-
-
Notifications
You must be signed in to change notification settings - Fork 13
/
xxh3.h
2320 lines (2122 loc) · 91.5 KB
/
xxh3.h
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
/*
* xxHash - Extremely Fast Hash algorithm
* Development source file for `xxh3`
* Copyright (C) 2019-2020 Yann Collet
*
* BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are
* met:
*
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * 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 COPYRIGHT HOLDERS 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 COPYRIGHT
* OWNER 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.
*
* You can contact the author at:
* - xxHash homepage: https://www.xxhash.com
* - xxHash source repository: https://github.com/Cyan4973/xxHash
*/
/*
* Note: This file is separated for development purposes.
* It will be integrated into `xxhash.h` when development stage is completed.
*
* Credit: most of the work on vectorial and asm variants comes from @easyaspi314
*/
#ifndef XXH3_H_1397135465
#define XXH3_H_1397135465
/* === Dependencies === */
#ifndef XXHASH_H_5627135585666179
/* special: when including `xxh3.h` directly, turn on XXH_INLINE_ALL */
# undef XXH_INLINE_ALL /* avoid redefinition */
# define XXH_INLINE_ALL
#endif
#include "xxhash.h"
/* === Compiler specifics === */
#if defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L /* >= C99 */
# define XXH_RESTRICT restrict
#else
/* Note: it might be useful to define __restrict or __restrict__ for some C++ compilers */
# define XXH_RESTRICT /* disable */
#endif
#if (defined(__GNUC__) && (__GNUC__ >= 3)) \
|| (defined(__INTEL_COMPILER) && (__INTEL_COMPILER >= 800)) \
|| defined(__clang__)
# define XXH_likely(x) __builtin_expect(x, 1)
# define XXH_unlikely(x) __builtin_expect(x, 0)
#else
# define XXH_likely(x) (x)
# define XXH_unlikely(x) (x)
#endif
#if defined(__GNUC__)
# if defined(__AVX2__)
# include <immintrin.h>
# elif defined(__SSE2__)
# include <emmintrin.h>
# elif defined(__ARM_NEON__) || defined(__ARM_NEON)
# define inline __inline__ /* clang bug */
# include <arm_neon.h>
# undef inline
# endif
#elif defined(_MSC_VER)
# include <intrin.h>
#endif
/*
* One goal of XXH3 is to make it fast on both 32-bit and 64-bit, while
* remaining a true 64-bit/128-bit hash function.
*
* This is done by prioritizing a subset of 64-bit operations that can be
* emulated without too many steps on the average 32-bit machine.
*
* For example, these two lines seem similar, and run equally fast on 64-bit:
*
* xxh_u64 x;
* x ^= (x >> 47); // good
* x ^= (x >> 13); // bad
*
* However, to a 32-bit machine, there is a major difference.
*
* x ^= (x >> 47) looks like this:
*
* x.lo ^= (x.hi >> (47 - 32));
*
* while x ^= (x >> 13) looks like this:
*
* // note: funnel shifts are not usually cheap.
* x.lo ^= (x.lo >> 13) | (x.hi << (32 - 13));
* x.hi ^= (x.hi >> 13);
*
* The first one is significantly faster than the second, simply because the
* shift is larger than 32. This means:
* - All the bits we need are in the upper 32 bits, so we can ignore the lower
* 32 bits in the shift.
* - The shift result will always fit in the lower 32 bits, and therefore,
* we can ignore the upper 32 bits in the xor.
*
* Thanks to this optimization, XXH3 only requires these features to be efficient:
*
* - Usable unaligned access
* - A 32-bit or 64-bit ALU
* - If 32-bit, a decent ADC instruction
* - A 32 or 64-bit multiply with a 64-bit result
* - For the 128-bit variant, a decent byteswap helps short inputs.
*
* The first two are already required by XXH32, and almost all 32-bit and 64-bit
* platforms which can run XXH32 can run XXH3 efficiently.
*
* Thumb-1, the classic 16-bit only subset of ARM's instruction set, is one
* notable exception.
*
* First of all, Thumb-1 lacks support for the UMULL instruction which
* performs the important long multiply. This means numerous __aeabi_lmul
* calls.
*
* Second of all, the 8 functional registers are just not enough.
* Setup for __aeabi_lmul, byteshift loads, pointers, and all arithmetic need
* Lo registers, and this shuffling results in thousands more MOVs than A32.
*
* A32 and T32 don't have this limitation. They can access all 14 registers,
* do a 32->64 multiply with UMULL, and the flexible operand allowing free
* shifts is helpful, too.
*
* Therefore, we do a quick sanity check.
*
* If compiling Thumb-1 for a target which supports ARM instructions, we will
* emit a warning, as it is not a "sane" platform to compile for.
*
* Usually, if this happens, it is because of an accident and you probably need
* to specify -march, as you likely meant to compile for a newer architecture.
