braid58 0.1.0

Fast fixed-width Bitcoin Base58 for 32- and 64-byte values
Documentation
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
/*
 * AVX2 Bitcoin Base58 encoder for 32-byte inputs.
 *
 * AVX2 has four 64-bit lanes per YMM and no packed 64-bit multiply. The
 * matrix therefore uses radix 58^5:
 *
 *   encoder: ten base-2^26 chunks -> nine base-58^5 cells.
 *
 * Since 58^5 is below 2^30, every matrix coefficient fits in 32 bits and
 * VPMULUDQ produces complete 64-bit products. Raw columns remain below 2^58.
 *
 * ISA: Haswell-class x86-64. Backend selection occurs at compile time.
 */

#include "braid58_internal.h"

#include <immintrin.h>
#include <stddef.h>
#include <stdint.h>
#include <string.h>

#define B58_B5 UINT64_C(656356768)
#define B58_B3 UINT32_C(195112)
#define B58_B2 UINT32_C(3364)
#define B58_CHUNK26_MASK UINT64_C(0x3ffffff)

/*
 * Row i is the little-endian radix-58^5 expansion of 2^(26*i).
 * Row zero and row one are folded into the scalar low-52-bit seed.
 */
_Alignas(32) static const uint64_t B58_W5_ENCODE[10][8] = {
    {UINT64_C(1), 0, 0, 0, 0, 0, 0, 0},
    {UINT64_C(67108864), 0, 0, 0, 0, 0, 0, 0},
    {UINT64_C(443814048), UINT64_C(6861511), 0, 0, 0, 0, 0, 0},
    {UINT64_C(437973984), UINT64_C(506963877), UINT64_C(701551), 0,
     0, 0, 0, 0},
    {UINT64_C(192414240), UINT64_C(584944171), UINT64_C(527870455),
     UINT64_C(71729), 0, 0, 0, 0},
    {UINT64_C(85356032), UINT64_C(237320048), UINT64_C(230466711),
     UINT64_C(641497958), UINT64_C(7333), 0, 0, 0},
    {UINT64_C(59086336), UINT64_C(425716400), UINT64_C(469620603),
     UINT64_C(369434287), UINT64_C(563670112), UINT64_C(749), 0, 0},
    {UINT64_C(249488768), UINT64_C(563819044), UINT64_C(543108756),
     UINT64_C(589143927), UINT64_C(28959436), UINT64_C(439056932),
     UINT64_C(76), 0},
    {UINT64_C(147129216), UINT64_C(329522580), UINT64_C(449746271),
     UINT64_C(218521078), UINT64_C(590921969), UINT64_C(502289454),
     UINT64_C(550667440), UINT64_C(7)},
    {UINT64_C(576074656), UINT64_C(432829647), UINT64_C(553198108),
     UINT64_C(378465102), UINT64_C(185459504), UINT64_C(598754100),
     UINT64_C(313589673), UINT64_C(526064760)},
};

/* Expand a four-bit scalar carry mask into four byte values 0/1. */
_Alignas(64) static const uint32_t B58_EXPAND4[16] = {
    0x00000000,0x00000001,0x00000100,0x00000101,
    0x00010000,0x00010001,0x00010100,0x00010101,
    0x01000000,0x01000001,0x01000100,0x01000101,
    0x01010000,0x01010001,0x01010100,0x01010101,
};

/* Four 16-entry alphabet quarters, duplicated for both AVX2 half-lanes. */
_Alignas(32) static const uint8_t B58_MAP0[] =
    "123456789ABCDEFG123456789ABCDEFG";
_Alignas(32) static const uint8_t B58_MAP1[] =
    "HJKLMNPQRSTUVWXYHJKLMNPQRSTUVWXY";
_Alignas(32) static const uint8_t B58_MAP2[] =
    "ZabcdefghijkmnopZabcdefghijkmnop";
_Alignas(32) static const uint8_t B58_MAP3[] =
    "qrstuvwxyz\0\0\0\0\0\0qrstuvwxyz\0\0\0\0\0\0";

