-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy pathyescrypt-opt.c
1531 lines (1362 loc) · 40.8 KB
/
yescrypt-opt.c
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
/*-
* Copyright 2009 Colin Percival
* Copyright 2012-2025 Alexander Peslyak
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*
* This file was originally written by Colin Percival as part of the Tarsnap
* online backup system.
*/
/*
* AVX and especially XOP speed up Salsa20 a lot, but this mostly matters for
* classic scrypt and for YESCRYPT_WORM (which use 8 rounds of Salsa20 per
* sub-block), and much less so for YESCRYPT_RW (which uses 2 rounds of Salsa20
* per block except during pwxform S-box initialization).
*/
#ifdef __GNUC__
#ifdef __XOP__
#warning "Note: XOP is enabled. That's great."
#elif defined(__AVX512VL__)
#warning "Note: AVX512VL is enabled. That's great."
#elif defined(__AVX__)
#warning "Note: AVX is enabled, which is great for classic scrypt and YESCRYPT_WORM, but is sometimes slightly slower than plain SSE2 for YESCRYPT_RW"
#elif defined(__SSE2__)
#warning "Note: AVX and XOP are not enabled, which is great for YESCRYPT_RW, but they would substantially improve performance at classic scrypt and YESCRYPT_WORM"
#elif defined(__x86_64__) || defined(__i386__)
#warning "SSE2 not enabled. Expect poor performance."
#else
#warning "Note: building generic code for non-x86. That's OK."
#endif
#endif
/*
* The SSE4 code version has fewer instructions than the generic SSE2 version,
* but all of the instructions are SIMD, thereby wasting the scalar execution
* units. Thus, the generic SSE2 version below actually runs faster on some
* CPUs due to its balanced mix of SIMD and scalar instructions.
*/
#undef USE_SSE4_FOR_32BIT
#ifdef __SSE2__
/*
* GCC before 4.9 would by default unnecessarily use store/load (without
* SSE4.1) or (V)PEXTR (with SSE4.1 or AVX) instead of simply (V)MOV.
* This was tracked as GCC bug 54349.
* "-mtune=corei7" works around this, but is only supported for GCC 4.6+.
* We use inline asm for pre-4.6 GCC, further down this file.
*/
#if __GNUC__ == 4 && __GNUC_MINOR__ >= 6 && __GNUC_MINOR__ < 9 && \
!defined(__clang__) && !defined(__ICC)
#pragma GCC target ("tune=corei7")
#endif
#include <emmintrin.h>
#ifdef __XOP__
#include <x86intrin.h>
#elif defined(__AVX512VL__)
#include <immintrin.h>
#endif
#elif defined(__SSE__)
#include <xmmintrin.h>
#endif
#include <errno.h>
#include <stdint.h>
#include <stdlib.h>
#include <string.h>
#include "insecure_memzero.h"
#include "sha256.h"
#include "sysendian.h"
#define YESCRYPT_INTERNAL
#include "yescrypt.h"
#include "yescrypt-platform.c"
#if __STDC_VERSION__ >= 199901L
/* Have restrict */
#elif defined(__GNUC__)
#define restrict __restrict
#else
#define restrict
#endif
#ifdef __GNUC__
#define unlikely(exp) __builtin_expect(exp, 0)
#else
#define unlikely(exp) (exp)
#endif
#ifdef __SSE__
#define PREFETCH(x, hint) _mm_prefetch((const char *)(x), (hint));
#else
#undef PREFETCH
#endif
typedef union {
uint32_t w[16];
uint64_t d[8];
#ifdef __SSE2__
__m128i q[4];
#endif
} salsa20_blk_t;
static inline void salsa20_simd_shuffle(const salsa20_blk_t *Bin,
salsa20_blk_t *Bout)
{
#define COMBINE(out, in1, in2) \
