vendor: update github.com/klauspost/compress from v1.10.1 to v1.10.3

go.mod

@@ -7,7 +7,7 @@ require (
 	github.com/aws/aws-sdk-go v1.29.10
 	github.com/cespare/xxhash/v2 v2.1.1
 	github.com/golang/snappy v0.0.1
-	github.com/klauspost/compress v1.10.1
+	github.com/klauspost/compress v1.10.3
 	github.com/valyala/fasthttp v1.9.0
 	github.com/valyala/fastjson v1.5.0
 	github.com/valyala/fastrand v1.0.0

go.sum

@@ -105,8 +105,8 @@ github.com/kisielk/gotool v1.0.0/go.mod h1:XhKaO+MFFWcvkIS/tQcRk01m1F5IRFswLeQ+o
 github.com/klauspost/compress v1.4.0/go.mod h1:RyIbtBH6LamlWaDj8nUwkbUhJ87Yi3uG0guNDohfE1A=
 github.com/klauspost/compress v1.4.1/go.mod h1:RyIbtBH6LamlWaDj8nUwkbUhJ87Yi3uG0guNDohfE1A=
 github.com/klauspost/compress v1.8.2/go.mod h1:RyIbtBH6LamlWaDj8nUwkbUhJ87Yi3uG0guNDohfE1A=
-github.com/klauspost/compress v1.10.1 h1:a/QY0o9S6wCi0XhxaMX/QmusicNUqCqFugR6WKPOSoQ=
+github.com/klauspost/compress v1.10.3 h1:OP96hzwJVBIHYU52pVTI6CczrxPvrGfgqF9N5eTO0Q8=
-github.com/klauspost/compress v1.10.1/go.mod h1:aoV0uJVorq1K+umq18yTdKaF57EivdYsUV+/s2qKfXs=
+github.com/klauspost/compress v1.10.3/go.mod h1:aoV0uJVorq1K+umq18yTdKaF57EivdYsUV+/s2qKfXs=
 github.com/klauspost/cpuid v0.0.0-20180405133222-e7e905edc00e/go.mod h1:Pj4uuM528wm8OyEC2QMXAi2YiTZ96dNQPGgoMS4s3ek=
 github.com/klauspost/cpuid v1.2.0/go.mod h1:Pj4uuM528wm8OyEC2QMXAi2YiTZ96dNQPGgoMS4s3ek=
 github.com/klauspost/cpuid v1.2.1/go.mod h1:Pj4uuM528wm8OyEC2QMXAi2YiTZ96dNQPGgoMS4s3ek=

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

@@ -48,6 +48,8 @@ const (
 	maxHashOffset       = 1 << 24
 
 	skipNever = math.MaxInt32
+
+	debugDeflate = false
 )
 
 type compressionLevel struct {
@@ -365,7 +367,7 @@ func (d *compressor) deflateLazy() {
 	// Sanity enables additional runtime tests.
 	// It's intended to be used during development
 	// to supplement the currently ad-hoc unit tests.
-	const sanity = false
+	const sanity = debugDeflate
 
 	if d.windowEnd-s.index < minMatchLength+maxMatchLength && !d.sync {
 		return

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

@@ -35,16 +35,16 @@ func newFastEnc(level int) fastEnc {
 }
 
 const (
-	tableBits       = 16             // Bits used in the table
+	tableBits       = 15             // Bits used in the table
 	tableSize       = 1 << tableBits // Size of the table
 	tableShift      = 32 - tableBits // Right-shift to get the tableBits most significant bits of a uint32.
 	baseMatchOffset = 1              // The smallest match offset
 	baseMatchLength = 3              // The smallest match length per the RFC section 3.2.5
 	maxMatchOffset  = 1 << 15        // The largest match offset
 
-	bTableBits   = 18                                               // Bits used in the big tables
+	bTableBits   = 17                                               // Bits used in the big tables
 	bTableSize   = 1 << bTableBits                                  // Size of the table
-	allocHistory = maxStoreBlockSize * 20                           // Size to preallocate for history.
+	allocHistory = maxStoreBlockSize * 10                           // Size to preallocate for history.
 	bufferReset  = (1 << 31) - allocHistory - maxStoreBlockSize - 1 // Reset the buffer offset when reaching this.
 )
 
@@ -92,7 +92,6 @@ func hash(u uint32) uint32 {
 }
 
 type tableEntry struct {
-	val    uint32
 	offset int32
 }
 

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

@@ -0,0 +1,274 @@
+// +build generate
+
+//go:generate go run $GOFILE && gofmt -w inflate_gen.go
+
+package main
+
+import (
+	"os"
+	"strings"
+)
+
+func main() {
+	f, err := os.Create("inflate_gen.go")
+	if err != nil {
+		panic(err)
+	}
+	defer f.Close()
+	types := []string{"*bytes.Buffer", "*bytes.Reader", "*bufio.Reader", "*strings.Reader"}
+	names := []string{"BytesBuffer", "BytesReader", "BufioReader", "StringsReader"}
+	imports := []string{"bytes", "bufio", "io", "strings", "math/bits"}
+	f.WriteString(`// Code generated by go generate gen_inflate.go. DO NOT EDIT.
+
+package flate
+
+import (
+`)
+
+	for _, imp := range imports {
+		f.WriteString("\t\"" + imp + "\"\n")
+	}
+	f.WriteString(")\n\n")
+
+	template := `
+
+// Decode a single Huffman block from f.
+// hl and hd are the Huffman states for the lit/length values
+// and the distance values, respectively. If hd == nil, using the
+// fixed distance encoding associated with fixed Huffman blocks.
+func (f *decompressor) $FUNCNAME$() {
+	const (
+		stateInit = iota // Zero value must be stateInit
+		stateDict
+	)
+	fr := f.r.($TYPE$)
+	moreBits := func() error {
+		c, err := fr.ReadByte()
+		if err != nil {
+			return noEOF(err)
+		}
+		f.roffset++
+		f.b |= uint32(c) << f.nb
+		f.nb += 8
+		return nil
+	}
+
+	switch f.stepState {
+	case stateInit:
+		goto readLiteral
+	case stateDict:
+		goto copyHistory
+	}
+
+readLiteral:
+	// Read literal and/or (length, distance) according to RFC section 3.2.3.
+	{
+		var v int
+		{
+			// Inlined v, err := f.huffSym(f.hl)
+			// Since a huffmanDecoder can be empty or be composed of a degenerate tree
+			// with single element, huffSym must error on these two edge cases. In both
+			// cases, the chunks slice will be 0 for the invalid sequence, leading it
+			// satisfy the n == 0 check below.
+			n := uint(f.hl.maxRead)
+			// Optimization. Compiler isn't smart enough to keep f.b,f.nb in registers,
+			// but is smart enough to keep local variables in registers, so use nb and b,
+			// inline call to moreBits and reassign b,nb back to f on return.
+			nb, b := f.nb, f.b
+			for {
+				for nb < n {
+					c, err := fr.ReadByte()
+					if err != nil {
+						f.b = b
+						f.nb = nb
+						f.err = noEOF(err)
+						return
+					}
+					f.roffset++
+					b |= uint32(c) << (nb & 31)
+					nb += 8
+				}
+				chunk := f.hl.chunks[b&(huffmanNumChunks-1)]
+				n = uint(chunk & huffmanCountMask)
+				if n > huffmanChunkBits {
+					chunk = f.hl.links[chunk>>huffmanValueShift][(b>>huffmanChunkBits)&f.hl.linkMask]
+					n = uint(chunk & huffmanCountMask)
+				}
+				if n <= nb {
+					if n == 0 {
+						f.b = b
+						f.nb = nb
+						if debugDecode {
+							fmt.Println("huffsym: n==0")
+						}
+						f.err = CorruptInputError(f.roffset)
+						return
+					}
+					f.b = b >> (n & 31)
+					f.nb = nb - n
+					v = int(chunk >> huffmanValueShift)
+					break
+				}
+			}
+		}
+
+		var n uint // number of bits extra
+		var length int
+		var err error
+		switch {
+		case v < 256:
+			f.dict.writeByte(byte(v))
+			if f.dict.availWrite() == 0 {
+				f.toRead = f.dict.readFlush()
+				f.step = (*decompressor).$FUNCNAME$
+				f.stepState = stateInit
+				return
+			}
+			goto readLiteral
+		case v == 256:
+			f.finishBlock()
+			return
+		// otherwise, reference to older data
+		case v < 265:
+			length = v - (257 - 3)
+			n = 0
+		case v < 269:
+			length = v*2 - (265*2 - 11)
+			n = 1
+		case v < 273:
+			length = v*4 - (269*4 - 19)
+			n = 2
+		case v < 277:
+			length = v*8 - (273*8 - 35)
+			n = 3
+		case v < 281:
+			length = v*16 - (277*16 - 67)
+			n = 4
+		case v < 285:
+			length = v*32 - (281*32 - 131)
+			n = 5
+		case v < maxNumLit:
+			length = 258
+			n = 0
+		default:
+			if debugDecode {
+				fmt.Println(v, ">= maxNumLit")
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+		if n > 0 {
+			for f.nb < n {
+				if err = moreBits(); err != nil {
+					if debugDecode {
+						fmt.Println("morebits n>0:", err)
+					}
+					f.err = err
+					return
+				}
+			}
+			length += int(f.b & uint32(1<<n-1))
+			f.b >>= n
+			f.nb -= n
+		}
+
+		var dist int
+		if f.hd == nil {
+			for f.nb < 5 {
+				if err = moreBits(); err != nil {
+					if debugDecode {
+						fmt.Println("morebits f.nb<5:", err)
+					}
+					f.err = err
+					return
+				}
+			}
+			dist = int(bits.Reverse8(uint8(f.b & 0x1F << 3)))
+			f.b >>= 5
+			f.nb -= 5
+		} else {
+			if dist, err = f.huffSym(f.hd); err != nil {
+				if debugDecode {
+					fmt.Println("huffsym:", err)
+				}
+				f.err = err
+				return
+			}
+		}
+
+		switch {
+		case dist < 4:
+			dist++
+		case dist < maxNumDist:
+			nb := uint(dist-2) >> 1
+			// have 1 bit in bottom of dist, need nb more.
+			extra := (dist & 1) << nb
+			for f.nb < nb {
+				if err = moreBits(); err != nil {
+					if debugDecode {
+						fmt.Println("morebits f.nb<nb:", err)
+					}
+					f.err = err
+					return
+				}
+			}
+			extra |= int(f.b & uint32(1<<nb-1))
+			f.b >>= nb
+			f.nb -= nb
+			dist = 1<<(nb+1) + 1 + extra
+		default:
+			if debugDecode {
+				fmt.Println("dist too big:", dist, maxNumDist)
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		// No check on length; encoding can be prescient.
+		if dist > f.dict.histSize() {
+			if debugDecode {
+				fmt.Println("dist > f.dict.histSize():", dist, f.dict.histSize())
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		f.copyLen, f.copyDist = length, dist
+		goto copyHistory
+	}
+
+copyHistory:
+	// Perform a backwards copy according to RFC section 3.2.3.
+	{
+		cnt := f.dict.tryWriteCopy(f.copyDist, f.copyLen)
+		if cnt == 0 {
+			cnt = f.dict.writeCopy(f.copyDist, f.copyLen)
+		}
+		f.copyLen -= cnt
+
+		if f.dict.availWrite() == 0 || f.copyLen > 0 {
+			f.toRead = f.dict.readFlush()
+			f.step = (*decompressor).$FUNCNAME$ // We need to continue this work
+			f.stepState = stateDict
+			return
+		}
+		goto readLiteral
+	}
+}
+
+`
+	for i, t := range types {
+		s := strings.Replace(template, "$FUNCNAME$", "huffman"+names[i], -1)
+		s = strings.Replace(s, "$TYPE$", t, -1)
+		f.WriteString(s)
+	}
+	f.WriteString("func (f *decompressor) huffmanBlockDecoder() func() {\n")
+	f.WriteString("\tswitch f.r.(type) {\n")
+	for i, t := range types {
+		f.WriteString("\t\tcase " + t + ":\n")
+		f.WriteString("\t\t\treturn f.huffman" + names[i] + "\n")
+	}
+	f.WriteString("\t\tdefault:\n")
+	f.WriteString("\t\t\treturn f.huffmanBlockGeneric")
+	f.WriteString("\t}\n}\n")
+}

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

@@ -484,6 +484,9 @@ func (w *huffmanBitWriter) writeDynamicHeader(numLiterals int, numOffsets int, n
 	}
 }
 
+// writeStoredHeader will write a stored header.
+// If the stored block is only used for EOF,
+// it is replaced with a fixed huffman block.
 func (w *huffmanBitWriter) writeStoredHeader(length int, isEof bool) {
 	if w.err != nil {
 		return
@@ -493,6 +496,16 @@ func (w *huffmanBitWriter) writeStoredHeader(length int, isEof bool) {
 		w.writeCode(w.literalEncoding.codes[endBlockMarker])
 		w.lastHeader = 0
 	}
+
+	// To write EOF, use a fixed encoding block. 10 bits instead of 5 bytes.
+	if length == 0 && isEof {
+		w.writeFixedHeader(isEof)
+		// EOB: 7 bits, value: 0
+		w.writeBits(0, 7)
+		w.flush()
+		return
+	}
+
 	var flag int32
 	if isEof {
 		flag = 1

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

@@ -109,8 +109,8 @@ func generateFixedOffsetEncoding() *huffmanEncoder {
 	return h
 }
 
-var fixedLiteralEncoding *huffmanEncoder = generateFixedLiteralEncoding()
+var fixedLiteralEncoding = generateFixedLiteralEncoding()
-var fixedOffsetEncoding *huffmanEncoder = generateFixedOffsetEncoding()
+var fixedOffsetEncoding = generateFixedOffsetEncoding()
 
 func (h *huffmanEncoder) bitLength(freq []uint16) int {
 	var total int

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

@@ -106,7 +106,7 @@ const (
 )
 
 type huffmanDecoder struct {
-	min      int                       // the minimum code length
+	maxRead  int                       // the maximum number of bits we can read and not overread
 	chunks   *[huffmanNumChunks]uint16 // chunks as described above
 	links    [][]uint16                // overflow links
 	linkMask uint32                    // mask the width of the link table
@@ -126,12 +126,12 @@ func (h *huffmanDecoder) init(lengths []int) bool {
 	if h.chunks == nil {
 		h.chunks = &[huffmanNumChunks]uint16{}
 	}
-	if h.min != 0 {
+	if h.maxRead != 0 {
 		*h = huffmanDecoder{chunks: h.chunks, links: h.links}
 	}
 
 	// Count number of codes of each length,
-	// compute min and max length.
+	// compute maxRead and max length.
 	var count [maxCodeLen]int
 	var min, max int
 	for _, n := range lengths {
@@ -178,7 +178,7 @@ func (h *huffmanDecoder) init(lengths []int) bool {
 		return false
 	}
 
