authorgravatar for 124872+jedisct1@users.noreply.github.comFrank Denis <124872+jedisct1@users.noreply.github.com> 2020-09-23 10:18:17+02:00
committergravatar for andrew@ziglang.orgAndrew Kelley <andrew@ziglang.org> 2020-09-24 13:16:00-04:00
logbd89bd6fdbcc0ce5ea7763a8043fd46099022b19
treefb79f3598f44badcf0842ee3cef9853d44be2cf9
parentbba4576281241e8824de3d48b65c55fa1b1d4c9c

Revamp crypto/aes

* Reorganize crypto/aes in order to separate parameters, implementations and modes. * Add a zero-cost abstraction over the internal representation of a block, so that blocks can be kept in vector registers in optimized implementations. * Add architecture-independent aesenc/aesdec/aesenclast/aesdeclast operations, so that any AES-based primitive can be implemented, including these that don't use the original key schedule (AES-PRF, AEGIS, MeowHash...) * Add support for parallelization/wide blocks to take advantage of hardware implementations. * Align T-tables to cache lines in the software implementations to slightly reduce side channels. * Add an optimized implementation for modern Intel CPUs with AES-NI. * Add new tests (AES256 key expansion). * Reimplement the counter mode to work with any block cipher, any endianness and to take advantage of wide blocks. * Add benchmarks for AES.

6 files changed, 1363 insertions(+), 643 deletions(-)

lib/std/crypto.zig+8
...@@ -59,6 +59,13 @@ pub const pwhash = struct {...@@ -59,6 +59,13 @@ pub const pwhash = struct {
59pub const core = struct {59pub const core = struct {
60 pub const aes = @import("crypto/aes.zig");60 pub const aes = @import("crypto/aes.zig");
61 pub const Gimli = @import("crypto/gimli.zig").State;61 pub const Gimli = @import("crypto/gimli.zig").State;
62
63 /// Modes are generic compositions to construct encryption/decryption functions from block ciphers and permutations.
64 ///
65 /// These modes are designed to be building blocks for higher-level constructions, and should generally not be used directly by applications, as they may not provide the expected properties and security guarantees.
66 ///
67 /// Most applications may want to use AEADs instead.
68 pub const modes = @import("crypto/modes.zig");
62};69};
6370
64/// Elliptic-curve arithmetic.71/// Elliptic-curve arithmetic.
...@@ -111,6 +118,7 @@ test "crypto" {...@@ -111,6 +118,7 @@ test "crypto" {
111 _ = @import("crypto/gimli.zig");118 _ = @import("crypto/gimli.zig");
112 _ = @import("crypto/hmac.zig");119 _ = @import("crypto/hmac.zig");
113 _ = @import("crypto/md5.zig");120 _ = @import("crypto/md5.zig");
121 _ = @import("crypto/modes.zig");
114 _ = @import("crypto/pbkdf2.zig");122 _ = @import("crypto/pbkdf2.zig");
115 _ = @import("crypto/poly1305.zig");123 _ = @import("crypto/poly1305.zig");
116 _ = @import("crypto/sha1.zig");124 _ = @import("crypto/sha1.zig");
lib/std/crypto/aes.zig+72-638
...@@ -3,249 +3,44 @@...@@ -3,249 +3,44 @@
3// This file is part of [zig](https://ziglang.org/), which is MIT licensed.3// This file is part of [zig](https://ziglang.org/), which is MIT licensed.
4// The MIT license requires this copyright notice to be included in all copies4// The MIT license requires this copyright notice to be included in all copies
5// and substantial portions of the software.5// and substantial portions of the software.
6// Based on Go stdlib implementation
76
8const std = @import("../std.zig");7const std = @import("../std.zig");
9const mem = std.mem;
10const testing = std.testing;8const testing = std.testing;
9const builtin = std.builtin;
1110
12// Apply sbox0 to each byte in w.11const has_aesni = comptime std.Target.x86.featureSetHas(std.Target.current.cpu.features, .aes);
13fn subw(w: u32) u32 {12const has_avx = comptime std.Target.x86.featureSetHas(std.Target.current.cpu.features, .avx);
14 return @as(u32, sbox0[w >> 24]) << 24 | @as(u32, sbox0[w >> 16 & 0xff]) << 16 | @as(u32, sbox0[w >> 8 & 0xff]) << 8 | @as(u32, sbox0[w & 0xff]);13const impl = if (std.Target.current.cpu.arch == .x86_64 and has_aesni and has_avx) @import("aes/aesni.zig") else @import("aes/soft.zig");
15}
16
17fn rotw(w: u32) u32 {
18 return w << 8 | w >> 24;
19}
20
21// Encrypt one block from src into dst, using the expanded key xk.
22fn encryptBlock(xk: []const u32, dst: []u8, src: []const u8) void {
23 var s0 = mem.readIntBig(u32, src[0..4]);
24 var s1 = mem.readIntBig(u32, src[4..8]);
25 var s2 = mem.readIntBig(u32, src[8..12]);
26 var s3 = mem.readIntBig(u32, src[12..16]);
27
28 // First round just XORs input with key.
29 s0 ^= xk[0];
30 s1 ^= xk[1];
31 s2 ^= xk[2];
32 s3 ^= xk[3];
33
34 // Middle rounds shuffle using tables.
35 // Number of rounds is set by length of expanded key.
36 var nr = xk.len / 4 - 2; // - 2: one above, one more below
37 var k: usize = 4;
38 var t0: u32 = undefined;
39 var t1: u32 = undefined;
40 var t2: u32 = undefined;
41 var t3: u32 = undefined;
42 var r: usize = 0;
43 while (r < nr) : (r += 1) {
44 t0 = xk[k + 0] ^ te0[@truncate(u8, s0 >> 24)] ^ te1[@truncate(u8, s1 >> 16)] ^ te2[@truncate(u8, s2 >> 8)] ^ te3[@truncate(u8, s3)];
45 t1 = xk[k + 1] ^ te0[@truncate(u8, s1 >> 24)] ^ te1[@truncate(u8, s2 >> 16)] ^ te2[@truncate(u8, s3 >> 8)] ^ te3[@truncate(u8, s0)];
46 t2 = xk[k + 2] ^ te0[@truncate(u8, s2 >> 24)] ^ te1[@truncate(u8, s3 >> 16)] ^ te2[@truncate(u8, s0 >> 8)] ^ te3[@truncate(u8, s1)];
47 t3 = xk[k + 3] ^ te0[@truncate(u8, s3 >> 24)] ^ te1[@truncate(u8, s0 >> 16)] ^ te2[@truncate(u8, s1 >> 8)] ^ te3[@truncate(u8, s2)];
48 k += 4;
49 s0 = t0;
50 s1 = t1;
51 s2 = t2;
52 s3 = t3;
53 }
54
55 // Last round uses s-box directly and XORs to produce output.
56 s0 = @as(u32, sbox0[t0 >> 24]) << 24 | @as(u32, sbox0[t1 >> 16 & 0xff]) << 16 | @as(u32, sbox0[t2 >> 8 & 0xff]) << 8 | @as(u32, sbox0[t3 & 0xff]);
57 s1 = @as(u32, sbox0[t1 >> 24]) << 24 | @as(u32, sbox0[t2 >> 16 & 0xff]) << 16 | @as(u32, sbox0[t3 >> 8 & 0xff]) << 8 | @as(u32, sbox0[t0 & 0xff]);
58 s2 = @as(u32, sbox0[t2 >> 24]) << 24 | @as(u32, sbox0[t3 >> 16 & 0xff]) << 16 | @as(u32, sbox0[t0 >> 8 & 0xff]) << 8 | @as(u32, sbox0[t1 & 0xff]);
59 s3 = @as(u32, sbox0[t3 >> 24]) << 24 | @as(u32, sbox0[t0 >> 16 & 0xff]) << 16 | @as(u32, sbox0[t1 >> 8 & 0xff]) << 8 | @as(u32, sbox0[t2 & 0xff]);
60
61 s0 ^= xk[k + 0];
62 s1 ^= xk[k + 1];
63 s2 ^= xk[k + 2];
64 s3 ^= xk[k + 3];
65
66 mem.writeIntBig(u32, dst[0..4], s0);
67 mem.writeIntBig(u32, dst[4..8], s1);
68 mem.writeIntBig(u32, dst[8..12], s2);
69 mem.writeIntBig(u32, dst[12..16], s3);
70}
71
72// Decrypt one block from src into dst, using the expanded key xk.
73pub fn decryptBlock(xk: []const u32, dst: []u8, src: []const u8) void {
74 var s0 = mem.readIntBig(u32, src[0..4]);
75 var s1 = mem.readIntBig(u32, src[4..8]);
76 var s2 = mem.readIntBig(u32, src[8..12]);
77 var s3 = mem.readIntBig(u32, src[12..16]);
78
79 // First round just XORs input with key.
80 s0 ^= xk[0];
81 s1 ^= xk[1];
82 s2 ^= xk[2];
83 s3 ^= xk[3];
84
85 // Middle rounds shuffle using tables.
86 // Number of rounds is set by length of expanded key.
87 var nr = xk.len / 4 - 2; // - 2: one above, one more below
88 var k: usize = 4;
89 var t0: u32 = undefined;
90 var t1: u32 = undefined;
91 var t2: u32 = undefined;
92 var t3: u32 = undefined;
93 var r: usize = 0;
94 while (r < nr) : (r += 1) {
95 t0 = xk[k + 0] ^ td0[@truncate(u8, s0 >> 24)] ^ td1[@truncate(u8, s3 >> 16)] ^ td2[@truncate(u8, s2 >> 8)] ^ td3[@truncate(u8, s1)];
96 t1 = xk[k + 1] ^ td0[@truncate(u8, s1 >> 24)] ^ td1[@truncate(u8, s0 >> 16)] ^ td2[@truncate(u8, s3 >> 8)] ^ td3[@truncate(u8, s2)];
97 t2 = xk[k + 2] ^ td0[@truncate(u8, s2 >> 24)] ^ td1[@truncate(u8, s1 >> 16)] ^ td2[@truncate(u8, s0 >> 8)] ^ td3[@truncate(u8, s3)];
98 t3 = xk[k + 3] ^ td0[@truncate(u8, s3 >> 24)] ^ td1[@truncate(u8, s2 >> 16)] ^ td2[@truncate(u8, s1 >> 8)] ^ td3[@truncate(u8, s0)];
99 k += 4;
100 s0 = t0;
101 s1 = t1;
102 s2 = t2;
103 s3 = t3;
104 }
105
106 // Last round uses s-box directly and XORs to produce output.
107 s0 = @as(u32, sbox1[t0 >> 24]) << 24 | @as(u32, sbox1[t3 >> 16 & 0xff]) << 16 | @as(u32, sbox1[t2 >> 8 & 0xff]) << 8 | @as(u32, sbox1[t1 & 0xff]);
108 s1 = @as(u32, sbox1[t1 >> 24]) << 24 | @as(u32, sbox1[t0 >> 16 & 0xff]) << 16 | @as(u32, sbox1[t3 >> 8 & 0xff]) << 8 | @as(u32, sbox1[t2 & 0xff]);
109 s2 = @as(u32, sbox1[t2 >> 24]) << 24 | @as(u32, sbox1[t1 >> 16 & 0xff]) << 16 | @as(u32, sbox1[t0 >> 8 & 0xff]) << 8 | @as(u32, sbox1[t3 & 0xff]);
110 s3 = @as(u32, sbox1[t3 >> 24]) << 24 | @as(u32, sbox1[t2 >> 16 & 0xff]) << 16 | @as(u32, sbox1[t1 >> 8 & 0xff]) << 8 | @as(u32, sbox1[t0 & 0xff]);
111
112 s0 ^= xk[k + 0];
113 s1 ^= xk[k + 1];
114 s2 ^= xk[k + 2];
115 s3 ^= xk[k + 3];
116
117 mem.writeIntBig(u32, dst[0..4], s0);
118 mem.writeIntBig(u32, dst[4..8], s1);
119 mem.writeIntBig(u32, dst[8..12], s2);
120 mem.writeIntBig(u32, dst[12..16], s3);
121}
122
123fn xorBytes(dst: []u8, a: []const u8, b: []const u8) usize {
124 var n = std.math.min(dst.len, std.math.min(a.len, b.len));
125 for (dst[0..n]) |_, i| {
126 dst[i] = a[i] ^ b[i];
127 }
128 return n;
129}
130
131pub const AES128 = AES(128);
132pub const AES256 = AES(256);
133
134fn AES(comptime keysize: usize) type {
135 return struct {
136 const Self = @This();
137
138 pub const Encrypt = AESEncrypt(keysize);
139 pub const Decrypt = AESDecrypt(keysize);
140
141 const nn = (keysize / 8) + 28;
142 enc: Encrypt,
143 dec: Decrypt,
144
145 pub fn init(key: [keysize / 8]u8) Self {
146 var ctx: Self = undefined;
147 ctx.enc = Encrypt.init(key);
148 ctx.dec = ctx.enc.toDecrypt();
149 return ctx;
150 }
15114
152 pub fn encrypt(ctx: Self, dst: []u8, src: []const u8) void {15pub const Block = impl.Block;
153 ctx.enc.encrypt(dst, src);16pub const AESEncryptCtx = impl.AESEncryptCtx;
154 }17pub const AESDecryptCtx = impl.AESDecryptCtx;
155 pub fn decrypt(ctx: Self, dst: []u8, src: []const u8) void {18pub const AES128 = impl.AES128;
156 ctx.dec.decrypt(dst, src);19pub const AES256 = impl.AES256;
157 }
158 pub fn ctr(ctx: Self, dst: []u8, src: []const u8, iv: [16]u8) void {
159 ctx.enc.ctr(dst, src, iv);
160 }
161 };
162}
163
164fn AESEncrypt(comptime keysize: usize) type {
165 return struct {
166 const Self = @This();
167
168 const Decrypt = AESDecrypt(keysize);
169
170 const nn = (keysize / 8) + 28;
171 enc: [nn]u32,
17220
173 pub fn init(key: [keysize / 8]u8) Self {21test "ctr" {
174 var ctx: Self = undefined;22 // NIST SP 800-38A pp 55-58
175 expandKeyEncrypt(&key, ctx.enc[0..]);23 const ctr = @import("modes.zig").ctr;
176 return ctx;
177 }
178
179 pub fn toDecrypt(ctx: Self) Decrypt {
180 var dec: Decrypt = undefined;
181 expandKeyDecrypt(ctx.enc[0..], dec.dec[0..]);
182 return dec;
183 }
184
185 pub fn encrypt(ctx: Self, dst: []u8, src: []const u8) void {
186 encryptBlock(ctx.enc[0..], dst, src);
187 }
188 pub fn ctr(ctx: Self, dst: []u8, src: []const u8, iv: [16]u8) void {
189 std.debug.assert(dst.len >= src.len);
190
191 var keystream: [16]u8 = undefined;
192 var ctrbuf = iv;
193 var n: usize = 0;
194 while (n < src.len) {
195 ctx.encrypt(keystream[0..], ctrbuf[0..]);
196 var ctr_i = std.mem.readIntBig(u128, ctrbuf[0..]);
197 std.mem.writeIntBig(u128, ctrbuf[0..], ctr_i +% 1);
19824
199 n += xorBytes(dst[n..], src[n..], &keystream);25 const key = [_]u8{ 0x2b, 0x7e, 0x15, 0x16, 0x28, 0xae, 0xd2, 0xa6, 0xab, 0xf7, 0x15, 0x88, 0x09, 0xcf, 0x4f, 0x3c };
200 }26 const iv = [_]u8{ 0xf0, 0xf1, 0xf2, 0xf3, 0xf4, 0xf5, 0xf6, 0xf7, 0xf8, 0xf9, 0xfa, 0xfb, 0xfc, 0xfd, 0xfe, 0xff };
201 }27 const in = [_]u8{
28 0x6b, 0xc1, 0xbe, 0xe2, 0x2e, 0x40, 0x9f, 0x96, 0xe9, 0x3d, 0x7e, 0x11, 0x73, 0x93, 0x17, 0x2a,
29 0xae, 0x2d, 0x8a, 0x57, 0x1e, 0x03, 0xac, 0x9c, 0x9e, 0xb7, 0x6f, 0xac, 0x45, 0xaf, 0x8e, 0x51,
30 0x30, 0xc8, 0x1c, 0x46, 0xa3, 0x5c, 0xe4, 0x11, 0xe5, 0xfb, 0xc1, 0x19, 0x1a, 0x0a, 0x52, 0xef,
31 0xf6, 0x9f, 0x24, 0x45, 0xdf, 0x4f, 0x9b, 0x17, 0xad, 0x2b, 0x41, 0x7b, 0xe6, 0x6c, 0x37, 0x10,
202 };32 };
203}33 const exp_out = [_]u8{
20434 0x87, 0x4d, 0x61, 0x91, 0xb6, 0x20, 0xe3, 0x26, 0x1b, 0xef, 0x68, 0x64, 0x99, 0x0d, 0xb6, 0xce,
205fn AESDecrypt(comptime keysize: usize) type {35 0x98, 0x06, 0xf6, 0x6b, 0x79, 0x70, 0xfd, 0xff, 0x86, 0x17, 0x18, 0x7b, 0xb9, 0xff, 0xfd, 0xff,
206 return struct {36 0x5a, 0xe4, 0xdf, 0x3e, 0xdb, 0xd5, 0xd3, 0x5e, 0x5b, 0x4f, 0x09, 0x02, 0x0d, 0xb0, 0x3e, 0xab,
207 const Self = @This();37 0x1e, 0x03, 0x1d, 0xda, 0x2f, 0xbe, 0x03, 0xd1, 0x79, 0x21, 0x70, 0xa0, 0xf3, 0x00, 0x9c, 0xee,
208
209 const nn = (keysize / 8) + 28;
210 dec: [nn]u32,
211
212 pub fn init(key: [keysize / 8]u8) Self {
213 var ctx: Self = undefined;
214 var enc: [nn]u32 = undefined;
215 expandKeyEncrypt(key[0..], enc[0..]);
216 expandKeyDecrypt(enc[0..], ctx.dec[0..]);
217 return ctx;
218 }
219
220 pub fn decrypt(ctx: Self, dst: []u8, src: []const u8) void {
221 decryptBlock(ctx.dec[0..], dst, src);
222 }
223 };38 };
224}
225
226test "ctr" {
227 // NIST SP 800-38A pp 55-58
228 {
229 const key = [_]u8{ 0x2b, 0x7e, 0x15, 0x16, 0x28, 0xae, 0xd2, 0xa6, 0xab, 0xf7, 0x15, 0x88, 0x09, 0xcf, 0x4f, 0x3c };
230 const iv = [_]u8{ 0xf0, 0xf1, 0xf2, 0xf3, 0xf4, 0xf5, 0xf6, 0xf7, 0xf8, 0xf9, 0xfa, 0xfb, 0xfc, 0xfd, 0xfe, 0xff };
231 const in = [_]u8{
232 0x6b, 0xc1, 0xbe, 0xe2, 0x2e, 0x40, 0x9f, 0x96, 0xe9, 0x3d, 0x7e, 0x11, 0x73, 0x93, 0x17, 0x2a,
233 0xae, 0x2d, 0x8a, 0x57, 0x1e, 0x03, 0xac, 0x9c, 0x9e, 0xb7, 0x6f, 0xac, 0x45, 0xaf, 0x8e, 0x51,
234 0x30, 0xc8, 0x1c, 0x46, 0xa3, 0x5c, 0xe4, 0x11, 0xe5, 0xfb, 0xc1, 0x19, 0x1a, 0x0a, 0x52, 0xef,
235 0xf6, 0x9f, 0x24, 0x45, 0xdf, 0x4f, 0x9b, 0x17, 0xad, 0x2b, 0x41, 0x7b, 0xe6, 0x6c, 0x37, 0x10,
236 };
237 const exp_out = [_]u8{
238 0x87, 0x4d, 0x61, 0x91, 0xb6, 0x20, 0xe3, 0x26, 0x1b, 0xef, 0x68, 0x64, 0x99, 0x0d, 0xb6, 0xce,
239 0x98, 0x06, 0xf6, 0x6b, 0x79, 0x70, 0xfd, 0xff, 0x86, 0x17, 0x18, 0x7b, 0xb9, 0xff, 0xfd, 0xff,
240 0x5a, 0xe4, 0xdf, 0x3e, 0xdb, 0xd5, 0xd3, 0x5e, 0x5b, 0x4f, 0x09, 0x02, 0x0d, 0xb0, 0x3e, 0xab,
241 0x1e, 0x03, 0x1d, 0xda, 0x2f, 0xbe, 0x03, 0xd1, 0x79, 0x21, 0x70, 0xa0, 0xf3, 0x00, 0x9c, 0xee,
242 };
24339
244 var out: [exp_out.len]u8 = undefined;40 var out: [exp_out.len]u8 = undefined;
245 var aes = AES128.init(key);41 var ctx = AES128.initEnc(key);
246 aes.ctr(out[0..], in[0..], iv);42 ctr(AESEncryptCtx(AES128), ctx, out[0..], in[0..], iv, builtin.Endian.Big);
247 testing.expectEqualSlices(u8, exp_out[0..], out[0..]);43 testing.expectEqualSlices(u8, exp_out[0..], out[0..]);
248 }
249}44}
25045
251test "encrypt" {46test "encrypt" {
...@@ -256,8 +51,8 @@ test "encrypt" {...@@ -256,8 +51,8 @@ test "encrypt" {
256 const exp_out = [_]u8{ 0x39, 0x25, 0x84, 0x1d, 0x02, 0xdc, 0x09, 0xfb, 0xdc, 0x11, 0x85, 0x97, 0x19, 0x6a, 0x0b, 0x32 };51 const exp_out = [_]u8{ 0x39, 0x25, 0x84, 0x1d, 0x02, 0xdc, 0x09, 0xfb, 0xdc, 0x11, 0x85, 0x97, 0x19, 0x6a, 0x0b, 0x32 };
25752
258 var out: [exp_out.len]u8 = undefined;53 var out: [exp_out.len]u8 = undefined;
