Update dependencies (#5518)

vendor/github.com/klauspost/compress/huff0/README.md

@@ -12,11 +12,11 @@ but it can be used as a secondary step to compressors (like Snappy) that does no
 
 * [Godoc documentation](https://godoc.org/github.com/klauspost/compress/huff0)
 
-THIS PACKAGE IS NOT CONSIDERED STABLE AND API OR ENCODING MAY CHANGE IN THE FUTURE.
-
 ## News
 
- * Mar 2018: First implementation released. Consider this beta software for now.
+This is used as part of the [zstandard](https://github.com/klauspost/compress/tree/master/zstd#zstd) compression and decompression package.
+
+This ensures that most functionality is well tested.
 
 # Usage
 
@@ -75,6 +75,8 @@ which can be given to the decompressor.
 Decompressing is done by calling the [`Decompress1X`](https://godoc.org/github.com/klauspost/compress/huff0#Scratch.Decompress1X) 
 or [`Decompress4X`](https://godoc.org/github.com/klauspost/compress/huff0#Scratch.Decompress4X) function.
 
+For concurrently decompressing content with a fixed table a stateless [`Decoder`](https://godoc.org/github.com/klauspost/compress/huff0#Decoder) can be requested which will remain correct as long as the scratch is unchanged. The capacity of the provided slice indicates the expected output size.
+
 You must provide the output from the compression stage, at exactly the size you got back. If you receive an error back
 your input was likely corrupted. 
 
@@ -84,4 +86,4 @@ There are no integrity checks, so relying on errors from the decompressor does n
 # Contributing
 
 Contributions are always welcome. Be aware that adding public functions will require good justification and breaking 
-changes will likely not be accepted. If in doubt open an issue before writing the PR.
+changes will likely not be accepted. If in doubt open an issue before writing the PR.

vendor/github.com/klauspost/compress/huff0/bitreader.go

@@ -6,6 +6,7 @@
 package huff0
 
 import (
+	"encoding/binary"
 	"errors"
 	"io"
 )
@@ -34,29 +35,16 @@ func (b *bitReader) init(in []byte) error {
 	}
 	b.bitsRead = 64
 	b.value = 0
-	b.fill()
-	b.fill()
+	if len(in) >= 8 {
+		b.fillFastStart()
+	} else {
+		b.fill()
+		b.fill()
+	}
 	b.bitsRead += 8 - uint8(highBit32(uint32(v)))
 	return nil
 }
 
-// getBits will return n bits. n can be 0.
-func (b *bitReader) getBits(n uint8) uint16 {
-	if n == 0 || b.bitsRead >= 64 {
-		return 0
-	}
-	return b.getBitsFast(n)
-}
-
-// getBitsFast requires that at least one bit is requested every time.
-// There are no checks if the buffer is filled.
-func (b *bitReader) getBitsFast(n uint8) uint16 {
-	const regMask = 64 - 1
-	v := uint16((b.value << (b.bitsRead & regMask)) >> ((regMask + 1 - n) & regMask))
-	b.bitsRead += n
-	return v
-}
-
 // peekBitsFast requires that at least one bit is requested every time.
 // There are no checks if the buffer is filled.
 func (b *bitReader) peekBitsFast(n uint8) uint16 {
@@ -71,21 +59,36 @@ func (b *bitReader) fillFast() {
 	if b.bitsRead < 32 {
 		return
 	}
-	// Do single re-slice to avoid bounds checks.
+
+	// 2 bounds checks.
 	v := b.in[b.off-4 : b.off]
+	v = v[:4]
 	low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
 	b.value = (b.value << 32) | uint64(low)
 	b.bitsRead -= 32
 	b.off -= 4
 }
 