Bout->d[out] = Bin->w[in1 * 2] | ((uint64_t)Bin->w[in2 * 2 + 1] << 32);
COMBINE(0, 0, 2)
COMBINE(1, 5, 7)
COMBINE(2, 2, 4)
COMBINE(3, 7, 1)
COMBINE(4, 4, 6)
COMBINE(5, 1, 3)
COMBINE(6, 6, 0)
COMBINE(7, 3, 5)
#undef COMBINE
}
static inline void salsa20_simd_unshuffle(const salsa20_blk_t *Bin,
salsa20_blk_t *Bout)
{
#define UNCOMBINE(out, in1, in2) \
Bout->w[out * 2] = Bin->d[in1]; \
Bout->w[out * 2 + 1] = Bin->d[in2] >> 32;
UNCOMBINE(0, 0, 6)
UNCOMBINE(1, 5, 3)
UNCOMBINE(2, 2, 0)
UNCOMBINE(3, 7, 5)
UNCOMBINE(4, 4, 2)
UNCOMBINE(5, 1, 7)
UNCOMBINE(6, 6, 4)
UNCOMBINE(7, 3, 1)
#undef UNCOMBINE
}
#ifdef __SSE2__
#define DECL_X \
__m128i X0, X1, X2, X3;
#define DECL_Y \
__m128i Y0, Y1, Y2, Y3;
#define READ_X(in) \
X0 = (in).q[0]; X1 = (in).q[1]; X2 = (in).q[2]; X3 = (in).q[3];
#define WRITE_X(out) \
(out).q[0] = X0; (out).q[1] = X1; (out).q[2] = X2; (out).q[3] = X3;
#ifdef __XOP__
#define ARX(out, in1, in2, s) \
out = _mm_xor_si128(out, _mm_roti_epi32(_mm_add_epi32(in1, in2), s));
#elif defined(__AVX512VL__)
#define ARX(out, in1, in2, s) \
out = _mm_xor_si128(out, _mm_rol_epi32(_mm_add_epi32(in1, in2), s));
#else
#define ARX(out, in1, in2, s) { \
__m128i tmp = _mm_add_epi32(in1, in2); \
out = _mm_xor_si128(out, _mm_slli_epi32(tmp, s)); \
out = _mm_xor_si128(out, _mm_srli_epi32(tmp, 32 - s)); \
}
#endif
#define SALSA20_2ROUNDS \
/* Operate on "columns" */ \
ARX(X1, X0, X3, 7) \
ARX(X2, X1, X0, 9) \
ARX(X3, X2, X1, 13) \
ARX(X0, X3, X2, 18) \
/* Rearrange data */ \
X1 = _mm_shuffle_epi32(X1, 0x93); \
X2 = _mm_shuffle_epi32(X2, 0x4E); \
X3 = _mm_shuffle_epi32(X3, 0x39); \
/* Operate on "rows" */ \
ARX(X3, X0, X1, 7) \
ARX(X2, X3, X0, 9) \
ARX(X1, X2, X3, 13) \
ARX(X0, X1, X2, 18) \
/* Rearrange data */ \
X1 = _mm_shuffle_epi32(X1, 0x39); \
X2 = _mm_shuffle_epi32(X2, 0x4E); \
X3 = _mm_shuffle_epi32(X3, 0x93);
/**
* Apply the Salsa20 core to the block provided in (X0 ... X3).
*/
#define SALSA20_wrapper(out, rounds) { \
__m128i Z0 = X0, Z1 = X1, Z2 = X2, Z3 = X3; \
rounds \
(out).q[0] = X0 = _mm_add_epi32(X0, Z0); \
(out).q[1] = X1 = _mm_add_epi32(X1, Z1); \
(out).q[2] = X2 = _mm_add_epi32(X2, Z2); \
(out).q[3] = X3 = _mm_add_epi32(X3, Z3); \
}
/**
* Apply the Salsa20/2 core to the block provided in X.
*/
#define SALSA20_2(out) \
SALSA20_wrapper(out, SALSA20_2ROUNDS)
#define SALSA20_8ROUNDS \
SALSA20_2ROUNDS SALSA20_2ROUNDS SALSA20_2ROUNDS SALSA20_2ROUNDS
#define XOR_X(in) \
X0 = _mm_xor_si128(X0, (in).q[0]); \
X1 = _mm_xor_si128(X1, (in).q[1]); \
X2 = _mm_xor_si128(X2, (in).q[2]); \
X3 = _mm_xor_si128(X3, (in).q[3]);
#define XOR_X_2(in1, in2) \
X0 = _mm_xor_si128((in1).q[0], (in2).q[0]); \
X1 = _mm_xor_si128((in1).q[1], (in2).q[1]); \
X2 = _mm_xor_si128((in1).q[2], (in2).q[2]); \
X3 = _mm_xor_si128((in1).q[3], (in2).q[3]);
#define XOR_X_WRITE_XOR_Y_2(out, in) \
(out).q[0] = Y0 = _mm_xor_si128((out).q[0], (in).q[0]); \
(out).q[1] = Y1 = _mm_xor_si128((out).q[1], (in).q[1]); \
(out).q[2] = Y2 = _mm_xor_si128((out).q[2], (in).q[2]); \
(out).q[3] = Y3 = _mm_xor_si128((out).q[3], (in).q[3]); \
X0 = _mm_xor_si128(X0, Y0); \
X1 = _mm_xor_si128(X1, Y1); \
X2 = _mm_xor_si128(X2, Y2); \
X3 = _mm_xor_si128(X3, Y3);
/**
* Apply the Salsa20/8 core to the block provided in X ^ in.
