crypto/
aessafe.rs

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
// option. This file may not be copied, modified, or distributed
// except according to those terms.

/*!

The `aessafe` module implements the AES algorithm completely in software without using any table
lookups or other timing dependant mechanisms. This module actually contains two seperate
implementations - an implementation that works on a single block at a time and a second
implementation that processes 8 blocks in parallel. Some block encryption modes really only work if
you are processing a single blocks (CFB, OFB, and CBC encryption for example) while other modes
are trivially parallelizable (CTR and CBC decryption). Processing more blocks at once allows for
greater efficiency, especially when using wide registers, such as the XMM registers available in
x86 processors.

## AES Algorithm

There are lots of places to go to on the internet for an involved description of how AES works. For
the purposes of this description, it sufficies to say that AES is just a block cipher that takes
a key of 16, 24, or 32 bytes and uses that to either encrypt or decrypt a block of 16 bytes. An
encryption or decryption operation consists of a number of rounds which involve some combination of
the following 4 basic operations:

* ShiftRows
* MixColumns
* SubBytes
* AddRoundKey

## Timing problems

Most software implementations of AES use a large set of lookup tables - generally at least the
SubBytes step is implemented via lookup tables; faster implementations generally implement the
MixColumns step this way as well. This is largely a design flaw in the AES implementation as it was
not realized during the NIST standardization process that table lookups can lead to security
problems [1]. The issue is that not all table lookups occur in constant time - an address that was
recently used is looked up much faster than one that hasn't been used in a while. A careful
adversary can measure the amount of time that each AES operation takes and use that information to
help determine the secret key or plain text information. More specifically, its not table lookups
that lead to these types of timing attacks - the issue is table lookups that use secret information
as part of the address to lookup. A table lookup that is performed the exact same way every time
regardless of the key or plaintext doesn't leak any information. This implementation uses no data
dependant table lookups.

## Bit Slicing

Bit Slicing is a technique that is basically a software emulation of hardware implementation
techniques. One of the earliest implementations of this technique was for a DES implementation [4].
In hardware, table lookups do not present the same timing problems as they do in software, however
they present other problems - namely that a 256 byte S-box table takes up a huge amount of space on
a chip. Hardware implementations, thus, tend to avoid table lookups and instead calculate the
contents of the S-Boxes as part of every operation. So, the key to an efficient Bit Sliced software
implementation is to re-arrange all of the bits of data to process into a form that can easily be
applied in much the same way that it would be in hardeware. It is fortunate, that AES was designed
such that these types of hardware implementations could be very efficient - the contents of the
S-boxes are defined by a mathematical formula.

A hardware implementation works on single bits at a time. Unlike adding variables in software,
however, that occur generally one at a time, hardware implementations are extremely parallel and
operate on many, many bits at once. Bit Slicing emulates that by moving all "equivalent" bits into
common registers and then operating on large groups of bits all at once. Calculating the S-box value
for a single bit is extremely expensive, but its much cheaper when you can amortize that cost over
128 bits (as in an XMM register). This implementation follows the same strategy as in [5] and that
is an excellent source for more specific details. However, a short description follows.

The input data is simply a collection of bytes. Each byte is comprised of 8 bits, a low order bit
(bit 0) through a high order bit (bit 7). Bit slicing the input data simply takes all of the low
order bits (bit 0) from the input data, and moves them into a single register (eg: XMM0). Next, all
of them 2nd lowest bits are moved into their own register (eg: XMM1), and so on. After completion,
we're left with 8 variables, each of which contains an equivalent set of bits. The exact order of
those bits is irrevent for the implementation of the SubBytes step, however, it is very important
for the MixColumns step. Again, see [5] for details. Due to the design of AES, its them possible to
execute the entire AES operation using just bitwise exclusive ors and rotates once we have Bit
Sliced the input data. After the completion of the AES operation, we then un-Bit Slice the data
to give us our output. Clearly, the more bits that we can process at once, the faster this will go -
thus, the version that processes 8 blocks at once is roughly 8 times faster than processing just a
single block at a time.

The ShiftRows step is fairly straight-forward to implement on the Bit Sliced state. The MixColumns
and especially the SubBytes steps are more complicated. This implementation draws heavily on the
formulas from [5], [6], and [7] to implement these steps.

## Implementation

Both implementations work basically the same way and share pretty much all of their code. The key
is first processed to create all of the round keys where each round key is just a 16 byte chunk of
data that is combined into the AES state by the AddRoundKey step as part of each encryption or
decryption round. Processing the round key can be expensive, so this is done before encryption or
decryption. Before encrypting or decrypting data, the data to be processed by be Bit Sliced into 8
seperate variables where each variable holds equivalent bytes from the state. This Bit Sliced state
is stored as a Bs8State<T>, where T is the type that stores each set of bits. The first
implementation stores these bits in a u32 which permits up to 8 * 32 = 1024 bits of data to be
processed at once. This implementation only processes a single block at a time, so, in reality, only
512 bits are processed at once and the remaining 512 bits of the variables are unused. The 2nd
implementation uses u32x4s - vectors of 4 u32s. Thus, we can process 8 * 128 = 4096 bits at once,
which corresponds exactly to 8 blocks.

The Bs8State struct implements the AesOps trait, which contains methods for each of the 4 main steps
of the AES algorithm. The types, T, each implement the AesBitValueOps trait, which containts methods
necessary for processing a collection or bit values and the AesOps trait relies heavily on this
trait to perform its operations.

The Bs4State and Bs2State struct implement operations of various subfields of the full GF(2^8)
finite field which allows for efficient computation of the AES S-Boxes. See [7] for details.

## References

[1] - "Cache-Collision Timing Attacks Against AES". Joseph Bonneau and Ilya Mironov.
      http://www.jbonneau.com/doc/BM06-CHES-aes_cache_timing.pdf
[2] - "Software mitigations to hedge AES against cache-based software side channel vulnerabilities".
      Ernie Brickell, et al. http://eprint.iacr.org/2006/052.pdf.
[3] - "Cache Attacks and Countermeasures: the Case of AES (Extended Version)".
      Dag Arne Osvik, et al. tau.ac.il/~tromer/papers/cache.pdf‎.
[4] - "A Fast New DES Implementation in Software". Eli Biham.
      http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.52.5429&rep=rep1&type=pdf.
[5] - "Faster and Timing-Attack Resistant AES-GCM". Emilia K ̈asper and Peter Schwabe.
      http://www.chesworkshop.org/ches2009/presentations/01_Session_1/CHES2009_ekasper.pdf.
[6] - "FAST AES DECRYPTION". Vinit Azad. http://webcache.googleusercontent.com/
      search?q=cache:ld_f8pSgURcJ:csusdspace.calstate.edu/bitstream/handle/10211.9/1224/
      Vinit_Azad_MS_Report.doc%3Fsequence%3D2+&cd=4&hl=en&ct=clnk&gl=us&client=ubuntu.
[7] - "A Very Compact Rijndael S-box". D. Canright.
      http://www.dtic.mil/cgi-bin/GetTRDoc?AD=ADA434781.
*/

use std::ops::{BitAnd, BitXor, Not};
use std::default::Default;

