core: implement Pattern<&[T]> for &[T]; get rid of SlicePattern

Implement Haystack<&[T]> and corresponding Pattern<&[T]> for &[T].
That is, provide implementation for searching for subslices in slices.
This replaces SlicePattern type.

To make use of this new implementations, provide a few new methods on
[T] type modelling them after types on str.  Specifically, introduce
{starts,ends}_with_pattern, find, rfind, [T]::{split,rsplit}_once and
trim{,_start,_end}_matches.

Note that due to existing starts_with and ends_with methods, the
_pattern suffix had to be used.  This is unfortunate but the type
of starts_with’s argument cannot be changed without affecting type
inference and thus breaking the API.

This change doesn’t implement functions returning iterators such as
split_pattern or matches which in str type are built on top of the
Pattern API.

Issue: rust-lang#49802
Issue: rust-lang#56345

Author: mina86

library/core/src/slice/cmp.rs

@@ -227,34 +227,286 @@ impl_marker_for!(BytewiseEquality,
                  u8 i8 u16 i16 u32 i32 u64 i64 u128 i128 usize isize char bool);
 
 pub(super) trait SliceContains: Sized {
-    fn slice_contains(&self, x: &[Self]) -> bool;
+    fn slice_contains_element(hs: &[Self], needle: &Self) -> bool;
+    fn slice_contains_slice(hs: &[Self], needle: &[Self]) -> bool;
 }
 
 impl<T> SliceContains for T
 where
     T: PartialEq,
 {
-    default fn slice_contains(&self, x: &[Self]) -> bool {
-        x.iter().any(|y| *y == *self)
+    default fn slice_contains_element(hs: &[Self], needle: &Self) -> bool {
+        hs.iter().any(|element| *element == *needle)
+    }
+
+    default fn slice_contains_slice(hs: &[Self], needle: &[Self]) -> bool {
+        default_slice_contains_slice(hs, needle)
     }
 }
 
 impl SliceContains for u8 {
     #[inline]
-    fn slice_contains(&self, x: &[Self]) -> bool {
-        memchr::memchr(*self, x).is_some()
+    fn slice_contains_element(hs: &[Self], needle: &Self) -> bool {
+        memchr::memchr(*needle, hs).is_some()
+    }
+
+    #[inline]
+    fn slice_contains_slice(hs: &[Self], needle: &[Self]) -> bool {
+        if needle.len() <= 32 {
+            if let Some(result) = simd_contains(hs, needle) {
+                return result;
+            }
+        }
+        default_slice_contains_slice(hs, needle)
     }
 }
 
