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
// SPDX-License-Identifier: Apache-2.0 OR ISC OR MIT-0
//! The steps of P-256's generator multiplication, using aws-lc's
//! `p256-x86_64-asm` routines, which aws-lc's `ecp_nistz256_point_mul_base`
//! runs over its 7-bit generator table.
//!
//! The assembly is aws-lc's, imported by `admin/import-aws-lc`. The row scan
//! reads all 64 points of a row and masks off all but one, and the mixed
//! addition selects its special cases with masks, so neither branches on or
//! indexes by its inputs. Neither is formally verified: their correctness
//! rests on their use and testing in OpenSSL, BoringSSL, and aws-lc, and on
//! rind's tests of them against s2n-bignum's verified arithmetic,
//! Wycheproof, and aws-lc-rs.
use super::cpu::{HaveAvx2, HaveBmi2Adx};
/// One row of the generator table: 64 affine points with Montgomery
/// coordinates, `x` then `y`, each 4 little-endian words. The scans load it
/// with `movdqa` and `vmovdqa`.
#[repr(C, align(64))]
pub(crate) struct P256GeneratorRow(pub(crate) [u64; 512]);
impl core::ops::Deref for P256GeneratorRow {
type Target = [u64; 512];
fn deref(&self) -> &Self::Target {
&self.0
}
}
/// Returns `row`'s point `index - 1`, or zeros for an `index` of 0 or above
/// 64, reading every point.
///
/// `index` is secret, so not even a debug build asserts on it.
pub(crate) fn p256_select_w7(row: &P256GeneratorRow, index: u8) -> [u64; 8] {
let mut r = [0; 8];
// SAFETY: `r` is one writable point and `row` 64 readable points, as
// aligned as the routines load them. The AVX2 routine runs only where the
// token proves AVX2, and the other needs only SSE2.
unsafe {
match HaveAvx2::check() {
Some(_) => ecp_nistz256_avx2_select_w7(&mut r, row, index.into()),
None => ecp_nistz256_select_w7(&mut r, row, index.into()),
}
}
r
}
/// Returns the Jacobian `p1` plus the affine `p2`, both in Montgomery form.
///
/// Either may be infinity: `p1` with a zero `z`, `p2` as all zeros. They must
/// not be the same point, which this does not double.
pub(crate) fn p256_point_add_affine(p1: &[u64; 12], p2: &[u64; 8]) -> [u64; 12] {
let mut r = [0; 12];
// SAFETY: `r` is one writable Jacobian point, `p1` one readable Jacobian
// point, and `p2` one readable affine point; the routines load them
// unaligned. The BMI2 and ADX routine runs only where the token proves
// both, and the other needs only SSE2.
unsafe {
match HaveBmi2Adx::check() {
Some(_) => ecp_nistz256_point_add_affinex(&mut r, p1, p2),
None => ecp_nistz256_point_add_affine(&mut r, p1, p2),
}
}
r
}
/// aws-lc's `.Lpoly`, the field prime, as the multiplications load it.
#[repr(C, align(16))]
struct Poly([u64; 4]);
/// aws-lc's `.LOne`, `.LTwo`, and `.LThree`, the scans' counters, as aligned
/// as the AVX2 scan loads them.
#[repr(C, align(32))]
struct One([u32; 24]);
/// aws-lc's `.LONE_mont`, one in Montgomery form, which the addition masks
/// with `pand`. It runs on through `.Lord` and `.LordK`, which the import
/// keeps but these routines never read.
#[repr(C, align(16))]
struct OneMont([u64; 9]);
static POLY: Poly = Poly(include!("aws_lc/p256-x86_64-asm.poly.rs"));
static ONE: One = One(include!("aws_lc/p256-x86_64-asm.One.rs"));
static ONE_MONT: OneMont = OneMont(include!("aws_lc/p256-x86_64-asm.ONE_mont.rs"));
/// Defines a naked function running an imported routine, which reaches its
/// tables and helpers through the `sym` operands listed. Each takes the System
/// V ABI on every OS, since its text is aws-lc's ELF and Mach-O flavors.
