[chore]: Bump github.com/prometheus/client_golang from 1.19.1 to 1.20.0 (#3210)

48 files changed, 18686 insertions, 0 deletions

diff --git a/vendor/github.com/klauspost/compress/zstd/README.md b/vendor/github.com/klauspost/compress/zstd/README.md
new file mode 100644
index 000000000..92e2347bb
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/README.md
@@ -0,0 +1,441 @@
+# zstd 
+
+[Zstandard](https://facebook.github.io/zstd/) is a real-time compression algorithm, providing high compression ratios. 
+It offers a very wide range of compression / speed trade-off, while being backed by a very fast decoder.
+A high performance compression algorithm is implemented. For now focused on speed. 
+
+This package provides [compression](#Compressor) to and [decompression](#Decompressor) of Zstandard content. 
+
+This package is pure Go and without use of "unsafe". 
+
+The `zstd` package is provided as open source software using a Go standard license.
+
+Currently the package is heavily optimized for 64 bit processors and will be significantly slower on 32 bit processors.
+
+For seekable zstd streams, see [this excellent package](https://github.com/SaveTheRbtz/zstd-seekable-format-go).
+
+## Installation
+
+Install using `go get -u github.com/klauspost/compress`. The package is located in `github.com/klauspost/compress/zstd`.
+
+[![Go Reference](https://pkg.go.dev/badge/github.com/klauspost/compress/zstd.svg)](https://pkg.go.dev/github.com/klauspost/compress/zstd)
+
+## Compressor
+
+### Status: 
+
+STABLE - there may always be subtle bugs, a wide variety of content has been tested and the library is actively 
+used by several projects. This library is being [fuzz-tested](https://github.com/klauspost/compress-fuzz) for all updates.
+
+There may still be specific combinations of data types/size/settings that could lead to edge cases, 
+so as always, testing is recommended.  
+
+For now, a high speed (fastest) and medium-fast (default) compressor has been implemented. 
+
+* The "Fastest" compression ratio is roughly equivalent to zstd level 1. 
+* The "Default" compression ratio is roughly equivalent to zstd level 3 (default).
+* The "Better" compression ratio is roughly equivalent to zstd level 7.
+* The "Best" compression ratio is roughly equivalent to zstd level 11.
+
+In terms of speed, it is typically 2x as fast as the stdlib deflate/gzip in its fastest mode. 
+The compression ratio compared to stdlib is around level 3, but usually 3x as fast.
+
+ 
+### Usage
+
+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 that can be used for both.
+
+To create a writer with default options, do like this:
+
+```Go
+// Compress input to output.
+func Compress(in io.Reader, out io.Writer) error {
+    enc, err := zstd.NewWriter(out)
+    if err != nil {
+        return err
+    }
+    _, err = io.Copy(enc, in)
+    if err != nil {
+        enc.Close()
+        return err
+    }
+    return enc.Close()
+}
+```
+
+Now you can encode by writing data to `enc`. The output will be finished writing when `Close()` is called.
+Even if your encode fails, you should still call `Close()` to release any resources that may be held up.  
+
+The above is fine for big encodes. However, whenever possible try to *reuse* the writer.
+
+To reuse the encoder, you can use the `Reset(io.Writer)` function to change to another output. 
+This will allow the encoder to reuse all resources and avoid wasteful allocations. 
+
+Currently stream encoding has 'light' concurrency, meaning up to 2 goroutines can be working on part 
+of a stream. This is independent of the `WithEncoderConcurrency(n)`, but that is likely to change 
+in the future. So if you want to limit concurrency for future updates, specify the concurrency
+you would like.
+
+If you would like stream encoding to be done without spawning async goroutines, use `WithEncoderConcurrency(1)`
+which will compress input as each block is completed, blocking on writes until each has completed.
+
+You can specify your desired compression level using `WithEncoderLevel()` option. Currently only pre-defined 
+compression settings can be specified.
+
+#### Future Compatibility Guarantees
+
+This will be an evolving project. When using this package it is important to note that both the compression efficiency and speed may change.
+
+The goal will be to keep the default efficiency at the default zstd (level 3). 
+However the encoding should never be assumed to remain the same, 
+and you should not use hashes of compressed output for similarity checks.
+
+The Encoder can be assumed to produce the same output from the exact same code version.
+However, the may be modes in the future that break this, 
+although they will not be enabled without an explicit option.   
+
+This encoder is not designed to (and will probably never) output the exact same bitstream as the reference encoder.
+
+Also note, that the cgo decompressor currently does not [report all errors on invalid input](https://github.com/DataDog/zstd/issues/59),
+[omits error checks](https://github.com/DataDog/zstd/issues/61), [ignores checksums](https://github.com/DataDog/zstd/issues/43) 
+and seems to ignore concatenated streams, even though [it is part of the spec](https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#frames).
+
+#### Blocks
+
+For compressing small blocks, the returned encoder has a function called `EncodeAll(src, dst []byte) []byte`.
+
+`EncodeAll` will encode all input in src and append it to dst.
+This function can be called concurrently. 
+Each call will only run on a same goroutine as the caller.
+
+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`.
+
+Especially when encoding blocks you should take special care to reuse the encoder. 
+This will effectively make it run without allocations after a warmup period. 
+To make it run completely without allocations, supply a destination buffer with space for all content.   
+
+```Go
+import "github.com/klauspost/compress/zstd"
+
+// Create a writer that caches compressors.
+// For this operation type we supply a nil Reader.
+var encoder, _ = zstd.NewWriter(nil)
+
+// Compress a buffer. 
+// If you have a destination buffer, the allocation in the call can also be eliminated.
+func Compress(src []byte) []byte {
+    return encoder.EncodeAll(src, make([]byte, 0, len(src)))
+} 
+```
+
+You can control the maximum number of concurrent encodes using the `WithEncoderConcurrency(n)` 
+option when creating the writer.
+
+Using the Encoder for both a stream and individual blocks concurrently is safe. 
+
+### Performance
+
+I have collected some speed examples to compare speed and compression against other compressors.
+
+* `file` is the input file.
+* `out` is the compressor used. `zskp` is this package. `zstd` is the Datadog cgo library. `gzstd/gzkp` is gzip standard and this library.
+* `level` is the compression level used. For `zskp` level 1 is "fastest", level 2 is "default"; 3 is "better", 4 is "best".
+* `insize`/`outsize` is the input/output size.
+* `millis` is the number of milliseconds used for compression.
+* `mb/s` is megabytes (2^20 bytes) per second.
+
+```
+Silesia Corpus:
+http://sun.aei.polsl.pl/~sdeor/corpus/silesia.zip
+
+This package:
+file    out     level   insize      outsize     millis  mb/s
+silesia.tar zskp    1   211947520   73821326    634     318.47
+silesia.tar zskp    2   211947520   67655404    1508    133.96
+silesia.tar zskp    3   211947520   64746933    3000    67.37
+silesia.tar zskp    4   211947520   60073508    16926   11.94
+
+cgo zstd:
+silesia.tar zstd    1   211947520   73605392    543     371.56
+silesia.tar zstd    3   211947520   66793289    864     233.68
+silesia.tar zstd    6   211947520   62916450    1913    105.66
+silesia.tar zstd    9   211947520   60212393    5063    39.92
+
+gzip, stdlib/this package:
+silesia.tar gzstd   1   211947520   80007735    1498    134.87
+silesia.tar gzkp    1   211947520   80088272    1009    200.31
+
+GOB stream of binary data. Highly compressible.
+https://files.klauspost.com/compress/gob-stream.7z
+
+file        out     level   insize  outsize     millis  mb/s
+gob-stream  zskp    1   1911399616  233948096   3230    564.34
+gob-stream  zskp    2   1911399616  203997694   4997    364.73
+gob-stream  zskp    3   1911399616  173526523   13435   135.68
+gob-stream  zskp    4   1911399616  162195235   47559   38.33
+
+gob-stream  zstd    1   1911399616  249810424   2637    691.26
+gob-stream  zstd    3   1911399616  208192146   3490    522.31
+gob-stream  zstd    6   1911399616  193632038   6687    272.56
+gob-stream  zstd    9   1911399616  177620386   16175   112.70
+
+gob-stream  gzstd   1   1911399616  357382013   9046    201.49
+gob-stream  gzkp    1   1911399616  359136669   4885    373.08
+
+The test data for the Large Text Compression Benchmark is the first
+10^9 bytes of the English Wikipedia dump on Mar. 3, 2006.
+http://mattmahoney.net/dc/textdata.html
+
+file    out level   insize      outsize     millis  mb/s
+enwik9  zskp    1   1000000000  343833605   3687    258.64
+enwik9  zskp    2   1000000000  317001237   7672    124.29
+enwik9  zskp    3   1000000000  291915823   15923   59.89
+enwik9  zskp    4   1000000000  261710291   77697   12.27
+
+enwik9  zstd    1   1000000000  358072021   3110    306.65
+enwik9  zstd    3   1000000000  313734672   4784    199.35
+enwik9  zstd    6   1000000000  295138875   10290   92.68
+enwik9  zstd    9   1000000000  278348700   28549   33.40
+
+enwik9  gzstd   1   1000000000  382578136   8608    110.78
+enwik9  gzkp    1   1000000000  382781160   5628    169.45
+
+Highly compressible JSON file.
+https://files.klauspost.com/compress/github-june-2days-2019.json.zst
+
+file                        out level   insize      outsize     millis  mb/s
+github-june-2days-2019.json zskp    1   6273951764  697439532   9789    611.17
+github-june-2days-2019.json zskp    2   6273951764  610876538   18553   322.49
+github-june-2days-2019.json zskp    3   6273951764  517662858   44186   135.41
+github-june-2days-2019.json zskp    4   6273951764  464617114   165373  36.18
+
+github-june-2days-2019.json zstd    1   6273951764  766284037   8450    708.00
+github-june-2days-2019.json zstd    3   6273951764  661889476   10927   547.57
+github-june-2days-2019.json zstd    6   6273951764  642756859   22996   260.18
+github-june-2days-2019.json zstd    9   6273951764  601974523   52413   114.16
+
+github-june-2days-2019.json gzstd   1   6273951764  1164397768  26793   223.32
+github-june-2days-2019.json gzkp    1   6273951764  1120631856  17693   338.16
+
+VM Image, Linux mint with a few installed applications:
+https://files.klauspost.com/compress/rawstudio-mint14.7z
+
+file                    out level   insize      outsize     millis  mb/s
+rawstudio-mint14.tar    zskp    1   8558382592  3718400221  18206   448.29
+rawstudio-mint14.tar    zskp    2   8558382592  3326118337  37074   220.15
+rawstudio-mint14.tar    zskp    3   8558382592  3163842361  87306   93.49
+rawstudio-mint14.tar    zskp    4   8558382592  2970480650  783862  10.41
+
+rawstudio-mint14.tar    zstd    1   8558382592  3609250104  17136   476.27
+rawstudio-mint14.tar    zstd    3   8558382592  3341679997  29262   278.92
+rawstudio-mint14.tar    zstd    6   8558382592  3235846406  77904   104.77
+rawstudio-mint14.tar    zstd    9   8558382592  3160778861  140946  57.91
+
+rawstudio-mint14.tar    gzstd   1   8558382592  3926234992  51345   158.96
+rawstudio-mint14.tar    gzkp    1   8558382592  3960117298  36722   222.26
+
+CSV data:
+https://files.klauspost.com/compress/nyc-taxi-data-10M.csv.zst
+
+file                    out level   insize      outsize     millis  mb/s
+nyc-taxi-data-10M.csv   zskp    1   3325605752  641319332   9462    335.17
+nyc-taxi-data-10M.csv   zskp    2   3325605752  588976126   17570   180.50
+nyc-taxi-data-10M.csv   zskp    3   3325605752  529329260   32432   97.79
+nyc-taxi-data-10M.csv   zskp    4   3325605752  474949772   138025  22.98
+
+nyc-taxi-data-10M.csv   zstd    1   3325605752  687399637   8233    385.18
+nyc-taxi-data-10M.csv   zstd    3   3325605752  598514411   10065   315.07
+nyc-taxi-data-10M.csv   zstd    6   3325605752  570522953   20038   158.27
+nyc-taxi-data-10M.csv   zstd    9   3325605752  517554797   64565   49.12
+
+nyc-taxi-data-10M.csv   gzstd   1   3325605752  928654908   21270   149.11
+nyc-taxi-data-10M.csv   gzkp    1   3325605752  922273214   13929   227.68
+```
+
+## Decompressor
+
+Status: STABLE - there may still be subtle bugs, but a wide variety of content has been tested.
+
+This library is being continuously [fuzz-tested](https://github.com/klauspost/compress-fuzz),
+kindly supplied by [fuzzit.dev](https://fuzzit.dev/). 
+The main purpose of the fuzz testing is to ensure that it is not possible to crash the decoder, 
+or run it past its limits with ANY input provided.  
+ 
+### Usage
+
+The package has been designed for two main usages, big streams of data and smaller in-memory buffers. 
+There are two main usages of the package for these. Both of them are accessed by creating a `Decoder`.
+
+For streaming use a simple setup could look like this:
+
+```Go
+import "github.com/klauspost/compress/zstd"
+
+func Decompress(in io.Reader, out io.Writer) error {
+    d, err := zstd.NewReader(in)
+    if err != nil {
+        return err
+    }
+    defer d.Close()
+    
+    // Copy content...
+    _, err = io.Copy(out, d)
+    return err
+}
+```
+
+It is important to use the "Close" function when you no longer need the Reader to stop running goroutines, 
+when running with default settings.
+Goroutines will exit once an error has been returned, including `io.EOF` at the end of a stream.
+
+Streams are decoded concurrently in 4 asynchronous stages to give the best possible throughput.
+However, if you prefer synchronous decompression, use `WithDecoderConcurrency(1)` which will decompress data 
+as it is being requested only.
+
+For decoding buffers, it could look something like this:
+
+```Go
+import "github.com/klauspost/compress/zstd"
+
+// Create a reader that caches decompressors.
+// For this operation type we supply a nil Reader.
+var decoder, _ = zstd.NewReader(nil, zstd.WithDecoderConcurrency(0))
+
+// Decompress a buffer. We don't supply a destination buffer,
+// so it will be allocated by the decoder.
+func Decompress(src []byte) ([]byte, error) {
+    return decoder.DecodeAll(src, nil)
+} 
+```
+
+Both of these cases should provide the functionality needed. 
+The decoder can be used for *concurrent* decompression of multiple buffers.
+By default 4 decompressors will be created. 
+
+It will only allow a certain number of concurrent operations to run. 
+To tweak that yourself use the `WithDecoderConcurrency(n)` option when creating the decoder.
+It is possible to use `WithDecoderConcurrency(0)` to create GOMAXPROCS decoders.
+
+### Dictionaries
+
+Data compressed with [dictionaries](https://github.com/facebook/zstd#the-case-for-small-data-compression) can be decompressed.
+
+Dictionaries are added individually to Decoders.
+Dictionaries are generated by the `zstd --train` command and contains an initial state for the decoder.
+To add a dictionary use the `WithDecoderDicts(dicts ...[]byte)` option with the dictionary data.
+Several dictionaries can be added at once.
+
+The dictionary will be used automatically for the data that specifies them.
+A re-used Decoder will still contain the dictionaries registered.
+
+When registering multiple dictionaries with the same ID, the last one will be used.
+
+It is possible to use dictionaries when compressing data.
+
+To enable a dictionary use `WithEncoderDict(dict []byte)`. Here only one dictionary will be used 
+and it will likely be used even if it doesn't improve compression. 
+
+The used dictionary must be used to decompress the content.
+
+For any real gains, the dictionary should be built with similar data. 
+If an unsuitable dictionary is used the output may be slightly larger than using no dictionary.
+Use the [zstd commandline tool](https://github.com/facebook/zstd/releases) to build a dictionary from sample data.
+For information see [zstd dictionary information](https://github.com/facebook/zstd#the-case-for-small-data-compression). 
+
+For now there is a fixed startup performance penalty for compressing content with dictionaries. 
+This will likely be improved over time. Just be aware to test performance when implementing.  
+
+### Allocation-less operation
+
+The decoder has been designed to operate without allocations after a warmup. 
+
+This means that you should *store* the decoder for best performance. 
+To re-use a stream decoder, use the `Reset(r io.Reader) error` to switch to another stream.
+A decoder can safely be re-used even if the previous stream failed.
+
+To release the resources, you must call the `Close()` function on a decoder.
+After this it can *no longer be reused*, but all running goroutines will be stopped.
+So you *must* use this if you will no longer need the Reader.
+
+For decompressing smaller buffers a single decoder can be used.
+When decoding buffers, you can supply a destination slice with length 0 and your expected capacity.
+In this case no unneeded allocations should be made. 
+
+### Concurrency
+
+The buffer decoder does everything on the same goroutine and does nothing concurrently.
+It can however decode several buffers concurrently. Use `WithDecoderConcurrency(n)` to limit that.
+
+The stream decoder will create goroutines that:
+
+1) Reads input and splits the input into blocks.
+2) Decompression of literals.
+3) Decompression of sequences.
+4) Reconstruction of output stream.
+
+So effectively this also means the decoder will "read ahead" and prepare data to always be available for output.
+
+The concurrency level will, for streams, determine how many blocks ahead the compression will start.
+
+Since "blocks" are quite dependent on the output of the previous block stream decoding will only have limited concurrency.
+
+In practice this means that concurrency is often limited to utilizing about 3 cores effectively.
+  
+### Benchmarks
+
+The first two are streaming decodes and the last are smaller inputs. 
+
+Running on AMD Ryzen 9 3950X 16-Core Processor. AMD64 assembly used.
+
+```
+BenchmarkDecoderSilesia-32    	                   5	 206878840 ns/op	1024.50 MB/s	   49808 B/op	      43 allocs/op
+BenchmarkDecoderEnwik9-32                          1	1271809000 ns/op	 786.28 MB/s	   72048 B/op	      52 allocs/op
+
+Concurrent blocks, performance:
+
+BenchmarkDecoder_DecodeAllParallel/kppkn.gtb.zst-32         	   67356	     17857 ns/op	10321.96 MB/s	        22.48 pct	     102 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/geo.protodata.zst-32     	  266656	      4421 ns/op	26823.21 MB/s	        11.89 pct	      19 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/plrabn12.txt.zst-32      	   20992	     56842 ns/op	8477.17 MB/s	        39.90 pct	     754 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/lcet10.txt.zst-32        	   27456	     43932 ns/op	9714.01 MB/s	        33.27 pct	     524 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/asyoulik.txt.zst-32      	   78432	     15047 ns/op	8319.15 MB/s	        40.34 pct	      66 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/alice29.txt.zst-32       	   65800	     18436 ns/op	8249.63 MB/s	        37.75 pct	      88 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/html_x_4.zst-32          	  102993	     11523 ns/op	35546.09 MB/s	         3.637 pct	     143 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/paper-100k.pdf.zst-32    	 1000000	      1070 ns/op	95720.98 MB/s	        80.53 pct	       3 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/fireworks.jpeg.zst-32    	  749802	      1752 ns/op	70272.35 MB/s	       100.0 pct	       5 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/urls.10K.zst-32          	   22640	     52934 ns/op	13263.37 MB/s	        26.25 pct	    1014 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/html.zst-32              	  226412	      5232 ns/op	19572.27 MB/s	        14.49 pct	      20 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/comp-data.bin.zst-32     	  923041	      1276 ns/op	3194.71 MB/s	        31.26 pct	       0 B/op	       0 allocs/op
+```
+
+This reflects the performance around May 2022, but this may be out of date.
+
+## Zstd inside ZIP files
+
+It is possible to use zstandard to compress individual files inside zip archives.
+While this isn't widely supported it can be useful for internal files.
+
+To support the compression and decompression of these files you must register a compressor and decompressor.
+
+It is highly recommended registering the (de)compressors on individual zip Reader/Writer and NOT
+use the global registration functions. The main reason for this is that 2 registrations from 
+different packages will result in a panic.
+
+It is a good idea to only have a single compressor and decompressor, since they can be used for multiple zip
+files concurrently, and using a single instance will allow reusing some resources.
+
+See [this example](https://pkg.go.dev/github.com/klauspost/compress/zstd#example-ZipCompressor) for 
+how to compress and decompress files inside zip archives.
+
+# Contributions
+
+Contributions are always welcome. 
+For new features/fixes, remember to add tests and for performance enhancements include benchmarks.
+
+For general feedback and experience reports, feel free to open an issue or write me on [Twitter](https://twitter.com/sh0dan).
+
+This package includes the excellent [`github.com/cespare/xxhash`](https://github.com/cespare/xxhash) package Copyright (c) 2016 Caleb Spare.
diff --git a/vendor/github.com/klauspost/compress/zstd/bitreader.go b/vendor/github.com/klauspost/compress/zstd/bitreader.go
new file mode 100644
index 000000000..25ca98394
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/bitreader.go
@@ -0,0 +1,136 @@
+// 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"
+	"math/bits"
+)
+
+// 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
+	value    uint64 // Maybe use [16]byte, but shifting is awkward.
+	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
+	// 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) int {
+	if n == 0 /*|| b.bitsRead >= 64 */ {
+		return 0
+	}
+	return int(b.get32BitsFast(n))
+}
+
+// get32BitsFast requires that at least one bit is requested every time.
+// There are no checks if the buffer is filled.
+func (b *bitReader) get32BitsFast(n uint8) uint32 {
+	const regMask = 64 - 1
+	v := uint32((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
+	}
+	v := b.in[len(b.in)-4:]
+	b.in = b.in[:len(b.in)-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
+}
+
+// fillFastStart() assumes the bitreader is empty and there is at least 8 bytes to read.
+func (b *bitReader) fillFastStart() {
+	v := b.in[len(b.in)-8:]
+	b.in = b.in[:len(b.in)-8]
+	b.value = binary.LittleEndian.Uint64(v)
+	b.bitsRead = 0
+}
+
+// fill() will make sure at least 32 bits are available.
+func (b *bitReader) fill() {
+	if b.bitsRead < 32 {
+		return
+	}
+	if len(b.in) >= 4 {
+		v := b.in[len(b.in)-4:]
+		b.in = b.in[:len(b.in)-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
+		return
+	}
+
+	b.bitsRead -= uint8(8 * len(b.in))
+	for len(b.in) > 0 {
+		b.value = (b.value << 8) | uint64(b.in[len(b.in)-1])
+		b.in = b.in[:len(b.in)-1]
+	}
+}
+
+// finished returns true if all bits have been read from the bit stream.
+func (b *bitReader) finished() bool {
+	return len(b.in) == 0 && b.bitsRead >= 64
+}
+
+// overread returns true if more bits have been requested than is on the stream.
+func (b *bitReader) overread() bool {
+	return b.bitsRead > 64
+}
+
+// remain returns the number of bits remaining.
+func (b *bitReader) remain() uint {
+	return 8*uint(len(b.in)) + 64 - uint(b.bitsRead)
+}
+
+// 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.finished() {
+		return fmt.Errorf("%d extra bits on block, should be 0", b.remain())
+	}
+	if b.bitsRead > 64 {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}
+
+func highBits(val uint32) (n uint32) {
+	return uint32(bits.Len32(val) - 1)
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/bitwriter.go b/vendor/github.com/klauspost/compress/zstd/bitwriter.go
new file mode 100644
index 000000000..1952f175b
--- /dev/null
+++ b/vendor/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/vendor/github.com/klauspost/compress/zstd/blockdec.go b/vendor/github.com/klauspost/compress/zstd/blockdec.go
new file mode 100644
index 000000000..03744fbc7
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/blockdec.go
@@ -0,0 +1,729 @@
+// 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 (
+	"bytes"
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"hash/crc32"
+	"io"
+	"os"
+	"path/filepath"
+	"sync"
+
+	"github.com/klauspost/compress/huff0"
+	"github.com/klauspost/compress/zstd/internal/xxhash"
+)
+
+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
+)
+
+var (
+	huffDecoderPool = sync.Pool{New: func() interface{} {
+		return &huff0.Scratch{}
+	}}
+
+	fseDecoderPool = sync.Pool{New: func() interface{} {
+		return &fseDecoder{}
+	}}
+)
+
+type blockDec struct {
+	// Raw source data of the block.
+	data        []byte
+	dataStorage []byte
+
+	// Destination of the decoded data.
+	dst []byte
+
+	// Buffer for literals data.
+	literalBuf []byte
+
+	// Window size of the block.
+	WindowSize uint64
+
+	err error
+
+	// Check against this crc, if hasCRC is true.
+	checkCRC uint32
+	hasCRC   bool
+
+	// Frame to use for singlethreaded decoding.
+	// Should not be used by the decoder itself since parent may be another frame.
+	localFrame *frameDec
+
+	sequence []seqVals
+
+	async struct {
+		newHist  *history
+		literals []byte
+		seqData  []byte
+		seqSize  int // Size of uncompressed sequences
+		fcs      uint64
+	}
+
+	// Block is RLE, this is the size.
+	RLESize uint32
+
+	Type blockType
+
+	// Is this the last block of a frame?
+	Last bool
+
+	// Use less memory
+	lowMem bool
+}
+
+func (b *blockDec) String() string {
+	if b == nil {
+		return "<nil>"
+	}
+	return fmt.Sprintf("Steam Size: %d, Type: %v, Last: %t, Window: %d", len(b.data), b.Type, b.Last, b.WindowSize)
+}
+
+func newBlockDec(lowMem bool) *blockDec {
+	b := blockDec{
+		lowMem: lowMem,
+	}
+	return &b
+}
+
+// reset will reset the block.
+// Input must be a start of a block and will be at the end of the block when returned.
+func (b *blockDec) reset(br byteBuffer, windowSize uint64) error {
+	b.WindowSize = windowSize
+	tmp, err := br.readSmall(3)
+	if err != nil {
+		println("Reading block header:", err)
+		return err
+	}
+	bh := uint32(tmp[0]) | (uint32(tmp[1]) << 8) | (uint32(tmp[2]) << 16)
+	b.Last = bh&1 != 0
+	b.Type = blockType((bh >> 1) & 3)
+	// find size.
+	cSize := int(bh >> 3)
+	maxSize := maxCompressedBlockSizeAlloc
+	switch b.Type {
+	case blockTypeReserved:
+		return ErrReservedBlockType
+	case blockTypeRLE:
+		if cSize > maxCompressedBlockSize || cSize > int(b.WindowSize) {
+			if debugDecoder {
+				printf("rle block too big: csize:%d block: %+v\n", uint64(cSize), b)
+			}
+			return ErrWindowSizeExceeded
+		}
+		b.RLESize = uint32(cSize)
+		if b.lowMem {
+			maxSize = cSize
+		}
+		cSize = 1
+	case blockTypeCompressed:
+		if debugDecoder {
+			println("Data size on stream:", cSize)
+		}
+		b.RLESize = 0
+		maxSize = maxCompressedBlockSizeAlloc
+		if windowSize < maxCompressedBlockSize && b.lowMem {
+			maxSize = int(windowSize) + compressedBlockOverAlloc
+		}
+		if cSize > maxCompressedBlockSize || uint64(cSize) > b.WindowSize {
+			if debugDecoder {
+				printf("compressed block too big: csize:%d block: %+v\n", uint64(cSize), b)
+			}
+			return ErrCompressedSizeTooBig
+		}
+		// Empty compressed blocks must at least be 2 bytes
+		// for Literals_Block_Type and one for Sequences_Section_Header.
