diff --git a/internal/github.com/klauspost/compress/LICENSE b/internal/github.com/klauspost/compress/LICENSE new file mode 100644 index 000000000..1eb75ef68 --- /dev/null +++ b/internal/github.com/klauspost/compress/LICENSE @@ -0,0 +1,28 @@ +Copyright (c) 2012 The Go Authors. All rights reserved. +Copyright (c) 2019 Klaus Post. All rights reserved. + +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are +met: + + * Redistributions of source code must retain the above copyright +notice, this list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above +copyright notice, this list of conditions and the following disclaimer +in the documentation and/or other materials provided with the +distribution. + * Neither the name of Google Inc. nor the names of its +contributors may be used to endorse or promote products derived from +this software without specific prior written permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. diff --git a/internal/github.com/klauspost/compress/README.md b/internal/github.com/klauspost/compress/README.md new file mode 100644 index 000000000..a50360c83 --- /dev/null +++ b/internal/github.com/klauspost/compress/README.md @@ -0,0 +1,9 @@ +This source code is a stripped down version of zstd from the https://github.com/klauspost/compress/tree/517288e9a6e1dd4dea10ad42ffe2829c58dadf51/zstd. + +Motivation: https://github.com/kubernetes/kubernetes/pull/130569#discussion_r1981503174 + +Changes: +* Remove all but things necessary to use and create zstd.NewWriter for SpeedFastest mode. +* Use github.com/cespare/xxhash/v2 instead of vendored copy. + +The goal is to remove this once stdlib will support zstd. diff --git a/internal/github.com/klauspost/compress/fse/bitreader.go b/internal/github.com/klauspost/compress/fse/bitreader.go new file mode 100644 index 000000000..f65eb3909 --- /dev/null +++ b/internal/github.com/klauspost/compress/fse/bitreader.go @@ -0,0 +1,122 @@ +// Copyright 2018 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. +// Based on work Copyright (c) 2013, Yann Collet, released under BSD License. + +package fse + +import ( + "encoding/binary" + "errors" + "io" +) + +// bitReader reads a bitstream in reverse. +// The last set bit indicates the start of the stream and is used +// for aligning the input. +type bitReader struct { + in []byte + off uint // next byte to read is at in[off - 1] + value uint64 + bitsRead uint8 +} + +// init initializes and resets the bit reader. +func (b *bitReader) init(in []byte) error { + if len(in) < 1 { + return errors.New("corrupt stream: too short") + } + b.in = in + b.off = uint(len(in)) + // The highest bit of the last byte indicates where to start + v := in[len(in)-1] + if v == 0 { + return errors.New("corrupt stream, did not find end of stream") + } + b.bitsRead = 64 + b.value = 0 + if len(in) >= 8 { + b.fillFastStart() + } else { + b.fill() + b.fill() + } + b.bitsRead += 8 - uint8(highBits(uint32(v))) + return nil +} + +// getBits will return n bits. n can be 0. +func (b *bitReader) getBits(n uint8) uint16 { + if n == 0 || b.bitsRead >= 64 { + return 0 + } + return b.getBitsFast(n) +} + +// getBitsFast requires that at least one bit is requested every time. +// There are no checks if the buffer is filled. +func (b *bitReader) getBitsFast(n uint8) uint16 { + const regMask = 64 - 1 + v := uint16((b.value << (b.bitsRead & regMask)) >> ((regMask + 1 - n) & regMask)) + b.bitsRead += n + return v +} + +// fillFast() will make sure at least 32 bits are available. +// There must be at least 4 bytes available. +func (b *bitReader) fillFast() { + if b.bitsRead < 32 { + return + } + // 2 bounds checks. + v := b.in[b.off-4:] + v = v[:4] + low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24) + b.value = (b.value << 32) | uint64(low) + b.bitsRead -= 32 + b.off -= 4 +} + +// fill() will make sure at least 32 bits are available. +func (b *bitReader) fill() { + if b.bitsRead < 32 { + return + } + if b.off > 4 { + v := b.in[b.off-4:] + v = v[:4] + low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24) + b.value = (b.value << 32) | uint64(low) + b.bitsRead -= 32 + b.off -= 4 + return + } + for b.off > 0 { + b.value = (b.value << 8) | uint64(b.in[b.off-1]) + b.bitsRead -= 8 + b.off-- + } +} + +// fillFastStart() assumes the bitreader is empty and there is at least 8 bytes to read. +func (b *bitReader) fillFastStart() { + // Do single re-slice to avoid bounds checks. + b.value = binary.LittleEndian.Uint64(b.in[b.off-8:]) + b.bitsRead = 0 + b.off -= 8 +} + +// finished returns true if all bits have been read from the bit stream. +func (b *bitReader) finished() bool { + return b.bitsRead >= 64 && b.off == 0 +} + +// close the bitstream and returns an error if out-of-buffer reads occurred. +func (b *bitReader) close() error { + // Release reference. + b.in = nil + if b.bitsRead > 64 { + return io.ErrUnexpectedEOF + } + return nil +} diff --git a/internal/github.com/klauspost/compress/fse/bitwriter.go b/internal/github.com/klauspost/compress/fse/bitwriter.go new file mode 100644 index 000000000..e82fa3bb7 --- /dev/null +++ b/internal/github.com/klauspost/compress/fse/bitwriter.go @@ -0,0 +1,167 @@ +// Copyright 2018 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. +// Based on work Copyright (c) 2013, Yann Collet, released under BSD License. + +package fse + +import "fmt" + +// bitWriter will write bits. +// First bit will be LSB of the first byte of output. +type bitWriter struct { + bitContainer uint64 + nBits uint8 + out []byte +} + +// bitMask16 is bitmasks. Has extra to avoid bounds check. +var bitMask16 = [32]uint16{ + 0, 1, 3, 7, 0xF, 0x1F, + 0x3F, 0x7F, 0xFF, 0x1FF, 0x3FF, 0x7FF, + 0xFFF, 0x1FFF, 0x3FFF, 0x7FFF, 0xFFFF, 0xFFFF, + 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, + 0xFFFF, 0xFFFF} /* up to 16 bits */ + +// addBits16NC will add up to 16 bits. +// It will not check if there is space for them, +// so the caller must ensure that it has flushed recently. +func (b *bitWriter) addBits16NC(value uint16, bits uint8) { + b.bitContainer |= uint64(value&bitMask16[bits&31]) << (b.nBits & 63) + b.nBits += bits +} + +// addBits16Clean will add up to 16 bits. value may not contain more set bits than indicated. +// It will not check if there is space for them, so the caller must ensure that it has flushed recently. +func (b *bitWriter) addBits16Clean(value uint16, bits uint8) { + b.bitContainer |= uint64(value) << (b.nBits & 63) + b.nBits += bits +} + +// addBits16ZeroNC will add up to 16 bits. +// It will not check if there is space for them, +// so the caller must ensure that it has flushed recently. +// This is fastest if bits can be zero. +func (b *bitWriter) addBits16ZeroNC(value uint16, bits uint8) { + if bits == 0 { + return + } + value <<= (16 - bits) & 15 + value >>= (16 - bits) & 15 + b.bitContainer |= uint64(value) << (b.nBits & 63) + b.nBits += bits +} + +// flush will flush all pending full bytes. +// There will be at least 56 bits available for writing when this has been called. +// Using flush32 is faster, but leaves less space for writing. +func (b *bitWriter) flush() { + v := b.nBits >> 3 + switch v { + case 0: + case 1: + b.out = append(b.out, + byte(b.bitContainer), + ) + case 2: + b.out = append(b.out, + byte(b.bitContainer), + byte(b.bitContainer>>8), + ) + case 3: + b.out = append(b.out, + byte(b.bitContainer), + byte(b.bitContainer>>8), + byte(b.bitContainer>>16), + ) + case 4: + b.out = append(b.out, + byte(b.bitContainer), + byte(b.bitContainer>>8), + byte(b.bitContainer>>16), + byte(b.bitContainer>>24), + ) + case 5: + b.out = append(b.out, + byte(b.bitContainer), + byte(b.bitContainer>>8), + byte(b.bitContainer>>16), + byte(b.bitContainer>>24), + byte(b.bitContainer>>32), + ) + case 6: + b.out = append(b.out, + byte(b.bitContainer), + byte(b.bitContainer>>8), + byte(b.bitContainer>>16), + byte(b.bitContainer>>24), + byte(b.bitContainer>>32), + byte(b.bitContainer>>40), + ) + case 7: + b.out = append(b.out, + byte(b.bitContainer), + byte(b.bitContainer>>8), + byte(b.bitContainer>>16), + byte(b.bitContainer>>24), + byte(b.bitContainer>>32), + byte(b.bitContainer>>40), + byte(b.bitContainer>>48), + ) + case 8: + b.out = append(b.out, + byte(b.bitContainer), + byte(b.bitContainer>>8), + byte(b.bitContainer>>16), + byte(b.bitContainer>>24), + byte(b.bitContainer>>32), + byte(b.bitContainer>>40), + byte(b.bitContainer>>48), + byte(b.bitContainer>>56), + ) + default: + panic(fmt.Errorf("bits (%d) > 64", b.nBits)) + } + b.bitContainer >>= v << 3 + b.nBits &= 7 +} + +// flush32 will flush out, so there are at least 32 bits available for writing. +func (b *bitWriter) flush32() { + if b.nBits < 32 { + return + } + b.out = append(b.out, + byte(b.bitContainer), + byte(b.bitContainer>>8), + byte(b.bitContainer>>16), + byte(b.bitContainer>>24)) + b.nBits -= 32 + b.bitContainer >>= 32 +} + +// flushAlign will flush remaining full bytes and align to next byte boundary. +func (b *bitWriter) flushAlign() { + nbBytes := (b.nBits + 7) >> 3 + for i := uint8(0); i < nbBytes; i++ { + b.out = append(b.out, byte(b.bitContainer>>(i*8))) + } + b.nBits = 0 + b.bitContainer = 0 +} + +// close will write the alignment bit and write the final byte(s) +// to the output. +func (b *bitWriter) close() { + // End mark + b.addBits16Clean(1, 1) + // flush until next byte. + b.flushAlign() +} + +// reset and continue writing by appending to out. +func (b *bitWriter) reset(out []byte) { + b.bitContainer = 0 + b.nBits = 0 + b.out = out +} diff --git a/internal/github.com/klauspost/compress/fse/bytereader.go b/internal/github.com/klauspost/compress/fse/bytereader.go new file mode 100644 index 000000000..abade2d60 --- /dev/null +++ b/internal/github.com/klauspost/compress/fse/bytereader.go @@ -0,0 +1,47 @@ +// Copyright 2018 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. +// Based on work Copyright (c) 2013, Yann Collet, released under BSD License. + +package fse + +// byteReader provides a byte reader that reads +// little endian values from a byte stream. +// The input stream is manually advanced. +// The reader performs no bounds checks. +type byteReader struct { + b []byte + off int +} + +// init will initialize the reader and set the input. +func (b *byteReader) init(in []byte) { + b.b = in + b.off = 0 +} + +// advance the stream b n bytes. +func (b *byteReader) advance(n uint) { + b.off += int(n) +} + +// Uint32 returns a little endian uint32 starting at current offset. +func (b byteReader) Uint32() uint32 { + b2 := b.b[b.off:] + b2 = b2[:4] + v3 := uint32(b2[3]) + v2 := uint32(b2[2]) + v1 := uint32(b2[1]) + v0 := uint32(b2[0]) + return v0 | (v1 << 8) | (v2 << 16) | (v3 << 24) +} + +// unread returns the unread portion of the input. +func (b byteReader) unread() []byte { + return b.b[b.off:] +} + +// remain will return the number of bytes remaining. +func (b byteReader) remain() int { + return len(b.b) - b.off +} diff --git a/internal/github.com/klauspost/compress/fse/compress.go b/internal/github.com/klauspost/compress/fse/compress.go new file mode 100644 index 000000000..26e36a3ba --- /dev/null +++ b/internal/github.com/klauspost/compress/fse/compress.go @@ -0,0 +1,687 @@ +// Copyright 2018 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. +// Based on work Copyright (c) 2013, Yann Collet, released under BSD License. + +package fse + +import ( + "errors" + "fmt" +) + +const ( + tablelogAbsoluteMax = 15 +) + +// Compress the input bytes. Input must be < 2GB. +// Provide a Scratch buffer to avoid memory allocations. +// Note that the output is also kept in the scratch buffer. +// If input is too hard to compress, ErrIncompressible is returned. +// If input is a single byte value repeated ErrUseRLE is returned. +func Compress(in []byte, s *Scratch) ([]byte, error) { + if len(in) <= 1 { + return nil, ErrIncompressible + } + if len(in) > (2<<30)-1 { + return nil, errors.New("input too big, must be < 2GB") + } + s, err := s.prepare(in) + if err != nil { + return nil, err + } + + // Create histogram, if none was provided. + maxCount := s.maxCount + if maxCount == 0 { + maxCount = s.countSimple(in) + } + // Reset for next run. + s.clearCount = true + s.maxCount = 0 + if maxCount == len(in) { + // One symbol, use RLE + return nil, ErrUseRLE + } + if maxCount == 1 || maxCount < (len(in)>>7) { + // Each symbol present maximum once or too well distributed. + return nil, ErrIncompressible + } + s.optimalTableLog() + err = s.normalizeCount() + if err != nil { + return nil, err + } + err = s.writeCount() + if err != nil { + return nil, err + } + + if false { + err = s.validateNorm() + if err != nil { + return nil, err + } + } + + err = s.buildCTable() + if err != nil { + return nil, err + } + err = s.compress(in) + if err != nil { + return nil, err + } + s.Out = s.bw.out + // Check if we compressed. + if len(s.Out) >= len(in) { + return nil, ErrIncompressible + } + return s.Out, nil +} + +// cState contains the compression state of a stream. +type cState struct { + bw *bitWriter + stateTable []uint16 + state uint16 +} + +// init will initialize the compression state to the first symbol of the stream. +func (c *cState) init(bw *bitWriter, ct *cTable, tableLog uint8, first symbolTransform) { + c.bw = bw + c.stateTable = ct.stateTable + + nbBitsOut := (first.deltaNbBits + (1 << 15)) >> 16 + im := int32((nbBitsOut << 16) - first.deltaNbBits) + lu := (im >> nbBitsOut) + first.deltaFindState + c.state = c.stateTable[lu] +} + +// encode the output symbol provided and write it to the bitstream. +func (c *cState) encode(symbolTT symbolTransform) { + nbBitsOut := (uint32(c.state) + symbolTT.deltaNbBits) >> 16 + dstState := int32(c.state>>(nbBitsOut&15)) + symbolTT.deltaFindState + c.bw.addBits16NC(c.state, uint8(nbBitsOut)) + c.state = c.stateTable[dstState] +} + +// encode the output symbol provided and write it to the bitstream. +func (c *cState) encodeZero(symbolTT symbolTransform) { + nbBitsOut := (uint32(c.state) + symbolTT.deltaNbBits) >> 16 + dstState := int32(c.state>>(nbBitsOut&15)) + symbolTT.deltaFindState + c.bw.addBits16ZeroNC(c.state, uint8(nbBitsOut)) + c.state = c.stateTable[dstState] +} + +// flush will write the tablelog to the output and flush the remaining full bytes. +func (c *cState) flush(tableLog uint8) { + c.bw.flush32() + c.bw.addBits16NC(c.state, tableLog) + c.bw.flush() +} + +// compress is the main compression loop that will encode the input from the last byte to the first. +func (s *Scratch) compress(src []byte) error { + if len(src) <= 2 { + return errors.New("compress: src too small") + } + tt := s.ct.symbolTT[:256] + s.bw.reset(s.Out) + + // Our two states each encodes every second byte. + // Last byte encoded (first byte decoded) will always be encoded by c1. + var c1, c2 cState + + // Encode so remaining size is divisible by 4. + ip := len(src) + if ip&1 == 1 { + c1.init(&s.bw, &s.ct, s.actualTableLog, tt[src[ip-1]]) + c2.init(&s.bw, &s.ct, s.actualTableLog, tt[src[ip-2]]) + c1.encodeZero(tt[src[ip-3]]) + ip -= 3 + } else { + c2.init(&s.bw, &s.ct, s.actualTableLog, tt[src[ip-1]]) + c1.init(&s.bw, &s.ct, s.actualTableLog, tt[src[ip-2]]) + ip -= 2 + } + if ip&2 != 0 { + c2.encodeZero(tt[src[ip-1]]) + c1.encodeZero(tt[src[ip-2]]) + ip -= 2 + } + src = src[:ip] + + // Main compression loop. + switch { + case !s.zeroBits && s.actualTableLog <= 8: + // We can encode 4 symbols without requiring a flush. + // We do not need to check if any output is 0 bits. + for ; len(src) >= 4; src = src[:len(src)-4] { + s.bw.flush32() + v3, v2, v1, v0 := src[len(src)-4], src[len(src)-3], src[len(src)-2], src[len(src)-1] + c2.encode(tt[v0]) + c1.encode(tt[v1]) + c2.encode(tt[v2]) + c1.encode(tt[v3]) + } + case !s.zeroBits: + // We do not need to check if any output is 0 bits. + for ; len(src) >= 4; src = src[:len(src)-4] { + s.bw.flush32() + v3, v2, v1, v0 := src[len(src)-4], src[len(src)-3], src[len(src)-2], src[len(src)-1] + c2.encode(tt[v0]) + c1.encode(tt[v1]) + s.bw.flush32() + c2.encode(tt[v2]) + c1.encode(tt[v3]) + } + case s.actualTableLog <= 8: + // We can encode 4 symbols without requiring a flush + for ; len(src) >= 4; src = src[:len(src)-4] { + s.bw.flush32() + v3, v2, v1, v0 := src[len(src)-4], src[len(src)-3], src[len(src)-2], src[len(src)-1] + c2.encodeZero(tt[v0]) + c1.encodeZero(tt[v1]) + c2.encodeZero(tt[v2]) + c1.encodeZero(tt[v3]) + } + default: + for ; len(src) >= 4; src = src[:len(src)-4] { + s.bw.flush32() + v3, v2, v1, v0 := src[len(src)-4], src[len(src)-3], src[len(src)-2], src[len(src)-1] + c2.encodeZero(tt[v0]) + c1.encodeZero(tt[v1]) + s.bw.flush32() + c2.encodeZero(tt[v2]) + c1.encodeZero(tt[v3]) + } + } + + // Flush final state. + // Used to initialize state when decoding. + c2.flush(s.actualTableLog) + c1.flush(s.actualTableLog) + + s.bw.close() + return nil +} + +// writeCount will write the normalized histogram count to header. +// This is read back by readNCount. +func (s *Scratch) writeCount() error { + var ( + tableLog = s.actualTableLog + tableSize = 1 << tableLog + previous0 bool + charnum uint16 + + maxHeaderSize = ((int(s.symbolLen)*int(tableLog) + 4 + 2) >> 3) + 3 + + // Write Table Size + bitStream = uint32(tableLog - minTablelog) + bitCount = uint(4) + remaining = int16(tableSize + 1) /* +1 for extra accuracy */ + threshold = int16(tableSize) + nbBits = uint(tableLog + 1) + ) + if cap(s.Out) < maxHeaderSize { + s.Out = make([]byte, 0, s.br.remain()+maxHeaderSize) + } + outP := uint(0) + out := s.Out[:maxHeaderSize] + + // stops at 1 + for remaining > 1 { + if previous0 { + start := charnum + for s.norm[charnum] == 0 { + charnum++ + } + for charnum >= start+24 { + start += 24 + bitStream += uint32(0xFFFF) << bitCount + out[outP] = byte(bitStream) + out[outP+1] = byte(bitStream >> 8) + outP += 2 + bitStream >>= 16 + } + for charnum >= start+3 { + start += 3 + bitStream += 3 << bitCount + bitCount += 2 + } + bitStream += uint32(charnum-start) << bitCount + bitCount += 2 + if bitCount > 16 { + out[outP] = byte(bitStream) + out[outP+1] = byte(bitStream >> 8) + outP += 2 + bitStream >>= 16 + bitCount -= 16 + } + } + + count := s.norm[charnum] + charnum++ + max := (2*threshold - 1) - remaining + if count < 0 { + remaining += count + } else { + remaining -= count + } + count++ // +1 for extra accuracy + if count >= threshold { + count += max // [0..max[ [max..threshold[ (...) [threshold+max 2*threshold[ + } + bitStream += uint32(count) << bitCount + bitCount += nbBits + if count < max { + bitCount-- + } + + previous0 = count == 1 + if remaining < 1 { + return errors.New("internal error: remaining<1") + } + for remaining < threshold { + nbBits-- + threshold >>= 1 + } + + if bitCount > 16 { + out[outP] = byte(bitStream) + out[outP+1] = byte(bitStream >> 8) + outP += 2 + bitStream >>= 16 + bitCount -= 16 + } + } + + out[outP] = byte(bitStream) + out[outP+1] = byte(bitStream >> 8) + outP += (bitCount + 7) / 8 + + if charnum > s.symbolLen { + return errors.New("internal error: charnum > s.symbolLen") + } + s.Out = out[:outP] + return nil +} + +// symbolTransform contains the state transform for a symbol. +type symbolTransform struct { + deltaFindState int32 + deltaNbBits uint32 +} + +// String prints values as a human readable string. +func (s symbolTransform) String() string { + return fmt.Sprintf("dnbits: %08x, fs:%d", s.deltaNbBits, s.deltaFindState) +} + +// cTable contains tables used for compression. +type cTable struct { + tableSymbol []byte + stateTable []uint16 + symbolTT []symbolTransform +} + +// allocCtable will allocate tables needed for compression. +// If existing tables a re big enough, they are simply re-used. +func (s *Scratch) allocCtable() { + tableSize := 1 << s.actualTableLog + // get tableSymbol that is big enough. + if cap(s.ct.tableSymbol) < tableSize { + s.ct.tableSymbol = make([]byte, tableSize) + } + s.ct.tableSymbol = s.ct.tableSymbol[:tableSize] + + ctSize := tableSize + if cap(s.ct.stateTable) < ctSize { + s.ct.stateTable = make([]uint16, ctSize) + } + s.ct.stateTable = s.ct.stateTable[:ctSize] + + if cap(s.ct.symbolTT) < 256 { + s.ct.symbolTT = make([]symbolTransform, 256) + } + s.ct.symbolTT = s.ct.symbolTT[:256] +} + +// buildCTable will populate the compression table so it is ready to be used. +func (s *Scratch) buildCTable() error { + tableSize := uint32(1 << s.actualTableLog) + highThreshold := tableSize - 1 + var cumul [maxSymbolValue + 2]int16 + + s.allocCtable() + tableSymbol := s.ct.tableSymbol[:tableSize] + // symbol start positions + { + cumul[0] = 0 + for ui, v := range s.norm[:s.symbolLen-1] { + u := byte(ui) // one less than reference + if v == -1 { + // Low proba symbol + cumul[u+1] = cumul[u] + 1 + tableSymbol[highThreshold] = u + highThreshold-- + } else { + cumul[u+1] = cumul[u] + v + } + } + // Encode last symbol separately to avoid overflowing u + u := int(s.symbolLen - 1) + v := s.norm[s.symbolLen-1] + if v == -1 { + // Low proba symbol + cumul[u+1] = cumul[u] + 1 + tableSymbol[highThreshold] = byte(u) + highThreshold-- + } else { + cumul[u+1] = cumul[u] + v + } + if uint32(cumul[s.symbolLen]) != tableSize { + return fmt.Errorf("internal error: expected cumul[s.symbolLen] (%d) == tableSize (%d)", cumul[s.symbolLen], tableSize) + } + cumul[s.symbolLen] = int16(tableSize) + 1 + } + // Spread symbols + s.zeroBits = false + { + step := tableStep(tableSize) + tableMask := tableSize - 1 + var position uint32 + // if any symbol > largeLimit, we may have 0 bits output. + largeLimit := int16(1 << (s.actualTableLog - 1)) + for ui, v := range s.norm[:s.symbolLen] { + symbol := byte(ui) + if v > largeLimit { + s.zeroBits = true + } + for nbOccurrences := int16(0); nbOccurrences < v; nbOccurrences++ { + tableSymbol[position] = symbol + position = (position + step) & tableMask + for position > highThreshold { + position = (position + step) & tableMask + } /* Low proba area */ + } + } + + // Check if we have gone through all positions + if position != 0 { + return errors.New("position!