sevenhill/QSfera
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Bump github.com/nats-io/nats-server/v2 from 2.9.4 to 2.9.17 (#6373)

Browse Source 
Bumps [github.com/nats-io/nats-server/v2](https://github.com/nats-io/nats-server) from 2.9.4 to 2.9.17.
- [Release notes](https://github.com/nats-io/nats-server/releases)
- [Changelog](https://github.com/nats-io/nats-server/blob/main/.goreleaser.yml)
- [Commits](https://github.com/nats-io/nats-server/compare/v2.9.4...v2.9.17)

---
updated-dependencies:
- dependency-name: github.com/nats-io/nats-server/v2
  dependency-type: direct:production
  update-type: version-update:semver-patch
...

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
This commit is contained in:
dependabot[bot]
2023-05-27 19:06:42 +02:00
committed by GitHub
co-authored by dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
parent c0665975b3
commit 1ad0637254
113 changed files with 33307 additions and 13806 deletions
Download Patch File Download Diff File Expand all files Collapse all files
vendor/github.com/klauspost/compress/flate/deflate.go
Generated Vendored
+989
View File
@@ -0,0 +1,989 @@
// Copyright 2009 The Go Authors. All rights reserved.
// Copyright (c) 2015 Klaus Post
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package flate
import (
"encoding/binary"
"fmt"
"io"
"math"
)
const (
NoCompression = 0
BestSpeed = 1
BestCompression = 9
DefaultCompression = -1
// HuffmanOnly disables Lempel-Ziv match searching and only performs Huffman
// entropy encoding. This mode is useful in compressing data that has
// already been compressed with an LZ style algorithm (e.g. Snappy or LZ4)
// that lacks an entropy encoder. Compression gains are achieved when
// certain bytes in the input stream occur more frequently than others.
//
// Note that HuffmanOnly produces a compressed output that is
// RFC 1951 compliant. That is, any valid DEFLATE decompressor will
// continue to be able to decompress this output.
HuffmanOnly = -2
ConstantCompression = HuffmanOnly // compatibility alias.
logWindowSize = 15
windowSize = 1 << logWindowSize
windowMask = windowSize - 1
logMaxOffsetSize = 15 // Standard DEFLATE
minMatchLength = 4 // The smallest match that the compressor looks for
maxMatchLength = 258 // The longest match for the compressor
minOffsetSize = 1 // The shortest offset that makes any sense
// The maximum number of tokens we will encode at the time.
// Smaller sizes usually creates less optimal blocks.
// Bigger can make context switching slow.
// We use this for levels 7-9, so we make it big.
maxFlateBlockTokens = 1 << 15
maxStoreBlockSize = 65535
hashBits = 17 // After 17 performance degrades
hashSize = 1 << hashBits
hashMask = (1 << hashBits) - 1
hashShift = (hashBits + minMatchLength - 1) / minMatchLength
maxHashOffset = 1 << 28
skipNever = math.MaxInt32
debugDeflate = false
)
type compressionLevel struct {
good, lazy, nice, chain, fastSkipHashing, level int
}
// Compression levels have been rebalanced from zlib deflate defaults
// to give a bigger spread in speed and compression.
// See https://blog.klauspost.com/rebalancing-deflate-compression-levels/
var levels = []compressionLevel{
{}, // 0
// Level 1-6 uses specialized algorithm - values not used
{0, 0, 0, 0, 0, 1},
{0, 0, 0, 0, 0, 2},
{0, 0, 0, 0, 0, 3},
{0, 0, 0, 0, 0, 4},
{0, 0, 0, 0, 0, 5},
{0, 0, 0, 0, 0, 6},
// Levels 7-9 use increasingly more lazy matching
// and increasingly stringent conditions for "good enough".
{8, 12, 16, 24, skipNever, 7},
{16, 30, 40, 64, skipNever, 8},
{32, 258, 258, 1024, skipNever, 9},
}
// advancedState contains state for the advanced levels, with bigger hash tables, etc.
type advancedState struct {
// deflate state
length int
offset int
maxInsertIndex int
chainHead int
hashOffset int
ii uint16 // position of last match, intended to overflow to reset.
// input window: unprocessed data is window[index:windowEnd]
index int
estBitsPerByte int
hashMatch [maxMatchLength + minMatchLength]uint32
// Input hash chains
// hashHead[hashValue] contains the largest inputIndex with the specified hash value
// If hashHead[hashValue] is within the current window, then
// hashPrev[hashHead[hashValue] & windowMask] contains the previous index
// with the same hash value.
hashHead [hashSize]uint32
hashPrev [windowSize]uint32
}
type compressor struct {
compressionLevel
h *huffmanEncoder
w *huffmanBitWriter
// compression algorithm
fill func(*compressor, []byte) int // copy data to window
step func(*compressor) // process window
window []byte
windowEnd int
blockStart int // window index where current tokens start
err error
// queued output tokens
tokens tokens
fast fastEnc
state *advancedState
sync bool // requesting flush
byteAvailable bool // if true, still need to process window[index-1].
}
func (d *compressor) fillDeflate(b []byte) int {
s := d.state
if s.index >= 2*windowSize-(minMatchLength+maxMatchLength) {
// shift the window by windowSize
//copy(d.window[:], d.window[windowSize:2*windowSize])
*(*[windowSize]byte)(d.window) = *(*[windowSize]byte)(d.window[windowSize:])
s.index -= windowSize
d.windowEnd -= windowSize
if d.blockStart >= windowSize {
d.blockStart -= windowSize
} else {
d.blockStart = math.MaxInt32
}
s.hashOffset += windowSize
if s.hashOffset > maxHashOffset {
delta := s.hashOffset - 1
s.hashOffset -= delta
s.chainHead -= delta
// Iterate over slices instead of arrays to avoid copying
// the entire table onto the stack (Issue #18625).
for i, v := range s.hashPrev[:] {
if int(v) > delta {
s.hashPrev[i] = uint32(int(v) - delta)
} else {
s.hashPrev[i] = 0
}
}
for i, v := range s.hashHead[:] {
if int(v) > delta {
s.hashHead[i] = uint32(int(v) - delta)
} else {
s.hashHead[i] = 0
}
}
}
}
n := copy(d.window[d.windowEnd:], b)
d.windowEnd += n
return n
}
func (d *compressor) writeBlock(tok *tokens, index int, eof bool) error {
if index > 0 || eof {
var window []byte
if d.blockStart <= index {
window = d.window[d.blockStart:index]
}
d.blockStart = index
//d.w.writeBlock(tok, eof, window)
d.w.writeBlockDynamic(tok, eof, window, d.sync)
return d.w.err
}
return nil
}
// writeBlockSkip writes the current block and uses the number of tokens
// to determine if the block should be stored on no matches, or
// only huffman encoded.
func (d *compressor) writeBlockSkip(tok *tokens, index int, eof bool) error {
if index > 0 || eof {
if d.blockStart <= index {
window := d.window[d.blockStart:index]
// If we removed less than a 64th of all literals
// we huffman compress the block.
if int(tok.n) > len(window)-int(tok.n>>6) {
d.w.writeBlockHuff(eof, window, d.sync)
} else {
// Write a dynamic huffman block.
d.w.writeBlockDynamic(tok, eof, window, d.sync)
}
} else {
d.w.writeBlock(tok, eof, nil)
}
d.blockStart = index
return d.w.err
}
return nil
}
// fillWindow will fill the current window with the supplied
// dictionary and calculate all hashes.
// This is much faster than doing a full encode.
// Should only be used after a start/reset.
func (d *compressor) fillWindow(b []byte) {
// Do not fill window if we are in store-only or huffman mode.
if d.level <= 0 {
return
}
if d.fast != nil {
// encode the last data, but discard the result
if len(b) > maxMatchOffset {
b = b[len(b)-maxMatchOffset:]
}
d.fast.Encode(&d.tokens, b)
d.tokens.Reset()
return
}
s := d.state
// If we are given too much, cut it.
if len(b) > windowSize {
b = b[len(b)-windowSize:]
}
// Add all to window.
n := copy(d.window[d.windowEnd:], b)
// Calculate 256 hashes at the time (more L1 cache hits)
loops := (n + 256 - minMatchLength) / 256
for j := 0; j < loops; j++ {
startindex := j * 256
end := startindex + 256 + minMatchLength - 1
if end > n {
end = n
}
tocheck := d.window[startindex:end]
dstSize := len(tocheck) - minMatchLength + 1
if dstSize <= 0 {
continue
}
dst := s.hashMatch[:dstSize]
bulkHash4(tocheck, dst)
var newH uint32
for i, val := range dst {
di := i + startindex
newH = val & hashMask
// Get previous value with the same hash.
// Our chain should point to the previous value.
s.hashPrev[di&windowMask] = s.hashHead[newH]
// Set the head of the hash chain to us.
s.hashHead[newH] = uint32(di + s.hashOffset)
}
}
// Update window information.
d.windowEnd += n
s.index = n
}
// Try to find a match starting at index whose length is greater than prevSize.
// We only look at chainCount possibilities before giving up.
// pos = s.index, prevHead = s.chainHead-s.hashOffset, prevLength=minMatchLength-1, lookahead
func (d *compressor) findMatch(pos int, prevHead int, lookahead int) (length, offset int, ok bool) {
minMatchLook := maxMatchLength
if lookahead < minMatchLook {
minMatchLook = lookahead
}
win := d.window[0 : pos+minMatchLook]
// We quit when we get a match that's at least nice long
nice := len(win) - pos
if d.nice < nice {
nice = d.nice
}
// If we've got a match that's good enough, only look in 1/4 the chain.
tries := d.chain
length = minMatchLength - 1
wEnd := win[pos+length]
wPos := win[pos:]
minIndex := pos - windowSize
if minIndex < 0 {
minIndex = 0
}
offset = 0
if d.chain < 100 {
for i := prevHead; tries > 0; tries-- {
if wEnd == win[i+length] {
n := matchLen(win[i:i+minMatchLook], wPos)
if n > length {
length = n
offset = pos - i
ok = true
if n >= nice {
// The match is good enough that we don't try to find a better one.
break
}
wEnd = win[pos+n]
}
}
if i <= minIndex {
// hashPrev[i & windowMask] has already been overwritten, so stop now.
break
}
i = int(d.state.hashPrev[i&windowMask]) - d.state.hashOffset
if i < minIndex {
break
}
}
return
}
// Minimum gain to accept a match.
cGain := 4
// Some like it higher (CSV), some like it lower (JSON)
const baseCost = 3
// Base is 4 bytes at with an additional cost.
// Matches must be better than this.
for i := prevHead; tries > 0; tries-- {
if wEnd == win[i+length] {
n := matchLen(win[i:i+minMatchLook], wPos)
if n > length {
// Calculate gain. Estimate
newGain := d.h.bitLengthRaw(wPos[:n]) - int(offsetExtraBits[offsetCode(uint32(pos-i))]) - baseCost - int(lengthExtraBits[lengthCodes[(n-3)&255]])
//fmt.Println("gain:", newGain, "prev:", cGain, "raw:", d.h.bitLengthRaw(wPos[:n]), "this-len:", n, "prev-len:", length)
if newGain > cGain {
length = n
offset = pos - i
cGain = newGain
ok = true
if n >= nice {
// The match is good enough that we don't try to find a better one.
break
}
wEnd = win[pos+n]
}
}
}
if i <= minIndex {
// hashPrev[i & windowMask] has already been overwritten, so stop now.
break
}
i = int(d.state.hashPrev[i&windowMask]) - d.state.hashOffset
if i < minIndex {
break
}
}
return
}
func (d *compressor) writeStoredBlock(buf []byte) error {
if d.w.writeStoredHeader(len(buf), false); d.w.err != nil {
return d.w.err
}
d.w.writeBytes(buf)
return d.w.err
}
// hash4 returns a hash representation of the first 4 bytes
// of the supplied slice.
// The caller must ensure that len(b) >= 4.
func hash4(b []byte) uint32 {
return hash4u(binary.LittleEndian.Uint32(b), hashBits)
}
// hash4 returns the hash of u to fit in a hash table with h bits.
// Preferably h should be a constant and should always be <32.
func hash4u(u uint32, h uint8) uint32 {
return (u * prime4bytes) >> (32 - h)
}
// bulkHash4 will compute hashes using the same
// algorithm as hash4
func bulkHash4(b []byte, dst []uint32) {
if len(b) < 4 {
return
}
hb := binary.LittleEndian.Uint32(b)
dst[0] = hash4u(hb, hashBits)
end := len(b) - 4 + 1
for i := 1; i < end; i++ {
hb = (hb >> 8) | uint32(b[i+3])<<24
dst[i] = hash4u(hb, hashBits)
}
}
func (d *compressor) initDeflate() {
d.window = make([]byte, 2*windowSize)
d.byteAvailable = false
d.err = nil
if d.state == nil {
return
}
s := d.state
s.index = 0
s.hashOffset = 1
s.length = minMatchLength - 1
s.offset = 0
s.chainHead = -1
}
// deflateLazy is the same as deflate, but with d.fastSkipHashing == skipNever,
// meaning it always has lazy matching on.
func (d *compressor) deflateLazy() {
s := d.state
// Sanity enables additional runtime tests.
// It's intended to be used during development
// to supplement the currently ad-hoc unit tests.
const sanity = debugDeflate
if d.windowEnd-s.index < minMatchLength+maxMatchLength && !d.sync {
return
}
if d.windowEnd != s.index && d.chain > 100 {
// 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)
for {
if sanity && s.index > d.windowEnd {
panic("index > windowEnd")
}
lookahead := d.windowEnd - s.index
if lookahead < minMatchLength+maxMatchLength {
if !d.sync {
return
}
if sanity && s.index > d.windowEnd {
panic("index > windowEnd")
}
if lookahead == 0 {
// Flush current output block if any.
if d.byteAvailable {
// There is still one pending token that needs to be flushed
d.tokens.AddLiteral(d.window[s.index-1])
d.byteAvailable = false
}
if d.tokens.n > 0 {
if d.err = d.writeBlock(&d.tokens, s.index, false); d.err != nil {
return
}
d.tokens.Reset()
}
return
}
}
if s.index < s.maxInsertIndex {
// Update the hash
hash := hash4(d.window[s.index:])
ch := s.hashHead[hash]
s.chainHead = int(ch)
s.hashPrev[s.index&windowMask] = ch
s.hashHead[hash] = uint32(s.index + s.hashOffset)
}
prevLength := s.length
prevOffset := s.offset
s.length = minMatchLength - 1
s.offset = 0
minIndex := s.index - windowSize
if minIndex < 0 {
minIndex = 0
}
if s.chainHead-s.hashOffset >= minIndex && lookahead > prevLength && prevLength < d.lazy {
if newLength, newOffset, ok := d.findMatch(s.index, s.chainHead-s.hashOffset, lookahead); ok {
s.length = newLength
s.offset = newOffset
}
}
if prevLength >= minMatchLength && s.length <= prevLength {
// No better match, but check for better match at end...
//
// Skip forward a number of bytes.
// Offset of 2 seems to yield best results. 3 is sometimes better.
const checkOff = 2
// Check all, except full length
if prevLength < maxMatchLength-checkOff {
prevIndex := s.index - 1
if prevIndex+prevLength < s.maxInsertIndex {
end := lookahead
if lookahead > maxMatchLength+checkOff {
end = maxMatchLength + checkOff
}
end += prevIndex
// Hash at match end.
h := hash4(d.window[prevIndex+prevLength:])
ch2 := int(s.hashHead[h]) - s.hashOffset - prevLength
if prevIndex-ch2 != prevOffset && ch2 > minIndex+checkOff {
length := matchLen(d.window[prevIndex+checkOff:end], d.window[ch2+checkOff:])
// It seems like a pure length metric is best.
if length > prevLength {
prevLength = length
prevOffset = prevIndex - ch2
// Extend back...
for i := checkOff - 1; i >= 0; i-- {
if prevLength >= maxMatchLength || d.window[prevIndex+i] != d.window[ch2+i] {
// Emit tokens we "owe"
for j := 0; j <= i; j++ {
d.tokens.AddLiteral(d.window[prevIndex+j])
if d.tokens.n == maxFlateBlockTokens {
// The block includes the current character
if d.err = d.writeBlock(&d.tokens, s.index, false); d.err != nil {
return
}
d.tokens.Reset()
}
s.index++
if s.index < s.maxInsertIndex {
h := hash4(d.window[s.index:])
ch := s.hashHead[h]
s.chainHead = int(ch)
s.hashPrev[s.index&windowMask] = ch
s.hashHead[h] = uint32(s.index + s.hashOffset)
}
}
break
} else {
prevLength++
}
}
} else if false {
// Check one further ahead.
// Only rarely better, disabled for now.
prevIndex++
h := hash4(d.window[prevIndex+prevLength:])
ch2 := int(s.hashHead[h]) - s.hashOffset - prevLength
if prevIndex-ch2 != prevOffset && ch2 > minIndex+checkOff {
length := matchLen(d.window[prevIndex+checkOff:end], d.window[ch2+checkOff:])
// It seems like a pure length metric is best.
if length > prevLength+checkOff {
prevLength = length
prevOffset = prevIndex - ch2
prevIndex--
// Extend back...
for i := checkOff; i >= 0; i-- {
if prevLength >= maxMatchLength || d.window[prevIndex+i] != d.window[ch2+i-1] {
// Emit tokens we "owe"
for j := 0; j <= i; j++ {
d.tokens.AddLiteral(d.window[prevIndex+j])
if d.tokens.n == maxFlateBlockTokens {
// The block includes the current character
if d.err = d.writeBlock(&d.tokens, s.index, false); d.err != nil {
return
}
d.tokens.Reset()
}
s.index++
if s.index < s.maxInsertIndex {
h := hash4(d.window[s.index:])
ch := s.hashHead[h]
s.chainHead = int(ch)
s.hashPrev[s.index&windowMask] = ch
s.hashHead[h] = uint32(s.index + s.hashOffset)
}
}
break
} else {
prevLength++
}
}
}
}
}
}
}
}
// There was a match at the previous step, and the current match is
// not better. Output the previous match.
d.tokens.AddMatch(uint32(prevLength-3), uint32(prevOffset-minOffsetSize))
// Insert in the hash table all strings up to the end of the match.
// index and index-1 are already inserted. If there is not enough
// lookahead, the last two strings are not inserted into the hash
// table.
newIndex := s.index + prevLength - 1
// Calculate missing hashes
end := newIndex
if end > s.maxInsertIndex {
end = s.maxInsertIndex
}
end += minMatchLength - 1
startindex := s.index + 1
if startindex > s.maxInsertIndex {
startindex = s.maxInsertIndex
}
tocheck := d.window[startindex:end]
dstSize := len(tocheck) - minMatchLength + 1
if dstSize > 0 {
dst := s.hashMatch[:dstSize]
bulkHash4(tocheck, dst)
var newH uint32
for i, val := range dst {
di := i + startindex
newH = val & hashMask
// Get previous value with the same hash.
// Our chain should point to the previous value.
s.hashPrev[di&windowMask] = s.hashHead[newH]
// Set the head of the hash chain to us.
s.hashHead[newH] = uint32(di + s.hashOffset)
}
}
s.index = newIndex
d.byteAvailable = false
s.length = minMatchLength - 1
if d.tokens.n == maxFlateBlockTokens {
// The block includes the current character
if d.err = d.writeBlock(&d.tokens, s.index, false); d.err != nil {
return
}
d.tokens.Reset()
}
s.ii = 0
} else {
// Reset, if we got a match this run.
if s.length >= minMatchLength {
s.ii = 0
}
// We have a byte waiting. Emit it.
if d.byteAvailable {
s.ii++
d.tokens.AddLiteral(d.window[s.index-1])
if d.tokens.n == maxFlateBlockTokens {
if d.err = d.writeBlock(&d.tokens, s.index, false); d.err != nil {
return
}
d.tokens.Reset()
}
s.index++
// If we have a long run of no matches, skip additional bytes
// Resets when s.ii overflows after 64KB.
if n := int(s.ii) - d.chain; n > 0 {
n = 1 + int(n>>6)
for j := 0; j < n; j++ {
if s.index >= d.windowEnd-1 {
break
}
d.tokens.AddLiteral(d.window[s.index-1])
if d.tokens.n == maxFlateBlockTokens {
if d.err = d.writeBlock(&d.tokens, s.index, false); d.err != nil {
return
}
d.tokens.Reset()
}
// Index...
if s.index < s.maxInsertIndex {
h := hash4(d.window[s.index:])
ch := s.hashHead[h]
s.chainHead = int(ch)
s.hashPrev[s.index&windowMask] = ch
s.hashHead[h] = uint32(s.index + s.hashOffset)
}
s.index++
}
// Flush last byte
d.tokens.AddLiteral(d.window[s.index-1])
d.byteAvailable = false
// s.length = minMatchLength - 1 // not needed, since s.ii is reset above, so it should never be > minMatchLength
if d.tokens.n == maxFlateBlockTokens {
if d.err = d.writeBlock(&d.tokens, s.index, false); d.err != nil {
return
}
d.tokens.Reset()
}
}
} else {
s.index++
d.byteAvailable = true
}
}
}
}
func (d *compressor) store() {
if d.windowEnd > 0 && (d.windowEnd == maxStoreBlockSize || d.sync) {
d.err = d.writeStoredBlock(d.window[:d.windowEnd])
d.windowEnd = 0
}
}
// fillWindow will fill the buffer with data for huffman-only compression.
// The number of bytes copied is returned.
func (d *compressor) fillBlock(b []byte) int {
n := copy(d.window[d.windowEnd:], b)
d.windowEnd += n
return n
}
// storeHuff will compress and store the currently added data,
// if enough has been accumulated or we at the end of the stream.
// Any error that occurred will be in d.err
func (d *compressor) storeHuff() {
if d.windowEnd < len(d.window) && !d.sync || d.windowEnd == 0 {
return
}
d.w.writeBlockHuff(false, d.window[:d.windowEnd], d.sync)
d.err = d.w.err
d.windowEnd = 0
}
// storeFast will compress and store the currently added data,
// if enough has been accumulated or we at the end of the stream.
// Any error that occurred will be in d.err
func (d *compressor) storeFast() {
// We only compress if we have maxStoreBlockSize.
if d.windowEnd < len(d.window) {
if !d.sync {
return
}
// Handle extremely small sizes.
if d.windowEnd < 128 {
if d.windowEnd == 0 {
return
}
if d.windowEnd <= 32 {
d.err = d.writeStoredBlock(d.window[:d.windowEnd])
} else {
d.w.writeBlockHuff(false, d.window[:d.windowEnd], true)
d.err = d.w.err
}
d.tokens.Reset()
d.windowEnd = 0
d.fast.Reset()
return
}
}
d.fast.Encode(&d.tokens, d.window[:d.windowEnd])
// If we made zero matches, store the block as is.
if d.tokens.n == 0 {
d.err = d.writeStoredBlock(d.window[:d.windowEnd])
// If we removed less than 1/16th, huffman compress the block.
} else if int(d.tokens.n) > d.windowEnd-(d.windowEnd>>4) {
d.w.writeBlockHuff(false, d.window[:d.windowEnd], d.sync)
d.err = d.w.err
} else {
d.w.writeBlockDynamic(&d.tokens, false, d.window[:d.windowEnd], d.sync)
d.err = d.w.err
}
d.tokens.Reset()
d.windowEnd = 0
}
// write will add input byte to the stream.
// Unless an error occurs all bytes will be consumed.
func (d *compressor) write(b []byte) (n int, err error) {
if d.err != nil {
return 0, d.err
}
n = len(b)
for len(b) > 0 {
if d.windowEnd == len(d.window) || d.sync {
d.step(d)
}
b = b[d.fill(d, b):]
if d.err != nil {
return 0, d.err
}
}
return n, d.err
}
func (d *compressor) syncFlush() error {
d.sync = true
if d.err != nil {
return d.err
}
d.step(d)
if d.err == nil {
d.w.writeStoredHeader(0, false)
d.w.flush()
d.err = d.w.err
}
d.sync = false
return d.err
}
func (d *compressor) init(w io.Writer, level int) (err error) {
d.w = newHuffmanBitWriter(w)
switch {
case level == NoCompression:
d.window = make([]byte, maxStoreBlockSize)
d.fill = (*compressor).fillBlock
d.step = (*compressor).store
case level == ConstantCompression:
d.w.logNewTablePenalty = 10
d.window = make([]byte, 32<<10)
d.fill = (*compressor).fillBlock
d.step = (*compressor).storeHuff
case level == DefaultCompression:
level = 5
fallthrough
case level >= 1 && level <= 6:
d.w.logNewTablePenalty = 7
d.fast = newFastEnc(level)
d.window = make([]byte, maxStoreBlockSize)
d.fill = (*compressor).fillBlock
d.step = (*compressor).storeFast
case 7 <= level && level <= 9:
d.w.logNewTablePenalty = 8
d.state = &advancedState{}
d.compressionLevel = levels[level]
d.initDeflate()
d.fill = (*compressor).fillDeflate
d.step = (*compressor).deflateLazy
default:
return fmt.Errorf("flate: invalid compression level %d: want value in range [-2, 9]", level)
}
d.level = level
return nil
}
// reset the state of the compressor.
func (d *compressor) reset(w io.Writer) {
d.w.reset(w)
d.sync = false
d.err = nil
// We only need to reset a few things for Snappy.
if d.fast != nil {
d.fast.Reset()
d.windowEnd = 0
d.tokens.Reset()
return
}
switch d.compressionLevel.chain {
case 0:
// level was NoCompression or ConstantCompresssion.
d.windowEnd = 0
default:
s := d.state
s.chainHead = -1
for i := range s.hashHead {
s.hashHead[i] = 0
}
for i := range s.hashPrev {
s.hashPrev[i] = 0
}
s.hashOffset = 1
s.index, d.windowEnd = 0, 0
d.blockStart, d.byteAvailable = 0, false
d.tokens.Reset()
s.length = minMatchLength - 1
s.offset = 0
s.ii = 0
s.maxInsertIndex = 0
}
}
func (d *compressor) close() error {
if d.err != nil {
return d.err
}
d.sync = true
d.step(d)
if d.err != nil {
return d.err
}
if d.w.writeStoredHeader(0, true); d.w.err != nil {
return d.w.err
}
d.w.flush()
d.w.reset(nil)
return d.w.err
}
// NewWriter returns a new Writer compressing data at the given level.
// Following zlib, levels range from 1 (BestSpeed) to 9 (BestCompression);
// higher levels typically run slower but compress more.
// Level 0 (NoCompression) does not attempt any compression; it only adds the
// necessary DEFLATE framing.
// Level -1 (DefaultCompression) uses the default compression level.
// Level -2 (ConstantCompression) will use Huffman compression only, giving
// a very fast compression for all types of input, but sacrificing considerable
// compression efficiency.
//
// If level is in the range [-2, 9] then the error returned will be nil.
// Otherwise the error returned will be non-nil.
func NewWriter(w io.Writer, level int) (*Writer, error) {
var dw Writer
if err := dw.d.init(w, level); err != nil {
return nil, err
}
return &dw, nil
}
// NewWriterDict is like NewWriter but initializes the new
// Writer with a preset dictionary. The returned Writer behaves
// as if the dictionary had been written to it without producing
// any compressed output. The compressed data written to w
// can only be decompressed by a Reader initialized with the
// same dictionary.
func NewWriterDict(w io.Writer, level int, dict []byte) (*Writer, error) {
zw, err := NewWriter(w, level)
if err != nil {
return nil, err
}
zw.d.fillWindow(dict)
zw.dict = append(zw.dict, dict...) // duplicate dictionary for Reset method.
return zw, err
}
// A Writer takes data written to it and writes the compressed
// form of that data to an underlying writer (see NewWriter).
type Writer struct {
d compressor
dict []byte
}
// Write writes data to w, which will eventually write the
// compressed form of data to its underlying writer.
func (w *Writer) Write(data []byte) (n int, err error) {
return w.d.write(data)
}
// Flush flushes any pending data to the underlying writer.
// It is useful mainly in compressed network protocols, to ensure that
// a remote reader has enough data to reconstruct a packet.
// Flush does not return until the data has been written.
// Calling Flush when there is no pending data still causes the Writer
// to emit a sync marker of at least 4 bytes.
// If the underlying writer returns an error, Flush returns that error.
