core/slice/
ascii.rs

1//! Operations on ASCII `[u8]`.
2
3#[cfg(not(feature = "ferrocene_subset"))]
4use core::ascii::EscapeDefault;
5
6#[cfg(not(feature = "ferrocene_subset"))]
7use crate::fmt::{self, Write};
8#[cfg(not(any(
9    all(target_arch = "x86_64", target_feature = "sse2"),
10    all(target_arch = "loongarch64", target_feature = "lsx")
11)))]
12#[cfg(not(feature = "ferrocene_subset"))]
13use crate::intrinsics::const_eval_select;
14#[cfg(not(feature = "ferrocene_subset"))]
15use crate::{ascii, iter, ops};
16
17impl [u8] {
18    /// Checks if all bytes in this slice are within the ASCII range.
19    ///
20    /// An empty slice returns `true`.
21    #[cfg(not(feature = "ferrocene_subset"))]
22    #[stable(feature = "ascii_methods_on_intrinsics", since = "1.23.0")]
23    #[rustc_const_stable(feature = "const_slice_is_ascii", since = "1.74.0")]
24    #[must_use]
25    #[inline]
26    pub const fn is_ascii(&self) -> bool {
27        is_ascii(self)
28    }
29
30    /// If this slice [`is_ascii`](Self::is_ascii), returns it as a slice of
31    /// [ASCII characters](`ascii::Char`), otherwise returns `None`.
32    #[cfg(not(feature = "ferrocene_subset"))]
33    #[unstable(feature = "ascii_char", issue = "110998")]
34    #[must_use]
35    #[inline]
36    pub const fn as_ascii(&self) -> Option<&[ascii::Char]> {
37        if self.is_ascii() {
38            // SAFETY: Just checked that it's ASCII
39            Some(unsafe { self.as_ascii_unchecked() })
40        } else {
41            None
42        }
43    }
44
45    /// Converts this slice of bytes into a slice of ASCII characters,
46    /// without checking whether they're valid.
47    ///
48    /// # Safety
49    ///
50    /// Every byte in the slice must be in `0..=127`, or else this is UB.
51    #[cfg(not(feature = "ferrocene_subset"))]
52    #[unstable(feature = "ascii_char", issue = "110998")]
53    #[must_use]
54    #[inline]
55    pub const unsafe fn as_ascii_unchecked(&self) -> &[ascii::Char] {
56        let byte_ptr: *const [u8] = self;
57        let ascii_ptr = byte_ptr as *const [ascii::Char];
58        // SAFETY: The caller promised all the bytes are ASCII
59        unsafe { &*ascii_ptr }
60    }
61
62    /// Checks that two slices are an ASCII case-insensitive match.
63    ///
64    /// Same as `to_ascii_lowercase(a) == to_ascii_lowercase(b)`,
65    /// but without allocating and copying temporaries.
66    #[stable(feature = "ascii_methods_on_intrinsics", since = "1.23.0")]
67    #[rustc_const_stable(feature = "const_eq_ignore_ascii_case", since = "1.89.0")]
68    #[must_use]
69    #[inline]
70    pub const fn eq_ignore_ascii_case(&self, other: &[u8]) -> bool {
71        if self.len() != other.len() {
72            return false;
73        }
74
75        // FIXME(const-hack): This implementation can be reverted when
76        // `core::iter::zip` is allowed in const. The original implementation:
77        //  self.len() == other.len() && iter::zip(self, other).all(|(a, b)| a.eq_ignore_ascii_case(b))
78        let mut a = self;
79        let mut b = other;
80
81        while let ([first_a, rest_a @ ..], [first_b, rest_b @ ..]) = (a, b) {
82            if first_a.eq_ignore_ascii_case(&first_b) {
83                a = rest_a;
84                b = rest_b;
85            } else {
86                return false;
87            }
88        }
89
90        true
91    }
92
93    /// Converts this slice to its ASCII upper case equivalent in-place.
94    ///
95    /// ASCII letters 'a' to 'z' are mapped to 'A' to 'Z',
96    /// but non-ASCII letters are unchanged.
97    ///
98    /// To return a new uppercased value without modifying the existing one, use
99    /// [`to_ascii_uppercase`].
100    ///
101    /// [`to_ascii_uppercase`]: #method.to_ascii_uppercase
102    #[cfg(not(feature = "ferrocene_subset"))]
103    #[stable(feature = "ascii_methods_on_intrinsics", since = "1.23.0")]
104    #[rustc_const_stable(feature = "const_make_ascii", since = "1.84.0")]
105    #[inline]
106    pub const fn make_ascii_uppercase(&mut self) {
107        // FIXME(const-hack): We would like to simply iterate using `for` loops but this isn't currently allowed in constant expressions.
