fast-md5 1.0.0

Optimized MD5 with hand-written x86_64 and aarch64 assembly cores, plus a portable Rust fallback.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168

use crate::{BLOCK_SIZE, STATE_WORDS};

const K: [u32; BLOCK_SIZE] = [
    0xd76aa478, 0xe8c7b756, 0x242070db, 0xc1bdceee, 0xf57c0faf, 0x4787c62a, 0xa8304613, 0xfd469501,
    0x698098d8, 0x8b44f7af, 0xffff5bb1, 0x895cd7be, 0x6b901122, 0xfd987193, 0xa679438e, 0x49b40821,
    0xf61e2562, 0xc040b340, 0x265e5a51, 0xe9b6c7aa, 0xd62f105d, 0x02441453, 0xd8a1e681, 0xe7d3fbc8,
    0x21e1cde6, 0xc33707d6, 0xf4d50d87, 0x455a14ed, 0xa9e3e905, 0xfcefa3f8, 0x676f02d9, 0x8d2a4c8a,
    0xfffa3942, 0x8771f681, 0x6d9d6122, 0xfde5380c, 0xa4beea44, 0x4bdecfa9, 0xf6bb4b60, 0xbebfbc70,
    0x289b7ec6, 0xeaa127fa, 0xd4ef3085, 0x04881d05, 0xd9d4d039, 0xe6db99e5, 0x1fa27cf8, 0xc4ac5665,
    0xf4292244, 0x432aff97, 0xab9423a7, 0xfc93a039, 0x655b59c3, 0x8f0ccc92, 0xffeff47d, 0x85845dd1,
    0x6fa87e4f, 0xfe2ce6e0, 0xa3014314, 0x4e0811a1, 0xf7537e82, 0xbd3af235, 0x2ad7d2bb, 0xeb86d391,
];

#[allow(clippy::inline_always)] // hot per-block call site; must inline into Md5::update
#[allow(clippy::too_many_lines)] // 64 unrolled MD5 rounds; splitting defeats the unroll
#[allow(clippy::many_single_char_names)] // a/b/c/d/m are standard MD5 register names
#[allow(clippy::cast_ptr_alignment)] // read_unaligned() is used immediately after the cast
#[inline(always)]
pub(crate) fn transform(state: &mut [u32; STATE_WORDS], block: &[u8; BLOCK_SIZE]) {
    // Per-round on-demand input loads via a closure.  Letting each round
    // pull its input word individually allows LLVM to interleave the LDR
    // with the prior round's arithmetic, hiding L1 load-use latency —
    // unlike a front-loaded `[u32; 16]` array which forces all 8 LDPs
    // to retire before round 1 starts.
    let m = block.as_ptr().cast::<u32>();
    let mi = |i: usize| -> u32 {
        // SAFETY: block is 64 bytes, i in 0..16
        u32::from_le(unsafe { m.add(i).read_unaligned() })
    };

    let (a0, b0, c0, d0) = (state[0], state[1], state[2], state[3]);
    let (mut a, mut b, mut c, mut d) = (a0, b0, c0, d0);

    // Per-round `asm!` for rotate: forces LLVM to materialize the accumulator
    // in a concrete 32-bit register at each round boundary, which guides
    // register allocation more tightly than giving LLVM full freedom via
    // `rotate_left()`.  The `pure, nomem, nostack` options tell LLVM the
    // result depends only on the input, so it can still schedule the surrounding
    // arithmetic freely within each round.
    macro_rules! ror {
        ($x:expr, $r:expr) => {{
            let mut v: u32 = $x;
            unsafe {
                core::arch::asm!(
                    "ror {v:w}, {v:w}, #{n}",
                    v = inout(reg) v,
                    n = const (32u32 - $r),
                    options(pure, nomem, nostack),
                );
            }
            v
        }};
    }

    // F(b,c,d) = D ^ (B & (C^D))
    macro_rules! f {
        ($a:ident, $b:ident, $c:ident, $d:ident, $m:expr, $k:expr, $r:expr) => {
            $a = $a
                .wrapping_add($d ^ ($b & ($c ^ $d)))
                .wrapping_add($m)
                .wrapping_add($k);
            $a = ror!($a, $r).wrapping_add($b);
        };
    }
    // G(b,c,d) = (~D & C) + (D & B)   — maps to BIC + AND on AArch64
    macro_rules! g {
        ($a:ident, $b:ident, $c:ident, $d:ident, $m:expr, $k:expr, $r:expr) => {
            $a = $a
                .wrapping_add((!$d & $c).wrapping_add($d & $b))
                .wrapping_add($m)
                .wrapping_add($k);
            $a = ror!($a, $r).wrapping_add($b);
        };
    }
    // H(b,c,d) = B ^ C ^ D
    macro_rules! h {
        ($a:ident, $b:ident, $c:ident, $d:ident, $m:expr, $k:expr, $r:expr) => {
            $a = $a
                .wrapping_add($b ^ $c ^ $d)
                .wrapping_add($m)
                .wrapping_add($k);
            $a = ror!($a, $r).wrapping_add($b);
        };
    }
    // I(b,c,d) = C ^ (B | ~D)   — maps to ORN on AArch64
    macro_rules! i {
        ($a:ident, $b:ident, $c:ident, $d:ident, $m:expr, $k:expr, $r:expr) => {
            $a = $a
                .wrapping_add($c ^ ($b | !$d))
                .wrapping_add($m)
                .wrapping_add($k);
            $a = ror!($a, $r).wrapping_add($b);
        };
    }