=0") + } + } + + // Build table + table := s.ct.stateTable + { + tsi := int(tableSize) + for u, v := range tableSymbol { + // TableU16 : sorted by symbol order; gives next state value + table[cumul[v]] = uint16(tsi + u) + cumul[v]++ + } + } + + // Build Symbol Transformation Table + { + total := int16(0) + symbolTT := s.ct.symbolTT[:s.symbolLen] + tableLog := s.actualTableLog + tl := (uint32(tableLog) << 16) - (1 << tableLog) + for i, v := range s.norm[:s.symbolLen] { + switch v { + case 0: + case -1, 1: + symbolTT[i].deltaNbBits = tl + symbolTT[i].deltaFindState = int32(total - 1) + total++ + default: + maxBitsOut := uint32(tableLog) - highBits(uint32(v-1)) + minStatePlus := uint32(v) << maxBitsOut + symbolTT[i].deltaNbBits = (maxBitsOut << 16) - minStatePlus + symbolTT[i].deltaFindState = int32(total - v) + total += v + } + } + if total != int16(tableSize) { + return fmt.Errorf("total mismatch %d (got) != %d (want)", total, tableSize) + } + } + return nil +} + +// countSimple will create a simple histogram in s.count. +// Returns the biggest count. +// Does not update s.clearCount. +func (s *Scratch) countSimple(in []byte) (max int) { + for _, v := range in { + s.count[v]++ + } + m, symlen := uint32(0), s.symbolLen + for i, v := range s.count[:] { + if v == 0 { + continue + } + if v > m { + m = v + } + symlen = uint16(i) + 1 + } + s.symbolLen = symlen + return int(m) +} + +// minTableLog provides the minimum logSize to safely represent a distribution. +func (s *Scratch) minTableLog() uint8 { + minBitsSrc := highBits(uint32(s.br.remain()-1)) + 1 + minBitsSymbols := highBits(uint32(s.symbolLen-1)) + 2 + if minBitsSrc < minBitsSymbols { + return uint8(minBitsSrc) + } + return uint8(minBitsSymbols) +} + +// optimalTableLog calculates and sets the optimal tableLog in s.actualTableLog +func (s *Scratch) optimalTableLog() { + tableLog := s.TableLog + minBits := s.minTableLog() + maxBitsSrc := uint8(highBits(uint32(s.br.remain()-1))) - 2 + if maxBitsSrc < tableLog { + // Accuracy can be reduced + tableLog = maxBitsSrc + } + if minBits > tableLog { + tableLog = minBits + } + // Need a minimum to safely represent all symbol values + if tableLog < minTablelog { + tableLog = minTablelog + } + if tableLog > maxTableLog { + tableLog = maxTableLog + } + s.actualTableLog = tableLog +} + +var rtbTable = [...]uint32{0, 473195, 504333, 520860, 550000, 700000, 750000, 830000} + +// normalizeCount will normalize the count of the symbols so +// the total is equal to the table size. +func (s *Scratch) normalizeCount() error { + var ( + tableLog = s.actualTableLog + scale = 62 - uint64(tableLog) + step = (1 << 62) / uint64(s.br.remain()) + vStep = uint64(1) << (scale - 20) + stillToDistribute = int16(1 << tableLog) + largest int + largestP int16 + lowThreshold = (uint32)(s.br.remain() >> tableLog) + ) + + for i, cnt := range s.count[:s.symbolLen] { + // already handled + // if (count[s] == s.length) return 0; /* rle special case */ + + if cnt == 0 { + s.norm[i] = 0 + continue + } + if cnt <= lowThreshold { + s.norm[i] = -1 + stillToDistribute-- + } else { + proba := (int16)((uint64(cnt) * step) >> scale) + if proba < 8 { + restToBeat := vStep * uint64(rtbTable[proba]) + v := uint64(cnt)*step - (uint64(proba) << scale) + if v > restToBeat { + proba++ + } + } + if proba > largestP { + largestP = proba + largest = i + } + s.norm[i] = proba + stillToDistribute -= proba + } + } + + if -stillToDistribute >= (s.norm[largest] >> 1) { + // corner case, need another normalization method + return s.normalizeCount2() + } + s.norm[largest] += stillToDistribute + return nil +} + +// Secondary normalization method. +// To be used when primary method fails. +func (s *Scratch) normalizeCount2() error { + const notYetAssigned = -2 + var ( + distributed uint32 + total = uint32(s.br.remain()) + tableLog = s.actualTableLog + lowThreshold = total >> tableLog + lowOne = (total * 3) >> (tableLog + 1) + ) + for i, cnt := range s.count[:s.symbolLen] { + if cnt == 0 { + s.norm[i] = 0 + continue + } + if cnt <= lowThreshold { + s.norm[i] = -1 + distributed++ + total -= cnt + continue + } + if cnt <= lowOne { + s.norm[i] = 1 + distributed++ + total -= cnt + continue + } + s.norm[i] = notYetAssigned + } + toDistribute := (1 << tableLog) - distributed + + if (total / toDistribute) > lowOne { + // risk of rounding to zero + lowOne = (total * 3) / (toDistribute * 2) + for i, cnt := range s.count[:s.symbolLen] { + if (s.norm[i] == notYetAssigned) && (cnt <= lowOne) { + s.norm[i] = 1 + distributed++ + total -= cnt + continue + } + } + toDistribute = (1 << tableLog) - distributed + } + if distributed == uint32(s.symbolLen)+1 { + // all values are pretty poor; + // probably incompressible data (should have already been detected); + // find max, then give all remaining points to max + var maxV int + var maxC uint32 + for i, cnt := range s.count[:s.symbolLen] { + if cnt > maxC { + maxV = i + maxC = cnt + } + } + s.norm[maxV] += int16(toDistribute) + return nil + } + + if total == 0 { + // all of the symbols were low enough for the lowOne or lowThreshold + for i := uint32(0); toDistribute > 0; i = (i + 1) % (uint32(s.symbolLen)) { + if s.norm[i] > 0 { + toDistribute-- + s.norm[i]++ + } + } + return nil + } + + var ( + vStepLog = 62 - uint64(tableLog) + mid = uint64((1 << (vStepLog - 1)) - 1) + rStep = (((1 << vStepLog) * uint64(toDistribute)) + mid) / uint64(total) // scale on remaining + tmpTotal = mid + ) + for i, cnt := range s.count[:s.symbolLen] { + if s.norm[i] == notYetAssigned { + var ( + end = tmpTotal + uint64(cnt)*rStep + sStart = uint32(tmpTotal >> vStepLog) + sEnd = uint32(end >> vStepLog) + weight = sEnd - sStart + ) + if weight < 1 { + return errors.New("weight < 1") + } + s.norm[i] = int16(weight) + tmpTotal = end + } + } + return nil +} + +// validateNorm validates the normalized histogram table. +func (s *Scratch) validateNorm() (err error) { + var total int + for _, v := range s.norm[:s.symbolLen] { + if v >= 0 { + total += int(v) + } else { + total -= int(v) + } + } + defer func() { + if err == nil { + return + } + fmt.Printf("selected TableLog: %d, Symbol length: %d\n", s.actualTableLog, s.symbolLen) + for i, v := range s.norm[:s.symbolLen] { + fmt.Printf("%3d: %5d -> %4d \n", i, s.count[i], v) + } + }() + if total != (1 << s.actualTableLog) { + return fmt.Errorf("warning: Total == %d != %d", total, 1<2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.) + * Increasing memory usage improves compression ratio + * Reduced memory usage can improve speed, due to cache effect + * Recommended max value is 14, for 16KB, which nicely fits into Intel x86 L1 cache */ + maxMemoryUsage = 14 + defaultMemoryUsage = 13 + + maxTableLog = maxMemoryUsage - 2 + maxTablesize = 1 << maxTableLog + defaultTablelog = defaultMemoryUsage - 2 + minTablelog = 5 + maxSymbolValue = 255 +) + +var ( + // ErrIncompressible is returned when input is judged to be too hard to compress. + ErrIncompressible = errors.New("input is not compressible") + + // ErrUseRLE is returned from the compressor when the input is a single byte value repeated. + ErrUseRLE = errors.New("input is single value repeated") +) + +// Scratch provides temporary storage for compression and decompression. +type Scratch struct { + // Private + count [maxSymbolValue + 1]uint32 + norm [maxSymbolValue + 1]int16 + br byteReader + bits bitReader + bw bitWriter + ct cTable // Compression tables. + maxCount int // count of the most probable symbol + + // Per block parameters. + // These can be used to override compression parameters of the block. + // Do not touch, unless you know what you are doing. + + // Out is output buffer. + // If the scratch is re-used before the caller is done processing the output, + // set this field to nil. + // Otherwise the output buffer will be re-used for next Compression/Decompression step + // and allocation will be avoided. + Out []byte + + // DecompressLimit limits the maximum decoded size acceptable. + // If > 0 decompression will stop when approximately this many bytes + // has been decoded. + // If 0, maximum size will be 2GB. + DecompressLimit int + + symbolLen uint16 // Length of active part of the symbol table. + actualTableLog uint8 // Selected tablelog. + zeroBits bool // no bits has prob > 50%. + clearCount bool // clear count + + // MaxSymbolValue will override the maximum symbol value of the next block. + MaxSymbolValue uint8 + + // TableLog will attempt to override the tablelog for the next block. + TableLog uint8 +} + +// Histogram allows to populate the histogram and skip that step in the compression, +// It otherwise allows to inspect the histogram when compression is done. +// To indicate that you have populated the histogram call HistogramFinished +// with the value of the highest populated symbol, as well as the number of entries +// in the most populated entry. These are accepted at face value. +// The returned slice will always be length 256. +func (s *Scratch) Histogram() []uint32 { + return s.count[:] +} + +// HistogramFinished can be called to indicate that the histogram has been populated. +// maxSymbol is the index of the highest set symbol of the next data segment. +// maxCount is the number of entries in the most populated entry. +// These are accepted at face value. +func (s *Scratch) HistogramFinished(maxSymbol uint8, maxCount int) { + s.maxCount = maxCount + s.symbolLen = uint16(maxSymbol) + 1 + s.clearCount = maxCount != 0 +} + +// prepare will prepare and allocate scratch tables used for both compression and decompression. +func (s *Scratch) prepare(in []byte) (*Scratch, error) { + if s == nil { + s = &Scratch{} + } + if s.MaxSymbolValue == 0 { + s.MaxSymbolValue = 255 + } + if s.TableLog == 0 { + s.TableLog = defaultTablelog + } + if s.TableLog > maxTableLog { + return nil, fmt.Errorf("tableLog (%d) > maxTableLog (%d)", s.TableLog, maxTableLog) + } + if cap(s.Out) == 0 { + s.Out = make([]byte, 0, len(in)) + } + if s.clearCount && s.maxCount == 0 { + for i := range s.count { + s.count[i] = 0 + } + s.clearCount = false + } + s.br.init(in) + if s.DecompressLimit == 0 { + // Max size 2GB. + s.DecompressLimit = (2 << 30) - 1 + } + + return s, nil +} + +// tableStep returns the next table index. +func tableStep(tableSize uint32) uint32 { + return (tableSize >> 1) + (tableSize >> 3) + 3 +} + +func highBits(val uint32) (n uint32) { + return uint32(bits.Len32(val) - 1) +} diff --git a/internal/github.com/klauspost/compress/huff0/.gitignore b/internal/github.com/klauspost/compress/huff0/.gitignore new file mode 100644 index 000000000..b3d262958 --- /dev/null +++ b/internal/github.com/klauspost/compress/huff0/.gitignore @@ -0,0 +1 @@ +/huff0-fuzz.zip diff --git a/internal/github.com/klauspost/compress/huff0/bitwriter.go b/internal/github.com/klauspost/compress/huff0/bitwriter.go new file mode 100644 index 000000000..0ebc9aaac --- /dev/null +++ b/internal/github.com/klauspost/compress/huff0/bitwriter.go @@ -0,0 +1,102 @@ +// Copyright 2018 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. +// Based on work Copyright (c) 2013, Yann Collet, released under BSD License. + +package huff0 + +// bitWriter will write bits. +// First bit will be LSB of the first byte of output. +type bitWriter struct { + bitContainer uint64 + nBits uint8 + out []byte +} + +// addBits16Clean will add up to 16 bits. value may not contain more set bits than indicated. +// It will not check if there is space for them, so the caller must ensure that it has flushed recently. +func (b *bitWriter) addBits16Clean(value uint16, bits uint8) { + b.bitContainer |= uint64(value) << (b.nBits & 63) + b.nBits += bits +} + +// encSymbol will add up to 16 bits. value may not contain more set bits than indicated. +// It will not check if there is space for them, so the caller must ensure that it has flushed recently. +func (b *bitWriter) encSymbol(ct cTable, symbol byte) { + enc := ct[symbol] + b.bitContainer |= uint64(enc.val) << (b.nBits & 63) + if false { + if enc.nBits == 0 { + panic("nbits 0") + } + } + b.nBits += enc.nBits +} + +// encTwoSymbols will add up to 32 bits. value may not contain more set bits than indicated. +// It will not check if there is space for them, so the caller must ensure that it has flushed recently. +func (b *bitWriter) encTwoSymbols(ct cTable, av, bv byte) { + encA := ct[av] + encB := ct[bv] + sh := b.nBits & 63 + combined := uint64(encA.val) | (uint64(encB.val) << (encA.nBits & 63)) + b.bitContainer |= combined << sh + if false { + if encA.nBits == 0 { + panic("nbitsA 0") + } + if encB.nBits == 0 { + panic("nbitsB 0") + } + } + b.nBits += encA.nBits + encB.nBits +} + +// encFourSymbols adds up to 32 bits from four symbols. +// It will not check if there is space for them, +// so the caller must ensure that b has been flushed recently. +func (b *bitWriter) encFourSymbols(encA, encB, encC, encD cTableEntry) { + bitsA := encA.nBits + bitsB := bitsA + encB.nBits + bitsC := bitsB + encC.nBits + bitsD := bitsC + encD.nBits + combined := uint64(encA.val) | + (uint64(encB.val) << (bitsA & 63)) | + (uint64(encC.val) << (bitsB & 63)) | + (uint64(encD.val) << (bitsC & 63)) + b.bitContainer |= combined << (b.nBits & 63) + b.nBits += bitsD +} + +// flush32 will flush out, so there are at least 32 bits available for writing. +func (b *bitWriter) flush32() { + if b.nBits < 32 { + return + } + b.out = append(b.out, + byte(b.bitContainer), + byte(b.bitContainer>>8), + byte(b.bitContainer>>16), + byte(b.bitContainer>>24)) + b.nBits -= 32 + b.bitContainer >>= 32 +} + +// flushAlign will flush remaining full bytes and align to next byte boundary. +func (b *bitWriter) flushAlign() { + nbBytes := (b.nBits + 7) >> 3 + for i := uint8(0); i < nbBytes; i++ { + b.out = append(b.out, byte(b.bitContainer>>(i*8))) + } + b.nBits = 0 + b.bitContainer = 0 +} + +// close will write the alignment bit and write the final byte(s) +// to the output. +func (b *bitWriter) close() { + // End mark + b.addBits16Clean(1, 1) + // flush until next byte. + b.flushAlign() +} diff --git a/internal/github.com/klauspost/compress/huff0/compress.go b/internal/github.com/klauspost/compress/huff0/compress.go new file mode 100644 index 000000000..9bad07505 --- /dev/null +++ b/internal/github.com/klauspost/compress/huff0/compress.go @@ -0,0 +1,678 @@ +package huff0 + +import ( + "fmt" + "math" + "runtime" + "sync" +) + +// Compress1X will compress the input. +// The output can be decoded using Decompress1X. +// Supply a Scratch object. The scratch object contains state about re-use, +// So when sharing across independent encodes, be sure to set the re-use policy. +func Compress1X(in []byte, s *Scratch) (out []byte, reUsed bool, err error) { + s, err = s.prepare(in) + if err != nil { + return nil, false, err + } + return compress(in, s, s.compress1X) +} + +// Compress4X will compress the input. The input is split into 4 independent blocks +// and compressed similar to Compress1X. +// The output can be decoded using Decompress4X. +// Supply a Scratch object. The scratch object contains state about re-use, +// So when sharing across independent encodes, be sure to set the re-use policy. +func Compress4X(in []byte, s *Scratch) (out []byte, reUsed bool, err error) { + s, err = s.prepare(in) + if err != nil { + return nil, false, err + } + if false { + // TODO: compress4Xp only slightly faster. + const parallelThreshold = 8 << 10 + if len(in) < parallelThreshold || runtime.GOMAXPROCS(0) == 1 { + return compress(in, s, s.compress4X) + } + return compress(in, s, s.compress4Xp) + } + return compress(in, s, s.compress4X) +} + +func compress(in []byte, s *Scratch, compressor func(src []byte) ([]byte, error)) (out []byte, reUsed bool, err error) { + // Nuke previous table if we cannot reuse anyway. + if s.Reuse == ReusePolicyNone { + s.prevTable = s.prevTable[:0] + } + + // Create histogram, if none was provided. + maxCount := s.maxCount + var canReuse = false + if maxCount == 0 { + maxCount, canReuse = s.countSimple(in) + } else { + canReuse = s.canUseTable(s.prevTable) + } + + // We want the output size to be less than this: + wantSize := len(in) + if s.WantLogLess > 0 { + wantSize -= wantSize >> s.WantLogLess + } + + // Reset for next run. + s.clearCount = true + s.maxCount = 0 + if maxCount >= len(in) { + if maxCount > len(in) { + return nil, false, fmt.Errorf("maxCount (%d) > length (%d)", maxCount, len(in)) + } + if len(in) == 1 { + return nil, false, ErrIncompressible + } + // One symbol, use RLE + return nil, false, ErrUseRLE + } + if maxCount == 1 || maxCount < (len(in)>>7) { + // Each symbol present maximum once or too well distributed. + return nil, false, ErrIncompressible + } + if s.Reuse == ReusePolicyMust && !canReuse { + // We must reuse, but we can't. + return nil, false, ErrIncompressible + } + if (s.Reuse == ReusePolicyPrefer || s.Reuse == ReusePolicyMust) && canReuse { + keepTable := s.cTable + keepTL := s.actualTableLog + s.cTable = s.prevTable + s.actualTableLog = s.prevTableLog + s.Out, err = compressor(in) + s.cTable = keepTable + s.actualTableLog = keepTL + if err == nil && len(s.Out) < wantSize { + s.OutData = s.Out + return s.Out, true, nil + } + if s.Reuse == ReusePolicyMust { + return nil, false, ErrIncompressible + } + // Do not attempt to re-use later. + s.prevTable = s.prevTable[:0] + } + + // Calculate new table. + err = s.buildCTable() + if err != nil { + return nil, false, err + } + + if false && !s.canUseTable(s.cTable) { + panic("invalid table generated") + } + + if s.Reuse == ReusePolicyAllow && canReuse { + hSize := len(s.Out) + oldSize := s.prevTable.estimateSize(s.count[:s.symbolLen]) + newSize := s.cTable.estimateSize(s.count[:s.symbolLen]) + if oldSize <= hSize+newSize || hSize+12 >= wantSize { + // Retain cTable even if we re-use. + keepTable := s.cTable + keepTL := s.actualTableLog + + s.cTable = s.prevTable + s.actualTableLog = s.prevTableLog + s.Out, err = compressor(in) + + // Restore ctable. + s.cTable = keepTable + s.actualTableLog = keepTL + if err != nil { + return nil, false, err + } + if len(s.Out) >= wantSize { + return nil, false, ErrIncompressible + } + s.OutData = s.Out + return s.Out, true, nil + } + } + + // Use new table + err = s.cTable.write(s) + if err != nil { + s.OutTable = nil + return nil, false, err + } + s.OutTable = s.Out + + // Compress using new table + s.Out, err = compressor(in) + if err != nil { + s.OutTable = nil + return nil, false, err + } + if len(s.Out) >= wantSize { + s.OutTable = nil + return nil, false, ErrIncompressible + } + // Move current table into previous. + s.prevTable, s.prevTableLog, s.cTable = s.cTable, s.actualTableLog, s.prevTable[:0] + s.OutData = s.Out[len(s.OutTable):] + return s.Out, false, nil +} + +func (s *Scratch) compress1X(src []byte) ([]byte, error) { + return s.compress1xDo(s.Out, src), nil +} + +func (s *Scratch) compress1xDo(dst, src []byte) []byte { + var bw = bitWriter{out: dst} + + // N is length divisible by 4. + n := len(src) + n -= n & 3 + cTable := s.cTable[:256] + + // Encode last bytes. + for i := len(src) & 3; i > 0; i-- { + bw.encSymbol(cTable, src[n+i-1]) + } + n -= 4 + if s.actualTableLog <= 8 { + for ; n >= 0; n -= 4 { + tmp := src[n : n+4] + // tmp should be len 4 + bw.flush32() + bw.encFourSymbols(cTable[tmp[3]], cTable[tmp[2]], cTable[tmp[1]], cTable[tmp[0]]) + } + } else { + for ; n >= 0; n -= 4 { + tmp := src[n : n+4] + // tmp should be len 4 + bw.flush32() + bw.encTwoSymbols(cTable, tmp[3], tmp[2]) + bw.flush32() + bw.encTwoSymbols(cTable, tmp[1], tmp[0]) + } + } + bw.close() + return bw.out +} + +var sixZeros [6]byte + +func (s *Scratch) compress4X(src []byte) ([]byte, error) { + if len(src) < 12 { + return nil, ErrIncompressible + } + segmentSize := (len(src) + 3) / 4 + + // Add placeholder for output length + offsetIdx := len(s.Out) + s.Out = append(s.Out, sixZeros[:]...) + + for i := 0; i < 4; i++ { + toDo := src + if len(toDo) > segmentSize { + toDo = toDo[:segmentSize] + } + src = src[len(toDo):] + + idx := len(s.Out) + s.Out = s.compress1xDo(s.Out, toDo) + if len(s.Out)-idx > math.MaxUint16 { + // We cannot store the size in the jump table + return nil, ErrIncompressible + } + // Write compressed length as little endian before block. + if i < 3 { + // Last length is not written. + length := len(s.Out) - idx + s.Out[i*2+offsetIdx] = byte(length) + s.Out[i*2+offsetIdx+1] = byte(length >> 8) + } + } + + return s.Out, nil +} + +// compress4Xp will compress 4 streams using separate goroutines. +func (s *Scratch) compress4Xp(src []byte) ([]byte, error) { + if len(src) < 12 { + return nil, ErrIncompressible + } + // Add placeholder for output length + s.Out = s.Out[:6] + + segmentSize := (len(src) + 3) / 4 + var wg sync.WaitGroup + wg.Add(4) + for i := 0; i < 4; i++ { + toDo := src + if len(toDo) > segmentSize { + toDo = toDo[:segmentSize] + } + src = src[len(toDo):] + + // Separate goroutine for each block. + go func(i int) { + s.tmpOut[i] = s.compress1xDo(s.tmpOut[i][:0], toDo) + wg.Done() + }(i) + } + wg.Wait() + for i := 0; i < 4; i++ { + o := s.tmpOut[i] + if len(o) > math.MaxUint16 { + // We cannot store the size in the jump table + return nil, ErrIncompressible + } + // Write compressed length as little endian before block. + if i < 3 { + // Last length is not written. + s.Out[i*2] = byte(len(o)) + s.Out[i*2+1] = byte(len(o) >> 8) + } + + // Write output. + s.Out = append(s.Out, o...) + } + return s.Out, nil +} + +// countSimple will create a simple histogram in s.count. +// Returns the biggest count. +// Does not update s.clearCount. +func (s *Scratch) countSimple(in []byte) (max int, reuse bool) { + reuse = true + _ = s.count // Assert that s != nil to speed up the following loop. + for _, v := range in { + s.count[v]++ + } + m := uint32(0) + if len(s.prevTable) > 0 { + for i, v := range s.count[:] { + if v == 0 { + continue + } + if v > m { + m = v + } + s.symbolLen = uint16(i) + 1 + if i >= len(s.prevTable) { + reuse = false + } else if s.prevTable[i].nBits == 0 { + reuse = false + } + } + return int(m), reuse + } + for i, v := range s.count[:] { + if v == 0 { + continue + } + if v > m { + m = v + } + s.symbolLen = uint16(i) + 1 + } + return int(m), false +} + +func (s *Scratch) canUseTable(c cTable) bool { + if len(c) < int(s.symbolLen) { + return false + } + for i, v := range s.count[:s.symbolLen] { + if v != 0 && c[i].nBits == 0 { + return false + } + } + return true +} + +//lint:ignore U1000 used for debugging +func (s *Scratch) validateTable(c cTable) bool { + if len(c) < int(s.symbolLen) { + return false + } + for i, v := range s.count[:s.symbolLen] { + if v != 0 { + if c[i].nBits == 0 { + return false + } + if c[i].nBits > s.actualTableLog { + return false + } + } + } + return true +} + +// minTableLog provides the minimum logSize to safely represent a distribution. +func (s *Scratch) minTableLog() uint8 { + minBitsSrc := highBit32(uint32(s.srcLen)) + 1 + minBitsSymbols := highBit32(uint32(s.symbolLen-1)) + 2 + if minBitsSrc < minBitsSymbols { + return uint8(minBitsSrc) + } + return uint8(minBitsSymbols) +} + +// optimalTableLog calculates and sets the optimal tableLog in s.actualTableLog +func (s *Scratch) optimalTableLog() { + tableLog := s.TableLog + minBits := s.minTableLog() + maxBitsSrc := uint8(highBit32(uint32(s.srcLen-1))) - 1 + if maxBitsSrc < tableLog { + // Accuracy can be reduced + tableLog = maxBitsSrc + } + if minBits > tableLog { + tableLog = minBits + } + // Need a minimum to safely represent all symbol values + if tableLog < minTablelog { + tableLog = minTablelog + } + if tableLog > tableLogMax { + tableLog = tableLogMax + } + s.actualTableLog = tableLog +} + +type cTableEntry struct { + val uint16 + nBits uint8 + // We have 8 bits extra +} + +const huffNodesMask = huffNodesLen - 1 + +func (s *Scratch) buildCTable() error { + s.optimalTableLog() + s.huffSort() + if cap(s.cTable) < maxSymbolValue+1 { + s.cTable = make([]cTableEntry, s.symbolLen, maxSymbolValue+1) + } else { + s.cTable = s.cTable[:s.symbolLen] + for i := range s.cTable { + s.cTable[i] = cTableEntry{} + } + } + + var startNode = int16(s.symbolLen) + nonNullRank := s.symbolLen - 1 + + nodeNb := startNode + huffNode := s.nodes[1 : huffNodesLen+1] + + // This overlays the slice above, but allows "-1" index lookups. + // Different from reference implementation. + huffNode0 := s.nodes[0 : huffNodesLen+1] + + for huffNode[nonNullRank].count() == 0 { + nonNullRank-- + } + + lowS := int16(nonNullRank) + nodeRoot := nodeNb + lowS - 1 + lowN := nodeNb + huffNode[nodeNb].setCount(huffNode[lowS].count() + huffNode[lowS-1].count()) + huffNode[lowS].setParent(nodeNb) + huffNode[lowS-1].setParent(nodeNb) + nodeNb++ + lowS -= 2 + for n := nodeNb; n <= nodeRoot; n++ { + huffNode[n].setCount(1 << 30) + } + // fake entry, strong barrier + huffNode0[0].setCount(1 << 31) + + // create parents + for nodeNb <= nodeRoot { + var n1, n2 int16 + if huffNode0[lowS+1].count() < huffNode0[lowN+1].count() { + n1 = lowS + lowS-- + } else { + n1 = lowN + lowN++ + } + if huffNode0[lowS+1].count() < huffNode0[lowN+1].count() { + n2 = lowS + lowS-- + } else { + n2 = lowN + lowN++ + } + + huffNode[nodeNb].setCount(huffNode0[n1+1].count() + huffNode0[n2+1].count()) + huffNode0[n1+1].setParent(nodeNb) + huffNode0[n2+1].setParent(nodeNb) + nodeNb++ + } + + // distribute weights (unlimited tree height) + huffNode[nodeRoot].setNbBits(0) + for n := nodeRoot - 1; n >= startNode; n-- { + huffNode[n].setNbBits(huffNode[huffNode[n].parent()].nbBits() + 1) + } + for n := uint16(0); n <= nonNullRank; n++ { + huffNode[n].setNbBits(huffNode[huffNode[n].parent()].nbBits() + 1) + } + s.actualTableLog = s.setMaxHeight(int(nonNullRank)) + maxNbBits := s.actualTableLog + + // fill result into tree (val, nbBits) + if maxNbBits > tableLogMax { + return fmt.Errorf("internal error: maxNbBits (%d) > tableLogMax (%d)", maxNbBits, tableLogMax) + } + var nbPerRank [tableLogMax + 1]uint16 + var valPerRank [16]uint16 + for _, v := range huffNode[:nonNullRank+1] { + nbPerRank[v.nbBits()]++ + } + // determine stating value per rank + { + min := uint16(0) + for n := maxNbBits; n > 0; n-- { + // get starting value within each rank + valPerRank[n] = min + min += nbPerRank[n] + min >>= 1 + } + } + + // push nbBits per symbol, symbol order + for _, v := range huffNode[:nonNullRank+1] { + s.cTable[v.symbol()].nBits = v.nbBits() + } + + // assign value within rank, symbol order + t := s.cTable[:s.symbolLen] + for n, val := range t { + nbits := val.nBits & 15 + v := valPerRank[nbits] + t[n].val = v + valPerRank[nbits] = v + 1 + } + + return nil +} + +// huffSort will sort symbols, decreasing order. +func (s *Scratch) huffSort() { + type rankPos struct { + base uint32 + current uint32 + } + + // Clear nodes + nodes := s.nodes[:huffNodesLen+1] + s.nodes = nodes + nodes = nodes[1 : huffNodesLen+1] + + // Sort into buckets based on length of symbol count. + var rank [32]rankPos + for _, v := range s.count[:s.symbolLen] { + r := highBit32(v+1) & 31 + rank[r].base++ + } + // maxBitLength is log2(BlockSizeMax) + 1 + const maxBitLength = 18 + 1 + for n := maxBitLength; n > 0; n-- { + rank[n-1].base += rank[n].base + } + for n := range rank[:maxBitLength] { + rank[n].current = rank[n].base + } + for n, c := range s.count[:s.symbolLen] { + r := (highBit32(c+1) + 1) & 31 + pos := rank[r].current + rank[r].current++ + prev := nodes[(pos-1)&huffNodesMask] + for pos > rank[r].base && c > prev.count() { + nodes[pos&huffNodesMask] = prev + pos-- + prev = nodes[(pos-1)&huffNodesMask] + } + nodes[pos&huffNodesMask] = makeNodeElt(c, byte(n)) + } +} + +func (s *Scratch) setMaxHeight(lastNonNull int) uint8 { + maxNbBits := s.actualTableLog + huffNode := s.nodes[1 : huffNodesLen+1] + //huffNode = huffNode[: huffNodesLen] + + largestBits := huffNode[lastNonNull].nbBits() + + // early exit : no elt > maxNbBits + if largestBits <= maxNbBits { + return largestBits + } + totalCost := int(0) + baseCost := int(1) << (largestBits - maxNbBits) + n := uint32(lastNonNull) + + for huffNode[n].nbBits() > maxNbBits { + totalCost += baseCost - (1 << (largestBits - huffNode[n].nbBits())) + huffNode[n].setNbBits(maxNbBits) + n-- + } + // n stops at huffNode[n].nbBits <= maxNbBits + + for huffNode[n].nbBits() == maxNbBits { + n-- + } + // n end at index of smallest symbol using < maxNbBits + + // renorm totalCost + totalCost >>= largestBits - maxNbBits /* note : totalCost is necessarily a multiple of baseCost */ + + // repay normalized cost + { + const noSymbol = 0xF0F0F0F0 + var rankLast [tableLogMax + 2]uint32 + + for i := range rankLast[:] { + rankLast[i] = noSymbol + } + + // Get pos of last (smallest) symbol per rank + { + currentNbBits := maxNbBits + for pos := int(n); pos >= 0; pos-- { + if huffNode[pos].nbBits() >= currentNbBits { + continue + } + currentNbBits = huffNode[pos].nbBits() // < maxNbBits + rankLast[maxNbBits-currentNbBits] = uint32(pos) + } + } + + for totalCost > 0 { + nBitsToDecrease := uint8(highBit32(uint32(totalCost))) + 1 + + for ; nBitsToDecrease > 1; nBitsToDecrease-- { + highPos := rankLast[nBitsToDecrease] + lowPos := rankLast[nBitsToDecrease-1] + if highPos == noSymbol { + continue + } + if lowPos == noSymbol { + break + } + highTotal := huffNode[highPos].count() + lowTotal := 2 * huffNode[lowPos].count() + if highTotal <= lowTotal { + break + } + } + // only triggered when no more rank 1 symbol left => find closest one (note : there is necessarily at least one !) + // HUF_MAX_TABLELOG test just to please gcc 5+; but it should not be necessary + // FIXME: try to remove + for (nBitsToDecrease <= tableLogMax) && (rankLast[nBitsToDecrease] == noSymbol) { + nBitsToDecrease++ + } + totalCost -= 1 << (nBitsToDecrease - 1) + if rankLast[nBitsToDecrease-1] == noSymbol { + // this rank is no longer empty + rankLast[nBitsToDecrease-1] = rankLast[nBitsToDecrease] + } + huffNode[rankLast[nBitsToDecrease]].setNbBits(1 + + huffNode[rankLast[nBitsToDecrease]].nbBits()) + if rankLast[nBitsToDecrease] == 0 { + /* special case, reached largest symbol */ + rankLast[nBitsToDecrease] = noSymbol + } else { + rankLast[nBitsToDecrease]-- + if huffNode[rankLast[nBitsToDecrease]].nbBits() != maxNbBits-nBitsToDecrease { + rankLast[nBitsToDecrease] = noSymbol /* this rank is now empty */ + } + } + } + + for totalCost < 0 { /* Sometimes, cost correction overshoot */ + if rankLast[1] == noSymbol { /* special case : no rank 1 symbol (using maxNbBits-1); let's create one from largest rank 0 (using maxNbBits) */ + for huffNode[n].nbBits() == maxNbBits { + n-- + } + huffNode[n+1].setNbBits(huffNode[n+1].nbBits() - 1) + rankLast[1] = n + 1 + totalCost++ + continue + } + huffNode[rankLast[1]+1].setNbBits(huffNode[rankLast[1]+1].nbBits() - 1) + rankLast[1]++ + totalCost++ + } + } + return maxNbBits +} + +// A nodeElt is the fields +// +// count uint32 +// parent uint16 +// symbol byte +// nbBits uint8 +// +// in some order, all squashed into an integer so that the compiler +// always loads and stores entire nodeElts instead of separate fields. +type nodeElt uint64 + +func makeNodeElt(count uint32, symbol byte) nodeElt { + return nodeElt(count) | nodeElt(symbol)<<48 +} + +func (e *nodeElt) count() uint32 { return uint32(*e) } +func (e *nodeElt) parent() uint16 { return uint16(*e >> 32) } +func (e *nodeElt) symbol() byte { return byte(*e >> 48) } +func (e *nodeElt) nbBits() uint8 { return uint8(*e >> 56) } + +func (e *nodeElt) setCount(c uint32) { *e = (*e)&0xffffffff00000000 | nodeElt(c) } +func (e *nodeElt) setParent(p int16) { *e = (*e)&0xffff0000ffffffff | nodeElt(uint16(p))<<32 } +func (e *nodeElt) setNbBits(n uint8) { *e = (*e)&0x00ffffffffffffff | nodeElt(n)<<56 } diff --git a/internal/github.com/klauspost/compress/huff0/huff0.go b/internal/github.com/klauspost/compress/huff0/huff0.go new file mode 100644 index 000000000..4fa6cfd9e --- /dev/null +++ b/internal/github.com/klauspost/compress/huff0/huff0.go @@ -0,0 +1,337 @@ +// Package huff0 provides fast huffman encoding as used in zstd. +// +// See README.md at https://github.com/klauspost/compress/tree/master/huff0 for details. + +package huff0 + +import ( + "errors" + "fmt" + "math" + "math/bits" + "sync" + + "github.com/prometheus/client_golang/internal/github.com/klauspost/compress/fse" +) + +const ( + maxSymbolValue = 255 + + // zstandard limits tablelog to 11, see: + // https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#huffman-tree-description + tableLogMax = 11 + tableLogDefault = 11 + minTablelog = 5 + huffNodesLen = 512 + + // BlockSizeMax is maximum input size for a single block uncompressed. + BlockSizeMax = 1<<18 - 1 +) + +var ( + // ErrIncompressible is returned when input is judged to be too hard to compress. + ErrIncompressible = errors.New("input is not compressible") + + // ErrUseRLE is returned from the compressor when the input is a single byte value repeated. + ErrUseRLE = errors.New("input is single value repeated") + + // ErrTooBig is return if input is too large for a single block. + ErrTooBig = errors.New("input too big") + + // ErrMaxDecodedSizeExceeded is return if input is too large for a single block. + ErrMaxDecodedSizeExceeded = errors.New("maximum output size exceeded") +) + +type ReusePolicy uint8 + +const ( + // ReusePolicyAllow will allow reuse if it produces smaller output. + ReusePolicyAllow ReusePolicy = iota + + // ReusePolicyPrefer will re-use aggressively if possible. + // This will not check if a new table will produce smaller output, + // except if the current table is impossible to use or + // compressed output is bigger than input. + ReusePolicyPrefer + + // ReusePolicyNone will disable re-use of tables. + // This is slightly faster than ReusePolicyAllow but may produce larger output. + ReusePolicyNone + + // ReusePolicyMust must allow reuse and produce smaller output. + ReusePolicyMust +) + +type Scratch struct { + count [maxSymbolValue + 1]uint32 + + // Per block parameters. + // These can be used to override compression parameters of the block. + // Do not touch, unless you know what you are doing. + + // Out is output buffer. + // If the scratch is re-used before the caller is done processing the output, + // set this field to nil. + // Otherwise the output buffer will be re-used for next Compression/Decompression step + // and allocation will be avoided. + Out []byte + + // OutTable will contain the table data only, if a new table has been generated. + // Slice of the returned data. + OutTable []byte + + // OutData will contain the compressed data. + // Slice of the returned data. + OutData []byte + + // MaxDecodedSize will set the maximum allowed output size. + // This value will automatically be set to BlockSizeMax if not set. + // Decoders will return ErrMaxDecodedSizeExceeded is this limit is exceeded. + MaxDecodedSize int + + srcLen int + + // MaxSymbolValue will override the maximum symbol value of the next block. + MaxSymbolValue uint8 + + // TableLog will attempt to override the tablelog for the next block. + // Must be <= 11 and >= 5. + TableLog uint8 + + // Reuse will specify the reuse policy + Reuse ReusePolicy + + // WantLogLess allows to specify a log 2 reduction that should at least be achieved, + // otherwise the block will be returned as incompressible. + // The reduction should then at least be (input size >> WantLogLess) + // If WantLogLess == 0 any improvement will do. + WantLogLess uint8 + + symbolLen uint16 // Length of active part of the symbol table. + maxCount int // count of the most probable symbol + clearCount bool // clear count + actualTableLog uint8 // Selected tablelog. + prevTableLog uint8 // Tablelog for previous table + prevTable cTable // Table used for previous compression. + cTable cTable // compression table + nodes []nodeElt + tmpOut [4][]byte + fse *fse.Scratch + decPool sync.Pool // *[4][256]byte buffers. + huffWeight [maxSymbolValue + 1]byte +} + +// TransferCTable will transfer the previously used compression table. +func (s *Scratch) TransferCTable(src *Scratch) { + if cap(s.prevTable) < len(src.prevTable) { + s.prevTable = make(cTable, 0, maxSymbolValue+1) + } + s.prevTable = s.prevTable[:len(src.prevTable)] + copy(s.prevTable, src.prevTable) + s.prevTableLog = src.prevTableLog +} + +func (s *Scratch) prepare(in []byte) (*Scratch, error) { + if len(in) > BlockSizeMax { + return nil, ErrTooBig + } + if s == nil { + s = &Scratch{} + } + if s.MaxSymbolValue == 0 { + s.MaxSymbolValue = maxSymbolValue + } + if s.TableLog == 0 { + s.TableLog = tableLogDefault + } + if s.TableLog > tableLogMax || s.TableLog < minTablelog { + return nil, fmt.Errorf(" invalid tableLog %d (%d -> %d)", s.TableLog, minTablelog, tableLogMax) + } + if s.MaxDecodedSize <= 0 || s.MaxDecodedSize > BlockSizeMax { + s.MaxDecodedSize = BlockSizeMax + } + if s.clearCount && s.maxCount == 0 { + for i := range s.count { + s.count[i] = 0 + } + s.clearCount = false + } + if cap(s.Out) == 0 { + s.Out = make([]byte, 0, len(in)) + } + s.Out = s.Out[:0] + + s.OutTable = nil + s.OutData = nil + if cap(s.nodes) < huffNodesLen+1 { + s.nodes = make([]nodeElt, 0, huffNodesLen+1) + } + s.nodes = s.nodes[:0] + if s.fse == nil { + s.fse = &fse.Scratch{} + } + s.srcLen = len(in) + + return s, nil +} + +type cTable []cTableEntry + +func (c cTable) write(s *Scratch) error { + var ( + // precomputed conversion table + bitsToWeight [tableLogMax + 1]byte + huffLog = s.actualTableLog + // last weight is not saved. + maxSymbolValue = uint8(s.symbolLen - 1) + huffWeight = s.huffWeight[:256] + ) + const ( + maxFSETableLog = 6 + ) + // convert to weight + bitsToWeight[0] = 0 + for n := uint8(1); n < huffLog+1; n++ { + bitsToWeight[n] = huffLog + 1 - n + } + + // Acquire histogram for FSE. + hist := s.fse.Histogram() + hist = hist[:256] + for i := range hist[:16] { + hist[i] = 0 + } + for n := uint8(0); n < maxSymbolValue; n++ { + v := bitsToWeight[c[n].nBits] & 15 + huffWeight[n] = v + hist[v]++ + } + + // FSE compress if feasible. + if maxSymbolValue >= 2 { + huffMaxCnt := uint32(0) + huffMax := uint8(0) + for i, v := range hist[:16] { + if v == 0 { + continue + } + huffMax = byte(i) + if v > huffMaxCnt { + huffMaxCnt = v + } + } + s.fse.HistogramFinished(huffMax, int(huffMaxCnt)) + s.fse.TableLog = maxFSETableLog + b, err := fse.Compress(huffWeight[:maxSymbolValue], s.fse) + if err == nil && len(b) < int(s.symbolLen>>1) { + s.Out = append(s.Out, uint8(len(b))) + s.Out = append(s.Out, b...) + return nil + } + // Unable to compress (RLE/uncompressible) + } + // write raw values as 4-bits (max : 15) + if maxSymbolValue > (256 - 128) { + // should not happen : likely means source cannot be compressed + return ErrIncompressible + } + op := s.Out + // special case, pack weights 4 bits/weight. + op = append(op, 128|(maxSymbolValue-1)) + // be sure it doesn't cause msan issue in final combination + huffWeight[maxSymbolValue] = 0 + for n := uint16(0); n < uint16(maxSymbolValue); n += 2 { + op = append(op, (huffWeight[n]<<4)|huffWeight[n+1]) + } + s.Out = op + return nil +} + +func (c cTable) estTableSize(s *Scratch) (sz int, err error) { + var ( + // precomputed conversion table + bitsToWeight [tableLogMax + 1]byte + huffLog = s.actualTableLog + // last weight is not saved. + maxSymbolValue = uint8(s.symbolLen - 1) + huffWeight = s.huffWeight[:256] + ) + const ( + maxFSETableLog = 6 + ) + // convert to weight + bitsToWeight[0] = 0 + for n := uint8(1); n < huffLog+1; n++ { + bitsToWeight[n] = huffLog + 1 - n + } + + // Acquire histogram for FSE. + hist := s.fse.Histogram() + hist = hist[:256] + for i := range hist[:16] { + hist[i] = 0 + } + for n := uint8(0); n < maxSymbolValue; n++ { + v := bitsToWeight[c[n].nBits] & 15 + huffWeight[n] = v + hist[v]++ + } + + // FSE compress if feasible. + if maxSymbolValue >= 2 { + huffMaxCnt := uint32(0) + huffMax := uint8(0) + for i, v := range hist[:16] { + if v == 0 { + continue + } + huffMax = byte(i) + if v > huffMaxCnt { + huffMaxCnt = v + } + } + s.fse.HistogramFinished(huffMax, int(huffMaxCnt)) + s.fse.TableLog = maxFSETableLog + b, err := fse.Compress(huffWeight[:maxSymbolValue], s.fse) + if err == nil && len(b) < int(s.symbolLen>>1) { + sz += 1 + len(b) + return sz, nil + } + // Unable to compress (RLE/uncompressible) + } + // write raw values as 4-bits (max : 15) + if maxSymbolValue > (256 - 128) { + // should not happen : likely means source cannot be compressed + return 0, ErrIncompressible + } + // special case, pack weights 4 bits/weight. + sz += 1 + int(maxSymbolValue/2) + return sz, nil +} + +// estimateSize returns the estimated size in bytes of the input represented in the +// histogram supplied. +func (c cTable) estimateSize(hist []uint32) int { + nbBits := uint32(7) + for i, v := range c[:len(hist)] { + nbBits += uint32(v.nBits) * hist[i] + } + return int(nbBits >> 3) +} + +// minSize returns the minimum possible size considering the shannon limit. +func (s *Scratch) minSize(total int) int { + nbBits := float64(7) + fTotal := float64(total) + for _, v := range s.count[:s.symbolLen] { + n := float64(v) + if n > 0 { + nbBits += math.Log2(fTotal/n) * n + } + } + return int(nbBits) >> 3 +} + +func highBit32(val uint32) (n uint32) { + return uint32(bits.Len32(val) - 1) +} diff --git a/internal/github.com/klauspost/compress/internal/le/le.go b/internal/github.com/klauspost/compress/internal/le/le.go new file mode 100644 index 000000000..e54909e16 --- /dev/null +++ b/internal/github.com/klauspost/compress/internal/le/le.go @@ -0,0 +1,5 @@ +package le + +type Indexer interface { + int | int8 | int16 | int32 | int64 | uint | uint8 | uint16 | uint32 | uint64 +} diff --git a/internal/github.com/klauspost/compress/internal/le/unsafe_disabled.go b/internal/github.com/klauspost/compress/internal/le/unsafe_disabled.go new file mode 100644 index 000000000..830a79bae --- /dev/null +++ b/internal/github.com/klauspost/compress/internal/le/unsafe_disabled.go @@ -0,0 +1,40 @@ +package le + +import ( + "encoding/binary" +) + +// Load8 will load from b at index i. +func Load8[I Indexer](b []byte, i I) byte { + return b[i] +} + +// Load16 will load from b at index i. +func Load16[I Indexer](b []byte, i I) uint16 { + return binary.LittleEndian.Uint16(b[i:]) +} + +// Load32 will load from b at index i. +func Load32[I Indexer](b []byte, i I) uint32 { + return binary.LittleEndian.Uint32(b[i:]) +} + +// Load64 will load from b at index i. +func Load64[I Indexer](b []byte, i I) uint64 { + return binary.LittleEndian.Uint64(b[i:]) +} + +// Store16 will store v at b. +func Store16(b []byte, v uint16) { + binary.LittleEndian.PutUint16(b, v) +} + +// Store32 will store v at b. +func Store32(b []byte, v uint32) { + binary.LittleEndian.PutUint32(b, v) +} + +// Store64 will store v at b. +func Store64(b []byte, v uint64) { + binary.LittleEndian.PutUint64(b, v) +} diff --git a/internal/github.com/klauspost/compress/zstd/bitwriter.go b/internal/github.com/klauspost/compress/zstd/bitwriter.go new file mode 100644 index 000000000..1952f175b --- /dev/null +++ b/internal/github.com/klauspost/compress/zstd/bitwriter.go @@ -0,0 +1,112 @@ +// Copyright 2018 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. +// Based on work Copyright (c) 2013, Yann Collet, released under BSD License. + +package zstd + +// bitWriter will write bits. +// First bit will be LSB of the first byte of output. +type bitWriter struct { + bitContainer uint64 + nBits uint8 + out []byte +} + +// bitMask16 is bitmasks. Has extra to avoid bounds check. +var bitMask16 = [32]uint16{ + 0, 1, 3, 7, 0xF, 0x1F, + 0x3F, 0x7F, 0xFF, 0x1FF, 0x3FF, 0x7FF, + 0xFFF, 0x1FFF, 0x3FFF, 0x7FFF, 0xFFFF, 0xFFFF, + 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, + 0xFFFF, 0xFFFF} /* up to 16 bits */ + +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 + +// addBits16NC will add up to 16 bits. +// It will not check if there is space for them, +// so the caller must ensure that it has flushed recently. +func (b *bitWriter) addBits16NC(value uint16, bits uint8) { + b.bitContainer |= uint64(value&bitMask16[bits&31]) << (b.nBits & 63) + b.nBits += bits +} + +// addBits32NC will add up to 31 bits. +// It will not check if there is space for them, +// so the caller must ensure that it has flushed recently. +func (b *bitWriter) addBits32NC(value uint32, bits uint8) { + b.bitContainer |= uint64(value&bitMask32[bits&31]) << (b.nBits & 63) + b.nBits += bits +} + +// addBits64NC will add up to 64 bits. +// There must be space for 32 bits. +func (b *bitWriter) addBits64NC(value uint64, bits uint8) { + if bits <= 31 { + b.addBits32Clean(uint32(value), bits) + return + } + b.addBits32Clean(uint32(value), 32) + b.flush32() + b.addBits32Clean(uint32(value>>32), bits-32) +} + +// addBits32Clean will add up to 32 bits. +// It will not check if there is space for them. +// The input must not contain more bits than specified. +func (b *bitWriter) addBits32Clean(value uint32, bits uint8) { + b.bitContainer |= uint64(value) << (b.nBits & 63) + b.nBits += bits +} + +// addBits16Clean will add up to 16 bits. value may not contain more set bits than indicated. +// It will not check if there is space for them, so the caller must ensure that it has flushed recently. +func (b *bitWriter) addBits16Clean(value uint16, bits uint8) { + b.bitContainer |= uint64(value) << (b.nBits & 63) + b.nBits += bits +} + +// flush32 will flush out, so there are at least 32 bits available for writing. +func (b *bitWriter) flush32() { + if b.nBits < 32 { + return + } + b.out = append(b.out, + byte(b.bitContainer), + byte(b.bitContainer>>8), + byte(b.bitContainer>>16), + byte(b.bitContainer>>24)) + b.nBits -= 32 + b.bitContainer >>= 32 +} + +// flushAlign will flush remaining full bytes and align to next byte boundary. +func (b *bitWriter) flushAlign() { + nbBytes := (b.nBits + 7) >> 3 + for i := uint8(0); i < nbBytes; i++ { + b.out = append(b.out, byte(b.bitContainer>>(i*8))) + } + b.nBits = 0 + b.bitContainer = 0 +} + +// close will write the alignment bit and write the final byte(s) +// to the output. +func (b *bitWriter) close() { + // End mark + b.addBits16Clean(1, 1) + // flush until next byte. + b.flushAlign() +} + +// reset and continue writing by appending to out. +func (b *bitWriter) reset(out []byte) { + b.bitContainer = 0 + b.nBits = 0 + b.out = out +} diff --git a/internal/github.com/klauspost/compress/zstd/blockenc.go b/internal/github.com/klauspost/compress/zstd/blockenc.go new file mode 100644 index 000000000..9d7962495 --- /dev/null +++ b/internal/github.com/klauspost/compress/zstd/blockenc.go @@ -0,0 +1,887 @@ +// Copyright 2019+ Klaus Post. All rights reserved. +// License information can be found in the LICENSE file. +// Based on work by Yann Collet, released under BSD License. + +package zstd + +import ( + "errors" + "fmt" + "math" + "math/bits" + "slices" + + "github.com/prometheus/client_golang/internal/github.com/klauspost/compress/huff0" +) + +type blockType uint8 + +//go:generate stringer -type=blockType,literalsBlockType,seqCompMode,tableIndex + +const ( + blockTypeRaw blockType = iota + blockTypeRLE + blockTypeCompressed + blockTypeReserved +) + +type literalsBlockType uint8 + +const ( + literalsBlockRaw literalsBlockType = iota + literalsBlockRLE + literalsBlockCompressed + literalsBlockTreeless +) + +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 + + maxMatchLen = 131074 + maxSequences = 0x7f00 + 0xffff + + // We support slightly less than the reference decoder to be able to + // use ints on 32 bit archs. + maxOffsetBits = 30 +) + +type blockEnc struct { + size int + literals []byte + sequences []seq + coders seqCoders + litEnc *huff0.Scratch + dictLitEnc *huff0.Scratch + wr bitWriter + + extraLits int + output []byte + recentOffsets [3]uint32 + prevRecentOffsets [3]uint32 + + last bool + lowMem bool +} + +// init should be used once the block has been created. +// If called more than once, the effect is the same as calling reset. +func (b *blockEnc) init() { + if b.lowMem { + // 1K literals + if cap(b.literals) < 1<<10 { + b.literals = make([]byte, 0, 1<<10) + } + const defSeqs = 20 + if cap(b.sequences) < defSeqs { + b.sequences = make([]seq, 0, defSeqs) + } + // 1K + if cap(b.output) < 1<<10 { + b.output = make([]byte, 0, 1<<10) + } + } else { + if cap(b.literals) < maxCompressedBlockSize { + b.literals = make([]byte, 0, maxCompressedBlockSize) + } + const defSeqs = 2000 + if cap(b.sequences) < defSeqs { + b.sequences = make([]seq, 0, defSeqs) + } + if cap(b.output) < maxCompressedBlockSize { + b.output = make([]byte, 0, maxCompressedBlockSize) + } + } + + if b.coders.mlEnc == nil { + b.coders.mlEnc = &fseEncoder{} + b.coders.mlPrev = &fseEncoder{} + b.coders.ofEnc = &fseEncoder{} + b.coders.ofPrev = &fseEncoder{} + b.coders.llEnc = &fseEncoder{} + b.coders.llPrev = &fseEncoder{} + } + b.litEnc = &huff0.Scratch{WantLogLess: 4} + b.reset(nil) +} + +// initNewEncode can be used to reset offsets and encoders to the initial state. +func (b *blockEnc) initNewEncode() { + b.recentOffsets = [3]uint32{1, 4, 8} + b.litEnc.Reuse = huff0.ReusePolicyNone + b.coders.setPrev(nil, nil, nil) +} + +// reset will reset the block for a new encode, but in the same stream, +// meaning that state will be carried over, but the block content is reset. +// If a previous block is provided, the recent offsets are carried over. +func (b *blockEnc) reset(prev *blockEnc) { + b.extraLits = 0 + b.literals = b.literals[:0] + b.size = 0 + b.sequences = b.sequences[:0] + b.output = b.output[:0] + b.last = false + if prev != nil { + b.recentOffsets = prev.prevRecentOffsets + } + b.dictLitEnc = nil +} + +// reset will reset the block for a new encode, but in the same stream, +// meaning that state will be carried over, but the block content is reset. +// If a previous block is provided, the recent offsets are carried over. +func (b *blockEnc) swapEncoders(prev *blockEnc) { + b.coders.swap(&prev.coders) + b.litEnc, prev.litEnc = prev.litEnc, b.litEnc +} + +// blockHeader contains the information for a block header. +type blockHeader uint32 + +// setLast sets the 'last' indicator on a block. +func (h *blockHeader) setLast(b bool) { + if b { + *h = *h | 1 + } else { + const mask = (1 << 24) - 2 + *h = *h & mask + } +} + +// setSize will store the compressed size of a block. +func (h *blockHeader) setSize(v uint32) { + const mask = 7 + *h = (*h)&mask | blockHeader(v<<3) +} + +// setType sets the block type. +func (h *blockHeader) setType(t blockType) { + const mask = 1 | (((1 << 24) - 1) ^ 7) + *h = (*h & mask) | blockHeader(t<<1) +} + +// appendTo will append the block header to a slice. +func (h blockHeader) appendTo(b []byte) []byte { + return append(b, uint8(h), uint8(h>>8), uint8(h>>16)) +} + +// String returns a string representation of the block. +func (h blockHeader) String() string { + return fmt.Sprintf("Type: %d, Size: %d, Last:%t", (h>>1)&3, h>>3, h&1 == 1) +} + +// literalsHeader contains literals header information. +type literalsHeader uint64 + +// setType can be used to set the type of literal block. +func (h *literalsHeader) setType(t literalsBlockType) { + const mask = math.MaxUint64 - 3 + *h = (*h & mask) | literalsHeader(t) +} + +// setSize can be used to set a single size, for uncompressed and RLE content. +func (h *literalsHeader) setSize(regenLen int) { + inBits := bits.Len32(uint32(regenLen)) + // Only retain 2 bits + const mask = 3 + lh := uint64(*h & mask) + switch { + case inBits < 5: + lh |= (uint64(regenLen) << 3) | (1 << 60) + if debugEncoder { + got := int(lh>>3) & 0xff + if got != regenLen { + panic(fmt.Sprint("litRegenSize = ", regenLen, "(want) != ", got, "(got)")) + } + } + case inBits < 12: + lh |= (1 << 2) | (uint64(regenLen) << 4) | (2 << 60) + case inBits < 20: + lh |= (3 << 2) | (uint64(regenLen) << 4) | (3 << 60) + default: + panic(fmt.Errorf("internal error: block too big (%d)", regenLen)) + } + *h = literalsHeader(lh) +} + +// setSizes will set the size of a compressed literals section and the input length. +func (h *literalsHeader) setSizes(compLen, inLen int, single bool) { + compBits, inBits := bits.Len32(uint32(compLen)), bits.Len32(uint32(inLen)) + // Only retain 2 bits + const mask = 3 + lh := uint64(*h & mask) + switch { + case compBits <= 10 && inBits <= 10: + if !single { + lh |= 1 << 2 + } + lh |= (uint64(inLen) << 4) | (uint64(compLen) << (10 + 4)) | (3 << 60) + if debugEncoder { + const mmask = (1 << 24) - 1 + n := (lh >> 4) & mmask + if int(n&1023) != inLen { + panic(fmt.Sprint("regensize:", int(n&1023), "!