+		if cSize < 2 {
+			return ErrBlockTooSmall
+		}
+	case blockTypeRaw:
+		if cSize > maxCompressedBlockSize || cSize > int(b.WindowSize) {
+			if debugDecoder {
+				printf("rle block too big: csize:%d block: %+v\n", uint64(cSize), b)
+			}
+			return ErrWindowSizeExceeded
+		}
+
+		b.RLESize = 0
+		// We do not need a destination for raw blocks.
+		maxSize = -1
+	default:
+		panic("Invalid block type")
+	}
+
+	// Read block data.
+	if _, ok := br.(*byteBuf); !ok && cap(b.dataStorage) < cSize {
+		// byteBuf doesn't need a destination buffer.
+		if b.lowMem || cSize > maxCompressedBlockSize {
+			b.dataStorage = make([]byte, 0, cSize+compressedBlockOverAlloc)
+		} else {
+			b.dataStorage = make([]byte, 0, maxCompressedBlockSizeAlloc)
+		}
+	}
+	b.data, err = br.readBig(cSize, b.dataStorage)
+	if err != nil {
+		if debugDecoder {
+			println("Reading block:", err, "(", cSize, ")", len(b.data))
+			printf("%T", br)
+		}
+		return err
+	}
+	if cap(b.dst) <= maxSize {
+		b.dst = make([]byte, 0, maxSize+1)
+	}
+	return nil
+}
+
+// sendEOF will make the decoder send EOF on this frame.
+func (b *blockDec) sendErr(err error) {
+	b.Last = true
+	b.Type = blockTypeReserved
+	b.err = err
+}
+
+// Close will release resources.
+// Closed blockDec cannot be reset.
+func (b *blockDec) Close() {
+}
+
+// decodeBuf
+func (b *blockDec) decodeBuf(hist *history) error {
+	switch b.Type {
+	case blockTypeRLE:
+		if cap(b.dst) < int(b.RLESize) {
+			if b.lowMem {
+				b.dst = make([]byte, b.RLESize)
+			} else {
+				b.dst = make([]byte, maxCompressedBlockSize)
+			}
+		}
+		b.dst = b.dst[:b.RLESize]
+		v := b.data[0]
+		for i := range b.dst {
+			b.dst[i] = v
+		}
+		hist.appendKeep(b.dst)
+		return nil
+	case blockTypeRaw:
+		hist.appendKeep(b.data)
+		return nil
+	case blockTypeCompressed:
+		saved := b.dst
+		// Append directly to history
+		if hist.ignoreBuffer == 0 {
+			b.dst = hist.b
+			hist.b = nil
+		} else {
+			b.dst = b.dst[:0]
+		}
+		err := b.decodeCompressed(hist)
+		if debugDecoder {
+			println("Decompressed to total", len(b.dst), "bytes, hash:", xxhash.Sum64(b.dst), "error:", err)
+		}
+		if hist.ignoreBuffer == 0 {
+			hist.b = b.dst
+			b.dst = saved
+		} else {
+			hist.appendKeep(b.dst)
+		}
+		return err
+	case blockTypeReserved:
+		// Used for returning errors.
+		return b.err
+	default:
+		panic("Invalid block type")
+	}
+}
+
+func (b *blockDec) decodeLiterals(in []byte, hist *history) (remain []byte, err error) {
+	// There must be at least one byte for Literals_Block_Type and one for Sequences_Section_Header
+	if len(in) < 2 {
+		return in, ErrBlockTooSmall
+	}
+
+	litType := literalsBlockType(in[0] & 3)
+	var litRegenSize int
+	var litCompSize int
+	sizeFormat := (in[0] >> 2) & 3
+	var fourStreams bool
+	var literals []byte
+	switch litType {
+	case literalsBlockRaw, literalsBlockRLE:
+		switch sizeFormat {
+		case 0, 2:
+			// Regenerated_Size uses 5 bits (0-31). Literals_Section_Header uses 1 byte.
+			litRegenSize = int(in[0] >> 3)
+			in = in[1:]
+		case 1:
+			// Regenerated_Size uses 12 bits (0-4095). Literals_Section_Header uses 2 bytes.
+			litRegenSize = int(in[0]>>4) + (int(in[1]) << 4)
+			in = in[2:]
+		case 3:
+			//  Regenerated_Size uses 20 bits (0-1048575). Literals_Section_Header uses 3 bytes.
+			if len(in) < 3 {
+				println("too small: litType:", litType, " sizeFormat", sizeFormat, len(in))
+				return in, ErrBlockTooSmall
+			}
+			litRegenSize = int(in[0]>>4) + (int(in[1]) << 4) + (int(in[2]) << 12)
+			in = in[3:]
+		}
+	case literalsBlockCompressed, literalsBlockTreeless:
+		switch sizeFormat {
+		case 0, 1:
+			// Both Regenerated_Size and Compressed_Size use 10 bits (0-1023).
+			if len(in) < 3 {
+				println("too small: litType:", litType, " sizeFormat", sizeFormat, len(in))
+				return in, ErrBlockTooSmall
+			}
+			n := uint64(in[0]>>4) + (uint64(in[1]) << 4) + (uint64(in[2]) << 12)
+			litRegenSize = int(n & 1023)
+			litCompSize = int(n >> 10)
+			fourStreams = sizeFormat == 1
+			in = in[3:]
+		case 2:
+			fourStreams = true
+			if len(in) < 4 {
+				println("too small: litType:", litType, " sizeFormat", sizeFormat, len(in))
+				return in, ErrBlockTooSmall
+			}
+			n := uint64(in[0]>>4) + (uint64(in[1]) << 4) + (uint64(in[2]) << 12) + (uint64(in[3]) << 20)
+			litRegenSize = int(n & 16383)
+			litCompSize = int(n >> 14)
+			in = in[4:]
+		case 3:
+			fourStreams = true
+			if len(in) < 5 {
+				println("too small: litType:", litType, " sizeFormat", sizeFormat, len(in))
+				return in, ErrBlockTooSmall
+			}
+			n := uint64(in[0]>>4) + (uint64(in[1]) << 4) + (uint64(in[2]) << 12) + (uint64(in[3]) << 20) + (uint64(in[4]) << 28)
+			litRegenSize = int(n & 262143)
+			litCompSize = int(n >> 18)
+			in = in[5:]
+		}
+	}
+	if debugDecoder {
+		println("literals type:", litType, "litRegenSize:", litRegenSize, "litCompSize:", litCompSize, "sizeFormat:", sizeFormat, "4X:", fourStreams)
+	}
+	if litRegenSize > int(b.WindowSize) || litRegenSize > maxCompressedBlockSize {
+		return in, ErrWindowSizeExceeded
+	}
+
+	switch litType {
+	case literalsBlockRaw:
+		if len(in) < litRegenSize {
+			println("too small: litType:", litType, " sizeFormat", sizeFormat, "remain:", len(in), "want:", litRegenSize)
+			return in, ErrBlockTooSmall
+		}
+		literals = in[:litRegenSize]
+		in = in[litRegenSize:]
+		//printf("Found %d uncompressed literals\n", litRegenSize)
+	case literalsBlockRLE:
+		if len(in) < 1 {
+			println("too small: litType:", litType, " sizeFormat", sizeFormat, "remain:", len(in), "want:", 1)
+			return in, ErrBlockTooSmall
+		}
+		if cap(b.literalBuf) < litRegenSize {
+			if b.lowMem {
+				b.literalBuf = make([]byte, litRegenSize, litRegenSize+compressedBlockOverAlloc)
+			} else {
+				b.literalBuf = make([]byte, litRegenSize, maxCompressedBlockSize+compressedBlockOverAlloc)
+			}
+		}
+		literals = b.literalBuf[:litRegenSize]
+		v := in[0]
+		for i := range literals {
+			literals[i] = v
+		}
+		in = in[1:]
+		if debugDecoder {
+			printf("Found %d RLE compressed literals\n", litRegenSize)
+		}
+	case literalsBlockTreeless:
+		if len(in) < litCompSize {
+			println("too small: litType:", litType, " sizeFormat", sizeFormat, "remain:", len(in), "want:", litCompSize)
+			return in, ErrBlockTooSmall
+		}
+		// Store compressed literals, so we defer decoding until we get history.
+		literals = in[:litCompSize]
+		in = in[litCompSize:]
+		if debugDecoder {
+			printf("Found %d compressed literals\n", litCompSize)
+		}
+		huff := hist.huffTree
+		if huff == nil {
+			return in, errors.New("literal block was treeless, but no history was defined")
+		}
+		// Ensure we have space to store it.
+		if cap(b.literalBuf) < litRegenSize {
+			if b.lowMem {
+				b.literalBuf = make([]byte, 0, litRegenSize+compressedBlockOverAlloc)
+			} else {
+				b.literalBuf = make([]byte, 0, maxCompressedBlockSize+compressedBlockOverAlloc)
+			}
+		}
+		var err error
+		// Use our out buffer.
+		huff.MaxDecodedSize = litRegenSize
+		if fourStreams {
+			literals, err = huff.Decoder().Decompress4X(b.literalBuf[:0:litRegenSize], literals)
+		} else {
+			literals, err = huff.Decoder().Decompress1X(b.literalBuf[:0:litRegenSize], literals)
+		}
+		// Make sure we don't leak our literals buffer
+		if err != nil {
+			println("decompressing literals:", err)
+			return in, err
+		}
+		if len(literals) != litRegenSize {
+			return in, fmt.Errorf("literal output size mismatch want %d, got %d", litRegenSize, len(literals))
+		}
+
+	case literalsBlockCompressed:
+		if len(in) < litCompSize {
+			println("too small: litType:", litType, " sizeFormat", sizeFormat, "remain:", len(in), "want:", litCompSize)
+			return in, ErrBlockTooSmall
+		}
+		literals = in[:litCompSize]
+		in = in[litCompSize:]
+		// Ensure we have space to store it.
+		if cap(b.literalBuf) < litRegenSize {
+			if b.lowMem {
+				b.literalBuf = make([]byte, 0, litRegenSize+compressedBlockOverAlloc)
+			} else {
+				b.literalBuf = make([]byte, 0, maxCompressedBlockSize+compressedBlockOverAlloc)
+			}
+		}
+		huff := hist.huffTree
+		if huff == nil || (hist.dict != nil && huff == hist.dict.litEnc) {
+			huff = huffDecoderPool.Get().(*huff0.Scratch)
+			if huff == nil {
+				huff = &huff0.Scratch{}
+			}
+		}
+		var err error
+		if debugDecoder {
+			println("huff table input:", len(literals), "CRC:", crc32.ChecksumIEEE(literals))
+		}
+		huff, literals, err = huff0.ReadTable(literals, huff)
+		if err != nil {
+			println("reading huffman table:", err)
+			return in, err
+		}
+		hist.huffTree = huff
+		huff.MaxDecodedSize = litRegenSize
+		// Use our out buffer.
+		if fourStreams {
+			literals, err = huff.Decoder().Decompress4X(b.literalBuf[:0:litRegenSize], literals)
+		} else {
+			literals, err = huff.Decoder().Decompress1X(b.literalBuf[:0:litRegenSize], literals)
+		}
+		if err != nil {
+			println("decoding compressed literals:", err)
+			return in, err
+		}
+		// Make sure we don't leak our literals buffer
+		if len(literals) != litRegenSize {
+			return in, fmt.Errorf("literal output size mismatch want %d, got %d", litRegenSize, len(literals))
+		}
+		// Re-cap to get extra size.
+		literals = b.literalBuf[:len(literals)]
+		if debugDecoder {
+			printf("Decompressed %d literals into %d bytes\n", litCompSize, litRegenSize)
+		}
+	}
+	hist.decoders.literals = literals
+	return in, nil
+}
+
+// decodeCompressed will start decompressing a block.
+func (b *blockDec) decodeCompressed(hist *history) error {
+	in := b.data
+	in, err := b.decodeLiterals(in, hist)
+	if err != nil {
+		return err
+	}
+	err = b.prepareSequences(in, hist)
+	if err != nil {
+		return err
+	}
+	if hist.decoders.nSeqs == 0 {
+		b.dst = append(b.dst, hist.decoders.literals...)
+		return nil
+	}
+	before := len(hist.decoders.out)
+	err = hist.decoders.decodeSync(hist.b[hist.ignoreBuffer:])
+	if err != nil {
+		return err
+	}
+	if hist.decoders.maxSyncLen > 0 {
+		hist.decoders.maxSyncLen += uint64(before)
+		hist.decoders.maxSyncLen -= uint64(len(hist.decoders.out))
+	}
+	b.dst = hist.decoders.out
+	hist.recentOffsets = hist.decoders.prevOffset
+	return nil
+}
+
+func (b *blockDec) prepareSequences(in []byte, hist *history) (err error) {
+	if debugDecoder {
+		printf("prepareSequences: %d byte(s) input\n", len(in))
+	}
+	// Decode Sequences
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#sequences-section
+	if len(in) < 1 {
+		return ErrBlockTooSmall
+	}
+	var nSeqs int
+	seqHeader := in[0]
+	switch {
+	case seqHeader < 128:
+		nSeqs = int(seqHeader)
+		in = in[1:]
+	case seqHeader < 255:
+		if len(in) < 2 {
+			return ErrBlockTooSmall
+		}
+		nSeqs = int(seqHeader-128)<<8 | int(in[1])
+		in = in[2:]
+	case seqHeader == 255:
+		if len(in) < 3 {
+			return ErrBlockTooSmall
+		}
+		nSeqs = 0x7f00 + int(in[1]) + (int(in[2]) << 8)
+		in = in[3:]
+	}
+	if nSeqs == 0 && len(in) != 0 {
+		// When no sequences, there should not be any more data...
+		if debugDecoder {
+			printf("prepareSequences: 0 sequences, but %d byte(s) left on stream\n", len(in))
+		}
+		return ErrUnexpectedBlockSize
+	}
+
+	var seqs = &hist.decoders
+	seqs.nSeqs = nSeqs
+	if nSeqs > 0 {
+		if len(in) < 1 {
+			return ErrBlockTooSmall
+		}
+		br := byteReader{b: in, off: 0}
+		compMode := br.Uint8()
+		br.advance(1)
+		if debugDecoder {
+			printf("Compression modes: 0b%b", compMode)
+		}
+		if compMode&3 != 0 {
+			return errors.New("corrupt block: reserved bits not zero")
+		}
+		for i := uint(0); i < 3; i++ {
+			mode := seqCompMode((compMode >> (6 - i*2)) & 3)
+			if debugDecoder {
+				println("Table", tableIndex(i), "is", mode)
+			}
+			var seq *sequenceDec
+			switch tableIndex(i) {
+			case tableLiteralLengths:
+				seq = &seqs.litLengths
+			case tableOffsets:
+				seq = &seqs.offsets
+			case tableMatchLengths:
+				seq = &seqs.matchLengths
+			default:
+				panic("unknown table")
+			}
+			switch mode {
+			case compModePredefined:
+				if seq.fse != nil && !seq.fse.preDefined {
+					fseDecoderPool.Put(seq.fse)
+				}
+				seq.fse = &fsePredef[i]
+			case compModeRLE:
+				if br.remain() < 1 {
+					return ErrBlockTooSmall
+				}
+				v := br.Uint8()
+				br.advance(1)
+				if seq.fse == nil || seq.fse.preDefined {
+					seq.fse = fseDecoderPool.Get().(*fseDecoder)
+				}
+				symb, err := decSymbolValue(v, symbolTableX[i])
+				if err != nil {
+					printf("RLE Transform table (%v) error: %v", tableIndex(i), err)
+					return err
+				}
+				seq.fse.setRLE(symb)
+				if debugDecoder {
+					printf("RLE set to 0x%x, code: %v", symb, v)
+				}
+			case compModeFSE:
+				println("Reading table for", tableIndex(i))
+				if seq.fse == nil || seq.fse.preDefined {
+					seq.fse = fseDecoderPool.Get().(*fseDecoder)
+				}
+				err := seq.fse.readNCount(&br, uint16(maxTableSymbol[i]))
+				if err != nil {
+					println("Read table error:", err)
+					return err
+				}
+				err = seq.fse.transform(symbolTableX[i])
+				if err != nil {
+					println("Transform table error:", err)
+					return err
+				}
+				if debugDecoder {
+					println("Read table ok", "symbolLen:", seq.fse.symbolLen)
+				}
+			case compModeRepeat:
+				seq.repeat = true
+			}
+			if br.overread() {
+				return io.ErrUnexpectedEOF
+			}
+		}
+		in = br.unread()
+	}
+	if debugDecoder {
+		println("Literals:", len(seqs.literals), "hash:", xxhash.Sum64(seqs.literals), "and", seqs.nSeqs, "sequences.")
+	}
+
+	if nSeqs == 0 {
+		if len(b.sequence) > 0 {
+			b.sequence = b.sequence[:0]
+		}
+		return nil
+	}
+	br := seqs.br
+	if br == nil {
+		br = &bitReader{}
+	}
+	if err := br.init(in); err != nil {
+		return err
+	}
+
+	if err := seqs.initialize(br, hist, b.dst); err != nil {
+		println("initializing sequences:", err)
+		return err
+	}
+	// Extract blocks...
+	if false && hist.dict == nil {
+		fatalErr := func(err error) {
+			if err != nil {
+				panic(err)
+			}
+		}
+		fn := fmt.Sprintf("n-%d-lits-%d-prev-%d-%d-%d-win-%d.blk", hist.decoders.nSeqs, len(hist.decoders.literals), hist.recentOffsets[0], hist.recentOffsets[1], hist.recentOffsets[2], hist.windowSize)
+		var buf bytes.Buffer
+		fatalErr(binary.Write(&buf, binary.LittleEndian, hist.decoders.litLengths.fse))
+		fatalErr(binary.Write(&buf, binary.LittleEndian, hist.decoders.matchLengths.fse))
+		fatalErr(binary.Write(&buf, binary.LittleEndian, hist.decoders.offsets.fse))
+		buf.Write(in)
+		os.WriteFile(filepath.Join("testdata", "seqs", fn), buf.Bytes(), os.ModePerm)
+	}
+
+	return nil
+}
+
+func (b *blockDec) decodeSequences(hist *history) error {
+	if cap(b.sequence) < hist.decoders.nSeqs {
+		if b.lowMem {
+			b.sequence = make([]seqVals, 0, hist.decoders.nSeqs)
+		} else {
+			b.sequence = make([]seqVals, 0, 0x7F00+0xffff)
+		}
+	}
+	b.sequence = b.sequence[:hist.decoders.nSeqs]
+	if hist.decoders.nSeqs == 0 {
+		hist.decoders.seqSize = len(hist.decoders.literals)
+		return nil
+	}
+	hist.decoders.windowSize = hist.windowSize
+	hist.decoders.prevOffset = hist.recentOffsets
+
+	err := hist.decoders.decode(b.sequence)
+	hist.recentOffsets = hist.decoders.prevOffset
+	return err
+}
+
+func (b *blockDec) executeSequences(hist *history) error {
+	hbytes := hist.b
+	if len(hbytes) > hist.windowSize {
+		hbytes = hbytes[len(hbytes)-hist.windowSize:]
+		// We do not need history anymore.
+		if hist.dict != nil {
+			hist.dict.content = nil
+		}
+	}
+	hist.decoders.windowSize = hist.windowSize
+	hist.decoders.out = b.dst[:0]
+	err := hist.decoders.execute(b.sequence, hbytes)
+	if err != nil {
+		return err
+	}
+	return b.updateHistory(hist)
+}
+
+func (b *blockDec) updateHistory(hist *history) error {
+	if len(b.data) > maxCompressedBlockSize {
+		return fmt.Errorf("compressed block size too large (%d)", len(b.data))
+	}
+	// Set output and release references.
+	b.dst = hist.decoders.out
+	hist.recentOffsets = hist.decoders.prevOffset
+
+	if b.Last {
+		// if last block we don't care about history.
+		println("Last block, no history returned")
+		hist.b = hist.b[:0]
+		return nil
+	} else {
+		hist.append(b.dst)
+		if debugDecoder {
+			println("Finished block with ", len(b.sequence), "sequences. Added", len(b.dst), "to history, now length", len(hist.b))
+		}
+	}
+	hist.decoders.out, hist.decoders.literals = nil, nil
+
+	return nil
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/blockenc.go b/vendor/github.com/klauspost/compress/zstd/blockenc.go
new file mode 100644
index 000000000..32a7f401d
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/blockenc.go
@@ -0,0 +1,909 @@
+// 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"
+
+	"github.com/klauspost/compress/huff0"
+)
+
+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)
+}
+
+// fuzzFseEncoder can be used to fuzz the FSE encoder.
+func fuzzFseEncoder(data []byte) int {
+	if len(data) > maxSequences || len(data) < 2 {
+		return 0
+	}
+	enc := fseEncoder{}
+	hist := enc.Histogram()
+	maxSym := uint8(0)
+	for i, v := range data {
+		v = v & 63
+		data[i] = v
+		hist[v]++
+		if v > maxSym {
+			maxSym = v
+		}
+	}
+	if maxSym == 0 {
+		// All 0
+		return 0
+	}
+	maxCount := func(a []uint32) int {
+		var max uint32
+		for _, v := range a {
+			if v > max {
+				max = v
+			}
+		}
+		return int(max)
+	}
+	cnt := maxCount(hist[:maxSym])
+	if cnt == len(data) {
+		// RLE
+		return 0
+	}
+	enc.HistogramFinished(maxSym, cnt)
+	err := enc.normalizeCount(len(data))
+	if err != nil {
+		return 0
+	}
+	_, err = enc.writeCount(nil)
+	if err != nil {
+		panic(err)
+	}
+	return 1
+}
+
+// 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)
+			if debugEncoder {
+				_, _, err := huff0.ReadTable(out, nil)
+				if err != nil {
+					panic(err)
+				}
+			}
+		}
+		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<<outBits | uint64(s.matchLen&bitMask32[outBits])
+		extraBitsN += outBits
+
+		llB := llTT[s.llCode]
+		//ll.encode(llB)
+		nbBitsOut = (uint32(ll.state) + llB.deltaNbBits) >> 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<<outBits | uint64(s.litLen&bitMask32[outBits])
+		extraBitsN += outBits
+
+		wr.flush32()
+		wr.addBits64NC(extraBits, extraBitsN)
+
+		if debugSequences {
+			println("Encoded seq", seq, s)
+		}
+
+		seq--
+	}
+	ml.flush(mlEnc.actualTableLog)
+	of.flush(ofEnc.actualTableLog)
+	ll.flush(llEnc.actualTableLog)
+	wr.close()
+	b.output = wr.out
+
+	// Maybe even add a bigger margin.
+	if len(b.output)-3-bhOffset >= 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))
+			}
+		}
+	}
+	maxCount := func(a []uint32) int {
+		var max uint32
+		for _, v := range a {
+			if v > max {
+				max = v
+			}
+		}
+		return int(max)
+	}
+	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, maxCount(mlH[:mlMax+1]))
+	b.coders.ofEnc.HistogramFinished(ofMax, maxCount(ofH[:ofMax+1]))
+	b.coders.llEnc.HistogramFinished(llMax, maxCount(llH[:llMax+1]))
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/blocktype_string.go b/vendor/github.com/klauspost/compress/zstd/blocktype_string.go
new file mode 100644
index 000000000..01a01e486
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/blocktype_string.go
@@ -0,0 +1,85 @@
+// Code generated by "stringer -type=blockType,literalsBlockType,seqCompMode,tableIndex"; DO NOT EDIT.
+
+package zstd
+
+import "strconv"
+
+func _() {
+	// An "invalid array index" compiler error signifies that the constant values have changed.
+	// Re-run the stringer command to generate them again.
+	var x [1]struct{}
+	_ = x[blockTypeRaw-0]
+	_ = x[blockTypeRLE-1]
+	_ = x[blockTypeCompressed-2]
+	_ = x[blockTypeReserved-3]
+}
+
+const _blockType_name = "blockTypeRawblockTypeRLEblockTypeCompressedblockTypeReserved"
+
+var _blockType_index = [...]uint8{0, 12, 24, 43, 60}
+
+func (i blockType) String() string {
+	if i >= blockType(len(_blockType_index)-1) {
+		return "blockType(" + strconv.FormatInt(int64(i), 10) + ")"
+	}
+	return _blockType_name[_blockType_index[i]:_blockType_index[i+1]]
+}
+func _() {
+	// An "invalid array index" compiler error signifies that the constant values have changed.
+	// Re-run the stringer command to generate them again.
+	var x [1]struct{}
+	_ = x[literalsBlockRaw-0]
+	_ = x[literalsBlockRLE-1]
+	_ = x[literalsBlockCompressed-2]
+	_ = x[literalsBlockTreeless-3]
+}
+
+const _literalsBlockType_name = "literalsBlockRawliteralsBlockRLEliteralsBlockCompressedliteralsBlockTreeless"
+
+var _literalsBlockType_index = [...]uint8{0, 16, 32, 55, 76}
+
+func (i literalsBlockType) String() string {
+	if i >= literalsBlockType(len(_literalsBlockType_index)-1) {
+		return "literalsBlockType(" + strconv.FormatInt(int64(i), 10) + ")"
+	}
+	return _literalsBlockType_name[_literalsBlockType_index[i]:_literalsBlockType_index[i+1]]
+}
+func _() {
+	// An "invalid array index" compiler error signifies that the constant values have changed.
+	// Re-run the stringer command to generate them again.
+	var x [1]struct{}
+	_ = x[compModePredefined-0]
+	_ = x[compModeRLE-1]
+	_ = x[compModeFSE-2]
+	_ = x[compModeRepeat-3]
+}
+
+const _seqCompMode_name = "compModePredefinedcompModeRLEcompModeFSEcompModeRepeat"
+
+var _seqCompMode_index = [...]uint8{0, 18, 29, 40, 54}
+
+func (i seqCompMode) String() string {
+	if i >= seqCompMode(len(_seqCompMode_index)-1) {
+		return "seqCompMode(" + strconv.FormatInt(int64(i), 10) + ")"
+	}
+	return _seqCompMode_name[_seqCompMode_index[i]:_seqCompMode_index[i+1]]
+}
+func _() {
+	// An "invalid array index" compiler error signifies that the constant values have changed.
+	// Re-run the stringer command to generate them again.
+	var x [1]struct{}
+	_ = x[tableLiteralLengths-0]
+	_ = x[tableOffsets-1]
+	_ = x[tableMatchLengths-2]
+}
+
+const _tableIndex_name = "tableLiteralLengthstableOffsetstableMatchLengths"
+
+var _tableIndex_index = [...]uint8{0, 19, 31, 48}
+
+func (i tableIndex) String() string {
+	if i >= tableIndex(len(_tableIndex_index)-1) {
+		return "tableIndex(" + strconv.FormatInt(int64(i), 10) + ")"
+	}
+	return _tableIndex_name[_tableIndex_index[i]:_tableIndex_index[i+1]]
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/bytebuf.go b/vendor/github.com/klauspost/compress/zstd/bytebuf.go
new file mode 100644
index 000000000..55a388553
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/bytebuf.go
@@ -0,0 +1,131 @@
+// 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"
+	"io"
+)
+
+type byteBuffer interface {
+	// Read up to 8 bytes.
+	// Returns io.ErrUnexpectedEOF if this cannot be satisfied.
+	readSmall(n int) ([]byte, error)
+
+	// Read >8 bytes.
+	// MAY use the destination slice.
+	readBig(n int, dst []byte) ([]byte, error)
+
+	// Read a single byte.
+	readByte() (byte, error)
+
+	// Skip n bytes.