+func (b *bitReader) advance(n uint8) {
+	b.bitsRead += n
+}
+
+// fillFastStart() assumes the bitreader is empty and there is at least 8 bytes to read.
+func (b *bitReader) fillFastStart() {
+	// Do single re-slice to avoid bounds checks.
+	b.value = binary.LittleEndian.Uint64(b.in[b.off-8:])
+	b.bitsRead = 0
+	b.off -= 8
+}
+
 // fill() will make sure at least 32 bits are available.
 func (b *bitReader) fill() {
 	if b.bitsRead < 32 {
 		return
 	}
 	if b.off > 4 {
-		v := b.in[b.off-4 : b.off]
+		v := b.in[b.off-4:]
+		v = v[:4]
 		low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
 		b.value = (b.value << 32) | uint64(low)
 		b.bitsRead -= 32
@@ -113,3 +116,214 @@ func (b *bitReader) close() error {
 	}
 	return nil
 }
+
+// bitReader reads a bitstream in reverse.
+// The last set bit indicates the start of the stream and is used
+// for aligning the input.
+type bitReaderBytes struct {
+	in       []byte
+	off      uint // next byte to read is at in[off - 1]
+	value    uint64
+	bitsRead uint8
+}
+
+// init initializes and resets the bit reader.
+func (b *bitReaderBytes) init(in []byte) error {
+	if len(in) < 1 {
+		return errors.New("corrupt stream: too short")
+	}
+	b.in = in
+	b.off = uint(len(in))
+	// The highest bit of the last byte indicates where to start
+	v := in[len(in)-1]
+	if v == 0 {
+		return errors.New("corrupt stream, did not find end of stream")
+	}
+	b.bitsRead = 64
+	b.value = 0
+	if len(in) >= 8 {
+		b.fillFastStart()
+	} else {
+		b.fill()
+		b.fill()
+	}
+	b.advance(8 - uint8(highBit32(uint32(v))))
+	return nil
+}
+
+// peekBitsFast requires that at least one bit is requested every time.
+// There are no checks if the buffer is filled.
+func (b *bitReaderBytes) peekByteFast() uint8 {
+	got := uint8(b.value >> 56)
+	return got
+}
+
+func (b *bitReaderBytes) advance(n uint8) {
+	b.bitsRead += n
+	b.value <<= n & 63
+}
+
+// fillFast() will make sure at least 32 bits are available.
+// There must be at least 4 bytes available.
+func (b *bitReaderBytes) fillFast() {
+	if b.bitsRead < 32 {
+		return
+	}
+
+	// 2 bounds checks.
+	v := b.in[b.off-4 : b.off]
+	v = v[:4]
+	low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	b.value |= uint64(low) << (b.bitsRead - 32)
+	b.bitsRead -= 32
+	b.off -= 4
+}
+
+// fillFastStart() assumes the bitReaderBytes is empty and there is at least 8 bytes to read.
+func (b *bitReaderBytes) fillFastStart() {
+	// Do single re-slice to avoid bounds checks.
+	b.value = binary.LittleEndian.Uint64(b.in[b.off-8:])
+	b.bitsRead = 0
+	b.off -= 8
+}
+
+// fill() will make sure at least 32 bits are available.
+func (b *bitReaderBytes) fill() {
+	if b.bitsRead < 32 {
+		return
+	}
+	if b.off > 4 {
+		v := b.in[b.off-4:]
+		v = v[:4]
+		low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+		b.value |= uint64(low) << (b.bitsRead - 32)
+		b.bitsRead -= 32
+		b.off -= 4
+		return
+	}
+	for b.off > 0 {
+		b.value |= uint64(b.in[b.off-1]) << (b.bitsRead - 8)
+		b.bitsRead -= 8
+		b.off--
+	}
+}
+
+// finished returns true if all bits have been read from the bit stream.
+func (b *bitReaderBytes) finished() bool {
+	return b.off == 0 && b.bitsRead >= 64
+}
+
+// close the bitstream and returns an error if out-of-buffer reads occurred.
+func (b *bitReaderBytes) close() error {
+	// Release reference.
+	b.in = nil
+	if b.bitsRead > 64 {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}
+
+// bitReaderShifted reads a bitstream in reverse.
+// The last set bit indicates the start of the stream and is used
+// for aligning the input.
+type bitReaderShifted struct {
+	in       []byte
+	off      uint // next byte to read is at in[off - 1]
+	value    uint64
+	bitsRead uint8
+}
+
+// init initializes and resets the bit reader.
+func (b *bitReaderShifted) init(in []byte) error {
+	if len(in) < 1 {
+		return errors.New("corrupt stream: too short")
+	}
+	b.in = in
+	b.off = uint(len(in))
+	// The highest bit of the last byte indicates where to start
+	v := in[len(in)-1]
+	if v == 0 {
+		return errors.New("corrupt stream, did not find end of stream")
+	}
+	b.bitsRead = 64
+	b.value = 0
+	if len(in) >= 8 {
+		b.fillFastStart()
+	} else {
+		b.fill()
+		b.fill()
+	}
+	b.advance(8 - uint8(highBit32(uint32(v))))
+	return nil
+}
+
+// peekBitsFast requires that at least one bit is requested every time.
+// There are no checks if the buffer is filled.
+func (b *bitReaderShifted) peekBitsFast(n uint8) uint16 {
+	return uint16(b.value >> ((64 - n) & 63))
+}
+
+func (b *bitReaderShifted) advance(n uint8) {
+	b.bitsRead += n
+	b.value <<= n & 63
+}
+
+// fillFast() will make sure at least 32 bits are available.
+// There must be at least 4 bytes available.
+func (b *bitReaderShifted) fillFast() {
+	if b.bitsRead < 32 {
+		return
+	}
+
+	// 2 bounds checks.
+	v := b.in[b.off-4 : b.off]
+	v = v[:4]
+	low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	b.value |= uint64(low) << ((b.bitsRead - 32) & 63)
+	b.bitsRead -= 32
+	b.off -= 4
+}
+
+// fillFastStart() assumes the bitReaderShifted is empty and there is at least 8 bytes to read.
+func (b *bitReaderShifted) fillFastStart() {
+	// Do single re-slice to avoid bounds checks.
+	b.value = binary.LittleEndian.Uint64(b.in[b.off-8:])
+	b.bitsRead = 0
+	b.off -= 8
+}
+
+// fill() will make sure at least 32 bits are available.
+func (b *bitReaderShifted) fill() {
+	if b.bitsRead < 32 {
+		return
+	}
+	if b.off > 4 {
+		v := b.in[b.off-4:]
+		v = v[:4]
+		low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+		b.value |= uint64(low) << ((b.bitsRead - 32) & 63)
+		b.bitsRead -= 32
+		b.off -= 4
+		return
+	}
+	for b.off > 0 {
+		b.value |= uint64(b.in[b.off-1]) << ((b.bitsRead - 8) & 63)
+		b.bitsRead -= 8
+		b.off--
+	}
+}
+
+// finished returns true if all bits have been read from the bit stream.
+func (b *bitReaderShifted) finished() bool {
+	return b.off == 0 && b.bitsRead >= 64
+}
+
+// close the bitstream and returns an error if out-of-buffer reads occurred.
+func (b *bitReaderShifted) close() error {
+	// Release reference.
+	b.in = nil
+	if b.bitsRead > 64 {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}