*/
#define SALSA20_8_XOR_MEM(in, out) \
XOR_X(in) \
SALSA20_wrapper(out, SALSA20_8ROUNDS)
#define INTEGERIFY _mm_cvtsi128_si32(X0)
#else /* !defined(__SSE2__) */
#define DECL_X \
salsa20_blk_t X;
#define DECL_Y \
salsa20_blk_t Y;
#define COPY(out, in) \
(out).d[0] = (in).d[0]; \
(out).d[1] = (in).d[1]; \
(out).d[2] = (in).d[2]; \
(out).d[3] = (in).d[3]; \
(out).d[4] = (in).d[4]; \
(out).d[5] = (in).d[5]; \
(out).d[6] = (in).d[6]; \
(out).d[7] = (in).d[7];
#define READ_X(in) COPY(X, in)
#define WRITE_X(out) COPY(out, X)
/**
* salsa20(B):
* Apply the Salsa20 core to the provided block.
*/
static inline void salsa20(salsa20_blk_t *restrict B,
salsa20_blk_t *restrict Bout, uint32_t doublerounds)
{
salsa20_blk_t X;
#define x X.w
salsa20_simd_unshuffle(B, &X);
do {
#define R(a,b) (((a) << (b)) | ((a) >> (32 - (b))))
/* Operate on columns */
x[ 4] ^= R(x[ 0]+x[12], 7); x[ 8] ^= R(x[ 4]+x[ 0], 9);
x[12] ^= R(x[ 8]+x[ 4],13); x[ 0] ^= R(x[12]+x[ 8],18);
x[ 9] ^= R(x[ 5]+x[ 1], 7); x[13] ^= R(x[ 9]+x[ 5], 9);
x[ 1] ^= R(x[13]+x[ 9],13); x[ 5] ^= R(x[ 1]+x[13],18);
x[14] ^= R(x[10]+x[ 6], 7); x[ 2] ^= R(x[14]+x[10], 9);
x[ 6] ^= R(x[ 2]+x[14],13); x[10] ^= R(x[ 6]+x[ 2],18);
x[ 3] ^= R(x[15]+x[11], 7); x[ 7] ^= R(x[ 3]+x[15], 9);
x[11] ^= R(x[ 7]+x[ 3],13); x[15] ^= R(x[11]+x[ 7],18);
/* Operate on rows */
x[ 1] ^= R(x[ 0]+x[ 3], 7); x[ 2] ^= R(x[ 1]+x[ 0], 9);
x[ 3] ^= R(x[ 2]+x[ 1],13); x[ 0] ^= R(x[ 3]+x[ 2],18);
x[ 6] ^= R(x[ 5]+x[ 4], 7); x[ 7] ^= R(x[ 6]+x[ 5], 9);
x[ 4] ^= R(x[ 7]+x[ 6],13); x[ 5] ^= R(x[ 4]+x[ 7],18);
x[11] ^= R(x[10]+x[ 9], 7); x[ 8] ^= R(x[11]+x[10], 9);
x[ 9] ^= R(x[ 8]+x[11],13); x[10] ^= R(x[ 9]+x[ 8],18);
x[12] ^= R(x[15]+x[14], 7); x[13] ^= R(x[12]+x[15], 9);
x[14] ^= R(x[13]+x[12],13); x[15] ^= R(x[14]+x[13],18);
#undef R
} while (--doublerounds);
#undef x
{
uint32_t i;
salsa20_simd_shuffle(&X, Bout);
for (i = 0; i < 16; i += 4) {
B->w[i] = Bout->w[i] += B->w[i];
B->w[i + 1] = Bout->w[i + 1] += B->w[i + 1];
B->w[i + 2] = Bout->w[i + 2] += B->w[i + 2];
B->w[i + 3] = Bout->w[i + 3] += B->w[i + 3];
}
}
#if 0
/* Too expensive */
insecure_memzero(&X, sizeof(X));
#endif
}
/**
* Apply the Salsa20/2 core to the block provided in X.
*/
#define SALSA20_2(out) \
salsa20(&X, &out, 1);
#define XOR(out, in1, in2) \
(out).d[0] = (in1).d[0] ^ (in2).d[0]; \
(out).d[1] = (in1).d[1] ^ (in2).d[1]; \
(out).d[2] = (in1).d[2] ^ (in2).d[2]; \
(out).d[3] = (in1).d[3] ^ (in2).d[3]; \
(out).d[4] = (in1).d[4] ^ (in2).d[4]; \
(out).d[5] = (in1).d[5] ^ (in2).d[5]; \
(out).d[6] = (in1).d[6] ^ (in2).d[6]; \
(out).d[7] = (in1).d[7] ^ (in2).d[7];
#define XOR_X(in) XOR(X, X, in)
#define XOR_X_2(in1, in2) XOR(X, in1, in2)
#define XOR_X_WRITE_XOR_Y_2(out, in) \
XOR(Y, out, in) \
COPY(out, Y) \
XOR(X, X, Y)
/**
* Apply the Salsa20/8 core to the block provided in X ^ in.
*/
#define SALSA20_8_XOR_MEM(in, out) \
XOR_X(in); \
salsa20(&X, &out, 4);
#define INTEGERIFY (uint32_t)X.d[0]
#endif
/**
* blockmix_salsa8(Bin, Bout, r):
* Compute Bout = BlockMix_{salsa20/8, r}(Bin). The input Bin must be 128r
* bytes in length; the output Bout must also be the same size.
