rscrypto 0.1.0

Rust crypto with zero default deps: BLAKE3, Ed25519/X25519, hashes, MACs, KDFs, AEADs, and checksums with SIMD/ASM acceleration.
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
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240

use core::arch::x86_64::{
  __m256i, _mm256_add_epi32, _mm256_loadu_si256, _mm256_or_si256, _mm256_permute2x128_si256, _mm256_set_epi8,
  _mm256_set1_epi32, _mm256_setr_epi32, _mm256_shuffle_epi8, _mm256_slli_epi32, _mm256_srli_epi32, _mm256_storeu_si256,
  _mm256_unpackhi_epi32, _mm256_unpackhi_epi64, _mm256_unpacklo_epi32, _mm256_unpacklo_epi64, _mm256_xor_si256,
};

use super::{BLOCK_SIZE, KEY_SIZE, NONCE_SIZE, load_u32_le, x86_ssse3_x4, xor_keystream_portable};

const BLOCKS_PER_BATCH: usize = 8;

#[inline]
pub(super) fn xor_keystream(key: &[u8; KEY_SIZE], initial_counter: u32, nonce: &[u8; NONCE_SIZE], buffer: &mut [u8]) {
  // SAFETY: Backend selection guarantees AVX2 is available before this wrapper is chosen.
  unsafe { xor_keystream_impl(key, initial_counter, nonce, buffer) }
}

#[target_feature(enable = "avx2")]
unsafe fn xor_keystream_impl(key: &[u8; KEY_SIZE], initial_counter: u32, nonce: &[u8; NONCE_SIZE], buffer: &mut [u8]) {
  // vpshufb masks for byte-aligned rotations (16-bit and 8-bit).
  let rot16 = _mm256_set_epi8(
    13, 12, 15, 14, 9, 8, 11, 10, 5, 4, 7, 6, 1, 0, 3, 2, 13, 12, 15, 14, 9, 8, 11, 10, 5, 4, 7, 6, 1, 0, 3, 2,
  );
  let rot8 = _mm256_set_epi8(
    14, 13, 12, 15, 10, 9, 8, 11, 6, 5, 4, 7, 2, 1, 0, 3, 14, 13, 12, 15, 10, 9, 8, 11, 6, 5, 4, 7, 2, 1, 0, 3,
  );

  let mut counter = initial_counter;
  let mut batches = buffer.chunks_exact_mut(BLOCK_SIZE * BLOCKS_PER_BATCH);
  for chunk in &mut batches {
    debug_assert!(counter.checked_add((BLOCKS_PER_BATCH - 1) as u32).is_some());

    let mut x0 = _mm256_set1_epi32(0x6170_7865u32 as i32);
    let mut x1 = _mm256_set1_epi32(0x3320_646eu32 as i32);
    let mut x2 = _mm256_set1_epi32(0x7962_2d32u32 as i32);
    let mut x3 = _mm256_set1_epi32(0x6b20_6574u32 as i32);
    let mut x4 = _mm256_set1_epi32(load_u32_le(&key[0..4]) as i32);
    let mut x5 = _mm256_set1_epi32(load_u32_le(&key[4..8]) as i32);
    let mut x6 = _mm256_set1_epi32(load_u32_le(&key[8..12]) as i32);
    let mut x7 = _mm256_set1_epi32(load_u32_le(&key[12..16]) as i32);
    let mut x8 = _mm256_set1_epi32(load_u32_le(&key[16..20]) as i32);
    let mut x9 = _mm256_set1_epi32(load_u32_le(&key[20..24]) as i32);
    let mut x10 = _mm256_set1_epi32(load_u32_le(&key[24..28]) as i32);
    let mut x11 = _mm256_set1_epi32(load_u32_le(&key[28..32]) as i32);
    let mut x12 = _mm256_setr_epi32(
      counter as i32,
      counter.wrapping_add(1) as i32,
      counter.wrapping_add(2) as i32,
      counter.wrapping_add(3) as i32,
      counter.wrapping_add(4) as i32,
      counter.wrapping_add(5) as i32,
      counter.wrapping_add(6) as i32,
      counter.wrapping_add(7) as i32,
    );
    let mut x13 = _mm256_set1_epi32(load_u32_le(&nonce[0..4]) as i32);
    let mut x14 = _mm256_set1_epi32(load_u32_le(&nonce[4..8]) as i32);
    let mut x15 = _mm256_set1_epi32(load_u32_le(&nonce[8..12]) as i32);

    let o0 = x0;
    let o1 = x1;
    let o2 = x2;
    let o3 = x3;
    let o4 = x4;
    let o5 = x5;
    let o6 = x6;
    let o7 = x7;
    let o8 = x8;
    let o9 = x9;
    let o10 = x10;
    let o11 = x11;
    let o12 = x12;
    let o13 = x13;
    let o14 = x14;
    let o15 = x15;

    let mut round = 0usize;
    while round < 10 {
      quarter_round(&mut x0, &mut x4, &mut x8, &mut x12, rot16, rot8);
      quarter_round(&mut x1, &mut x5, &mut x9, &mut x13, rot16, rot8);
      quarter_round(&mut x2, &mut x6, &mut x10, &mut x14, rot16, rot8);
      quarter_round(&mut x3, &mut x7, &mut x11, &mut x15, rot16, rot8);

