[feature] add per-uri dereferencer locks (#2291)

This commit is contained in:
kim
2023-10-31 11:12:22 +00:00
committed by GitHub
parent 51d0a0bba5
commit ce71a5a790
54 changed files with 2432 additions and 2719 deletions

View File

@ -7,6 +7,7 @@ package flate
import (
"encoding/binary"
"errors"
"fmt"
"io"
"math"
@ -833,6 +834,12 @@ func (d *compressor) init(w io.Writer, level int) (err error) {
d.initDeflate()
d.fill = (*compressor).fillDeflate
d.step = (*compressor).deflateLazy
case -level >= MinCustomWindowSize && -level <= MaxCustomWindowSize:
d.w.logNewTablePenalty = 7
d.fast = &fastEncL5Window{maxOffset: int32(-level), cur: maxStoreBlockSize}
d.window = make([]byte, maxStoreBlockSize)
d.fill = (*compressor).fillBlock
d.step = (*compressor).storeFast
default:
return fmt.Errorf("flate: invalid compression level %d: want value in range [-2, 9]", level)
}
@ -929,6 +936,28 @@ func NewWriterDict(w io.Writer, level int, dict []byte) (*Writer, error) {
return zw, err
}
// MinCustomWindowSize is the minimum window size that can be sent to NewWriterWindow.
const MinCustomWindowSize = 32
// MaxCustomWindowSize is the maximum custom window that can be sent to NewWriterWindow.
const MaxCustomWindowSize = windowSize
// NewWriterWindow returns a new Writer compressing data with a custom window size.
// windowSize must be from MinCustomWindowSize to MaxCustomWindowSize.
func NewWriterWindow(w io.Writer, windowSize int) (*Writer, error) {
if windowSize < MinCustomWindowSize {
return nil, errors.New("flate: requested window size less than MinWindowSize")
}
if windowSize > MaxCustomWindowSize {
return nil, errors.New("flate: requested window size bigger than MaxCustomWindowSize")
}
var dw Writer
if err := dw.d.init(w, -windowSize); err != nil {
return nil, err
}
return &dw, nil
}
// A Writer takes data written to it and writes the compressed
// form of that data to an underlying writer (see NewWriter).
type Writer struct {

View File

@ -8,7 +8,6 @@ package flate
import (
"encoding/binary"
"fmt"
"math/bits"
)
type fastEnc interface {
@ -192,25 +191,3 @@ func (e *fastGen) Reset() {
}
e.hist = e.hist[:0]
}
// matchLen returns the maximum length.
// 'a' must be the shortest of the two.
func matchLen(a, b []byte) int {
var checked int
for len(a) >= 8 {
if diff := binary.LittleEndian.Uint64(a) ^ binary.LittleEndian.Uint64(b); diff != 0 {
return checked + (bits.TrailingZeros64(diff) >> 3)
}
checked += 8
a = a[8:]
b = b[8:]
}
b = b[:len(a)]
for i := range a {
if a[i] != b[i] {
return i + checked
}
}
return len(a) + checked
}

