Bump github.com/containers/storage from 1.37.0 to 1.38.0

Bumps [github.com/containers/storage](https://github.com/containers/storage) from 1.37.0 to 1.38.0.
- [Release notes](https://github.com/containers/storage/releases)
- [Changelog](https://github.com/containers/storage/blob/main/docs/containers-storage-changes.md)
- [Commits](https://github.com/containers/storage/compare/v1.37.0...v1.38.0)

---
updated-dependencies:
- dependency-name: github.com/containers/storage
  dependency-type: direct:production
  update-type: version-update:semver-minor
...

Signed-off-by: dependabot[bot] <support@github.com>

vendor/github.com/klauspost/compress/flate/deflate.go

@@ -6,9 +6,13 @@
 package flate
 
 import (
+	"encoding/binary"
 	"fmt"
 	"io"
 	"math"
+	"math/bits"
+
+	comp "github.com/klauspost/compress"
 )
 
 const (
@@ -37,15 +41,17 @@ const (
 	maxMatchLength   = 258 // The longest match for the compressor
 	minOffsetSize    = 1   // The shortest offset that makes any sense
 
-	// The maximum number of tokens we put into a single flat block, just too
-	// stop things from getting too large.
-	maxFlateBlockTokens = 1 << 14
+	// The maximum number of tokens we will encode at the time.
+	// Smaller sizes usually creates less optimal blocks.
+	// Bigger can make context switching slow.
+	// We use this for levels 7-9, so we make it big.
+	maxFlateBlockTokens = 1 << 15
 	maxStoreBlockSize   = 65535
 	hashBits            = 17 // After 17 performance degrades
 	hashSize            = 1 << hashBits
 	hashMask            = (1 << hashBits) - 1
 	hashShift           = (hashBits + minMatchLength - 1) / minMatchLength
-	maxHashOffset       = 1 << 24
+	maxHashOffset       = 1 << 28
 
 	skipNever = math.MaxInt32
 
@@ -70,9 +76,9 @@ var levels = []compressionLevel{
 	{0, 0, 0, 0, 0, 6},
 	// Levels 7-9 use increasingly more lazy matching
 	// and increasingly stringent conditions for "good enough".
-	{8, 8, 24, 16, skipNever, 7},
-	{10, 16, 24, 64, skipNever, 8},
-	{32, 258, 258, 4096, skipNever, 9},
+	{6, 10, 12, 16, skipNever, 7},
+	{10, 24, 32, 64, skipNever, 8},
+	{32, 258, 258, 1024, skipNever, 9},
 }
 
 // advancedState contains state for the advanced levels, with bigger hash tables, etc.
@@ -93,8 +99,9 @@ type advancedState struct {
 	hashOffset int
 
 	// input window: unprocessed data is window[index:windowEnd]
-	index     int
-	hashMatch [maxMatchLength + minMatchLength]uint32
+	index          int
+	estBitsPerByte int
+	hashMatch      [maxMatchLength + minMatchLength]uint32
 
 	hash uint32
 	ii   uint16 // position of last match, intended to overflow to reset.
@@ -170,7 +177,8 @@ func (d *compressor) writeBlock(tok *tokens, index int, eof bool) error {
 			window = d.window[d.blockStart:index]
 		}
 		d.blockStart = index
-		d.w.writeBlock(tok, eof, window)
+		//d.w.writeBlock(tok, eof, window)
+		d.w.writeBlockDynamic(tok, eof, window, d.sync)
 		return d.w.err
 	}
 	return nil
@@ -263,7 +271,7 @@ func (d *compressor) fillWindow(b []byte) {
 // Try to find a match starting at index whose length is greater than prevSize.
 // We only look at chainCount possibilities before giving up.
 // pos = s.index, prevHead = s.chainHead-s.hashOffset, prevLength=minMatchLength-1, lookahead
-func (d *compressor) findMatch(pos int, prevHead int, prevLength int, lookahead int) (length, offset int, ok bool) {
+func (d *compressor) findMatch(pos int, prevHead int, lookahead, bpb int) (length, offset int, ok bool) {
 	minMatchLook := maxMatchLength
 	if lookahead < minMatchLook {
 		minMatchLook = lookahead
@@ -279,36 +287,43 @@ func (d *compressor) findMatch(pos int, prevHead int, prevLength int, lookahead
 