use cryptoutil::{read_u32v_le, write_u32_le};
use simd::u32x4;
use step_by::RangeExt;
use symmetriccipher::{BlockEncryptor, BlockEncryptorX8, BlockDecryptor, BlockDecryptorX8};

const U32X4_0: u32x4 = u32x4(0, 0, 0, 0);
const U32X4_1: u32x4 = u32x4(0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff);

macro_rules! define_aes_struct(
    (
        $name:ident,
        $rounds:expr
    ) => (
        #[derive(Clone, Copy)]
        pub struct $name {
            sk: [Bs8State<u16>; ($rounds + 1)]
        }
    )
);

macro_rules! define_aes_impl(
    (
        $name:ident,
        $mode:ident,
        $rounds:expr,
        $key_size:expr
    ) => (
        impl $name {
            pub fn new(key: &[u8]) -> $name {
                let mut a =  $name {
                    sk: [Bs8State(0, 0, 0, 0, 0, 0, 0, 0); ($rounds + 1)]
                };
                let mut tmp = [[0u32; 4]; ($rounds + 1)];
                create_round_keys(key, KeyType::$mode, &mut tmp);
                for i in 0..$rounds + 1 {
                    a.sk[i] = bit_slice_4x4_with_u16(tmp[i][0], tmp[i][1], tmp[i][2], tmp[i][3]);
                }
                a
            }
        }
    )
);

macro_rules! define_aes_enc(
    (
        $name:ident,
        $rounds:expr
    ) => (
        impl BlockEncryptor for $name {
            fn block_size(&self) -> usize { 16 }
            fn encrypt_block(&self, input: &[u8], output: &mut [u8]) {
                let mut bs = bit_slice_1x16_with_u16(input);
                bs = encrypt_core(&bs, &self.sk);
                un_bit_slice_1x16_with_u16(&bs, output);
            }
        }
    )
);

macro_rules! define_aes_dec(
    (
        $name:ident,
        $rounds:expr
    ) => (
        impl BlockDecryptor for $name {
            fn block_size(&self) -> usize { 16 }
            fn decrypt_block(&self, input: &[u8], output: &mut [u8]) {
                let mut bs = bit_slice_1x16_with_u16(input);
                bs = decrypt_core(&bs, &self.sk);
                un_bit_slice_1x16_with_u16(&bs, output);
            }
        }
    )
);

define_aes_struct!(AesSafe128Encryptor, 10);
define_aes_struct!(AesSafe128Decryptor, 10);
define_aes_impl!(AesSafe128Encryptor, Encryption, 10, 16);
define_aes_impl!(AesSafe128Decryptor, Decryption, 10, 16);
define_aes_enc!(AesSafe128Encryptor, 10);
define_aes_dec!(AesSafe128Decryptor, 10);

define_aes_struct!(AesSafe192Encryptor, 12);
define_aes_struct!(AesSafe192Decryptor, 12);
define_aes_impl!(AesSafe192Encryptor, Encryption, 12, 24);
define_aes_impl!(AesSafe192Decryptor, Decryption, 12, 24);
define_aes_enc!(AesSafe192Encryptor, 12);
define_aes_dec!(AesSafe192Decryptor, 12);

define_aes_struct!(AesSafe256Encryptor, 14);
define_aes_struct!(AesSafe256Decryptor, 14);
define_aes_impl!(AesSafe256Encryptor, Encryption, 14, 32);
define_aes_impl!(AesSafe256Decryptor, Decryption, 14, 32);
define_aes_enc!(AesSafe256Encryptor, 14);
define_aes_dec!(AesSafe256Decryptor, 14);

macro_rules! define_aes_struct_x8(
    (
        $name:ident,
        $rounds:expr
    ) => (
        #[derive(Clone, Copy)]
        pub struct $name {
            sk: [Bs8State<u32x4>; ($rounds + 1)]
        }
    )
);

macro_rules! define_aes_impl_x8(
    (
        $name:ident,
        $mode:ident,
        $rounds:expr,
        $key_size:expr
    ) => (
        impl $name {
            pub fn new(key: &[u8]) -> $name {
                let mut a =  $name {
                    sk: [
                        Bs8State(
                            U32X4_0,
                            U32X4_0,
                            U32X4_0,
                            U32X4_0,
                            U32X4_0,
                            U32X4_0,
                            U32X4_0,
                            U32X4_0);
                        ($rounds + 1)]
                };
                let mut tmp = [[0u32; 4]; ($rounds + 1)];
                create_round_keys(key, KeyType::$mode, &mut tmp);
                for i in 0..$rounds + 1 {
                    a.sk[i] = bit_slice_fill_4x4_with_u32x4(
                        tmp[i][0],
                        tmp[i][1],
                        tmp[i][2],
                        tmp[i][3]);
                }
                a
            }
        }
    )
);

macro_rules! define_aes_enc_x8(
    (
        $name:ident,
        $rounds:expr
    ) => (
        impl BlockEncryptorX8 for $name {
            fn block_size(&self) -> usize { 16 }
            fn encrypt_block_x8(&self, input: &[u8], output: &mut [u8]) {
                let bs = bit_slice_1x128_with_u32x4(input);
                let bs2 = encrypt_core(&bs, &self.sk);
                un_bit_slice_1x128_with_u32x4(bs2, output);
            }
        }
    )
);

macro_rules! define_aes_dec_x8(
    (
        $name:ident,
        $rounds:expr
    ) => (
        impl BlockDecryptorX8 for $name {
            fn block_size(&self) -> usize { 16 }
            fn decrypt_block_x8(&self, input: &[u8], output: &mut [u8]) {
                let bs = bit_slice_1x128_with_u32x4(input);
                let bs2 = decrypt_core(&bs, &self.sk);
                un_bit_slice_1x128_with_u32x4(bs2, output);
            }
        }
    )
);

define_aes_struct_x8!(AesSafe128EncryptorX8, 10);
define_aes_struct_x8!(AesSafe128DecryptorX8, 10);
define_aes_impl_x8!(AesSafe128EncryptorX8, Encryption, 10, 16);
define_aes_impl_x8!(AesSafe128DecryptorX8, Decryption, 10, 16);
define_aes_enc_x8!(AesSafe128EncryptorX8, 10);
define_aes_dec_x8!(AesSafe128DecryptorX8, 10);

define_aes_struct_x8!(AesSafe192EncryptorX8, 12);
define_aes_struct_x8!(AesSafe192DecryptorX8, 12);
define_aes_impl_x8!(AesSafe192EncryptorX8, Encryption, 12, 24);
define_aes_impl_x8!(AesSafe192DecryptorX8, Decryption, 12, 24);
define_aes_enc_x8!(AesSafe192EncryptorX8, 12);
define_aes_dec_x8!(AesSafe192DecryptorX8, 12);

define_aes_struct_x8!(AesSafe256EncryptorX8, 14);
define_aes_struct_x8!(AesSafe256DecryptorX8, 14);
define_aes_impl_x8!(AesSafe256EncryptorX8, Encryption, 14, 32);
define_aes_impl_x8!(AesSafe256DecryptorX8, Decryption, 14, 32);
define_aes_enc_x8!(AesSafe256EncryptorX8, 14);
define_aes_dec_x8!(AesSafe256DecryptorX8, 14);