View File

@ -120,8 +120,9 @@ func (h *huffmanDecoder) init(lengths []int) bool {
const sanity = false
if h.chunks == nil {
h.chunks = &[huffmanNumChunks]uint16{}
h.chunks = new([huffmanNumChunks]uint16)
}
if h.maxRead != 0 {
*h = huffmanDecoder{chunks: h.chunks, links: h.links}
}
@ -175,6 +176,7 @@ func (h *huffmanDecoder) init(lengths []int) bool {
}
h.maxRead = min
chunks := h.chunks[:]
for i := range chunks {
chunks[i] = 0
@ -202,8 +204,7 @@ func (h *huffmanDecoder) init(lengths []int) bool {
if cap(h.links[off]) < numLinks {
h.links[off] = make([]uint16, numLinks)
} else {
links := h.links[off][:0]
h.links[off] = links[:numLinks]
h.links[off] = h.links[off][:numLinks]
}
}
} else {
@ -277,7 +278,7 @@ func (h *huffmanDecoder) init(lengths []int) bool {
return true
}
// The actual read interface needed by NewReader.
// Reader is the actual read interface needed by NewReader.
// If the passed in io.Reader does not also have ReadByte,
// the NewReader will introduce its own buffering.
type Reader interface {
@ -285,6 +286,18 @@ type Reader interface {
io.ByteReader
}
type step uint8
const (
copyData step = iota + 1
nextBlock
huffmanBytesBuffer
huffmanBytesReader
huffmanBufioReader
huffmanStringsReader
huffmanGenericReader
)
// Decompress state.
type decompressor struct {
// Input source.
@ -303,7 +316,7 @@ type decompressor struct {
// Next step in the decompression,
// and decompression state.
step func(*decompressor)
step step
stepState int
err error
toRead []byte
@ -342,7 +355,7 @@ func (f *decompressor) nextBlock() {
// compressed, fixed Huffman tables
f.hl = &fixedHuffmanDecoder
f.hd = nil
f.huffmanBlockDecoder()()
f.huffmanBlockDecoder()
if debugDecode {
fmt.Println("predefinied huffman block")
}
@ -353,7 +366,7 @@ func (f *decompressor) nextBlock() {
}
f.hl = &f.h1
f.hd = &f.h2
f.huffmanBlockDecoder()()
f.huffmanBlockDecoder()
if debugDecode {
fmt.Println("dynamic huffman block")
}
@ -379,14 +392,16 @@ func (f *decompressor) Read(b []byte) (int, error) {
if f.err != nil {
return 0, f.err
}
f.step(f)
f.doStep()
if f.err != nil && len(f.toRead) == 0 {
f.toRead = f.dict.readFlush() // Flush what's left in case of error
}
}
}
// Support the io.WriteTo interface for io.Copy and friends.
// WriteTo implements the io.WriteTo interface for io.Copy and friends.
func (f *decompressor) WriteTo(w io.Writer) (int64, error) {
total := int64(0)
flushed := false
@ -410,7 +425,7 @@ func (f *decompressor) WriteTo(w io.Writer) (int64, error) {
return total, f.err
}
if f.err == nil {
f.step(f)
f.doStep()
}
if len(f.toRead) == 0 && f.err != nil && !flushed {
f.toRead = f.dict.readFlush() // Flush what's left in case of error
@ -631,7 +646,7 @@ func (f *decompressor) copyData() {
if f.dict.availWrite() == 0 || f.copyLen > 0 {
f.toRead = f.dict.readFlush()
f.step = (*decompressor).copyData
f.step = copyData
return
}
f.finishBlock()
@ -644,7 +659,28 @@ func (f *decompressor) finishBlock() {
}
f.err = io.EOF
}
f.step = (*decompressor).nextBlock
f.step = nextBlock
}
func (f *decompressor) doStep() {
switch f.step {
case copyData:
f.copyData()
case nextBlock:
f.nextBlock()
case huffmanBytesBuffer:
f.huffmanBytesBuffer()
case huffmanBytesReader:
f.huffmanBytesReader()
case huffmanBufioReader:
f.huffmanBufioReader()
case huffmanStringsReader:
f.huffmanStringsReader()
case huffmanGenericReader:
f.huffmanGenericReader()
default:
panic("BUG: unexpected step state")
}
}
// noEOF returns err, unless err == io.EOF, in which case it returns io.ErrUnexpectedEOF.
@ -747,7 +783,7 @@ func (f *decompressor) Reset(r io.Reader, dict []byte) error {
h1: f.h1,
h2: f.h2,
dict: f.dict,
step: (*decompressor).nextBlock,
step: nextBlock,
}
f.dict.init(maxMatchOffset, dict)
return nil
@ -768,7 +804,7 @@ func NewReader(r io.Reader) io.ReadCloser {
f.r = makeReader(r)
f.bits = new([maxNumLit + maxNumDist]int)
f.codebits = new([numCodes]int)
f.step = (*decompressor).nextBlock
f.step = nextBlock
f.dict.init(maxMatchOffset, nil)
return &f
}
@ -787,7 +823,7 @@ func NewReaderDict(r io.Reader, dict []byte) io.ReadCloser {
f.r = makeReader(r)
f.bits = new([maxNumLit + maxNumDist]int)
f.codebits = new([numCodes]int)
f.step = (*decompressor).nextBlock
f.step = nextBlock
f.dict.init(maxMatchOffset, dict)
return &f
}