 	// If we've got a match that's good enough, only look in 1/4 the chain.
 	tries := d.chain
-	length = prevLength
-	if length >= d.good {
-		tries >>= 2
-	}
+	length = minMatchLength - 1
 
 	wEnd := win[pos+length]
 	wPos := win[pos:]
 	minIndex := pos - windowSize
+	if minIndex < 0 {
+		minIndex = 0
+	}
+	offset = 0
 
+	// Base is 4 bytes at with an additional cost.
+	// Matches must be better than this.
+	cGain := minMatchLength*bpb - 12
 	for i := prevHead; tries > 0; tries-- {
 		if wEnd == win[i+length] {
 			n := matchLen(win[i:i+minMatchLook], wPos)
-
-			if n > length && (n > minMatchLength || pos-i <= 4096) {
-				length = n
-				offset = pos - i
-				ok = true
-				if n >= nice {
-					// The match is good enough that we don't try to find a better one.
-					break
+			if n > length {
+				newGain := n*bpb - bits.Len32(uint32(pos-i))
+				if newGain > cGain {
+					length = n
+					offset = pos - i
+					cGain = newGain
+					ok = true
+					if n >= nice {
+						// The match is good enough that we don't try to find a better one.
+						break
+					}
+					wEnd = win[pos+n]
 				}
-				wEnd = win[pos+n]
 			}
 		}
-		if i == minIndex {
+		if i <= minIndex {
 			// hashPrev[i & windowMask] has already been overwritten, so stop now.
 			break
 		}
 		i = int(d.state.hashPrev[i&windowMask]) - d.state.hashOffset
-		if i < minIndex || i < 0 {
+		if i < minIndex {
 			break
 		}
 	}
@@ -327,8 +342,7 @@ func (d *compressor) writeStoredBlock(buf []byte) error {
 // of the supplied slice.
 // The caller must ensure that len(b) >= 4.
 func hash4(b []byte) uint32 {
-	b = b[:4]
-	return hash4u(uint32(b[3])|uint32(b[2])<<8|uint32(b[1])<<16|uint32(b[0])<<24, hashBits)
+	return hash4u(binary.LittleEndian.Uint32(b), hashBits)
 }
 
 // bulkHash4 will compute hashes using the same
@@ -337,11 +351,12 @@ func bulkHash4(b []byte, dst []uint32) {
 	if len(b) < 4 {
 		return
 	}
-	hb := uint32(b[3]) | uint32(b[2])<<8 | uint32(b[1])<<16 | uint32(b[0])<<24
+	hb := binary.LittleEndian.Uint32(b)
+
 	dst[0] = hash4u(hb, hashBits)
 	end := len(b) - 4 + 1
 	for i := 1; i < end; i++ {
-		hb = (hb << 8) | uint32(b[i+3])
+		hb = (hb >> 8) | uint32(b[i+3])<<24
 		dst[i] = hash4u(hb, hashBits)
 	}
 }
@@ -374,10 +389,15 @@ func (d *compressor) deflateLazy() {
 	if d.windowEnd-s.index < minMatchLength+maxMatchLength && !d.sync {
 		return
 	}
+	s.estBitsPerByte = 8
+	if !d.sync {
+		s.estBitsPerByte = comp.ShannonEntropyBits(d.window[s.index:d.windowEnd])
+		s.estBitsPerByte = int(1 + float64(s.estBitsPerByte)/float64(d.windowEnd-s.index))
+	}
 
 	s.maxInsertIndex = d.windowEnd - (minMatchLength - 1)
 	if s.index < s.maxInsertIndex {
-		s.hash = hash4(d.window[s.index : s.index+minMatchLength])
+		s.hash = hash4(d.window[s.index:])
 	}
 