/* Cell-field assembly masks.  The split produces fields in four-cell
 * half-lanes; these shuffles emit cells 7..0 in five-digit order. */
_Alignas(32) static const uint8_t B58_FIELDS_A[32] = {
     4, 5, 6, 7,12,13,14,15, 0x80,0x80,0x80,0x80, 0,1,2,3,
     4, 5, 6, 7,12,13,14,15, 0x80,0x80,0x80,0x80, 0,1,2,3,
};
_Alignas(32) static const uint8_t B58_FIELDS_B[32] = {
    0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80, 4,5,6,7,
    0x80,0x80,0x80,0x80, 0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,
    4,5,6,7, 0x80,0x80,0x80,0x80,
};
_Alignas(32) static const uint8_t B58_HEAD_A[32] = {
    0x80,15,11,7,3, 0x80,14,10,6,2, 0x80,13,9,5,1, 0x80,
    0x80,15,11,7,3, 0x80,14,10,6,2, 0x80,13,9,5,1, 0x80,
};
_Alignas(32) static const uint8_t B58_HEAD_B[32] = {
    3,0x80,0x80,0x80,0x80, 2,0x80,0x80,0x80,0x80,
    1,0x80,0x80,0x80,0x80, 0,
    3,0x80,0x80,0x80,0x80, 2,0x80,0x80,0x80,0x80,
    1,0x80,0x80,0x80,0x80, 0,
};
_Alignas(32) static const uint8_t B58_TAIL_A[32] = {
    12,8,4,0, 0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,
    0x80,0x80,0x80,0x80,
    12,8,4,0, 0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,
    0x80,0x80,0x80,0x80,
};

static inline uint64_t
b58_load_be64(const void *src) {
  uint64_t value;
  memcpy(&value, src, sizeof(value));
  return __builtin_bswap64(value);
}

/* Exact floor(x/58) for every uint32_t x.  Spell out the reciprocal so
 * target-specific compilers do not choose a scalar IDIV when both quotient
 * and remainder are consumed by the top-cell splitter. */
static inline uint32_t
b58_div58_u32(uint32_t x) {
  return (uint32_t)(((uint64_t)x * UINT64_C(0x8d3dcb09)) >> 37);
}

/* Exact floor(x/B58_B5), remainder, for every encoder raw-column bound.
 * M=ceil(2^61/B).  Since x<2^58, the estimate has error below 0.063 and is
 * either exact or one high. */
static inline void
b58_div_b5_8(__m256i xlo, __m256i xhi,
             __m256i *qlo_out, __m256i *qhi_out,
             __m256i *rlo_out, __m256i *rhi_out) {
  const __m256i magic = _mm256_set1_epi64x((long long)UINT64_C(3513093979));
  const __m256i divisor = _mm256_set1_epi64x((long long)B58_B5);
  const __m256i lo0 = _mm256_mul_epu32(xlo, magic);
  const __m256i lo1 = _mm256_mul_epu32(_mm256_srli_epi64(xlo, 32), magic);
  const __m256i hi0 = _mm256_mul_epu32(xhi, magic);
  const __m256i hi1 = _mm256_mul_epu32(_mm256_srli_epi64(xhi, 32), magic);
  __m256i qlo = _mm256_srli_epi64(
      _mm256_add_epi64(lo1, _mm256_srli_epi64(lo0, 32)), 29);
  __m256i qhi = _mm256_srli_epi64(
      _mm256_add_epi64(hi1, _mm256_srli_epi64(hi0, 32)), 29);
  __m256i plo = _mm256_mul_epu32(qlo, divisor);
  __m256i phi = _mm256_mul_epu32(qhi, divisor);
  const __m256i hlo = _mm256_cmpgt_epi64(plo, xlo);
  const __m256i hhi = _mm256_cmpgt_epi64(phi, xhi);
  const __m256i any = _mm256_or_si256(hlo, hhi);
  if (__builtin_expect(!_mm256_testz_si256(any, any), 0)) {
    qlo = _mm256_add_epi64(qlo, hlo); /* masks are -1 in corrected lanes */
    qhi = _mm256_add_epi64(qhi, hhi);
    plo = _mm256_sub_epi64(plo, _mm256_and_si256(hlo, divisor));
    phi = _mm256_sub_epi64(phi, _mm256_and_si256(hhi, divisor));
  }
  *qlo_out = qlo;
  *qhi_out = qhi;
  *rlo_out = _mm256_sub_epi64(xlo, plo);
  *rhi_out = _mm256_sub_epi64(xhi, phi);
}