-	h.min = min
+	h.maxRead = min
 	chunks := h.chunks[:]
 	for i := range chunks {
 		chunks[i] = 0
@@ -342,7 +342,7 @@ func (f *decompressor) nextBlock() {
 		// compressed, fixed Huffman tables
 		f.hl = &fixedHuffmanDecoder
 		f.hd = nil
-		f.huffmanBlock()
+		f.huffmanBlockDecoder()()
 	case 2:
 		// compressed, dynamic Huffman tables
 		if f.err = f.readHuffman(); f.err != nil {
@@ -350,7 +350,7 @@ func (f *decompressor) nextBlock() {
 		}
 		f.hl = &f.h1
 		f.hd = &f.h2
-		f.huffmanBlock()
+		f.huffmanBlockDecoder()()
 	default:
 		// 3 is reserved.
 		if debugDecode {
@@ -543,12 +543,18 @@ func (f *decompressor) readHuffman() error {
 		return CorruptInputError(f.roffset)
 	}
 
-	// As an optimization, we can initialize the min bits to read at a time
+	// As an optimization, we can initialize the maxRead bits to read at a time
 	// for the HLIT tree to the length of the EOB marker since we know that
 	// every block must terminate with one. This preserves the property that
 	// we never read any extra bytes after the end of the DEFLATE stream.
-	if f.h1.min < f.bits[endBlockMarker] {
+	if f.h1.maxRead < f.bits[endBlockMarker] {
-		f.h1.min = f.bits[endBlockMarker]
+		f.h1.maxRead = f.bits[endBlockMarker]
+	}
+	if !f.final {
+		// If not the final block, the smallest block possible is
+		// a predefined table, BTYPE=01, with a single EOB marker.
+		// This will take up 3 + 7 bits.
+		f.h1.maxRead += 10
 	}
 
 	return nil
@@ -558,7 +564,7 @@ func (f *decompressor) readHuffman() error {
 // hl and hd are the Huffman states for the lit/length values
 // and the distance values, respectively. If hd == nil, using the
 // fixed distance encoding associated with fixed Huffman blocks.
-func (f *decompressor) huffmanBlock() {
+func (f *decompressor) huffmanBlockGeneric() {
 	const (
 		stateInit = iota // Zero value must be stateInit
 		stateDict
@@ -574,19 +580,64 @@ func (f *decompressor) huffmanBlock() {
 readLiteral:
 	// Read literal and/or (length, distance) according to RFC section 3.2.3.
 	{
-		v, err := f.huffSym(f.hl)
+		var v int
-		if err != nil {
+		{
-			f.err = err
+			// Inlined v, err := f.huffSym(f.hl)
-			return
+			// Since a huffmanDecoder can be empty or be composed of a degenerate tree
+			// with single element, huffSym must error on these two edge cases. In both
+			// cases, the chunks slice will be 0 for the invalid sequence, leading it
+			// satisfy the n == 0 check below.
+			n := uint(f.hl.maxRead)
+			// Optimization. Compiler isn't smart enough to keep f.b,f.nb in registers,
+			// but is smart enough to keep local variables in registers, so use nb and b,
+			// inline call to moreBits and reassign b,nb back to f on return.
+			nb, b := f.nb, f.b
+			for {
+				for nb < n {
+					c, err := f.r.ReadByte()
+					if err != nil {
+						f.b = b
+						f.nb = nb
+						f.err = noEOF(err)
+						return
+					}
+					f.roffset++
+					b |= uint32(c) << (nb & 31)
+					nb += 8
+				}
+				chunk := f.hl.chunks[b&(huffmanNumChunks-1)]
+				n = uint(chunk & huffmanCountMask)
+				if n > huffmanChunkBits {
+					chunk = f.hl.links[chunk>>huffmanValueShift][(b>>huffmanChunkBits)&f.hl.linkMask]
+					n = uint(chunk & huffmanCountMask)
+				}
+				if n <= nb {
+					if n == 0 {
+						f.b = b
+						f.nb = nb
+						if debugDecode {
+							fmt.Println("huffsym: n==0")
+						}
+						f.err = CorruptInputError(f.roffset)
+						return
+					}
+					f.b = b >> (n & 31)
+					f.nb = nb - n
+					v = int(chunk >> huffmanValueShift)
+					break
+				}
+			}
 		}
+
 		var n uint // number of bits extra
 		var length int
+		var err error
 		switch {
 		case v < 256:
 			f.dict.writeByte(byte(v))
 			if f.dict.availWrite() == 0 {
 				f.toRead = f.dict.readFlush()
-				f.step = (*decompressor).huffmanBlock
+				f.step = (*decompressor).huffmanBlockGeneric
 				f.stepState = stateInit
 				return
 			}
@@ -714,7 +765,7 @@ copyHistory:
 
 		if f.dict.availWrite() == 0 || f.copyLen > 0 {
 			f.toRead = f.dict.readFlush()
-			f.step = (*decompressor).huffmanBlock // We need to continue this work
+			f.step = (*decompressor).huffmanBlockGeneric // We need to continue this work
 			f.stepState = stateDict
 			return
 		}
@@ -726,21 +777,33 @@ copyHistory:
 func (f *decompressor) dataBlock() {
 	// Uncompressed.
 	// Discard current half-byte.
-	f.nb = 0
+	left := (f.nb) & 7
-	f.b = 0
+	f.nb -= left
+	f.b >>= left
+
+	offBytes := f.nb >> 3
+	// Unfilled values will be overwritten.
+	f.buf[0] = uint8(f.b)
+	f.buf[1] = uint8(f.b >> 8)
+	f.buf[2] = uint8(f.b >> 16)
+	f.buf[3] = uint8(f.b >> 24)
+
+	f.roffset += int64(offBytes)
+	f.nb, f.b = 0, 0
 
 	// Length then ones-complement of length.
-	nr, err := io.ReadFull(f.r, f.buf[0:4])
+	nr, err := io.ReadFull(f.r, f.buf[offBytes:4])
 	f.roffset += int64(nr)
 	if err != nil {
 		f.err = noEOF(err)
 		return
 	}
-	n := int(f.buf[0]) | int(f.buf[1])<<8
+	n := uint16(f.buf[0]) | uint16(f.buf[1])<<8
-	nn := int(f.buf[2]) | int(f.buf[3])<<8
+	nn := uint16(f.buf[2]) | uint16(f.buf[3])<<8
-	if uint16(nn) != uint16(^n) {
+	if nn != ^n {
 		if debugDecode {
-			fmt.Println("uint16(nn) != uint16(^n)", nn, ^n)
+			ncomp := ^n
+			fmt.Println("uint16(nn) != uint16(^n)", nn, ncomp)
 		}
 		f.err = CorruptInputError(f.roffset)
 		return
@@ -752,7 +815,7 @@ func (f *decompressor) dataBlock() {
 		return
 	}
 
-	f.copyLen = n
+	f.copyLen = int(n)
 	f.copyData()
 }
 
@@ -816,7 +879,7 @@ func (f *decompressor) huffSym(h *huffmanDecoder) (int, error) {
 	// with single element, huffSym must error on these two edge cases. In both
 	// cases, the chunks slice will be 0 for the invalid sequence, leading it
 	// satisfy the n == 0 check below.
-	n := uint(h.min)
+	n := uint(h.maxRead)
 	// Optimization. Compiler isn't smart enough to keep f.b,f.nb in registers,
 	// but is smart enough to keep local variables in registers, so use nb and b,
 	// inline call to moreBits and reassign b,nb back to f on return.