*/
#if defined(__thumb__) && !defined(__thumb2__) && defined(__ARM_ARCH_ISA_ARM)
# warning "XXH3 is highly inefficient without ARM or Thumb-2."
#endif
/* ==========================================
* Vectorization detection
* ========================================== */
#define XXH_SCALAR 0 /* Portable scalar version */
#define XXH_SSE2 1 /* SSE2 for Pentium 4 and all x86_64 */
#define XXH_AVX2 2 /* AVX2 for Haswell and Bulldozer */
#define XXH_NEON 3 /* NEON for most ARMv7-A and all AArch64 */
#define XXH_VSX 4 /* VSX and ZVector for POWER8/z13 */
#ifndef XXH_VECTOR /* can be defined on command line */
# if defined(__AVX2__)
# define XXH_VECTOR XXH_AVX2
# elif defined(__SSE2__) || defined(_M_AMD64) || defined(_M_X64) || (defined(_M_IX86_FP) && (_M_IX86_FP == 2))
# define XXH_VECTOR XXH_SSE2
# elif defined(__GNUC__) /* msvc support maybe later */ \
&& (defined(__ARM_NEON__) || defined(__ARM_NEON)) \
&& (defined(__LITTLE_ENDIAN__) /* We only support little endian NEON */ \
|| (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__))
# define XXH_VECTOR XXH_NEON
# elif (defined(__PPC64__) && defined(__POWER8_VECTOR__)) \
|| (defined(__s390x__) && defined(__VEC__)) \
&& defined(__GNUC__) /* TODO: IBM XL */
# define XXH_VECTOR XXH_VSX
# else
# define XXH_VECTOR XXH_SCALAR
# endif
#endif
/*
* Controls the alignment of the accumulator.
* This is for compatibility with aligned vector loads, which are usually faster.
*/
#ifndef XXH_ACC_ALIGN
# if XXH_VECTOR == XXH_SCALAR /* scalar */
# define XXH_ACC_ALIGN 8
# elif XXH_VECTOR == XXH_SSE2 /* sse2 */
# define XXH_ACC_ALIGN 16
# elif XXH_VECTOR == XXH_AVX2 /* avx2 */
# define XXH_ACC_ALIGN 32
# elif XXH_VECTOR == XXH_NEON /* neon */
# define XXH_ACC_ALIGN 16
# elif XXH_VECTOR == XXH_VSX /* vsx */
# define XXH_ACC_ALIGN 16
# endif
#endif
/*
* UGLY HACK:
* GCC usually generates the best code with -O3 for xxHash.
*
* However, when targeting AVX2, it is overzealous in its unrolling resulting
* in code roughly 3/4 the speed of Clang.
*
* There are other issues, such as GCC splitting _mm256_loadu_si256 into
* _mm_loadu_si128 + _mm256_inserti128_si256. This is an optimization which
* only applies to Sandy and Ivy Bridge... which don't even support AVX2.
*
* That is why when compiling the AVX2 version, it is recommended to use either
* -O2 -mavx2 -march=haswell
* or
* -O2 -mavx2 -mno-avx256-split-unaligned-load
* for decent performance, or to use Clang instead.
*
* Fortunately, we can control the first one with a pragma that forces GCC into
* -O2, but the other one we can't control without "failed to inline always
* inline function due to target mismatch" warnings.
*/
#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
&& defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
&& defined(__OPTIMIZE__) && !defined(__OPTIMIZE_SIZE__) /* respect -O0 and -Os */
# pragma GCC push_options
# pragma GCC optimize("-O2")
#endif
#if XXH_VECTOR == XXH_NEON
/*
* NEON's setup for vmlal_u32 is a little more complicated than it is on
* SSE2, AVX2, and VSX.
*
* While PMULUDQ and VMULEUW both perform a mask, VMLAL.U32 performs an upcast.
*
* To do the same operation, the 128-bit 'Q' register needs to be split into
* two 64-bit 'D' registers, performing this operation::
*
* [ a | b ]
* | '---------. .--------' |
* | x |
* | .---------' '--------. |
* [ a & 0xFFFFFFFF | b & 0xFFFFFFFF ],[ a >> 32 | b >> 32 ]
*
* Due to significant changes in aarch64, the fastest method for aarch64 is
* completely different than the fastest method for ARMv7-A.
*
* ARMv7-A treats D registers as unions overlaying Q registers, so modifying
* D11 will modify the high half of Q5. This is similar to how modifying AH
* will only affect bits 8-15 of AX on x86.
*
* VZIP takes two registers, and puts even lanes in one register and odd lanes
* in the other.
*
* On ARMv7-A, this strangely modifies both parameters in place instead of
* taking the usual 3-operand form.
*
* Therefore, if we want to do this, we can simply use a D-form VZIP.32 on the
* lower and upper halves of the Q register to end up with the high and low
* halves where we want - all in one instruction.
*
* vzip.32 d10, d11 @ d10 = { d10[0], d11[0] }; d11 = { d10[1], d11[1] }
*
* Unfortunately we need inline assembly for this: Instructions modifying two
* registers at once is not possible in GCC or Clang's IR, and they have to
* create a copy.
*
* aarch64 requires a different approach.
*
* In order to make it easier to write a decent compiler for aarch64, many
* quirks were removed, such as conditional execution.
*
* NEON was also affected by this.
*
* aarch64 cannot access the high bits of a Q-form register, and writes to a
* D-form register zero the high bits, similar to how writes to W-form scalar
* registers (or DWORD registers on x86_64) work.
