Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

inflate: 10-15% faster decompression #293

Merged
merged 2 commits into from
Nov 10, 2020
Merged

inflate: 10-15% faster decompression #293

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
325e0f9
inflate: 10-15% faster decompression
klauspost Nov 9, 2020
d981cb7
Correct comment
klauspost Nov 9, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
114 changes: 66 additions & 48 deletions flate/gen_inflate.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,16 +42,6 @@ func (f *decompressor) $FUNCNAME$() {
stateDict
)
fr := f.r.($TYPE$)
moreBits := func() error {
c, err := fr.ReadByte()
if err != nil {
return noEOF(err)
}
f.roffset++
f.b |= uint32(c) << f.nb
f.nb += 8
return nil
}

switch f.stepState {
case stateInit:
Expand Down Expand Up @@ -112,9 +102,7 @@ readLiteral:
}
}

var n uint // number of bits extra
var length int
var err error
switch {
case v < 256:
f.dict.writeByte(byte(v))
Expand All @@ -131,71 +119,97 @@ readLiteral:
// otherwise, reference to older data
case v < 265:
length = v - (257 - 3)
n = 0
case v < 269:
length = v*2 - (265*2 - 11)
n = 1
case v < 273:
length = v*4 - (269*4 - 19)
n = 2
case v < 277:
length = v*8 - (273*8 - 35)
n = 3
case v < 281:
length = v*16 - (277*16 - 67)
n = 4
case v < 285:
length = v*32 - (281*32 - 131)
n = 5
case v < maxNumLit:
length = 258
n = 0
default:
if debugDecode {
fmt.Println(v, ">= maxNumLit")
}
f.err = CorruptInputError(f.roffset)
return
}
if n > 0 {
val := decCodeToLen[(v - 257)]
length = int(val.length) + 3
n := uint(val.extra)
for f.nb < n {
if err = moreBits(); err != nil {
c, err := fr.ReadByte()
if err != nil {
if debugDecode {
fmt.Println("morebits n>0:", err)
}
f.err = err
return
}
f.roffset++
f.b |= uint32(c) << f.nb
f.nb += 8
}
length += int(f.b & uint32(1<<(n&regSizeMaskUint32)-1))
f.b >>= n & regSizeMaskUint32
f.nb -= n
default:
if debugDecode {
fmt.Println(v, ">= maxNumLit")
}
f.err = CorruptInputError(f.roffset)
return
}

var dist uint32
if f.hd == nil {
for f.nb < 5 {
if err = f.moreBits(); err != nil {
c, err := fr.ReadByte()
if err != nil {
if debugDecode {
fmt.Println("morebits f.nb<5:", err)
}
f.err = err
return
}
f.roffset++
f.b |= uint32(c) << f.nb
f.nb += 8
}
dist = uint32(bits.Reverse8(uint8(f.b & 0x1F << 3)))
f.b >>= 5
f.nb -= 5
} else {
sym, err := f.huffSym(f.hd)
if err != nil {
if debugDecode {
fmt.Println("huffsym:", err)
// 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(f.hd.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 := fr.ReadByte()
if err != nil {
f.b = b
f.nb = nb
f.err = noEOF(err)
return
}
f.roffset++
b |= uint32(c) << (nb & regSizeMaskUint32)
nb += 8
}
chunk := f.hd.chunks[b&(huffmanNumChunks-1)]
n = uint(chunk & huffmanCountMask)
if n > huffmanChunkBits {
chunk = f.hd.links[chunk>>huffmanValueShift][(b>>huffmanChunkBits)&f.hd.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
}
f.b = b >> (n & regSizeMaskUint32)
f.nb = nb - n
dist = uint32(chunk >> huffmanValueShift)
break
}
f.err = err
return
}
dist = uint32(sym)
}

switch {
Expand All @@ -206,13 +220,17 @@ readLiteral:
// have 1 bit in bottom of dist, need nb more.
extra := (dist & 1) << (nb & regSizeMaskUint32)
for f.nb < nb {
if err = f.moreBits(); err != nil {
c, err := fr.ReadByte()
if err != nil {
if debugDecode {
fmt.Println("morebits f.nb<nb:", err)
}
f.err = err
return
}
f.roffset++
f.b |= uint32(c) << f.nb
f.nb += 8
}
extra |= f.b & uint32(1<<(nb&regSizeMaskUint32)-1)
f.b >>= nb & regSizeMaskUint32
Expand Down
7 changes: 7 additions & 0 deletions flate/inflate.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,6 +29,13 @@ const (
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}}

// Initialize the fixedHuffmanDecoder only once upon first use.
var fixedOnce sync.Once
var fixedHuffmanDecoder huffmanDecoder
Expand Down
Loading