| //! SIMD byte scanning logic. |
| //! |
| //! This module provides functions that allow walking through byteslices, calling |
| //! provided callback functions on special bytes and their indices using SIMD. |
| //! The byteset is defined in `compute_lookup`. |
| //! |
| //! The idea is to load in a chunk of 16 bytes and perform a lookup into a set of |
| //! bytes on all the bytes in this chunk simultaneously. We produce a 16 bit bitmask |
| //! from this and call the callback on every index corresponding to a 1 in this mask |
| //! before moving on to the next chunk. This allows us to move quickly when there |
| //! are no or few matches. |
| //! |
| //! The table lookup is inspired by this [great overview]. However, since all of the |
| //! bytes we're interested in are ASCII, we don't quite need the full generality of |
| //! the universal algorithm and are hence able to skip a few instructions. |
| //! |
| //! [great overview]: http://0x80.pl/articles/simd-byte-lookup.html |
| |
| use crate::parse::LoopInstruction; |
| use core::arch::x86_64::*; |
| |
| const VECTOR_SIZE: usize = std::mem::size_of::<__m128i>(); |
| |
| /// Generates a lookup table containing the bitmaps for our |
| /// special marker bytes. This is effectively a 128 element 2d bitvector, |
| /// that can be indexed by a four bit row index (the lower nibble) |
| /// and a three bit column index (upper nibble). |
| const fn compute_lookup() -> [u8; 16] { |
| let mut lookup = [0u8; 16]; |
| lookup[(b'\n' & 0x0f) as usize] |= 1 << (b'\n' >> 4); |
| lookup[(b'\r' & 0x0f) as usize] |= 1 << (b'\r' >> 4); |
| lookup[(b'*' & 0x0f) as usize] |= 1 << (b'*' >> 4); |
| lookup[(b'_' & 0x0f) as usize] |= 1 << (b'_' >> 4); |
| lookup[(b'~' & 0x0f) as usize] |= 1 << (b'~' >> 4); |
| lookup[(b'|' & 0x0f) as usize] |= 1 << (b'|' >> 4); |
| lookup[(b'&' & 0x0f) as usize] |= 1 << (b'&' >> 4); |
| lookup[(b'\\' & 0x0f) as usize] |= 1 << (b'\\' >> 4); |
| lookup[(b'[' & 0x0f) as usize] |= 1 << (b'[' >> 4); |
| lookup[(b']' & 0x0f) as usize] |= 1 << (b']' >> 4); |
| lookup[(b'<' & 0x0f) as usize] |= 1 << (b'<' >> 4); |
| lookup[(b'!' & 0x0f) as usize] |= 1 << (b'!' >> 4); |
| lookup[(b'`' & 0x0f) as usize] |= 1 << (b'`' >> 4); |
| lookup |
| } |
| |
| /// Computes a bit mask for the given byteslice starting from the given index, |
| /// where the 16 least significant bits indicate (by value of 1) whether or not |
| /// there is a special character at that byte position. The least significant bit |
| /// corresponds to `bytes[ix]` and the most significant bit corresponds to |
| /// `bytes[ix + 15]`. |
| /// It is only safe to call this function when `bytes.len() >= ix + VECTOR_SIZE`. |
| #[target_feature(enable = "ssse3")] |
| #[inline] |
| unsafe fn compute_mask(bytes: &[u8], ix: usize) -> i32 { |
| debug_assert!(bytes.len() >= ix + VECTOR_SIZE); |
| |
| let lookup = compute_lookup(); |
| let bitmap = _mm_loadu_si128(lookup.as_ptr() as *const __m128i); |
| // Small lookup table to compute single bit bitshifts |
| // for 16 bytes at once. |
| let bitmask_lookup = |
| _mm_setr_epi8(1, 2, 4, 8, 16, 32, 64, -128, -1, -1, -1, -1, -1, -1, -1, -1); |
| |
| // Load input from memory. |
| let raw_ptr = bytes.as_ptr().add(ix) as *const __m128i; |
| let input = _mm_loadu_si128(raw_ptr); |
| // Compute the bitmap using the bottom nibble as an index |
| // into the lookup table. Note that non-ascii bytes will have |
| // their most significant bit set and will map to lookup[0]. |
| let bitset = _mm_shuffle_epi8(bitmap, input); |
| // Compute the high nibbles of the input using a 16-bit rightshift of four |
| // and a mask to prevent most-significant bit issues. |
| let higher_nibbles = _mm_and_si128(_mm_srli_epi16(input, 4), _mm_set1_epi8(0x0f)); |
| // Create a bitmask for the bitmap by perform a left shift of the value |
| // of the higher nibble. Bytes with their most significant set are mapped |
| // to -1 (all ones). |
| let bitmask = _mm_shuffle_epi8(bitmask_lookup, higher_nibbles); |
| // Test the bit of the bitmap by AND'ing the bitmap and the mask together. |
| let tmp = _mm_and_si128(bitset, bitmask); |
| // Check whether the result was not null. NEQ is not a SIMD intrinsic, |
| // but comparing to the bitmask is logically equivalent. This also prevents us |