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

@@ -0,0 +1,922 @@
+// Code generated by go generate gen_inflate.go. DO NOT EDIT.
+
+package flate
+
+import (
+	"bufio"
+	"bytes"
+	"fmt"
+	"math/bits"
+	"strings"
+)
+
+// Decode a single Huffman block from f.
+// hl and hd are the Huffman states for the lit/length values
+// and the distance values, respectively. If hd == nil, using the
+// fixed distance encoding associated with fixed Huffman blocks.
+func (f *decompressor) huffmanBytesBuffer() {
+	const (
+		stateInit = iota // Zero value must be stateInit
+		stateDict
+	)
+	fr := f.r.(*bytes.Buffer)
+	moreBits := func() error {
+		c, err := fr.ReadByte()
+		if err != nil {
+			return noEOF(err)
+		}
+		f.roffset++
+		f.b |= uint32(c) << f.nb
+		f.nb += 8
+		return nil
+	}
+
+	switch f.stepState {
+	case stateInit:
+		goto readLiteral
+	case stateDict:
+		goto copyHistory
+	}
+
+readLiteral:
+	// Read literal and/or (length, distance) according to RFC section 3.2.3.
+	{
+		var v int
+		{
+			// Inlined v, err := f.huffSym(f.hl)
+			// Since a huffmanDecoder can be empty or be composed of a degenerate tree
+			// with single element, huffSym must error on these two edge cases. In both
+			// cases, the chunks slice will be 0 for the invalid sequence, leading it
+			// satisfy the n == 0 check below.
+			n := uint(f.hl.maxRead)
+			// Optimization. Compiler isn't smart enough to keep f.b,f.nb in registers,
+			// but is smart enough to keep local variables in registers, so use nb and b,
+			// inline call to moreBits and reassign b,nb back to f on return.
+			nb, b := f.nb, f.b
+			for {
+				for nb < n {
+					c, err := fr.ReadByte()
+					if err != nil {
+						f.b = b
+						f.nb = nb
+						f.err = noEOF(err)
+						return
+					}
+					f.roffset++
+					b |= uint32(c) << (nb & 31)
+					nb += 8
+				}
+				chunk := f.hl.chunks[b&(huffmanNumChunks-1)]
+				n = uint(chunk & huffmanCountMask)
+				if n > huffmanChunkBits {
+					chunk = f.hl.links[chunk>>huffmanValueShift][(b>>huffmanChunkBits)&f.hl.linkMask]
+					n = uint(chunk & huffmanCountMask)
+				}
+				if n <= nb {
+					if n == 0 {
+						f.b = b
+						f.nb = nb
+						if debugDecode {
+							fmt.Println("huffsym: n==0")
+						}
+						f.err = CorruptInputError(f.roffset)
+						return
+					}
+					f.b = b >> (n & 31)
+					f.nb = nb - n
+					v = int(chunk >> huffmanValueShift)
+					break
+				}
+			}
+		}
+
+		var n uint // number of bits extra
+		var length int
+		var err error
+		switch {
+		case v < 256:
+			f.dict.writeByte(byte(v))
+			if f.dict.availWrite() == 0 {
+				f.toRead = f.dict.readFlush()
+				f.step = (*decompressor).huffmanBytesBuffer
+				f.stepState = stateInit
+				return
+			}
+			goto readLiteral
+		case v == 256:
+			f.finishBlock()
+			return
+		// otherwise, reference to older data
+		case v < 265:
+			length = v - (257 - 3)
+			n = 0
+		case v < 269:
+			length = v*2 - (265*2 - 11)
+			n = 1
+		case v < 273:
+			length = v*4 - (269*4 - 19)
+			n = 2
+		case v < 277:
+			length = v*8 - (273*8 - 35)
+			n = 3
+		case v < 281:
+			length = v*16 - (277*16 - 67)
+			n = 4
+		case v < 285:
+			length = v*32 - (281*32 - 131)
+			n = 5
+		case v < maxNumLit:
+			length = 258
+			n = 0
+		default:
+			if debugDecode {
+				fmt.Println(v, ">= maxNumLit")
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+		if n > 0 {
+			for f.nb < n {
+				if err = moreBits(); err != nil {
+					if debugDecode {
+						fmt.Println("morebits n>0:", err)
+					}
+					f.err = err
+					return
+				}
+			}
+			length += int(f.b & uint32(1<<n-1))
+			f.b >>= n
+			f.nb -= n
+		}
+
+		var dist int
+		if f.hd == nil {
+			for f.nb < 5 {
+				if err = moreBits(); err != nil {
+					if debugDecode {
+						fmt.Println("morebits f.nb<5:", err)
+					}
+					f.err = err
+					return
+				}
+			}
+			dist = int(bits.Reverse8(uint8(f.b & 0x1F << 3)))
+			f.b >>= 5
+			f.nb -= 5
+		} else {
+			if dist, err = f.huffSym(f.hd); err != nil {
+				if debugDecode {
+					fmt.Println("huffsym:", err)
+				}
+				f.err = err
+				return
+			}
+		}
+
+		switch {
+		case dist < 4:
+			dist++
+		case dist < maxNumDist:
+			nb := uint(dist-2) >> 1
+			// have 1 bit in bottom of dist, need nb more.
+			extra := (dist & 1) << nb
+			for f.nb < nb {
+				if err = moreBits(); err != nil {
+					if debugDecode {
+						fmt.Println("morebits f.nb<nb:", err)
+					}
+					f.err = err
+					return
+				}
+			}
+			extra |= int(f.b & uint32(1<<nb-1))
+			f.b >>= nb
+			f.nb -= nb
+			dist = 1<<(nb+1) + 1 + extra
+		default:
+			if debugDecode {
+				fmt.Println("dist too big:", dist, maxNumDist)
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		// No check on length; encoding can be prescient.
+		if dist > f.dict.histSize() {
+			if debugDecode {
+				fmt.Println("dist > f.dict.histSize():", dist, f.dict.histSize())
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		f.copyLen, f.copyDist = length, dist
+		goto copyHistory
+	}
+
+copyHistory:
+	// Perform a backwards copy according to RFC section 3.2.3.
+	{
+		cnt := f.dict.tryWriteCopy(f.copyDist, f.copyLen)
+		if cnt == 0 {
+			cnt = f.dict.writeCopy(f.copyDist, f.copyLen)
+		}
+		f.copyLen -= cnt
+
+		if f.dict.availWrite() == 0 || f.copyLen > 0 {
+			f.toRead = f.dict.readFlush()
+			f.step = (*decompressor).huffmanBytesBuffer // We need to continue this work
+			f.stepState = stateDict
+			return
+		}
+		goto readLiteral
+	}
+}
+
+// Decode a single Huffman block from f.
+// hl and hd are the Huffman states for the lit/length values
+// and the distance values, respectively. If hd == nil, using the
+// fixed distance encoding associated with fixed Huffman blocks.
+func (f *decompressor) huffmanBytesReader() {
+	const (
+		stateInit = iota // Zero value must be stateInit
+		stateDict
+	)
+	fr := f.r.(*bytes.Reader)
+	moreBits := func() error {
+		c, err := fr.ReadByte()
+		if err != nil {
+			return noEOF(err)
+		}
+		f.roffset++
+		f.b |= uint32(c) << f.nb
+		f.nb += 8
+		return nil
+	}
+
+	switch f.stepState {
+	case stateInit:
+		goto readLiteral
+	case stateDict:
+		goto copyHistory
+	}
+
+readLiteral:
+	// Read literal and/or (length, distance) according to RFC section 3.2.3.
+	{
+		var v int
+		{
+			// Inlined v, err := f.huffSym(f.hl)
+			// Since a huffmanDecoder can be empty or be composed of a degenerate tree
+			// with single element, huffSym must error on these two edge cases. In both
+			// cases, the chunks slice will be 0 for the invalid sequence, leading it
+			// satisfy the n == 0 check below.
+			n := uint(f.hl.maxRead)
+			// Optimization. Compiler isn't smart enough to keep f.b,f.nb in registers,
+			// but is smart enough to keep local variables in registers, so use nb and b,
+			// inline call to moreBits and reassign b,nb back to f on return.
+			nb, b := f.nb, f.b
+			for {
+				for nb < n {
+					c, err := fr.ReadByte()
+					if err != nil {
+						f.b = b
+						f.nb = nb
+						f.err = noEOF(err)
+						return
+					}
+					f.roffset++
+					b |= uint32(c) << (nb & 31)
+					nb += 8
+				}
+				chunk := f.hl.chunks[b&(huffmanNumChunks-1)]
+				n = uint(chunk & huffmanCountMask)
+				if n > huffmanChunkBits {
+					chunk = f.hl.links[chunk>>huffmanValueShift][(b>>huffmanChunkBits)&f.hl.linkMask]
+					n = uint(chunk & huffmanCountMask)
+				}
+				if n <= nb {
+					if n == 0 {
+						f.b = b
+						f.nb = nb
+						if debugDecode {
+							fmt.Println("huffsym: n==0")
+						}
+						f.err = CorruptInputError(f.roffset)
+						return
+					}
+					f.b = b >> (n & 31)
+					f.nb = nb - n
+					v = int(chunk >> huffmanValueShift)
+					break
+				}
+			}
+		}
+
+		var n uint // number of bits extra
+		var length int
+		var err error
+		switch {
+		case v < 256:
+			f.dict.writeByte(byte(v))
+			if f.dict.availWrite() == 0 {
+				f.toRead = f.dict.readFlush()
+				f.step = (*decompressor).huffmanBytesReader
+				f.stepState = stateInit
+				return
+			}
+			goto readLiteral
+		case v == 256:
+			f.finishBlock()
+			return
+		// otherwise, reference to older data
+		case v < 265:
+			length = v - (257 - 3)
+			n = 0
+		case v < 269:
+			length = v*2 - (265*2 - 11)
+			n = 1
+		case v < 273:
+			length = v*4 - (269*4 - 19)
+			n = 2
+		case v < 277:
+			length = v*8 - (273*8 - 35)
+			n = 3
+		case v < 281:
+			length = v*16 - (277*16 - 67)
+			n = 4
+		case v < 285:
+			length = v*32 - (281*32 - 131)
+			n = 5
+		case v < maxNumLit:
+			length = 258
+			n = 0
+		default:
+			if debugDecode {
+				fmt.Println(v, ">= maxNumLit")
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+		if n > 0 {
+			for f.nb < n {
+				if err = moreBits(); err != nil {
+					if debugDecode {
+						fmt.Println("morebits n>0:", err)
+					}
+					f.err = err
+					return
+				}
+			}
+			length += int(f.b & uint32(1<<n-1))
+			f.b >>= n
+			f.nb -= n
+		}
+
+		var dist int
+		if f.hd == nil {
+			for f.nb < 5 {
+				if err = moreBits(); err != nil {
+					if debugDecode {
+						fmt.Println("morebits f.nb<5:", err)
+					}
+					f.err = err
+					return
+				}
+			}
+			dist = int(bits.Reverse8(uint8(f.b & 0x1F << 3)))
+			f.b >>= 5
+			f.nb -= 5
+		} else {
+			if dist, err = f.huffSym(f.hd); err != nil {
+				if debugDecode {
+					fmt.Println("huffsym:", err)
+				}
+				f.err = err
+				return
+			}
+		}
+
+		switch {
+		case dist < 4:
+			dist++
+		case dist < maxNumDist:
+			nb := uint(dist-2) >> 1
+			// have 1 bit in bottom of dist, need nb more.
+			extra := (dist & 1) << nb
+			for f.nb < nb {
+				if err = moreBits(); err != nil {
+					if debugDecode {
+						fmt.Println("morebits f.nb<nb:", err)
+					}
+					f.err = err
+					return
+				}
+			}
+			extra |= int(f.b & uint32(1<<nb-1))
+			f.b >>= nb
+			f.nb -= nb
+			dist = 1<<(nb+1) + 1 + extra
+		default:
+			if debugDecode {
+				fmt.Println("dist too big:", dist, maxNumDist)
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		// No check on length; encoding can be prescient.
+		if dist > f.dict.histSize() {
+			if debugDecode {
+				fmt.Println("dist > f.dict.histSize():", dist, f.dict.histSize())
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		f.copyLen, f.copyDist = length, dist
+		goto copyHistory
+	}
+
+copyHistory:
+	// Perform a backwards copy according to RFC section 3.2.3.
+	{
+		cnt := f.dict.tryWriteCopy(f.copyDist, f.copyLen)
+		if cnt == 0 {
+			cnt = f.dict.writeCopy(f.copyDist, f.copyLen)
+		}
+		f.copyLen -= cnt
+
+		if f.dict.availWrite() == 0 || f.copyLen > 0 {
+			f.toRead = f.dict.readFlush()
+			f.step = (*decompressor).huffmanBytesReader // We need to continue this work
+			f.stepState = stateDict
+			return
+		}
+		goto readLiteral
+	}
+}
+
+// Decode a single Huffman block from f.
+// hl and hd are the Huffman states for the lit/length values
+// and the distance values, respectively. If hd == nil, using the
+// fixed distance encoding associated with fixed Huffman blocks.
+func (f *decompressor) huffmanBufioReader() {
+	const (
+		stateInit = iota // Zero value must be stateInit
+		stateDict
+	)
+	fr := f.r.(*bufio.Reader)
+	moreBits := func() error {
+		c, err := fr.ReadByte()
+		if err != nil {
+			return noEOF(err)
+		}
+		f.roffset++
+		f.b |= uint32(c) << f.nb
+		f.nb += 8
+		return nil
+	}
+
+	switch f.stepState {
+	case stateInit:
+		goto readLiteral
+	case stateDict:
+		goto copyHistory
+	}
+
+readLiteral:
+	// Read literal and/or (length, distance) according to RFC section 3.2.3.
+	{
+		var v int
+		{
+			// Inlined v, err := f.huffSym(f.hl)
+			// Since a huffmanDecoder can be empty or be composed of a degenerate tree
+			// with single element, huffSym must error on these two edge cases. In both
+			// cases, the chunks slice will be 0 for the invalid sequence, leading it
+			// satisfy the n == 0 check below.
+			n := uint(f.hl.maxRead)
+			// Optimization. Compiler isn't smart enough to keep f.b,f.nb in registers,
+			// but is smart enough to keep local variables in registers, so use nb and b,
+			// inline call to moreBits and reassign b,nb back to f on return.
+			nb, b := f.nb, f.b
+			for {
+				for nb < n {
+					c, err := fr.ReadByte()
+					if err != nil {
+						f.b = b
+						f.nb = nb
+						f.err = noEOF(err)
+						return
+					}
+					f.roffset++
+					b |= uint32(c) << (nb & 31)
+					nb += 8
+				}
+				chunk := f.hl.chunks[b&(huffmanNumChunks-1)]
+				n = uint(chunk & huffmanCountMask)
+				if n > huffmanChunkBits {
+					chunk = f.hl.links[chunk>>huffmanValueShift][(b>>huffmanChunkBits)&f.hl.linkMask]
+					n = uint(chunk & huffmanCountMask)
+				}
+				if n <= nb {
+					if n == 0 {
+						f.b = b
+						f.nb = nb
+						if debugDecode {
+							fmt.Println("huffsym: n==0")
+						}
+						f.err = CorruptInputError(f.roffset)
+						return
+					}
+					f.b = b >> (n & 31)
+					f.nb = nb - n
+					v = int(chunk >> huffmanValueShift)
+					break
+				}
+			}
+		}
+
+		var n uint // number of bits extra
+		var length int
+		var err error
+		switch {
+		case v < 256:
+			f.dict.writeByte(byte(v))
+			if f.dict.availWrite() == 0 {
+				f.toRead = f.dict.readFlush()
+				f.step = (*decompressor).huffmanBufioReader
+				f.stepState = stateInit
+				return
+			}
+			goto readLiteral
+		case v == 256:
+			f.finishBlock()
+			return
+		// otherwise, reference to older data
+		case v < 265:
+			length = v - (257 - 3)
+			n = 0
+		case v < 269:
+			length = v*2 - (265*2 - 11)
+			n = 1
+		case v < 273:
+			length = v*4 - (269*4 - 19)
+			n = 2
+		case v < 277:
+			length = v*8 - (273*8 - 35)
+			n = 3
+		case v < 281:
+			length = v*16 - (277*16 - 67)
+			n = 4
+		case v < 285:
+			length = v*32 - (281*32 - 131)
+			n = 5
+		case v < maxNumLit:
+			length = 258
+			n = 0
+		default:
+			if debugDecode {
+				fmt.Println(v, ">= maxNumLit")
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+		if n > 0 {
+			for f.nb < n {
+				if err = moreBits(); err != nil {
+					if debugDecode {
+						fmt.Println("morebits n>0:", err)
+					}
+					f.err = err
+					return
+				}
+			}
+			length += int(f.b & uint32(1<<n-1))
+			f.b >>= n
+			f.nb -= n
+		}
+
+		var dist int
+		if f.hd == nil {
+			for f.nb < 5 {
+				if err = moreBits(); err != nil {
+					if debugDecode {
+						fmt.Println("morebits f.nb<5:", err)
+					}
+					f.err = err
+					return
+				}
+			}
+			dist = int(bits.Reverse8(uint8(f.b & 0x1F << 3)))
+			f.b >>= 5
+			f.nb -= 5
+		} else {
+			if dist, err = f.huffSym(f.hd); err != nil {
+				if debugDecode {
+					fmt.Println("huffsym:", err)
+				}
+				f.err = err
+				return
+			}
+		}
+
+		switch {
+		case dist < 4:
+			dist++
+		case dist < maxNumDist:
+			nb := uint(dist-2) >> 1
+			// have 1 bit in bottom of dist, need nb more.
+			extra := (dist & 1) << nb
+			for f.nb < nb {
+				if err = moreBits(); err != nil {
+					if debugDecode {
+						fmt.Println("morebits f.nb<nb:", err)
+					}
+					f.err = err
+					return
+				}
+			}
+			extra |= int(f.b & uint32(1<<nb-1))
+			f.b >>= nb
+			f.nb -= nb
+			dist = 1<<(nb+1) + 1 + extra
+		default:
+			if debugDecode {
+				fmt.Println("dist too big:", dist, maxNumDist)
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		// No check on length; encoding can be prescient.
+		if dist > f.dict.histSize() {
+			if debugDecode {
+				fmt.Println("dist > f.dict.histSize():", dist, f.dict.histSize())
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		f.copyLen, f.copyDist = length, dist
+		goto copyHistory
+	}
+
+copyHistory:
+	// Perform a backwards copy according to RFC section 3.2.3.
+	{
+		cnt := f.dict.tryWriteCopy(f.copyDist, f.copyLen)
+		if cnt == 0 {
+			cnt = f.dict.writeCopy(f.copyDist, f.copyLen)
+		}
+		f.copyLen -= cnt
+
+		if f.dict.availWrite() == 0 || f.copyLen > 0 {
+			f.toRead = f.dict.readFlush()
+			f.step = (*decompressor).huffmanBufioReader // We need to continue this work
+			f.stepState = stateDict
+			return
+		}
+		goto readLiteral
+	}
+}
+
+// Decode a single Huffman block from f.
+// hl and hd are the Huffman states for the lit/length values
+// and the distance values, respectively. If hd == nil, using the
+// fixed distance encoding associated with fixed Huffman blocks.
+func (f *decompressor) huffmanStringsReader() {
+	const (
+		stateInit = iota // Zero value must be stateInit
+		stateDict
+	)
+	fr := f.r.(*strings.Reader)
+	moreBits := func() error {
+		c, err := fr.ReadByte()
+		if err != nil {
+			return noEOF(err)
+		}
+		f.roffset++
+		f.b |= uint32(c) << f.nb
+		f.nb += 8
+		return nil
+	}
+
+	switch f.stepState {
+	case stateInit:
+		goto readLiteral
+	case stateDict:
+		goto copyHistory
+	}
+
+readLiteral:
+	// Read literal and/or (length, distance) according to RFC section 3.2.3.
+	{
+		var v int
+		{
+			// Inlined v, err := f.huffSym(f.hl)
+			// Since a huffmanDecoder can be empty or be composed of a degenerate tree
+			// with single element, huffSym must error on these two edge cases. In both
+			// cases, the chunks slice will be 0 for the invalid sequence, leading it
+			// satisfy the n == 0 check below.
+			n := uint(f.hl.maxRead)
+			// Optimization. Compiler isn't smart enough to keep f.b,f.nb in registers,
+			// but is smart enough to keep local variables in registers, so use nb and b,
+			// inline call to moreBits and reassign b,nb back to f on return.
+			nb, b := f.nb, f.b
+			for {
+				for nb < n {
+					c, err := fr.ReadByte()
+					if err != nil {
+						f.b = b
+						f.nb = nb
+						f.err = noEOF(err)
+						return
+					}
+					f.roffset++
+					b |= uint32(c) << (nb & 31)
+					nb += 8
+				}
+				chunk := f.hl.chunks[b&(huffmanNumChunks-1)]
+				n = uint(chunk & huffmanCountMask)
+				if n > huffmanChunkBits {
+					chunk = f.hl.links[chunk>>huffmanValueShift][(b>>huffmanChunkBits)&f.hl.linkMask]
+					n = uint(chunk & huffmanCountMask)
+				}
+				if n <= nb {
+					if n == 0 {
+						f.b = b
+						f.nb = nb
+						if debugDecode {
+							fmt.Println("huffsym: n==0")
+						}
+						f.err = CorruptInputError(f.roffset)
+						return
+					}
+					f.b = b >> (n & 31)
+					f.nb = nb - n
+					v = int(chunk >> huffmanValueShift)
+					break
+				}
+			}
+		}
+
+		var n uint // number of bits extra
+		var length int
+		var err error
+		switch {
+		case v < 256:
+			f.dict.writeByte(byte(v))
+			if f.dict.availWrite() == 0 {
+				f.toRead = f.dict.readFlush()
+				f.step = (*decompressor).huffmanStringsReader
+				f.stepState = stateInit
+				return
+			}
+			goto readLiteral
+		case v == 256:
+			f.finishBlock()
+			return
+		// otherwise, reference to older data
+		case v < 265:
+			length = v - (257 - 3)
+			n = 0
+		case v < 269:
+			length = v*2 - (265*2 - 11)
+			n = 1
+		case v < 273:
+			length = v*4 - (269*4 - 19)
+			n = 2
+		case v < 277:
+			length = v*8 - (273*8 - 35)
+			n = 3
+		case v < 281:
+			length = v*16 - (277*16 - 67)
+			n = 4
+		case v < 285:
+			length = v*32 - (281*32 - 131)
+			n = 5
+		case v < maxNumLit:
+			length = 258
+			n = 0
+		default:
+			if debugDecode {
+				fmt.Println(v, ">= maxNumLit")
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+		if n > 0 {
+			for f.nb < n {
+				if err = moreBits(); err != nil {
+					if debugDecode {
+						fmt.Println("morebits n>0:", err)
+					}
+					f.err = err
+					return
+				}
+			}
+			length += int(f.b & uint32(1<<n-1))
+			f.b >>= n
+			f.nb -= n
+		}
+
+		var dist int
+		if f.hd == nil {
+			for f.nb < 5 {
+				if err = moreBits(); err != nil {
+					if debugDecode {
+						fmt.Println("morebits f.nb<5:", err)
+					}
+					f.err = err
+					return
+				}
+			}
+			dist = int(bits.Reverse8(uint8(f.b & 0x1F << 3)))
+			f.b >>= 5
+			f.nb -= 5
+		} else {
+			if dist, err = f.huffSym(f.hd); err != nil {
+				if debugDecode {
+					fmt.Println("huffsym:", err)
+				}
+				f.err = err
+				return
+			}
+		}
+
+		switch {
+		case dist < 4:
+			dist++
+		case dist < maxNumDist:
+			nb := uint(dist-2) >> 1
+			// have 1 bit in bottom of dist, need nb more.
+			extra := (dist & 1) << nb
+			for f.nb < nb {
+				if err = moreBits(); err != nil {
+					if debugDecode {
+						fmt.Println("morebits f.nb<nb:", err)
+					}
+					f.err = err
+					return
+				}
+			}
+			extra |= int(f.b & uint32(1<<nb-1))
+			f.b >>= nb
+			f.nb -= nb
+			dist = 1<<(nb+1) + 1 + extra
+		default:
+			if debugDecode {
+				fmt.Println("dist too big:", dist, maxNumDist)
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		// No check on length; encoding can be prescient.
+		if dist > f.dict.histSize() {
+			if debugDecode {
+				fmt.Println("dist > f.dict.histSize():", dist, f.dict.histSize())
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		f.copyLen, f.copyDist = length, dist
+		goto copyHistory
+	}
+
+copyHistory:
+	// Perform a backwards copy according to RFC section 3.2.3.
+	{
+		cnt := f.dict.tryWriteCopy(f.copyDist, f.copyLen)
+		if cnt == 0 {
+			cnt = f.dict.writeCopy(f.copyDist, f.copyLen)
+		}
+		f.copyLen -= cnt
+
+		if f.dict.availWrite() == 0 || f.copyLen > 0 {
+			f.toRead = f.dict.readFlush()
+			f.step = (*decompressor).huffmanStringsReader // We need to continue this work
+			f.stepState = stateDict
+			return
+		}
+		goto readLiteral
+	}
+}
+
+func (f *decompressor) huffmanBlockDecoder() func() {
+	switch f.r.(type) {
+	case *bytes.Buffer:
+		return f.huffmanBytesBuffer
+	case *bytes.Reader:
+		return f.huffmanBytesReader
+	case *bufio.Reader:
+		return f.huffmanBufioReader
+	case *strings.Reader:
+		return f.huffmanStringsReader
+	default:
+		return f.huffmanBlockGeneric
+	}
+}