static inline __m256i
b58_div195112_u32(__m256i value) {
  const __m256i magic = _mm256_set1_epi32(721316415);
  const __m256i even = _mm256_mul_epu32(value, magic);
  const __m256i odd =
      _mm256_mul_epu32(_mm256_srli_epi64(value, 32), magic);
  const __m256i high = _mm256_blend_epi32(
      _mm256_shuffle_epi32(even, 0xf5), odd, 0xaa);
  return _mm256_srli_epi32(high, 15);
}

typedef struct {
  uint64_t packed;
  uint32_t first;
} b58_top_digits;

static inline b58_top_digits
b58_split_top_simd(uint32_t top_cell) {
  const uint32_t top2 = top_cell / B58_B2;
  const uint32_t rem2 = top_cell - top2 * B58_B2;
  const __m128i pair = _mm_cvtsi32_si128((int)(top2 | (rem2 << 16)));
  const __m128i quotient = _mm_mulhi_epu16(pair, _mm_set1_epi16(1130));
  const __m128i remainder = _mm_sub_epi16(
      pair, _mm_mullo_epi16(quotient, _mm_set1_epi16(58)));
  const __m128i interleaved = _mm_unpacklo_epi16(quotient, remainder);
  const __m128i bytes = _mm_packus_epi16(interleaved, _mm_setzero_si128());
  const b58_top_digits result = {
      (uint64_t)(uint32_t)_mm_cvtsi128_si32(bytes) << 8,
      (uint32_t)_mm_cvtsi128_si32(quotient) & 0xffffU};
  return result;
}

static inline __m256i
b58_map58(__m256i digit) {
  const __m256i index =
      _mm256_and_si256(digit, _mm256_set1_epi8(15));
  const __m256i group_bit4 = _mm256_slli_epi16(digit, 3);
  const __m256i group_bit5 = _mm256_slli_epi16(digit, 2);
  const __m256i low_half = _mm256_blendv_epi8(
      _mm256_shuffle_epi8(
          _mm256_load_si256((const __m256i *)(const void *)B58_MAP0), index),
      _mm256_shuffle_epi8(
          _mm256_load_si256((const __m256i *)(const void *)B58_MAP1), index),
      group_bit4);
  const __m256i high_half = _mm256_blendv_epi8(
      _mm256_shuffle_epi8(
          _mm256_load_si256((const __m256i *)(const void *)B58_MAP2), index),
      _mm256_shuffle_epi8(
          _mm256_load_si256((const __m256i *)(const void *)B58_MAP3), index),
      group_bit4);
  return _mm256_blendv_epi8(low_half, high_half, group_bit5);
}

typedef struct {
  __m256i lo;
  __m256i hi;
} b58_raw_columns;

/* Branch-free input rechunk and radix-matrix stage. */
static inline b58_raw_columns
b58_make_raw_columns(const uint8_t input[32]) {
  const uint64_t word0 = b58_load_be64(input + 24);
  const uint64_t word1 = b58_load_be64(input + 16);
  const uint64_t word2 = b58_load_be64(input + 8);
  const uint64_t word3 = b58_load_be64(input + 0);
  const uint64_t a2 =
      ((word0 >> 52) | (word1 << 12)) & B58_CHUNK26_MASK;
  const uint64_t a3 = (word1 >> 14) & B58_CHUNK26_MASK;
  const uint64_t a4 =
      ((word1 >> 40) | (word2 << 24)) & B58_CHUNK26_MASK;
  const uint64_t a5 = (word2 >> 2) & B58_CHUNK26_MASK;
  const uint64_t a6 = (word2 >> 28) & B58_CHUNK26_MASK;
  const uint64_t a7 =
      ((word2 >> 54) | (word3 << 10)) & B58_CHUNK26_MASK;
  const uint64_t a8 = (word3 >> 16) & B58_CHUNK26_MASK;
  const uint64_t a9 = word3 >> 42;

