| 1 | use crate::iter::Bytes; |
| 2 | |
| 3 | #[inline ] |
| 4 | #[target_feature (enable = "avx2" )] |
| 5 | pub unsafe fn match_uri_vectored(bytes: &mut Bytes) { |
| 6 | while bytes.as_ref().len() >= 32 { |
| 7 | |
| 8 | let advance: usize = match_url_char_32_avx(buf:bytes.as_ref()); |
| 9 | |
| 10 | bytes.advance(advance); |
| 11 | |
| 12 | if advance != 32 { |
| 13 | return; |
| 14 | } |
| 15 | } |
| 16 | // NOTE: use SWAR for <32B, more efficient than falling back to SSE4.2 |
| 17 | super::swar::match_uri_vectored(bytes) |
| 18 | } |
| 19 | |
| 20 | #[inline (always)] |
| 21 | #[allow (non_snake_case, overflowing_literals)] |
| 22 | #[allow (unused)] |
| 23 | unsafe fn match_url_char_32_avx(buf: &[u8]) -> usize { |
| 24 | // NOTE: This check might be not necessary since this function is only used in |
| 25 | // `match_uri_vectored` where buffer overflow is taken care of. |
| 26 | debug_assert!(buf.len() >= 32); |
| 27 | |
| 28 | #[cfg (target_arch = "x86" )] |
| 29 | use core::arch::x86::*; |
| 30 | #[cfg (target_arch = "x86_64" )] |
| 31 | use core::arch::x86_64::*; |
| 32 | |
| 33 | // pointer to buffer |
| 34 | let ptr = buf.as_ptr(); |
| 35 | |
| 36 | // %x21-%x7e %x80-%xff |
| 37 | // |
| 38 | // Character ranges allowed by this function, can also be interpreted as: |
| 39 | // 33 =< (x != 127) =< 255 |
| 40 | // |
| 41 | // Create a vector full of DEL (0x7f) characters. |
| 42 | let DEL: __m256i = _mm256_set1_epi8(0x7f); |
| 43 | // Create a vector full of exclamation mark (!) (0x21) characters. |
| 44 | // Used as lower threshold, characters in URLs cannot be smaller than this. |
| 45 | let LOW: __m256i = _mm256_set1_epi8(0x21); |
| 46 | |
| 47 | // Load a chunk of 32 bytes from `ptr` as a vector. |
| 48 | // We can check 32 bytes in parallel at most with AVX2 since |
| 49 | // YMM registers can only have 256 bits most. |
| 50 | let dat = _mm256_lddqu_si256(ptr as *const _); |
| 51 | |
| 52 | // unsigned comparison dat >= LOW |
| 53 | // |
| 54 | // `_mm256_max_epu8` creates a new vector by comparing vectors `dat` and `LOW` |
| 55 | // and picks the max. values from each for all indices. |
| 56 | // So if a byte in `dat` is <= 32, it'll be represented as 33 |
| 57 | // which is the smallest valid character. |
| 58 | // |
| 59 | // Then, we compare the new vector with `dat` for equality. |
| 60 | // |
| 61 | // `_mm256_cmpeq_epi8` returns a new vector where; |
| 62 | // * matching bytes are set to 0xFF (all bits set), |
| 63 | // * nonmatching bytes are set to 0 (no bits set). |
| 64 | let low = _mm256_cmpeq_epi8(_mm256_max_epu8(dat, LOW), dat); |
| 65 | // Similar to what we did before, but now invalid characters are set to 0xFF. |
| 66 | let del = _mm256_cmpeq_epi8(dat, DEL); |
| 67 | |
| 68 | // We glue the both comparisons via `_mm256_andnot_si256`. |
| 69 | // |
| 70 | // Since the representation of truthiness differ in these comparisons, |
| 71 | // we are in need of bitwise NOT to convert valid characters of `del`. |
| 72 | let bit = _mm256_andnot_si256(del, low); |
| 73 | // This creates a bitmask from the most significant bit of each byte. |
| 74 | // Simply, we're converting a vector value to scalar value here. |
| 75 | let res = _mm256_movemask_epi8(bit) as u32; |
| 76 | |
| 77 | // Count trailing zeros to find the first encountered invalid character. |
| 78 | // Bitwise NOT is required once again to flip truthiness. |
| 79 | // TODO: use .trailing_ones() once MSRV >= 1.46 |
| 80 | (!res).trailing_zeros() as usize |
| 81 | } |
| 82 | |
| 83 | #[target_feature (enable = "avx2" )] |
| 84 | pub unsafe fn match_header_value_vectored(bytes: &mut Bytes) { |
| 85 | while bytes.as_ref().len() >= 32 { |
| 86 | let advance: usize = match_header_value_char_32_avx(buf:bytes.as_ref()); |
| 87 | bytes.advance(advance); |
| 88 | |
| 89 | if advance != 32 { |
| 90 | return; |
| 91 | } |
| 92 | } |
| 93 | // NOTE: use SWAR for <32B, more efficient than falling back to SSE4.2 |
| 94 | super::swar::match_header_value_vectored(bytes) |
| 95 | } |
| 96 | |
| 97 | #[inline (always)] |