*/
static void blockmix_salsa8(const salsa20_blk_t *restrict Bin,
salsa20_blk_t *restrict Bout, size_t r)
{
size_t i;
DECL_X
READ_X(Bin[r * 2 - 1])
for (i = 0; i < r; i++) {
SALSA20_8_XOR_MEM(Bin[i * 2], Bout[i])
SALSA20_8_XOR_MEM(Bin[i * 2 + 1], Bout[r + i])
}
}
static uint32_t blockmix_salsa8_xor(const salsa20_blk_t *restrict Bin1,
const salsa20_blk_t *restrict Bin2, salsa20_blk_t *restrict Bout,
size_t r)
{
size_t i;
DECL_X
#ifdef PREFETCH
PREFETCH(&Bin2[r * 2 - 1], _MM_HINT_T0)
for (i = 0; i < r - 1; i++) {
PREFETCH(&Bin2[i * 2], _MM_HINT_T0)
PREFETCH(&Bin2[i * 2 + 1], _MM_HINT_T0)
}
PREFETCH(&Bin2[i * 2], _MM_HINT_T0)
#endif
XOR_X_2(Bin1[r * 2 - 1], Bin2[r * 2 - 1])
for (i = 0; i < r; i++) {
XOR_X(Bin1[i * 2])
SALSA20_8_XOR_MEM(Bin2[i * 2], Bout[i])
XOR_X(Bin1[i * 2 + 1])
SALSA20_8_XOR_MEM(Bin2[i * 2 + 1], Bout[r + i])
}
return INTEGERIFY;
}
/* This is tunable */
#define Swidth 8
/* Not tunable in this implementation, hard-coded in a few places */
#define PWXsimple 2
#define PWXgather 4
/* Derived values. Not tunable except via Swidth above. */
#define PWXbytes (PWXgather * PWXsimple * 8)
#define Sbytes (3 * (1 << Swidth) * PWXsimple * 8)
#define Smask (((1 << Swidth) - 1) * PWXsimple * 8)
#define Smask2 (((uint64_t)Smask << 32) | Smask)
#define DECL_SMASK2REG /* empty */
#define FORCE_REGALLOC_3 /* empty */
#define MAYBE_MEMORY_BARRIER /* empty */
#ifdef __SSE2__
/*
* (V)PSRLDQ and (V)PSHUFD have higher throughput than (V)PSRLQ on some CPUs
* starting with Sandy Bridge. Additionally, PSHUFD uses separate source and
* destination registers, whereas the shifts would require an extra move
* instruction for our code when building without AVX. Unfortunately, PSHUFD
* is much slower on Conroe (4 cycles latency vs. 1 cycle latency for PSRLQ)
* and somewhat slower on some non-Intel CPUs (luckily not including AMD
* Bulldozer and Piledriver).
*/
#ifdef __AVX__
#define HI32(X) \
_mm_srli_si128((X), 4)
#elif 1 /* As an option, check for __SSE4_1__ here not to hurt Conroe */
#define HI32(X) \
_mm_shuffle_epi32((X), _MM_SHUFFLE(2,3,0,1))
#else
#define HI32(X) \
_mm_srli_epi64((X), 32)
#endif
#if defined(__x86_64__) && \
__GNUC__ == 4 && __GNUC_MINOR__ < 6 && !defined(__ICC)
#ifdef __AVX__
#define MOVQ "vmovq"
#else
/* "movq" would be more correct, but "movd" is supported by older binutils
* due to an error in AMD's spec for x86-64. */
#define MOVQ "movd"
#endif
#define EXTRACT64(X) ({ \
uint64_t result; \
__asm__(MOVQ " %1, %0" : "=r" (result) : "x" (X)); \
result; \
})
#elif defined(__x86_64__) && !defined(_MSC_VER) && !defined(__OPEN64__)
/* MSVC and Open64 had bugs */
#define EXTRACT64(X) _mm_cvtsi128_si64(X)
#elif defined(__x86_64__) && defined(__SSE4_1__)
/* No known bugs for this intrinsic */
#include <smmintrin.h>
#define EXTRACT64(X) _mm_extract_epi64((X), 0)
#elif defined(USE_SSE4_FOR_32BIT) && defined(__SSE4_1__)
/* 32-bit */
#include <smmintrin.h>
#if 0
/* This is currently unused by the code below, which instead uses these two
* intrinsics explicitly when (!defined(__x86_64__) && defined(__SSE4_1__)) */
#define EXTRACT64(X) \
((uint64_t)(uint32_t)_mm_cvtsi128_si32(X) | \
((uint64_t)(uint32_t)_mm_extract_epi32((X), 1) << 32))
#endif
#else
/* 32-bit or compilers with known past bugs in _mm_cvtsi128_si64() */
#define EXTRACT64(X) \
((uint64_t)(uint32_t)_mm_cvtsi128_si32(X) | \
((uint64_t)(uint32_t)_mm_cvtsi128_si32(HI32(X)) << 32))
#endif
#if defined(__x86_64__) && (defined(__AVX__) || !defined(__GNUC__))
/* 64-bit with AVX */
/* Force use of 64-bit AND instead of two 32-bit ANDs */
#undef DECL_SMASK2REG
#if defined(__GNUC__) && !defined(__ICC)
