Vendor in containers/(common, storage, image)

Signed-off-by: Daniel J Walsh <dwalsh@redhat.com>
This commit is contained in:
Daniel J Walsh
2022-05-03 09:59:43 -04:00
parent b8b0e9937b
commit 3c286dd1d1
152 changed files with 9926 additions and 1403 deletions

View File

@@ -386,47 +386,31 @@ In practice this means that concurrency is often limited to utilizing about 3 co
### Benchmarks
These are some examples of performance compared to [datadog cgo library](https://github.com/DataDog/zstd).
The first two are streaming decodes and the last are smaller inputs.
Running on AMD Ryzen 9 3950X 16-Core Processor. AMD64 assembly used.
```
BenchmarkDecoderSilesia-8 3 385000067 ns/op 550.51 MB/s 5498 B/op 8 allocs/op
BenchmarkDecoderSilesiaCgo-8 6 197666567 ns/op 1072.25 MB/s 270672 B/op 8 allocs/op
BenchmarkDecoderSilesia-32 5 206878840 ns/op 1024.50 MB/s 49808 B/op 43 allocs/op
BenchmarkDecoderEnwik9-32 1 1271809000 ns/op 786.28 MB/s 72048 B/op 52 allocs/op
BenchmarkDecoderEnwik9-8 1 2027001600 ns/op 493.34 MB/s 10496 B/op 18 allocs/op
BenchmarkDecoderEnwik9Cgo-8 2 979499200 ns/op 1020.93 MB/s 270672 B/op 8 allocs/op
Concurrent blocks, performance:
Concurrent performance:
BenchmarkDecoder_DecodeAllParallel/kppkn.gtb.zst-16 28915 42469 ns/op 4340.07 MB/s 114 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/geo.protodata.zst-16 116505 9965 ns/op 11900.16 MB/s 16 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/plrabn12.txt.zst-16 8952 134272 ns/op 3588.70 MB/s 915 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/lcet10.txt.zst-16 11820 102538 ns/op 4161.90 MB/s 594 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/asyoulik.txt.zst-16 34782 34184 ns/op 3661.88 MB/s 60 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/alice29.txt.zst-16 27712 43447 ns/op 3500.58 MB/s 99 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/html_x_4.zst-16 62826 18750 ns/op 21845.10 MB/s 104 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/paper-100k.pdf.zst-16 631545 1794 ns/op 57078.74 MB/s 2 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/fireworks.jpeg.zst-16 1690140 712 ns/op 172938.13 MB/s 1 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/urls.10K.zst-16 10432 113593 ns/op 6180.73 MB/s 1143 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/html.zst-16 113206 10671 ns/op 9596.27 MB/s 15 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/comp-data.bin.zst-16 1530615 779 ns/op 5229.49 MB/s 0 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/kppkn.gtb.zst-16 65217 16192 ns/op 11383.34 MB/s 46 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/geo.protodata.zst-16 292671 4039 ns/op 29363.19 MB/s 6 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/plrabn12.txt.zst-16 26314 46021 ns/op 10470.43 MB/s 293 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/lcet10.txt.zst-16 33897 34900 ns/op 12227.96 MB/s 205 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/asyoulik.txt.zst-16 104348 11433 ns/op 10949.01 MB/s 20 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/alice29.txt.zst-16 75949 15510 ns/op 9805.60 MB/s 32 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/html_x_4.zst-16 173910 6756 ns/op 60624.29 MB/s 37 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/paper-100k.pdf.zst-16 923076 1339 ns/op 76474.87 MB/s 1 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/fireworks.jpeg.zst-16 922920 1351 ns/op 91102.57 MB/s 2 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/urls.10K.zst-16 27649 43618 ns/op 16096.19 MB/s 407 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/html.zst-16 279073 4160 ns/op 24614.18 MB/s 6 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/comp-data.bin.zst-16 749938 1579 ns/op 2581.71 MB/s 0 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/kppkn.gtb.zst-32 67356 17857 ns/op 10321.96 MB/s 22.48 pct 102 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/geo.protodata.zst-32 266656 4421 ns/op 26823.21 MB/s 11.89 pct 19 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/plrabn12.txt.zst-32 20992 56842 ns/op 8477.17 MB/s 39.90 pct 754 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/lcet10.txt.zst-32 27456 43932 ns/op 9714.01 MB/s 33.27 pct 524 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/asyoulik.txt.zst-32 78432 15047 ns/op 8319.15 MB/s 40.34 pct 66 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/alice29.txt.zst-32 65800 18436 ns/op 8249.63 MB/s 37.75 pct 88 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/html_x_4.zst-32 102993 11523 ns/op 35546.09 MB/s 3.637 pct 143 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/paper-100k.pdf.zst-32 1000000 1070 ns/op 95720.98 MB/s 80.53 pct 3 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/fireworks.jpeg.zst-32 749802 1752 ns/op 70272.35 MB/s 100.0 pct 5 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/urls.10K.zst-32 22640 52934 ns/op 13263.37 MB/s 26.25 pct 1014 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/html.zst-32 226412 5232 ns/op 19572.27 MB/s 14.49 pct 20 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/comp-data.bin.zst-32 923041 1276 ns/op 3194.71 MB/s 31.26 pct 0 B/op 0 allocs/op
```
This reflects the performance around May 2020, but this may be out of date.
This reflects the performance around May 2022, but this may be out of date.
## Zstd inside ZIP files