fn ffmulx(x: u32) -> u32 {
    let m1: u32 = 0x80808080;
    let m2: u32 = 0x7f7f7f7f;
    let m3: u32 = 0x0000001b;
    ((x & m2) << 1) ^ (((x & m1) >> 7) * m3)
}

fn inv_mcol(x: u32) -> u32 {
    let f2 = ffmulx(x);
    let f4 = ffmulx(f2);
    let f8 = ffmulx(f4);
    let f9 = x ^ f8;

    f2 ^ f4 ^ f8 ^ (f2 ^ f9).rotate_right(8) ^ (f4 ^ f9).rotate_right(16) ^ f9.rotate_right(24)
}

fn sub_word(x: u32) -> u32 {
    let bs = bit_slice_4x1_with_u16(x).sub_bytes();
    un_bit_slice_4x1_with_u16(&bs)
}

enum KeyType {
    Encryption,
    Decryption
}

// This array is not accessed in any key-dependant way, so there are no timing problems inherent in
// using it.
static RCON: [u32; 10] = [0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x1b, 0x36];

// The round keys are created without bit-slicing the key data. The individual implementations bit
// slice the round keys returned from this function. This function, and the few functions above, are
// derived from the BouncyCastle AES implementation.
fn create_round_keys(key: &[u8], key_type: KeyType, round_keys: &mut [[u32; 4]]) {
    let (key_words, rounds) = match key.len() {
        16 => (4, 10),
        24 => (6, 12),
        32 => (8, 14),
        _ => panic!("Invalid AES key size.")
    };

    // The key is copied directly into the first few round keys
    let mut j = 0;
    for i in (0..key.len()).step_up(4) {
        round_keys[j / 4][j % 4] =
            (key[i] as u32) |
            ((key[i+1] as u32) << 8) |
            ((key[i+2] as u32) << 16) |
            ((key[i+3] as u32) << 24);
        j += 1;
    };

    // Calculate the rest of the round keys
    for i in key_words..(rounds + 1) * 4 {
        let mut tmp = round_keys[(i - 1) / 4][(i - 1) % 4];
        if (i % key_words) == 0 {
            tmp = sub_word(tmp.rotate_right(8)) ^ RCON[(i / key_words) - 1];
        } else if (key_words == 8) && ((i % key_words) == 4) {
            // This is only necessary for AES-256 keys
            tmp = sub_word(tmp);
        }
        round_keys[i / 4][i % 4] = round_keys[(i - key_words) / 4][(i - key_words) % 4] ^ tmp;
    }

    // Decryption round keys require extra processing
    match key_type {
        KeyType::Decryption => {
            for j in 1..rounds {
                for i in 0..4 {
                    round_keys[j][i] = inv_mcol(round_keys[j][i]);
                }
            }
        },
        KeyType::Encryption => { }
    }
}

// This trait defines all of the operations needed for a type to be processed as part of an AES
// encryption or decryption operation.
trait AesOps {
    fn sub_bytes(self) -> Self;
    fn inv_sub_bytes(self) -> Self;
    fn shift_rows(self) -> Self;
    fn inv_shift_rows(self) -> Self;
    fn mix_columns(self) -> Self;
    fn inv_mix_columns(self) -> Self;
    fn add_round_key(self, rk: &Self) -> Self;
}

fn encrypt_core<S: AesOps + Copy>(state: &S, sk: &[S]) -> S {
    // Round 0 - add round key
    let mut tmp = state.add_round_key(&sk[0]);

    // Remaining rounds (except last round)
    for i in 1..sk.len() - 1 {
        tmp = tmp.sub_bytes();
        tmp = tmp.shift_rows();
        tmp = tmp.mix_columns();
        tmp = tmp.add_round_key(&sk[i]);
    }

    // Last round
    tmp = tmp.sub_bytes();
    tmp = tmp.shift_rows();
    tmp = tmp.add_round_key(&sk[sk.len() - 1]);

    tmp
}

fn decrypt_core<S: AesOps + Copy>(state: &S, sk: &[S]) -> S {
    // Round 0 - add round key
    let mut tmp = state.add_round_key(&sk[sk.len() - 1]);

    // Remaining rounds (except last round)
    for i in 1..sk.len() - 1 {
        tmp = tmp.inv_sub_bytes();
        tmp = tmp.inv_shift_rows();
        tmp = tmp.inv_mix_columns();
        tmp = tmp.add_round_key(&sk[sk.len() - 1 - i]);
    }

    // Last round
    tmp = tmp.inv_sub_bytes();
    tmp = tmp.inv_shift_rows();
    tmp = tmp.add_round_key(&sk[0]);

    tmp
}

#[derive(Clone, Copy)]
struct Bs8State<T>(T, T, T, T, T, T, T, T);

impl <T: Copy> Bs8State<T> {
    fn split(self) -> (Bs4State<T>, Bs4State<T>) {
        let Bs8State(x0, x1, x2, x3, x4, x5, x6, x7) = self;
        (Bs4State(x0, x1, x2, x3), Bs4State(x4, x5, x6, x7))
    }
}

impl <T: BitXor<Output = T> + Copy> Bs8State<T> {
    fn xor(self, rhs: Bs8State<T>) -> Bs8State<T> {
        let Bs8State(a0, a1, a2, a3, a4, a5, a6, a7) = self;
        let Bs8State(b0, b1, b2, b3, b4, b5, b6, b7) = rhs;
        Bs8State(a0 ^ b0, a1 ^ b1, a2 ^ b2, a3 ^ b3, a4 ^ b4, a5 ^ b5, a6 ^ b6, a7 ^ b7)
    }