vendor/github.com/klauspost/compress/flate/level1.go

@@ -16,7 +16,7 @@ func (e *fastEncL1) Encode(dst *tokens, src []byte) {
 		inputMargin            = 12 - 1
 		minNonLiteralBlockSize = 1 + 1 + inputMargin
 	)
-	if debugDecode && e.cur < 0 {
+	if debugDeflate && e.cur < 0 {
 		panic(fmt.Sprint("e.cur < 0: ", e.cur))
 	}
 
@@ -81,12 +81,12 @@ func (e *fastEncL1) Encode(dst *tokens, src []byte) {
 			}
 
 			now := load6432(src, nextS)
-			e.table[nextHash] = tableEntry{offset: s + e.cur, val: cv}
+			e.table[nextHash] = tableEntry{offset: s + e.cur}
 			nextHash = hash(uint32(now))
 
 			offset := s - (candidate.offset - e.cur)
-			if offset < maxMatchOffset && cv == candidate.val {
+			if offset < maxMatchOffset && cv == load3232(src, candidate.offset-e.cur) {
-				e.table[nextHash] = tableEntry{offset: nextS + e.cur, val: uint32(now)}
+				e.table[nextHash] = tableEntry{offset: nextS + e.cur}
 				break
 			}
 
@@ -96,11 +96,11 @@ func (e *fastEncL1) Encode(dst *tokens, src []byte) {
 			nextS++
 			candidate = e.table[nextHash]
 			now >>= 8
-			e.table[nextHash] = tableEntry{offset: s + e.cur, val: cv}
+			e.table[nextHash] = tableEntry{offset: s + e.cur}
 
 			offset = s - (candidate.offset - e.cur)
-			if offset < maxMatchOffset && cv == candidate.val {
+			if offset < maxMatchOffset && cv == load3232(src, candidate.offset-e.cur) {
-				e.table[nextHash] = tableEntry{offset: nextS + e.cur, val: uint32(now)}
+				e.table[nextHash] = tableEntry{offset: nextS + e.cur}
 				break
 			}
 			cv = uint32(now)
@@ -139,7 +139,7 @@ func (e *fastEncL1) Encode(dst *tokens, src []byte) {
 				// Index first pair after match end.
 				if int(s+l+4) < len(src) {
 					cv := load3232(src, s)
-					e.table[hash(cv)] = tableEntry{offset: s + e.cur, val: cv}
+					e.table[hash(cv)] = tableEntry{offset: s + e.cur}
 				}
 				goto emitRemainder
 			}
@@ -153,14 +153,14 @@ func (e *fastEncL1) Encode(dst *tokens, src []byte) {
 			x := load6432(src, s-2)
 			o := e.cur + s - 2
 			prevHash := hash(uint32(x))
-			e.table[prevHash] = tableEntry{offset: o, val: uint32(x)}
+			e.table[prevHash] = tableEntry{offset: o}
 			x >>= 16
 			currHash := hash(uint32(x))
 			candidate = e.table[currHash]
-			e.table[currHash] = tableEntry{offset: o + 2, val: uint32(x)}
+			e.table[currHash] = tableEntry{offset: o + 2}
 
 			offset := s - (candidate.offset - e.cur)
-			if offset > maxMatchOffset || uint32(x) != candidate.val {
+			if offset > maxMatchOffset || uint32(x) != load3232(src, candidate.offset-e.cur) {
 				cv = uint32(x >> 8)
 				s++
 				break

vendor/github.com/klauspost/compress/flate/level2.go

@@ -18,7 +18,7 @@ func (e *fastEncL2) Encode(dst *tokens, src []byte) {
 		minNonLiteralBlockSize = 1 + 1 + inputMargin
 	)
 
-	if debugDecode && e.cur < 0 {
+	if debugDeflate && e.cur < 0 {
 		panic(fmt.Sprint("e.cur < 0: ", e.cur))
 	}
 
@@ -83,12 +83,12 @@ func (e *fastEncL2) Encode(dst *tokens, src []byte) {
 			}
 			candidate = e.table[nextHash]
 			now := load6432(src, nextS)
-			e.table[nextHash] = tableEntry{offset: s + e.cur, val: cv}
+			e.table[nextHash] = tableEntry{offset: s + e.cur}
 			nextHash = hash4u(uint32(now), bTableBits)
 
 			offset := s - (candidate.offset - e.cur)
-			if offset < maxMatchOffset && cv == candidate.val {
+			if offset < maxMatchOffset && cv == load3232(src, candidate.offset-e.cur) {
-				e.table[nextHash] = tableEntry{offset: nextS + e.cur, val: uint32(now)}
+				e.table[nextHash] = tableEntry{offset: nextS + e.cur}
 				break
 			}
 
@@ -98,10 +98,10 @@ func (e *fastEncL2) Encode(dst *tokens, src []byte) {
 			nextS++
 			candidate = e.table[nextHash]
 			now >>= 8
-			e.table[nextHash] = tableEntry{offset: s + e.cur, val: cv}
+			e.table[nextHash] = tableEntry{offset: s + e.cur}
 
 			offset = s - (candidate.offset - e.cur)
-			if offset < maxMatchOffset && cv == candidate.val {
+			if offset < maxMatchOffset && cv == load3232(src, candidate.offset-e.cur) {
 				break
 			}
 			cv = uint32(now)
@@ -148,7 +148,7 @@ func (e *fastEncL2) Encode(dst *tokens, src []byte) {
 				// Index first pair after match end.
 				if int(s+l+4) < len(src) {
 					cv := load3232(src, s)
-					e.table[hash4u(cv, bTableBits)] = tableEntry{offset: s + e.cur, val: cv}
+					e.table[hash4u(cv, bTableBits)] = tableEntry{offset: s + e.cur}
 				}
 				goto emitRemainder
 			}
@@ -157,15 +157,15 @@ func (e *fastEncL2) Encode(dst *tokens, src []byte) {
 			for i := s - l + 2; i < s-5; i += 7 {
 				x := load6432(src, int32(i))
 				nextHash := hash4u(uint32(x), bTableBits)
-				e.table[nextHash] = tableEntry{offset: e.cur + i, val: uint32(x)}
+				e.table[nextHash] = tableEntry{offset: e.cur + i}
 				// Skip one
 				x >>= 16
 				nextHash = hash4u(uint32(x), bTableBits)
-				e.table[nextHash] = tableEntry{offset: e.cur + i + 2, val: uint32(x)}
+				e.table[nextHash] = tableEntry{offset: e.cur + i + 2}
 				// Skip one
 				x >>= 16
 				nextHash = hash4u(uint32(x), bTableBits)
-				e.table[nextHash] = tableEntry{offset: e.cur + i + 4, val: uint32(x)}
+				e.table[nextHash] = tableEntry{offset: e.cur + i + 4}
 			}
 
 			// We could immediately start working at s now, but to improve
@@ -178,14 +178,14 @@ func (e *fastEncL2) Encode(dst *tokens, src []byte) {
 			o := e.cur + s - 2
 			prevHash := hash4u(uint32(x), bTableBits)
 			prevHash2 := hash4u(uint32(x>>8), bTableBits)
-			e.table[prevHash] = tableEntry{offset: o, val: uint32(x)}
+			e.table[prevHash] = tableEntry{offset: o}
-			e.table[prevHash2] = tableEntry{offset: o + 1, val: uint32(x >> 8)}
+			e.table[prevHash2] = tableEntry{offset: o + 1}
 			currHash := hash4u(uint32(x>>16), bTableBits)
 			candidate = e.table[currHash]
-			e.table[currHash] = tableEntry{offset: o + 2, val: uint32(x >> 16)}
+			e.table[currHash] = tableEntry{offset: o + 2}
 
 			offset := s - (candidate.offset - e.cur)
-			if offset > maxMatchOffset || uint32(x>>16) != candidate.val {
+			if offset > maxMatchOffset || uint32(x>>16) != load3232(src, candidate.offset-e.cur) {
 				cv = uint32(x >> 24)
 				s++
 				break

vendor/github.com/klauspost/compress/flate/level3.go

@@ -15,7 +15,7 @@ func (e *fastEncL3) Encode(dst *tokens, src []byte) {
 		minNonLiteralBlockSize = 1 + 1 + inputMargin
 	)
 
-	if debugDecode && e.cur < 0 {
+	if debugDeflate && e.cur < 0 {
 		panic(fmt.Sprint("e.cur < 0: ", e.cur))
 	}
 
@@ -81,22 +81,26 @@ func (e *fastEncL3) Encode(dst *tokens, src []byte) {
 			}
 			candidates := e.table[nextHash]
 			now := load3232(src, nextS)
-			e.table[nextHash] = tableEntryPrev{Prev: candidates.Cur, Cur: tableEntry{offset: s + e.cur, val: cv}}
+
+			// Safe offset distance until s + 4...
+			minOffset := e.cur + s - (maxMatchOffset - 4)
+			e.table[nextHash] = tableEntryPrev{Prev: candidates.Cur, Cur: tableEntry{offset: s + e.cur}}
 
 			// Check both candidates
 			candidate = candidates.Cur
-			offset := s - (candidate.offset - e.cur)
+			if candidate.offset < minOffset {
-			if cv == candidate.val {
+				cv = now
-				if offset > maxMatchOffset {
+				// Previous will also be invalid, we have nothing.
-					cv = now
+				continue
-					// Previous will also be invalid, we have nothing.
+			}
-					continue
+
-				}
+			if cv == load3232(src, candidate.offset-e.cur) {
-				o2 := s - (candidates.Prev.offset - e.cur)
+				if candidates.Prev.offset < minOffset || cv != load3232(src, candidates.Prev.offset-e.cur) {
-				if cv != candidates.Prev.val || o2 > maxMatchOffset {
 					break
 				}
 				// Both match and are valid, pick longest.
+				offset := s - (candidate.offset - e.cur)
+				o2 := s - (candidates.Prev.offset - e.cur)
 				l1, l2 := matchLen(src[s+4:], src[s-offset+4:]), matchLen(src[s+4:], src[s-o2+4:])
 				if l2 > l1 {
 					candidate = candidates.Prev
@@ -106,11 +110,8 @@ func (e *fastEncL3) Encode(dst *tokens, src []byte) {
 				// We only check if value mismatches.
 				// Offset will always be invalid in other cases.
 				candidate = candidates.Prev
-				if cv == candidate.val {
+				if candidate.offset > minOffset && cv == load3232(src, candidate.offset-e.cur) {
-					offset := s - (candidate.offset - e.cur)
+					break
-					if offset <= maxMatchOffset {
-						break
-					}
 				}
 			}
 			cv = now
@@ -158,7 +159,7 @@ func (e *fastEncL3) Encode(dst *tokens, src []byte) {
 					nextHash := hash(cv)
 					e.table[nextHash] = tableEntryPrev{
 						Prev: e.table[nextHash].Cur,
-						Cur:  tableEntry{offset: e.cur + t, val: cv},
+						Cur:  tableEntry{offset: e.cur + t},
 					}
 				}
 				goto emitRemainder
@@ -170,21 +171,21 @@ func (e *fastEncL3) Encode(dst *tokens, src []byte) {
 			prevHash := hash(uint32(x))
 			e.table[prevHash] = tableEntryPrev{
 				Prev: e.table[prevHash].Cur,
-				Cur:  tableEntry{offset: e.cur + s - 3, val: uint32(x)},
+				Cur:  tableEntry{offset: e.cur + s - 3},
 			}
 			x >>= 8
 			prevHash = hash(uint32(x))
 
 			e.table[prevHash] = tableEntryPrev{
 				Prev: e.table[prevHash].Cur,
-				Cur:  tableEntry{offset: e.cur + s - 2, val: uint32(x)},
+				Cur:  tableEntry{offset: e.cur + s - 2},
 			}
 			x >>= 8
 			prevHash = hash(uint32(x))
 
 			e.table[prevHash] = tableEntryPrev{
 				Prev: e.table[prevHash].Cur,
-				Cur:  tableEntry{offset: e.cur + s - 1, val: uint32(x)},
+				Cur:  tableEntry{offset: e.cur + s - 1},
 			}
 			x >>= 8
 			currHash := hash(uint32(x))
@@ -192,21 +193,18 @@ func (e *fastEncL3) Encode(dst *tokens, src []byte) {
 			cv = uint32(x)
 			e.table[currHash] = tableEntryPrev{
 				Prev: candidates.Cur,
-				Cur:  tableEntry{offset: s + e.cur, val: cv},
+				Cur:  tableEntry{offset: s + e.cur},
 			}
 