View File

@@ -5,9 +5,14 @@
package zstd
import (
"bytes"
"encoding/binary"
"errors"
"fmt"
"io"
"io/ioutil"
"os"
"path/filepath"
"sync"
"github.com/klauspost/compress/huff0"
@@ -38,6 +43,9 @@ const (
// maxCompressedBlockSize is the biggest allowed compressed block size (128KB)
maxCompressedBlockSize = 128 << 10
compressedBlockOverAlloc = 16
maxCompressedBlockSizeAlloc = 128<<10 + compressedBlockOverAlloc
// Maximum possible block size (all Raw+Uncompressed).
maxBlockSize = (1 << 21) - 1
@@ -136,7 +144,7 @@ func (b *blockDec) reset(br byteBuffer, windowSize uint64) error {
b.Type = blockType((bh >> 1) & 3)
// find size.
cSize := int(bh >> 3)
maxSize := maxBlockSize
maxSize := maxCompressedBlockSizeAlloc
switch b.Type {
case blockTypeReserved:
return ErrReservedBlockType
@@ -157,9 +165,9 @@ func (b *blockDec) reset(br byteBuffer, windowSize uint64) error {
println("Data size on stream:", cSize)
}
b.RLESize = 0
maxSize = maxCompressedBlockSize
maxSize = maxCompressedBlockSizeAlloc
if windowSize < maxCompressedBlockSize && b.lowMem {
maxSize = int(windowSize)
maxSize = int(windowSize) + compressedBlockOverAlloc
}
if cSize > maxCompressedBlockSize || uint64(cSize) > b.WindowSize {
if debugDecoder {
@@ -190,9 +198,9 @@ func (b *blockDec) reset(br byteBuffer, windowSize uint64) error {
// Read block data.
if cap(b.dataStorage) < cSize {
if b.lowMem || cSize > maxCompressedBlockSize {
b.dataStorage = make([]byte, 0, cSize)
b.dataStorage = make([]byte, 0, cSize+compressedBlockOverAlloc)
} else {
b.dataStorage = make([]byte, 0, maxCompressedBlockSize)
b.dataStorage = make([]byte, 0, maxCompressedBlockSizeAlloc)
}
}
if cap(b.dst) <= maxSize {
@@ -486,10 +494,15 @@ func (b *blockDec) decodeCompressed(hist *history) error {
b.dst = append(b.dst, hist.decoders.literals...)
return nil
}
err = hist.decoders.decodeSync(hist)
before := len(hist.decoders.out)
err = hist.decoders.decodeSync(hist.b[hist.ignoreBuffer:])
if err != nil {
return err
}
if hist.decoders.maxSyncLen > 0 {
hist.decoders.maxSyncLen += uint64(before)
hist.decoders.maxSyncLen -= uint64(len(hist.decoders.out))
}
b.dst = hist.decoders.out
hist.recentOffsets = hist.decoders.prevOffset
return nil
@@ -632,6 +645,22 @@ func (b *blockDec) prepareSequences(in []byte, hist *history) (err error) {
println("initializing sequences:", err)
return err
}
// Extract blocks...
if false && hist.dict == nil {
fatalErr := func(err error) {
if err != nil {
panic(err)
}
}
fn := fmt.Sprintf("n-%d-lits-%d-prev-%d-%d-%d-win-%d.blk", hist.decoders.nSeqs, len(hist.decoders.literals), hist.recentOffsets[0], hist.recentOffsets[1], hist.recentOffsets[2], hist.windowSize)
var buf bytes.Buffer
fatalErr(binary.Write(&buf, binary.LittleEndian, hist.decoders.litLengths.fse))
fatalErr(binary.Write(&buf, binary.LittleEndian, hist.decoders.matchLengths.fse))
fatalErr(binary.Write(&buf, binary.LittleEndian, hist.decoders.offsets.fse))
buf.Write(in)
ioutil.WriteFile(filepath.Join("testdata", "seqs", fn), buf.Bytes(), os.ModePerm)
}
return nil
}
@@ -650,6 +679,7 @@ func (b *blockDec) decodeSequences(hist *history) error {
}
hist.decoders.windowSize = hist.windowSize
hist.decoders.prevOffset = hist.recentOffsets
err := hist.decoders.decode(b.sequence)
hist.recentOffsets = hist.decoders.prevOffset
return err

View File

@@ -347,18 +347,20 @@ func (d *Decoder) DecodeAll(input, dst []byte) ([]byte, error) {
}
frame.history.setDict(&dict)
}
if frame.FrameContentSize != fcsUnknown && frame.FrameContentSize > d.o.maxDecodedSize-uint64(len(dst)) {
return dst, ErrDecoderSizeExceeded
if frame.WindowSize > d.o.maxWindowSize {
return dst, ErrWindowSizeExceeded
}
if frame.FrameContentSize < 1<<30 {
// Never preallocate more than 1 GB up front.
if frame.FrameContentSize != fcsUnknown {
if frame.FrameContentSize > d.o.maxDecodedSize-uint64(len(dst)) {
return dst, ErrDecoderSizeExceeded
}
if cap(dst)-len(dst) < int(frame.FrameContentSize) {
dst2 := make([]byte, len(dst), len(dst)+int(frame.FrameContentSize))
dst2 := make([]byte, len(dst), len(dst)+int(frame.FrameContentSize)+compressedBlockOverAlloc)
copy(dst2, dst)
dst = dst2
}
}
if cap(dst) == 0 {
// Allocate len(input) * 2 by default if nothing is provided
// and we didn't get frame content size.