View File

@ -85,7 +85,7 @@ readLiteral:
dict.writeByte(byte(v))
if dict.availWrite() == 0 {
f.toRead = dict.readFlush()
f.step = (*decompressor).huffmanBytesBuffer
f.step = huffmanBytesBuffer
f.stepState = stateInit
f.b, f.nb = fb, fnb
return
@ -251,7 +251,7 @@ copyHistory:
if dict.availWrite() == 0 || f.copyLen > 0 {
f.toRead = dict.readFlush()
f.step = (*decompressor).huffmanBytesBuffer // We need to continue this work
f.step = huffmanBytesBuffer // We need to continue this work
f.stepState = stateDict
f.b, f.nb = fb, fnb
return
@ -336,7 +336,7 @@ readLiteral:
dict.writeByte(byte(v))
if dict.availWrite() == 0 {
f.toRead = dict.readFlush()
f.step = (*decompressor).huffmanBytesReader
f.step = huffmanBytesReader
f.stepState = stateInit
f.b, f.nb = fb, fnb
return
@ -502,7 +502,7 @@ copyHistory:
if dict.availWrite() == 0 || f.copyLen > 0 {
f.toRead = dict.readFlush()
f.step = (*decompressor).huffmanBytesReader // We need to continue this work
f.step = huffmanBytesReader // We need to continue this work
f.stepState = stateDict
f.b, f.nb = fb, fnb
return
@ -587,7 +587,7 @@ readLiteral:
dict.writeByte(byte(v))
if dict.availWrite() == 0 {
f.toRead = dict.readFlush()
f.step = (*decompressor).huffmanBufioReader
f.step = huffmanBufioReader
f.stepState = stateInit
f.b, f.nb = fb, fnb
return
@ -753,7 +753,7 @@ copyHistory:
if dict.availWrite() == 0 || f.copyLen > 0 {
f.toRead = dict.readFlush()
f.step = (*decompressor).huffmanBufioReader // We need to continue this work
f.step = huffmanBufioReader // We need to continue this work
f.stepState = stateDict
f.b, f.nb = fb, fnb
return
@ -838,7 +838,7 @@ readLiteral:
dict.writeByte(byte(v))
if dict.availWrite() == 0 {
f.toRead = dict.readFlush()
f.step = (*decompressor).huffmanStringsReader
f.step = huffmanStringsReader
f.stepState = stateInit
f.b, f.nb = fb, fnb
return
@ -1004,7 +1004,7 @@ copyHistory:
if dict.availWrite() == 0 || f.copyLen > 0 {
f.toRead = dict.readFlush()
f.step = (*decompressor).huffmanStringsReader // We need to continue this work
f.step = huffmanStringsReader // We need to continue this work
f.stepState = stateDict
f.b, f.nb = fb, fnb
return
@ -1089,7 +1089,7 @@ readLiteral:
dict.writeByte(byte(v))
if dict.availWrite() == 0 {
f.toRead = dict.readFlush()
f.step = (*decompressor).huffmanGenericReader
f.step = huffmanGenericReader
f.stepState = stateInit
f.b, f.nb = fb, fnb
return
@ -1255,7 +1255,7 @@ copyHistory:
if dict.availWrite() == 0 || f.copyLen > 0 {
f.toRead = dict.readFlush()
f.step = (*decompressor).huffmanGenericReader // We need to continue this work
f.step = huffmanGenericReader // We need to continue this work
f.stepState = stateDict
f.b, f.nb = fb, fnb
return
@ -1265,19 +1265,19 @@ copyHistory:
// Not reached
}
func (f *decompressor) huffmanBlockDecoder() func() {
func (f *decompressor) huffmanBlockDecoder() {
switch f.r.(type) {
case *bytes.Buffer:
return f.huffmanBytesBuffer
f.huffmanBytesBuffer()
case *bytes.Reader:
return f.huffmanBytesReader
f.huffmanBytesReader()
case *bufio.Reader:
return f.huffmanBufioReader
f.huffmanBufioReader()
case *strings.Reader:
return f.huffmanStringsReader
f.huffmanStringsReader()
case Reader:
return f.huffmanGenericReader
f.huffmanGenericReader()
default:
return f.huffmanGenericReader
f.huffmanGenericReader()
}
}