vendor/github.com/klauspost/compress/huff0/bitwriter.go

@@ -43,6 +43,11 @@ func (b *bitWriter) addBits16Clean(value uint16, bits uint8) {
 func (b *bitWriter) encSymbol(ct cTable, symbol byte) {
 	enc := ct[symbol]
 	b.bitContainer |= uint64(enc.val) << (b.nBits & 63)
+	if false {
+		if enc.nBits == 0 {
+			panic("nbits 0")
+		}
+	}
 	b.nBits += enc.nBits
 }
 
@@ -54,6 +59,14 @@ func (b *bitWriter) encTwoSymbols(ct cTable, av, bv byte) {
 	sh := b.nBits & 63
 	combined := uint64(encA.val) | (uint64(encB.val) << (encA.nBits & 63))
 	b.bitContainer |= combined << sh
+	if false {
+		if encA.nBits == 0 {
+			panic("nbitsA 0")
+		}
+		if encB.nBits == 0 {
+			panic("nbitsB 0")
+		}
+	}
 	b.nBits += encA.nBits + encB.nBits
 }
 

vendor/github.com/klauspost/compress/huff0/compress.go

@@ -77,22 +77,30 @@ func compress(in []byte, s *Scratch, compressor func(src []byte) ([]byte, error)
 		// Each symbol present maximum once or too well distributed.
 		return nil, false, ErrIncompressible
 	}
-
-	if s.Reuse == ReusePolicyPrefer && canReuse {
+	if s.Reuse == ReusePolicyMust && !canReuse {
+		// We must reuse, but we can't.
+		return nil, false, ErrIncompressible
+	}
+	if (s.Reuse == ReusePolicyPrefer || s.Reuse == ReusePolicyMust) && canReuse {
 		keepTable := s.cTable
+		keepTL := s.actualTableLog
 		s.cTable = s.prevTable
+		s.actualTableLog = s.prevTableLog
 		s.Out, err = compressor(in)
 		s.cTable = keepTable
+		s.actualTableLog = keepTL
 		if err == nil && len(s.Out) < wantSize {
 			s.OutData = s.Out
 			return s.Out, true, nil
 		}
+		if s.Reuse == ReusePolicyMust {
+			return nil, false, ErrIncompressible
+		}
 		// Do not attempt to re-use later.
 		s.prevTable = s.prevTable[:0]
 	}
 
