| ... | ... | @@ -38,7 +38,35 @@ pub const State = struct { |
| 38 | 38 | return mem.sliceAsBytes(self.data[0..]); |
| 39 | 39 | } |
| 40 | 40 | |
| 41 | | pub fn permute(self: *Self) void { |
| 41 | fn permute_unrolled(self: *Self) void { |
| 42 | const state = &self.data; |
| 43 | comptime var round = @as(u32, 24); |
| 44 | inline while (round > 0) : (round -= 1) { |
| 45 | var column = @as(usize, 0); |
| 46 | while (column < 4) : (column += 1) { |
| 47 | const x = math.rotl(u32, state[column], 24); |
| 48 | const y = math.rotl(u32, state[4 + column], 9); |
| 49 | const z = state[8 + column]; |
| 50 | state[8 + column] = ((x ^ (z << 1)) ^ ((y & z) << 2)); |
| 51 | state[4 + column] = ((y ^ x) ^ ((x | z) << 1)); |
| 52 | state[column] = ((z ^ y) ^ ((x & y) << 3)); |
| 53 | } |
| 54 | switch (round & 3) { |
| 55 | 0 => { |
| 56 | mem.swap(u32, &state[0], &state[1]); |
| 57 | mem.swap(u32, &state[2], &state[3]); |
| 58 | state[0] ^= round | 0x9e377900; |
| 59 | }, |
| 60 | 2 => { |
| 61 | mem.swap(u32, &state[0], &state[2]); |
| 62 | mem.swap(u32, &state[1], &state[3]); |
| 63 | }, |
| 64 | else => {}, |
| 65 | } |
| 66 | } |
| 67 | } |
| 68 | |
| 69 | fn permute_small(self: *Self) void { |
| 42 | 70 | const state = &self.data; |
| 43 | 71 | var round = @as(u32, 24); |
| 44 | 72 | while (round > 0) : (round -= 1) { |
| ... | ... | @@ -66,6 +94,8 @@ pub const State = struct { |
| 66 | 94 | } |
| 67 | 95 | } |
| 68 | 96 | |
| 97 | pub const permute = if (std.builtin.mode == .ReleaseSmall) permute_small else permute_unrolled; |
| 98 | |
| 69 | 99 | pub fn squeeze(self: *Self, out: []u8) void { |
| 70 | 100 | var i = @as(usize, 0); |
| 71 | 101 | while (i + RATE <= out.len) : (i += RATE) { |
| ... | ... | @@ -249,15 +279,15 @@ pub const Aead = struct { |
| 249 | 279 | in = in[State.RATE..]; |
| 250 | 280 | out = out[State.RATE..]; |
| 251 | 281 | }) { |
| 252 | | for (buf[0..State.RATE]) |*p, i| { |
| 253 | | p.* ^= in[i]; |
| 254 | | out[i] = p.*; |
| 282 | for (in[0..State.RATE]) |v, i| { |
| 283 | buf[i] ^= v; |
| 255 | 284 | } |
| 285 | mem.copy(u8, out[0..State.RATE], buf[0..State.RATE]); |
| 256 | 286 | state.permute(); |
| 257 | 287 | } |
| 258 | | for (buf[0..in.len]) |*p, i| { |
| 259 | | p.* ^= in[i]; |
| 260 | | out[i] = p.*; |
| 288 | for (in[0..]) |v, i| { |
| 289 | buf[i] ^= v; |
| 290 | out[i] = buf[i]; |
| 261 | 291 | } |
| 262 | 292 | |
| 263 | 293 | // XOR 1 into the next byte of the state |
| ... | ... | @@ -291,15 +321,17 @@ pub const Aead = struct { |
| 291 | 321 | in = in[State.RATE..]; |
| 292 | 322 | out = out[State.RATE..]; |
| 293 | 323 | }) { |
| 294 | | for (buf[0..State.RATE]) |*p, i| { |
| 295 | | out[i] = p.* ^ in[i]; |
| 296 | | p.* = in[i]; |
| 324 | const d = in[0..State.RATE].*; |
| 325 | for (d) |v, i| { |
| 326 | out[i] = buf[i] ^ v; |
| 297 | 327 | } |
| 328 | mem.copy(u8, buf[0..State.RATE], d[0..State.RATE]); |
| 298 | 329 | state.permute(); |
| 299 | 330 | } |
| 300 | 331 | for (buf[0..in.len]) |*p, i| { |
| 301 | | out[i] = p.* ^ in[i]; |
| 302 | | p.* = in[i]; |
| 332 | const d = in[i]; |
| 333 | out[i] = p.* ^ d; |
| 334 | p.* = d; |
| 303 | 335 | } |
| 304 | 336 | |
| 305 | 337 | // XOR 1 into the next byte of the state |