 	for {
@@ -410,7 +430,7 @@ func (d *compressor) deflateLazy() {
 		}
 		if s.index < s.maxInsertIndex {
 			// Update the hash
-			s.hash = hash4(d.window[s.index : s.index+minMatchLength])
+			s.hash = hash4(d.window[s.index:])
 			ch := s.hashHead[s.hash&hashMask]
 			s.chainHead = int(ch)
 			s.hashPrev[s.index&windowMask] = ch
@@ -426,12 +446,37 @@ func (d *compressor) deflateLazy() {
 		}
 
 		if s.chainHead-s.hashOffset >= minIndex && lookahead > prevLength && prevLength < d.lazy {
-			if newLength, newOffset, ok := d.findMatch(s.index, s.chainHead-s.hashOffset, minMatchLength-1, lookahead); ok {
+			if newLength, newOffset, ok := d.findMatch(s.index, s.chainHead-s.hashOffset, lookahead, s.estBitsPerByte); ok {
 				s.length = newLength
 				s.offset = newOffset
 			}
 		}
+
 		if prevLength >= minMatchLength && s.length <= prevLength {
+			// Check for better match at end...
+			//
+			// checkOff must be >=2 since we otherwise risk checking s.index
+			// Offset of 2 seems to yield best results.
+			const checkOff = 2
+			prevIndex := s.index - 1
+			if prevIndex+prevLength+checkOff < s.maxInsertIndex {
+				end := lookahead
+				if lookahead > maxMatchLength {
+					end = maxMatchLength
+				}
+				end += prevIndex
+				idx := prevIndex + prevLength - (4 - checkOff)
+				h := hash4(d.window[idx:])
+				ch2 := int(s.hashHead[h&hashMask]) - s.hashOffset - prevLength + (4 - checkOff)
+				if ch2 > minIndex {
+					length := matchLen(d.window[prevIndex:end], d.window[ch2:])
+					// It seems like a pure length metric is best.
+					if length > prevLength {
+						prevLength = length
+						prevOffset = prevIndex - ch2
+					}
+				}
+			}
 			// There was a match at the previous step, and the current match is
 			// not better. Output the previous match.
 			d.tokens.AddMatch(uint32(prevLength-3), uint32(prevOffset-minOffsetSize))
@@ -479,6 +524,7 @@ func (d *compressor) deflateLazy() {
 				}
 				d.tokens.Reset()
 			}
+			s.ii = 0
 		} else {
 			// Reset, if we got a match this run.
 			if s.length >= minMatchLength {
@@ -498,13 +544,12 @@ func (d *compressor) deflateLazy() {
 
 				// If we have a long run of no matches, skip additional bytes
 				// Resets when s.ii overflows after 64KB.
-				if s.ii > 31 {
-					n := int(s.ii >> 5)
+				if n := int(s.ii) - d.chain; n > 0 {
+					n = 1 + int(n>>6)
 					for j := 0; j < n; j++ {
 						if s.index >= d.windowEnd-1 {
 							break
 						}
-
 						d.tokens.AddLiteral(d.window[s.index-1])
 						if d.tokens.n == maxFlateBlockTokens {
 							if d.err = d.writeBlock(&d.tokens, s.index, false); d.err != nil {
@@ -512,6 +557,14 @@ func (d *compressor) deflateLazy() {
 							}
 							d.tokens.Reset()
 						}
+						// Index...
+						if s.index < s.maxInsertIndex {
+							h := hash4(d.window[s.index:])
+							ch := s.hashHead[h]
+							s.chainHead = int(ch)
+							s.hashPrev[s.index&windowMask] = ch
+							s.hashHead[h] = uint32(s.index + s.hashOffset)
+						}
 						s.index++
 					}
 					// Flush last byte
@@ -611,7 +664,9 @@ func (d *compressor) write(b []byte) (n int, err error) {
 	}
 	n = len(b)
 	for len(b) > 0 {
-		d.step(d)
+		if d.windowEnd == len(d.window) || d.sync {
+			d.step(d)
+		}
 		b = b[d.fill(d, b):]
 		if d.err != nil {
 			return 0, d.err
@@ -652,13 +707,13 @@ func (d *compressor) init(w io.Writer, level int) (err error) {
 		level = 5
 		fallthrough
 	case level >= 1 && level <= 6:
-		d.w.logNewTablePenalty = 8
+		d.w.logNewTablePenalty = 7
 		d.fast = newFastEnc(level)
 		d.window = make([]byte, maxStoreBlockSize)
 		d.fill = (*compressor).fillBlock
 		d.step = (*compressor).storeFast
 	case 7 <= level && level <= 9:
-		d.w.logNewTablePenalty = 10
+		d.w.logNewTablePenalty = 8
 		d.state = &advancedState{}
 		d.compressionLevel = levels[level]
 		d.initDeflate()