 			// Check both candidates
 			candidate = candidates.Cur
-			if cv == candidate.val {
+			minOffset := e.cur + s - (maxMatchOffset - 4)
-				offset := s - (candidate.offset - e.cur)
+
-				if offset <= maxMatchOffset {
+			if candidate.offset > minOffset && cv != load3232(src, candidate.offset-e.cur) {
-					continue
-				}
-			} else {
 				// We only check if value mismatches.
 				// Offset will always be invalid in other cases.
 				candidate = candidates.Prev
-				if cv == candidate.val {
+				if candidate.offset > minOffset && cv == load3232(src, candidate.offset-e.cur) {
 					offset := s - (candidate.offset - e.cur)
 					if offset <= maxMatchOffset {
 						continue

vendor/github.com/klauspost/compress/flate/level4.go

@@ -13,7 +13,7 @@ func (e *fastEncL4) Encode(dst *tokens, src []byte) {
 		inputMargin            = 12 - 1
 		minNonLiteralBlockSize = 1 + 1 + inputMargin
 	)
-	if debugDecode && e.cur < 0 {
+	if debugDeflate && e.cur < 0 {
 		panic(fmt.Sprint("e.cur < 0: ", e.cur))
 	}
 	// Protect against e.cur wraparound.
@@ -92,24 +92,24 @@ func (e *fastEncL4) Encode(dst *tokens, src []byte) {
 			sCandidate := e.table[nextHashS]
 			lCandidate := e.bTable[nextHashL]
 			next := load6432(src, nextS)
-			entry := tableEntry{offset: s + e.cur, val: uint32(cv)}
+			entry := tableEntry{offset: s + e.cur}
 			e.table[nextHashS] = entry
 			e.bTable[nextHashL] = entry
 
 			t = lCandidate.offset - e.cur
-			if s-t < maxMatchOffset && uint32(cv) == lCandidate.val {
+			if s-t < maxMatchOffset && uint32(cv) == load3232(src, lCandidate.offset-e.cur) {
 				// We got a long match. Use that.
 				break
 			}
 
 			t = sCandidate.offset - e.cur
-			if s-t < maxMatchOffset && uint32(cv) == sCandidate.val {
+			if s-t < maxMatchOffset && uint32(cv) == load3232(src, sCandidate.offset-e.cur) {
 				// Found a 4 match...
 				lCandidate = e.bTable[hash7(next, tableBits)]
 
 				// If the next long is a candidate, check if we should use that instead...
 				lOff := nextS - (lCandidate.offset - e.cur)
-				if lOff < maxMatchOffset && lCandidate.val == uint32(next) {
+				if lOff < maxMatchOffset && load3232(src, lCandidate.offset-e.cur) == uint32(next) {
 					l1, l2 := matchLen(src[s+4:], src[t+4:]), matchLen(src[nextS+4:], src[nextS-lOff+4:])
 					if l2 > l1 {
 						s = nextS
@@ -137,7 +137,7 @@ func (e *fastEncL4) Encode(dst *tokens, src []byte) {
 		if nextEmit < s {
 			emitLiteral(dst, src[nextEmit:s])
 		}
-		if false {
+		if debugDeflate {
 			if t >= s {
 				panic("s-t")
 			}
@@ -160,8 +160,8 @@ func (e *fastEncL4) Encode(dst *tokens, src []byte) {
 			// Index first pair after match end.
 			if int(s+8) < len(src) {
 				cv := load6432(src, s)
-				e.table[hash4x64(cv, tableBits)] = tableEntry{offset: s + e.cur, val: uint32(cv)}
+				e.table[hash4x64(cv, tableBits)] = tableEntry{offset: s + e.cur}
-				e.bTable[hash7(cv, tableBits)] = tableEntry{offset: s + e.cur, val: uint32(cv)}
+				e.bTable[hash7(cv, tableBits)] = tableEntry{offset: s + e.cur}
 			}
 			goto emitRemainder
 		}
@@ -171,20 +171,20 @@ func (e *fastEncL4) Encode(dst *tokens, src []byte) {
 			i := nextS
 			if i < s-1 {
 				cv := load6432(src, i)
-				t := tableEntry{offset: i + e.cur, val: uint32(cv)}
+				t := tableEntry{offset: i + e.cur}
-				t2 := tableEntry{val: uint32(cv >> 8), offset: t.offset + 1}
+				t2 := tableEntry{offset: t.offset + 1}
 				e.bTable[hash7(cv, tableBits)] = t
 				e.bTable[hash7(cv>>8, tableBits)] = t2
-				e.table[hash4u(t2.val, tableBits)] = t2
+				e.table[hash4u(uint32(cv>>8), tableBits)] = t2
 
 				i += 3
 				for ; i < s-1; i += 3 {
 					cv := load6432(src, i)
-					t := tableEntry{offset: i + e.cur, val: uint32(cv)}
+					t := tableEntry{offset: i + e.cur}
-					t2 := tableEntry{val: uint32(cv >> 8), offset: t.offset + 1}
+					t2 := tableEntry{offset: t.offset + 1}
 					e.bTable[hash7(cv, tableBits)] = t
 					e.bTable[hash7(cv>>8, tableBits)] = t2
-					e.table[hash4u(t2.val, tableBits)] = t2
+					e.table[hash4u(uint32(cv>>8), tableBits)] = t2
 				}
 			}
 		}
@@ -195,8 +195,8 @@ func (e *fastEncL4) Encode(dst *tokens, src []byte) {
 		o := e.cur + s - 1
 		prevHashS := hash4x64(x, tableBits)
 		prevHashL := hash7(x, tableBits)
-		e.table[prevHashS] = tableEntry{offset: o, val: uint32(x)}
+		e.table[prevHashS] = tableEntry{offset: o}
-		e.bTable[prevHashL] = tableEntry{offset: o, val: uint32(x)}
+		e.bTable[prevHashL] = tableEntry{offset: o}
 		cv = x >> 8
 	}
 

vendor/github.com/klauspost/compress/flate/level5.go

@@ -13,7 +13,7 @@ func (e *fastEncL5) Encode(dst *tokens, src []byte) {
 		inputMargin            = 12 - 1
 		minNonLiteralBlockSize = 1 + 1 + inputMargin
 	)
-	if debugDecode && e.cur < 0 {
+	if debugDeflate && e.cur < 0 {
 		panic(fmt.Sprint("e.cur < 0: ", e.cur))
 	}
 
@@ -100,7 +100,7 @@ func (e *fastEncL5) Encode(dst *tokens, src []byte) {
 			sCandidate := e.table[nextHashS]
 			lCandidate := e.bTable[nextHashL]
 			next := load6432(src, nextS)
-			entry := tableEntry{offset: s + e.cur, val: uint32(cv)}
+			entry := tableEntry{offset: s + e.cur}
 			e.table[nextHashS] = entry
 			eLong := &e.bTable[nextHashL]
 			eLong.Cur, eLong.Prev = entry, eLong.Cur
@@ -110,14 +110,14 @@ func (e *fastEncL5) Encode(dst *tokens, src []byte) {
 
 			t = lCandidate.Cur.offset - e.cur
 			if s-t < maxMatchOffset {
-				if uint32(cv) == lCandidate.Cur.val {
+				if uint32(cv) == load3232(src, lCandidate.Cur.offset-e.cur) {
 					// Store the next match
-					e.table[nextHashS] = tableEntry{offset: nextS + e.cur, val: uint32(next)}
+					e.table[nextHashS] = tableEntry{offset: nextS + e.cur}
 					eLong := &e.bTable[nextHashL]
-					eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur, val: uint32(next)}, eLong.Cur
+					eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur}, eLong.Cur
 
 					t2 := lCandidate.Prev.offset - e.cur
-					if s-t2 < maxMatchOffset && uint32(cv) == lCandidate.Prev.val {
+					if s-t2 < maxMatchOffset && uint32(cv) == load3232(src, lCandidate.Prev.offset-e.cur) {
 						l = e.matchlen(s+4, t+4, src) + 4
 						ml1 := e.matchlen(s+4, t2+4, src) + 4
 						if ml1 > l {
@@ -129,30 +129,30 @@ func (e *fastEncL5) Encode(dst *tokens, src []byte) {
 					break
 				}
 				t = lCandidate.Prev.offset - e.cur
-				if s-t < maxMatchOffset && uint32(cv) == lCandidate.Prev.val {
+				if s-t < maxMatchOffset && uint32(cv) == load3232(src, lCandidate.Prev.offset-e.cur) {
 					// Store the next match
-					e.table[nextHashS] = tableEntry{offset: nextS + e.cur, val: uint32(next)}
+					e.table[nextHashS] = tableEntry{offset: nextS + e.cur}
 					eLong := &e.bTable[nextHashL]
-					eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur, val: uint32(next)}, eLong.Cur
+					eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur}, eLong.Cur
 					break
 				}
 			}
 
 			t = sCandidate.offset - e.cur
-			if s-t < maxMatchOffset && uint32(cv) == sCandidate.val {
+			if s-t < maxMatchOffset && uint32(cv) == load3232(src, sCandidate.offset-e.cur) {
 				// Found a 4 match...
 				l = e.matchlen(s+4, t+4, src) + 4
 				lCandidate = e.bTable[nextHashL]
 				// Store the next match
 
-				e.table[nextHashS] = tableEntry{offset: nextS + e.cur, val: uint32(next)}
+				e.table[nextHashS] = tableEntry{offset: nextS + e.cur}
 				eLong := &e.bTable[nextHashL]
-				eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur, val: uint32(next)}, eLong.Cur
+				eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur}, eLong.Cur
 
 				// If the next long is a candidate, use that...
 				t2 := lCandidate.Cur.offset - e.cur
 				if nextS-t2 < maxMatchOffset {
-					if lCandidate.Cur.val == uint32(next) {
+					if load3232(src, lCandidate.Cur.offset-e.cur) == uint32(next) {
 						ml := e.matchlen(nextS+4, t2+4, src) + 4
 						if ml > l {
 							t = t2
@@ -163,7 +163,7 @@ func (e *fastEncL5) Encode(dst *tokens, src []byte) {
 					}
 					// If the previous long is a candidate, use that...
 					t2 = lCandidate.Prev.offset - e.cur
-					if nextS-t2 < maxMatchOffset && lCandidate.Prev.val == uint32(next) {
+					if nextS-t2 < maxMatchOffset && load3232(src, lCandidate.Prev.offset-e.cur) == uint32(next) {
 						ml := e.matchlen(nextS+4, t2+4, src) + 4
 						if ml > l {
 							t = t2
@@ -197,7 +197,7 @@ func (e *fastEncL5) Encode(dst *tokens, src []byte) {
 		if nextEmit < s {
 			emitLiteral(dst, src[nextEmit:s])
 		}
-		if false {
+		if debugDeflate {
 			if t >= s {
 				panic(fmt.Sprintln("s-t", s, t))
 			}
@@ -226,31 +226,31 @@ func (e *fastEncL5) Encode(dst *tokens, src []byte) {
 			i := s - l + 1
 			if i < s-1 {
 				cv := load6432(src, i)
-				t := tableEntry{offset: i + e.cur, val: uint32(cv)}
+				t := tableEntry{offset: i + e.cur}
 				e.table[hash4x64(cv, tableBits)] = t
 				eLong := &e.bTable[hash7(cv, tableBits)]
 				eLong.Cur, eLong.Prev = t, eLong.Cur
 
 				// Do an long at i+1
 				cv >>= 8
-				t = tableEntry{offset: t.offset + 1, val: uint32(cv)}
+				t = tableEntry{offset: t.offset + 1}
 				eLong = &e.bTable[hash7(cv, tableBits)]
 				eLong.Cur, eLong.Prev = t, eLong.Cur
 
 				// We only have enough bits for a short entry at i+2
 				cv >>= 8
-				t = tableEntry{offset: t.offset + 1, val: uint32(cv)}
+				t = tableEntry{offset: t.offset + 1}
 				e.table[hash4x64(cv, tableBits)] = t
 
 				// Skip one - otherwise we risk hitting 's'
 				i += 4
 				for ; i < s-1; i += hashEvery {
 					cv := load6432(src, i)
-					t := tableEntry{offset: i + e.cur, val: uint32(cv)}
+					t := tableEntry{offset: i + e.cur}
-					t2 := tableEntry{offset: t.offset + 1, val: uint32(cv >> 8)}
+					t2 := tableEntry{offset: t.offset + 1}
 					eLong := &e.bTable[hash7(cv, tableBits)]
 					eLong.Cur, eLong.Prev = t, eLong.Cur
-					e.table[hash4u(t2.val, tableBits)] = t2
+					e.table[hash4u(uint32(cv>>8), tableBits)] = t2
 				}
 			}
 		}
@@ -261,9 +261,9 @@ func (e *fastEncL5) Encode(dst *tokens, src []byte) {
 		o := e.cur + s - 1
 		prevHashS := hash4x64(x, tableBits)
 		prevHashL := hash7(x, tableBits)
-		e.table[prevHashS] = tableEntry{offset: o, val: uint32(x)}
+		e.table[prevHashS] = tableEntry{offset: o}
 		eLong := &e.bTable[prevHashL]
-		eLong.Cur, eLong.Prev = tableEntry{offset: o, val: uint32(x)}, eLong.Cur
+		eLong.Cur, eLong.Prev = tableEntry{offset: o}, eLong.Cur
 		cv = x >> 8
 	}
 

vendor/github.com/klauspost/compress/flate/level6.go

@@ -13,7 +13,7 @@ func (e *fastEncL6) Encode(dst *tokens, src []byte) {
 		inputMargin            = 12 - 1
 		minNonLiteralBlockSize = 1 + 1 + inputMargin
 	)
-	if debugDecode && e.cur < 0 {
+	if debugDeflate && e.cur < 0 {
 		panic(fmt.Sprint("e.cur < 0: ", e.cur))
 	}
 
@@ -101,7 +101,7 @@ func (e *fastEncL6) Encode(dst *tokens, src []byte) {
 			sCandidate := e.table[nextHashS]
 			lCandidate := e.bTable[nextHashL]
 			next := load6432(src, nextS)
-			entry := tableEntry{offset: s + e.cur, val: uint32(cv)}
+			entry := tableEntry{offset: s + e.cur}
 			e.table[nextHashS] = entry
 			eLong := &e.bTable[nextHashL]
 			eLong.Cur, eLong.Prev = entry, eLong.Cur
@@ -112,17 +112,17 @@ func (e *fastEncL6) Encode(dst *tokens, src []byte) {
 
 			t = lCandidate.Cur.offset - e.cur
 			if s-t < maxMatchOffset {
-				if uint32(cv) == lCandidate.Cur.val {
+				if uint32(cv) == load3232(src, lCandidate.Cur.offset-e.cur) {
 					// Long candidate matches at least 4 bytes.
 