259 var aes = AES128.init(key);54 var ctx = AES128.initEnc(key);
260 aes.encrypt(out[0..], in[0..]);55 ctx.encrypt(out[0..], in[0..]);
261 testing.expectEqualSlices(u8, exp_out[0..], out[0..]);56 testing.expectEqualSlices(u8, exp_out[0..], out[0..]);
262 }57 }
26358
...@@ -271,8 +66,8 @@ test "encrypt" {...@@ -271,8 +66,8 @@ test "encrypt" {
271 const exp_out = [_]u8{ 0x8e, 0xa2, 0xb7, 0xca, 0x51, 0x67, 0x45, 0xbf, 0xea, 0xfc, 0x49, 0x90, 0x4b, 0x49, 0x60, 0x89 };66 const exp_out = [_]u8{ 0x8e, 0xa2, 0xb7, 0xca, 0x51, 0x67, 0x45, 0xbf, 0xea, 0xfc, 0x49, 0x90, 0x4b, 0x49, 0x60, 0x89 };
27267
273 var out: [exp_out.len]u8 = undefined;68 var out: [exp_out.len]u8 = undefined;
274 var aes = AES256.init(key);69 var ctx = AES256.initEnc(key);
275 aes.encrypt(out[0..], in[0..]);70 ctx.encrypt(out[0..], in[0..]);
276 testing.expectEqualSlices(u8, exp_out[0..], out[0..]);71 testing.expectEqualSlices(u8, exp_out[0..], out[0..]);
277 }72 }
278}73}
...@@ -285,8 +80,8 @@ test "decrypt" {...@@ -285,8 +80,8 @@ test "decrypt" {
285 const exp_out = [_]u8{ 0x32, 0x43, 0xf6, 0xa8, 0x88, 0x5a, 0x30, 0x8d, 0x31, 0x31, 0x98, 0xa2, 0xe0, 0x37, 0x07, 0x34 };80 const exp_out = [_]u8{ 0x32, 0x43, 0xf6, 0xa8, 0x88, 0x5a, 0x30, 0x8d, 0x31, 0x31, 0x98, 0xa2, 0xe0, 0x37, 0x07, 0x34 };
28681
287 var out: [exp_out.len]u8 = undefined;82 var out: [exp_out.len]u8 = undefined;
288 var aes = AES128.init(key);83 var ctx = AES128.initDec(key);
289 aes.decrypt(out[0..], in[0..]);84 ctx.decrypt(out[0..], in[0..]);
290 testing.expectEqualSlices(u8, exp_out[0..], out[0..]);85 testing.expectEqualSlices(u8, exp_out[0..], out[0..]);
291 }86 }
29287
...@@ -300,413 +95,52 @@ test "decrypt" {...@@ -300,413 +95,52 @@ test "decrypt" {
300 const exp_out = [_]u8{ 0x00, 0x11, 0x22, 0x33, 0x44, 0x55, 0x66, 0x77, 0x88, 0x99, 0xaa, 0xbb, 0xcc, 0xdd, 0xee, 0xff };95 const exp_out = [_]u8{ 0x00, 0x11, 0x22, 0x33, 0x44, 0x55, 0x66, 0x77, 0x88, 0x99, 0xaa, 0xbb, 0xcc, 0xdd, 0xee, 0xff };
30196
302 var out: [exp_out.len]u8 = undefined;97 var out: [exp_out.len]u8 = undefined;
303 var aes = AES256.init(key);98 var ctx = AES256.initDec(key);
304 aes.decrypt(out[0..], in[0..]);99 ctx.decrypt(out[0..], in[0..]);
305 testing.expectEqualSlices(u8, exp_out[0..], out[0..]);100 testing.expectEqualSlices(u8, exp_out[0..], out[0..]);
306 }101 }
307}102}
308103
309// Key expansion algorithm. See FIPS-197, Figure 11.104test "expand 128-bit key" {
310fn expandKeyEncrypt(key: []const u8, enc: []u32) void {105 const key = [_]u8{ 0x2b, 0x7e, 0x15, 0x16, 0x28, 0xae, 0xd2, 0xa6, 0xab, 0xf7, 0x15, 0x88, 0x09, 0xcf, 0x4f, 0x3c };
311 var i: usize = 0;106 const exp_enc = [_]*const [32:0]u8{
312 var nk = key.len / 4;107 "2b7e151628aed2a6abf7158809cf4f3c", "a0fafe1788542cb123a339392a6c7605", "f2c295f27a96b9435935807a7359f67f", "3d80477d4716fe3e1e237e446d7a883b", "ef44a541a8525b7fb671253bdb0bad00", "d4d1c6f87c839d87caf2b8bc11f915bc", "6d88a37a110b3efddbf98641ca0093fd", "4e54f70e5f5fc9f384a64fb24ea6dc4f", "ead27321b58dbad2312bf5607f8d292f", "ac7766f319fadc2128d12941575c006e", "d014f9a8c9ee2589e13f0cc8b6630ca6",
313 while (i < nk) : (i += 1) {108 };
314 enc[i] = mem.readIntBig(u32, key[4 * i ..][0..4]);109 const exp_dec = [_]*const [32:0]u8{
315 }110 "2b7e151628aed2a6abf7158809cf4f3c", "a0fafe1788542cb123a339392a6c7605", "f2c295f27a96b9435935807a7359f67f", "3d80477d4716fe3e1e237e446d7a883b", "ef44a541a8525b7fb671253bdb0bad00", "d4d1c6f87c839d87caf2b8bc11f915bc", "6d88a37a110b3efddbf98641ca0093fd", "4e54f70e5f5fc9f384a64fb24ea6dc4f", "ead27321b58dbad2312bf5607f8d292f", "ac7766f319fadc2128d12941575c006e", "d014f9a8c9ee2589e13f0cc8b6630ca6",
316 while (i < enc.len) : (i += 1) {111 };
317 var t = enc[i - 1];112 const enc = AES128.initEnc(key);
318 if (i % nk == 0) {113 const dec = AES128.initDec(key);
319 t = subw(rotw(t)) ^ (@as(u32, powx[i / nk - 1]) << 24);114 var exp: [16]u8 = undefined;
320 } else if (nk > 6 and i % nk == 4) {
321 t = subw(t);
322 }
323 enc[i] = enc[i - nk] ^ t;
324 }
325}
326115
327fn expandKeyDecrypt(enc: []const u32, dec: []u32) void {116 for (enc.key_schedule.round_keys) |round_key, i| {
328 var i: usize = 0;117 try std.fmt.hexToBytes(&exp, exp_enc[i]);
329 var n = enc.len;118 testing.expectEqualSlices(u8, &exp, &round_key.toBytes());
330 while (i < n) : (i += 4) {119 }
331 var ei = n - i - 4;120 for (enc.key_schedule.round_keys) |round_key, i| {
332 var j: usize = 0;121 try std.fmt.hexToBytes(&exp, exp_dec[i]);
333 while (j < 4) : (j += 1) {122 testing.expectEqualSlices(u8, &exp, &round_key.toBytes());
334 var x = enc[ei + j];
335 if (i > 0 and i + 4 < n) {
336 x = td0[sbox0[x >> 24]] ^ td1[sbox0[x >> 16 & 0xff]] ^ td2[sbox0[x >> 8 & 0xff]] ^ td3[sbox0[x & 0xff]];
337 }
338 dec[i + j] = x;
339 }
340 }123 }
341}124}
342125
343test "expand key" {126test "expand 256-bit key" {
344 const key = [_]u8{ 0x2b, 0x7e, 0x15, 0x16, 0x28, 0xae, 0xd2, 0xa6, 0xab, 0xf7, 0x15, 0x88, 0x09, 0xcf, 0x4f, 0x3c };127 const key = [_]u8{ 0x60, 0x3d, 0xeb, 0x10, 0x15, 0xca, 0x71, 0xbe, 0x2b, 0x73, 0xae, 0xf0, 0x85, 0x7d, 0x77, 0x81, 0x1f, 0x35, 0x2c, 0x07, 0x3b, 0x61, 0x08, 0xd7, 0x2d, 0x98, 0x10, 0xa3, 0x09, 0x14, 0xdf, 0xf4 };
345 const exp_enc = [_]u32{128 const exp_enc = [_]*const [32:0]u8{
346 0x2b7e1516, 0x28aed2a6, 0xabf71588, 0x09cf4f3c,129 "603deb1015ca71be2b73aef0857d7781", "1f352c073b6108d72d9810a30914dff4", "9ba354118e6925afa51a8b5f2067fcde", "a8b09c1a93d194cdbe49846eb75d5b9a", "d59aecb85bf3c917fee94248de8ebe96", "b5a9328a2678a647983122292f6c79b3", "812c81addadf48ba24360af2fab8b464", "98c5bfc9bebd198e268c3ba709e04214", "68007bacb2df331696e939e46c518d80", "c814e20476a9fb8a5025c02d59c58239", "de1369676ccc5a71fa2563959674ee15", "5886ca5d2e2f31d77e0af1fa27cf73c3", "749c47ab18501ddae2757e4f7401905a", "cafaaae3e4d59b349adf6acebd10190d", "fe4890d1e6188d0b046df344706c631e",
347 0xa0fafe17, 0x88542cb1, 0x23a33939, 0x2a6c7605,
348 0xf2c295f2, 0x7a96b943, 0x5935807a, 0x7359f67f,
349 0x3d80477d, 0x4716fe3e, 0x1e237e44, 0x6d7a883b,
350 0xef44a541, 0xa8525b7f, 0xb671253b, 0xdb0bad00,
351 0xd4d1c6f8, 0x7c839d87, 0xcaf2b8bc, 0x11f915bc,
352 0x6d88a37a, 0x110b3efd, 0xdbf98641, 0xca0093fd,
353 0x4e54f70e, 0x5f5fc9f3, 0x84a64fb2, 0x4ea6dc4f,
354 0xead27321, 0xb58dbad2, 0x312bf560, 0x7f8d292f,
355 0xac7766f3, 0x19fadc21, 0x28d12941, 0x575c006e,
356 0xd014f9a8, 0xc9ee2589, 0xe13f0cc8, 0xb6630ca6,
357 };130 };
358 const exp_dec = [_]u32{131 const exp_dec = [_]*const [32:0]u8{
359 0xd014f9a8, 0xc9ee2589, 0xe13f0cc8, 0xb6630ca6,132 "fe4890d1e6188d0b046df344706c631e", "ada23f4963e23b2455427c8a5c709104", "57c96cf6074f07c0706abb07137f9241", "b668b621ce40046d36a047ae0932ed8e", "34ad1e4450866b367725bcc763152946", "32526c367828b24cf8e043c33f92aa20", "c440b289642b757227a3d7f114309581", "d669a7334a7ade7a80c8f18fc772e9e3", "25ba3c22a06bc7fb4388a28333934270", "54fb808b9c137949cab22ff547ba186c", "6c3d632985d1fbd9e3e36578701be0f3", "4a7459f9c8e8f9c256a156bc8d083799", "42107758e9ec98f066329ea193f8858b", "8ec6bff6829ca03b9e49af7edba96125", "603deb1015ca71be2b73aef0857d7781",
360 0xc7b5a63, 0x1319eafe, 0xb0398890, 0x664cfbb4,
361 0xdf7d925a, 0x1f62b09d, 0xa320626e, 0xd6757324,
362 0x12c07647, 0xc01f22c7, 0xbc42d2f3, 0x7555114a,
363 0x6efcd876, 0xd2df5480, 0x7c5df034, 0xc917c3b9,
364 0x6ea30afc, 0xbc238cf6, 0xae82a4b4, 0xb54a338d,
365 0x90884413, 0xd280860a, 0x12a12842, 0x1bc89739,
366 0x7c1f13f7, 0x4208c219, 0xc021ae48, 0x969bf7b,
367 0xcc7505eb, 0x3e17d1ee, 0x82296c51, 0xc9481133,
368 0x2b3708a7, 0xf262d405, 0xbc3ebdbf, 0x4b617d62,
369 0x2b7e1516, 0x28aed2a6, 0xabf71588, 0x9cf4f3c,
370 };133 };
371 var enc: [exp_enc.len]u32 = undefined;134 const enc = AES256.initEnc(key);
372 var dec: [exp_dec.len]u32 = undefined;135 const dec = AES256.initDec(key);
373 expandKeyEncrypt(key[0..], enc[0..]);136 var exp: [16]u8 = undefined;
374 expandKeyDecrypt(enc[0..], dec[0..]);
375 testing.expectEqualSlices(u32, exp_enc[0..], enc[0..]);
376 testing.expectEqualSlices(u32, exp_dec[0..], dec[0..]);
377}
378137
379// constants138 for (enc.key_schedule.round_keys) |round_key, i| {
380139 try std.fmt.hexToBytes(&exp, exp_enc[i]);
381const poly = 1 << 8 | 1 << 4 | 1 << 3 | 1 << 1 | 1 << 0;140 testing.expectEqualSlices(u8, &exp, &round_key.toBytes());
382141 }
383const powx = [16]u8{142 for (dec.key_schedule.round_keys) |round_key, i| {
384 0x01,143 try std.fmt.hexToBytes(&exp, exp_dec[i]);
385 0x02,144 testing.expectEqualSlices(u8, &exp, &round_key.toBytes());
386 0x04,145 }
387 0x08,146}
388 0x10,
389 0x20,
390 0x40,
391 0x80,
392 0x1b,
393 0x36,
394 0x6c,
395 0xd8,
396 0xab,
397 0x4d,
398 0x9a,
399 0x2f,
400};
401
402const sbox0 = [256]u8{
403 0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5, 0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76,
404 0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0, 0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0,
405 0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc, 0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15,
406 0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a, 0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75,
407 0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0, 0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84,
408 0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b, 0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf,
409 0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85, 0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8,
410 0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5, 0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2,
411 0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17, 0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73,
412 0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88, 0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb,
413 0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c, 0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79,
414 0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9, 0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08,
415 0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6, 0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a,
416 0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e, 0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e,
417 0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94, 0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf,
418 0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68, 0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16,
419};
420
421const sbox1 = [256]u8{
422 0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38, 0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb,
423 0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87, 0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb,
424 0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d, 0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e,
425 0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2, 0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25,
426 0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16, 0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92,
427 0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda, 0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84,
428 0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a, 0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06,
429 0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02, 0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b,
430 0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea, 0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73,
431 0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85, 0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e,
432 0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89, 0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b,
433 0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20, 0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4,
434 0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31, 0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f,
435 0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d, 0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef,
436 0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0, 0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61,
437 0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26, 0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d,
438};
439
440const te0 = [256]u32{
441 0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d, 0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554,
442 0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d, 0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a,
443 0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87, 0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b,
444 0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea, 0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b,
445 0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a, 0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f,
446 0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108, 0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f,
447 0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e, 0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5,
448 0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d, 0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f,
449 0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e, 0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb,
450 0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce, 0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497,
451 0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c, 0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed,
452 0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b, 0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a,
453 0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16, 0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594,
454 0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81, 0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3,
455 0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a, 0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504,
456 0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163, 0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d,
457 0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f, 0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739,
458 0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47, 0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395,
459 0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f, 0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883,