vendor/github.com/klauspost/compress/flate/fast_encoder.go

@@ -213,11 +213,9 @@ func (e *fastGen) Reset() {
 // matchLen returns the maximum length.
 // 'a' must be the shortest of the two.
 func matchLen(a, b []byte) int {
-	b = b[:len(a)]
 	var checked int
 
 	for len(a) >= 8 {
-		b = b[:len(a)]
 		if diff := binary.LittleEndian.Uint64(a) ^ binary.LittleEndian.Uint64(b); diff != 0 {
 			return checked + (bits.TrailingZeros64(diff) >> 3)
 		}

vendor/github.com/klauspost/compress/flate/huffman_bit_writer.go

@@ -155,37 +155,33 @@ func (w *huffmanBitWriter) reset(writer io.Writer) {
 	w.lastHuffMan = false
 }
 
-func (w *huffmanBitWriter) canReuse(t *tokens) (offsets, lits bool) {
-	offsets, lits = true, true
+func (w *huffmanBitWriter) canReuse(t *tokens) (ok bool) {
 	a := t.offHist[:offsetCodeCount]
-	b := w.offsetFreq[:len(a)]
-	for i := range a {
-		if b[i] == 0 && a[i] != 0 {
-			offsets = false
-			break
+	b := w.offsetEncoding.codes
+	b = b[:len(a)]
+	for i, v := range a {
+		if v != 0 && b[i].len == 0 {
+			return false
 		}
 	}
 
 	a = t.extraHist[:literalCount-256]
-	b = w.literalFreq[256:literalCount]
+	b = w.literalEncoding.codes[256:literalCount]
 	b = b[:len(a)]
-	for i := range a {
-		if b[i] == 0 && a[i] != 0 {
-			lits = false
-			break
+	for i, v := range a {
+		if v != 0 && b[i].len == 0 {
+			return false
 		}
 	}
-	if lits {
-		a = t.litHist[:]
-		b = w.literalFreq[:len(a)]
-		for i := range a {
-			if b[i] == 0 && a[i] != 0 {
-				lits = false
-				break
-			}
+
+	a = t.litHist[:256]
+	b = w.literalEncoding.codes[:len(a)]
+	for i, v := range a {
+		if v != 0 && b[i].len == 0 {
+			return false
 		}
 	}
-	return
+	return true
 }
 
 func (w *huffmanBitWriter) flush() {
@@ -222,7 +218,7 @@ func (w *huffmanBitWriter) write(b []byte) {
 }
 
 func (w *huffmanBitWriter) writeBits(b int32, nb uint16) {
-	w.bits |= uint64(b) << w.nbits
+	w.bits |= uint64(b) << (w.nbits & 63)
 	w.nbits += nb
 	if w.nbits >= 48 {
 		w.writeOutBits()
@@ -423,7 +419,7 @@ func (w *huffmanBitWriter) storedSize(in []byte) (int, bool) {
 
 func (w *huffmanBitWriter) writeCode(c hcode) {
 	// The function does not get inlined if we "& 63" the shift.
-	w.bits |= uint64(c.code) << w.nbits
+	w.bits |= uint64(c.code) << (w.nbits & 63)
 	w.nbits += c.len
 	if w.nbits >= 48 {
 		w.writeOutBits()
@@ -566,7 +562,7 @@ func (w *huffmanBitWriter) writeBlock(tokens *tokens, eof bool, input []byte) {
 		w.lastHeader = 0
 	}
 	numLiterals, numOffsets := w.indexTokens(tokens, false)
-	w.generate(tokens)
+	w.generate()
 	var extraBits int
 	storedSize, storable := w.storedSize(input)
 	if storable {
@@ -595,7 +591,7 @@ func (w *huffmanBitWriter) writeBlock(tokens *tokens, eof bool, input []byte) {
 	}
 