108        let mut i = 0;
109        while i < self.len() {
110            let byte = &mut self[i];
111            byte.make_ascii_uppercase();
112            i += 1;
113        }
114    }
115
116    /// Converts this slice to its ASCII lower case equivalent in-place.
117    ///
118    /// ASCII letters 'A' to 'Z' are mapped to 'a' to 'z',
119    /// but non-ASCII letters are unchanged.
120    ///
121    /// To return a new lowercased value without modifying the existing one, use
122    /// [`to_ascii_lowercase`].
123    ///
124    /// [`to_ascii_lowercase`]: #method.to_ascii_lowercase
125    #[cfg(not(feature = "ferrocene_subset"))]
126    #[stable(feature = "ascii_methods_on_intrinsics", since = "1.23.0")]
127    #[rustc_const_stable(feature = "const_make_ascii", since = "1.84.0")]
128    #[inline]
129    pub const fn make_ascii_lowercase(&mut self) {
130        // FIXME(const-hack): We would like to simply iterate using `for` loops but this isn't currently allowed in constant expressions.
131        let mut i = 0;
132        while i < self.len() {
133            let byte = &mut self[i];
134            byte.make_ascii_lowercase();
135            i += 1;
136        }
137    }
138
139    /// Returns an iterator that produces an escaped version of this slice,
140    /// treating it as an ASCII string.
141    ///
142    /// # Examples
143    ///
144    /// ```
145    /// let s = b"0\t\r\n'\"\\\x9d";
146    /// let escaped = s.escape_ascii().to_string();
147    /// assert_eq!(escaped, "0\\t\\r\\n\\'\\\"\\\\\\x9d");
148    /// ```
149    #[cfg(not(feature = "ferrocene_subset"))]
150    #[must_use = "this returns the escaped bytes as an iterator, \
151                  without modifying the original"]
152    #[stable(feature = "inherent_ascii_escape", since = "1.60.0")]
153    pub fn escape_ascii(&self) -> EscapeAscii<'_> {
154        EscapeAscii { inner: self.iter().flat_map(EscapeByte) }
155    }
156
157    /// Returns a byte slice with leading ASCII whitespace bytes removed.
158    ///
159    /// 'Whitespace' refers to the definition used by
160    /// [`u8::is_ascii_whitespace`].
161    ///
162    /// # Examples
163    ///
164    /// ```
165    /// assert_eq!(b" \t hello world\n".trim_ascii_start(), b"hello world\n");
166    /// assert_eq!(b"  ".trim_ascii_start(), b"");
167    /// assert_eq!(b"".trim_ascii_start(), b"");
168    /// ```
169    #[cfg(not(feature = "ferrocene_subset"))]
170    #[stable(feature = "byte_slice_trim_ascii", since = "1.80.0")]
171    #[rustc_const_stable(feature = "byte_slice_trim_ascii", since = "1.80.0")]
172    #[inline]
173    pub const fn trim_ascii_start(&self) -> &[u8] {
174        let mut bytes = self;
175        // Note: A pattern matching based approach (instead of indexing) allows
176        // making the function const.
177        while let [first, rest @ ..] = bytes {
178            if first.is_ascii_whitespace() {
179                bytes = rest;
180            } else {
181                break;
182            }
183        }
184        bytes
185    }
186
187    /// Returns a byte slice with trailing ASCII whitespace bytes removed.
188    ///
189    /// 'Whitespace' refers to the definition used by
190    /// [`u8::is_ascii_whitespace`].
191    ///
192    /// # Examples
193    ///
194    /// ```
195    /// assert_eq!(b"\r hello world\n ".trim_ascii_end(), b"\r hello world");
196    /// assert_eq!(b"  ".trim_ascii_end(), b"");
197    /// assert_eq!(b"".trim_ascii_end(), b"");
198    /// ```
199    #[cfg(not(feature = "ferrocene_subset"))]
200    #[stable(feature = "byte_slice_trim_ascii", since = "1.80.0")]
201    #[rustc_const_stable(feature = "byte_slice_trim_ascii", since = "1.80.0")]
202    #[inline]
203    pub const fn trim_ascii_end(&self) -> &[u8] {
204        let mut bytes = self;
205        // Note: A pattern matching based approach (instead of indexing) allows
206        // making the function const.