460 0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c, 0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76,
461 0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e, 0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4,
462 0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6, 0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b,
463 0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7, 0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0,
464 0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25, 0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818,
465 0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72, 0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651,
466 0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21, 0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85,
467 0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa, 0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12,
468 0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0, 0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9,
469 0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133, 0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7,
470 0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920, 0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a,
471 0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17, 0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8,
472 0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11, 0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a,
473};
474const te1 = [256]u32{
475 0xa5c66363, 0x84f87c7c, 0x99ee7777, 0x8df67b7b, 0x0dfff2f2, 0xbdd66b6b, 0xb1de6f6f, 0x5491c5c5,
476 0x50603030, 0x03020101, 0xa9ce6767, 0x7d562b2b, 0x19e7fefe, 0x62b5d7d7, 0xe64dabab, 0x9aec7676,
477 0x458fcaca, 0x9d1f8282, 0x4089c9c9, 0x87fa7d7d, 0x15effafa, 0xebb25959, 0xc98e4747, 0x0bfbf0f0,
478 0xec41adad, 0x67b3d4d4, 0xfd5fa2a2, 0xea45afaf, 0xbf239c9c, 0xf753a4a4, 0x96e47272, 0x5b9bc0c0,
479 0xc275b7b7, 0x1ce1fdfd, 0xae3d9393, 0x6a4c2626, 0x5a6c3636, 0x417e3f3f, 0x02f5f7f7, 0x4f83cccc,
480 0x5c683434, 0xf451a5a5, 0x34d1e5e5, 0x08f9f1f1, 0x93e27171, 0x73abd8d8, 0x53623131, 0x3f2a1515,
481 0x0c080404, 0x5295c7c7, 0x65462323, 0x5e9dc3c3, 0x28301818, 0xa1379696, 0x0f0a0505, 0xb52f9a9a,
482 0x090e0707, 0x36241212, 0x9b1b8080, 0x3ddfe2e2, 0x26cdebeb, 0x694e2727, 0xcd7fb2b2, 0x9fea7575,
483 0x1b120909, 0x9e1d8383, 0x74582c2c, 0x2e341a1a, 0x2d361b1b, 0xb2dc6e6e, 0xeeb45a5a, 0xfb5ba0a0,
484 0xf6a45252, 0x4d763b3b, 0x61b7d6d6, 0xce7db3b3, 0x7b522929, 0x3edde3e3, 0x715e2f2f, 0x97138484,
485 0xf5a65353, 0x68b9d1d1, 0x00000000, 0x2cc1eded, 0x60402020, 0x1fe3fcfc, 0xc879b1b1, 0xedb65b5b,
486 0xbed46a6a, 0x468dcbcb, 0xd967bebe, 0x4b723939, 0xde944a4a, 0xd4984c4c, 0xe8b05858, 0x4a85cfcf,
487 0x6bbbd0d0, 0x2ac5efef, 0xe54faaaa, 0x16edfbfb, 0xc5864343, 0xd79a4d4d, 0x55663333, 0x94118585,
488 0xcf8a4545, 0x10e9f9f9, 0x06040202, 0x81fe7f7f, 0xf0a05050, 0x44783c3c, 0xba259f9f, 0xe34ba8a8,
489 0xf3a25151, 0xfe5da3a3, 0xc0804040, 0x8a058f8f, 0xad3f9292, 0xbc219d9d, 0x48703838, 0x04f1f5f5,
490 0xdf63bcbc, 0xc177b6b6, 0x75afdada, 0x63422121, 0x30201010, 0x1ae5ffff, 0x0efdf3f3, 0x6dbfd2d2,
491 0x4c81cdcd, 0x14180c0c, 0x35261313, 0x2fc3ecec, 0xe1be5f5f, 0xa2359797, 0xcc884444, 0x392e1717,
492 0x5793c4c4, 0xf255a7a7, 0x82fc7e7e, 0x477a3d3d, 0xacc86464, 0xe7ba5d5d, 0x2b321919, 0x95e67373,
493 0xa0c06060, 0x98198181, 0xd19e4f4f, 0x7fa3dcdc, 0x66442222, 0x7e542a2a, 0xab3b9090, 0x830b8888,
494 0xca8c4646, 0x29c7eeee, 0xd36bb8b8, 0x3c281414, 0x79a7dede, 0xe2bc5e5e, 0x1d160b0b, 0x76addbdb,
495 0x3bdbe0e0, 0x56643232, 0x4e743a3a, 0x1e140a0a, 0xdb924949, 0x0a0c0606, 0x6c482424, 0xe4b85c5c,
496 0x5d9fc2c2, 0x6ebdd3d3, 0xef43acac, 0xa6c46262, 0xa8399191, 0xa4319595, 0x37d3e4e4, 0x8bf27979,
497 0x32d5e7e7, 0x438bc8c8, 0x596e3737, 0xb7da6d6d, 0x8c018d8d, 0x64b1d5d5, 0xd29c4e4e, 0xe049a9a9,
498 0xb4d86c6c, 0xfaac5656, 0x07f3f4f4, 0x25cfeaea, 0xafca6565, 0x8ef47a7a, 0xe947aeae, 0x18100808,
499 0xd56fbaba, 0x88f07878, 0x6f4a2525, 0x725c2e2e, 0x24381c1c, 0xf157a6a6, 0xc773b4b4, 0x5197c6c6,
500 0x23cbe8e8, 0x7ca1dddd, 0x9ce87474, 0x213e1f1f, 0xdd964b4b, 0xdc61bdbd, 0x860d8b8b, 0x850f8a8a,
501 0x90e07070, 0x427c3e3e, 0xc471b5b5, 0xaacc6666, 0xd8904848, 0x05060303, 0x01f7f6f6, 0x121c0e0e,
502 0xa3c26161, 0x5f6a3535, 0xf9ae5757, 0xd069b9b9, 0x91178686, 0x5899c1c1, 0x273a1d1d, 0xb9279e9e,
503 0x38d9e1e1, 0x13ebf8f8, 0xb32b9898, 0x33221111, 0xbbd26969, 0x70a9d9d9, 0x89078e8e, 0xa7339494,
504 0xb62d9b9b, 0x223c1e1e, 0x92158787, 0x20c9e9e9, 0x4987cece, 0xffaa5555, 0x78502828, 0x7aa5dfdf,
505 0x8f038c8c, 0xf859a1a1, 0x80098989, 0x171a0d0d, 0xda65bfbf, 0x31d7e6e6, 0xc6844242, 0xb8d06868,
506 0xc3824141, 0xb0299999, 0x775a2d2d, 0x111e0f0f, 0xcb7bb0b0, 0xfca85454, 0xd66dbbbb, 0x3a2c1616,
507};
508const te2 = [256]u32{
509 0x63a5c663, 0x7c84f87c, 0x7799ee77, 0x7b8df67b, 0xf20dfff2, 0x6bbdd66b, 0x6fb1de6f, 0xc55491c5,
510 0x30506030, 0x01030201, 0x67a9ce67, 0x2b7d562b, 0xfe19e7fe, 0xd762b5d7, 0xabe64dab, 0x769aec76,
511 0xca458fca, 0x829d1f82, 0xc94089c9, 0x7d87fa7d, 0xfa15effa, 0x59ebb259, 0x47c98e47, 0xf00bfbf0,
512 0xadec41ad, 0xd467b3d4, 0xa2fd5fa2, 0xafea45af, 0x9cbf239c, 0xa4f753a4, 0x7296e472, 0xc05b9bc0,
513 0xb7c275b7, 0xfd1ce1fd, 0x93ae3d93, 0x266a4c26, 0x365a6c36, 0x3f417e3f, 0xf702f5f7, 0xcc4f83cc,
514 0x345c6834, 0xa5f451a5, 0xe534d1e5, 0xf108f9f1, 0x7193e271, 0xd873abd8, 0x31536231, 0x153f2a15,
515 0x040c0804, 0xc75295c7, 0x23654623, 0xc35e9dc3, 0x18283018, 0x96a13796, 0x050f0a05, 0x9ab52f9a,
516 0x07090e07, 0x12362412, 0x809b1b80, 0xe23ddfe2, 0xeb26cdeb, 0x27694e27, 0xb2cd7fb2, 0x759fea75,
517 0x091b1209, 0x839e1d83, 0x2c74582c, 0x1a2e341a, 0x1b2d361b, 0x6eb2dc6e, 0x5aeeb45a, 0xa0fb5ba0,
518 0x52f6a452, 0x3b4d763b, 0xd661b7d6, 0xb3ce7db3, 0x297b5229, 0xe33edde3, 0x2f715e2f, 0x84971384,
519 0x53f5a653, 0xd168b9d1, 0x00000000, 0xed2cc1ed, 0x20604020, 0xfc1fe3fc, 0xb1c879b1, 0x5bedb65b,
520 0x6abed46a, 0xcb468dcb, 0xbed967be, 0x394b7239, 0x4ade944a, 0x4cd4984c, 0x58e8b058, 0xcf4a85cf,
521 0xd06bbbd0, 0xef2ac5ef, 0xaae54faa, 0xfb16edfb, 0x43c58643, 0x4dd79a4d, 0x33556633, 0x85941185,
522 0x45cf8a45, 0xf910e9f9, 0x02060402, 0x7f81fe7f, 0x50f0a050, 0x3c44783c, 0x9fba259f, 0xa8e34ba8,
523 0x51f3a251, 0xa3fe5da3, 0x40c08040, 0x8f8a058f, 0x92ad3f92, 0x9dbc219d, 0x38487038, 0xf504f1f5,
524 0xbcdf63bc, 0xb6c177b6, 0xda75afda, 0x21634221, 0x10302010, 0xff1ae5ff, 0xf30efdf3, 0xd26dbfd2,
525 0xcd4c81cd, 0x0c14180c, 0x13352613, 0xec2fc3ec, 0x5fe1be5f, 0x97a23597, 0x44cc8844, 0x17392e17,
526 0xc45793c4, 0xa7f255a7, 0x7e82fc7e, 0x3d477a3d, 0x64acc864, 0x5de7ba5d, 0x192b3219, 0x7395e673,
527 0x60a0c060, 0x81981981, 0x4fd19e4f, 0xdc7fa3dc, 0x22664422, 0x2a7e542a, 0x90ab3b90, 0x88830b88,
528 0x46ca8c46, 0xee29c7ee, 0xb8d36bb8, 0x143c2814, 0xde79a7de, 0x5ee2bc5e, 0x0b1d160b, 0xdb76addb,
529 0xe03bdbe0, 0x32566432, 0x3a4e743a, 0x0a1e140a, 0x49db9249, 0x060a0c06, 0x246c4824, 0x5ce4b85c,
530 0xc25d9fc2, 0xd36ebdd3, 0xacef43ac, 0x62a6c462, 0x91a83991, 0x95a43195, 0xe437d3e4, 0x798bf279,
531 0xe732d5e7, 0xc8438bc8, 0x37596e37, 0x6db7da6d, 0x8d8c018d, 0xd564b1d5, 0x4ed29c4e, 0xa9e049a9,
532 0x6cb4d86c, 0x56faac56, 0xf407f3f4, 0xea25cfea, 0x65afca65, 0x7a8ef47a, 0xaee947ae, 0x08181008,
533 0xbad56fba, 0x7888f078, 0x256f4a25, 0x2e725c2e, 0x1c24381c, 0xa6f157a6, 0xb4c773b4, 0xc65197c6,
534 0xe823cbe8, 0xdd7ca1dd, 0x749ce874, 0x1f213e1f, 0x4bdd964b, 0xbddc61bd, 0x8b860d8b, 0x8a850f8a,
535 0x7090e070, 0x3e427c3e, 0xb5c471b5, 0x66aacc66, 0x48d89048, 0x03050603, 0xf601f7f6, 0x0e121c0e,
536 0x61a3c261, 0x355f6a35, 0x57f9ae57, 0xb9d069b9, 0x86911786, 0xc15899c1, 0x1d273a1d, 0x9eb9279e,
537 0xe138d9e1, 0xf813ebf8, 0x98b32b98, 0x11332211, 0x69bbd269, 0xd970a9d9, 0x8e89078e, 0x94a73394,
538 0x9bb62d9b, 0x1e223c1e, 0x87921587, 0xe920c9e9, 0xce4987ce, 0x55ffaa55, 0x28785028, 0xdf7aa5df,
539 0x8c8f038c, 0xa1f859a1, 0x89800989, 0x0d171a0d, 0xbfda65bf, 0xe631d7e6, 0x42c68442, 0x68b8d068,
540 0x41c38241, 0x99b02999, 0x2d775a2d, 0x0f111e0f, 0xb0cb7bb0, 0x54fca854, 0xbbd66dbb, 0x163a2c16,
541};
542const te3 = [256]u32{
543 0x6363a5c6, 0x7c7c84f8, 0x777799ee, 0x7b7b8df6, 0xf2f20dff, 0x6b6bbdd6, 0x6f6fb1de, 0xc5c55491,
544 0x30305060, 0x01010302, 0x6767a9ce, 0x2b2b7d56, 0xfefe19e7, 0xd7d762b5, 0xababe64d, 0x76769aec,
545 0xcaca458f, 0x82829d1f, 0xc9c94089, 0x7d7d87fa, 0xfafa15ef, 0x5959ebb2, 0x4747c98e, 0xf0f00bfb,
546 0xadadec41, 0xd4d467b3, 0xa2a2fd5f, 0xafafea45, 0x9c9cbf23, 0xa4a4f753, 0x727296e4, 0xc0c05b9b,
547 0xb7b7c275, 0xfdfd1ce1, 0x9393ae3d, 0x26266a4c, 0x36365a6c, 0x3f3f417e, 0xf7f702f5, 0xcccc4f83,
548 0x34345c68, 0xa5a5f451, 0xe5e534d1, 0xf1f108f9, 0x717193e2, 0xd8d873ab, 0x31315362, 0x15153f2a,
549 0x04040c08, 0xc7c75295, 0x23236546, 0xc3c35e9d, 0x18182830, 0x9696a137, 0x05050f0a, 0x9a9ab52f,
550 0x0707090e, 0x12123624, 0x80809b1b, 0xe2e23ddf, 0xebeb26cd, 0x2727694e, 0xb2b2cd7f, 0x75759fea,
551 0x09091b12, 0x83839e1d, 0x2c2c7458, 0x1a1a2e34, 0x1b1b2d36, 0x6e6eb2dc, 0x5a5aeeb4, 0xa0a0fb5b,
552 0x5252f6a4, 0x3b3b4d76, 0xd6d661b7, 0xb3b3ce7d, 0x29297b52, 0xe3e33edd, 0x2f2f715e, 0x84849713,
553 0x5353f5a6, 0xd1d168b9, 0x00000000, 0xeded2cc1, 0x20206040, 0xfcfc1fe3, 0xb1b1c879, 0x5b5bedb6,
554 0x6a6abed4, 0xcbcb468d, 0xbebed967, 0x39394b72, 0x4a4ade94, 0x4c4cd498, 0x5858e8b0, 0xcfcf4a85,
555 0xd0d06bbb, 0xefef2ac5, 0xaaaae54f, 0xfbfb16ed, 0x4343c586, 0x4d4dd79a, 0x33335566, 0x85859411,
556 0x4545cf8a, 0xf9f910e9, 0x02020604, 0x7f7f81fe, 0x5050f0a0, 0x3c3c4478, 0x9f9fba25, 0xa8a8e34b,
557 0x5151f3a2, 0xa3a3fe5d, 0x4040c080, 0x8f8f8a05, 0x9292ad3f, 0x9d9dbc21, 0x38384870, 0xf5f504f1,
558 0xbcbcdf63, 0xb6b6c177, 0xdada75af, 0x21216342, 0x10103020, 0xffff1ae5, 0xf3f30efd, 0xd2d26dbf,
559 0xcdcd4c81, 0x0c0c1418, 0x13133526, 0xecec2fc3, 0x5f5fe1be, 0x9797a235, 0x4444cc88, 0x1717392e,
560 0xc4c45793, 0xa7a7f255, 0x7e7e82fc, 0x3d3d477a, 0x6464acc8, 0x5d5de7ba, 0x19192b32, 0x737395e6,
561 0x6060a0c0, 0x81819819, 0x4f4fd19e, 0xdcdc7fa3, 0x22226644, 0x2a2a7e54, 0x9090ab3b, 0x8888830b,
562 0x4646ca8c, 0xeeee29c7, 0xb8b8d36b, 0x14143c28, 0xdede79a7, 0x5e5ee2bc, 0x0b0b1d16, 0xdbdb76ad,
563 0xe0e03bdb, 0x32325664, 0x3a3a4e74, 0x0a0a1e14, 0x4949db92, 0x06060a0c, 0x24246c48, 0x5c5ce4b8,
564 0xc2c25d9f, 0xd3d36ebd, 0xacacef43, 0x6262a6c4, 0x9191a839, 0x9595a431, 0xe4e437d3, 0x79798bf2,
565 0xe7e732d5, 0xc8c8438b, 0x3737596e, 0x6d6db7da, 0x8d8d8c01, 0xd5d564b1, 0x4e4ed29c, 0xa9a9e049,
566 0x6c6cb4d8, 0x5656faac, 0xf4f407f3, 0xeaea25cf, 0x6565afca, 0x7a7a8ef4, 0xaeaee947, 0x08081810,
567 0xbabad56f, 0x787888f0, 0x25256f4a, 0x2e2e725c, 0x1c1c2438, 0xa6a6f157, 0xb4b4c773, 0xc6c65197,
568 0xe8e823cb, 0xdddd7ca1, 0x74749ce8, 0x1f1f213e, 0x4b4bdd96, 0xbdbddc61, 0x8b8b860d, 0x8a8a850f,
569 0x707090e0, 0x3e3e427c, 0xb5b5c471, 0x6666aacc, 0x4848d890, 0x03030506, 0xf6f601f7, 0x0e0e121c,
570 0x6161a3c2, 0x35355f6a, 0x5757f9ae, 0xb9b9d069, 0x86869117, 0xc1c15899, 0x1d1d273a, 0x9e9eb927,
571 0xe1e138d9, 0xf8f813eb, 0x9898b32b, 0x11113322, 0x6969bbd2, 0xd9d970a9, 0x8e8e8907, 0x9494a733,
572 0x9b9bb62d, 0x1e1e223c, 0x87879215, 0xe9e920c9, 0xcece4987, 0x5555ffaa, 0x28287850, 0xdfdf7aa5,
573 0x8c8c8f03, 0xa1a1f859, 0x89898009, 0x0d0d171a, 0xbfbfda65, 0xe6e631d7, 0x4242c684, 0x6868b8d0,
574 0x4141c382, 0x9999b029, 0x2d2d775a, 0x0f0f111e, 0xb0b0cb7b, 0x5454fca8, 0xbbbbd66d, 0x16163a2c,
575};
576
577const td0 = [256]u32{
578 0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96, 0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393,
579 0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25, 0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f,
580 0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1, 0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6,
581 0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da, 0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844,
582 0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd, 0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4,
583 0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45, 0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94,
584 0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7, 0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a,
585 0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5, 0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c,
586 0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1, 0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a,
587 0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75, 0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051,
588 0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46, 0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff,
589 0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77, 0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb,
590 0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000, 0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e,
591 0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927, 0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a,
592 0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e, 0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16,
593 0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d, 0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8,
594 0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd, 0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34,
595 0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163, 0xd731dcca, 0x42638510, 0x13972240, 0x84c61120,
596 0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d, 0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0,
597 0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422, 0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef,
598 0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36, 0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4,
599 0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662, 0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5,
600 0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3, 0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b,
601 0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8, 0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6,
602 0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6, 0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0,
603 0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815, 0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f,
604 0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df, 0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f,
605 0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e, 0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713,
606 0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89, 0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c,
607 0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf, 0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86,