+	skipN(n int64) error
+}
+
+// in-memory buffer
+type byteBuf []byte
+
+func (b *byteBuf) readSmall(n int) ([]byte, error) {
+	if debugAsserts && n > 8 {
+		panic(fmt.Errorf("small read > 8 (%d). use readBig", n))
+	}
+	bb := *b
+	if len(bb) < n {
+		return nil, io.ErrUnexpectedEOF
+	}
+	r := bb[:n]
+	*b = bb[n:]
+	return r, nil
+}
+
+func (b *byteBuf) readBig(n int, dst []byte) ([]byte, error) {
+	bb := *b
+	if len(bb) < n {
+		return nil, io.ErrUnexpectedEOF
+	}
+	r := bb[:n]
+	*b = bb[n:]
+	return r, nil
+}
+
+func (b *byteBuf) readByte() (byte, error) {
+	bb := *b
+	if len(bb) < 1 {
+		return 0, io.ErrUnexpectedEOF
+	}
+	r := bb[0]
+	*b = bb[1:]
+	return r, nil
+}
+
+func (b *byteBuf) skipN(n int64) error {
+	bb := *b
+	if n < 0 {
+		return fmt.Errorf("negative skip (%d) requested", n)
+	}
+	if int64(len(bb)) < n {
+		return io.ErrUnexpectedEOF
+	}
+	*b = bb[n:]
+	return nil
+}
+
+// wrapper around a reader.
+type readerWrapper struct {
+	r   io.Reader
+	tmp [8]byte
+}
+
+func (r *readerWrapper) readSmall(n int) ([]byte, error) {
+	if debugAsserts && n > 8 {
+		panic(fmt.Errorf("small read > 8 (%d). use readBig", n))
+	}
+	n2, err := io.ReadFull(r.r, r.tmp[:n])
+	// We only really care about the actual bytes read.
+	if err != nil {
+		if err == io.EOF {
+			return nil, io.ErrUnexpectedEOF
+		}
+		if debugDecoder {
+			println("readSmall: got", n2, "want", n, "err", err)
+		}
+		return nil, err
+	}
+	return r.tmp[:n], nil
+}
+
+func (r *readerWrapper) readBig(n int, dst []byte) ([]byte, error) {
+	if cap(dst) < n {
+		dst = make([]byte, n)
+	}
+	n2, err := io.ReadFull(r.r, dst[:n])
+	if err == io.EOF && n > 0 {
+		err = io.ErrUnexpectedEOF
+	}
+	return dst[:n2], err
+}
+
+func (r *readerWrapper) readByte() (byte, error) {
+	n2, err := io.ReadFull(r.r, r.tmp[:1])
+	if err != nil {
+		if err == io.EOF {
+			err = io.ErrUnexpectedEOF
+		}
+		return 0, err
+	}
+	if n2 != 1 {
+		return 0, io.ErrUnexpectedEOF
+	}
+	return r.tmp[0], nil
+}
+
+func (r *readerWrapper) skipN(n int64) error {
+	n2, err := io.CopyN(io.Discard, r.r, n)
+	if n2 != n {
+		err = io.ErrUnexpectedEOF
+	}
+	return err
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/bytereader.go b/vendor/github.com/klauspost/compress/zstd/bytereader.go
new file mode 100644
index 000000000..0e59a242d
--- /dev/null
+++ b/vendor/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/vendor/github.com/klauspost/compress/zstd/decodeheader.go b/vendor/github.com/klauspost/compress/zstd/decodeheader.go
new file mode 100644
index 000000000..6a5a2988b
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/decodeheader.go
@@ -0,0 +1,261 @@
+// Copyright 2020+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+
+package zstd
+
+import (
+	"encoding/binary"
+	"errors"
+	"io"
+)
+
+// HeaderMaxSize is the maximum size of a Frame and Block Header.
+// If less is sent to Header.Decode it *may* still contain enough information.
+const HeaderMaxSize = 14 + 3
+
+// Header contains information about the first frame and block within that.
+type Header struct {
+	// SingleSegment specifies whether the data is to be decompressed into a
+	// single contiguous memory segment.
+	// It implies that WindowSize is invalid and that FrameContentSize is valid.
+	SingleSegment bool
+
+	// WindowSize is the window of data to keep while decoding.
+	// Will only be set if SingleSegment is false.
+	WindowSize uint64
+
+	// Dictionary ID.
+	// If 0, no dictionary.
+	DictionaryID uint32
+
+	// HasFCS specifies whether FrameContentSize has a valid value.
+	HasFCS bool
+
+	// FrameContentSize is the expected uncompressed size of the entire frame.
+	FrameContentSize uint64
+
+	// Skippable will be true if the frame is meant to be skipped.
+	// This implies that FirstBlock.OK is false.
+	Skippable bool
+
+	// SkippableID is the user-specific ID for the skippable frame.
+	// Valid values are between 0 to 15, inclusive.
+	SkippableID int
+
+	// SkippableSize is the length of the user data to skip following
+	// the header.
+	SkippableSize uint32
+
+	// HeaderSize is the raw size of the frame header.
+	//
+	// For normal frames, it includes the size of the magic number and
+	// the size of the header (per section 3.1.1.1).
+	// It does not include the size for any data blocks (section 3.1.1.2) nor
+	// the size for the trailing content checksum.
+	//
+	// For skippable frames, this counts the size of the magic number
+	// along with the size of the size field of the payload.
+	// It does not include the size of the skippable payload itself.
+	// The total frame size is the HeaderSize plus the SkippableSize.
+	HeaderSize int
+
+	// First block information.
+	FirstBlock struct {
+		// OK will be set if first block could be decoded.
+		OK bool
+
+		// Is this the last block of a frame?
+		Last bool
+
+		// Is the data compressed?
+		// If true CompressedSize will be populated.
+		// Unfortunately DecompressedSize cannot be determined
+		// without decoding the blocks.
+		Compressed bool
+
+		// DecompressedSize is the expected decompressed size of the block.
+		// Will be 0 if it cannot be determined.
+		DecompressedSize int
+
+		// CompressedSize of the data in the block.
+		// Does not include the block header.
+		// Will be equal to DecompressedSize if not Compressed.
+		CompressedSize int
+	}
+
+	// If set there is a checksum present for the block content.
+	// The checksum field at the end is always 4 bytes long.
+	HasCheckSum bool
+}
+
+// Decode the header from the beginning of the stream.
+// This will decode the frame header and the first block header if enough bytes are provided.
+// It is recommended to provide at least HeaderMaxSize bytes.
+// If the frame header cannot be read an error will be returned.
+// If there isn't enough input, io.ErrUnexpectedEOF is returned.
+// The FirstBlock.OK will indicate if enough information was available to decode the first block header.
+func (h *Header) Decode(in []byte) error {
+	_, err := h.DecodeAndStrip(in)
+	return err
+}
+
+// DecodeAndStrip will decode the header from the beginning of the stream
+// and on success return the remaining bytes.
+// This will decode the frame header and the first block header if enough bytes are provided.
+// It is recommended to provide at least HeaderMaxSize bytes.
+// If the frame header cannot be read an error will be returned.
+// If there isn't enough input, io.ErrUnexpectedEOF is returned.
+// The FirstBlock.OK will indicate if enough information was available to decode the first block header.
+func (h *Header) DecodeAndStrip(in []byte) (remain []byte, err error) {
+	*h = Header{}
+	if len(in) < 4 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	h.HeaderSize += 4
+	b, in := in[:4], in[4:]
+	if string(b) != frameMagic {
+		if string(b[1:4]) != skippableFrameMagic || b[0]&0xf0 != 0x50 {
+			return nil, ErrMagicMismatch
+		}
+		if len(in) < 4 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		h.HeaderSize += 4
+		h.Skippable = true
+		h.SkippableID = int(b[0] & 0xf)
+		h.SkippableSize = binary.LittleEndian.Uint32(in)
+		return in[4:], nil
+	}
+
+	// Read Window_Descriptor
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#window_descriptor
+	if len(in) < 1 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	fhd, in := in[0], in[1:]
+	h.HeaderSize++
+	h.SingleSegment = fhd&(1<<5) != 0
+	h.HasCheckSum = fhd&(1<<2) != 0
+	if fhd&(1<<3) != 0 {
+		return nil, errors.New("reserved bit set on frame header")
+	}
+
+	if !h.SingleSegment {
+		if len(in) < 1 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		var wd byte
+		wd, in = in[0], in[1:]
+		h.HeaderSize++
+		windowLog := 10 + (wd >> 3)
+		windowBase := uint64(1) << windowLog
+		windowAdd := (windowBase / 8) * uint64(wd&0x7)
+		h.WindowSize = windowBase + windowAdd
+	}
+
+	// Read Dictionary_ID
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#dictionary_id
+	if size := fhd & 3; size != 0 {
+		if size == 3 {
+			size = 4
+		}
+		if len(in) < int(size) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b, in = in[:size], in[size:]
+		h.HeaderSize += int(size)
+		switch len(b) {
+		case 1:
+			h.DictionaryID = uint32(b[0])
+		case 2:
+			h.DictionaryID = uint32(b[0]) | (uint32(b[1]) << 8)
+		case 4:
+			h.DictionaryID = uint32(b[0]) | (uint32(b[1]) << 8) | (uint32(b[2]) << 16) | (uint32(b[3]) << 24)
+		}
+	}
+
+	// Read Frame_Content_Size
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#frame_content_size
+	var fcsSize int
+	v := fhd >> 6
+	switch v {
+	case 0:
+		if h.SingleSegment {
+			fcsSize = 1
+		}
+	default:
+		fcsSize = 1 << v
+	}
+
+	if fcsSize > 0 {
+		h.HasFCS = true
+		if len(in) < fcsSize {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b, in = in[:fcsSize], in[fcsSize:]
+		h.HeaderSize += int(fcsSize)
+		switch len(b) {
+		case 1:
+			h.FrameContentSize = uint64(b[0])
+		case 2:
+			// When FCS_Field_Size is 2, the offset of 256 is added.
+			h.FrameContentSize = uint64(b[0]) | (uint64(b[1]) << 8) + 256
+		case 4:
+			h.FrameContentSize = uint64(b[0]) | (uint64(b[1]) << 8) | (uint64(b[2]) << 16) | (uint64(b[3]) << 24)
+		case 8:
+			d1 := uint32(b[0]) | (uint32(b[1]) << 8) | (uint32(b[2]) << 16) | (uint32(b[3]) << 24)
+			d2 := uint32(b[4]) | (uint32(b[5]) << 8) | (uint32(b[6]) << 16) | (uint32(b[7]) << 24)
+			h.FrameContentSize = uint64(d1) | (uint64(d2) << 32)
+		}
+	}
+
+	// Frame Header done, we will not fail from now on.
+	if len(in) < 3 {
+		return in, nil
+	}
+	tmp := in[:3]
+	bh := uint32(tmp[0]) | (uint32(tmp[1]) << 8) | (uint32(tmp[2]) << 16)
+	h.FirstBlock.Last = bh&1 != 0
+	blockType := blockType((bh >> 1) & 3)
+	// find size.
+	cSize := int(bh >> 3)
+	switch blockType {
+	case blockTypeReserved:
+		return in, nil
+	case blockTypeRLE:
+		h.FirstBlock.Compressed = true
+		h.FirstBlock.DecompressedSize = cSize
+		h.FirstBlock.CompressedSize = 1
+	case blockTypeCompressed:
+		h.FirstBlock.Compressed = true
+		h.FirstBlock.CompressedSize = cSize
+	case blockTypeRaw:
+		h.FirstBlock.DecompressedSize = cSize
+		h.FirstBlock.CompressedSize = cSize
+	default:
+		panic("Invalid block type")
+	}
+
+	h.FirstBlock.OK = true
+	return in, nil
+}
+
+// AppendTo will append the encoded header to the dst slice.
+// There is no error checking performed on the header values.
+func (h *Header) AppendTo(dst []byte) ([]byte, error) {
+	if h.Skippable {
+		magic := [4]byte{0x50, 0x2a, 0x4d, 0x18}
+		magic[0] |= byte(h.SkippableID & 0xf)
+		dst = append(dst, magic[:]...)
+		f := h.SkippableSize
+		return append(dst, uint8(f), uint8(f>>8), uint8(f>>16), uint8(f>>24)), nil
+	}
+	f := frameHeader{
+		ContentSize:   h.FrameContentSize,
+		WindowSize:    uint32(h.WindowSize),
+		SingleSegment: h.SingleSegment,
+		Checksum:      h.HasCheckSum,
+		DictID:        h.DictionaryID,
+	}
+	return f.appendTo(dst), nil
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/decoder.go b/vendor/github.com/klauspost/compress/zstd/decoder.go
new file mode 100644
index 000000000..bbca17234
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/decoder.go
@@ -0,0 +1,948 @@
+// 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 (
+	"context"
+	"encoding/binary"
+	"io"
+	"sync"
+
+	"github.com/klauspost/compress/zstd/internal/xxhash"
+)
+
+// Decoder provides decoding of zstandard streams.
+// The decoder has been designed to operate without allocations after a warmup.
+// This means that you should store the decoder for best performance.
+// To re-use a stream decoder, use the Reset(r io.Reader) error to switch to another stream.
+// A decoder can safely be re-used even if the previous stream failed.
+// To release the resources, you must call the Close() function on a decoder.
+type Decoder struct {
+	o decoderOptions
+
+	// Unreferenced decoders, ready for use.
+	decoders chan *blockDec
+
+	// Current read position used for Reader functionality.
+	current decoderState
+
+	// sync stream decoding
+	syncStream struct {
+		decodedFrame uint64
+		br           readerWrapper
+		enabled      bool
+		inFrame      bool
+		dstBuf       []byte
+	}
+
+	frame *frameDec
+
+	// Custom dictionaries.
+	dicts map[uint32]*dict
+
+	// streamWg is the waitgroup for all streams
+	streamWg sync.WaitGroup
+}
+
+// decoderState is used for maintaining state when the decoder
+// is used for streaming.
+type decoderState struct {
+	// current block being written to stream.
+	decodeOutput
+
+	// output in order to be written to stream.
+	output chan decodeOutput
+
+	// cancel remaining output.
+	cancel context.CancelFunc
+
+	// crc of current frame
+	crc *xxhash.Digest
+
+	flushed bool
+}
+
+var (
+	// Check the interfaces we want to support.
+	_ = io.WriterTo(&Decoder{})
+	_ = io.Reader(&Decoder{})
+)
+
+// NewReader creates a new decoder.
+// A nil Reader can be provided in which case Reset can be used to start a decode.
+//
+// A Decoder can be used in two modes:
+//
+// 1) As a stream, or
+// 2) For stateless decoding using DecodeAll.
+//
+// Only a single stream can be decoded concurrently, but the same decoder
+// can run multiple concurrent stateless decodes. It is even possible to
+// use stateless decodes while a stream is being decoded.
+//
+// The Reset function can be used to initiate a new stream, which will considerably
+// reduce the allocations normally caused by NewReader.
+func NewReader(r io.Reader, opts ...DOption) (*Decoder, error) {
+	initPredefined()
+	var d Decoder
+	d.o.setDefault()
+	for _, o := range opts {
+		err := o(&d.o)
+		if err != nil {
+			return nil, err
+		}
+	}
+	d.current.crc = xxhash.New()
+	d.current.flushed = true
+
+	if r == nil {
+		d.current.err = ErrDecoderNilInput
+	}
+
+	// Transfer option dicts.
+	d.dicts = make(map[uint32]*dict, len(d.o.dicts))
+	for _, dc := range d.o.dicts {
+		d.dicts[dc.id] = dc
+	}
+	d.o.dicts = nil
+
+	// Create decoders
+	d.decoders = make(chan *blockDec, d.o.concurrent)
+	for i := 0; i < d.o.concurrent; i++ {
+		dec := newBlockDec(d.o.lowMem)
+		dec.localFrame = newFrameDec(d.o)
+		d.decoders <- dec
+	}
+
+	if r == nil {
+		return &d, nil
+	}
+	return &d, d.Reset(r)
+}
+
+// Read bytes from the decompressed stream into p.
+// Returns the number of bytes written and any error that occurred.
+// When the stream is done, io.EOF will be returned.
+func (d *Decoder) Read(p []byte) (int, error) {
+	var n int
+	for {
+		if len(d.current.b) > 0 {
+			filled := copy(p, d.current.b)
+			p = p[filled:]
+			d.current.b = d.current.b[filled:]
+			n += filled
+		}
+		if len(p) == 0 {
+			break
+		}
+		if len(d.current.b) == 0 {
+			// We have an error and no more data
+			if d.current.err != nil {
+				break
+			}
+			if !d.nextBlock(n == 0) {
+				return n, d.current.err
+			}
+		}
+	}
+	if len(d.current.b) > 0 {
+		if debugDecoder {
+			println("returning", n, "still bytes left:", len(d.current.b))
+		}
+		// Only return error at end of block
+		return n, nil
+	}
+	if d.current.err != nil {
+		d.drainOutput()
+	}
+	if debugDecoder {
+		println("returning", n, d.current.err, len(d.decoders))
+	}
+	return n, d.current.err
+}
+
+// Reset will reset the decoder the supplied stream after the current has finished processing.
+// Note that this functionality cannot be used after Close has been called.
+// Reset can be called with a nil reader to release references to the previous reader.
+// After being called with a nil reader, no other operations than Reset or DecodeAll or Close
+// should be used.
+func (d *Decoder) Reset(r io.Reader) error {
+	if d.current.err == ErrDecoderClosed {
+		return d.current.err
+	}
+
+	d.drainOutput()
+
+	d.syncStream.br.r = nil
+	if r == nil {
+		d.current.err = ErrDecoderNilInput
+		if len(d.current.b) > 0 {
+			d.current.b = d.current.b[:0]
+		}
+		d.current.flushed = true
+		return nil
+	}
+
+	// If bytes buffer and < 5MB, do sync decoding anyway.
+	if bb, ok := r.(byter); ok && bb.Len() < d.o.decodeBufsBelow && !d.o.limitToCap {
+		bb2 := bb
+		if debugDecoder {
+			println("*bytes.Buffer detected, doing sync decode, len:", bb.Len())
+		}
+		b := bb2.Bytes()
+		var dst []byte
+		if cap(d.syncStream.dstBuf) > 0 {
+			dst = d.syncStream.dstBuf[:0]
+		}
+
+		dst, err := d.DecodeAll(b, dst)
+		if err == nil {
+			err = io.EOF
+		}
+		// Save output buffer
+		d.syncStream.dstBuf = dst
+		d.current.b = dst
+		d.current.err = err
+		d.current.flushed = true
+		if debugDecoder {
+			println("sync decode to", len(dst), "bytes, err:", err)
+		}
+		return nil
+	}
+	// Remove current block.
+	d.stashDecoder()
+	d.current.decodeOutput = decodeOutput{}
+	d.current.err = nil
+	d.current.flushed = false
+	d.current.d = nil
+	d.syncStream.dstBuf = nil
+
+	// Ensure no-one else is still running...
+	d.streamWg.Wait()
+	if d.frame == nil {
+		d.frame = newFrameDec(d.o)
+	}
+
+	if d.o.concurrent == 1 {
+		return d.startSyncDecoder(r)
+	}
+
+	d.current.output = make(chan decodeOutput, d.o.concurrent)
+	ctx, cancel := context.WithCancel(context.Background())
+	d.current.cancel = cancel
+	d.streamWg.Add(1)
+	go d.startStreamDecoder(ctx, r, d.current.output)
+
+	return nil
+}
+
+// drainOutput will drain the output until errEndOfStream is sent.
+func (d *Decoder) drainOutput() {
+	if d.current.cancel != nil {
+		if debugDecoder {
+			println("cancelling current")
+		}
+		d.current.cancel()
+		d.current.cancel = nil
+	}
+	if d.current.d != nil {
+		if debugDecoder {
+			printf("re-adding current decoder %p, decoders: %d", d.current.d, len(d.decoders))
+		}
+		d.decoders <- d.current.d
+		d.current.d = nil
+		d.current.b = nil
+	}
+	if d.current.output == nil || d.current.flushed {
+		println("current already flushed")
+		return
+	}
+	for v := range d.current.output {
+		if v.d != nil {
+			if debugDecoder {
+				printf("re-adding decoder %p", v.d)
+			}
+			d.decoders <- v.d
+		}
+	}
+	d.current.output = nil
+	d.current.flushed = true
+}
+
+// WriteTo writes data to w until there's no more data to write or when an error occurs.
+// The return value n is the number of bytes written.
+// Any error encountered during the write is also returned.
+func (d *Decoder) WriteTo(w io.Writer) (int64, error) {
+	var n int64
+	for {
+		if len(d.current.b) > 0 {
+			n2, err2 := w.Write(d.current.b)
+			n += int64(n2)
+			if err2 != nil && (d.current.err == nil || d.current.err == io.EOF) {
+				d.current.err = err2
+			} else if n2 != len(d.current.b) {
+				d.current.err = io.ErrShortWrite
+			}
+		}
+		if d.current.err != nil {
+			break
+		}
+		d.nextBlock(true)
+	}
+	err := d.current.err
+	if err != nil {
+		d.drainOutput()
+	}
+	if err == io.EOF {
+		err = nil
+	}
+	return n, err
+}
+
+// DecodeAll allows stateless decoding of a blob of bytes.
+// Output will be appended to dst, so if the destination size is known
+// you can pre-allocate the destination slice to avoid allocations.
+// DecodeAll can be used concurrently.
+// The Decoder concurrency limits will be respected.
+func (d *Decoder) DecodeAll(input, dst []byte) ([]byte, error) {
+	if d.decoders == nil {
+		return dst, ErrDecoderClosed
+	}
+
+	// Grab a block decoder and frame decoder.
+	block := <-d.decoders
+	frame := block.localFrame
+	initialSize := len(dst)
+	defer func() {
+		if debugDecoder {
+			printf("re-adding decoder: %p", block)
+		}
+		frame.rawInput = nil
+		frame.bBuf = nil
+		if frame.history.decoders.br != nil {
+			frame.history.decoders.br.in = nil
+		}
+		d.decoders <- block
+	}()
+	frame.bBuf = input
+
+	for {
+		frame.history.reset()
+		err := frame.reset(&frame.bBuf)
+		if err != nil {
+			if err == io.EOF {
+				if debugDecoder {
+					println("frame reset return EOF")
+				}
+				return dst, nil
+			}
+			return dst, err
+		}
+		if err = d.setDict(frame); err != nil {
+			return nil, err
+		}
+		if frame.WindowSize > d.o.maxWindowSize {
+			if debugDecoder {
+				println("window size exceeded:", frame.WindowSize, ">", d.o.maxWindowSize)
+			}
+			return dst, ErrWindowSizeExceeded
+		}
+		if frame.FrameContentSize != fcsUnknown {
+			if frame.FrameContentSize > d.o.maxDecodedSize-uint64(len(dst)-initialSize) {
+				if debugDecoder {
+					println("decoder size exceeded; fcs:", frame.FrameContentSize, "> mcs:", d.o.maxDecodedSize-uint64(len(dst)-initialSize), "len:", len(dst))
+				}
+				return dst, ErrDecoderSizeExceeded
+			}
+			if d.o.limitToCap && frame.FrameContentSize > uint64(cap(dst)-len(dst)) {
+				if debugDecoder {
+					println("decoder size exceeded; fcs:", frame.FrameContentSize, "> (cap-len)", cap(dst)-len(dst))
+				}
+				return dst, ErrDecoderSizeExceeded
+			}
+			if cap(dst)-len(dst) < int(frame.FrameContentSize) {
+				dst2 := make([]byte, len(dst), len(dst)+int(frame.FrameContentSize)+compressedBlockOverAlloc)
+				copy(dst2, dst)
+				dst = dst2
+			}
+		}
+
+		if cap(dst) == 0 && !d.o.limitToCap {
+			// Allocate len(input) * 2 by default if nothing is provided
+			// and we didn't get frame content size.
+			size := len(input) * 2
+			// Cap to 1 MB.
+			if size > 1<<20 {
+				size = 1 << 20
+			}
+			if uint64(size) > d.o.maxDecodedSize {
+				size = int(d.o.maxDecodedSize)
+			}
+			dst = make([]byte, 0, size)
+		}
+
+		dst, err = frame.runDecoder(dst, block)
+		if err != nil {
+			return dst, err
+		}
+		if uint64(len(dst)-initialSize) > d.o.maxDecodedSize {
+			return dst, ErrDecoderSizeExceeded
+		}
+		if len(frame.bBuf) == 0 {
+			if debugDecoder {
+				println("frame dbuf empty")
+			}
+			break
+		}
+	}
+	return dst, nil
+}
+
+// nextBlock returns the next block.
+// If an error occurs d.err will be set.
+// Optionally the function can block for new output.
+// If non-blocking mode is used the returned boolean will be false
+// if no data was available without blocking.
+func (d *Decoder) nextBlock(blocking bool) (ok bool) {
+	if d.current.err != nil {
+		// Keep error state.
+		return false
+	}
+	d.current.b = d.current.b[:0]
+
+	// SYNC:
+	if d.syncStream.enabled {
+		if !blocking {
+			return false
+		}
+		ok = d.nextBlockSync()
+		if !ok {
+			d.stashDecoder()
+		}
+		return ok
+	}
+
+	//ASYNC:
+	d.stashDecoder()
+	if blocking {
+		d.current.decodeOutput, ok = <-d.current.output
+	} else {
+		select {
+		case d.current.decodeOutput, ok = <-d.current.output:
+		default:
+			return false
+		}
+	}
+	if !ok {
+		// This should not happen, so signal error state...