//
// In the terminology of the zlib library, Flush is equivalent to Z_SYNC_FLUSH.
func (w *Writer) Flush() error {
// For more about flushing:
// http://www.bolet.org/~pornin/deflate-flush.html
return w.d.syncFlush()
}
// Close flushes and closes the writer.
func (w *Writer) Close() error {
return w.d.close()
}
// Reset discards the writer's state and makes it equivalent to
// the result of NewWriter or NewWriterDict called with dst
// and w's level and dictionary.
func (w *Writer) Reset(dst io.Writer) {
if len(w.dict) > 0 {
// w was created with NewWriterDict
w.d.reset(dst)
if dst != nil {
w.d.fillWindow(w.dict)
}
} else {
// w was created with NewWriter
w.d.reset(dst)
}
}
// ResetDict discards the writer's state and makes it equivalent to
// the result of NewWriter or NewWriterDict called with dst
// and w's level, but sets a specific dictionary.
func (w *Writer) ResetDict(dst io.Writer, dict []byte) {
w.dict = dict
w.d.reset(dst)
w.d.fillWindow(w.dict)
}
vendor/github.com/klauspost/compress/flate/dict_decoder.go
Generated Vendored
+184
View File
@@ -0,0 +1,184 @@
// Copyright 2016 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package flate
// dictDecoder implements the LZ77 sliding dictionary as used in decompression.
// LZ77 decompresses data through sequences of two forms of commands:
//
// - Literal insertions: Runs of one or more symbols are inserted into the data
// stream as is. This is accomplished through the writeByte method for a
// single symbol, or combinations of writeSlice/writeMark for multiple symbols.
// Any valid stream must start with a literal insertion if no preset dictionary
// is used.
//
// - Backward copies: Runs of one or more symbols are copied from previously
// emitted data. Backward copies come as the tuple (dist, length) where dist
// determines how far back in the stream to copy from and length determines how
// many bytes to copy. Note that it is valid for the length to be greater than
// the distance. Since LZ77 uses forward copies, that situation is used to
// perform a form of run-length encoding on repeated runs of symbols.
// The writeCopy and tryWriteCopy are used to implement this command.
//
// For performance reasons, this implementation performs little to no sanity
// checks about the arguments. As such, the invariants documented for each
// method call must be respected.
type dictDecoder struct {
hist []byte // Sliding window history
// Invariant: 0 <= rdPos <= wrPos <= len(hist)
wrPos int // Current output position in buffer
rdPos int // Have emitted hist[:rdPos] already
full bool // Has a full window length been written yet?
}
// init initializes dictDecoder to have a sliding window dictionary of the given
// size. If a preset dict is provided, it will initialize the dictionary with
// the contents of dict.
func (dd *dictDecoder) init(size int, dict []byte) {
*dd = dictDecoder{hist: dd.hist}
if cap(dd.hist) < size {
dd.hist = make([]byte, size)
}
dd.hist = dd.hist[:size]
if len(dict) > len(dd.hist) {
dict = dict[len(dict)-len(dd.hist):]
}
dd.wrPos = copy(dd.hist, dict)
if dd.wrPos == len(dd.hist) {
dd.wrPos = 0
dd.full = true
}
dd.rdPos = dd.wrPos
}
// histSize reports the total amount of historical data in the dictionary.
func (dd *dictDecoder) histSize() int {
if dd.full {
return len(dd.hist)
}
return dd.wrPos
}
// availRead reports the number of bytes that can be flushed by readFlush.
func (dd *dictDecoder) availRead() int {
return dd.wrPos - dd.rdPos
}
// availWrite reports the available amount of output buffer space.
func (dd *dictDecoder) availWrite() int {
return len(dd.hist) - dd.wrPos
}
// writeSlice returns a slice of the available buffer to write data to.
//
// This invariant will be kept: len(s) <= availWrite()
func (dd *dictDecoder) writeSlice() []byte {
return dd.hist[dd.wrPos:]
}
// writeMark advances the writer pointer by cnt.
//
// This invariant must be kept: 0 <= cnt <= availWrite()
func (dd *dictDecoder) writeMark(cnt int) {
dd.wrPos += cnt
}
// writeByte writes a single byte to the dictionary.
//
// This invariant must be kept: 0 < availWrite()
func (dd *dictDecoder) writeByte(c byte) {
dd.hist[dd.wrPos] = c
dd.wrPos++
}
// writeCopy copies a string at a given (dist, length) to the output.
// This returns the number of bytes copied and may be less than the requested
// length if the available space in the output buffer is too small.
//
// This invariant must be kept: 0 < dist <= histSize()
func (dd *dictDecoder) writeCopy(dist, length int) int {
dstBase := dd.wrPos
dstPos := dstBase
srcPos := dstPos - dist
endPos := dstPos + length
if endPos > len(dd.hist) {
endPos = len(dd.hist)
}
// Copy non-overlapping section after destination position.
//
// This section is non-overlapping in that the copy length for this section
// is always less than or equal to the backwards distance. This can occur
// if a distance refers to data that wraps-around in the buffer.
// Thus, a backwards copy is performed here; that is, the exact bytes in
// the source prior to the copy is placed in the destination.
if srcPos < 0 {
srcPos += len(dd.hist)
dstPos += copy(dd.hist[dstPos:endPos], dd.hist[srcPos:])
srcPos = 0
}
// Copy possibly overlapping section before destination position.
//
// This section can overlap if the copy length for this section is larger
// than the backwards distance. This is allowed by LZ77 so that repeated
// strings can be succinctly represented using (dist, length) pairs.
// Thus, a forwards copy is performed here; that is, the bytes copied is
// possibly dependent on the resulting bytes in the destination as the copy
// progresses along. This is functionally equivalent to the following:
//
// for i := 0; i < endPos-dstPos; i++ {
// dd.hist[dstPos+i] = dd.hist[srcPos+i]
// }
// dstPos = endPos
//
for dstPos < endPos {
dstPos += copy(dd.hist[dstPos:endPos], dd.hist[srcPos:dstPos])
}
dd.wrPos = dstPos
return dstPos - dstBase
}
// tryWriteCopy tries to copy a string at a given (distance, length) to the
// output. This specialized version is optimized for short distances.
//
// This method is designed to be inlined for performance reasons.
//
// This invariant must be kept: 0 < dist <= histSize()
func (dd *dictDecoder) tryWriteCopy(dist, length int) int {
dstPos := dd.wrPos
endPos := dstPos + length
if dstPos < dist || endPos > len(dd.hist) {
return 0
}
dstBase := dstPos
srcPos := dstPos - dist
// Copy possibly overlapping section before destination position.
loop:
dstPos += copy(dd.hist[dstPos:endPos], dd.hist[srcPos:dstPos])
if dstPos < endPos {
goto loop // Avoid for-loop so that this function can be inlined
}
dd.wrPos = dstPos
return dstPos - dstBase
}
// readFlush returns a slice of the historical buffer that is ready to be
// emitted to the user. The data returned by readFlush must be fully consumed
// before calling any other dictDecoder methods.
func (dd *dictDecoder) readFlush() []byte {
toRead := dd.hist[dd.rdPos:dd.wrPos]
dd.rdPos = dd.wrPos
if dd.wrPos == len(dd.hist) {
dd.wrPos, dd.rdPos = 0, 0
dd.full = true
}
return toRead
}
vendor/github.com/klauspost/compress/flate/fast_encoder.go
Generated Vendored
+216
View File
@@ -0,0 +1,216 @@
// Copyright 2011 The Snappy-Go Authors. All rights reserved.
// Modified for deflate by Klaus Post (c) 2015.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package flate
import (
"encoding/binary"
"fmt"
"math/bits"
)
type fastEnc interface {
Encode(dst *tokens, src []byte)
Reset()
}
func newFastEnc(level int) fastEnc {
switch level {
case 1:
return &fastEncL1{fastGen: fastGen{cur: maxStoreBlockSize}}
case 2:
return &fastEncL2{fastGen: fastGen{cur: maxStoreBlockSize}}
case 3:
return &fastEncL3{fastGen: fastGen{cur: maxStoreBlockSize}}
case 4:
return &fastEncL4{fastGen: fastGen{cur: maxStoreBlockSize}}
case 5:
return &fastEncL5{fastGen: fastGen{cur: maxStoreBlockSize}}
case 6:
return &fastEncL6{fastGen: fastGen{cur: maxStoreBlockSize}}
default:
panic("invalid level specified")
}
}
const (
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 = 17 // Bits used in the big tables
bTableSize = 1 << bTableBits // Size of the table
allocHistory = maxStoreBlockSize * 5 // Size to preallocate for history.
bufferReset = (1 << 31) - allocHistory - maxStoreBlockSize - 1 // Reset the buffer offset when reaching this.
)
const (
prime3bytes = 506832829
prime4bytes = 2654435761
prime5bytes = 889523592379
prime6bytes = 227718039650203
prime7bytes = 58295818150454627
prime8bytes = 0xcf1bbcdcb7a56463
)
func load3232(b []byte, i int32) uint32 {
return binary.LittleEndian.Uint32(b[i:])
}
func load6432(b []byte, i int32) uint64 {
return binary.LittleEndian.Uint64(b[i:])
}
type tableEntry struct {
offset int32
}
// fastGen maintains the table for matches,
// and the previous byte block for level 2.
// This is the generic implementation.
type fastGen struct {
hist []byte
cur int32
}
func (e *fastGen) addBlock(src []byte) int32 {
// check if we have space already
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) < maxMatchOffset*2 {
panic("unexpected buffer size")
}
// Move down
offset := int32(len(e.hist)) - maxMatchOffset
// copy(e.hist[0:maxMatchOffset], e.hist[offset:])
*(*[maxMatchOffset]byte)(e.hist) = *(*[maxMatchOffset]byte)(e.hist[offset:])
e.cur += offset
e.hist = e.hist[:maxMatchOffset]
}
}
s := int32(len(e.hist))
e.hist = append(e.hist, src...)
return s
}
type tableEntryPrev struct {
Cur tableEntry
Prev tableEntry
}
// hash7 returns the hash of the lowest 7 bytes of u to fit in a hash table with h bits.
// Preferably h should be a constant and should always be <64.
func hash7(u uint64, h uint8) uint32 {
return uint32(((u << (64 - 56)) * prime7bytes) >> ((64 - h) & reg8SizeMask64))
}
// hashLen returns a hash of the lowest mls bytes of with length output bits.
// mls must be >=3 and <=8. Any other value will return hash for 4 bytes.
// length should always be < 32.
// Preferably length and mls should be a constant for inlining.
func hashLen(u uint64, length, mls uint8) uint32 {
switch mls {
case 3:
return (uint32(u<<8) * prime3bytes) >> (32 - length)
case 5:
return uint32(((u << (64 - 40)) * prime5bytes) >> (64 - length))
case 6:
return uint32(((u << (64 - 48)) * prime6bytes) >> (64 - length))
case 7:
return uint32(((u << (64 - 56)) * prime7bytes) >> (64 - length))
case 8:
return uint32((u * prime8bytes) >> (64 - length))
default:
return (uint32(u) * prime4bytes) >> (32 - length)
}
}
// 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 *fastGen) 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 > maxMatchOffset {
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 *fastGen) 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 > maxMatchOffset {
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:]))
}
// Reset the encoding table.
func (e *fastGen) Reset() {
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 <= bufferReset {
e.cur += maxMatchOffset + int32(len(e.hist))
}
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
}
vendor/github.com/klauspost/compress/flate/huffman_bit_writer.go
Generated Vendored
+1187
View File
File diff suppressed because it is too large Load Diff
vendor/github.com/klauspost/compress/flate/huffman_code.go
Generated Vendored
+417
View File
@@ -0,0 +1,417 @@
// Copyright 2009 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package flate
import (
"math"
"math/bits"
)
const (
maxBitsLimit = 16
// number of valid literals
literalCount = 286
)
// hcode is a huffman code with a bit code and bit length.
type hcode uint32
func (h hcode) len() uint8 {
return uint8(h)
}
func (h hcode) code64() uint64 {
return uint64(h >> 8)
}
func (h hcode) zero() bool {
return h == 0
}
type huffmanEncoder struct {
codes []hcode
bitCount [17]int32
// Allocate a reusable buffer with the longest possible frequency table.
// Possible lengths are codegenCodeCount, offsetCodeCount and literalCount.
// The largest of these is literalCount, so we allocate for that case.
freqcache [literalCount + 1]literalNode
}
type literalNode struct {
literal uint16
freq uint16
}
// A levelInfo describes the state of the constructed tree for a given depth.
type levelInfo struct {
// Our level. for better printing
level int32
// The frequency of the last node at this level
lastFreq int32
// The frequency of the next character to add to this level
nextCharFreq int32
// The frequency of the next pair (from level below) to add to this level.
// Only valid if the "needed" value of the next lower level is 0.
nextPairFreq int32
// The number of chains remaining to generate for this level before moving
// up to the next level
needed int32
}
// set sets the code and length of an hcode.
func (h *hcode) set(code uint16, length uint8) {
*h = hcode(length) | (hcode(code) << 8)
}
func newhcode(code uint16, length uint8) hcode {
return hcode(length) | (hcode(code) << 8)
}
func reverseBits(number uint16, bitLength byte) uint16 {
return bits.Reverse16(number << ((16 - bitLength) & 15))
}
func maxNode() literalNode { return literalNode{math.MaxUint16, math.MaxUint16} }
func newHuffmanEncoder(size int) *huffmanEncoder {
// Make capacity to next power of two.
c := uint(bits.Len32(uint32(size - 1)))
return &huffmanEncoder{codes: make([]hcode, size, 1<<c)}
}
// Generates a HuffmanCode corresponding to the fixed literal table
func generateFixedLiteralEncoding() *huffmanEncoder {
h := newHuffmanEncoder(literalCount)
codes := h.codes
var ch uint16
for ch = 0; ch < literalCount; ch++ {
var bits uint16
var size uint8
switch {
case ch < 144:
// size 8, 000110000 .. 10111111
bits = ch + 48
size = 8
case ch < 256:
// size 9, 110010000 .. 111111111
bits = ch + 400 - 144
size = 9
case ch < 280:
// size 7, 0000000 .. 0010111
bits = ch - 256
size = 7
default:
// size 8, 11000000 .. 11000111
bits = ch + 192 - 280
size = 8
}
codes[ch] = newhcode(reverseBits(bits, size), size)
}
return h
}
func generateFixedOffsetEncoding() *huffmanEncoder {
h := newHuffmanEncoder(30)
codes := h.codes
for ch := range codes {
codes[ch] = newhcode(reverseBits(uint16(ch), 5), 5)
}
return h
}
var fixedLiteralEncoding = generateFixedLiteralEncoding()
var fixedOffsetEncoding = generateFixedOffsetEncoding()
func (h *huffmanEncoder) bitLength(freq []uint16) int {
var total int
for i, f := range freq {
if f != 0 {
total += int(f) * int(h.codes[i].len())
}
}
return total
}
func (h *huffmanEncoder) bitLengthRaw(b []byte) int {
var total int
for _, f := range b {
total += int(h.codes[f].len())
}
return total
}
// canReuseBits returns the number of bits or math.MaxInt32 if the encoder cannot be reused.
func (h *huffmanEncoder) canReuseBits(freq []uint16) int {
var total int
for i, f := range freq {
if f != 0 {
code := h.codes[i]
if code.zero() {
return math.MaxInt32
}
total += int(f) * int(code.len())
}
}
return total
}
// Return the number of literals assigned to each bit size in the Huffman encoding
//
// This method is only called when list.length >= 3
// The cases of 0, 1, and 2 literals are handled by special case code.
//
// list An array of the literals with non-zero frequencies
//
// and their associated frequencies. The array is in order of increasing
// frequency, and has as its last element a special element with frequency
// MaxInt32
//
// maxBits The maximum number of bits that should be used to encode any literal.
//
// Must be less than 16.
//
// return An integer array in which array[i] indicates the number of literals
//
// that should be encoded in i bits.
func (h *huffmanEncoder) bitCounts(list []literalNode, maxBits int32) []int32 {
if maxBits >= maxBitsLimit {
panic("flate: maxBits too large")
}
n := int32(len(list))
list = list[0 : n+1]
list[n] = maxNode()
// The tree can't have greater depth than n - 1, no matter what. This
// saves a little bit of work in some small cases
if maxBits > n-1 {
maxBits = n - 1
}
// Create information about each of the levels.
// A bogus "Level 0" whose sole purpose is so that
// level1.prev.needed==0. This makes level1.nextPairFreq
// be a legitimate value that never gets chosen.
var levels [maxBitsLimit]levelInfo
// leafCounts[i] counts the number of literals at the left
// of ancestors of the rightmost node at level i.
// leafCounts[i][j] is the number of literals at the left
// of the level j ancestor.
var leafCounts [maxBitsLimit][maxBitsLimit]int32
// Descending to only have 1 bounds check.
l2f := int32(list[2].freq)
l1f := int32(list[1].freq)
l0f := int32(list[0].freq) + int32(list[1].freq)
for level := int32(1); level <= maxBits; level++ {
// For every level, the first two items are the first two characters.
// We initialize the levels as if we had already figured this out.
levels[level] = levelInfo{
level: level,
lastFreq: l1f,
nextCharFreq: l2f,
nextPairFreq: l0f,
}
leafCounts[level][level] = 2
if level == 1 {
levels[level].nextPairFreq = math.MaxInt32
}
}
// We need a total of 2*n - 2 items at top level and have already generated 2.
levels[maxBits].needed = 2*n - 4
level := uint32(maxBits)
for level < 16 {
l := &levels[level]
if l.nextPairFreq == math.MaxInt32 && l.nextCharFreq == math.MaxInt32 {
// We've run out of both leafs and pairs.
// End all calculations for this level.
// To make sure we never come back to this level or any lower level,
// set nextPairFreq impossibly large.
l.needed = 0
levels[level+1].nextPairFreq = math.MaxInt32
level++
continue
}
prevFreq := l.lastFreq
if l.nextCharFreq < l.nextPairFreq {
// The next item on this row is a leaf node.
n := leafCounts[level][level] + 1
l.lastFreq = l.nextCharFreq
// Lower leafCounts are the same of the previous node.
leafCounts[level][level] = n
e := list[n]
if e.literal < math.MaxUint16 {
l.nextCharFreq = int32(e.freq)
} else {
l.nextCharFreq = math.MaxInt32
}
} else {
// The next item on this row is a pair from the previous row.
// nextPairFreq isn't valid until we generate two
// more values in the level below
l.lastFreq = l.nextPairFreq
// Take leaf counts from the lower level, except counts[level] remains the same.
if true {
save := leafCounts[level][level]
leafCounts[level] = leafCounts[level-1]
leafCounts[level][level] = save
} else {
copy(leafCounts[level][:level], leafCounts[level-1][:level])
}
levels[l.level-1].needed = 2
}
if l.needed--; l.needed == 0 {
// We've done everything we need to do for this level.
// Continue calculating one level up. Fill in nextPairFreq
// of that level with the sum of the two nodes we've just calculated on
// this level.
if l.level == maxBits {
// All done!
break
}
levels[l.level+1].nextPairFreq = prevFreq + l.lastFreq
level++
} else {
// If we stole from below, move down temporarily to replenish it.
for levels[level-1].needed > 0 {
level--
}
}
}
// Somethings is wrong if at the end, the top level is null or hasn't used
// all of the leaves.
if leafCounts[maxBits][maxBits] != n {
panic("leafCounts[maxBits][maxBits] != n")
}
bitCount := h.bitCount[:maxBits+1]
bits := 1
counts := &leafCounts[maxBits]
for level := maxBits; level > 0; level-- {
// chain.leafCount gives the number of literals requiring at least "bits"
// bits to encode.
bitCount[bits] = counts[level] - counts[level-1]
bits++
}
return bitCount
}
// Look at the leaves and assign them a bit count and an encoding as specified
// in RFC 1951 3.2.2
func (h *huffmanEncoder) assignEncodingAndSize(bitCount []int32, list []literalNode) {
code := uint16(0)
for n, bits := range bitCount {
code <<= 1
if n == 0 || bits == 0 {
continue
}
// The literals list[len(list)-bits] .. list[len(list)-bits]
// are encoded using "bits" bits, and get the values
// code, code + 1, .... The code values are
// assigned in literal order (not frequency order).
chunk := list[len(list)-int(bits):]
sortByLiteral(chunk)
for _, node := range chunk {
h.codes[node.literal] = newhcode(reverseBits(code, uint8(n)), uint8(n))
code++
}
list = list[0 : len(list)-int(bits)]
}
}
// Update this Huffman Code object to be the minimum code for the specified frequency count.
//
// freq An array of frequencies, in which frequency[i] gives the frequency of literal i.
// maxBits The maximum number of bits to use for any literal.
func (h *huffmanEncoder) generate(freq []uint16, maxBits int32) {
list := h.freqcache[:len(freq)+1]
codes := h.codes[:len(freq)]
// Number of non-zero literals
count := 0
// Set list to be the set of all non-zero literals and their frequencies
for i, f := range freq {
if f != 0 {
list[count] = literalNode{uint16(i), f}
count++
} else {
codes[i] = 0
}
}
list[count] = literalNode{}
list = list[:count]
if count <= 2 {
// Handle the small cases here, because they are awkward for the general case code. With
// two or fewer literals, everything has bit length 1.
for i, node := range list {
// "list" is in order of increasing literal value.
h.codes[node.literal].set(uint16(i), 1)
}
return
}
sortByFreq(list)
// Get the number of literals for each bit count
bitCount := h.bitCounts(list, maxBits)
// And do the assignment
h.assignEncodingAndSize(bitCount, list)
}
// atLeastOne clamps the result between 1 and 15.
func atLeastOne(v float32) float32 {
if v < 1 {
return 1
}
if v > 15 {
return 15
}
return v
}
func histogram(b []byte, h []uint16) {
if true && len(b) >= 8<<10 {
// Split for bigger inputs
histogramSplit(b, h)
} else {
h = h[:256]
for _, t := range b {
h[t]++
}
}
}
func histogramSplit(b []byte, h []uint16) {
// Tested, and slightly faster than 2-way.
// Writing to separate arrays and combining is also slightly slower.
h = h[:256]
for len(b)&3 != 0 {
h[b[0]]++
b = b[1:]
}
n := len(b) / 4
x, y, z, w := b[:n], b[n:], b[n+n:], b[n+n+n:]
y, z, w = y[:len(x)], z[:len(x)], w[:len(x)]
for i, t := range x {
v0 := &h[t]
v1 := &h[y[i]]
v3 := &h[w[i]]
v2 := &h[z[i]]
*v0++
*v1++
*v2++
*v3++
}
}
vendor/github.com/klauspost/compress/flate/huffman_sortByFreq.go
Generated Vendored
+178
View File
@@ -0,0 +1,178 @@
// Copyright 2009 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package flate
// Sort sorts data.
// It makes one call to data.Len to determine n, and O(n*log(n)) calls to
// data.Less and data.Swap. The sort is not guaranteed to be stable.
func sortByFreq(data []literalNode) {
n := len(data)
quickSortByFreq(data, 0, n, maxDepth(n))
}
func quickSortByFreq(data []literalNode, a, b, maxDepth int) {
for b-a > 12 { // Use ShellSort for slices <= 12 elements
if maxDepth == 0 {
heapSort(data, a, b)
return
}
maxDepth--
mlo, mhi := doPivotByFreq(data, a, b)
// Avoiding recursion on the larger subproblem guarantees
// a stack depth of at most lg(b-a).
if mlo-a < b-mhi {
quickSortByFreq(data, a, mlo, maxDepth)
a = mhi // i.e., quickSortByFreq(data, mhi, b)
} else {
quickSortByFreq(data, mhi, b, maxDepth)
b = mlo // i.e., quickSortByFreq(data, a, mlo)
}
}
if b-a > 1 {
// Do ShellSort pass with gap 6
// It could be written in this simplified form cause b-a <= 12
for i := a + 6; i < b; i++ {
if data[i].freq == data[i-6].freq && data[i].literal < data[i-6].literal || data[i].freq < data[i-6].freq {
data[i], data[i-6] = data[i-6], data[i]
}
}
insertionSortByFreq(data, a, b)
}
}
// siftDownByFreq implements the heap property on data[lo, hi).
// first is an offset into the array where the root of the heap lies.
func siftDownByFreq(data []literalNode, lo, hi, first int) {
root := lo
for {
child := 2*root + 1
if child >= hi {
break
}
if child+1 < hi && (data[first+child].freq == data[first+child+1].freq && data[first+child].literal < data[first+child+1].literal || data[first+child].freq < data[first+child+1].freq) {
child++
}
if data[first+root].freq == data[first+child].freq && data[first+root].literal > data[first+child].literal || data[first+root].freq > data[first+child].freq {
return
}
data[first+root], data[first+child] = data[first+child], data[first+root]
root = child
}
}
func doPivotByFreq(data []literalNode, lo, hi int) (midlo, midhi int) {
m := int(uint(lo+hi) >> 1) // Written like this to avoid integer overflow.
if hi-lo > 40 {
// Tukey's ``Ninther,'' median of three medians of three.
s := (hi - lo) / 8
medianOfThreeSortByFreq(data, lo, lo+s, lo+2*s)
medianOfThreeSortByFreq(data, m, m-s, m+s)
medianOfThreeSortByFreq(data, hi-1, hi-1-s, hi-1-2*s)
}
medianOfThreeSortByFreq(data, lo, m, hi-1)
// Invariants are:
// data[lo] = pivot (set up by ChoosePivot)
// data[lo < i < a] < pivot
// data[a <= i < b] <= pivot
// data[b <= i < c] unexamined
// data[c <= i < hi-1] > pivot
// data[hi-1] >= pivot
pivot := lo
a, c := lo+1, hi-1
for ; a < c && (data[a].freq == data[pivot].freq && data[a].literal < data[pivot].literal || data[a].freq < data[pivot].freq); a++ {
}
b := a
for {
for ; b < c && (data[pivot].freq == data[b].freq && data[pivot].literal > data[b].literal || data[pivot].freq > data[b].freq); b++ { // data[b] <= pivot
}
for ; b < c && (data[pivot].freq == data[c-1].freq && data[pivot].literal < data[c-1].literal || data[pivot].freq < data[c-1].freq); c-- { // data[c-1] > pivot
}
if b >= c {
break
}
// data[b] > pivot; data[c-1] <= pivot
data[b], data[c-1] = data[c-1], data[b]
b++
c--
}
// If hi-c<3 then there are duplicates (by property of median of nine).
// Let's be a bit more conservative, and set border to 5.
protect := hi-c < 5
if !protect && hi-c < (hi-lo)/4 {
// Lets test some points for equality to pivot
dups := 0
if data[pivot].freq == data[hi-1].freq && data[pivot].literal > data[hi-1].literal || data[pivot].freq > data[hi-1].freq { // data[hi-1] = pivot
data[c], data[hi-1] = data[hi-1], data[c]
c++
dups++
}
if data[b-1].freq == data[pivot].freq && data[b-1].literal > data[pivot].literal || data[b-1].freq > data[pivot].freq { // data[b-1] = pivot
b--
dups++
}
// m-lo = (hi-lo)/2 > 6
// b-lo > (hi-lo)*3/4-1 > 8
// ==> m < b ==> data[m] <= pivot
if data[m].freq == data[pivot].freq && data[m].literal > data[pivot].literal || data[m].freq > data[pivot].freq { // data[m] = pivot
data[m], data[b-1] = data[b-1], data[m]
b--
dups++
}
// if at least 2 points are equal to pivot, assume skewed distribution
protect = dups > 1
}
if protect {
// Protect against a lot of duplicates
// Add invariant:
// data[a <= i < b] unexamined
// data[b <= i < c] = pivot
for {
for ; a < b && (data[b-1].freq == data[pivot].freq && data[b-1].literal > data[pivot].literal || data[b-1].freq > data[pivot].freq); b-- { // data[b] == pivot
}
for ; a < b && (data[a].freq == data[pivot].freq && data[a].literal < data[pivot].literal || data[a].freq < data[pivot].freq); a++ { // data[a] < pivot
}
if a >= b {
break
}
// data[a] == pivot; data[b-1] < pivot
data[a], data[b-1] = data[b-1], data[a]
a++
b--
}
}
// Swap pivot into middle
data[pivot], data[b-1] = data[b-1], data[pivot]
return b - 1, c
}
// Insertion sort
func insertionSortByFreq(data []literalNode, a, b int) {
for i := a + 1; i < b; i++ {
for j := i; j > a && (data[j].freq == data[j-1].freq && data[j].literal < data[j-1].literal || data[j].freq < data[j-1].freq); j-- {
data[j], data[j-1] = data[j-1], data[j]
}
}
}
// quickSortByFreq, loosely following Bentley and McIlroy,
// ``Engineering a Sort Function,'' SP&E November 1993.