View File

@ -308,3 +308,401 @@ emitRemainder:
emitLiteral(dst, src[nextEmit:])
}
}
// fastEncL5Window is a level 5 encoder,
// but with a custom window size.
type fastEncL5Window struct {
hist []byte
cur int32
maxOffset int32
table [tableSize]tableEntry
bTable [tableSize]tableEntryPrev
}
func (e *fastEncL5Window) Encode(dst *tokens, src []byte) {
const (
inputMargin = 12 - 1
minNonLiteralBlockSize = 1 + 1 + inputMargin
hashShortBytes = 4
)
maxMatchOffset := e.maxOffset
if debugDeflate && e.cur < 0 {
panic(fmt.Sprint("e.cur < 0: ", e.cur))
}
// 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.bTable[:] {
e.bTable[i] = tableEntryPrev{}
}
e.cur = maxMatchOffset
break
}
// Shift down everything in the table that isn't already too far away.
minOff := e.cur + int32(len(e.hist)) - maxMatchOffset
for i := range e.table[:] {
v := e.table[i].offset
if v <= minOff {
v = 0
} else {
v = v - e.cur + maxMatchOffset
}
e.table[i].offset = v
}
for i := range e.bTable[:] {
v := e.bTable[i]
if v.Cur.offset <= minOff {
v.Cur.offset = 0
v.Prev.offset = 0
} else {
v.Cur.offset = v.Cur.offset - e.cur + maxMatchOffset
if v.Prev.offset <= minOff {
v.Prev.offset = 0
} else {
v.Prev.offset = v.Prev.offset - e.cur + maxMatchOffset
}
}
e.bTable[i] = v
}
e.cur = maxMatchOffset
}
s := e.addBlock(src)
// This check isn't in the Snappy implementation, but there, the caller
// instead of the callee handles this case.
if len(src) < minNonLiteralBlockSize {
// We do not fill the token table.
// This will be picked up by caller.
dst.n = uint16(len(src))
return
}
// Override src
src = e.hist
nextEmit := s
// sLimit is when to stop looking for offset/length copies. The inputMargin
// lets us use a fast path for emitLiteral in the main loop, while we are
// looking for copies.
sLimit := int32(len(src) - inputMargin)
// nextEmit is where in src the next emitLiteral should start from.
cv := load6432(src, s)
for {
const skipLog = 6
const doEvery = 1
nextS := s
var l int32
var t int32
for {
nextHashS := hashLen(cv, tableBits, hashShortBytes)
nextHashL := hash7(cv, tableBits)
s = nextS
nextS = s + doEvery + (s-nextEmit)>>skipLog
if nextS > sLimit {
goto emitRemainder
}
// Fetch a short+long candidate
sCandidate := e.table[nextHashS]
lCandidate := e.bTable[nextHashL]
next := load6432(src, nextS)
entry := tableEntry{offset: s + e.cur}
e.table[nextHashS] = entry
eLong := &e.bTable[nextHashL]
eLong.Cur, eLong.Prev = entry, eLong.Cur
nextHashS = hashLen(next, tableBits, hashShortBytes)
nextHashL = hash7(next, tableBits)
t = lCandidate.Cur.offset - e.cur
if s-t < maxMatchOffset {
if uint32(cv) == load3232(src, lCandidate.Cur.offset-e.cur) {
// Store the next match
e.table[nextHashS] = tableEntry{offset: nextS + e.cur}
eLong := &e.bTable[nextHashL]
eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur}, eLong.Cur
t2 := lCandidate.Prev.offset - e.cur
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 {