+		d.current.err = io.ErrUnexpectedEOF
+		return false
+	}
+	next := d.current.decodeOutput
+	if next.d != nil && next.d.async.newHist != nil {
+		d.current.crc.Reset()
+	}
+	if debugDecoder {
+		var tmp [4]byte
+		binary.LittleEndian.PutUint32(tmp[:], uint32(xxhash.Sum64(next.b)))
+		println("got", len(d.current.b), "bytes, error:", d.current.err, "data crc:", tmp)
+	}
+
+	if d.o.ignoreChecksum {
+		return true
+	}
+
+	if len(next.b) > 0 {
+		d.current.crc.Write(next.b)
+	}
+	if next.err == nil && next.d != nil && next.d.hasCRC {
+		got := uint32(d.current.crc.Sum64())
+		if got != next.d.checkCRC {
+			if debugDecoder {
+				printf("CRC Check Failed: %08x (got) != %08x (on stream)\n", got, next.d.checkCRC)
+			}
+			d.current.err = ErrCRCMismatch
+		} else {
+			if debugDecoder {
+				printf("CRC ok %08x\n", got)
+			}
+		}
+	}
+
+	return true
+}
+
+func (d *Decoder) nextBlockSync() (ok bool) {
+	if d.current.d == nil {
+		d.current.d = <-d.decoders
+	}
+	for len(d.current.b) == 0 {
+		if !d.syncStream.inFrame {
+			d.frame.history.reset()
+			d.current.err = d.frame.reset(&d.syncStream.br)
+			if d.current.err == nil {
+				d.current.err = d.setDict(d.frame)
+			}
+			if d.current.err != nil {
+				return false
+			}
+			if d.frame.WindowSize > d.o.maxDecodedSize || d.frame.WindowSize > d.o.maxWindowSize {
+				d.current.err = ErrDecoderSizeExceeded
+				return false
+			}
+
+			d.syncStream.decodedFrame = 0
+			d.syncStream.inFrame = true
+		}
+		d.current.err = d.frame.next(d.current.d)
+		if d.current.err != nil {
+			return false
+		}
+		d.frame.history.ensureBlock()
+		if debugDecoder {
+			println("History trimmed:", len(d.frame.history.b), "decoded already:", d.syncStream.decodedFrame)
+		}
+		histBefore := len(d.frame.history.b)
+		d.current.err = d.current.d.decodeBuf(&d.frame.history)
+
+		if d.current.err != nil {
+			println("error after:", d.current.err)
+			return false
+		}
+		d.current.b = d.frame.history.b[histBefore:]
+		if debugDecoder {
+			println("history after:", len(d.frame.history.b))
+		}
+
+		// Check frame size (before CRC)
+		d.syncStream.decodedFrame += uint64(len(d.current.b))
+		if d.syncStream.decodedFrame > d.frame.FrameContentSize {
+			if debugDecoder {
+				printf("DecodedFrame (%d) > FrameContentSize (%d)\n", d.syncStream.decodedFrame, d.frame.FrameContentSize)
+			}
+			d.current.err = ErrFrameSizeExceeded
+			return false
+		}
+
+		// Check FCS
+		if d.current.d.Last && d.frame.FrameContentSize != fcsUnknown && d.syncStream.decodedFrame != d.frame.FrameContentSize {
+			if debugDecoder {
+				printf("DecodedFrame (%d) != FrameContentSize (%d)\n", d.syncStream.decodedFrame, d.frame.FrameContentSize)
+			}
+			d.current.err = ErrFrameSizeMismatch
+			return false
+		}
+
+		// Update/Check CRC
+		if d.frame.HasCheckSum {
+			if !d.o.ignoreChecksum {
+				d.frame.crc.Write(d.current.b)
+			}
+			if d.current.d.Last {
+				if !d.o.ignoreChecksum {
+					d.current.err = d.frame.checkCRC()
+				} else {
+					d.current.err = d.frame.consumeCRC()
+				}
+				if d.current.err != nil {
+					println("CRC error:", d.current.err)
+					return false
+				}
+			}
+		}
+		d.syncStream.inFrame = !d.current.d.Last
+	}
+	return true
+}
+
+func (d *Decoder) stashDecoder() {
+	if d.current.d != nil {
+		if debugDecoder {
+			printf("re-adding current decoder %p", d.current.d)
+		}
+		d.decoders <- d.current.d
+		d.current.d = nil
+	}
+}
+
+// Close will release all resources.
+// It is NOT possible to reuse the decoder after this.
+func (d *Decoder) Close() {
+	if d.current.err == ErrDecoderClosed {
+		return
+	}
+	d.drainOutput()
+	if d.current.cancel != nil {
+		d.current.cancel()
+		d.streamWg.Wait()
+		d.current.cancel = nil
+	}
+	if d.decoders != nil {
+		close(d.decoders)
+		for dec := range d.decoders {
+			dec.Close()
+		}
+		d.decoders = nil
+	}
+	if d.current.d != nil {
+		d.current.d.Close()
+		d.current.d = nil
+	}
+	d.current.err = ErrDecoderClosed
+}
+
+// IOReadCloser returns the decoder as an io.ReadCloser for convenience.
+// Any changes to the decoder will be reflected, so the returned ReadCloser
+// can be reused along with the decoder.
+// io.WriterTo is also supported by the returned ReadCloser.
+func (d *Decoder) IOReadCloser() io.ReadCloser {
+	return closeWrapper{d: d}
+}
+
+// closeWrapper wraps a function call as a closer.
+type closeWrapper struct {
+	d *Decoder
+}
+
+// WriteTo forwards WriteTo calls to the decoder.
+func (c closeWrapper) WriteTo(w io.Writer) (n int64, err error) {
+	return c.d.WriteTo(w)
+}
+
+// Read forwards read calls to the decoder.
+func (c closeWrapper) Read(p []byte) (n int, err error) {
+	return c.d.Read(p)
+}
+
+// Close closes the decoder.
+func (c closeWrapper) Close() error {
+	c.d.Close()
+	return nil
+}
+
+type decodeOutput struct {
+	d   *blockDec
+	b   []byte
+	err error
+}
+
+func (d *Decoder) startSyncDecoder(r io.Reader) error {
+	d.frame.history.reset()
+	d.syncStream.br = readerWrapper{r: r}
+	d.syncStream.inFrame = false
+	d.syncStream.enabled = true
+	d.syncStream.decodedFrame = 0
+	return nil
+}
+
+// Create Decoder:
+// ASYNC:
+// Spawn 3 go routines.
+// 0: Read frames and decode block literals.
+// 1: Decode sequences.
+// 2: Execute sequences, send to output.
+func (d *Decoder) startStreamDecoder(ctx context.Context, r io.Reader, output chan decodeOutput) {
+	defer d.streamWg.Done()
+	br := readerWrapper{r: r}
+
+	var seqDecode = make(chan *blockDec, d.o.concurrent)
+	var seqExecute = make(chan *blockDec, d.o.concurrent)
+
+	// Async 1: Decode sequences...
+	go func() {
+		var hist history
+		var hasErr bool
+
+		for block := range seqDecode {
+			if hasErr {
+				if block != nil {
+					seqExecute <- block
+				}
+				continue
+			}
+			if block.async.newHist != nil {
+				if debugDecoder {
+					println("Async 1: new history, recent:", block.async.newHist.recentOffsets)
+				}
+				hist.reset()
+				hist.decoders = block.async.newHist.decoders
+				hist.recentOffsets = block.async.newHist.recentOffsets
+				hist.windowSize = block.async.newHist.windowSize
+				if block.async.newHist.dict != nil {
+					hist.setDict(block.async.newHist.dict)
+				}
+			}
+			if block.err != nil || block.Type != blockTypeCompressed {
+				hasErr = block.err != nil
+				seqExecute <- block
+				continue
+			}
+
+			hist.decoders.literals = block.async.literals
+			block.err = block.prepareSequences(block.async.seqData, &hist)
+			if debugDecoder && block.err != nil {
+				println("prepareSequences returned:", block.err)
+			}
+			hasErr = block.err != nil
+			if block.err == nil {
+				block.err = block.decodeSequences(&hist)
+				if debugDecoder && block.err != nil {
+					println("decodeSequences returned:", block.err)
+				}
+				hasErr = block.err != nil
+				//				block.async.sequence = hist.decoders.seq[:hist.decoders.nSeqs]
+				block.async.seqSize = hist.decoders.seqSize
+			}
+			seqExecute <- block
+		}
+		close(seqExecute)
+		hist.reset()
+	}()
+
+	var wg sync.WaitGroup
+	wg.Add(1)
+
+	// Async 3: Execute sequences...
+	frameHistCache := d.frame.history.b
+	go func() {
+		var hist history
+		var decodedFrame uint64
+		var fcs uint64
+		var hasErr bool
+		for block := range seqExecute {
+			out := decodeOutput{err: block.err, d: block}
+			if block.err != nil || hasErr {
+				hasErr = true
+				output <- out
+				continue
+			}
+			if block.async.newHist != nil {
+				if debugDecoder {
+					println("Async 2: new history")
+				}
+				hist.reset()
+				hist.windowSize = block.async.newHist.windowSize
+				hist.allocFrameBuffer = block.async.newHist.allocFrameBuffer
+				if block.async.newHist.dict != nil {
+					hist.setDict(block.async.newHist.dict)
+				}
+
+				if cap(hist.b) < hist.allocFrameBuffer {
+					if cap(frameHistCache) >= hist.allocFrameBuffer {
+						hist.b = frameHistCache
+					} else {
+						hist.b = make([]byte, 0, hist.allocFrameBuffer)
+						println("Alloc history sized", hist.allocFrameBuffer)
+					}
+				}
+				hist.b = hist.b[:0]
+				fcs = block.async.fcs
+				decodedFrame = 0
+			}
+			do := decodeOutput{err: block.err, d: block}
+			switch block.Type {
+			case blockTypeRLE:
+				if debugDecoder {
+					println("add rle block length:", block.RLESize)
+				}
+
+				if cap(block.dst) < int(block.RLESize) {
+					if block.lowMem {
+						block.dst = make([]byte, block.RLESize)
+					} else {
+						block.dst = make([]byte, maxCompressedBlockSize)
+					}
+				}
+				block.dst = block.dst[:block.RLESize]
+				v := block.data[0]
+				for i := range block.dst {
+					block.dst[i] = v
+				}
+				hist.append(block.dst)
+				do.b = block.dst
+			case blockTypeRaw:
+				if debugDecoder {
+					println("add raw block length:", len(block.data))
+				}
+				hist.append(block.data)
+				do.b = block.data
+			case blockTypeCompressed:
+				if debugDecoder {
+					println("execute with history length:", len(hist.b), "window:", hist.windowSize)
+				}
+				hist.decoders.seqSize = block.async.seqSize
+				hist.decoders.literals = block.async.literals
+				do.err = block.executeSequences(&hist)
+				hasErr = do.err != nil
+				if debugDecoder && hasErr {
+					println("executeSequences returned:", do.err)
+				}
+				do.b = block.dst
+			}
+			if !hasErr {
+				decodedFrame += uint64(len(do.b))
+				if decodedFrame > fcs {
+					println("fcs exceeded", block.Last, fcs, decodedFrame)
+					do.err = ErrFrameSizeExceeded
+					hasErr = true
+				} else if block.Last && fcs != fcsUnknown && decodedFrame != fcs {
+					do.err = ErrFrameSizeMismatch
+					hasErr = true
+				} else {
+					if debugDecoder {
+						println("fcs ok", block.Last, fcs, decodedFrame)
+					}
+				}
+			}
+			output <- do
+		}
+		close(output)
+		frameHistCache = hist.b
+		wg.Done()
+		if debugDecoder {
+			println("decoder goroutines finished")
+		}
+		hist.reset()
+	}()
+
+	var hist history
+decodeStream:
+	for {
+		var hasErr bool
+		hist.reset()
+		decodeBlock := func(block *blockDec) {
+			if hasErr {
+				if block != nil {
+					seqDecode <- block
+				}
+				return
+			}
+			if block.err != nil || block.Type != blockTypeCompressed {
+				hasErr = block.err != nil
+				seqDecode <- block
+				return
+			}
+
+			remain, err := block.decodeLiterals(block.data, &hist)
+			block.err = err
+			hasErr = block.err != nil
+			if err == nil {
+				block.async.literals = hist.decoders.literals
+				block.async.seqData = remain
+			} else if debugDecoder {
+				println("decodeLiterals error:", err)
+			}
+			seqDecode <- block
+		}
+		frame := d.frame
+		if debugDecoder {
+			println("New frame...")
+		}
+		var historySent bool
+		frame.history.reset()
+		err := frame.reset(&br)
+		if debugDecoder && err != nil {
+			println("Frame decoder returned", err)
+		}
+		if err == nil {
+			err = d.setDict(frame)
+		}
+		if err == nil && d.frame.WindowSize > d.o.maxWindowSize {
+			if debugDecoder {
+				println("decoder size exceeded, fws:", d.frame.WindowSize, "> mws:", d.o.maxWindowSize)
+			}
+
+			err = ErrDecoderSizeExceeded
+		}
+		if err != nil {
+			select {
+			case <-ctx.Done():
+			case dec := <-d.decoders:
+				dec.sendErr(err)
+				decodeBlock(dec)
+			}
+			break decodeStream
+		}
+
+		// Go through all blocks of the frame.
+		for {
+			var dec *blockDec
+			select {
+			case <-ctx.Done():
+				break decodeStream
+			case dec = <-d.decoders:
+				// Once we have a decoder, we MUST return it.
+			}
+			err := frame.next(dec)
+			if !historySent {
+				h := frame.history
+				if debugDecoder {
+					println("Alloc History:", h.allocFrameBuffer)
+				}
+				hist.reset()
+				if h.dict != nil {
+					hist.setDict(h.dict)
+				}
+				dec.async.newHist = &h
+				dec.async.fcs = frame.FrameContentSize
+				historySent = true
+			} else {
+				dec.async.newHist = nil
+			}
+			if debugDecoder && err != nil {
+				println("next block returned error:", err)
+			}
+			dec.err = err
+			dec.hasCRC = false
+			if dec.Last && frame.HasCheckSum && err == nil {
+				crc, err := frame.rawInput.readSmall(4)
+				if len(crc) < 4 {
+					if err == nil {
+						err = io.ErrUnexpectedEOF
+
+					}
+					println("CRC missing?", err)
+					dec.err = err
+				} else {
+					dec.checkCRC = binary.LittleEndian.Uint32(crc)
+					dec.hasCRC = true
+					if debugDecoder {
+						printf("found crc to check: %08x\n", dec.checkCRC)
+					}
+				}
+			}
+			err = dec.err
+			last := dec.Last
+			decodeBlock(dec)
+			if err != nil {
+				break decodeStream
+			}
+			if last {
+				break
+			}
+		}
+	}
+	close(seqDecode)
+	wg.Wait()
+	hist.reset()
+	d.frame.history.b = frameHistCache
+}
+
+func (d *Decoder) setDict(frame *frameDec) (err error) {
+	dict, ok := d.dicts[frame.DictionaryID]
+	if ok {
+		if debugDecoder {
+			println("setting dict", frame.DictionaryID)
+		}
+		frame.history.setDict(dict)
+	} else if frame.DictionaryID != 0 {
+		// A zero or missing dictionary id is ambiguous:
+		// either dictionary zero, or no dictionary. In particular,
+		// zstd --patch-from uses this id for the source file,
+		// so only return an error if the dictionary id is not zero.
+		err = ErrUnknownDictionary
+	}
+	return err
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/decoder_options.go b/vendor/github.com/klauspost/compress/zstd/decoder_options.go
new file mode 100644
index 000000000..774c5f00f
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/decoder_options.go
@@ -0,0 +1,169 @@
+// 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/bits"
+	"runtime"
+)
+
+// DOption is an option for creating a decoder.
+type DOption func(*decoderOptions) error
+
+// options retains accumulated state of multiple options.
+type decoderOptions struct {
+	lowMem          bool
+	concurrent      int
+	maxDecodedSize  uint64
+	maxWindowSize   uint64
+	dicts           []*dict
+	ignoreChecksum  bool
+	limitToCap      bool
+	decodeBufsBelow int
+}
+
+func (o *decoderOptions) setDefault() {
+	*o = decoderOptions{
+		// use less ram: true for now, but may change.
+		lowMem:          true,
+		concurrent:      runtime.GOMAXPROCS(0),
+		maxWindowSize:   MaxWindowSize,
+		decodeBufsBelow: 128 << 10,
+	}
+	if o.concurrent > 4 {
+		o.concurrent = 4
+	}
+	o.maxDecodedSize = 64 << 30
+}
+
+// WithDecoderLowmem will set whether to use a lower amount of memory,
+// but possibly have to allocate more while running.
+func WithDecoderLowmem(b bool) DOption {
+	return func(o *decoderOptions) error { o.lowMem = b; return nil }
+}
+
+// WithDecoderConcurrency sets the number of created decoders.
+// When decoding block with DecodeAll, this will limit the number
+// of possible concurrently running decodes.
+// When decoding streams, this will limit the number of
+// inflight blocks.
+// When decoding streams and setting maximum to 1,
+// no async decoding will be done.
+// When a value of 0 is provided GOMAXPROCS will be used.
+// By default this will be set to 4 or GOMAXPROCS, whatever is lower.
+func WithDecoderConcurrency(n int) DOption {
+	return func(o *decoderOptions) error {
+		if n < 0 {
+			return errors.New("concurrency must be at least 1")
+		}
+		if n == 0 {
+			o.concurrent = runtime.GOMAXPROCS(0)
+		} else {
+			o.concurrent = n
+		}
+		return nil
+	}
+}
+
+// WithDecoderMaxMemory allows to set a maximum decoded size for in-memory
+// non-streaming operations or maximum window size for streaming operations.
+// This can be used to control memory usage of potentially hostile content.
+// Maximum is 1 << 63 bytes. Default is 64GiB.
+func WithDecoderMaxMemory(n uint64) DOption {
+	return func(o *decoderOptions) error {
+		if n == 0 {
+			return errors.New("WithDecoderMaxMemory must be at least 1")
+		}
+		if n > 1<<63 {
+			return errors.New("WithDecoderMaxmemory must be less than 1 << 63")
+		}
+		o.maxDecodedSize = n
+		return nil
+	}
+}
+
+// WithDecoderDicts allows to register one or more dictionaries for the decoder.
+//
+// Each slice in dict must be in the [dictionary format] produced by
+// "zstd --train" from the Zstandard reference implementation.
+//
+// If several dictionaries with the same ID are provided, the last one will be used.
+//
+// [dictionary format]: https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#dictionary-format
+func WithDecoderDicts(dicts ...[]byte) DOption {
+	return func(o *decoderOptions) error {
+		for _, b := range dicts {
+			d, err := loadDict(b)
+			if err != nil {
+				return err
+			}
+			o.dicts = append(o.dicts, d)
+		}
+		return nil
+	}
+}
+
+// WithDecoderDictRaw registers a dictionary that may be used by the decoder.
+// The slice content can be arbitrary data.
+func WithDecoderDictRaw(id uint32, content []byte) DOption {
+	return func(o *decoderOptions) error {
+		if bits.UintSize > 32 && uint(len(content)) > dictMaxLength {
+			return fmt.Errorf("dictionary of size %d > 2GiB too large", len(content))
+		}
+		o.dicts = append(o.dicts, &dict{id: id, content: content, offsets: [3]int{1, 4, 8}})
+		return nil
+	}
+}
+
+// WithDecoderMaxWindow allows to set a maximum window size for decodes.
+// This allows rejecting packets that will cause big memory usage.
+// The Decoder will likely allocate more memory based on the WithDecoderLowmem setting.
+// If WithDecoderMaxMemory is set to a lower value, that will be used.
+// Default is 512MB, Maximum is ~3.75 TB as per zstandard spec.
+func WithDecoderMaxWindow(size uint64) DOption {
+	return func(o *decoderOptions) error {
+		if size < MinWindowSize {
+			return errors.New("WithMaxWindowSize must be at least 1KB, 1024 bytes")
+		}
+		if size > (1<<41)+7*(1<<38) {
+			return errors.New("WithMaxWindowSize must be less than (1<<41) + 7*(1<<38) ~ 3.75TB")
+		}
+		o.maxWindowSize = size
+		return nil
+	}
+}
+
+// WithDecodeAllCapLimit will limit DecodeAll to decoding cap(dst)-len(dst) bytes,
+// or any size set in WithDecoderMaxMemory.
+// This can be used to limit decoding to a specific maximum output size.
+// Disabled by default.
+func WithDecodeAllCapLimit(b bool) DOption {
+	return func(o *decoderOptions) error {
+		o.limitToCap = b
+		return nil
+	}
+}
+
+// WithDecodeBuffersBelow will fully decode readers that have a
+// `Bytes() []byte` and `Len() int` interface similar to bytes.Buffer.
+// This typically uses less allocations but will have the full decompressed object in memory.
+// Note that DecodeAllCapLimit will disable this, as well as giving a size of 0 or less.
+// Default is 128KiB.
+func WithDecodeBuffersBelow(size int) DOption {
+	return func(o *decoderOptions) error {
+		o.decodeBufsBelow = size
+		return nil
+	}
+}
+
+// IgnoreChecksum allows to forcibly ignore checksum checking.
+func IgnoreChecksum(b bool) DOption {
+	return func(o *decoderOptions) error {
+		o.ignoreChecksum = b
+		return nil
+	}
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/dict.go b/vendor/github.com/klauspost/compress/zstd/dict.go
new file mode 100644
index 000000000..b7b83164b
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/dict.go
@@ -0,0 +1,565 @@
+package zstd
+
+import (
+	"bytes"
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+	"math"
+	"sort"
+
+	"github.com/klauspost/compress/huff0"
+)
+
+type dict struct {
+	id uint32
+
+	litEnc              *huff0.Scratch
+	llDec, ofDec, mlDec sequenceDec
+	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
+}
+
+// Load a dictionary as described in
+// https://github.com/facebook/zstd/blob/master/doc/zstd_compression_format.md#dictionary-format
+func loadDict(b []byte) (*dict, error) {
+	// Check static field size.
+	if len(b) <= 8+(3*4) {
+		return nil, io.ErrUnexpectedEOF
+	}
+	d := dict{
+		llDec: sequenceDec{fse: &fseDecoder{}},
+		ofDec: sequenceDec{fse: &fseDecoder{}},
+		mlDec: sequenceDec{fse: &fseDecoder{}},
+	}
+	if string(b[:4]) != dictMagic {
+		return nil, ErrMagicMismatch
+	}
+	d.id = binary.LittleEndian.Uint32(b[4:8])
+	if d.id == 0 {
+		return nil, errors.New("dictionaries cannot have ID 0")
+	}
+
+	// Read literal table
+	var err error
+	d.litEnc, b, err = huff0.ReadTable(b[8:], nil)
+	if err != nil {
+		return nil, fmt.Errorf("loading literal table: %w", err)
+	}
+	d.litEnc.Reuse = huff0.ReusePolicyMust
+
+	br := byteReader{
+		b:   b,
+		off: 0,
+	}
+	readDec := func(i tableIndex, dec *fseDecoder) error {
+		if err := dec.readNCount(&br, uint16(maxTableSymbol[i])); err != nil {
+			return err
+		}
+		if br.overread() {
+			return io.ErrUnexpectedEOF
+		}
+		err = dec.transform(symbolTableX[i])
+		if err != nil {
+			println("Transform table error:", err)
+			return err
+		}
+		if debugDecoder || debugEncoder {
+			println("Read table ok", "symbolLen:", dec.symbolLen)
+		}
+		// Set decoders as predefined so they aren't reused.
+		dec.preDefined = true
+		return nil
+	}
+
+	if err := readDec(tableOffsets, d.ofDec.fse); err != nil {
+		return nil, err
+	}
+	if err := readDec(tableMatchLengths, d.mlDec.fse); err != nil {
+		return nil, err
+	}
+	if err := readDec(tableLiteralLengths, d.llDec.fse); err != nil {
+		return nil, err
+	}
+	if br.remain() < 12 {
+		return nil, io.ErrUnexpectedEOF
+	}
+
+	d.offsets[0] = int(br.Uint32())
+	br.advance(4)
+	d.offsets[1] = int(br.Uint32())
+	br.advance(4)
+	d.offsets[2] = int(br.Uint32())
+	br.advance(4)
+	if d.offsets[0] <= 0 || d.offsets[1] <= 0 || d.offsets[2] <= 0 {
+		return nil, errors.New("invalid offset in dictionary")
+	}
+	d.content = make([]byte, br.remain())
+	copy(d.content, br.unread())
+	if d.offsets[0] > len(d.content) || d.offsets[1] > len(d.content) || d.offsets[2] > len(d.content) {
+		return nil, fmt.Errorf("initial offset bigger than dictionary content size %d, offsets: %v", len(d.content), d.offsets)
+	}
+
+	return &d, nil
+}
+
+// InspectDictionary loads a zstd dictionary and provides functions to inspect the content.
+func InspectDictionary(b []byte) (interface {
+	ID() uint32
+	ContentSize() int
+	Content() []byte
+	Offsets() [3]int
+	LitEncoder() *huff0.Scratch
+}, error) {
+	initPredefined()
+	d, err := loadDict(b)
+	return d, err
+}
+
+type BuildDictOptions struct {
+	// Dictionary ID.
+	ID uint32
+
+	// Content to use to create dictionary tables.
+	Contents [][]byte
+
+	// History to use for all blocks.
+	History []byte
+
+	// Offsets to use.
+	Offsets [3]int
+
+	// CompatV155 will make the dictionary compatible with Zstd v1.5.5 and earlier.
+	// See https://github.com/facebook/zstd/issues/3724
+	CompatV155 bool
+
+	// Use the specified encoder level.
+	// The dictionary will be built using the specified encoder level,
+	// which will reflect speed and make the dictionary tailored for that level.
+	// If not set SpeedBestCompression will be used.
+	Level EncoderLevel
+
+	// DebugOut will write stats and other details here if set.
+	DebugOut io.Writer
+}
+
+func BuildDict(o BuildDictOptions) ([]byte, error) {
+	initPredefined()
+	hist := o.History
+	contents := o.Contents
+	debug := o.DebugOut != nil
+	println := func(args ...interface{}) {
+		if o.DebugOut != nil {
+			fmt.Fprintln(o.DebugOut, args...)
+		}
+	}
+	printf := func(s string, args ...interface{}) {
+		if o.DebugOut != nil {
+			fmt.Fprintf(o.DebugOut, s, args...)
+		}
+	}
+	print := func(args ...interface{}) {
+		if o.DebugOut != nil {
+			fmt.Fprint(o.DebugOut, args...)
+		}
+	}
+
+	if int64(len(hist)) > dictMaxLength {
+		return nil, fmt.Errorf("dictionary of size %d > %d", len(hist), int64(dictMaxLength))
+	}
+	if len(hist) < 8 {
+		return nil, fmt.Errorf("dictionary of size %d < %d", len(hist), 8)
+	}
+	if len(contents) == 0 {
+		return nil, errors.New("no content provided")
+	}
+	d := dict{
+		id:      o.ID,
+		litEnc:  nil,
+		llDec:   sequenceDec{},
+		ofDec:   sequenceDec{},
+		mlDec:   sequenceDec{},
+		offsets: o.Offsets,
+		content: hist,
+	}
+	block := blockEnc{lowMem: false}
+	block.init()
+	enc := encoder(&bestFastEncoder{fastBase: fastBase{maxMatchOff: int32(maxMatchLen), bufferReset: math.MaxInt32 - int32(maxMatchLen*2), lowMem: false}})
+	if o.Level != 0 {
+		eOpts := encoderOptions{
+			level:      o.Level,
+			blockSize:  maxMatchLen,
+			windowSize: maxMatchLen,
+			dict:       &d,
+			lowMem:     false,
+		}
+		enc = eOpts.encoder()
+	} else {
+		o.Level = SpeedBestCompression
+	}
+	var (
+		remain [256]int
+		ll     [256]int
+		ml     [256]int
+		of     [256]int
+	)
+	addValues := func(dst *[256]int, src []byte) {
+		for _, v := range src {
+			dst[v]++
+		}
+	}
+	addHist := func(dst *[256]int, src *[256]uint32) {
+		for i, v := range src {
+			dst[i] += int(v)
+		}
+	}
+	seqs := 0
+	nUsed := 0
+	litTotal := 0
+	newOffsets := make(map[uint32]int, 1000)
+	for _, b := range contents {
+		block.reset(nil)
+		if len(b) < 8 {
+			continue
+		}
+		nUsed++
+		enc.Reset(&d, true)
+		enc.Encode(&block, b)
+		addValues(&remain, block.literals)
+		litTotal += len(block.literals)
+		if len(block.sequences) == 0 {
+			continue
+		}
+		seqs += len(block.sequences)
+		block.genCodes()
+		addHist(&ll, block.coders.llEnc.Histogram())
+		addHist(&ml, block.coders.mlEnc.Histogram())
+		addHist(&of, block.coders.ofEnc.Histogram())
+		for i, seq := range block.sequences {
+			if i > 3 {
+				break
+			}
+			offset := seq.offset
+			if offset == 0 {
+				continue
+			}
+			if int(offset) >= len(o.History) {
+				continue
+			}
+			if offset > 3 {
+				newOffsets[offset-3]++
+			} else {
+				newOffsets[uint32(o.Offsets[offset-1])]++
+			}
+		}
+	}
+	// Find most used offsets.