View File

@@ -31,7 +31,7 @@ func (o *decoderOptions) setDefault() {
if o.concurrent > 4 {
o.concurrent = 4
}
o.maxDecodedSize = 1 << 63
o.maxDecodedSize = 64 << 30
}
// WithDecoderLowmem will set whether to use a lower amount of memory,
@@ -66,7 +66,7 @@ func WithDecoderConcurrency(n int) DOption {
// WithDecoderMaxMemory allows to set a maximum decoded size for in-memory
// non-streaming operations or maximum window size for streaming operations.
// This can be used to control memory usage of potentially hostile content.
// Maximum and default is 1 << 63 bytes.
// Maximum is 1 << 63 bytes. Default is 64GiB.
func WithDecoderMaxMemory(n uint64) DOption {
return func(o *decoderOptions) error {
if n == 0 {

View File

@@ -326,6 +326,19 @@ func (d *frameDec) runDecoder(dst []byte, dec *blockDec) ([]byte, error) {
d.history.ignoreBuffer = len(dst)
// Store input length, so we only check new data.
crcStart := len(dst)
d.history.decoders.maxSyncLen = 0
if d.FrameContentSize != fcsUnknown {
d.history.decoders.maxSyncLen = d.FrameContentSize + uint64(len(dst))
if d.history.decoders.maxSyncLen > d.o.maxDecodedSize {
return dst, ErrDecoderSizeExceeded
}
if uint64(cap(dst)) < d.history.decoders.maxSyncLen {
// Alloc for output
dst2 := make([]byte, len(dst), d.history.decoders.maxSyncLen+compressedBlockOverAlloc)
copy(dst2, dst)
dst = dst2
}
}
var err error
for {
err = dec.reset(d.rawInput, d.WindowSize)

View File

@@ -5,8 +5,10 @@
package zstd
import (
"encoding/binary"
"errors"
"fmt"
"io"
)
const (
@@ -182,6 +184,29 @@ func (s *fseDecoder) readNCount(b *byteReader, maxSymbol uint16) error {
return s.buildDtable()
}
func (s *fseDecoder) mustReadFrom(r io.Reader) {
fatalErr := func(err error) {
if err != nil {
panic(err)
}
}
// dt [maxTablesize]decSymbol // Decompression table.
// symbolLen uint16 // Length of active part of the symbol table.
// actualTableLog uint8 // Selected tablelog.
// maxBits uint8 // Maximum number of additional bits
// // used for table creation to avoid allocations.
// stateTable [256]uint16
// norm [maxSymbolValue + 1]int16
// preDefined bool
fatalErr(binary.Read(r, binary.LittleEndian, &s.dt))
fatalErr(binary.Read(r, binary.LittleEndian, &s.symbolLen))
fatalErr(binary.Read(r, binary.LittleEndian, &s.actualTableLog))
fatalErr(binary.Read(r, binary.LittleEndian, &s.maxBits))
fatalErr(binary.Read(r, binary.LittleEndian, &s.stateTable))
fatalErr(binary.Read(r, binary.LittleEndian, &s.norm))
fatalErr(binary.Read(r, binary.LittleEndian, &s.preDefined))
}
// decSymbol contains information about a state entry,
// Including the state offset base, the output symbol and
// the number of bits to read for the low part of the destination state.