#define DECL_SMASK2REG uint64_t Smask2reg = Smask2;
/* Force use of lower-numbered registers to reduce number of prefixes, relying
* on out-of-order execution and register renaming. */
#define FORCE_REGALLOC_1 \
__asm__("" : "=a" (x), "+d" (Smask2reg), "+S" (S0), "+D" (S1));
#define FORCE_REGALLOC_2 \
__asm__("" : : "c" (lo));
#else
static volatile uint64_t Smask2var = Smask2;
#define DECL_SMASK2REG uint64_t Smask2reg = Smask2var;
#define FORCE_REGALLOC_1 /* empty */
#define FORCE_REGALLOC_2 /* empty */
#endif
#define PWXFORM_SIMD(X) { \
uint64_t x; \
FORCE_REGALLOC_1 \
uint32_t lo = x = EXTRACT64(X) & Smask2reg; \
FORCE_REGALLOC_2 \
uint32_t hi = x >> 32; \
X = _mm_mul_epu32(HI32(X), X); \
X = _mm_add_epi64(X, *(__m128i *)(S0 + lo)); \
X = _mm_xor_si128(X, *(__m128i *)(S1 + hi)); \
}
#elif defined(__x86_64__)
/* 64-bit without AVX. This relies on out-of-order execution and register
* renaming. It may actually be fastest on CPUs with AVX(2) as well - e.g.,
* it runs great on Haswell. */
#warning "Note: using x86-64 inline assembly for YESCRYPT_RW. That's great."
/* We need a compiler memory barrier between sub-blocks to ensure that none of
* the writes into what was S2 during processing of the previous sub-block are
* postponed until after a read from S0 or S1 in the inline asm code below. */
#undef MAYBE_MEMORY_BARRIER
#define MAYBE_MEMORY_BARRIER \
__asm__("" : : : "memory");
#ifdef __ILP32__ /* x32 */
#define REGISTER_PREFIX "e"
#else
#define REGISTER_PREFIX "r"
#endif
#define PWXFORM_SIMD(X) { \
__m128i H; \
__asm__( \
"movd %0, %%rax\n\t" \
"pshufd $0xb1, %0, %1\n\t" \
"andq %2, %%rax\n\t" \
"pmuludq %1, %0\n\t" \
"movl %%eax, %%ecx\n\t" \
"shrq $0x20, %%rax\n\t" \
"paddq (%3,%%" REGISTER_PREFIX "cx), %0\n\t" \
"pxor (%4,%%" REGISTER_PREFIX "ax), %0\n\t" \
: "+x" (X), "=x" (H) \
: "d" (Smask2), "S" (S0), "D" (S1) \
: "cc", "ax", "cx"); \
}
#elif defined(USE_SSE4_FOR_32BIT) && defined(__SSE4_1__)
/* 32-bit with SSE4.1 */
#define PWXFORM_SIMD(X) { \
__m128i x = _mm_and_si128(X, _mm_set1_epi64x(Smask2)); \
__m128i s0 = *(__m128i *)(S0 + (uint32_t)_mm_cvtsi128_si32(x)); \
__m128i s1 = *(__m128i *)(S1 + (uint32_t)_mm_extract_epi32(x, 1)); \
X = _mm_mul_epu32(HI32(X), X); \
X = _mm_add_epi64(X, s0); \
X = _mm_xor_si128(X, s1); \
}
#else
/* 32-bit without SSE4.1 */
#define PWXFORM_SIMD(X) { \
uint64_t x = EXTRACT64(X) & Smask2; \
__m128i s0 = *(__m128i *)(S0 + (uint32_t)x); \
__m128i s1 = *(__m128i *)(S1 + (x >> 32)); \
X = _mm_mul_epu32(HI32(X), X); \
X = _mm_add_epi64(X, s0); \
X = _mm_xor_si128(X, s1); \
}
#endif
#define PWXFORM_ROUND \
PWXFORM_SIMD(X0) \
PWXFORM_SIMD(X1) \
PWXFORM_SIMD(X2) \
PWXFORM_SIMD(X3)
#if defined(__x86_64__) && defined(__GNUC__) && !defined(__ICC)
#undef FORCE_REGALLOC_3
#define FORCE_REGALLOC_3 __asm__("" : : "b" (Sw));
#endif
#else /* !defined(__SSE2__) */
#define PWXFORM_SIMD(x0, x1) { \
uint64_t x = x0 & Smask2; \
uint64_t *p0 = (uint64_t *)(S0 + (uint32_t)x); \
uint64_t *p1 = (uint64_t *)(S1 + (x >> 32)); \
x0 = ((x0 >> 32) * (uint32_t)x0 + p0[0]) ^ p1[0]; \
x1 = ((x1 >> 32) * (uint32_t)x1 + p0[1]) ^ p1[1]; \
}
#define PWXFORM_ROUND \
PWXFORM_SIMD(X.d[0], X.d[1]) \
PWXFORM_SIMD(X.d[2], X.d[3]) \
PWXFORM_SIMD(X.d[4], X.d[5]) \
PWXFORM_SIMD(X.d[6], X.d[7])
#endif
/*
* This offset helps address the 256-byte write block via the single-byte
* displacements encodable in x86(-64) instructions. It is needed because the
* displacements are signed. Without it, we'd get 4-byte displacements for
* half of the writes. Setting it to 0x80 instead of 0x7c would avoid needing
* a displacement for one of the writes, but then the LEA instruction would
* need a 4-byte displacement.