      quarter_round(&mut x0, &mut x5, &mut x10, &mut x15, rot16, rot8);
      quarter_round(&mut x1, &mut x6, &mut x11, &mut x12, rot16, rot8);
      quarter_round(&mut x2, &mut x7, &mut x8, &mut x13, rot16, rot8);
      quarter_round(&mut x3, &mut x4, &mut x9, &mut x14, rot16, rot8);

      round = round.strict_add(1);
    }

    x0 = _mm256_add_epi32(x0, o0);
    x1 = _mm256_add_epi32(x1, o1);
    x2 = _mm256_add_epi32(x2, o2);
    x3 = _mm256_add_epi32(x3, o3);
    x4 = _mm256_add_epi32(x4, o4);
    x5 = _mm256_add_epi32(x5, o5);
    x6 = _mm256_add_epi32(x6, o6);
    x7 = _mm256_add_epi32(x7, o7);
    x8 = _mm256_add_epi32(x8, o8);
    x9 = _mm256_add_epi32(x9, o9);
    x10 = _mm256_add_epi32(x10, o10);
    x11 = _mm256_add_epi32(x11, o11);
    x12 = _mm256_add_epi32(x12, o12);
    x13 = _mm256_add_epi32(x13, o13);
    x14 = _mm256_add_epi32(x14, o14);
    x15 = _mm256_add_epi32(x15, o15);

    // 16×8 matrix transpose: convert from word-major (x_i = word i for 8 blocks)
    // to block-major (each pair of YMM registers = one 64-byte block).
    //
    // Stage 1: 32-bit interleave.
    // SAFETY: AVX2 intrinsics are valid under the enclosing target_feature.
    let s1_0 = _mm256_unpacklo_epi32(x0, x1);
    let s1_1 = _mm256_unpackhi_epi32(x0, x1);
    let s1_2 = _mm256_unpacklo_epi32(x2, x3);
    let s1_3 = _mm256_unpackhi_epi32(x2, x3);
    let s1_4 = _mm256_unpacklo_epi32(x4, x5);
    let s1_5 = _mm256_unpackhi_epi32(x4, x5);
    let s1_6 = _mm256_unpacklo_epi32(x6, x7);
    let s1_7 = _mm256_unpackhi_epi32(x6, x7);
    let s1_8 = _mm256_unpacklo_epi32(x8, x9);
    let s1_9 = _mm256_unpackhi_epi32(x8, x9);
    let s1_10 = _mm256_unpacklo_epi32(x10, x11);
    let s1_11 = _mm256_unpackhi_epi32(x10, x11);
    let s1_12 = _mm256_unpacklo_epi32(x12, x13);
    let s1_13 = _mm256_unpackhi_epi32(x12, x13);
    let s1_14 = _mm256_unpacklo_epi32(x14, x15);
    let s1_15 = _mm256_unpackhi_epi32(x14, x15);

    // Stage 2: 64-bit interleave. After this, each 128-bit lane holds 4 consecutive
    // words from one block (lo lane = blocks 0-3, hi lane = blocks 4-7).
    let s2_0 = _mm256_unpacklo_epi64(s1_0, s1_2);
    let s2_1 = _mm256_unpackhi_epi64(s1_0, s1_2);
    let s2_2 = _mm256_unpacklo_epi64(s1_1, s1_3);
    let s2_3 = _mm256_unpackhi_epi64(s1_1, s1_3);
    let s2_4 = _mm256_unpacklo_epi64(s1_4, s1_6);
    let s2_5 = _mm256_unpackhi_epi64(s1_4, s1_6);
    let s2_6 = _mm256_unpacklo_epi64(s1_5, s1_7);
    let s2_7 = _mm256_unpackhi_epi64(s1_5, s1_7);
    let s2_8 = _mm256_unpacklo_epi64(s1_8, s1_10);
    let s2_9 = _mm256_unpackhi_epi64(s1_8, s1_10);
    let s2_10 = _mm256_unpacklo_epi64(s1_9, s1_11);
    let s2_11 = _mm256_unpackhi_epi64(s1_9, s1_11);
    let s2_12 = _mm256_unpacklo_epi64(s1_12, s1_14);
    let s2_13 = _mm256_unpackhi_epi64(s1_12, s1_14);
    let s2_14 = _mm256_unpacklo_epi64(s1_13, s1_15);
    let s2_15 = _mm256_unpackhi_epi64(s1_13, s1_15);