 	// Calculate new table.
-	s.optimalTableLog()
 	err = s.buildCTable()
 	if err != nil {
 		return nil, false, err
@@ -109,9 +117,15 @@ func compress(in []byte, s *Scratch, compressor func(src []byte) ([]byte, error)
 		if oldSize <= hSize+newSize || hSize+12 >= wantSize {
 			// Retain cTable even if we re-use.
 			keepTable := s.cTable
+			keepTL := s.actualTableLog
+
 			s.cTable = s.prevTable
+			s.actualTableLog = s.prevTableLog
 			s.Out, err = compressor(in)
+
+			// Restore ctable.
 			s.cTable = keepTable
+			s.actualTableLog = keepTL
 			if err != nil {
 				return nil, false, err
 			}
@@ -142,11 +156,75 @@ func compress(in []byte, s *Scratch, compressor func(src []byte) ([]byte, error)
 		return nil, false, ErrIncompressible
 	}
 	// Move current table into previous.
-	s.prevTable, s.cTable = s.cTable, s.prevTable[:0]
+	s.prevTable, s.prevTableLog, s.cTable = s.cTable, s.actualTableLog, s.prevTable[:0]
 	s.OutData = s.Out[len(s.OutTable):]
 	return s.Out, false, nil
 }
 
+// EstimateSizes will estimate the data sizes
+func EstimateSizes(in []byte, s *Scratch) (tableSz, dataSz, reuseSz int, err error) {
+	s, err = s.prepare(in)
+	if err != nil {
+		return 0, 0, 0, err
+	}
+
+	// Create histogram, if none was provided.
+	tableSz, dataSz, reuseSz = -1, -1, -1
+	maxCount := s.maxCount
+	var canReuse = false
+	if maxCount == 0 {
+		maxCount, canReuse = s.countSimple(in)
+	} else {
+		canReuse = s.canUseTable(s.prevTable)
+	}
+
+	// We want the output size to be less than this:
+	wantSize := len(in)
+	if s.WantLogLess > 0 {
+		wantSize -= wantSize >> s.WantLogLess
+	}
+
+	// Reset for next run.
+	s.clearCount = true
+	s.maxCount = 0
+	if maxCount >= len(in) {
+		if maxCount > len(in) {
+			return 0, 0, 0, fmt.Errorf("maxCount (%d) > length (%d)", maxCount, len(in))
+		}
+		if len(in) == 1 {
+			return 0, 0, 0, ErrIncompressible
+		}
+		// One symbol, use RLE
+		return 0, 0, 0, ErrUseRLE
+	}
+	if maxCount == 1 || maxCount < (len(in)>>7) {
+		// Each symbol present maximum once or too well distributed.
+		return 0, 0, 0, ErrIncompressible
+	}
+
+	// Calculate new table.
+	err = s.buildCTable()
+	if err != nil {
+		return 0, 0, 0, err
+	}
+
+	if false && !s.canUseTable(s.cTable) {
+		panic("invalid table generated")
+	}
+
+	tableSz, err = s.cTable.estTableSize(s)
+	if err != nil {
+		return 0, 0, 0, err
+	}
+	if canReuse {
+		reuseSz = s.prevTable.estimateSize(s.count[:s.symbolLen])
+	}
+	dataSz = s.cTable.estimateSize(s.count[:s.symbolLen])
+
+	// Restore
+	return tableSz, dataSz, reuseSz, nil
+}
+
 func (s *Scratch) compress1X(src []byte) ([]byte, error) {
 	return s.compress1xDo(s.Out, src)
 }
@@ -317,9 +395,26 @@ func (s *Scratch) canUseTable(c cTable) bool {
 	return true
 }
 