*
* The formerly free vget_high intrinsics now require a vext (with a few
* exceptions)
*
* Additionally, VZIP was replaced by ZIP1 and ZIP2, which are the equivalent
* of PUNPCKL* and PUNPCKH* in SSE, respectively, in order to only modify one
* operand.
*
* The equivalent of the VZIP.32 on the lower and upper halves would be this
* mess:
*
* ext v2.4s, v0.4s, v0.4s, #2 // v2 = { v0[2], v0[3], v0[0], v0[1] }
* zip1 v1.2s, v0.2s, v2.2s // v1 = { v0[0], v2[0] }
* zip2 v0.2s, v0.2s, v1.2s // v0 = { v0[1], v2[1] }
*
* Instead, we use a literal downcast, vmovn_u64 (XTN), and vshrn_n_u64 (SHRN):
*
* shrn v1.2s, v0.2d, #32 // v1 = (uint32x2_t)(v0 >> 32);
* xtn v0.2s, v0.2d // v0 = (uint32x2_t)(v0 & 0xFFFFFFFF);
*
* This is available on ARMv7-A, but is less efficient than a single VZIP.32.
*/
/*
* Function-like macro:
* void XXH_SPLIT_IN_PLACE(uint64x2_t &in, uint32x2_t &outLo, uint32x2_t &outHi)
* {
* outLo = (uint32x2_t)(in & 0xFFFFFFFF);
* outHi = (uint32x2_t)(in >> 32);
* in = UNDEFINED;
* }
*/
# if !defined(XXH_NO_VZIP_HACK) /* define to disable */ \
&& defined(__GNUC__) \
&& !defined(__aarch64__) && !defined(__arm64__)
# define XXH_SPLIT_IN_PLACE(in, outLo, outHi) \
do { \
/* Undocumented GCC/Clang operand modifier: %e0 = lower D half, %f0 = upper D half */ \
/* https://github.com/gcc-mirror/gcc/blob/38cf91e5/gcc/config/arm/arm.c#L22486 */ \
/* https://github.com/llvm-mirror/llvm/blob/2c4ca683/lib/Target/ARM/ARMAsmPrinter.cpp#L399 */ \
__asm__("vzip.32 %e0, %f0" : "+w" (in)); \
(outLo) = vget_low_u32 (vreinterpretq_u32_u64(in)); \
(outHi) = vget_high_u32(vreinterpretq_u32_u64(in)); \
} while (0)
# else
# define XXH_SPLIT_IN_PLACE(in, outLo, outHi) \
do { \
(outLo) = vmovn_u64 (in); \
(outHi) = vshrn_n_u64 ((in), 32); \
} while (0)
# endif
#endif /* XXH_VECTOR == XXH_NEON */
/*
* VSX and Z Vector helpers.
*
* This is very messy, and any pull requests to clean this up are welcome.
*
* There are a lot of problems with supporting VSX and s390x, due to
* inconsistent intrinsics, spotty coverage, and multiple endiannesses.
*/
#if XXH_VECTOR == XXH_VSX
# if defined(__s390x__)
# include <s390intrin.h>
# else
# include <altivec.h>
# endif
# undef vector /* Undo the pollution */
typedef __vector unsigned long long xxh_u64x2;
typedef __vector unsigned char xxh_u8x16;
typedef __vector unsigned xxh_u32x4;
# ifndef XXH_VSX_BE
# if defined(__BIG_ENDIAN__) \
|| (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
# define XXH_VSX_BE 1
# elif defined(__VEC_ELEMENT_REG_ORDER__) && __VEC_ELEMENT_REG_ORDER__ == __ORDER_BIG_ENDIAN__
# warning "-maltivec=be is not recommended. Please use native endianness."
# define XXH_VSX_BE 1
# else
# define XXH_VSX_BE 0
# endif
# endif /* !defined(XXH_VSX_BE) */
# if XXH_VSX_BE
/* A wrapper for POWER9's vec_revb. */
# if defined(__POWER9_VECTOR__) || (defined(__clang__) && defined(__s390x__))
# define XXH_vec_revb vec_revb
# else
XXH_FORCE_INLINE xxh_u64x2 XXH_vec_revb(xxh_u64x2 val)
{
xxh_u8x16 const vByteSwap = { 0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00,
0x0F, 0x0E, 0x0D, 0x0C, 0x0B, 0x0A, 0x09, 0x08 };
return vec_perm(val, val, vByteSwap);
}
# endif
# endif /* XXH_VSX_BE */
/*
* Performs an unaligned load and byte swaps it on big endian.
*/
XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr)
{
xxh_u64x2 ret;
memcpy(&ret, ptr, sizeof(xxh_u64x2));
# if XXH_VSX_BE
ret = XXH_vec_revb(ret);
# endif
return ret;
}
/*
* vec_mulo and vec_mule are very problematic intrinsics on PowerPC
*
* These intrinsics weren't added until GCC 8, despite existing for a while,
* and they are endian dependent. Also, their meaning swap depending on version.
