as_bytes().align_to::<usize>() };
60
61 // This should be quite rare, and basically exists to handle the degenerate
62 // cases where align_to fails (as well as miri under symbolic alignment
63 // mode).
64 //
65 // The `unlikely` helps discourage LLVM from inlining the body, which is
66 // nice, as we would rather not mark the `char_count_general_case` function
67 // as cold.
68 if unlikely(body.is_empty() || head.len() > USIZE_SIZE || tail.len() > USIZE_SIZE) {
69 return char_count_general_case(s.as_bytes());
70 }
71
72 let mut total = char_count_general_case(head) + char_count_general_case(tail);
73 // Split `body` into `CHUNK_SIZE` chunks to reduce the frequency with which
74 // we call `sum_bytes_in_usize`.
75 for chunk in body.chunks(CHUNK_SIZE) {
76 // We accumulate intermediate sums in `counts`, where each byte contains
77 // a subset of the sum of this chunk, like a `[u8; size_of::<usize>()]`.
78 let mut counts = 0;
79
80 let (unrolled_chunks, remainder) = chunk.as_chunks::<UNROLL_INNER>();
81 for unrolled in unrolled_chunks {
82 for &word in unrolled {
83 // Because `CHUNK_SIZE` is < 256, this addition can't cause the
84 // count in any of the bytes to overflow into a subsequent byte.
85 counts += contains_non_continuation_byte(word);
86 }
87 }
88
89 // Sum the values in `counts` (which, again, is conceptually a `[u8;
90 // size_of::<usize>()]`), and accumulate the result into `total`.
91 total += sum_bytes_in_usize(counts);
92
93 // If there's any data in `remainder`, then handle it. This will only
94 // happen for the last `chunk` in `body.chunks()` (because `CHUNK_SIZE`
95 // is divisible by `UNROLL_INNER`), so we explicitly break at the end
96 // (which seems to help LLVM out).
97 if !remainder.is_empty() {
98 // Accumulate all the data in the remainder.
99 let mut counts = 0;
100 for &word in remainder {
101 counts += contains_non_continuation_byte(word);
102 }
103 total += sum_bytes_in_usize(counts);
104 break;
105 }
106 }
107 total
108}
109
110// Checks each byte of `w` to see if it contains the first byte in a UTF-8
111// sequence. Bytes in `w` which are continuation bytes are left as `0x00` (e.g.
112// false), and bytes which are non-continuation bytes are left as `0x01` (e.g.
113// true)
114#[inline]
115fn contains_non_continuation_byte(w: usize) -> usize {
116 const LSB: usize = usize::repeat_u8(0x01);
117 ((!w >> 7) | (w >> 6)) & LSB
118}
119
120// Morally equivalent to `values.to_ne_bytes().into_iter().sum::<usize>()`, but
121// more efficient.
122#[inline]
123fn sum_bytes_in_usize(values: usize) -> usize {
124 const LSB_SHORTS: usize = usize::repeat_u16(0x0001);
125 const SKIP_BYTES: usize = usize::repeat_u16(0x00ff);
126
127 let pair_sum: usize = (values & SKIP_BYTES) + ((values >> 8) & SKIP_BYTES);
128 pair_sum.wrapping_mul(LSB_SHORTS) >> ((USIZE_SIZE - 2) * 8)
129}
130
131// This is the most direct implementation of the concept of "count the number of
132// bytes in the string which are not continuation bytes", and is used for the
133// head and tail of the input string (the first and last item in the tuple
134// returned by `slice::align_to`).
135fn char_count_general_case(s: &[u8]) -> usize {
136 s.iter().filter(|&&byte: u8| !super::validations::utf8_is_cont_byte(byte)).count()
137}
138