  __m256i column_lo = _mm256_setr_epi64x(
      (long long)(word0 & ((UINT64_C(1) << 52) - 1)), 0, 0, 0);
  __m256i column_hi = _mm256_setzero_si256();
#define B58_ADD_LO(row, value)                                                \
  column_lo = _mm256_add_epi64(                                              \
      column_lo, _mm256_mul_epu32(                                           \
          _mm256_set1_epi64x((long long)(value)),                            \
          _mm256_load_si256(                                                 \
              (const __m256i *)(const void *)&B58_W5_ENCODE[(row)][0])))
#define B58_ADD_HI(row, value)                                                \
  column_hi = _mm256_add_epi64(                                              \
      column_hi, _mm256_mul_epu32(                                           \
          _mm256_set1_epi64x((long long)(value)),                            \
          _mm256_load_si256(                                                 \
              (const __m256i *)(const void *)&B58_W5_ENCODE[(row)][4])))
  B58_ADD_LO(2, a2);
  B58_ADD_LO(3, a3);
  B58_ADD_LO(4, a4);
  B58_ADD_LO(5, a5); B58_ADD_HI(5, a5);
  B58_ADD_LO(6, a6); B58_ADD_HI(6, a6);
  B58_ADD_LO(7, a7); B58_ADD_HI(7, a7);
  B58_ADD_LO(8, a8); B58_ADD_HI(8, a8);
  B58_ADD_LO(9, a9); B58_ADD_HI(9, a9);
#undef B58_ADD_LO
#undef B58_ADD_HI

  const b58_raw_columns result = {column_lo, column_hi};
  return result;
}

/* Normalize raw columns into cells 0..7; return cell 8 through top_cell. */
static inline __m256i
b58_finish_cells(b58_raw_columns raw, uint32_t *top_cell) {
  const __m256i column_lo = raw.lo;
  const __m256i column_hi = raw.hi;

  __m256i qlo, qhi, rlo, rhi;
  b58_div_b5_8(column_lo, column_hi, &qlo, &qhi, &rlo, &rhi);

  const __m256i zero = _mm256_setzero_si256();
  __m256i qprev_lo = _mm256_permute4x64_epi64(qlo, 0x90);
  qprev_lo = _mm256_blend_epi32(zero, qprev_lo, 0xfc);
  const __m256i cross = _mm256_permute2x128_si256(qlo, qhi, 0x21);
  const __m256i qprev_hi = _mm256_alignr_epi8(qhi, cross, 8);
  __m256i sum_lo = _mm256_add_epi64(rlo, qprev_lo);
  __m256i sum_hi = _mm256_add_epi64(rhi, qprev_hi);

  const __m256i propagate = _mm256_set1_epi64x((long long)(B58_B5 - 1));
  const __m256i initial_generate_lo =
      _mm256_cmpgt_epi64(sum_lo, propagate);
  const __m256i initial_generate_hi =
      _mm256_cmpgt_epi64(sum_hi, propagate);
  unsigned g = (unsigned)_mm256_movemask_pd(_mm256_castsi256_pd(
      initial_generate_lo));
  g |= (unsigned)_mm256_movemask_pd(_mm256_castsi256_pd(
           initial_generate_hi)) << 4;
  const __m256i propagate_lo = _mm256_cmpeq_epi64(sum_lo, propagate);
  const __m256i propagate_hi = _mm256_cmpeq_epi64(sum_hi, propagate);
  const __m256i any_propagate =
      _mm256_or_si256(propagate_lo, propagate_hi);
  const unsigned had_propagate = (unsigned)!_mm256_testz_si256(
      any_propagate, any_propagate);
  if (__builtin_expect(had_propagate != 0U, 0)) {
    unsigned p = (unsigned)_mm256_movemask_pd(
        _mm256_castsi256_pd(propagate_lo));
    p |= (unsigned)_mm256_movemask_pd(
             _mm256_castsi256_pd(propagate_hi)) << 4;
    g |= p & (g << 1); p &= p << 1;
    g |= p & (g << 2); p &= p << 2;
    g |= p & (g << 4);
  }
  g &= 0xffU;
  const unsigned carry_in = (g << 1) & 0xffU;
  const __m256i carry_lo = _mm256_cvtepu8_epi64(
      _mm_cvtsi32_si128((int)B58_EXPAND4[carry_in & 15U]));
  const __m256i carry_hi = _mm256_cvtepu8_epi64(
      _mm_cvtsi32_si128((int)B58_EXPAND4[carry_in >> 4]));
  sum_lo = _mm256_add_epi64(sum_lo, carry_lo);
  sum_hi = _mm256_add_epi64(sum_hi, carry_hi);
  const __m256i divisor = _mm256_set1_epi64x((long long)B58_B5);
  __m256i generate_lo = initial_generate_lo;
  __m256i generate_hi = initial_generate_hi;
  if (__builtin_expect(had_propagate != 0U, 0)) {
    generate_lo = _mm256_cmpgt_epi64(sum_lo, propagate);
    generate_hi = _mm256_cmpgt_epi64(sum_hi, propagate);
  }
  sum_lo = _mm256_sub_epi64(
      sum_lo, _mm256_and_si256(generate_lo, divisor));
  sum_hi = _mm256_sub_epi64(
      sum_hi, _mm256_and_si256(generate_hi, divisor));