* */
# if defined(__s390x__)
/* s390x is always big endian, no issue on this platform */
# define XXH_vec_mulo vec_mulo
# define XXH_vec_mule vec_mule
# elif defined(__clang__) && __has_builtin(__builtin_altivec_vmuleuw)
/* Clang has a better way to control this, we can just use the builtin which doesn't swap. */
# define XXH_vec_mulo __builtin_altivec_vmulouw
# define XXH_vec_mule __builtin_altivec_vmuleuw
# else
/* gcc needs inline assembly */
/* Adapted from https://github.com/google/highwayhash/blob/master/highwayhash/hh_vsx.h. */
XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mulo(xxh_u32x4 a, xxh_u32x4 b)
{
xxh_u64x2 result;
__asm__("vmulouw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
return result;
}
XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b)
{
xxh_u64x2 result;
__asm__("vmuleuw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
return result;
}
# endif /* XXH_vec_mulo, XXH_vec_mule */
#endif /* XXH_VECTOR == XXH_VSX */
/* prefetch
* can be disabled, by declaring XXH_NO_PREFETCH build macro */
#if defined(XXH_NO_PREFETCH)
# define XXH_PREFETCH(ptr) (void)(ptr) /* disabled */
#else
# if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_I86)) /* _mm_prefetch() is not defined outside of x86/x64 */
# include <mmintrin.h> /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
# define XXH_PREFETCH(ptr) _mm_prefetch((const char*)(ptr), _MM_HINT_T0)
# elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) )
# define XXH_PREFETCH(ptr) __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */)
# else
# define XXH_PREFETCH(ptr) (void)(ptr) /* disabled */
# endif
#endif /* XXH_NO_PREFETCH */
/* ==========================================
* XXH3 default settings
* ========================================== */
#define XXH_SECRET_DEFAULT_SIZE 192 /* minimum XXH3_SECRET_SIZE_MIN */
#if (XXH_SECRET_DEFAULT_SIZE < XXH3_SECRET_SIZE_MIN)
# error "default keyset is not large enough"
#endif
/* Pseudorandom secret taken directly from FARSH */
XXH_ALIGN(64) static const xxh_u8 kSecret[XXH_SECRET_DEFAULT_SIZE] = {
0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, 0xf7, 0x21, 0xad, 0x1c,
0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f,
0xcb, 0x79, 0xe6, 0x4e, 0xcc, 0xc0, 0xe5, 0x78, 0x82, 0x5a, 0xd0, 0x7d, 0xcc, 0xff, 0x72, 0x21,
0xb8, 0x08, 0x46, 0x74, 0xf7, 0x43, 0x24, 0x8e, 0xe0, 0x35, 0x90, 0xe6, 0x81, 0x3a, 0x26, 0x4c,
0x3c, 0x28, 0x52, 0xbb, 0x91, 0xc3, 0x00, 0xcb, 0x88, 0xd0, 0x65, 0x8b, 0x1b, 0x53, 0x2e, 0xa3,
0x71, 0x64, 0x48, 0x97, 0xa2, 0x0d, 0xf9, 0x4e, 0x38, 0x19, 0xef, 0x46, 0xa9, 0xde, 0xac, 0xd8,
0xa8, 0xfa, 0x76, 0x3f, 0xe3, 0x9c, 0x34, 0x3f, 0xf9, 0xdc, 0xbb, 0xc7, 0xc7, 0x0b, 0x4f, 0x1d,
0x8a, 0x51, 0xe0, 0x4b, 0xcd, 0xb4, 0x59, 0x31, 0xc8, 0x9f, 0x7e, 0xc9, 0xd9, 0x78, 0x73, 0x64,
0xea, 0xc5, 0xac, 0x83, 0x34, 0xd3, 0xeb, 0xc3, 0xc5, 0x81, 0xa0, 0xff, 0xfa, 0x13, 0x63, 0xeb,
0x17, 0x0d, 0xdd, 0x51, 0xb7, 0xf0, 0xda, 0x49, 0xd3, 0x16, 0x55, 0x26, 0x29, 0xd4, 0x68, 0x9e,
0x2b, 0x16, 0xbe, 0x58, 0x7d, 0x47, 0xa1, 0xfc, 0x8f, 0xf8, 0xb8, 0xd1, 0x7a, 0xd0, 0x31, 0xce,
0x45, 0xcb, 0x3a, 0x8f, 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, 0xbb, 0x4b, 0x40, 0x7e,
};
/*
* Calculates a 32-bit to 64-bit long multiply.
*
* Wraps __emulu on MSVC x86 because it tends to call __allmul when it doesn't
* need to (but it shouldn't need to anyways, it is about 7 instructions to do
* a 64x64 multiply...). Since we know that this will _always_ emit MULL, we
* use that instead of the normal method.
*
* If you are compiling for platforms like Thumb-1 and don't have a better option,
* you may also want to write your own long multiply routine here.
*
* XXH_FORCE_INLINE xxh_u64 XXH_mult32to64(xxh_u64 x, xxh_u64 y)
* {
* return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF);
* }
*/
#if defined(_MSC_VER) && defined(_M_IX86)
# include <intrin.h>
# define XXH_mult32to64(x, y) __emulu((unsigned)(x), (unsigned)(y))
#else
/*
* Downcast + upcast is usually better than masking on older compilers like
* GCC 4.2 (especially 32-bit ones), all without affecting newer compilers.