    // We need to be able to convert a Bs8State to and from a polynomial basis and a normal
    // basis. That transformation could be done via pseudocode that roughly looks like the
    // following:
    //
    // for x in 0..8 {
    //     for y in 0..8 {
    //         result.x ^= input.y & MATRIX[7 - y][x]
    //     }
    // }
    //
    // Where the MATRIX is one of the following depending on the conversion being done.
    // (The affine transformation step is included in all of these matrices):
    //
    // A2X = [
    //     [ 0,  0,  0, -1, -1,  0,  0, -1],
    //     [-1, -1,  0,  0, -1, -1, -1, -1],
    //     [ 0, -1,  0,  0, -1, -1, -1, -1],
    //     [ 0,  0,  0, -1,  0,  0, -1,  0],
    //     [-1,  0,  0, -1,  0,  0,  0,  0],
    //     [-1,  0,  0,  0,  0,  0,  0, -1],
    //     [-1,  0,  0, -1,  0, -1,  0, -1],
    //     [-1, -1, -1, -1, -1, -1, -1, -1]
    // ];
    //
    // X2A = [
    //     [ 0,  0, -1,  0,  0, -1, -1,  0],
    //     [ 0,  0,  0, -1, -1, -1, -1,  0],
    //     [ 0, -1, -1, -1,  0, -1, -1,  0],
    //     [ 0,  0, -1, -1,  0,  0,  0, -1],
    //     [ 0,  0,  0, -1,  0, -1, -1,  0],
    //     [-1,  0,  0, -1,  0, -1,  0,  0],
    //     [ 0, -1, -1, -1, -1,  0, -1, -1],
    //     [ 0,  0,  0,  0,  0, -1, -1,  0],
    // ];
    //
    // X2S = [
    //     [ 0,  0,  0, -1, -1,  0, -1,  0],
    //     [-1,  0, -1, -1,  0, -1,  0,  0],
    //     [ 0, -1, -1, -1, -1,  0,  0, -1],
    //     [-1, -1,  0, -1,  0,  0,  0,  0],
    //     [ 0,  0, -1, -1, -1,  0, -1, -1],
    //     [ 0,  0, -1,  0,  0,  0,  0,  0],
    //     [-1, -1,  0,  0,  0,  0,  0,  0],
    //     [ 0,  0, -1,  0,  0, -1,  0,  0],
    // ];
    //
    // S2X = [
    //     [ 0,  0, -1, -1,  0,  0,  0, -1],
    //     [-1,  0,  0, -1, -1, -1, -1,  0],
    //     [-1,  0, -1,  0,  0,  0,  0,  0],
    //     [-1, -1,  0, -1,  0, -1, -1, -1],
    //     [ 0, -1,  0,  0, -1,  0,  0,  0],
    //     [ 0,  0, -1,  0,  0,  0,  0,  0],
    //     [-1,  0,  0,  0, -1,  0, -1,  0],
    //     [-1, -1,  0,  0, -1,  0, -1,  0],
    // ];
    //
    // Looking at the pseudocode implementation, we see that there is no point
    // in processing any of the elements in those matrices that have zero values
    // since a logical AND with 0 will produce 0 which will have no effect when it
    // is XORed into the result.
    //
    // LLVM doesn't appear to be able to fully unroll the loops in the pseudocode
    // above and to eliminate processing of the 0 elements. So, each transformation is
    // implemented independently directly in fully unrolled form with the 0 elements
    // removed.
    //
    // As an optimization, elements that are XORed together multiple times are
    // XORed just once and then used multiple times. I wrote a simple program that
    // greedily looked for terms to combine to create the implementations below.
    // It is likely that this could be optimized more.

    fn change_basis_a2x(&self) -> Bs8State<T> {
        let t06 = self.6 ^ self.0;
        let t056 = self.5 ^ t06;
        let t0156 = t056 ^ self.1;
        let t13 = self.1 ^ self.3;

        let x0 = self.2 ^ t06 ^ t13;
        let x1 = t056;
        let x2 = self.0;
        let x3 = self.0 ^ self.4 ^ self.7 ^ t13;
        let x4 = self.7 ^ t056;
        let x5 = t0156;
        let x6 = self.4 ^ t056;
        let x7 = self.2 ^ self.7 ^ t0156;

        Bs8State(x0, x1, x2, x3, x4, x5, x6, x7)
    }

    fn change_basis_x2s(&self) -> Bs8State<T> {
        let t46 = self.4 ^ self.6;
        let t35 = self.3 ^ self.5;
        let t06 = self.0 ^ self.6;
        let t357 = t35 ^ self.7;

        let x0 = self.1 ^ t46;
        let x1 = self.1 ^ self.4 ^ self.5;
        let x2 = self.2 ^ t35 ^ t06;
        let x3 = t46 ^ t357;
        let x4 = t357;
        let x5 = t06;
        let x6 = self.3 ^ self.7;
        let x7 = t35;

        Bs8State(x0, x1, x2, x3, x4, x5, x6, x7)
    }

    fn change_basis_x2a(&self) -> Bs8State<T> {
        let t15 = self.1 ^ self.5;
        let t36 = self.3 ^ self.6;
        let t1356 = t15 ^ t36;
        let t07 = self.0 ^ self.7;

        let x0 = self.2;
        let x1 = t15;
        let x2 = self.4 ^ self.7 ^ t15;
        let x3 = self.2 ^ self.4 ^ t1356;
        let x4 = self.1 ^ self.6;
        let x5 = self.2 ^ self.5 ^ t36 ^ t07;
        let x6 = t1356 ^ t07;
        let x7 = self.1 ^ self.4;

        Bs8State(x0, x1, x2, x3, x4, x5, x6, x7)
    }

    fn change_basis_s2x(&self) -> Bs8State<T> {
        let t46 = self.4 ^ self.6;
        let t01 = self.0 ^ self.1;
        let t0146 = t01 ^ t46;

        let x0 = self.5 ^ t0146;
        let x1 = self.0 ^ self.3 ^ self.4;
        let x2 = self.2 ^ self.5 ^ self.7;
        let x3 = self.7 ^ t46;
        let x4 = self.3 ^ self.6 ^ t01;
        let x5 = t46;
        let x6 = t0146;
        let x7 = self.4 ^ self.7;

        Bs8State(x0, x1, x2, x3, x4, x5, x6, x7)
    }
}

impl <T: Not<Output = T> + Copy> Bs8State<T> {
    // The special value "x63" is used as part of the sub_bytes and inv_sub_bytes
    // steps. It is conceptually a Bs8State value where the 0th, 1st, 5th, and 6th
    // elements are all 1s and the other elements are all 0s. The only thing that
    // we do with the "x63" value is to XOR a Bs8State with it. We optimize that XOR
    // below into just inverting 4 of the elements and leaving the other 4 elements
    // untouched.
    fn xor_x63(self) -> Bs8State<T> {
        Bs8State (
            !self.0,
            !self.1,
            self.2,
            self.3,
            self.4,
            !self.5,
            !self.6,
            self.7)
    }
}

#[derive(Clone, Copy)]
struct Bs4State<T>(T, T, T, T);

impl <T: Copy> Bs4State<T> {
    fn split(self) -> (Bs2State<T>, Bs2State<T>) {
        let Bs4State(x0, x1, x2, x3) = self;
        (Bs2State(x0, x1), Bs2State(x2, x3))
    }

    fn join(self, rhs: Bs4State<T>) -> Bs8State<T> {
        let Bs4State(a0, a1, a2, a3) = self;
        let Bs4State(b0, b1, b2, b3) = rhs;
        Bs8State(a0, a1, a2, a3, b0, b1, b2, b3)
    }
}

impl <T: BitXor<Output = T> + Copy> Bs4State<T> {
    fn xor(self, rhs: Bs4State<T>) -> Bs4State<T> {
        let Bs4State(a0, a1, a2, a3) = self;
        let Bs4State(b0, b1, b2, b3) = rhs;
        Bs4State(a0 ^ b0, a1 ^ b1, a2 ^ b2, a3 ^ b3)
    }
}

#[derive(Clone, Copy)]
struct Bs2State<T>(T, T);

impl <T> Bs2State<T> {
    fn split(self) -> (T, T) {
        let Bs2State(x0, x1) = self;
        (x0, x1)
    }

    fn join(self, rhs: Bs2State<T>) -> Bs4State<T> {
        let Bs2State(a0, a1) = self;
        let Bs2State(b0, b1) = rhs;
        Bs4State(a0, a1, b0, b1)
    }
}

impl <T: BitXor<Output = T> + Copy> Bs2State<T> {
    fn xor(self, rhs: Bs2State<T>) -> Bs2State<T> {
        let Bs2State(a0, a1) = self;
        let Bs2State(b0, b1) = rhs;
        Bs2State(a0 ^ b0, a1 ^ b1)
    }
}