// medianOfThreeSortByFreq moves the median of the three values data[m0], data[m1], data[m2] into data[m1].
func medianOfThreeSortByFreq(data []literalNode, m1, m0, m2 int) {
// sort 3 elements
if data[m1].freq == data[m0].freq && data[m1].literal < data[m0].literal || data[m1].freq < data[m0].freq {
data[m1], data[m0] = data[m0], data[m1]
}
// data[m0] <= data[m1]
if data[m2].freq == data[m1].freq && data[m2].literal < data[m1].literal || data[m2].freq < data[m1].freq {
data[m2], data[m1] = data[m1], data[m2]
// data[m0] <= data[m2] && data[m1] < data[m2]
if data[m1].freq == data[m0].freq && data[m1].literal < data[m0].literal || data[m1].freq < data[m0].freq {
data[m1], data[m0] = data[m0], data[m1]
}
}
// now data[m0] <= data[m1] <= data[m2]
}
vendor/github.com/klauspost/compress/flate/huffman_sortByLiteral.go
Generated Vendored
+201
View File
@@ -0,0 +1,201 @@
// Copyright 2009 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package flate
// Sort sorts data.
// It makes one call to data.Len to determine n, and O(n*log(n)) calls to
// data.Less and data.Swap. The sort is not guaranteed to be stable.
func sortByLiteral(data []literalNode) {
n := len(data)
quickSort(data, 0, n, maxDepth(n))
}
func quickSort(data []literalNode, a, b, maxDepth int) {
for b-a > 12 { // Use ShellSort for slices <= 12 elements
if maxDepth == 0 {
heapSort(data, a, b)
return
}
maxDepth--
mlo, mhi := doPivot(data, a, b)
// Avoiding recursion on the larger subproblem guarantees
// a stack depth of at most lg(b-a).
if mlo-a < b-mhi {
quickSort(data, a, mlo, maxDepth)
a = mhi // i.e., quickSort(data, mhi, b)
} else {
quickSort(data, mhi, b, maxDepth)
b = mlo // i.e., quickSort(data, a, mlo)
}
}
if b-a > 1 {
// Do ShellSort pass with gap 6
// It could be written in this simplified form cause b-a <= 12
for i := a + 6; i < b; i++ {
if data[i].literal < data[i-6].literal {
data[i], data[i-6] = data[i-6], data[i]
}
}
insertionSort(data, a, b)
}
}
func heapSort(data []literalNode, a, b int) {
first := a
lo := 0
hi := b - a
// Build heap with greatest element at top.
for i := (hi - 1) / 2; i >= 0; i-- {
siftDown(data, i, hi, first)
}
// Pop elements, largest first, into end of data.
for i := hi - 1; i >= 0; i-- {
data[first], data[first+i] = data[first+i], data[first]
siftDown(data, lo, i, first)
}
}
// siftDown implements the heap property on data[lo, hi).
// first is an offset into the array where the root of the heap lies.
func siftDown(data []literalNode, lo, hi, first int) {
root := lo
for {
child := 2*root + 1
if child >= hi {
break
}
if child+1 < hi && data[first+child].literal < data[first+child+1].literal {
child++
}
if data[first+root].literal > data[first+child].literal {
return
}
data[first+root], data[first+child] = data[first+child], data[first+root]
root = child
}
}
func doPivot(data []literalNode, lo, hi int) (midlo, midhi int) {
m := int(uint(lo+hi) >> 1) // Written like this to avoid integer overflow.
if hi-lo > 40 {
// Tukey's ``Ninther,'' median of three medians of three.
s := (hi - lo) / 8
medianOfThree(data, lo, lo+s, lo+2*s)
medianOfThree(data, m, m-s, m+s)
medianOfThree(data, hi-1, hi-1-s, hi-1-2*s)
}
medianOfThree(data, lo, m, hi-1)
// Invariants are:
// data[lo] = pivot (set up by ChoosePivot)
// data[lo < i < a] < pivot
// data[a <= i < b] <= pivot
// data[b <= i < c] unexamined
// data[c <= i < hi-1] > pivot
// data[hi-1] >= pivot
pivot := lo
a, c := lo+1, hi-1
for ; a < c && data[a].literal < data[pivot].literal; a++ {
}
b := a
for {
for ; b < c && data[pivot].literal > data[b].literal; b++ { // data[b] <= pivot
}
for ; b < c && data[pivot].literal < data[c-1].literal; c-- { // data[c-1] > pivot
}
if b >= c {
break
}
// data[b] > pivot; data[c-1] <= pivot
data[b], data[c-1] = data[c-1], data[b]
b++
c--
}
// If hi-c<3 then there are duplicates (by property of median of nine).
// Let's be a bit more conservative, and set border to 5.
protect := hi-c < 5
if !protect && hi-c < (hi-lo)/4 {
// Lets test some points for equality to pivot
dups := 0
if data[pivot].literal > data[hi-1].literal { // data[hi-1] = pivot
data[c], data[hi-1] = data[hi-1], data[c]
c++
dups++
}
if data[b-1].literal > data[pivot].literal { // data[b-1] = pivot
b--
dups++
}
// m-lo = (hi-lo)/2 > 6
// b-lo > (hi-lo)*3/4-1 > 8
// ==> m < b ==> data[m] <= pivot
if data[m].literal > data[pivot].literal { // data[m] = pivot
data[m], data[b-1] = data[b-1], data[m]
b--
dups++
}
// if at least 2 points are equal to pivot, assume skewed distribution
protect = dups > 1
}
if protect {
// Protect against a lot of duplicates
// Add invariant:
// data[a <= i < b] unexamined
// data[b <= i < c] = pivot
for {
for ; a < b && data[b-1].literal > data[pivot].literal; b-- { // data[b] == pivot
}
for ; a < b && data[a].literal < data[pivot].literal; a++ { // data[a] < pivot
}
if a >= b {
break
}
// data[a] == pivot; data[b-1] < pivot
data[a], data[b-1] = data[b-1], data[a]
a++
b--
}
}
// Swap pivot into middle
data[pivot], data[b-1] = data[b-1], data[pivot]
return b - 1, c
}
// Insertion sort
func insertionSort(data []literalNode, a, b int) {
for i := a + 1; i < b; i++ {
for j := i; j > a && data[j].literal < data[j-1].literal; j-- {
data[j], data[j-1] = data[j-1], data[j]
}
}
}
// maxDepth returns a threshold at which quicksort should switch
// to heapsort. It returns 2*ceil(lg(n+1)).
func maxDepth(n int) int {
var depth int
for i := n; i > 0; i >>= 1 {
depth++
}
return depth * 2
}
// medianOfThree moves the median of the three values data[m0], data[m1], data[m2] into data[m1].
func medianOfThree(data []literalNode, m1, m0, m2 int) {
// sort 3 elements
if data[m1].literal < data[m0].literal {
data[m1], data[m0] = data[m0], data[m1]
}
// data[m0] <= data[m1]
if data[m2].literal < data[m1].literal {
data[m2], data[m1] = data[m1], data[m2]
// data[m0] <= data[m2] && data[m1] < data[m2]
if data[m1].literal < data[m0].literal {
data[m1], data[m0] = data[m0], data[m1]
}
}
// now data[m0] <= data[m1] <= data[m2]
}
vendor/github.com/klauspost/compress/flate/inflate.go
Generated Vendored
+793
View File
@@ -0,0 +1,793 @@
// Copyright 2009 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
// Package flate implements the DEFLATE compressed data format, described in
// RFC 1951. The gzip and zlib packages implement access to DEFLATE-based file
// formats.
package flate
import (
"bufio"
"compress/flate"
"fmt"
"io"
"math/bits"
"sync"
)
const (
maxCodeLen = 16 // max length of Huffman code
maxCodeLenMask = 15 // mask for max length of Huffman code
// The next three numbers come from the RFC section 3.2.7, with the
// additional proviso in section 3.2.5 which implies that distance codes
// 30 and 31 should never occur in compressed data.
maxNumLit = 286
maxNumDist = 30
numCodes = 19 // number of codes in Huffman meta-code
debugDecode = false
)
// Value of length - 3 and extra bits.
type lengthExtra struct {
length, extra uint8
}
var decCodeToLen = [32]lengthExtra{{length: 0x0, extra: 0x0}, {length: 0x1, extra: 0x0}, {length: 0x2, extra: 0x0}, {length: 0x3, extra: 0x0}, {length: 0x4, extra: 0x0}, {length: 0x5, extra: 0x0}, {length: 0x6, extra: 0x0}, {length: 0x7, extra: 0x0}, {length: 0x8, extra: 0x1}, {length: 0xa, extra: 0x1}, {length: 0xc, extra: 0x1}, {length: 0xe, extra: 0x1}, {length: 0x10, extra: 0x2}, {length: 0x14, extra: 0x2}, {length: 0x18, extra: 0x2}, {length: 0x1c, extra: 0x2}, {length: 0x20, extra: 0x3}, {length: 0x28, extra: 0x3}, {length: 0x30, extra: 0x3}, {length: 0x38, extra: 0x3}, {length: 0x40, extra: 0x4}, {length: 0x50, extra: 0x4}, {length: 0x60, extra: 0x4}, {length: 0x70, extra: 0x4}, {length: 0x80, extra: 0x5}, {length: 0xa0, extra: 0x5}, {length: 0xc0, extra: 0x5}, {length: 0xe0, extra: 0x5}, {length: 0xff, extra: 0x0}, {length: 0x0, extra: 0x0}, {length: 0x0, extra: 0x0}, {length: 0x0, extra: 0x0}}
var bitMask32 = [32]uint32{
0, 1, 3, 7, 0xF, 0x1F, 0x3F, 0x7F, 0xFF,
0x1FF, 0x3FF, 0x7FF, 0xFFF, 0x1FFF, 0x3FFF, 0x7FFF, 0xFFFF,
0x1ffff, 0x3ffff, 0x7FFFF, 0xfFFFF, 0x1fFFFF, 0x3fFFFF, 0x7fFFFF, 0xffFFFF,
0x1ffFFFF, 0x3ffFFFF, 0x7ffFFFF, 0xfffFFFF, 0x1fffFFFF, 0x3fffFFFF, 0x7fffFFFF,
} // up to 32 bits
// Initialize the fixedHuffmanDecoder only once upon first use.
var fixedOnce sync.Once
var fixedHuffmanDecoder huffmanDecoder
// A CorruptInputError reports the presence of corrupt input at a given offset.
type CorruptInputError = flate.CorruptInputError
// An InternalError reports an error in the flate code itself.
type InternalError string
func (e InternalError) Error() string { return "flate: internal error: " + string(e) }
// A ReadError reports an error encountered while reading input.
//
// Deprecated: No longer returned.
type ReadError = flate.ReadError
// A WriteError reports an error encountered while writing output.
//
// Deprecated: No longer returned.
type WriteError = flate.WriteError
// Resetter resets a ReadCloser returned by NewReader or NewReaderDict to
// to switch to a new underlying Reader. This permits reusing a ReadCloser
// instead of allocating a new one.
type Resetter interface {
// Reset discards any buffered data and resets the Resetter as if it was
// newly initialized with the given reader.
Reset(r io.Reader, dict []byte) error
}
// The data structure for decoding Huffman tables is based on that of
// zlib. There is a lookup table of a fixed bit width (huffmanChunkBits),
// For codes smaller than the table width, there are multiple entries
// (each combination of trailing bits has the same value). For codes
// larger than the table width, the table contains a link to an overflow
// table. The width of each entry in the link table is the maximum code
// size minus the chunk width.
//
// Note that you can do a lookup in the table even without all bits
// filled. Since the extra bits are zero, and the DEFLATE Huffman codes
// have the property that shorter codes come before longer ones, the
// bit length estimate in the result is a lower bound on the actual
// number of bits.
//
// See the following:
// http://www.gzip.org/algorithm.txt
// chunk & 15 is number of bits
// chunk >> 4 is value, including table link
const (
huffmanChunkBits = 9
huffmanNumChunks = 1 << huffmanChunkBits
huffmanCountMask = 15
huffmanValueShift = 4
)
type huffmanDecoder struct {
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
}
// Initialize Huffman decoding tables from array of code lengths.
// Following this function, h is guaranteed to be initialized into a complete
// tree (i.e., neither over-subscribed nor under-subscribed). The exception is a
// degenerate case where the tree has only a single symbol with length 1. Empty
// trees are permitted.
func (h *huffmanDecoder) init(lengths []int) bool {
// Sanity enables additional runtime tests during Huffman
// table construction. It's intended to be used during
// development to supplement the currently ad-hoc unit tests.
const sanity = false
if h.chunks == nil {
h.chunks = &[huffmanNumChunks]uint16{}
}
if h.maxRead != 0 {
*h = huffmanDecoder{chunks: h.chunks, links: h.links}
}
// Count number of codes of each length,
// compute maxRead and max length.
var count [maxCodeLen]int
var min, max int
for _, n := range lengths {
if n == 0 {
continue
}
if min == 0 || n < min {
min = n
}
if n > max {
max = n
}
count[n&maxCodeLenMask]++
}
// Empty tree. The decompressor.huffSym function will fail later if the tree
// is used. Technically, an empty tree is only valid for the HDIST tree and
// not the HCLEN and HLIT tree. However, a stream with an empty HCLEN tree
// is guaranteed to fail since it will attempt to use the tree to decode the
// codes for the HLIT and HDIST trees. Similarly, an empty HLIT tree is
// guaranteed to fail later since the compressed data section must be
// composed of at least one symbol (the end-of-block marker).
if max == 0 {
return true
}
code := 0
var nextcode [maxCodeLen]int
for i := min; i <= max; i++ {
code <<= 1
nextcode[i&maxCodeLenMask] = code
code += count[i&maxCodeLenMask]
}
// Check that the coding is complete (i.e., that we've
// assigned all 2-to-the-max possible bit sequences).
// Exception: To be compatible with zlib, we also need to
// accept degenerate single-code codings. See also
// TestDegenerateHuffmanCoding.
if code != 1<<uint(max) && !(code == 1 && max == 1) {
if debugDecode {
fmt.Println("coding failed, code, max:", code, max, code == 1<<uint(max), code == 1 && max == 1, "(one should be true)")
}
return false
}
h.maxRead = min
chunks := h.chunks[:]
for i := range chunks {
chunks[i] = 0
}
if max > huffmanChunkBits {
numLinks := 1 << (uint(max) - huffmanChunkBits)
h.linkMask = uint32(numLinks - 1)
// create link tables
link := nextcode[huffmanChunkBits+1] >> 1
if cap(h.links) < huffmanNumChunks-link {
h.links = make([][]uint16, huffmanNumChunks-link)
} else {
h.links = h.links[:huffmanNumChunks-link]
}
for j := uint(link); j < huffmanNumChunks; j++ {
reverse := int(bits.Reverse16(uint16(j)))
reverse >>= uint(16 - huffmanChunkBits)
off := j - uint(link)
if sanity && h.chunks[reverse] != 0 {
panic("impossible: overwriting existing chunk")
}
h.chunks[reverse] = uint16(off<<huffmanValueShift | (huffmanChunkBits + 1))
if cap(h.links[off]) < numLinks {
h.links[off] = make([]uint16, numLinks)
} else {
links := h.links[off][:0]
h.links[off] = links[:numLinks]
}
}
} else {
h.links = h.links[:0]
}
for i, n := range lengths {
if n == 0 {
continue
}
code := nextcode[n]
nextcode[n]++
chunk := uint16(i<<huffmanValueShift | n)
reverse := int(bits.Reverse16(uint16(code)))
reverse >>= uint(16 - n)
if n <= huffmanChunkBits {
for off := reverse; off < len(h.chunks); off += 1 << uint(n) {
// We should never need to overwrite
// an existing chunk. Also, 0 is
// never a valid chunk, because the
// lower 4 "count" bits should be
// between 1 and 15.
if sanity && h.chunks[off] != 0 {
panic("impossible: overwriting existing chunk")
}
h.chunks[off] = chunk
}
} else {
j := reverse & (huffmanNumChunks - 1)
if sanity && h.chunks[j]&huffmanCountMask != huffmanChunkBits+1 {
// Longer codes should have been
// associated with a link table above.
panic("impossible: not an indirect chunk")
}
value := h.chunks[j] >> huffmanValueShift
linktab := h.links[value]
reverse >>= huffmanChunkBits
for off := reverse; off < len(linktab); off += 1 << uint(n-huffmanChunkBits) {
if sanity && linktab[off] != 0 {
panic("impossible: overwriting existing chunk")
}
linktab[off] = chunk
}
}
}
if sanity {
// Above we've sanity checked that we never overwrote
// an existing entry. Here we additionally check that
// we filled the tables completely.
for i, chunk := range h.chunks {
if chunk == 0 {
// As an exception, in the degenerate
// single-code case, we allow odd
// chunks to be missing.
if code == 1 && i%2 == 1 {
continue
}
panic("impossible: missing chunk")
}
}
for _, linktab := range h.links {
for _, chunk := range linktab {
if chunk == 0 {
panic("impossible: missing chunk")
}
}
}
}
return true
}
// 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 {
io.Reader
io.ByteReader
}
// Decompress state.
type decompressor struct {
// Input source.
r Reader
roffset int64
// Huffman decoders for literal/length, distance.
h1, h2 huffmanDecoder
// Length arrays used to define Huffman codes.
bits *[maxNumLit + maxNumDist]int
codebits *[numCodes]int
// Output history, buffer.
dict dictDecoder
// Next step in the decompression,
// and decompression state.
step func(*decompressor)
stepState int
err error
toRead []byte
hl, hd *huffmanDecoder
copyLen int
copyDist int
// Temporary buffer (avoids repeated allocation).
buf [4]byte
// Input bits, in top of b.
b uint32
nb uint
final bool
}
func (f *decompressor) nextBlock() {
for f.nb < 1+2 {
if f.err = f.moreBits(); f.err != nil {
return
}
}
f.final = f.b&1 == 1
f.b >>= 1
typ := f.b & 3
f.b >>= 2
f.nb -= 1 + 2
switch typ {
case 0:
f.dataBlock()
if debugDecode {
fmt.Println("stored block")
}
case 1:
// compressed, fixed Huffman tables
f.hl = &fixedHuffmanDecoder
f.hd = nil
f.huffmanBlockDecoder()()
if debugDecode {
fmt.Println("predefinied huffman block")
}
case 2:
// compressed, dynamic Huffman tables
if f.err = f.readHuffman(); f.err != nil {
break
}
f.hl = &f.h1
f.hd = &f.h2
f.huffmanBlockDecoder()()
if debugDecode {
fmt.Println("dynamic huffman block")
}
default:
// 3 is reserved.
if debugDecode {
fmt.Println("reserved data block encountered")
}
f.err = CorruptInputError(f.roffset)
}
}
func (f *decompressor) Read(b []byte) (int, error) {
for {
if len(f.toRead) > 0 {
n := copy(b, f.toRead)
f.toRead = f.toRead[n:]
if len(f.toRead) == 0 {
return n, f.err
}
return n, nil
}
if f.err != nil {
return 0, f.err
}
f.step(f)
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.
func (f *decompressor) WriteTo(w io.Writer) (int64, error) {
total := int64(0)
flushed := false
for {
if len(f.toRead) > 0 {
n, err := w.Write(f.toRead)
total += int64(n)
if err != nil {
f.err = err
return total, err
}
if n != len(f.toRead) {
return total, io.ErrShortWrite
}
f.toRead = f.toRead[:0]
}
if f.err != nil && flushed {
if f.err == io.EOF {
return total, nil
}
return total, f.err
}
if f.err == nil {
f.step(f)
}
if len(f.toRead) == 0 && f.err != nil && !flushed {
f.toRead = f.dict.readFlush() // Flush what's left in case of error
flushed = true
}
}
}
func (f *decompressor) Close() error {
if f.err == io.EOF {
return nil
}
return f.err
}
// RFC 1951 section 3.2.7.
// Compression with dynamic Huffman codes
var codeOrder = [...]int{16, 17, 18, 0, 8, 7, 9, 6, 10, 5, 11, 4, 12, 3, 13, 2, 14, 1, 15}
func (f *decompressor) readHuffman() error {
// HLIT[5], HDIST[5], HCLEN[4].
for f.nb < 5+5+4 {
if err := f.moreBits(); err != nil {
return err
}
}
nlit := int(f.b&0x1F) + 257
if nlit > maxNumLit {
if debugDecode {
fmt.Println("nlit > maxNumLit", nlit)
}
return CorruptInputError(f.roffset)
}
f.b >>= 5
ndist := int(f.b&0x1F) + 1
if ndist > maxNumDist {
if debugDecode {
fmt.Println("ndist > maxNumDist", ndist)
}
return CorruptInputError(f.roffset)
}
f.b >>= 5
nclen := int(f.b&0xF) + 4
// numCodes is 19, so nclen is always valid.
f.b >>= 4
f.nb -= 5 + 5 + 4
// (HCLEN+4)*3 bits: code lengths in the magic codeOrder order.
for i := 0; i < nclen; i++ {
for f.nb < 3 {
if err := f.moreBits(); err != nil {
return err
}
}
f.codebits[codeOrder[i]] = int(f.b & 0x7)
f.b >>= 3
f.nb -= 3
}
for i := nclen; i < len(codeOrder); i++ {
f.codebits[codeOrder[i]] = 0
}
if !f.h1.init(f.codebits[0:]) {
if debugDecode {
fmt.Println("init codebits failed")
}
return CorruptInputError(f.roffset)
}
// HLIT + 257 code lengths, HDIST + 1 code lengths,
// using the code length Huffman code.
for i, n := 0, nlit+ndist; i < n; {
x, err := f.huffSym(&f.h1)
if err != nil {
return err
}
if x < 16 {
// Actual length.
f.bits[i] = x
i++
continue
}
// Repeat previous length or zero.
var rep int
var nb uint
var b int
switch x {
default:
return InternalError("unexpected length code")
case 16:
rep = 3
nb = 2
if i == 0 {
if debugDecode {
fmt.Println("i==0")
}
return CorruptInputError(f.roffset)
}
b = f.bits[i-1]
case 17:
rep = 3
nb = 3
b = 0
case 18:
rep = 11
nb = 7
b = 0
}
for f.nb < nb {
if err := f.moreBits(); err != nil {
if debugDecode {
fmt.Println("morebits:", err)
}
return err
}
}
rep += int(f.b & uint32(1<<(nb&regSizeMaskUint32)-1))
f.b >>= nb & regSizeMaskUint32
f.nb -= nb
if i+rep > n {
if debugDecode {
fmt.Println("i+rep > n", i, rep, n)
}
return CorruptInputError(f.roffset)
}
for j := 0; j < rep; j++ {
f.bits[i] = b
i++
}
}
if !f.h1.init(f.bits[0:nlit]) || !f.h2.init(f.bits[nlit:nlit+ndist]) {
if debugDecode {
fmt.Println("init2 failed")
}
return CorruptInputError(f.roffset)
}
// 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.maxRead < 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
}
// Copy a single uncompressed data block from input to output.
func (f *decompressor) dataBlock() {
// Uncompressed.
// Discard current half-byte.
left := (f.nb) & 7
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[offBytes:4])
f.roffset += int64(nr)
if err != nil {
f.err = noEOF(err)
return
}
n := uint16(f.buf[0]) | uint16(f.buf[1])<<8
nn := uint16(f.buf[2]) | uint16(f.buf[3])<<8
if nn != ^n {
if debugDecode {
ncomp := ^n
fmt.Println("uint16(nn) != uint16(^n)", nn, ncomp)
}
f.err = CorruptInputError(f.roffset)
return
}
if n == 0 {
f.toRead = f.dict.readFlush()
f.finishBlock()
return
}
f.copyLen = int(n)
f.copyData()
}
// copyData copies f.copyLen bytes from the underlying reader into f.hist.
// It pauses for reads when f.hist is full.
func (f *decompressor) copyData() {
buf := f.dict.writeSlice()
if len(buf) > f.copyLen {
buf = buf[:f.copyLen]
}
cnt, err := io.ReadFull(f.r, buf)
f.roffset += int64(cnt)
f.copyLen -= cnt
f.dict.writeMark(cnt)
if err != nil {
f.err = noEOF(err)
return
}
if f.dict.availWrite() == 0 || f.copyLen > 0 {
f.toRead = f.dict.readFlush()
f.step = (*decompressor).copyData
return
}
f.finishBlock()
}
func (f *decompressor) finishBlock() {
if f.final {
if f.dict.availRead() > 0 {
f.toRead = f.dict.readFlush()
}
f.err = io.EOF
}
f.step = (*decompressor).nextBlock
}
// noEOF returns err, unless err == io.EOF, in which case it returns io.ErrUnexpectedEOF.
func noEOF(e error) error {
if e == io.EOF {
return io.ErrUnexpectedEOF
}
return e
}
func (f *decompressor) moreBits() error {
c, err := f.r.ReadByte()
if err != nil {
return noEOF(err)
}
f.roffset++
f.b |= uint32(c) << (f.nb & regSizeMaskUint32)
f.nb += 8
return nil
}
// Read the next Huffman-encoded symbol from f according to h.
func (f *decompressor) huffSym(h *huffmanDecoder) (int, error) {
// 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(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.
nb, b := f.nb, f.b
for {
for nb < n {
c, err := f.r.ReadByte()
if err != nil {
f.b = b
f.nb = nb
return 0, noEOF(err)
}
f.roffset++
b |= uint32(c) << (nb & regSizeMaskUint32)
nb += 8
}
chunk := h.chunks[b&(huffmanNumChunks-1)]
n = uint(chunk & huffmanCountMask)
if n > huffmanChunkBits {
chunk = h.links[chunk>>huffmanValueShift][(b>>huffmanChunkBits)&h.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 0, f.err
}
f.b = b >> (n & regSizeMaskUint32)
f.nb = nb - n
return int(chunk >> huffmanValueShift), nil
}
}
}
func makeReader(r io.Reader) Reader {
if rr, ok := r.(Reader); ok {
return rr
}
return bufio.NewReader(r)
}
func fixedHuffmanDecoderInit() {
fixedOnce.Do(func() {
// These come from the RFC section 3.2.6.
var bits [288]int
for i := 0; i < 144; i++ {
bits[i] = 8
}
for i := 144; i < 256; i++ {
bits[i] = 9
}
for i := 256; i < 280; i++ {
bits[i] = 7
}
for i := 280; i < 288; i++ {
bits[i] = 8
}
fixedHuffmanDecoder.init(bits[:])
})
}
func (f *decompressor) Reset(r io.Reader, dict []byte) error {
*f = decompressor{
r: makeReader(r),
bits: f.bits,
codebits: f.codebits,
h1: f.h1,
h2: f.h2,
dict: f.dict,
step: (*decompressor).nextBlock,
}
f.dict.init(maxMatchOffset, dict)
return nil
}
// NewReader returns a new ReadCloser that can be used
// to read the uncompressed version of r.
// If r does not also implement io.ByteReader,
// the decompressor may read more data than necessary from r.
// It is the caller's responsibility to call Close on the ReadCloser
// when finished reading.
//
// The ReadCloser returned by NewReader also implements Resetter.
func NewReader(r io.Reader) io.ReadCloser {
fixedHuffmanDecoderInit()
var f decompressor
f.r = makeReader(r)
f.bits = new([maxNumLit + maxNumDist]int)
f.codebits = new([numCodes]int)
f.step = (*decompressor).nextBlock
f.dict.init(maxMatchOffset, nil)
return &f
}
// NewReaderDict is like NewReader but initializes the reader
// with a preset dictionary. The returned Reader behaves as if
// the uncompressed data stream started with the given dictionary,
// which has already been read. NewReaderDict is typically used
// to read data compressed by NewWriterDict.
//
// The ReadCloser returned by NewReader also implements Resetter.
func NewReaderDict(r io.Reader, dict []byte) io.ReadCloser {
fixedHuffmanDecoderInit()
var f decompressor
f.r = makeReader(r)
f.bits = new([maxNumLit + maxNumDist]int)
f.codebits = new([numCodes]int)
f.step = (*decompressor).nextBlock
f.dict.init(maxMatchOffset, dict)
return &f
}
vendor/github.com/klauspost/compress/flate/inflate_gen.go
Generated Vendored
+1283
View File
File diff suppressed because it is too large Load Diff
vendor/github.com/klauspost/compress/flate/level1.go
Generated Vendored
+241
View File
@@ -0,0 +1,241 @@
package flate
import (
"encoding/binary"
"fmt"
"math/bits"
)
// fastGen maintains the table for matches,
// and the previous byte block for level 2.