207        while let [rest @ .., last] = bytes {
208            if last.is_ascii_whitespace() {
209                bytes = rest;
210            } else {
211                break;
212            }
213        }
214        bytes
215    }
216
217    /// Returns a byte slice with leading and trailing ASCII whitespace bytes
218    /// removed.
219    ///
220    /// 'Whitespace' refers to the definition used by
221    /// [`u8::is_ascii_whitespace`].
222    ///
223    /// # Examples
224    ///
225    /// ```
226    /// assert_eq!(b"\r hello world\n ".trim_ascii(), b"hello world");
227    /// assert_eq!(b"  ".trim_ascii(), b"");
228    /// assert_eq!(b"".trim_ascii(), b"");
229    /// ```
230    #[cfg(not(feature = "ferrocene_subset"))]
231    #[stable(feature = "byte_slice_trim_ascii", since = "1.80.0")]
232    #[rustc_const_stable(feature = "byte_slice_trim_ascii", since = "1.80.0")]
233    #[inline]
234    pub const fn trim_ascii(&self) -> &[u8] {
235        self.trim_ascii_start().trim_ascii_end()
236    }
237}
238
239#[cfg(not(feature = "ferrocene_subset"))]
240impl_fn_for_zst! {
241    #[derive(Clone)]
242    struct EscapeByte impl Fn = |byte: &u8| -> ascii::EscapeDefault {
243        ascii::escape_default(*byte)
244    };
245}
246
247/// An iterator over the escaped version of a byte slice.
248///
249/// This `struct` is created by the [`slice::escape_ascii`] method. See its
250/// documentation for more information.
251#[stable(feature = "inherent_ascii_escape", since = "1.60.0")]
252#[derive(Clone)]
253#[must_use = "iterators are lazy and do nothing unless consumed"]
254#[cfg(not(feature = "ferrocene_subset"))]
255pub struct EscapeAscii<'a> {
256    inner: iter::FlatMap<super::Iter<'a, u8>, ascii::EscapeDefault, EscapeByte>,
257}
258
259#[cfg(not(feature = "ferrocene_subset"))]
260#[stable(feature = "inherent_ascii_escape", since = "1.60.0")]
261impl<'a> iter::Iterator for EscapeAscii<'a> {
262    type Item = u8;
263    #[inline]
264    fn next(&mut self) -> Option<u8> {
265        self.inner.next()
266    }
267    #[inline]
268    fn size_hint(&self) -> (usize, Option<usize>) {
269        self.inner.size_hint()
270    }
271    #[inline]
272    fn try_fold<Acc, Fold, R>(&mut self, init: Acc, fold: Fold) -> R
273    where
274        Fold: FnMut(Acc, Self::Item) -> R,
275        R: ops::Try<Output = Acc>,
276    {
277        self.inner.try_fold(init, fold)
278    }
279    #[inline]
280    fn fold<Acc, Fold>(self, init: Acc, fold: Fold) -> Acc
281    where
282        Fold: FnMut(Acc, Self::Item) -> Acc,
283    {
284        self.inner.fold(init, fold)
285    }
286    #[inline]
287    fn last(mut self) -> Option<u8> {
288        self.next_back()
289    }
290}
291
292#[cfg(not(feature = "ferrocene_subset"))]
293#[stable(feature = "inherent_ascii_escape", since = "1.60.0")]
294impl<'a> iter::DoubleEndedIterator for EscapeAscii<'a> {
295    fn next_back(&mut self) -> Option<u8> {
296        self.inner.next_back()
297    }
298}
299#[cfg(not(feature = "ferrocene_subset"))]
300#[stable(feature = "inherent_ascii_escape", since = "1.60.0")]
301impl<'a> iter::FusedIterator for EscapeAscii<'a> {}
302#[cfg(not(feature = "ferrocene_subset"))]
303#[stable(feature = "inherent_ascii_escape", since = "1.60.0")]
304impl<'a> fmt::Display for EscapeAscii<'a> {
305    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
306        // disassemble iterator, including front/back parts of flatmap in case it has been partially consumed
307        let (front, slice, back) = self.clone().inner.into_parts();
308        let front = front.unwrap_or(EscapeDefault::empty());
309        let mut bytes = slice.unwrap_or_default().as_slice();
310        let back = back.unwrap_or(EscapeDefault::empty());
311
312        // usually empty, so the formatter won't have to do any work
313        for byte in front {
314            f.write_char(byte as char)?;
315        }
316
317        fn needs_escape(b: u8) -> bool {
318            b > 0x7E || b < 0x20 || b == b'\\' || b == b'\'' || b == b'"'
319        }
320
321        while bytes.len() > 0 {
322            // fast path for the printable, non-escaped subset of ascii
323            let prefix = bytes.iter().take_while(|&&b| !needs_escape(b)).count();
324            // SAFETY: prefix length was derived by counting bytes in the same splice, so it's in-bounds
325            let (prefix, remainder) = unsafe { bytes.split_at_unchecked(prefix) };
326            // SAFETY: prefix is a valid utf8 sequence, as it's a subset of ASCII
327            let prefix = unsafe { crate::str::from_utf8_unchecked(prefix) };
328
329            f.write_str(prefix)?; // the fast part
330
331            bytes = remainder;
332
333            if let Some(&b) = bytes.first() {
334                // guaranteed to be non-empty, better to write it as a str
335                fmt::Display::fmt(&ascii::escape_default(b), f)?;
336                bytes = &bytes[1..];
337            }
338        }
339
340        // also usually empty
341        for byte in back {
342            f.write_char(byte as char)?;
343        }
344        Ok(())
345    }
346}
347#[cfg(not(feature = "ferrocene_subset"))]
348#[stable(feature = "inherent_ascii_escape", since = "1.60.0")]
349impl<'a> fmt::Debug for EscapeAscii<'a> {
350    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
351        f.debug_struct("EscapeAscii").finish_non_exhaustive()
352    }
353}
354
355/// ASCII test *without* the chunk-at-a-time optimizations.
