summaryrefslogtreecommitdiff
path: root/vendor/github.com/klauspost/compress/zstd/blockenc.go
blob: 507757d525a1a4ad912244f97c11247a22ae873f (plain)
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
// Copyright 2019+ Klaus Post. All rights reserved.
// License information can be found in the LICENSE file.
// Based on work by Yann Collet, released under BSD License.

package zstd

import (
	"errors"
	"fmt"
	"math"
	"math/bits"

	"github.com/klauspost/compress/huff0"
)

type blockEnc struct {
	size      int
	literals  []byte
	sequences []seq
	coders    seqCoders
	litEnc    *huff0.Scratch
	wr        bitWriter

	extraLits int
	last      bool

	output            []byte
	recentOffsets     [3]uint32
	prevRecentOffsets [3]uint32
}

// init should be used once the block has been created.
// If called more than once, the effect is the same as calling reset.
func (b *blockEnc) init() {
	if cap(b.literals) < maxCompressedLiteralSize {
		b.literals = make([]byte, 0, maxCompressedLiteralSize)
	}
	const defSeqs = 200
	b.literals = b.literals[:0]
	if cap(b.sequences) < defSeqs {
		b.sequences = make([]seq, 0, defSeqs)
	}
	if cap(b.output) < maxCompressedBlockSize {
		b.output = make([]byte, 0, maxCompressedBlockSize)
	}
	if b.coders.mlEnc == nil {
		b.coders.mlEnc = &fseEncoder{}
		b.coders.mlPrev = &fseEncoder{}
		b.coders.ofEnc = &fseEncoder{}
		b.coders.ofPrev = &fseEncoder{}
		b.coders.llEnc = &fseEncoder{}
		b.coders.llPrev = &fseEncoder{}
	}
	b.litEnc = &huff0.Scratch{WantLogLess: 4}
	b.reset(nil)
}

// initNewEncode can be used to reset offsets and encoders to the initial state.
func (b *blockEnc) initNewEncode() {
	b.recentOffsets = [3]uint32{1, 4, 8}
	b.litEnc.Reuse = huff0.ReusePolicyNone
	b.coders.setPrev(nil, nil, nil)
}

// reset will reset the block for a new encode, but in the same stream,
// meaning that state will be carried over, but the block content is reset.
// If a previous block is provided, the recent offsets are carried over.
func (b *blockEnc) reset(prev *blockEnc) {
	b.extraLits = 0
	b.literals = b.literals[:0]
	b.size = 0
	b.sequences = b.sequences[:0]
	b.output = b.output[:0]
	b.last = false
	if prev != nil {
		b.recentOffsets = prev.prevRecentOffsets
	}
}

// reset will reset the block for a new encode, but in the same stream,
// meaning that state will be carried over, but the block content is reset.
// If a previous block is provided, the recent offsets are carried over.
func (b *blockEnc) swapEncoders(prev *blockEnc) {
	b.coders.swap(&prev.coders)
	b.litEnc, prev.litEnc = prev.litEnc, b.litEnc
}

// blockHeader contains the information for a block header.
type blockHeader uint32

// setLast sets the 'last' indicator on a block.
func (h *blockHeader) setLast(b bool) {
	if b {
		*h = *h | 1
	} else {
		const mask = (1 << 24) - 2
		*h = *h & mask
	}
}

// setSize will store the compressed size of a block.
func (h *blockHeader) setSize(v uint32) {
	const mask = 7
	*h = (*h)&mask | blockHeader(v<<3)
}

// setType sets the block type.
func (h *blockHeader) setType(t blockType) {
	const mask = 1 | (((1 << 24) - 1) ^ 7)
	*h = (*h & mask) | blockHeader(t<<1)
}

// appendTo will append the block header to a slice.
func (h blockHeader) appendTo(b []byte) []byte {
	return append(b, uint8(h), uint8(h>>8), uint8(h>>16))
}

// String returns a string representation of the block.
func (h blockHeader) String() string {
	return fmt.Sprintf("Type: %d, Size: %d, Last:%t", (h>>1)&3, h>>3, h&1 == 1)
}

