aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
-rw-r--r--src/accessor.rs58
-rw-r--r--src/lib.rs88
2 files changed, 132 insertions, 14 deletions
diff --git a/src/accessor.rs b/src/accessor.rs
index 35d2890..e7e6de6 100644
--- a/src/accessor.rs
+++ b/src/accessor.rs
@@ -8,6 +8,8 @@ pub struct BlockAccessor<'a> {
num_block_pairs: usize,
}
+const BIG_CHUNK_SIZE: usize = 8 * 32;
+
impl<'a> BlockAccessor<'a> {
pub fn new(message: &'a mut [u8]) -> Self {
let num_block_pairs = (message.len() - 16 - 16) / 32;
@@ -78,4 +80,60 @@ impl<'a> BlockAccessor<'a> {
.map(move |x| x.split_at_mut(16))
.map(move |(x, y)| (x.try_into().unwrap(), y.try_into().unwrap()))
}
+
+ pub fn suffix_8_mut(&mut self) -> impl Iterator<Item = (&mut [u8; 16], &mut [u8; 16])> {
+ let start = self.suffix_start() / BIG_CHUNK_SIZE * BIG_CHUNK_SIZE;
+ let stop = self.suffix_start();
+ self.data[start..stop]
+ .chunks_exact_mut(32)
+ .map(move |x| x.split_at_mut(16))
+ .map(move |(x, y)| (x.try_into().unwrap(), y.try_into().unwrap()))
+ }
+
+ pub fn pairs_8_mut(
+ &mut self,
+ ) -> impl Iterator<Item = ([&mut [u8; 16]; 8], [&mut [u8; 16]; 8])> {
+ let stop = self.suffix_start() / BIG_CHUNK_SIZE * BIG_CHUNK_SIZE;
+ self.data[..stop]
+ .chunks_exact_mut(BIG_CHUNK_SIZE)
+ .map(move |x| {
+ let (b0, b1) = x.split_at_mut(BIG_CHUNK_SIZE / 2);
+ let (b00, b01) = b0.split_at_mut(BIG_CHUNK_SIZE / 4);
+ let (b10, b11) = b1.split_at_mut(BIG_CHUNK_SIZE / 4);
+ let (b000, b001) = b00.split_at_mut(BIG_CHUNK_SIZE / 8);
+ let (b010, b011) = b01.split_at_mut(BIG_CHUNK_SIZE / 8);
+ let (b100, b101) = b10.split_at_mut(BIG_CHUNK_SIZE / 8);
+ let (b110, b111) = b11.split_at_mut(BIG_CHUNK_SIZE / 8);
+ let (b0000, b0001) = b000.split_at_mut(16);
+ let (b0010, b0011) = b001.split_at_mut(16);
+ let (b0100, b0101) = b010.split_at_mut(16);
+ let (b0110, b0111) = b011.split_at_mut(16);
+ let (b1000, b1001) = b100.split_at_mut(16);
+ let (b1010, b1011) = b101.split_at_mut(16);
+ let (b1100, b1101) = b110.split_at_mut(16);
+ let (b1110, b1111) = b111.split_at_mut(16);
+ (
+ [
+ b0000.try_into().unwrap(),
+ b0010.try_into().unwrap(),
+ b0100.try_into().unwrap(),
+ b0110.try_into().unwrap(),
+ b1000.try_into().unwrap(),
+ b1010.try_into().unwrap(),
+ b1100.try_into().unwrap(),
+ b1110.try_into().unwrap(),
+ ],
+ [
+ b0001.try_into().unwrap(),
+ b0011.try_into().unwrap(),
+ b0101.try_into().unwrap(),
+ b0111.try_into().unwrap(),
+ b1001.try_into().unwrap(),
+ b1011.try_into().unwrap(),
+ b1101.try_into().unwrap(),
+ b1111.try_into().unwrap(),
+ ],
+ )
+ })
+ }
}
diff --git a/src/lib.rs b/src/lib.rs
index 5759046..919e91b 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -481,6 +481,54 @@ fn cipher_aez_tiny(mode: Mode, aez: &Aez, tweaks: Tweak, message: &mut [u8]) {
}
}
+macro_rules! unroll_pairs {
+ (
+ $accessor:expr;
+ setup_unrolled => $setup_unrolled:block;
+ setup_single => $setup_single:block;
+ roll ($left:ident, $right:ident) => $roll:block;
+ ) => {
+ for (left, right) in $accessor.pairs_8_mut() {
+ $setup_unrolled;
+
+ let [l0, l1, l2, l3, l4, l5, l6, l7] = left;
+ let [r0, r1, r2, r3, r4, r5, r6, r7] = right;
+
+ let $left = l0;
+ let $right = r0;
+ $roll;
+ let $left = l1;
+ let $right = r1;
+ $roll;
+ let $left = l2;
+ let $right = r2;
+ $roll;
+ let $left = l3;
+ let $right = r3;
+ $roll;
+ let $left = l4;
+ let $right = r4;
+ $roll;
+ let $left = l5;
+ let $right = r5;
+ $roll;
+ let $left = l6;
+ let $right = r6;
+ $roll;
+ let $left = l7;
+ let $right = r7;
+ $roll;
+ }
+
+ for (left, right) in $accessor.suffix_8_mut() {
+ let $left = left;
+ let $right = right;
+ $setup_single;
+ $roll;
+ }
+ };
+}
+
fn pass_one(aez: &Aez, blocks: &mut BlockAccessor) -> Block {
let mut x = Block::null();
let mut e1_eval = E::new(1, 0, aez);
@@ -507,20 +555,32 @@ fn pass_two(aez: &Aez, blocks: &mut BlockAccessor, s: Block) -> Block {
let e2_eval = E::new(2, 0, aez);
let mut e1_eval = E::new(1, 0, aez);
let e0_eval = E::new(0, 0, aez);
+ let mut evals_for_s = e2_eval.evals_for(s);
- for ((raw_wi, raw_xi), s_) in blocks.pairs_mut().zip(e2_eval.evals_for(s)) {
- e1_eval.advance();
- let wi = Block::from(*raw_wi);
- let xi = Block::from(*raw_xi);
- let yi = wi ^ s_;
- let zi = xi ^ s_;
- let ci_ = yi ^ e0_eval.eval(zi);
- let ci = zi ^ e1_eval.eval(ci_);
-
- ci.write_to(raw_wi);
- ci_.write_to(raw_xi);
-
- y = y ^ yi;
+ unroll_pairs! { blocks;
+ setup_unrolled => {
+ evals_for_s.refill();
+ };
+ setup_single => {
+ if evals_for_s.len == 0 {
+ evals_for_s.refill();
+ }
+ };
+ roll (raw_wi, raw_xi) => {
+ e1_eval.advance();
+ let wi = Block::from(*raw_wi);
+ let xi = Block::from(*raw_xi);
+ let yi = wi ^ evals_for_s.blocks[8 - evals_for_s.len];
+ let zi = xi ^ evals_for_s.blocks[8 - evals_for_s.len];
+ let ci_ = yi ^ e0_eval.eval(zi);
+ let ci = zi ^ e1_eval.eval(ci_);
+
+ ci.write_to(raw_wi);
+ ci_.write_to(raw_xi);
+
+ y = y ^ yi;
+ evals_for_s.len -= 1;
+ };
}
y
@@ -692,7 +752,7 @@ impl<'a> E<'a> {
self.aez.aes.aes4(block ^ delta)
}
- fn evals_for(self, block: Block) -> impl Iterator<Item = Block> {
+ fn evals_for(self, block: Block) -> Eiter<'a> {
Eiter::new(self, block)
}