608 0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f, 0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541,
609 0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190, 0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742,
610};
611const td1 = [256]u32{
612 0x5051f4a7, 0x537e4165, 0xc31a17a4, 0x963a275e, 0xcb3bab6b, 0xf11f9d45, 0xabacfa58, 0x934be303,
613 0x552030fa, 0xf6ad766d, 0x9188cc76, 0x25f5024c, 0xfc4fe5d7, 0xd7c52acb, 0x80263544, 0x8fb562a3,
614 0x49deb15a, 0x6725ba1b, 0x9845ea0e, 0xe15dfec0, 0x02c32f75, 0x12814cf0, 0xa38d4697, 0xc66bd3f9,
615 0xe7038f5f, 0x9515929c, 0xebbf6d7a, 0xda955259, 0x2dd4be83, 0xd3587421, 0x2949e069, 0x448ec9c8,
616 0x6a75c289, 0x78f48e79, 0x6b99583e, 0xdd27b971, 0xb6bee14f, 0x17f088ad, 0x66c920ac, 0xb47dce3a,
617 0x1863df4a, 0x82e51a31, 0x60975133, 0x4562537f, 0xe0b16477, 0x84bb6bae, 0x1cfe81a0, 0x94f9082b,
618 0x58704868, 0x198f45fd, 0x8794de6c, 0xb7527bf8, 0x23ab73d3, 0xe2724b02, 0x57e31f8f, 0x2a6655ab,
619 0x07b2eb28, 0x032fb5c2, 0x9a86c57b, 0xa5d33708, 0xf2302887, 0xb223bfa5, 0xba02036a, 0x5ced1682,
620 0x2b8acf1c, 0x92a779b4, 0xf0f307f2, 0xa14e69e2, 0xcd65daf4, 0xd50605be, 0x1fd13462, 0x8ac4a6fe,
621 0x9d342e53, 0xa0a2f355, 0x32058ae1, 0x75a4f6eb, 0x390b83ec, 0xaa4060ef, 0x065e719f, 0x51bd6e10,
622 0xf93e218a, 0x3d96dd06, 0xaedd3e05, 0x464de6bd, 0xb591548d, 0x0571c45d, 0x6f0406d4, 0xff605015,
623 0x241998fb, 0x97d6bde9, 0xcc894043, 0x7767d99e, 0xbdb0e842, 0x8807898b, 0x38e7195b, 0xdb79c8ee,
624 0x47a17c0a, 0xe97c420f, 0xc9f8841e, 0x00000000, 0x83098086, 0x48322bed, 0xac1e1170, 0x4e6c5a72,
625 0xfbfd0eff, 0x560f8538, 0x1e3daed5, 0x27362d39, 0x640a0fd9, 0x21685ca6, 0xd19b5b54, 0x3a24362e,
626 0xb10c0a67, 0x0f9357e7, 0xd2b4ee96, 0x9e1b9b91, 0x4f80c0c5, 0xa261dc20, 0x695a774b, 0x161c121a,
627 0x0ae293ba, 0xe5c0a02a, 0x433c22e0, 0x1d121b17, 0x0b0e090d, 0xadf28bc7, 0xb92db6a8, 0xc8141ea9,
628 0x8557f119, 0x4caf7507, 0xbbee99dd, 0xfda37f60, 0x9ff70126, 0xbc5c72f5, 0xc544663b, 0x345bfb7e,
629 0x768b4329, 0xdccb23c6, 0x68b6edfc, 0x63b8e4f1, 0xcad731dc, 0x10426385, 0x40139722, 0x2084c611,
630 0x7d854a24, 0xf8d2bb3d, 0x11aef932, 0x6dc729a1, 0x4b1d9e2f, 0xf3dcb230, 0xec0d8652, 0xd077c1e3,
631 0x6c2bb316, 0x99a970b9, 0xfa119448, 0x2247e964, 0xc4a8fc8c, 0x1aa0f03f, 0xd8567d2c, 0xef223390,
632 0xc787494e, 0xc1d938d1, 0xfe8ccaa2, 0x3698d40b, 0xcfa6f581, 0x28a57ade, 0x26dab78e, 0xa43fadbf,
633 0xe42c3a9d, 0x0d507892, 0x9b6a5fcc, 0x62547e46, 0xc2f68d13, 0xe890d8b8, 0x5e2e39f7, 0xf582c3af,
634 0xbe9f5d80, 0x7c69d093, 0xa96fd52d, 0xb3cf2512, 0x3bc8ac99, 0xa710187d, 0x6ee89c63, 0x7bdb3bbb,
635 0x09cd2678, 0xf46e5918, 0x01ec9ab7, 0xa8834f9a, 0x65e6956e, 0x7eaaffe6, 0x0821bccf, 0xe6ef15e8,
636 0xd9bae79b, 0xce4a6f36, 0xd4ea9f09, 0xd629b07c, 0xaf31a4b2, 0x312a3f23, 0x30c6a594, 0xc035a266,
637 0x37744ebc, 0xa6fc82ca, 0xb0e090d0, 0x1533a7d8, 0x4af10498, 0xf741ecda, 0x0e7fcd50, 0x2f1791f6,
638 0x8d764dd6, 0x4d43efb0, 0x54ccaa4d, 0xdfe49604, 0xe39ed1b5, 0x1b4c6a88, 0xb8c12c1f, 0x7f466551,
639 0x049d5eea, 0x5d018c35, 0x73fa8774, 0x2efb0b41, 0x5ab3671d, 0x5292dbd2, 0x33e91056, 0x136dd647,
640 0x8c9ad761, 0x7a37a10c, 0x8e59f814, 0x89eb133c, 0xeecea927, 0x35b761c9, 0xede11ce5, 0x3c7a47b1,
641 0x599cd2df, 0x3f55f273, 0x791814ce, 0xbf73c737, 0xea53f7cd, 0x5b5ffdaa, 0x14df3d6f, 0x867844db,
642 0x81caaff3, 0x3eb968c4, 0x2c382434, 0x5fc2a340, 0x72161dc3, 0x0cbce225, 0x8b283c49, 0x41ff0d95,
643 0x7139a801, 0xde080cb3, 0x9cd8b4e4, 0x906456c1, 0x617bcb84, 0x70d532b6, 0x74486c5c, 0x42d0b857,
644};
645const td2 = [256]u32{
646 0xa75051f4, 0x65537e41, 0xa4c31a17, 0x5e963a27, 0x6bcb3bab, 0x45f11f9d, 0x58abacfa, 0x03934be3,
647 0xfa552030, 0x6df6ad76, 0x769188cc, 0x4c25f502, 0xd7fc4fe5, 0xcbd7c52a, 0x44802635, 0xa38fb562,
648 0x5a49deb1, 0x1b6725ba, 0x0e9845ea, 0xc0e15dfe, 0x7502c32f, 0xf012814c, 0x97a38d46, 0xf9c66bd3,
649 0x5fe7038f, 0x9c951592, 0x7aebbf6d, 0x59da9552, 0x832dd4be, 0x21d35874, 0x692949e0, 0xc8448ec9,
650 0x896a75c2, 0x7978f48e, 0x3e6b9958, 0x71dd27b9, 0x4fb6bee1, 0xad17f088, 0xac66c920, 0x3ab47dce,
651 0x4a1863df, 0x3182e51a, 0x33609751, 0x7f456253, 0x77e0b164, 0xae84bb6b, 0xa01cfe81, 0x2b94f908,
652 0x68587048, 0xfd198f45, 0x6c8794de, 0xf8b7527b, 0xd323ab73, 0x02e2724b, 0x8f57e31f, 0xab2a6655,
653 0x2807b2eb, 0xc2032fb5, 0x7b9a86c5, 0x08a5d337, 0x87f23028, 0xa5b223bf, 0x6aba0203, 0x825ced16,
654 0x1c2b8acf, 0xb492a779, 0xf2f0f307, 0xe2a14e69, 0xf4cd65da, 0xbed50605, 0x621fd134, 0xfe8ac4a6,
655 0x539d342e, 0x55a0a2f3, 0xe132058a, 0xeb75a4f6, 0xec390b83, 0xefaa4060, 0x9f065e71, 0x1051bd6e,
656 0x8af93e21, 0x063d96dd, 0x05aedd3e, 0xbd464de6, 0x8db59154, 0x5d0571c4, 0xd46f0406, 0x15ff6050,
657 0xfb241998, 0xe997d6bd, 0x43cc8940, 0x9e7767d9, 0x42bdb0e8, 0x8b880789, 0x5b38e719, 0xeedb79c8,
658 0x0a47a17c, 0x0fe97c42, 0x1ec9f884, 0x00000000, 0x86830980, 0xed48322b, 0x70ac1e11, 0x724e6c5a,
659 0xfffbfd0e, 0x38560f85, 0xd51e3dae, 0x3927362d, 0xd9640a0f, 0xa621685c, 0x54d19b5b, 0x2e3a2436,
660 0x67b10c0a, 0xe70f9357, 0x96d2b4ee, 0x919e1b9b, 0xc54f80c0, 0x20a261dc, 0x4b695a77, 0x1a161c12,
661 0xba0ae293, 0x2ae5c0a0, 0xe0433c22, 0x171d121b, 0x0d0b0e09, 0xc7adf28b, 0xa8b92db6, 0xa9c8141e,
662 0x198557f1, 0x074caf75, 0xddbbee99, 0x60fda37f, 0x269ff701, 0xf5bc5c72, 0x3bc54466, 0x7e345bfb,
663 0x29768b43, 0xc6dccb23, 0xfc68b6ed, 0xf163b8e4, 0xdccad731, 0x85104263, 0x22401397, 0x112084c6,
664 0x247d854a, 0x3df8d2bb, 0x3211aef9, 0xa16dc729, 0x2f4b1d9e, 0x30f3dcb2, 0x52ec0d86, 0xe3d077c1,
665 0x166c2bb3, 0xb999a970, 0x48fa1194, 0x642247e9, 0x8cc4a8fc, 0x3f1aa0f0, 0x2cd8567d, 0x90ef2233,
666 0x4ec78749, 0xd1c1d938, 0xa2fe8cca, 0x0b3698d4, 0x81cfa6f5, 0xde28a57a, 0x8e26dab7, 0xbfa43fad,
667 0x9de42c3a, 0x920d5078, 0xcc9b6a5f, 0x4662547e, 0x13c2f68d, 0xb8e890d8, 0xf75e2e39, 0xaff582c3,
668 0x80be9f5d, 0x937c69d0, 0x2da96fd5, 0x12b3cf25, 0x993bc8ac, 0x7da71018, 0x636ee89c, 0xbb7bdb3b,
669 0x7809cd26, 0x18f46e59, 0xb701ec9a, 0x9aa8834f, 0x6e65e695, 0xe67eaaff, 0xcf0821bc, 0xe8e6ef15,
670 0x9bd9bae7, 0x36ce4a6f, 0x09d4ea9f, 0x7cd629b0, 0xb2af31a4, 0x23312a3f, 0x9430c6a5, 0x66c035a2,
671 0xbc37744e, 0xcaa6fc82, 0xd0b0e090, 0xd81533a7, 0x984af104, 0xdaf741ec, 0x500e7fcd, 0xf62f1791,
672 0xd68d764d, 0xb04d43ef, 0x4d54ccaa, 0x04dfe496, 0xb5e39ed1, 0x881b4c6a, 0x1fb8c12c, 0x517f4665,
673 0xea049d5e, 0x355d018c, 0x7473fa87, 0x412efb0b, 0x1d5ab367, 0xd25292db, 0x5633e910, 0x47136dd6,
674 0x618c9ad7, 0x0c7a37a1, 0x148e59f8, 0x3c89eb13, 0x27eecea9, 0xc935b761, 0xe5ede11c, 0xb13c7a47,
675 0xdf599cd2, 0x733f55f2, 0xce791814, 0x37bf73c7, 0xcdea53f7, 0xaa5b5ffd, 0x6f14df3d, 0xdb867844,
676 0xf381caaf, 0xc43eb968, 0x342c3824, 0x405fc2a3, 0xc372161d, 0x250cbce2, 0x498b283c, 0x9541ff0d,
677 0x017139a8, 0xb3de080c, 0xe49cd8b4, 0xc1906456, 0x84617bcb, 0xb670d532, 0x5c74486c, 0x5742d0b8,
678};
679const td3 = [256]u32{
680 0xf4a75051, 0x4165537e, 0x17a4c31a, 0x275e963a, 0xab6bcb3b, 0x9d45f11f, 0xfa58abac, 0xe303934b,
681 0x30fa5520, 0x766df6ad, 0xcc769188, 0x024c25f5, 0xe5d7fc4f, 0x2acbd7c5, 0x35448026, 0x62a38fb5,
682 0xb15a49de, 0xba1b6725, 0xea0e9845, 0xfec0e15d, 0x2f7502c3, 0x4cf01281, 0x4697a38d, 0xd3f9c66b,
683 0x8f5fe703, 0x929c9515, 0x6d7aebbf, 0x5259da95, 0xbe832dd4, 0x7421d358, 0xe0692949, 0xc9c8448e,
684 0xc2896a75, 0x8e7978f4, 0x583e6b99, 0xb971dd27, 0xe14fb6be, 0x88ad17f0, 0x20ac66c9, 0xce3ab47d,
685 0xdf4a1863, 0x1a3182e5, 0x51336097, 0x537f4562, 0x6477e0b1, 0x6bae84bb, 0x81a01cfe, 0x082b94f9,
686 0x48685870, 0x45fd198f, 0xde6c8794, 0x7bf8b752, 0x73d323ab, 0x4b02e272, 0x1f8f57e3, 0x55ab2a66,
687 0xeb2807b2, 0xb5c2032f, 0xc57b9a86, 0x3708a5d3, 0x2887f230, 0xbfa5b223, 0x036aba02, 0x16825ced,
688 0xcf1c2b8a, 0x79b492a7, 0x07f2f0f3, 0x69e2a14e, 0xdaf4cd65, 0x05bed506, 0x34621fd1, 0xa6fe8ac4,
689 0x2e539d34, 0xf355a0a2, 0x8ae13205, 0xf6eb75a4, 0x83ec390b, 0x60efaa40, 0x719f065e, 0x6e1051bd,
690 0x218af93e, 0xdd063d96, 0x3e05aedd, 0xe6bd464d, 0x548db591, 0xc45d0571, 0x06d46f04, 0x5015ff60,
691 0x98fb2419, 0xbde997d6, 0x4043cc89, 0xd99e7767, 0xe842bdb0, 0x898b8807, 0x195b38e7, 0xc8eedb79,
692 0x7c0a47a1, 0x420fe97c, 0x841ec9f8, 0x00000000, 0x80868309, 0x2bed4832, 0x1170ac1e, 0x5a724e6c,
693 0x0efffbfd, 0x8538560f, 0xaed51e3d, 0x2d392736, 0x0fd9640a, 0x5ca62168, 0x5b54d19b, 0x362e3a24,
694 0x0a67b10c, 0x57e70f93, 0xee96d2b4, 0x9b919e1b, 0xc0c54f80, 0xdc20a261, 0x774b695a, 0x121a161c,
695 0x93ba0ae2, 0xa02ae5c0, 0x22e0433c, 0x1b171d12, 0x090d0b0e, 0x8bc7adf2, 0xb6a8b92d, 0x1ea9c814,
696 0xf1198557, 0x75074caf, 0x99ddbbee, 0x7f60fda3, 0x01269ff7, 0x72f5bc5c, 0x663bc544, 0xfb7e345b,
697 0x4329768b, 0x23c6dccb, 0xedfc68b6, 0xe4f163b8, 0x31dccad7, 0x63851042, 0x97224013, 0xc6112084,
698 0x4a247d85, 0xbb3df8d2, 0xf93211ae, 0x29a16dc7, 0x9e2f4b1d, 0xb230f3dc, 0x8652ec0d, 0xc1e3d077,
699 0xb3166c2b, 0x70b999a9, 0x9448fa11, 0xe9642247, 0xfc8cc4a8, 0xf03f1aa0, 0x7d2cd856, 0x3390ef22,
700 0x494ec787, 0x38d1c1d9, 0xcaa2fe8c, 0xd40b3698, 0xf581cfa6, 0x7ade28a5, 0xb78e26da, 0xadbfa43f,
701 0x3a9de42c, 0x78920d50, 0x5fcc9b6a, 0x7e466254, 0x8d13c2f6, 0xd8b8e890, 0x39f75e2e, 0xc3aff582,
702 0x5d80be9f, 0xd0937c69, 0xd52da96f, 0x2512b3cf, 0xac993bc8, 0x187da710, 0x9c636ee8, 0x3bbb7bdb,
703 0x267809cd, 0x5918f46e, 0x9ab701ec, 0x4f9aa883, 0x956e65e6, 0xffe67eaa, 0xbccf0821, 0x15e8e6ef,
704 0xe79bd9ba, 0x6f36ce4a, 0x9f09d4ea, 0xb07cd629, 0xa4b2af31, 0x3f23312a, 0xa59430c6, 0xa266c035,
705 0x4ebc3774, 0x82caa6fc, 0x90d0b0e0, 0xa7d81533, 0x04984af1, 0xecdaf741, 0xcd500e7f, 0x91f62f17,
706 0x4dd68d76, 0xefb04d43, 0xaa4d54cc, 0x9604dfe4, 0xd1b5e39e, 0x6a881b4c, 0x2c1fb8c1, 0x65517f46,
707 0x5eea049d, 0x8c355d01, 0x877473fa, 0x0b412efb, 0x671d5ab3, 0xdbd25292, 0x105633e9, 0xd647136d,
708 0xd7618c9a, 0xa10c7a37, 0xf8148e59, 0x133c89eb, 0xa927eece, 0x61c935b7, 0x1ce5ede1, 0x47b13c7a,
709 0xd2df599c, 0xf2733f55, 0x14ce7918, 0xc737bf73, 0xf7cdea53, 0xfdaa5b5f, 0x3d6f14df, 0x44db8678,
710 0xaff381ca, 0x68c43eb9, 0x24342c38, 0xa3405fc2, 0x1dc37216, 0xe2250cbc, 0x3c498b28, 0x0d9541ff,
711 0xa8017139, 0x0cb3de08, 0xb4e49cd8, 0x56c19064, 0xcb84617b, 0x32b670d5, 0x6c5c7448, 0xb85742d0,
712};
lib/std/crypto/aes/aesni.zig created+420
...@@ -0,0 +1,420 @@
1// SPDX-License-Identifier: MIT
2// Copyright (c) 2015-2020 Zig Contributors
3// This file is part of [zig](https://ziglang.org/), which is MIT licensed.
4// The MIT license requires this copyright notice to be included in all copies
5// and substantial portions of the software.
6// Based on Go stdlib implementation
7
8const std = @import("../../std.zig");
9const mem = std.mem;
10const debug = std.debug;
11const Vector = std.meta.Vector;
12
13const BlockVec = Vector(2, u64);
14
15/// A single AES block.
16pub const Block = struct {
17 pub const block_size: usize = 16;
18
19 /// Internal representation of a block.
20 repr: BlockVec,
21
22 /// Convert a byte sequence into an internal representation.
23 pub inline fn fromBytes(bytes: *const [16]u8) Block {
24 const repr = mem.bytesToValue(BlockVec, bytes);
25 return Block{ .repr = repr };
26 }
27
28 /// Convert the internal representation of a block into a byte sequence.
29 pub inline fn toBytes(block: Block) [16]u8 {
30 return mem.toBytes(block.repr);
31 }
32
33 /// XOR the block with a byte sequence.
34 pub inline fn xorBytes(block: Block, bytes: *const [16]u8) [16]u8 {
35 const x = block.repr ^ fromBytes(bytes).repr;
36 return mem.toBytes(x);
37 }
38
39 /// Encrypt a block with a round key.
40 pub inline fn encrypt(block: Block, round_key: Block) Block {
41 return Block{
42 .repr = asm (
43 \\ vaesenc %[rk], %[in], %[out]
44 : [out] "=x" (-> BlockVec)
45 : [in] "x" (block.repr),
46 [rk] "x" (round_key.repr)
47 ),
48 };
49 }
50
51 /// Encrypt a block with the last round key.
52 pub inline fn encryptLast(block: Block, round_key: Block) Block {
53 return Block{
54 .repr = asm (
55 \\ vaesenclast %[rk], %[in], %[out]
56 : [out] "=x" (-> BlockVec)
57 : [in] "x" (block.repr),
58 [rk] "x" (round_key.repr)
59 ),
60 };
61 }
62
63 /// Decrypt a block with a round key.
64 pub inline fn decrypt(block: Block, inv_round_key: Block) Block {
65 return Block{
66 .repr = asm (
67 \\ vaesdec %[rk], %[in], %[out]
68 : [out] "=x" (-> BlockVec)
69 : [in] "x" (block.repr),
70 [rk] "x" (inv_round_key.repr)
71 ),
72 };
73 }
74
75 /// Decrypt a block with the last round key.
76 pub inline fn decryptLast(block: Block, inv_round_key: Block) Block {
77 return Block{
78 .repr = asm (
79 \\ vaesdeclast %[rk], %[in], %[out]
80 : [out] "=x" (-> BlockVec)
81 : [in] "x" (block.repr),
82 [rk] "x" (inv_round_key.repr)
83 ),
84 };
85 }
86
87 /// XOR the content of two blocks.
88 pub inline fn xor(block1: Block, block2: Block) Block {
89 return Block{ .repr = block1.repr ^ block2.repr };
90 }
91
92 /// Perform operations on multiple blocks in parallel.
93 pub const parallel = struct {
94 /// The recommended number of AES encryption/decryption to perform in parallel for the chosen implementation.
95 pub const optimal_parallel_blocks = 8;
96
97 /// Encrypt multiple blocks in parallel, each their own round key.
98 pub inline fn encryptParallel(comptime count: usize, blocks: [count]Block, round_keys: [count]Block) [count]Block {
99 comptime var i = 0;
100 var out: [count]Block = undefined;
101 inline while (i < count) : (i += 1) {
102 out[i] = blocks[i].encrypt(round_keys[i]);
103 }
104 return out;
105 }
106
107 /// Decrypt multiple blocks in parallel, each their own round key.
108 pub inline fn decryptParallel(comptime count: usize, blocks: [count]Block, round_keys: [count]Block) [count]Block {
109 comptime var i = 0;
110 var out: [count]Block = undefined;
111 inline while (i < count) : (i += 1) {
112 out[i] = blocks[i].decrypt(round_keys[i]);
113 }
114 return out;
115 }
116
117 /// Encrypt multple blocks in parallel with the same round key.
118 pub inline fn encryptWide(comptime count: usize, blocks: [count]Block, round_key: Block) [count]Block {
119 comptime var i = 0;
120 var out: [count]Block = undefined;
121 inline while (i < count) : (i += 1) {
122 out[i] = blocks[i].encrypt(round_key);
123 }
124 return out;
125 }
126
127 /// Decrypt multple blocks in parallel with the same round key.
128 pub inline fn decryptWide(comptime count: usize, blocks: [count]Block, round_key: Block) [count]Block {
129 comptime var i = 0;
130 var out: [count]Block = undefined;
131 inline while (i < count) : (i += 1) {
132 out[i] = blocks[i].decrypt(round_key);
133 }
134 return out;
135 }
136
137 /// Encrypt multple blocks in parallel with the same last round key.
138 pub inline fn encryptLastWide(comptime count: usize, blocks: [count]Block, round_key: Block) [count]Block {
139 comptime var i = 0;
140 var out: [count]Block = undefined;
141 inline while (i < count) : (i += 1) {
142 out[i] = blocks[i].encryptLast(round_key);
143 }
144 return out;
145 }
146
147 /// Decrypt multple blocks in parallel with the same last round key.