t = t2
l = ml1
break
}
}
break
}
t = lCandidate.Prev.offset - e.cur
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}
eLong := &e.bTable[nextHashL]
eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur}, eLong.Cur
break
}
}
t = sCandidate.offset - e.cur
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}
eLong := &e.bTable[nextHashL]
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 load3232(src, lCandidate.Cur.offset-e.cur) == uint32(next) {
ml := e.matchlen(nextS+4, t2+4, src) + 4
if ml > l {
t = t2
s = nextS
l = ml
break
}
}
// If the previous long is a candidate, use that...
t2 = lCandidate.Prev.offset - e.cur
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
s = nextS
l = ml
break
}
}
}
break
}
cv = next
}
// A 4-byte match has been found. We'll later see if more than 4 bytes
// match. But, prior to the match, src[nextEmit:s] are unmatched. Emit
// them as literal bytes.
if l == 0 {
// Extend the 4-byte match as long as possible.
l = e.matchlenLong(s+4, t+4, src) + 4
} else if l == maxMatchLength {
l += e.matchlenLong(s+l, t+l, src)
}
// Try to locate a better match by checking the end of best match...
if sAt := s + l; l < 30 && sAt < sLimit {
// Allow some bytes at the beginning to mismatch.
// Sweet spot is 2/3 bytes depending on input.
// 3 is only a little better when it is but sometimes a lot worse.
// The skipped bytes are tested in Extend backwards,
// and still picked up as part of the match if they do.
const skipBeginning = 2
eLong := e.bTable[hash7(load6432(src, sAt), tableBits)].Cur.offset
t2 := eLong - e.cur - l + skipBeginning
s2 := s + skipBeginning
off := s2 - t2
if t2 >= 0 && off < maxMatchOffset && off > 0 {
if l2 := e.matchlenLong(s2, t2, src); l2 > l {
t = t2
l = l2
s = s2
}
}
}
// Extend backwards
for t > 0 && s > nextEmit && src[t-1] == src[s-1] {
s--
t--
l++
}
if nextEmit < s {
if false {
emitLiteral(dst, src[nextEmit:s])
} else {
for _, v := range src[nextEmit:s] {
dst.tokens[dst.n] = token(v)
dst.litHist[v]++
dst.n++
}
}
}
if debugDeflate {
if t >= s {
panic(fmt.Sprintln("s-t", s, t))
}
if (s - t) > maxMatchOffset {
panic(fmt.Sprintln("mmo", s-t))
}
if l < baseMatchLength {
panic("bml")
}
}
dst.AddMatchLong(l, uint32(s-t-baseMatchOffset))
s += l
nextEmit = s
if nextS >= s {
s = nextS + 1
}
if s >= sLimit {
goto emitRemainder
}
// Store every 3rd hash in-between.
if true {
const hashEvery = 3
i := s - l + 1
if i < s-1 {
cv := load6432(src, i)
t := tableEntry{offset: i + e.cur}
e.table[hashLen(cv, tableBits, hashShortBytes)] = 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}
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}
e.table[hashLen(cv, tableBits, hashShortBytes)] = 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}
t2 := tableEntry{offset: t.offset + 1}
eLong := &e.bTable[hash7(cv, tableBits)]
eLong.Cur, eLong.Prev = t, eLong.Cur
