flate: Improve level 7-9 (again)

github-june-2days-2019.json 899480979 -> 898541970
nyc-taxi-data-10M.csv 709158534 -> 706632036
enwik9 324078757 -> 322251720
github-ranks-backup.bin 435530626 -> 435809465
gob-stream 279987188 -> 278406897
silesia.tar 68817867 -> 68772170

flate/deflate.go

@@ -10,9 +10,6 @@ import (
 	"fmt"
 	"io"
 	"math"
-	"math/bits"
-
-	comp "github.com/klauspost/compress"
 )
 
 const (
@@ -110,6 +107,7 @@ type advancedState struct {
 type compressor struct {
 	compressionLevel
 
+	h *huffmanEncoder
 	w *huffmanBitWriter
 
 	// compression algorithm
@@ -299,12 +297,13 @@ func (d *compressor) findMatch(pos int, prevHead int, lookahead, bpb int) (lengt
 
 	// Base is 4 bytes at with an additional cost.
 	// Matches must be better than this.
-	cGain := minMatchLength*bpb - 12
+	cGain := 0
 	for i := prevHead; tries > 0; tries-- {
 		if wEnd == win[i+length] {
 			n := matchLen(win[i:i+minMatchLook], wPos)
 			if n > length {
-				newGain := n*bpb - bits.Len32(uint32(pos-i))
+				// Calculate gain. Estimate
+				newGain := d.h.bitLengthRaw(wPos[:n]) - int(offsetExtraBits[offsetCode(uint32(pos-i))]) - bpb - int(lengthExtraBits[lengthCodes[(n-3)&255]])
 				if newGain > cGain {
 					length = n
 					offset = pos - i
@@ -389,10 +388,16 @@ func (d *compressor) deflateLazy() {
 	if d.windowEnd-s.index < minMatchLength+maxMatchLength && !d.sync {
 		return
 	}
-	s.estBitsPerByte = 8
-	if !d.sync {
-		s.estBitsPerByte = comp.ShannonEntropyBits(d.window[s.index:d.windowEnd])
-		s.estBitsPerByte = int(1 + float64(s.estBitsPerByte)/float64(d.windowEnd-s.index))
+	if d.windowEnd != s.index {
+		// Get literal huffman coder.
+		if d.h == nil {
+			d.h = newHuffmanEncoder(maxFlateBlockTokens)
+		}
+		var tmp [256]uint16
+		for _, v := range d.window[s.index:d.windowEnd] {
+			tmp[v]++
+		}
+		d.h.generate(tmp[:], 15)
 	}
 
 	s.maxInsertIndex = d.windowEnd - (minMatchLength - 1)
@@ -446,7 +451,12 @@ func (d *compressor) deflateLazy() {
 		}
 
 		if s.chainHead-s.hashOffset >= minIndex && lookahead > prevLength && prevLength < d.lazy {
-			if newLength, newOffset, ok := d.findMatch(s.index, s.chainHead-s.hashOffset, lookahead, s.estBitsPerByte); ok {
+			e := 8
+			if d.tokens.nLits*4 > d.tokens.nMatch {
+				// We have many literals, emit any size.
+				e = -100
+			}
+			if newLength, newOffset, ok := d.findMatch(s.index, s.chainHead-s.hashOffset, lookahead, e); ok {
 				s.length = newLength
 				s.offset = newOffset
 			}

flate/token.go

@@ -129,11 +129,13 @@ var offsetCodes14 = [256]uint32{
 type token uint32
 
 type tokens struct {
-	nLits     int
 	extraHist [32]uint16  // codes 256->maxnumlit
 	offHist   [32]uint16  // offset codes
 	litHist   [256]uint16 // codes 0->255
-	n         uint16      // Must be able to contain maxStoreBlockSize
+	nFilled   int
+	nLits     int
+	nMatch    int
+	n         uint16 // Must be able to contain maxStoreBlockSize
 	tokens    [maxStoreBlockSize + 1]token
 }
 
@@ -142,7 +144,9 @@ func (t *tokens) Reset() {
 		return
 	}
 	t.n = 0
+	t.nFilled = 0
 	t.nLits = 0
+	t.nMatch = 0
 	for i := range t.litHist[:] {
 		t.litHist[i] = 0
 	}
@@ -161,12 +165,12 @@ func (t *tokens) Fill() {
 	for i, v := range t.litHist[:] {
 		if v == 0 {
 			t.litHist[i] = 1
-			t.nLits++
+			t.nFilled++
 		}
 	}
 	for i, v := range t.extraHist[:literalCount-256] {
 		if v == 0 {
-			t.nLits++
+			t.nFilled++
 			t.extraHist[i] = 1
 		}
 	}
@@ -230,8 +234,9 @@ func (t *tokens) EstimatedBits() int {
 	shannon := float32(0)
 	bits := int(0)
 	nMatches := 0
-	if t.nLits > 0 {
-		invTotal := 1.0 / float32(t.nLits)
+	total := int(t.n) + t.nFilled
+	if total > 0 {
+		invTotal := 1.0 / float32(total)
 		for _, v := range t.litHist[:] {
 			if v > 0 {
 				n := float32(v)
@@ -275,9 +280,9 @@ func (t *tokens) AddMatch(xlength uint32, xoffset uint32) {
 	}
 	oCode := offsetCode(xoffset)
 	xoffset |= oCode << 16
-	t.nLits++
 
 	t.extraHist[lengthCodes1[uint8(xlength)]]++
+	t.nMatch += int(xlength) + baseMatchLength
 	t.offHist[oCode]++
 	t.tokens[t.n] = token(matchType | xlength<<lengthShift | xoffset)
 	t.n++
@@ -301,7 +306,6 @@ func (t *tokens) AddMatchLong(xlength int32, xoffset uint32) {
 		}
 		xlength -= xl
 		xl -= baseMatchLength
-		t.nLits++
 		t.extraHist[lengthCodes1[uint8(xl)]]++
 		t.offHist[oc]++
 		t.tokens[t.n] = token(matchType | uint32(xl)<<lengthShift | xoffset)