// This is the generic implementation.
type fastEncL1 struct {
fastGen
table [tableSize]tableEntry
}
// EncodeL1 uses a similar algorithm to level 1
func (e *fastEncL1) Encode(dst *tokens, src []byte) {
const (
inputMargin = 12 - 1
minNonLiteralBlockSize = 1 + 1 + inputMargin
hashBytes = 5
)
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{}
}
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
}
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 = 5
const doEvery = 2
nextS := s
var candidate tableEntry
for {
nextHash := hashLen(cv, tableBits, hashBytes)
candidate = e.table[nextHash]
nextS = s + doEvery + (s-nextEmit)>>skipLog
if nextS > sLimit {
goto emitRemainder
}
now := load6432(src, nextS)
e.table[nextHash] = tableEntry{offset: s + e.cur}
nextHash = hashLen(now, tableBits, hashBytes)
offset := s - (candidate.offset - e.cur)
if offset < maxMatchOffset && uint32(cv) == load3232(src, candidate.offset-e.cur) {
e.table[nextHash] = tableEntry{offset: nextS + e.cur}
break
}
// Do one right away...
cv = now
s = nextS
nextS++
candidate = e.table[nextHash]
now >>= 8
e.table[nextHash] = tableEntry{offset: s + e.cur}
offset = s - (candidate.offset - e.cur)
if offset < maxMatchOffset && uint32(cv) == load3232(src, candidate.offset-e.cur) {
e.table[nextHash] = tableEntry{offset: nextS + e.cur}
break
}
cv = now
s = nextS
}
// 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.
for {
// Invariant: we have a 4-byte match at s, and no need to emit any
// literal bytes prior to s.
// Extend the 4-byte match as long as possible.
t := candidate.offset - e.cur
var l = int32(4)
if false {
l = e.matchlenLong(s+4, t+4, src) + 4
} else {
// inlined:
a := src[s+4:]
b := src[t+4:]
for len(a) >= 8 {
if diff := binary.LittleEndian.Uint64(a) ^ binary.LittleEndian.Uint64(b); diff != 0 {
l += int32(bits.TrailingZeros64(diff) >> 3)
break
}
l += 8
a = a[8:]
b = b[8:]
}
if len(a) < 8 {
b = b[:len(a)]
for i := range a {
if a[i] != b[i] {
break
}
l++
}
}
}
// 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++
}
}
}
// Save the match found
if false {
dst.AddMatchLong(l, uint32(s-t-baseMatchOffset))
} else {
// Inlined...
xoffset := uint32(s - t - baseMatchOffset)
xlength := l
oc := offsetCode(xoffset)
xoffset |= oc << 16
for xlength > 0 {
xl := xlength
if xl > 258 {
if xl > 258+baseMatchLength {
xl = 258
} else {
xl = 258 - baseMatchLength
}
}
xlength -= xl
xl -= baseMatchLength
dst.extraHist[lengthCodes1[uint8(xl)]]++
dst.offHist[oc]++
dst.tokens[dst.n] = token(matchType | uint32(xl)<<lengthShift | xoffset)
dst.n++
}
}
s += l
nextEmit = s
if nextS >= s {
s = nextS + 1
}
if s >= sLimit {
// Index first pair after match end.
if int(s+l+8) < len(src) {
cv := load6432(src, s)
e.table[hashLen(cv, tableBits, hashBytes)] = tableEntry{offset: s + e.cur}
}
goto emitRemainder
}
// We could immediately start working at s now, but to improve
// compression we first update the hash table at s-2 and at s. If
// another emitCopy is not our next move, also calculate nextHash
// at s+1. At least on GOARCH=amd64, these three hash calculations
// are faster as one load64 call (with some shifts) instead of
// three load32 calls.
x := load6432(src, s-2)
o := e.cur + s - 2
prevHash := hashLen(x, tableBits, hashBytes)
e.table[prevHash] = tableEntry{offset: o}
x >>= 16
currHash := hashLen(x, tableBits, hashBytes)
candidate = e.table[currHash]
e.table[currHash] = tableEntry{offset: o + 2}
offset := s - (candidate.offset - e.cur)
if offset > maxMatchOffset || uint32(x) != load3232(src, candidate.offset-e.cur) {
cv = x >> 8
s++
break
}
}
}
emitRemainder:
if int(nextEmit) < len(src) {
// If nothing was added, don't encode literals.
if dst.n == 0 {
return
}
emitLiteral(dst, src[nextEmit:])
}
}
vendor/github.com/klauspost/compress/flate/level2.go
Generated Vendored
+214
View File
@@ -0,0 +1,214 @@
package flate
import "fmt"
// fastGen maintains the table for matches,
// and the previous byte block for level 2.
// This is the generic implementation.
type fastEncL2 struct {
fastGen
table [bTableSize]tableEntry
}
// EncodeL2 uses a similar algorithm to level 1, but is capable
// of matching across blocks giving better compression at a small slowdown.
func (e *fastEncL2) Encode(dst *tokens, src []byte) {
const (
inputMargin = 12 - 1
minNonLiteralBlockSize = 1 + 1 + inputMargin
hashBytes = 5
)
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{}
}
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
}
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 {
// When should we start skipping if we haven't found matches in a long while.
const skipLog = 5
const doEvery = 2
nextS := s
var candidate tableEntry
for {
nextHash := hashLen(cv, bTableBits, hashBytes)
s = nextS
nextS = s + doEvery + (s-nextEmit)>>skipLog
if nextS > sLimit {
goto emitRemainder
}
candidate = e.table[nextHash]
now := load6432(src, nextS)
e.table[nextHash] = tableEntry{offset: s + e.cur}
nextHash = hashLen(now, bTableBits, hashBytes)
offset := s - (candidate.offset - e.cur)
if offset < maxMatchOffset && uint32(cv) == load3232(src, candidate.offset-e.cur) {
e.table[nextHash] = tableEntry{offset: nextS + e.cur}
break
}
// Do one right away...
cv = now
s = nextS
nextS++
candidate = e.table[nextHash]
now >>= 8
e.table[nextHash] = tableEntry{offset: s + e.cur}
offset = s - (candidate.offset - e.cur)
if offset < maxMatchOffset && uint32(cv) == load3232(src, candidate.offset-e.cur) {
break
}
cv = now
}
// 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.
// Call emitCopy, and then see if another emitCopy could be our next
// move. Repeat until we find no match for the input immediately after
// what was consumed by the last emitCopy call.
//
// If we exit this loop normally then we need to call emitLiteral next,
// though we don't yet know how big the literal will be. We handle that
// by proceeding to the next iteration of the main loop. We also can
// exit this loop via goto if we get close to exhausting the input.
for {
// Invariant: we have a 4-byte match at s, and no need to emit any
// literal bytes prior to s.
// Extend the 4-byte match as long as possible.
t := candidate.offset - e.cur
l := e.matchlenLong(s+4, t+4, src) + 4
// 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++
}
}
}
dst.AddMatchLong(l, uint32(s-t-baseMatchOffset))
s += l
nextEmit = s
if nextS >= s {
s = nextS + 1
}
if s >= sLimit {
// Index first pair after match end.
if int(s+l+8) < len(src) {
cv := load6432(src, s)
e.table[hashLen(cv, bTableBits, hashBytes)] = tableEntry{offset: s + e.cur}
}
goto emitRemainder
}
// Store every second hash in-between, but offset by 1.
for i := s - l + 2; i < s-5; i += 7 {
x := load6432(src, i)
nextHash := hashLen(x, bTableBits, hashBytes)
e.table[nextHash] = tableEntry{offset: e.cur + i}
// Skip one
x >>= 16
nextHash = hashLen(x, bTableBits, hashBytes)
e.table[nextHash] = tableEntry{offset: e.cur + i + 2}
// Skip one
x >>= 16
nextHash = hashLen(x, bTableBits, hashBytes)
e.table[nextHash] = tableEntry{offset: e.cur + i + 4}
}
// We could immediately start working at s now, but to improve
// compression we first update the hash table at s-2 to s. If
// another emitCopy is not our next move, also calculate nextHash
// at s+1. At least on GOARCH=amd64, these three hash calculations
// are faster as one load64 call (with some shifts) instead of
// three load32 calls.
x := load6432(src, s-2)
o := e.cur + s - 2
prevHash := hashLen(x, bTableBits, hashBytes)
prevHash2 := hashLen(x>>8, bTableBits, hashBytes)
e.table[prevHash] = tableEntry{offset: o}
e.table[prevHash2] = tableEntry{offset: o + 1}
currHash := hashLen(x>>16, bTableBits, hashBytes)
candidate = e.table[currHash]
e.table[currHash] = tableEntry{offset: o + 2}
offset := s - (candidate.offset - e.cur)
if offset > maxMatchOffset || uint32(x>>16) != load3232(src, candidate.offset-e.cur) {
cv = x >> 24
s++
break
}
}
}
emitRemainder:
if int(nextEmit) < len(src) {
// If nothing was added, don't encode literals.
if dst.n == 0 {
return
}
emitLiteral(dst, src[nextEmit:])
}
}
vendor/github.com/klauspost/compress/flate/level3.go
Generated Vendored
+241
View File
@@ -0,0 +1,241 @@
package flate
import "fmt"
// fastEncL3
type fastEncL3 struct {
fastGen
table [1 << 16]tableEntryPrev
}
// Encode uses a similar algorithm to level 2, will check up to two candidates.
func (e *fastEncL3) Encode(dst *tokens, src []byte) {
const (
inputMargin = 12 - 1
minNonLiteralBlockSize = 1 + 1 + inputMargin
tableBits = 16
tableSize = 1 << tableBits
hashBytes = 5
)
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] = 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]
if v.Cur.offset <= minOff {
v.Cur.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.table[i] = v
}
e.cur = maxMatchOffset
}
s := e.addBlock(src)
// Skip if too small.
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 = 7
nextS := s
var candidate tableEntry
for {
nextHash := hashLen(cv, tableBits, hashBytes)
s = nextS
nextS = s + 1 + (s-nextEmit)>>skipLog
if nextS > sLimit {
goto emitRemainder
}
candidates := e.table[nextHash]
now := load6432(src, nextS)
// 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
if candidate.offset < minOffset {
cv = now
// Previous will also be invalid, we have nothing.
continue
}
if uint32(cv) == load3232(src, candidate.offset-e.cur) {
if candidates.Prev.offset < minOffset || uint32(cv) != load3232(src, candidates.Prev.offset-e.cur) {
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
}
break
} else {
// We only check if value mismatches.
// Offset will always be invalid in other cases.
candidate = candidates.Prev
if candidate.offset > minOffset && uint32(cv) == load3232(src, candidate.offset-e.cur) {
break
}
}
cv = now
}
// Call emitCopy, and then see if another emitCopy could be our next
// move. Repeat until we find no match for the input immediately after
// what was consumed by the last emitCopy call.
//
// If we exit this loop normally then we need to call emitLiteral next,
// though we don't yet know how big the literal will be. We handle that
// by proceeding to the next iteration of the main loop. We also can
// exit this loop via goto if we get close to exhausting the input.
for {
// Invariant: we have a 4-byte match at s, and no need to emit any
// literal bytes prior to s.
// Extend the 4-byte match as long as possible.
//
t := candidate.offset - e.cur
l := e.matchlenLong(s+4, t+4, src) + 4
// 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++
}
}
}
dst.AddMatchLong(l, uint32(s-t-baseMatchOffset))
s += l
nextEmit = s
if nextS >= s {
s = nextS + 1
}
if s >= sLimit {
t += l
// Index first pair after match end.
if int(t+8) < len(src) && t > 0 {
cv = load6432(src, t)
nextHash := hashLen(cv, tableBits, hashBytes)
e.table[nextHash] = tableEntryPrev{
Prev: e.table[nextHash].Cur,
Cur: tableEntry{offset: e.cur + t},
}
}
goto emitRemainder
}
// Store every 5th hash in-between.
for i := s - l + 2; i < s-5; i += 6 {
nextHash := hashLen(load6432(src, i), tableBits, hashBytes)
e.table[nextHash] = tableEntryPrev{
Prev: e.table[nextHash].Cur,
Cur: tableEntry{offset: e.cur + i}}
}
// We could immediately start working at s now, but to improve
// compression we first update the hash table at s-2 to s.
x := load6432(src, s-2)
prevHash := hashLen(x, tableBits, hashBytes)
e.table[prevHash] = tableEntryPrev{
Prev: e.table[prevHash].Cur,
Cur: tableEntry{offset: e.cur + s - 2},
}
x >>= 8
prevHash = hashLen(x, tableBits, hashBytes)
e.table[prevHash] = tableEntryPrev{
Prev: e.table[prevHash].Cur,
Cur: tableEntry{offset: e.cur + s - 1},
}
x >>= 8
currHash := hashLen(x, tableBits, hashBytes)
candidates := e.table[currHash]
cv = x
e.table[currHash] = tableEntryPrev{
Prev: candidates.Cur,
Cur: tableEntry{offset: s + e.cur},
}
// Check both candidates
candidate = candidates.Cur
minOffset := e.cur + s - (maxMatchOffset - 4)
if candidate.offset > minOffset {
if uint32(cv) == load3232(src, candidate.offset-e.cur) {
// Found a match...
continue
}
candidate = candidates.Prev
if candidate.offset > minOffset && uint32(cv) == load3232(src, candidate.offset-e.cur) {
// Match at prev...
continue
}
}
cv = x >> 8
s++
break
}
}
emitRemainder:
if int(nextEmit) < len(src) {
// If nothing was added, don't encode literals.
if dst.n == 0 {
return
}
emitLiteral(dst, src[nextEmit:])
}
}
vendor/github.com/klauspost/compress/flate/level4.go
Generated Vendored
+221
View File
@@ -0,0 +1,221 @@
package flate
import "fmt"
type fastEncL4 struct {
fastGen
table [tableSize]tableEntry
bTable [tableSize]tableEntry
}
func (e *fastEncL4) Encode(dst *tokens, src []byte) {
const (
inputMargin = 12 - 1
minNonLiteralBlockSize = 1 + 1 + inputMargin
hashShortBytes = 4
)
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] = tableEntry{}
}
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].offset
if v <= minOff {
v = 0
} else {
v = v - e.cur + maxMatchOffset
}
e.bTable[i].offset = 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 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
e.bTable[nextHashL] = entry
t = lCandidate.offset - e.cur
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) == 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 && 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
t = lCandidate.offset - e.cur
}
}
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.
// Extend the 4-byte match as long as possible.
l := e.matchlenLong(s+4, t+4, src) + 4
// 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("s-t")
}
if (s - t) > maxMatchOffset {
panic(fmt.Sprintln("mmo", 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 {
// Index first pair after match end.
if int(s+8) < len(src) {
cv := load6432(src, s)
e.table[hashLen(cv, tableBits, hashShortBytes)] = tableEntry{offset: s + e.cur}
e.bTable[hash7(cv, tableBits)] = tableEntry{offset: s + e.cur}
}
goto emitRemainder
}
// Store every 3rd hash in-between
if true {
i := nextS
if i < s-1 {
cv := load6432(src, i)
t := tableEntry{offset: i + e.cur}
t2 := tableEntry{offset: t.offset + 1}
e.bTable[hash7(cv, tableBits)] = t
e.bTable[hash7(cv>>8, tableBits)] = t2
e.table[hashLen(cv>>8, tableBits, hashShortBytes)] = t2
i += 3
for ; i < s-1; i += 3 {
cv := load6432(src, i)
t := tableEntry{offset: i + e.cur}
t2 := tableEntry{offset: t.offset + 1}
e.bTable[hash7(cv, tableBits)] = t
e.bTable[hash7(cv>>8, tableBits)] = t2
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}
e.bTable[prevHashL] = tableEntry{offset: o}
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:])
}
}
vendor/github.com/klauspost/compress/flate/level5.go
Generated Vendored
+310
View File
@@ -0,0 +1,310 @@
package flate
import "fmt"
type fastEncL5 struct {
fastGen
table [tableSize]tableEntry
bTable [tableSize]tableEntryPrev
}
func (e *fastEncL5) Encode(dst *tokens, src []byte) {
const (
inputMargin = 12 - 1
minNonLiteralBlockSize = 1 + 1 + inputMargin
hashShortBytes = 4
)
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:])
}
}
vendor/github.com/klauspost/compress/flate/level6.go
Generated Vendored
+325
View File
@@ -0,0 +1,325 @@
package flate
import "fmt"
type fastEncL6 struct {
fastGen
table [tableSize]tableEntry
bTable [tableSize]tableEntryPrev
}
func (e *fastEncL6) Encode(dst *tokens, src []byte) {
const (
inputMargin = 12 - 1
minNonLiteralBlockSize = 1 + 1 + inputMargin
hashShortBytes = 4
)
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)
// Repeat MUST be > 1 and within range
repeat := int32(1)
for {
const skipLog = 7
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
// Calculate hashes of 'next'
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) {
// Long candidate matches at least 4 bytes.
// 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
// Check the previous long candidate as well.
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
}
// Current value did not match, but check if previous long value does.
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
// Look up next long candidate (at nextS)
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
// Check repeat at s + repOff
const repOff = 1
t2 := s - repeat + repOff
if load3232(src, t2) == uint32(cv>>(8*repOff)) {
ml := e.matchlen(s+4+repOff, t2+4, src) + 4
if ml > l {
t = t2
l = ml
s += repOff
// Not worth checking more.
break
}
}
// 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
// This is ok, but check previous as well.
}
}
// 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.
// Extend the 4-byte match as long as possible.
if l == 0 {
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-match...
if sAt := s + l; 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)]
// Test current
t2 := eLong.Cur.offset - e.cur - l + skipBeginning
s2 := s + skipBeginning
off := s2 - t2
if off < maxMatchOffset {
if off > 0 && t2 >= 0 {
if l2 := e.matchlenLong(s2, t2, src); l2 > l {
t = t2
l = l2
s = s2
}
}
// Test next:
t2 = eLong.Prev.offset - e.cur - l + skipBeginning
off := s2 - t2
if off > 0 && off < maxMatchOffset && t2 >= 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 false {
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))
repeat = s - t
s += l
nextEmit = s
if nextS >= s {
s = nextS + 1
}
if s >= sLimit {
// Index after match end.
for i := nextS + 1; i < int32(len(src))-8; i += 2 {
cv := load6432(src, i)
e.table[hashLen(cv, tableBits, hashShortBytes)] = tableEntry{offset: i + e.cur}
eLong := &e.bTable[hash7(cv, tableBits)]
eLong.Cur, eLong.Prev = tableEntry{offset: i + e.cur}, eLong.Cur
}
goto emitRemainder
}
// Store every long hash in-between and every second short.
if true {
for i := nextS + 1; i < s-1; i += 2 {
cv := load6432(src, i)
t := tableEntry{offset: i + e.cur}
t2 := tableEntry{offset: t.offset + 1}
eLong := &e.bTable[hash7(cv, tableBits)]
eLong2 := &e.bTable[hash7(cv>>8, tableBits)]
e.table[hashLen(cv, tableBits, hashShortBytes)] = t
eLong.Cur, eLong.Prev = t, eLong.Cur
eLong2.Cur, eLong2.Prev = t2, eLong2.Cur
}
}
// We could immediately start working at s now, but to improve
// compression we first update the hash table at s-1 and at s.
cv = load6432(src, s)
}
emitRemainder:
if int(nextEmit) < len(src) {
// If nothing was added, don't encode literals.
if dst.n == 0 {
return
}
emitLiteral(dst, src[nextEmit:])
}
}
vendor/github.com/klauspost/compress/flate/regmask_amd64.go
Generated Vendored
+37
View File
@@ -0,0 +1,37 @@
package flate
const (
// Masks for shifts with register sizes of the shift value.
// This can be used to work around the x86 design of shifting by mod register size.
// It can be used when a variable shift is always smaller than the register size.
// reg8SizeMaskX - shift value is 8 bits, shifted is X
reg8SizeMask8 = 7
reg8SizeMask16 = 15
reg8SizeMask32 = 31
reg8SizeMask64 = 63
// reg16SizeMaskX - shift value is 16 bits, shifted is X
reg16SizeMask8 = reg8SizeMask8
reg16SizeMask16 = reg8SizeMask16
reg16SizeMask32 = reg8SizeMask32
reg16SizeMask64 = reg8SizeMask64
// reg32SizeMaskX - shift value is 32 bits, shifted is X
reg32SizeMask8 = reg8SizeMask8
reg32SizeMask16 = reg8SizeMask16
reg32SizeMask32 = reg8SizeMask32
reg32SizeMask64 = reg8SizeMask64
// reg64SizeMaskX - shift value is 64 bits, shifted is X
reg64SizeMask8 = reg8SizeMask8
reg64SizeMask16 = reg8SizeMask16
reg64SizeMask32 = reg8SizeMask32
reg64SizeMask64 = reg8SizeMask64
// regSizeMaskUintX - shift value is uint, shifted is X
regSizeMaskUint8 = reg8SizeMask8
regSizeMaskUint16 = reg8SizeMask16
regSizeMaskUint32 = reg8SizeMask32
regSizeMaskUint64 = reg8SizeMask64
)
vendor/github.com/klauspost/compress/flate/regmask_other.go
Generated Vendored
+40
View File
@@ -0,0 +1,40 @@
//go:build !amd64
// +build !amd64
package flate
const (
// Masks for shifts with register sizes of the shift value.
// This can be used to work around the x86 design of shifting by mod register size.
// It can be used when a variable shift is always smaller than the register size.
// reg8SizeMaskX - shift value is 8 bits, shifted is X
reg8SizeMask8 = 0xff
reg8SizeMask16 = 0xff
reg8SizeMask32 = 0xff
reg8SizeMask64 = 0xff
// reg16SizeMaskX - shift value is 16 bits, shifted is X
reg16SizeMask8 = 0xffff
reg16SizeMask16 = 0xffff
reg16SizeMask32 = 0xffff
reg16SizeMask64 = 0xffff
// reg32SizeMaskX - shift value is 32 bits, shifted is X
reg32SizeMask8 = 0xffffffff
reg32SizeMask16 = 0xffffffff
reg32SizeMask32 = 0xffffffff
reg32SizeMask64 = 0xffffffff
// reg64SizeMaskX - shift value is 64 bits, shifted is X
reg64SizeMask8 = 0xffffffffffffffff
reg64SizeMask16 = 0xffffffffffffffff
reg64SizeMask32 = 0xffffffffffffffff
reg64SizeMask64 = 0xffffffffffffffff
// regSizeMaskUintX - shift value is uint, shifted is X
regSizeMaskUint8 = ^uint(0)
regSizeMaskUint16 = ^uint(0)
regSizeMaskUint32 = ^uint(0)
regSizeMaskUint64 = ^uint(0)
)
vendor/github.com/klauspost/compress/flate/stateless.go
Generated Vendored
+318
View File
@@ -0,0 +1,318 @@
package flate
import (
"io"
"math"
"sync"
)
const (
maxStatelessBlock = math.MaxInt16
// dictionary will be taken from maxStatelessBlock, so limit it.
maxStatelessDict = 8 << 10
slTableBits = 13
slTableSize = 1 << slTableBits
slTableShift = 32 - slTableBits
)
type statelessWriter struct {
dst io.Writer
closed bool
}
func (s *statelessWriter) Close() error {
if s.closed {
return nil
}
s.closed = true
// Emit EOF block
return StatelessDeflate(s.dst, nil, true, nil)
}
func (s *statelessWriter) Write(p []byte) (n int, err error) {
err = StatelessDeflate(s.dst, p, false, nil)
if err != nil {
return 0, err
}
return len(p), nil
}
func (s *statelessWriter) Reset(w io.Writer) {
s.dst = w
s.closed = false
}
// NewStatelessWriter will do compression but without maintaining any state
// between Write calls.
// There will be no memory kept between Write calls,
// but compression and speed will be suboptimal.
// Because of this, the size of actual Write calls will affect output size.
func NewStatelessWriter(dst io.Writer) io.WriteCloser {
return &statelessWriter{dst: dst}
}
// bitWriterPool contains bit writers that can be reused.
var bitWriterPool = sync.Pool{
New: func() interface{} {
return newHuffmanBitWriter(nil)
},
}
// StatelessDeflate allows compressing directly to a Writer without retaining state.
// When returning everything will be flushed.
// Up to 8KB of an optional dictionary can be given which is presumed to precede the block.
// Longer dictionaries will be truncated and will still produce valid output.
// Sending nil dictionary is perfectly fine.
func StatelessDeflate(out io.Writer, in []byte, eof bool, dict []byte) error {
var dst tokens
bw := bitWriterPool.Get().(*huffmanBitWriter)
bw.reset(out)
defer func() {
// don't keep a reference to our output
bw.reset(nil)
bitWriterPool.Put(bw)
}()
if eof && len(in) == 0 {
// Just write an EOF block.
// Could be faster...
bw.writeStoredHeader(0, true)
bw.flush()
return bw.err
}
// Truncate dict
if len(dict) > maxStatelessDict {
dict = dict[len(dict)-maxStatelessDict:]
}
// For subsequent loops, keep shallow dict reference to avoid alloc+copy.
var inDict []byte
for len(in) > 0 {
todo := in
if len(inDict) > 0 {
if len(todo) > maxStatelessBlock-maxStatelessDict {
todo = todo[:maxStatelessBlock-maxStatelessDict]
}
} else if len(todo) > maxStatelessBlock-len(dict) {
todo = todo[:maxStatelessBlock-len(dict)]
}
inOrg := in
in = in[len(todo):]
uncompressed := todo
if len(dict) > 0 {
// combine dict and source
bufLen := len(todo) + len(dict)
combined := make([]byte, bufLen)
copy(combined, dict)
copy(combined[len(dict):], todo)
todo = combined
}
// Compress
if len(inDict) == 0 {
statelessEnc(&dst, todo, int16(len(dict)))
} else {
statelessEnc(&dst, inDict[:maxStatelessDict+len(todo)], maxStatelessDict)
}
isEof := eof && len(in) == 0
if dst.n == 0 {
bw.writeStoredHeader(len(uncompressed), isEof)
if bw.err != nil {
return bw.err
}
bw.writeBytes(uncompressed)
} else if int(dst.n) > len(uncompressed)-len(uncompressed)>>4 {
// If we removed less than 1/16th, huffman compress the block.
bw.writeBlockHuff(isEof, uncompressed, len(in) == 0)
} else {
bw.writeBlockDynamic(&dst, isEof, uncompressed, len(in) == 0)
}
if len(in) > 0 {
// Retain a dict if we have more
inDict = inOrg[len(uncompressed)-maxStatelessDict:]
dict = nil
dst.Reset()
}
if bw.err != nil {
return bw.err
}
}
if !eof {
// Align, only a stored block can do that.
bw.writeStoredHeader(0, false)
}
bw.flush()
return bw.err
}
func hashSL(u uint32) uint32 {
return (u * 0x1e35a7bd) >> slTableShift
}
func load3216(b []byte, i int16) uint32 {
// Help the compiler eliminate bounds checks on the read so it can be done in a single read.
b = b[i:]
b = b[:4]
return uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24
}
func load6416(b []byte, i int16) uint64 {
// Help the compiler eliminate bounds checks on the read so it can be done in a single read.
b = b[i:]
b = b[:8]
return uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 |
uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56
}
func statelessEnc(dst *tokens, src []byte, startAt int16) {
const (
inputMargin = 12 - 1
minNonLiteralBlockSize = 1 + 1 + inputMargin
)
type tableEntry struct {
offset int16
}
var table [slTableSize]tableEntry
// This check isn't in the Snappy implementation, but there, the caller
// instead of the callee handles this case.
if len(src)-int(startAt) < minNonLiteralBlockSize {
// We do not fill the token table.
// This will be picked up by caller.
dst.n = 0
return
}
// Index until startAt
if startAt > 0 {
cv := load3232(src, 0)
for i := int16(0); i < startAt; i++ {
table[hashSL(cv)] = tableEntry{offset: i}
cv = (cv >> 8) | (uint32(src[i+4]) << 24)
}
}
s := startAt + 1
nextEmit := startAt
// 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 := int16(len(src) - inputMargin)
// nextEmit is where in src the next emitLiteral should start from.
cv := load3216(src, s)
for {
const skipLog = 5
const doEvery = 2
nextS := s
var candidate tableEntry
for {
nextHash := hashSL(cv)
candidate = table[nextHash]
nextS = s + doEvery + (s-nextEmit)>>skipLog
if nextS > sLimit || nextS <= 0 {
goto emitRemainder
}
now := load6416(src, nextS)
table[nextHash] = tableEntry{offset: s}
nextHash = hashSL(uint32(now))
if cv == load3216(src, candidate.offset) {
table[nextHash] = tableEntry{offset: nextS}
break
}
// Do one right away...
cv = uint32(now)
s = nextS
nextS++
candidate = table[nextHash]
now >>= 8
table[nextHash] = tableEntry{offset: s}
if cv == load3216(src, candidate.offset) {
table[nextHash] = tableEntry{offset: nextS}
break
}
cv = uint32(now)
s = nextS
}
// 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.
for {
// Invariant: we have a 4-byte match at s, and no need to emit any
// literal bytes prior to s.