*/
#define PWXFORM_WRITE_OFFSET 0x7c
#define PWXFORM_WRITE \
WRITE_X(*(salsa20_blk_t *)(Sw - PWXFORM_WRITE_OFFSET)) \
Sw += 64;
#define PWXFORM { \
uint8_t *Sw = S2 + w + PWXFORM_WRITE_OFFSET; \
FORCE_REGALLOC_3 \
MAYBE_MEMORY_BARRIER \
PWXFORM_ROUND \
PWXFORM_ROUND PWXFORM_WRITE \
PWXFORM_ROUND PWXFORM_WRITE \
PWXFORM_ROUND PWXFORM_WRITE \
PWXFORM_ROUND PWXFORM_WRITE \
PWXFORM_ROUND \
w = (w + 64 * 4) & Smask2; \
{ \
uint8_t *Stmp = S2; \
S2 = S1; \
S1 = S0; \
S0 = Stmp; \
} \
}
typedef struct {
uint8_t *S0, *S1, *S2;
size_t w;
} pwxform_ctx_t;
#define Salloc (Sbytes + ((sizeof(pwxform_ctx_t) + 63) & ~63U))
/**
* blockmix_pwxform(Bin, Bout, r, S):
* Compute Bout = BlockMix_pwxform{salsa20/2, r, S}(Bin). The input Bin must
* be 128r bytes in length; the output Bout must also be the same size.
*/
static void blockmix(const salsa20_blk_t *restrict Bin,
salsa20_blk_t *restrict Bout, size_t r, pwxform_ctx_t *restrict ctx)
{
uint8_t *S0 = ctx->S0, *S1 = ctx->S1, *S2 = ctx->S2;
size_t w = ctx->w;
size_t i;
DECL_X
/* Convert count of 128-byte blocks to max index of 64-byte block */
r = r * 2 - 1;
READ_X(Bin[r])
DECL_SMASK2REG
i = 0;
do {
XOR_X(Bin[i])
PWXFORM
if (unlikely(i >= r))
break;
WRITE_X(Bout[i])
i++;
} while (1);
ctx->S0 = S0; ctx->S1 = S1; ctx->S2 = S2;
ctx->w = w;
SALSA20_2(Bout[i])
}
static uint32_t blockmix_xor(const salsa20_blk_t *Bin1,
const salsa20_blk_t *restrict Bin2, salsa20_blk_t *Bout,
size_t r, int Bin2_in_ROM, pwxform_ctx_t *restrict ctx)
{
uint8_t *S0 = ctx->S0, *S1 = ctx->S1, *S2 = ctx->S2;
size_t w = ctx->w;
size_t i;
DECL_X
/* Convert count of 128-byte blocks to max index of 64-byte block */
r = r * 2 - 1;
#ifdef PREFETCH
if (Bin2_in_ROM) {
PREFETCH(&Bin2[r], _MM_HINT_NTA)
for (i = 0; i < r; i++) {
PREFETCH(&Bin2[i], _MM_HINT_NTA)
}
} else {
PREFETCH(&Bin2[r], _MM_HINT_T0)
for (i = 0; i < r; i++) {
PREFETCH(&Bin2[i], _MM_HINT_T0)
}
}
#else
(void)Bin2_in_ROM; /* unused */
#endif
XOR_X_2(Bin1[r], Bin2[r])
DECL_SMASK2REG
i = 0;
r--;
do {
XOR_X(Bin1[i])
XOR_X(Bin2[i])
PWXFORM
WRITE_X(Bout[i])
XOR_X(Bin1[i + 1])
XOR_X(Bin2[i + 1])
PWXFORM
if (unlikely(i >= r))
break;
WRITE_X(Bout[i + 1])
i += 2;
} while (1);
i++;
ctx->S0 = S0; ctx->S1 = S1; ctx->S2 = S2;
ctx->w = w;
SALSA20_2(Bout[i])
return INTEGERIFY;
}
static uint32_t blockmix_xor_save(salsa20_blk_t *restrict Bin1out,
salsa20_blk_t *restrict Bin2,
size_t r, pwxform_ctx_t *restrict ctx)
{
uint8_t *S0 = ctx->S0, *S1 = ctx->S1, *S2 = ctx->S2;
size_t w = ctx->w;
size_t i;
DECL_X
DECL_Y
/* Convert count of 128-byte blocks to max index of 64-byte block */
r = r * 2 - 1;
#ifdef PREFETCH
PREFETCH(&Bin2[r], _MM_HINT_T0)
for (i = 0; i < r; i++) {
PREFETCH(&Bin2[i], _MM_HINT_T0)
}
#endif
XOR_X_2(Bin1out[r], Bin2[r])
DECL_SMASK2REG
i = 0;
r--;
do {
XOR_X_WRITE_XOR_Y_2(Bin2[i], Bin1out[i])
PWXFORM
WRITE_X(Bin1out[i])
XOR_X_WRITE_XOR_Y_2(Bin2[i + 1], Bin1out[i + 1])