+	var sortedOffsets []uint32
+	for k := range newOffsets {
+		sortedOffsets = append(sortedOffsets, k)
+	}
+	sort.Slice(sortedOffsets, func(i, j int) bool {
+		a, b := sortedOffsets[i], sortedOffsets[j]
+		if a == b {
+			// Prefer the longer offset
+			return sortedOffsets[i] > sortedOffsets[j]
+		}
+		return newOffsets[sortedOffsets[i]] > newOffsets[sortedOffsets[j]]
+	})
+	if len(sortedOffsets) > 3 {
+		if debug {
+			print("Offsets:")
+			for i, v := range sortedOffsets {
+				if i > 20 {
+					break
+				}
+				printf("[%d: %d],", v, newOffsets[v])
+			}
+			println("")
+		}
+
+		sortedOffsets = sortedOffsets[:3]
+	}
+	for i, v := range sortedOffsets {
+		o.Offsets[i] = int(v)
+	}
+	if debug {
+		println("New repeat offsets", o.Offsets)
+	}
+
+	if nUsed == 0 || seqs == 0 {
+		return nil, fmt.Errorf("%d blocks, %d sequences found", nUsed, seqs)
+	}
+	if debug {
+		println("Sequences:", seqs, "Blocks:", nUsed, "Literals:", litTotal)
+	}
+	if seqs/nUsed < 512 {
+		// Use 512 as minimum.
+		nUsed = seqs / 512
+		if nUsed == 0 {
+			nUsed = 1
+		}
+	}
+	copyHist := func(dst *fseEncoder, src *[256]int) ([]byte, error) {
+		hist := dst.Histogram()
+		var maxSym uint8
+		var maxCount int
+		var fakeLength int
+		for i, v := range src {
+			if v > 0 {
+				v = v / nUsed
+				if v == 0 {
+					v = 1
+				}
+			}
+			if v > maxCount {
+				maxCount = v
+			}
+			if v != 0 {
+				maxSym = uint8(i)
+			}
+			fakeLength += v
+			hist[i] = uint32(v)
+		}
+
+		// Ensure we aren't trying to represent RLE.
+		if maxCount == fakeLength {
+			for i := range hist {
+				if uint8(i) == maxSym {
+					fakeLength++
+					maxSym++
+					hist[i+1] = 1
+					if maxSym > 1 {
+						break
+					}
+				}
+				if hist[0] == 0 {
+					fakeLength++
+					hist[i] = 1
+					if maxSym > 1 {
+						break
+					}
+				}
+			}
+		}
+
+		dst.HistogramFinished(maxSym, maxCount)
+		dst.reUsed = false
+		dst.useRLE = false
+		err := dst.normalizeCount(fakeLength)
+		if err != nil {
+			return nil, err
+		}
+		if debug {
+			println("RAW:", dst.count[:maxSym+1], "NORM:", dst.norm[:maxSym+1], "LEN:", fakeLength)
+		}
+		return dst.writeCount(nil)
+	}
+	if debug {
+		print("Literal lengths: ")
+	}
+	llTable, err := copyHist(block.coders.llEnc, &ll)
+	if err != nil {
+		return nil, err
+	}
+	if debug {
+		print("Match lengths: ")
+	}
+	mlTable, err := copyHist(block.coders.mlEnc, &ml)
+	if err != nil {
+		return nil, err
+	}
+	if debug {
+		print("Offsets: ")
+	}
+	ofTable, err := copyHist(block.coders.ofEnc, &of)
+	if err != nil {
+		return nil, err
+	}
+
+	// Literal table
+	avgSize := litTotal
+	if avgSize > huff0.BlockSizeMax/2 {
+		avgSize = huff0.BlockSizeMax / 2
+	}
+	huffBuff := make([]byte, 0, avgSize)
+	// Target size
+	div := litTotal / avgSize
+	if div < 1 {
+		div = 1
+	}
+	if debug {
+		println("Huffman weights:")
+	}
+	for i, n := range remain[:] {
+		if n > 0 {
+			n = n / div
+			// Allow all entries to be represented.
+			if n == 0 {
+				n = 1
+			}
+			huffBuff = append(huffBuff, bytes.Repeat([]byte{byte(i)}, n)...)
+			if debug {
+				printf("[%d: %d], ", i, n)
+			}
+		}
+	}
+	if o.CompatV155 && remain[255]/div == 0 {
+		huffBuff = append(huffBuff, 255)
+	}
+	scratch := &huff0.Scratch{TableLog: 11}
+	for tries := 0; tries < 255; tries++ {
+		scratch = &huff0.Scratch{TableLog: 11}
+		_, _, err = huff0.Compress1X(huffBuff, scratch)
+		if err == nil {
+			break
+		}
+		if debug {
+			printf("Try %d: Huffman error: %v\n", tries+1, err)
+		}
+		huffBuff = huffBuff[:0]
+		if tries == 250 {
+			if debug {
+				println("Huffman: Bailing out with predefined table")
+			}
+
+			// Bail out.... Just generate something
+			huffBuff = append(huffBuff, bytes.Repeat([]byte{255}, 10000)...)
+			for i := 0; i < 128; i++ {
+				huffBuff = append(huffBuff, byte(i))
+			}
+			continue
+		}
+		if errors.Is(err, huff0.ErrIncompressible) {
+			// Try truncating least common.
+			for i, n := range remain[:] {
+				if n > 0 {
+					n = n / (div * (i + 1))
+					if n > 0 {
+						huffBuff = append(huffBuff, bytes.Repeat([]byte{byte(i)}, n)...)
+					}
+				}
+			}
+			if o.CompatV155 && len(huffBuff) > 0 && huffBuff[len(huffBuff)-1] != 255 {
+				huffBuff = append(huffBuff, 255)
+			}
+			if len(huffBuff) == 0 {
+				huffBuff = append(huffBuff, 0, 255)
+			}
+		}
+		if errors.Is(err, huff0.ErrUseRLE) {
+			for i, n := range remain[:] {
+				n = n / (div * (i + 1))
+				// Allow all entries to be represented.
+				if n == 0 {
+					n = 1
+				}
+				huffBuff = append(huffBuff, bytes.Repeat([]byte{byte(i)}, n)...)
+			}
+		}
+	}
+
+	var out bytes.Buffer
+	out.Write([]byte(dictMagic))
+	out.Write(binary.LittleEndian.AppendUint32(nil, o.ID))
+	out.Write(scratch.OutTable)
+	if debug {
+		println("huff table:", len(scratch.OutTable), "bytes")
+		println("of table:", len(ofTable), "bytes")
+		println("ml table:", len(mlTable), "bytes")
+		println("ll table:", len(llTable), "bytes")
+	}
+	out.Write(ofTable)
+	out.Write(mlTable)
+	out.Write(llTable)
+	out.Write(binary.LittleEndian.AppendUint32(nil, uint32(o.Offsets[0])))
+	out.Write(binary.LittleEndian.AppendUint32(nil, uint32(o.Offsets[1])))
+	out.Write(binary.LittleEndian.AppendUint32(nil, uint32(o.Offsets[2])))
+	out.Write(hist)
+	if debug {
+		_, err := loadDict(out.Bytes())
+		if err != nil {
+			panic(err)
+		}
+		i, err := InspectDictionary(out.Bytes())
+		if err != nil {
+			panic(err)
+		}
+		println("ID:", i.ID())
+		println("Content size:", i.ContentSize())
+		println("Encoder:", i.LitEncoder() != nil)
+		println("Offsets:", i.Offsets())
+		var totalSize int
+		for _, b := range contents {
+			totalSize += len(b)
+		}
+
+		encWith := func(opts ...EOption) int {
+			enc, err := NewWriter(nil, opts...)
+			if err != nil {
+				panic(err)
+			}
+			defer enc.Close()
+			var dst []byte
+			var totalSize int
+			for _, b := range contents {
+				dst = enc.EncodeAll(b, dst[:0])
+				totalSize += len(dst)
+			}
+			return totalSize
+		}
+		plain := encWith(WithEncoderLevel(o.Level))
+		withDict := encWith(WithEncoderLevel(o.Level), WithEncoderDict(out.Bytes()))
+		println("Input size:", totalSize)
+		println("Plain Compressed:", plain)
+		println("Dict Compressed:", withDict)
+		println("Saved:", plain-withDict, (plain-withDict)/len(contents), "bytes per input (rounded down)")
+	}
+	return out.Bytes(), nil
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/enc_base.go b/vendor/github.com/klauspost/compress/zstd/enc_base.go
new file mode 100644
index 000000000..5ca46038a
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/enc_base.go
@@ -0,0 +1,173 @@
+package zstd
+
+import (
+	"fmt"
+	"math/bits"
+
+	"github.com/klauspost/compress/zstd/internal/xxhash"
+)
+
+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("s (%d) < 0", s)
+			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/vendor/github.com/klauspost/compress/zstd/enc_best.go b/vendor/github.com/klauspost/compress/zstd/enc_best.go
new file mode 100644
index 000000000..4613724e9
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/enc_best.go
@@ -0,0 +1,560 @@
+// 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 (
+	"bytes"
+	"fmt"
+
+	"github.com/klauspost/compress"
+)
+
+const (
+	bestLongTableBits = 22                     // Bits used in the long match table
+	bestLongTableSize = 1 << bestLongTableBits // Size of the table
+	bestLongLen       = 8                      // Bytes used for table hash
+
+	// Note: Increasing the short table bits or making the hash shorter
+	// can actually lead to compression degradation since it will 'steal' more from the
+	// long match table and match offsets are quite big.
+	// This greatly depends on the type of input.
+	bestShortTableBits = 18                      // Bits used in the short match table
+	bestShortTableSize = 1 << bestShortTableBits // Size of the table
+	bestShortLen       = 4                       // Bytes used for table hash
+
+)
+
+type match struct {
+	offset int32
+	s      int32
+	length int32
+	rep    int32
+	est    int32
+}
+
+const highScore = maxMatchLen * 8
+
+// estBits will estimate output bits from predefined tables.
+func (m *match) estBits(bitsPerByte int32) {
+	mlc := mlCode(uint32(m.length - zstdMinMatch))
+	var ofc uint8
+	if m.rep < 0 {
+		ofc = ofCode(uint32(m.s-m.offset) + 3)
+	} else {
+		ofc = ofCode(uint32(m.rep) & 3)
+	}
+	// Cost, excluding
+	ofTT, mlTT := fsePredefEnc[tableOffsets].ct.symbolTT[ofc], fsePredefEnc[tableMatchLengths].ct.symbolTT[mlc]
+
+	// Add cost of match encoding...
+	m.est = int32(ofTT.outBits + mlTT.outBits)
+	m.est += int32(ofTT.deltaNbBits>>16 + mlTT.deltaNbBits>>16)
+	// Subtract savings compared to literal encoding...
+	m.est -= (m.length * bitsPerByte) >> 10
+	if m.est > 0 {
+		// Unlikely gain..
+		m.length = 0
+		m.est = highScore
+	}
+}
+
+// bestFastEncoder uses 2 tables, one for short matches (5 bytes) and one for long matches.
+// The long match table contains the previous entry with the same hash,
+// effectively making it a "chain" of length 2.
+// When we find a long match we choose between the two values and select the longest.
+// When we find a short match, after checking the long, we check if we can find a long at n+1
+// and that it is longer (lazy matching).
+type bestFastEncoder struct {
+	fastBase
+	table         [bestShortTableSize]prevEntry
+	longTable     [bestLongTableSize]prevEntry
+	dictTable     []prevEntry
+	dictLongTable []prevEntry
+}
+
+// Encode improves compression...
+func (e *bestFastEncoder) Encode(blk *blockEnc, src []byte) {
+	const (
+		// Input margin is the number of bytes we read (8)
+		// and the maximum we will read ahead (2)
+		inputMargin            = 8 + 4
+		minNonLiteralBlockSize = 16
+	)
+
+	// Protect against e.cur wraparound.
+	for e.cur >= e.bufferReset-int32(len(e.hist)) {
+		if len(e.hist) == 0 {
+			e.table = [bestShortTableSize]prevEntry{}
+			e.longTable = [bestLongTableSize]prevEntry{}
+			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
+			v2 := e.table[i].prev
+			if v < minOff {
+				v = 0
+				v2 = 0
+			} else {
+				v = v - e.cur + e.maxMatchOff
+				if v2 < minOff {
+					v2 = 0
+				} else {
+					v2 = v2 - e.cur + e.maxMatchOff
+				}
+			}
+			e.table[i] = prevEntry{
+				offset: v,
+				prev:   v2,
+			}
+		}
+		for i := range e.longTable[:] {
+			v := e.longTable[i].offset
+			v2 := e.longTable[i].prev
+			if v < minOff {
+				v = 0
+				v2 = 0
+			} else {
+				v = v - e.cur + e.maxMatchOff
+				if v2 < minOff {
+					v2 = 0
+				} else {
+					v2 = v2 - e.cur + e.maxMatchOff
+				}
+			}
+			e.longTable[i] = prevEntry{
+				offset: v,
+				prev:   v2,
+			}
+		}
+		e.cur = e.maxMatchOff
+		break
+	}
+
+	// Add block to history
+	s := e.addBlock(src)
+	blk.size = len(src)
+
+	// Check RLE first
+	if len(src) > zstdMinMatch {
+		ml := matchLen(src[1:], src)
+		if ml == len(src)-1 {
+			blk.literals = append(blk.literals, src[0])
+			blk.sequences = append(blk.sequences, seq{litLen: 1, matchLen: uint32(len(src)-1) - zstdMinMatch, offset: 1 + 3})
+			return
+		}
+	}
+
+	if len(src) < minNonLiteralBlockSize {
+		blk.extraLits = len(src)
+		blk.literals = blk.literals[:len(src)]
+		copy(blk.literals, src)
+		return
+	}
+
+	// Use this to estimate literal cost.
+	// Scaled by 10 bits.
+	bitsPerByte := int32((compress.ShannonEntropyBits(src) * 1024) / len(src))
+	// Huffman can never go < 1 bit/byte
+	if bitsPerByte < 1024 {
+		bitsPerByte = 1024
+	}
+
+	// Override src
+	src = e.hist
+	sLimit := int32(len(src)) - inputMargin
+	const kSearchStrength = 10
+
+	// nextEmit is where in src the next emitLiteral should start from.
+	nextEmit := s
+
+	// Relative offsets
+	offset1 := int32(blk.recentOffsets[0])
+	offset2 := int32(blk.recentOffsets[1])
+	offset3 := int32(blk.recentOffsets[2])
+
+	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 {
+		// We allow the encoder to optionally turn off repeat offsets across blocks
+		canRepeat := len(blk.sequences) > 2
+
+		if debugAsserts && canRepeat && offset1 == 0 {
+			panic("offset0 was 0")
+		}
+
+		const goodEnough = 250
+
+		cv := load6432(src, s)
+
+		nextHashL := hashLen(cv, bestLongTableBits, bestLongLen)
+		nextHashS := hashLen(cv, bestShortTableBits, bestShortLen)
+		candidateL := e.longTable[nextHashL]
+		candidateS := e.table[nextHashS]
+
+		// Set m to a match at offset if it looks like that will improve compression.
+		improve := func(m *match, offset int32, s int32, first uint32, rep int32) {
+			delta := s - offset
+			if delta >= e.maxMatchOff || delta <= 0 || load3232(src, offset) != first {
+				return
+			}
+			// Try to quick reject if we already have a long match.
+			if m.length > 16 {
+				left := len(src) - int(m.s+m.length)
+				// If we are too close to the end, keep as is.
+				if left <= 0 {
+					return
+				}
+				checkLen := m.length - (s - m.s) - 8
+				if left > 2 && checkLen > 4 {
+					// Check 4 bytes, 4 bytes from the end of the current match.
+					a := load3232(src, offset+checkLen)
+					b := load3232(src, s+checkLen)
+					if a != b {
+						return
+					}
+				}
+			}
+			l := 4 + e.matchlen(s+4, offset+4, src)
+			if m.rep <= 0 {
+				// Extend candidate match backwards as far as possible.
+				// Do not extend repeats as we can assume they are optimal
+				// and offsets change if s == nextEmit.
+				tMin := s - e.maxMatchOff
+				if tMin < 0 {
+					tMin = 0
+				}
+				for offset > tMin && s > nextEmit && src[offset-1] == src[s-1] && l < maxMatchLength {
+					s--
+					offset--
+					l++
+				}
+			}
+			if debugAsserts {
+				if offset >= s {
+					panic(fmt.Sprintf("offset: %d - s:%d - rep: %d - cur :%d - max: %d", offset, s, rep, e.cur, e.maxMatchOff))
+				}
+				if !bytes.Equal(src[s:s+l], src[offset:offset+l]) {
+					panic(fmt.Sprintf("second match mismatch: %v != %v, first: %08x", src[s:s+4], src[offset:offset+4], first))
+				}
+			}
+			cand := match{offset: offset, s: s, length: l, rep: rep}
+			cand.estBits(bitsPerByte)
+			if m.est >= highScore || cand.est-m.est+(cand.s-m.s)*bitsPerByte>>10 < 0 {
+				*m = cand
+			}
+		}
+
+		best := match{s: s, est: highScore}
+		improve(&best, candidateL.offset-e.cur, s, uint32(cv), -1)
+		improve(&best, candidateL.prev-e.cur, s, uint32(cv), -1)
+		improve(&best, candidateS.offset-e.cur, s, uint32(cv), -1)
+		improve(&best, candidateS.prev-e.cur, s, uint32(cv), -1)
+
+		if canRepeat && best.length < goodEnough {
+			if s == nextEmit {
+				// Check repeats straight after a match.
+				improve(&best, s-offset2, s, uint32(cv), 1|4)
+				improve(&best, s-offset3, s, uint32(cv), 2|4)
+				if offset1 > 1 {
+					improve(&best, s-(offset1-1), s, uint32(cv), 3|4)
+				}
+			}
+
+			// If either no match or a non-repeat match, check at + 1
+			if best.rep <= 0 {
+				cv32 := uint32(cv >> 8)
+				spp := s + 1
+				improve(&best, spp-offset1, spp, cv32, 1)
+				improve(&best, spp-offset2, spp, cv32, 2)
+				improve(&best, spp-offset3, spp, cv32, 3)
+				if best.rep < 0 {
+					cv32 = uint32(cv >> 24)
+					spp += 2
+					improve(&best, spp-offset1, spp, cv32, 1)
+					improve(&best, spp-offset2, spp, cv32, 2)
+					improve(&best, spp-offset3, spp, cv32, 3)
+				}
+			}
+		}
+		// Load next and check...
+		e.longTable[nextHashL] = prevEntry{offset: s + e.cur, prev: candidateL.offset}
+		e.table[nextHashS] = prevEntry{offset: s + e.cur, prev: candidateS.offset}
+		index0 := s + 1
+
+		// Look far ahead, unless we have a really long match already...
+		if best.length < goodEnough {
+			// No match found, move forward on input, no need to check forward...
+			if best.length < 4 {
+				s += 1 + (s-nextEmit)>>(kSearchStrength-1)
+				if s >= sLimit {
+					break encodeLoop
+				}
+				continue
+			}
+
+			candidateS = e.table[hashLen(cv>>8, bestShortTableBits, bestShortLen)]
+			cv = load6432(src, s+1)
+			cv2 := load6432(src, s+2)
+			candidateL = e.longTable[hashLen(cv, bestLongTableBits, bestLongLen)]
+			candidateL2 := e.longTable[hashLen(cv2, bestLongTableBits, bestLongLen)]
+
+			// Short at s+1
+			improve(&best, candidateS.offset-e.cur, s+1, uint32(cv), -1)
+			// Long at s+1, s+2
+			improve(&best, candidateL.offset-e.cur, s+1, uint32(cv), -1)
+			improve(&best, candidateL.prev-e.cur, s+1, uint32(cv), -1)
+			improve(&best, candidateL2.offset-e.cur, s+2, uint32(cv2), -1)
+			improve(&best, candidateL2.prev-e.cur, s+2, uint32(cv2), -1)
+			if false {
+				// Short at s+3.
+				// Too often worse...
+				improve(&best, e.table[hashLen(cv2>>8, bestShortTableBits, bestShortLen)].offset-e.cur, s+3, uint32(cv2>>8), -1)
+			}
+
+			// Start check at a fixed offset to allow for a few mismatches.
+			// For this compression level 2 yields the best results.
+			// We cannot do this if we have already indexed this position.
+			const skipBeginning = 2
+			if best.s > s-skipBeginning {
+				// See if we can find a better match by checking where the current best ends.
+				// Use that offset to see if we can find a better full match.
+				if sAt := best.s + best.length; sAt < sLimit {
+					nextHashL := hashLen(load6432(src, sAt), bestLongTableBits, bestLongLen)
+					candidateEnd := e.longTable[nextHashL]
+
+					if off := candidateEnd.offset - e.cur - best.length + skipBeginning; off >= 0 {
+						improve(&best, off, best.s+skipBeginning, load3232(src, best.s+skipBeginning), -1)
+						if off := candidateEnd.prev - e.cur - best.length + skipBeginning; off >= 0 {
+							improve(&best, off, best.s+skipBeginning, load3232(src, best.s+skipBeginning), -1)
+						}
+					}
+				}
+			}
+		}
+
+		if debugAsserts {
+			if best.offset >= best.s {
+				panic(fmt.Sprintf("best.offset > s: %d >= %d", best.offset, best.s))
+			}
+			if best.s < nextEmit {
+				panic(fmt.Sprintf("s %d < nextEmit %d", best.s, nextEmit))
+			}
+			if best.offset < s-e.maxMatchOff {
+				panic(fmt.Sprintf("best.offset < s-e.maxMatchOff: %d < %d", best.offset, s-e.maxMatchOff))
+			}
+			if !bytes.Equal(src[best.s:best.s+best.length], src[best.offset:best.offset+best.length]) {
+				panic(fmt.Sprintf("match mismatch: %v != %v", src[best.s:best.s+best.length], src[best.offset:best.offset+best.length]))
+			}
+		}
+
+		// We have a match, we can store the forward value
+		s = best.s
+		if best.rep > 0 {
+			var seq seq
+			seq.matchLen = uint32(best.length - zstdMinMatch)
+			addLiterals(&seq, best.s)
+
+			// Repeat. If bit 4 is set, this is a non-lit repeat.
+			seq.offset = uint32(best.rep & 3)
+			if debugSequences {
+				println("repeat sequence", seq, "next s:", best.s, "off:", best.s-best.offset)
+			}
+			blk.sequences = append(blk.sequences, seq)
+
+			// Index old s + 1 -> s - 1
+			s = best.s + best.length
+			nextEmit = s
+
+			// Index skipped...
+			end := s
+			if s > sLimit+4 {
+				end = sLimit + 4
+			}
+			off := index0 + e.cur
+			for index0 < end {
+				cv0 := load6432(src, index0)
+				h0 := hashLen(cv0, bestLongTableBits, bestLongLen)
+				h1 := hashLen(cv0, bestShortTableBits, bestShortLen)
+				e.longTable[h0] = prevEntry{offset: off, prev: e.longTable[h0].offset}
+				e.table[h1] = prevEntry{offset: off, prev: e.table[h1].offset}
+				off++
+				index0++
+			}
+
+			switch best.rep {
+			case 2, 4 | 1:
+				offset1, offset2 = offset2, offset1
+			case 3, 4 | 2:
+				offset1, offset2, offset3 = offset3, offset1, offset2
+			case 4 | 3:
+				offset1, offset2, offset3 = offset1-1, offset1, offset2
+			}
+			if s >= sLimit {
+				if debugEncoder {
+					println("repeat ended", s, best.length)
+				}
+				break encodeLoop
+			}
+			continue
+		}
+
+		// A 4-byte match has been found. Update recent offsets.
+		// We'll later see if more than 4 bytes.
+		t := best.offset
+		offset1, offset2, offset3 = s-t, offset1, offset2
+
+		if debugAsserts && s <= t {
+			panic(fmt.Sprintf("s (%d) <= t (%d)", s, t))
+		}
+
+		if debugAsserts && int(offset1) > len(src) {
+			panic("invalid offset")
+		}
+
+		// Write our sequence
+		var seq seq
+		l := best.length
+		seq.litLen = uint32(s - nextEmit)
+		seq.matchLen = uint32(l - zstdMinMatch)
+		if seq.litLen > 0 {
+			blk.literals = append(blk.literals, src[nextEmit:s]...)
+		}
+		seq.offset = uint32(s-t) + 3
+		s += l
+		if debugSequences {
+			println("sequence", seq, "next s:", s)
+		}
+		blk.sequences = append(blk.sequences, seq)
+		nextEmit = s
+
+		// Index old s + 1 -> s - 1 or sLimit
+		end := s
+		if s > sLimit-4 {
+			end = sLimit - 4
+		}
+
+		off := index0 + e.cur
+		for index0 < end {
+			cv0 := load6432(src, index0)
+			h0 := hashLen(cv0, bestLongTableBits, bestLongLen)
+			h1 := hashLen(cv0, bestShortTableBits, bestShortLen)
+			e.longTable[h0] = prevEntry{offset: off, prev: e.longTable[h0].offset}
+			e.table[h1] = prevEntry{offset: off, prev: e.table[h1].offset}
+			index0++
+			off++
+		}
+		if s >= sLimit {
+			break encodeLoop
+		}
+	}
+
+	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)
+	blk.recentOffsets[2] = uint32(offset3)
+	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 *bestFastEncoder) EncodeNoHist(blk *blockEnc, src []byte) {
+	e.ensureHist(len(src))
+	e.Encode(blk, src)
+}
+
+// Reset will reset and set a dictionary if not nil
+func (e *bestFastEncoder) 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([]prevEntry, len(e.table))
+		}
+		end := int32(len(d.content)) - 8 + e.maxMatchOff
+		for i := e.maxMatchOff; i < end; i += 4 {
+			const hashLog = bestShortTableBits
+
+			cv := load6432(d.content, i-e.maxMatchOff)
+			nextHash := hashLen(cv, hashLog, bestShortLen)      // 0 -> 4
+			nextHash1 := hashLen(cv>>8, hashLog, bestShortLen)  // 1 -> 5
+			nextHash2 := hashLen(cv>>16, hashLog, bestShortLen) // 2 -> 6
+			nextHash3 := hashLen(cv>>24, hashLog, bestShortLen) // 3 -> 7
+			e.dictTable[nextHash] = prevEntry{
+				prev:   e.dictTable[nextHash].offset,
+				offset: i,
+			}
+			e.dictTable[nextHash1] = prevEntry{
+				prev:   e.dictTable[nextHash1].offset,
+				offset: i + 1,
+			}
+			e.dictTable[nextHash2] = prevEntry{
+				prev:   e.dictTable[nextHash2].offset,
+				offset: i + 2,
+			}
+			e.dictTable[nextHash3] = prevEntry{
+				prev:   e.dictTable[nextHash3].offset,
+				offset: i + 3,
+			}
+		}
+		e.lastDictID = d.id
+	}
+
+	// Init or copy dict table
+	if len(e.dictLongTable) != len(e.longTable) || d.id != e.lastDictID {
+		if len(e.dictLongTable) != len(e.longTable) {
+			e.dictLongTable = make([]prevEntry, len(e.longTable))
+		}
+		if len(d.content) >= 8 {
+			cv := load6432(d.content, 0)
+			h := hashLen(cv, bestLongTableBits, bestLongLen)
+			e.dictLongTable[h] = prevEntry{
+				offset: e.maxMatchOff,
+				prev:   e.dictLongTable[h].offset,
+			}
+
+			end := int32(len(d.content)) - 8 + e.maxMatchOff
+			off := 8 // First to read
+			for i := e.maxMatchOff + 1; i < end; i++ {
+				cv = cv>>8 | (uint64(d.content[off]) << 56)
+				h := hashLen(cv, bestLongTableBits, bestLongLen)
+				e.dictLongTable[h] = prevEntry{
+					offset: i,
+					prev:   e.dictLongTable[h].offset,
+				}
+				off++
+			}
+		}
+		e.lastDictID = d.id
+	}
+	// Reset table to initial state
+	copy(e.longTable[:], e.dictLongTable)
+
+	e.cur = e.maxMatchOff
+	// Reset table to initial state
+	copy(e.table[:], e.dictTable)
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/enc_better.go b/vendor/github.com/klauspost/compress/zstd/enc_better.go
new file mode 100644
index 000000000..a4f5bf91f
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/enc_better.go
@@ -0,0 +1,1252 @@
+// 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 (
+	betterLongTableBits = 19                       // Bits used in the long match table
+	betterLongTableSize = 1 << betterLongTableBits // Size of the table
+	betterLongLen       = 8                        // Bytes used for table hash
+
+	// Note: Increasing the short table bits or making the hash shorter
+	// can actually lead to compression degradation since it will 'steal' more from the
+	// long match table and match offsets are quite big.
