philbin 1.0.1

A pure Rust AEGIS library with SIMD and runtime CPU detection
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
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217

use crate::{
  aligned_buffer::{AlignedBufHolder, AlignedBufRouter},
  arch::Simd,
  base::{aegis::Aegis, block::Block},
  careful::Nonce256,
  easy::{AuthTag, AuthTag128, AuthTag256, Key256},
  utils::num_bits,
};
use zerocopy::{FromZeros, transmute};

/// The state for the AEGIS-256 algorithm.
///
/// `StateBlock` represents the algo block storage and low-level block
/// manipulation functions, making the `Aegis256` type generic across
/// architectures, SIMD sizes, CPU features and even AEGIS-256 variants.
///
/// AEGIS-256X2 and AEGIS-256X4 also use this type; only the `StateBlock` is
/// different!
///
/// See [`Block`].
pub struct Aegis256<S: Simd, StateBlock, const OUTPUT_RATE_BYTES: usize> {
  s0: StateBlock,
  s1: StateBlock,
  s2: StateBlock,
  s3: StateBlock,
  s4: StateBlock,
  s5: StateBlock,
  simd: S,
}

// NOTE: There is no need for an assert (static or otherwise) checking that
// size_of::<StateBlock>() == OUTPUT_RATE_BYTES because we only impl
// Block::SelfArray with *exact* array sizes. In other words, we already get a
// compile-time error for all values of OUTPUT_RATE_BYTES other than the
// supported ones.
impl<S, StateBlock, const OUTPUT_RATE_BYTES: usize>
  Aegis256<S, StateBlock, OUTPUT_RATE_BYTES>
where
  S: Simd,
  StateBlock: Block<SelfArray = [u8; OUTPUT_RATE_BYTES], Simd = S>,
{
  /// Updates the internal state with the provided blocks.
  #[inline(always)]
  fn update(&mut self, m: StateBlock) {
    let orig_s5 = self.s5;
    self.s5 = StateBlock::aes_encrypt_round(self.s4, self.s5);
    self.s4 = StateBlock::aes_encrypt_round(self.s3, self.s4);
    self.s3 = StateBlock::aes_encrypt_round(self.s2, self.s3);
    self.s2 = StateBlock::aes_encrypt_round(self.s1, self.s2);
    self.s1 = StateBlock::aes_encrypt_round(self.s0, self.s1);
    self.s0 = StateBlock::aes_encrypt_round(orig_s5, self.s0);

    // PERF: Pulling this out improves perf for 256x2. Callgrind metrics get
    // worse for 256x4, but runtime perf doesn't change.
    self.s0 = self.s0 ^ m;
  }
}

impl<S, StateBlock, const OUTPUT_RATE_BYTES: usize> Aegis<OUTPUT_RATE_BYTES>
  for Aegis256<S, StateBlock, OUTPUT_RATE_BYTES>
where
  S: Simd,
  StateBlock: Block<SelfArray = [u8; OUTPUT_RATE_BYTES], Simd = S>,
  AlignedBufHolder: AlignedBufRouter<OUTPUT_RATE_BYTES>,
{
  type AlignedBuf =
    <AlignedBufHolder as AlignedBufRouter<OUTPUT_RATE_BYTES>>::AlignedBuf;
  type Simd = S;
  type Key = Key256;
  type Nonce = Nonce256;

  #[inline(always)]
  fn init(simd: S, key: &Key256, nonce: Nonce256) -> Self {
    // SECURITY: `transmute!` expands to code that has no branches at runtime.
    let (k0_raw, k1_raw) = transmute!(*key.expose_secret());
    let (n0_raw, n1_raw) = transmute!(nonce.into_array());

    let (k0, k1, n0, n1) = (
      StateBlock::from_128_bits(simd, k0_raw),
      StateBlock::from_128_bits(simd, k1_raw),
      StateBlock::from_128_bits(simd, n0_raw),
      StateBlock::from_128_bits(simd, n1_raw),
    );

    let k0_xor_n0 = k0 ^ n0;
    let k1_xor_n1 = k1 ^ n1;

    let mut state = Self {
      s0: k0_xor_n0,
      s1: k1_xor_n1,
      s2: StateBlock::c1(simd),
      s3: StateBlock::c0(simd),
      s4: k0 ^ StateBlock::c0(simd),
      s5: k1 ^ StateBlock::c1(simd),
      simd,
    };

    for _ in 0..4 {
      for x in [k0, k1, k0_xor_n0, k1_xor_n1] {
        state.s3 = state.s3 ^ StateBlock::ctx(simd);
        state.s5 = state.s5 ^ StateBlock::ctx(simd);
        state.update(x);
      }
    }
    state
  }