 	// Stored bytes?
-	if storable && storedSize < size {
+	if storable && storedSize <= size {
 		w.writeStoredHeader(len(input), eof)
 		w.writeBytes(input)
 		return
@@ -634,22 +630,39 @@ func (w *huffmanBitWriter) writeBlockDynamic(tokens *tokens, eof bool, input []b
 		w.lastHeader = 0
 		w.lastHuffMan = false
 	}
-	if !sync {
-		tokens.Fill()
+
+	// fillReuse enables filling of empty values.
+	// This will make encodings always reusable without testing.
+	// However, this does not appear to benefit on most cases.
+	const fillReuse = false
+
+	// Check if we can reuse...
+	if !fillReuse && w.lastHeader > 0 && !w.canReuse(tokens) {
+		w.writeCode(w.literalEncoding.codes[endBlockMarker])
+		w.lastHeader = 0
 	}
+
 	numLiterals, numOffsets := w.indexTokens(tokens, !sync)
+	extraBits := 0
+	ssize, storable := w.storedSize(input)
+
+	const usePrefs = true
+	if storable || w.lastHeader > 0 {
+		extraBits = w.extraBitSize()
+	}
 
 	var size int
+
 	// Check if we should reuse.
 	if w.lastHeader > 0 {
 		// Estimate size for using a new table.
 		// Use the previous header size as the best estimate.
 		newSize := w.lastHeader + tokens.EstimatedBits()
-		newSize += newSize >> w.logNewTablePenalty
+		newSize += int(w.literalEncoding.codes[endBlockMarker].len) + newSize>>w.logNewTablePenalty
 
 		// The estimated size is calculated as an optimal table.
 		// We add a penalty to make it more realistic and re-use a bit more.
-		reuseSize := w.dynamicReuseSize(w.literalEncoding, w.offsetEncoding) + w.extraBitSize()
+		reuseSize := w.dynamicReuseSize(w.literalEncoding, w.offsetEncoding) + extraBits
 
 		// Check if a new table is better.
 		if newSize < reuseSize {
@@ -660,35 +673,79 @@ func (w *huffmanBitWriter) writeBlockDynamic(tokens *tokens, eof bool, input []b
 		} else {
 			size = reuseSize
 		}
+
+		if preSize := w.fixedSize(extraBits) + 7; usePrefs && preSize < size {
+			// Check if we get a reasonable size decrease.
+			if storable && ssize <= size {
+				w.writeStoredHeader(len(input), eof)
+				w.writeBytes(input)
+				return
+			}
+			w.writeFixedHeader(eof)
+			if !sync {
+				tokens.AddEOB()
+			}
+			w.writeTokens(tokens.Slice(), fixedLiteralEncoding.codes, fixedOffsetEncoding.codes)
+			return
+		}
 		// Check if we get a reasonable size decrease.
-		if ssize, storable := w.storedSize(input); storable && ssize < (size+size>>4) {
+		if storable && ssize <= size {
 			w.writeStoredHeader(len(input), eof)
 			w.writeBytes(input)
-			w.lastHeader = 0
 			return
 		}
 	}
 