  *top_cell = (uint32_t)(uint64_t)_mm256_extract_epi64(qhi, 3)
            + ((g >> 7) & 1U);
  const __m256i index = _mm256_setr_epi32(0,2,4,6,0,0,0,0);
  const __m256i low_dwords = _mm256_permutevar8x32_epi32(sum_lo, index);
  const __m256i high_dwords = _mm256_permutevar8x32_epi32(sum_hi, index);
  return _mm256_permute2x128_si256(low_dwords, high_dwords, 0x20);
}

/* Emit a cell vector for the common case where the most-significant input
 * byte is nonzero.  In that domain the canonical skip is exactly 1 or 2. */
static inline size_t
b58_emit_common(__m256i cell, uint32_t top_cell, char output[45],
                int simd_top) {
  const __m256i zero = _mm256_setzero_si256();
  const __m256i top_pair = b58_div195112_u32(cell);
  const __m256i bottom3 = _mm256_sub_epi32(
      cell, _mm256_mullo_epi32(top_pair, _mm256_set1_epi32((int)B58_B3)));
  const __m256i quarter = _mm256_srli_epi32(bottom3, 2);
  const __m256i pair = _mm256_packus_epi32(top_pair, quarter);
  const __m256i first_magic = _mm256_setr_epi16(
      1130,1130,1130,1130,(short)39899,(short)39899,(short)39899,(short)39899,
      1130,1130,1130,1130,(short)39899,(short)39899,(short)39899,(short)39899);
  const __m256i first_high = _mm256_mulhi_epu16(pair, first_magic);
  const __m256i high = _mm256_blend_epi16(
      first_high, _mm256_srli_epi16(first_high, 9), 0xf0);
  const __m256i first_coefficient = _mm256_setr_epi16(
      58,58,58,58,841,841,841,841,58,58,58,58,841,841,841,841);
  const __m256i first_remainder = _mm256_sub_epi16(
      pair, _mm256_mullo_epi16(high, first_coefficient));
  const __m256i scale = _mm256_setr_epi16(
      1,1,1,1,4,4,4,4,1,1,1,1,4,4,4,4);
  __m256i pair58 = _mm256_mullo_epi16(first_remainder, scale);
  const __m256i low2 =
      _mm256_and_si256(bottom3, _mm256_set1_epi32(3));
  pair58 = _mm256_add_epi16(
      pair58, _mm256_packus_epi32(zero, low2));
  const __m256i middle =
      _mm256_mulhi_epu16(pair58, _mm256_set1_epi16(1130));
  const __m256i low = _mm256_sub_epi16(
      pair58, _mm256_mullo_epi16(middle, _mm256_set1_epi16(58)));

  const __m256i byte_fields = _mm256_packus_epi16(low, middle);
  const __m256i top_field = _mm256_packus_epi16(high, zero);
  const __m256i fields = _mm256_or_si256(
      _mm256_shuffle_epi8(
          byte_fields,
          _mm256_load_si256((const __m256i *)(const void *)B58_FIELDS_A)),
      _mm256_shuffle_epi8(
          top_field,
          _mm256_load_si256((const __m256i *)(const void *)B58_FIELDS_B)));
  const __m256i head = _mm256_or_si256(
      _mm256_shuffle_epi8(
          fields,
          _mm256_load_si256((const __m256i *)(const void *)B58_HEAD_A)),
      _mm256_shuffle_epi8(
          top_field,
          _mm256_load_si256((const __m256i *)(const void *)B58_HEAD_B)));
  const __m256i tail = _mm256_shuffle_epi8(
      fields, _mm256_load_si256((const __m256i *)(const void *)B58_TAIL_A));

