201 lines
6.6 KiB
Go
201 lines
6.6 KiB
Go
// Copyright 2015, Joe Tsai. All rights reserved.
|
|
// Use of this source code is governed by a BSD-style
|
|
// license that can be found in the LICENSE.md file.
|
|
|
|
package brotli
|
|
|
|
// The algorithm used to decode variable length codes is based on the lookup
|
|
// method in zlib. If the code is less-than-or-equal to prefixMaxChunkBits,
|
|
// then the symbol can be decoded using a single lookup into the chunks table.
|
|
// Otherwise, the links table will be used for a second level lookup.
|
|
//
|
|
// The chunks slice is keyed by the contents of the bit buffer ANDed with
|
|
// the chunkMask to avoid a out-of-bounds lookup. The value of chunks is a tuple
|
|
// that is decoded as follow:
|
|
//
|
|
// var length = chunks[bitBuffer&chunkMask] & prefixCountMask
|
|
// var symbol = chunks[bitBuffer&chunkMask] >> prefixCountBits
|
|
//
|
|
// If the decoded length is larger than chunkBits, then an overflow link table
|
|
// must be used for further decoding. In this case, the symbol is actually the
|
|
// index into the links tables. The second-level links table returned is
|
|
// processed in the same way as the chunks table.
|
|
//
|
|
// if length > chunkBits {
|
|
// var index = symbol // Previous symbol is index into links tables
|
|
// length = links[index][bitBuffer>>chunkBits & linkMask] & prefixCountMask
|
|
// symbol = links[index][bitBuffer>>chunkBits & linkMask] >> prefixCountBits
|
|
// }
|
|
//
|
|
// See the following:
|
|
// http://www.gzip.org/algorithm.txt
|
|
|
|
const (
|
|
// These values add up to the width of a uint32 integer.
|
|
prefixCountBits = 5 // Number of bits to store the bit-width of the code
|
|
prefixSymbolBits = 27 // Number of bits to store the symbol value
|
|
|
|
prefixCountMask = (1 << prefixCountBits) - 1
|
|
prefixMaxChunkBits = 9 // This can be tuned for better performance
|
|
)
|
|
|
|
type prefixDecoder struct {
|
|
chunks []uint32 // First-level lookup map
|
|
links [][]uint32 // Second-level lookup map
|
|
chunkMask uint32 // Mask the width of the chunks table
|
|
linkMask uint32 // Mask the width of the link table
|
|
chunkBits uint32 // Bit-width of the chunks table
|
|
minBits uint32 // The minimum number of bits to safely make progress
|
|
numSyms uint32 // Number of symbols
|
|
}
|
|
|
|
// Init initializes prefixDecoder according to the codes provided.
|
|
// The symbols provided must be unique and in ascending order.
|
|
//
|
|
// If assignCodes is true, then generate a canonical prefix tree using the
|
|
// prefixCode.len field and assign the generated value to prefixCode.val.
|
|
//
|
|
// If assignCodes is false, then initialize using the information inside the
|
|
// codes themselves. The input codes must form a valid prefix tree.
|
|
func (pd *prefixDecoder) Init(codes []prefixCode, assignCodes bool) {
|
|
// Handle special case trees.
|
|
if len(codes) <= 1 {
|
|
switch {
|
|
case len(codes) == 0: // Empty tree (should error if used later)
|
|
*pd = prefixDecoder{chunks: pd.chunks[:0], links: pd.links[:0], numSyms: 0}
|
|
case len(codes) == 1: // Single code tree (bit-width of zero)
|
|
*pd = prefixDecoder{
|
|
chunks: append(pd.chunks[:0], codes[0].sym<<prefixCountBits|0),
|
|
links: pd.links[:0],
|
|
numSyms: 1,
|
|
}
|
|
}
|
|
return
|
|
}
|
|
|
|
// Compute basic statistics on the symbols.
|
|
var bitCnts [maxPrefixBits + 1]uint
|
|
c0 := codes[0]
|
|
bitCnts[c0.len]++
|
|
minBits, maxBits, symLast := c0.len, c0.len, int(c0.sym)
|
|
for _, c := range codes[1:] {
|
|
if int(c.sym) <= symLast {
|
|
panic(ErrCorrupt) // Non-unique or non-monotonically increasing
|
|
}
|
|
if minBits > c.len {
|
|
minBits = c.len
|
|
}
|
|
if maxBits < c.len {
|
|
maxBits = c.len
|
|
}
|
|
bitCnts[c.len]++ // Histogram of bit counts
|
|
symLast = int(c.sym) // Keep track of last symbol
|
|
}
|
|
if maxBits >= 1<<prefixCountBits || minBits == 0 {
|
|
panic(ErrCorrupt) // Bit-width is too long or too short
|
|
}
|
|
if symLast >= 1<<prefixSymbolBits {
|
|
panic(ErrCorrupt) // Alphabet cardinality too large
|
|
}
|
|
|
|
// Compute the next code for a symbol of a given bit length.