 	// We want a new block/table
 	if w.lastHeader == 0 {
-		w.generate(tokens)
+		if fillReuse && !sync {
+			w.fillTokens()
+			numLiterals, numOffsets = maxNumLit, maxNumDist
+		} else {
+			w.literalFreq[endBlockMarker] = 1
+		}
+
+		w.generate()
 		// Generate codegen and codegenFrequencies, which indicates how to encode
 		// the literalEncoding and the offsetEncoding.
 		w.generateCodegen(numLiterals, numOffsets, w.literalEncoding, w.offsetEncoding)
 		w.codegenEncoding.generate(w.codegenFreq[:], 7)
+
 		var numCodegens int
-		size, numCodegens = w.dynamicSize(w.literalEncoding, w.offsetEncoding, w.extraBitSize())
-		// Store bytes, if we don't get a reasonable improvement.
-		if ssize, storable := w.storedSize(input); storable && ssize < (size+size>>4) {
+		if fillReuse && !sync {
+			// Reindex for accurate size...
+			w.indexTokens(tokens, true)
+		}
+		size, numCodegens = w.dynamicSize(w.literalEncoding, w.offsetEncoding, extraBits)
+
+		// Store predefined, if we don't get a reasonable improvement.
+		if preSize := w.fixedSize(extraBits); usePrefs && preSize <= size {
+			// Store bytes, if we don't get an improvement.
+			if storable && ssize <= preSize {
+				w.writeStoredHeader(len(input), eof)
+				w.writeBytes(input)
+				return
+			}
+			w.writeFixedHeader(eof)
+			if !sync {
+				tokens.AddEOB()
+			}
+			w.writeTokens(tokens.Slice(), fixedLiteralEncoding.codes, fixedOffsetEncoding.codes)
+			return
+		}
+
+		if storable && ssize <= size {
+			// Store bytes, if we don't get an improvement.
 			w.writeStoredHeader(len(input), eof)
 			w.writeBytes(input)
-			w.lastHeader = 0
 			return
 		}
 
 		// Write Huffman table.
 		w.writeDynamicHeader(numLiterals, numOffsets, numCodegens, eof)
-		w.lastHeader, _ = w.headerSize()
+		if !sync {
+			w.lastHeader, _ = w.headerSize()
+		}
 		w.lastHuffMan = false
 	}
 
@@ -699,6 +756,19 @@ func (w *huffmanBitWriter) writeBlockDynamic(tokens *tokens, eof bool, input []b
 	w.writeTokens(tokens.Slice(), w.literalEncoding.codes, w.offsetEncoding.codes)
 }
 
+func (w *huffmanBitWriter) fillTokens() {
+	for i, v := range w.literalFreq[:literalCount] {
+		if v == 0 {
+			w.literalFreq[i] = 1
+		}
+	}
+	for i, v := range w.offsetFreq[:offsetCodeCount] {
+		if v == 0 {
+			w.offsetFreq[i] = 1
+		}
+	}
+}
+
 // indexTokens indexes a slice of tokens, and updates
 // literalFreq and offsetFreq, and generates literalEncoding
 // and offsetEncoding.
@@ -733,7 +803,7 @@ func (w *huffmanBitWriter) indexTokens(t *tokens, filled bool) (numLiterals, num
 	return
 }
 