148 pub inline fn decryptLastWide(comptime count: usize, blocks: [count]Block, round_key: Block) [count]Block {
149 comptime var i = 0;
150 var out: [count]Block = undefined;
151 inline while (i < count) : (i += 1) {
152 out[i] = blocks[i].decryptLast(round_key);
153 }
154 return out;
155 }
156 };
157};
158
159fn KeySchedule(comptime AES: type) type {
160 std.debug.assert(AES.rounds == 10 or AES.rounds == 14);
161 const rounds = AES.rounds;
162
163 return struct {
164 const Self = @This();
165 round_keys: [rounds + 1]Block,
166
167 fn drc(comptime second: bool, comptime rc: u8, t: BlockVec, tx: BlockVec) BlockVec {
168 var s: BlockVec = undefined;
169 var ts: BlockVec = undefined;
170 return asm (
171 \\ vaeskeygenassist %[rc], %[t], %[s]
172 \\ vpslldq $4, %[tx], %[ts]
173 \\ vpxor %[ts], %[tx], %[r]
174 \\ vpslldq $8, %[r], %[ts]
175 \\ vpxor %[ts], %[r], %[r]
176 \\ vpshufd %[mask], %[s], %[ts]
177 \\ vpxor %[ts], %[r], %[r]
178 : [r] "=&x" (-> BlockVec),
179 [s] "=&x" (s),
180 [ts] "=&x" (ts)
181 : [rc] "n" (rc),
182 [t] "x" (t),
183 [tx] "x" (tx),
184 [mask] "n" (@as(u8, if (second) 0xaa else 0xff))
185 );
186 }
187
188 fn expand128(t1: *Block) Self {
189 var round_keys: [11]Block = undefined;
190 const rcs = [_]u8{ 1, 2, 4, 8, 16, 32, 64, 128, 27, 54 };
191 inline for (rcs) |rc, round| {
192 round_keys[round] = t1.*;
193 t1.repr = drc(false, rc, t1.repr, t1.repr);
194 }
195 round_keys[rcs.len] = t1.*;
196 return Self{ .round_keys = round_keys };
197 }
198
199 fn expand256(t1: *Block, t2: *Block) Self {
200 var round_keys: [15]Block = undefined;
201 const rcs = [_]u8{ 1, 2, 4, 8, 16, 32 };
202 round_keys[0] = t1.*;
203 inline for (rcs) |rc, round| {
204 round_keys[round * 2 + 1] = t2.*;
205 t1.repr = drc(false, rc, t2.repr, t1.repr);
206 round_keys[round * 2 + 2] = t1.*;
207 t2.repr = drc(true, rc, t1.repr, t2.repr);
208 }
209 round_keys[rcs.len * 2 + 1] = t2.*;
210 t1.repr = drc(false, 64, t2.repr, t1.repr);
211 round_keys[rcs.len * 2 + 2] = t1.*;
212 return Self{ .round_keys = round_keys };
213 }
214
215 /// Invert the key schedule.
216 pub fn invert(key_schedule: Self) Self {
217 const round_keys = &key_schedule.round_keys;
218 var inv_round_keys: [rounds + 1]Block = undefined;
219 inv_round_keys[0] = round_keys[rounds];
220 comptime var i = 1;
221 inline while (i < rounds) : (i += 1) {
222 inv_round_keys[i] = Block{
223 .repr = asm (
224 \\ vaesimc %[rk], %[inv_rk]
225 : [inv_rk] "=x" (-> BlockVec)
226 : [rk] "x" (round_keys[rounds - i].repr)
227 ),
228 };
229 }
230 inv_round_keys[rounds] = round_keys[0];
231 return Self{ .round_keys = inv_round_keys };
232 }
233 };
234}
235
236/// A context to perform encryption using the standard AES key schedule.
237pub fn AESEncryptCtx(comptime AES: type) type {
238 std.debug.assert(AES.key_bits == 128 or AES.key_bits == 256);
239 const rounds = AES.rounds;
240
241 return struct {
242 const Self = @This();
243 pub const block = AES.block;
244 pub const block_size = block.block_size;
245 key_schedule: KeySchedule(AES),
246
247 /// Create a new encryption context with the given key.
248 pub fn init(key: [AES.key_bits / 8]u8) Self {
249 var t1 = Block.fromBytes(key[0..16]);
250 const key_schedule = if (AES.key_bits == 128) ks: {
251 break :ks KeySchedule(AES).expand128(&t1);
252 } else ks: {
253 var t2 = Block.fromBytes(key[16..32]);
254 break :ks KeySchedule(AES).expand256(&t1, &t2);
255 };
256 return Self{
257 .key_schedule = key_schedule,
258 };
259 }
260
261 /// Encrypt a single block.
262 pub fn encrypt(ctx: Self, dst: *[16]u8, src: *const [16]u8) void {
263 const round_keys = ctx.key_schedule.round_keys;
264 var t = Block.fromBytes(src).xor(round_keys[0]);
265 comptime var i = 1;
266 inline while (i < rounds) : (i += 1) {
267 t = t.encrypt(round_keys[i]);
268 }
269 t = t.encryptLast(round_keys[rounds]);
270 dst.* = t.toBytes();
271 }
272
273 /// Encrypt+XOR a single block.
274 pub fn xor(ctx: Self, dst: *[16]u8, src: *const [16]u8, counter: [16]u8) void {
275 const round_keys = ctx.key_schedule.round_keys;
276 var t = Block.fromBytes(&counter).xor(round_keys[0]);
277 comptime var i = 1;
278 inline while (i < rounds) : (i += 1) {
279 t = t.encrypt(round_keys[i]);
280 }
281 t = t.encryptLast(round_keys[rounds]);
282 dst.* = t.xorBytes(src);
283 }
284
285 /// Encrypt multiple blocks, possibly leveraging parallelization.
286 pub fn encryptWide(ctx: Self, comptime count: usize, dst: *[16 * count]u8, src: *const [16 * count]u8) void {
287 const round_keys = ctx.key_schedule.round_keys;
288 var ts: [count]Block = undefined;
289 comptime var j = 0;
290 inline while (j < count) : (j += 1) {
291 ts[j] = Block.fromBytes(src[j * 16 .. j * 16 + 16][0..16]).xor(round_keys[0]);
292 }
293 comptime var i = 1;
294 inline while (i < rounds) : (i += 1) {
295 ts = Block.parallel.encryptWide(count, ts, round_keys[i]);
296 }
297 i = 1;
298 inline while (i < count) : (i += 1) {
299 ts = Block.parallel.encryptLastWide(count, ts, round_keys[i]);
300 }
301 j = 0;
302 inline while (j < count) : (j += 1) {
303 dst[16 * j .. 16 * j + 16].* = ts[j].toBytes();
304 }
305 }
306
307 /// Encrypt+XOR multiple blocks, possibly leveraging parallelization.
308 pub fn xorWide(ctx: Self, comptime count: usize, dst: *[16 * count]u8, src: *const [16 * count]u8, counters: [16 * count]u8) void {
309 const round_keys = ctx.key_schedule.round_keys;
310 var ts: [count]Block = undefined;
311 comptime var j = 0;
312 inline while (j < count) : (j += 1) {
313 ts[j] = Block.fromBytes(counters[j * 16 .. j * 16 + 16][0..16]).xor(round_keys[0]);
314 }
315 comptime var i = 1;
316 inline while (i < rounds) : (i += 1) {
317 ts = Block.parallel.encryptWide(count, ts, round_keys[i]);
318 }
319 ts = Block.parallel.encryptLastWide(count, ts, round_keys[i]);
320 j = 0;
321 inline while (j < count) : (j += 1) {
322 dst[16 * j .. 16 * j + 16].* = ts[j].xorBytes(src[16 * j .. 16 * j + 16]);
323 }
324 }
325 };
326}
327
328/// A context to perform decryption using the standard AES key schedule.
329pub fn AESDecryptCtx(comptime AES: type) type {
330 std.debug.assert(AES.key_bits == 128 or AES.key_bits == 256);
331 const rounds = AES.rounds;
332
333 return struct {
334 const Self = @This();
335 pub const block = AES.block;
336 pub const block_size = block.block_size;
337 key_schedule: KeySchedule(AES),
338
339 /// Create a decryption context from an existing encryption context.
340 pub fn initFromEnc(ctx: AESEncryptCtx(AES)) Self {
341 return Self{
342 .key_schedule = ctx.key_schedule.invert(),
343 };
344 }
345
346 /// Create a new decryption context with the given key.
347 pub fn init(key: [AES.key_bits / 8]u8) Self {
348 const enc_ctx = AESEncryptCtx(AES).init(key);
349 return initFromEnc(enc_ctx);
350 }
351
352 /// Decrypt a single block.
353 pub fn decrypt(ctx: Self, dst: *[16]u8, src: *const [16]u8) void {
354 const inv_round_keys = ctx.key_schedule.round_keys;
355 var t = Block.fromBytes(src).xor(inv_round_keys[0]);
356 comptime var i = 1;
357 inline while (i < rounds) : (i += 1) {
358 t = t.decrypt(inv_round_keys[i]);
359 }
360 t = t.decryptLast(inv_round_keys[rounds]);
361 dst.* = t.toBytes();
362 }
363
364 /// Decrypt multiple blocks, possibly leveraging parallelization.
365 pub fn decryptWide(ctx: Self, comptime count: usize, dst: *[16 * count]u8, src: *const [16 * count]u8) void {
366 const inv_round_keys = ctx.key_schedule.round_keys;
367 var ts: [count]Block = undefined;
368 comptime var j = 0;
369 inline while (j < count) : (j += 1) {
370 ts[j] = Block.fromBytes(src[j * 16 .. j * 16 + 16][0..16]).xor(inv_round_keys[0]);
371 }
372 comptime var i = 1;
373 inline while (i < rounds) : (i += 1) {
374 ts = Block.parallel.decryptWide(count, ts, inv_round_keys[i]);
375 }
376 i = 1;
377 inline while (i < count) : (i += 1) {
378 ts = Block.parallel.decryptLastWide(count, ts, inv_round_keys[i]);
379 }
380 j = 0;
381 inline while (j < count) : (j += 1) {
382 dst[16 * j .. 16 * j + 16].* = ts[j].toBytes();
383 }
384 }
385 };
386}
387
388/// AES-128 with the standard key schedule.
389pub const AES128 = struct {
390 pub const key_bits: usize = 128;
391 pub const rounds = ((key_bits - 64) / 32 + 8);
392 pub const block = Block;
393
394 /// Create a new context for encryption.
395 pub fn initEnc(key: [key_bits / 8]u8) AESEncryptCtx(AES128) {
396 return AESEncryptCtx(AES128).init(key);
397 }
398
399 /// Create a new context for decryption.
400 pub fn initDec(key: [key_bits / 8]u8) AESDecryptCtx(AES128) {
401 return AESDecryptCtx(AES128).init(key);
402 }
403};
404
405/// AES-256 with the standard key schedule.
406pub const AES256 = struct {
407 pub const key_bits: usize = 256;
408 pub const rounds = ((key_bits - 64) / 32 + 8);
409 pub const block = Block;
410
411 /// Create a new context for encryption.
412 pub fn initEnc(key: [key_bits / 8]u8) AESEncryptCtx(AES256) {
413 return AESEncryptCtx(AES256).init(key);
414 }
415
416 /// Create a new context for decryption.
417 pub fn initDec(key: [key_bits / 8]u8) AESDecryptCtx(AES256) {
418 return AESDecryptCtx(AES256).init(key);
419 }
420};
lib/std/crypto/aes/soft.zig created+735
...@@ -0,0 +1,735 @@
1// SPDX-License-Identifier: MIT
2// Copyright (c) 2015-2020 Zig Contributors
3// This file is part of [zig](https://ziglang.org/), which is MIT licensed.
4// The MIT license requires this copyright notice to be included in all copies
5// and substantial portions of the software.
6// Based on Go stdlib implementation
7
8const std = @import("../../std.zig");
9const mem = std.mem;
10
11const BlockVec = [4]u32;
12
13/// A single AES block.
14pub const Block = struct {
15 pub const block_size: usize = 16;
16
17 /// Internal representation of a block.
18 repr: BlockVec align(16),
19
20 /// Convert a byte sequence into an internal representation.
21 pub inline fn fromBytes(bytes: *const [16]u8) Block {
22 const s0 = mem.readIntBig(u32, bytes[0..4]);
23 const s1 = mem.readIntBig(u32, bytes[4..8]);
24 const s2 = mem.readIntBig(u32, bytes[8..12]);
25 const s3 = mem.readIntBig(u32, bytes[12..16]);
26 return Block{ .repr = BlockVec{ s0, s1, s2, s3 } };
27 }
28
29 /// Convert the internal representation of a block into a byte sequence.
30 pub inline fn toBytes(block: Block) [16]u8 {
31 var bytes: [16]u8 = undefined;
32 mem.writeIntBig(u32, bytes[0..4], block.repr[0]);
33 mem.writeIntBig(u32, bytes[4..8], block.repr[1]);
34 mem.writeIntBig(u32, bytes[8..12], block.repr[2]);
35 mem.writeIntBig(u32, bytes[12..16], block.repr[3]);
36 return bytes;
37 }
38
39 /// XOR the block with a byte sequence.
40 pub inline fn xorBytes(block: Block, bytes: *const [16]u8) [16]u8 {
41 const block_bytes = block.toBytes();
42 var x: [16]u8 = undefined;
43 comptime var i: usize = 0;
44 inline while (i < 16) : (i += 1) {
45 x[i] = block_bytes[i] ^ bytes[i];
46 }
47 return x;
48 }
49
50 /// Encrypt a block with a round key.
51 pub inline fn encrypt(block: Block, round_key: Block) Block {
52 const src = &block.repr;
53
54 const s0 = block.repr[0];
55 const s1 = block.repr[1];
56 const s2 = block.repr[2];
57 const s3 = block.repr[3];
58
59 const t0 = round_key.repr[0] ^ te0[@truncate(u8, s0 >> 24)] ^ te1[@truncate(u8, s1 >> 16)] ^ te2[@truncate(u8, s2 >> 8)] ^ te3[@truncate(u8, s3)];
60 const t1 = round_key.repr[1] ^ te0[@truncate(u8, s1 >> 24)] ^ te1[@truncate(u8, s2 >> 16)] ^ te2[@truncate(u8, s3 >> 8)] ^ te3[@truncate(u8, s0)];
61 const t2 = round_key.repr[2] ^ te0[@truncate(u8, s2 >> 24)] ^ te1[@truncate(u8, s3 >> 16)] ^ te2[@truncate(u8, s0 >> 8)] ^ te3[@truncate(u8, s1)];
62 const t3 = round_key.repr[3] ^ te0[@truncate(u8, s3 >> 24)] ^ te1[@truncate(u8, s0 >> 16)] ^ te2[@truncate(u8, s1 >> 8)] ^ te3[@truncate(u8, s2)];
63
64 return Block{ .repr = BlockVec{ t0, t1, t2, t3 } };
65 }
66
67 /// Encrypt a block with the last round key.
68 pub inline fn encryptLast(block: Block, round_key: Block) Block {
69 const src = &block.repr;
70
71 const t0 = block.repr[0];
72 const t1 = block.repr[1];
73 const t2 = block.repr[2];
74 const t3 = block.repr[3];
75
76 // Last round uses s-box directly and XORs to produce output.
77 var s0 = @as(u32, sbox0[t0 >> 24]) << 24 | @as(u32, sbox0[t1 >> 16 & 0xff]) << 16 | @as(u32, sbox0[t2 >> 8 & 0xff]) << 8 | @as(u32, sbox0[t3 & 0xff]);
78 var s1 = @as(u32, sbox0[t1 >> 24]) << 24 | @as(u32, sbox0[t2 >> 16 & 0xff]) << 16 | @as(u32, sbox0[t3 >> 8 & 0xff]) << 8 | @as(u32, sbox0[t0 & 0xff]);
79 var s2 = @as(u32, sbox0[t2 >> 24]) << 24 | @as(u32, sbox0[t3 >> 16 & 0xff]) << 16 | @as(u32, sbox0[t0 >> 8 & 0xff]) << 8 | @as(u32, sbox0[t1 & 0xff]);
80 var s3 = @as(u32, sbox0[t3 >> 24]) << 24 | @as(u32, sbox0[t0 >> 16 & 0xff]) << 16 | @as(u32, sbox0[t1 >> 8 & 0xff]) << 8 | @as(u32, sbox0[t2 & 0xff]);
81 s0 ^= round_key.repr[0];
82 s1 ^= round_key.repr[1];
83 s2 ^= round_key.repr[2];
84 s3 ^= round_key.repr[3];
85
86 return Block{ .repr = BlockVec{ s0, s1, s2, s3 } };
87 }
88
89 /// Decrypt a block with a round key.
90 pub inline fn decrypt(block: Block, round_key: Block) Block {
91 const src = &block.repr;
92
93 const s0 = block.repr[0];
94 const s1 = block.repr[1];
95 const s2 = block.repr[2];
96 const s3 = block.repr[3];
97
98 const t0 = round_key.repr[0] ^ td0[@truncate(u8, s0 >> 24)] ^ td1[@truncate(u8, s3 >> 16)] ^ td2[@truncate(u8, s2 >> 8)] ^ td3[@truncate(u8, s1)];
99 const t1 = round_key.repr[1] ^ td0[@truncate(u8, s1 >> 24)] ^ td1[@truncate(u8, s0 >> 16)] ^ td2[@truncate(u8, s3 >> 8)] ^ td3[@truncate(u8, s2)];
100 const t2 = round_key.repr[2] ^ td0[@truncate(u8, s2 >> 24)] ^ td1[@truncate(u8, s1 >> 16)] ^ td2[@truncate(u8, s0 >> 8)] ^ td3[@truncate(u8, s3)];
101 const t3 = round_key.repr[3] ^ td0[@truncate(u8, s3 >> 24)] ^ td1[@truncate(u8, s2 >> 16)] ^ td2[@truncate(u8, s1 >> 8)] ^ td3[@truncate(u8, s0)];
102
103 return Block{ .repr = BlockVec{ t0, t1, t2, t3 } };
104 }
105
106 /// Decrypt a block with the last round key.
107 pub inline fn decryptLast(block: Block, round_key: Block) Block {
108 const src = &block.repr;
109
110 const t0 = block.repr[0];
111 const t1 = block.repr[1];
112 const t2 = block.repr[2];
113 const t3 = block.repr[3];
114
115 // Last round uses s-box directly and XORs to produce output.
116 var s0 = @as(u32, sbox1[t0 >> 24]) << 24 | @as(u32, sbox1[t3 >> 16 & 0xff]) << 16 | @as(u32, sbox1[t2 >> 8 & 0xff]) << 8 | @as(u32, sbox1[t1 & 0xff]);
117 var s1 = @as(u32, sbox1[t1 >> 24]) << 24 | @as(u32, sbox1[t0 >> 16 & 0xff]) << 16 | @as(u32, sbox1[t3 >> 8 & 0xff]) << 8 | @as(u32, sbox1[t2 & 0xff]);
118 var s2 = @as(u32, sbox1[t2 >> 24]) << 24 | @as(u32, sbox1[t1 >> 16 & 0xff]) << 16 | @as(u32, sbox1[t0 >> 8 & 0xff]) << 8 | @as(u32, sbox1[t3 & 0xff]);
119 var s3 = @as(u32, sbox1[t3 >> 24]) << 24 | @as(u32, sbox1[t2 >> 16 & 0xff]) << 16 | @as(u32, sbox1[t1 >> 8 & 0xff]) << 8 | @as(u32, sbox1[t0 & 0xff]);
120 s0 ^= round_key.repr[0];
121 s1 ^= round_key.repr[1];
122 s2 ^= round_key.repr[2];
123 s3 ^= round_key.repr[3];
124
125 return Block{ .repr = BlockVec{ s0, s1, s2, s3 } };
126 }
127
128 /// XOR the content of two blocks.
129 pub inline fn xor(block1: Block, block2: Block) Block {
130 var x: BlockVec = undefined;
131 comptime var i = 0;
132 inline while (i < 4) : (i += 1) {
133 x[i] = block1.repr[i] ^ block2.repr[i];
134 }
135 return Block{ .repr = x };
136 }
137
138 /// Perform operations on multiple blocks in parallel.