+	// This greatly depends on the type of input.
+	betterShortTableBits = 13                        // Bits used in the short match table
+	betterShortTableSize = 1 << betterShortTableBits // Size of the table
+	betterShortLen       = 5                         // Bytes used for table hash
+
+	betterLongTableShardCnt  = 1 << (betterLongTableBits - dictShardBits)    // Number of shards in the table
+	betterLongTableShardSize = betterLongTableSize / betterLongTableShardCnt // Size of an individual shard
+
+	betterShortTableShardCnt  = 1 << (betterShortTableBits - dictShardBits)     // Number of shards in the table
+	betterShortTableShardSize = betterShortTableSize / betterShortTableShardCnt // Size of an individual shard
+)
+
+type prevEntry struct {
+	offset int32
+	prev   int32
+}
+
+// betterFastEncoder uses 2 tables, one for short matches (5 bytes) and one for long matches.
+// The long match table contains the previous entry with the same hash,
+// effectively making it a "chain" of length 2.
+// When we find a long match we choose between the two values and select the longest.
+// When we find a short match, after checking the long, we check if we can find a long at n+1
+// and that it is longer (lazy matching).
+type betterFastEncoder struct {
+	fastBase
+	table     [betterShortTableSize]tableEntry
+	longTable [betterLongTableSize]prevEntry
+}
+
+type betterFastEncoderDict struct {
+	betterFastEncoder
+	dictTable            []tableEntry
+	dictLongTable        []prevEntry
+	shortTableShardDirty [betterShortTableShardCnt]bool
+	longTableShardDirty  [betterLongTableShardCnt]bool
+	allDirty             bool
+}
+
+// Encode improves compression...
+func (e *betterFastEncoder) Encode(blk *blockEnc, src []byte) {
+	const (
+		// Input margin is the number of bytes we read (8)
+		// and the maximum we will read ahead (2)
+		inputMargin            = 8 + 2
+		minNonLiteralBlockSize = 16
+	)
+
+	// Protect against e.cur wraparound.
+	for e.cur >= e.bufferReset-int32(len(e.hist)) {
+		if len(e.hist) == 0 {
+			e.table = [betterShortTableSize]tableEntry{}
+			e.longTable = [betterLongTableSize]prevEntry{}
+			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
+		}
+		for i := range e.longTable[:] {
+			v := e.longTable[i].offset
+			v2 := e.longTable[i].prev
+			if v < minOff {
+				v = 0
+				v2 = 0
+			} else {
+				v = v - e.cur + e.maxMatchOff
+				if v2 < minOff {
+					v2 = 0
+				} else {
+					v2 = v2 - e.cur + e.maxMatchOff
+				}
+			}
+			e.longTable[i] = prevEntry{
+				offset: v,
+				prev:   v2,
+			}
+		}
+		e.cur = e.maxMatchOff
+		break
+	}
+	// Add block to history
+	s := e.addBlock(src)
+	blk.size = len(src)
+
+	// Check RLE first
+	if len(src) > zstdMinMatch {
+		ml := matchLen(src[1:], src)
+		if ml == len(src)-1 {
+			blk.literals = append(blk.literals, src[0])
+			blk.sequences = append(blk.sequences, seq{litLen: 1, matchLen: uint32(len(src)-1) - zstdMinMatch, offset: 1 + 3})
+			return
+		}
+	}
+
+	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 >= 1.
+	const stepSize = 1
+
+	const kSearchStrength = 9
+
+	// 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 {
+		var t int32
+		// We allow the encoder to optionally turn off repeat offsets across blocks
+		canRepeat := len(blk.sequences) > 2
+		var matched, index0 int32
+
+		for {
+			if debugAsserts && canRepeat && offset1 == 0 {
+				panic("offset0 was 0")
+			}
+
+			nextHashL := hashLen(cv, betterLongTableBits, betterLongLen)
+			nextHashS := hashLen(cv, betterShortTableBits, betterShortLen)
+			candidateL := e.longTable[nextHashL]
+			candidateS := e.table[nextHashS]
+
+			const repOff = 1
+			repIndex := s - offset1 + repOff
+			off := s + e.cur
+			e.longTable[nextHashL] = prevEntry{offset: off, prev: candidateL.offset}
+			e.table[nextHashS] = tableEntry{offset: off, val: uint32(cv)}
+			index0 = s + 1
+
+			if canRepeat {
+				if repIndex >= 0 && load3232(src, repIndex) == uint32(cv>>(repOff*8)) {
+					// Consider history as well.
+					var seq seq
+					lenght := 4 + e.matchlen(s+4+repOff, repIndex+4, src)
+
+					seq.matchLen = uint32(lenght - zstdMinMatch)
+
+					// We might be able to match backwards.
+					// Extend as long as we can.
+					start := s + repOff
+					// We end the search early, so we don't risk 0 literals
+					// and have to do special offset treatment.
+					startLimit := nextEmit + 1
+
+					tMin := s - e.maxMatchOff
+					if tMin < 0 {
+						tMin = 0
+					}
+					for repIndex > tMin && start > startLimit && src[repIndex-1] == src[start-1] && seq.matchLen < maxMatchLength-zstdMinMatch-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)
+
+					// Index match start+1 (long) -> s - 1
+					index0 := s + repOff
+					s += lenght + repOff
+
+					nextEmit = s
+					if s >= sLimit {
+						if debugEncoder {
+							println("repeat ended", s, lenght)
+
+						}
+						break encodeLoop
+					}
+					// Index skipped...
+					for index0 < s-1 {
+						cv0 := load6432(src, index0)
+						cv1 := cv0 >> 8
+						h0 := hashLen(cv0, betterLongTableBits, betterLongLen)
+						off := index0 + e.cur
+						e.longTable[h0] = prevEntry{offset: off, prev: e.longTable[h0].offset}
+						e.table[hashLen(cv1, betterShortTableBits, betterShortLen)] = tableEntry{offset: off + 1, val: uint32(cv1)}
+						index0 += 2
+					}
+					cv = load6432(src, s)
+					continue
+				}
+				const repOff2 = 1
+
+				// We deviate from the reference encoder and also check offset 2.
+				// Still slower and not much better, so disabled.
+				// repIndex = s - offset2 + repOff2
+				if false && repIndex >= 0 && load6432(src, repIndex) == load6432(src, s+repOff) {
+					// Consider history as well.
+					var seq seq
+					lenght := 8 + e.matchlen(s+8+repOff2, repIndex+8, src)
+
+					seq.matchLen = uint32(lenght - zstdMinMatch)
+
+					// We might be able to match backwards.
+					// Extend as long as we can.
+					start := s + repOff2
+					// We end the search early, so we don't risk 0 literals
+					// and have to do special offset treatment.
+					startLimit := nextEmit + 1
+
+					tMin := s - e.maxMatchOff
+					if tMin < 0 {
+						tMin = 0
+					}
+					for repIndex > tMin && start > startLimit && src[repIndex-1] == src[start-1] && seq.matchLen < maxMatchLength-zstdMinMatch-1 {
+						repIndex--
+						start--
+						seq.matchLen++
+					}
+					addLiterals(&seq, start)
+
+					// rep 2
+					seq.offset = 2
+					if debugSequences {
+						println("repeat sequence 2", seq, "next s:", s)
+					}
+					blk.sequences = append(blk.sequences, seq)
+
+					s += lenght + repOff2
+					nextEmit = s
+					if s >= sLimit {
+						if debugEncoder {
+							println("repeat ended", s, lenght)
+
+						}
+						break encodeLoop
+					}
+
+					// Index skipped...
+					for index0 < s-1 {
+						cv0 := load6432(src, index0)
+						cv1 := cv0 >> 8
+						h0 := hashLen(cv0, betterLongTableBits, betterLongLen)
+						off := index0 + e.cur
+						e.longTable[h0] = prevEntry{offset: off, prev: e.longTable[h0].offset}
+						e.table[hashLen(cv1, betterShortTableBits, betterShortLen)] = tableEntry{offset: off + 1, val: uint32(cv1)}
+						index0 += 2
+					}
+					cv = load6432(src, s)
+					// Swap offsets
+					offset1, offset2 = offset2, offset1
+					continue
+				}
+			}
+			// Find the offsets of our two matches.
+			coffsetL := candidateL.offset - e.cur
+			coffsetLP := candidateL.prev - e.cur
+
+			// Check if we have a long match.
+			if s-coffsetL < e.maxMatchOff && cv == load6432(src, coffsetL) {
+				// Found a long match, at least 8 bytes.
+				matched = e.matchlen(s+8, coffsetL+8, src) + 8
+				t = coffsetL
+				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 debugMatches {
+					println("long match")
+				}
+
+				if s-coffsetLP < e.maxMatchOff && cv == load6432(src, coffsetLP) {
+					// Found a long match, at least 8 bytes.
+					prevMatch := e.matchlen(s+8, coffsetLP+8, src) + 8
+					if prevMatch > matched {
+						matched = prevMatch
+						t = coffsetLP
+					}
+					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 debugMatches {
+						println("long match")
+					}
+				}
+				break
+			}
+
+			// Check if we have a long match on prev.
+			if s-coffsetLP < e.maxMatchOff && cv == load6432(src, coffsetLP) {
+				// Found a long match, at least 8 bytes.
+				matched = e.matchlen(s+8, coffsetLP+8, src) + 8
+				t = coffsetLP
+				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 debugMatches {
+					println("long match")
+				}
+				break
+			}
+
+			coffsetS := candidateS.offset - e.cur
+
+			// Check if we have a short match.
+			if s-coffsetS < e.maxMatchOff && uint32(cv) == candidateS.val {
+				// found a regular match
+				matched = e.matchlen(s+4, coffsetS+4, src) + 4
+
+				// See if we can find a long match at s+1
+				const checkAt = 1
+				cv := load6432(src, s+checkAt)
+				nextHashL = hashLen(cv, betterLongTableBits, betterLongLen)
+				candidateL = e.longTable[nextHashL]
+				coffsetL = candidateL.offset - e.cur
+
+				// We can store it, since we have at least a 4 byte match.
+				e.longTable[nextHashL] = prevEntry{offset: s + checkAt + e.cur, prev: candidateL.offset}
+				if s-coffsetL < e.maxMatchOff && cv == load6432(src, coffsetL) {
+					// Found a long match, at least 8 bytes.
+					matchedNext := e.matchlen(s+8+checkAt, coffsetL+8, src) + 8
+					if matchedNext > matched {
+						t = coffsetL
+						s += checkAt
+						matched = matchedNext
+						if debugMatches {
+							println("long match (after short)")
+						}
+						break
+					}
+				}
+
+				// Check prev long...
+				coffsetL = candidateL.prev - e.cur
+				if s-coffsetL < e.maxMatchOff && cv == load6432(src, coffsetL) {
+					// Found a long match, at least 8 bytes.
+					matchedNext := e.matchlen(s+8+checkAt, coffsetL+8, src) + 8
+					if matchedNext > matched {
+						t = coffsetL
+						s += checkAt
+						matched = matchedNext
+						if debugMatches {
+							println("prev long match (after short)")
+						}
+						break
+					}
+				}
+				t = coffsetS
+				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")
+				}
+				if debugMatches {
+					println("short match")
+				}
+				break
+			}
+
+			// No match found, move forward in input.
+			s += stepSize + ((s - nextEmit) >> (kSearchStrength - 1))
+			if s >= sLimit {
+				break encodeLoop
+			}
+			cv = load6432(src, s)
+		}
+
+		// Try to find a better match by searching for a long match at the end of the current best match
+		if s+matched < sLimit {
+			// Allow some bytes at the beginning to mismatch.
+			// Sweet spot is around 3 bytes, but depends on input.
+			// The skipped bytes are tested in Extend backwards,
+			// and still picked up as part of the match if they do.
+			const skipBeginning = 3
+
+			nextHashL := hashLen(load6432(src, s+matched), betterLongTableBits, betterLongLen)
+			s2 := s + skipBeginning
+			cv := load3232(src, s2)
+			candidateL := e.longTable[nextHashL]
+			coffsetL := candidateL.offset - e.cur - matched + skipBeginning
+			if coffsetL >= 0 && coffsetL < s2 && s2-coffsetL < e.maxMatchOff && cv == load3232(src, coffsetL) {
+				// Found a long match, at least 4 bytes.
+				matchedNext := e.matchlen(s2+4, coffsetL+4, src) + 4
+				if matchedNext > matched {
+					t = coffsetL
+					s = s2
+					matched = matchedNext
+					if debugMatches {
+						println("long match at end-of-match")
+					}
+				}
+			}
+
+			// Check prev long...
+			if true {
+				coffsetL = candidateL.prev - e.cur - matched + skipBeginning
+				if coffsetL >= 0 && coffsetL < s2 && s2-coffsetL < e.maxMatchOff && cv == load3232(src, coffsetL) {
+					// Found a long match, at least 4 bytes.
+					matchedNext := e.matchlen(s2+4, coffsetL+4, src) + 4
+					if matchedNext > matched {
+						t = coffsetL
+						s = s2
+						matched = matchedNext
+						if debugMatches {
+							println("prev long match at end-of-match")
+						}
+					}
+				}
+			}
+		}
+		// A match has been found. Update recent offsets.
+		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 n-byte match as long as possible.
+		l := matched
+
+		// 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]...)
+		}
+		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
+		}
+
+		// Index match start+1 (long) -> s - 1
+		off := index0 + e.cur
+		for index0 < s-1 {
+			cv0 := load6432(src, index0)
+			cv1 := cv0 >> 8
+			h0 := hashLen(cv0, betterLongTableBits, betterLongLen)
+			e.longTable[h0] = prevEntry{offset: off, prev: e.longTable[h0].offset}
+			e.table[hashLen(cv1, betterShortTableBits, betterShortLen)] = tableEntry{offset: off + 1, val: uint32(cv1)}
+			index0 += 2
+			off += 2
+		}
+
+		cv = load6432(src, s)
+		if !canRepeat {
+			continue
+		}
+
+		// Check offset 2
+		for {
+			o2 := s - offset2
+			if load3232(src, o2) != uint32(cv) {
+				// Do regular search
+				break
+			}
+
+			// Store this, since we have it.
+			nextHashL := hashLen(cv, betterLongTableBits, betterLongLen)
+			nextHashS := hashLen(cv, betterShortTableBits, betterShortLen)
+
+			// 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)
+
+			e.longTable[nextHashL] = prevEntry{offset: s + e.cur, prev: e.longTable[nextHashL].offset}
+			e.table[nextHashS] = 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 {
+				// Finished
+				break encodeLoop
+			}
+			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 *betterFastEncoder) EncodeNoHist(blk *blockEnc, src []byte) {
+	e.ensureHist(len(src))
+	e.Encode(blk, src)
+}
+
+// Encode improves compression...
+func (e *betterFastEncoderDict) Encode(blk *blockEnc, src []byte) {
+	const (
+		// Input margin is the number of bytes we read (8)
+		// and the maximum we will read ahead (2)
+		inputMargin            = 8 + 2
+		minNonLiteralBlockSize = 16
+	)
+
+	// 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{}
+			}
+			for i := range e.longTable[:] {
+				e.longTable[i] = prevEntry{}
+			}
+			e.cur = e.maxMatchOff
+			e.allDirty = true
+			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
+		}
+		for i := range e.longTable[:] {
+			v := e.longTable[i].offset
+			v2 := e.longTable[i].prev
+			if v < minOff {
+				v = 0
+				v2 = 0
+			} else {
+				v = v - e.cur + e.maxMatchOff
+				if v2 < minOff {
+					v2 = 0
+				} else {
+					v2 = v2 - e.cur + e.maxMatchOff
+				}
+			}
+			e.longTable[i] = prevEntry{
+				offset: v,
+				prev:   v2,
+			}
+		}
+		e.allDirty = true
+		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 >= 1.
+	const stepSize = 1
+
+	const kSearchStrength = 9
+
+	// 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 {
+		var t int32
+		// We allow the encoder to optionally turn off repeat offsets across blocks
+		canRepeat := len(blk.sequences) > 2
+		var matched, index0 int32
+
+		for {
+			if debugAsserts && canRepeat && offset1 == 0 {
+				panic("offset0 was 0")
+			}
+
+			nextHashL := hashLen(cv, betterLongTableBits, betterLongLen)
+			nextHashS := hashLen(cv, betterShortTableBits, betterShortLen)
+			candidateL := e.longTable[nextHashL]
+			candidateS := e.table[nextHashS]
+
+			const repOff = 1
+			repIndex := s - offset1 + repOff
+			off := s + e.cur
+			e.longTable[nextHashL] = prevEntry{offset: off, prev: candidateL.offset}
+			e.markLongShardDirty(nextHashL)
+			e.table[nextHashS] = tableEntry{offset: off, val: uint32(cv)}
+			e.markShortShardDirty(nextHashS)
+			index0 = s + 1
+
+			if canRepeat {
+				if repIndex >= 0 && load3232(src, repIndex) == uint32(cv>>(repOff*8)) {
+					// Consider history as well.
+					var seq seq
+					lenght := 4 + e.matchlen(s+4+repOff, repIndex+4, src)
+
+					seq.matchLen = uint32(lenght - zstdMinMatch)
+
+					// We might be able to match backwards.
+					// Extend as long as we can.
+					start := s + repOff
+					// We end the search early, so we don't risk 0 literals
+					// and have to do special offset treatment.
+					startLimit := nextEmit + 1
+
+					tMin := s - e.maxMatchOff
+					if tMin < 0 {
+						tMin = 0
+					}
+					for repIndex > tMin && start > startLimit && src[repIndex-1] == src[start-1] && seq.matchLen < maxMatchLength-zstdMinMatch-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)
+
+					// Index match start+1 (long) -> s - 1
+					s += lenght + repOff
+
+					nextEmit = s
+					if s >= sLimit {
+						if debugEncoder {
+							println("repeat ended", s, lenght)
+
+						}
+						break encodeLoop
+					}
+					// Index skipped...
+					for index0 < s-1 {
+						cv0 := load6432(src, index0)
+						cv1 := cv0 >> 8
+						h0 := hashLen(cv0, betterLongTableBits, betterLongLen)
+						off := index0 + e.cur
+						e.longTable[h0] = prevEntry{offset: off, prev: e.longTable[h0].offset}
+						e.markLongShardDirty(h0)
+						h1 := hashLen(cv1, betterShortTableBits, betterShortLen)
+						e.table[h1] = tableEntry{offset: off + 1, val: uint32(cv1)}
+						e.markShortShardDirty(h1)
+						index0 += 2
+					}
+					cv = load6432(src, s)
+					continue
+				}
+				const repOff2 = 1
+
+				// We deviate from the reference encoder and also check offset 2.
+				// Still slower and not much better, so disabled.
+				// repIndex = s - offset2 + repOff2
+				if false && repIndex >= 0 && load6432(src, repIndex) == load6432(src, s+repOff) {
+					// Consider history as well.
+					var seq seq
+					lenght := 8 + e.matchlen(s+8+repOff2, repIndex+8, src)
+
+					seq.matchLen = uint32(lenght - zstdMinMatch)
+
+					// We might be able to match backwards.
+					// Extend as long as we can.
+					start := s + repOff2
+					// We end the search early, so we don't risk 0 literals
+					// and have to do special offset treatment.
+					startLimit := nextEmit + 1
+
+					tMin := s - e.maxMatchOff
+					if tMin < 0 {
+						tMin = 0
+					}
+					for repIndex > tMin && start > startLimit && src[repIndex-1] == src[start-1] && seq.matchLen < maxMatchLength-zstdMinMatch-1 {
+						repIndex--
+						start--
+						seq.matchLen++
+					}
+					addLiterals(&seq, start)
+
+					// rep 2
+					seq.offset = 2
+					if debugSequences {
+						println("repeat sequence 2", seq, "next s:", s)
+					}
+					blk.sequences = append(blk.sequences, seq)
+
+					s += lenght + repOff2
+					nextEmit = s
+					if s >= sLimit {
+						if debugEncoder {
+							println("repeat ended", s, lenght)
+
+						}
+						break encodeLoop
+					}
+
+					// Index skipped...
+					for index0 < s-1 {
+						cv0 := load6432(src, index0)
+						cv1 := cv0 >> 8
+						h0 := hashLen(cv0, betterLongTableBits, betterLongLen)
+						off := index0 + e.cur
+						e.longTable[h0] = prevEntry{offset: off, prev: e.longTable[h0].offset}
+						e.markLongShardDirty(h0)
+						h1 := hashLen(cv1, betterShortTableBits, betterShortLen)
+						e.table[h1] = tableEntry{offset: off + 1, val: uint32(cv1)}
+						e.markShortShardDirty(h1)
+						index0 += 2
+					}
+					cv = load6432(src, s)
+					// Swap offsets
+					offset1, offset2 = offset2, offset1
+					continue
+				}
+			}
+			// Find the offsets of our two matches.
+			coffsetL := candidateL.offset - e.cur
+			coffsetLP := candidateL.prev - e.cur
+
+			// Check if we have a long match.
+			if s-coffsetL < e.maxMatchOff && cv == load6432(src, coffsetL) {
+				// Found a long match, at least 8 bytes.
+				matched = e.matchlen(s+8, coffsetL+8, src) + 8
+				t = coffsetL
+				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 debugMatches {
+					println("long match")
+				}
+
+				if s-coffsetLP < e.maxMatchOff && cv == load6432(src, coffsetLP) {
+					// Found a long match, at least 8 bytes.
+					prevMatch := e.matchlen(s+8, coffsetLP+8, src) + 8
+					if prevMatch > matched {
+						matched = prevMatch
+						t = coffsetLP
+					}
+					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 debugMatches {
+						println("long match")
+					}
+				}
+				break
+			}
+
+			// Check if we have a long match on prev.
+			if s-coffsetLP < e.maxMatchOff && cv == load6432(src, coffsetLP) {
+				// Found a long match, at least 8 bytes.
+				matched = e.matchlen(s+8, coffsetLP+8, src) + 8
+				t = coffsetLP
+				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 debugMatches {
+					println("long match")
+				}
+				break
+			}
+
+			coffsetS := candidateS.offset - e.cur
+
+			// Check if we have a short match.
+			if s-coffsetS < e.maxMatchOff && uint32(cv) == candidateS.val {
+				// found a regular match
+				matched = e.matchlen(s+4, coffsetS+4, src) + 4
+
+				// See if we can find a long match at s+1
+				const checkAt = 1
+				cv := load6432(src, s+checkAt)
+				nextHashL = hashLen(cv, betterLongTableBits, betterLongLen)
+				candidateL = e.longTable[nextHashL]
+				coffsetL = candidateL.offset - e.cur
+
+				// We can store it, since we have at least a 4 byte match.
+				e.longTable[nextHashL] = prevEntry{offset: s + checkAt + e.cur, prev: candidateL.offset}
+				e.markLongShardDirty(nextHashL)
+				if s-coffsetL < e.maxMatchOff && cv == load6432(src, coffsetL) {
+					// Found a long match, at least 8 bytes.
+					matchedNext := e.matchlen(s+8+checkAt, coffsetL+8, src) + 8
+					if matchedNext > matched {
+						t = coffsetL
+						s += checkAt
+						matched = matchedNext
+						if debugMatches {
+							println("long match (after short)")
+						}
+						break
+					}
+				}
+
+				// Check prev long...
+				coffsetL = candidateL.prev - e.cur
+				if s-coffsetL < e.maxMatchOff && cv == load6432(src, coffsetL) {
+					// Found a long match, at least 8 bytes.
+					matchedNext := e.matchlen(s+8+checkAt, coffsetL+8, src) + 8
+					if matchedNext > matched {
+						t = coffsetL
+						s += checkAt
+						matched = matchedNext
+						if debugMatches {
+							println("prev long match (after short)")
+						}
+						break
+					}
+				}
+				t = coffsetS
+				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")
+				}
+				if debugMatches {
+					println("short match")
+				}
+				break
+			}
+
+			// No match found, move forward in input.
+			s += stepSize + ((s - nextEmit) >> (kSearchStrength - 1))
+			if s >= sLimit {
+				break encodeLoop
+			}
+			cv = load6432(src, s)
+		}
+		// Try to find a better match by searching for a long match at the end of the current best match
+		if s+matched < sLimit {
+			nextHashL := hashLen(load6432(src, s+matched), betterLongTableBits, betterLongLen)
+			cv := load3232(src, s)
+			candidateL := e.longTable[nextHashL]
+			coffsetL := candidateL.offset - e.cur - matched
+			if coffsetL >= 0 && coffsetL < s && s-coffsetL < e.maxMatchOff && cv == load3232(src, coffsetL) {
+				// Found a long match, at least 4 bytes.
+				matchedNext := e.matchlen(s+4, coffsetL+4, src) + 4
+				if matchedNext > matched {
+					t = coffsetL
+					matched = matchedNext
+					if debugMatches {
+						println("long match at end-of-match")
+					}
+				}
+			}
+
+			// Check prev long...
+			if true {
+				coffsetL = candidateL.prev - e.cur - matched
+				if coffsetL >= 0 && coffsetL < s && s-coffsetL < e.maxMatchOff && cv == load3232(src, coffsetL) {
+					// Found a long match, at least 4 bytes.
+					matchedNext := e.matchlen(s+4, coffsetL+4, src) + 4
+					if matchedNext > matched {
+						t = coffsetL
+						matched = matchedNext
+						if debugMatches {
+							println("prev long match at end-of-match")
+						}
+					}
+				}
+			}
+		}
+		// A match has been found. Update recent offsets.
+		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 n-byte match as long as possible.
+		l := matched
+
+		// 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]...)
+		}
+		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
+		}
+
+		// Index match start+1 (long) -> s - 1
+		off := index0 + e.cur
+		for index0 < s-1 {
+			cv0 := load6432(src, index0)
+			cv1 := cv0 >> 8
+			h0 := hashLen(cv0, betterLongTableBits, betterLongLen)
+			e.longTable[h0] = prevEntry{offset: off, prev: e.longTable[h0].offset}
+			e.markLongShardDirty(h0)
+			h1 := hashLen(cv1, betterShortTableBits, betterShortLen)
+			e.table[h1] = tableEntry{offset: off + 1, val: uint32(cv1)}
+			e.markShortShardDirty(h1)
+			index0 += 2
+			off += 2
+		}
+
+		cv = load6432(src, s)
+		if !canRepeat {
+			continue
+		}
+
+		// Check offset 2
+		for {
+			o2 := s - offset2
+			if load3232(src, o2) != uint32(cv) {
+				// Do regular search
+				break
+			}
+
+			// Store this, since we have it.