// Extend the 4-byte match as long as possible.
t := candidate.offset
l := int16(matchLen(src[s+4:], src[t+4:]) + 4)
// 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++
}
}
}
// Save the match found
dst.AddMatchLong(int32(l), uint32(s-t-baseMatchOffset))
s += l
nextEmit = s
if nextS >= s {
s = nextS + 1
}
if s >= sLimit {
goto emitRemainder
}
// We could immediately start working at s now, but to improve
// compression we first update the hash table at s-2 and at s. If
// another emitCopy is not our next move, also calculate nextHash
// at s+1. At least on GOARCH=amd64, these three hash calculations
// are faster as one load64 call (with some shifts) instead of
// three load32 calls.
x := load6416(src, s-2)
o := s - 2
prevHash := hashSL(uint32(x))
table[prevHash] = tableEntry{offset: o}
x >>= 16
currHash := hashSL(uint32(x))
candidate = table[currHash]
table[currHash] = tableEntry{offset: o + 2}
if uint32(x) != load3216(src, candidate.offset) {
cv = uint32(x >> 8)
s++
break
}
}
}
emitRemainder:
if int(nextEmit) < len(src) {
// If nothing was added, don't encode literals.
if dst.n == 0 {
return
}
emitLiteral(dst, src[nextEmit:])
}
}
vendor/github.com/klauspost/compress/flate/token.go
Generated Vendored
+379
View File
@@ -0,0 +1,379 @@
// Copyright 2009 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package flate
import (
"bytes"
"encoding/binary"
"fmt"
"io"
"math"
)
const (
// bits 0-16 xoffset = offset - MIN_OFFSET_SIZE, or literal - 16 bits
// bits 16-22 offsetcode - 5 bits
// bits 22-30 xlength = length - MIN_MATCH_LENGTH - 8 bits
// bits 30-32 type 0 = literal 1=EOF 2=Match 3=Unused - 2 bits
lengthShift = 22
offsetMask = 1<<lengthShift - 1
typeMask = 3 << 30
literalType = 0 << 30
matchType = 1 << 30
matchOffsetOnlyMask = 0xffff
)
// The length code for length X (MIN_MATCH_LENGTH <= X <= MAX_MATCH_LENGTH)
// is lengthCodes[length - MIN_MATCH_LENGTH]
var lengthCodes = [256]uint8{
0, 1, 2, 3, 4, 5, 6, 7, 8, 8,
9, 9, 10, 10, 11, 11, 12, 12, 12, 12,
13, 13, 13, 13, 14, 14, 14, 14, 15, 15,
15, 15, 16, 16, 16, 16, 16, 16, 16, 16,
17, 17, 17, 17, 17, 17, 17, 17, 18, 18,
18, 18, 18, 18, 18, 18, 19, 19, 19, 19,
19, 19, 19, 19, 20, 20, 20, 20, 20, 20,
20, 20, 20, 20, 20, 20, 20, 20, 20, 20,
21, 21, 21, 21, 21, 21, 21, 21, 21, 21,
21, 21, 21, 21, 21, 21, 22, 22, 22, 22,
22, 22, 22, 22, 22, 22, 22, 22, 22, 22,
22, 22, 23, 23, 23, 23, 23, 23, 23, 23,
23, 23, 23, 23, 23, 23, 23, 23, 24, 24,
24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
25, 25, 26, 26, 26, 26, 26, 26, 26, 26,
26, 26, 26, 26, 26, 26, 26, 26, 26, 26,
26, 26, 26, 26, 26, 26, 26, 26, 26, 26,
26, 26, 26, 26, 27, 27, 27, 27, 27, 27,
27, 27, 27, 27, 27, 27, 27, 27, 27, 27,
27, 27, 27, 27, 27, 27, 27, 27, 27, 27,
27, 27, 27, 27, 27, 28,
}
// lengthCodes1 is length codes, but starting at 1.
var lengthCodes1 = [256]uint8{
1, 2, 3, 4, 5, 6, 7, 8, 9, 9,
10, 10, 11, 11, 12, 12, 13, 13, 13, 13,
14, 14, 14, 14, 15, 15, 15, 15, 16, 16,
16, 16, 17, 17, 17, 17, 17, 17, 17, 17,
18, 18, 18, 18, 18, 18, 18, 18, 19, 19,
19, 19, 19, 19, 19, 19, 20, 20, 20, 20,
20, 20, 20, 20, 21, 21, 21, 21, 21, 21,
21, 21, 21, 21, 21, 21, 21, 21, 21, 21,
22, 22, 22, 22, 22, 22, 22, 22, 22, 22,
22, 22, 22, 22, 22, 22, 23, 23, 23, 23,
23, 23, 23, 23, 23, 23, 23, 23, 23, 23,
23, 23, 24, 24, 24, 24, 24, 24, 24, 24,
24, 24, 24, 24, 24, 24, 24, 24, 25, 25,
25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
26, 26, 26, 26, 26, 26, 26, 26, 26, 26,
26, 26, 26, 26, 26, 26, 26, 26, 26, 26,
26, 26, 26, 26, 26, 26, 26, 26, 26, 26,
26, 26, 27, 27, 27, 27, 27, 27, 27, 27,
27, 27, 27, 27, 27, 27, 27, 27, 27, 27,
27, 27, 27, 27, 27, 27, 27, 27, 27, 27,
27, 27, 27, 27, 28, 28, 28, 28, 28, 28,
28, 28, 28, 28, 28, 28, 28, 28, 28, 28,
28, 28, 28, 28, 28, 28, 28, 28, 28, 28,
28, 28, 28, 28, 28, 29,
}
var offsetCodes = [256]uint32{
0, 1, 2, 3, 4, 4, 5, 5, 6, 6, 6, 6, 7, 7, 7, 7,
8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9,
10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12,
12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12,
13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13,
13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13,
14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14,
14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14,
14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14,
14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14,
15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15,
15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15,
15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15,
15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15,
}
// offsetCodes14 are offsetCodes, but with 14 added.
var offsetCodes14 = [256]uint32{
14, 15, 16, 17, 18, 18, 19, 19, 20, 20, 20, 20, 21, 21, 21, 21,
22, 22, 22, 22, 22, 22, 22, 22, 23, 23, 23, 23, 23, 23, 23, 23,
24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26,
26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26,
27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27,
27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27,
28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28,
28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28,
28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28,
28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28,
29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29,
29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29,
29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29,
29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29,
}
type token uint32
type tokens struct {
extraHist [32]uint16 // codes 256->maxnumlit
offHist [32]uint16 // offset codes
litHist [256]uint16 // codes 0->255
nFilled int
n uint16 // Must be able to contain maxStoreBlockSize
tokens [maxStoreBlockSize + 1]token
}
func (t *tokens) Reset() {
if t.n == 0 {
return
}
t.n = 0
t.nFilled = 0
for i := range t.litHist[:] {
t.litHist[i] = 0
}
for i := range t.extraHist[:] {
t.extraHist[i] = 0
}
for i := range t.offHist[:] {
t.offHist[i] = 0
}
}
func (t *tokens) Fill() {
if t.n == 0 {
return
}
for i, v := range t.litHist[:] {
if v == 0 {
t.litHist[i] = 1
t.nFilled++
}
}
for i, v := range t.extraHist[:literalCount-256] {
if v == 0 {
t.nFilled++
t.extraHist[i] = 1
}
}
for i, v := range t.offHist[:offsetCodeCount] {
if v == 0 {
t.offHist[i] = 1
}
}
}
func indexTokens(in []token) tokens {
var t tokens
t.indexTokens(in)
return t
}
func (t *tokens) indexTokens(in []token) {
t.Reset()
for _, tok := range in {
if tok < matchType {
t.AddLiteral(tok.literal())
continue
}
t.AddMatch(uint32(tok.length()), tok.offset()&matchOffsetOnlyMask)
}
}
// emitLiteral writes a literal chunk and returns the number of bytes written.
func emitLiteral(dst *tokens, lit []byte) {
for _, v := range lit {
dst.tokens[dst.n] = token(v)
dst.litHist[v]++
dst.n++
}
}
func (t *tokens) AddLiteral(lit byte) {
t.tokens[t.n] = token(lit)
t.litHist[lit]++
t.n++
}
// from https://stackoverflow.com/a/28730362
func mFastLog2(val float32) float32 {
ux := int32(math.Float32bits(val))
log2 := (float32)(((ux >> 23) & 255) - 128)
ux &= -0x7f800001
ux += 127 << 23
uval := math.Float32frombits(uint32(ux))
log2 += ((-0.34484843)*uval+2.02466578)*uval - 0.67487759
return log2
}
// EstimatedBits will return an minimum size estimated by an *optimal*
// compression of the block.
// The size of the block
func (t *tokens) EstimatedBits() int {
shannon := float32(0)
bits := int(0)
nMatches := 0
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)
shannon += atLeastOne(-mFastLog2(n*invTotal)) * n
}
}
// Just add 15 for EOB
shannon += 15
for i, v := range t.extraHist[1 : literalCount-256] {
if v > 0 {
n := float32(v)
shannon += atLeastOne(-mFastLog2(n*invTotal)) * n
bits += int(lengthExtraBits[i&31]) * int(v)
nMatches += int(v)
}
}
}
if nMatches > 0 {
invTotal := 1.0 / float32(nMatches)
for i, v := range t.offHist[:offsetCodeCount] {
if v > 0 {
n := float32(v)
shannon += atLeastOne(-mFastLog2(n*invTotal)) * n
bits += int(offsetExtraBits[i&31]) * int(v)
}
}
}
return int(shannon) + bits
}
// 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 debugDeflate {
if xlength >= maxMatchLength+baseMatchLength {
panic(fmt.Errorf("invalid length: %v", xlength))
}
if xoffset >= maxMatchOffset+baseMatchOffset {
panic(fmt.Errorf("invalid offset: %v", xoffset))
}
}
oCode := offsetCode(xoffset)
xoffset |= oCode << 16
t.extraHist[lengthCodes1[uint8(xlength)]]++
t.offHist[oCode&31]++
t.tokens[t.n] = token(matchType | xlength<<lengthShift | xoffset)
t.n++
}
// 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 debugDeflate {
if xoffset >= maxMatchOffset+baseMatchOffset {
panic(fmt.Errorf("invalid offset: %v", xoffset))
}
}
oc := offsetCode(xoffset)
xoffset |= oc << 16
for xlength > 0 {
xl := xlength
if xl > 258 {
// We need to have at least baseMatchLength left over for next loop.
if xl > 258+baseMatchLength {
xl = 258
} else {
xl = 258 - baseMatchLength
}
}
xlength -= xl
xl -= baseMatchLength
t.extraHist[lengthCodes1[uint8(xl)]]++
t.offHist[oc&31]++
t.tokens[t.n] = token(matchType | uint32(xl)<<lengthShift | xoffset)
t.n++
}
}
func (t *tokens) AddEOB() {
t.tokens[t.n] = token(endBlockMarker)
t.extraHist[0]++
t.n++
}
func (t *tokens) Slice() []token {
return t.tokens[:t.n]
}
// VarInt returns the tokens as varint encoded bytes.
func (t *tokens) VarInt() []byte {
var b = make([]byte, binary.MaxVarintLen32*int(t.n))
var off int
for _, v := range t.tokens[:t.n] {
off += binary.PutUvarint(b[off:], uint64(v))
}
return b[:off]
}
// FromVarInt restores t to the varint encoded tokens provided.
// Any data in t is removed.
func (t *tokens) FromVarInt(b []byte) error {
var buf = bytes.NewReader(b)
var toks []token
for {
r, err := binary.ReadUvarint(buf)
if err == io.EOF {
break
}
if err != nil {
return err
}
toks = append(toks, token(r))
}
t.indexTokens(toks)
return nil
}
// Returns the type of a token
func (t token) typ() uint32 { return uint32(t) & typeMask }
// Returns the literal of a literal token
func (t token) literal() uint8 { return uint8(t) }
// Returns the extra offset of a match token
func (t token) offset() uint32 { return uint32(t) & offsetMask }
func (t token) length() uint8 { return uint8(t >> lengthShift) }
// Convert length to code.
func lengthCode(len uint8) uint8 { return lengthCodes[len] }
// Returns the offset code corresponding to a specific offset
func offsetCode(off uint32) uint32 {
if false {
if off < uint32(len(offsetCodes)) {
return offsetCodes[off&255]
} else if off>>7 < uint32(len(offsetCodes)) {
return offsetCodes[(off>>7)&255] + 14
} else {
return offsetCodes[(off>>14)&255] + 28
}
}
if off < uint32(len(offsetCodes)) {
return offsetCodes[uint8(off)]
}
return offsetCodes14[uint8(off>>7)]
}
vendor/github.com/klauspost/compress/s2/README.md
Generated Vendored
+185 -1
View File
@@ -20,11 +20,12 @@ This is important, so you don't have to worry about spending CPU cycles on alrea
* Concurrent stream compression
* Faster decompression, even for Snappy compatible content
* Concurrent Snappy/S2 stream decompression
* Ability to quickly skip forward in compressed stream
* Skip forward in compressed stream
* Random seeking with indexes
* Compatible with reading Snappy compressed content
* Smaller block size overhead on incompressible blocks
* Block concatenation
* Block Dictionary support
* Uncompressed stream mode
* Automatic stream size padding
* Snappy compatible block compression
@@ -594,6 +595,123 @@ Best... 10737418240 -> 4210602774 [39.21%]; 42.96s, 254.4MB/s
Decompression speed should be around the same as using the 'better' compression mode.
## Dictionaries
*Note: S2 dictionary compression is currently at an early implementation stage, with no assembly for
neither encoding nor decoding. Performance improvements can be expected in the future.*
Adding dictionaries allow providing a custom dictionary that will serve as lookup in the beginning of blocks.
The same dictionary *must* be used for both encoding and decoding.
S2 does not keep track of whether the same dictionary is used,
and using the wrong dictionary will most often not result in an error when decompressing.
Blocks encoded *without* dictionaries can be decompressed seamlessly *with* a dictionary.
This means it is possible to switch from an encoding without dictionaries to an encoding with dictionaries
and treat the blocks similarly.
Similar to [zStandard dictionaries](https://github.com/facebook/zstd#the-case-for-small-data-compression),
the same usage scenario applies to S2 dictionaries.
> Training works if there is some correlation in a family of small data samples. The more data-specific a dictionary is, the more efficient it is (there is no universal dictionary). Hence, deploying one dictionary per type of data will provide the greatest benefits. Dictionary gains are mostly effective in the first few KB. Then, the compression algorithm will gradually use previously decoded content to better compress the rest of the file.
S2 further limits the dictionary to only be enabled on the first 64KB of a block.
This will remove any negative (speed) impacts of the dictionaries on bigger blocks.
### Compression
Using the [github_users_sample_set](https://github.com/facebook/zstd/releases/download/v1.1.3/github_users_sample_set.tar.zst)
and a 64KB dictionary trained with zStandard the following sizes can be achieved.
| | Default | Better | Best |
|--------------------|------------------|------------------|-----------------------|
| Without Dictionary | 3362023 (44.92%) | 3083163 (41.19%) | 3057944 (40.86%) |
| With Dictionary | 921524 (12.31%) | 873154 (11.67%) | 785503 bytes (10.49%) |
So for highly repetitive content, this case provides an almost 3x reduction in size.
For less uniform data we will use the Go source code tree.
Compressing First 64KB of all `.go` files in `go/src`, Go 1.19.5, 8912 files, 51253563 bytes input:
| | Default | Better | Best |
|--------------------|-------------------|-------------------|-------------------|
| Without Dictionary | 22955767 (44.79%) | 20189613 (39.39% | 19482828 (38.01%) |
| With Dictionary | 19654568 (38.35%) | 16289357 (31.78%) | 15184589 (29.63%) |
| Saving/file | 362 bytes | 428 bytes | 472 bytes |
### Creating Dictionaries
There are no tools to create dictionaries in S2.
However, there are multiple ways to create a useful dictionary:
#### Using a Sample File
If your input is very uniform, you can just use a sample file as the dictionary.
For example in the `github_users_sample_set` above, the average compression only goes up from
10.49% to 11.48% by using the first file as dictionary compared to using a dedicated dictionary.
```Go
// Read a sample
sample, err := os.ReadFile("sample.json")
// Create a dictionary.
dict := s2.MakeDict(sample, nil)
// b := dict.Bytes() will provide a dictionary that can be saved
// and reloaded with s2.NewDict(b).
// To encode:
encoded := dict.Encode(nil, file)
// To decode:
decoded, err := dict.Decode(nil, file)
```
#### Using Zstandard
Zstandard dictionaries can easily be converted to S2 dictionaries.
This can be helpful to generate dictionaries for files that don't have a fixed structure.
Example, with training set files placed in `./training-set`:
`λ zstd -r --train-fastcover training-set/* --maxdict=65536 -o name.dict`
This will create a dictionary of 64KB, that can be converted to a dictionary like this:
```Go
// Decode the Zstandard dictionary.
insp, err := zstd.InspectDictionary(zdict)
if err != nil {
panic(err)
}
// We are only interested in the contents.
// Assume that files start with "// Copyright (c) 2023".
// Search for the longest match for that.
// This may save a few bytes.
dict := s2.MakeDict(insp.Content(), []byte("// Copyright (c) 2023"))
// b := dict.Bytes() will provide a dictionary that can be saved
// and reloaded with s2.NewDict(b).
// We can now encode using this dictionary
encodedWithDict := dict.Encode(nil, payload)
// To decode content:
decoded, err := dict.Decode(nil, encodedWithDict)
```
It is recommended to save the dictionary returned by ` b:= dict.Bytes()`, since that will contain only the S2 dictionary.
This dictionary can later be loaded using `s2.NewDict(b)`. The dictionary then no longer requires `zstd` to be initialized.
Also note how `s2.MakeDict` allows you to search for a common starting sequence of your files.
This can be omitted, at the expense of a few bytes.
# Snappy Compatibility
S2 now offers full compatibility with Snappy.
@@ -929,6 +1047,72 @@ The first copy of a block cannot be a repeat offset and the offset is reset on e
Default streaming block size is 1MB.
# Dictionary Encoding
Adding dictionaries allow providing a custom dictionary that will serve as lookup in the beginning of blocks.
A dictionary provides an initial repeat value that can be used to point to a common header.
Other than that the dictionary contains values that can be used as back-references.
Often used data should be placed at the *end* of the dictionary since offsets < 2048 bytes will be smaller.
## Format
Dictionary *content* must at least 16 bytes and less or equal to 64KiB (65536 bytes).
Encoding: `[repeat value (uvarint)][dictionary content...]`
Before the dictionary content, an unsigned base-128 (uvarint) encoded value specifying the initial repeat offset.
This value is an offset into the dictionary content and not a back-reference offset,
so setting this to 0 will make the repeat value point to the first value of the dictionary.
The value must be less than the dictionary length-8
## Encoding
From the decoder point of view the dictionary content is seen as preceding the encoded content.
`[dictionary content][decoded output]`
Backreferences to the dictionary are encoded as ordinary backreferences that have an offset before the start of the decoded block.
Matches copying from the dictionary are **not** allowed to cross from the dictionary into the decoded data.
However, if a copy ends at the end of the dictionary the next repeat will point to the start of the decoded buffer, which is allowed.
The first match can be a repeat value, which will use the repeat offset stored in the dictionary.
When 64KB (65536 bytes) has been en/decoded it is no longer allowed to reference the dictionary,
neither by a copy nor repeat operations.
If the boundary is crossed while copying from the dictionary, the operation should complete,
but the next instruction is not allowed to reference the dictionary.
Valid blocks encoded *without* a dictionary can be decoded with any dictionary.
There are no checks whether the supplied dictionary is the correct for a block.
Because of this there is no overhead by using a dictionary.
## Example
This is the dictionary content. Elements are separated by `[]`.
Dictionary: `[0x0a][Yesterday 25 bananas were added to Benjamins brown bag]`.
Initial repeat offset is set at 10, which is the letter `2`.
Encoded `[LIT "10"][REPEAT len=10][LIT "hich"][MATCH off=50 len=6][MATCH off=31 len=6][MATCH off=61 len=10]`
Decoded: `[10][ bananas w][hich][ were ][brown ][were added]`
Output: `10 bananas which were brown were added`
## Streams
For streams each block can use the dictionary.
The dictionary cannot not currently be provided on the stream.
# LICENSE
This code is based on the [Snappy-Go](https://github.com/golang/snappy) implementation.
vendor/github.com/klauspost/compress/s2/decode.go
Generated Vendored
+325 -934
View File
File diff suppressed because it is too large Load Diff
vendor/github.com/klauspost/compress/s2/decode_other.go
Generated Vendored
+13
View File
@@ -57,6 +57,9 @@ func s2Decode(dst, src []byte) int {
}
length = int(x) + 1
if length > len(dst)-d || length > len(src)-s || (strconv.IntSize == 32 && length <= 0) {
if debug {
fmt.Println("corrupt: lit size", length)
}
return decodeErrCodeCorrupt
}
if debug {
@@ -109,6 +112,10 @@ func s2Decode(dst, src []byte) int {
}
if offset <= 0 || d < offset || length > len(dst)-d {
if debug {
fmt.Println("corrupt: match, length", length, "offset:", offset, "dst avail:", len(dst)-d, "dst pos:", d)
}
return decodeErrCodeCorrupt
}
@@ -175,6 +182,9 @@ func s2Decode(dst, src []byte) int {
}
length = int(x) + 1
if length > len(dst)-d || length > len(src)-s || (strconv.IntSize == 32 && length <= 0) {
if debug {
fmt.Println("corrupt: lit size", length)
}
return decodeErrCodeCorrupt
}
if debug {
@@ -241,6 +251,9 @@ func s2Decode(dst, src []byte) int {
}
if offset <= 0 || d < offset || length > len(dst)-d {
if debug {
fmt.Println("corrupt: match, length", length, "offset:", offset, "dst avail:", len(dst)-d, "dst pos:", d)
}
return decodeErrCodeCorrupt
}
vendor/github.com/klauspost/compress/s2/dict.go
Generated Vendored
+331
View File
@@ -0,0 +1,331 @@
// Copyright (c) 2022+ Klaus Post. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package s2
import (
"bytes"
"encoding/binary"
"sync"
)
const (
// MinDictSize is the minimum dictionary size when repeat has been read.
MinDictSize = 16
// MaxDictSize is the maximum dictionary size when repeat has been read.
MaxDictSize = 65536
// MaxDictSrcOffset is the maximum offset where a dictionary entry can start.
MaxDictSrcOffset = 65535
)
// Dict contains a dictionary that can be used for encoding and decoding s2
type Dict struct {
dict []byte
repeat int // Repeat as index of dict
fast, better, best sync.Once
fastTable *[1 << 14]uint16
betterTableShort *[1 << 14]uint16
betterTableLong *[1 << 17]uint16
bestTableShort *[1 << 16]uint32
bestTableLong *[1 << 19]uint32
}
// NewDict will read a dictionary.
// It will return nil if the dictionary is invalid.
func NewDict(dict []byte) *Dict {
if len(dict) == 0 {
return nil
}
var d Dict
// Repeat is the first value of the dict
r, n := binary.Uvarint(dict)
if n <= 0 {
return nil
}
dict = dict[n:]
d.dict = dict
if cap(d.dict) < len(d.dict)+16 {
d.dict = append(make([]byte, 0, len(d.dict)+16), d.dict...)
}
if len(dict) < MinDictSize || len(dict) > MaxDictSize {
return nil
}
d.repeat = int(r)
if d.repeat > len(dict) {
return nil
}
return &d
}
// Bytes will return a serialized version of the dictionary.
// The output can be sent to NewDict.
func (d *Dict) Bytes() []byte {
dst := make([]byte, binary.MaxVarintLen16+len(d.dict))
return append(dst[:binary.PutUvarint(dst, uint64(d.repeat))], d.dict...)
}
// MakeDict will create a dictionary.
// 'data' must be at least MinDictSize.
// If data is longer than MaxDictSize only the last MaxDictSize bytes will be used.
// If searchStart is set the start repeat value will be set to the last
// match of this content.
// If no matches are found, it will attempt to find shorter matches.
// This content should match the typical start of a block.
// If at least 4 bytes cannot be matched, repeat is set to start of block.
func MakeDict(data []byte, searchStart []byte) *Dict {
if len(data) == 0 {
return nil
}
if len(data) > MaxDictSize {
data = data[len(data)-MaxDictSize:]
}
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...)
}
if len(dict) < MinDictSize {
return nil
}
// Find the longest match possible, last entry if multiple.
for s := len(searchStart); s > 4; s-- {
if idx := bytes.LastIndex(data, searchStart[:s]); idx >= 0 && idx <= len(data)-8 {
d.repeat = idx
break
}
}
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.
//
// The dst and src must not overlap. It is valid to pass a nil dst.
//
// The blocks will require the same amount of memory to decode as encoding,
// and does not make for concurrent decoding.
// Also note that blocks do not contain CRC information, so corruption may be undetected.
//
// If you need to encode larger amounts of data, consider using
// the streaming interface which gives all of these features.
func (d *Dict) Encode(dst, src []byte) []byte {
if n := MaxEncodedLen(len(src)); n < 0 {
panic(ErrTooLarge)
} else if cap(dst) < n {
dst = make([]byte, n)
} else {
dst = dst[:n]
}
// The block starts with the varint-encoded length of the decompressed bytes.
dstP := binary.PutUvarint(dst, uint64(len(src)))
if len(src) == 0 {
return dst[:dstP]
}
if len(src) < minNonLiteralBlockSize {
dstP += emitLiteral(dst[dstP:], src)
return dst[:dstP]
}
n := encodeBlockDictGo(dst[dstP:], src, d)
if n > 0 {
dstP += n
return dst[:dstP]
}
// Not compressible
dstP += emitLiteral(dst[dstP:], src)
return dst[:dstP]
}
// EncodeBetter 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.
//
// EncodeBetter compresses better than Encode but typically with a
// 10-40% speed decrease on both compression and decompression.
//
// The dst and src must not overlap. It is valid to pass a nil dst.
//
// The blocks will require the same amount of memory to decode as encoding,
// and does not make for concurrent decoding.
// Also note that blocks do not contain CRC information, so corruption may be undetected.
//
// If you need to encode larger amounts of data, consider using
// the streaming interface which gives all of these features.
func (d *Dict) EncodeBetter(dst, src []byte) []byte {
if n := MaxEncodedLen(len(src)); n < 0 {
panic(ErrTooLarge)
} else if len(dst) < n {
dst = make([]byte, n)
}
// The block starts with the varint-encoded length of the decompressed bytes.
dstP := binary.PutUvarint(dst, uint64(len(src)))
if len(src) == 0 {
return dst[:dstP]
}
if len(src) < minNonLiteralBlockSize {
dstP += emitLiteral(dst[dstP:], src)
return dst[:dstP]
}
n := encodeBlockBetterDict(dst[dstP:], src, d)
if n > 0 {
dstP += n
return dst[:dstP]
}
// Not compressible
dstP += emitLiteral(dst[dstP:], src)
return dst[:dstP]
}
// EncodeBest 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.
//
// EncodeBest compresses as good as reasonably possible but with a
// big speed decrease.
//
// The dst and src must not overlap. It is valid to pass a nil dst.
//
// The blocks will require the same amount of memory to decode as encoding,
// and does not make for concurrent decoding.
// Also note that blocks do not contain CRC information, so corruption may be undetected.
//
// If you need to encode larger amounts of data, consider using
// the streaming interface which gives all of these features.
func (d *Dict) EncodeBest(dst, src []byte) []byte {
if n := MaxEncodedLen(len(src)); n < 0 {
panic(ErrTooLarge)
} else if len(dst) < n {
dst = make([]byte, n)
}
// The block starts with the varint-encoded length of the decompressed bytes.
dstP := binary.PutUvarint(dst, uint64(len(src)))
if len(src) == 0 {
return dst[:dstP]
}
if len(src) < minNonLiteralBlockSize {
dstP += emitLiteral(dst[dstP:], src)
return dst[:dstP]
}
n := encodeBlockBest(dst[dstP:], src, d)
if n > 0 {
dstP += n
return dst[:dstP]
}
// Not compressible
dstP += emitLiteral(dst[dstP:], src)
return dst[:dstP]
}
// Decode returns the decoded form of src. The returned slice may be a sub-
// slice of dst if dst was large enough to hold the entire decoded block.
// Otherwise, a newly allocated slice will be returned.