 					// Store the next match
-					e.table[nextHashS] = tableEntry{offset: nextS + e.cur, val: uint32(next)}
+					e.table[nextHashS] = tableEntry{offset: nextS + e.cur}
 					eLong := &e.bTable[nextHashL]
-					eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur, val: uint32(next)}, eLong.Cur
+					eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur}, eLong.Cur
 
 					// Check the previous long candidate as well.
 					t2 := lCandidate.Prev.offset - e.cur
-					if s-t2 < maxMatchOffset && uint32(cv) == lCandidate.Prev.val {
+					if s-t2 < maxMatchOffset && uint32(cv) == load3232(src, lCandidate.Prev.offset-e.cur) {
 						l = e.matchlen(s+4, t+4, src) + 4
 						ml1 := e.matchlen(s+4, t2+4, src) + 4
 						if ml1 > l {
@@ -135,17 +135,17 @@ func (e *fastEncL6) Encode(dst *tokens, src []byte) {
 				}
 				// Current value did not match, but check if previous long value does.
 				t = lCandidate.Prev.offset - e.cur
-				if s-t < maxMatchOffset && uint32(cv) == lCandidate.Prev.val {
+				if s-t < maxMatchOffset && uint32(cv) == load3232(src, lCandidate.Prev.offset-e.cur) {
 					// Store the next match
-					e.table[nextHashS] = tableEntry{offset: nextS + e.cur, val: uint32(next)}
+					e.table[nextHashS] = tableEntry{offset: nextS + e.cur}
 					eLong := &e.bTable[nextHashL]
-					eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur, val: uint32(next)}, eLong.Cur
+					eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur}, eLong.Cur
 					break
 				}
 			}
 
 			t = sCandidate.offset - e.cur
-			if s-t < maxMatchOffset && uint32(cv) == sCandidate.val {
+			if s-t < maxMatchOffset && uint32(cv) == load3232(src, sCandidate.offset-e.cur) {
 				// Found a 4 match...
 				l = e.matchlen(s+4, t+4, src) + 4
 
@@ -153,9 +153,9 @@ func (e *fastEncL6) Encode(dst *tokens, src []byte) {
 				lCandidate = e.bTable[nextHashL]
 
 				// Store the next match
-				e.table[nextHashS] = tableEntry{offset: nextS + e.cur, val: uint32(next)}
+				e.table[nextHashS] = tableEntry{offset: nextS + e.cur}
 				eLong := &e.bTable[nextHashL]
-				eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur, val: uint32(next)}, eLong.Cur
+				eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur}, eLong.Cur
 
 				// Check repeat at s + repOff
 				const repOff = 1
@@ -174,7 +174,7 @@ func (e *fastEncL6) Encode(dst *tokens, src []byte) {
 				// If the next long is a candidate, use that...
 				t2 = lCandidate.Cur.offset - e.cur
 				if nextS-t2 < maxMatchOffset {
-					if lCandidate.Cur.val == uint32(next) {
+					if load3232(src, lCandidate.Cur.offset-e.cur) == uint32(next) {
 						ml := e.matchlen(nextS+4, t2+4, src) + 4
 						if ml > l {
 							t = t2
@@ -185,7 +185,7 @@ func (e *fastEncL6) Encode(dst *tokens, src []byte) {
 					}
 					// If the previous long is a candidate, use that...
 					t2 = lCandidate.Prev.offset - e.cur
-					if nextS-t2 < maxMatchOffset && lCandidate.Prev.val == uint32(next) {
+					if nextS-t2 < maxMatchOffset && load3232(src, lCandidate.Prev.offset-e.cur) == uint32(next) {
 						ml := e.matchlen(nextS+4, t2+4, src) + 4
 						if ml > l {
 							t = t2
@@ -244,9 +244,9 @@ func (e *fastEncL6) Encode(dst *tokens, src []byte) {
 			// Index after match end.
 			for i := nextS + 1; i < int32(len(src))-8; i += 2 {
 				cv := load6432(src, i)
-				e.table[hash4x64(cv, tableBits)] = tableEntry{offset: i + e.cur, val: uint32(cv)}
+				e.table[hash4x64(cv, tableBits)] = tableEntry{offset: i + e.cur}
 				eLong := &e.bTable[hash7(cv, tableBits)]
-				eLong.Cur, eLong.Prev = tableEntry{offset: i + e.cur, val: uint32(cv)}, eLong.Cur
+				eLong.Cur, eLong.Prev = tableEntry{offset: i + e.cur}, eLong.Cur
 			}
 			goto emitRemainder
 		}
@@ -255,8 +255,8 @@ func (e *fastEncL6) Encode(dst *tokens, src []byte) {
 		if true {
 			for i := nextS + 1; i < s-1; i += 2 {
 				cv := load6432(src, i)
-				t := tableEntry{offset: i + e.cur, val: uint32(cv)}
+				t := tableEntry{offset: i + e.cur}
-				t2 := tableEntry{offset: t.offset + 1, val: uint32(cv >> 8)}
+				t2 := tableEntry{offset: t.offset + 1}
 				eLong := &e.bTable[hash7(cv, tableBits)]
 				eLong2 := &e.bTable[hash7(cv>>8, tableBits)]
 				e.table[hash4x64(cv, tableBits)] = t

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

@@ -262,7 +262,7 @@ func (t *tokens) EstimatedBits() int {
 // AddMatch adds a match to the tokens.
 // This function is very sensitive to inlining and right on the border.
 func (t *tokens) AddMatch(xlength uint32, xoffset uint32) {
-	if debugDecode {
+	if debugDeflate {
 		if xlength >= maxMatchLength+baseMatchLength {
 			panic(fmt.Errorf("invalid length: %v", xlength))
 		}
@@ -281,7 +281,7 @@ func (t *tokens) AddMatch(xlength uint32, xoffset uint32) {
 // AddMatchLong adds a match to the tokens, potentially longer than max match length.
 // Length should NOT have the base subtracted, only offset should.
 func (t *tokens) AddMatchLong(xlength int32, xoffset uint32) {
-	if debugDecode {
+	if debugDeflate {
 		if xoffset >= maxMatchOffset+baseMatchOffset {
 			panic(fmt.Errorf("invalid offset: %v", xoffset))
 		}

vendor/github.com/klauspost/compress/zstd/decoder.go

@@ -66,7 +66,7 @@ var (
 // A Decoder can be used in two modes:
 //
 // 1) As a stream, or
-// 2) For stateless decoding using DecodeAll or DecodeBuffer.
+// 2) For stateless decoding using DecodeAll.
 //
 // Only a single stream can be decoded concurrently, but the same decoder
 // can run multiple concurrent stateless decodes. It is even possible to
@@ -315,7 +315,7 @@ func (d *Decoder) DecodeAll(input, dst []byte) ([]byte, error) {
 			if size > 1<<20 {
 				size = 1 << 20
 			}
-			dst = make([]byte, 0, frame.WindowSize)
+			dst = make([]byte, 0, size)
 		}
 
 		dst, err = frame.runDecoder(dst, block)