  const __m128i head_lo = _mm256_castsi256_si128(head);
  const __m128i head_hi = _mm256_extracti128_si256(head, 1);
  const __m128i tail_lo = _mm256_castsi256_si128(tail);
  const __m128i tail_hi = _mm256_extracti128_si256(tail, 1);
  uint32_t top0;
  uint64_t top_digits;
  if (simd_top) {
    const b58_top_digits top = b58_split_top_simd(top_cell);
    top0 = top.first;
    top_digits = top.packed;
  } else {
    const uint32_t top2 = top_cell / B58_B2;
    const uint32_t top_remainder = top_cell - top2 * B58_B2;
    top0 = top2 / 58U;
    const uint32_t top1 = top2 - top0 * 58U;
    const uint32_t top3 = top_remainder / 58U;
    const uint32_t top4 = top_remainder - top3 * 58U;
    top_digits = ((uint64_t)top0 << 8)
               | ((uint64_t)top1 << 16)
               | ((uint64_t)top3 << 24)
               | ((uint64_t)top4 << 32);
  }
  const __m128i digit0_lo = _mm_or_si128(
      _mm_cvtsi64_si128((long long)top_digits), _mm_slli_si128(head_hi, 5));
  const __m128i digit0_hi = _mm_or_si128(
      _mm_srli_si128(head_hi, 11),
      _mm_or_si128(_mm_slli_si128(tail_hi, 5),
                   _mm_slli_si128(head_lo, 9)));
  const __m128i digit1_lo = _mm_or_si128(
      _mm_srli_si128(head_lo, 7), _mm_slli_si128(tail_lo, 9));
  const __m256i digit0 = _mm256_inserti128_si256(
      _mm256_castsi128_si256(digit0_lo), digit0_hi, 1);
  const __m256i digit1 = _mm256_inserti128_si256(
      _mm256_castsi128_si256(digit1_lo), _mm_setzero_si128(), 1);
  const __m256i ascii0 = b58_map58(digit0);
  const __m256i ascii1 = b58_map58(digit1);
  const __m128i ascii0_lo = _mm256_castsi256_si128(ascii0);
  const __m128i ascii0_hi = _mm256_extracti128_si256(ascii0, 1);
  const __m128i ascii1_lo = _mm256_castsi256_si128(ascii1);
  __m128i output_lo;
  __m128i output_hi;
  unsigned skip;
  if (__builtin_expect(top0 != 0U, 1)) {
    output_lo = _mm_alignr_epi8(ascii0_hi, ascii0_lo, 1);
    output_hi = _mm_alignr_epi8(ascii1_lo, ascii0_hi, 1);
    skip = 1U;
  } else {
    output_lo = _mm_alignr_epi8(ascii0_hi, ascii0_lo, 2);
    output_hi = _mm_alignr_epi8(ascii1_lo, ascii0_hi, 2);
    skip = 2U;
  }
  const __m256i output_vector = _mm256_inserti128_si256(
      _mm256_castsi128_si256(output_lo), output_hi, 1);
  const __m128i output_tail =
      _mm_alignr_epi8(ascii1_lo, ascii0_hi, 13);
  const size_t length = 45U - skip;
  _mm256_storeu_si256((__m256i *)(void *)output, output_vector);
  _mm_storeu_si128((__m128i *)(void *)(output + length - 16U), output_tail);
  output[length] = '\0';
  return length;
}

static __attribute__((noinline)) size_t
b58_emit_general(const uint8_t input[32], __m256i cell, uint32_t top_cell,
                 char output[45]) {
  _Alignas(32) uint8_t ascii[64];
  const __m256i zero = _mm256_setzero_si256();