    f!(a, b, c, d, mi(0), K[0], 7);
    f!(d, a, b, c, mi(1), K[1], 12);
    f!(c, d, a, b, mi(2), K[2], 17);
    f!(b, c, d, a, mi(3), K[3], 22);
    f!(a, b, c, d, mi(4), K[4], 7);
    f!(d, a, b, c, mi(5), K[5], 12);
    f!(c, d, a, b, mi(6), K[6], 17);
    f!(b, c, d, a, mi(7), K[7], 22);
    f!(a, b, c, d, mi(8), K[8], 7);
    f!(d, a, b, c, mi(9), K[9], 12);
    f!(c, d, a, b, mi(10), K[10], 17);
    f!(b, c, d, a, mi(11), K[11], 22);
    f!(a, b, c, d, mi(12), K[12], 7);
    f!(d, a, b, c, mi(13), K[13], 12);
    f!(c, d, a, b, mi(14), K[14], 17);
    f!(b, c, d, a, mi(15), K[15], 22);

    g!(a, b, c, d, mi(1), K[16], 5);
    g!(d, a, b, c, mi(6), K[17], 9);
    g!(c, d, a, b, mi(11), K[18], 14);
    g!(b, c, d, a, mi(0), K[19], 20);
    g!(a, b, c, d, mi(5), K[20], 5);
    g!(d, a, b, c, mi(10), K[21], 9);
    g!(c, d, a, b, mi(15), K[22], 14);
    g!(b, c, d, a, mi(4), K[23], 20);
    g!(a, b, c, d, mi(9), K[24], 5);
    g!(d, a, b, c, mi(14), K[25], 9);
    g!(c, d, a, b, mi(3), K[26], 14);
    g!(b, c, d, a, mi(8), K[27], 20);
    g!(a, b, c, d, mi(13), K[28], 5);
    g!(d, a, b, c, mi(2), K[29], 9);
    g!(c, d, a, b, mi(7), K[30], 14);
    g!(b, c, d, a, mi(12), K[31], 20);

    h!(a, b, c, d, mi(5), K[32], 4);
    h!(d, a, b, c, mi(8), K[33], 11);
    h!(c, d, a, b, mi(11), K[34], 16);
    h!(b, c, d, a, mi(14), K[35], 23);
    h!(a, b, c, d, mi(1), K[36], 4);
    h!(d, a, b, c, mi(4), K[37], 11);
    h!(c, d, a, b, mi(7), K[38], 16);
    h!(b, c, d, a, mi(10), K[39], 23);
    h!(a, b, c, d, mi(13), K[40], 4);
    h!(d, a, b, c, mi(0), K[41], 11);
    h!(c, d, a, b, mi(3), K[42], 16);
    h!(b, c, d, a, mi(6), K[43], 23);
    h!(a, b, c, d, mi(9), K[44], 4);
    h!(d, a, b, c, mi(12), K[45], 11);
    h!(c, d, a, b, mi(15), K[46], 16);
    h!(b, c, d, a, mi(2), K[47], 23);

    i!(a, b, c, d, mi(0), K[48], 6);
    i!(d, a, b, c, mi(7), K[49], 10);
    i!(c, d, a, b, mi(14), K[50], 15);
    i!(b, c, d, a, mi(5), K[51], 21);
    i!(a, b, c, d, mi(12), K[52], 6);
    i!(d, a, b, c, mi(3), K[53], 10);
    i!(c, d, a, b, mi(10), K[54], 15);
    i!(b, c, d, a, mi(1), K[55], 21);
    i!(a, b, c, d, mi(8), K[56], 6);
    i!(d, a, b, c, mi(15), K[57], 10);
    i!(c, d, a, b, mi(6), K[58], 15);
    i!(b, c, d, a, mi(13), K[59], 21);
    i!(a, b, c, d, mi(4), K[60], 6);
    i!(d, a, b, c, mi(11), K[61], 10);
    i!(c, d, a, b, mi(2), K[62], 15);
    i!(b, c, d, a, mi(9), K[63], 21);

    state[0] = a0.wrapping_add(a);
    state[1] = b0.wrapping_add(b);
    state[2] = c0.wrapping_add(c);
    state[3] = d0.wrapping_add(d);
}