PWXFORM
if (unlikely(i >= r))
break;
WRITE_X(Bin1out[i + 1])
i += 2;
} while (1);
i++;
ctx->S0 = S0; ctx->S1 = S1; ctx->S2 = S2;
ctx->w = w;
SALSA20_2(Bin1out[i])
return INTEGERIFY;
}
/**
* integerify(B, r):
* Return the result of parsing B_{2r-1} as a little-endian integer.
*/
static inline uint32_t integerify(const salsa20_blk_t *B, size_t r)
{
/*
* Our 64-bit words are in host byte order, which is why we don't just read
* w[0] here (would be wrong on big-endian). Also, our 32-bit words are
* SIMD-shuffled (so the next 32 bits would be part of d[6]), but currently
* this does not matter as we only care about the least significant 32 bits.
*/
return (uint32_t)B[2 * r - 1].d[0];
}
/**
* smix1(B, r, N, flags, V, NROM, VROM, XY, ctx):
* Compute first loop of B = SMix_r(B, N). The input B must be 128r bytes in
* length; the temporary storage V must be 128rN bytes in length; the temporary
* storage XY must be 128r+64 bytes in length. N must be even and at least 4.
* The array V must be aligned to a multiple of 64 bytes, and arrays B and XY
* to a multiple of at least 16 bytes.
*/
static void smix1(uint8_t *B, size_t r, uint32_t N, yescrypt_flags_t flags,
salsa20_blk_t *V, uint32_t NROM, const salsa20_blk_t *VROM,
salsa20_blk_t *XY, pwxform_ctx_t *ctx)
{
size_t s = 2 * r;
salsa20_blk_t *X = V, *Y = &V[s];
uint32_t i, j;
for (i = 0; i < 2 * r; i++) {
const salsa20_blk_t *src = (salsa20_blk_t *)&B[i * 64];
salsa20_blk_t *tmp = Y;
salsa20_blk_t *dst = &X[i];
size_t k;
for (k = 0; k < 16; k++)
tmp->w[k] = le32dec(&src->w[k]);
salsa20_simd_shuffle(tmp, dst);
}
if (VROM) {
uint32_t n;
const salsa20_blk_t *V_j;
V_j = &VROM[(NROM - 1) * s];
j = blockmix_xor(X, V_j, Y, r, 1, ctx) & (NROM - 1);
V_j = &VROM[j * s];
X = Y + s;
j = blockmix_xor(Y, V_j, X, r, 1, ctx);
for (n = 2; n < N; n <<= 1) {
uint32_t m = (n < N / 2) ? n : (N - 1 - n);
for (i = 1; i < m; i += 2) {
j &= n - 1;
j += i - 1;
V_j = &V[j * s];
Y = X + s;
j = blockmix_xor(X, V_j, Y, r, 0, ctx) & (NROM - 1);
V_j = &VROM[j * s];
X = Y + s;
j = blockmix_xor(Y, V_j, X, r, 1, ctx);
}
}
n >>= 1;
j &= n - 1;
j += N - 2 - n;
V_j = &V[j * s];
Y = X + s;
j = blockmix_xor(X, V_j, Y, r, 0, ctx) & (NROM - 1);
V_j = &VROM[j * s];
blockmix_xor(Y, V_j, XY, r, 1, ctx);
} else if (flags & YESCRYPT_RW) {
uint32_t n;
salsa20_blk_t *V_j;
blockmix(X, Y, r, ctx);
X = Y + s;
blockmix(Y, X, r, ctx);
j = integerify(X, r);
for (n = 2; n < N; n <<= 1) {
uint32_t m = (n < N / 2) ? n : (N - 1 - n);
for (i = 1; i < m; i += 2) {
Y = X + s;
j &= n - 1;
j += i - 1;
V_j = &V[j * s];
j = blockmix_xor(X, V_j, Y, r, 0, ctx);
j &= n - 1;
j += i;
V_j = &V[j * s];
X = Y + s;
j = blockmix_xor(Y, V_j, X, r, 0, ctx);
}
}
n >>= 1;
j &= n - 1;
j += N - 2 - n;
V_j = &V[j * s];
Y = X + s;
j = blockmix_xor(X, V_j, Y, r, 0, ctx);
j &= n - 1;
j += N - 1 - n;
V_j = &V[j * s];
blockmix_xor(Y, V_j, XY, r, 0, ctx);
} else {
N -= 2;