139 pub const parallel = struct {
140 /// The recommended number of AES encryption/decryption to perform in parallel for the chosen implementation.
141 pub const optimal_parallel_blocks = 1;
142
143 /// Encrypt multiple blocks in parallel, each their own round key.
144 pub fn encryptParallel(comptime count: usize, blocks: [count]Block, round_keys: [count]Block) [count]Block {
145 var i = 0;
146 var out: [count]Block = undefined;
147 while (i < count) : (i += 1) {
148 out[i] = blocks[i].encrypt(round_keys[i]);
149 }
150 return out;
151 }
152
153 /// Decrypt multiple blocks in parallel, each their own round key.
154 pub fn decryptParallel(comptime count: usize, blocks: [count]Block, round_keys: [count]Block) [count]Block {
155 var i = 0;
156 var out: [count]Block = undefined;
157 while (i < count) : (i += 1) {
158 out[i] = blocks[i].decrypt(round_keys[i]);
159 }
160 return out;
161 }
162
163 /// Encrypt multple blocks in parallel with the same round key.
164 pub fn encryptWide(comptime count: usize, blocks: [count]Block, round_key: Block) [count]Block {
165 var i = 0;
166 var out: [count]Block = undefined;
167 while (i < count) : (i += 1) {
168 out[i] = blocks[i].encrypt(round_key);
169 }
170 return out;
171 }
172
173 /// Decrypt multple blocks in parallel with the same round key.
174 pub fn decryptWide(comptime count: usize, blocks: [count]Block, round_key: Block) [count]Block {
175 var i = 0;
176 var out: [count]Block = undefined;
177 while (i < count) : (i += 1) {
178 out[i] = blocks[i].decrypt(round_key);
179 }
180 return out;
181 }
182
183 /// Encrypt multple blocks in parallel with the same last round key.
184 pub fn encryptLastWide(comptime count: usize, blocks: [count]Block, round_key: Block) [count]Block {
185 var i = 0;
186 var out: [count]Block = undefined;
187 while (i < count) : (i += 1) {
188 out[i] = blocks[i].encryptLast(round_key);
189 }
190 return out;
191 }
192
193 /// Decrypt multple blocks in parallel with the same last round key.
194 pub fn decryptLastWide(comptime count: usize, blocks: [count]Block, round_key: Block) [count]Block {
195 var i = 0;
196 var out: [count]Block = undefined;
197 while (i < count) : (i += 1) {
198 out[i] = blocks[i].decryptLast(round_key);
199 }
200 return out;
201 }
202 };
203};
204
205fn KeySchedule(comptime AES: type) type {
206 std.debug.assert(AES.rounds == 10 or AES.rounds == 14);
207 const key_size = AES.key_bits / 8;
208 const rounds = AES.rounds;
209
210 return struct {
211 const Self = @This();
212 const words_in_key = key_size / 4;
213
214 round_keys: [rounds + 1]Block,
215
216 // Key expansion algorithm. See FIPS-197, Figure 11.
217 fn expandKey(key: [key_size]u8) Self {
218 const subw = struct {
219 // Apply sbox0 to each byte in w.
220 fn func(w: u32) u32 {
221 return @as(u32, sbox0[w >> 24]) << 24 | @as(u32, sbox0[w >> 16 & 0xff]) << 16 | @as(u32, sbox0[w >> 8 & 0xff]) << 8 | @as(u32, sbox0[w & 0xff]);
222 }
223 }.func;
224
225 var round_keys: [rounds + 1]Block = undefined;
226 comptime var i: usize = 0;
227 inline while (i < words_in_key) : (i += 1) {
228 round_keys[i / 4].repr[i % 4] = mem.readIntBig(u32, key[4 * i ..][0..4]);
229 }
230 inline while (i < round_keys.len * 4) : (i += 1) {
231 var t = round_keys[(i - 1) / 4].repr[(i - 1) % 4];
232 if (i % words_in_key == 0) {
233 t = subw(std.math.rotl(u32, t, 8)) ^ (@as(u32, powx[i / words_in_key - 1]) << 24);
234 } else if (words_in_key > 6 and i % words_in_key == 4) {
235 t = subw(t);
236 }
237 round_keys[i / 4].repr[i % 4] = round_keys[(i - words_in_key) / 4].repr[(i - words_in_key) % 4] ^ t;
238 }
239 return Self{ .round_keys = round_keys };
240 }
241
242 /// Invert the key schedule.
243 pub fn invert(key_schedule: Self) Self {
244 const round_keys = &key_schedule.round_keys;
245 var inv_round_keys: [rounds + 1]Block = undefined;
246 const total_words = 4 * round_keys.len;
247 var i: usize = 0;
248 while (i < total_words) : (i += 4) {
249 const ei = total_words - i - 4;
250 comptime var j: usize = 0;
251 inline while (j < 4) : (j += 1) {
252 var x = round_keys[(ei + j) / 4].repr[(ei + j) % 4];
253 if (i > 0 and i + 4 < total_words) {
254 x = td0[sbox0[x >> 24]] ^ td1[sbox0[x >> 16 & 0xff]] ^ td2[sbox0[x >> 8 & 0xff]] ^ td3[sbox0[x & 0xff]];
255 }
256 inv_round_keys[(i + j) / 4].repr[(i + j) % 4] = x;
257 }
258 }
259 return Self{ .round_keys = inv_round_keys };
260 }
261 };
262}
263
264/// A context to perform encryption using the standard AES key schedule.
265pub fn AESEncryptCtx(comptime AES: type) type {
266 std.debug.assert(AES.key_bits == 128 or AES.key_bits == 256);
267 const rounds = AES.rounds;
268
269 return struct {
270 const Self = @This();
271 pub const block = AES.block;
272 pub const block_size = block.block_size;
273 key_schedule: KeySchedule(AES),
274
275 /// Create a new encryption context with the given key.
276 pub fn init(key: [AES.key_bits / 8]u8) Self {
277 const key_schedule = KeySchedule(AES).expandKey(key);
278 return Self{
279 .key_schedule = key_schedule,
280 };
281 }
282
283 /// Encrypt a single block.
284 pub fn encrypt(ctx: Self, dst: *[16]u8, src: *const [16]u8) void {
285 const round_keys = ctx.key_schedule.round_keys;
286 var t = Block.fromBytes(src).xor(round_keys[0]);
287 comptime var i = 1;
288 inline while (i < rounds) : (i += 1) {
289 t = t.encrypt(round_keys[i]);
290 }
291 t = t.encryptLast(round_keys[rounds]);
292 dst.* = t.toBytes();
293 }
294
295 /// Encrypt+XOR a single block.
296 pub fn xor(ctx: Self, dst: *[16]u8, src: *const [16]u8, counter: [16]u8) void {
297 const round_keys = ctx.key_schedule.round_keys;
298 var t = Block.fromBytes(&counter).xor(round_keys[0]);
299 comptime var i = 1;
300 inline while (i < rounds) : (i += 1) {
301 t = t.encrypt(round_keys[i]);
302 }
303 t = t.encryptLast(round_keys[rounds]);
304 dst.* = t.xorBytes(src);
305 }
306
307 /// Encrypt multiple blocks, possibly leveraging parallelization.
308 pub fn encryptWide(ctx: Self, comptime count: usize, dst: *[16 * count]u8, src: *const [16 * count]u8) void {
309 var i: usize = 0;
310 while (i < count) : (i += 1) {
311 ctx.encrypt(dst[16 * i .. 16 * i + 16][0..16], src[16 * i .. 16 * i + 16][0..16]);
312 }
313 }
314
315 /// Encrypt+XOR multiple blocks, possibly leveraging parallelization.
316 pub fn xorWide(ctx: Self, comptime count: usize, dst: *[16 * count]u8, src: *const [16 * count]u8, counters: [16 * count]u8) void {
317 var i: usize = 0;
318 while (i < count) : (i += 1) {
319 ctx.xor(dst[16 * i .. 16 * i + 16][0..16], src[16 * i .. 16 * i + 16][0..16], counters[16 * i .. 16 * i + 16][0..16].*);
320 }
321 }
322 };
323}
324
325/// A context to perform decryption using the standard AES key schedule.
326pub fn AESDecryptCtx(comptime AES: type) type {
327 std.debug.assert(AES.key_bits == 128 or AES.key_bits == 256);
328 const rounds = AES.rounds;
329
330 return struct {
331 const Self = @This();
332 pub const block = AES.block;
333 pub const block_size = block.block_size;
334 key_schedule: KeySchedule(AES),
335
336 /// Create a decryption context from an existing encryption context.
337 pub fn initFromEnc(ctx: AESEncryptCtx(AES)) Self {
338 return Self{
339 .key_schedule = ctx.key_schedule.invert(),
340 };
341 }
342
343 /// Create a new decryption context with the given key.
344 pub fn init(key: [AES.key_bits / 8]u8) Self {
345 const enc_ctx = AESEncryptCtx(AES).init(key);
346 return initFromEnc(enc_ctx);
347 }
348
349 /// Decrypt a single block.
350 pub fn decrypt(ctx: Self, dst: *[16]u8, src: *const [16]u8) void {
351 const inv_round_keys = ctx.key_schedule.round_keys;
352 var t = Block.fromBytes(src).xor(inv_round_keys[0]);
353 comptime var i = 1;
354 inline while (i < rounds) : (i += 1) {
355 t = t.decrypt(inv_round_keys[i]);
356 }
357 t = t.decryptLast(inv_round_keys[rounds]);
358 dst.* = t.toBytes();
359 }
360
361 /// Decrypt multiple blocks, possibly leveraging parallelization.
362 pub fn decryptWide(ctx: Self, comptime count: usize, dst: *[16 * count]u8, src: *const [16 * count]u8) void {
363 var i: usize = 0;
364 while (i < count) : (i += 1) {
365 ctx.decrypt(dst[16 * i .. 16 * i + 16][0..16], src[16 * i .. 16 * i + 16][0..16]);
366 }
367 }
368 };
369}
370
371/// AES-128 with the standard key schedule.
372pub const AES128 = struct {
373 pub const key_bits: usize = 128;
374 pub const rounds = ((key_bits - 64) / 32 + 8);
375 pub const block = Block;
376
377 /// Create a new context for encryption.
378 pub fn initEnc(key: [key_bits / 8]u8) AESEncryptCtx(AES128) {
379 return AESEncryptCtx(AES128).init(key);
380 }
381
382 /// Create a new context for decryption.
383 pub fn initDec(key: [key_bits / 8]u8) AESDecryptCtx(AES128) {
384 return AESDecryptCtx(AES128).init(key);
385 }
386};
387
388/// AES-256 with the standard key schedule.
389pub const AES256 = struct {
390 pub const key_bits: usize = 256;
391 pub const rounds = ((key_bits - 64) / 32 + 8);
392 pub const block = Block;
393
394 /// Create a new context for encryption.
395 pub fn initEnc(key: [key_bits / 8]u8) AESEncryptCtx(AES256) {
396 return AESEncryptCtx(AES256).init(key);
397 }
398
399 /// Create a new context for decryption.
400 pub fn initDec(key: [key_bits / 8]u8) AESDecryptCtx(AES256) {
401 return AESDecryptCtx(AES256).init(key);
402 }
403};
404
405// constants
406const powx = [16]u8{
407 0x01,
408 0x02,
409 0x04,
410 0x08,
411 0x10,
412 0x20,
413 0x40,
414 0x80,
415 0x1b,
416 0x36,
417 0x6c,
418 0xd8,
419 0xab,
420 0x4d,
421 0x9a,
422 0x2f,
423};
424
425const sbox0 align(64) = [256]u8{
426 0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5, 0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76,
427 0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0, 0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0,
428 0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc, 0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15,
429 0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a, 0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75,
430 0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0, 0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84,
431 0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b, 0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf,
432 0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85, 0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8,
433 0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5, 0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2,
434 0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17, 0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73,
435 0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88, 0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb,
436 0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c, 0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79,
437 0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9, 0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08,
438 0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6, 0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a,
439 0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e, 0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e,
440 0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94, 0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf,
441 0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68, 0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16,
442};
443
444const sbox1 align(64) = [256]u8{
445 0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38, 0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb,
446 0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87, 0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb,
447 0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d, 0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e,
448 0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2, 0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25,
449 0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16, 0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92,
450 0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda, 0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84,
451 0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a, 0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06,
452 0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02, 0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b,
453 0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea, 0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73,
454 0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85, 0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e,
455 0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89, 0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b,
456 0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20, 0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4,
457 0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31, 0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f,
458 0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d, 0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef,
459 0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0, 0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61,
460 0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26, 0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d,
461};
462
463const te0 align(64) = [256]u32{
464 0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d, 0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554,
465 0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d, 0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a,
466 0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87, 0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b,
467 0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea, 0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b,
468 0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a, 0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f,
469 0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108, 0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f,
470 0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e, 0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5,
471 0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d, 0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f,
472 0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e, 0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb,
473 0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce, 0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497,
474 0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c, 0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed,
475 0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b, 0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a,
476 0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16, 0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594,
477 0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81, 0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3,
478 0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a, 0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504,
479 0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163, 0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d,
480 0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f, 0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739,
481 0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47, 0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395,
482 0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f, 0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883,
483 0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c, 0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76,
484 0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e, 0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4,
485 0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6, 0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b,
486 0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7, 0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0,