=", inLen, inBits)) + } + if int(n>>10) != compLen { + panic(fmt.Sprint("compsize:", int(n>>10), "!=", compLen, compBits)) + } + } + case compBits <= 14 && inBits <= 14: + lh |= (2 << 2) | (uint64(inLen) << 4) | (uint64(compLen) << (14 + 4)) | (4 << 60) + if single { + panic("single stream used with more than 10 bits length.") + } + case compBits <= 18 && inBits <= 18: + lh |= (3 << 2) | (uint64(inLen) << 4) | (uint64(compLen) << (18 + 4)) | (5 << 60) + if single { + panic("single stream used with more than 10 bits length.") + } + default: + panic("internal error: block too big") + } + *h = literalsHeader(lh) +} + +// appendTo will append the literals header to a byte slice. +func (h literalsHeader) appendTo(b []byte) []byte { + size := uint8(h >> 60) + switch size { + case 1: + b = append(b, uint8(h)) + case 2: + b = append(b, uint8(h), uint8(h>>8)) + case 3: + b = append(b, uint8(h), uint8(h>>8), uint8(h>>16)) + case 4: + b = append(b, uint8(h), uint8(h>>8), uint8(h>>16), uint8(h>>24)) + case 5: + b = append(b, uint8(h), uint8(h>>8), uint8(h>>16), uint8(h>>24), uint8(h>>32)) + default: + panic(fmt.Errorf("internal error: literalsHeader has invalid size (%d)", size)) + } + return b +} + +// size returns the output size with currently set values. +func (h literalsHeader) size() int { + return int(h >> 60) +} + +func (h literalsHeader) String() string { + return fmt.Sprintf("Type: %d, SizeFormat: %d, Size: 0x%d, Bytes:%d", literalsBlockType(h&3), (h>>2)&3, h&((1<<60)-1)>>4, h>>60) +} + +// pushOffsets will push the recent offsets to the backup store. +func (b *blockEnc) pushOffsets() { + b.prevRecentOffsets = b.recentOffsets +} + +// pushOffsets will push the recent offsets to the backup store. +func (b *blockEnc) popOffsets() { + b.recentOffsets = b.prevRecentOffsets +} + +// matchOffset will adjust recent offsets and return the adjusted one, +// if it matches a previous offset. +func (b *blockEnc) matchOffset(offset, lits uint32) uint32 { + // Check if offset is one of the recent offsets. + // Adjusts the output offset accordingly. + // Gives a tiny bit of compression, typically around 1%. + if true { + if lits > 0 { + switch offset { + case b.recentOffsets[0]: + offset = 1 + case b.recentOffsets[1]: + b.recentOffsets[1] = b.recentOffsets[0] + b.recentOffsets[0] = offset + offset = 2 + case b.recentOffsets[2]: + b.recentOffsets[2] = b.recentOffsets[1] + b.recentOffsets[1] = b.recentOffsets[0] + b.recentOffsets[0] = offset + offset = 3 + default: + b.recentOffsets[2] = b.recentOffsets[1] + b.recentOffsets[1] = b.recentOffsets[0] + b.recentOffsets[0] = offset + offset += 3 + } + } else { + switch offset { + case b.recentOffsets[1]: + b.recentOffsets[1] = b.recentOffsets[0] + b.recentOffsets[0] = offset + offset = 1 + case b.recentOffsets[2]: + b.recentOffsets[2] = b.recentOffsets[1] + b.recentOffsets[1] = b.recentOffsets[0] + b.recentOffsets[0] = offset + offset = 2 + case b.recentOffsets[0] - 1: + b.recentOffsets[2] = b.recentOffsets[1] + b.recentOffsets[1] = b.recentOffsets[0] + b.recentOffsets[0] = offset + offset = 3 + default: + b.recentOffsets[2] = b.recentOffsets[1] + b.recentOffsets[1] = b.recentOffsets[0] + b.recentOffsets[0] = offset + offset += 3 + } + } + } else { + offset += 3 + } + return offset +} + +// encodeRaw can be used to set the output to a raw representation of supplied bytes. +func (b *blockEnc) encodeRaw(a []byte) { + var bh blockHeader + bh.setLast(b.last) + bh.setSize(uint32(len(a))) + bh.setType(blockTypeRaw) + b.output = bh.appendTo(b.output[:0]) + b.output = append(b.output, a...) + if debugEncoder { + println("Adding RAW block, length", len(a), "last:", b.last) + } +} + +// encodeRaw can be used to set the output to a raw representation of supplied bytes. +func (b *blockEnc) encodeRawTo(dst, src []byte) []byte { + var bh blockHeader + bh.setLast(b.last) + bh.setSize(uint32(len(src))) + bh.setType(blockTypeRaw) + dst = bh.appendTo(dst) + dst = append(dst, src...) + if debugEncoder { + println("Adding RAW block, length", len(src), "last:", b.last) + } + return dst +} + +// encodeLits can be used if the block is only litLen. +func (b *blockEnc) encodeLits(lits []byte, raw bool) error { + var bh blockHeader + bh.setLast(b.last) + bh.setSize(uint32(len(lits))) + + // Don't compress extremely small blocks + if len(lits) < 8 || (len(lits) < 32 && b.dictLitEnc == nil) || raw { + if debugEncoder { + println("Adding RAW block, length", len(lits), "last:", b.last) + } + bh.setType(blockTypeRaw) + b.output = bh.appendTo(b.output) + b.output = append(b.output, lits...) + return nil + } + + var ( + out []byte + reUsed, single bool + err error + ) + if b.dictLitEnc != nil { + b.litEnc.TransferCTable(b.dictLitEnc) + b.litEnc.Reuse = huff0.ReusePolicyAllow + b.dictLitEnc = nil + } + if len(lits) >= 1024 { + // Use 4 Streams. + out, reUsed, err = huff0.Compress4X(lits, b.litEnc) + } else if len(lits) > 16 { + // Use 1 stream + single = true + out, reUsed, err = huff0.Compress1X(lits, b.litEnc) + } else { + err = huff0.ErrIncompressible + } + if err == nil && len(out)+5 > len(lits) { + // If we are close, we may still be worse or equal to raw. + var lh literalsHeader + lh.setSizes(len(out), len(lits), single) + if len(out)+lh.size() >= len(lits) { + err = huff0.ErrIncompressible + } + } + switch err { + case huff0.ErrIncompressible: + if debugEncoder { + println("Adding RAW block, length", len(lits), "last:", b.last) + } + bh.setType(blockTypeRaw) + b.output = bh.appendTo(b.output) + b.output = append(b.output, lits...) + return nil + case huff0.ErrUseRLE: + if debugEncoder { + println("Adding RLE block, length", len(lits)) + } + bh.setType(blockTypeRLE) + b.output = bh.appendTo(b.output) + b.output = append(b.output, lits[0]) + return nil + case nil: + default: + return err + } + // Compressed... + // Now, allow reuse + b.litEnc.Reuse = huff0.ReusePolicyAllow + bh.setType(blockTypeCompressed) + var lh literalsHeader + if reUsed { + if debugEncoder { + println("Reused tree, compressed to", len(out)) + } + lh.setType(literalsBlockTreeless) + } else { + if debugEncoder { + println("New tree, compressed to", len(out), "tree size:", len(b.litEnc.OutTable)) + } + lh.setType(literalsBlockCompressed) + } + // Set sizes + lh.setSizes(len(out), len(lits), single) + bh.setSize(uint32(len(out) + lh.size() + 1)) + + // Write block headers. + b.output = bh.appendTo(b.output) + b.output = lh.appendTo(b.output) + // Add compressed data. + b.output = append(b.output, out...) + // No sequences. + b.output = append(b.output, 0) + return nil +} + +// encodeRLE will encode an RLE block. +func (b *blockEnc) encodeRLE(val byte, length uint32) { + var bh blockHeader + bh.setLast(b.last) + bh.setSize(length) + bh.setType(blockTypeRLE) + b.output = bh.appendTo(b.output) + b.output = append(b.output, val) +} + +// encode will encode the block and append the output in b.output. +// Previous offset codes must be pushed if more blocks are expected. +func (b *blockEnc) encode(org []byte, raw, rawAllLits bool) error { + if len(b.sequences) == 0 { + return b.encodeLits(b.literals, rawAllLits) + } + if len(b.sequences) == 1 && len(org) > 0 && len(b.literals) <= 1 { + // Check common RLE cases. + seq := b.sequences[0] + if seq.litLen == uint32(len(b.literals)) && seq.offset-3 == 1 { + // Offset == 1 and 0 or 1 literals. + b.encodeRLE(org[0], b.sequences[0].matchLen+zstdMinMatch+seq.litLen) + return nil + } + } + + // We want some difference to at least account for the headers. + saved := b.size - len(b.literals) - (b.size >> 6) + if saved < 16 { + if org == nil { + return errIncompressible + } + b.popOffsets() + return b.encodeLits(org, rawAllLits) + } + + var bh blockHeader + var lh literalsHeader + bh.setLast(b.last) + bh.setType(blockTypeCompressed) + // Store offset of the block header. Needed when we know the size. + bhOffset := len(b.output) + b.output = bh.appendTo(b.output) + + var ( + out []byte + reUsed, single bool + err error + ) + if b.dictLitEnc != nil { + b.litEnc.TransferCTable(b.dictLitEnc) + b.litEnc.Reuse = huff0.ReusePolicyAllow + b.dictLitEnc = nil + } + if len(b.literals) >= 1024 && !raw { + // Use 4 Streams. + out, reUsed, err = huff0.Compress4X(b.literals, b.litEnc) + } else if len(b.literals) > 16 && !raw { + // Use 1 stream + single = true + out, reUsed, err = huff0.Compress1X(b.literals, b.litEnc) + } else { + err = huff0.ErrIncompressible + } + + if err == nil && len(out)+5 > len(b.literals) { + // If we are close, we may still be worse or equal to raw. + var lh literalsHeader + lh.setSize(len(b.literals)) + szRaw := lh.size() + lh.setSizes(len(out), len(b.literals), single) + szComp := lh.size() + if len(out)+szComp >= len(b.literals)+szRaw { + err = huff0.ErrIncompressible + } + } + switch err { + case huff0.ErrIncompressible: + lh.setType(literalsBlockRaw) + lh.setSize(len(b.literals)) + b.output = lh.appendTo(b.output) + b.output = append(b.output, b.literals...) + if debugEncoder { + println("Adding literals RAW, length", len(b.literals)) + } + case huff0.ErrUseRLE: + lh.setType(literalsBlockRLE) + lh.setSize(len(b.literals)) + b.output = lh.appendTo(b.output) + b.output = append(b.output, b.literals[0]) + if debugEncoder { + println("Adding literals RLE") + } + case nil: + // Compressed litLen... + if reUsed { + if debugEncoder { + println("reused tree") + } + lh.setType(literalsBlockTreeless) + } else { + if debugEncoder { + println("new tree, size:", len(b.litEnc.OutTable)) + } + lh.setType(literalsBlockCompressed) + } + lh.setSizes(len(out), len(b.literals), single) + if debugEncoder { + printf("Compressed %d literals to %d bytes", len(b.literals), len(out)) + println("Adding literal header:", lh) + } + b.output = lh.appendTo(b.output) + b.output = append(b.output, out...) + b.litEnc.Reuse = huff0.ReusePolicyAllow + if debugEncoder { + println("Adding literals compressed") + } + default: + if debugEncoder { + println("Adding literals ERROR:", err) + } + return err + } + // Sequence compression + + // Write the number of sequences + switch { + case len(b.sequences) < 128: + b.output = append(b.output, uint8(len(b.sequences))) + case len(b.sequences) < 0x7f00: // TODO: this could be wrong + n := len(b.sequences) + b.output = append(b.output, 128+uint8(n>>8), uint8(n)) + default: + n := len(b.sequences) - 0x7f00 + b.output = append(b.output, 255, uint8(n), uint8(n>>8)) + } + if debugEncoder { + println("Encoding", len(b.sequences), "sequences") + } + b.genCodes() + llEnc := b.coders.llEnc + ofEnc := b.coders.ofEnc + mlEnc := b.coders.mlEnc + err = llEnc.normalizeCount(len(b.sequences)) + if err != nil { + return err + } + err = ofEnc.normalizeCount(len(b.sequences)) + if err != nil { + return err + } + err = mlEnc.normalizeCount(len(b.sequences)) + if err != nil { + return err + } + + // Choose the best compression mode for each type. + // Will evaluate the new vs predefined and previous. + chooseComp := func(cur, prev, preDef *fseEncoder) (*fseEncoder, seqCompMode) { + // See if predefined/previous is better + hist := cur.count[:cur.symbolLen] + nSize := cur.approxSize(hist) + cur.maxHeaderSize() + predefSize := preDef.approxSize(hist) + prevSize := prev.approxSize(hist) + + // Add a small penalty for new encoders. + // Don't bother with extremely small (<2 byte gains). + nSize = nSize + (nSize+2*8*16)>>4 + switch { + case predefSize <= prevSize && predefSize <= nSize || forcePreDef: + if debugEncoder { + println("Using predefined", predefSize>>3, "<=", nSize>>3) + } + return preDef, compModePredefined + case prevSize <= nSize: + if debugEncoder { + println("Using previous", prevSize>>3, "<=", nSize>>3) + } + return prev, compModeRepeat + default: + if debugEncoder { + println("Using new, predef", predefSize>>3, ". previous:", prevSize>>3, ">", nSize>>3, "header max:", cur.maxHeaderSize()>>3, "bytes") + println("tl:", cur.actualTableLog, "symbolLen:", cur.symbolLen, "norm:", cur.norm[:cur.symbolLen], "hist", cur.count[:cur.symbolLen]) + } + return cur, compModeFSE + } + } + + // Write compression mode + var mode uint8 + if llEnc.useRLE { + mode |= uint8(compModeRLE) << 6 + llEnc.setRLE(b.sequences[0].llCode) + if debugEncoder { + println("llEnc.useRLE") + } + } else { + var m seqCompMode + llEnc, m = chooseComp(llEnc, b.coders.llPrev, &fsePredefEnc[tableLiteralLengths]) + mode |= uint8(m) << 6 + } + if ofEnc.useRLE { + mode |= uint8(compModeRLE) << 4 + ofEnc.setRLE(b.sequences[0].ofCode) + if debugEncoder { + println("ofEnc.useRLE") + } + } else { + var m seqCompMode + ofEnc, m = chooseComp(ofEnc, b.coders.ofPrev, &fsePredefEnc[tableOffsets]) + mode |= uint8(m) << 4 + } + + if mlEnc.useRLE { + mode |= uint8(compModeRLE) << 2 + mlEnc.setRLE(b.sequences[0].mlCode) + if debugEncoder { + println("mlEnc.useRLE, code: ", b.sequences[0].mlCode, "value", b.sequences[0].matchLen) + } + } else { + var m seqCompMode + mlEnc, m = chooseComp(mlEnc, b.coders.mlPrev, &fsePredefEnc[tableMatchLengths]) + mode |= uint8(m) << 2 + } + b.output = append(b.output, mode) + if debugEncoder { + printf("Compression modes: 0b%b", mode) + } + b.output, err = llEnc.writeCount(b.output) + if err != nil { + return err + } + start := len(b.output) + b.output, err = ofEnc.writeCount(b.output) + if err != nil { + return err + } + if false { + println("block:", b.output[start:], "tablelog", ofEnc.actualTableLog, "maxcount:", ofEnc.maxCount) + fmt.Printf("selected TableLog: %d, Symbol length: %d\n", ofEnc.actualTableLog, ofEnc.symbolLen) + for i, v := range ofEnc.norm[:ofEnc.symbolLen] { + fmt.Printf("%3d: %5d -> %4d \n", i, ofEnc.count[i], v) + } + } + b.output, err = mlEnc.writeCount(b.output) + if err != nil { + return err + } + + // Maybe in block? + wr := &b.wr + wr.reset(b.output) + + var ll, of, ml cState + + // Current sequence + seq := len(b.sequences) - 1 + s := b.sequences[seq] + llEnc.setBits(llBitsTable[:]) + mlEnc.setBits(mlBitsTable[:]) + ofEnc.setBits(nil) + + llTT, ofTT, mlTT := llEnc.ct.symbolTT[:256], ofEnc.ct.symbolTT[:256], mlEnc.ct.symbolTT[:256] + + // We have 3 bounds checks here (and in the loop). + // Since we are iterating backwards it is kinda hard to avoid. + llB, ofB, mlB := llTT[s.llCode], ofTT[s.ofCode], mlTT[s.mlCode] + ll.init(wr, &llEnc.ct, llB) + of.init(wr, &ofEnc.ct, ofB) + wr.flush32() + ml.init(wr, &mlEnc.ct, mlB) + + // Each of these lookups also generates a bounds check. + wr.addBits32NC(s.litLen, llB.outBits) + wr.addBits32NC(s.matchLen, mlB.outBits) + wr.flush32() + wr.addBits32NC(s.offset, ofB.outBits) + if debugSequences { + println("Encoded seq", seq, s, "codes:", s.llCode, s.mlCode, s.ofCode, "states:", ll.state, ml.state, of.state, "bits:", llB, mlB, ofB) + } + seq-- + // Store sequences in reverse... + for seq >= 0 { + s = b.sequences[seq] + + ofB := ofTT[s.ofCode] + wr.flush32() // tablelog max is below 8 for each, so it will fill max 24 bits. + //of.encode(ofB) + nbBitsOut := (uint32(of.state) + ofB.deltaNbBits) >> 16 + dstState := int32(of.state>>(nbBitsOut&15)) + int32(ofB.deltaFindState) + wr.addBits16NC(of.state, uint8(nbBitsOut)) + of.state = of.stateTable[dstState] + + // Accumulate extra bits. + outBits := ofB.outBits & 31 + extraBits := uint64(s.offset & bitMask32[outBits]) + extraBitsN := outBits + + mlB := mlTT[s.mlCode] + //ml.encode(mlB) + nbBitsOut = (uint32(ml.state) + mlB.deltaNbBits) >> 16 + dstState = int32(ml.state>>(nbBitsOut&15)) + int32(mlB.deltaFindState) + wr.addBits16NC(ml.state, uint8(nbBitsOut)) + ml.state = ml.stateTable[dstState] + + outBits = mlB.outBits & 31 + extraBits = extraBits<> 16 + dstState = int32(ll.state>>(nbBitsOut&15)) + int32(llB.deltaFindState) + wr.addBits16NC(ll.state, uint8(nbBitsOut)) + ll.state = ll.stateTable[dstState] + + outBits = llB.outBits & 31 + extraBits = extraBits<= b.size { + // Discard and encode as raw block. + b.output = b.encodeRawTo(b.output[:bhOffset], org) + b.popOffsets() + b.litEnc.Reuse = huff0.ReusePolicyNone + return nil + } + + // Size is output minus block header. + bh.setSize(uint32(len(b.output)-bhOffset) - 3) + if debugEncoder { + println("Rewriting block header", bh) + } + _ = bh.appendTo(b.output[bhOffset:bhOffset]) + b.coders.setPrev(llEnc, mlEnc, ofEnc) + return nil +} + +var errIncompressible = errors.New("incompressible") + +func (b *blockEnc) genCodes() { + if len(b.sequences) == 0 { + // nothing to do + return + } + if len(b.sequences) > math.MaxUint16 { + panic("can only encode up to 64K sequences") + } + // No bounds checks after here: + llH := b.coders.llEnc.Histogram() + ofH := b.coders.ofEnc.Histogram() + mlH := b.coders.mlEnc.Histogram() + for i := range llH { + llH[i] = 0 + } + for i := range ofH { + ofH[i] = 0 + } + for i := range mlH { + mlH[i] = 0 + } + + var llMax, ofMax, mlMax uint8 + for i := range b.sequences { + seq := &b.sequences[i] + v := llCode(seq.litLen) + seq.llCode = v + llH[v]++ + if v > llMax { + llMax = v + } + + v = ofCode(seq.offset) + seq.ofCode = v + ofH[v]++ + if v > ofMax { + ofMax = v + } + + v = mlCode(seq.matchLen) + seq.mlCode = v + mlH[v]++ + if v > mlMax { + mlMax = v + if debugAsserts && mlMax > maxMatchLengthSymbol { + panic(fmt.Errorf("mlMax > maxMatchLengthSymbol (%d), matchlen: %d", mlMax, seq.matchLen)) + } + } + } + if debugAsserts && mlMax > maxMatchLengthSymbol { + panic(fmt.Errorf("mlMax > maxMatchLengthSymbol (%d)", mlMax)) + } + if debugAsserts && ofMax > maxOffsetBits { + panic(fmt.Errorf("ofMax > maxOffsetBits (%d)", ofMax)) + } + if debugAsserts && llMax > maxLiteralLengthSymbol { + panic(fmt.Errorf("llMax > maxLiteralLengthSymbol (%d)", llMax)) + } + + b.coders.mlEnc.HistogramFinished(mlMax, int(slices.Max(mlH[:mlMax+1]))) + b.coders.ofEnc.HistogramFinished(ofMax, int(slices.Max(ofH[:ofMax+1]))) + b.coders.llEnc.HistogramFinished(llMax, int(slices.Max(llH[:llMax+1]))) +} diff --git a/internal/github.com/klauspost/compress/zstd/bytereader.go b/internal/github.com/klauspost/compress/zstd/bytereader.go new file mode 100644 index 000000000..0e59a242d --- /dev/null +++ b/internal/github.com/klauspost/compress/zstd/bytereader.go @@ -0,0 +1,82 @@ +// Copyright 2019+ Klaus Post. All rights reserved. +// License information can be found in the LICENSE file. +// Based on work by Yann Collet, released under BSD License. + +package zstd + +// byteReader provides a byte reader that reads +// little endian values from a byte stream. +// The input stream is manually advanced. +// The reader performs no bounds checks. +type byteReader struct { + b []byte + off int +} + +// advance the stream b n bytes. +func (b *byteReader) advance(n uint) { + b.off += int(n) +} + +// overread returns whether we have advanced too far. +func (b *byteReader) overread() bool { + return b.off > len(b.b) +} + +// Int32 returns a little endian int32 starting at current offset. +func (b byteReader) Int32() int32 { + b2 := b.b[b.off:] + b2 = b2[:4] + v3 := int32(b2[3]) + v2 := int32(b2[2]) + v1 := int32(b2[1]) + v0 := int32(b2[0]) + return v0 | (v1 << 8) | (v2 << 16) | (v3 << 24) +} + +// Uint8 returns the next byte +func (b *byteReader) Uint8() uint8 { + v := b.b[b.off] + return v +} + +// Uint32 returns a little endian uint32 starting at current offset. +func (b byteReader) Uint32() uint32 { + if r := b.remain(); r < 4 { + // Very rare + v := uint32(0) + for i := 1; i <= r; i++ { + v = (v << 8) | uint32(b.b[len(b.b)-i]) + } + return v + } + b2 := b.b[b.off:] + b2 = b2[:4] + v3 := uint32(b2[3]) + v2 := uint32(b2[2]) + v1 := uint32(b2[1]) + v0 := uint32(b2[0]) + return v0 | (v1 << 8) | (v2 << 16) | (v3 << 24) +} + +// Uint32NC returns a little endian uint32 starting at current offset. +// The caller must be sure if there are at least 4 bytes left. +func (b byteReader) Uint32NC() uint32 { + b2 := b.b[b.off:] + b2 = b2[:4] + v3 := uint32(b2[3]) + v2 := uint32(b2[2]) + v1 := uint32(b2[1]) + v0 := uint32(b2[0]) + return v0 | (v1 << 8) | (v2 << 16) | (v3 << 24) +} + +// unread returns the unread portion of the input. +func (b byteReader) unread() []byte { + return b.b[b.off:] +} + +// remain will return the number of bytes remaining. +func (b byteReader) remain() int { + return len(b.b) - b.off +} diff --git a/internal/github.com/klauspost/compress/zstd/dict.go b/internal/github.com/klauspost/compress/zstd/dict.go new file mode 100644 index 000000000..954b75bd0 --- /dev/null +++ b/internal/github.com/klauspost/compress/zstd/dict.go @@ -0,0 +1,58 @@ +package zstd + +import ( + "github.com/prometheus/client_golang/internal/github.com/klauspost/compress/huff0" +) + +type dict struct { + id uint32 + + litEnc *huff0.Scratch + offsets [3]int + content []byte +} + +const dictMagic = "\x37\xa4\x30\xec" + +// Maximum dictionary size for the reference implementation (1.5.3) is 2 GiB. +const dictMaxLength = 1 << 31 + +// ID returns the dictionary id or 0 if d is nil. +func (d *dict) ID() uint32 { + if d == nil { + return 0 + } + return d.id +} + +// ContentSize returns the dictionary content size or 0 if d is nil. +func (d *dict) ContentSize() int { + if d == nil { + return 0 + } + return len(d.content) +} + +// Content returns the dictionary content. +func (d *dict) Content() []byte { + if d == nil { + return nil + } + return d.content +} + +// Offsets returns the initial offsets. +func (d *dict) Offsets() [3]int { + if d == nil { + return [3]int{} + } + return d.offsets +} + +// LitEncoder returns the literal encoder. +func (d *dict) LitEncoder() *huff0.Scratch { + if d == nil { + return nil + } + return d.litEnc +} diff --git a/internal/github.com/klauspost/compress/zstd/enc_base.go b/internal/github.com/klauspost/compress/zstd/enc_base.go new file mode 100644 index 000000000..64ab0afb7 --- /dev/null +++ b/internal/github.com/klauspost/compress/zstd/enc_base.go @@ -0,0 +1,173 @@ +package zstd + +import ( + "fmt" + "math/bits" + + "github.com/cespare/xxhash/v2" +) + +const ( + dictShardBits = 6 +) + +type fastBase struct { + // cur is the offset at the start of hist + cur int32 + // maximum offset. Should be at least 2x block size. + maxMatchOff int32 + bufferReset int32 + hist []byte + crc *xxhash.Digest + tmp [8]byte + blk *blockEnc + lastDictID uint32 + lowMem bool +} + +// CRC returns the underlying CRC writer. +func (e *fastBase) CRC() *xxhash.Digest { + return e.crc +} + +// AppendCRC will append the CRC to the destination slice and return it. +func (e *fastBase) AppendCRC(dst []byte) []byte { + crc := e.crc.Sum(e.tmp[:0]) + dst = append(dst, crc[7], crc[6], crc[5], crc[4]) + return dst +} + +// WindowSize returns the window size of the encoder, +// or a window size small enough to contain the input size, if > 0. +func (e *fastBase) WindowSize(size int64) int32 { + if size > 0 && size < int64(e.maxMatchOff) { + b := int32(1) << uint(bits.Len(uint(size))) + // Keep minimum window. + if b < 1024 { + b = 1024 + } + return b + } + return e.maxMatchOff +} + +// Block returns the current block. +func (e *fastBase) Block() *blockEnc { + return e.blk +} + +func (e *fastBase) addBlock(src []byte) int32 { + if debugAsserts && e.cur > e.bufferReset { + panic(fmt.Sprintf("ecur (%d) > buffer reset (%d)", e.cur, e.bufferReset)) + } + // check if we have space already + if len(e.hist)+len(src) > cap(e.hist) { + if cap(e.hist) == 0 { + e.ensureHist(len(src)) + } else { + if cap(e.hist) < int(e.maxMatchOff+maxCompressedBlockSize) { + panic(fmt.Errorf("unexpected buffer cap %d, want at least %d with window %d", cap(e.hist), e.maxMatchOff+maxCompressedBlockSize, e.maxMatchOff)) + } + // Move down + offset := int32(len(e.hist)) - e.maxMatchOff + copy(e.hist[0:e.maxMatchOff], e.hist[offset:]) + e.cur += offset + e.hist = e.hist[:e.maxMatchOff] + } + } + s := int32(len(e.hist)) + e.hist = append(e.hist, src...) + return s +} + +// ensureHist will ensure that history can keep at least this many bytes. +func (e *fastBase) ensureHist(n int) { + if cap(e.hist) >= n { + return + } + l := e.maxMatchOff + if (e.lowMem && e.maxMatchOff > maxCompressedBlockSize) || e.maxMatchOff <= maxCompressedBlockSize { + l += maxCompressedBlockSize + } else { + l += e.maxMatchOff + } + // Make it at least 1MB. + if l < 1<<20 && !e.lowMem { + l = 1 << 20 + } + // Make it at least the requested size. + if l < int32(n) { + l = int32(n) + } + e.hist = make([]byte, 0, l) +} + +// useBlock will replace the block with the provided one, +// but transfer recent offsets from the previous. +func (e *fastBase) UseBlock(enc *blockEnc) { + enc.reset(e.blk) + e.blk = enc +} + +func (e *fastBase) matchlen(s, t int32, src []byte) int32 { + if debugAsserts { + if s < 0 { + err := fmt.Sprintf("s (%d) < 0", s) + panic(err) + } + if t < 0 { + err := fmt.Sprintf("t (%d) < 0", t) + panic(err) + } + if s-t > e.maxMatchOff { + err := fmt.Sprintf("s (%d) - t (%d) > maxMatchOff (%d)", s, t, e.maxMatchOff) + panic(err) + } + if len(src)-int(s) > maxCompressedBlockSize { + panic(fmt.Sprintf("len(src)-s (%d) > maxCompressedBlockSize (%d)", len(src)-int(s), maxCompressedBlockSize)) + } + } + return int32(matchLen(src[s:], src[t:])) +} + +// Reset the encoding table. +func (e *fastBase) resetBase(d *dict, singleBlock bool) { + if e.blk == nil { + e.blk = &blockEnc{lowMem: e.lowMem} + e.blk.init() + } else { + e.blk.reset(nil) + } + e.blk.initNewEncode() + if e.crc == nil { + e.crc = xxhash.New() + } else { + e.crc.Reset() + } + e.blk.dictLitEnc = nil + if d != nil { + low := e.lowMem + if singleBlock { + e.lowMem = true + } + e.ensureHist(d.ContentSize() + maxCompressedBlockSize) + e.lowMem = low + } + + // We offset current position so everything will be out of reach. + // If above reset line, history will be purged. + if e.cur < e.bufferReset { + e.cur += e.maxMatchOff + int32(len(e.hist)) + } + e.hist = e.hist[:0] + if d != nil { + // Set offsets (currently not used) + for i, off := range d.offsets { + e.blk.recentOffsets[i] = uint32(off) + e.blk.prevRecentOffsets[i] = e.blk.recentOffsets[i] + } + // Transfer litenc. + e.blk.dictLitEnc = d.litEnc + e.hist = append(e.hist, d.content...) + } +} diff --git a/internal/github.com/klauspost/compress/zstd/enc_fast.go b/internal/github.com/klauspost/compress/zstd/enc_fast.go new file mode 100644 index 000000000..f45a3da7d --- /dev/null +++ b/internal/github.com/klauspost/compress/zstd/enc_fast.go @@ -0,0 +1,891 @@ +// Copyright 2019+ Klaus Post. All rights reserved. +// License information can be found in the LICENSE file. +// Based on work by Yann Collet, released under BSD License. + +package zstd + +import ( + "fmt" +) + +const ( + tableBits = 15 // Bits used in the table + tableSize = 1 << tableBits // Size of the table + tableShardCnt = 1 << (tableBits - dictShardBits) // Number of shards in the table + tableShardSize = tableSize / tableShardCnt // Size of an individual shard + tableFastHashLen = 6 + tableMask = tableSize - 1 // Mask for table indices. Redundant, but can eliminate bounds checks. + maxMatchLength = 131074 +) + +type tableEntry struct { + val uint32 + offset int32 +} + +type fastEncoder struct { + fastBase + table [tableSize]tableEntry +} + +type fastEncoderDict struct { + fastEncoder + dictTable []tableEntry + tableShardDirty [tableShardCnt]bool + allDirty bool +} + +// Encode mimmics functionality in zstd_fast.c +func (e *fastEncoder) Encode(blk *blockEnc, src []byte) { + const ( + inputMargin = 8 + minNonLiteralBlockSize = 1 + 1 + inputMargin + ) + + // Protect against e.cur wraparound. + for e.cur >= e.bufferReset-int32(len(e.hist)) { + if len(e.hist) == 0 { + for i := range e.table[:] { + e.table[i] = tableEntry{} + } + e.cur = e.maxMatchOff + break + } + // Shift down everything in the table that isn't already too far away. + minOff := e.cur + int32(len(e.hist)) - e.maxMatchOff + for i := range e.table[:] { + v := e.table[i].offset + if v < minOff { + v = 0 + } else { + v = v - e.cur + e.maxMatchOff + } + e.table[i].offset = v + } + e.cur = e.maxMatchOff + break + } + + s := e.addBlock(src) + blk.size = len(src) + if len(src) < minNonLiteralBlockSize { + blk.extraLits = len(src) + blk.literals = blk.literals[:len(src)] + copy(blk.literals, src) + return + } + + // Override src + src = e.hist + sLimit := int32(len(src)) - inputMargin + // stepSize is the number of bytes to skip on every main loop iteration. + // It should be >= 2. + const stepSize = 2 + + // TEMPLATE + const hashLog = tableBits + // seems global, but would be nice to tweak. + const kSearchStrength = 6 + + // nextEmit is where in src the next emitLiteral should start from. + nextEmit := s + cv := load6432(src, s) + + // Relative offsets + offset1 := int32(blk.recentOffsets[0]) + offset2 := int32(blk.recentOffsets[1]) + + addLiterals := func(s *seq, until int32) { + if until == nextEmit { + return + } + blk.literals = append(blk.literals, src[nextEmit:until]...) + s.litLen = uint32(until - nextEmit) + } + if debugEncoder { + println("recent offsets:", blk.recentOffsets) + } + +encodeLoop: + for { + // t will contain the match offset when we find one. + // When existing the search loop, we have already checked 4 bytes. + var t int32 + + // We will not use repeat offsets across blocks. + // By not using them for the first 3 matches + canRepeat := len(blk.sequences) > 2 + + for { + if debugAsserts && canRepeat && offset1 == 0 { + panic("offset0 was 0") + } + + nextHash := hashLen(cv, hashLog, tableFastHashLen) + nextHash2 := hashLen(cv>>8, hashLog, tableFastHashLen) + candidate := e.table[nextHash] + candidate2 := e.table[nextHash2] + repIndex := s - offset1 + 2 + + e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)} + e.table[nextHash2] = tableEntry{offset: s + e.cur + 1, val: uint32(cv >> 8)} + + if canRepeat && repIndex >= 0 && load3232(src, repIndex) == uint32(cv>>16) { + // Consider history as well. + var seq seq + length := 4 + e.matchlen(s+6, repIndex+4, src) + seq.matchLen = uint32(length - zstdMinMatch) + + // We might be able to match backwards. + // Extend as long as we can. + start := s + 2 + // We end the search early, so we don't risk 0 literals + // and have to do special offset treatment. + startLimit := nextEmit + 1 + + sMin := s - e.maxMatchOff + if sMin < 0 { + sMin = 0 + } + for repIndex > sMin && start > startLimit && src[repIndex-1] == src[start-1] && seq.matchLen < maxMatchLength-zstdMinMatch { + repIndex-- + start-- + seq.matchLen++ + } + addLiterals(&seq, start) + + // rep 0 + seq.offset = 1 + if debugSequences { + println("repeat sequence", seq, "next s:", s) + } + blk.sequences = append(blk.sequences, seq) + s += length + 2 + nextEmit = s + if s >= sLimit { + if debugEncoder { + println("repeat ended", s, length) + + } + break encodeLoop + } + cv = load6432(src, s) + continue + } + coffset0 := s - (candidate.offset - e.cur) + coffset1 := s - (candidate2.offset - e.cur) + 1 + if coffset0 < e.maxMatchOff && uint32(cv) == candidate.val { + // found a regular match + t = candidate.offset - e.cur + if debugAsserts && s <= t { + panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) + } + if debugAsserts && s-t > e.maxMatchOff { + panic("s - t >e.maxMatchOff") + } + break + } + + if coffset1 < e.maxMatchOff && uint32(cv>>8) == candidate2.val { + // found a regular match + t = candidate2.offset - e.cur + s++ + if debugAsserts && s <= t { + panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) + } + if debugAsserts && s-t > e.maxMatchOff { + panic("s - t >e.maxMatchOff") + } + if debugAsserts && t < 0 { + panic("t<0") + } + break + } + s += stepSize + ((s - nextEmit) >> (kSearchStrength - 1)) + if s >= sLimit { + break encodeLoop + } + cv = load6432(src, s) + } + // A 4-byte match has been found. We'll later see if more than 4 bytes. + offset2 = offset1 + offset1 = s - t + + if debugAsserts && s <= t { + panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) + } + + if debugAsserts && canRepeat && int(offset1) > len(src) { + panic("invalid offset") + } + + // Extend the 4-byte match as long as possible. + l := e.matchlen(s+4, t+4, src) + 4 + + // Extend backwards + tMin := s - e.maxMatchOff + if tMin < 0 { + tMin = 0 + } + for t > tMin && s > nextEmit && src[t-1] == src[s-1] && l < maxMatchLength { + s-- + t-- + l++ + } + + // Write our sequence. + var seq seq + seq.litLen = uint32(s - nextEmit) + seq.matchLen = uint32(l - zstdMinMatch) + if seq.litLen > 0 { + blk.literals = append(blk.literals, src[nextEmit:s]...) + } + // Don't use repeat offsets + seq.offset = uint32(s-t) + 3 + s += l + if debugSequences { + println("sequence", seq, "next s:", s) + } + blk.sequences = append(blk.sequences, seq) + nextEmit = s + if s >= sLimit { + break encodeLoop + } + cv = load6432(src, s) + + // Check offset 2 + if o2 := s - offset2; canRepeat && load3232(src, o2) == uint32(cv) { + // We have at least 4 byte match. + // No need to check backwards. We come straight from a match + l := 4 + e.matchlen(s+4, o2+4, src) + + // Store this, since we have it. + nextHash := hashLen(cv, hashLog, tableFastHashLen) + e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)} + seq.matchLen = uint32(l) - zstdMinMatch + seq.litLen = 0 + // Since litlen is always 0, this is offset 1. + seq.offset = 1 + s += l + nextEmit = s + if debugSequences { + println("sequence", seq, "next s:", s) + } + blk.sequences = append(blk.sequences, seq) + + // Swap offset 1 and 2. + offset1, offset2 = offset2, offset1 + if s >= sLimit { + break encodeLoop + } + // Prepare next loop. + cv = load6432(src, s) + } + } + + if int(nextEmit) < len(src) { + blk.literals = append(blk.literals, src[nextEmit:]...) + blk.extraLits = len(src) - int(nextEmit) + } + blk.recentOffsets[0] = uint32(offset1) + blk.recentOffsets[1] = uint32(offset2) + if debugEncoder { + println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits) + } +} + +// EncodeNoHist will encode a block with no history and no following blocks. +// Most notable difference is that src will not be copied for history and +// we do not need to check for max match length. +func (e *fastEncoder) EncodeNoHist(blk *blockEnc, src []byte) { + const ( + inputMargin = 8 + minNonLiteralBlockSize = 1 + 1 + inputMargin + ) + if debugEncoder { + if len(src) > maxCompressedBlockSize { + panic("src too big") + } + } + + // Protect against e.cur wraparound. + if e.cur >= e.bufferReset { + for i := range e.table[:] { + e.table[i] = tableEntry{} + } + e.cur = e.maxMatchOff + } + + s := int32(0) + blk.size = len(src) + if len(src) < minNonLiteralBlockSize { + blk.extraLits = len(src) + blk.literals = blk.literals[:len(src)] + copy(blk.literals, src) + return + } + + sLimit := int32(len(src)) - inputMargin + // stepSize is the number of bytes to skip on every main loop iteration. + // It should be >= 2. + const stepSize = 2 + + // TEMPLATE + const hashLog = tableBits + // seems global, but would be nice to tweak. + const kSearchStrength = 6 + + // nextEmit is where in src the next emitLiteral should start from. + nextEmit := s + cv := load6432(src, s) + + // Relative offsets + offset1 := int32(blk.recentOffsets[0]) + offset2 := int32(blk.recentOffsets[1]) + + addLiterals := func(s *seq, until int32) { + if until == nextEmit { + return + } + blk.literals = append(blk.literals, src[nextEmit:until]...) + s.litLen = uint32(until - nextEmit) + } + if debugEncoder { + println("recent offsets:", blk.recentOffsets) + } + +encodeLoop: + for { + // t will contain the match offset when we find one. + // When existing the search loop, we have already checked 4 bytes. + var t int32 + + // We will not use repeat offsets across blocks. + // By not using them for the first 3 matches + + for { + nextHash := hashLen(cv, hashLog, tableFastHashLen) + nextHash2 := hashLen(cv>>8, hashLog, tableFastHashLen) + candidate := e.table[nextHash] + candidate2 := e.table[nextHash2] + repIndex := s - offset1 + 2 + + e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)} + e.table[nextHash2] = tableEntry{offset: s + e.cur + 1, val: uint32(cv >> 8)} + + if len(blk.sequences) > 2 && load3232(src, repIndex) == uint32(cv>>16) { + // Consider history as well. + var seq seq + length := 4 + e.matchlen(s+6, repIndex+4, src) + + seq.matchLen = uint32(length - zstdMinMatch) + + // We might be able to match backwards. + // Extend as long as we can. + start := s + 2 + // We end the search early, so we don't risk 0 literals + // and have to do special offset treatment. + startLimit := nextEmit + 1 + + sMin := s - e.maxMatchOff + if sMin < 0 { + sMin = 0 + } + for repIndex > sMin && start > startLimit && src[repIndex-1] == src[start-1] { + repIndex-- + start-- + seq.matchLen++ + } + addLiterals(&seq, start) + + // rep 0 + seq.offset = 1 + if debugSequences { + println("repeat sequence", seq, "next s:", s) + } + blk.sequences = append(blk.sequences, seq) + s += length + 2 + nextEmit = s + if s >= sLimit { + if debugEncoder { + println("repeat ended", s, length) + + } + break encodeLoop + } + cv = load6432(src, s) + continue + } + coffset0 := s - (candidate.offset - e.cur) + coffset1 := s - (candidate2.offset - e.cur) + 1 + if coffset0 < e.maxMatchOff && uint32(cv) == candidate.val { + // found a regular match + t = candidate.offset - e.cur + if debugAsserts && s <= t { + panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) + } + if debugAsserts && s-t > e.maxMatchOff { + panic("s - t >e.maxMatchOff") + } + if debugAsserts && t < 0 { + panic(fmt.Sprintf("t (%d) < 0, candidate.offset: %d, e.cur: %d, coffset0: %d, e.maxMatchOff: %d", t, candidate.offset, e.cur, coffset0, e.maxMatchOff)) + } + break + } + + if coffset1 < e.maxMatchOff && uint32(cv>>8) == candidate2.val { + // found a regular match + t = candidate2.offset - e.cur + s++ + if debugAsserts && s <= t { + panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) + } + if debugAsserts && s-t > e.maxMatchOff { + panic("s - t >e.maxMatchOff") + } + if debugAsserts && t < 0 { + panic("t<0") + } + break + } + s += stepSize + ((s - nextEmit) >> (kSearchStrength - 1)) + if s >= sLimit { + break encodeLoop + } + cv = load6432(src, s) + } + // A 4-byte match has been found. We'll later see if more than 4 bytes. + offset2 = offset1 + offset1 = s - t + + if debugAsserts && s <= t { + panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) + } + + if debugAsserts && t < 0 { + panic(fmt.Sprintf("t (%d) < 0 ", t)) + } + // Extend the 4-byte match as long as possible. + l := e.matchlen(s+4, t+4, src) + 4 + + // Extend backwards + tMin := s - e.maxMatchOff + if tMin < 0 { + tMin = 0 + } + for t > tMin && s > nextEmit && src[t-1] == src[s-1] { + s-- + t-- + l++ + } + + // Write our sequence. + var seq seq + seq.litLen = uint32(s - nextEmit) + seq.matchLen = uint32(l - zstdMinMatch) + if seq.litLen > 0 { + blk.literals = append(blk.literals, src[nextEmit:s]...) + } + // Don't use repeat offsets + seq.offset = uint32(s-t) + 3 + s += l + if debugSequences { + println("sequence", seq, "next s:", s) + } + blk.sequences = append(blk.sequences, seq) + nextEmit = s + if s >= sLimit { + break encodeLoop + } + cv = load6432(src, s) + + // Check offset 2 + if o2 := s - offset2; len(blk.sequences) > 2 && load3232(src, o2) == uint32(cv) { + // We have at least 4 byte match. + // No need to check backwards. We come straight from a match + l := 4 + e.matchlen(s+4, o2+4, src) + + // Store this, since we have it. + nextHash := hashLen(cv, hashLog, tableFastHashLen) + e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)} + seq.matchLen = uint32(l) - zstdMinMatch + seq.litLen = 0 + // Since litlen is always 0, this is offset 1. + seq.offset = 1 + s += l + nextEmit = s + if debugSequences { + println("sequence", seq, "next s:", s) + } + blk.sequences = append(blk.sequences, seq) + + // Swap offset 1 and 2. + offset1, offset2 = offset2, offset1 + if s >= sLimit { + break encodeLoop + } + // Prepare next loop. + cv = load6432(src, s) + } + } + + if int(nextEmit) < len(src) { + blk.literals = append(blk.literals, src[nextEmit:]...) + blk.extraLits = len(src) - int(nextEmit) + } + if debugEncoder { + println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits) + } + // We do not store history, so we must offset e.cur to avoid false matches for next user. + if e.cur < e.bufferReset { + e.cur += int32(len(src)) + } +} + +// Encode will encode the content, with a dictionary if initialized for it. +func (e *fastEncoderDict) Encode(blk *blockEnc, src []byte) { + const ( + inputMargin = 8 + minNonLiteralBlockSize = 1 + 1 + inputMargin + ) + if e.allDirty || len(src) > 32<<10 { + e.fastEncoder.Encode(blk, src) + e.allDirty = true + return + } + // Protect against e.cur wraparound. + for e.cur >= e.bufferReset-int32(len(e.hist)) { + if len(e.hist) == 0 { + e.table = [tableSize]tableEntry{} + e.cur = e.maxMatchOff + break + } + // Shift down everything in the table that isn't already too far away. + minOff := e.cur + int32(len(e.hist)) - e.maxMatchOff + for i := range e.table[:] { + v := e.table[i].offset + if v < minOff { + v = 0 + } else { + v = v - e.cur + e.maxMatchOff + } + e.table[i].offset = v + } + e.cur = e.maxMatchOff + break + } + + s := e.addBlock(src) + blk.size = len(src) + if len(src) < minNonLiteralBlockSize { + blk.extraLits = len(src) + blk.literals = blk.literals[:len(src)] + copy(blk.literals, src) + return + } + + // Override src + src = e.hist + sLimit := int32(len(src)) - inputMargin + // stepSize is the number of bytes to skip on every main loop iteration. + // It should be >= 2. + const stepSize = 2 + + // TEMPLATE + const hashLog = tableBits + // seems global, but would be nice to tweak. + const kSearchStrength = 7 + + // nextEmit is where in src the next emitLiteral should start from. + nextEmit := s + cv := load6432(src, s) + + // Relative offsets + offset1 := int32(blk.recentOffsets[0]) + offset2 := int32(blk.recentOffsets[1]) + + addLiterals := func(s *seq, until int32) { + if until == nextEmit { + return + } + blk.literals = append(blk.literals, src[nextEmit:until]...) + s.litLen = uint32(until - nextEmit) + } + if debugEncoder { + println("recent offsets:", blk.recentOffsets) + } + +encodeLoop: + for { + // t will contain the match offset when we find one. + // When existing the search loop, we have already checked 4 bytes. + var t int32 + + // We will not use repeat offsets across blocks. + // By not using them for the first 3 matches + canRepeat := len(blk.sequences) > 2 + + for { + if debugAsserts && canRepeat && offset1 == 0 { + panic("offset0 was 0") + } + + nextHash := hashLen(cv, hashLog, tableFastHashLen) + nextHash2 := hashLen(cv>>8, hashLog, tableFastHashLen) + candidate := e.table[nextHash] + candidate2 := e.table[nextHash2] + repIndex := s - offset1 + 2 + + e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)} + e.markShardDirty(nextHash) + e.table[nextHash2] = tableEntry{offset: s + e.cur + 1, val: uint32(cv >> 8)} + e.markShardDirty(nextHash2) + + if canRepeat && repIndex >= 0 && load3232(src, repIndex) == uint32(cv>>16) { + // Consider history as well. + var seq seq + length := 4 + e.matchlen(s+6, repIndex+4, src) + + seq.matchLen = uint32(length - zstdMinMatch) + + // We might be able to match backwards. + // Extend as long as we can. + start := s + 2 + // We end the search early, so we don't risk 0 literals + // and have to do special offset treatment. + startLimit := nextEmit + 1 + + sMin := s - e.maxMatchOff + if sMin < 0 { + sMin = 0 + } + for repIndex > sMin && start > startLimit && src[repIndex-1] == src[start-1] && seq.matchLen < maxMatchLength-zstdMinMatch { + repIndex-- + start-- + seq.matchLen++ + } + addLiterals(&seq, start) + + // rep 0 + seq.offset = 1 + if debugSequences { + println("repeat sequence", seq, "next s:", s) + } + blk.sequences = append(blk.sequences, seq) + s += length + 2 + nextEmit = s + if s >= sLimit { + if debugEncoder { + println("repeat ended", s, length) + + } + break encodeLoop + } + cv = load6432(src, s) + continue + } + coffset0 := s - (candidate.offset - e.cur) + coffset1 := s - (candidate2.offset - e.cur) + 1 + if coffset0 < e.maxMatchOff && uint32(cv) == candidate.val { + // found a regular match + t = candidate.offset - e.cur + if debugAsserts && s <= t { + panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) + } + if debugAsserts && s-t > e.maxMatchOff { + panic("s - t >e.maxMatchOff") + } + break + } + + if coffset1 < e.maxMatchOff && uint32(cv>>8) == candidate2.val { + // found a regular match + t = candidate2.offset - e.cur + s++ + if debugAsserts && s <= t { + panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) + } + if debugAsserts && s-t > e.maxMatchOff { + panic("s - t >e.maxMatchOff") + } + if debugAsserts && t < 0 { + panic("t<0") + } + break + } + s += stepSize + ((s - nextEmit) >> (kSearchStrength - 1)) + if s >= sLimit { + break encodeLoop + } + cv = load6432(src, s) + } + // A 4-byte match has been found. We'll later see if more than 4 bytes. + offset2 = offset1 + offset1 = s - t + + if debugAsserts && s <= t { + panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) + } + + if debugAsserts && canRepeat && int(offset1) > len(src) { + panic("invalid offset") + } + + // Extend the 4-byte match as long as possible. + l := e.matchlen(s+4, t+4, src) + 4 + + // Extend backwards + tMin := s - e.maxMatchOff + if tMin < 0 { + tMin = 0 + } + for t > tMin && s > nextEmit && src[t-1] == src[s-1] && l < maxMatchLength { + s-- + t-- + l++ + } + + // Write our sequence. + var seq seq + seq.litLen = uint32(s - nextEmit) + seq.matchLen = uint32(l - zstdMinMatch) + if seq.litLen > 0 { + blk.literals = append(blk.literals, src[nextEmit:s]...) + } + // Don't use repeat offsets + seq.offset = uint32(s-t) + 3 + s += l + if debugSequences { + println("sequence", seq, "next s:", s) + } + blk.sequences = append(blk.sequences, seq) + nextEmit = s + if s >= sLimit { + break encodeLoop + } + cv = load6432(src, s) + + // Check offset 2 + if o2 := s - offset2; canRepeat && load3232(src, o2) == uint32(cv) { + // We have at least 4 byte match. + // No need to check backwards. We come straight from a match + l := 4 + e.matchlen(s+4, o2+4, src) + + // Store this, since we have it. + nextHash := hashLen(cv, hashLog, tableFastHashLen) + e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)} + e.markShardDirty(nextHash) + seq.matchLen = uint32(l) - zstdMinMatch + seq.litLen = 0 + // Since litlen is always 0, this is offset 1. + seq.offset = 1 + s += l + nextEmit = s + if debugSequences { + println("sequence", seq, "next s:", s) + } + blk.sequences = append(blk.sequences, seq) + + // Swap offset 1 and 2. + offset1, offset2 = offset2, offset1 + if s >= sLimit { + break encodeLoop + } + // Prepare next loop. + cv = load6432(src, s) + } + } + + if int(nextEmit) < len(src) { + blk.literals = append(blk.literals, src[nextEmit:]...) + blk.extraLits = len(src) - int(nextEmit) + } + blk.recentOffsets[0] = uint32(offset1) + blk.recentOffsets[1] = uint32(offset2) + if debugEncoder { + println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits) + } +} + +// ResetDict will reset and set a dictionary if not nil +func (e *fastEncoder) Reset(d *dict, singleBlock bool) { + e.resetBase(d, singleBlock) + if d != nil { + panic("fastEncoder: Reset with dict") + } +} + +// ResetDict will reset and set a dictionary if not nil +func (e *fastEncoderDict) Reset(d *dict, singleBlock bool) { + e.resetBase(d, singleBlock) + if d == nil { + return + } + + // Init or copy dict table + if len(e.dictTable) != len(e.table) || d.id != e.lastDictID { + if len(e.dictTable) != len(e.table) { + e.dictTable = make([]tableEntry, len(e.table)) + } + if true { + end := e.maxMatchOff + int32(len(d.content)) - 8 + for i := e.maxMatchOff; i < end; i += 2 { + const hashLog = tableBits + + cv := load6432(d.content, i-e.maxMatchOff) + nextHash := hashLen(cv, hashLog, tableFastHashLen) // 0 -> 6 + nextHash1 := hashLen(cv>>8, hashLog, tableFastHashLen) // 1 -> 7 + e.dictTable[nextHash] = tableEntry{ + val: uint32(cv), + offset: i, + } + e.dictTable[nextHash1] = tableEntry{ + val: uint32(cv >> 8), + offset: i + 1, + } + } + } + e.lastDictID = d.id + e.allDirty = true + } + + e.cur = e.maxMatchOff + dirtyShardCnt := 0 + if !e.allDirty { + for i := range e.tableShardDirty { + if e.tableShardDirty[i] { + dirtyShardCnt++ + } + } + } + + const shardCnt = tableShardCnt + const shardSize = tableShardSize + if e.allDirty || dirtyShardCnt > shardCnt*4/6 { + //copy(e.table[:], e.dictTable) + e.table = *(*[tableSize]tableEntry)(e.dictTable) + for i := range e.tableShardDirty { + e.tableShardDirty[i] = false + } + e.allDirty = false + return + } + for i := range e.tableShardDirty { + if !e.tableShardDirty[i] { + continue + } + + //copy(e.table[i*shardSize:(i+1)*shardSize], e.dictTable[i*shardSize:(i+1)*shardSize]) + *(*[shardSize]tableEntry)(e.table[i*shardSize:]) = *(*[shardSize]tableEntry)(e.dictTable[i*shardSize:]) + e.tableShardDirty[i] = false + } + e.allDirty = false +} + +func (e *fastEncoderDict) markAllShardsDirty() { + e.allDirty = true +} + +func (e *fastEncoderDict) markShardDirty(entryNum uint32) { + e.tableShardDirty[entryNum/tableShardSize] = true +} diff --git a/internal/github.com/klauspost/compress/zstd/encoder.go b/internal/github.com/klauspost/compress/zstd/encoder.go new file mode 100644 index 000000000..26e0a8c10 --- /dev/null +++ b/internal/github.com/klauspost/compress/zstd/encoder.go @@ -0,0 +1,642 @@ +// Copyright 2019+ Klaus Post. All rights reserved. +// License information can be found in the LICENSE file. +// Based on work by Yann Collet, released under BSD License. + +package zstd + +import ( + "crypto/rand" + "errors" + "fmt" + "io" + "math" + rdebug "runtime/debug" + "sync" + + "github.com/cespare/xxhash/v2" +) + +// Encoder provides encoding to Zstandard. +// An Encoder can be used for either compressing a stream via the +// io.WriteCloser interface supported by the Encoder or as multiple independent +// tasks via the EncodeAll function. +// Smaller encodes are encouraged to use the EncodeAll function. +// Use NewWriter to create a new instance. +type Encoder struct { + o encoderOptions + encoders chan encoder + state encoderState + init sync.Once +} + +type encoder interface { + Encode(blk *blockEnc, src []byte) + EncodeNoHist(blk *blockEnc, src []byte) + Block() *blockEnc + CRC() *xxhash.Digest + AppendCRC([]byte) []byte + WindowSize(size int64) int32 + UseBlock(*blockEnc) + Reset(d *dict, singleBlock bool) +} + +type encoderState struct { + w io.Writer + filling []byte + current []byte + previous []byte + encoder encoder + writing *blockEnc + err error + writeErr error + nWritten int64 + nInput int64 + frameContentSize int64 + headerWritten bool + eofWritten bool + fullFrameWritten bool + + // This waitgroup indicates an encode is running. + wg sync.WaitGroup + // This waitgroup indicates we have a block encoding/writing. + wWg sync.WaitGroup +} + +// NewWriter will create a new Zstandard encoder. +// If the encoder will be used for encoding blocks a nil writer can be used. +func NewWriter(w io.Writer, opts ...EOption) (*Encoder, error) { + initPredefined() + var e Encoder + e.o.setDefault() + for _, o := range opts { + err := o(&e.o) + if err != nil { + return nil, err + } + } + if w != nil { + e.Reset(w) + } + return &e, nil +} + +func (e *Encoder) initialize() { + if e.o.concurrent == 0 { + e.o.setDefault() + } + e.encoders = make(chan encoder, e.o.concurrent) + for i := 0; i < e.o.concurrent; i++ { + enc := e.o.encoder() + e.encoders <- enc + } +} + +// Reset will re-initialize the writer and new writes will encode to the supplied writer +// as a new, independent stream. +func (e *Encoder) Reset(w io.Writer) { + s := &e.state + s.wg.Wait() + s.wWg.Wait() + if cap(s.filling) == 0 { + s.filling = make([]byte, 0, e.o.blockSize) + } + if e.o.concurrent > 1 { + if cap(s.current) == 0 { + s.current = make([]byte, 0, e.o.blockSize) + } + if cap(s.previous) == 0 { + s.previous = make([]byte, 0, e.o.blockSize) + } + s.current = s.current[:0] + s.previous = s.previous[:0] + if s.writing == nil { + s.writing = &blockEnc{lowMem: e.o.lowMem} + s.writing.init() + } + s.writing.initNewEncode() + } + if s.encoder == nil { + s.encoder = e.o.encoder() + } + s.filling = s.filling[:0] + s.encoder.Reset(e.o.dict, false) + s.headerWritten = false + s.eofWritten = false + s.fullFrameWritten = false + s.w = w + s.err = nil + s.nWritten = 0 + s.nInput = 0 + s.writeErr = nil + s.frameContentSize = 0 +} + +// ResetContentSize will reset and set a content size for the next stream. +// If the bytes written does not match the size given an error will be returned +// when calling Close(). +// This is removed when Reset is called. +// Sizes <= 0 results in no content size set. +func (e *Encoder) ResetContentSize(w io.Writer, size int64) { + e.Reset(w) + if size >= 0 { + e.state.frameContentSize = size + } +} + +// Write data to the encoder. +// Input data will be buffered and as the buffer fills up +// content will be compressed and written to the output. +// When done writing, use Close to flush the remaining output +// and write CRC if requested. +func (e *Encoder) Write(p []byte) (n int, err error) { + s := &e.state + if s.eofWritten { + return 0, ErrEncoderClosed + } + for len(p) > 0 { + if len(p)+len(s.filling) < e.o.blockSize { + if e.o.crc { + _, _ = s.encoder.CRC().Write(p) + } + s.filling = append(s.filling, p...) + return n + len(p), nil + } + add := p + if len(p)+len(s.filling) > e.o.blockSize { + add = add[:e.o.blockSize-len(s.filling)] + } + if e.o.crc { + _, _ = s.encoder.CRC().Write(add) + } + s.filling = append(s.filling, add...) + p = p[len(add):] + n += len(add) + if len(s.filling) < e.o.blockSize { + return n, nil + } + err := e.nextBlock(false) + if err != nil { + return n, err + } + if debugAsserts && len(s.filling) > 0 { + panic(len(s.filling)) + } + } + return n, nil +} + +// nextBlock will synchronize and start compressing input in e.state.filling. +// If an error has occurred during encoding it will be returned. +func (e *Encoder) nextBlock(final bool) error { + s := &e.state + // Wait for current block. + s.wg.Wait() + if s.err != nil { + return s.err + } + if len(s.filling) > e.o.blockSize { + return fmt.Errorf("block > maxStoreBlockSize") + } + if !s.headerWritten { + // If we have a single block encode, do a sync compression. + if final && len(s.filling) == 0 && !e.o.fullZero { + s.headerWritten = true + s.fullFrameWritten = true + s.eofWritten = true + return nil + } + if final && len(s.filling) > 0 { + s.current = e.encodeAll(s.encoder, s.filling, s.current[:0]) + var n2 int + n2, s.err = s.w.Write(s.current) + if s.err != nil { + return s.err + } + s.nWritten += int64(n2) + s.nInput += int64(len(s.filling)) + s.current = s.current[:0] + s.filling = s.filling[:0] + s.headerWritten = true + s.fullFrameWritten = true + s.eofWritten = true + return nil + } + + var tmp [maxHeaderSize]byte + fh := frameHeader{ + ContentSize: uint64(s.frameContentSize), + WindowSize: uint32(s.encoder.WindowSize(s.frameContentSize)), + SingleSegment: false, + Checksum: e.o.crc, + DictID: e.o.dict.ID(), + } + + dst := fh.appendTo(tmp[:0]) + s.headerWritten = true + s.wWg.Wait() + var n2 int + n2, s.err = s.w.Write(dst) + if s.err != nil { + return s.err + } + s.nWritten += int64(n2) + } + if s.eofWritten { + // Ensure we only write it once. + final = false + } + + if len(s.filling) == 0 { + // Final block, but no data. + if final { + enc := s.encoder + blk := enc.Block() + blk.reset(nil) + blk.last = true + blk.encodeRaw(nil) + s.wWg.Wait() + _, s.err = s.w.Write(blk.output) + s.nWritten += int64(len(blk.output)) + s.eofWritten = true + } + return s.err + } + + // SYNC: + if e.o.concurrent == 1 { + src := s.filling + s.nInput += int64(len(s.filling)) + if debugEncoder { + println("Adding sync block,", len(src), "bytes, final:", final) + } + enc := s.encoder + blk := enc.Block() + blk.reset(nil) + enc.Encode(blk, src) + blk.last = final + if final { + s.eofWritten = true + } + + s.err = blk.encode(src, e.o.noEntropy, !e.o.allLitEntropy) + if s.err != nil { + return s.err + } + _, s.err = s.w.Write(blk.output) + s.nWritten += int64(len(blk.output)) + s.filling = s.filling[:0] + return s.err + } + + // Move blocks forward. + s.filling, s.current, s.previous = s.previous[:0], s.filling, s.current + s.nInput += int64(len(s.current)) + s.wg.Add(1) + if final { + s.eofWritten = true + } + go func(src []byte) { + if debugEncoder { + println("Adding block,", len(src), "bytes, final:", final) + } + defer func() { + if r := recover(); r != nil { + s.err = fmt.Errorf("panic while encoding: %v", r) + rdebug.PrintStack() + } + s.wg.Done() + }() + enc := s.encoder + blk := enc.Block() + enc.Encode(blk, src) + blk.last = final + // Wait for pending writes. + s.wWg.Wait() + if s.writeErr != nil { + s.err = s.writeErr + return + } + // Transfer encoders from previous write block. + blk.swapEncoders(s.writing) + // Transfer recent offsets to next. + enc.UseBlock(s.writing) + s.writing = blk + s.wWg.Add(1) + go func() { + defer func() { + if r := recover(); r != nil { + s.writeErr = fmt.Errorf("panic while encoding/writing: %v", r) + rdebug.PrintStack() + } + s.wWg.Done() + }() + s.writeErr = blk.encode(src, e.o.noEntropy, !e.o.allLitEntropy) + if s.writeErr != nil { + return + } + _, s.writeErr = s.w.Write(blk.output) + s.nWritten += int64(len(blk.output)) + }() + }(s.current) + return nil +} + +// ReadFrom reads data from r until EOF or error. +// The return value n is the number of bytes read. +// Any error except io.EOF encountered during the read is also returned. +// +// The Copy function uses ReaderFrom if available. +func (e *Encoder) ReadFrom(r io.Reader) (n int64, err error) { + if debugEncoder { + println("Using ReadFrom") + } + + // Flush any current writes. + if len(e.state.filling) > 0 { + if err := e.nextBlock(false); err != nil { + return 0, err + } + } + e.state.filling = e.state.filling[:e.o.blockSize] + src := e.state.filling + for { + n2, err := r.Read(src) + if e.o.crc { + _, _ = e.state.encoder.CRC().Write(src[:n2]) + } + // src is now the unfilled part... + src = src[n2:] + n += int64(n2) + switch err { + case io.EOF: + e.state.filling = e.state.filling[:len(e.state.filling)-len(src)] + if debugEncoder { + println("ReadFrom: got EOF final block:", len(e.state.filling)) + } + return n, nil + case nil: + default: + if debugEncoder { + println("ReadFrom: got error:", err) + } + e.state.err = err + return n, err + } + if len(src) > 0 { + if debugEncoder { + println("ReadFrom: got space left in source:", len(src)) + } + continue + } + err = e.nextBlock(false) + if err != nil { + return n, err + } + e.state.filling = e.state.filling[:e.o.blockSize] + src = e.state.filling + } +} + +// Flush will send the currently written data to output +// and block until everything has been written. +// This should only be used on rare occasions where pushing the currently queued data is critical. +func (e *Encoder) Flush() error { + s := &e.state + if len(s.filling) > 0 { + err := e.nextBlock(false) + if err != nil { + // Ignore Flush after Close. + if errors.Is(s.err, ErrEncoderClosed) { + return nil + } + return err + } + } + s.wg.Wait() + s.wWg.Wait() + if s.err != nil { + // Ignore Flush after Close. + if errors.Is(s.err, ErrEncoderClosed) { + return nil + } + return s.err + } + return s.writeErr +} + +// Close will flush the final output and close the stream. +// The function will block until everything has been written. +// The Encoder can still be re-used after calling this. +func (e *Encoder) Close() error { + s := &e.state + if s.encoder == nil { + return nil + } + err := e.nextBlock(true) + if err != nil { + if errors.Is(s.err, ErrEncoderClosed) { + return nil + } + return err + } + if s.frameContentSize > 0 { + if s.nInput != s.frameContentSize { + return fmt.Errorf("frame content size %d given, but %d bytes was written", s.frameContentSize, s.nInput) + } + } + if e.state.fullFrameWritten { + return s.err + } + s.wg.Wait() + s.wWg.Wait() + + if s.err != nil { + return s.err + } + if s.writeErr != nil { + return s.writeErr + } + + // Write CRC + if e.o.crc && s.err == nil { + // heap alloc. + var tmp [4]byte + _, s.err = s.w.Write(s.encoder.AppendCRC(tmp[:0])) + s.nWritten += 4 + } + + // Add padding with content from crypto/rand.Reader + if s.err == nil && e.o.pad > 0 { + add := calcSkippableFrame(s.nWritten, int64(e.o.pad)) + frame, err := skippableFrame(s.filling[:0], add, rand.Reader) + if err != nil { + return err + } + _, s.err = s.w.Write(frame) + } + if s.err == nil { + s.err = ErrEncoderClosed + return nil + } + + return s.err +} + +// EncodeAll will encode all input in src and append it to dst. +// This function can be called concurrently, but each call will only run on a single goroutine. +// If empty input is given, nothing is returned, unless WithZeroFrames is specified. +// Encoded blocks can be concatenated and the result will be the combined input stream. +// Data compressed with EncodeAll can be decoded with the Decoder, +// using either a stream or DecodeAll. +func (e *Encoder) EncodeAll(src, dst []byte) []byte { + e.init.Do(e.initialize) + enc := <-e.encoders + defer func() { + e.encoders <- enc + }() + return e.encodeAll(enc, src, dst) +} + +func (e *Encoder) encodeAll(enc encoder, src, dst []byte) []byte { + if len(src) == 0 { + if e.o.fullZero { + // Add frame header. + fh := frameHeader{ + ContentSize: 0, + WindowSize: MinWindowSize, + SingleSegment: true, + // Adding a checksum would be a waste of space. + Checksum: false, + DictID: 0, + } + dst = fh.appendTo(dst) + + // Write raw block as last one only. + var blk blockHeader + blk.setSize(0) + blk.setType(blockTypeRaw) + blk.setLast(true) + dst = blk.appendTo(dst) + } + return dst + } + + // Use single segments when above minimum window and below window size. + single := len(src) <= e.o.windowSize && len(src) > MinWindowSize + if e.o.single != nil { + single = *e.o.single + } + fh := frameHeader{ + ContentSize: uint64(len(src)), + WindowSize: uint32(enc.WindowSize(int64(len(src)))), + SingleSegment: single, + Checksum: e.o.crc, + DictID: e.o.dict.ID(), + } + + // If less than 1MB, allocate a buffer up front. + if len(dst) == 0 && cap(dst) == 0 && len(src) < 1<<20 && !e.o.lowMem { + dst = make([]byte, 0, len(src)) + } + dst = fh.appendTo(dst) + + // If we can do everything in one block, prefer that. + if len(src) <= e.o.blockSize { + enc.Reset(e.o.dict, true) + // Slightly faster with no history and everything in one block. + if e.o.crc { + _, _ = enc.CRC().Write(src) + } + blk := enc.Block() + blk.last = true + if e.o.dict == nil { + enc.EncodeNoHist(blk, src) + } else { + enc.Encode(blk, src) + } + + // If we got the exact same number of literals as input, + // assume the literals cannot be compressed. + oldout := blk.output + // Output directly to dst + blk.output = dst + + err := blk.encode(src, e.o.noEntropy, !e.o.allLitEntropy) + if err != nil { + panic(err) + } + dst = blk.output + blk.output = oldout + } else { + enc.Reset(e.o.dict, false) + blk := enc.Block() + for len(src) > 0 { + todo := src + if len(todo) > e.o.blockSize { + todo = todo[:e.o.blockSize] + } + src = src[len(todo):] + if e.o.crc { + _, _ = enc.CRC().Write(todo) + } + blk.pushOffsets() + enc.Encode(blk, todo) + if len(src) == 0 { + blk.last = true + } + err := blk.encode(todo, e.o.noEntropy, !e.o.allLitEntropy) + if err != nil { + panic(err) + } + dst = append(dst, blk.output...) + blk.reset(nil) + } + } + if e.o.crc { + dst = enc.AppendCRC(dst) + } + // Add padding with content from crypto/rand.Reader + if e.o.pad > 0 { + add := calcSkippableFrame(int64(len(dst)), int64(e.o.pad)) + var err error + dst, err = skippableFrame(dst, add, rand.Reader) + if err != nil { + panic(err) + } + } + return dst +} + +// MaxEncodedSize returns the expected maximum +// size of an encoded block or stream. +func (e *Encoder) MaxEncodedSize(size int) int { + frameHeader := 4 + 2 // magic + frame header & window descriptor + if e.o.dict != nil { + frameHeader += 4 + } + // Frame content size: + if size < 256 { + frameHeader++ + } else if size < 65536+256 { + frameHeader += 2 + } else if size < math.MaxInt32 { + frameHeader += 4 + } else { + frameHeader += 8 + } + // Final crc + if e.o.crc { + frameHeader += 4 + } + + // Max overhead is 3 bytes/block. + // There cannot be 0 blocks. + blocks := (size + e.o.blockSize) / e.o.blockSize + + // Combine, add padding. + maxSz := frameHeader + 3*blocks + size + if e.o.pad > 1 { + maxSz += calcSkippableFrame(int64(maxSz), int64(e.o.pad)) + } + return maxSz +} diff --git a/internal/github.com/klauspost/compress/zstd/encoder_options.go b/internal/github.com/klauspost/compress/zstd/encoder_options.go new file mode 100644 index 000000000..80f736f6f --- /dev/null +++ b/internal/github.com/klauspost/compress/zstd/encoder_options.go @@ -0,0 +1,154 @@ +package zstd + +import ( + "fmt" + "math" + "runtime" + "strings" +) + +// EOption is an option for creating a encoder. +type EOption func(*encoderOptions) error + +// options retains accumulated state of multiple options. +type encoderOptions struct { + concurrent int + level EncoderLevel + single *bool + pad int + blockSize int + windowSize int + crc bool + fullZero bool + noEntropy bool + allLitEntropy bool + customWindow bool + customALEntropy bool + customBlockSize bool + lowMem bool + dict *dict +} + +func (o *encoderOptions) setDefault() { + *o = encoderOptions{ + concurrent: runtime.GOMAXPROCS(0), + crc: true, + single: nil, + blockSize: maxCompressedBlockSize, + windowSize: 8 << 20, + level: SpeedDefault, + allLitEntropy: false, + lowMem: false, + } +} + +// encoder returns an encoder with the selected options. +func (o encoderOptions) encoder() encoder { + // NOTE(bwplotka): No actual choice to vendor minimal amount of code. + return &fastEncoderDict{fastEncoder: fastEncoder{fastBase: fastBase{maxMatchOff: int32(o.windowSize), bufferReset: math.MaxInt32 - int32(o.windowSize*2), lowMem: o.lowMem}}} +} + +// EncoderLevel predefines encoder compression levels. +// Only use the constants made available, since the actual mapping +// of these values are very likely to change and your compression could change +// unpredictably when upgrading the library. +type EncoderLevel int + +const ( + speedNotSet EncoderLevel = iota + + // SpeedFastest will choose the fastest reasonable compression. + // This is roughly equivalent to the fastest Zstandard mode. + SpeedFastest + + // SpeedDefault is the default "pretty fast" compression option. + // This is roughly equivalent to the default Zstandard mode (level 3). + SpeedDefault + + // SpeedBetterCompression will yield better compression than the default. + // Currently it is about zstd level 7-8 with ~ 2x-3x the default CPU usage. + // By using this, notice that CPU usage may go up in the future. + SpeedBetterCompression + + // SpeedBestCompression will choose the best available compression option. + // This will offer the best compression no matter the CPU cost. + SpeedBestCompression + + // speedLast should be kept as the last actual compression option. + // The is not for external usage, but is used to keep track of the valid options. + speedLast +) + +// EncoderLevelFromString will convert a string representation of an encoding level back +// to a compression level. The compare is not case sensitive. +// If the string wasn't recognized, (false, SpeedDefault) will be returned. +func EncoderLevelFromString(s string) (bool, EncoderLevel) { + for l := speedNotSet + 1; l < speedLast; l++ { + if strings.EqualFold(s, l.String()) { + return true, l + } + } + return false, SpeedDefault +} + +// EncoderLevelFromZstd will return an encoder level that closest matches the compression +// ratio of a specific zstd compression level. +// Many input values will provide the same compression level. +func EncoderLevelFromZstd(level int) EncoderLevel { + switch { + case level < 3: + return SpeedFastest + case level >= 3 && level < 6: + return SpeedDefault + case level >= 6 && level < 10: + return SpeedBetterCompression + default: + return SpeedBestCompression + } +} + +// String provides a string representation of the compression level. +func (e EncoderLevel) String() string { + switch e { + case SpeedFastest: + return "fastest" + case SpeedDefault: + return "default" + case SpeedBetterCompression: + return "better" + case SpeedBestCompression: + return "best" + default: + return "invalid" + } +} + +// WithEncoderLevel specifies a predefined compression level. +func WithEncoderLevel(l EncoderLevel) EOption { + return func(o *encoderOptions) error { + // NOTE(bwlotka): To minimize the vendored size, only SpeedFastest is possible. + if l != SpeedFastest { + return fmt.Errorf("unknown encoder level %v", l) + } + o.level = l + if !o.customWindow { + switch o.level { + case SpeedFastest: + o.windowSize = 4 << 20 + if !o.customBlockSize { + o.blockSize = 1 << 16 + } + case SpeedDefault: + o.windowSize = 8 << 20 + case SpeedBetterCompression: + o.windowSize = 8 << 20 + case SpeedBestCompression: + o.windowSize = 8 << 20 + } + } + if !o.customALEntropy { + o.allLitEntropy = l > SpeedDefault + } + return nil + } +} diff --git a/internal/github.com/klauspost/compress/zstd/frameenc.go b/internal/github.com/klauspost/compress/zstd/frameenc.go new file mode 100644 index 000000000..2e7307058 --- /dev/null +++ b/internal/github.com/klauspost/compress/zstd/frameenc.go @@ -0,0 +1,147 @@ +// Copyright 2019+ Klaus Post. All rights reserved. +// License information can be found in the LICENSE file. +// Based on work by Yann Collet, released under BSD License. + +package zstd + +import ( + "encoding/binary" + "fmt" + "io" + "math" + "math/bits" +) + +type frameHeader struct { + ContentSize uint64 + WindowSize uint32 + SingleSegment bool + Checksum bool + DictID uint32 +} + +const ( + maxHeaderSize = 14 + frameMagic = "\x28\xb5\x2f\xfd" + + // MinWindowSize is the minimum Window Size, which is 1 