-func (w *huffmanBitWriter) generate(t *tokens) {
+func (w *huffmanBitWriter) generate() {
 	w.literalEncoding.generate(w.literalFreq[:literalCount], 15)
 	w.offsetEncoding.generate(w.offsetFreq[:offsetCodeCount], 15)
 }
@@ -768,7 +838,7 @@ func (w *huffmanBitWriter) writeTokens(tokens []token, leCodes, oeCodes []hcode)
 		if t < matchType {
 			//w.writeCode(lits[t.literal()])
 			c := lits[t.literal()]
-			bits |= uint64(c.code) << nbits
+			bits |= uint64(c.code) << (nbits & 63)
 			nbits += c.len
 			if nbits >= 48 {
 				binary.LittleEndian.PutUint64(w.bytes[nbytes:], bits)
@@ -796,7 +866,7 @@ func (w *huffmanBitWriter) writeTokens(tokens []token, leCodes, oeCodes []hcode)
 		} else {
 			// inlined
 			c := lengths[lengthCode&31]
-			bits |= uint64(c.code) << nbits
+			bits |= uint64(c.code) << (nbits & 63)
 			nbits += c.len
 			if nbits >= 48 {
 				binary.LittleEndian.PutUint64(w.bytes[nbytes:], bits)
@@ -819,7 +889,7 @@ func (w *huffmanBitWriter) writeTokens(tokens []token, leCodes, oeCodes []hcode)
 		if extraLengthBits > 0 {
 			//w.writeBits(extraLength, extraLengthBits)
 			extraLength := int32(length - lengthBase[lengthCode&31])
-			bits |= uint64(extraLength) << nbits
+			bits |= uint64(extraLength) << (nbits & 63)
 			nbits += extraLengthBits
 			if nbits >= 48 {
 				binary.LittleEndian.PutUint64(w.bytes[nbytes:], bits)
@@ -846,7 +916,7 @@ func (w *huffmanBitWriter) writeTokens(tokens []token, leCodes, oeCodes []hcode)
 		} else {
 			// inlined
 			c := offs[offsetCode]
-			bits |= uint64(c.code) << nbits
+			bits |= uint64(c.code) << (nbits & 63)
 			nbits += c.len
 			if nbits >= 48 {
 				binary.LittleEndian.PutUint64(w.bytes[nbytes:], bits)
@@ -867,7 +937,7 @@ func (w *huffmanBitWriter) writeTokens(tokens []token, leCodes, oeCodes []hcode)
 		offsetComb := offsetCombined[offsetCode]
 		if offsetComb > 1<<16 {
 			//w.writeBits(extraOffset, extraOffsetBits)
-			bits |= uint64(offset&matchOffsetOnlyMask-(offsetComb&0xffff)) << nbits
+			bits |= uint64(offset-(offsetComb&0xffff)) << (nbits & 63)
 			nbits += uint16(offsetComb >> 16)
 			if nbits >= 48 {
 				binary.LittleEndian.PutUint64(w.bytes[nbytes:], bits)
@@ -996,10 +1066,41 @@ func (w *huffmanBitWriter) writeBlockHuff(eof bool, input []byte, sync bool) {
 	encoding := w.literalEncoding.codes[:256]
 	// Go 1.16 LOVES having these on stack. At least 1.5x the speed.
 	bits, nbits, nbytes := w.bits, w.nbits, w.nbytes
+
+	// Unroll, write 3 codes/loop.
+	// Fastest number of unrolls.
+	for len(input) > 3 {
+		// We must have at least 48 bits free.
+		if nbits >= 8 {
+			n := nbits >> 3
+			binary.LittleEndian.PutUint64(w.bytes[nbytes:], bits)
+			bits >>= (n * 8) & 63
+			nbits -= n * 8
+			nbytes += uint8(n)
+		}
+		if nbytes >= bufferFlushSize {
+			if w.err != nil {
+				nbytes = 0
+				return
+			}
+			_, w.err = w.writer.Write(w.bytes[:nbytes])
+			nbytes = 0
+		}
+		a, b := encoding[input[0]], encoding[input[1]]
+		bits |= uint64(a.code) << (nbits & 63)
+		bits |= uint64(b.code) << ((nbits + a.len) & 63)
+		c := encoding[input[2]]
+		nbits += b.len + a.len
+		bits |= uint64(c.code) << (nbits & 63)
+		nbits += c.len
+		input = input[3:]
+	}
+
+	// Remaining...
 	for _, t := range input {
 		// Bitwriting inlined, ~30% speedup
 		c := encoding[t]
-		bits |= uint64(c.code) << nbits
+		bits |= uint64(c.code) << (nbits & 63)
 		nbits += c.len
 		if debugDeflate {
 			count += int(c.len)

vendor/github.com/klauspost/compress/flate/inflate.go

@@ -328,11 +328,17 @@ func (f *decompressor) nextBlock() {
 	switch typ {
 	case 0:
 		f.dataBlock()
+		if debugDecode {
+			fmt.Println("stored block")
+		}
 	case 1:
 		// compressed, fixed Huffman tables
 		f.hl = &fixedHuffmanDecoder
 		f.hd = nil
 		f.huffmanBlockDecoder()()
+		if debugDecode {
+			fmt.Println("predefinied huffman block")
+		}
 	case 2:
 		// compressed, dynamic Huffman tables
 		if f.err = f.readHuffman(); f.err != nil {
@@ -341,6 +347,9 @@ func (f *decompressor) nextBlock() {
 		f.hl = &f.h1
 		f.hd = &f.h2
 		f.huffmanBlockDecoder()()
+		if debugDecode {
+			fmt.Println("dynamic huffman block")
+		}
 	default:
 		// 3 is reserved.
 		if debugDecode {