+func (s *Scratch) validateTable(c cTable) bool {
+	if len(c) < int(s.symbolLen) {
+		return false
+	}
+	for i, v := range s.count[:s.symbolLen] {
+		if v != 0 {
+			if c[i].nBits == 0 {
+				return false
+			}
+			if c[i].nBits > s.actualTableLog {
+				return false
+			}
+		}
+	}
+	return true
+}
+
 // minTableLog provides the minimum logSize to safely represent a distribution.
 func (s *Scratch) minTableLog() uint8 {
-	minBitsSrc := highBit32(uint32(s.br.remain()-1)) + 1
+	minBitsSrc := highBit32(uint32(s.br.remain())) + 1
 	minBitsSymbols := highBit32(uint32(s.symbolLen-1)) + 2
 	if minBitsSrc < minBitsSymbols {
 		return uint8(minBitsSrc)
@@ -331,7 +426,7 @@ func (s *Scratch) minTableLog() uint8 {
 func (s *Scratch) optimalTableLog() {
 	tableLog := s.TableLog
 	minBits := s.minTableLog()
-	maxBitsSrc := uint8(highBit32(uint32(s.br.remain()-1))) - 2
+	maxBitsSrc := uint8(highBit32(uint32(s.br.remain()-1))) - 1
 	if maxBitsSrc < tableLog {
 		// Accuracy can be reduced
 		tableLog = maxBitsSrc
@@ -358,6 +453,7 @@ type cTableEntry struct {
 const huffNodesMask = huffNodesLen - 1
 
 func (s *Scratch) buildCTable() error {
+	s.optimalTableLog()
 	s.huffSort()
 	if cap(s.cTable) < maxSymbolValue+1 {
 		s.cTable = make([]cTableEntry, s.symbolLen, maxSymbolValue+1)
@@ -371,7 +467,7 @@ func (s *Scratch) buildCTable() error {
 	var startNode = int16(s.symbolLen)
 	nonNullRank := s.symbolLen - 1
 
-	nodeNb := int16(startNode)
+	nodeNb := startNode
 	huffNode := s.nodes[1 : huffNodesLen+1]
 
 	// This overlays the slice above, but allows "-1" index lookups.
@@ -504,7 +600,6 @@ func (s *Scratch) huffSort() {
 		}
 		nodes[pos&huffNodesMask] = nodeElt{count: c, symbol: byte(n)}
 	}
-	return
 }
 
 func (s *Scratch) setMaxHeight(lastNonNull int) uint8 {
@@ -548,7 +643,7 @@ func (s *Scratch) setMaxHeight(lastNonNull int) uint8 {
 
 		// Get pos of last (smallest) symbol per rank
 		{
-			currentNbBits := uint8(maxNbBits)
+			currentNbBits := maxNbBits
 			for pos := int(n); pos >= 0; pos-- {
 				if huffNode[pos].nbBits >= currentNbBits {
 					continue

vendor/github.com/klauspost/compress/huff0/huff0.go

@@ -55,6 +55,9 @@ const (
 	// ReusePolicyNone will disable re-use of tables.
 	// This is slightly faster than ReusePolicyAllow but may produce larger output.
 	ReusePolicyNone
+
+	// ReusePolicyMust must allow reuse and produce smaller output.
+	ReusePolicyMust
 )
 
 type Scratch struct {
@@ -79,11 +82,18 @@ type Scratch struct {
 	// Slice of the returned data.
 	OutData []byte
 
+	// MaxDecodedSize will set the maximum allowed output size.
+	// This value will automatically be set to BlockSizeMax if not set.
+	// Decoders will return ErrMaxDecodedSizeExceeded is this limit is exceeded.
+	MaxDecodedSize int
+
+	br byteReader
+
 	// MaxSymbolValue will override the maximum symbol value of the next block.
 	MaxSymbolValue uint8
 
 	// TableLog will attempt to override the tablelog for the next block.
-	// Must be <= 11.
+	// Must be <= 11 and >= 5.
 	TableLog uint8
 
 	// Reuse will specify the reuse policy
@@ -95,16 +105,11 @@ type Scratch struct {
 	// If WantLogLess == 0 any improvement will do.
 	WantLogLess uint8
 