//
// The dst and src must not overlap. It is valid to pass a nil dst.
func (d *Dict) Decode(dst, src []byte) ([]byte, error) {
dLen, s, err := decodedLen(src)
if err != nil {
return nil, err
}
if dLen <= cap(dst) {
dst = dst[:dLen]
} else {
dst = make([]byte, dLen)
}
if s2DecodeDict(dst, src[s:], d) != 0 {
return nil, ErrCorrupt
}
return dst, nil
}
func (d *Dict) initFast() {
d.fast.Do(func() {
const (
tableBits = 14
maxTableSize = 1 << tableBits
)
var table [maxTableSize]uint16
// We stop so any entry of length 8 can always be read.
for i := 0; i < len(d.dict)-8-2; i += 3 {
x0 := load64(d.dict, i)
h0 := hash6(x0, tableBits)
h1 := hash6(x0>>8, tableBits)
h2 := hash6(x0>>16, tableBits)
table[h0] = uint16(i)
table[h1] = uint16(i + 1)
table[h2] = uint16(i + 2)
}
d.fastTable = &table
})
}
func (d *Dict) initBetter() {
d.better.Do(func() {
const (
// Long hash matches.
lTableBits = 17
maxLTableSize = 1 << lTableBits
// Short hash matches.
sTableBits = 14
maxSTableSize = 1 << sTableBits
)
var lTable [maxLTableSize]uint16
var sTable [maxSTableSize]uint16
// We stop so any entry of length 8 can always be read.
for i := 0; i < len(d.dict)-8; i++ {
cv := load64(d.dict, i)
lTable[hash7(cv, lTableBits)] = uint16(i)
sTable[hash4(cv, sTableBits)] = uint16(i)
}
d.betterTableShort = &sTable
d.betterTableLong = &lTable
})
}
func (d *Dict) initBest() {
d.best.Do(func() {
const (
// Long hash matches.
lTableBits = 19
maxLTableSize = 1 << lTableBits
// Short hash matches.
sTableBits = 16
maxSTableSize = 1 << sTableBits
)
var lTable [maxLTableSize]uint32
var sTable [maxSTableSize]uint32
// We stop so any entry of length 8 can always be read.
for i := 0; i < len(d.dict)-8; i++ {
cv := load64(d.dict, i)
hashL := hash8(cv, lTableBits)
hashS := hash4(cv, sTableBits)
candidateL := lTable[hashL]
candidateS := sTable[hashS]
lTable[hashL] = uint32(i) | candidateL<<16
sTable[hashS] = uint32(i) | candidateS<<16
}
d.bestTableShort = &sTable
d.bestTableLong = &lTable
})
}
vendor/github.com/klauspost/compress/s2/encode.go
Generated Vendored
+47 -995
View File
File diff suppressed because it is too large Load Diff
vendor/github.com/klauspost/compress/s2/encode_all.go
Generated Vendored
+592
View File
@@ -8,6 +8,7 @@ package s2
import (
"bytes"
"encoding/binary"
"fmt"
"math/bits"
)
@@ -455,3 +456,594 @@ emitRemainder:
}
return d
}
// encodeBlockGo encodes a non-empty src to a guaranteed-large-enough dst. It
// assumes that the varint-encoded length of the decompressed bytes has already
// been written.
//
// It also assumes that:
//
// len(dst) >= MaxEncodedLen(len(src)) &&
// minNonLiteralBlockSize <= len(src) && len(src) <= maxBlockSize
func encodeBlockDictGo(dst, src []byte, dict *Dict) (d int) {
// Initialize the hash table.
const (
tableBits = 14
maxTableSize = 1 << tableBits
maxAhead = 8 // maximum bytes ahead without checking sLimit
debug = false
)
dict.initFast()
var table [maxTableSize]uint32
// 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 := len(src) - inputMargin
if sLimit > MaxDictSrcOffset-maxAhead {
sLimit = MaxDictSrcOffset - maxAhead
}
// Bail if we can't compress to at least this.
dstLimit := len(src) - len(src)>>5 - 5
// nextEmit is where in src the next emitLiteral should start from.
nextEmit := 0
// The encoded form can start with a dict entry (copy or repeat).
s := 0
// Convert dict repeat to offset
repeat := len(dict.dict) - dict.repeat
cv := load64(src, 0)
// While in dict
searchDict:
for {
// Next src position to check
nextS := s + (s-nextEmit)>>6 + 4
hash0 := hash6(cv, tableBits)
hash1 := hash6(cv>>8, tableBits)
if nextS > sLimit {
if debug {
fmt.Println("slimit reached", s, nextS)
}
break searchDict
}
candidateDict := int(dict.fastTable[hash0])
candidateDict2 := int(dict.fastTable[hash1])
candidate2 := int(table[hash1])
candidate := int(table[hash0])
table[hash0] = uint32(s)
table[hash1] = uint32(s + 1)
hash2 := hash6(cv>>16, tableBits)
// Check repeat at offset checkRep.
const checkRep = 1
if repeat > s {
candidate := len(dict.dict) - repeat + s
if repeat-s >= 4 && uint32(cv) == load32(dict.dict, candidate) {
// Extend back
base := s
for i := candidate; base > nextEmit && i > 0 && dict.dict[i-1] == src[base-1]; {
i--
base--
}
d += emitLiteral(dst[d:], src[nextEmit:base])
if debug && nextEmit != base {
fmt.Println("emitted ", base-nextEmit, "literals")
}
s += 4
candidate += 4
for candidate < len(dict.dict)-8 && s <= len(src)-8 {
if diff := load64(src, s) ^ load64(dict.dict, candidate); diff != 0 {
s += bits.TrailingZeros64(diff) >> 3
break
}
s += 8
candidate += 8
}
d += emitRepeat(dst[d:], repeat, s-base)
if debug {
fmt.Println("emitted dict repeat length", s-base, "offset:", repeat, "s:", s)
}
nextEmit = s
if s >= sLimit {
break searchDict
}
cv = load64(src, s)
continue
}
} else if uint32(cv>>(checkRep*8)) == load32(src, s-repeat+checkRep) {
base := s + checkRep
// Extend back
for i := base - repeat; base > nextEmit && i > 0 && src[i-1] == src[base-1]; {
i--
base--
}
d += emitLiteral(dst[d:], src[nextEmit:base])
if debug && nextEmit != base {
fmt.Println("emitted ", base-nextEmit, "literals")
}
// Extend forward
candidate := s - repeat + 4 + checkRep
s += 4 + checkRep
for s <= sLimit {
if diff := load64(src, s) ^ load64(src, candidate); diff != 0 {
s += bits.TrailingZeros64(diff) >> 3
break
}
s += 8
candidate += 8
}
if debug {
// Validate match.
if s <= candidate {
panic("s <= candidate")
}
a := src[base:s]
b := src[base-repeat : base-repeat+(s-base)]
if !bytes.Equal(a, b) {
panic("mismatch")
}
}
if nextEmit > 0 {
// same as `add := emitCopy(dst[d:], repeat, s-base)` but skips storing offset.
d += emitRepeat(dst[d:], repeat, s-base)
} else {
// First match, cannot be repeat.
d += emitCopy(dst[d:], repeat, s-base)
}
nextEmit = s
if s >= sLimit {
break searchDict
}
if debug {
fmt.Println("emitted reg repeat", s-base, "s:", s)
}
cv = load64(src, s)
continue searchDict
}
if s == 0 {
cv = load64(src, nextS)
s = nextS
continue searchDict
}
// Start with table. These matches will always be closer.
if uint32(cv) == load32(src, candidate) {
goto emitMatch
}
candidate = int(table[hash2])
if uint32(cv>>8) == load32(src, candidate2) {
table[hash2] = uint32(s + 2)
candidate = candidate2
s++
goto emitMatch
}
// Check dict. Dicts have longer offsets, so we want longer matches.
if cv == load64(dict.dict, candidateDict) {
table[hash2] = uint32(s + 2)
goto emitDict
}
candidateDict = int(dict.fastTable[hash2])
// Check if upper 7 bytes match
if candidateDict2 >= 1 {
if cv^load64(dict.dict, candidateDict2-1) < (1 << 8) {
table[hash2] = uint32(s + 2)
candidateDict = candidateDict2
s++
goto emitDict
}
}
table[hash2] = uint32(s + 2)
if uint32(cv>>16) == load32(src, candidate) {
s += 2
goto emitMatch
}
if candidateDict >= 2 {
// Check if upper 6 bytes match
if cv^load64(dict.dict, candidateDict-2) < (1 << 16) {
s += 2
goto emitDict
}
}
cv = load64(src, nextS)
s = nextS
continue searchDict
emitDict:
{
if debug {
if load32(dict.dict, candidateDict) != load32(src, s) {
panic("dict emit mismatch")
}
}
// Extend backwards.
// The top bytes will be rechecked to get the full match.
for candidateDict > 0 && s > nextEmit && dict.dict[candidateDict-1] == src[s-1] {
candidateDict--
s--
}
// Bail if we exceed the maximum size.
if d+(s-nextEmit) > dstLimit {
return 0
}
// 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.
d += emitLiteral(dst[d:], src[nextEmit:s])
if debug && nextEmit != s {
fmt.Println("emitted ", s-nextEmit, "literals")
}
{
// Invariant: we have a 4-byte match at s, and no need to emit any
// literal bytes prior to s.
base := s
repeat = s + (len(dict.dict)) - candidateDict
// Extend the 4-byte match as long as possible.
s += 4
candidateDict += 4
for s <= len(src)-8 && len(dict.dict)-candidateDict >= 8 {
if diff := load64(src, s) ^ load64(dict.dict, candidateDict); diff != 0 {
s += bits.TrailingZeros64(diff) >> 3
break
}
s += 8
candidateDict += 8
}
// Matches longer than 64 are split.
if s <= sLimit || s-base < 8 {
d += emitCopy(dst[d:], repeat, s-base)
} else {
// Split to ensure we don't start a copy within next block
d += emitCopy(dst[d:], repeat, 4)
d += emitRepeat(dst[d:], repeat, s-base-4)
}
if false {
// Validate match.
if s <= candidate {
panic("s <= candidate")
}
a := src[base:s]
b := dict.dict[base-repeat : base-repeat+(s-base)]
if !bytes.Equal(a, b) {
panic("mismatch")
}
}
if debug {
fmt.Println("emitted dict copy, length", s-base, "offset:", repeat, "s:", s)
}
nextEmit = s
if s >= sLimit {
break searchDict
}
if d > dstLimit {
// Do we have space for more, if not bail.
return 0
}
// Index and continue loop to try new candidate.
x := load64(src, s-2)
m2Hash := hash6(x, tableBits)
currHash := hash6(x>>8, tableBits)
candidate = int(table[currHash])
table[m2Hash] = uint32(s - 2)
table[currHash] = uint32(s - 1)
cv = load64(src, s)
}
continue
}
emitMatch:
// Extend backwards.
// The top bytes will be rechecked to get the full match.
for candidate > 0 && s > nextEmit && src[candidate-1] == src[s-1] {
candidate--
s--
}
// Bail if we exceed the maximum size.
if d+(s-nextEmit) > dstLimit {
return 0
}
// 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.
d += emitLiteral(dst[d:], src[nextEmit:s])
if debug && nextEmit != s {
fmt.Println("emitted ", s-nextEmit, "literals")
}
// Call emitCopy, and then see if another emitCopy could be our next
// move. Repeat until we find no match for the input immediately after
// what was consumed by the last emitCopy call.
//
// If we exit this loop normally then we need to call emitLiteral next,
// though we don't yet know how big the literal will be. We handle that
// by proceeding to the next iteration of the main loop. We also can
// exit this loop via goto if we get close to exhausting the input.
for {
// Invariant: we have a 4-byte match at s, and no need to emit any
// literal bytes prior to s.
base := s
repeat = base - candidate
// Extend the 4-byte match as long as possible.
s += 4
candidate += 4
for s <= len(src)-8 {
if diff := load64(src, s) ^ load64(src, candidate); diff != 0 {
s += bits.TrailingZeros64(diff) >> 3
break
}
s += 8
candidate += 8
}
d += emitCopy(dst[d:], repeat, s-base)
if debug {
// Validate match.
if s <= candidate {
panic("s <= candidate")
}
a := src[base:s]
b := src[base-repeat : base-repeat+(s-base)]
if !bytes.Equal(a, b) {
panic("mismatch")
}
}
if debug {
fmt.Println("emitted src copy, length", s-base, "offset:", repeat, "s:", s)
}
nextEmit = s
if s >= sLimit {
break searchDict
}
if d > dstLimit {
// Do we have space for more, if not bail.
return 0
}
// Check for an immediate match, otherwise start search at s+1
x := load64(src, s-2)
m2Hash := hash6(x, tableBits)
currHash := hash6(x>>16, tableBits)
candidate = int(table[currHash])
table[m2Hash] = uint32(s - 2)
table[currHash] = uint32(s)
if debug && s == candidate {
panic("s == candidate")
}
if uint32(x>>16) != load32(src, candidate) {
cv = load64(src, s+1)
s++
break
}
}
}
// Search without dict:
if repeat > s {
repeat = 0
}
// No more dict
sLimit = len(src) - inputMargin
if s >= sLimit {
goto emitRemainder
}
if debug {
fmt.Println("non-dict matching at", s, "repeat:", repeat)
}
cv = load64(src, s)
if debug {
fmt.Println("now", s, "->", sLimit, "out:", d, "left:", len(src)-s, "nextemit:", nextEmit, "dstLimit:", dstLimit, "s:", s)
}
for {
candidate := 0
for {
// Next src position to check
nextS := s + (s-nextEmit)>>6 + 4
if nextS > sLimit {
goto emitRemainder
}
hash0 := hash6(cv, tableBits)
hash1 := hash6(cv>>8, tableBits)
candidate = int(table[hash0])
candidate2 := int(table[hash1])
table[hash0] = uint32(s)
table[hash1] = uint32(s + 1)
hash2 := hash6(cv>>16, tableBits)
// Check repeat at offset checkRep.
const checkRep = 1
if repeat > 0 && uint32(cv>>(checkRep*8)) == load32(src, s-repeat+checkRep) {
base := s + checkRep
// Extend back
for i := base - repeat; base > nextEmit && i > 0 && src[i-1] == src[base-1]; {
i--
base--
}
d += emitLiteral(dst[d:], src[nextEmit:base])
if debug && nextEmit != base {
fmt.Println("emitted ", base-nextEmit, "literals")
}
// Extend forward
candidate := s - repeat + 4 + checkRep
s += 4 + checkRep
for s <= sLimit {
if diff := load64(src, s) ^ load64(src, candidate); diff != 0 {
s += bits.TrailingZeros64(diff) >> 3
break
}
s += 8
candidate += 8
}
if debug {
// Validate match.
if s <= candidate {
panic("s <= candidate")
}
a := src[base:s]
b := src[base-repeat : base-repeat+(s-base)]
if !bytes.Equal(a, b) {
panic("mismatch")
}
}
if nextEmit > 0 {
// same as `add := emitCopy(dst[d:], repeat, s-base)` but skips storing offset.
d += emitRepeat(dst[d:], repeat, s-base)
} else {
// First match, cannot be repeat.
d += emitCopy(dst[d:], repeat, s-base)
}
if debug {
fmt.Println("emitted src repeat length", s-base, "offset:", repeat, "s:", s)
}
nextEmit = s
if s >= sLimit {
goto emitRemainder
}
cv = load64(src, s)
continue
}
if uint32(cv) == load32(src, candidate) {
break
}
candidate = int(table[hash2])
if uint32(cv>>8) == load32(src, candidate2) {
table[hash2] = uint32(s + 2)
candidate = candidate2
s++
break
}
table[hash2] = uint32(s + 2)
if uint32(cv>>16) == load32(src, candidate) {
s += 2
break
}
cv = load64(src, nextS)
s = nextS
}
// Extend backwards.
// The top bytes will be rechecked to get the full match.
for candidate > 0 && s > nextEmit && src[candidate-1] == src[s-1] {
candidate--
s--
}
// Bail if we exceed the maximum size.
if d+(s-nextEmit) > dstLimit {
return 0
}
// 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.
d += emitLiteral(dst[d:], src[nextEmit:s])
if debug && nextEmit != s {
fmt.Println("emitted ", s-nextEmit, "literals")
}
// Call emitCopy, and then see if another emitCopy could be our next
// move. Repeat until we find no match for the input immediately after
// what was consumed by the last emitCopy call.
//
// If we exit this loop normally then we need to call emitLiteral next,
// though we don't yet know how big the literal will be. We handle that
// by proceeding to the next iteration of the main loop. We also can
// exit this loop via goto if we get close to exhausting the input.
for {
// Invariant: we have a 4-byte match at s, and no need to emit any
// literal bytes prior to s.
base := s
repeat = base - candidate
// Extend the 4-byte match as long as possible.
s += 4
candidate += 4
for s <= len(src)-8 {
if diff := load64(src, s) ^ load64(src, candidate); diff != 0 {
s += bits.TrailingZeros64(diff) >> 3
break
}
s += 8
candidate += 8
}
d += emitCopy(dst[d:], repeat, s-base)
if debug {
// Validate match.
if s <= candidate {
panic("s <= candidate")
}
a := src[base:s]
b := src[base-repeat : base-repeat+(s-base)]
if !bytes.Equal(a, b) {
panic("mismatch")
}
}
if debug {
fmt.Println("emitted src copy, length", s-base, "offset:", repeat, "s:", s)
}
nextEmit = s
if s >= sLimit {
goto emitRemainder
}
if d > dstLimit {
// Do we have space for more, if not bail.
return 0
}
// Check for an immediate match, otherwise start search at s+1
x := load64(src, s-2)
m2Hash := hash6(x, tableBits)
currHash := hash6(x>>16, tableBits)
candidate = int(table[currHash])
table[m2Hash] = uint32(s - 2)
table[currHash] = uint32(s)
if debug && s == candidate {
panic("s == candidate")
}
if uint32(x>>16) != load32(src, candidate) {
cv = load64(src, s+1)
s++
break
}
}
}
emitRemainder:
if nextEmit < len(src) {
// Bail if we exceed the maximum size.
if d+len(src)-nextEmit > dstLimit {
return 0
}
d += emitLiteral(dst[d:], src[nextEmit:])
if debug && nextEmit != s {
fmt.Println("emitted ", len(src)-nextEmit, "literals")
}
}
return d
}
vendor/github.com/klauspost/compress/s2/encode_amd64.go
Generated Vendored
+2
View File
@@ -3,6 +3,8 @@
package s2
const hasAmd64Asm = true
// encodeBlock encodes a non-empty src to a guaranteed-large-enough dst. It
// assumes that the varint-encoded length of the decompressed bytes has already
// been written.
vendor/github.com/klauspost/compress/s2/encode_best.go
Generated Vendored
+166 -24
View File
@@ -7,6 +7,7 @@ package s2
import (
"fmt"
"math"
"math/bits"
)
@@ -18,7 +19,7 @@ import (
//
// len(dst) >= MaxEncodedLen(len(src)) &&
// minNonLiteralBlockSize <= len(src) && len(src) <= maxBlockSize
func encodeBlockBest(dst, src []byte) (d int) {
func encodeBlockBest(dst, src []byte, dict *Dict) (d int) {
// Initialize the hash tables.
const (
// Long hash matches.
@@ -30,6 +31,8 @@ func encodeBlockBest(dst, src []byte) (d int) {
maxSTableSize = 1 << sTableBits
inputMargin = 8 + 2
debug = false
)
// sLimit is when to stop looking for offset/length copies. The inputMargin
@@ -39,6 +42,10 @@ func encodeBlockBest(dst, src []byte) (d int) {
if len(src) < minNonLiteralBlockSize {
return 0
}
sLimitDict := len(src) - inputMargin
if sLimitDict > MaxDictSrcOffset-inputMargin {
sLimitDict = MaxDictSrcOffset - inputMargin
}
var lTable [maxLTableSize]uint64
var sTable [maxSTableSize]uint64
@@ -52,10 +59,15 @@ func encodeBlockBest(dst, src []byte) (d int) {
// The encoded form must start with a literal, as there are no previous
// bytes to copy, so we start looking for hash matches at s == 1.
s := 1
repeat := 1
if dict != nil {
dict.initBest()
s = 0
repeat = len(dict.dict) - dict.repeat
}
cv := load64(src, s)
// We search for a repeat at -1, but don't output repeats when nextEmit == 0
repeat := 1
const lowbitMask = 0xffffffff
getCur := func(x uint64) int {
return int(x & lowbitMask)
@@ -67,11 +79,11 @@ func encodeBlockBest(dst, src []byte) (d int) {
for {
type match struct {
offset int
s int
length int
score int
rep bool
offset int
s int
length int
score int
rep, dict bool
}
var best match
for {
@@ -85,6 +97,12 @@ func encodeBlockBest(dst, src []byte) (d int) {
if nextS > sLimit {
goto emitRemainder
}
if dict != nil && s >= MaxDictSrcOffset {
dict = nil
if repeat > s {
repeat = math.MinInt32
}
}
hashL := hash8(cv, lTableBits)
hashS := hash4(cv, sTableBits)
candidateL := lTable[hashL]
@@ -114,7 +132,15 @@ func encodeBlockBest(dst, src []byte) (d int) {
}
m := match{offset: offset, s: s, length: 4 + offset, rep: rep}
s += 4
for s <= sLimit {
for s < len(src) {
if len(src)-s < 8 {
if src[s] == src[m.length] {
m.length++
s++
continue
}
break
}
if diff := load64(src, s) ^ load64(src, m.length); diff != 0 {
m.length += bits.TrailingZeros64(diff) >> 3
break
@@ -130,6 +156,62 @@ func encodeBlockBest(dst, src []byte) (d int) {
}
return m
}
matchDict := func(candidate, s int, first uint32, rep bool) match {
// 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 {
// Don't retest if we have the same offset.
return match{offset: offset, s: s}
}
if load32(dict.dict, candidate) != first {
return match{offset: offset, s: s}
}
m := match{offset: offset, s: s, length: 4 + candidate, rep: rep, dict: true}
s += 4
if !rep {
for s < sLimitDict && m.length < len(dict.dict) {
if len(src)-s < 8 || len(dict.dict)-m.length < 8 {
if src[s] == dict.dict[m.length] {
m.length++
s++
continue
}
break
}
if diff := load64(src, s) ^ load64(dict.dict, m.length); diff != 0 {
m.length += bits.TrailingZeros64(diff) >> 3
break
}
s += 8
m.length += 8
}
} else {
for s < len(src) && m.length < len(dict.dict) {
if len(src)-s < 8 || len(dict.dict)-m.length < 8 {
if src[s] == dict.dict[m.length] {
m.length++
s++
continue
}
break
}
if diff := load64(src, s) ^ load64(dict.dict, m.length); diff != 0 {
m.length += bits.TrailingZeros64(diff) >> 3
break
}
s += 8
m.length += 8
}
}
m.length -= candidate
m.score = score(m)
if m.score <= -m.s {
// Eliminate if no savings, we might find a better one.
m.length = 0
}
return m
}
bestOf := func(a, b match) match {
if b.length == 0 {
@@ -146,35 +228,82 @@ func encodeBlockBest(dst, src []byte) (d int) {
return b
}
best = bestOf(matchAt(getCur(candidateL), s, uint32(cv), false), matchAt(getPrev(candidateL), s, uint32(cv), false))
best = bestOf(best, matchAt(getCur(candidateS), s, uint32(cv), false))
best = bestOf(best, matchAt(getPrev(candidateS), s, uint32(cv), false))
if s > 0 {
best = bestOf(matchAt(getCur(candidateL), s, uint32(cv), false), matchAt(getPrev(candidateL), s, uint32(cv), false))
best = bestOf(best, matchAt(getCur(candidateS), s, uint32(cv), false))
best = bestOf(best, matchAt(getPrev(candidateS), s, uint32(cv), false))
}
if dict != nil {
candidateL := dict.bestTableLong[hashL]
candidateS := dict.bestTableShort[hashS]
best = bestOf(best, matchDict(int(candidateL&0xffff), s, uint32(cv), false))
best = bestOf(best, matchDict(int(candidateL>>16), s, uint32(cv), false))
best = bestOf(best, matchDict(int(candidateS&0xffff), s, uint32(cv), false))
best = bestOf(best, matchDict(int(candidateS>>16), s, uint32(cv), false))
}
{
best = bestOf(best, matchAt(s-repeat+1, s+1, uint32(cv>>8), true))
if (dict == nil || repeat <= s) && repeat > 0 {
best = bestOf(best, matchAt(s-repeat+1, s+1, uint32(cv>>8), true))
} else if s-repeat < -4 && dict != nil {
candidate := len(dict.dict) - (repeat - s)
best = bestOf(best, matchDict(candidate, s, uint32(cv), true))
candidate++
best = bestOf(best, matchDict(candidate, s+1, uint32(cv>>8), true))
}
if best.length > 0 {
hashS := hash4(cv>>8, sTableBits)
// s+1
nextShort := sTable[hash4(cv>>8, sTableBits)]
nextShort := sTable[hashS]
s := s + 1
cv := load64(src, s)
nextLong := lTable[hash8(cv, lTableBits)]
hashL := hash8(cv, lTableBits)
nextLong := lTable[hashL]
best = bestOf(best, matchAt(getCur(nextShort), s, uint32(cv), false))
best = bestOf(best, matchAt(getPrev(nextShort), s, uint32(cv), false))
best = bestOf(best, matchAt(getCur(nextLong), s, uint32(cv), false))
best = bestOf(best, matchAt(getPrev(nextLong), s, uint32(cv), false))
// Repeat at + 2
best = bestOf(best, matchAt(s-repeat+1, s+1, uint32(cv>>8), true))
// Dict at + 1
if dict != nil {
candidateL := dict.bestTableLong[hashL]
candidateS := dict.bestTableShort[hashS]
best = bestOf(best, matchDict(int(candidateL&0xffff), s, uint32(cv), false))
best = bestOf(best, matchDict(int(candidateS&0xffff), s, uint32(cv), false))
}
// s+2
if true {
nextShort = sTable[hash4(cv>>8, sTableBits)]
hashS := hash4(cv>>8, sTableBits)
nextShort = sTable[hashS]
s++
cv = load64(src, s)
nextLong = lTable[hash8(cv, lTableBits)]
hashL := hash8(cv, lTableBits)
nextLong = lTable[hashL]
if (dict == nil || repeat <= s) && repeat > 0 {
// Repeat at + 2
best = bestOf(best, matchAt(s-repeat, s, uint32(cv), true))
} else if repeat-s > 4 && dict != nil {
candidate := len(dict.dict) - (repeat - s)
best = bestOf(best, matchDict(candidate, s, uint32(cv), true))
}
best = bestOf(best, matchAt(getCur(nextShort), s, uint32(cv), false))
best = bestOf(best, matchAt(getPrev(nextShort), s, uint32(cv), false))
best = bestOf(best, matchAt(getCur(nextLong), s, uint32(cv), false))
best = bestOf(best, matchAt(getPrev(nextLong), s, uint32(cv), false))
// Dict at +2
// Very small gain
if dict != nil {
candidateL := dict.bestTableLong[hashL]
candidateS := dict.bestTableShort[hashS]
best = bestOf(best, matchDict(int(candidateL&0xffff), s, uint32(cv), false))
best = bestOf(best, matchDict(int(candidateS&0xffff), s, uint32(cv), false))
}
}
// Search for a match at best match end, see if that is better.
// Allow some bytes at the beginning to mismatch.
@@ -227,7 +356,7 @@ func encodeBlockBest(dst, src []byte) (d int) {
// Extend backwards, not needed for repeats...
s = best.s
if !best.rep {
if !best.rep && !best.dict {
for best.offset > 0 && s > nextEmit && src[best.offset-1] == src[s-1] {
best.offset--
best.length++
@@ -244,7 +373,6 @@ func encodeBlockBest(dst, src []byte) (d int) {
base := s
offset := s - best.offset
s += best.length
if offset > 65535 && s-base <= 5 && !best.rep {
@@ -256,16 +384,28 @@ func encodeBlockBest(dst, src []byte) (d int) {
cv = load64(src, s)
continue
}
if debug && nextEmit != base {
fmt.Println("EMIT", base-nextEmit, "literals. base-after:", base)
}
d += emitLiteral(dst[d:], src[nextEmit:base])
if best.rep {
if nextEmit > 0 {
if nextEmit > 0 || best.dict {
if debug {
fmt.Println("REPEAT, length", best.length, "offset:", offset, "s-after:", s, "dict:", best.dict, "best:", best)
}
// same as `add := emitCopy(dst[d:], repeat, s-base)` but skips storing offset.
d += emitRepeat(dst[d:], offset, best.length)
} else {
// First match, cannot be repeat.
// First match without dict cannot be a repeat.
if debug {
fmt.Println("COPY, length", best.length, "offset:", offset, "s-after:", s, "dict:", best.dict, "best:", best)
}
d += emitCopy(dst[d:], offset, best.length)
}
} else {
if debug {
fmt.Println("COPY, length", best.length, "offset:", offset, "s-after:", s, "dict:", best.dict, "best:", best)
}
d += emitCopy(dst[d:], offset, best.length)
}
repeat = offset
@@ -296,6 +436,9 @@ emitRemainder:
if d+len(src)-nextEmit > dstLimit {
return 0
}
if debug && nextEmit != s {
fmt.Println("emitted ", len(src)-nextEmit, "literals")
}
d += emitLiteral(dst[d:], src[nextEmit:])
}
return d
@@ -642,7 +785,6 @@ func emitRepeatSize(offset, length int) int {
left := 0
if length > maxRepeat {
left = length - maxRepeat + 4
length = maxRepeat - 4
}
if left > 0 {
return 5 + emitRepeatSize(offset, left)
vendor/github.com/klauspost/compress/s2/encode_better.go
Generated Vendored
+622
View File
@@ -6,6 +6,8 @@
package s2
import (
"bytes"
"fmt"
"math/bits"
)
@@ -476,3 +478,623 @@ emitRemainder:
}
return d
}
// encodeBlockBetterDict encodes a non-empty src to a guaranteed-large-enough dst. It
// assumes that the varint-encoded length of the decompressed bytes has already
// been written.
