1 | use crate::iter::Bytes; |
2 | |
3 | pub unsafe fn parse_uri_batch_16(bytes: &mut Bytes) { |
4 | while bytes.as_ref().len() >= 16 { |
5 | let advance: usize = match_url_char_16_sse(buf:bytes.as_ref()); |
6 | bytes.advance(advance); |
7 | |
8 | if advance != 16 { |
9 | break; |
10 | } |
11 | } |
12 | } |
13 | |
14 | #[target_feature (enable = "sse4.2" )] |
15 | #[allow (non_snake_case, overflowing_literals)] |
16 | unsafe fn match_url_char_16_sse(buf: &[u8]) -> usize { |
17 | debug_assert!(buf.len() >= 16); |
18 | |
19 | #[cfg (target_arch = "x86" )] |
20 | use core::arch::x86::*; |
21 | #[cfg (target_arch = "x86_64" )] |
22 | use core::arch::x86_64::*; |
23 | |
24 | let ptr = buf.as_ptr(); |
25 | |
26 | let LSH: __m128i = _mm_set1_epi8(0x0f); |
27 | |
28 | // The first 0xf8 corresponds to the 8 first rows of the first column |
29 | // of URI_MAP in the crate's root, with the first row corresponding to bit 0 |
30 | // and the 8th row corresponding to bit 7. |
31 | // The 8 first rows give 0 0 0 1 1 1 1 1, which is 0xf8 (with least |
32 | // significant digit on the left). |
33 | // |
34 | // Another example just to drive the point home: in column 15, '>' is |
35 | // rejected, so the values are 0 0 1 0 1 1 1 1, which gives us 0xf4. |
36 | // |
37 | // Thanks to Vlad Krasnov for explaining this stuff to us mere mortals in |
38 | // a GitHub comment! |
39 | // |
40 | // https://github.com/seanmonstar/httparse/pull/89#issuecomment-807039219 |
41 | |
42 | let URI: __m128i = _mm_setr_epi8( |
43 | 0xf8, 0xfc, 0xfc, 0xfc, 0xfc, 0xfc, 0xfc, 0xfc, |
44 | 0xfc, 0xfc, 0xfc, 0xfc, 0xf4, 0xfc, 0xf4, 0x7c, |
45 | ); |
46 | let ARF: __m128i = _mm_setr_epi8( |
47 | 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, |
48 | 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, |
49 | ); |
50 | |
51 | let data = _mm_lddqu_si128(ptr as *const _); |
52 | let rbms = _mm_shuffle_epi8(URI, data); |
53 | let cols = _mm_and_si128(LSH, _mm_srli_epi16(data, 4)); |
54 | let bits = _mm_and_si128(_mm_shuffle_epi8(ARF, cols), rbms); |
55 | |
56 | let v = _mm_cmpeq_epi8(bits, _mm_setzero_si128()); |
57 | let r = 0xffff_0000 | _mm_movemask_epi8(v) as u32; |
58 | |
59 | _tzcnt_u32(r) as usize |
60 | } |
61 | |
62 | pub unsafe fn match_header_value_batch_16(bytes: &mut Bytes) { |
63 | while bytes.as_ref().len() >= 16 { |
64 | let advance: usize = match_header_value_char_16_sse(buf:bytes.as_ref()); |
65 | bytes.advance(advance); |
66 | |
67 | if advance != 16 { |
68 | break; |
69 | } |
70 | } |
71 | } |
72 | |
73 | #[target_feature (enable = "sse4.2" )] |
74 | #[allow (non_snake_case)] |
75 | unsafe fn match_header_value_char_16_sse(buf: &[u8]) -> usize { |
76 | debug_assert!(buf.len() >= 16); |
77 | |
78 | #[cfg (target_arch = "x86" )] |
79 | use core::arch::x86::*; |
80 | #[cfg (target_arch = "x86_64" )] |
81 | use core::arch::x86_64::*; |
82 | |
83 | let ptr = buf.as_ptr(); |
84 | |
85 | // %x09 %x20-%x7e %x80-%xff |
86 | let TAB: __m128i = _mm_set1_epi8(0x09); |
87 | let DEL: __m128i = _mm_set1_epi8(0x7f); |
88 | let LOW: __m128i = _mm_set1_epi8(0x20); |
89 | |
90 | let dat = _mm_lddqu_si128(ptr as *const _); |
91 | // unsigned comparison dat >= LOW |
92 | let low = _mm_cmpeq_epi8(_mm_max_epu8(dat, LOW), dat); |
93 | let tab = _mm_cmpeq_epi8(dat, TAB); |
94 | let del = _mm_cmpeq_epi8(dat, DEL); |
95 | let bit = _mm_andnot_si128(del, _mm_or_si128(low, tab)); |
96 | let rev = _mm_cmpeq_epi8(bit, _mm_setzero_si128()); |
97 | let res = 0xffff_0000 | _mm_movemask_epi8(rev) as u32; |
98 | |
99 | _tzcnt_u32(res) as usize |
100 | } |
101 | |
102 | #[test ] |
103 | fn sse_code_matches_uri_chars_table() { |
104 | match super::detect() { |
105 | super::SSE_42 | super::AVX_2_AND_SSE_42 => {}, |
106 | _ => return, |
107 | } |
108 | |
109 | unsafe { |
110 | assert!(byte_is_allowed(b'_' , parse_uri_batch_16)); |
111 | |
112 | for (b: usize, allowed: bool) in crate::URI_MAP.iter().cloned().enumerate() { |
113 | assert_eq!( |
114 | byte_is_allowed(b as u8, parse_uri_batch_16), allowed, |
115 | "byte_is_allowed( {:?}) should be {:?}" , b, allowed, |
116 | ); |
117 | } |
118 | } |
119 | } |
120 | |
121 | #[test ] |
122 | fn sse_code_matches_header_value_chars_table() { |
123 | match super::detect() { |
124 | super::SSE_42 | super::AVX_2_AND_SSE_42 => {}, |
125 | _ => return, |
126 | } |
127 | |
128 | unsafe { |
129 | assert!(byte_is_allowed(b'_' , match_header_value_batch_16)); |
130 | |
131 | for (b: usize, allowed: bool) in crate::HEADER_VALUE_MAP.iter().cloned().enumerate() { |
132 | assert_eq!( |
133 | byte_is_allowed(b as u8, match_header_value_batch_16), allowed, |
134 | "byte_is_allowed( {:?}) should be {:?}" , b, allowed, |
135 | ); |
136 | } |
137 | } |
138 | } |
139 | |
140 | #[cfg (test)] |
141 | unsafe fn byte_is_allowed(byte: u8, f: unsafe fn(bytes: &mut Bytes<'_>)) -> bool { |
142 | let slice = [ |
143 | b'_' , b'_' , b'_' , b'_' , |
144 | b'_' , b'_' , b'_' , b'_' , |
145 | b'_' , b'_' , byte, b'_' , |
146 | b'_' , b'_' , b'_' , b'_' , |
147 | ]; |
148 | let mut bytes = Bytes::new(&slice); |
149 | |
150 | f(&mut bytes); |
151 | |
152 | match bytes.pos() { |
153 | 16 => true, |
154 | 10 => false, |
155 | _ => unreachable!(), |
156 | } |
157 | } |
158 | |