// literalsHeader contains literals header information.
type literalsHeader uint64

// setType can be used to set the type of literal block.
func (h *literalsHeader) setType(t literalsBlockType) {
	const mask = math.MaxUint64 - 3
	*h = (*h & mask) | literalsHeader(t)
}

// setSize can be used to set a single size, for uncompressed and RLE content.
func (h *literalsHeader) setSize(regenLen int) {
	inBits := bits.Len32(uint32(regenLen))
	// Only retain 2 bits
	const mask = 3
	lh := uint64(*h & mask)
	switch {
	case inBits < 5:
		lh |= (uint64(regenLen) << 3) | (1 << 60)
		if debug {
			got := int(lh>>3) & 0xff
			if got != regenLen {
				panic(fmt.Sprint("litRegenSize = ", regenLen, "(want) != ", got, "(got)"))
			}
		}
	case inBits < 12:
		lh |= (1 << 2) | (uint64(regenLen) << 4) | (2 << 60)
	case inBits < 20:
		lh |= (3 << 2) | (uint64(regenLen) << 4) | (3 << 60)
	default:
		panic(fmt.Errorf("internal error: block too big (%d)", regenLen))
	}
	*h = literalsHeader(lh)
}

// setSizes will set the size of a compressed literals section and the input length.
func (h *literalsHeader) setSizes(compLen, inLen int, single bool) {
	compBits, inBits := bits.Len32(uint32(compLen)), bits.Len32(uint32(inLen))
	// Only retain 2 bits
	const mask = 3
	lh := uint64(*h & mask)
	switch {
	case compBits <= 10 && inBits <= 10:
		if !single {
			lh |= 1 << 2
		}
		lh |= (uint64(inLen) << 4) | (uint64(compLen) << (10 + 4)) | (3 << 60)
		if debug {
			const mmask = (1 << 24) - 1
			n := (lh >> 4) & mmask
			if int(n&1023) != inLen {
				panic(fmt.Sprint("regensize:", int(n&1023), "!=", inLen, inBits))
			}
			if int(n>>10) != compLen {
				panic(fmt.Sprint("compsize:", int(n>>10), "!=", compLen, compBits))
			}
		}
	case compBits <= 14 && inBits <= 14:
		lh |= (2 << 2) | (uint64(inLen) << 4) | (uint64(compLen) << (14 + 4)) | (4 << 60)
		if single {
			panic("single stream used with more than 10 bits length.")
		}
	case compBits <= 18 && inBits <= 18:
		lh |= (3 << 2) | (uint64(inLen) << 4) | (uint64(compLen) << (18 + 4)) | (5 << 60)
		if single {
			panic("single stream used with more than 10 bits length.")
		}
	default:
		panic("internal error: block too big")
	}
	*h = literalsHeader(lh)
}

// appendTo will append the literals header to a byte slice.
func (h literalsHeader) appendTo(b []byte) []byte {
	size := uint8(h >> 60)
	switch size {
	case 1:
		b = append(b, uint8(h))
	case 2:
		b = append(b, uint8(h), uint8(h>>8))
	case 3:
		b = append(b, uint8(h), uint8(h>>8), uint8(h>>16))
	case 4:
		b = append(b, uint8(h), uint8(h>>8), uint8(h>>16), uint8(h>>24))
	case 5:
		b = append(b, uint8(h), uint8(h>>8), uint8(h>>16), uint8(h>>24), uint8(h>>32))
	default:
		panic(fmt.Errorf("internal error: literalsHeader has invalid size (%d)", size))
	}
	return b
}

// size returns the output size with currently set values.
func (h literalsHeader) size() int {
	return int(h >> 60)
}

func (h literalsHeader) String() string {
	return fmt.Sprintf("Type: %d, SizeFormat: %d, Size: 0x%d, Bytes:%d", literalsBlockType(h&3), (h>>2)&3, h&((1<<60)-1)>>4, h>>60)
}