356///
357/// This is carefully structured to produce nice small code -- it's smaller in
358/// `-O` than what the "obvious" ways produces under `-C opt-level=s`.  If you
359/// touch it, be sure to run (and update if needed) the assembly test.
360#[cfg(not(feature = "ferrocene_subset"))]
361#[unstable(feature = "str_internals", issue = "none")]
362#[doc(hidden)]
363#[inline]
364pub const fn is_ascii_simple(mut bytes: &[u8]) -> bool {
365    while let [rest @ .., last] = bytes {
366        if !last.is_ascii() {
367            break;
368        }
369        bytes = rest;
370    }
371    bytes.is_empty()
372}
373
374/// Optimized ASCII test that will use usize-at-a-time operations instead of
375/// byte-at-a-time operations (when possible).
376///
377/// The algorithm we use here is pretty simple. If `s` is too short, we just
378/// check each byte and be done with it. Otherwise:
379///
380/// - Read the first word with an unaligned load.
381/// - Align the pointer, read subsequent words until end with aligned loads.
382/// - Read the last `usize` from `s` with an unaligned load.
383///
384/// If any of these loads produces something for which `contains_nonascii`
385/// (above) returns true, then we know the answer is false.
386#[cfg(not(feature = "ferrocene_subset"))]
387#[cfg(not(any(
388    all(target_arch = "x86_64", target_feature = "sse2"),
389    all(target_arch = "loongarch64", target_feature = "lsx")
390)))]
391#[inline]
392#[rustc_allow_const_fn_unstable(const_eval_select)] // fallback impl has same behavior
393const fn is_ascii(s: &[u8]) -> bool {
394    // The runtime version behaves the same as the compiletime version, it's
395    // just more optimized.
396    const_eval_select!(
397        @capture { s: &[u8] } -> bool:
398        if const {
399            is_ascii_simple(s)
400        } else {
401            /// Returns `true` if any byte in the word `v` is nonascii (>= 128). Snarfed
402            /// from `../str/mod.rs`, which does something similar for utf8 validation.
403            const fn contains_nonascii(v: usize) -> bool {
404                const NONASCII_MASK: usize = usize::repeat_u8(0x80);
405                (NONASCII_MASK & v) != 0
406            }
407
408            const USIZE_SIZE: usize = size_of::<usize>();
409
410            let len = s.len();
411            let align_offset = s.as_ptr().align_offset(USIZE_SIZE);
412
413            // If we wouldn't gain anything from the word-at-a-time implementation, fall
414            // back to a scalar loop.
415            //
416            // We also do this for architectures where `size_of::<usize>()` isn't
417            // sufficient alignment for `usize`, because it's a weird edge case.
418            if len < USIZE_SIZE || len < align_offset || USIZE_SIZE < align_of::<usize>() {
419                return is_ascii_simple(s);
420            }
421
422            // We always read the first word unaligned, which means `align_offset` is
423            // 0, we'd read the same value again for the aligned read.
424            let offset_to_aligned = if align_offset == 0 { USIZE_SIZE } else { align_offset };
425
426            let start = s.as_ptr();
427            // SAFETY: We verify `len < USIZE_SIZE` above.