//
// It also assumes that:
//
// len(dst) >= MaxEncodedLen(len(src)) &&
// minNonLiteralBlockSize <= len(src) && len(src) <= maxBlockSize
func encodeBlockBetterDict(dst, src []byte, dict *Dict) (d int) {
// 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.
// Initialize the hash tables.
const (
// Long hash matches.
lTableBits = 17
maxLTableSize = 1 << lTableBits
// Short hash matches.
sTableBits = 14
maxSTableSize = 1 << sTableBits
maxAhead = 8 // maximum bytes ahead without checking sLimit
debug = false
)
sLimit := len(src) - inputMargin
if sLimit > MaxDictSrcOffset-maxAhead {
sLimit = MaxDictSrcOffset - maxAhead
}
if len(src) < minNonLiteralBlockSize {
return 0
}
dict.initBetter()
var lTable [maxLTableSize]uint32
var sTable [maxSTableSize]uint32
// Bail if we can't compress to at least this.
dstLimit := len(src) - len(src)>>5 - 6
// nextEmit is where in src the next emitLiteral should start from.
nextEmit := 0
// The encoded form must start with a literal, as there are no previous
// bytes to copy, so we start looking for hash matches at s == 1.
s := 0
cv := load64(src, s)
// We initialize repeat to 0, so we never match on first attempt
repeat := len(dict.dict) - dict.repeat
// While in dict
searchDict:
for {
candidateL := 0
nextS := 0
for {
// Next src position to check
nextS = s + (s-nextEmit)>>7 + 1
if nextS > sLimit {
break searchDict
}
hashL := hash7(cv, lTableBits)
hashS := hash4(cv, sTableBits)
candidateL = int(lTable[hashL])
candidateS := int(sTable[hashS])
dictL := int(dict.betterTableLong[hashL])
dictS := int(dict.betterTableShort[hashS])
lTable[hashL] = uint32(s)
sTable[hashS] = uint32(s)
valLong := load64(src, candidateL)
valShort := load64(src, candidateS)
// If long matches at least 8 bytes, use that.
if s != 0 {
if cv == valLong {
goto emitMatch
}
if cv == valShort {
candidateL = candidateS
goto emitMatch
}
}
// Check dict repeat.
if repeat >= s+4 {
candidate := len(dict.dict) - repeat + s
if candidate > 0 && uint32(cv) == load32(dict.dict, candidate) {
// Extend back
base := s
for i := candidate; base > nextEmit && i > 0 && dict.dict[i-1] == src[base-1]; {
i--
base--
}
d += emitLiteral(dst[d:], src[nextEmit:base])
if debug && nextEmit != base {
fmt.Println("emitted ", base-nextEmit, "literals")
}
s += 4
candidate += 4
for candidate < len(dict.dict)-8 && s <= len(src)-8 {
if diff := load64(src, s) ^ load64(dict.dict, candidate); diff != 0 {
s += bits.TrailingZeros64(diff) >> 3
break
}
s += 8
candidate += 8
}
d += emitRepeat(dst[d:], repeat, s-base)
if debug {
fmt.Println("emitted dict repeat length", s-base, "offset:", repeat, "s:", s)
}
nextEmit = s
if s >= sLimit {
break searchDict
}
cv = load64(src, s)
// Index in-between
index0 := base + 1
index1 := s - 2
cv = load64(src, s)
for index0 < index1 {
cv0 := load64(src, index0)
cv1 := load64(src, index1)
lTable[hash7(cv0, lTableBits)] = uint32(index0)
sTable[hash4(cv0>>8, sTableBits)] = uint32(index0 + 1)
lTable[hash7(cv1, lTableBits)] = uint32(index1)
sTable[hash4(cv1>>8, sTableBits)] = uint32(index1 + 1)
index0 += 2
index1 -= 2
}
continue
}
}
// Don't try to find match at s==0
if s == 0 {
cv = load64(src, nextS)
s = nextS
continue
}
// Long likely matches 7, so take that.
if uint32(cv) == uint32(valLong) {
goto emitMatch
}
// Long dict...
if uint32(cv) == load32(dict.dict, dictL) {
candidateL = dictL
goto emitDict
}
// Check our short candidate
if uint32(cv) == uint32(valShort) {
// Try a long candidate at s+1
hashL = hash7(cv>>8, lTableBits)
candidateL = int(lTable[hashL])
lTable[hashL] = uint32(s + 1)
if uint32(cv>>8) == load32(src, candidateL) {
s++
goto emitMatch
}
// Use our short candidate.
candidateL = candidateS
goto emitMatch
}
if uint32(cv) == load32(dict.dict, dictS) {
// Try a long candidate at s+1
hashL = hash7(cv>>8, lTableBits)
candidateL = int(lTable[hashL])
lTable[hashL] = uint32(s + 1)
if uint32(cv>>8) == load32(src, candidateL) {
s++
goto emitMatch
}
candidateL = dictS
goto emitDict
}
cv = load64(src, nextS)
s = nextS
}
emitDict:
{
if debug {
if load32(dict.dict, candidateL) != load32(src, s) {
panic("dict emit mismatch")
}
}
// Extend backwards.
// The top bytes will be rechecked to get the full match.
for candidateL > 0 && s > nextEmit && dict.dict[candidateL-1] == src[s-1] {
candidateL--
s--
}
// Bail if we exceed the maximum size.
if d+(s-nextEmit) > dstLimit {
return 0
}
// 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.
d += emitLiteral(dst[d:], src[nextEmit:s])
if debug && nextEmit != s {
fmt.Println("emitted ", s-nextEmit, "literals")
}
{
// Invariant: we have a 4-byte match at s, and no need to emit any
// literal bytes prior to s.
base := s
offset := s + (len(dict.dict)) - candidateL
// Extend the 4-byte match as long as possible.
s += 4
candidateL += 4
for s <= len(src)-8 && len(dict.dict)-candidateL >= 8 {
if diff := load64(src, s) ^ load64(dict.dict, candidateL); diff != 0 {
s += bits.TrailingZeros64(diff) >> 3
break
}
s += 8
candidateL += 8
}
if repeat == offset {
if debug {
fmt.Println("emitted dict repeat, length", s-base, "offset:", offset, "s:", s, "dict offset:", candidateL)
}
d += emitRepeat(dst[d:], offset, s-base)
} else {
if debug {
fmt.Println("emitted dict copy, length", s-base, "offset:", offset, "s:", s, "dict offset:", candidateL)
}
// Matches longer than 64 are split.
if s <= sLimit || s-base < 8 {
d += emitCopy(dst[d:], offset, s-base)
} else {
// Split to ensure we don't start a copy within next block.
d += emitCopy(dst[d:], offset, 4)
d += emitRepeat(dst[d:], offset, s-base-4)
}
repeat = offset
}
if false {
// Validate match.
if s <= candidateL {
panic("s <= candidate")
}
a := src[base:s]
b := dict.dict[base-repeat : base-repeat+(s-base)]
if !bytes.Equal(a, b) {
panic("mismatch")
}
}
nextEmit = s
if s >= sLimit {
break searchDict
}
if d > dstLimit {
// Do we have space for more, if not bail.
return 0
}
// Index short & long
index0 := base + 1
index1 := s - 2
cv0 := load64(src, index0)
cv1 := load64(src, index1)
lTable[hash7(cv0, lTableBits)] = uint32(index0)
sTable[hash4(cv0>>8, sTableBits)] = uint32(index0 + 1)
lTable[hash7(cv1, lTableBits)] = uint32(index1)
sTable[hash4(cv1>>8, sTableBits)] = uint32(index1 + 1)
index0 += 1
index1 -= 1
cv = load64(src, s)
// index every second long in between.
for index0 < index1 {
lTable[hash7(load64(src, index0), lTableBits)] = uint32(index0)
lTable[hash7(load64(src, index1), lTableBits)] = uint32(index1)
index0 += 2
index1 -= 2
}
}
continue
}
emitMatch:
// Extend backwards
for candidateL > 0 && s > nextEmit && src[candidateL-1] == src[s-1] {
candidateL--
s--
}
// Bail if we exceed the maximum size.
if d+(s-nextEmit) > dstLimit {
return 0
}
base := s
offset := base - candidateL
// Extend the 4-byte match as long as possible.
s += 4
candidateL += 4
for s < len(src) {
if len(src)-s < 8 {
if src[s] == src[candidateL] {
s++
candidateL++
continue
}
break
}
if diff := load64(src, s) ^ load64(src, candidateL); diff != 0 {
s += bits.TrailingZeros64(diff) >> 3
break
}
s += 8
candidateL += 8
}
if offset > 65535 && s-base <= 5 && repeat != offset {
// Bail if the match is equal or worse to the encoding.
s = nextS + 1
if s >= sLimit {
goto emitRemainder
}
cv = load64(src, s)
continue
}
d += emitLiteral(dst[d:], src[nextEmit:base])
if debug && nextEmit != s {
fmt.Println("emitted ", s-nextEmit, "literals")
}
if repeat == offset {
if debug {
fmt.Println("emitted match repeat, length", s-base, "offset:", offset, "s:", s)
}
d += emitRepeat(dst[d:], offset, s-base)
} else {
if debug {
fmt.Println("emitted match copy, length", s-base, "offset:", offset, "s:", s)
}
d += emitCopy(dst[d:], offset, s-base)
repeat = offset
}
nextEmit = s
if s >= sLimit {
goto emitRemainder
}
if d > dstLimit {
// Do we have space for more, if not bail.
return 0
}
// Index short & long
index0 := base + 1
index1 := s - 2
cv0 := load64(src, index0)
cv1 := load64(src, index1)
lTable[hash7(cv0, lTableBits)] = uint32(index0)
sTable[hash4(cv0>>8, sTableBits)] = uint32(index0 + 1)
lTable[hash7(cv1, lTableBits)] = uint32(index1)
sTable[hash4(cv1>>8, sTableBits)] = uint32(index1 + 1)
index0 += 1
index1 -= 1
cv = load64(src, s)
// index every second long in between.
for index0 < index1 {
lTable[hash7(load64(src, index0), lTableBits)] = uint32(index0)
lTable[hash7(load64(src, index1), lTableBits)] = uint32(index1)
index0 += 2
index1 -= 2
}
}
// Search without dict:
if repeat > s {
repeat = 0
}
// No more dict
sLimit = len(src) - inputMargin
if s >= sLimit {
goto emitRemainder
}
cv = load64(src, s)
if debug {
fmt.Println("now", s, "->", sLimit, "out:", d, "left:", len(src)-s, "nextemit:", nextEmit, "dstLimit:", dstLimit, "s:", s)
}
for {
candidateL := 0
nextS := 0
for {
// Next src position to check
nextS = s + (s-nextEmit)>>7 + 1
if nextS > sLimit {
goto emitRemainder
}
hashL := hash7(cv, lTableBits)
hashS := hash4(cv, sTableBits)
candidateL = int(lTable[hashL])
candidateS := int(sTable[hashS])
lTable[hashL] = uint32(s)
sTable[hashS] = uint32(s)
valLong := load64(src, candidateL)
valShort := load64(src, candidateS)
// If long matches at least 8 bytes, use that.
if cv == valLong {
break
}
if cv == valShort {
candidateL = candidateS
break
}
// Check repeat at offset checkRep.
const checkRep = 1
// Minimum length of a repeat. Tested with various values.
// While 4-5 offers improvements in some, 6 reduces
// regressions significantly.
const wantRepeatBytes = 6
const repeatMask = ((1 << (wantRepeatBytes * 8)) - 1) << (8 * checkRep)
if false && repeat > 0 && cv&repeatMask == load64(src, s-repeat)&repeatMask {
base := s + checkRep
// Extend back
for i := base - repeat; base > nextEmit && i > 0 && src[i-1] == src[base-1]; {
i--
base--
}
d += emitLiteral(dst[d:], src[nextEmit:base])
// Extend forward
candidate := s - repeat + wantRepeatBytes + checkRep
s += wantRepeatBytes + checkRep
for s < len(src) {
if len(src)-s < 8 {
if src[s] == src[candidate] {
s++
candidate++
continue
}
break
}
if diff := load64(src, s) ^ load64(src, candidate); diff != 0 {
s += bits.TrailingZeros64(diff) >> 3
break
}
s += 8
candidate += 8
}
// same as `add := emitCopy(dst[d:], repeat, s-base)` but skips storing offset.
d += emitRepeat(dst[d:], repeat, s-base)
nextEmit = s
if s >= sLimit {
goto emitRemainder
}
// Index in-between
index0 := base + 1
index1 := s - 2
cv = load64(src, s)
for index0 < index1 {
cv0 := load64(src, index0)
cv1 := load64(src, index1)
lTable[hash7(cv0, lTableBits)] = uint32(index0)
sTable[hash4(cv0>>8, sTableBits)] = uint32(index0 + 1)
lTable[hash7(cv1, lTableBits)] = uint32(index1)
sTable[hash4(cv1>>8, sTableBits)] = uint32(index1 + 1)
index0 += 2
index1 -= 2
}
cv = load64(src, s)
continue
}
// Long likely matches 7, so take that.
if uint32(cv) == uint32(valLong) {
break
}
// Check our short candidate
if uint32(cv) == uint32(valShort) {
// Try a long candidate at s+1
hashL = hash7(cv>>8, lTableBits)
candidateL = int(lTable[hashL])
lTable[hashL] = uint32(s + 1)
if uint32(cv>>8) == load32(src, candidateL) {
s++
break
}
// Use our short candidate.
candidateL = candidateS
break
}
cv = load64(src, nextS)
s = nextS
}
// Extend backwards
for candidateL > 0 && s > nextEmit && src[candidateL-1] == src[s-1] {
candidateL--
s--
}
// Bail if we exceed the maximum size.
if d+(s-nextEmit) > dstLimit {
return 0
}
base := s
offset := base - candidateL
// Extend the 4-byte match as long as possible.
s += 4
candidateL += 4
for s < len(src) {
if len(src)-s < 8 {
if src[s] == src[candidateL] {
s++
candidateL++
continue
}
break
}
if diff := load64(src, s) ^ load64(src, candidateL); diff != 0 {
s += bits.TrailingZeros64(diff) >> 3
break
}
s += 8
candidateL += 8
}
if offset > 65535 && s-base <= 5 && repeat != offset {
// Bail if the match is equal or worse to the encoding.
s = nextS + 1
if s >= sLimit {
goto emitRemainder
}
cv = load64(src, s)
continue
}
d += emitLiteral(dst[d:], src[nextEmit:base])
if repeat == offset {
d += emitRepeat(dst[d:], offset, s-base)
} else {
d += emitCopy(dst[d:], offset, s-base)
repeat = offset
}
nextEmit = s
if s >= sLimit {
goto emitRemainder
}
if d > dstLimit {
// Do we have space for more, if not bail.
return 0
}
// Index short & long
index0 := base + 1
index1 := s - 2
cv0 := load64(src, index0)
cv1 := load64(src, index1)
lTable[hash7(cv0, lTableBits)] = uint32(index0)
sTable[hash4(cv0>>8, sTableBits)] = uint32(index0 + 1)
lTable[hash7(cv1, lTableBits)] = uint32(index1)
sTable[hash4(cv1>>8, sTableBits)] = uint32(index1 + 1)
index0 += 1
index1 -= 1
cv = load64(src, s)
// index every second long in between.
for index0 < index1 {
lTable[hash7(load64(src, index0), lTableBits)] = uint32(index0)
lTable[hash7(load64(src, index1), lTableBits)] = uint32(index1)
index0 += 2
index1 -= 2
}
}
emitRemainder:
if nextEmit < len(src) {
// Bail if we exceed the maximum size.
if d+len(src)-nextEmit > dstLimit {
return 0
}
d += emitLiteral(dst[d:], src[nextEmit:])
}
return d
}
vendor/github.com/klauspost/compress/s2/encode_go.go
Generated Vendored
+413
View File
@@ -4,9 +4,12 @@
package s2
import (
"bytes"
"math/bits"
)
const hasAmd64Asm = false
// encodeBlock encodes a non-empty src to a guaranteed-large-enough dst. It
// assumes that the varint-encoded length of the decompressed bytes has already
// been written.
@@ -312,3 +315,413 @@ func matchLen(a []byte, b []byte) int {
}
return len(a) + checked
}
func calcBlockSize(src []byte) (d int) {
// Initialize the hash table.
const (
tableBits = 13
maxTableSize = 1 << tableBits
)
var table [maxTableSize]uint32
// 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 := len(src) - inputMargin
// Bail if we can't compress to at least this.
dstLimit := len(src) - len(src)>>5 - 5
// nextEmit is where in src the next emitLiteral should start from.
nextEmit := 0
// The encoded form must start with a literal, as there are no previous
// bytes to copy, so we start looking for hash matches at s == 1.
s := 1
cv := load64(src, s)
// We search for a repeat at -1, but don't output repeats when nextEmit == 0
repeat := 1
for {
candidate := 0
for {
// Next src position to check
nextS := s + (s-nextEmit)>>6 + 4
if nextS > sLimit {
goto emitRemainder
}
hash0 := hash6(cv, tableBits)
hash1 := hash6(cv>>8, tableBits)
candidate = int(table[hash0])
candidate2 := int(table[hash1])
table[hash0] = uint32(s)
table[hash1] = uint32(s + 1)
hash2 := hash6(cv>>16, tableBits)
// Check repeat at offset checkRep.
const checkRep = 1
if uint32(cv>>(checkRep*8)) == load32(src, s-repeat+checkRep) {
base := s + checkRep
// Extend back
for i := base - repeat; base > nextEmit && i > 0 && src[i-1] == src[base-1]; {
i--
base--
}
d += emitLiteralSize(src[nextEmit:base])
// Extend forward
candidate := s - repeat + 4 + checkRep
s += 4 + checkRep
for s <= sLimit {
if diff := load64(src, s) ^ load64(src, candidate); diff != 0 {
s += bits.TrailingZeros64(diff) >> 3
break
}
s += 8
candidate += 8
}
d += emitCopyNoRepeatSize(repeat, s-base)
nextEmit = s
if s >= sLimit {
goto emitRemainder
}
cv = load64(src, s)
continue
}
if uint32(cv) == load32(src, candidate) {
break
}
candidate = int(table[hash2])
if uint32(cv>>8) == load32(src, candidate2) {
table[hash2] = uint32(s + 2)
candidate = candidate2
s++
break
}
table[hash2] = uint32(s + 2)
if uint32(cv>>16) == load32(src, candidate) {
s += 2
break
}
cv = load64(src, nextS)
s = nextS
}
// Extend backwards
for candidate > 0 && s > nextEmit && src[candidate-1] == src[s-1] {
candidate--
s--
}
// Bail if we exceed the maximum size.
if d+(s-nextEmit) > dstLimit {
return 0
}
// 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.
d += emitLiteralSize(src[nextEmit:s])
// Call emitCopy, and then see if another emitCopy could be our next
// move. Repeat until we find no match for the input immediately after
// what was consumed by the last emitCopy call.
//
// If we exit this loop normally then we need to call emitLiteral next,
// though we don't yet know how big the literal will be. We handle that
// by proceeding to the next iteration of the main loop. We also can
// exit this loop via goto if we get close to exhausting the input.
for {
// Invariant: we have a 4-byte match at s, and no need to emit any
// literal bytes prior to s.
base := s
repeat = base - candidate
// Extend the 4-byte match as long as possible.
s += 4
candidate += 4
for s <= len(src)-8 {
if diff := load64(src, s) ^ load64(src, candidate); diff != 0 {
s += bits.TrailingZeros64(diff) >> 3
break
}
s += 8
candidate += 8
}
d += emitCopyNoRepeatSize(repeat, s-base)
if false {
// Validate match.
a := src[base:s]
b := src[base-repeat : base-repeat+(s-base)]
if !bytes.Equal(a, b) {
panic("mismatch")
}
}
nextEmit = s
if s >= sLimit {
goto emitRemainder
}
if d > dstLimit {
// Do we have space for more, if not bail.
return 0
}
// Check for an immediate match, otherwise start search at s+1
x := load64(src, s-2)
m2Hash := hash6(x, tableBits)
currHash := hash6(x>>16, tableBits)
candidate = int(table[currHash])
table[m2Hash] = uint32(s - 2)
table[currHash] = uint32(s)
if uint32(x>>16) != load32(src, candidate) {
cv = load64(src, s+1)
s++
break
}
}
}
emitRemainder:
if nextEmit < len(src) {
// Bail if we exceed the maximum size.
if d+len(src)-nextEmit > dstLimit {
return 0
}
d += emitLiteralSize(src[nextEmit:])
}
return d
}
func calcBlockSizeSmall(src []byte) (d int) {
// Initialize the hash table.
const (
tableBits = 9
maxTableSize = 1 << tableBits
)
var table [maxTableSize]uint32
// 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 := len(src) - inputMargin
// Bail if we can't compress to at least this.
dstLimit := len(src) - len(src)>>5 - 5
// nextEmit is where in src the next emitLiteral should start from.
nextEmit := 0
// The encoded form must start with a literal, as there are no previous
// bytes to copy, so we start looking for hash matches at s == 1.
s := 1
cv := load64(src, s)
// We search for a repeat at -1, but don't output repeats when nextEmit == 0
repeat := 1
for {
candidate := 0
for {
// Next src position to check
nextS := s + (s-nextEmit)>>6 + 4
if nextS > sLimit {
goto emitRemainder
}
hash0 := hash6(cv, tableBits)
hash1 := hash6(cv>>8, tableBits)
candidate = int(table[hash0])
candidate2 := int(table[hash1])
table[hash0] = uint32(s)
table[hash1] = uint32(s + 1)
hash2 := hash6(cv>>16, tableBits)
// Check repeat at offset checkRep.
const checkRep = 1
if uint32(cv>>(checkRep*8)) == load32(src, s-repeat+checkRep) {
base := s + checkRep
// Extend back
for i := base - repeat; base > nextEmit && i > 0 && src[i-1] == src[base-1]; {
i--
base--
}
d += emitLiteralSize(src[nextEmit:base])
// Extend forward
candidate := s - repeat + 4 + checkRep
s += 4 + checkRep
for s <= sLimit {
if diff := load64(src, s) ^ load64(src, candidate); diff != 0 {
s += bits.TrailingZeros64(diff) >> 3
break
}
s += 8
candidate += 8
}
d += emitCopyNoRepeatSize(repeat, s-base)
nextEmit = s
if s >= sLimit {
goto emitRemainder
}
cv = load64(src, s)
continue
}
if uint32(cv) == load32(src, candidate) {
break
}
candidate = int(table[hash2])
if uint32(cv>>8) == load32(src, candidate2) {
table[hash2] = uint32(s + 2)
candidate = candidate2
s++
break
}
table[hash2] = uint32(s + 2)
if uint32(cv>>16) == load32(src, candidate) {
s += 2
break
}
cv = load64(src, nextS)
s = nextS
}
// Extend backwards
for candidate > 0 && s > nextEmit && src[candidate-1] == src[s-1] {
candidate--
s--
}
// Bail if we exceed the maximum size.
if d+(s-nextEmit) > dstLimit {
return 0
}
// 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.
d += emitLiteralSize(src[nextEmit:s])
// Call emitCopy, and then see if another emitCopy could be our next
// move. Repeat until we find no match for the input immediately after
// what was consumed by the last emitCopy call.
//
// If we exit this loop normally then we need to call emitLiteral next,
// though we don't yet know how big the literal will be. We handle that
// by proceeding to the next iteration of the main loop. We also can
// exit this loop via goto if we get close to exhausting the input.
for {
// Invariant: we have a 4-byte match at s, and no need to emit any
// literal bytes prior to s.
base := s
repeat = base - candidate
// Extend the 4-byte match as long as possible.
s += 4
candidate += 4
for s <= len(src)-8 {
if diff := load64(src, s) ^ load64(src, candidate); diff != 0 {
s += bits.TrailingZeros64(diff) >> 3
break
}
s += 8
candidate += 8
}
d += emitCopyNoRepeatSize(repeat, s-base)
if false {
// Validate match.
a := src[base:s]
b := src[base-repeat : base-repeat+(s-base)]
if !bytes.Equal(a, b) {
panic("mismatch")
}
}
nextEmit = s
if s >= sLimit {
goto emitRemainder
}
if d > dstLimit {
// Do we have space for more, if not bail.
return 0
}
// Check for an immediate match, otherwise start search at s+1
x := load64(src, s-2)
m2Hash := hash6(x, tableBits)
currHash := hash6(x>>16, tableBits)
candidate = int(table[currHash])
table[m2Hash] = uint32(s - 2)
table[currHash] = uint32(s)
if uint32(x>>16) != load32(src, candidate) {
cv = load64(src, s+1)
s++
break
}
}
}
emitRemainder:
if nextEmit < len(src) {
// Bail if we exceed the maximum size.
if d+len(src)-nextEmit > dstLimit {
return 0
}
d += emitLiteralSize(src[nextEmit:])
}
return d
}
// emitLiteral writes a literal chunk and returns the number of bytes written.
//
// It assumes that:
//
// dst is long enough to hold the encoded bytes
// 0 <= len(lit) && len(lit) <= math.MaxUint32
func emitLiteralSize(lit []byte) int {
if len(lit) == 0 {
return 0
}
switch {
case len(lit) <= 60:
return len(lit) + 1
case len(lit) <= 1<<8:
return len(lit) + 2
case len(lit) <= 1<<16:
return len(lit) + 3
case len(lit) <= 1<<24:
return len(lit) + 4
default:
return len(lit) + 5
}
}
func cvtLZ4BlockAsm(dst []byte, src []byte) (uncompressed int, dstUsed int) {
panic("cvtLZ4BlockAsm should be unreachable")
}
func cvtLZ4BlockSnappyAsm(dst []byte, src []byte) (uncompressed int, dstUsed int) {
panic("cvtLZ4BlockSnappyAsm should be unreachable")
}
func cvtLZ4sBlockAsm(dst []byte, src []byte) (uncompressed int, dstUsed int) {
panic("cvtLZ4sBlockAsm should be unreachable")
}
func cvtLZ4sBlockSnappyAsm(dst []byte, src []byte) (uncompressed int, dstUsed int) {
panic("cvtLZ4sBlockSnappyAsm should be unreachable")
}
vendor/github.com/klauspost/compress/s2/encodeblock_amd64.go
Generated Vendored
+34
View File
@@ -146,6 +146,20 @@ func encodeSnappyBetterBlockAsm10B(dst []byte, src []byte) int
//go:noescape
func encodeSnappyBetterBlockAsm8B(dst []byte, src []byte) int
// calcBlockSize encodes a non-empty src to a guaranteed-large-enough dst.
// Maximum input 4294967295 bytes.
// It assumes that the varint-encoded length of the decompressed bytes has already been written.
//
//go:noescape
func calcBlockSize(src []byte) int
// calcBlockSizeSmall encodes a non-empty src to a guaranteed-large-enough dst.
// Maximum input 1024 bytes.
// It assumes that the varint-encoded length of the decompressed bytes has already been written.
//
//go:noescape
func calcBlockSizeSmall(src []byte) int
// emitLiteral writes a literal chunk and returns the number of bytes written.
//
// It assumes that:
@@ -192,3 +206,23 @@ func emitCopyNoRepeat(dst []byte, offset int, length int) int
//
//go:noescape
func matchLen(a []byte, b []byte) int
// cvtLZ4Block converts an LZ4 block to S2
//
//go:noescape
func cvtLZ4BlockAsm(dst []byte, src []byte) (uncompressed int, dstUsed int)
// cvtLZ4sBlock converts an LZ4s block to S2
//
//go:noescape
func cvtLZ4sBlockAsm(dst []byte, src []byte) (uncompressed int, dstUsed int)
// cvtLZ4Block converts an LZ4 block to Snappy
//
//go:noescape
func cvtLZ4BlockSnappyAsm(dst []byte, src []byte) (uncompressed int, dstUsed int)
// cvtLZ4sBlock converts an LZ4s block to Snappy
//
//go:noescape
func cvtLZ4sBlockSnappyAsm(dst []byte, src []byte) (uncompressed int, dstUsed int)
vendor/github.com/klauspost/compress/s2/encodeblock_amd64.s
Generated Vendored
+11474 -8908
View File
File diff suppressed because it is too large Load Diff
vendor/github.com/klauspost/compress/s2/lz4convert.go
Generated Vendored
+585
View File
@@ -0,0 +1,585 @@
// Copyright (c) 2022 Klaus Post. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package s2
import (
"encoding/binary"
"errors"
"fmt"
)
// LZ4Converter provides conversion from LZ4 blocks as defined here:
// https://github.com/lz4/lz4/blob/dev/doc/lz4_Block_format.md
type LZ4Converter struct {
}
// ErrDstTooSmall is returned when provided destination is too small.
var ErrDstTooSmall = errors.New("s2: destination too small")
// ConvertBlock will convert an LZ4 block and append it as an S2
// block without block length to dst.