View File

@@ -73,6 +73,7 @@ type sequenceDecs struct {
seqSize int
windowSize int
maxBits uint8
maxSyncLen uint64
}
// initialize all 3 decoders from the stream input.
@@ -98,153 +99,13 @@ func (s *sequenceDecs) initialize(br *bitReader, hist *history, out []byte) erro
return nil
}
// decode sequences from the stream with the provided history.
func (s *sequenceDecs) decode(seqs []seqVals) error {
br := s.br
// Grab full sizes tables, to avoid bounds checks.
llTable, mlTable, ofTable := s.litLengths.fse.dt[:maxTablesize], s.matchLengths.fse.dt[:maxTablesize], s.offsets.fse.dt[:maxTablesize]
llState, mlState, ofState := s.litLengths.state.state, s.matchLengths.state.state, s.offsets.state.state
s.seqSize = 0
litRemain := len(s.literals)
maxBlockSize := maxCompressedBlockSize
if s.windowSize < maxBlockSize {
maxBlockSize = s.windowSize
}
for i := range seqs {
var ll, mo, ml int
if br.off > 4+((maxOffsetBits+16+16)>>3) {
// inlined function:
// ll, mo, ml = s.nextFast(br, llState, mlState, ofState)
// Final will not read from stream.
var llB, mlB, moB uint8
ll, llB = llState.final()
ml, mlB = mlState.final()
mo, moB = ofState.final()
// extra bits are stored in reverse order.
br.fillFast()
mo += br.getBits(moB)
if s.maxBits > 32 {
br.fillFast()
}
ml += br.getBits(mlB)
ll += br.getBits(llB)
if moB > 1 {
s.prevOffset[2] = s.prevOffset[1]
s.prevOffset[1] = s.prevOffset[0]
s.prevOffset[0] = mo
} else {
// mo = s.adjustOffset(mo, ll, moB)
// Inlined for rather big speedup
if ll == 0 {
// There is an exception though, when current sequence's literals_length = 0.
// In this case, repeated offsets are shifted by one, so an offset_value of 1 means Repeated_Offset2,
// an offset_value of 2 means Repeated_Offset3, and an offset_value of 3 means Repeated_Offset1 - 1_byte.
mo++
}
if mo == 0 {
mo = s.prevOffset[0]
} else {
var temp int
if mo == 3 {
temp = s.prevOffset[0] - 1
} else {
temp = s.prevOffset[mo]
}
if temp == 0 {
// 0 is not valid; input is corrupted; force offset to 1
println("WARNING: temp was 0")
temp = 1
}
if mo != 1 {
s.prevOffset[2] = s.prevOffset[1]
}
s.prevOffset[1] = s.prevOffset[0]
s.prevOffset[0] = temp
mo = temp
}
}
br.fillFast()
} else {
if br.overread() {
if debugDecoder {
printf("reading sequence %d, exceeded available data\n", i)
}
return io.ErrUnexpectedEOF
}
ll, mo, ml = s.next(br, llState, mlState, ofState)
br.fill()
}
if debugSequences {
println("Seq", i, "Litlen:", ll, "mo:", mo, "(abs) ml:", ml)
}
// Evaluate.
// We might be doing this async, so do it early.
if mo == 0 && ml > 0 {
return fmt.Errorf("zero matchoff and matchlen (%d) > 0", ml)
}
if ml > maxMatchLen {
return fmt.Errorf("match len (%d) bigger than max allowed length", ml)
}
s.seqSize += ll + ml
if s.seqSize > maxBlockSize {
return fmt.Errorf("output (%d) bigger than max block size (%d)", s.seqSize, maxBlockSize)
}
litRemain -= ll
if litRemain < 0 {
return fmt.Errorf("unexpected literal count, want %d bytes, but only %d is available", ll, litRemain+ll)
}
seqs[i] = seqVals{
ll: ll,
ml: ml,
mo: mo,
}
if i == len(seqs)-1 {
// This is the last sequence, so we shouldn't update state.
break
}
// Manually inlined, ~ 5-20% faster
// Update all 3 states at once. Approx 20% faster.
nBits := llState.nbBits() + mlState.nbBits() + ofState.nbBits()
if nBits == 0 {
llState = llTable[llState.newState()&maxTableMask]
mlState = mlTable[mlState.newState()&maxTableMask]
ofState = ofTable[ofState.newState()&maxTableMask]
} else {
bits := br.get32BitsFast(nBits)
lowBits := uint16(bits >> ((ofState.nbBits() + mlState.nbBits()) & 31))
llState = llTable[(llState.newState()+lowBits)&maxTableMask]
lowBits = uint16(bits >> (ofState.nbBits() & 31))
lowBits &= bitMask[mlState.nbBits()&15]
mlState = mlTable[(mlState.newState()+lowBits)&maxTableMask]
lowBits = uint16(bits) & bitMask[ofState.nbBits()&15]
ofState = ofTable[(ofState.newState()+lowBits)&maxTableMask]
}
}
s.seqSize += litRemain
if s.seqSize > maxBlockSize {
return fmt.Errorf("output (%d) bigger than max block size (%d)", s.seqSize, maxBlockSize)
}
err := br.close()
if err != nil {
printf("Closing sequences: %v, %+v\n", err, *br)
}
return err
}
// execute will execute the decoded sequence with the provided history.
// The sequence must be evaluated before being sent.
func (s *sequenceDecs) execute(seqs []seqVals, hist []byte) error {
if len(s.dict) == 0 {
return s.executeSimple(seqs, hist)
}
// Ensure we have enough output size...
if len(s.out)+s.seqSize > cap(s.out) {
addBytes := s.seqSize + len(s.out)
@@ -341,14 +202,19 @@ func (s *sequenceDecs) execute(seqs []seqVals, hist []byte) error {
}
// decode sequences from the stream with the provided history.
func (s *sequenceDecs) decodeSync(history *history) error {
func (s *sequenceDecs) decodeSync(hist []byte) error {
if true {
supported, err := s.decodeSyncSimple(hist)
if supported {
return err
}
}
br := s.br
seqs := s.nSeqs
startSize := len(s.out)
// Grab full sizes tables, to avoid bounds checks.
llTable, mlTable, ofTable := s.litLengths.fse.dt[:maxTablesize], s.matchLengths.fse.dt[:maxTablesize], s.offsets.fse.dt[:maxTablesize]
llState, mlState, ofState := s.litLengths.state.state, s.matchLengths.state.state, s.offsets.state.state
hist := history.b[history.ignoreBuffer:]
out := s.out
maxBlockSize := maxCompressedBlockSize
if s.windowSize < maxBlockSize {
@@ -433,7 +299,7 @@ func (s *sequenceDecs) decodeSync(history *history) error {
}
size := ll + ml + len(out)
if size-startSize > maxBlockSize {
return fmt.Errorf("output (%d) bigger than max block size (%d)", size, maxBlockSize)
return fmt.Errorf("output (%d) bigger than max block size (%d)", size-startSize, maxBlockSize)
}
if size > cap(out) {
// Not enough size, which can happen under high volume block streaming conditions
@@ -463,13 +329,13 @@ func (s *sequenceDecs) decodeSync(history *history) error {
if mo > len(out)+len(hist) || mo > s.windowSize {
if len(s.dict) == 0 {
return fmt.Errorf("match offset (%d) bigger than current history (%d)", mo, len(out)+len(hist))
return fmt.Errorf("match offset (%d) bigger than current history (%d)", mo, len(out)+len(hist)-startSize)
}
// we may be in dictionary.
dictO := len(s.dict) - (mo - (len(out) + len(hist)))
if dictO < 0 || dictO >= len(s.dict) {
return fmt.Errorf("match offset (%d) bigger than current history (%d)", mo, len(out)+len(hist))
return fmt.Errorf("match offset (%d) bigger than current history (%d)", mo, len(out)+len(hist)-startSize)
}
end := dictO + ml
if end > len(s.dict) {
@@ -543,8 +409,8 @@ func (s *sequenceDecs) decodeSync(history *history) error {
}
// Check if space for literals
if len(s.literals)+len(s.out)-startSize > maxBlockSize {
return fmt.Errorf("output (%d) bigger than max block size (%d)", len(s.out), maxBlockSize)
if size := len(s.literals) + len(s.out) - startSize; size > maxBlockSize {
return fmt.Errorf("output (%d) bigger than max block size (%d)", size, maxBlockSize)
}
// Add final literals