e.table[hashLen(cv>>8, tableBits, hashShortBytes)] = t2
}
}
}
// We could immediately start working at s now, but to improve
// compression we first update the hash table at s-1 and at s.
x := load6432(src, s-1)
o := e.cur + s - 1
prevHashS := hashLen(x, tableBits, hashShortBytes)
prevHashL := hash7(x, tableBits)
e.table[prevHashS] = tableEntry{offset: o}
eLong := &e.bTable[prevHashL]
eLong.Cur, eLong.Prev = tableEntry{offset: o}, eLong.Cur
cv = x >> 8
}
emitRemainder:
if int(nextEmit) < len(src) {
// If nothing was added, don't encode literals.
if dst.n == 0 {
return
}
emitLiteral(dst, src[nextEmit:])
}
}
// Reset the encoding table.
func (e *fastEncL5Window) Reset() {
// We keep the same allocs, since we are compressing the same block sizes.
if cap(e.hist) < allocHistory {
e.hist = make([]byte, 0, allocHistory)
}
// We offset current position so everything will be out of reach.
// If we are above the buffer reset it will be cleared anyway since len(hist) == 0.
if e.cur <= int32(bufferReset) {
e.cur += e.maxOffset + int32(len(e.hist))
}
e.hist = e.hist[:0]
}
func (e *fastEncL5Window) addBlock(src []byte) int32 {
// check if we have space already
maxMatchOffset := e.maxOffset
if len(e.hist)+len(src) > cap(e.hist) {
if cap(e.hist) == 0 {
e.hist = make([]byte, 0, allocHistory)
} else {
if cap(e.hist) < int(maxMatchOffset*2) {
panic("unexpected buffer size")
}
// Move down
offset := int32(len(e.hist)) - maxMatchOffset
copy(e.hist[0:maxMatchOffset], e.hist[offset:])
e.cur += offset
e.hist = e.hist[:maxMatchOffset]
}
}
s := int32(len(e.hist))
e.hist = append(e.hist, src...)
return s
}
// matchlen will return the match length between offsets and t in src.
// The maximum length returned is maxMatchLength - 4.
// It is assumed that s > t, that t >=0 and s < len(src).
func (e *fastEncL5Window) matchlen(s, t int32, src []byte) int32 {
if debugDecode {
if t >= s {
panic(fmt.Sprint("t >=s:", t, s))
}
if int(s) >= len(src) {
panic(fmt.Sprint("s >= len(src):", s, len(src)))
}
if t < 0 {
panic(fmt.Sprint("t < 0:", t))
}
if s-t > e.maxOffset {
panic(fmt.Sprint(s, "-", t, "(", s-t, ") > maxMatchLength (", maxMatchOffset, ")"))
}
}
s1 := int(s) + maxMatchLength - 4
if s1 > len(src) {
s1 = len(src)
}
// Extend the match to be as long as possible.
return int32(matchLen(src[s:s1], src[t:]))
}
// matchlenLong will return the match length between offsets and t in src.
// It is assumed that s > t, that t >=0 and s < len(src).
func (e *fastEncL5Window) matchlenLong(s, t int32, src []byte) int32 {
if debugDeflate {
if t >= s {
panic(fmt.Sprint("t >=s:", t, s))
}
if int(s) >= len(src) {
panic(fmt.Sprint("s >= len(src):", s, len(src)))
}
if t < 0 {
panic(fmt.Sprint("t < 0:", t))
}
if s-t > e.maxOffset {
panic(fmt.Sprint(s, "-", t, "(", s-t, ") > maxMatchLength (", maxMatchOffset, ")"))
}
}
// Extend the match to be as long as possible.
return int32(matchLen(src[s:], src[t:]))
}