  #[inline(always)]
  fn absorb(&mut self, block: &[u8; OUTPUT_RATE_BYTES]) {
    self.update(StateBlock::new(self.simd, *block));
  }

  #[inline(always)]
  fn enc(
    &mut self,
    plaintext_chunk: &[u8; OUTPUT_RATE_BYTES],
    ciphertext_chunk: &mut [u8; OUTPUT_RATE_BYTES],
  ) {
    // PERF: We reorder the operations for (much) better codegen.
    let xi = StateBlock::new(self.simd, *plaintext_chunk);
    let z = xi ^ self.s1 ^ self.s4 ^ self.s5 ^ (self.s2 & self.s3);
    *ciphertext_chunk = z.into_bytes();
    self.update(xi); // PERF: Making Update() last is important.
  }

  #[inline(always)]
  fn dec(
    &mut self,
    ciphertext_chunk: &[u8; OUTPUT_RATE_BYTES],
    plaintext_chunk: &mut [u8; OUTPUT_RATE_BYTES],
  ) {
    // PERF: We reorder the operations for (much) better codegen.
    // (Outputs are the same; only perf improves.)
    let z = StateBlock::new(self.simd, *ciphertext_chunk)
      ^ self.s1
      ^ self.s4
      ^ self.s5
      ^ (self.s2 & self.s3);
    *plaintext_chunk = z.into_bytes();
    self.update(z); // PERF: Making Update() last is important.
  }

  #[inline(always)]
  fn dec_partial(
    &mut self,
    ciphertext_chunk: &[u8; OUTPUT_RATE_BYTES],
    plaintext_remainder: &mut [u8],
  ) {
    // PERF: We reorder the operations for (much) better codegen.
    // (Outputs are the same; only perf improves.)

    let z = StateBlock::new(self.simd, *ciphertext_chunk)
      ^ self.s1
      ^ self.s4
      ^ self.s5
      ^ (self.s2 & self.s3);

    let mut out = z.into_bytes();
    // SECURITY:
    // - The length of the ciphertext and plaintext are not secret.
    // - We don't need the fancy `zeroize()` here since the bytes we're throwing
    //   away are useless. `zero()` (from `zerocopy` crate) is good enough.
    out[plaintext_remainder.len()..].zero(); // ZeroPad(xn) in spec
    plaintext_remainder.copy_from_slice(&out[..plaintext_remainder.len()]);

    self.update(StateBlock::new(self.simd, out));
  }

  #[inline(always)]
  fn finalize<Tag: AuthTag>(
    &mut self,
    plaintext: &[u8],
    associated_data: &[u8],
  ) -> Tag {
    let lengths = StateBlock::from_ints(
      self.simd,
      num_bits(associated_data),
      num_bits(plaintext),
    );

    let t = self.s3 ^ lengths;
    for _ in 0..7 {
      self.update(t);
    }

    let mut tag = Tag::default();
    // NOTE: Due to:
    // - size_of::<Tag>() being a compile-time constant after
    //   monomorphization,
    // - the compiler using dead code elimination,
    // there will be no branching in the generated assembly.
    //
    // SECURITY: We do not depend on the absence of branching since the auth
    // tag size is not secret.
    //
    // While Rust likes to pretend that `size_of::<Tag>()` isn't known at
    // compile time (for complex technical reasons), the optimizer doesn't
    // care and eats those dead branches for breakfast.
    match size_of::<Tag>() {
      AuthTag128::BYTES => {
        // PERF: Adding parens or pulling out subexpressions doesn't even change
        // the generated assembly, let alone improve perf.
        let block = self.s0 ^ self.s1 ^ self.s2 ^ self.s3 ^ self.s4 ^ self.s5;
        // Can't panic as `tag` is guaranteed to be [u8; 16] here.
        tag.as_mut().copy_from_slice(&block.xor_down());
      },
      AuthTag256::BYTES => {
        let high = (self.s0 ^ self.s1 ^ self.s2).xor_down();
        let low = (self.s3 ^ self.s4 ^ self.s5).xor_down();
        // Can't panic as `tag` is guaranteed to be [u8; 32] here.
        tag.as_mut().copy_from_slice([high, low].as_flattened());
      },
      _ => unreachable!("Tag limited to 16 or 32 bytes at compile-time."),
    }
    tag
  }
}