// pushOffsets will push the recent offsets to the backup store.
func (b *blockEnc) pushOffsets() {
	b.prevRecentOffsets = b.recentOffsets
}

// pushOffsets will push the recent offsets to the backup store.
func (b *blockEnc) popOffsets() {
	b.recentOffsets = b.prevRecentOffsets
}

// matchOffset will adjust recent offsets and return the adjusted one,
// if it matches a previous offset.
func (b *blockEnc) matchOffset(offset, lits uint32) uint32 {
	// Check if offset is one of the recent offsets.
	// Adjusts the output offset accordingly.
	// Gives a tiny bit of compression, typically around 1%.
	if true {
		if lits > 0 {
			switch offset {
			case b.recentOffsets[0]:
				offset = 1
			case b.recentOffsets[1]:
				b.recentOffsets[1] = b.recentOffsets[0]
				b.recentOffsets[0] = offset
				offset = 2
			case b.recentOffsets[2]:
				b.recentOffsets[2] = b.recentOffsets[1]
				b.recentOffsets[1] = b.recentOffsets[0]
				b.recentOffsets[0] = offset
				offset = 3
			default:
				b.recentOffsets[2] = b.recentOffsets[1]
				b.recentOffsets[1] = b.recentOffsets[0]
				b.recentOffsets[0] = offset
				offset += 3
			}
		} else {
			switch offset {
			case b.recentOffsets[1]:
				b.recentOffsets[1] = b.recentOffsets[0]
				b.recentOffsets[0] = offset
				offset = 1
			case b.recentOffsets[2]:
				b.recentOffsets[2] = b.recentOffsets[1]
				b.recentOffsets[1] = b.recentOffsets[0]
				b.recentOffsets[0] = offset
				offset = 2
			case b.recentOffsets[0] - 1:
				b.recentOffsets[2] = b.recentOffsets[1]
				b.recentOffsets[1] = b.recentOffsets[0]
				b.recentOffsets[0] = offset
				offset = 3
			default:
				b.recentOffsets[2] = b.recentOffsets[1]
				b.recentOffsets[1] = b.recentOffsets[0]
				b.recentOffsets[0] = offset
				offset += 3
			}
		}
	} else {
		offset += 3
	}
	return offset
}

// encodeRaw can be used to set the output to a raw representation of supplied bytes.
func (b *blockEnc) encodeRaw(a []byte) {
	var bh blockHeader
	bh.setLast(b.last)
	bh.setSize(uint32(len(a)))
	bh.setType(blockTypeRaw)
	b.output = bh.appendTo(b.output[:0])
	b.output = append(b.output, a...)
	if debug {
		println("Adding RAW block, length", len(a))
	}
}

// encodeRaw can be used to set the output to a raw representation of supplied bytes.
func (b *blockEnc) encodeRawTo(dst, src []byte) []byte {
	var bh blockHeader
	bh.setLast(b.last)
	bh.setSize(uint32(len(src)))
	bh.setType(blockTypeRaw)
	dst = bh.appendTo(dst)
	dst = append(dst, src...)
	if debug {
		println("Adding RAW block, length", len(src))
	}
	return dst
}

// encodeLits can be used if the block is only litLen.
func (b *blockEnc) encodeLits(raw bool) error {
	var bh blockHeader
	bh.setLast(b.last)
	bh.setSize(uint32(len(b.literals)))

	// Don't compress extremely small blocks
	if len(b.literals) < 32 || raw {
		if debug {
			println("Adding RAW block, length", len(b.literals))
		}
		bh.setType(blockTypeRaw)
		b.output = bh.appendTo(b.output)
		b.output = append(b.output, b.literals...)
		return nil
	}

	var (
		out            []byte
		reUsed, single bool
		err            error
	)
	if len(b.literals) >= 1024 {
		// Use 4 Streams.
		out, reUsed, err = huff0.Compress4X(b.literals, b.litEnc)
	} else if len(b.literals) > 32 {
		// Use 1 stream
		single = true
		out, reUsed, err = huff0.Compress1X(b.literals, b.litEnc)
	} else {
		err = huff0.ErrIncompressible
	}