View File

@ -0,0 +1,16 @@
//go:build amd64 && !appengine && !noasm && gc
// +build amd64,!appengine,!noasm,gc
// Copyright 2019+ Klaus Post. All rights reserved.
// License information can be found in the LICENSE file.
package flate
// matchLen returns how many bytes match in a and b
//
// It assumes that:
//
// len(a) <= len(b) and len(a) > 0
//
//go:noescape
func matchLen(a []byte, b []byte) int

View File

@ -0,0 +1,68 @@
// Copied from S2 implementation.
//go:build !appengine && !noasm && gc && !noasm
#include "textflag.h"
// func matchLen(a []byte, b []byte) int
// Requires: BMI
TEXT ·matchLen(SB), NOSPLIT, $0-56
MOVQ a_base+0(FP), AX
MOVQ b_base+24(FP), CX
MOVQ a_len+8(FP), DX
// matchLen
XORL SI, SI
CMPL DX, $0x08
JB matchlen_match4_standalone
matchlen_loopback_standalone:
MOVQ (AX)(SI*1), BX
XORQ (CX)(SI*1), BX
TESTQ BX, BX
JZ matchlen_loop_standalone
#ifdef GOAMD64_v3
TZCNTQ BX, BX
#else
BSFQ BX, BX
#endif
SARQ $0x03, BX
LEAL (SI)(BX*1), SI
JMP gen_match_len_end
matchlen_loop_standalone:
LEAL -8(DX), DX
LEAL 8(SI), SI
CMPL DX, $0x08
JAE matchlen_loopback_standalone
matchlen_match4_standalone:
CMPL DX, $0x04
JB matchlen_match2_standalone
MOVL (AX)(SI*1), BX
CMPL (CX)(SI*1), BX
JNE matchlen_match2_standalone
LEAL -4(DX), DX
LEAL 4(SI), SI
matchlen_match2_standalone:
CMPL DX, $0x02
JB matchlen_match1_standalone
MOVW (AX)(SI*1), BX
CMPW (CX)(SI*1), BX
JNE matchlen_match1_standalone
LEAL -2(DX), DX
LEAL 2(SI), SI
matchlen_match1_standalone:
CMPL DX, $0x01
JB gen_match_len_end
MOVB (AX)(SI*1), BL
CMPB (CX)(SI*1), BL
JNE gen_match_len_end
INCL SI
gen_match_len_end:
MOVQ SI, ret+48(FP)
RET

View File

@ -0,0 +1,33 @@
//go:build !amd64 || appengine || !gc || noasm
// +build !amd64 appengine !gc noasm
// Copyright 2019+ Klaus Post. All rights reserved.
// License information can be found in the LICENSE file.
package flate
import (
"encoding/binary"
"math/bits"
)
// matchLen returns the maximum common prefix length of a and b.
// a must be the shortest of the two.
func matchLen(a, b []byte) (n int) {
for ; len(a) >= 8 && len(b) >= 8; a, b = a[8:], b[8:] {
diff := binary.LittleEndian.Uint64(a) ^ binary.LittleEndian.Uint64(b)
if diff != 0 {
return n + bits.TrailingZeros64(diff)>>3
}
n += 8
}
for i := range a {
if a[i] != b[i] {
break
}
n++
}
return n
}

View File

@ -106,6 +106,7 @@ func (z *Reader) Reset(r io.Reader) error {
*z = Reader{
decompressor: z.decompressor,
multistream: true,
br: z.br,
}
if rr, ok := r.(flate.Reader); ok {
z.r = rr

View File

@ -74,6 +74,27 @@ func NewWriterLevel(w io.Writer, level int) (*Writer, error) {
return z, nil
}
// MinCustomWindowSize is the minimum window size that can be sent to NewWriterWindow.
const MinCustomWindowSize = flate.MinCustomWindowSize
// MaxCustomWindowSize is the maximum custom window that can be sent to NewWriterWindow.
const MaxCustomWindowSize = flate.MaxCustomWindowSize
// NewWriterWindow returns a new Writer compressing data with a custom window size.
// windowSize must be from MinCustomWindowSize to MaxCustomWindowSize.
func NewWriterWindow(w io.Writer, windowSize int) (*Writer, error) {
if windowSize < MinCustomWindowSize {
return nil, errors.New("gzip: requested window size less than MinWindowSize")
}
if windowSize > MaxCustomWindowSize {
return nil, errors.New("gzip: requested window size bigger than MaxCustomWindowSize")
}
z := new(Writer)
z.init(w, -windowSize)
return z, nil
}
func (z *Writer) init(w io.Writer, level int) {
compressor := z.compressor
if level != StatelessCompression {