| // from matching any non-ASCII bytes since none of the bitmaps were all ones |
| // (-1). |
| let result = _mm_cmpeq_epi8(tmp, bitmask); |
| |
| // Return the resulting bitmask. |
| _mm_movemask_epi8(result) |
| } |
| |
| /// Calls callback on byte indices and their value. |
| /// Breaks when callback returns LoopInstruction::BreakAtWith(ix, val). And skips the |
| /// number of bytes in callback return value otherwise. |
| /// Returns the final index and a possible break value. |
| pub(crate) fn iterate_special_bytes<F, T>( |
| bytes: &[u8], |
| ix: usize, |
| callback: F, |
| ) -> (usize, Option<T>) |
| where |
| F: FnMut(usize, u8) -> LoopInstruction<Option<T>>, |
| { |
| if is_x86_feature_detected!("ssse3") && bytes.len() >= VECTOR_SIZE { |
| unsafe { simd_iterate_special_bytes(bytes, ix, callback) } |
| } else { |
| crate::parse::scalar_iterate_special_bytes(bytes, ix, callback) |
| } |
| } |
| |
| /// Calls the callback function for every 1 in the given bitmask with |
| /// the index `offset + ix`, where `ix` is the position of the 1 in the mask. |
| /// Returns `Ok(ix)` to continue from index `ix`, `Err((end_ix, opt_val)` to break with |
| /// final index `end_ix` and optional value `opt_val`. |
| unsafe fn process_mask<F, T>( |
| mut mask: i32, |
| bytes: &[u8], |
| mut offset: usize, |
| callback: &mut F, |
| ) -> Result<usize, (usize, Option<T>)> |
| where |
| F: FnMut(usize, u8) -> LoopInstruction<Option<T>>, |
| { |
| while mask != 0 { |
| let mask_ix = mask.trailing_zeros() as usize; |
| offset += mask_ix; |
| match callback(offset, *bytes.get_unchecked(offset)) { |
| LoopInstruction::ContinueAndSkip(skip) => { |
| offset += skip + 1; |
| mask >>= skip + 1 + mask_ix; |
| } |
| LoopInstruction::BreakAtWith(ix, val) => return Err((ix, val)), |
| } |
| } |
| Ok(offset) |
| } |
| |
| #[target_feature(enable = "ssse3")] |
| /// Important: only call this function when `bytes.len() >= 16`. Doing |
| /// so otherwise may exhibit undefined behaviour. |
| unsafe fn simd_iterate_special_bytes<F, T>( |
| bytes: &[u8], |
| mut ix: usize, |
| mut callback: F, |
| ) -> (usize, Option<T>) |
| where |
| F: FnMut(usize, u8) -> LoopInstruction<Option<T>>, |
| { |
| debug_assert!(bytes.len() >= VECTOR_SIZE); |
| let upperbound = bytes.len() - VECTOR_SIZE; |
| |
| while ix < upperbound { |
| let mask = compute_mask(bytes, ix); |
| let block_start = ix; |
| ix = match process_mask(mask, bytes, ix, &mut callback) { |
| Ok(ix) => std::cmp::max(ix, VECTOR_SIZE + block_start), |
| Err((end_ix, val)) => return (end_ix, val), |
| }; |
| } |
| |
| if bytes.len() > ix { |
| // shift off the bytes at start we have already scanned |
| let mask = compute_mask(bytes, upperbound) >> ix - upperbound; |
| if let Err((end_ix, val)) = process_mask(mask, bytes, ix, &mut callback) { |
| return (end_ix, val); |
| } |
| } |
| |
| (bytes.len(), None) |
| } |
| |
| #[cfg(test)] |
| mod simd_test { |
| use super::{iterate_special_bytes, LoopInstruction}; |
| |
| fn check_expected_indices(bytes: &[u8], expected: &[usize], skip: usize) { |
| let mut indices = vec![]; |
| |
| iterate_special_bytes::<_, i32>(bytes, 0, |ix, _byte_ty| { |
| indices.push(ix); |
| LoopInstruction::ContinueAndSkip(skip) |
| }); |
| |
| assert_eq!(&indices[..], expected); |
| } |
| |
| #[test] |
| fn simple_no_match() { |
| check_expected_indices("abcdef0123456789".as_bytes(), &[], 0); |
| } |
| |
| #[test] |
| fn simple_match() { |
| check_expected_indices("*bcd&f0123456789".as_bytes(), &[0, 4], 0); |
| } |
| |
| #[test] |
| fn single_open_fish() { |
| check_expected_indices("<".as_bytes(), &[0], 0); |
| } |
| |
| #[test] |
| fn long_match() { |
| check_expected_indices("0123456789abcde~*bcd&f0".as_bytes(), &[15, 16, 20], 0); |
| } |
| |
| #[test] |
| fn border_skip() { |
| check_expected_indices("0123456789abcde~~~~d&f0".as_bytes(), &[15, 20], 3); |
| } |
| |
| #[test] |
| fn exhaustive_search() { |
| let chars = [ |
| b'\n', b'\r', b'*', b'_', b'~', b'|', b'&', b'\\', b'[', b']', b'<', b'!', b'`', |
| ]; |
| |
| for &c in &chars { |
| for i in 0u8..=255 { |
| if !chars.contains(&i) { |
| // full match |
| let mut buf = [i; 18]; |
| buf[3] = c; |
| buf[6] = c; |
| |
| check_expected_indices(&buf[..], &[3, 6], 0); |
| } |
| } |
| } |
| } |
| } |