// Bit Slice data in the form of 4 u32s in column-major order
fn bit_slice_4x4_with_u16(a: u32, b: u32, c: u32, d: u32) -> Bs8State<u16> {
    fn pb(x: u32, bit: u32, shift: u32) -> u16 {
        (((x >> bit) & 1) as u16) << shift
    }

    fn construct(a: u32, b: u32, c: u32, d: u32, bit: u32) -> u16 {
        pb(a, bit, 0)       | pb(b, bit, 1)       | pb(c, bit, 2)       | pb(d, bit, 3)       |
        pb(a, bit + 8, 4)   | pb(b, bit + 8, 5)   | pb(c, bit + 8, 6)   | pb(d, bit + 8, 7)   |
        pb(a, bit + 16, 8)  | pb(b, bit + 16, 9)  | pb(c, bit + 16, 10) | pb(d, bit + 16, 11) |
        pb(a, bit + 24, 12) | pb(b, bit + 24, 13) | pb(c, bit + 24, 14) | pb(d, bit + 24, 15)
    }

    let x0 = construct(a, b, c, d, 0);
    let x1 = construct(a, b, c, d, 1);
    let x2 = construct(a, b, c, d, 2);
    let x3 = construct(a, b, c, d, 3);
    let x4 = construct(a, b, c, d, 4);
    let x5 = construct(a, b, c, d, 5);
    let x6 = construct(a, b, c, d, 6);
    let x7 = construct(a, b, c, d, 7);

    Bs8State(x0, x1, x2, x3, x4, x5, x6, x7)
}

// Bit slice a single u32 value - this is used to calculate the SubBytes step when creating the
// round keys.
fn bit_slice_4x1_with_u16(a: u32) -> Bs8State<u16> {
    bit_slice_4x4_with_u16(a, 0, 0, 0)
}

// Bit slice a 16 byte array in column major order
fn bit_slice_1x16_with_u16(data: &[u8]) -> Bs8State<u16> {
    let mut n = [0u32; 4];
    read_u32v_le(&mut n, data);

    let a = n[0];
    let b = n[1];
    let c = n[2];
    let d = n[3];

    bit_slice_4x4_with_u16(a, b, c, d)
}

// Un Bit Slice into a set of 4 u32s
fn un_bit_slice_4x4_with_u16(bs: &Bs8State<u16>) -> (u32, u32, u32, u32) {
    fn pb(x: u16, bit: u32, shift: u32) -> u32 {
        (((x >> bit) & 1) as u32) << shift
    }

    fn deconstruct(bs: &Bs8State<u16>, bit: u32) -> u32 {
        let Bs8State(x0, x1, x2, x3, x4, x5, x6, x7) = *bs;

        pb(x0, bit, 0) | pb(x1, bit, 1) | pb(x2, bit, 2) | pb(x3, bit, 3) |
        pb(x4, bit, 4) | pb(x5, bit, 5) | pb(x6, bit, 6) | pb(x7, bit, 7) |

        pb(x0, bit + 4, 8)  | pb(x1, bit + 4, 9)  | pb(x2, bit + 4, 10) | pb(x3, bit + 4, 11) |
        pb(x4, bit + 4, 12) | pb(x5, bit + 4, 13) | pb(x6, bit + 4, 14) | pb(x7, bit + 4, 15) |

        pb(x0, bit + 8, 16) | pb(x1, bit + 8, 17) | pb(x2, bit + 8, 18) | pb(x3, bit + 8, 19) |
        pb(x4, bit + 8, 20) | pb(x5, bit + 8, 21) | pb(x6, bit + 8, 22) | pb(x7, bit + 8, 23) |

        pb(x0, bit + 12, 24) | pb(x1, bit + 12, 25) | pb(x2, bit + 12, 26) | pb(x3, bit + 12, 27) |
        pb(x4, bit + 12, 28) | pb(x5, bit + 12, 29) | pb(x6, bit + 12, 30) | pb(x7, bit + 12, 31)
    }

    let a = deconstruct(bs, 0);
    let b = deconstruct(bs, 1);
    let c = deconstruct(bs, 2);
    let d = deconstruct(bs, 3);

    (a, b, c, d)
}

// Un Bit Slice into a single u32. This is used when creating the round keys.
fn un_bit_slice_4x1_with_u16(bs: &Bs8State<u16>) -> u32 {
    let (a, _, _, _) = un_bit_slice_4x4_with_u16(bs);
    a
}

// Un Bit Slice into a 16 byte array
fn un_bit_slice_1x16_with_u16(bs: &Bs8State<u16>, output: &mut [u8]) {
    let (a, b, c, d) = un_bit_slice_4x4_with_u16(bs);

    write_u32_le(&mut output[0..4], a);
    write_u32_le(&mut output[4..8], b);
    write_u32_le(&mut output[8..12], c);
    write_u32_le(&mut output[12..16], d);
}

// Bit Slice a 128 byte array of eight 16 byte blocks. Each block is in column major order.
fn bit_slice_1x128_with_u32x4(data: &[u8]) -> Bs8State<u32x4> {
    let bit0 = u32x4(0x01010101, 0x01010101, 0x01010101, 0x01010101);
    let bit1 = u32x4(0x02020202, 0x02020202, 0x02020202, 0x02020202);
    let bit2 = u32x4(0x04040404, 0x04040404, 0x04040404, 0x04040404);
    let bit3 = u32x4(0x08080808, 0x08080808, 0x08080808, 0x08080808);
    let bit4 = u32x4(0x10101010, 0x10101010, 0x10101010, 0x10101010);
    let bit5 = u32x4(0x20202020, 0x20202020, 0x20202020, 0x20202020);
    let bit6 = u32x4(0x40404040, 0x40404040, 0x40404040, 0x40404040);
    let bit7 = u32x4(0x80808080, 0x80808080, 0x80808080, 0x80808080);

    fn read_row_major(data: &[u8]) -> u32x4 {
        u32x4(
            (data[0] as u32) |
            ((data[4] as u32) << 8) |
            ((data[8] as u32) << 16) |
            ((data[12] as u32) << 24),
            (data[1] as u32) |
            ((data[5] as u32) << 8) |
            ((data[9] as u32) << 16) |
            ((data[13] as u32) << 24),
            (data[2] as u32) |
            ((data[6] as u32) << 8) |
            ((data[10] as u32) << 16) |
            ((data[14] as u32) << 24),
            (data[3] as u32) |
            ((data[7] as u32) << 8) |
            ((data[11] as u32) << 16) |
            ((data[15] as u32) << 24))
    }

    let t0 = read_row_major(&data[0..16]);
    let t1 = read_row_major(&data[16..32]);
    let t2 = read_row_major(&data[32..48]);
    let t3 = read_row_major(&data[48..64]);
    let t4 = read_row_major(&data[64..80]);
    let t5 = read_row_major(&data[80..96]);
    let t6 = read_row_major(&data[96..112]);
    let t7 = read_row_major(&data[112..128]);