  /* Eight B5 cells -> five byte fields each.  The r3/3364 operation is
   * performed as floor(floor(r3/4)/841), with the two low bits restored. */
  const __m256i top_pair = b58_div195112_u32(cell);
  const __m256i bottom3 = _mm256_sub_epi32(
      cell, _mm256_mullo_epi32(top_pair, _mm256_set1_epi32((int)B58_B3)));
  const __m256i quarter = _mm256_srli_epi32(bottom3, 2);
  const __m256i pair = _mm256_packus_epi32(top_pair, quarter);
  const __m256i first_magic = _mm256_setr_epi16(
      1130,1130,1130,1130,(short)39899,(short)39899,(short)39899,(short)39899,
      1130,1130,1130,1130,(short)39899,(short)39899,(short)39899,(short)39899);
  const __m256i first_high = _mm256_mulhi_epu16(pair, first_magic);
  const __m256i high = _mm256_blend_epi16(
      first_high, _mm256_srli_epi16(first_high, 9), 0xf0);
  const __m256i first_coefficient = _mm256_setr_epi16(
      58,58,58,58,841,841,841,841,58,58,58,58,841,841,841,841);
  const __m256i first_remainder = _mm256_sub_epi16(
      pair, _mm256_mullo_epi16(high, first_coefficient));
  const __m256i scale = _mm256_setr_epi16(
      1,1,1,1,4,4,4,4,1,1,1,1,4,4,4,4);
  __m256i pair58 = _mm256_mullo_epi16(first_remainder, scale);
  const __m256i low2 =
      _mm256_and_si256(bottom3, _mm256_set1_epi32(3));
  pair58 = _mm256_add_epi16(
      pair58, _mm256_packus_epi32(zero, low2));
  const __m256i middle =
      _mm256_mulhi_epu16(pair58, _mm256_set1_epi16(1130));
  const __m256i low = _mm256_sub_epi16(
      pair58, _mm256_mullo_epi16(middle, _mm256_set1_epi16(58)));

  const __m256i byte_fields = _mm256_packus_epi16(low, middle);
  const __m256i top_field = _mm256_packus_epi16(high, zero);
  const __m256i fields = _mm256_or_si256(
      _mm256_shuffle_epi8(
          byte_fields,
          _mm256_load_si256((const __m256i *)(const void *)B58_FIELDS_A)),
      _mm256_shuffle_epi8(
          top_field,
          _mm256_load_si256((const __m256i *)(const void *)B58_FIELDS_B)));
  const __m256i head = _mm256_or_si256(
      _mm256_shuffle_epi8(
          fields, _mm256_load_si256(
                      (const __m256i *)(const void *)B58_HEAD_A)),
      _mm256_shuffle_epi8(
          top_field, _mm256_load_si256(
                         (const __m256i *)(const void *)B58_HEAD_B)));
  const __m256i tail = _mm256_shuffle_epi8(
      fields, _mm256_load_si256((const __m256i *)(const void *)B58_TAIL_A));

  const __m128i head_lo = _mm256_castsi256_si128(head);
  const __m128i head_hi = _mm256_extracti128_si256(head, 1);
  const __m128i tail_lo = _mm256_castsi256_si128(tail);
  const __m128i tail_hi = _mm256_extracti128_si256(tail, 1);

  /* The ninth cell is below 58^4, so its first field is always zero. */
  const uint32_t top2 = top_cell / B58_B2;
  const uint32_t top_remainder = top_cell - top2 * B58_B2;
  const uint32_t top0 = b58_div58_u32(top2);
  const uint32_t top1 = top2 - top0 * 58U;
  const uint32_t top3 = b58_div58_u32(top_remainder);
  const uint32_t top4 = top_remainder - top3 * 58U;
  const uint64_t top_digits = ((uint64_t)top0 << 8)
                            | ((uint64_t)top1 << 16)
                            | ((uint64_t)top3 << 24)
                            | ((uint64_t)top4 << 32);
  const __m128i digit0_lo = _mm_or_si128(
      _mm_cvtsi64_si128((long long)top_digits), _mm_slli_si128(head_hi, 5));
  const __m128i digit0_hi = _mm_or_si128(
      _mm_srli_si128(head_hi, 11),
      _mm_or_si128(_mm_slli_si128(tail_hi, 5),
                   _mm_slli_si128(head_lo, 9)));
  const __m128i digit1_lo = _mm_or_si128(
      _mm_srli_si128(head_lo, 7), _mm_slli_si128(tail_lo, 9));
  const __m256i digit0 = _mm256_inserti128_si256(
      _mm256_castsi128_si256(digit0_lo), digit0_hi, 1);
  const __m256i digit1 = _mm256_inserti128_si256(
      _mm256_castsi128_si256(digit1_lo), _mm_setzero_si128(), 1);
  const __m256i ascii0 = b58_map58(digit0);
  const __m256i ascii1 = b58_map58(digit1);