vendor/github.com/klauspost/compress/zstd/enc_better.go

@@ -0,0 +1,521 @@
+// 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 "fmt"
+
+const (
+	betterLongTableBits = 19                       // Bits used in the long match table
+	betterLongTableSize = 1 << betterLongTableBits // Size of the table
+
+	// Note: Increasing the short table bits or making the hash shorter
+	// can actually lead to compression degradation since it will 'steal' more from the
+	// long match table and match offsets are quite big.
+	// This greatly depends on the type of input.
+	betterShortTableBits = 13                        // Bits used in the short match table
+	betterShortTableSize = 1 << betterShortTableBits // Size of the table
+)
+
+type prevEntry struct {
+	offset int32
+	prev   int32
+}
+
+// betterFastEncoder uses 2 tables, one for short matches (5 bytes) and one for long matches.
+// The long match table contains the previous entry with the same hash,
+// effectively making it a "chain" of length 2.
+// When we find a long match we choose between the two values and select the longest.
+// When we find a short match, after checking the long, we check if we can find a long at n+1
+// and that it is longer (lazy matching).
+type betterFastEncoder struct {
+	fastBase
+	table     [betterShortTableSize]tableEntry
+	longTable [betterLongTableSize]prevEntry
+}
+
+// Encode improves compression...
+func (e *betterFastEncoder) Encode(blk *blockEnc, src []byte) {
+	const (
+		// Input margin is the number of bytes we read (8)
+		// and the maximum we will read ahead (2)
+		inputMargin            = 8 + 2
+		minNonLiteralBlockSize = 16
+	)
+
+	// Protect against e.cur wraparound.
+	for e.cur >= bufferReset {
+		if len(e.hist) == 0 {
+			for i := range e.table[:] {
+				e.table[i] = tableEntry{}
+			}
+			for i := range e.longTable[:] {
+				e.longTable[i] = prevEntry{}
+			}
+			e.cur = e.maxMatchOff
+			break
+		}
+		// Shift down everything in the table that isn't already too far away.
+		minOff := e.cur + int32(len(e.hist)) - e.maxMatchOff
+		for i := range e.table[:] {
+			v := e.table[i].offset
+			if v < minOff {
+				v = 0
+			} else {
+				v = v - e.cur + e.maxMatchOff
+			}
+			e.table[i].offset = v
+		}
+		for i := range e.longTable[:] {
+			v := e.longTable[i].offset
+			v2 := e.longTable[i].prev
+			if v < minOff {
+				v = 0
+				v2 = 0
+			} else {
+				v = v - e.cur + e.maxMatchOff
+				if v2 < minOff {
+					v2 = 0
+				} else {
+					v2 = v2 - e.cur + e.maxMatchOff
+				}
+			}
+			e.longTable[i] = prevEntry{
+				offset: v,
+				prev:   v2,
+			}
+		}
+		e.cur = e.maxMatchOff
+		break
+	}
+
+	s := e.addBlock(src)
+	blk.size = len(src)
+	if len(src) < minNonLiteralBlockSize {
+		blk.extraLits = len(src)
+		blk.literals = blk.literals[:len(src)]
+		copy(blk.literals, src)
+		return
+	}
+
+	// Override src
+	src = e.hist
+	sLimit := int32(len(src)) - inputMargin
+	// stepSize is the number of bytes to skip on every main loop iteration.
+	// It should be >= 1.
+	stepSize := int32(e.o.targetLength)
+	if stepSize == 0 {
+		stepSize++
+	}
+
+	const kSearchStrength = 9
+
+	// nextEmit is where in src the next emitLiteral should start from.
+	nextEmit := s
+	cv := load6432(src, s)
+
+	// Relative offsets
+	offset1 := int32(blk.recentOffsets[0])
+	offset2 := int32(blk.recentOffsets[1])
+
+	addLiterals := func(s *seq, until int32) {
+		if until == nextEmit {
+			return
+		}
+		blk.literals = append(blk.literals, src[nextEmit:until]...)
+		s.litLen = uint32(until - nextEmit)
+	}
+	if debug {
+		println("recent offsets:", blk.recentOffsets)
+	}
+
+encodeLoop:
+	for {
+		var t int32
+		// We allow the encoder to optionally turn off repeat offsets across blocks
+		canRepeat := len(blk.sequences) > 2
+		var matched int32
+
+		for {
+			if debugAsserts && canRepeat && offset1 == 0 {
+				panic("offset0 was 0")
+			}
+
+			nextHashS := hash5(cv, betterShortTableBits)
+			nextHashL := hash8(cv, betterLongTableBits)
+			candidateL := e.longTable[nextHashL]
+			candidateS := e.table[nextHashS]
+
+			const repOff = 1
+			repIndex := s - offset1 + repOff
+			off := s + e.cur
+			e.longTable[nextHashL] = prevEntry{offset: off, prev: candidateL.offset}
+			e.table[nextHashS] = tableEntry{offset: off, val: uint32(cv)}
+
+			if canRepeat {
+				if repIndex >= 0 && load3232(src, repIndex) == uint32(cv>>(repOff*8)) {
+					// Consider history as well.
+					var seq seq
+					lenght := 4 + e.matchlen(s+4+repOff, repIndex+4, src)
+
+					seq.matchLen = uint32(lenght - zstdMinMatch)
+
+					// We might be able to match backwards.
+					// Extend as long as we can.
+					start := s + repOff
+					// We end the search early, so we don't risk 0 literals
+					// and have to do special offset treatment.
+					startLimit := nextEmit + 1
+
+					tMin := s - e.maxMatchOff
+					if tMin < 0 {
+						tMin = 0
+					}
+					for repIndex > tMin && start > startLimit && src[repIndex-1] == src[start-1] && seq.matchLen < maxMatchLength-zstdMinMatch-1 {
+						repIndex--
+						start--
+						seq.matchLen++
+					}
+					addLiterals(&seq, start)
+
+					// rep 0
+					seq.offset = 1
+					if debugSequences {
+						println("repeat sequence", seq, "next s:", s)
+					}
+					blk.sequences = append(blk.sequences, seq)
+
+					// Index match start+1 (long) -> s - 1
+					index0 := s + repOff
+					s += lenght + repOff
+
+					nextEmit = s
+					if s >= sLimit {
+						if debug {
+							println("repeat ended", s, lenght)
+
+						}
+						break encodeLoop
+					}
+					// Index skipped...
+					for index0 < s-1 {
+						cv0 := load6432(src, index0)
+						cv1 := cv0 >> 8
+						h0 := hash8(cv0, betterLongTableBits)
+						off := index0 + e.cur
+						e.longTable[h0] = prevEntry{offset: off, prev: e.longTable[h0].offset}
+						e.table[hash5(cv1, betterShortTableBits)] = tableEntry{offset: off + 1, val: uint32(cv1)}
+						index0 += 2
+					}
+					cv = load6432(src, s)
+					continue
+				}
+				const repOff2 = 1
+
+				// We deviate from the reference encoder and also check offset 2.
+				// Still slower and not much better, so disabled.
+				// repIndex = s - offset2 + repOff2
+				if false && repIndex >= 0 && load6432(src, repIndex) == load6432(src, s+repOff) {
+					// Consider history as well.
+					var seq seq
+					lenght := 8 + e.matchlen(s+8+repOff2, repIndex+8, src)
+
+					seq.matchLen = uint32(lenght - zstdMinMatch)
+
+					// We might be able to match backwards.
+					// Extend as long as we can.
+					start := s + repOff2
+					// We end the search early, so we don't risk 0 literals
+					// and have to do special offset treatment.
+					startLimit := nextEmit + 1
+
+					tMin := s - e.maxMatchOff
+					if tMin < 0 {
+						tMin = 0
+					}
+					for repIndex > tMin && start > startLimit && src[repIndex-1] == src[start-1] && seq.matchLen < maxMatchLength-zstdMinMatch-1 {
+						repIndex--
+						start--
+						seq.matchLen++
+					}
+					addLiterals(&seq, start)
+
+					// rep 2
+					seq.offset = 2
+					if debugSequences {
+						println("repeat sequence 2", seq, "next s:", s)
+					}
+					blk.sequences = append(blk.sequences, seq)
+
+					index0 := s + repOff2
+					s += lenght + repOff2
+					nextEmit = s
+					if s >= sLimit {
+						if debug {
+							println("repeat ended", s, lenght)
+
+						}
+						break encodeLoop
+					}
+
+					// Index skipped...
+					for index0 < s-1 {
+						cv0 := load6432(src, index0)
+						cv1 := cv0 >> 8
+						h0 := hash8(cv0, betterLongTableBits)
+						off := index0 + e.cur
+						e.longTable[h0] = prevEntry{offset: off, prev: e.longTable[h0].offset}
+						e.table[hash5(cv1, betterShortTableBits)] = tableEntry{offset: off + 1, val: uint32(cv1)}
+						index0 += 2
+					}
+					cv = load6432(src, s)
+					// Swap offsets
+					offset1, offset2 = offset2, offset1
+					continue
+				}
+			}
+			// Find the offsets of our two matches.
+			coffsetL := candidateL.offset - e.cur
+			coffsetLP := candidateL.prev - e.cur
+
+			// Check if we have a long match.
+			if s-coffsetL < e.maxMatchOff && cv == load6432(src, coffsetL) {
+				// Found a long match, at least 8 bytes.
+				matched = e.matchlen(s+8, coffsetL+8, src) + 8
+				t = coffsetL
+				if debugAsserts && s <= t {
+					panic(fmt.Sprintf("s (%d) <= t (%d)", s, t))
+				}
+				if debugAsserts && s-t > e.maxMatchOff {
+					panic("s - t >e.maxMatchOff")
+				}
+				if debugMatches {
+					println("long match")
+				}
+
+				if s-coffsetLP < e.maxMatchOff && cv == load6432(src, coffsetLP) {
+					// Found a long match, at least 8 bytes.
+					prevMatch := e.matchlen(s+8, coffsetLP+8, src) + 8
+					if prevMatch > matched {
+						matched = prevMatch
+						t = coffsetLP
+					}
+					if debugAsserts && s <= t {
+						panic(fmt.Sprintf("s (%d) <= t (%d)", s, t))
+					}
+					if debugAsserts && s-t > e.maxMatchOff {
+						panic("s - t >e.maxMatchOff")
+					}
+					if debugMatches {
+						println("long match")
+					}
+				}
+				break
+			}
+
+			// Check if we have a long match on prev.
+			if s-coffsetLP < e.maxMatchOff && cv == load6432(src, coffsetLP) {
+				// Found a long match, at least 8 bytes.
+				matched = e.matchlen(s+8, coffsetLP+8, src) + 8
+				t = coffsetLP
+				if debugAsserts && s <= t {
+					panic(fmt.Sprintf("s (%d) <= t (%d)", s, t))
+				}
+				if debugAsserts && s-t > e.maxMatchOff {
+					panic("s - t >e.maxMatchOff")
+				}
+				if debugMatches {
+					println("long match")
+				}
+				break
+			}
+
+			coffsetS := candidateS.offset - e.cur
+
+			// Check if we have a short match.
+			if s-coffsetS < e.maxMatchOff && uint32(cv) == candidateS.val {
+				// found a regular match
+				matched = e.matchlen(s+4, coffsetS+4, src) + 4
+
+				// See if we can find a long match at s+1
+				const checkAt = 1
+				cv := load6432(src, s+checkAt)
+				nextHashL = hash8(cv, betterLongTableBits)
+				candidateL = e.longTable[nextHashL]
+				coffsetL = candidateL.offset - e.cur
+
+				// We can store it, since we have at least a 4 byte match.
+				e.longTable[nextHashL] = prevEntry{offset: s + checkAt + e.cur, prev: candidateL.offset}
+				if s-coffsetL < e.maxMatchOff && cv == load6432(src, coffsetL) {
+					// Found a long match, at least 8 bytes.
+					matchedNext := e.matchlen(s+8+checkAt, coffsetL+8, src) + 8
+					if matchedNext > matched {
+						t = coffsetL
+						s += checkAt
+						matched = matchedNext
+						if debugMatches {
+							println("long match (after short)")
+						}
+						break
+					}
+				}
+
+				// Check prev long...
+				coffsetL = candidateL.prev - e.cur
+				if s-coffsetL < e.maxMatchOff && cv == load6432(src, coffsetL) {
+					// Found a long match, at least 8 bytes.
+					matchedNext := e.matchlen(s+8+checkAt, coffsetL+8, src) + 8
+					if matchedNext > matched {
+						t = coffsetL
+						s += checkAt
+						matched = matchedNext
+						if debugMatches {
+							println("prev long match (after short)")
+						}
+						break
+					}
+				}
+				t = coffsetS
+				if debugAsserts && s <= t {
+					panic(fmt.Sprintf("s (%d) <= t (%d)", s, t))
+				}
+				if debugAsserts && s-t > e.maxMatchOff {
+					panic("s - t >e.maxMatchOff")
+				}
+				if debugAsserts && t < 0 {
+					panic("t<0")
+				}
+				if debugMatches {
+					println("short match")
+				}
+				break
+			}
+
+			// No match found, move forward in input.
+			s += stepSize + ((s - nextEmit) >> (kSearchStrength - 1))
+			if s >= sLimit {
+				break encodeLoop
+			}
+			cv = load6432(src, s)
+		}
+
+		// A 4-byte match has been found. Update recent offsets.
+		// We'll later see if more than 4 bytes.
+		offset2 = offset1
+		offset1 = s - t
+
+		if debugAsserts && s <= t {
+			panic(fmt.Sprintf("s (%d) <= t (%d)", s, t))
+		}
+
+		if debugAsserts && canRepeat && int(offset1) > len(src) {
+			panic("invalid offset")
+		}
+
+		// Extend the n-byte match as long as possible.
+		l := matched
+
+		// Extend backwards
+		tMin := s - e.maxMatchOff
+		if tMin < 0 {
+			tMin = 0
+		}
+		for t > tMin && s > nextEmit && src[t-1] == src[s-1] && l < maxMatchLength {
+			s--
+			t--
+			l++
+		}
+
+		// Write our sequence
+		var seq seq
+		seq.litLen = uint32(s - nextEmit)
+		seq.matchLen = uint32(l - zstdMinMatch)
+		if seq.litLen > 0 {
+			blk.literals = append(blk.literals, src[nextEmit:s]...)
+		}
+		seq.offset = uint32(s-t) + 3
+		s += l
+		if debugSequences {
+			println("sequence", seq, "next s:", s)
+		}
+		blk.sequences = append(blk.sequences, seq)
+		nextEmit = s
+		if s >= sLimit {
+			break encodeLoop
+		}
+
+		// Index match start+1 (long) -> s - 1
+		index0 := s - l + 1
+		for index0 < s-1 {
+			cv0 := load6432(src, index0)
+			cv1 := cv0 >> 8
+			h0 := hash8(cv0, betterLongTableBits)
+			off := index0 + e.cur
+			e.longTable[h0] = prevEntry{offset: off, prev: e.longTable[h0].offset}
+			e.table[hash5(cv1, betterShortTableBits)] = tableEntry{offset: off + 1, val: uint32(cv1)}
+			index0 += 2
+		}
+
+		cv = load6432(src, s)
+		if !canRepeat {
+			continue
+		}
+
+		// Check offset 2
+		for {
+			o2 := s - offset2
+			if load3232(src, o2) != uint32(cv) {
+				// Do regular search
+				break
+			}
+
+			// Store this, since we have it.
+			nextHashS := hash5(cv, betterShortTableBits)
+			nextHashL := hash8(cv, betterLongTableBits)
+
+			// We have at least 4 byte match.
+			// No need to check backwards. We come straight from a match
+			l := 4 + e.matchlen(s+4, o2+4, src)
+
+			e.longTable[nextHashL] = prevEntry{offset: s + e.cur, prev: e.longTable[nextHashL].offset}
+			e.table[nextHashS] = tableEntry{offset: s + e.cur, val: uint32(cv)}
+			seq.matchLen = uint32(l) - zstdMinMatch
+			seq.litLen = 0
+
+			// Since litlen is always 0, this is offset 1.
+			seq.offset = 1
+			s += l
+			nextEmit = s
+			if debugSequences {
+				println("sequence", seq, "next s:", s)
+			}
+			blk.sequences = append(blk.sequences, seq)
+
+			// Swap offset 1 and 2.
+			offset1, offset2 = offset2, offset1
+			if s >= sLimit {
+				// Finished
+				break encodeLoop
+			}
+			cv = load6432(src, s)
+		}
+	}
+
+	if int(nextEmit) < len(src) {
+		blk.literals = append(blk.literals, src[nextEmit:]...)
+		blk.extraLits = len(src) - int(nextEmit)
+	}
+	blk.recentOffsets[0] = uint32(offset1)
+	blk.recentOffsets[1] = uint32(offset2)
+	if debug {
+		println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits)
+	}
+}
+
+// EncodeNoHist will encode a block with no history and no following blocks.
+// Most notable difference is that src will not be copied for history and
+// we do not need to check for max match length.
+func (e *betterFastEncoder) EncodeNoHist(blk *blockEnc, src []byte) {
+	e.Encode(blk, src)
+}

vendor/github.com/klauspost/compress/zstd/enc_dfast.go

@@ -172,55 +172,6 @@ encodeLoop:
 					cv = load6432(src, s)
 					continue
 				}
-				const repOff2 = 1
-				// We deviate from the reference encoder and also check offset 2.
-				// Slower and not consistently better, so disabled.
-				// repIndex = s - offset2 + repOff2
-				if false && repIndex >= 0 && load3232(src, repIndex) == uint32(cv>>(repOff2*8)) {
-					// Consider history as well.
-					var seq seq
-					lenght := 4 + e.matchlen(s+4+repOff2, repIndex+4, src)
-
-					seq.matchLen = uint32(lenght - zstdMinMatch)
-
-					// We might be able to match backwards.
-					// Extend as long as we can.
-					start := s + repOff2
-					// We end the search early, so we don't risk 0 literals
-					// and have to do special offset treatment.
-					startLimit := nextEmit + 1
-
-					tMin := s - e.maxMatchOff
-					if tMin < 0 {
-						tMin = 0
-					}
-					for repIndex > tMin && start > startLimit && src[repIndex-1] == src[start-1] && seq.matchLen < maxMatchLength-zstdMinMatch-1 {
-						repIndex--
-						start--
-						seq.matchLen++
-					}
-					addLiterals(&seq, start)
-
-					// rep 2
-					seq.offset = 2
-					if debugSequences {
-						println("repeat sequence 2", seq, "next s:", s)
-					}
-					blk.sequences = append(blk.sequences, seq)
-					s += lenght + repOff2
-					nextEmit = s
-					if s >= sLimit {
-						if debug {
-							println("repeat ended", s, lenght)
-
-						}
-						break encodeLoop
-					}
-					cv = load6432(src, s)
-					// Swap offsets
-					offset1, offset2 = offset2, offset1
-					continue
-				}
 			}
 			// Find the offsets of our two matches.
 			coffsetL := s - (candidateL.offset - e.cur)
@@ -372,7 +323,7 @@ encodeLoop:
 			}
 
 			// Store this, since we have it.
-			nextHashS := hash5(cv1>>8, dFastShortTableBits)
+			nextHashS := hash5(cv, dFastShortTableBits)
 			nextHashL := hash8(cv, dFastLongTableBits)
 
 			// We have at least 4 byte match.

vendor/github.com/klauspost/compress/zstd/enc_fast.go

@@ -6,6 +6,7 @@ package zstd
 
 import (
 	"fmt"
+	"math"
 	"math/bits"
 
 	"github.com/klauspost/compress/zstd/internal/xxhash"
@@ -23,7 +24,7 @@ type tableEntry struct {
 	offset int32
 }
 
-type fastEncoder struct {
+type fastBase struct {
 	o encParams
 	// cur is the offset at the start of hist
 	cur int32
@@ -31,18 +32,22 @@ type fastEncoder struct {
 	maxMatchOff int32
 	hist        []byte
 	crc         *xxhash.Digest
-	table       [tableSize]tableEntry
 	tmp         [8]byte
 	blk         *blockEnc
 }
 
+type fastEncoder struct {
+	fastBase
+	table [tableSize]tableEntry
+}
+
 // CRC returns the underlying CRC writer.
-func (e *fastEncoder) CRC() *xxhash.Digest {
+func (e *fastBase) CRC() *xxhash.Digest {
 	return e.crc
 }
 
 // AppendCRC will append the CRC to the destination slice and return it.
-func (e *fastEncoder) AppendCRC(dst []byte) []byte {
+func (e *fastBase) AppendCRC(dst []byte) []byte {
 	crc := e.crc.Sum(e.tmp[:0])
 	dst = append(dst, crc[7], crc[6], crc[5], crc[4])
 	return dst
@@ -50,7 +55,7 @@ func (e *fastEncoder) AppendCRC(dst []byte) []byte {
 
 // WindowSize returns the window size of the encoder,
 // or a window size small enough to contain the input size, if > 0.
-func (e *fastEncoder) WindowSize(size int) int32 {
+func (e *fastBase) WindowSize(size int) int32 {
 	if size > 0 && size < int(e.maxMatchOff) {
 		b := int32(1) << uint(bits.Len(uint(size)))
 		// Keep minimum window.
@@ -63,7 +68,7 @@ func (e *fastEncoder) WindowSize(size int) int32 {
 }
 
 // Block returns the current block.
-func (e *fastEncoder) Block() *blockEnc {
+func (e *fastBase) Block() *blockEnc {
 	return e.blk
 }
 
@@ -169,9 +174,22 @@ encodeLoop:
 			if canRepeat && repIndex >= 0 && load3232(src, repIndex) == uint32(cv>>16) {
 				// Consider history as well.
 				var seq seq
-				lenght := 4 + e.matchlen(s+6, repIndex+4, src)
+				var length int32
+				// length = 4 + e.matchlen(s+6, repIndex+4, src)
+				{
+					a := src[s+6:]
+					b := src[repIndex+4:]
+					endI := len(a) & (math.MaxInt32 - 7)
+					length = int32(endI) + 4
+					for i := 0; i < endI; i += 8 {
+						if diff := load64(a, i) ^ load64(b, i); diff != 0 {
+							length = int32(i+bits.TrailingZeros64(diff)>>3) + 4
+							break
+						}
+					}
+				}
 
-				seq.matchLen = uint32(lenght - zstdMinMatch)
+				seq.matchLen = uint32(length - zstdMinMatch)
 
 				// We might be able to match backwards.
 				// Extend as long as we can.
@@ -197,11 +215,11 @@ encodeLoop:
 					println("repeat sequence", seq, "next s:", s)
 				}
 				blk.sequences = append(blk.sequences, seq)
-				s += lenght + 2
+				s += length + 2
 				nextEmit = s
 				if s >= sLimit {
 					if debug {
-						println("repeat ended", s, lenght)
+						println("repeat ended", s, length)
 