+			nextHashL := hashLen(cv, betterLongTableBits, betterLongLen)
+			nextHashS := hashLen(cv, betterShortTableBits, betterShortLen)
+
+			// 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)
+
+			e.longTable[nextHashL] = prevEntry{offset: s + e.cur, prev: e.longTable[nextHashL].offset}
+			e.markLongShardDirty(nextHashL)
+			e.table[nextHashS] = tableEntry{offset: s + e.cur, val: uint32(cv)}
+			e.markShortShardDirty(nextHashS)
+			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 {
+				// Finished
+				break encodeLoop
+			}
+			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 *betterFastEncoder) Reset(d *dict, singleBlock bool) {
+	e.resetBase(d, singleBlock)
+	if d != nil {
+		panic("betterFastEncoder: Reset with dict")
+	}
+}
+
+// ResetDict will reset and set a dictionary if not nil
+func (e *betterFastEncoderDict) 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))
+		}
+		end := int32(len(d.content)) - 8 + e.maxMatchOff
+		for i := e.maxMatchOff; i < end; i += 4 {
+			const hashLog = betterShortTableBits
+
+			cv := load6432(d.content, i-e.maxMatchOff)
+			nextHash := hashLen(cv, hashLog, betterShortLen)      // 0 -> 4
+			nextHash1 := hashLen(cv>>8, hashLog, betterShortLen)  // 1 -> 5
+			nextHash2 := hashLen(cv>>16, hashLog, betterShortLen) // 2 -> 6
+			nextHash3 := hashLen(cv>>24, hashLog, betterShortLen) // 3 -> 7
+			e.dictTable[nextHash] = tableEntry{
+				val:    uint32(cv),
+				offset: i,
+			}
+			e.dictTable[nextHash1] = tableEntry{
+				val:    uint32(cv >> 8),
+				offset: i + 1,
+			}
+			e.dictTable[nextHash2] = tableEntry{
+				val:    uint32(cv >> 16),
+				offset: i + 2,
+			}
+			e.dictTable[nextHash3] = tableEntry{
+				val:    uint32(cv >> 24),
+				offset: i + 3,
+			}
+		}
+		e.lastDictID = d.id
+		e.allDirty = true
+	}
+
+	// Init or copy dict table
+	if len(e.dictLongTable) != len(e.longTable) || d.id != e.lastDictID {
+		if len(e.dictLongTable) != len(e.longTable) {
+			e.dictLongTable = make([]prevEntry, len(e.longTable))
+		}
+		if len(d.content) >= 8 {
+			cv := load6432(d.content, 0)
+			h := hashLen(cv, betterLongTableBits, betterLongLen)
+			e.dictLongTable[h] = prevEntry{
+				offset: e.maxMatchOff,
+				prev:   e.dictLongTable[h].offset,
+			}
+
+			end := int32(len(d.content)) - 8 + e.maxMatchOff
+			off := 8 // First to read
+			for i := e.maxMatchOff + 1; i < end; i++ {
+				cv = cv>>8 | (uint64(d.content[off]) << 56)
+				h := hashLen(cv, betterLongTableBits, betterLongLen)
+				e.dictLongTable[h] = prevEntry{
+					offset: i,
+					prev:   e.dictLongTable[h].offset,
+				}
+				off++
+			}
+		}
+		e.lastDictID = d.id
+		e.allDirty = true
+	}
+
+	// Reset table to initial state
+	{
+		dirtyShardCnt := 0
+		if !e.allDirty {
+			for i := range e.shortTableShardDirty {
+				if e.shortTableShardDirty[i] {
+					dirtyShardCnt++
+				}
+			}
+		}
+		const shardCnt = betterShortTableShardCnt
+		const shardSize = betterShortTableShardSize
+		if e.allDirty || dirtyShardCnt > shardCnt*4/6 {
+			copy(e.table[:], e.dictTable)
+			for i := range e.shortTableShardDirty {
+				e.shortTableShardDirty[i] = false
+			}
+		} else {
+			for i := range e.shortTableShardDirty {
+				if !e.shortTableShardDirty[i] {
+					continue
+				}
+
+				copy(e.table[i*shardSize:(i+1)*shardSize], e.dictTable[i*shardSize:(i+1)*shardSize])
+				e.shortTableShardDirty[i] = false
+			}
+		}
+	}
+	{
+		dirtyShardCnt := 0
+		if !e.allDirty {
+			for i := range e.shortTableShardDirty {
+				if e.shortTableShardDirty[i] {
+					dirtyShardCnt++
+				}
+			}
+		}
+		const shardCnt = betterLongTableShardCnt
+		const shardSize = betterLongTableShardSize
+		if e.allDirty || dirtyShardCnt > shardCnt*4/6 {
+			copy(e.longTable[:], e.dictLongTable)
+			for i := range e.longTableShardDirty {
+				e.longTableShardDirty[i] = false
+			}
+		} else {
+			for i := range e.longTableShardDirty {
+				if !e.longTableShardDirty[i] {
+					continue
+				}
+
+				copy(e.longTable[i*shardSize:(i+1)*shardSize], e.dictLongTable[i*shardSize:(i+1)*shardSize])
+				e.longTableShardDirty[i] = false
+			}
+		}
+	}
+	e.cur = e.maxMatchOff
+	e.allDirty = false
+}
+
+func (e *betterFastEncoderDict) markLongShardDirty(entryNum uint32) {
+	e.longTableShardDirty[entryNum/betterLongTableShardSize] = true
+}
+
+func (e *betterFastEncoderDict) markShortShardDirty(entryNum uint32) {
+	e.shortTableShardDirty[entryNum/betterShortTableShardSize] = true
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/enc_dfast.go b/vendor/github.com/klauspost/compress/zstd/enc_dfast.go
new file mode 100644
index 000000000..a154c18f7
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/enc_dfast.go
@@ -0,0 +1,1123 @@
+// 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 (
+	dFastLongTableBits = 17                      // Bits used in the long match table
+	dFastLongTableSize = 1 << dFastLongTableBits // Size of the table
+	dFastLongTableMask = dFastLongTableSize - 1  // Mask for table indices. Redundant, but can eliminate bounds checks.
+	dFastLongLen       = 8                       // Bytes used for table hash
+
+	dLongTableShardCnt  = 1 << (dFastLongTableBits - dictShardBits) // Number of shards in the table
+	dLongTableShardSize = dFastLongTableSize / tableShardCnt        // Size of an individual shard
+
+	dFastShortTableBits = tableBits                // Bits used in the short match table
+	dFastShortTableSize = 1 << dFastShortTableBits // Size of the table
+	dFastShortTableMask = dFastShortTableSize - 1  // Mask for table indices. Redundant, but can eliminate bounds checks.
+	dFastShortLen       = 5                        // Bytes used for table hash
+
+)
+
+type doubleFastEncoder struct {
+	fastEncoder
+	longTable [dFastLongTableSize]tableEntry
+}
+
+type doubleFastEncoderDict struct {
+	fastEncoderDict
+	longTable           [dFastLongTableSize]tableEntry
+	dictLongTable       []tableEntry
+	longTableShardDirty [dLongTableShardCnt]bool
+}
+
+// Encode mimmics functionality in zstd_dfast.c
+func (e *doubleFastEncoder) Encode(blk *blockEnc, src []byte) {
+	const (
+		// Input margin is the number of bytes we read (8)
+		// and the maximum we will read ahead (2)
+		inputMargin            = 8 + 2
+		minNonLiteralBlockSize = 16
+	)
+
+	// Protect against e.cur wraparound.
+	for e.cur >= e.bufferReset-int32(len(e.hist)) {
+		if len(e.hist) == 0 {
+			e.table = [dFastShortTableSize]tableEntry{}
+			e.longTable = [dFastLongTableSize]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
+		}
+		for i := range e.longTable[:] {
+			v := e.longTable[i].offset
+			if v < minOff {
+				v = 0
+			} else {
+				v = v - e.cur + e.maxMatchOff
+			}
+			e.longTable[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 >= 1.
+	const stepSize = 1
+
+	const kSearchStrength = 8
+
+	// 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 {
+		var t int32
+		// We allow the encoder to optionally turn off repeat offsets across blocks
+		canRepeat := len(blk.sequences) > 2
+
+		for {
+			if debugAsserts && canRepeat && offset1 == 0 {
+				panic("offset0 was 0")
+			}
+
+			nextHashL := hashLen(cv, dFastLongTableBits, dFastLongLen)
+			nextHashS := hashLen(cv, dFastShortTableBits, dFastShortLen)
+			candidateL := e.longTable[nextHashL]
+			candidateS := e.table[nextHashS]
+
+			const repOff = 1
+			repIndex := s - offset1 + repOff
+			entry := tableEntry{offset: s + e.cur, val: uint32(cv)}
+			e.longTable[nextHashL] = entry
+			e.table[nextHashS] = entry
+
+			if canRepeat {
+				if repIndex >= 0 && load3232(src, repIndex) == uint32(cv>>(repOff*8)) {
+					// Consider history as well.
+					var seq seq
+					lenght := 4 + e.matchlen(s+4+repOff, repIndex+4, src)
+
+					seq.matchLen = uint32(lenght - zstdMinMatch)
+
+					// We might be able to match backwards.
+					// Extend as long as we can.
+					start := s + repOff
+					// We end the search early, so we don't risk 0 literals
+					// and have to do special offset treatment.
+					startLimit := nextEmit + 1
+
+					tMin := s - e.maxMatchOff
+					if tMin < 0 {
+						tMin = 0
+					}
+					for repIndex > tMin && start > startLimit && src[repIndex-1] == src[start-1] && seq.matchLen < maxMatchLength-zstdMinMatch-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 += lenght + repOff
+					nextEmit = s
+					if s >= sLimit {
+						if debugEncoder {
+							println("repeat ended", s, lenght)
+
+						}
+						break encodeLoop
+					}
+					cv = load6432(src, s)
+					continue
+				}
+			}
+			// Find the offsets of our two matches.
+			coffsetL := s - (candidateL.offset - e.cur)
+			coffsetS := s - (candidateS.offset - e.cur)
+
+			// Check if we have a long match.
+			if coffsetL < e.maxMatchOff && uint32(cv) == candidateL.val {
+				// Found a long match, likely at least 8 bytes.
+				// Reference encoder checks all 8 bytes, we only check 4,
+				// but the likelihood of both the first 4 bytes and the hash matching should be enough.
+				t = candidateL.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 debugMatches {
+					println("long match")
+				}
+				break
+			}
+
+			// Check if we have a short match.
+			if coffsetS < e.maxMatchOff && uint32(cv) == candidateS.val {
+				// found a regular match
+				// See if we can find a long match at s+1
+				const checkAt = 1
+				cv := load6432(src, s+checkAt)
+				nextHashL = hashLen(cv, dFastLongTableBits, dFastLongLen)
+				candidateL = e.longTable[nextHashL]
+				coffsetL = s - (candidateL.offset - e.cur) + checkAt
+
+				// We can store it, since we have at least a 4 byte match.
+				e.longTable[nextHashL] = tableEntry{offset: s + checkAt + e.cur, val: uint32(cv)}
+				if coffsetL < e.maxMatchOff && uint32(cv) == candidateL.val {
+					// Found a long match, likely at least 8 bytes.
+					// Reference encoder checks all 8 bytes, we only check 4,
+					// but the likelihood of both the first 4 bytes and the hash matching should be enough.
+					t = candidateL.offset - e.cur
+					s += checkAt
+					if debugMatches {
+						println("long match (after short)")
+					}
+					break
+				}
+
+				t = candidateS.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("t<0")
+				}
+				if debugMatches {
+					println("short match")
+				}
+				break
+			}
+
+			// No match found, move forward in input.
+			s += stepSize + ((s - nextEmit) >> (kSearchStrength - 1))
+			if s >= sLimit {
+				break encodeLoop
+			}
+			cv = load6432(src, s)
+		}
+
+		// A 4-byte match has been found. Update recent offsets.
+		// 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]...)
+		}
+		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
+		}
+
+		// Index match start+1 (long) and start+2 (short)
+		index0 := s - l + 1
+		// Index match end-2 (long) and end-1 (short)
+		index1 := s - 2
+
+		cv0 := load6432(src, index0)
+		cv1 := load6432(src, index1)
+		te0 := tableEntry{offset: index0 + e.cur, val: uint32(cv0)}
+		te1 := tableEntry{offset: index1 + e.cur, val: uint32(cv1)}
+		e.longTable[hashLen(cv0, dFastLongTableBits, dFastLongLen)] = te0
+		e.longTable[hashLen(cv1, dFastLongTableBits, dFastLongLen)] = te1
+		cv0 >>= 8
+		cv1 >>= 8
+		te0.offset++
+		te1.offset++
+		te0.val = uint32(cv0)
+		te1.val = uint32(cv1)
+		e.table[hashLen(cv0, dFastShortTableBits, dFastShortLen)] = te0
+		e.table[hashLen(cv1, dFastShortTableBits, dFastShortLen)] = te1
+
+		cv = load6432(src, s)
+
+		if !canRepeat {
+			continue
+		}
+
+		// Check offset 2
+		for {
+			o2 := s - offset2
+			if load3232(src, o2) != uint32(cv) {
+				// Do regular search
+				break
+			}
+
+			// Store this, since we have it.
+			nextHashS := hashLen(cv, dFastShortTableBits, dFastShortLen)
+			nextHashL := hashLen(cv, dFastLongTableBits, dFastLongLen)
+
+			// 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)
+
+			entry := tableEntry{offset: s + e.cur, val: uint32(cv)}
+			e.longTable[nextHashL] = entry
+			e.table[nextHashS] = entry
+			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 {
+				// Finished
+				break encodeLoop
+			}
+			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 *doubleFastEncoder) EncodeNoHist(blk *blockEnc, src []byte) {
+	const (
+		// Input margin is the number of bytes we read (8)
+		// and the maximum we will read ahead (2)
+		inputMargin            = 8 + 2
+		minNonLiteralBlockSize = 16
+	)
+
+	// Protect against e.cur wraparound.
+	if e.cur >= e.bufferReset {
+		for i := range e.table[:] {
+			e.table[i] = tableEntry{}
+		}
+		for i := range e.longTable[:] {
+			e.longTable[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
+	}
+
+	// Override src
+	sLimit := int32(len(src)) - inputMargin
+	// stepSize is the number of bytes to skip on every main loop iteration.
+	// It should be >= 1.
+	const stepSize = 1
+
+	const kSearchStrength = 8
+
+	// 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 {
+		var t int32
+		for {
+
+			nextHashL := hashLen(cv, dFastLongTableBits, dFastLongLen)
+			nextHashS := hashLen(cv, dFastShortTableBits, dFastShortLen)
+			candidateL := e.longTable[nextHashL]
+			candidateS := e.table[nextHashS]
+
+			const repOff = 1
+			repIndex := s - offset1 + repOff
+			entry := tableEntry{offset: s + e.cur, val: uint32(cv)}
+			e.longTable[nextHashL] = entry
+			e.table[nextHashS] = entry
+
+			if len(blk.sequences) > 2 {
+				if load3232(src, repIndex) == uint32(cv>>(repOff*8)) {
+					// Consider history as well.
+					var seq seq
+					//length := 4 + e.matchlen(s+4+repOff, repIndex+4, src)
+					length := 4 + int32(matchLen(src[s+4+repOff:], src[repIndex+4:]))
+
+					seq.matchLen = uint32(length - zstdMinMatch)
+
+					// We might be able to match backwards.
+					// Extend as long as we can.
+					start := s + repOff
+					// We end the search early, so we don't risk 0 literals
+					// and have to do special offset treatment.
+					startLimit := nextEmit + 1
+
+					tMin := s - e.maxMatchOff
+					if tMin < 0 {
+						tMin = 0
+					}
+					for repIndex > tMin && 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 + repOff
+					nextEmit = s
+					if s >= sLimit {
+						if debugEncoder {
+							println("repeat ended", s, length)
+
+						}
+						break encodeLoop
+					}
+					cv = load6432(src, s)
+					continue
+				}
+			}
+			// Find the offsets of our two matches.
+			coffsetL := s - (candidateL.offset - e.cur)
+			coffsetS := s - (candidateS.offset - e.cur)
+
+			// Check if we have a long match.
+			if coffsetL < e.maxMatchOff && uint32(cv) == candidateL.val {
+				// Found a long match, likely at least 8 bytes.
+				// Reference encoder checks all 8 bytes, we only check 4,
+				// but the likelihood of both the first 4 bytes and the hash matching should be enough.
+				t = candidateL.offset - e.cur
+				if debugAsserts && s <= t {
+					panic(fmt.Sprintf("s (%d) <= t (%d). cur: %d", s, t, e.cur))
+				}
+				if debugAsserts && s-t > e.maxMatchOff {
+					panic("s - t >e.maxMatchOff")
+				}
+				if debugMatches {
+					println("long match")
+				}
+				break
+			}
+
+			// Check if we have a short match.
+			if coffsetS < e.maxMatchOff && uint32(cv) == candidateS.val {
+				// found a regular match
+				// See if we can find a long match at s+1
+				const checkAt = 1
+				cv := load6432(src, s+checkAt)
+				nextHashL = hashLen(cv, dFastLongTableBits, dFastLongLen)
+				candidateL = e.longTable[nextHashL]
+				coffsetL = s - (candidateL.offset - e.cur) + checkAt
+
+				// We can store it, since we have at least a 4 byte match.
+				e.longTable[nextHashL] = tableEntry{offset: s + checkAt + e.cur, val: uint32(cv)}
+				if coffsetL < e.maxMatchOff && uint32(cv) == candidateL.val {
+					// Found a long match, likely at least 8 bytes.
+					// Reference encoder checks all 8 bytes, we only check 4,
+					// but the likelihood of both the first 4 bytes and the hash matching should be enough.
+					t = candidateL.offset - e.cur
+					s += checkAt
+					if debugMatches {
+						println("long match (after short)")
+					}
+					break
+				}
+
+				t = candidateS.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("t<0")
+				}
+				if debugMatches {
+					println("short match")
+				}
+				break
+			}
+
+			// No match found, move forward in input.
+			s += stepSize + ((s - nextEmit) >> (kSearchStrength - 1))
+			if s >= sLimit {
+				break encodeLoop
+			}
+			cv = load6432(src, s)
+		}
+
+		// A 4-byte match has been found. Update recent offsets.
+		// 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))
+		}
+
+		// Extend the 4-byte match as long as possible.
+		//l := e.matchlen(s+4, t+4, src) + 4
+		l := int32(matchLen(src[s+4:], src[t+4:])) + 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]...)
+		}
+		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
+		}
+
+		// Index match start+1 (long) and start+2 (short)
+		index0 := s - l + 1
+		// Index match end-2 (long) and end-1 (short)
+		index1 := s - 2
+
+		cv0 := load6432(src, index0)
+		cv1 := load6432(src, index1)
+		te0 := tableEntry{offset: index0 + e.cur, val: uint32(cv0)}
+		te1 := tableEntry{offset: index1 + e.cur, val: uint32(cv1)}
+		e.longTable[hashLen(cv0, dFastLongTableBits, dFastLongLen)] = te0
+		e.longTable[hashLen(cv1, dFastLongTableBits, dFastLongLen)] = te1
+		cv0 >>= 8
+		cv1 >>= 8
+		te0.offset++
+		te1.offset++
+		te0.val = uint32(cv0)
+		te1.val = uint32(cv1)
+		e.table[hashLen(cv0, dFastShortTableBits, dFastShortLen)] = te0
+		e.table[hashLen(cv1, dFastShortTableBits, dFastShortLen)] = te1
+
+		cv = load6432(src, s)
+
+		if len(blk.sequences) <= 2 {
+			continue
+		}
+
+		// Check offset 2
+		for {
+			o2 := s - offset2
+			if load3232(src, o2) != uint32(cv) {
+				// Do regular search
+				break
+			}
+
+			// Store this, since we have it.
+			nextHashS := hashLen(cv1>>8, dFastShortTableBits, dFastShortLen)
+			nextHashL := hashLen(cv, dFastLongTableBits, dFastLongLen)
+
+			// 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)
+			l := 4 + int32(matchLen(src[s+4:], src[o2+4:]))
+
+			entry := tableEntry{offset: s + e.cur, val: uint32(cv)}
+			e.longTable[nextHashL] = entry
+			e.table[nextHashS] = entry
+			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 {
+				// Finished
+				break encodeLoop
+			}
+			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 *doubleFastEncoderDict) Encode(blk *blockEnc, src []byte) {
+	const (
+		// Input margin is the number of bytes we read (8)
+		// and the maximum we will read ahead (2)
+		inputMargin            = 8 + 2
+		minNonLiteralBlockSize = 16
+	)
+
+	// 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{}
+			}
+			for i := range e.longTable[:] {
+				e.longTable[i] = tableEntry{}
+			}
+			e.markAllShardsDirty()
+			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
+		}
+		for i := range e.longTable[:] {
+			v := e.longTable[i].offset
+			if v < minOff {
+				v = 0
+			} else {
+				v = v - e.cur + e.maxMatchOff
+			}
+			e.longTable[i].offset = v
+		}
+		e.markAllShardsDirty()
+		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 >= 1.
+	const stepSize = 1
+
+	const kSearchStrength = 8
+
+	// 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 {
+		var t int32
+		// We allow the encoder to optionally turn off repeat offsets across blocks
+		canRepeat := len(blk.sequences) > 2
+
+		for {
+			if debugAsserts && canRepeat && offset1 == 0 {
+				panic("offset0 was 0")
+			}
+
+			nextHashL := hashLen(cv, dFastLongTableBits, dFastLongLen)
+			nextHashS := hashLen(cv, dFastShortTableBits, dFastShortLen)
+			candidateL := e.longTable[nextHashL]
+			candidateS := e.table[nextHashS]
+
+			const repOff = 1
+			repIndex := s - offset1 + repOff
+			entry := tableEntry{offset: s + e.cur, val: uint32(cv)}
+			e.longTable[nextHashL] = entry
+			e.markLongShardDirty(nextHashL)
+			e.table[nextHashS] = entry
+			e.markShardDirty(nextHashS)
+
+			if canRepeat {
+				if repIndex >= 0 && load3232(src, repIndex) == uint32(cv>>(repOff*8)) {
+					// Consider history as well.
+					var seq seq
+					lenght := 4 + e.matchlen(s+4+repOff, repIndex+4, src)
+
+					seq.matchLen = uint32(lenght - zstdMinMatch)
+
+					// We might be able to match backwards.
+					// Extend as long as we can.
+					start := s + repOff
+					// We end the search early, so we don't risk 0 literals
+					// and have to do special offset treatment.
+					startLimit := nextEmit + 1
+
+					tMin := s - e.maxMatchOff
+					if tMin < 0 {
+						tMin = 0
+					}
+					for repIndex > tMin && start > startLimit && src[repIndex-1] == src[start-1] && seq.matchLen < maxMatchLength-zstdMinMatch-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 += lenght + repOff
+					nextEmit = s
+					if s >= sLimit {
+						if debugEncoder {
+							println("repeat ended", s, lenght)
+
+						}
+						break encodeLoop
+					}
+					cv = load6432(src, s)
+					continue
+				}
+			}
+			// Find the offsets of our two matches.
+			coffsetL := s - (candidateL.offset - e.cur)
+			coffsetS := s - (candidateS.offset - e.cur)
+
+			// Check if we have a long match.
+			if coffsetL < e.maxMatchOff && uint32(cv) == candidateL.val {
+				// Found a long match, likely at least 8 bytes.
+				// Reference encoder checks all 8 bytes, we only check 4,
+				// but the likelihood of both the first 4 bytes and the hash matching should be enough.
+				t = candidateL.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 debugMatches {
+					println("long match")
+				}
+				break
+			}
+
+			// Check if we have a short match.
+			if coffsetS < e.maxMatchOff && uint32(cv) == candidateS.val {
+				// found a regular match
+				// See if we can find a long match at s+1
+				const checkAt = 1
+				cv := load6432(src, s+checkAt)
+				nextHashL = hashLen(cv, dFastLongTableBits, dFastLongLen)
+				candidateL = e.longTable[nextHashL]
+				coffsetL = s - (candidateL.offset - e.cur) + checkAt
+
+				// We can store it, since we have at least a 4 byte match.
+				e.longTable[nextHashL] = tableEntry{offset: s + checkAt + e.cur, val: uint32(cv)}
+				e.markLongShardDirty(nextHashL)
+				if coffsetL < e.maxMatchOff && uint32(cv) == candidateL.val {
+					// Found a long match, likely at least 8 bytes.
+					// Reference encoder checks all 8 bytes, we only check 4,
+					// but the likelihood of both the first 4 bytes and the hash matching should be enough.
+					t = candidateL.offset - e.cur
+					s += checkAt
+					if debugMatches {
+						println("long match (after short)")
+					}
+					break
+				}
+
+				t = candidateS.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("t<0")
+				}
+				if debugMatches {
+					println("short match")
+				}
+				break
+			}
+
+			// No match found, move forward in input.
+			s += stepSize + ((s - nextEmit) >> (kSearchStrength - 1))
+			if s >= sLimit {
+				break encodeLoop
+			}
+			cv = load6432(src, s)
+		}
+
+		// A 4-byte match has been found. Update recent offsets.
+		// 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]...)
+		}
+		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
+		}
+
+		// Index match start+1 (long) and start+2 (short)
+		index0 := s - l + 1
+		// Index match end-2 (long) and end-1 (short)
+		index1 := s - 2
+
+		cv0 := load6432(src, index0)
+		cv1 := load6432(src, index1)
+		te0 := tableEntry{offset: index0 + e.cur, val: uint32(cv0)}
+		te1 := tableEntry{offset: index1 + e.cur, val: uint32(cv1)}
+		longHash1 := hashLen(cv0, dFastLongTableBits, dFastLongLen)
+		longHash2 := hashLen(cv1, dFastLongTableBits, dFastLongLen)
+		e.longTable[longHash1] = te0
+		e.longTable[longHash2] = te1
+		e.markLongShardDirty(longHash1)
+		e.markLongShardDirty(longHash2)
+		cv0 >>= 8
+		cv1 >>= 8
+		te0.offset++
+		te1.offset++
+		te0.val = uint32(cv0)
+		te1.val = uint32(cv1)
+		hashVal1 := hashLen(cv0, dFastShortTableBits, dFastShortLen)
+		hashVal2 := hashLen(cv1, dFastShortTableBits, dFastShortLen)
+		e.table[hashVal1] = te0
+		e.markShardDirty(hashVal1)
+		e.table[hashVal2] = te1
+		e.markShardDirty(hashVal2)
+
+		cv = load6432(src, s)
+
+		if !canRepeat {
+			continue
+		}
+
+		// Check offset 2
+		for {
+			o2 := s - offset2
+			if load3232(src, o2) != uint32(cv) {
+				// Do regular search
+				break
+			}
+
+			// Store this, since we have it.
+			nextHashL := hashLen(cv, dFastLongTableBits, dFastLongLen)
+			nextHashS := hashLen(cv, dFastShortTableBits, dFastShortLen)
+
+			// 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)
+
+			entry := tableEntry{offset: s + e.cur, val: uint32(cv)}
+			e.longTable[nextHashL] = entry
+			e.markLongShardDirty(nextHashL)
+			e.table[nextHashS] = entry
+			e.markShardDirty(nextHashS)
+			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 {
+				// Finished
+				break encodeLoop
+			}
+			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)
+	}
+	// If we encoded more than 64K mark all dirty.