*
* The other method, (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF), will AND both operands
* and perform a full 64x64 multiply -- entirely redundant on 32-bit.
*/
# define XXH_mult32to64(x, y) ((xxh_u64)(xxh_u32)(x) * (xxh_u64)(xxh_u32)(y))
#endif
/*
* Calculates a 64->128-bit long multiply.
*
* Uses __uint128_t and _umul128 if available, otherwise uses a scalar version.
*/
static XXH128_hash_t
XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs)
{
/*
* GCC/Clang __uint128_t method.
*
* On most 64-bit targets, GCC and Clang define a __uint128_t type.
* This is usually the best way as it usually uses a native long 64-bit
* multiply, such as MULQ on x86_64 or MUL + UMULH on aarch64.
*
* Usually.
*
* Despite being a 32-bit platform, Clang (and emscripten) define this type
* despite not having the arithmetic for it. This results in a laggy
* compiler builtin call which calculates a full 128-bit multiply.
* In that case it is best to use the portable one.
* https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677
*/
#if defined(__GNUC__) && !defined(__wasm__) \
&& defined(__SIZEOF_INT128__) \
|| (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128)
__uint128_t const product = (__uint128_t)lhs * (__uint128_t)rhs;
XXH128_hash_t r128;
r128.low64 = (xxh_u64)(product);
r128.high64 = (xxh_u64)(product >> 64);
return r128;
/*
* MSVC for x64's _umul128 method.
*
* xxh_u64 _umul128(xxh_u64 Multiplier, xxh_u64 Multiplicand, xxh_u64 *HighProduct);
*
* This compiles to single operand MUL on x64.
*/
#elif defined(_M_X64) || defined(_M_IA64)
#ifndef _MSC_VER
# pragma intrinsic(_umul128)
#endif
xxh_u64 product_high;
xxh_u64 const product_low = _umul128(lhs, rhs, &product_high);
XXH128_hash_t r128;
r128.low64 = product_low;
r128.high64 = product_high;
return r128;
#else
/*
* Portable scalar method. Optimized for 32-bit and 64-bit ALUs.
*
* This is a fast and simple grade school multiply, which is shown below
* with base 10 arithmetic instead of base 0x100000000.
*
* 9 3 // D2 lhs = 93
* x 7 5 // D2 rhs = 75
* ----------
* 1 5 // D2 lo_lo = (93 % 10) * (75 % 10) = 15
* 4 5 | // D2 hi_lo = (93 / 10) * (75 % 10) = 45
* 2 1 | // D2 lo_hi = (93 % 10) * (75 / 10) = 21
* + 6 3 | | // D2 hi_hi = (93 / 10) * (75 / 10) = 63
* ---------
* 2 7 | // D2 cross = (15 / 10) + (45 % 10) + 21 = 27
* + 6 7 | | // D2 upper = (27 / 10) + (45 / 10) + 63 = 67
* ---------
* 6 9 7 5 // D4 res = (27 * 10) + (15 % 10) + (67 * 100) = 6975
*
* The reasons for adding the products like this are:
* 1. It avoids manual carry tracking. Just like how
* (9 * 9) + 9 + 9 = 99, the same applies with this for UINT64_MAX.
* This avoids a lot of complexity.
*
* 2. It hints for, and on Clang, compiles to, the powerful UMAAL
* instruction available in ARM's Digital Signal Processing extension
* in 32-bit ARMv6 and later, which is shown below:
*
* void UMAAL(xxh_u32 *RdLo, xxh_u32 *RdHi, xxh_u32 Rn, xxh_u32 Rm)
* {
* xxh_u64 product = (xxh_u64)*RdLo * (xxh_u64)*RdHi + Rn + Rm;
* *RdLo = (xxh_u32)(product & 0xFFFFFFFF);
* *RdHi = (xxh_u32)(product >> 32);
* }
*
* This instruction was designed for efficient long multiplication, and
* allows this to be calculated in only 4 instructions at speeds
* comparable to some 64-bit ALUs.
*
* 3. It isn't terrible on other platforms. Usually this will be a couple
* of 32-bit ADD/ADCs.
*/
/* First calculate all of the cross products. */
xxh_u64 const lo_lo = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs & 0xFFFFFFFF);
xxh_u64 const hi_lo = XXH_mult32to64(lhs >> 32, rhs & 0xFFFFFFFF);
xxh_u64 const lo_hi = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs >> 32);
xxh_u64 const hi_hi = XXH_mult32to64(lhs >> 32, rhs >> 32);
/* Now add the products together. These will never overflow. */
xxh_u64 const cross = (lo_lo >> 32) + (hi_lo & 0xFFFFFFFF) + lo_hi;
xxh_u64 const upper = (hi_lo >> 32) + (cross >> 32) + hi_hi;
xxh_u64 const lower = (cross << 32) | (lo_lo & 0xFFFFFFFF);
XXH128_hash_t r128;
r128.low64 = lower;
r128.high64 = upper;
return r128;
#endif
}
/*
* Does a 64-bit to 128-bit multiply, then XOR folds it.