428            let first_word = unsafe { (start as *const usize).read_unaligned() };
429
430            if contains_nonascii(first_word) {
431                return false;
432            }
433            // We checked this above, somewhat implicitly. Note that `offset_to_aligned`
434            // is either `align_offset` or `USIZE_SIZE`, both of are explicitly checked
435            // above.
436            debug_assert!(offset_to_aligned <= len);
437
438            // SAFETY: word_ptr is the (properly aligned) usize ptr we use to read the
439            // middle chunk of the slice.
440            let mut word_ptr = unsafe { start.add(offset_to_aligned) as *const usize };
441
442            // `byte_pos` is the byte index of `word_ptr`, used for loop end checks.
443            let mut byte_pos = offset_to_aligned;
444
445            // Paranoia check about alignment, since we're about to do a bunch of
446            // unaligned loads. In practice this should be impossible barring a bug in
447            // `align_offset` though.
448            // While this method is allowed to spuriously fail in CTFE, if it doesn't
449            // have alignment information it should have given a `usize::MAX` for
450            // `align_offset` earlier, sending things through the scalar path instead of
451            // this one, so this check should pass if it's reachable.
452            debug_assert!(word_ptr.is_aligned_to(align_of::<usize>()));
453
454            // Read subsequent words until the last aligned word, excluding the last
455            // aligned word by itself to be done in tail check later, to ensure that
456            // tail is always one `usize` at most to extra branch `byte_pos == len`.
457            while byte_pos < len - USIZE_SIZE {
458                // Sanity check that the read is in bounds
459                debug_assert!(byte_pos + USIZE_SIZE <= len);
460                // And that our assumptions about `byte_pos` hold.
461                debug_assert!(word_ptr.cast::<u8>() == start.wrapping_add(byte_pos));
462
463                // SAFETY: We know `word_ptr` is properly aligned (because of
464                // `align_offset`), and we know that we have enough bytes between `word_ptr` and the end
465                let word = unsafe { word_ptr.read() };
466                if contains_nonascii(word) {
467                    return false;
468                }
469
470                byte_pos += USIZE_SIZE;
471                // SAFETY: We know that `byte_pos <= len - USIZE_SIZE`, which means that
472                // after this `add`, `word_ptr` will be at most one-past-the-end.
473                word_ptr = unsafe { word_ptr.add(1) };
474            }
475
476            // Sanity check to ensure there really is only one `usize` left. This should
477            // be guaranteed by our loop condition.
478            debug_assert!(byte_pos <= len && len - byte_pos <= USIZE_SIZE);
479
480            // SAFETY: This relies on `len >= USIZE_SIZE`, which we check at the start.
481            let last_word = unsafe { (start.add(len - USIZE_SIZE) as *const usize).read_unaligned() };
482
483            !contains_nonascii(last_word)
484        }
485    )
486}
487
488/// ASCII test optimized to use the `pmovmskb` instruction on `x86-64` and the
489/// `vmskltz.b` instruction on `loongarch64`.
490///
491/// Other platforms are not likely to benefit from this code structure, so they
492/// use SWAR techniques to test for ASCII in `usize`-sized chunks.
493#[cfg(not(feature = "ferrocene_subset"))]
494#[cfg(any(
495    all(target_arch = "x86_64", target_feature = "sse2"),
496    all(target_arch = "loongarch64", target_feature = "lsx")
497))]
498#[inline]
499const fn is_ascii(bytes: &[u8]) -> bool {
500    // Process chunks of 32 bytes at a time in the fast path to enable
501    // auto-vectorization and use of `pmovmskb`. Two 128-bit vector registers
502    // can be OR'd together and then the resulting vector can be tested for
503    // non-ASCII bytes.
504    const CHUNK_SIZE: usize = 32;
505
506    let mut i = 0;
507
508    while i + CHUNK_SIZE <= bytes.len() {
509        let chunk_end = i + CHUNK_SIZE;
510
511        // Get LLVM to produce a `pmovmskb` instruction on x86-64 which
512        // creates a mask from the most significant bit of each byte.
513        // ASCII bytes are less than 128 (0x80), so their most significant
514        // bit is unset.
515        let mut count = 0;
516        while i < chunk_end {
517            count += bytes[i].is_ascii() as u8;
518            i += 1;
519        }
520
521        // All bytes should be <= 127 so count is equal to chunk size.
522        if count != CHUNK_SIZE as u8 {
523            return false;
524        }
525    }
526
527    // Process the remaining `bytes.len() % N` bytes.
528    let mut is_ascii = true;
529    while i < bytes.len() {
530        is_ascii &= bytes[i].is_ascii();
531        i += 1;
532    }
533
534    is_ascii
535}