View File

@@ -0,0 +1,350 @@
//go:build amd64 && !appengine && !noasm && gc
// +build amd64,!appengine,!noasm,gc
package zstd
import (
"fmt"
"github.com/klauspost/compress/internal/cpuinfo"
)
type decodeSyncAsmContext struct {
llTable []decSymbol
mlTable []decSymbol
ofTable []decSymbol
llState uint64
mlState uint64
ofState uint64
iteration int
litRemain int
out []byte
outPosition int
literals []byte
litPosition int
history []byte
windowSize int
ll int // set on error (not for all errors, please refer to _generate/gen.go)
ml int // set on error (not for all errors, please refer to _generate/gen.go)
mo int // set on error (not for all errors, please refer to _generate/gen.go)
}
// sequenceDecs_decodeSync_amd64 implements the main loop of sequenceDecs.decodeSync in x86 asm.
//
// Please refer to seqdec_generic.go for the reference implementation.
//go:noescape
func sequenceDecs_decodeSync_amd64(s *sequenceDecs, br *bitReader, ctx *decodeSyncAsmContext) int
// sequenceDecs_decodeSync_bmi2 implements the main loop of sequenceDecs.decodeSync in x86 asm with BMI2 extensions.
//go:noescape
func sequenceDecs_decodeSync_bmi2(s *sequenceDecs, br *bitReader, ctx *decodeSyncAsmContext) int
// sequenceDecs_decodeSync_safe_amd64 does the same as above, but does not write more than output buffer.
//go:noescape
func sequenceDecs_decodeSync_safe_amd64(s *sequenceDecs, br *bitReader, ctx *decodeSyncAsmContext) int
// sequenceDecs_decodeSync_safe_bmi2 does the same as above, but does not write more than output buffer.
//go:noescape
func sequenceDecs_decodeSync_safe_bmi2(s *sequenceDecs, br *bitReader, ctx *decodeSyncAsmContext) int
// decode sequences from the stream with the provided history but without a dictionary.
func (s *sequenceDecs) decodeSyncSimple(hist []byte) (bool, error) {
if len(s.dict) > 0 {
return false, nil
}
if s.maxSyncLen == 0 && cap(s.out)-len(s.out) < maxCompressedBlockSize {
return false, nil
}
useSafe := false
if s.maxSyncLen == 0 && cap(s.out)-len(s.out) < maxCompressedBlockSizeAlloc {
useSafe = true
}
if s.maxSyncLen > 0 && cap(s.out)-len(s.out)-compressedBlockOverAlloc < int(s.maxSyncLen) {
useSafe = true
}
br := s.br
maxBlockSize := maxCompressedBlockSize
if s.windowSize < maxBlockSize {
maxBlockSize = s.windowSize
}
ctx := decodeSyncAsmContext{
llTable: s.litLengths.fse.dt[:maxTablesize],
mlTable: s.matchLengths.fse.dt[:maxTablesize],
ofTable: s.offsets.fse.dt[:maxTablesize],
llState: uint64(s.litLengths.state.state),
mlState: uint64(s.matchLengths.state.state),
ofState: uint64(s.offsets.state.state),
iteration: s.nSeqs - 1,
litRemain: len(s.literals),
out: s.out,
outPosition: len(s.out),
literals: s.literals,
windowSize: s.windowSize,
history: hist,
}
s.seqSize = 0
startSize := len(s.out)
var errCode int
if cpuinfo.HasBMI2() {
if useSafe {
errCode = sequenceDecs_decodeSync_safe_bmi2(s, br, &ctx)
} else {
errCode = sequenceDecs_decodeSync_bmi2(s, br, &ctx)
}
} else {
if useSafe {
errCode = sequenceDecs_decodeSync_safe_amd64(s, br, &ctx)
} else {
errCode = sequenceDecs_decodeSync_amd64(s, br, &ctx)
}
}
switch errCode {
case noError:
break
case errorMatchLenOfsMismatch:
return true, fmt.Errorf("zero matchoff and matchlen (%d) > 0", ctx.ml)
case errorMatchLenTooBig:
return true, fmt.Errorf("match len (%d) bigger than max allowed length", ctx.ml)
case errorMatchOffTooBig:
return true, fmt.Errorf("match offset (%d) bigger than current history (%d)",
ctx.mo, ctx.outPosition+len(hist)-startSize)
case errorNotEnoughLiterals:
return true, fmt.Errorf("unexpected literal count, want %d bytes, but only %d is available",
ctx.ll, ctx.litRemain+ctx.ll)
case errorNotEnoughSpace:
size := ctx.outPosition + ctx.ll + ctx.ml
if debugDecoder {
println("msl:", s.maxSyncLen, "cap", cap(s.out), "bef:", startSize, "sz:", size-startSize, "mbs:", maxBlockSize, "outsz:", cap(s.out)-startSize)
}
return true, fmt.Errorf("output (%d) bigger than max block size (%d)", size-startSize, maxBlockSize)
default:
return true, fmt.Errorf("sequenceDecs_decode returned erronous code %d", errCode)
}
s.seqSize += ctx.litRemain
if s.seqSize > maxBlockSize {
return true, fmt.Errorf("output (%d) bigger than max block size (%d)", s.seqSize, maxBlockSize)
}
err := br.close()
if err != nil {
printf("Closing sequences: %v, %+v\n", err, *br)
return true, err
}
s.literals = s.literals[ctx.litPosition:]
t := ctx.outPosition
s.out = s.out[:t]
// Add final literals
s.out = append(s.out, s.literals...)
if debugDecoder {
t += len(s.literals)
if t != len(s.out) {
panic(fmt.Errorf("length mismatch, want %d, got %d", len(s.out), t))
}
}
return true, nil
}
// --------------------------------------------------------------------------------
type decodeAsmContext struct {
llTable []decSymbol
mlTable []decSymbol
ofTable []decSymbol
llState uint64
mlState uint64
ofState uint64
iteration int
seqs []seqVals
litRemain int
}
const noError = 0
// error reported when mo == 0 && ml > 0
const errorMatchLenOfsMismatch = 1