    // Stage 3: 128-bit lane permute. Separate lo/hi lanes to form complete blocks.
    // Each block = 2 YMM registers (32 + 32 = 64 bytes).
    // Process block pairs (j, j+4) and XOR+store immediately to ease register pressure.
    let ptr = chunk.as_mut_ptr();
    // SAFETY: each `chunk` slice has exactly BLOCKS_PER_BATCH * BLOCK_SIZE = 512
    // bytes, so all 8 × 64-byte load/store pairs are in bounds.
    unsafe {
      // Blocks 0 and 4 — from s2[0], s2[4], s2[8], s2[12]
      xor_block_pair(ptr, 0, 4, s2_0, s2_4, s2_8, s2_12);
      // Blocks 1 and 5 — from s2[1], s2[5], s2[9], s2[13]
      xor_block_pair(ptr, 1, 5, s2_1, s2_5, s2_9, s2_13);
      // Blocks 2 and 6 — from s2[2], s2[6], s2[10], s2[14]
      xor_block_pair(ptr, 2, 6, s2_2, s2_6, s2_10, s2_14);
      // Blocks 3 and 7 — from s2[3], s2[7], s2[11], s2[15]
      xor_block_pair(ptr, 3, 7, s2_3, s2_7, s2_11, s2_15);
    }

    counter = counter.wrapping_add(BLOCKS_PER_BATCH as u32);
  }

  let remainder = batches.into_remainder();
  let mut x4_batches = remainder.chunks_exact_mut(BLOCK_SIZE * x86_ssse3_x4::BLOCKS_PER_BATCH);
  for chunk in &mut x4_batches {
    // SAFETY: AVX2-capable CPUs provide the SSSE3 instructions used by the
    // 4-block tail kernel, and `chunk` is exactly 4 ChaCha20 blocks.
    unsafe { x86_ssse3_x4::xor_blocks(key, counter, nonce, chunk) };
    counter = counter.wrapping_add(x86_ssse3_x4::BLOCKS_PER_BATCH as u32);
  }

  let remainder = x4_batches.into_remainder();
  if !remainder.is_empty() {
    xor_keystream_portable(key, counter, nonce, remainder);
  }
}

/// Permute stage-2 results into two complete blocks (lo_idx and hi_idx) and
/// XOR+store them in-place. Each block is 64 bytes = 2 × YMM.
#[inline(always)]
unsafe fn xor_block_pair(
  buf: *mut u8,
  lo_idx: usize,
  hi_idx: usize,
  w03: __m256i,
  w47: __m256i,
  w811: __m256i,
  w1215: __m256i,
) {
  // SAFETY: caller guarantees `buf` points to a 512-byte chunk and indices < 8.
  unsafe {
    // lo_idx block: select lo 128-bit lanes (0x20 = [a_lo, b_lo])
    let blk_lo_a = _mm256_permute2x128_si256::<0x20>(w03, w47);
    let blk_lo_b = _mm256_permute2x128_si256::<0x20>(w811, w1215);
    // hi_idx block: select hi 128-bit lanes (0x31 = [a_hi, b_hi])
    let blk_hi_a = _mm256_permute2x128_si256::<0x31>(w03, w47);
    let blk_hi_b = _mm256_permute2x128_si256::<0x31>(w811, w1215);

    // XOR and store lo_idx block
    let p_lo = buf.add(lo_idx.strict_mul(BLOCK_SIZE));
    let pt0 = _mm256_loadu_si256(p_lo.cast());
    let pt1 = _mm256_loadu_si256(p_lo.add(32).cast());
    _mm256_storeu_si256(p_lo.cast(), _mm256_xor_si256(pt0, blk_lo_a));
    _mm256_storeu_si256(p_lo.add(32).cast(), _mm256_xor_si256(pt1, blk_lo_b));

    // XOR and store hi_idx block
    let p_hi = buf.add(hi_idx.strict_mul(BLOCK_SIZE));
    let pt2 = _mm256_loadu_si256(p_hi.cast());
    let pt3 = _mm256_loadu_si256(p_hi.add(32).cast());
    _mm256_storeu_si256(p_hi.cast(), _mm256_xor_si256(pt2, blk_hi_a));
    _mm256_storeu_si256(p_hi.add(32).cast(), _mm256_xor_si256(pt3, blk_hi_b));
  }
}

#[inline(always)]
fn quarter_round(a: &mut __m256i, b: &mut __m256i, c: &mut __m256i, d: &mut __m256i, rot16: __m256i, rot8: __m256i) {
  // SAFETY: this helper is only reached from the AVX2-enabled backend.
  unsafe {
    *a = _mm256_add_epi32(*a, *b);
    *d = _mm256_shuffle_epi8(_mm256_xor_si256(*d, *a), rot16);
    *c = _mm256_add_epi32(*c, *d);
    *b = rotl::<12, 20>(_mm256_xor_si256(*b, *c));
    *a = _mm256_add_epi32(*a, *b);
    *d = _mm256_shuffle_epi8(_mm256_xor_si256(*d, *a), rot8);
    *c = _mm256_add_epi32(*c, *d);
    *b = rotl::<7, 25>(_mm256_xor_si256(*b, *c));
  }
}

#[inline(always)]
fn rotl<const LEFT: i32, const RIGHT: i32>(value: __m256i) -> __m256i {
  const { assert!(LEFT + RIGHT == 32, "rotation amounts must sum to 32") };
  // SAFETY: only reached from the AVX2-enabled backend.
  unsafe { _mm256_or_si256(_mm256_slli_epi32(value, LEFT), _mm256_srli_epi32(value, RIGHT)) }
}