	switch err {
	case huff0.ErrIncompressible:
		if debug {
			println("Adding RAW block, length", len(b.literals))
		}
		bh.setType(blockTypeRaw)
		b.output = bh.appendTo(b.output)
		b.output = append(b.output, b.literals...)
		return nil
	case huff0.ErrUseRLE:
		if debug {
			println("Adding RLE block, length", len(b.literals))
		}
		bh.setType(blockTypeRLE)
		b.output = bh.appendTo(b.output)
		b.output = append(b.output, b.literals[0])
		return nil
	default:
		return err
	case nil:
	}
	// Compressed...
	// Now, allow reuse
	b.litEnc.Reuse = huff0.ReusePolicyAllow
	bh.setType(blockTypeCompressed)
	var lh literalsHeader
	if reUsed {
		if debug {
			println("Reused tree, compressed to", len(out))
		}
		lh.setType(literalsBlockTreeless)
	} else {
		if debug {
			println("New tree, compressed to", len(out), "tree size:", len(b.litEnc.OutTable))
		}
		lh.setType(literalsBlockCompressed)
	}
	// Set sizes
	lh.setSizes(len(out), len(b.literals), single)
	bh.setSize(uint32(len(out) + lh.size() + 1))

	// Write block headers.
	b.output = bh.appendTo(b.output)
	b.output = lh.appendTo(b.output)
	// Add compressed data.
	b.output = append(b.output, out...)
	// No sequences.
	b.output = append(b.output, 0)
	return nil
}

// fuzzFseEncoder can be used to fuzz the FSE encoder.
func fuzzFseEncoder(data []byte) int {
	if len(data) > maxSequences || len(data) < 2 {
		return 0
	}
	enc := fseEncoder{}
	hist := enc.Histogram()[:256]
	maxSym := uint8(0)
	for i, v := range data {
		v = v & 63
		data[i] = v
		hist[v]++
		if v > maxSym {
			maxSym = v
		}
	}
	if maxSym == 0 {
		// All 0
		return 0
	}
	maxCount := func(a []uint32) int {
		var max uint32
		for _, v := range a {
			if v > max {
				max = v
			}
		}
		return int(max)
	}
	cnt := maxCount(hist[:maxSym])
	if cnt == len(data) {
		// RLE
		return 0
	}
	enc.HistogramFinished(maxSym, cnt)
	err := enc.normalizeCount(len(data))
	if err != nil {
		return 0
	}
	_, err = enc.writeCount(nil)
	if err != nil {
		panic(err)
	}
	return 1
}

// encode will encode the block and append the output in b.output.
func (b *blockEnc) encode(raw bool) error {
	if len(b.sequences) == 0 {
		return b.encodeLits(raw)
	}
	// We want some difference
	if len(b.literals) > (b.size - (b.size >> 5)) {
		return errIncompressible
	}

	var bh blockHeader
	var lh literalsHeader
	bh.setLast(b.last)
	bh.setType(blockTypeCompressed)
	// Store offset of the block header. Needed when we know the size.
	bhOffset := len(b.output)
	b.output = bh.appendTo(b.output)

	var (
		out            []byte
		reUsed, single bool
		err            error
	)
	if len(b.literals) >= 1024 && !raw {
		// Use 4 Streams.
		out, reUsed, err = huff0.Compress4X(b.literals, b.litEnc)
	} else if len(b.literals) > 32 && !raw {
		// Use 1 stream
		single = true
		out, reUsed, err = huff0.Compress1X(b.literals, b.litEnc)
	} else {
		err = huff0.ErrIncompressible
	}