*
* The reason for the separate function is to prevent passing too many structs
* around by value. This will hopefully inline the multiply, but we don't force it.
*/
static xxh_u64
XXH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs)
{
XXH128_hash_t product = XXH_mult64to128(lhs, rhs);
return product.low64 ^ product.high64;
}
/* Seems to produce slightly better code on GCC for some reason. */
XXH_FORCE_INLINE xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift)
{
XXH_ASSERT(0 <= shift && shift < 64);
return v64 ^ (v64 >> shift);
}
/*
* We don't need to (or want to) mix as much as XXH64.
*
* Short hashes are more evenly distributed, so it isn't necessary.
*/
static XXH64_hash_t XXH3_avalanche(xxh_u64 h64)
{
h64 = XXH_xorshift64(h64, 37);
h64 *= 0x165667919E3779F9ULL;
h64 = XXH_xorshift64(h64, 32);
return h64;
}
/* ==========================================
* Short keys
* ==========================================
* One of the shortcomings of XXH32 and XXH64 was that their performance was
* sub-optimal on short lengths. It used an iterative algorithm which strongly
* favored lengths that were a multiple of 4 or 8.
*
* Instead of iterating over individual inputs, we use a set of single shot
* functions which piece together a range of lengths and operate in constant time.
*
* Additionally, the number of multiplies has been significantly reduced. This
* reduces latency, especially when emulating 64-bit multiplies on 32-bit.
*
* Depending on the platform, this may or may not be faster than XXH32, but it
* is almost guaranteed to be faster than XXH64.
*/
/*
* At very short lengths, there isn't enough input to fully hide secrets, or use
* the entire secret.
*
* There is also only a limited amount of mixing we can do before significantly
* impacting performance.
*
* Therefore, we use different sections of the secret and always mix two secret
* samples with an XOR. This should have no effect on performance on the
* seedless or withSeed variants because everything _should_ be constant folded
* by modern compilers.
*
* The XOR mixing hides individual parts of the secret and increases entropy.
*
* This adds an extra layer of strength for custom secrets.
*/
XXH_FORCE_INLINE XXH64_hash_t
XXH3_len_1to3_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
{
XXH_ASSERT(input != NULL);
XXH_ASSERT(1 <= len && len <= 3);
XXH_ASSERT(secret != NULL);
/*
* len = 1: combined = { input[0], 0x01, input[0], input[0] }
* len = 2: combined = { input[1], 0x02, input[0], input[1] }
* len = 3: combined = { input[2], 0x03, input[0], input[1] }
*/
{ xxh_u8 const c1 = input[0];
xxh_u8 const c2 = input[len >> 1];
xxh_u8 const c3 = input[len - 1];
xxh_u32 const combined = ((xxh_u32)c1 << 16) | ((xxh_u32)c2 << 24)
| ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8);
xxh_u64 const bitflip = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
xxh_u64 const keyed = (xxh_u64)combined ^ bitflip;
xxh_u64 const mixed = keyed * PRIME64_1;
return XXH3_avalanche(mixed);
}
}
XXH_FORCE_INLINE XXH64_hash_t
XXH3_len_4to8_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
{
XXH_ASSERT(input != NULL);
XXH_ASSERT(secret != NULL);
XXH_ASSERT(4 <= len && len < 8);
seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32;
{ xxh_u32 const input1 = XXH_readLE32(input);
xxh_u32 const input2 = XXH_readLE32(input + len - 4);
xxh_u64 const bitflip = (XXH_readLE64(secret+8) ^ XXH_readLE64(secret+16)) - seed;
xxh_u64 const input64 = input2 + (((xxh_u64)input1) << 32);
xxh_u64 x = input64 ^ bitflip;
/* this mix is inspired by Pelle Evensen's rrmxmx */
x ^= XXH_rotl64(x, 49) ^ XXH_rotl64(x, 24);
x *= 0x9FB21C651E98DF25ULL;
x ^= (x >> 35) + len ;
x *= 0x9FB21C651E98DF25ULL;
return XXH_xorshift64(x, 28);
}
}
XXH_FORCE_INLINE XXH64_hash_t
XXH3_len_9to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
{
XXH_ASSERT(input != NULL);
XXH_ASSERT(secret != NULL);
XXH_ASSERT(8 <= len && len <= 16);
{ xxh_u64 const bitflip1 = (XXH_readLE64(secret+24) ^ XXH_readLE64(secret+32)) + seed;
xxh_u64 const bitflip2 = (XXH_readLE64(secret+40) ^ XXH_readLE64(secret+48)) - seed;
xxh_u64 const input_lo = XXH_readLE64(input) ^ bitflip1;
xxh_u64 const input_hi = XXH_readLE64(input + len - 8) ^ bitflip2;
xxh_u64 const acc = len
+ XXH_swap64(input_lo) + input_hi
+ XXH3_mul128_fold64(input_lo, input_hi);
return XXH3_avalanche(acc);
}
}
XXH_FORCE_INLINE XXH64_hash_t
XXH3_len_0to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
{
XXH_ASSERT(len <= 16);
{ if (XXH_likely(len > 8)) return XXH3_len_9to16_64b(input, len, secret, seed);
if (XXH_likely(len >= 4)) return XXH3_len_4to8_64b(input, len, secret, seed);
if (len) return XXH3_len_1to3_64b(input, len, secret, seed);
return XXH3_avalanche((PRIME64_1 + seed) ^ (XXH_readLE64(secret+56) ^ XXH_readLE64(secret+64)));
}
}
/*
* DISCLAIMER: There are known *seed-dependent* multicollisions here due to
* multiplication by zero, affecting hashes of lengths 17 to 240.