// error reported when ml > maxMatchLen
const errorMatchLenTooBig = 2
// error reported when mo > available history or mo > s.windowSize
const errorMatchOffTooBig = 3
// error reported when the sum of literal lengths exeeceds the literal buffer size
const errorNotEnoughLiterals = 4
// error reported when capacity of `out` is too small
const errorNotEnoughSpace = 5
// sequenceDecs_decode implements the main loop of sequenceDecs in x86 asm.
//
// Please refer to seqdec_generic.go for the reference implementation.
//go:noescape
func sequenceDecs_decode_amd64(s *sequenceDecs, br *bitReader, ctx *decodeAsmContext) int
// sequenceDecs_decode implements the main loop of sequenceDecs in x86 asm.
//
// Please refer to seqdec_generic.go for the reference implementation.
//go:noescape
func sequenceDecs_decode_56_amd64(s *sequenceDecs, br *bitReader, ctx *decodeAsmContext) int
// sequenceDecs_decode implements the main loop of sequenceDecs in x86 asm with BMI2 extensions.
//go:noescape
func sequenceDecs_decode_bmi2(s *sequenceDecs, br *bitReader, ctx *decodeAsmContext) int
// sequenceDecs_decode implements the main loop of sequenceDecs in x86 asm with BMI2 extensions.
//go:noescape
func sequenceDecs_decode_56_bmi2(s *sequenceDecs, br *bitReader, ctx *decodeAsmContext) int
// decode sequences from the stream without the provided history.
func (s *sequenceDecs) decode(seqs []seqVals) error {
br := s.br
maxBlockSize := maxCompressedBlockSize
if s.windowSize < maxBlockSize {
maxBlockSize = s.windowSize
}
ctx := decodeAsmContext{
llTable: s.litLengths.fse.dt[:maxTablesize],
mlTable: s.matchLengths.fse.dt[:maxTablesize],
ofTable: s.offsets.fse.dt[:maxTablesize],
llState: uint64(s.litLengths.state.state),
mlState: uint64(s.matchLengths.state.state),
ofState: uint64(s.offsets.state.state),
seqs: seqs,
iteration: len(seqs) - 1,
litRemain: len(s.literals),
}
s.seqSize = 0
lte56bits := s.maxBits+s.offsets.fse.actualTableLog+s.matchLengths.fse.actualTableLog+s.litLengths.fse.actualTableLog <= 56
var errCode int
if cpuinfo.HasBMI2() {
if lte56bits {
errCode = sequenceDecs_decode_56_bmi2(s, br, &ctx)
} else {
errCode = sequenceDecs_decode_bmi2(s, br, &ctx)
}
} else {
if lte56bits {
errCode = sequenceDecs_decode_56_amd64(s, br, &ctx)
} else {
errCode = sequenceDecs_decode_amd64(s, br, &ctx)
}
}
if errCode != 0 {
i := len(seqs) - ctx.iteration - 1
switch errCode {
case errorMatchLenOfsMismatch:
ml := ctx.seqs[i].ml
return fmt.Errorf("zero matchoff and matchlen (%d) > 0", ml)
case errorMatchLenTooBig:
ml := ctx.seqs[i].ml
return fmt.Errorf("match len (%d) bigger than max allowed length", ml)
case errorNotEnoughLiterals:
ll := ctx.seqs[i].ll
return fmt.Errorf("unexpected literal count, want %d bytes, but only %d is available", ll, ctx.litRemain+ll)
}
return fmt.Errorf("sequenceDecs_decode_amd64 returned erronous code %d", errCode)
}
if ctx.litRemain < 0 {
return fmt.Errorf("literal count is too big: total available %d, total requested %d",
len(s.literals), len(s.literals)-ctx.litRemain)
}
s.seqSize += ctx.litRemain
if s.seqSize > maxBlockSize {
return fmt.Errorf("output (%d) bigger than max block size (%d)", s.seqSize, maxBlockSize)
}
err := br.close()
if err != nil {
printf("Closing sequences: %v, %+v\n", err, *br)
}
return err
}
// --------------------------------------------------------------------------------
type executeAsmContext struct {
seqs []seqVals
seqIndex int
out []byte
history []byte
literals []byte
outPosition int
litPosition int
windowSize int
}
// sequenceDecs_executeSimple_amd64 implements the main loop of sequenceDecs.executeSimple in x86 asm.
//
// Returns false if a match offset is too big.
//
// Please refer to seqdec_generic.go for the reference implementation.
//go:noescape
func sequenceDecs_executeSimple_amd64(ctx *executeAsmContext) bool
// executeSimple handles cases when dictionary is not used.
func (s *sequenceDecs) executeSimple(seqs []seqVals, hist []byte) error {
// Ensure we have enough output size...
if len(s.out)+s.seqSize+compressedBlockOverAlloc > cap(s.out) {
addBytes := s.seqSize + len(s.out) + compressedBlockOverAlloc
s.out = append(s.out, make([]byte, addBytes)...)
s.out = s.out[:len(s.out)-addBytes]
}
if debugDecoder {
printf("Execute %d seqs with literals: %d into %d bytes\n", len(seqs), len(s.literals), s.seqSize)
}
var t = len(s.out)
out := s.out[:t+s.seqSize]
ctx := executeAsmContext{
seqs: seqs,
seqIndex: 0,
out: out,
history: hist,
outPosition: t,
litPosition: 0,
literals: s.literals,
windowSize: s.windowSize,
}
ok := sequenceDecs_executeSimple_amd64(&ctx)
if !ok {
return fmt.Errorf("match offset (%d) bigger than current history (%d)",
seqs[ctx.seqIndex].mo, ctx.outPosition+len(hist))
}
s.literals = s.literals[ctx.litPosition:]
t = ctx.outPosition
// Add final literals
copy(out[t:], s.literals)
if debugDecoder {
t += len(s.literals)
if t != len(out) {
panic(fmt.Errorf("length mismatch, want %d, got %d, ss: %d", len(out), t, s.seqSize))
}
}
s.out = out
return nil
}