|
|
var nextCodes [maxPrefixBits + 1]uint
|
|
var code uint
|
|
for i := minBits; i <= maxBits; i++ {
|
|
code <<= 1
|
|
nextCodes[i] = code
|
|
code += bitCnts[i]
|
|
}
|
|
if code != 1<<maxBits {
|
|
panic(ErrCorrupt) // Tree is under or over subscribed
|
|
}
|
|
|
|
// Allocate chunks table if necessary.
|
|
pd.numSyms = uint32(len(codes))
|
|
pd.minBits = minBits
|
|
pd.chunkBits = maxBits
|
|
if pd.chunkBits > prefixMaxChunkBits {
|
|
pd.chunkBits = prefixMaxChunkBits
|
|
}
|
|
numChunks := 1 << pd.chunkBits
|
|
pd.chunks = allocUint32s(pd.chunks, numChunks)
|
|
pd.chunkMask = uint32(numChunks - 1)
|
|
|
|
// Allocate links tables if necessary.
|
|
pd.links = pd.links[:0]
|
|
pd.linkMask = 0
|
|
if pd.chunkBits < maxBits {
|
|
numLinks := 1 << (maxBits - pd.chunkBits)
|
|
pd.linkMask = uint32(numLinks - 1)
|
|
|
|
if assignCodes {
|
|
baseCode := nextCodes[pd.chunkBits+1] >> 1
|
|
pd.links = extendSliceUints32s(pd.links, numChunks-int(baseCode))
|
|
for linkIdx := range pd.links {
|
|
code := reverseBits(uint32(baseCode)+uint32(linkIdx), uint(pd.chunkBits))
|
|
pd.links[linkIdx] = allocUint32s(pd.links[linkIdx], numLinks)
|
|
pd.chunks[code] = uint32(linkIdx<<prefixCountBits) | uint32(pd.chunkBits+1)
|
|
}
|
|
} else {
|
|
for i := range pd.chunks {
|
|
pd.chunks[i] = 0 // Logic below relies zero value as uninitialized
|
|
}
|
|
for _, c := range codes {
|
|
if c.len <= pd.chunkBits {
|
|
continue // Ignore symbols that don't require links
|
|
}
|
|
code := c.val & pd.chunkMask
|
|
if pd.chunks[code] > 0 {
|
|
continue // Link table already initialized
|
|
}
|
|
linkIdx := len(pd.links)
|
|
pd.links = extendSliceUints32s(pd.links, len(pd.links)+1)
|
|
pd.links[linkIdx] = allocUint32s(pd.links[linkIdx], numLinks)
|
|
pd.chunks[code] = uint32(linkIdx<<prefixCountBits) | uint32(pd.chunkBits+1)
|
|
}
|
|
}
|
|
}
|
|
|
|
// Fill out chunks and links tables with values.
|
|
for i, c := range codes {
|
|
chunk := c.sym<<prefixCountBits | uint32(c.len)
|
|
if assignCodes {
|
|
codes[i].val = reverseBits(uint32(nextCodes[c.len]), uint(c.len))
|
|
nextCodes[c.len]++
|
|
c = codes[i]
|
|
}
|
|
|
|
if c.len <= pd.chunkBits {
|
|
skip := 1 << uint(c.len)
|
|
for j := int(c.val); j < len(pd.chunks); j += skip {
|
|
pd.chunks[j] = chunk
|
|
}
|
|
} else {
|
|
linkIdx := pd.chunks[c.val&pd.chunkMask] >> prefixCountBits
|
|
links := pd.links[linkIdx]
|
|
skip := 1 << uint(c.len-pd.chunkBits)
|
|
for j := int(c.val >> pd.chunkBits); j < len(links); j += skip {
|
|
links[j] = chunk
|
|
}
|
|
}
|
|
}
|
|
|
|
if debug && !checkPrefixes(codes) {
|
|
panic(ErrCorrupt) // The codes do not form a valid prefix tree.
|
|
}
|
|
}
|
|
|
|
// checkPrefixes reports whether any codes have overlapping prefixes.
|
|
// This check is expensive and runs in O(n^2) time!
|
|
func checkPrefixes(codes []prefixCode) bool {
|
|
for i, c1 := range codes {
|
|
for j, c2 := range codes {
|
|
mask := uint32(1)<<c1.len - 1
|
|
if i != j && c1.len <= c2.len && c1.val&mask == c2.val&mask {
|
|
return false
|
|
}
|
|
}
|
|
}
|
|
return true
|
|
}
|