macro_rules! p256_routine {
($name:ident($($arg:ident: $ty:ty),*) { $($op:ident = $sym:ident),* }) => {
#[unsafe(naked)]
unsafe extern "sysv64" fn $name($($arg: $ty),*) {
#[cfg(not(target_vendor = "apple"))]
core::arch::naked_asm!(
include_str!(concat!("aws_lc/", stringify!($name), ".linux.s"))
$(, $op = sym $sym)*,
options(att_syntax),
);
#[cfg(target_vendor = "apple")]
core::arch::naked_asm!(
include_str!(concat!("aws_lc/", stringify!($name), ".apple.s"))
$(, $op = sym $sym)*,
options(att_syntax),
);
}
};
}
p256_routine!(ecp_nistz256_select_w7(out: *mut [u64; 8], row: *const P256GeneratorRow, index: u32) {
One = ONE
});
p256_routine!(ecp_nistz256_avx2_select_w7(out: *mut [u64; 8], row: *const P256GeneratorRow, index: u32) {
One = ONE
});
p256_routine!(ecp_nistz256_point_add_affine(r: *mut [u64; 12], a: *const [u64; 12], b: *const [u64; 8]) {
ONE_mont = ONE_MONT,
ecp_nistz256_mul_montq = ecp_nistz256_mul_montq,
ecp_nistz256_sqr_montq = ecp_nistz256_sqr_montq,
ecp_nistz256_sub_fromq = ecp_nistz256_sub_fromq,
ecp_nistz256_subq = ecp_nistz256_subq
});
p256_routine!(ecp_nistz256_point_add_affinex(r: *mut [u64; 12], a: *const [u64; 12], b: *const [u64; 8]) {
ONE_mont = ONE_MONT,
ecp_nistz256_mul_montx = ecp_nistz256_mul_montx,
ecp_nistz256_sqr_montx = ecp_nistz256_sqr_montx,
ecp_nistz256_sub_fromx = ecp_nistz256_sub_fromx,
ecp_nistz256_subx = ecp_nistz256_subx
});
// The routines' helpers, which pass their arguments in registers of their
// own choosing rather than by the C ABI, so only the routines call them.
p256_routine!(ecp_nistz256_mul_montq() { poly = POLY });
p256_routine!(ecp_nistz256_sqr_montq() { poly = POLY });
p256_routine!(ecp_nistz256_sub_fromq() {});
p256_routine!(ecp_nistz256_subq() {});
p256_routine!(ecp_nistz256_mul_montx() { poly = POLY });
p256_routine!(ecp_nistz256_sqr_montx() { poly = POLY });
p256_routine!(ecp_nistz256_sub_fromx() {});
p256_routine!(ecp_nistz256_subx() {});
#[cfg(test)]
#[cfg_attr(coverage_nightly, coverage(off))]
mod tests {
use super::super::adx_mux::tests::Rng;
use super::*;
// The other tests run only the twin this CPU picks, so each pair is
// checked against each other here, where the CPU runs both, and the
// mixed addition against s2n-bignum's verified one.
#[test]
fn twins_agree() {
let mut rng = Rng(0x243f_6a88_85a3_08d3);
let mut row = P256GeneratorRow([0; 512]);
for word in row.0.iter_mut() {
*word = rng.next();
}
if HaveAvx2::check().is_some() {
for index in 0..=255 {
let (mut sse2, mut avx2) = ([1; 8], [1; 8]);
// SAFETY: as in `p256_select_w7`, on a CPU with AVX2.
unsafe {
ecp_nistz256_select_w7(&mut sse2, &row, index);
ecp_nistz256_avx2_select_w7(&mut avx2, &row, index);
}
assert_eq!(sse2, avx2);
let want = match index {
0 | 65.. => [0; 8],
i => row.0[(i as usize - 1) * 8..][..8].try_into().unwrap(),
};
assert_eq!(sse2, want);
}
}
if HaveBmi2Adx::check().is_some() {
// The addition is straight-line field arithmetic, so off-curve
// points exercise it as well as real ones, and a zero `z` or
// `p2` its infinity cases.
for i in 0..256 {
let mut p1: [u64; 12] = rng.field(255);
let mut p2: [u64; 8] = rng.field(255);
match i % 8 {
0 => p1[8..].fill(0),
1 => p2.fill(0),
_ => {}
}
let (mut mulq, mut mulx) = ([0; 12], [0; 12]);
// SAFETY: as in `p256_point_add_affine`, on a CPU with BMI2
// and ADX.
unsafe {
ecp_nistz256_point_add_affine(&mut mulq, &p1, &p2);
ecp_nistz256_point_add_affinex(&mut mulx, &p1, &p2);
}
assert_eq!(mulq, mulx);
}
}
// `p1` is not at infinity, which s2n-bignum's addition requires.
for _ in 0..256 {
let mut p1: [u64; 12] = rng.field(255);
let p2: [u64; 8] = rng.field(255);
p1[8] |= 1;
let mut s2n = [0; 12];
super::super::adx_mux::p256_montjmixadd(&mut s2n, &p1, &p2);
assert_eq!(p256_point_add_affine(&p1, &p2), s2n);
}
}
}