-	// MaxDecodedSize will set the maximum allowed output size.
-	// This value will automatically be set to BlockSizeMax if not set.
-	// Decoders will return ErrMaxDecodedSizeExceeded is this limit is exceeded.
-	MaxDecodedSize int
-
-	br             byteReader
 	symbolLen      uint16 // Length of active part of the symbol table.
 	maxCount       int    // count of the most probable symbol
 	clearCount     bool   // clear count
 	actualTableLog uint8  // Selected tablelog.
+	prevTableLog   uint8  // Tablelog for previous table
 	prevTable      cTable // Table used for previous compression.
 	cTable         cTable // compression table
 	dt             dTable // decompression table
@@ -114,6 +119,16 @@ type Scratch struct {
 	huffWeight     [maxSymbolValue + 1]byte
 }
 
+// TransferCTable will transfer the previously used compression table.
+func (s *Scratch) TransferCTable(src *Scratch) {
+	if cap(s.prevTable) < len(src.prevTable) {
+		s.prevTable = make(cTable, 0, maxSymbolValue+1)
+	}
+	s.prevTable = s.prevTable[:len(src.prevTable)]
+	copy(s.prevTable, src.prevTable)
+	s.prevTableLog = src.prevTableLog
+}
+
 func (s *Scratch) prepare(in []byte) (*Scratch, error) {
 	if len(in) > BlockSizeMax {
 		return nil, ErrTooBig
@@ -127,8 +142,8 @@ func (s *Scratch) prepare(in []byte) (*Scratch, error) {
 	if s.TableLog == 0 {
 		s.TableLog = tableLogDefault
 	}
-	if s.TableLog > tableLogMax {
-		return nil, fmt.Errorf("tableLog (%d) > maxTableLog (%d)", s.TableLog, tableLogMax)
+	if s.TableLog > tableLogMax || s.TableLog < minTablelog {
+		return nil, fmt.Errorf(" invalid tableLog %d (%d -> %d)", s.TableLog, minTablelog, tableLogMax)
 	}
 	if s.MaxDecodedSize <= 0 || s.MaxDecodedSize > BlockSizeMax {
 		s.MaxDecodedSize = BlockSizeMax
@@ -230,6 +245,68 @@ func (c cTable) write(s *Scratch) error {
 	return nil
 }
 
+func (c cTable) estTableSize(s *Scratch) (sz int, err error) {
+	var (
+		// precomputed conversion table
+		bitsToWeight [tableLogMax + 1]byte
+		huffLog      = s.actualTableLog
+		// last weight is not saved.
+		maxSymbolValue = uint8(s.symbolLen - 1)
+		huffWeight     = s.huffWeight[:256]
+	)
+	const (
+		maxFSETableLog = 6
+	)
+	// convert to weight
+	bitsToWeight[0] = 0
+	for n := uint8(1); n < huffLog+1; n++ {
+		bitsToWeight[n] = huffLog + 1 - n
+	}
+
+	// Acquire histogram for FSE.
+	hist := s.fse.Histogram()
+	hist = hist[:256]
+	for i := range hist[:16] {
+		hist[i] = 0
+	}
+	for n := uint8(0); n < maxSymbolValue; n++ {
+		v := bitsToWeight[c[n].nBits] & 15
+		huffWeight[n] = v
+		hist[v]++
+	}
+
+	// FSE compress if feasible.
+	if maxSymbolValue >= 2 {
+		huffMaxCnt := uint32(0)
+		huffMax := uint8(0)
+		for i, v := range hist[:16] {
+			if v == 0 {
+				continue
+			}
+			huffMax = byte(i)
+			if v > huffMaxCnt {
+				huffMaxCnt = v
+			}
+		}
+		s.fse.HistogramFinished(huffMax, int(huffMaxCnt))
+		s.fse.TableLog = maxFSETableLog
+		b, err := fse.Compress(huffWeight[:maxSymbolValue], s.fse)
+		if err == nil && len(b) < int(s.symbolLen>>1) {
+			sz += 1 + len(b)
+			return sz, nil
+		}
+		// Unable to compress (RLE/uncompressible)
+	}
+	// write raw values as 4-bits (max : 15)
+	if maxSymbolValue > (256 - 128) {
+		// should not happen : likely means source cannot be compressed
+		return 0, ErrIncompressible
+	}
+	// special case, pack weights 4 bits/weight.
+	sz += 1 + int(maxSymbolValue/2)
+	return sz, nil
+}
+
 // estimateSize returns the estimated size in bytes of the input represented in the
 // histogram supplied.
 func (c cTable) estimateSize(hist []uint32) int {