KB. + MinWindowSize = 1 << 10 + + // MaxWindowSize is the maximum encoder window size + // and the default decoder maximum window size. + MaxWindowSize = 1 << 29 +) + +func (f frameHeader) appendTo(dst []byte) []byte { + dst = append(dst, frameMagic...) + var fhd uint8 + if f.Checksum { + fhd |= 1 << 2 + } + if f.SingleSegment { + fhd |= 1 << 5 + } + + var dictIDContent []byte + if f.DictID > 0 { + var tmp [4]byte + if f.DictID < 256 { + fhd |= 1 + tmp[0] = uint8(f.DictID) + dictIDContent = tmp[:1] + } else if f.DictID < 1<<16 { + fhd |= 2 + binary.LittleEndian.PutUint16(tmp[:2], uint16(f.DictID)) + dictIDContent = tmp[:2] + } else { + fhd |= 3 + binary.LittleEndian.PutUint32(tmp[:4], f.DictID) + dictIDContent = tmp[:4] + } + } + var fcs uint8 + if f.ContentSize >= 256 { + fcs++ + } + if f.ContentSize >= 65536+256 { + fcs++ + } + if f.ContentSize >= 0xffffffff { + fcs++ + } + + fhd |= fcs << 6 + + dst = append(dst, fhd) + if !f.SingleSegment { + const winLogMin = 10 + windowLog := (bits.Len32(f.WindowSize-1) - winLogMin) << 3 + dst = append(dst, uint8(windowLog)) + } + if f.DictID > 0 { + dst = append(dst, dictIDContent...) + } + switch fcs { + case 0: + if f.SingleSegment { + dst = append(dst, uint8(f.ContentSize)) + } + // Unless SingleSegment is set, framessizes < 256 are not stored. + case 1: + f.ContentSize -= 256 + dst = append(dst, uint8(f.ContentSize), uint8(f.ContentSize>>8)) + case 2: + dst = append(dst, uint8(f.ContentSize), uint8(f.ContentSize>>8), uint8(f.ContentSize>>16), uint8(f.ContentSize>>24)) + case 3: + dst = append(dst, uint8(f.ContentSize), uint8(f.ContentSize>>8), uint8(f.ContentSize>>16), uint8(f.ContentSize>>24), + uint8(f.ContentSize>>32), uint8(f.ContentSize>>40), uint8(f.ContentSize>>48), uint8(f.ContentSize>>56)) + default: + panic("invalid fcs") + } + return dst +} + +const skippableFrameHeader = 4 + 4 + +// calcSkippableFrame will return a total size to be added for written +// to be divisible by multiple. +// The value will always be > skippableFrameHeader. +// The function will panic if written < 0 or wantMultiple <= 0. +func calcSkippableFrame(written, wantMultiple int64) int { + if wantMultiple <= 0 { + panic("wantMultiple <= 0") + } + if written < 0 { + panic("written < 0") + } + leftOver := written % wantMultiple + if leftOver == 0 { + return 0 + } + toAdd := wantMultiple - leftOver + for toAdd < skippableFrameHeader { + toAdd += wantMultiple + } + return int(toAdd) +} + +// skippableFrame will add a skippable frame with a total size of bytes. +// total should be >= skippableFrameHeader and < math.MaxUint32. +func skippableFrame(dst []byte, total int, r io.Reader) ([]byte, error) { + if total == 0 { + return dst, nil + } + if total < skippableFrameHeader { + return dst, fmt.Errorf("requested skippable frame (%d) < 8", total) + } + if int64(total) > math.MaxUint32 { + return dst, fmt.Errorf("requested skippable frame (%d) > max uint32", total) + } + dst = append(dst, 0x50, 0x2a, 0x4d, 0x18) + f := uint32(total - skippableFrameHeader) + dst = append(dst, uint8(f), uint8(f>>8), uint8(f>>16), uint8(f>>24)) + start := len(dst) + dst = append(dst, make([]byte, f)...) + _, err := io.ReadFull(r, dst[start:]) + return dst, err +} diff --git a/internal/github.com/klauspost/compress/zstd/fse_decoder.go b/internal/github.com/klauspost/compress/zstd/fse_decoder.go new file mode 100644 index 000000000..d3f2587e1 --- /dev/null +++ b/internal/github.com/klauspost/compress/zstd/fse_decoder.go @@ -0,0 +1,295 @@ +// Copyright 2019+ Klaus Post. All rights reserved. +// License information can be found in the LICENSE file. +// Based on work by Yann Collet, released under BSD License. + +package zstd + +import ( + "encoding/binary" + "errors" + "fmt" + "io" +) + +const ( + tablelogAbsoluteMax = 9 +) + +const ( + /*!MEMORY_USAGE : + * Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.) + * Increasing memory usage improves compression ratio + * Reduced memory usage can improve speed, due to cache effect + * Recommended max value is 14, for 16KB, which nicely fits into Intel x86 L1 cache */ + maxMemoryUsage = tablelogAbsoluteMax + 2 + + maxTableLog = maxMemoryUsage - 2 + maxTablesize = 1 << maxTableLog + maxTableMask = (1 << maxTableLog) - 1 + minTablelog = 5 + maxSymbolValue = 255 +) + +// fseDecoder provides temporary storage for compression and decompression. +type fseDecoder struct { + 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 +} + +// tableStep returns the next table index. +func tableStep(tableSize uint32) uint32 { + return (tableSize >> 1) + (tableSize >> 3) + 3 +} + +// readNCount will read the symbol distribution so decoding tables can be constructed. +func (s *fseDecoder) readNCount(b *byteReader, maxSymbol uint16) error { + var ( + charnum uint16 + previous0 bool + ) + if b.remain() < 4 { + return errors.New("input too small") + } + bitStream := b.Uint32NC() + nbBits := uint((bitStream & 0xF) + minTablelog) // extract tableLog + if nbBits > tablelogAbsoluteMax { + println("Invalid tablelog:", nbBits) + return errors.New("tableLog too large") + } + bitStream >>= 4 + bitCount := uint(4) + + s.actualTableLog = uint8(nbBits) + remaining := int32((1 << nbBits) + 1) + threshold := int32(1 << nbBits) + gotTotal := int32(0) + nbBits++ + + for remaining > 1 && charnum <= maxSymbol { + if previous0 { + //println("prev0") + n0 := charnum + for (bitStream & 0xFFFF) == 0xFFFF { + //println("24 x 0") + n0 += 24 + if r := b.remain(); r > 5 { + b.advance(2) + // The check above should make sure we can read 32 bits + bitStream = b.Uint32NC() >> bitCount + } else { + // end of bit stream + bitStream >>= 16 + bitCount += 16 + } + } + //printf("bitstream: %d, 0b%b", bitStream&3, bitStream) + for (bitStream & 3) == 3 { + n0 += 3 + bitStream >>= 2 + bitCount += 2 + } + n0 += uint16(bitStream & 3) + bitCount += 2 + + if n0 > maxSymbolValue { + return errors.New("maxSymbolValue too small") + } + //println("inserting ", n0-charnum, "zeroes from idx", charnum, "ending before", n0) + for charnum < n0 { + s.norm[uint8(charnum)] = 0 + charnum++ + } + + if r := b.remain(); r >= 7 || r-int(bitCount>>3) >= 4 { + b.advance(bitCount >> 3) + bitCount &= 7 + // The check above should make sure we can read 32 bits + bitStream = b.Uint32NC() >> bitCount + } else { + bitStream >>= 2 + } + } + + max := (2*threshold - 1) - remaining + var count int32 + + if int32(bitStream)&(threshold-1) < max { + count = int32(bitStream) & (threshold - 1) + if debugAsserts && nbBits < 1 { + panic("nbBits underflow") + } + bitCount += nbBits - 1 + } else { + count = int32(bitStream) & (2*threshold - 1) + if count >= threshold { + count -= max + } + bitCount += nbBits + } + + // extra accuracy + count-- + if count < 0 { + // -1 means +1 + remaining += count + gotTotal -= count + } else { + remaining -= count + gotTotal += count + } + s.norm[charnum&0xff] = int16(count) + charnum++ + previous0 = count == 0 + for remaining < threshold { + nbBits-- + threshold >>= 1 + } + + if r := b.remain(); r >= 7 || r-int(bitCount>>3) >= 4 { + b.advance(bitCount >> 3) + bitCount &= 7 + // The check above should make sure we can read 32 bits + bitStream = b.Uint32NC() >> (bitCount & 31) + } else { + bitCount -= (uint)(8 * (len(b.b) - 4 - b.off)) + b.off = len(b.b) - 4 + bitStream = b.Uint32() >> (bitCount & 31) + } + } + s.symbolLen = charnum + if s.symbolLen <= 1 { + return fmt.Errorf("symbolLen (%d) too small", s.symbolLen) + } + if s.symbolLen > maxSymbolValue+1 { + return fmt.Errorf("symbolLen (%d) too big", s.symbolLen) + } + if remaining != 1 { + return fmt.Errorf("corruption detected (remaining %d != 1)", remaining) + } + if bitCount > 32 { + return fmt.Errorf("corruption detected (bitCount %d > 32)", bitCount) + } + if gotTotal != 1<> 3) + 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. +// Using a composite uint64 is faster than a struct with separate members. +type decSymbol uint64 + +func newDecSymbol(nbits, addBits uint8, newState uint16, baseline uint32) decSymbol { + return decSymbol(nbits) | (decSymbol(addBits) << 8) | (decSymbol(newState) << 16) | (decSymbol(baseline) << 32) +} + +func (d decSymbol) nbBits() uint8 { + return uint8(d) +} + +func (d decSymbol) addBits() uint8 { + return uint8(d >> 8) +} + +func (d decSymbol) newState() uint16 { + return uint16(d >> 16) +} + +func (d decSymbol) baselineInt() int { + return int(d >> 32) +} + +func (d *decSymbol) setNBits(nBits uint8) { + const mask = 0xffffffffffffff00 + *d = (*d & mask) | decSymbol(nBits) +} + +func (d *decSymbol) setAddBits(addBits uint8) { + const mask = 0xffffffffffff00ff + *d = (*d & mask) | (decSymbol(addBits) << 8) +} + +func (d *decSymbol) setNewState(state uint16) { + const mask = 0xffffffff0000ffff + *d = (*d & mask) | decSymbol(state)<<16 +} + +func (d *decSymbol) setExt(addBits uint8, baseline uint32) { + const mask = 0xffff00ff + *d = (*d & mask) | (decSymbol(addBits) << 8) | (decSymbol(baseline) << 32) +} + +// decSymbolValue returns the transformed decSymbol for the given symbol. +func decSymbolValue(symb uint8, t []baseOffset) (decSymbol, error) { + if int(symb) >= len(t) { + return 0, fmt.Errorf("rle symbol %d >= max %d", symb, len(t)) + } + lu := t[symb] + return newDecSymbol(0, lu.addBits, 0, lu.baseLine), nil +} + +// setRLE will set the decoder til RLE mode. +func (s *fseDecoder) setRLE(symbol decSymbol) { + s.actualTableLog = 0 + s.maxBits = symbol.addBits() + s.dt[0] = symbol +} + +// transform will transform the decoder table into a table usable for +// decoding without having to apply the transformation while decoding. +// The state will contain the base value and the number of bits to read. +func (s *fseDecoder) transform(t []baseOffset) error { + tableSize := uint16(1 << s.actualTableLog) + s.maxBits = 0 + for i, v := range s.dt[:tableSize] { + add := v.addBits() + if int(add) >= len(t) { + return fmt.Errorf("invalid decoding table entry %d, symbol %d >= max (%d)", i, v.addBits(), len(t)) + } + lu := t[add] + if lu.addBits > s.maxBits { + s.maxBits = lu.addBits + } + v.setExt(lu.addBits, lu.baseLine) + s.dt[i] = v + } + return nil +} + +// final returns the current state symbol without decoding the next. +func (s decSymbol) final() (int, uint8) { + return s.baselineInt(), s.addBits() +} diff --git a/internal/github.com/klauspost/compress/zstd/fse_decoder_generic.go b/internal/github.com/klauspost/compress/zstd/fse_decoder_generic.go new file mode 100644 index 000000000..b9818c497 --- /dev/null +++ b/internal/github.com/klauspost/compress/zstd/fse_decoder_generic.go @@ -0,0 +1,75 @@ +package zstd + +import ( + "errors" + "fmt" + "math/bits" +) + +// buildDtable will build the decoding table. +func (s *fseDecoder) buildDtable() error { + tableSize := uint32(1 << s.actualTableLog) + highThreshold := tableSize - 1 + symbolNext := s.stateTable[:256] + + // Init, lay down lowprob symbols + { + for i, v := range s.norm[:s.symbolLen] { + if v == -1 { + s.dt[highThreshold].setAddBits(uint8(i)) + highThreshold-- + v = 1 + } + symbolNext[i] = uint16(v) + } + } + + // Spread symbols + { + tableMask := tableSize - 1 + step := tableStep(tableSize) + position := uint32(0) + for ss, v := range s.norm[:s.symbolLen] { + for i := 0; i < int(v); i++ { + s.dt[position].setAddBits(uint8(ss)) + for { + // lowprob area + position = (position + step) & tableMask + if position <= highThreshold { + break + } + } + } + } + if position != 0 { + // position must reach all cells once, otherwise normalizedCounter is incorrect + return errors.New("corrupted input (position != 0)") + } + } + + // Build Decoding table + { + tableSize := uint16(1 << s.actualTableLog) + for u, v := range s.dt[:tableSize] { + symbol := v.addBits() + nextState := symbolNext[symbol] + symbolNext[symbol] = nextState + 1 + nBits := s.actualTableLog - byte(highBits(uint32(nextState))) + s.dt[u&maxTableMask].setNBits(nBits) + newState := (nextState << nBits) - tableSize + if newState > tableSize { + return fmt.Errorf("newState (%d) outside table size (%d)", newState, tableSize) + } + if newState == uint16(u) && nBits == 0 { + // Seems weird that this is possible with nbits > 0. + return fmt.Errorf("newState (%d) == oldState (%d) and no bits", newState, u) + } + s.dt[u&maxTableMask].setNewState(newState) + } + } + return nil +} + +func highBits(val uint32) (n uint32) { + return uint32(bits.Len32(val) - 1) +} diff --git a/internal/github.com/klauspost/compress/zstd/fse_encoder.go b/internal/github.com/klauspost/compress/zstd/fse_encoder.go new file mode 100644 index 000000000..ab26326a8 --- /dev/null +++ b/internal/github.com/klauspost/compress/zstd/fse_encoder.go @@ -0,0 +1,701 @@ +// Copyright 2019+ Klaus Post. All rights reserved. +// License information can be found in the LICENSE file. +// Based on work by Yann Collet, released under BSD License. + +package zstd + +import ( + "errors" + "fmt" + "math" +) + +const ( + // For encoding we only support up to + maxEncTableLog = 8 + maxEncTablesize = 1 << maxTableLog + maxEncTableMask = (1 << maxTableLog) - 1 + minEncTablelog = 5 + maxEncSymbolValue = maxMatchLengthSymbol +) + +// Scratch provides temporary storage for compression and decompression. +type fseEncoder struct { + symbolLen uint16 // Length of active part of the symbol table. + actualTableLog uint8 // Selected tablelog. + ct cTable // Compression tables. + maxCount int // count of the most probable symbol + zeroBits bool // no bits has prob > 50%. + clearCount bool // clear count + useRLE bool // This encoder is for RLE + preDefined bool // This encoder is predefined. + reUsed bool // Set to know when the encoder has been reused. + rleVal uint8 // RLE Symbol + maxBits uint8 // Maximum output bits after transform. + + // TODO: Technically zstd should be fine with 64 bytes. + count [256]uint32 + norm [256]int16 +} + +// cTable contains tables used for compression. +type cTable struct { + tableSymbol []byte + stateTable []uint16 + symbolTT []symbolTransform +} + +// symbolTransform contains the state transform for a symbol. +type symbolTransform struct { + deltaNbBits uint32 + deltaFindState int16 + outBits uint8 +} + +// String prints values as a human readable string. +func (s symbolTransform) String() string { + return fmt.Sprintf("{deltabits: %08x, findstate:%d outbits:%d}", s.deltaNbBits, s.deltaFindState, s.outBits) +} + +// Histogram allows to populate the histogram and skip that step in the compression, +// It otherwise allows to inspect the histogram when compression is done. +// To indicate that you have populated the histogram call HistogramFinished +// with the value of the highest populated symbol, as well as the number of entries +// in the most populated entry. These are accepted at face value. +func (s *fseEncoder) Histogram() *[256]uint32 { + return &s.count +} + +// HistogramFinished can be called to indicate that the histogram has been populated. +// maxSymbol is the index of the highest set symbol of the next data segment. +// maxCount is the number of entries in the most populated entry. +// These are accepted at face value. +func (s *fseEncoder) HistogramFinished(maxSymbol uint8, maxCount int) { + s.maxCount = maxCount + s.symbolLen = uint16(maxSymbol) + 1 + s.clearCount = maxCount != 0 +} + +// allocCtable will allocate tables needed for compression. +// If existing tables a re big enough, they are simply re-used. +func (s *fseEncoder) allocCtable() { + tableSize := 1 << s.actualTableLog + // get tableSymbol that is big enough. + if cap(s.ct.tableSymbol) < tableSize { + s.ct.tableSymbol = make([]byte, tableSize) + } + s.ct.tableSymbol = s.ct.tableSymbol[:tableSize] + + ctSize := tableSize + if cap(s.ct.stateTable) < ctSize { + s.ct.stateTable = make([]uint16, ctSize) + } + s.ct.stateTable = s.ct.stateTable[:ctSize] + + if cap(s.ct.symbolTT) < 256 { + s.ct.symbolTT = make([]symbolTransform, 256) + } + s.ct.symbolTT = s.ct.symbolTT[:256] +} + +// buildCTable will populate the compression table so it is ready to be used. +func (s *fseEncoder) buildCTable() error { + tableSize := uint32(1 << s.actualTableLog) + highThreshold := tableSize - 1 + var cumul [256]int16 + + s.allocCtable() + tableSymbol := s.ct.tableSymbol[:tableSize] + // symbol start positions + { + cumul[0] = 0 + for ui, v := range s.norm[:s.symbolLen-1] { + u := byte(ui) // one less than reference + if v == -1 { + // Low proba symbol + cumul[u+1] = cumul[u] + 1 + tableSymbol[highThreshold] = u + highThreshold-- + } else { + cumul[u+1] = cumul[u] + v + } + } + // Encode last symbol separately to avoid overflowing u + u := int(s.symbolLen - 1) + v := s.norm[s.symbolLen-1] + if v == -1 { + // Low proba symbol + cumul[u+1] = cumul[u] + 1 + tableSymbol[highThreshold] = byte(u) + highThreshold-- + } else { + cumul[u+1] = cumul[u] + v + } + if uint32(cumul[s.symbolLen]) != tableSize { + return fmt.Errorf("internal error: expected cumul[s.symbolLen] (%d) == tableSize (%d)", cumul[s.symbolLen], tableSize) + } + cumul[s.symbolLen] = int16(tableSize) + 1 + } + // Spread symbols + s.zeroBits = false + { + step := tableStep(tableSize) + tableMask := tableSize - 1 + var position uint32 + // if any symbol > largeLimit, we may have 0 bits output. + largeLimit := int16(1 << (s.actualTableLog - 1)) + for ui, v := range s.norm[:s.symbolLen] { + symbol := byte(ui) + if v > largeLimit { + s.zeroBits = true + } + for nbOccurrences := int16(0); nbOccurrences < v; nbOccurrences++ { + tableSymbol[position] = symbol + position = (position + step) & tableMask + for position > highThreshold { + position = (position + step) & tableMask + } /* Low proba area */ + } + } + + // Check if we have gone through all positions + if position != 0 { + return errors.New("position!=0") + } + } + + // Build table + table := s.ct.stateTable + { + tsi := int(tableSize) + for u, v := range tableSymbol { + // TableU16 : sorted by symbol order; gives next state value + table[cumul[v]] = uint16(tsi + u) + cumul[v]++ + } + } + + // Build Symbol Transformation Table + { + total := int16(0) + symbolTT := s.ct.symbolTT[:s.symbolLen] + tableLog := s.actualTableLog + tl := (uint32(tableLog) << 16) - (1 << tableLog) + for i, v := range s.norm[:s.symbolLen] { + switch v { + case 0: + case -1, 1: + symbolTT[i].deltaNbBits = tl + symbolTT[i].deltaFindState = total - 1 + total++ + default: + maxBitsOut := uint32(tableLog) - highBit(uint32(v-1)) + minStatePlus := uint32(v) << maxBitsOut + symbolTT[i].deltaNbBits = (maxBitsOut << 16) - minStatePlus + symbolTT[i].deltaFindState = total - v + total += v + } + } + if total != int16(tableSize) { + return fmt.Errorf("total mismatch %d (got) != %d (want)", total, tableSize) + } + } + return nil +} + +var rtbTable = [...]uint32{0, 473195, 504333, 520860, 550000, 700000, 750000, 830000} + +func (s *fseEncoder) setRLE(val byte) { + s.allocCtable() + s.actualTableLog = 0 + s.ct.stateTable = s.ct.stateTable[:1] + s.ct.symbolTT[val] = symbolTransform{ + deltaFindState: 0, + deltaNbBits: 0, + } + if debugEncoder { + println("setRLE: val", val, "symbolTT", s.ct.symbolTT[val]) + } + s.rleVal = val + s.useRLE = true +} + +// setBits will set output bits for the transform. +// if nil is provided, the number of bits is equal to the index. +func (s *fseEncoder) setBits(transform []byte) { + if s.reUsed || s.preDefined { + return + } + if s.useRLE { + if transform == nil { + s.ct.symbolTT[s.rleVal].outBits = s.rleVal + s.maxBits = s.rleVal + return + } + s.maxBits = transform[s.rleVal] + s.ct.symbolTT[s.rleVal].outBits = s.maxBits + return + } + if transform == nil { + for i := range s.ct.symbolTT[:s.symbolLen] { + s.ct.symbolTT[i].outBits = uint8(i) + } + s.maxBits = uint8(s.symbolLen - 1) + return + } + s.maxBits = 0 + for i, v := range transform[:s.symbolLen] { + s.ct.symbolTT[i].outBits = v + if v > s.maxBits { + // We could assume bits always going up, but we play safe. + s.maxBits = v + } + } +} + +// normalizeCount will normalize the count of the symbols so +// the total is equal to the table size. +// If successful, compression tables will also be made ready. +func (s *fseEncoder) normalizeCount(length int) error { + if s.reUsed { + return nil + } + s.optimalTableLog(length) + var ( + tableLog = s.actualTableLog + scale = 62 - uint64(tableLog) + step = (1 << 62) / uint64(length) + vStep = uint64(1) << (scale - 20) + stillToDistribute = int16(1 << tableLog) + largest int + largestP int16 + lowThreshold = (uint32)(length >> tableLog) + ) + if s.maxCount == length { + s.useRLE = true + return nil + } + s.useRLE = false + for i, cnt := range s.count[:s.symbolLen] { + // already handled + // if (count[s] == s.length) return 0; /* rle special case */ + + if cnt == 0 { + s.norm[i] = 0 + continue + } + if cnt <= lowThreshold { + s.norm[i] = -1 + stillToDistribute-- + } else { + proba := (int16)((uint64(cnt) * step) >> scale) + if proba < 8 { + restToBeat := vStep * uint64(rtbTable[proba]) + v := uint64(cnt)*step - (uint64(proba) << scale) + if v > restToBeat { + proba++ + } + } + if proba > largestP { + largestP = proba + largest = i + } + s.norm[i] = proba + stillToDistribute -= proba + } + } + + if -stillToDistribute >= (s.norm[largest] >> 1) { + // corner case, need another normalization method + err := s.normalizeCount2(length) + if err != nil { + return err + } + if debugAsserts { + err = s.validateNorm() + if err != nil { + return err + } + } + return s.buildCTable() + } + s.norm[largest] += stillToDistribute + if debugAsserts { + err := s.validateNorm() + if err != nil { + return err + } + } + return s.buildCTable() +} + +// Secondary normalization method. +// To be used when primary method fails. +func (s *fseEncoder) normalizeCount2(length int) error { + const notYetAssigned = -2 + var ( + distributed uint32 + total = uint32(length) + tableLog = s.actualTableLog + lowThreshold = total >> tableLog + lowOne = (total * 3) >> (tableLog + 1) + ) + for i, cnt := range s.count[:s.symbolLen] { + if cnt == 0 { + s.norm[i] = 0 + continue + } + if cnt <= lowThreshold { + s.norm[i] = -1 + distributed++ + total -= cnt + continue + } + if cnt <= lowOne { + s.norm[i] = 1 + distributed++ + total -= cnt + continue + } + s.norm[i] = notYetAssigned + } + toDistribute := (1 << tableLog) - distributed + + if (total / toDistribute) > lowOne { + // risk of rounding to zero + lowOne = (total * 3) / (toDistribute * 2) + for i, cnt := range s.count[:s.symbolLen] { + if (s.norm[i] == notYetAssigned) && (cnt <= lowOne) { + s.norm[i] = 1 + distributed++ + total -= cnt + continue + } + } + toDistribute = (1 << tableLog) - distributed + } + if distributed == uint32(s.symbolLen)+1 { + // all values are pretty poor; + // probably incompressible data (should have already been detected); + // find max, then give all remaining points to max + var maxV int + var maxC uint32 + for i, cnt := range s.count[:s.symbolLen] { + if cnt > maxC { + maxV = i + maxC = cnt + } + } + s.norm[maxV] += int16(toDistribute) + return nil + } + + if total == 0 { + // all of the symbols were low enough for the lowOne or lowThreshold + for i := uint32(0); toDistribute > 0; i = (i + 1) % (uint32(s.symbolLen)) { + if s.norm[i] > 0 { + toDistribute-- + s.norm[i]++ + } + } + return nil + } + + var ( + vStepLog = 62 - uint64(tableLog) + mid = uint64((1 << (vStepLog - 1)) - 1) + rStep = (((1 << vStepLog) * uint64(toDistribute)) + mid) / uint64(total) // scale on remaining + tmpTotal = mid + ) + for i, cnt := range s.count[:s.symbolLen] { + if s.norm[i] == notYetAssigned { + var ( + end = tmpTotal + uint64(cnt)*rStep + sStart = uint32(tmpTotal >> vStepLog) + sEnd = uint32(end >> vStepLog) + weight = sEnd - sStart + ) + if weight < 1 { + return errors.New("weight < 1") + } + s.norm[i] = int16(weight) + tmpTotal = end + } + } + return nil +} + +// optimalTableLog calculates and sets the optimal tableLog in s.actualTableLog +func (s *fseEncoder) optimalTableLog(length int) { + tableLog := uint8(maxEncTableLog) + minBitsSrc := highBit(uint32(length)) + 1 + minBitsSymbols := highBit(uint32(s.symbolLen-1)) + 2 + minBits := uint8(minBitsSymbols) + if minBitsSrc < minBitsSymbols { + minBits = uint8(minBitsSrc) + } + + maxBitsSrc := uint8(highBit(uint32(length-1))) - 2 + if maxBitsSrc < tableLog { + // Accuracy can be reduced + tableLog = maxBitsSrc + } + if minBits > tableLog { + tableLog = minBits + } + // Need a minimum to safely represent all symbol values + if tableLog < minEncTablelog { + tableLog = minEncTablelog + } + if tableLog > maxEncTableLog { + tableLog = maxEncTableLog + } + s.actualTableLog = tableLog +} + +// validateNorm validates the normalized histogram table. +func (s *fseEncoder) validateNorm() (err error) { + var total int + for _, v := range s.norm[:s.symbolLen] { + if v >= 0 { + total += int(v) + } else { + total -= int(v) + } + } + defer func() { + if err == nil { + return + } + fmt.Printf("selected TableLog: %d, Symbol length: %d\n", s.actualTableLog, s.symbolLen) + for i, v := range s.norm[:s.symbolLen] { + fmt.Printf("%3d: %5d -> %4d \n", i, s.count[i], v) + } + }() + if total != (1 << s.actualTableLog) { + return fmt.Errorf("warning: Total == %d != %d", total, 1<> 3) + 3 + 2 + + // Write Table Size + bitStream = uint32(tableLog - minEncTablelog) + bitCount = uint(4) + remaining = int16(tableSize + 1) /* +1 for extra accuracy */ + threshold = int16(tableSize) + nbBits = uint(tableLog + 1) + outP = len(out) + ) + if cap(out) < outP+maxHeaderSize { + out = append(out, make([]byte, maxHeaderSize*3)...) + out = out[:len(out)-maxHeaderSize*3] + } + out = out[:outP+maxHeaderSize] + + // stops at 1 + for remaining > 1 { + if previous0 { + start := charnum + for s.norm[charnum] == 0 { + charnum++ + } + for charnum >= start+24 { + start += 24 + bitStream += uint32(0xFFFF) << bitCount + out[outP] = byte(bitStream) + out[outP+1] = byte(bitStream >> 8) + outP += 2 + bitStream >>= 16 + } + for charnum >= start+3 { + start += 3 + bitStream += 3 << bitCount + bitCount += 2 + } + bitStream += uint32(charnum-start) << bitCount + bitCount += 2 + if bitCount > 16 { + out[outP] = byte(bitStream) + out[outP+1] = byte(bitStream >> 8) + outP += 2 + bitStream >>= 16 + bitCount -= 16 + } + } + + count := s.norm[charnum] + charnum++ + max := (2*threshold - 1) - remaining + if count < 0 { + remaining += count + } else { + remaining -= count + } + count++ // +1 for extra accuracy + if count >= threshold { + count += max // [0..max[ [max..threshold[ (...) [threshold+max 2*threshold[ + } + bitStream += uint32(count) << bitCount + bitCount += nbBits + if count < max { + bitCount-- + } + + previous0 = count == 1 + if remaining < 1 { + return nil, errors.New("internal error: remaining < 1") + } + for remaining < threshold { + nbBits-- + threshold >>= 1 + } + + if bitCount > 16 { + out[outP] = byte(bitStream) + out[outP+1] = byte(bitStream >> 8) + outP += 2 + bitStream >>= 16 + bitCount -= 16 + } + } + + if outP+2 > len(out) { + return nil, fmt.Errorf("internal error: %d > %d, maxheader: %d, sl: %d, tl: %d, normcount: %v", outP+2, len(out), maxHeaderSize, s.symbolLen, int(tableLog), s.norm[:s.symbolLen]) + } + out[outP] = byte(bitStream) + out[outP+1] = byte(bitStream >> 8) + outP += int((bitCount + 7) / 8) + + if charnum > s.symbolLen { + return nil, errors.New("internal error: charnum > s.symbolLen") + } + return out[:outP], nil +} + +// Approximate symbol cost, as fractional value, using fixed-point format (accuracyLog fractional bits) +// note 1 : assume symbolValue is valid (<= maxSymbolValue) +// note 2 : if freq[symbolValue]==0, @return a fake cost of tableLog+1 bits * +func (s *fseEncoder) bitCost(symbolValue uint8, accuracyLog uint32) uint32 { + minNbBits := s.ct.symbolTT[symbolValue].deltaNbBits >> 16 + threshold := (minNbBits + 1) << 16 + if debugAsserts { + if !(s.actualTableLog < 16) { + panic("!s.actualTableLog < 16") + } + // ensure enough room for renormalization double shift + if !(uint8(accuracyLog) < 31-s.actualTableLog) { + panic("!uint8(accuracyLog) < 31-s.actualTableLog") + } + } + tableSize := uint32(1) << s.actualTableLog + deltaFromThreshold := threshold - (s.ct.symbolTT[symbolValue].deltaNbBits + tableSize) + // linear interpolation (very approximate) + normalizedDeltaFromThreshold := (deltaFromThreshold << accuracyLog) >> s.actualTableLog + bitMultiplier := uint32(1) << accuracyLog + if debugAsserts { + if s.ct.symbolTT[symbolValue].deltaNbBits+tableSize > threshold { + panic("s.ct.symbolTT[symbolValue].deltaNbBits+tableSize > threshold") + } + if normalizedDeltaFromThreshold > bitMultiplier { + panic("normalizedDeltaFromThreshold > bitMultiplier") + } + } + return (minNbBits+1)*bitMultiplier - normalizedDeltaFromThreshold +} + +// Returns the cost in bits of encoding the distribution in count using ctable. +// Histogram should only be up to the last non-zero symbol. +// Returns an -1 if ctable cannot represent all the symbols in count. +func (s *fseEncoder) approxSize(hist []uint32) uint32 { + if int(s.symbolLen) < len(hist) { + // More symbols than we have. + return math.MaxUint32 + } + if s.useRLE { + // We will never reuse RLE encoders. + return math.MaxUint32 + } + const kAccuracyLog = 8 + badCost := (uint32(s.actualTableLog) + 1) << kAccuracyLog + var cost uint32 + for i, v := range hist { + if v == 0 { + continue + } + if s.norm[i] == 0 { + return math.MaxUint32 + } + bitCost := s.bitCost(uint8(i), kAccuracyLog) + if bitCost > badCost { + return math.MaxUint32 + } + cost += v * bitCost + } + return cost >> kAccuracyLog +} + +// maxHeaderSize returns the maximum header size in bits. +// This is not exact size, but we want a penalty for new tables anyway. +func (s *fseEncoder) maxHeaderSize() uint32 { + if s.preDefined { + return 0 + } + if s.useRLE { + return 8 + } + return (((uint32(s.symbolLen) * uint32(s.actualTableLog)) >> 3) + 3) * 8 +} + +// cState contains the compression state of a stream. +type cState struct { + bw *bitWriter + stateTable []uint16 + state uint16 +} + +// init will initialize the compression state to the first symbol of the stream. +func (c *cState) init(bw *bitWriter, ct *cTable, first symbolTransform) { + c.bw = bw + c.stateTable = ct.stateTable + if len(c.stateTable) == 1 { + // RLE + c.stateTable[0] = uint16(0) + c.state = 0 + return + } + nbBitsOut := (first.deltaNbBits + (1 << 15)) >> 16 + im := int32((nbBitsOut << 16) - first.deltaNbBits) + lu := (im >> nbBitsOut) + int32(first.deltaFindState) + c.state = c.stateTable[lu] +} + +// flush will write the tablelog to the output and flush the remaining full bytes. +func (c *cState) flush(tableLog uint8) { + c.bw.flush32() + c.bw.addBits16NC(c.state, tableLog) +} diff --git a/internal/github.com/klauspost/compress/zstd/fse_predefined.go b/internal/github.com/klauspost/compress/zstd/fse_predefined.go new file mode 100644 index 000000000..474cb77d2 --- /dev/null +++ b/internal/github.com/klauspost/compress/zstd/fse_predefined.go @@ -0,0 +1,158 @@ +// Copyright 2019+ Klaus Post. All rights reserved. +// License information can be found in the LICENSE file. +// Based on work by Yann Collet, released under BSD License. + +package zstd + +import ( + "fmt" + "math" + "sync" +) + +var ( + // fsePredef are the predefined fse tables as defined here: + // https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#default-distributions + // These values are already transformed. + fsePredef [3]fseDecoder + + // fsePredefEnc are the predefined encoder based on fse tables as defined here: + // https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#default-distributions + // These values are already transformed. + fsePredefEnc [3]fseEncoder + + // symbolTableX contain the transformations needed for each type as defined in + // https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#the-codes-for-literals-lengths-match-lengths-and-offsets + symbolTableX [3][]baseOffset + + // maxTableSymbol is the biggest supported symbol for each table type + // https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#the-codes-for-literals-lengths-match-lengths-and-offsets + maxTableSymbol = [3]uint8{tableLiteralLengths: maxLiteralLengthSymbol, tableOffsets: maxOffsetLengthSymbol, tableMatchLengths: maxMatchLengthSymbol} + + // bitTables is the bits table for each table. + bitTables = [3][]byte{tableLiteralLengths: llBitsTable[:], tableOffsets: nil, tableMatchLengths: mlBitsTable[:]} +) + +type tableIndex uint8 + +const ( + // indexes for fsePredef and symbolTableX + tableLiteralLengths tableIndex = 0 + tableOffsets tableIndex = 1 + tableMatchLengths tableIndex = 2 + + maxLiteralLengthSymbol = 35 + maxOffsetLengthSymbol = 30 + maxMatchLengthSymbol = 52 +) + +// baseOffset is used for calculating transformations. +type baseOffset struct { + baseLine uint32 + addBits uint8 +} + +// fillBase will precalculate base offsets with the given bit distributions. +func fillBase(dst []baseOffset, base uint32, bits ...uint8) { + if len(bits) != len(dst) { + panic(fmt.Sprintf("len(dst) (%d) != len(bits) (%d)", len(dst), len(bits))) + } + for i, bit := range bits { + if base > math.MaxInt32 { + panic("invalid decoding table, base overflows int32") + } + + dst[i] = baseOffset{ + baseLine: base, + addBits: bit, + } + base += 1 << bit + } +} + +var predef sync.Once + +func initPredefined() { + predef.Do(func() { + // Literals length codes + tmp := make([]baseOffset, 36) + for i := range tmp[:16] { + tmp[i] = baseOffset{ + baseLine: uint32(i), + addBits: 0, + } + } + fillBase(tmp[16:], 16, 1, 1, 1, 1, 2, 2, 3, 3, 4, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16) + symbolTableX[tableLiteralLengths] = tmp + + // Match length codes + tmp = make([]baseOffset, 53) + for i := range tmp[:32] { + tmp[i] = baseOffset{ + // The transformation adds the 3 length. + baseLine: uint32(i) + 3, + addBits: 0, + } + } + fillBase(tmp[32:], 35, 1, 1, 1, 1, 2, 2, 3, 3, 4, 4, 5, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16) + symbolTableX[tableMatchLengths] = tmp + + // Offset codes + tmp = make([]baseOffset, maxOffsetBits+1) + tmp[1] = baseOffset{ + baseLine: 1, + addBits: 1, + } + fillBase(tmp[2:], 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30) + symbolTableX[tableOffsets] = tmp + + // Fill predefined tables and transform them. + // https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#default-distributions + for i := range fsePredef[:] { + f := &fsePredef[i] + switch tableIndex(i) { + case tableLiteralLengths: + // https://github.com/facebook/zstd/blob/ededcfca57366461021c922720878c81a5854a0a/lib/decompress/zstd_decompress_block.c#L243 + f.actualTableLog = 6 + copy(f.norm[:], []int16{4, 3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 1, 1, 1, + 2, 2, 2, 2, 2, 2, 2, 2, 2, 3, 2, 1, 1, 1, 1, 1, + -1, -1, -1, -1}) + f.symbolLen = 36 + case tableOffsets: + // https://github.com/facebook/zstd/blob/ededcfca57366461021c922720878c81a5854a0a/lib/decompress/zstd_decompress_block.c#L281 + f.actualTableLog = 5 + copy(f.norm[:], []int16{ + 1, 1, 1, 1, 1, 1, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1, + 1, 1, 1, 1, 1, 1, 1, 1, -1, -1, -1, -1, -1}) + f.symbolLen = 29 + case tableMatchLengths: + //https://github.com/facebook/zstd/blob/ededcfca57366461021c922720878c81a5854a0a/lib/decompress/zstd_decompress_block.c#L304 + f.actualTableLog = 6 + copy(f.norm[:], []int16{ + 1, 4, 3, 2, 2, 2, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1, + 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, + 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, -1, -1, + -1, -1, -1, -1, -1}) + f.symbolLen = 53 + } + if err := f.buildDtable(); err != nil { + panic(fmt.Errorf("building table %v: %v", tableIndex(i), err)) + } + if err := f.transform(symbolTableX[i]); err != nil { + panic(fmt.Errorf("building table %v: %v", tableIndex(i), err)) + } + f.preDefined = true + + // Create encoder as well + enc := &fsePredefEnc[i] + copy(enc.norm[:], f.norm[:]) + enc.symbolLen = f.symbolLen + enc.actualTableLog = f.actualTableLog + if err := enc.buildCTable(); err != nil { + panic(fmt.Errorf("building encoding table %v: %v", tableIndex(i), err)) + } + enc.setBits(bitTables[i]) + enc.preDefined = true + } + }) +} diff --git a/internal/github.com/klauspost/compress/zstd/hash.go b/internal/github.com/klauspost/compress/zstd/hash.go new file mode 100644 index 000000000..5d73c21eb --- /dev/null +++ b/internal/github.com/klauspost/compress/zstd/hash.go @@ -0,0 +1,35 @@ +// Copyright 2019+ Klaus Post. All rights reserved. +// License information can be found in the LICENSE file. +// Based on work by Yann Collet, released under BSD License. + +package zstd + +const ( + prime3bytes = 506832829 + prime4bytes = 2654435761 + prime5bytes = 889523592379 + prime6bytes = 227718039650203 + prime7bytes = 58295818150454627 + prime8bytes = 0xcf1bbcdcb7a56463 +) + +// 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) + } +} diff --git a/internal/github.com/klauspost/compress/zstd/matchlen_generic.go b/internal/github.com/klauspost/compress/zstd/matchlen_generic.go new file mode 100644 index 000000000..206c4aef3 --- /dev/null +++ b/internal/github.com/klauspost/compress/zstd/matchlen_generic.go @@ -0,0 +1,35 @@ +// Copyright 2019+ Klaus Post. All rights reserved. +// License information can be found in the LICENSE file. + +package zstd + +import ( + "math/bits" + + "github.com/prometheus/client_golang/internal/github.com/klauspost/compress/internal/le" +) + +// matchLen returns the maximum common prefix length of a and b. +// a must be the shortest of the two. +func matchLen(a, b []byte) (n int) { + left := len(a) + for left >= 8 { + diff := le.Load64(a, n) ^ le.Load64(b, n) + if diff != 0 { + return n + bits.TrailingZeros64(diff)>>3 + } + n += 8 + left -= 8 + } + a = a[n:] + b = b[n:] + + for i := range a { + if a[i] != b[i] { + break + } + n++ + } + return n + +} diff --git a/internal/github.com/klauspost/compress/zstd/seqenc.go b/internal/github.com/klauspost/compress/zstd/seqenc.go new file mode 100644 index 000000000..d9855d8f7 --- /dev/null +++ b/internal/github.com/klauspost/compress/zstd/seqenc.go @@ -0,0 +1,144 @@ +// Copyright 2019+ Klaus Post. All rights reserved. +// License information can be found in the LICENSE file. +// Based on work by Yann Collet, released under BSD License. + +package zstd + +import ( + "fmt" + "math/bits" +) + +type seq struct { + litLen uint32 + matchLen uint32 + offset uint32 + + // Codes are stored here for the encoder + // so they only have to be looked up once. + llCode, mlCode, ofCode uint8 +} + +type seqCompMode uint8 + +const ( + compModePredefined seqCompMode = iota + compModeRLE + compModeFSE + compModeRepeat +) + +func (s seq) String() string { + if s.offset <= 3 { + if s.offset == 0 { + return fmt.Sprint("litLen:", s.litLen, ", matchLen:", s.matchLen+zstdMinMatch, ", offset: INVALID (0)") + } + return fmt.Sprint("litLen:", s.litLen, ", matchLen:", s.matchLen+zstdMinMatch, ", offset:", s.offset, " (repeat)") + } + return fmt.Sprint("litLen:", s.litLen, ", matchLen:", s.matchLen+zstdMinMatch, ", offset:", s.offset-3, " (new)") +} + +type seqCoders struct { + llEnc, ofEnc, mlEnc *fseEncoder + llPrev, ofPrev, mlPrev *fseEncoder +} + +// swap coders with another (block). +func (s *seqCoders) swap(other *seqCoders) { + *s, *other = *other, *s +} + +// setPrev will update the previous encoders to the actually used ones +// and make sure a fresh one is in the main slot. +func (s *seqCoders) setPrev(ll, ml, of *fseEncoder) { + compareSwap := func(used *fseEncoder, current, prev **fseEncoder) { + // We used the new one, more current to history and reuse the previous history + if *current == used { + *prev, *current = *current, *prev + c := *current + p := *prev + c.reUsed = false + p.reUsed = true + return + } + if used == *prev { + return + } + // Ensure we cannot reuse by accident + prevEnc := *prev + prevEnc.symbolLen = 0 + } + compareSwap(ll, &s.llEnc, &s.llPrev) + compareSwap(ml, &s.mlEnc, &s.mlPrev) + compareSwap(of, &s.ofEnc, &s.ofPrev) +} + +func highBit(val uint32) (n uint32) { + return uint32(bits.Len32(val) - 1) +} + +var llCodeTable = [64]byte{0, 1, 2, 3, 4, 5, 6, 7, + 8, 9, 10, 11, 12, 13, 14, 15, + 16, 16, 17, 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} + +// Up to 6 bits +const maxLLCode = 35 + +// llBitsTable translates from ll code to number of bits. +var llBitsTable = [maxLLCode + 1]byte{ + 0, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 0, 0, 0, 0, 0, 0, + 1, 1, 1, 1, 2, 2, 3, 3, + 4, 6, 7, 8, 9, 10, 11, 12, + 13, 14, 15, 16} + +// llCode returns the code that represents the literal length requested. +func llCode(litLength uint32) uint8 { + const llDeltaCode = 19 + if litLength <= 63 { + return llCodeTable[litLength&63] + } + return uint8(highBit(litLength)) + llDeltaCode +} + +var mlCodeTable = [128]byte{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, + 32, 32, 33, 33, 34, 34, 35, 35, 36, 36, 36, 36, 37, 37, 37, 37, + 38, 38, 38, 38, 38, 38, 38, 38, 39, 39, 39, 39, 39, 39, 39, 39, + 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, + 41, 41, 41, 41, 41, 41, 41, 41, 41, 41, 41, 41, 41, 41, 41, 41, + 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, + 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42} + +// Up to 6 bits +const maxMLCode = 52 + +// mlBitsTable translates from ml code to number of bits. +var mlBitsTable = [maxMLCode + 1]byte{ + 0, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 0, 0, 0, 0, 0, 0, + 1, 1, 1, 1, 2, 2, 3, 3, + 4, 4, 5, 7, 8, 9, 10, 11, + 12, 13, 14, 15, 16} + +// note : mlBase = matchLength - MINMATCH; +// because it's the format it's stored in seqStore->sequences +func mlCode(mlBase uint32) uint8 { + const mlDeltaCode = 36 + if mlBase <= 127 { + return mlCodeTable[mlBase&127] + } + return uint8(highBit(mlBase)) + mlDeltaCode +} + +func ofCode(offset uint32) uint8 { + // A valid offset will always be > 0. + return uint8(bits.Len32(offset) - 1) +} diff --git a/internal/github.com/klauspost/compress/zstd/zstd.go b/internal/github.com/klauspost/compress/zstd/zstd.go new file mode 100644 index 000000000..7fc27901f --- /dev/null +++ b/internal/github.com/klauspost/compress/zstd/zstd.go @@ -0,0 +1,126 @@ +// Package zstd provides decompression of zstandard files. +// +// For advanced usage and examples, go to the README: https://github.com/klauspost/compress/tree/master/zstd#zstd +package zstd + +import ( + "bytes" + "errors" + "log" + "math" + + "github.com/prometheus/client_golang/internal/github.com/klauspost/compress/internal/le" +) + +// enable debug printing +const debug = false + +// enable encoding debug printing +const debugEncoder = debug + +// enable decoding debug printing +const debugDecoder = debug + +// Enable extra assertions. +const debugAsserts = debug || false + +// print sequence details +const debugSequences = false + +// print detailed matching information +const debugMatches = false + +// force encoder to use predefined tables. +const forcePreDef = false + +// zstdMinMatch is the minimum zstd match length. +const zstdMinMatch = 3 + +// fcsUnknown is used for unknown frame content size. +const fcsUnknown = math.MaxUint64 + +var ( + // ErrReservedBlockType is returned when a reserved block type is found. + // Typically this indicates wrong or corrupted input. + ErrReservedBlockType = errors.New("invalid input: reserved block type encountered") + + // ErrCompressedSizeTooBig is returned when a block is bigger than allowed. + // Typically this indicates wrong or corrupted input. + ErrCompressedSizeTooBig = errors.New("invalid input: compressed size too big") + + // ErrBlockTooSmall is returned when a block is too small to be decoded. + // Typically returned on invalid input. + ErrBlockTooSmall = errors.New("block too small") + + // ErrUnexpectedBlockSize is returned when a block has unexpected size. + // Typically returned on invalid input. + ErrUnexpectedBlockSize = errors.New("unexpected block size") + + // ErrMagicMismatch is returned when a "magic" number isn't what is expected. + // Typically this indicates wrong or corrupted input. + ErrMagicMismatch = errors.New("invalid input: magic number mismatch") + + // ErrWindowSizeExceeded is returned when a reference exceeds the valid window size. + // Typically this indicates wrong or corrupted input. + ErrWindowSizeExceeded = errors.New("window size exceeded") + + // ErrWindowSizeTooSmall is returned when no window size is specified. + // Typically this indicates wrong or corrupted input. + ErrWindowSizeTooSmall = errors.New("invalid input: window size was too small") + + // ErrDecoderSizeExceeded is returned if decompressed size exceeds the configured limit. + ErrDecoderSizeExceeded = errors.New("decompressed size exceeds configured limit") + + // ErrUnknownDictionary is returned if the dictionary ID is unknown. + ErrUnknownDictionary = errors.New("unknown dictionary") + + // ErrFrameSizeExceeded is returned if the stated frame size is exceeded. + // This is only returned if SingleSegment is specified on the frame. + ErrFrameSizeExceeded = errors.New("frame size exceeded") + + // ErrFrameSizeMismatch is returned if the stated frame size does not match the expected size. + // This is only returned if SingleSegment is specified on the frame. + ErrFrameSizeMismatch = errors.New("frame size does not match size on stream") + + // ErrCRCMismatch is returned if CRC mismatches. + ErrCRCMismatch = errors.New("CRC check failed") + + // ErrDecoderClosed will be returned if the Decoder was used after + // Close has been called. + ErrDecoderClosed = errors.New("decoder used after Close") + + // ErrEncoderClosed will be returned if the Encoder was used after + // Close has been called. + ErrEncoderClosed = errors.New("encoder used after Close") + + // ErrDecoderNilInput is returned when a nil Reader was provided + // and an operation other than Reset/DecodeAll/Close was attempted. + ErrDecoderNilInput = errors.New("nil input provided as reader") +) + +func println(a ...interface{}) { + if debug || debugDecoder || debugEncoder { + log.Println(a...) + } +} + +func printf(format string, a ...interface{}) { + if debug || debugDecoder || debugEncoder { + log.Printf(format, a...) + } +} + +func load3232(b []byte, i int32) uint32 { + return le.Load32(b, i) +} + +func load6432(b []byte, i int32) uint64 { + return le.Load64(b, i) +} + +type byter interface { + Bytes() []byte + Len() int +} + +var _ byter = &bytes.Buffer{} diff --git a/prometheus/promhttp/http.go b/prometheus/promhttp/http.go index e598e66e6..f3cac13cc 100644 --- a/prometheus/promhttp/http.go +++ b/prometheus/promhttp/http.go @@ -41,10 +41,10 @@ import ( "sync" "time" - "github.com/klauspost/compress/zstd" "github.com/prometheus/common/expfmt" "github.com/prometheus/client_golang/internal/github.com/golang/gddo/httputil" + "github.com/prometheus/client_golang/internal/github.com/klauspost/compress/zstd" "github.com/prometheus/client_golang/prometheus" ) @@ -445,7 +445,7 @@ func negotiateEncodingWriter(r *http.Request, rw io.Writer, compressions []strin switch selected { case "zstd": - // TODO(mrueg): Replace klauspost/compress with stdlib implementation once https://github.com/golang/go/issues/62513 is implemented. + // TODO(mrueg): Replace klauspost/compress vendored lib with stdlib implementation once https://github.com/golang/go/issues/62513 is implemented. z, err := zstd.NewWriter(rw, zstd.WithEncoderLevel(zstd.SpeedFastest)) if err != nil { return nil, "", func() {}, err diff --git a/prometheus/promhttp/http_test.go b/prometheus/promhttp/http_test.go index 3ad2d1da8..3fa817c6e 100644 --- a/prometheus/promhttp/http_test.go +++ b/prometheus/promhttp/http_test.go @@ -26,7 +26,7 @@ import ( "testing" "time" - "github.com/klauspost/compress/zstd" + orignalzstd "github.com/klauspost/compress/zstd" dto "github.com/prometheus/client_model/go" "github.com/prometheus/client_golang/prometheus" @@ -90,7 +90,7 @@ func readCompressedBody(r io.Reader, comp Compression) (string, error) { got, err := io.ReadAll(reader) return string(got), err case Zstd: - reader, err := zstd.NewReader(r) + reader, err := orignalzstd.NewReader(r) if err != nil { return "", err }