3519
vendor/github.com/klauspost/compress/zstd/seqdec_amd64.s generated vendored Normal file

File diff suppressed because it is too large Load Diff

View File

@@ -0,0 +1,237 @@
//go:build !amd64 || appengine || !gc || noasm
// +build !amd64 appengine !gc noasm
package zstd
import (
"fmt"
"io"
)
// decode sequences from the stream with the provided history but without dictionary.
func (s *sequenceDecs) decodeSyncSimple(hist []byte) (bool, error) {
return false, nil
}
// decode sequences from the stream without the provided history.
func (s *sequenceDecs) decode(seqs []seqVals) error {
br := s.br
// Grab full sizes tables, to avoid bounds checks.
llTable, mlTable, ofTable := s.litLengths.fse.dt[:maxTablesize], s.matchLengths.fse.dt[:maxTablesize], s.offsets.fse.dt[:maxTablesize]
llState, mlState, ofState := s.litLengths.state.state, s.matchLengths.state.state, s.offsets.state.state
s.seqSize = 0
litRemain := len(s.literals)
maxBlockSize := maxCompressedBlockSize
if s.windowSize < maxBlockSize {
maxBlockSize = s.windowSize
}
for i := range seqs {
var ll, mo, ml int
if br.off > 4+((maxOffsetBits+16+16)>>3) {
// inlined function:
// ll, mo, ml = s.nextFast(br, llState, mlState, ofState)
// Final will not read from stream.
var llB, mlB, moB uint8
ll, llB = llState.final()
ml, mlB = mlState.final()
mo, moB = ofState.final()
// extra bits are stored in reverse order.
br.fillFast()
mo += br.getBits(moB)
if s.maxBits > 32 {
br.fillFast()
}
ml += br.getBits(mlB)
ll += br.getBits(llB)
if moB > 1 {
s.prevOffset[2] = s.prevOffset[1]
s.prevOffset[1] = s.prevOffset[0]
s.prevOffset[0] = mo
} else {
// mo = s.adjustOffset(mo, ll, moB)
// Inlined for rather big speedup
if ll == 0 {
// There is an exception though, when current sequence's literals_length = 0.
// In this case, repeated offsets are shifted by one, so an offset_value of 1 means Repeated_Offset2,
// an offset_value of 2 means Repeated_Offset3, and an offset_value of 3 means Repeated_Offset1 - 1_byte.
mo++
}
if mo == 0 {
mo = s.prevOffset[0]
} else {
var temp int
if mo == 3 {
temp = s.prevOffset[0] - 1
} else {
temp = s.prevOffset[mo]
}
if temp == 0 {
// 0 is not valid; input is corrupted; force offset to 1
println("WARNING: temp was 0")
temp = 1
}
if mo != 1 {
s.prevOffset[2] = s.prevOffset[1]
}
s.prevOffset[1] = s.prevOffset[0]
s.prevOffset[0] = temp
mo = temp
}
}
br.fillFast()
} else {
if br.overread() {
if debugDecoder {
printf("reading sequence %d, exceeded available data\n", i)
}
return io.ErrUnexpectedEOF
}
ll, mo, ml = s.next(br, llState, mlState, ofState)
br.fill()
}
if debugSequences {
println("Seq", i, "Litlen:", ll, "mo:", mo, "(abs) ml:", ml)
}
// Evaluate.
// We might be doing this async, so do it early.
if mo == 0 && ml > 0 {
return fmt.Errorf("zero matchoff and matchlen (%d) > 0", ml)
}
if ml > maxMatchLen {
return fmt.Errorf("match len (%d) bigger than max allowed length", ml)
}
s.seqSize += ll + ml
if s.seqSize > maxBlockSize {
return fmt.Errorf("output (%d) bigger than max block size (%d)", s.seqSize, maxBlockSize)
}
litRemain -= ll
if litRemain < 0 {
return fmt.Errorf("unexpected literal count, want %d bytes, but only %d is available", ll, litRemain+ll)
}
seqs[i] = seqVals{
ll: ll,
ml: ml,
mo: mo,
}
if i == len(seqs)-1 {
// This is the last sequence, so we shouldn't update state.
break
}
// Manually inlined, ~ 5-20% faster
// Update all 3 states at once. Approx 20% faster.
nBits := llState.nbBits() + mlState.nbBits() + ofState.nbBits()
if nBits == 0 {
llState = llTable[llState.newState()&maxTableMask]
mlState = mlTable[mlState.newState()&maxTableMask]
ofState = ofTable[ofState.newState()&maxTableMask]
} else {
bits := br.get32BitsFast(nBits)
lowBits := uint16(bits >> ((ofState.nbBits() + mlState.nbBits()) & 31))
llState = llTable[(llState.newState()+lowBits)&maxTableMask]
lowBits = uint16(bits >> (ofState.nbBits() & 31))
lowBits &= bitMask[mlState.nbBits()&15]
mlState = mlTable[(mlState.newState()+lowBits)&maxTableMask]
lowBits = uint16(bits) & bitMask[ofState.nbBits()&15]
ofState = ofTable[(ofState.newState()+lowBits)&maxTableMask]
}
}
s.seqSize += litRemain
if s.seqSize > maxBlockSize {
return fmt.Errorf("output (%d) bigger than max block size (%d)", s.seqSize, maxBlockSize)
}
err := br.close()
if err != nil {
printf("Closing sequences: %v, %+v\n", err, *br)
}
return err
}
// executeSimple handles cases when a dictionary is not used.
func (s *sequenceDecs) executeSimple(seqs []seqVals, hist []byte) error {
// Ensure we have enough output size...
if len(s.out)+s.seqSize > cap(s.out) {
addBytes := s.seqSize + len(s.out)
s.out = append(s.out, make([]byte, addBytes)...)
s.out = s.out[:len(s.out)-addBytes]
}
if debugDecoder {
printf("Execute %d seqs with literals: %d into %d bytes\n", len(seqs), len(s.literals), s.seqSize)
}
var t = len(s.out)
out := s.out[:t+s.seqSize]
for _, seq := range seqs {
// Add literals
copy(out[t:], s.literals[:seq.ll])
t += seq.ll
s.literals = s.literals[seq.ll:]
// Malformed input
if seq.mo > t+len(hist) || seq.mo > s.windowSize {
return fmt.Errorf("match offset (%d) bigger than current history (%d)", seq.mo, t+len(hist))
}
// Copy from history.
if v := seq.mo - t; v > 0 {
// v is the start position in history from end.
start := len(hist) - v
if seq.ml > v {
// Some goes into the current block.
// Copy remainder of history
copy(out[t:], hist[start:])
t += v
seq.ml -= v
} else {
copy(out[t:], hist[start:start+seq.ml])
t += seq.ml
continue
}
}
// We must be in the current buffer now
if seq.ml > 0 {
start := t - seq.mo
if seq.ml <= t-start {
// No overlap
copy(out[t:], out[start:start+seq.ml])
t += seq.ml
} else {
// Overlapping copy
// Extend destination slice and copy one byte at the time.
src := out[start : start+seq.ml]
dst := out[t:]
dst = dst[:len(src)]
t += len(src)
// Destination is the space we just added.
for i := range src {
dst[i] = src[i]
}
}
}
}
// Add final literals
copy(out[t:], s.literals)
if debugDecoder {
t += len(s.literals)
if t != len(out) {
panic(fmt.Errorf("length mismatch, want %d, got %d, ss: %d", len(out), t, s.seqSize))
}
}
s.out = out
return nil
}