do {
blockmix_salsa8(X, Y, r);
X = Y + s;
blockmix_salsa8(Y, X, r);
Y = X + s;
} while ((N -= 2));
blockmix_salsa8(X, Y, r);
blockmix_salsa8(Y, XY, r);
}
for (i = 0; i < 2 * r; i++) {
const salsa20_blk_t *src = &XY[i];
salsa20_blk_t *tmp = &XY[s];
salsa20_blk_t *dst = (salsa20_blk_t *)&B[i * 64];
size_t k;
for (k = 0; k < 16; k++)
le32enc(&tmp->w[k], src->w[k]);
salsa20_simd_unshuffle(tmp, dst);
}
}
/**
* smix2(B, r, N, Nloop, flags, V, NROM, VROM, XY, ctx):
* Compute second loop of B = SMix_r(B, N). The input B must be 128r bytes in
* length; the temporary storage V must be 128rN bytes in length; the temporary
* storage XY must be 256r bytes in length. N must be a power of 2 and at
* least 2. Nloop must be even. The array V must be aligned to a multiple of
* 64 bytes, and arrays B and XY to a multiple of at least 16 bytes.
*/
static void smix2(uint8_t *B, size_t r, uint32_t N, uint64_t Nloop,
yescrypt_flags_t flags, salsa20_blk_t *V, uint32_t NROM,
const salsa20_blk_t *VROM, salsa20_blk_t *XY, pwxform_ctx_t *ctx)
{
size_t s = 2 * r;
salsa20_blk_t *X = XY, *Y = &XY[s];
uint32_t i, j;
if (Nloop == 0)
return;
for (i = 0; i < 2 * r; i++) {
const salsa20_blk_t *src = (salsa20_blk_t *)&B[i * 64];
salsa20_blk_t *tmp = Y;
salsa20_blk_t *dst = &X[i];
size_t k;
for (k = 0; k < 16; k++)
tmp->w[k] = le32dec(&src->w[k]);
salsa20_simd_shuffle(tmp, dst);
}
j = integerify(X, r) & (N - 1);
/*
* Normally, VROM implies YESCRYPT_RW, but we check for these separately
* because our SMix resets YESCRYPT_RW for the smix2() calls operating on the
* entire V when p > 1.
*/
if (VROM && (flags & YESCRYPT_RW)) {
do {
salsa20_blk_t *V_j = &V[j * s];
const salsa20_blk_t *VROM_j;
j = blockmix_xor_save(X, V_j, r, ctx) & (NROM - 1);
VROM_j = &VROM[j * s];
j = blockmix_xor(X, VROM_j, X, r, 1, ctx) & (N - 1);
} while (Nloop -= 2);
} else if (VROM) {
do {
const salsa20_blk_t *V_j = &V[j * s];
j = blockmix_xor(X, V_j, X, r, 0, ctx) & (NROM - 1);
V_j = &VROM[j * s];
j = blockmix_xor(X, V_j, X, r, 1, ctx) & (N - 1);
} while (Nloop -= 2);
} else if (flags & YESCRYPT_RW) {
do {
salsa20_blk_t *V_j = &V[j * s];
j = blockmix_xor_save(X, V_j, r, ctx) & (N - 1);
V_j = &V[j * s];
j = blockmix_xor_save(X, V_j, r, ctx) & (N - 1);
} while (Nloop -= 2);
} else if (ctx) {
do {
const salsa20_blk_t *V_j = &V[j * s];
j = blockmix_xor(X, V_j, X, r, 0, ctx) & (N - 1);
V_j = &V[j * s];
j = blockmix_xor(X, V_j, X, r, 0, ctx) & (N - 1);
} while (Nloop -= 2);
} else {
do {
const salsa20_blk_t *V_j = &V[j * s];
j = blockmix_salsa8_xor(X, V_j, Y, r) & (N - 1);
V_j = &V[j * s];
j = blockmix_salsa8_xor(Y, V_j, X, r) & (N - 1);
} while (Nloop -= 2);
}
for (i = 0; i < 2 * r; i++) {
const salsa20_blk_t *src = &X[i];
salsa20_blk_t *tmp = Y;
salsa20_blk_t *dst = (salsa20_blk_t *)&B[i * 64];
size_t k;
for (k = 0; k < 16; k++)
le32enc(&tmp->w[k], src->w[k]);
salsa20_simd_unshuffle(tmp, dst);
}
}