*
* However, they are very unlikely.
*
* Keep this in mind when using the unseeded XXH3_64bits() variant: As with all
* unseeded non-cryptographic hashes, it does not attempt to defend itself
* against specially crafted inputs, only random inputs.
*
* Compared to classic UMAC where a 1 in 2^31 chance of 4 consecutive bytes
* cancelling out the secret is taken an arbitrary number of times (addressed
* in XXH3_accumulate_512), this collision is very unlikely with random inputs
* and/or proper seeding:
*
* This only has a 1 in 2^63 chance of 8 consecutive bytes cancelling out, in a
* function that is only called up to 16 times per hash with up to 240 bytes of
* input.
*
* This is not too bad for a non-cryptographic hash function, especially with
* only 64 bit outputs.
*
* The 128-bit variant (which trades some speed for strength) is NOT affected
* by this, although it is always a good idea to use a proper seed if you care
* about strength.
*/
XXH_FORCE_INLINE xxh_u64 XXH3_mix16B(const xxh_u8* XXH_RESTRICT input,
const xxh_u8* XXH_RESTRICT secret, xxh_u64 seed64)
{
#if defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
&& defined(__i386__) && defined(__SSE2__) /* x86 + SSE2 */ \
&& !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable like XXH32 hack */
/*
* UGLY HACK:
* GCC for x86 tends to autovectorize the 128-bit multiply, resulting in
* slower code.
*
* By forcing seed64 into a register, we disrupt the cost model and
* cause it to scalarize. See `XXH32_round()`
*
* FIXME: Clang's output is still _much_ faster -- On an AMD Ryzen 3600,
* XXH3_64bits @ len=240 runs at 4.6 GB/s with Clang 9, but 3.3 GB/s on
* GCC 9.2, despite both emitting scalar code.
*
* GCC generates much better scalar code than Clang for the rest of XXH3,
* which is why finding a more optimal codepath is an interest.
*/
__asm__ ("" : "+r" (seed64));
#endif
{ xxh_u64 const input_lo = XXH_readLE64(input);
xxh_u64 const input_hi = XXH_readLE64(input+8);
return XXH3_mul128_fold64(
input_lo ^ (XXH_readLE64(secret) + seed64),
input_hi ^ (XXH_readLE64(secret+8) - seed64)
);
}
}
/* For mid range keys, XXH3 uses a Mum-hash variant. */
XXH_FORCE_INLINE XXH64_hash_t
XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
XXH64_hash_t seed)
{
XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
XXH_ASSERT(16 < len && len <= 128);
{ xxh_u64 acc = len * PRIME64_1;
if (len > 32) {
if (len > 64) {
if (len > 96) {
acc += XXH3_mix16B(input+48, secret+96, seed);
acc += XXH3_mix16B(input+len-64, secret+112, seed);
}
acc += XXH3_mix16B(input+32, secret+64, seed);
acc += XXH3_mix16B(input+len-48, secret+80, seed);
}
acc += XXH3_mix16B(input+16, secret+32, seed);
acc += XXH3_mix16B(input+len-32, secret+48, seed);
}
acc += XXH3_mix16B(input+0, secret+0, seed);
acc += XXH3_mix16B(input+len-16, secret+16, seed);
return XXH3_avalanche(acc);
}
}
#define XXH3_MIDSIZE_MAX 240
XXH_NO_INLINE XXH64_hash_t
XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
XXH64_hash_t seed)
{
XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
#define XXH3_MIDSIZE_STARTOFFSET 3
#define XXH3_MIDSIZE_LASTOFFSET 17
{ xxh_u64 acc = len * PRIME64_1;
int const nbRounds = (int)len / 16;
int i;
for (i=0; i<8; i++) {
acc += XXH3_mix16B(input+(16*i), secret+(16*i), seed);
}
acc = XXH3_avalanche(acc);
XXH_ASSERT(nbRounds >= 8);
#if defined(__clang__) /* Clang */ \
&& (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \
&& !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable */
/*
* UGLY HACK:
* Clang for ARMv7-A tries to vectorize this loop, similar to GCC x86.
* In everywhere else, it uses scalar code.
*
* For 64->128-bit multiplies, even if the NEON was 100% optimal, it
* would still be slower than UMAAL (see XXH_mult64to128).
*
* Unfortunately, Clang doesn't handle the long multiplies properly and
* converts them to the nonexistent "vmulq_u64" intrinsic, which is then
* scalarized into an ugly mess of VMOV.32 instructions.
*
* This mess is difficult to avoid without turning autovectorization
* off completely, but they are usually relatively minor and/or not
* worth it to fix.
*
* This loop is the easiest to fix, as unlike XXH32, this pragma
* _actually works_ because it is a loop vectorization instead of an
* SLP vectorization.