View File

@@ -21,23 +21,34 @@ const ZipMethodPKWare = 20
var zipReaderPool sync.Pool
// newZipReader creates a pooled zip decompressor.
func newZipReader(r io.Reader) io.ReadCloser {
dec, ok := zipReaderPool.Get().(*Decoder)
if ok {
dec.Reset(r)
} else {
d, err := NewReader(r, WithDecoderConcurrency(1), WithDecoderLowmem(true))
if err != nil {
panic(err)
}
dec = d
func newZipReader(opts ...DOption) func(r io.Reader) io.ReadCloser {
pool := &zipReaderPool
if len(opts) > 0 {
opts = append([]DOption{WithDecoderLowmem(true), WithDecoderMaxWindow(128 << 20)}, opts...)
// Force concurrency 1
opts = append(opts, WithDecoderConcurrency(1))
// Create our own pool
pool = &sync.Pool{}
}
return func(r io.Reader) io.ReadCloser {
dec, ok := pool.Get().(*Decoder)
if ok {
dec.Reset(r)
} else {
d, err := NewReader(r, opts...)
if err != nil {
panic(err)
}
dec = d
}
return &pooledZipReader{dec: dec, pool: pool}
}
return &pooledZipReader{dec: dec}
}
type pooledZipReader struct {
mu sync.Mutex // guards Close and Read
dec *Decoder
mu sync.Mutex // guards Close and Read
pool *sync.Pool
dec *Decoder
}
func (r *pooledZipReader) Read(p []byte) (n int, err error) {
@@ -48,8 +59,8 @@ func (r *pooledZipReader) Read(p []byte) (n int, err error) {
}
dec, err := r.dec.Read(p)
if err == io.EOF {
err = r.dec.Reset(nil)
zipReaderPool.Put(r.dec)
r.dec.Reset(nil)
r.pool.Put(r.dec)
r.dec = nil
}
return dec, err
@@ -61,7 +72,7 @@ func (r *pooledZipReader) Close() error {
var err error
if r.dec != nil {
err = r.dec.Reset(nil)
zipReaderPool.Put(r.dec)
r.pool.Put(r.dec)
r.dec = nil
}
return err
@@ -115,6 +126,9 @@ func ZipCompressor(opts ...EOption) func(w io.Writer) (io.WriteCloser, error) {
// ZipDecompressor returns a decompressor that can be registered with zip libraries.
// See ZipCompressor for example.
func ZipDecompressor() func(r io.Reader) io.ReadCloser {
return newZipReader
// Options can be specified. WithDecoderConcurrency(1) is forced,
// and by default a 128MB maximum decompression window is specified.
// The window size can be overridden if required.
func ZipDecompressor(opts ...DOption) func(r io.Reader) io.ReadCloser {
return newZipReader(opts...)
}