  uint64_t nonzero_digits =
      (uint32_t)~(uint32_t)_mm256_movemask_epi8(
          _mm256_cmpeq_epi8(digit0, zero));
  nonzero_digits |= (uint64_t)(uint32_t)~(uint32_t)_mm256_movemask_epi8(
      _mm256_cmpeq_epi8(digit1, zero)) << 32;
  nonzero_digits &= (UINT64_C(1) << 45) - 1;
  const unsigned first_digit = nonzero_digits
      ? (unsigned)__builtin_ctzll(nonzero_digits) : 45U;

  _mm256_store_si256((__m256i *)(void *)ascii, ascii0);
  _mm256_store_si256((__m256i *)(void *)(ascii + 32), ascii1);

  const __m256i input_vector =
      _mm256_loadu_si256((const __m256i *)(const void *)input);
  const uint32_t nonzero_bytes =
      ~(uint32_t)_mm256_movemask_epi8(
          _mm256_cmpeq_epi8(input_vector, zero));
  const unsigned leading_zero_bytes = nonzero_bytes
      ? (unsigned)__builtin_ctz(nonzero_bytes) : 32U;

  const unsigned skip = first_digit - leading_zero_bytes;
  const size_t length = 45U - skip;
  _mm256_storeu_si256((__m256i *)(void *)output,
      _mm256_loadu_si256((const __m256i *)(const void *)(ascii + skip)));
  _mm_storeu_si128((__m128i *)(void *)(output + length - 16U),
      _mm_loadu_si128((const __m128i *)(const void *)(ascii + 29)));
  output[length] = '\0';
  return length;
}

void
braid58_encode_32x2_avx2(const uint8_t input[2][32],
                         char output[2][45], size_t output_len[2]) {
  const b58_raw_columns raw0 = b58_make_raw_columns(input[0]);
  const b58_raw_columns raw1 = b58_make_raw_columns(input[1]);
  uint32_t top0, top1;
  const __m256i cell0 = b58_finish_cells(raw0, &top0);
  const __m256i cell1 = b58_finish_cells(raw1, &top1);
  output_len[0] = input[0][0] != 0
      ? b58_emit_common(cell0, top0, output[0], 0)
      : b58_emit_general(input[0], cell0, top0, output[0]);
  output_len[1] = input[1][0] != 0
      ? b58_emit_common(cell1, top1, output[1], 0)
      : b58_emit_general(input[1], cell1, top1, output[1]);
}

size_t
braid58_encode_32_avx2(const uint8_t input[32], char output[45]) {
  const b58_raw_columns raw = b58_make_raw_columns(input);
  uint32_t top;
  const __m256i cell = b58_finish_cells(raw, &top);
  return input[0] != 0 ? b58_emit_common(cell, top, output, 1)
                       : b58_emit_general(input, cell, top, output);
}

void
braid58_encode_32x3_avx2(const uint8_t input[3][32],
                         char output[3][45], size_t output_len[3]) {
  const b58_raw_columns raw0 = b58_make_raw_columns(input[0]);
  const b58_raw_columns raw1 = b58_make_raw_columns(input[1]);
  const b58_raw_columns raw2 = b58_make_raw_columns(input[2]);
  uint32_t top0, top1, top2;
  const __m256i cell0 = b58_finish_cells(raw0, &top0);
  const __m256i cell1 = b58_finish_cells(raw1, &top1);
  const __m256i cell2 = b58_finish_cells(raw2, &top2);
  output_len[0] = input[0][0] != 0
      ? b58_emit_common(cell0, top0, output[0], 0)
      : b58_emit_general(input[0], cell0, top0, output[0]);
  output_len[1] = input[1][0] != 0
      ? b58_emit_common(cell1, top1, output[1], 0)
      : b58_emit_general(input[1], cell1, top1, output[1]);
  output_len[2] = input[2][0] != 0
      ? b58_emit_common(cell2, top2, output[2], 0)
      : b58_emit_general(input[2], cell2, top2, output[2]);
}