// The uncompressed size is returned as well.
// dst must have capacity to contain the entire compressed block.
func (l *LZ4Converter) ConvertBlock(dst, src []byte) ([]byte, int, error) {
if len(src) == 0 {
return dst, 0, nil
}
const debug = false
const inline = true
const lz4MinMatch = 4
s, d := 0, len(dst)
dst = dst[:cap(dst)]
if !debug && hasAmd64Asm {
res, sz := cvtLZ4BlockAsm(dst[d:], src)
if res < 0 {
const (
errCorrupt = -1
errDstTooSmall = -2
)
switch res {
case errCorrupt:
return nil, 0, ErrCorrupt
case errDstTooSmall:
return nil, 0, ErrDstTooSmall
default:
return nil, 0, fmt.Errorf("unexpected result: %d", res)
}
}
if d+sz > len(dst) {
return nil, 0, ErrDstTooSmall
}
return dst[:d+sz], res, nil
}
dLimit := len(dst) - 10
var lastOffset uint16
var uncompressed int
if debug {
fmt.Printf("convert block start: len(src): %d, len(dst):%d \n", len(src), len(dst))
}
for {
if s >= len(src) {
return dst[:d], 0, ErrCorrupt
}
// Read literal info
token := src[s]
ll := int(token >> 4)
ml := int(lz4MinMatch + (token & 0xf))
// If upper nibble is 15, literal length is extended
if token >= 0xf0 {
for {
s++
if s >= len(src) {
if debug {
fmt.Printf("error reading ll: s (%d) >= len(src) (%d)\n", s, len(src))
}
return dst[:d], 0, ErrCorrupt
}
val := src[s]
ll += int(val)
if val != 255 {
break
}
}
}
// Skip past token
if s+ll >= len(src) {
if debug {
fmt.Printf("error literals: s+ll (%d+%d) >= len(src) (%d)\n", s, ll, len(src))
}
return nil, 0, ErrCorrupt
}
s++
if ll > 0 {
if d+ll > dLimit {
return nil, 0, ErrDstTooSmall
}
if debug {
fmt.Printf("emit %d literals\n", ll)
}
d += emitLiteralGo(dst[d:], src[s:s+ll])
s += ll
uncompressed += ll
}
// Check if we are done...
if s == len(src) && ml == lz4MinMatch {
break
}
// 2 byte offset
if s >= len(src)-2 {
if debug {
fmt.Printf("s (%d) >= len(src)-2 (%d)", s, len(src)-2)
}
return nil, 0, ErrCorrupt
}
offset := binary.LittleEndian.Uint16(src[s:])
s += 2
if offset == 0 {
if debug {
fmt.Printf("error: offset 0, ml: %d, len(src)-s: %d\n", ml, len(src)-s)
}
return nil, 0, ErrCorrupt
}
if int(offset) > uncompressed {
if debug {
fmt.Printf("error: offset (%d)> uncompressed (%d)\n", offset, uncompressed)
}
return nil, 0, ErrCorrupt
}
if ml == lz4MinMatch+15 {
for {
if s >= len(src) {
if debug {
fmt.Printf("error reading ml: s (%d) >= len(src) (%d)\n", s, len(src))
}
return nil, 0, ErrCorrupt
}
val := src[s]
s++
ml += int(val)
if val != 255 {
if s >= len(src) {
if debug {
fmt.Printf("error reading ml: s (%d) >= len(src) (%d)\n", s, len(src))
}
return nil, 0, ErrCorrupt
}
break
}
}
}
if offset == lastOffset {
if debug {
fmt.Printf("emit repeat, length: %d, offset: %d\n", ml, offset)
}
if !inline {
d += emitRepeat16(dst[d:], offset, ml)
} else {
length := ml
dst := dst[d:]
for len(dst) > 5 {
// Repeat offset, make length cheaper
length -= 4
if length <= 4 {
dst[0] = uint8(length)<<2 | tagCopy1
dst[1] = 0
d += 2
break
}
if length < 8 && offset < 2048 {
// Encode WITH offset
dst[1] = uint8(offset)
dst[0] = uint8(offset>>8)<<5 | uint8(length)<<2 | tagCopy1
d += 2
break
}
if length < (1<<8)+4 {
length -= 4
dst[2] = uint8(length)
dst[1] = 0
dst[0] = 5<<2 | tagCopy1
d += 3
break
}
if length < (1<<16)+(1<<8) {
length -= 1 << 8
dst[3] = uint8(length >> 8)
dst[2] = uint8(length >> 0)
dst[1] = 0
dst[0] = 6<<2 | tagCopy1
d += 4
break
}
const maxRepeat = (1 << 24) - 1
length -= 1 << 16
left := 0
if length > maxRepeat {
left = length - maxRepeat + 4
length = maxRepeat - 4
}
dst[4] = uint8(length >> 16)
dst[3] = uint8(length >> 8)
dst[2] = uint8(length >> 0)
dst[1] = 0
dst[0] = 7<<2 | tagCopy1
if left > 0 {
d += 5 + emitRepeat16(dst[5:], offset, left)
break
}
d += 5
break
}
}
} else {
if debug {
fmt.Printf("emit copy, length: %d, offset: %d\n", ml, offset)
}
if !inline {
d += emitCopy16(dst[d:], offset, ml)
} else {
length := ml
dst := dst[d:]
for len(dst) > 5 {
// Offset no more than 2 bytes.
if length > 64 {
off := 3
if offset < 2048 {
// emit 8 bytes as tagCopy1, rest as repeats.
dst[1] = uint8(offset)
dst[0] = uint8(offset>>8)<<5 | uint8(8-4)<<2 | tagCopy1
length -= 8
off = 2
} else {
// Emit a length 60 copy, encoded as 3 bytes.
// Emit remaining as repeat value (minimum 4 bytes).
dst[2] = uint8(offset >> 8)
dst[1] = uint8(offset)
dst[0] = 59<<2 | tagCopy2
length -= 60
}
// Emit remaining as repeats, at least 4 bytes remain.
d += off + emitRepeat16(dst[off:], offset, length)
break
}
if length >= 12 || offset >= 2048 {
// Emit the remaining copy, encoded as 3 bytes.
dst[2] = uint8(offset >> 8)
dst[1] = uint8(offset)
dst[0] = uint8(length-1)<<2 | tagCopy2
d += 3
break
}
// Emit the remaining copy, encoded as 2 bytes.
dst[1] = uint8(offset)
dst[0] = uint8(offset>>8)<<5 | uint8(length-4)<<2 | tagCopy1
d += 2
break
}
}
lastOffset = offset
}
uncompressed += ml
if d > dLimit {
return nil, 0, ErrDstTooSmall
}
}
return dst[:d], uncompressed, nil
}
// ConvertBlockSnappy will convert an LZ4 block and append it
// as a Snappy block without block length to dst.
// The uncompressed size is returned as well.
// dst must have capacity to contain the entire compressed block.
func (l *LZ4Converter) ConvertBlockSnappy(dst, src []byte) ([]byte, int, error) {
if len(src) == 0 {
return dst, 0, nil
}
const debug = false
const lz4MinMatch = 4
s, d := 0, len(dst)
dst = dst[:cap(dst)]
// Use assembly when possible
if !debug && hasAmd64Asm {
res, sz := cvtLZ4BlockSnappyAsm(dst[d:], src)
if res < 0 {
const (
errCorrupt = -1
errDstTooSmall = -2
)
switch res {
case errCorrupt:
return nil, 0, ErrCorrupt
case errDstTooSmall:
return nil, 0, ErrDstTooSmall
default:
return nil, 0, fmt.Errorf("unexpected result: %d", res)
}
}
if d+sz > len(dst) {
return nil, 0, ErrDstTooSmall
}
return dst[:d+sz], res, nil
}
dLimit := len(dst) - 10
var uncompressed int
if debug {
fmt.Printf("convert block start: len(src): %d, len(dst):%d \n", len(src), len(dst))
}
for {
if s >= len(src) {
return nil, 0, ErrCorrupt
}
// Read literal info
token := src[s]
ll := int(token >> 4)
ml := int(lz4MinMatch + (token & 0xf))
// If upper nibble is 15, literal length is extended
if token >= 0xf0 {
for {
s++
if s >= len(src) {
if debug {
fmt.Printf("error reading ll: s (%d) >= len(src) (%d)\n", s, len(src))
}
return nil, 0, ErrCorrupt
}
val := src[s]
ll += int(val)
if val != 255 {
break
}
}
}
// Skip past token
if s+ll >= len(src) {
if debug {
fmt.Printf("error literals: s+ll (%d+%d) >= len(src) (%d)\n", s, ll, len(src))
}
return nil, 0, ErrCorrupt
}
s++
if ll > 0 {
if d+ll > dLimit {
return nil, 0, ErrDstTooSmall
}
if debug {
fmt.Printf("emit %d literals\n", ll)
}
d += emitLiteralGo(dst[d:], src[s:s+ll])
s += ll
uncompressed += ll
}
// Check if we are done...
if s == len(src) && ml == lz4MinMatch {
break
}
// 2 byte offset
if s >= len(src)-2 {
if debug {
fmt.Printf("s (%d) >= len(src)-2 (%d)", s, len(src)-2)
}
return nil, 0, ErrCorrupt
}
offset := binary.LittleEndian.Uint16(src[s:])
s += 2
if offset == 0 {
if debug {
fmt.Printf("error: offset 0, ml: %d, len(src)-s: %d\n", ml, len(src)-s)
}
return nil, 0, ErrCorrupt
}
if int(offset) > uncompressed {
if debug {
fmt.Printf("error: offset (%d)> uncompressed (%d)\n", offset, uncompressed)
}
return nil, 0, ErrCorrupt
}
if ml == lz4MinMatch+15 {
for {
if s >= len(src) {
if debug {
fmt.Printf("error reading ml: s (%d) >= len(src) (%d)\n", s, len(src))
}
return nil, 0, ErrCorrupt
}
val := src[s]
s++
ml += int(val)
if val != 255 {
if s >= len(src) {
if debug {
fmt.Printf("error reading ml: s (%d) >= len(src) (%d)\n", s, len(src))
}
return nil, 0, ErrCorrupt
}
break
}
}
}
if debug {
fmt.Printf("emit copy, length: %d, offset: %d\n", ml, offset)
}
length := ml
// d += emitCopyNoRepeat(dst[d:], int(offset), ml)
for length > 0 {
if d >= dLimit {
return nil, 0, ErrDstTooSmall
}
// Offset no more than 2 bytes.
if length > 64 {
// Emit a length 64 copy, encoded as 3 bytes.
dst[d+2] = uint8(offset >> 8)
dst[d+1] = uint8(offset)
dst[d+0] = 63<<2 | tagCopy2
length -= 64
d += 3
continue
}
if length >= 12 || offset >= 2048 || length < 4 {
// Emit the remaining copy, encoded as 3 bytes.
dst[d+2] = uint8(offset >> 8)
dst[d+1] = uint8(offset)
dst[d+0] = uint8(length-1)<<2 | tagCopy2
d += 3
break
}
// Emit the remaining copy, encoded as 2 bytes.
dst[d+1] = uint8(offset)
dst[d+0] = uint8(offset>>8)<<5 | uint8(length-4)<<2 | tagCopy1
d += 2
break
}
uncompressed += ml
if d > dLimit {
return nil, 0, ErrDstTooSmall
}
}
return dst[:d], uncompressed, nil
}
// emitRepeat writes a repeat chunk and returns the number of bytes written.
// Length must be at least 4 and < 1<<24
func emitRepeat16(dst []byte, offset uint16, length int) int {
// Repeat offset, make length cheaper
length -= 4
if length <= 4 {
dst[0] = uint8(length)<<2 | tagCopy1
dst[1] = 0
return 2
}
if length < 8 && offset < 2048 {
// Encode WITH offset
dst[1] = uint8(offset)
dst[0] = uint8(offset>>8)<<5 | uint8(length)<<2 | tagCopy1
return 2
}
if length < (1<<8)+4 {
length -= 4
dst[2] = uint8(length)
dst[1] = 0
dst[0] = 5<<2 | tagCopy1
return 3
}
if length < (1<<16)+(1<<8) {
length -= 1 << 8
dst[3] = uint8(length >> 8)
dst[2] = uint8(length >> 0)
dst[1] = 0
dst[0] = 6<<2 | tagCopy1
return 4
}
const maxRepeat = (1 << 24) - 1
length -= 1 << 16
left := 0
if length > maxRepeat {
left = length - maxRepeat + 4
length = maxRepeat - 4
}
dst[4] = uint8(length >> 16)
dst[3] = uint8(length >> 8)
dst[2] = uint8(length >> 0)
dst[1] = 0
dst[0] = 7<<2 | tagCopy1
if left > 0 {
return 5 + emitRepeat16(dst[5:], offset, left)
}
return 5
}
// emitCopy writes a copy chunk and returns the number of bytes written.
//
// It assumes that:
//
// dst is long enough to hold the encoded bytes
// 1 <= offset && offset <= math.MaxUint16
// 4 <= length && length <= math.MaxUint32
func emitCopy16(dst []byte, offset uint16, length int) int {
// Offset no more than 2 bytes.
if length > 64 {
off := 3
if offset < 2048 {
// emit 8 bytes as tagCopy1, rest as repeats.
dst[1] = uint8(offset)
dst[0] = uint8(offset>>8)<<5 | uint8(8-4)<<2 | tagCopy1
length -= 8
off = 2
} else {
// Emit a length 60 copy, encoded as 3 bytes.
// Emit remaining as repeat value (minimum 4 bytes).
dst[2] = uint8(offset >> 8)
dst[1] = uint8(offset)
dst[0] = 59<<2 | tagCopy2
length -= 60
}
// Emit remaining as repeats, at least 4 bytes remain.
return off + emitRepeat16(dst[off:], offset, length)
}
if length >= 12 || offset >= 2048 {
// Emit the remaining copy, encoded as 3 bytes.
dst[2] = uint8(offset >> 8)
dst[1] = uint8(offset)
dst[0] = uint8(length-1)<<2 | tagCopy2
return 3
}
// Emit the remaining copy, encoded as 2 bytes.
dst[1] = uint8(offset)
dst[0] = uint8(offset>>8)<<5 | uint8(length-4)<<2 | tagCopy1
return 2
}
// emitLiteral writes a literal chunk and returns the number of bytes written.
//
// It assumes that:
//
// dst is long enough to hold the encoded bytes
// 0 <= len(lit) && len(lit) <= math.MaxUint32
func emitLiteralGo(dst, lit []byte) int {
if len(lit) == 0 {
return 0
}
i, n := 0, uint(len(lit)-1)
switch {
case n < 60:
dst[0] = uint8(n)<<2 | tagLiteral
i = 1
case n < 1<<8:
dst[1] = uint8(n)
dst[0] = 60<<2 | tagLiteral
i = 2
case n < 1<<16:
dst[2] = uint8(n >> 8)
dst[1] = uint8(n)
dst[0] = 61<<2 | tagLiteral
i = 3
case n < 1<<24:
dst[3] = uint8(n >> 16)
dst[2] = uint8(n >> 8)
dst[1] = uint8(n)
dst[0] = 62<<2 | tagLiteral
i = 4
default:
dst[4] = uint8(n >> 24)
dst[3] = uint8(n >> 16)
dst[2] = uint8(n >> 8)
dst[1] = uint8(n)
dst[0] = 63<<2 | tagLiteral
i = 5
}
return i + copy(dst[i:], lit)
}
vendor/github.com/klauspost/compress/s2/lz4sconvert.go
Generated Vendored
+467
View File
@@ -0,0 +1,467 @@
// Copyright (c) 2022 Klaus Post. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package s2
import (
"encoding/binary"
"fmt"
)
// LZ4sConverter provides conversion from LZ4s.
// (Intel modified LZ4 Blocks)
// https://cdrdv2-public.intel.com/743912/743912-qat-programmers-guide-v2.0.pdf
// LZ4s is a variant of LZ4 block format. LZ4s should be considered as an intermediate compressed block format.
// The LZ4s format is selected when the application sets the compType to CPA_DC_LZ4S in CpaDcSessionSetupData.
// The LZ4s block returned by the Intel® QAT hardware can be used by an external
// software post-processing to generate other compressed data formats.
// The following table lists the differences between LZ4 and LZ4s block format. LZ4s block format uses
// the same high-level formatting as LZ4 block format with the following encoding changes:
// For Min Match of 4 bytes, Copy length value 1-15 means length 4-18 with 18 bytes adding an extra byte.
// ONLY "Min match of 4 bytes" is supported.
type LZ4sConverter struct {
}
// ConvertBlock will convert an LZ4s block and append it as an S2
// block without block length to dst.
// The uncompressed size is returned as well.
// dst must have capacity to contain the entire compressed block.
func (l *LZ4sConverter) ConvertBlock(dst, src []byte) ([]byte, int, error) {
if len(src) == 0 {
return dst, 0, nil
}
const debug = false
const inline = true
const lz4MinMatch = 3
s, d := 0, len(dst)
dst = dst[:cap(dst)]
if !debug && hasAmd64Asm {
res, sz := cvtLZ4sBlockAsm(dst[d:], src)
if res < 0 {
const (
errCorrupt = -1
errDstTooSmall = -2
)
switch res {
case errCorrupt:
return nil, 0, ErrCorrupt
case errDstTooSmall:
return nil, 0, ErrDstTooSmall
default:
return nil, 0, fmt.Errorf("unexpected result: %d", res)
}
}
if d+sz > len(dst) {
return nil, 0, ErrDstTooSmall
}
return dst[:d+sz], res, nil
}
dLimit := len(dst) - 10
var lastOffset uint16
var uncompressed int
if debug {
fmt.Printf("convert block start: len(src): %d, len(dst):%d \n", len(src), len(dst))
}
for {
if s >= len(src) {
return dst[:d], 0, ErrCorrupt
}
// Read literal info
token := src[s]
ll := int(token >> 4)
ml := int(lz4MinMatch + (token & 0xf))
// If upper nibble is 15, literal length is extended
if token >= 0xf0 {
for {
s++
if s >= len(src) {
if debug {
fmt.Printf("error reading ll: s (%d) >= len(src) (%d)\n", s, len(src))
}
return dst[:d], 0, ErrCorrupt
}
val := src[s]
ll += int(val)
if val != 255 {
break
}
}
}
// Skip past token
if s+ll >= len(src) {
if debug {
fmt.Printf("error literals: s+ll (%d+%d) >= len(src) (%d)\n", s, ll, len(src))
}
return nil, 0, ErrCorrupt
}
s++
if ll > 0 {
if d+ll > dLimit {
return nil, 0, ErrDstTooSmall
}
if debug {
fmt.Printf("emit %d literals\n", ll)
}
d += emitLiteralGo(dst[d:], src[s:s+ll])
s += ll
uncompressed += ll
}
// Check if we are done...
if ml == lz4MinMatch {
if s == len(src) {
break
}
// 0 bytes.
continue
}
// 2 byte offset
if s >= len(src)-2 {
if debug {
fmt.Printf("s (%d) >= len(src)-2 (%d)", s, len(src)-2)
}
return nil, 0, ErrCorrupt
}
offset := binary.LittleEndian.Uint16(src[s:])
s += 2
if offset == 0 {
if debug {
fmt.Printf("error: offset 0, ml: %d, len(src)-s: %d\n", ml, len(src)-s)
}
return nil, 0, ErrCorrupt
}
if int(offset) > uncompressed {
if debug {
fmt.Printf("error: offset (%d)> uncompressed (%d)\n", offset, uncompressed)
}
return nil, 0, ErrCorrupt
}
if ml == lz4MinMatch+15 {
for {
if s >= len(src) {
if debug {
fmt.Printf("error reading ml: s (%d) >= len(src) (%d)\n", s, len(src))
}
return nil, 0, ErrCorrupt
}
val := src[s]
s++
ml += int(val)
if val != 255 {
if s >= len(src) {
if debug {
fmt.Printf("error reading ml: s (%d) >= len(src) (%d)\n", s, len(src))
}
return nil, 0, ErrCorrupt
}
break
}
}
}
if offset == lastOffset {
if debug {
fmt.Printf("emit repeat, length: %d, offset: %d\n", ml, offset)
}
if !inline {
d += emitRepeat16(dst[d:], offset, ml)
} else {
length := ml
dst := dst[d:]
for len(dst) > 5 {
// Repeat offset, make length cheaper
length -= 4
if length <= 4 {
dst[0] = uint8(length)<<2 | tagCopy1
dst[1] = 0
d += 2
break
}
if length < 8 && offset < 2048 {
// Encode WITH offset
dst[1] = uint8(offset)
dst[0] = uint8(offset>>8)<<5 | uint8(length)<<2 | tagCopy1
d += 2
break
}
if length < (1<<8)+4 {
length -= 4
dst[2] = uint8(length)
dst[1] = 0
dst[0] = 5<<2 | tagCopy1
d += 3
break
}
if length < (1<<16)+(1<<8) {
length -= 1 << 8
dst[3] = uint8(length >> 8)
dst[2] = uint8(length >> 0)
dst[1] = 0
dst[0] = 6<<2 | tagCopy1
d += 4
break
}
const maxRepeat = (1 << 24) - 1
length -= 1 << 16
left := 0
if length > maxRepeat {
left = length - maxRepeat + 4
length = maxRepeat - 4
}
dst[4] = uint8(length >> 16)
dst[3] = uint8(length >> 8)
dst[2] = uint8(length >> 0)
dst[1] = 0
dst[0] = 7<<2 | tagCopy1
if left > 0 {
d += 5 + emitRepeat16(dst[5:], offset, left)
break
}
d += 5
break
}
}
} else {
if debug {
fmt.Printf("emit copy, length: %d, offset: %d\n", ml, offset)
}
if !inline {
d += emitCopy16(dst[d:], offset, ml)
} else {
length := ml
dst := dst[d:]
for len(dst) > 5 {
// Offset no more than 2 bytes.
if length > 64 {
off := 3
if offset < 2048 {
// emit 8 bytes as tagCopy1, rest as repeats.
dst[1] = uint8(offset)
dst[0] = uint8(offset>>8)<<5 | uint8(8-4)<<2 | tagCopy1
length -= 8
off = 2
} else {
// Emit a length 60 copy, encoded as 3 bytes.
// Emit remaining as repeat value (minimum 4 bytes).
dst[2] = uint8(offset >> 8)
dst[1] = uint8(offset)
dst[0] = 59<<2 | tagCopy2
length -= 60
}
// Emit remaining as repeats, at least 4 bytes remain.
d += off + emitRepeat16(dst[off:], offset, length)
break
}
if length >= 12 || offset >= 2048 {
// Emit the remaining copy, encoded as 3 bytes.
dst[2] = uint8(offset >> 8)
dst[1] = uint8(offset)
dst[0] = uint8(length-1)<<2 | tagCopy2
d += 3
break
}
// Emit the remaining copy, encoded as 2 bytes.
dst[1] = uint8(offset)
dst[0] = uint8(offset>>8)<<5 | uint8(length-4)<<2 | tagCopy1
d += 2
break
}
}
lastOffset = offset
}
uncompressed += ml
if d > dLimit {
return nil, 0, ErrDstTooSmall
}
}
return dst[:d], uncompressed, nil
}
// ConvertBlockSnappy will convert an LZ4s block and append it
// as a Snappy block without block length to dst.
// The uncompressed size is returned as well.
// dst must have capacity to contain the entire compressed block.
func (l *LZ4sConverter) ConvertBlockSnappy(dst, src []byte) ([]byte, int, error) {
if len(src) == 0 {
return dst, 0, nil
}
const debug = false
const lz4MinMatch = 3
s, d := 0, len(dst)
dst = dst[:cap(dst)]
// Use assembly when possible
if !debug && hasAmd64Asm {
res, sz := cvtLZ4sBlockSnappyAsm(dst[d:], src)
if res < 0 {
const (
errCorrupt = -1
errDstTooSmall = -2
)
switch res {
case errCorrupt:
return nil, 0, ErrCorrupt
case errDstTooSmall:
return nil, 0, ErrDstTooSmall
default:
return nil, 0, fmt.Errorf("unexpected result: %d", res)
}
}
if d+sz > len(dst) {
return nil, 0, ErrDstTooSmall
}
return dst[:d+sz], res, nil
}
dLimit := len(dst) - 10
var uncompressed int
if debug {
fmt.Printf("convert block start: len(src): %d, len(dst):%d \n", len(src), len(dst))
}
for {
if s >= len(src) {
return nil, 0, ErrCorrupt
}
// Read literal info
token := src[s]
ll := int(token >> 4)
ml := int(lz4MinMatch + (token & 0xf))
// If upper nibble is 15, literal length is extended
if token >= 0xf0 {
for {
s++
if s >= len(src) {
if debug {
fmt.Printf("error reading ll: s (%d) >= len(src) (%d)\n", s, len(src))
}
return nil, 0, ErrCorrupt
}
val := src[s]
ll += int(val)
if val != 255 {
break
}
}
}
// Skip past token
if s+ll >= len(src) {
if debug {
fmt.Printf("error literals: s+ll (%d+%d) >= len(src) (%d)\n", s, ll, len(src))
}
return nil, 0, ErrCorrupt
}
s++
if ll > 0 {
if d+ll > dLimit {
return nil, 0, ErrDstTooSmall
}
if debug {
fmt.Printf("emit %d literals\n", ll)
}
d += emitLiteralGo(dst[d:], src[s:s+ll])
s += ll
uncompressed += ll
}
// Check if we are done...
if ml == lz4MinMatch {
if s == len(src) {
break
}
// 0 bytes.
continue
}
// 2 byte offset
if s >= len(src)-2 {
if debug {
fmt.Printf("s (%d) >= len(src)-2 (%d)", s, len(src)-2)
}
return nil, 0, ErrCorrupt
}
offset := binary.LittleEndian.Uint16(src[s:])
s += 2
if offset == 0 {
if debug {
fmt.Printf("error: offset 0, ml: %d, len(src)-s: %d\n", ml, len(src)-s)
}
return nil, 0, ErrCorrupt
}
if int(offset) > uncompressed {
if debug {
fmt.Printf("error: offset (%d)> uncompressed (%d)\n", offset, uncompressed)
}
return nil, 0, ErrCorrupt
}
if ml == lz4MinMatch+15 {
for {
if s >= len(src) {
if debug {
fmt.Printf("error reading ml: s (%d) >= len(src) (%d)\n", s, len(src))
}
return nil, 0, ErrCorrupt
}
val := src[s]
s++
ml += int(val)
if val != 255 {
if s >= len(src) {
if debug {
fmt.Printf("error reading ml: s (%d) >= len(src) (%d)\n", s, len(src))
}
return nil, 0, ErrCorrupt
}
break
}
}
}
if debug {
fmt.Printf("emit copy, length: %d, offset: %d\n", ml, offset)
}
length := ml
// d += emitCopyNoRepeat(dst[d:], int(offset), ml)
for length > 0 {
if d >= dLimit {
return nil, 0, ErrDstTooSmall
}
// Offset no more than 2 bytes.
if length > 64 {
// Emit a length 64 copy, encoded as 3 bytes.
dst[d+2] = uint8(offset >> 8)
dst[d+1] = uint8(offset)
dst[d+0] = 63<<2 | tagCopy2
length -= 64
d += 3
continue
}
if length >= 12 || offset >= 2048 || length < 4 {
// Emit the remaining copy, encoded as 3 bytes.
dst[d+2] = uint8(offset >> 8)
dst[d+1] = uint8(offset)
dst[d+0] = uint8(length-1)<<2 | tagCopy2
d += 3
break
}
// Emit the remaining copy, encoded as 2 bytes.
dst[d+1] = uint8(offset)
dst[d+0] = uint8(offset>>8)<<5 | uint8(length-4)<<2 | tagCopy1
d += 2
break
}
uncompressed += ml
if d > dLimit {
return nil, 0, ErrDstTooSmall
}
}
return dst[:d], uncompressed, nil
}
vendor/github.com/klauspost/compress/s2/reader.go
Generated Vendored
+1055
View File
File diff suppressed because it is too large Load Diff
vendor/github.com/klauspost/compress/s2/writer.go
Generated Vendored
+1020
View File
File diff suppressed because it is too large Load Diff