	switch err {
	case huff0.ErrIncompressible:
		lh.setType(literalsBlockRaw)
		lh.setSize(len(b.literals))
		b.output = lh.appendTo(b.output)
		b.output = append(b.output, b.literals...)
		if debug {
			println("Adding literals RAW, length", len(b.literals))
		}
	case huff0.ErrUseRLE:
		lh.setType(literalsBlockRLE)
		lh.setSize(len(b.literals))
		b.output = lh.appendTo(b.output)
		b.output = append(b.output, b.literals[0])
		if debug {
			println("Adding literals RLE")
		}
	default:
		if debug {
			println("Adding literals ERROR:", err)
		}
		return err
	case nil:
		// Compressed litLen...
		if reUsed {
			if debug {
				println("reused tree")
			}
			lh.setType(literalsBlockTreeless)
		} else {
			if debug {
				println("new tree, size:", len(b.litEnc.OutTable))
			}
			lh.setType(literalsBlockCompressed)
			if debug {
				_, _, err := huff0.ReadTable(out, nil)
				if err != nil {
					panic(err)
				}
			}
		}
		lh.setSizes(len(out), len(b.literals), single)
		if debug {
			printf("Compressed %d literals to %d bytes", len(b.literals), len(out))
			println("Adding literal header:", lh)
		}
		b.output = lh.appendTo(b.output)
		b.output = append(b.output, out...)
		b.litEnc.Reuse = huff0.ReusePolicyAllow
		if debug {
			println("Adding literals compressed")
		}
	}
	// Sequence compression

	// Write the number of sequences
	switch {
	case len(b.sequences) < 128:
		b.output = append(b.output, uint8(len(b.sequences)))
	case len(b.sequences) < 0x7f00: // TODO: this could be wrong
		n := len(b.sequences)
		b.output = append(b.output, 128+uint8(n>>8), uint8(n))
	default:
		n := len(b.sequences) - 0x7f00
		b.output = append(b.output, 255, uint8(n), uint8(n>>8))
	}
	if debug {
		println("Encoding", len(b.sequences), "sequences")
	}
	b.genCodes()
	llEnc := b.coders.llEnc
	ofEnc := b.coders.ofEnc
	mlEnc := b.coders.mlEnc
	err = llEnc.normalizeCount(len(b.sequences))
	if err != nil {
		return err
	}
	err = ofEnc.normalizeCount(len(b.sequences))
	if err != nil {
		return err
	}
	err = mlEnc.normalizeCount(len(b.sequences))
	if err != nil {
		return err
	}

	// Choose the best compression mode for each type.
	// Will evaluate the new vs predefined and previous.
	chooseComp := func(cur, prev, preDef *fseEncoder) (*fseEncoder, seqCompMode) {
		// See if predefined/previous is better
		hist := cur.count[:cur.symbolLen]
		nSize := cur.approxSize(hist) + cur.maxHeaderSize()
		predefSize := preDef.approxSize(hist)
		prevSize := prev.approxSize(hist)

		// Add a small penalty for new encoders.
		// Don't bother with extremely small (<2 byte gains).
		nSize = nSize + (nSize+2*8*16)>>4
		switch {
		case predefSize <= prevSize && predefSize <= nSize || forcePreDef:
			if debug {
				println("Using predefined", predefSize>>3, "<=", nSize>>3)
			}
			return preDef, compModePredefined
		case prevSize <= nSize:
			if debug {
				println("Using previous", prevSize>>3, "<=", nSize>>3)
			}
			return prev, compModeRepeat
		default:
			if debug {
				println("Using new, predef", predefSize>>3, ". previous:", prevSize>>3, ">", nSize>>3, "header max:", cur.maxHeaderSize()>>3, "bytes")
				println("tl:", cur.actualTableLog, "symbolLen:", cur.symbolLen, "norm:", cur.norm[:cur.symbolLen], "hist", cur.count[:cur.symbolLen])
			}
			return cur, compModeFSE
		}
	}