487 0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25, 0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818,
488 0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72, 0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651,
489 0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21, 0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85,
490 0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa, 0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12,
491 0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0, 0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9,
492 0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133, 0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7,
493 0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920, 0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a,
494 0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17, 0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8,
495 0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11, 0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a,
496};
497const te1 align(64) = [256]u32{
498 0xa5c66363, 0x84f87c7c, 0x99ee7777, 0x8df67b7b, 0x0dfff2f2, 0xbdd66b6b, 0xb1de6f6f, 0x5491c5c5,
499 0x50603030, 0x03020101, 0xa9ce6767, 0x7d562b2b, 0x19e7fefe, 0x62b5d7d7, 0xe64dabab, 0x9aec7676,
500 0x458fcaca, 0x9d1f8282, 0x4089c9c9, 0x87fa7d7d, 0x15effafa, 0xebb25959, 0xc98e4747, 0x0bfbf0f0,
501 0xec41adad, 0x67b3d4d4, 0xfd5fa2a2, 0xea45afaf, 0xbf239c9c, 0xf753a4a4, 0x96e47272, 0x5b9bc0c0,
502 0xc275b7b7, 0x1ce1fdfd, 0xae3d9393, 0x6a4c2626, 0x5a6c3636, 0x417e3f3f, 0x02f5f7f7, 0x4f83cccc,
503 0x5c683434, 0xf451a5a5, 0x34d1e5e5, 0x08f9f1f1, 0x93e27171, 0x73abd8d8, 0x53623131, 0x3f2a1515,
504 0x0c080404, 0x5295c7c7, 0x65462323, 0x5e9dc3c3, 0x28301818, 0xa1379696, 0x0f0a0505, 0xb52f9a9a,
505 0x090e0707, 0x36241212, 0x9b1b8080, 0x3ddfe2e2, 0x26cdebeb, 0x694e2727, 0xcd7fb2b2, 0x9fea7575,
506 0x1b120909, 0x9e1d8383, 0x74582c2c, 0x2e341a1a, 0x2d361b1b, 0xb2dc6e6e, 0xeeb45a5a, 0xfb5ba0a0,
507 0xf6a45252, 0x4d763b3b, 0x61b7d6d6, 0xce7db3b3, 0x7b522929, 0x3edde3e3, 0x715e2f2f, 0x97138484,
508 0xf5a65353, 0x68b9d1d1, 0x00000000, 0x2cc1eded, 0x60402020, 0x1fe3fcfc, 0xc879b1b1, 0xedb65b5b,
509 0xbed46a6a, 0x468dcbcb, 0xd967bebe, 0x4b723939, 0xde944a4a, 0xd4984c4c, 0xe8b05858, 0x4a85cfcf,
510 0x6bbbd0d0, 0x2ac5efef, 0xe54faaaa, 0x16edfbfb, 0xc5864343, 0xd79a4d4d, 0x55663333, 0x94118585,
511 0xcf8a4545, 0x10e9f9f9, 0x06040202, 0x81fe7f7f, 0xf0a05050, 0x44783c3c, 0xba259f9f, 0xe34ba8a8,
512 0xf3a25151, 0xfe5da3a3, 0xc0804040, 0x8a058f8f, 0xad3f9292, 0xbc219d9d, 0x48703838, 0x04f1f5f5,
513 0xdf63bcbc, 0xc177b6b6, 0x75afdada, 0x63422121, 0x30201010, 0x1ae5ffff, 0x0efdf3f3, 0x6dbfd2d2,
514 0x4c81cdcd, 0x14180c0c, 0x35261313, 0x2fc3ecec, 0xe1be5f5f, 0xa2359797, 0xcc884444, 0x392e1717,
515 0x5793c4c4, 0xf255a7a7, 0x82fc7e7e, 0x477a3d3d, 0xacc86464, 0xe7ba5d5d, 0x2b321919, 0x95e67373,
516 0xa0c06060, 0x98198181, 0xd19e4f4f, 0x7fa3dcdc, 0x66442222, 0x7e542a2a, 0xab3b9090, 0x830b8888,
517 0xca8c4646, 0x29c7eeee, 0xd36bb8b8, 0x3c281414, 0x79a7dede, 0xe2bc5e5e, 0x1d160b0b, 0x76addbdb,
518 0x3bdbe0e0, 0x56643232, 0x4e743a3a, 0x1e140a0a, 0xdb924949, 0x0a0c0606, 0x6c482424, 0xe4b85c5c,
519 0x5d9fc2c2, 0x6ebdd3d3, 0xef43acac, 0xa6c46262, 0xa8399191, 0xa4319595, 0x37d3e4e4, 0x8bf27979,
520 0x32d5e7e7, 0x438bc8c8, 0x596e3737, 0xb7da6d6d, 0x8c018d8d, 0x64b1d5d5, 0xd29c4e4e, 0xe049a9a9,
521 0xb4d86c6c, 0xfaac5656, 0x07f3f4f4, 0x25cfeaea, 0xafca6565, 0x8ef47a7a, 0xe947aeae, 0x18100808,
522 0xd56fbaba, 0x88f07878, 0x6f4a2525, 0x725c2e2e, 0x24381c1c, 0xf157a6a6, 0xc773b4b4, 0x5197c6c6,
523 0x23cbe8e8, 0x7ca1dddd, 0x9ce87474, 0x213e1f1f, 0xdd964b4b, 0xdc61bdbd, 0x860d8b8b, 0x850f8a8a,
524 0x90e07070, 0x427c3e3e, 0xc471b5b5, 0xaacc6666, 0xd8904848, 0x05060303, 0x01f7f6f6, 0x121c0e0e,
525 0xa3c26161, 0x5f6a3535, 0xf9ae5757, 0xd069b9b9, 0x91178686, 0x5899c1c1, 0x273a1d1d, 0xb9279e9e,
526 0x38d9e1e1, 0x13ebf8f8, 0xb32b9898, 0x33221111, 0xbbd26969, 0x70a9d9d9, 0x89078e8e, 0xa7339494,
527 0xb62d9b9b, 0x223c1e1e, 0x92158787, 0x20c9e9e9, 0x4987cece, 0xffaa5555, 0x78502828, 0x7aa5dfdf,
528 0x8f038c8c, 0xf859a1a1, 0x80098989, 0x171a0d0d, 0xda65bfbf, 0x31d7e6e6, 0xc6844242, 0xb8d06868,
529 0xc3824141, 0xb0299999, 0x775a2d2d, 0x111e0f0f, 0xcb7bb0b0, 0xfca85454, 0xd66dbbbb, 0x3a2c1616,
530};
531const te2 align(64) = [256]u32{
532 0x63a5c663, 0x7c84f87c, 0x7799ee77, 0x7b8df67b, 0xf20dfff2, 0x6bbdd66b, 0x6fb1de6f, 0xc55491c5,
533 0x30506030, 0x01030201, 0x67a9ce67, 0x2b7d562b, 0xfe19e7fe, 0xd762b5d7, 0xabe64dab, 0x769aec76,
534 0xca458fca, 0x829d1f82, 0xc94089c9, 0x7d87fa7d, 0xfa15effa, 0x59ebb259, 0x47c98e47, 0xf00bfbf0,
535 0xadec41ad, 0xd467b3d4, 0xa2fd5fa2, 0xafea45af, 0x9cbf239c, 0xa4f753a4, 0x7296e472, 0xc05b9bc0,
536 0xb7c275b7, 0xfd1ce1fd, 0x93ae3d93, 0x266a4c26, 0x365a6c36, 0x3f417e3f, 0xf702f5f7, 0xcc4f83cc,
537 0x345c6834, 0xa5f451a5, 0xe534d1e5, 0xf108f9f1, 0x7193e271, 0xd873abd8, 0x31536231, 0x153f2a15,
538 0x040c0804, 0xc75295c7, 0x23654623, 0xc35e9dc3, 0x18283018, 0x96a13796, 0x050f0a05, 0x9ab52f9a,
539 0x07090e07, 0x12362412, 0x809b1b80, 0xe23ddfe2, 0xeb26cdeb, 0x27694e27, 0xb2cd7fb2, 0x759fea75,
540 0x091b1209, 0x839e1d83, 0x2c74582c, 0x1a2e341a, 0x1b2d361b, 0x6eb2dc6e, 0x5aeeb45a, 0xa0fb5ba0,
541 0x52f6a452, 0x3b4d763b, 0xd661b7d6, 0xb3ce7db3, 0x297b5229, 0xe33edde3, 0x2f715e2f, 0x84971384,
542 0x53f5a653, 0xd168b9d1, 0x00000000, 0xed2cc1ed, 0x20604020, 0xfc1fe3fc, 0xb1c879b1, 0x5bedb65b,
543 0x6abed46a, 0xcb468dcb, 0xbed967be, 0x394b7239, 0x4ade944a, 0x4cd4984c, 0x58e8b058, 0xcf4a85cf,
544 0xd06bbbd0, 0xef2ac5ef, 0xaae54faa, 0xfb16edfb, 0x43c58643, 0x4dd79a4d, 0x33556633, 0x85941185,
545 0x45cf8a45, 0xf910e9f9, 0x02060402, 0x7f81fe7f, 0x50f0a050, 0x3c44783c, 0x9fba259f, 0xa8e34ba8,
546 0x51f3a251, 0xa3fe5da3, 0x40c08040, 0x8f8a058f, 0x92ad3f92, 0x9dbc219d, 0x38487038, 0xf504f1f5,
547 0xbcdf63bc, 0xb6c177b6, 0xda75afda, 0x21634221, 0x10302010, 0xff1ae5ff, 0xf30efdf3, 0xd26dbfd2,
548 0xcd4c81cd, 0x0c14180c, 0x13352613, 0xec2fc3ec, 0x5fe1be5f, 0x97a23597, 0x44cc8844, 0x17392e17,
549 0xc45793c4, 0xa7f255a7, 0x7e82fc7e, 0x3d477a3d, 0x64acc864, 0x5de7ba5d, 0x192b3219, 0x7395e673,
550 0x60a0c060, 0x81981981, 0x4fd19e4f, 0xdc7fa3dc, 0x22664422, 0x2a7e542a, 0x90ab3b90, 0x88830b88,
551 0x46ca8c46, 0xee29c7ee, 0xb8d36bb8, 0x143c2814, 0xde79a7de, 0x5ee2bc5e, 0x0b1d160b, 0xdb76addb,
552 0xe03bdbe0, 0x32566432, 0x3a4e743a, 0x0a1e140a, 0x49db9249, 0x060a0c06, 0x246c4824, 0x5ce4b85c,
553 0xc25d9fc2, 0xd36ebdd3, 0xacef43ac, 0x62a6c462, 0x91a83991, 0x95a43195, 0xe437d3e4, 0x798bf279,
554 0xe732d5e7, 0xc8438bc8, 0x37596e37, 0x6db7da6d, 0x8d8c018d, 0xd564b1d5, 0x4ed29c4e, 0xa9e049a9,
555 0x6cb4d86c, 0x56faac56, 0xf407f3f4, 0xea25cfea, 0x65afca65, 0x7a8ef47a, 0xaee947ae, 0x08181008,
556 0xbad56fba, 0x7888f078, 0x256f4a25, 0x2e725c2e, 0x1c24381c, 0xa6f157a6, 0xb4c773b4, 0xc65197c6,
557 0xe823cbe8, 0xdd7ca1dd, 0x749ce874, 0x1f213e1f, 0x4bdd964b, 0xbddc61bd, 0x8b860d8b, 0x8a850f8a,
558 0x7090e070, 0x3e427c3e, 0xb5c471b5, 0x66aacc66, 0x48d89048, 0x03050603, 0xf601f7f6, 0x0e121c0e,
559 0x61a3c261, 0x355f6a35, 0x57f9ae57, 0xb9d069b9, 0x86911786, 0xc15899c1, 0x1d273a1d, 0x9eb9279e,
560 0xe138d9e1, 0xf813ebf8, 0x98b32b98, 0x11332211, 0x69bbd269, 0xd970a9d9, 0x8e89078e, 0x94a73394,
561 0x9bb62d9b, 0x1e223c1e, 0x87921587, 0xe920c9e9, 0xce4987ce, 0x55ffaa55, 0x28785028, 0xdf7aa5df,
562 0x8c8f038c, 0xa1f859a1, 0x89800989, 0x0d171a0d, 0xbfda65bf, 0xe631d7e6, 0x42c68442, 0x68b8d068,
563 0x41c38241, 0x99b02999, 0x2d775a2d, 0x0f111e0f, 0xb0cb7bb0, 0x54fca854, 0xbbd66dbb, 0x163a2c16,
564};
565const te3 align(64) = [256]u32{
566 0x6363a5c6, 0x7c7c84f8, 0x777799ee, 0x7b7b8df6, 0xf2f20dff, 0x6b6bbdd6, 0x6f6fb1de, 0xc5c55491,
567 0x30305060, 0x01010302, 0x6767a9ce, 0x2b2b7d56, 0xfefe19e7, 0xd7d762b5, 0xababe64d, 0x76769aec,
568 0xcaca458f, 0x82829d1f, 0xc9c94089, 0x7d7d87fa, 0xfafa15ef, 0x5959ebb2, 0x4747c98e, 0xf0f00bfb,
569 0xadadec41, 0xd4d467b3, 0xa2a2fd5f, 0xafafea45, 0x9c9cbf23, 0xa4a4f753, 0x727296e4, 0xc0c05b9b,
570 0xb7b7c275, 0xfdfd1ce1, 0x9393ae3d, 0x26266a4c, 0x36365a6c, 0x3f3f417e, 0xf7f702f5, 0xcccc4f83,
571 0x34345c68, 0xa5a5f451, 0xe5e534d1, 0xf1f108f9, 0x717193e2, 0xd8d873ab, 0x31315362, 0x15153f2a,
572 0x04040c08, 0xc7c75295, 0x23236546, 0xc3c35e9d, 0x18182830, 0x9696a137, 0x05050f0a, 0x9a9ab52f,
573 0x0707090e, 0x12123624, 0x80809b1b, 0xe2e23ddf, 0xebeb26cd, 0x2727694e, 0xb2b2cd7f, 0x75759fea,
574 0x09091b12, 0x83839e1d, 0x2c2c7458, 0x1a1a2e34, 0x1b1b2d36, 0x6e6eb2dc, 0x5a5aeeb4, 0xa0a0fb5b,
575 0x5252f6a4, 0x3b3b4d76, 0xd6d661b7, 0xb3b3ce7d, 0x29297b52, 0xe3e33edd, 0x2f2f715e, 0x84849713,
576 0x5353f5a6, 0xd1d168b9, 0x00000000, 0xeded2cc1, 0x20206040, 0xfcfc1fe3, 0xb1b1c879, 0x5b5bedb6,
577 0x6a6abed4, 0xcbcb468d, 0xbebed967, 0x39394b72, 0x4a4ade94, 0x4c4cd498, 0x5858e8b0, 0xcfcf4a85,
578 0xd0d06bbb, 0xefef2ac5, 0xaaaae54f, 0xfbfb16ed, 0x4343c586, 0x4d4dd79a, 0x33335566, 0x85859411,
579 0x4545cf8a, 0xf9f910e9, 0x02020604, 0x7f7f81fe, 0x5050f0a0, 0x3c3c4478, 0x9f9fba25, 0xa8a8e34b,
580 0x5151f3a2, 0xa3a3fe5d, 0x4040c080, 0x8f8f8a05, 0x9292ad3f, 0x9d9dbc21, 0x38384870, 0xf5f504f1,
581 0xbcbcdf63, 0xb6b6c177, 0xdada75af, 0x21216342, 0x10103020, 0xffff1ae5, 0xf3f30efd, 0xd2d26dbf,
582 0xcdcd4c81, 0x0c0c1418, 0x13133526, 0xecec2fc3, 0x5f5fe1be, 0x9797a235, 0x4444cc88, 0x1717392e,
583 0xc4c45793, 0xa7a7f255, 0x7e7e82fc, 0x3d3d477a, 0x6464acc8, 0x5d5de7ba, 0x19192b32, 0x737395e6,
584 0x6060a0c0, 0x81819819, 0x4f4fd19e, 0xdcdc7fa3, 0x22226644, 0x2a2a7e54, 0x9090ab3b, 0x8888830b,
585 0x4646ca8c, 0xeeee29c7, 0xb8b8d36b, 0x14143c28, 0xdede79a7, 0x5e5ee2bc, 0x0b0b1d16, 0xdbdb76ad,
586 0xe0e03bdb, 0x32325664, 0x3a3a4e74, 0x0a0a1e14, 0x4949db92, 0x06060a0c, 0x24246c48, 0x5c5ce4b8,
587 0xc2c25d9f, 0xd3d36ebd, 0xacacef43, 0x6262a6c4, 0x9191a839, 0x9595a431, 0xe4e437d3, 0x79798bf2,
588 0xe7e732d5, 0xc8c8438b, 0x3737596e, 0x6d6db7da, 0x8d8d8c01, 0xd5d564b1, 0x4e4ed29c, 0xa9a9e049,
589 0x6c6cb4d8, 0x5656faac, 0xf4f407f3, 0xeaea25cf, 0x6565afca, 0x7a7a8ef4, 0xaeaee947, 0x08081810,
590 0xbabad56f, 0x787888f0, 0x25256f4a, 0x2e2e725c, 0x1c1c2438, 0xa6a6f157, 0xb4b4c773, 0xc6c65197,
591 0xe8e823cb, 0xdddd7ca1, 0x74749ce8, 0x1f1f213e, 0x4b4bdd96, 0xbdbddc61, 0x8b8b860d, 0x8a8a850f,
592 0x707090e0, 0x3e3e427c, 0xb5b5c471, 0x6666aacc, 0x4848d890, 0x03030506, 0xf6f601f7, 0x0e0e121c,
593 0x6161a3c2, 0x35355f6a, 0x5757f9ae, 0xb9b9d069, 0x86869117, 0xc1c15899, 0x1d1d273a, 0x9e9eb927,
594 0xe1e138d9, 0xf8f813eb, 0x9898b32b, 0x11113322, 0x6969bbd2, 0xd9d970a9, 0x8e8e8907, 0x9494a733,
595 0x9b9bb62d, 0x1e1e223c, 0x87879215, 0xe9e920c9, 0xcece4987, 0x5555ffaa, 0x28287850, 0xdfdf7aa5,
596 0x8c8c8f03, 0xa1a1f859, 0x89898009, 0x0d0d171a, 0xbfbfda65, 0xe6e631d7, 0x4242c684, 0x6868b8d0,
597 0x4141c382, 0x9999b029, 0x2d2d775a, 0x0f0f111e, 0xb0b0cb7b, 0x5454fca8, 0xbbbbd66d, 0x16163a2c,
598};
599
600const td0 align(64) = [256]u32{
601 0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96, 0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393,
602 0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25, 0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f,
603 0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1, 0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6,
604 0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da, 0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844,
605 0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd, 0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4,
606 0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45, 0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94,
607 0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7, 0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a,
608 0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5, 0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c,
609 0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1, 0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a,
610 0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75, 0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051,
611 0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46, 0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff,
612 0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77, 0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb,
613 0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000, 0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e,
614 0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927, 0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a,
615 0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e, 0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16,
616 0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d, 0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8,
617 0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd, 0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34,
618 0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163, 0xd731dcca, 0x42638510, 0x13972240, 0x84c61120,
619 0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d, 0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0,
620 0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422, 0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef,
621 0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36, 0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4,
622 0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662, 0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5,
623 0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3, 0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b,
624 0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8, 0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6,
625 0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6, 0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0,
626 0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815, 0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f,
627 0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df, 0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f,
628 0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e, 0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713,
629 0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89, 0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c,
630 0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf, 0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86,
631 0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f, 0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541,
632 0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190, 0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742,
633};
634const td1 align(64) = [256]u32{
635 0x5051f4a7, 0x537e4165, 0xc31a17a4, 0x963a275e, 0xcb3bab6b, 0xf11f9d45, 0xabacfa58, 0x934be303,
636 0x552030fa, 0xf6ad766d, 0x9188cc76, 0x25f5024c, 0xfc4fe5d7, 0xd7c52acb, 0x80263544, 0x8fb562a3,
637 0x49deb15a, 0x6725ba1b, 0x9845ea0e, 0xe15dfec0, 0x02c32f75, 0x12814cf0, 0xa38d4697, 0xc66bd3f9,
638 0xe7038f5f, 0x9515929c, 0xebbf6d7a, 0xda955259, 0x2dd4be83, 0xd3587421, 0x2949e069, 0x448ec9c8,
639 0x6a75c289, 0x78f48e79, 0x6b99583e, 0xdd27b971, 0xb6bee14f, 0x17f088ad, 0x66c920ac, 0xb47dce3a,
640 0x1863df4a, 0x82e51a31, 0x60975133, 0x4562537f, 0xe0b16477, 0x84bb6bae, 0x1cfe81a0, 0x94f9082b,
641 0x58704868, 0x198f45fd, 0x8794de6c, 0xb7527bf8, 0x23ab73d3, 0xe2724b02, 0x57e31f8f, 0x2a6655ab,
642 0x07b2eb28, 0x032fb5c2, 0x9a86c57b, 0xa5d33708, 0xf2302887, 0xb223bfa5, 0xba02036a, 0x5ced1682,
643 0x2b8acf1c, 0x92a779b4, 0xf0f307f2, 0xa14e69e2, 0xcd65daf4, 0xd50605be, 0x1fd13462, 0x8ac4a6fe,
644 0x9d342e53, 0xa0a2f355, 0x32058ae1, 0x75a4f6eb, 0x390b83ec, 0xaa4060ef, 0x065e719f, 0x51bd6e10,
645 0xf93e218a, 0x3d96dd06, 0xaedd3e05, 0x464de6bd, 0xb591548d, 0x0571c45d, 0x6f0406d4, 0xff605015,
646 0x241998fb, 0x97d6bde9, 0xcc894043, 0x7767d99e, 0xbdb0e842, 0x8807898b, 0x38e7195b, 0xdb79c8ee,
647 0x47a17c0a, 0xe97c420f, 0xc9f8841e, 0x00000000, 0x83098086, 0x48322bed, 0xac1e1170, 0x4e6c5a72,
648 0xfbfd0eff, 0x560f8538, 0x1e3daed5, 0x27362d39, 0x640a0fd9, 0x21685ca6, 0xd19b5b54, 0x3a24362e,