*/
#pragma clang loop vectorize(disable)
#endif
for (i=8 ; i < nbRounds; i++) {
acc += XXH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXH3_MIDSIZE_STARTOFFSET, seed);
}
/* last bytes */
acc += XXH3_mix16B(input + len - 16, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, seed);
return XXH3_avalanche(acc);
}
}
/* === Long Keys === */
#define STRIPE_LEN 64
#define XXH_SECRET_CONSUME_RATE 8 /* nb of secret bytes consumed at each accumulation */
#define ACC_NB (STRIPE_LEN / sizeof(xxh_u64))
typedef enum { XXH3_acc_64bits, XXH3_acc_128bits } XXH3_accWidth_e;
/*
* XXH3_accumulate_512 is the tightest loop for long inputs, and it is the most optimized.
*
* It is a hardened version of UMAC, based off of FARSH's implementation.
*
* This was chosen because it adapts quite well to 32-bit, 64-bit, and SIMD
* implementations, and it is ridiculously fast.
*
* We harden it by mixing the original input to the accumulators as well as the product.
*
* This means that in the (relatively likely) case of a multiply by zero, the
* original input is preserved.
*
* On 128-bit inputs, we swap 64-bit pairs when we add the input to improve
* cross-pollination, as otherwise the upper and lower halves would be
* essentially independent.
*
* This doesn't matter on 64-bit hashes since they all get merged together in
* the end, so we skip the extra step.
*
* Both XXH3_64bits and XXH3_128bits use this subroutine.
*/
XXH_FORCE_INLINE void
XXH3_accumulate_512( void* XXH_RESTRICT acc,
const void* XXH_RESTRICT input,
const void* XXH_RESTRICT secret,
XXH3_accWidth_e accWidth)
{
#if (XXH_VECTOR == XXH_AVX2)
XXH_ASSERT((((size_t)acc) & 31) == 0);
{ XXH_ALIGN(32) __m256i* const xacc = (__m256i *) acc;
/* Unaligned. This is mainly for pointer arithmetic, and because
* _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
const __m256i* const xinput = (const __m256i *) input;
/* Unaligned. This is mainly for pointer arithmetic, and because
* _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
const __m256i* const xsecret = (const __m256i *) secret;
size_t i;
for (i=0; i < STRIPE_LEN/sizeof(__m256i); i++) {
/* data_vec = xinput[i]; */
__m256i const data_vec = _mm256_loadu_si256 (xinput+i);
/* key_vec = xsecret[i]; */
__m256i const key_vec = _mm256_loadu_si256 (xsecret+i);
/* data_key = data_vec ^ key_vec; */
__m256i const data_key = _mm256_xor_si256 (data_vec, key_vec);
/* data_key_lo = data_key >> 32; */
__m256i const data_key_lo = _mm256_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
/* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
__m256i const product = _mm256_mul_epu32 (data_key, data_key_lo);
if (accWidth == XXH3_acc_128bits) {
/* xacc[i] += swap(data_vec); */
__m256i const data_swap = _mm256_shuffle_epi32(data_vec, _MM_SHUFFLE(1, 0, 3, 2));
__m256i const sum = _mm256_add_epi64(xacc[i], data_swap);
/* xacc[i] += product; */
xacc[i] = _mm256_add_epi64(product, sum);
} else { /* XXH3_acc_64bits */
/* xacc[i] += data_vec; */
__m256i const sum = _mm256_add_epi64(xacc[i], data_vec);
/* xacc[i] += product; */
xacc[i] = _mm256_add_epi64(product, sum);
}
} }
#elif (XXH_VECTOR == XXH_SSE2)
/* SSE2 is just a half-scale version of the AVX2 version. */
XXH_ASSERT((((size_t)acc) & 15) == 0);
{ XXH_ALIGN(16) __m128i* const xacc = (__m128i *) acc;
/* Unaligned. This is mainly for pointer arithmetic, and because
* _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
const __m128i* const xinput = (const __m128i *) input;
/* Unaligned. This is mainly for pointer arithmetic, and because
* _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
const __m128i* const xsecret = (const __m128i *) secret;
size_t i;
for (i=0; i < STRIPE_LEN/sizeof(__m128i); i++) {
/* data_vec = xinput[i]; */
__m128i const data_vec = _mm_loadu_si128 (xinput+i);
/* key_vec = xsecret[i]; */
__m128i const key_vec = _mm_loadu_si128 (xsecret+i);
/* data_key = data_vec ^ key_vec; */
__m128i const data_key = _mm_xor_si128 (data_vec, key_vec);
/* data_key_lo = data_key >> 32; */
__m128i const data_key_lo = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
/* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
__m128i const product = _mm_mul_epu32 (data_key, data_key_lo);
if (accWidth == XXH3_acc_128bits) {
/* xacc[i] += swap(data_vec); */
__m128i const data_swap = _mm_shuffle_epi32(data_vec, _MM_SHUFFLE(1,0,3,2));
__m128i const sum = _mm_add_epi64(xacc[i], data_swap);
/* xacc[i] += product; */
xacc[i] = _mm_add_epi64(product, sum);
} else { /* XXH3_acc_64bits */