    let x0 = (t0 & bit0) | (t1.lsh(1) & bit1) | (t2.lsh(2) & bit2) | (t3.lsh(3) & bit3) |
        (t4.lsh(4) & bit4) | (t5.lsh(5) & bit5) | (t6.lsh(6) & bit6) | (t7.lsh(7) & bit7);
    let x1 = (t0.rsh(1) & bit0) | (t1 & bit1) | (t2.lsh(1) & bit2) | (t3.lsh(2) & bit3) |
        (t4.lsh(3) & bit4) | (t5.lsh(4) & bit5) | (t6.lsh(5) & bit6) | (t7.lsh(6) & bit7);
    let x2 = (t0.rsh(2) & bit0) | (t1.rsh(1) & bit1) | (t2 & bit2) | (t3.lsh(1) & bit3) |
        (t4.lsh(2) & bit4) | (t5.lsh(3) & bit5) | (t6.lsh(4) & bit6) | (t7.lsh(5) & bit7);
    let x3 = (t0.rsh(3) & bit0) | (t1.rsh(2) & bit1) | (t2.rsh(1) & bit2) | (t3 & bit3) |
        (t4.lsh(1) & bit4) | (t5.lsh(2) & bit5) | (t6.lsh(3) & bit6) | (t7.lsh(4) & bit7);
    let x4 = (t0.rsh(4) & bit0) | (t1.rsh(3) & bit1) | (t2.rsh(2) & bit2) | (t3.rsh(1) & bit3) |
        (t4 & bit4) | (t5.lsh(1) & bit5) | (t6.lsh(2) & bit6) | (t7.lsh(3) & bit7);
    let x5 = (t0.rsh(5) & bit0) | (t1.rsh(4) & bit1) | (t2.rsh(3) & bit2) | (t3.rsh(2) & bit3) |
        (t4.rsh(1) & bit4) | (t5 & bit5) | (t6.lsh(1) & bit6) | (t7.lsh(2) & bit7);
    let x6 = (t0.rsh(6) & bit0) | (t1.rsh(5) & bit1) | (t2.rsh(4) & bit2) | (t3.rsh(3) & bit3) |
        (t4.rsh(2) & bit4) | (t5.rsh(1) & bit5) | (t6 & bit6) | (t7.lsh(1) & bit7);
    let x7 = (t0.rsh(7) & bit0) | (t1.rsh(6) & bit1) | (t2.rsh(5) & bit2) | (t3.rsh(4) & bit3) |
        (t4.rsh(3) & bit4) | (t5.rsh(2) & bit5) | (t6.rsh(1) & bit6) | (t7 & bit7);

    Bs8State(x0, x1, x2, x3, x4, x5, x6, x7)
}

// Bit slice a set of 4 u32s by filling a full 128 byte data block with those repeated values. This
// is used as part of bit slicing the round keys.
fn bit_slice_fill_4x4_with_u32x4(a: u32, b: u32, c: u32, d: u32) -> Bs8State<u32x4> {
    let mut tmp = [0u8; 128];
    for i in 0..8 {
        write_u32_le(&mut tmp[i * 16..i * 16 + 4], a);
        write_u32_le(&mut tmp[i * 16 + 4..i * 16 + 8], b);
        write_u32_le(&mut tmp[i * 16 + 8..i * 16 + 12], c);
        write_u32_le(&mut tmp[i * 16 + 12..i * 16 + 16], d);
    }
    bit_slice_1x128_with_u32x4(&tmp)
}

// Un bit slice into a 128 byte buffer.
fn un_bit_slice_1x128_with_u32x4(bs: Bs8State<u32x4>, output: &mut [u8]) {
    let Bs8State(t0, t1, t2, t3, t4, t5, t6, t7) = bs;

    let bit0 = u32x4(0x01010101, 0x01010101, 0x01010101, 0x01010101);
    let bit1 = u32x4(0x02020202, 0x02020202, 0x02020202, 0x02020202);
    let bit2 = u32x4(0x04040404, 0x04040404, 0x04040404, 0x04040404);
    let bit3 = u32x4(0x08080808, 0x08080808, 0x08080808, 0x08080808);
    let bit4 = u32x4(0x10101010, 0x10101010, 0x10101010, 0x10101010);
    let bit5 = u32x4(0x20202020, 0x20202020, 0x20202020, 0x20202020);
    let bit6 = u32x4(0x40404040, 0x40404040, 0x40404040, 0x40404040);
    let bit7 = u32x4(0x80808080, 0x80808080, 0x80808080, 0x80808080);

    // decode the individual blocks, in row-major order
    // TODO: this is identical to the same block in bit_slice_1x128_with_u32x4
    let x0 = (t0 & bit0) | (t1.lsh(1) & bit1) | (t2.lsh(2) & bit2) | (t3.lsh(3) & bit3) |
        (t4.lsh(4) & bit4) | (t5.lsh(5) & bit5) | (t6.lsh(6) & bit6) | (t7.lsh(7) & bit7);
    let x1 = (t0.rsh(1) & bit0) | (t1 & bit1) | (t2.lsh(1) & bit2) | (t3.lsh(2) & bit3) |
        (t4.lsh(3) & bit4) | (t5.lsh(4) & bit5) | (t6.lsh(5) & bit6) | (t7.lsh(6) & bit7);
    let x2 = (t0.rsh(2) & bit0) | (t1.rsh(1) & bit1) | (t2 & bit2) | (t3.lsh(1) & bit3) |
        (t4.lsh(2) & bit4) | (t5.lsh(3) & bit5) | (t6.lsh(4) & bit6) | (t7.lsh(5) & bit7);
    let x3 = (t0.rsh(3) & bit0) | (t1.rsh(2) & bit1) | (t2.rsh(1) & bit2) | (t3 & bit3) |
        (t4.lsh(1) & bit4) | (t5.lsh(2) & bit5) | (t6.lsh(3) & bit6) | (t7.lsh(4) & bit7);
    let x4 = (t0.rsh(4) & bit0) | (t1.rsh(3) & bit1) | (t2.rsh(2) & bit2) | (t3.rsh(1) & bit3) |
        (t4 & bit4) | (t5.lsh(1) & bit5) | (t6.lsh(2) & bit6) | (t7.lsh(3) & bit7);
    let x5 = (t0.rsh(5) & bit0) | (t1.rsh(4) & bit1) | (t2.rsh(3) & bit2) | (t3.rsh(2) & bit3) |
        (t4.rsh(1) & bit4) | (t5 & bit5) | (t6.lsh(1) & bit6) | (t7.lsh(2) & bit7);
    let x6 = (t0.rsh(6) & bit0) | (t1.rsh(5) & bit1) | (t2.rsh(4) & bit2) | (t3.rsh(3) & bit3) |
        (t4.rsh(2) & bit4) | (t5.rsh(1) & bit5) | (t6 & bit6) | (t7.lsh(1) & bit7);
    let x7 = (t0.rsh(7) & bit0) | (t1.rsh(6) & bit1) | (t2.rsh(5) & bit2) | (t3.rsh(4) & bit3) |
        (t4.rsh(3) & bit4) | (t5.rsh(2) & bit5) | (t6.rsh(1) & bit6) | (t7 & bit7);