| 98 | #[allow (non_snake_case)] |
| 99 | #[allow (unused)] |
| 100 | unsafe fn match_header_value_char_32_avx(buf: &[u8]) -> usize { |
| 101 | debug_assert!(buf.len() >= 32); |
| 102 | |
| 103 | #[cfg (target_arch = "x86" )] |
| 104 | use core::arch::x86::*; |
| 105 | #[cfg (target_arch = "x86_64" )] |
| 106 | use core::arch::x86_64::*; |
| 107 | |
| 108 | let ptr = buf.as_ptr(); |
| 109 | |
| 110 | // %x09 %x20-%x7e %x80-%xff |
| 111 | // Create a vector full of horizontal tab (\t) (0x09) characters. |
| 112 | let TAB: __m256i = _mm256_set1_epi8(0x09); |
| 113 | // Create a vector full of DEL (0x7f) characters. |
| 114 | let DEL: __m256i = _mm256_set1_epi8(0x7f); |
| 115 | // Create a vector full of space (0x20) characters. |
| 116 | let LOW: __m256i = _mm256_set1_epi8(0x20); |
| 117 | |
| 118 | // Load a chunk of 32 bytes from `ptr` as a vector. |
| 119 | let dat = _mm256_lddqu_si256(ptr as *const _); |
| 120 | |
| 121 | // unsigned comparison dat >= LOW |
| 122 | // |
| 123 | // Same as what we do in `match_url_char_32_avx`. |
| 124 | // This time the lower threshold is set to space character though. |
| 125 | let low = _mm256_cmpeq_epi8(_mm256_max_epu8(dat, LOW), dat); |
| 126 | // Check if `dat` includes `TAB` characters. |
| 127 | let tab = _mm256_cmpeq_epi8(dat, TAB); |
| 128 | // Check if `dat` includes `DEL` characters. |
| 129 | let del = _mm256_cmpeq_epi8(dat, DEL); |
| 130 | |
| 131 | // Combine all comparisons together, notice that we're also using OR |
| 132 | // to connect `low` and `tab` but flip bits of `del`. |
| 133 | // |
| 134 | // In the end, this is simply: |
| 135 | // ~del & (low | tab) |
| 136 | let bit = _mm256_andnot_si256(del, _mm256_or_si256(low, tab)); |
| 137 | // This creates a bitmask from the most significant bit of each byte. |
| 138 | // Creates a scalar value from vector value. |
| 139 | let res = _mm256_movemask_epi8(bit) as u32; |
| 140 | |
| 141 | // Count trailing zeros to find the first encountered invalid character. |
| 142 | // Bitwise NOT is required once again to flip truthiness. |
| 143 | // TODO: use .trailing_ones() once MSRV >= 1.46 |
| 144 | (!res).trailing_zeros() as usize |
| 145 | } |
| 146 | |
| 147 | #[test ] |
| 148 | fn avx2_code_matches_uri_chars_table() { |
| 149 | if !is_x86_feature_detected!("avx2" ) { |
| 150 | return; |
| 151 | } |
| 152 | |
| 153 | #[allow (clippy::undocumented_unsafe_blocks)] |
| 154 | unsafe { |
| 155 | assert!(byte_is_allowed(b'_' , match_uri_vectored)); |
| 156 | |
| 157 | for (b, allowed) in crate::URI_MAP.iter().cloned().enumerate() { |
| 158 | assert_eq!( |
| 159 | byte_is_allowed(b as u8, match_uri_vectored), allowed, |
| 160 | "byte_is_allowed({:?}) should be {:?}" , b, allowed, |
| 161 | ); |
| 162 | } |
| 163 | } |
| 164 | } |
| 165 | |
| 166 | #[test ] |
| 167 | fn avx2_code_matches_header_value_chars_table() { |
| 168 | if !is_x86_feature_detected!("avx2" ) { |
| 169 | return; |
| 170 | } |
| 171 | |
| 172 | #[allow (clippy::undocumented_unsafe_blocks)] |
| 173 | unsafe { |
| 174 | assert!(byte_is_allowed(b'_' , match_header_value_vectored)); |
| 175 | |
| 176 | for (b, allowed) in crate::HEADER_VALUE_MAP.iter().cloned().enumerate() { |
| 177 | assert_eq!( |
| 178 | byte_is_allowed(b as u8, match_header_value_vectored), allowed, |
| 179 | "byte_is_allowed({:?}) should be {:?}" , b, allowed, |
| 180 | ); |
| 181 | } |
| 182 | } |
| 183 | } |
| 184 | |
| 185 | #[cfg (test)] |
| 186 | unsafe fn byte_is_allowed(byte: u8, f: unsafe fn(bytes: &mut Bytes<'_>)) -> bool { |
| 187 | let slice = [ |
| 188 | b'_' , b'_' , b'_' , b'_' , |
| 189 | b'_' , b'_' , b'_' , b'_' , |
| 190 | b'_' , b'_' , b'_' , b'_' , |
| 191 | b'_' , b'_' , b'_' , b'_' , |
| 192 | b'_' , b'_' , b'_' , b'_' , |
| 193 | b'_' , b'_' , b'_' , b'_' , |
| 194 | b'_' , b'_' , byte, b'_' , |
| 195 | b'_' , b'_' , b'_' , b'_' , |
| 196 | ]; |
| 197 | let mut bytes = Bytes::new(&slice); |
| 198 | |
| 199 | f(&mut bytes); |
| 200 | |
| 201 | match bytes.pos() { |
| 202 | 32 => true, |
| 203 | 26 => false, |
| 204 | _ => unreachable!(), |
| 205 | } |
| 206 | } |
| 207 | |