+unsafe fn bytes_of<T>(slice: &[T]) -> &[u8] {
+    // SAFETY: caller promises that `T` and `u8` have the same memory layout,
+    // thus casting `x.as_ptr()` as `*const u8` is safe.  The `x.as_ptr()` comes
+    // from a reference and is thus guaranteed to be valid for reads for the
+    // length of the slice `x.len()`, which cannot be larger than
+    // `isize::MAX`. The returned slice is never mutated.
+    unsafe { from_raw_parts(slice.as_ptr() as *const u8, slice.len()) }
+}
+
 impl SliceContains for i8 {
     #[inline]
-    fn slice_contains(&self, x: &[Self]) -> bool {
-        let byte = *self as u8;
-        // SAFETY: `i8` and `u8` have the same memory layout, thus casting `x.as_ptr()`
-        // as `*const u8` is safe. The `x.as_ptr()` comes from a reference and is thus guaranteed
-        // to be valid for reads for the length of the slice `x.len()`, which cannot be larger
-        // than `isize::MAX`. The returned slice is never mutated.
-        let bytes: &[u8] = unsafe { from_raw_parts(x.as_ptr() as *const u8, x.len()) };
-        memchr::memchr(byte, bytes).is_some()
+    fn slice_contains_element(hs: &[Self], needle: &Self) -> bool {
+        // SAFETY: i8 and u8 have the same memory layout
+        u8::slice_contains_element(unsafe { bytes_of(hs) }, &(*needle as u8))
+    }
+
+    #[inline]
+    fn slice_contains_slice(hs: &[Self], needle: &[Self]) -> bool {
+        // SAFETY: i8 and u8 have the same memory layout
+        unsafe { u8::slice_contains_slice(bytes_of(hs), bytes_of(needle)) }
+    }
+}
+
+impl SliceContains for bool {
+    #[inline]
+    fn slice_contains_element(hs: &[Self], needle: &Self) -> bool {
+        // SAFETY: bool and u8 have the same memory layout and all valid bool
+        // bit patterns are valid u8 bit patterns.
+        u8::slice_contains_element(unsafe { bytes_of(hs) }, &(*needle as u8))
+    }
+
+    #[inline]
+    fn slice_contains_slice(hs: &[Self], needle: &[Self]) -> bool {
+        // SAFETY: bool and u8 have the same memory layout and all valid bool
+        // bit patterns are valid u8 bit patterns.
+        unsafe { u8::slice_contains_slice(bytes_of(hs), bytes_of(needle)) }
+    }
+}
+
+fn default_slice_contains_slice<T: PartialEq>(hs: &[T], needle: &[T]) -> bool {
+    super::pattern::NaiveSearcherState::new(hs.len())
+        .next_match(hs, needle)
+        .is_some()
+}
+
+
+/// SIMD search for short needles based on
+/// Wojciech Muła's "SIMD-friendly algorithms for substring searching"[0]
+///
+/// It skips ahead by the vector width on each iteration (rather than the needle length as two-way
+/// does) by probing the first and last byte of the needle for the whole vector width
+/// and only doing full needle comparisons when the vectorized probe indicated potential matches.
+///
+/// Since the x86_64 baseline only offers SSE2 we only use u8x16 here.
+/// If we ever ship std with for x86-64-v3 or adapt this for other platforms then wider vectors
+/// should be evaluated.
+///
+/// For haystacks smaller than vector-size + needle length it falls back to
+/// a naive O(n*m) search so this implementation should not be called on larger needles.
+///
+/// [0]: http://0x80.pl/articles/simd-strfind.html#sse-avx2
+#[cfg(all(target_arch = "x86_64", target_feature = "sse2"))]
+#[inline]
+fn simd_contains(haystack: &[u8], needle: &[u8]) -> Option<bool> {
+    debug_assert!(needle.len() > 1);
+
+    use crate::ops::BitAnd;
+    use crate::simd::mask8x16 as Mask;
+    use crate::simd::u8x16 as Block;
+    use crate::simd::{SimdPartialEq, ToBitMask};
+
+    let first_probe = needle[0];
+    let last_byte_offset = needle.len() - 1;
+
+    // the offset used for the 2nd vector
+    let second_probe_offset = if needle.len() == 2 {
+        // never bail out on len=2 needles because the probes will fully cover them and have
+        // no degenerate cases.
+        1
+    } else {
+        // try a few bytes in case first and last byte of the needle are the same
+        let Some(second_probe_offset) = (needle.len().saturating_sub(4)..needle.len()).rfind(|&idx| needle[idx] != first_probe) else {
+            // fall back to other search methods if we can't find any different bytes
+            // since we could otherwise hit some degenerate cases
+            return None;
+        };
+        second_probe_offset
+    };
+
+    // do a naive search if the haystack is too small to fit
+    if haystack.len() < Block::LANES + last_byte_offset {
+        return Some(haystack.windows(needle.len()).any(|c| c == needle));
+    }
+
+    let first_probe: Block = Block::splat(first_probe);
+    let second_probe: Block = Block::splat(needle[second_probe_offset]);
+    // first byte are already checked by the outer loop. to verify a match only the
+    // remainder has to be compared.
+    let trimmed_needle = &needle[1..];
+
+    // this #[cold] is load-bearing, benchmark before removing it...
+    let check_mask = #[cold]
+    |idx, mask: u16, skip: bool| -> bool {
+        if skip {
+            return false;
+        }
+
+        // and so is this. optimizations are weird.
+        let mut mask = mask;
+
+        while mask != 0 {
+            let trailing = mask.trailing_zeros();
+            let offset = idx + trailing as usize + 1;
+            // SAFETY: mask is between 0 and 15 trailing zeroes, we skip one additional byte that was already compared
+            // and then take trimmed_needle.len() bytes. This is within the bounds defined by the outer loop