 					}
 					break encodeLoop
@@ -257,7 +275,20 @@ encodeLoop:
 		}
 
 		// Extend the 4-byte match as long as possible.
-		l := e.matchlen(s+4, t+4, src) + 4
+		//l := e.matchlen(s+4, t+4, src) + 4
+		var l int32
+		{
+			a := src[s+4:]
+			b := src[t+4:]
+			endI := len(a) & (math.MaxInt32 - 7)
+			l = int32(endI) + 4
+			for i := 0; i < endI; i += 8 {
+				if diff := load64(a, i) ^ load64(b, i); diff != 0 {
+					l = int32(i+bits.TrailingZeros64(diff)>>3) + 4
+					break
+				}
+			}
+		}
 
 		// Extend backwards
 		tMin := s - e.maxMatchOff
@@ -294,7 +325,20 @@ encodeLoop:
 		if o2 := s - offset2; canRepeat && load3232(src, o2) == uint32(cv) {
 			// We have at least 4 byte match.
 			// No need to check backwards. We come straight from a match
-			l := 4 + e.matchlen(s+4, o2+4, src)
+			//l := 4 + e.matchlen(s+4, o2+4, src)
+			var l int32
+			{
+				a := src[s+4:]
+				b := src[o2+4:]
+				endI := len(a) & (math.MaxInt32 - 7)
+				l = int32(endI) + 4
+				for i := 0; i < endI; i += 8 {
+					if diff := load64(a, i) ^ load64(b, i); diff != 0 {
+						l = int32(i+bits.TrailingZeros64(diff)>>3) + 4
+						break
+					}
+				}
+			}
 
 			// Store this, since we have it.
 			nextHash := hash6(cv, hashLog)
@@ -412,10 +456,23 @@ encodeLoop:
 			if len(blk.sequences) > 2 && load3232(src, repIndex) == uint32(cv>>16) {
 				// Consider history as well.
 				var seq seq
-				// lenght := 4 + e.matchlen(s+6, repIndex+4, src)
+				// length := 4 + e.matchlen(s+6, repIndex+4, src)
-				lenght := 4 + int32(matchLen(src[s+6:], src[repIndex+4:]))
+				// length := 4 + int32(matchLen(src[s+6:], src[repIndex+4:]))
+				var length int32
+				{
+					a := src[s+6:]
+					b := src[repIndex+4:]
+					endI := len(a) & (math.MaxInt32 - 7)
+					length = int32(endI) + 4
+					for i := 0; i < endI; i += 8 {
+						if diff := load64(a, i) ^ load64(b, i); diff != 0 {
+							length = int32(i+bits.TrailingZeros64(diff)>>3) + 4
+							break
+						}
+					}
+				}
 
-				seq.matchLen = uint32(lenght - zstdMinMatch)
+				seq.matchLen = uint32(length - zstdMinMatch)
 
 				// We might be able to match backwards.
 				// Extend as long as we can.
@@ -441,11 +498,11 @@ encodeLoop:
 					println("repeat sequence", seq, "next s:", s)
 				}
 				blk.sequences = append(blk.sequences, seq)
-				s += lenght + 2
+				s += length + 2
 				nextEmit = s
 				if s >= sLimit {
 					if debug {
-						println("repeat ended", s, lenght)
+						println("repeat ended", s, length)
 
 					}
 					break encodeLoop
@@ -498,7 +555,20 @@ encodeLoop:
 
 		// Extend the 4-byte match as long as possible.
 		//l := e.matchlenNoHist(s+4, t+4, src) + 4
-		l := int32(matchLen(src[s+4:], src[t+4:])) + 4
+		// l := int32(matchLen(src[s+4:], src[t+4:])) + 4
+		var l int32
+		{
+			a := src[s+4:]
+			b := src[t+4:]
+			endI := len(a) & (math.MaxInt32 - 7)
+			l = int32(endI) + 4
+			for i := 0; i < endI; i += 8 {
+				if diff := load64(a, i) ^ load64(b, i); diff != 0 {
+					l = int32(i+bits.TrailingZeros64(diff)>>3) + 4
+					break
+				}
+			}
+		}
 
 		// Extend backwards
 		tMin := s - e.maxMatchOff
@@ -536,7 +606,20 @@ encodeLoop:
 			// We have at least 4 byte match.
 			// No need to check backwards. We come straight from a match
 			//l := 4 + e.matchlenNoHist(s+4, o2+4, src)
-			l := 4 + int32(matchLen(src[s+4:], src[o2+4:]))
+			// l := 4 + int32(matchLen(src[s+4:], src[o2+4:]))
+			var l int32
+			{
+				a := src[s+4:]
+				b := src[o2+4:]
+				endI := len(a) & (math.MaxInt32 - 7)
+				l = int32(endI) + 4
+				for i := 0; i < endI; i += 8 {
+					if diff := load64(a, i) ^ load64(b, i); diff != 0 {
+						l = int32(i+bits.TrailingZeros64(diff)>>3) + 4
+						break
+					}
+				}
+			}
 
 			// Store this, since we have it.
 			nextHash := hash6(cv, hashLog)
@@ -571,7 +654,7 @@ encodeLoop:
 	}
 }
 
-func (e *fastEncoder) addBlock(src []byte) int32 {
+func (e *fastBase) addBlock(src []byte) int32 {
 	if debugAsserts && e.cur > bufferReset {
 		panic(fmt.Sprintf("ecur (%d) > buffer reset (%d)", e.cur, bufferReset))
 	}
@@ -602,17 +685,17 @@ func (e *fastEncoder) addBlock(src []byte) int32 {
 
 // useBlock will replace the block with the provided one,
 // but transfer recent offsets from the previous.
-func (e *fastEncoder) UseBlock(enc *blockEnc) {
+func (e *fastBase) UseBlock(enc *blockEnc) {
 	enc.reset(e.blk)
 	e.blk = enc
 }
 
-func (e *fastEncoder) matchlenNoHist(s, t int32, src []byte) int32 {
+func (e *fastBase) matchlenNoHist(s, t int32, src []byte) int32 {
 	// Extend the match to be as long as possible.
 	return int32(matchLen(src[s:], src[t:]))
 }
 
-func (e *fastEncoder) matchlen(s, t int32, src []byte) int32 {
+func (e *fastBase) matchlen(s, t int32, src []byte) int32 {
 	if debugAsserts {
 		if s < 0 {
 			err := fmt.Sprintf("s (%d) < 0", s)
@@ -626,18 +709,17 @@ func (e *fastEncoder) matchlen(s, t int32, src []byte) int32 {
 			err := fmt.Sprintf("s (%d) - t (%d) > maxMatchOff (%d)", s, t, e.maxMatchOff)
 			panic(err)
 		}
-	}
+		if len(src)-int(s) > maxCompressedBlockSize {
-	s1 := int(s) + maxMatchLength - 4
+			panic(fmt.Sprintf("len(src)-s (%d) > maxCompressedBlockSize (%d)", len(src)-int(s), maxCompressedBlockSize))
-	if s1 > len(src) {
+		}
-		s1 = len(src)
 	}
 
 	// Extend the match to be as long as possible.
-	return int32(matchLen(src[s:s1], src[t:]))
+	return int32(matchLen(src[s:], src[t:]))
 }
 
 // Reset the encoding table.
-func (e *fastEncoder) Reset() {
+func (e *fastBase) Reset() {
 	if e.blk == nil {
 		e.blk = &blockEnc{}
 		e.blk.init()

vendor/github.com/klauspost/compress/zstd/encoder.go

@@ -71,15 +71,14 @@ func NewWriter(w io.Writer, opts ...EOption) (*Encoder, error) {
 	}
 	if w != nil {
 		e.Reset(w)
-	} else {
-		e.init.Do(func() {
-			e.initialize()
-		})
 	}
 	return &e, nil
 }
 
 func (e *Encoder) initialize() {
+	if e.o.concurrent == 0 {
+		e.o.setDefault()
+	}
 	e.encoders = make(chan encoder, e.o.concurrent)
 	for i := 0; i < e.o.concurrent; i++ {
 		e.encoders <- e.o.encoder()
@@ -89,9 +88,6 @@ func (e *Encoder) initialize() {
 // Reset will re-initialize the writer and new writes will encode to the supplied writer
 // as a new, independent stream.
 func (e *Encoder) Reset(w io.Writer) {
-	e.init.Do(func() {
-		e.initialize()
-	})
 	s := &e.state
 	s.wg.Wait()
 	s.wWg.Wait()
@@ -422,10 +418,7 @@ func (e *Encoder) EncodeAll(src, dst []byte) []byte {
 		}
 		return dst
 	}
-	e.init.Do(func() {
+	e.init.Do(e.initialize)
-		e.o.setDefault()
-		e.initialize()
-	})
 	enc := <-e.encoders
 	defer func() {
 		// Release encoder reference to last block.

vendor/github.com/klauspost/compress/zstd/encoder_options.go

@@ -39,9 +39,11 @@ func (o *encoderOptions) setDefault() {
 func (o encoderOptions) encoder() encoder {
 	switch o.level {
 	case SpeedDefault:
-		return &doubleFastEncoder{fastEncoder: fastEncoder{maxMatchOff: int32(o.windowSize)}}
+		return &doubleFastEncoder{fastEncoder: fastEncoder{fastBase: fastBase{maxMatchOff: int32(o.windowSize)}}}
+	case SpeedBetterCompression:
+		return &betterFastEncoder{fastBase: fastBase{maxMatchOff: int32(o.windowSize)}}
 	case SpeedFastest:
-		return &fastEncoder{maxMatchOff: int32(o.windowSize)}
+		return &fastEncoder{fastBase: fastBase{maxMatchOff: int32(o.windowSize)}}
 	}
 	panic("unknown compression level")
 }
@@ -67,7 +69,7 @@ func WithEncoderConcurrency(n int) EOption {
 }
 
 // WithWindowSize will set the maximum allowed back-reference distance.
-// The value must be a power of two between WindowSizeMin and WindowSizeMax.
+// The value must be a power of two between MinWindowSize and MaxWindowSize.
 // A larger value will enable better compression but allocate more memory and,
 // for above-default values, take considerably longer.
 // The default value is determined by the compression level.
@@ -130,18 +132,18 @@ const (
 	// This is roughly equivalent to the default Zstandard mode (level 3).
 	SpeedDefault
 
+	// SpeedBetterCompression will yield better compression than the default.
+	// Currently it is about zstd level 7-8 with ~ 2x-3x the default CPU usage.
+	// By using this, notice that CPU usage may go up in the future.
+	SpeedBetterCompression
+
 	// speedLast should be kept as the last actual compression option.
 	// The is not for external usage, but is used to keep track of the valid options.
 	speedLast
 
-	// SpeedBetterCompression will (in the future) yield better compression than the default,
-	// but at approximately 4x the CPU usage of the default.
-	// For now this is not implemented.
-	SpeedBetterCompression = SpeedDefault
-
 	// SpeedBestCompression will choose the best available compression option.
 	// For now this is not implemented.
-	SpeedBestCompression = SpeedDefault
+	SpeedBestCompression = SpeedBetterCompression
 )
 
 // EncoderLevelFromString will convert a string representation of an encoding level back
@@ -163,8 +165,10 @@ func EncoderLevelFromZstd(level int) EncoderLevel {
 	switch {
 	case level < 3:
 		return SpeedFastest
-	case level >= 3:
+	case level >= 3 && level < 6:
 		return SpeedDefault
+	case level > 5:
+		return SpeedBetterCompression
 	}
 	return SpeedDefault
 }
@@ -176,6 +180,8 @@ func (e EncoderLevel) String() string {
 		return "fastest"
 	case SpeedDefault:
 		return "default"
+	case SpeedBetterCompression:
+		return "better"
 	default:
 		return "invalid"
 	}

vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_amd64.s

@@ -179,13 +179,13 @@ TEXT ·writeBlocks(SB), NOSPLIT, $0-40
 	MOVQ ·prime2v(SB), R14
 
 	// Load slice.
-	MOVQ b_base+8(FP), CX
+	MOVQ arg1_base+8(FP), CX
-	MOVQ b_len+16(FP), DX
+	MOVQ arg1_len+16(FP), DX
 	LEAQ (CX)(DX*1), BX
 	SUBQ $32, BX
 
 	// Load vN from d.
-	MOVQ d+0(FP), AX
+	MOVQ arg+0(FP), AX
 	MOVQ 0(AX), R8   // v1
 	MOVQ 8(AX), R9   // v2
 	MOVQ 16(AX), R10 // v3
@@ -209,7 +209,7 @@ blockLoop:
 	MOVQ R11, 24(AX)
 
 	// The number of bytes written is CX minus the old base pointer.
-	SUBQ b_base+8(FP), CX
+	SUBQ arg1_base+8(FP), CX
 	MOVQ CX, ret+32(FP)
 
 	RET

vendor/github.com/klauspost/compress/zstd/zstd.go

@@ -87,6 +87,17 @@ func printf(format string, a ...interface{}) {
 	}
 }
 
+// matchLenFast does matching, but will not match the last up to 7 bytes.
+func matchLenFast(a, b []byte) int {
+	endI := len(a) & (math.MaxInt32 - 7)
+	for i := 0; i < endI; i += 8 {
+		if diff := load64(a, i) ^ load64(b, i); diff != 0 {
+			return i + bits.TrailingZeros64(diff)>>3
+		}
+	}
+	return endI
+}
+
 // matchLen returns the maximum length.
 // a must be the shortest of the two.
 // The function also returns whether all bytes matched.
@@ -97,33 +108,18 @@ func matchLen(a, b []byte) int {
 			return i + (bits.TrailingZeros64(diff) >> 3)
 		}
 	}
+
 	checked := (len(a) >> 3) << 3
 	a = a[checked:]
 	b = b[checked:]
-	// TODO: We could do a 4 check.
 	for i := range a {
 		if a[i] != b[i] {
-			return int(i) + checked
+			return i + checked
 		}
 	}
 	return len(a) + checked
 }
 
-// matchLen returns a match length in src between index s and t
-func matchLenIn(src []byte, s, t int32) int32 {
-	s1 := len(src)
-	b := src[t:]
-	a := src[s:s1]
-	b = b[:len(a)]
-	// Extend the match to be as long as possible.
-	for i := range a {
-		if a[i] != b[i] {
-			return int32(i)
-		}
-	}
-	return int32(len(a))
-}
-
 func load3232(b []byte, i int32) uint32 {
 	// Help the compiler eliminate bounds checks on the read so it can be done in a single read.
 	b = b[i:]

vendor/modules.txt

@@ -85,7 +85,7 @@ github.com/jmespath/go-jmespath
 github.com/jstemmer/go-junit-report
 github.com/jstemmer/go-junit-report/formatter
 github.com/jstemmer/go-junit-report/parser
-# github.com/klauspost/compress v1.10.1
+# github.com/klauspost/compress v1.10.3
 github.com/klauspost/compress/flate
 github.com/klauspost/compress/fse
 github.com/klauspost/compress/gzip