	// Write compression mode
	var mode uint8
	if llEnc.useRLE {
		mode |= uint8(compModeRLE) << 6
		llEnc.setRLE(b.sequences[0].llCode)
		if debug {
			println("llEnc.useRLE")
		}
	} else {
		var m seqCompMode
		llEnc, m = chooseComp(llEnc, b.coders.llPrev, &fsePredefEnc[tableLiteralLengths])
		mode |= uint8(m) << 6
	}
	if ofEnc.useRLE {
		mode |= uint8(compModeRLE) << 4
		ofEnc.setRLE(b.sequences[0].ofCode)
		if debug {
			println("ofEnc.useRLE")
		}
	} else {
		var m seqCompMode
		ofEnc, m = chooseComp(ofEnc, b.coders.ofPrev, &fsePredefEnc[tableOffsets])
		mode |= uint8(m) << 4
	}

	if mlEnc.useRLE {
		mode |= uint8(compModeRLE) << 2
		mlEnc.setRLE(b.sequences[0].mlCode)
		if debug {
			println("mlEnc.useRLE, code: ", b.sequences[0].mlCode, "value", b.sequences[0].matchLen)
		}
	} else {
		var m seqCompMode
		mlEnc, m = chooseComp(mlEnc, b.coders.mlPrev, &fsePredefEnc[tableMatchLengths])
		mode |= uint8(m) << 2
	}
	b.output = append(b.output, mode)
	if debug {
		printf("Compression modes: 0b%b", mode)
	}
	b.output, err = llEnc.writeCount(b.output)
	if err != nil {
		return err
	}
	start := len(b.output)
	b.output, err = ofEnc.writeCount(b.output)
	if err != nil {
		return err
	}
	if false {
		println("block:", b.output[start:], "tablelog", ofEnc.actualTableLog, "maxcount:", ofEnc.maxCount)
		fmt.Printf("selected TableLog: %d, Symbol length: %d\n", ofEnc.actualTableLog, ofEnc.symbolLen)
		for i, v := range ofEnc.norm[:ofEnc.symbolLen] {
			fmt.Printf("%3d: %5d -> %4d \n", i, ofEnc.count[i], v)
		}
	}
	b.output, err = mlEnc.writeCount(b.output)
	if err != nil {
		return err
	}

	// Maybe in block?
	wr := &b.wr
	wr.reset(b.output)

	var ll, of, ml cState

	// Current sequence
	seq := len(b.sequences) - 1
	s := b.sequences[seq]
	llEnc.setBits(llBitsTable[:])
	mlEnc.setBits(mlBitsTable[:])
	ofEnc.setBits(nil)

	llTT, ofTT, mlTT := llEnc.ct.symbolTT[:256], ofEnc.ct.symbolTT[:256], mlEnc.ct.symbolTT[:256]

	// We have 3 bounds checks here (and in the loop).
	// Since we are iterating backwards it is kinda hard to avoid.
	llB, ofB, mlB := llTT[s.llCode], ofTT[s.ofCode], mlTT[s.mlCode]
	ll.init(wr, &llEnc.ct, llB)
	of.init(wr, &ofEnc.ct, ofB)
	wr.flush32()
	ml.init(wr, &mlEnc.ct, mlB)

	// Each of these lookups also generates a bounds check.
	wr.addBits32NC(s.litLen, llB.outBits)
	wr.addBits32NC(s.matchLen, mlB.outBits)
	wr.flush32()
	wr.addBits32NC(s.offset, ofB.outBits)
	if debugSequences {
		println("Encoded seq", seq, s, "codes:", s.llCode, s.mlCode, s.ofCode, "states:", ll.state, ml.state, of.state, "bits:", llB, mlB, ofB)
	}
	seq--
	if llEnc.maxBits+mlEnc.maxBits+ofEnc.maxBits <= 32 {
		// No need to flush (common)
		for seq >= 0 {
			s = b.sequences[seq]
			wr.flush32()
			llB, ofB, mlB := llTT[s.llCode], ofTT[s.ofCode], mlTT[s.mlCode]
			// tabelog max is 8 for all.
			of.encode(ofB)
			ml.encode(mlB)
			ll.encode(llB)
			wr.flush32()