+            unsafe {
+                let sub = haystack.get_unchecked(offset..).get_unchecked(..trimmed_needle.len());
+                if small_slice_eq(sub, trimmed_needle) {
+                    return true;
+                }
+            }
+            mask &= !(1 << trailing);
+        }
+        return false;
+    };
+
+    let test_chunk = |idx| -> u16 {
+        // SAFETY: this requires at least LANES bytes being readable at idx
+        // that is ensured by the loop ranges (see comments below)
+        let a: Block = unsafe { haystack.as_ptr().add(idx).cast::<Block>().read_unaligned() };
+        // SAFETY: this requires LANES + block_offset bytes being readable at idx
+        let b: Block = unsafe {
+            haystack.as_ptr().add(idx).add(second_probe_offset).cast::<Block>().read_unaligned()
+        };
+        let eq_first: Mask = a.simd_eq(first_probe);
+        let eq_last: Mask = b.simd_eq(second_probe);
+        let both = eq_first.bitand(eq_last);
+        let mask = both.to_bitmask();
+
+        return mask;
+    };
+
+    let mut i = 0;
+    let mut result = false;
+    // The loop condition must ensure that there's enough headroom to read LANE bytes,
+    // and not only at the current index but also at the index shifted by block_offset
+    const UNROLL: usize = 4;
+    while i + last_byte_offset + UNROLL * Block::LANES < haystack.len() && !result {
+        let mut masks = [0u16; UNROLL];
+        for j in 0..UNROLL {
+            masks[j] = test_chunk(i + j * Block::LANES);
+        }
+        for j in 0..UNROLL {
+            let mask = masks[j];
+            if mask != 0 {
+                result |= check_mask(i + j * Block::LANES, mask, result);
+            }
+        }
+        i += UNROLL * Block::LANES;
+    }
+    while i + last_byte_offset + Block::LANES < haystack.len() && !result {
+        let mask = test_chunk(i);
+        if mask != 0 {
+            result |= check_mask(i, mask, result);
+        }
+        i += Block::LANES;
+    }
+
+    // Process the tail that didn't fit into LANES-sized steps.
+    // This simply repeats the same procedure but as right-aligned chunk instead
+    // of a left-aligned one. The last byte must be exactly flush with the string end so
+    // we don't miss a single byte or read out of bounds.
+    let i = haystack.len() - last_byte_offset - Block::LANES;
+    let mask = test_chunk(i);
+    if mask != 0 {
+        result |= check_mask(i, mask, result);
+    }
+
+    Some(result)
+}
+
+/// Compares short slices for equality.
+///
+/// It avoids a call to libc's memcmp which is faster on long slices
+/// due to SIMD optimizations but it incurs a function call overhead.
+///
+/// # Safety
+///
+/// Both slices must have the same length.
+#[cfg(all(target_arch = "x86_64", target_feature = "sse2"))] // only called on x86
+#[inline]
+unsafe fn small_slice_eq(x: &[u8], y: &[u8]) -> bool {
+    debug_assert_eq!(x.len(), y.len());
+    // This function is adapted from
+    // https://github.com/BurntSushi/memchr/blob/8037d11b4357b0f07be2bb66dc2659d9cf28ad32/src/memmem/util.rs#L32
+
+    // If we don't have enough bytes to do 4-byte at a time loads, then
+    // fall back to the naive slow version.
+    //
+    // Potential alternative: We could do a copy_nonoverlapping combined with a mask instead
+    // of a loop. Benchmark it.
+    if x.len() < 4 {
+        for (&b1, &b2) in x.iter().zip(y) {
+            if b1 != b2 {
+                return false;
+            }
+        }
+        return true;
+    }
+    // When we have 4 or more bytes to compare, then proceed in chunks of 4 at
+    // a time using unaligned loads.
+    //
+    // Also, why do 4 byte loads instead of, say, 8 byte loads? The reason is
+    // that this particular version of memcmp is likely to be called with tiny
+    // needles. That means that if we do 8 byte loads, then a higher proportion
+    // of memcmp calls will use the slower variant above. With that said, this
+    // is a hypothesis and is only loosely supported by benchmarks. There's
+    // likely some improvement that could be made here. The main thing here
+    // though is to optimize for latency, not throughput.
+
+    // SAFETY: Via the conditional above, we know that both `px` and `py`
+    // have the same length, so `px < pxend` implies that `py < pyend`.
+    // Thus, derefencing both `px` and `py` in the loop below is safe.
+    //
+    // Moreover, we set `pxend` and `pyend` to be 4 bytes before the actual
+    // end of `px` and `py`. Thus, the final dereference outside of the
+    // loop is guaranteed to be valid. (The final comparison will overlap with
+    // the last comparison done in the loop for lengths that aren't multiples
+    // of four.)
+    //
+    // Finally, we needn't worry about alignment here, since we do unaligned
+    // loads.
+    unsafe {
+        let (mut px, mut py) = (x.as_ptr(), y.as_ptr());
+        let (pxend, pyend) = (px.add(x.len() - 4), py.add(y.len() - 4));
+        while px < pxend {
+            let vx = (px as *const u32).read_unaligned();
+            let vy = (py as *const u32).read_unaligned();
+            if vx != vy {
+                return false;
+            }
+            px = px.add(4);
+            py = py.add(4);
+        }
+        let vx = (pxend as *const u32).read_unaligned();
+        let vy = (pyend as *const u32).read_unaligned();
+        vx == vy
     }
 }