649 0xb10c0a67, 0x0f9357e7, 0xd2b4ee96, 0x9e1b9b91, 0x4f80c0c5, 0xa261dc20, 0x695a774b, 0x161c121a,
650 0x0ae293ba, 0xe5c0a02a, 0x433c22e0, 0x1d121b17, 0x0b0e090d, 0xadf28bc7, 0xb92db6a8, 0xc8141ea9,
651 0x8557f119, 0x4caf7507, 0xbbee99dd, 0xfda37f60, 0x9ff70126, 0xbc5c72f5, 0xc544663b, 0x345bfb7e,
652 0x768b4329, 0xdccb23c6, 0x68b6edfc, 0x63b8e4f1, 0xcad731dc, 0x10426385, 0x40139722, 0x2084c611,
653 0x7d854a24, 0xf8d2bb3d, 0x11aef932, 0x6dc729a1, 0x4b1d9e2f, 0xf3dcb230, 0xec0d8652, 0xd077c1e3,
654 0x6c2bb316, 0x99a970b9, 0xfa119448, 0x2247e964, 0xc4a8fc8c, 0x1aa0f03f, 0xd8567d2c, 0xef223390,
655 0xc787494e, 0xc1d938d1, 0xfe8ccaa2, 0x3698d40b, 0xcfa6f581, 0x28a57ade, 0x26dab78e, 0xa43fadbf,
656 0xe42c3a9d, 0x0d507892, 0x9b6a5fcc, 0x62547e46, 0xc2f68d13, 0xe890d8b8, 0x5e2e39f7, 0xf582c3af,
657 0xbe9f5d80, 0x7c69d093, 0xa96fd52d, 0xb3cf2512, 0x3bc8ac99, 0xa710187d, 0x6ee89c63, 0x7bdb3bbb,
658 0x09cd2678, 0xf46e5918, 0x01ec9ab7, 0xa8834f9a, 0x65e6956e, 0x7eaaffe6, 0x0821bccf, 0xe6ef15e8,
659 0xd9bae79b, 0xce4a6f36, 0xd4ea9f09, 0xd629b07c, 0xaf31a4b2, 0x312a3f23, 0x30c6a594, 0xc035a266,
660 0x37744ebc, 0xa6fc82ca, 0xb0e090d0, 0x1533a7d8, 0x4af10498, 0xf741ecda, 0x0e7fcd50, 0x2f1791f6,
661 0x8d764dd6, 0x4d43efb0, 0x54ccaa4d, 0xdfe49604, 0xe39ed1b5, 0x1b4c6a88, 0xb8c12c1f, 0x7f466551,
662 0x049d5eea, 0x5d018c35, 0x73fa8774, 0x2efb0b41, 0x5ab3671d, 0x5292dbd2, 0x33e91056, 0x136dd647,
663 0x8c9ad761, 0x7a37a10c, 0x8e59f814, 0x89eb133c, 0xeecea927, 0x35b761c9, 0xede11ce5, 0x3c7a47b1,
664 0x599cd2df, 0x3f55f273, 0x791814ce, 0xbf73c737, 0xea53f7cd, 0x5b5ffdaa, 0x14df3d6f, 0x867844db,
665 0x81caaff3, 0x3eb968c4, 0x2c382434, 0x5fc2a340, 0x72161dc3, 0x0cbce225, 0x8b283c49, 0x41ff0d95,
666 0x7139a801, 0xde080cb3, 0x9cd8b4e4, 0x906456c1, 0x617bcb84, 0x70d532b6, 0x74486c5c, 0x42d0b857,
667};
668const td2 align(64) = [256]u32{
669 0xa75051f4, 0x65537e41, 0xa4c31a17, 0x5e963a27, 0x6bcb3bab, 0x45f11f9d, 0x58abacfa, 0x03934be3,
670 0xfa552030, 0x6df6ad76, 0x769188cc, 0x4c25f502, 0xd7fc4fe5, 0xcbd7c52a, 0x44802635, 0xa38fb562,
671 0x5a49deb1, 0x1b6725ba, 0x0e9845ea, 0xc0e15dfe, 0x7502c32f, 0xf012814c, 0x97a38d46, 0xf9c66bd3,
672 0x5fe7038f, 0x9c951592, 0x7aebbf6d, 0x59da9552, 0x832dd4be, 0x21d35874, 0x692949e0, 0xc8448ec9,
673 0x896a75c2, 0x7978f48e, 0x3e6b9958, 0x71dd27b9, 0x4fb6bee1, 0xad17f088, 0xac66c920, 0x3ab47dce,
674 0x4a1863df, 0x3182e51a, 0x33609751, 0x7f456253, 0x77e0b164, 0xae84bb6b, 0xa01cfe81, 0x2b94f908,
675 0x68587048, 0xfd198f45, 0x6c8794de, 0xf8b7527b, 0xd323ab73, 0x02e2724b, 0x8f57e31f, 0xab2a6655,
676 0x2807b2eb, 0xc2032fb5, 0x7b9a86c5, 0x08a5d337, 0x87f23028, 0xa5b223bf, 0x6aba0203, 0x825ced16,
677 0x1c2b8acf, 0xb492a779, 0xf2f0f307, 0xe2a14e69, 0xf4cd65da, 0xbed50605, 0x621fd134, 0xfe8ac4a6,
678 0x539d342e, 0x55a0a2f3, 0xe132058a, 0xeb75a4f6, 0xec390b83, 0xefaa4060, 0x9f065e71, 0x1051bd6e,
679 0x8af93e21, 0x063d96dd, 0x05aedd3e, 0xbd464de6, 0x8db59154, 0x5d0571c4, 0xd46f0406, 0x15ff6050,
680 0xfb241998, 0xe997d6bd, 0x43cc8940, 0x9e7767d9, 0x42bdb0e8, 0x8b880789, 0x5b38e719, 0xeedb79c8,
681 0x0a47a17c, 0x0fe97c42, 0x1ec9f884, 0x00000000, 0x86830980, 0xed48322b, 0x70ac1e11, 0x724e6c5a,
682 0xfffbfd0e, 0x38560f85, 0xd51e3dae, 0x3927362d, 0xd9640a0f, 0xa621685c, 0x54d19b5b, 0x2e3a2436,
683 0x67b10c0a, 0xe70f9357, 0x96d2b4ee, 0x919e1b9b, 0xc54f80c0, 0x20a261dc, 0x4b695a77, 0x1a161c12,
684 0xba0ae293, 0x2ae5c0a0, 0xe0433c22, 0x171d121b, 0x0d0b0e09, 0xc7adf28b, 0xa8b92db6, 0xa9c8141e,
685 0x198557f1, 0x074caf75, 0xddbbee99, 0x60fda37f, 0x269ff701, 0xf5bc5c72, 0x3bc54466, 0x7e345bfb,
686 0x29768b43, 0xc6dccb23, 0xfc68b6ed, 0xf163b8e4, 0xdccad731, 0x85104263, 0x22401397, 0x112084c6,
687 0x247d854a, 0x3df8d2bb, 0x3211aef9, 0xa16dc729, 0x2f4b1d9e, 0x30f3dcb2, 0x52ec0d86, 0xe3d077c1,
688 0x166c2bb3, 0xb999a970, 0x48fa1194, 0x642247e9, 0x8cc4a8fc, 0x3f1aa0f0, 0x2cd8567d, 0x90ef2233,
689 0x4ec78749, 0xd1c1d938, 0xa2fe8cca, 0x0b3698d4, 0x81cfa6f5, 0xde28a57a, 0x8e26dab7, 0xbfa43fad,
690 0x9de42c3a, 0x920d5078, 0xcc9b6a5f, 0x4662547e, 0x13c2f68d, 0xb8e890d8, 0xf75e2e39, 0xaff582c3,
691 0x80be9f5d, 0x937c69d0, 0x2da96fd5, 0x12b3cf25, 0x993bc8ac, 0x7da71018, 0x636ee89c, 0xbb7bdb3b,
692 0x7809cd26, 0x18f46e59, 0xb701ec9a, 0x9aa8834f, 0x6e65e695, 0xe67eaaff, 0xcf0821bc, 0xe8e6ef15,
693 0x9bd9bae7, 0x36ce4a6f, 0x09d4ea9f, 0x7cd629b0, 0xb2af31a4, 0x23312a3f, 0x9430c6a5, 0x66c035a2,
694 0xbc37744e, 0xcaa6fc82, 0xd0b0e090, 0xd81533a7, 0x984af104, 0xdaf741ec, 0x500e7fcd, 0xf62f1791,
695 0xd68d764d, 0xb04d43ef, 0x4d54ccaa, 0x04dfe496, 0xb5e39ed1, 0x881b4c6a, 0x1fb8c12c, 0x517f4665,
696 0xea049d5e, 0x355d018c, 0x7473fa87, 0x412efb0b, 0x1d5ab367, 0xd25292db, 0x5633e910, 0x47136dd6,
697 0x618c9ad7, 0x0c7a37a1, 0x148e59f8, 0x3c89eb13, 0x27eecea9, 0xc935b761, 0xe5ede11c, 0xb13c7a47,
698 0xdf599cd2, 0x733f55f2, 0xce791814, 0x37bf73c7, 0xcdea53f7, 0xaa5b5ffd, 0x6f14df3d, 0xdb867844,
699 0xf381caaf, 0xc43eb968, 0x342c3824, 0x405fc2a3, 0xc372161d, 0x250cbce2, 0x498b283c, 0x9541ff0d,
700 0x017139a8, 0xb3de080c, 0xe49cd8b4, 0xc1906456, 0x84617bcb, 0xb670d532, 0x5c74486c, 0x5742d0b8,
701};
702const td3 align(64) = [256]u32{
703 0xf4a75051, 0x4165537e, 0x17a4c31a, 0x275e963a, 0xab6bcb3b, 0x9d45f11f, 0xfa58abac, 0xe303934b,
704 0x30fa5520, 0x766df6ad, 0xcc769188, 0x024c25f5, 0xe5d7fc4f, 0x2acbd7c5, 0x35448026, 0x62a38fb5,
705 0xb15a49de, 0xba1b6725, 0xea0e9845, 0xfec0e15d, 0x2f7502c3, 0x4cf01281, 0x4697a38d, 0xd3f9c66b,
706 0x8f5fe703, 0x929c9515, 0x6d7aebbf, 0x5259da95, 0xbe832dd4, 0x7421d358, 0xe0692949, 0xc9c8448e,
707 0xc2896a75, 0x8e7978f4, 0x583e6b99, 0xb971dd27, 0xe14fb6be, 0x88ad17f0, 0x20ac66c9, 0xce3ab47d,
708 0xdf4a1863, 0x1a3182e5, 0x51336097, 0x537f4562, 0x6477e0b1, 0x6bae84bb, 0x81a01cfe, 0x082b94f9,
709 0x48685870, 0x45fd198f, 0xde6c8794, 0x7bf8b752, 0x73d323ab, 0x4b02e272, 0x1f8f57e3, 0x55ab2a66,
710 0xeb2807b2, 0xb5c2032f, 0xc57b9a86, 0x3708a5d3, 0x2887f230, 0xbfa5b223, 0x036aba02, 0x16825ced,
711 0xcf1c2b8a, 0x79b492a7, 0x07f2f0f3, 0x69e2a14e, 0xdaf4cd65, 0x05bed506, 0x34621fd1, 0xa6fe8ac4,
712 0x2e539d34, 0xf355a0a2, 0x8ae13205, 0xf6eb75a4, 0x83ec390b, 0x60efaa40, 0x719f065e, 0x6e1051bd,
713 0x218af93e, 0xdd063d96, 0x3e05aedd, 0xe6bd464d, 0x548db591, 0xc45d0571, 0x06d46f04, 0x5015ff60,
714 0x98fb2419, 0xbde997d6, 0x4043cc89, 0xd99e7767, 0xe842bdb0, 0x898b8807, 0x195b38e7, 0xc8eedb79,
715 0x7c0a47a1, 0x420fe97c, 0x841ec9f8, 0x00000000, 0x80868309, 0x2bed4832, 0x1170ac1e, 0x5a724e6c,
716 0x0efffbfd, 0x8538560f, 0xaed51e3d, 0x2d392736, 0x0fd9640a, 0x5ca62168, 0x5b54d19b, 0x362e3a24,
717 0x0a67b10c, 0x57e70f93, 0xee96d2b4, 0x9b919e1b, 0xc0c54f80, 0xdc20a261, 0x774b695a, 0x121a161c,
718 0x93ba0ae2, 0xa02ae5c0, 0x22e0433c, 0x1b171d12, 0x090d0b0e, 0x8bc7adf2, 0xb6a8b92d, 0x1ea9c814,
719 0xf1198557, 0x75074caf, 0x99ddbbee, 0x7f60fda3, 0x01269ff7, 0x72f5bc5c, 0x663bc544, 0xfb7e345b,
720 0x4329768b, 0x23c6dccb, 0xedfc68b6, 0xe4f163b8, 0x31dccad7, 0x63851042, 0x97224013, 0xc6112084,
721 0x4a247d85, 0xbb3df8d2, 0xf93211ae, 0x29a16dc7, 0x9e2f4b1d, 0xb230f3dc, 0x8652ec0d, 0xc1e3d077,
722 0xb3166c2b, 0x70b999a9, 0x9448fa11, 0xe9642247, 0xfc8cc4a8, 0xf03f1aa0, 0x7d2cd856, 0x3390ef22,
723 0x494ec787, 0x38d1c1d9, 0xcaa2fe8c, 0xd40b3698, 0xf581cfa6, 0x7ade28a5, 0xb78e26da, 0xadbfa43f,
724 0x3a9de42c, 0x78920d50, 0x5fcc9b6a, 0x7e466254, 0x8d13c2f6, 0xd8b8e890, 0x39f75e2e, 0xc3aff582,
725 0x5d80be9f, 0xd0937c69, 0xd52da96f, 0x2512b3cf, 0xac993bc8, 0x187da710, 0x9c636ee8, 0x3bbb7bdb,
726 0x267809cd, 0x5918f46e, 0x9ab701ec, 0x4f9aa883, 0x956e65e6, 0xffe67eaa, 0xbccf0821, 0x15e8e6ef,
727 0xe79bd9ba, 0x6f36ce4a, 0x9f09d4ea, 0xb07cd629, 0xa4b2af31, 0x3f23312a, 0xa59430c6, 0xa266c035,
728 0x4ebc3774, 0x82caa6fc, 0x90d0b0e0, 0xa7d81533, 0x04984af1, 0xecdaf741, 0xcd500e7f, 0x91f62f17,
729 0x4dd68d76, 0xefb04d43, 0xaa4d54cc, 0x9604dfe4, 0xd1b5e39e, 0x6a881b4c, 0x2c1fb8c1, 0x65517f46,
730 0x5eea049d, 0x8c355d01, 0x877473fa, 0x0b412efb, 0x671d5ab3, 0xdbd25292, 0x105633e9, 0xd647136d,
731 0xd7618c9a, 0xa10c7a37, 0xf8148e59, 0x133c89eb, 0xa927eece, 0x61c935b7, 0x1ce5ede1, 0x47b13c7a,
732 0xd2df599c, 0xf2733f55, 0x14ce7918, 0xc737bf73, 0xf7cdea53, 0xfdaa5b5f, 0x3d6f14df, 0x44db8678,
733 0xaff381ca, 0x68c43eb9, 0x24342c38, 0xa3405fc2, 0x1dc37216, 0xe2250cbc, 0x3c498b28, 0x0d9541ff,
734 0xa8017139, 0x0cb3de08, 0xb4e49cd8, 0x56c19064, 0xcb84617b, 0x32b670d5, 0x6c5c7448, 0xb85742d0,
735};
lib/std/crypto/benchmark.zig+77-5
...@@ -179,6 +179,64 @@ pub fn benchmarkAead(comptime Aead: anytype, comptime bytes: comptime_int) !u64...@@ -179,6 +179,64 @@ pub fn benchmarkAead(comptime Aead: anytype, comptime bytes: comptime_int) !u64
179 return throughput;179 return throughput;
180}180}
181181
182const aes = [_]Crypto{
183 Crypto{ .ty = crypto.core.aes.AES128, .name = "aes128-single" },
184 Crypto{ .ty = crypto.core.aes.AES256, .name = "aes256-single" },
185};
186
187pub fn benchmarkAES(comptime AES: anytype, comptime count: comptime_int) !u64 {
188 var key: [AES.key_bits / 8]u8 = undefined;
189 prng.random.bytes(key[0..]);
190 const ctx = AES.initEnc(key);
191
192 var in = [_]u8{0} ** 16;
193
194 var timer = try Timer.start();
195 const start = timer.lap();
196 {
197 var i: usize = 0;
198 while (i < count) : (i += 1) {
199 ctx.encrypt(&in, &in);
200 }
201 }
202 mem.doNotOptimizeAway(&in);
203 const end = timer.read();
204
205 const elapsed_s = @intToFloat(f64, end - start) / time.ns_per_s;
206 const throughput = @floatToInt(u64, count / elapsed_s);
207
208 return throughput;
209}
210
211const aes8 = [_]Crypto{
212 Crypto{ .ty = crypto.core.aes.AES128, .name = "aes128-8" },
213 Crypto{ .ty = crypto.core.aes.AES256, .name = "aes256-8" },
214};
215
216pub fn benchmarkAES8(comptime AES: anytype, comptime count: comptime_int) !u64 {
217 var key: [AES.key_bits / 8]u8 = undefined;
218 prng.random.bytes(key[0..]);
219 const ctx = AES.initEnc(key);
220
221 var in = [_]u8{0} ** (8 * 16);
222
223 var timer = try Timer.start();
224 const start = timer.lap();
225 {
226 var i: usize = 0;
227 while (i < count) : (i += 1) {
228 ctx.encryptWide(8, &in, &in);
229 }
230 }
231 mem.doNotOptimizeAway(&in);
232 const end = timer.read();
233
234 const elapsed_s = @intToFloat(f64, end - start) / time.ns_per_s;
235 const throughput = @floatToInt(u64, 8 * count / elapsed_s);
236
237 return throughput;
238}
239
182fn usage() void {240fn usage() void {
183 std.debug.warn(241 std.debug.warn(
184 \\throughput_test [options]242 \\throughput_test [options]
...@@ -238,35 +296,49 @@ pub fn main() !void {...@@ -238,35 +296,49 @@ pub fn main() !void {
238 inline for (hashes) |H| {296 inline for (hashes) |H| {
239 if (filter == null or std.mem.indexOf(u8, H.name, filter.?) != null) {297 if (filter == null or std.mem.indexOf(u8, H.name, filter.?) != null) {
240 const throughput = try benchmarkHash(H.ty, mode(128 * MiB));298 const throughput = try benchmarkHash(H.ty, mode(128 * MiB));
241 try stdout.print("{:>17}: {:7} MiB/s\n", .{ H.name, throughput / (1 * MiB) });299 try stdout.print("{:>17}: {:10} MiB/s\n", .{ H.name, throughput / (1 * MiB) });
242 }300 }
243 }301 }
244302
245 inline for (macs) |M| {303 inline for (macs) |M| {
246 if (filter == null or std.mem.indexOf(u8, M.name, filter.?) != null) {304 if (filter == null or std.mem.indexOf(u8, M.name, filter.?) != null) {
247 const throughput = try benchmarkMac(M.ty, mode(128 * MiB));305 const throughput = try benchmarkMac(M.ty, mode(128 * MiB));
248 try stdout.print("{:>17}: {:7} MiB/s\n", .{ M.name, throughput / (1 * MiB) });306 try stdout.print("{:>17}: {:10} MiB/s\n", .{ M.name, throughput / (1 * MiB) });
249 }307 }
250 }308 }
251309
252 inline for (exchanges) |E| {310 inline for (exchanges) |E| {
253 if (filter == null or std.mem.indexOf(u8, E.name, filter.?) != null) {311 if (filter == null or std.mem.indexOf(u8, E.name, filter.?) != null) {
254 const throughput = try benchmarkKeyExchange(E.ty, mode(1000));312 const throughput = try benchmarkKeyExchange(E.ty, mode(1000));
255 try stdout.print("{:>17}: {:7} exchanges/s\n", .{ E.name, throughput });313 try stdout.print("{:>17}: {:10} exchanges/s\n", .{ E.name, throughput });
256 }314 }
257 }315 }
258316
259 inline for (signatures) |E| {317 inline for (signatures) |E| {
260 if (filter == null or std.mem.indexOf(u8, E.name, filter.?) != null) {318 if (filter == null or std.mem.indexOf(u8, E.name, filter.?) != null) {
261 const throughput = try benchmarkSignature(E.ty, mode(1000));319 const throughput = try benchmarkSignature(E.ty, mode(1000));
262 try stdout.print("{:>17}: {:7} signatures/s\n", .{ E.name, throughput });320 try stdout.print("{:>17}: {:10} signatures/s\n", .{ E.name, throughput });
263 }321 }
264 }322 }
265323
266 inline for (aeads) |E| {324 inline for (aeads) |E| {
267 if (filter == null or std.mem.indexOf(u8, E.name, filter.?) != null) {325 if (filter == null or std.mem.indexOf(u8, E.name, filter.?) != null) {
268 const throughput = try benchmarkAead(E.ty, mode(128 * MiB));326 const throughput = try benchmarkAead(E.ty, mode(128 * MiB));
269 try stdout.print("{:>17}: {:7} MiB/s\n", .{ E.name, throughput / (1 * MiB) });327 try stdout.print("{:>17}: {:10} MiB/s\n", .{ E.name, throughput / (1 * MiB) });
328 }
329 }
330
331 inline for (aes) |E| {
332 if (filter == null or std.mem.indexOf(u8, E.name, filter.?) != null) {
333 const throughput = try benchmarkAES(E.ty, mode(100000000));
334 try stdout.print("{:>17}: {:10} ops/s\n", .{ E.name, throughput });
335 }
336 }
337
338 inline for (aes8) |E| {
339 if (filter == null or std.mem.indexOf(u8, E.name, filter.?) != null) {
340 const throughput = try benchmarkAES8(E.ty, mode(10000000));
341 try stdout.print("{:>17}: {:10} ops/s\n", .{ E.name, throughput });
270 }342 }
271 }343 }
272}344}
lib/std/crypto/modes.zig created+51
...@@ -0,0 +1,51 @@
1// SPDX-License-Identifier: MIT
2// Copyright (c) 2015-2020 Zig Contributors
3// This file is part of [zig](https://ziglang.org/), which is MIT licensed.
4// The MIT license requires this copyright notice to be included in all copies
5// and substantial portions of the software.
6// Based on Go stdlib implementation
7
8const std = @import("../std.zig");
9const builtin = std.builtin;
10const mem = std.mem;
11const debug = std.debug;
12
13/// Counter mode.
14///
15/// This mode creates a key stream by encrypting an incrementing counter using a block cipher, and adding it to the source material.
16///
17/// Important: the counter mode doesn't provide authenticated encryption: the ciphertext can be trivially modified without this being detected.
18/// As a result, applications should generally never use it directly, but only in a construction that includes a MAC.
19pub fn ctr(comptime BlockCipher: anytype, block_cipher: BlockCipher, dst: []u8, src: []const u8, iv: [BlockCipher.block_size]u8, endian: comptime builtin.Endian) void {
20 debug.assert(dst.len >= src.len);
21 const block_size = BlockCipher.block_size;
22 var counter: [BlockCipher.block_size]u8 = undefined;
23 var counterInt = mem.readInt(u128, &iv, endian);
24 var i: usize = 0;
25
26 const parallel_count = BlockCipher.block.parallel.optimal_parallel_blocks;
27 const wide_block_size = parallel_count * 16;
28 if (src.len >= wide_block_size) {
29 var counters: [parallel_count * 16]u8 = undefined;
30 while (i + wide_block_size <= src.len) : (i += wide_block_size) {
31 comptime var j = 0;
32 inline while (j < parallel_count) : (j += 1) {
33 mem.writeInt(u128, counters[j * 16 .. j * 16 + 16], counterInt, endian);
34 counterInt +%= 1;
35 }
36 block_cipher.xorWide(parallel_count, dst[i .. i + wide_block_size][0..wide_block_size], src[i .. i + wide_block_size][0..wide_block_size], counters);
37 }
38 }
39 while (i + block_size <= src.len) : (i += block_size) {
40 mem.writeInt(u128, &counter, counterInt, endian);
41 counterInt +%= 1;
42 block_cipher.xor(dst[i .. i + block_size][0..block_size], src[i .. i + block_size][0..block_size], counter);
43 }
44 if (i < src.len) {
45 mem.writeInt(u128, &counter, counterInt, endian);
46 var pad = [_]u8{0} ** block_size;
47 mem.copy(u8, &pad, src[i..]);
48 block_cipher.xor(&pad, &pad, counter);
49 mem.copy(u8, dst[i..], pad[0 .. src.len - i]);
50 }
51}