View File

@ -106,6 +106,25 @@ func MakeDict(data []byte, searchStart []byte) *Dict {
return &d
}
// MakeDictManual will create a dictionary.
// 'data' must be at least MinDictSize and less than or equal to MaxDictSize.
// A manual first repeat index into data must be provided.
// It must be less than len(data)-8.
func MakeDictManual(data []byte, firstIdx uint16) *Dict {
if len(data) < MinDictSize || int(firstIdx) >= len(data)-8 || len(data) > MaxDictSize {
return nil
}
var d Dict
dict := data
d.dict = dict
if cap(d.dict) < len(d.dict)+16 {
d.dict = append(make([]byte, 0, len(d.dict)+16), d.dict...)
}
d.repeat = int(firstIdx)
return &d
}
// Encode returns the encoded form of src. The returned slice may be a sub-
// slice of dst if dst was large enough to hold the entire encoded block.
// Otherwise, a newly allocated slice will be returned.

View File

@ -57,7 +57,7 @@ func Encode(dst, src []byte) []byte {
// The function returns -1 if no improvement could be achieved.
// Using actual compression will most often produce better compression than the estimate.
func EstimateBlockSize(src []byte) (d int) {
if len(src) < 6 || int64(len(src)) > 0xffffffff {
if len(src) <= inputMargin || int64(len(src)) > 0xffffffff {
return -1
}
if len(src) <= 1024 {

View File

@ -157,6 +157,9 @@ func encodeBlockBest(dst, src []byte, dict *Dict) (d int) {
return m
}
matchDict := func(candidate, s int, first uint32, rep bool) match {
if s >= MaxDictSrcOffset {
return match{offset: candidate, s: s}
}
// Calculate offset as if in continuous array with s
offset := -len(dict.dict) + candidate
if best.length != 0 && best.s-best.offset == s-offset && !rep {

View File

@ -316,6 +316,7 @@ func matchLen(a []byte, b []byte) int {
return len(a) + checked
}
// input must be > inputMargin
func calcBlockSize(src []byte) (d int) {
// Initialize the hash table.
const (
@ -501,6 +502,7 @@ emitRemainder:
return d
}
// length must be > inputMargin.
func calcBlockSizeSmall(src []byte) (d int) {
// Initialize the hash table.
const (

File diff suppressed because it is too large Load Diff

View File

@ -511,24 +511,22 @@ func IndexStream(r io.Reader) ([]byte, error) {
// JSON returns the index as JSON text.
func (i *Index) JSON() []byte {
type offset struct {
CompressedOffset int64 `json:"compressed"`
UncompressedOffset int64 `json:"uncompressed"`
}
x := struct {
TotalUncompressed int64 `json:"total_uncompressed"` // Total Uncompressed size if known. Will be -1 if unknown.
TotalCompressed int64 `json:"total_compressed"` // Total Compressed size if known. Will be -1 if unknown.
Offsets []struct {
CompressedOffset int64 `json:"compressed"`
UncompressedOffset int64 `json:"uncompressed"`
} `json:"offsets"`
EstBlockUncomp int64 `json:"est_block_uncompressed"`
TotalUncompressed int64 `json:"total_uncompressed"` // Total Uncompressed size if known. Will be -1 if unknown.
TotalCompressed int64 `json:"total_compressed"` // Total Compressed size if known. Will be -1 if unknown.
Offsets []offset `json:"offsets"`
EstBlockUncomp int64 `json:"est_block_uncompressed"`
}{
TotalUncompressed: i.TotalUncompressed,
TotalCompressed: i.TotalCompressed,
EstBlockUncomp: i.estBlockUncomp,
}
for _, v := range i.info {
x.Offsets = append(x.Offsets, struct {
CompressedOffset int64 `json:"compressed"`
UncompressedOffset int64 `json:"uncompressed"`
}{CompressedOffset: v.compressedOffset, UncompressedOffset: v.uncompressedOffset})
x.Offsets = append(x.Offsets, offset{CompressedOffset: v.compressedOffset, UncompressedOffset: v.uncompressedOffset})
}
b, _ := json.MarshalIndent(x, "", " ")
return b