			// We checked that all can stay within 32 bits
			wr.addBits32NC(s.litLen, llB.outBits)
			wr.addBits32NC(s.matchLen, mlB.outBits)
			wr.addBits32NC(s.offset, ofB.outBits)

			if debugSequences {
				println("Encoded seq", seq, s)
			}

			seq--
		}
	} else {
		for seq >= 0 {
			s = b.sequences[seq]
			wr.flush32()
			llB, ofB, mlB := llTT[s.llCode], ofTT[s.ofCode], mlTT[s.mlCode]
			// tabelog max is below 8 for each.
			of.encode(ofB)
			ml.encode(mlB)
			ll.encode(llB)
			wr.flush32()

			// ml+ll = max 32 bits total
			wr.addBits32NC(s.litLen, llB.outBits)
			wr.addBits32NC(s.matchLen, mlB.outBits)
			wr.flush32()
			wr.addBits32NC(s.offset, ofB.outBits)

			if debugSequences {
				println("Encoded seq", seq, s)
			}

			seq--
		}
	}
	ml.flush(mlEnc.actualTableLog)
	of.flush(ofEnc.actualTableLog)
	ll.flush(llEnc.actualTableLog)
	err = wr.close()
	if err != nil {
		return err
	}
	b.output = wr.out

	if len(b.output)-3-bhOffset >= b.size {
		// Maybe even add a bigger margin.
		b.litEnc.Reuse = huff0.ReusePolicyNone
		return errIncompressible
	}

	// Size is output minus block header.
	bh.setSize(uint32(len(b.output)-bhOffset) - 3)
	if debug {
		println("Rewriting block header", bh)
	}
	_ = bh.appendTo(b.output[bhOffset:bhOffset])
	b.coders.setPrev(llEnc, mlEnc, ofEnc)
	return nil
}

var errIncompressible = errors.New("incompressible")

func (b *blockEnc) genCodes() {
	if len(b.sequences) == 0 {
		// nothing to do
		return
	}

	if len(b.sequences) > math.MaxUint16 {
		panic("can only encode up to 64K sequences")
	}
	// No bounds checks after here:
	llH := b.coders.llEnc.Histogram()[:256]
	ofH := b.coders.ofEnc.Histogram()[:256]
	mlH := b.coders.mlEnc.Histogram()[:256]
	for i := range llH {
		llH[i] = 0
	}
	for i := range ofH {
		ofH[i] = 0
	}
	for i := range mlH {
		mlH[i] = 0
	}

	var llMax, ofMax, mlMax uint8
	for i, seq := range b.sequences {
		v := llCode(seq.litLen)
		seq.llCode = v
		llH[v]++
		if v > llMax {
			llMax = v
		}

		v = ofCode(seq.offset)
		seq.ofCode = v
		ofH[v]++
		if v > ofMax {
			ofMax = v
		}

		v = mlCode(seq.matchLen)
		seq.mlCode = v
		mlH[v]++
		if v > mlMax {
			mlMax = v
			if debug && mlMax > maxMatchLengthSymbol {
				panic(fmt.Errorf("mlMax > maxMatchLengthSymbol (%d), matchlen: %d", mlMax, seq.matchLen))
			}
		}
		b.sequences[i] = seq
	}
	maxCount := func(a []uint32) int {
		var max uint32
		for _, v := range a {
			if v > max {
				max = v
			}
		}
		return int(max)
	}
	if mlMax > maxMatchLengthSymbol {
		panic(fmt.Errorf("mlMax > maxMatchLengthSymbol (%d)", mlMax))
	}
	if ofMax > maxOffsetBits {
		panic(fmt.Errorf("ofMax > maxOffsetBits (%d)", ofMax))
	}
	if llMax > maxLiteralLengthSymbol {
		panic(fmt.Errorf("llMax > maxLiteralLengthSymbol (%d)", llMax))
	}

	b.coders.mlEnc.HistogramFinished(mlMax, maxCount(mlH[:mlMax+1]))
	b.coders.ofEnc.HistogramFinished(ofMax, maxCount(ofH[:ofMax+1]))
	b.coders.llEnc.HistogramFinished(llMax, maxCount(llH[:llMax+1]))
}