    fn write_row_major(block: u32x4, output: &mut [u8]) {
        let u32x4(a0, a1, a2, a3) = block;
        output[0] = a0 as u8;
        output[1] = a1 as u8;
        output[2] = a2 as u8;
        output[3] = a3 as u8;
        output[4] = (a0 >> 8) as u8;
        output[5] = (a1 >> 8) as u8;
        output[6] = (a2 >> 8) as u8;
        output[7] = (a3 >> 8) as u8;
        output[8] = (a0 >> 16) as u8;
        output[9] = (a1 >> 16) as u8;
        output[10] = (a2 >> 16) as u8;
        output[11] = (a3 >> 16) as u8;
        output[12] = (a0 >> 24) as u8;
        output[13] = (a1 >> 24) as u8;
        output[14] = (a2 >> 24) as u8;
        output[15] = (a3 >> 24) as u8;
    }

    write_row_major(x0, &mut output[0..16]);
    write_row_major(x1, &mut output[16..32]);
    write_row_major(x2, &mut output[32..48]);
    write_row_major(x3, &mut output[48..64]);
    write_row_major(x4, &mut output[64..80]);
    write_row_major(x5, &mut output[80..96]);
    write_row_major(x6, &mut output[96..112]);
    write_row_major(x7, &mut output[112..128])
}

// The Gf2Ops, Gf4Ops, and Gf8Ops traits specify the functions needed to calculate the AES S-Box
// values. This particuar implementation of those S-Box values is taken from [7], so that is where
// to look for details on how all that all works. This includes the transformations matrices defined
// below for the change_basis operation on the u32 and u32x4 types.

// Operations in GF(2^2) using normal basis (Omega^2,Omega)
trait Gf2Ops {
    // multiply
    fn mul(self, y: Self) -> Self;

    // scale by N = Omega^2
    fn scl_n(self) -> Self;

    // scale by N^2 = Omega
    fn scl_n2(self) -> Self;

    // square
    fn sq(self) -> Self;

    // Same as sqaure
    fn inv(self) -> Self;
}

impl <T: BitXor<Output = T> + BitAnd<Output = T> + Copy> Gf2Ops for Bs2State<T> {
    fn mul(self, y: Bs2State<T>) -> Bs2State<T> {
        let (b, a) = self.split();
        let (d, c) = y.split();
        let e = (a ^ b) & (c ^ d);
        let p = (a & c) ^ e;
        let q = (b & d) ^ e;
        Bs2State(q, p)
    }

    fn scl_n(self) -> Bs2State<T> {
        let (b, a) = self.split();
        let q = a ^ b;
        Bs2State(q, b)
    }

    fn scl_n2(self) -> Bs2State<T> {
        let (b, a) = self.split();
        let p = a ^ b;
        let q = a;
        Bs2State(q, p)
    }

    fn sq(self) -> Bs2State<T> {
        let (b, a) = self.split();
        Bs2State(a, b)
    }

    fn inv(self) -> Bs2State<T> {
        self.sq()
    }
}

// Operations in GF(2^4) using normal basis (alpha^8,alpha^2)
trait Gf4Ops {
    // multiply
    fn mul(self, y: Self) -> Self;

    // square & scale by nu
    // nu = beta^8 = N^2*alpha^2, N = w^2
    fn sq_scl(self) -> Self;

    // inverse
    fn inv(self) -> Self;
}

impl <T: BitXor<Output = T> + BitAnd<Output = T> + Copy> Gf4Ops for Bs4State<T> {
    fn mul(self, y: Bs4State<T>) -> Bs4State<T> {
        let (b, a) = self.split();
        let (d, c) = y.split();
        let f = c.xor(d);
        let e = a.xor(b).mul(f).scl_n();
        let p = a.mul(c).xor(e);
        let q = b.mul(d).xor(e);
        q.join(p)
    }

    fn sq_scl(self) -> Bs4State<T> {
        let (b, a) = self.split();
        let p = a.xor(b).sq();
        let q = b.sq().scl_n2();
        q.join(p)
    }

    fn inv(self) -> Bs4State<T> {
        let (b, a) = self.split();
        let c = a.xor(b).sq().scl_n();
        let d = a.mul(b);
        let e = c.xor(d).inv();
        let p = e.mul(b);
        let q = e.mul(a);
        q.join(p)
    }
}

// Operations in GF(2^8) using normal basis (d^16,d)
trait Gf8Ops {
    // inverse
    fn inv(&self) -> Self;
}

impl <T: BitXor<Output = T> + BitAnd<Output = T> + Copy + Default> Gf8Ops for Bs8State<T> {
    fn inv(&self) -> Bs8State<T> {
        let (b, a) = self.split();
        let c = a.xor(b).sq_scl();
        let d = a.mul(b);
        let e = c.xor(d).inv();
        let p = e.mul(b);
        let q = e.mul(a);
        q.join(p)
    }
}

impl <T: AesBitValueOps + Copy + 'static> AesOps for Bs8State<T> {
    fn sub_bytes(self) -> Bs8State<T> {
        let nb: Bs8State<T> = self.change_basis_a2x();
        let inv = nb.inv();
        let nb2: Bs8State<T> = inv.change_basis_x2s();
        nb2.xor_x63()
    }

    fn inv_sub_bytes(self) -> Bs8State<T> {
        let t = self.xor_x63();
        let nb: Bs8State<T> = t.change_basis_s2x();
        let inv = nb.inv();
        inv.change_basis_x2a()
    }

    fn shift_rows(self) -> Bs8State<T> {
        let Bs8State(x0, x1, x2, x3, x4, x5, x6, x7) = self;
        Bs8State(
            x0.shift_row(),
            x1.shift_row(),
            x2.shift_row(),
            x3.shift_row(),
            x4.shift_row(),
            x5.shift_row(),
            x6.shift_row(),
            x7.shift_row())
    }

    fn inv_shift_rows(self) -> Bs8State<T> {
        let Bs8State(x0, x1, x2, x3, x4, x5, x6, x7) = self;
        Bs8State(
            x0.inv_shift_row(),
            x1.inv_shift_row(),
            x2.inv_shift_row(),
            x3.inv_shift_row(),
            x4.inv_shift_row(),
            x5.inv_shift_row(),
            x6.inv_shift_row(),
            x7.inv_shift_row())
    }

    // Formula from [5]
    fn mix_columns(self) -> Bs8State<T> {
        let Bs8State(x0, x1, x2, x3, x4, x5, x6, x7) = self;

        let x0out = x7 ^ x7.ror1() ^ x0.ror1() ^ (x0 ^ x0.ror1()).ror2();
        let x1out = x0 ^ x0.ror1() ^ x7 ^ x7.ror1() ^ x1.ror1() ^ (x1 ^ x1.ror1()).ror2();
        let x2out = x1 ^ x1.ror1() ^ x2.ror1() ^ (x2 ^ x2.ror1()).ror2();
        let x3out = x2 ^ x2.ror1() ^ x7 ^ x7.ror1() ^ x3.ror1() ^ (x3 ^ x3.ror1()).ror2();
        let x4out = x3 ^ x3.ror1() ^ x7 ^ x7.ror1() ^ x4.ror1() ^ (x4 ^ x4.ror1()).ror2();
        let x5out = x4 ^ x4.ror1() ^ x5.ror1() ^ (x5 ^ x5.ror1()).ror2();
        let x6out = x5 ^ x5.ror1() ^ x6.ror1() ^ (x6 ^ x6.ror1()).ror2();
        let x7out = x6 ^ x6.ror1() ^ x7.ror1() ^ (x7 ^ x7.ror1()).ror2();

        Bs8State(x0out, x1out, x2out, x3out, x4out, x5out, x6out, x7out)
    }

    // Formula from [6]
    fn inv_mix_columns(self) -> Bs8State<T> {
        let Bs8State(x0, x1, x2, x3, x4, x5, x6, x7) = self;

        let x0out = x5 ^ x6 ^ x7 ^
            (x5 ^ x7 ^ x0).ror1() ^
            (x0 ^ x5 ^ x6).ror2() ^
            (x5 ^ x0).ror3();
        let x1out = x5 ^ x0 ^
            (x6 ^ x5 ^ x0 ^ x7 ^ x1).ror1() ^
            (x1 ^ x7 ^ x5).ror2() ^
            (x6 ^ x5 ^ x1).ror3();
        let x2out = x6 ^ x0 ^ x1 ^
            (x7 ^ x6 ^ x1 ^ x2).ror1() ^
            (x0 ^ x2 ^ x6).ror2() ^
            (x7 ^ x6 ^ x2).ror3();
        let x3out = x0 ^ x5 ^ x1 ^ x6 ^ x2 ^
            (x0 ^ x5 ^ x2 ^ x3).ror1() ^
            (x0 ^ x1 ^ x3 ^ x5 ^ x6 ^ x7).ror2() ^
            (x0 ^ x5 ^ x7 ^ x3).ror3();
        let x4out = x1 ^ x5 ^ x2 ^ x3 ^
            (x1 ^ x6 ^ x5 ^ x3 ^ x7 ^ x4).ror1() ^
            (x1 ^ x2 ^ x4 ^ x5 ^ x7).ror2() ^
            (x1 ^ x5 ^ x6 ^ x4).ror3();
        let x5out = x2 ^ x6 ^ x3 ^ x4 ^
            (x2 ^ x7 ^ x6 ^ x4 ^ x5).ror1() ^
            (x2 ^ x3 ^ x5 ^ x6).ror2() ^
            (x2 ^ x6 ^ x7 ^ x5).ror3();
        let x6out =  x3 ^ x7 ^ x4 ^ x5 ^
            (x3 ^ x7 ^ x5 ^ x6).ror1() ^
            (x3 ^ x4 ^ x6 ^ x7).ror2() ^
            (x3 ^ x7 ^ x6).ror3();
        let x7out = x4 ^ x5 ^ x6 ^
            (x4 ^ x6 ^ x7).ror1() ^
            (x4 ^ x5 ^ x7).ror2() ^
            (x4 ^ x7).ror3();

        Bs8State(x0out, x1out, x2out, x3out, x4out, x5out, x6out, x7out)
    }

    fn add_round_key(self, rk: &Bs8State<T>) -> Bs8State<T> {
        self.xor(*rk)
    }
}

trait AesBitValueOps: BitXor<Output = Self> + BitAnd<Output = Self> + Not<Output = Self> + Default + Sized {
    fn shift_row(self) -> Self;
    fn inv_shift_row(self) -> Self;
    fn ror1(self) -> Self;
    fn ror2(self) -> Self;
    fn ror3(self) -> Self;
}

impl AesBitValueOps for u16 {
    fn shift_row(self) -> u16 {
        // first 4 bits represent first row - don't shift
        (self & 0x000f) |
        // next 4 bits represent 2nd row - left rotate 1 bit
        ((self & 0x00e0) >> 1) | ((self & 0x0010) << 3) |
        // next 4 bits represent 3rd row - left rotate 2 bits
        ((self & 0x0c00) >> 2) | ((self & 0x0300) << 2) |
        // next 4 bits represent 4th row - left rotate 3 bits
        ((self & 0x8000) >> 3) | ((self & 0x7000) << 1)
    }

    fn inv_shift_row(self) -> u16 {
        // first 4 bits represent first row - don't shift
        (self & 0x000f) |
        // next 4 bits represent 2nd row - right rotate 1 bit
        ((self & 0x0080) >> 3) | ((self & 0x0070) << 1) |
        // next 4 bits represent 3rd row - right rotate 2 bits
        ((self & 0x0c00) >> 2) | ((self & 0x0300) << 2) |
        // next 4 bits represent 4th row - right rotate 3 bits
        ((self & 0xe000) >> 1) | ((self & 0x1000) << 3)
    }

    fn ror1(self) -> u16 {
        self >> 4 | self << 12
    }

    fn ror2(self) -> u16 {
        self >> 8 | self << 8
    }

    fn ror3(self) -> u16 {
        self >> 12 | self << 4
    }
}

impl u32x4 {
    fn lsh(self, s: u32) -> u32x4 {
        let u32x4(a0, a1, a2, a3) = self;
        u32x4(
            a0 << s,
            (a1 << s) | (a0 >> (32 - s)),
            (a2 << s) | (a1 >> (32 - s)),
            (a3 << s) | (a2 >> (32 - s)))
    }

    fn rsh(self, s: u32) -> u32x4 {
        let u32x4(a0, a1, a2, a3) = self;
        u32x4(
            (a0 >> s) | (a1 << (32 - s)),
            (a1 >> s) | (a2 << (32 - s)),
            (a2 >> s) | (a3 << (32 - s)),
            a3 >> s)
    }
}

impl Not for u32x4 {
    type Output = u32x4;

    fn not(self) -> u32x4 {
        self ^ U32X4_1
    }
}

impl Default for u32x4 {
    fn default() -> u32x4 {
        u32x4(0, 0, 0, 0)
    }
}

impl AesBitValueOps for u32x4 {
    fn shift_row(self) -> u32x4 {
        let u32x4(a0, a1, a2, a3) = self;
        u32x4(a0, a1 >> 8 | a1 << 24, a2 >> 16 | a2 << 16, a3 >> 24 | a3 << 8)
    }

    fn inv_shift_row(self) -> u32x4 {
        let u32x4(a0, a1, a2, a3) = self;
        u32x4(a0, a1 >> 24 | a1 << 8, a2 >> 16 | a2 << 16, a3 >> 8 | a3 << 24)
    }

    fn ror1(self) -> u32x4 {
        let u32x4(a0, a1, a2, a3) = self;
        u32x4(a1, a2, a3, a0)
    }

    fn ror2(self) -> u32x4 {
        let u32x4(a0, a1, a2, a3) = self;
        u32x4(a2, a3, a0, a1)
    }

    fn ror3(self) -> u32x4 {
        let u32x4(a0, a1, a2, a3) = self;
        u32x4(a3, a0, a1, a2)
    }
}