+	if len(src) > 64<<10 {
+		e.markAllShardsDirty()
+	}
+}
+
+// ResetDict will reset and set a dictionary if not nil
+func (e *doubleFastEncoder) Reset(d *dict, singleBlock bool) {
+	e.fastEncoder.Reset(d, singleBlock)
+	if d != nil {
+		panic("doubleFastEncoder: Reset with dict not supported")
+	}
+}
+
+// ResetDict will reset and set a dictionary if not nil
+func (e *doubleFastEncoderDict) Reset(d *dict, singleBlock bool) {
+	allDirty := e.allDirty
+	e.fastEncoderDict.Reset(d, singleBlock)
+	if d == nil {
+		return
+	}
+
+	// Init or copy dict table
+	if len(e.dictLongTable) != len(e.longTable) || d.id != e.lastDictID {
+		if len(e.dictLongTable) != len(e.longTable) {
+			e.dictLongTable = make([]tableEntry, len(e.longTable))
+		}
+		if len(d.content) >= 8 {
+			cv := load6432(d.content, 0)
+			e.dictLongTable[hashLen(cv, dFastLongTableBits, dFastLongLen)] = tableEntry{
+				val:    uint32(cv),
+				offset: e.maxMatchOff,
+			}
+			end := int32(len(d.content)) - 8 + e.maxMatchOff
+			for i := e.maxMatchOff + 1; i < end; i++ {
+				cv = cv>>8 | (uint64(d.content[i-e.maxMatchOff+7]) << 56)
+				e.dictLongTable[hashLen(cv, dFastLongTableBits, dFastLongLen)] = tableEntry{
+					val:    uint32(cv),
+					offset: i,
+				}
+			}
+		}
+		e.lastDictID = d.id
+		allDirty = true
+	}
+	// Reset table to initial state
+	e.cur = e.maxMatchOff
+
+	dirtyShardCnt := 0
+	if !allDirty {
+		for i := range e.longTableShardDirty {
+			if e.longTableShardDirty[i] {
+				dirtyShardCnt++
+			}
+		}
+	}
+
+	if allDirty || dirtyShardCnt > dLongTableShardCnt/2 {
+		//copy(e.longTable[:], e.dictLongTable)
+		e.longTable = *(*[dFastLongTableSize]tableEntry)(e.dictLongTable)
+		for i := range e.longTableShardDirty {
+			e.longTableShardDirty[i] = false
+		}
+		return
+	}
+	for i := range e.longTableShardDirty {
+		if !e.longTableShardDirty[i] {
+			continue
+		}
+
+		// copy(e.longTable[i*dLongTableShardSize:(i+1)*dLongTableShardSize], e.dictLongTable[i*dLongTableShardSize:(i+1)*dLongTableShardSize])
+		*(*[dLongTableShardSize]tableEntry)(e.longTable[i*dLongTableShardSize:]) = *(*[dLongTableShardSize]tableEntry)(e.dictLongTable[i*dLongTableShardSize:])
+
+		e.longTableShardDirty[i] = false
+	}
+}
+
+func (e *doubleFastEncoderDict) markLongShardDirty(entryNum uint32) {
+	e.longTableShardDirty[entryNum/dLongTableShardSize] = true
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/enc_fast.go b/vendor/github.com/klauspost/compress/zstd/enc_fast.go
new file mode 100644
index 000000000..f45a3da7d
--- /dev/null
+++ b/vendor/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/vendor/github.com/klauspost/compress/zstd/encoder.go b/vendor/github.com/klauspost/compress/zstd/encoder.go
new file mode 100644
index 000000000..72af7ef0f
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/encoder.go
@@ -0,0 +1,619 @@
+// 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"
+	"fmt"
+	"io"
+	"math"
+	rdebug "runtime/debug"
+	"sync"
+
+	"github.com/klauspost/compress/zstd/internal/xxhash"
+)
+
+// 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
+	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.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)
+	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
+		if final {
+			s.eofWritten = true
+		}
+		// 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 {
+			return err
+		}
+	}
+	s.wg.Wait()
+	s.wWg.Wait()
+	if s.err != 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 {
+		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)
+	}
+	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 {
+	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
+	}
+	e.init.Do(e.initialize)
+	enc := <-e.encoders
+	defer func() {
+		// Release encoder reference to last block.
+		// If a non-single block is needed the encoder will reset again.
+		e.encoders <- enc
+	}()
+	// 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/vendor/github.com/klauspost/compress/zstd/encoder_options.go b/vendor/github.com/klauspost/compress/zstd/encoder_options.go
new file mode 100644
index 000000000..20671dcb9
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/encoder_options.go
@@ -0,0 +1,339 @@
+package zstd
+
+import (
+	"errors"
+	"fmt"
+	"math"
+	"math/bits"
+	"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 {
+	switch o.level {
+	case SpeedFastest:
+		if o.dict != nil {
+			return &fastEncoderDict{fastEncoder: fastEncoder{fastBase: fastBase{maxMatchOff: int32(o.windowSize), bufferReset: math.MaxInt32 - int32(o.windowSize*2), lowMem: o.lowMem}}}
+		}
+		return &fastEncoder{fastBase: fastBase{maxMatchOff: int32(o.windowSize), bufferReset: math.MaxInt32 - int32(o.windowSize*2), lowMem: o.lowMem}}
+
+	case SpeedDefault:
+		if o.dict != nil {
+			return &doubleFastEncoderDict{fastEncoderDict: fastEncoderDict{fastEncoder: fastEncoder{fastBase: fastBase{maxMatchOff: int32(o.windowSize), bufferReset: math.MaxInt32 - int32(o.windowSize*2), lowMem: o.lowMem}}}}
+		}
+		return &doubleFastEncoder{fastEncoder: fastEncoder{fastBase: fastBase{maxMatchOff: int32(o.windowSize), bufferReset: math.MaxInt32 - int32(o.windowSize*2), lowMem: o.lowMem}}}
+	case SpeedBetterCompression:
+		if o.dict != nil {
+			return &betterFastEncoderDict{betterFastEncoder: betterFastEncoder{fastBase: fastBase{maxMatchOff: int32(o.windowSize), bufferReset: math.MaxInt32 - int32(o.windowSize*2), lowMem: o.lowMem}}}
+		}
+		return &betterFastEncoder{fastBase: fastBase{maxMatchOff: int32(o.windowSize), bufferReset: math.MaxInt32 - int32(o.windowSize*2), lowMem: o.lowMem}}
+	case SpeedBestCompression:
+		return &bestFastEncoder{fastBase: fastBase{maxMatchOff: int32(o.windowSize), bufferReset: math.MaxInt32 - int32(o.windowSize*2), lowMem: o.lowMem}}
+	}
+	panic("unknown compression level")
+}
+
+// WithEncoderCRC will add CRC value to output.
+// Output will be 4 bytes larger.
+func WithEncoderCRC(b bool) EOption {
+	return func(o *encoderOptions) error { o.crc = b; return nil }
+}
+
+// WithEncoderConcurrency will set the concurrency,
+// meaning the maximum number of encoders to run concurrently.
+// The value supplied must be at least 1.
+// For streams, setting a value of 1 will disable async compression.
+// By default this will be set to GOMAXPROCS.
+func WithEncoderConcurrency(n int) EOption {
+	return func(o *encoderOptions) error {
+		if n <= 0 {
+			return fmt.Errorf("concurrency must be at least 1")
+		}
+		o.concurrent = n
+		return nil
+	}
+}
+
+// WithWindowSize will set the maximum allowed back-reference distance.
+// The value must be a power of two between MinWindowSize and MaxWindowSize.
+// A larger value will enable better compression but allocate more memory and,
+// for above-default values, take considerably longer.
+// The default value is determined by the compression level and max 8MB.
+func WithWindowSize(n int) EOption {
+	return func(o *encoderOptions) error {
+		switch {
+		case n < MinWindowSize:
+			return fmt.Errorf("window size must be at least %d", MinWindowSize)
+		case n > MaxWindowSize:
+			return fmt.Errorf("window size must be at most %d", MaxWindowSize)
+		case (n & (n - 1)) != 0:
+			return errors.New("window size must be a power of 2")
+		}
+
+		o.windowSize = n
+		o.customWindow = true
+		if o.blockSize > o.windowSize {
+			o.blockSize = o.windowSize
+			o.customBlockSize = true
+		}
+		return nil
+	}
+}
+
+// WithEncoderPadding will add padding to all output so the size will be a multiple of n.
+// This can be used to obfuscate the exact output size or make blocks of a certain size.
+// The contents will be a skippable frame, so it will be invisible by the decoder.
+// n must be > 0 and <= 1GB, 1<<30 bytes.
+// The padded area will be filled with data from crypto/rand.Reader.
+// If `EncodeAll` is used with data already in the destination, the total size will be multiple of this.
+func WithEncoderPadding(n int) EOption {
+	return func(o *encoderOptions) error {
+		if n <= 0 {
+			return fmt.Errorf("padding must be at least 1")
+		}
+		// No need to waste our time.
+		if n == 1 {
+			n = 0
+		}
+		if n > 1<<30 {
+			return fmt.Errorf("padding must less than 1GB (1<<30 bytes) ")
+		}
+		o.pad = n
+		return nil
+	}
+}
+
+// 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 {
+		switch {
+		case l <= speedNotSet || l >= speedLast:
+			return fmt.Errorf("unknown encoder level")
+		}
+		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
+	}
+}
+
+// WithZeroFrames will encode 0 length input as full frames.
+// This can be needed for compatibility with zstandard usage,
+// but is not needed for this package.
+func WithZeroFrames(b bool) EOption {
+	return func(o *encoderOptions) error {
+		o.fullZero = b
+		return nil
+	}
+}
+
+// WithAllLitEntropyCompression will apply entropy compression if no matches are found.
+// Disabling this will skip incompressible data faster, but in cases with no matches but
+// skewed character distribution compression is lost.
+// Default value depends on the compression level selected.
+func WithAllLitEntropyCompression(b bool) EOption {
+	return func(o *encoderOptions) error {
+		o.customALEntropy = true
+		o.allLitEntropy = b
+		return nil
+	}
+}
+
+// WithNoEntropyCompression will always skip entropy compression of literals.
+// This can be useful if content has matches, but unlikely to benefit from entropy
+// compression. Usually the slight speed improvement is not worth enabling this.
+func WithNoEntropyCompression(b bool) EOption {
+	return func(o *encoderOptions) error {
+		o.noEntropy = b
+		return nil
+	}
+}
+
+// WithSingleSegment will set the "single segment" flag when EncodeAll is used.
+// If this flag is set, data must be regenerated within a single continuous memory segment.
+// In this case, Window_Descriptor byte is skipped, but Frame_Content_Size is necessarily present.
+// As a consequence, the decoder must allocate a memory segment of size equal or larger than size of your content.
+// In order to preserve the decoder from unreasonable memory requirements,
+// a decoder is allowed to reject a compressed frame which requests a memory size beyond decoder's authorized range.
+// For broader compatibility, decoders are recommended to support memory sizes of at least 8 MB.
+// This is only a recommendation, each decoder is free to support higher or lower limits, depending on local limitations.
+// If this is not specified, block encodes will automatically choose this based on the input size and the window size.
+// This setting has no effect on streamed encodes.
+func WithSingleSegment(b bool) EOption {
+	return func(o *encoderOptions) error {
+		o.single = &b
+		return nil
+	}
+}
+
+// WithLowerEncoderMem will trade in some memory cases trade less memory usage for
+// slower encoding speed.
+// This will not change the window size which is the primary function for reducing
+// memory usage. See WithWindowSize.
+func WithLowerEncoderMem(b bool) EOption {
+	return func(o *encoderOptions) error {
+		o.lowMem = b
+		return nil
+	}
+}
+
+// WithEncoderDict allows to register a dictionary that will be used for the encode.
+//
+// The slice dict must be in the [dictionary format] produced by
+// "zstd --train" from the Zstandard reference implementation.
+//
+// The encoder *may* choose to use no dictionary instead for certain payloads.
+//
+// [dictionary format]: https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#dictionary-format
+func WithEncoderDict(dict []byte) EOption {
+	return func(o *encoderOptions) error {
+		d, err := loadDict(dict)
+		if err != nil {
+			return err
+		}
+		o.dict = d
+		return nil
+	}
+}
+
+// WithEncoderDictRaw registers a dictionary that may be used by the encoder.
+//
+// The slice content may contain arbitrary data. It will be used as an initial
+// history.
+func WithEncoderDictRaw(id uint32, content []byte) EOption {
+	return func(o *encoderOptions) error {
+		if bits.UintSize > 32 && uint(len(content)) > dictMaxLength {
+			return fmt.Errorf("dictionary of size %d > 2GiB too large", len(content))
+		}
+		o.dict = &dict{id: id, content: content, offsets: [3]int{1, 4, 8}}
+		return nil
+	}
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/framedec.go b/vendor/github.com/klauspost/compress/zstd/framedec.go
new file mode 100644
index 000000000..53e160f7e
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/framedec.go
@@ -0,0 +1,413 @@
+// 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"
+	"encoding/hex"
+	"errors"
+	"io"
+
+	"github.com/klauspost/compress/zstd/internal/xxhash"
+)
+
+type frameDec struct {
+	o   decoderOptions
+	crc *xxhash.Digest
+
+	WindowSize uint64
+
+	// Frame history passed between blocks
+	history history
+
+	rawInput byteBuffer
+
+	// Byte buffer that can be reused for small input blocks.
+	bBuf byteBuf
+
+	FrameContentSize uint64
+
+	DictionaryID  uint32
+	HasCheckSum   bool
+	SingleSegment bool
+}
+
+const (
+	// 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
+)
+
+const (
+	frameMagic          = "\x28\xb5\x2f\xfd"
+	skippableFrameMagic = "\x2a\x4d\x18"
+)
+
+func newFrameDec(o decoderOptions) *frameDec {
+	if o.maxWindowSize > o.maxDecodedSize {
+		o.maxWindowSize = o.maxDecodedSize
+	}
+	d := frameDec{
+		o: o,
+	}
+	return &d
+}
+
+// reset will read the frame header and prepare for block decoding.
+// If nothing can be read from the input, io.EOF will be returned.
+// Any other error indicated that the stream contained data, but
+// there was a problem.
+func (d *frameDec) reset(br byteBuffer) error {
+	d.HasCheckSum = false
+	d.WindowSize = 0
+	var signature [4]byte
+	for {
+		var err error
+		// Check if we can read more...
+		b, err := br.readSmall(1)
+		switch err {
+		case io.EOF, io.ErrUnexpectedEOF:
+			return io.EOF
+		case nil:
+			signature[0] = b[0]
+		default:
+			return err
+		}
+		// Read the rest, don't allow io.ErrUnexpectedEOF
+		b, err = br.readSmall(3)
+		switch err {
+		case io.EOF:
+			return io.EOF
+		case nil:
+			copy(signature[1:], b)
+		default:
+			return err
+		}
+
+		if string(signature[1:4]) != skippableFrameMagic || signature[0]&0xf0 != 0x50 {
+			if debugDecoder {
+				println("Not skippable", hex.EncodeToString(signature[:]), hex.EncodeToString([]byte(skippableFrameMagic)))
+			}
+			// Break if not skippable frame.
+			break
+		}
+		// Read size to skip
+		b, err = br.readSmall(4)
+		if err != nil {
+			if debugDecoder {
+				println("Reading Frame Size", err)
+			}
+			return err
+		}
+		n := uint32(b[0]) | (uint32(b[1]) << 8) | (uint32(b[2]) << 16) | (uint32(b[3]) << 24)
+		println("Skipping frame with", n, "bytes.")
+		err = br.skipN(int64(n))
+		if err != nil {
+			if debugDecoder {
+				println("Reading discarded frame", err)
+			}
+			return err
+		}
+	}
+	if string(signature[:]) != frameMagic {
+		if debugDecoder {
+			println("Got magic numbers: ", signature, "want:", []byte(frameMagic))
+		}
+		return ErrMagicMismatch
+	}
+
+	// Read Frame_Header_Descriptor
+	fhd, err := br.readByte()
+	if err != nil {
+		if debugDecoder {
+			println("Reading Frame_Header_Descriptor", err)
+		}
+		return err
+	}
+	d.SingleSegment = fhd&(1<<5) != 0
+
+	if fhd&(1<<3) != 0 {
+		return errors.New("reserved bit set on frame header")
+	}
+
+	// Read Window_Descriptor
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#window_descriptor
+	d.WindowSize = 0
+	if !d.SingleSegment {
+		wd, err := br.readByte()
+		if err != nil {
+			if debugDecoder {
+				println("Reading Window_Descriptor", err)
+			}
+			return err
+		}
+		printf("raw: %x, mantissa: %d, exponent: %d\n", wd, wd&7, wd>>3)
+		windowLog := 10 + (wd >> 3)
+		windowBase := uint64(1) << windowLog
+		windowAdd := (windowBase / 8) * uint64(wd&0x7)
+		d.WindowSize = windowBase + windowAdd
+	}
+
+	// Read Dictionary_ID
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#dictionary_id
+	d.DictionaryID = 0
+	if size := fhd & 3; size != 0 {
+		if size == 3 {
+			size = 4
+		}
+
+		b, err := br.readSmall(int(size))
+		if err != nil {
+			println("Reading Dictionary_ID", err)
+			return err
+		}
+		var id uint32
+		switch len(b) {
+		case 1:
+			id = uint32(b[0])
+		case 2:
+			id = uint32(b[0]) | (uint32(b[1]) << 8)
+		case 4:
+			id = uint32(b[0]) | (uint32(b[1]) << 8) | (uint32(b[2]) << 16) | (uint32(b[3]) << 24)
+		}
+		if debugDecoder {
+			println("Dict size", size, "ID:", id)
+		}
+		d.DictionaryID = id
+	}
+
+	// Read Frame_Content_Size
+	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#frame_content_size
+	var fcsSize int
+	v := fhd >> 6
+	switch v {
+	case 0:
+		if d.SingleSegment {
+			fcsSize = 1
+		}
+	default:
+		fcsSize = 1 << v
+	}
+	d.FrameContentSize = fcsUnknown
+	if fcsSize > 0 {
+		b, err := br.readSmall(fcsSize)
+		if err != nil {
+			println("Reading Frame content", err)
+			return err
+		}
+		switch len(b) {
+		case 1:
+			d.FrameContentSize = uint64(b[0])
+		case 2:
+			// When FCS_Field_Size is 2, the offset of 256 is added.
+			d.FrameContentSize = uint64(b[0]) | (uint64(b[1]) << 8) + 256
+		case 4:
+			d.FrameContentSize = uint64(b[0]) | (uint64(b[1]) << 8) | (uint64(b[2]) << 16) | (uint64(b[3]) << 24)
+		case 8:
+			d1 := uint32(b[0]) | (uint32(b[1]) << 8) | (uint32(b[2]) << 16) | (uint32(b[3]) << 24)
+			d2 := uint32(b[4]) | (uint32(b[5]) << 8) | (uint32(b[6]) << 16) | (uint32(b[7]) << 24)
+			d.FrameContentSize = uint64(d1) | (uint64(d2) << 32)
+		}
+		if debugDecoder {
+			println("Read FCS:", d.FrameContentSize)
+		}
+	}
+
+	// Move this to shared.
+	d.HasCheckSum = fhd&(1<<2) != 0
+	if d.HasCheckSum {
+		if d.crc == nil {
+			d.crc = xxhash.New()
+		}
+		d.crc.Reset()
+	}
+
+	if d.WindowSize > d.o.maxWindowSize {
+		if debugDecoder {
+			printf("window size %d > max %d\n", d.WindowSize, d.o.maxWindowSize)
+		}
+		return ErrWindowSizeExceeded
+	}
+
+	if d.WindowSize == 0 && d.SingleSegment {
+		// We may not need window in this case.
+		d.WindowSize = d.FrameContentSize
+		if d.WindowSize < MinWindowSize {
+			d.WindowSize = MinWindowSize
+		}
+		if d.WindowSize > d.o.maxDecodedSize {
+			if debugDecoder {
+				printf("window size %d > max %d\n", d.WindowSize, d.o.maxWindowSize)
+			}
+			return ErrDecoderSizeExceeded
+		}
+	}
+
+	// The minimum Window_Size is 1 KB.
+	if d.WindowSize < MinWindowSize {
+		if debugDecoder {
+			println("got window size: ", d.WindowSize)
+		}
+		return ErrWindowSizeTooSmall
+	}
+	d.history.windowSize = int(d.WindowSize)
+	if !d.o.lowMem || d.history.windowSize < maxBlockSize {
+		// Alloc 2x window size if not low-mem, or window size below 2MB.
+		d.history.allocFrameBuffer = d.history.windowSize * 2
+	} else {
+		if d.o.lowMem {
+			// Alloc with 1MB extra.
+			d.history.allocFrameBuffer = d.history.windowSize + maxBlockSize/2
+		} else {
+			// Alloc with 2MB extra.
+			d.history.allocFrameBuffer = d.history.windowSize + maxBlockSize
+		}
+	}
+
+	if debugDecoder {
+		println("Frame: Dict:", d.DictionaryID, "FrameContentSize:", d.FrameContentSize, "singleseg:", d.SingleSegment, "window:", d.WindowSize, "crc:", d.HasCheckSum)
+	}
+
+	// history contains input - maybe we do something
+	d.rawInput = br
+	return nil
+}
+
+// next will start decoding the next block from stream.
+func (d *frameDec) next(block *blockDec) error {
+	if debugDecoder {
+		println("decoding new block")
+	}
+	err := block.reset(d.rawInput, d.WindowSize)
+	if err != nil {
+		println("block error:", err)
+		// Signal the frame decoder we have a problem.
+		block.sendErr(err)
+		return err
+	}
+	return nil
+}
+
+// checkCRC will check the checksum, assuming the frame has one.
+// Will return ErrCRCMismatch if crc check failed, otherwise nil.
+func (d *frameDec) checkCRC() error {
+	// We can overwrite upper tmp now
+	buf, err := d.rawInput.readSmall(4)
+	if err != nil {
+		println("CRC missing?", err)
+		return err
+	}
+
+	want := binary.LittleEndian.Uint32(buf[:4])
+	got := uint32(d.crc.Sum64())
+
+	if got != want {
+		if debugDecoder {
+			printf("CRC check failed: got %08x, want %08x\n", got, want)
+		}
+		return ErrCRCMismatch
+	}
+	if debugDecoder {
+		printf("CRC ok %08x\n", got)
+	}
+	return nil
+}
+
+// consumeCRC skips over the checksum, assuming the frame has one.
+func (d *frameDec) consumeCRC() error {
+	_, err := d.rawInput.readSmall(4)
+	if err != nil {
+		println("CRC missing?", err)
+	}
+	return err
+}
+
+// runDecoder will run the decoder for the remainder of the frame.
+func (d *frameDec) runDecoder(dst []byte, dec *blockDec) ([]byte, error) {
+	saved := d.history.b
+
+	// We use the history for output to avoid copying it.
+	d.history.b = dst
+	d.history.ignoreBuffer = len(dst)
+	// Store input length, so we only check new data.
+	crcStart := len(dst)
+	d.history.decoders.maxSyncLen = 0
+	if d.o.limitToCap {
+		d.history.decoders.maxSyncLen = uint64(cap(dst) - len(dst))
+	}
+	if d.FrameContentSize != fcsUnknown {
+		if !d.o.limitToCap || d.FrameContentSize+uint64(len(dst)) < d.history.decoders.maxSyncLen {
+			d.history.decoders.maxSyncLen = d.FrameContentSize + uint64(len(dst))
+		}
+		if d.history.decoders.maxSyncLen > d.o.maxDecodedSize {
+			if debugDecoder {
+				println("maxSyncLen:", d.history.decoders.maxSyncLen, "> maxDecodedSize:", d.o.maxDecodedSize)
+			}
+			return dst, ErrDecoderSizeExceeded
+		}
+		if debugDecoder {
+			println("maxSyncLen:", d.history.decoders.maxSyncLen)
+		}
+		if !d.o.limitToCap && uint64(cap(dst)) < d.history.decoders.maxSyncLen {
+			// Alloc for output
+			dst2 := make([]byte, len(dst), d.history.decoders.maxSyncLen+compressedBlockOverAlloc)
+			copy(dst2, dst)
+			dst = dst2
+		}
+	}
+	var err error
+	for {
+		err = dec.reset(d.rawInput, d.WindowSize)
+		if err != nil {
+			break
+		}
+		if debugDecoder {
+			println("next block:", dec)
+		}
+		err = dec.decodeBuf(&d.history)
+		if err != nil {
+			break
+		}
+		if uint64(len(d.history.b)-crcStart) > d.o.maxDecodedSize {
+			println("runDecoder: maxDecodedSize exceeded", uint64(len(d.history.b)-crcStart), ">", d.o.maxDecodedSize)
+			err = ErrDecoderSizeExceeded
+			break
+		}
+		if d.o.limitToCap && len(d.history.b) > cap(dst) {
+			println("runDecoder: cap exceeded", uint64(len(d.history.b)), ">", cap(dst))
+			err = ErrDecoderSizeExceeded
+			break
+		}
+		if uint64(len(d.history.b)-crcStart) > d.FrameContentSize {
+			println("runDecoder: FrameContentSize exceeded", uint64(len(d.history.b)-crcStart), ">", d.FrameContentSize)
+			err = ErrFrameSizeExceeded
+			break
+		}
+		if dec.Last {
+			break
+		}
+		if debugDecoder {
+			println("runDecoder: FrameContentSize", uint64(len(d.history.b)-crcStart), "<=", d.FrameContentSize)
+		}
+	}
+	dst = d.history.b
+	if err == nil {
+		if d.FrameContentSize != fcsUnknown && uint64(len(d.history.b)-crcStart) != d.FrameContentSize {
+			err = ErrFrameSizeMismatch
+		} else if d.HasCheckSum {
+			if d.o.ignoreChecksum {
+				err = d.consumeCRC()
+			} else {
+				d.crc.Write(dst[crcStart:])
+				err = d.checkCRC()
+			}
+		}
+	}
+	d.history.b = saved
+	return dst, err
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/frameenc.go b/vendor/github.com/klauspost/compress/zstd/frameenc.go
new file mode 100644
index 000000000..667ca0679
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/frameenc.go
@@ -0,0 +1,137 @@
+// 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
+
+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/vendor/github.com/klauspost/compress/zstd/fse_decoder.go b/vendor/github.com/klauspost/compress/zstd/fse_decoder.go
new file mode 100644
index 000000000..2f8860a72
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/fse_decoder.go
@@ -0,0 +1,307 @@
+// 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<<s.actualTableLog {
+		return fmt.Errorf("corruption detected (total %d != %d)", gotTotal, 1<<s.actualTableLog)
+	}
+	b.advance((bitCount + 7) >> 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
+}
+
+type fseState struct {
+	dt    []decSymbol
+	state decSymbol
+}
+
+// Initialize and decodeAsync first state and symbol.
+func (s *fseState) init(br *bitReader, tableLog uint8, dt []decSymbol) {
+	s.dt = dt
+	br.fill()
+	s.state = dt[br.getBits(tableLog)]
+}
+
+// final returns the current state symbol without decoding the next.
+func (s decSymbol) final() (int, uint8) {
+	return s.baselineInt(), s.addBits()
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/fse_decoder_amd64.go b/vendor/github.com/klauspost/compress/zstd/fse_decoder_amd64.go
new file mode 100644
index 000000000..d04a829b0
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/fse_decoder_amd64.go
@@ -0,0 +1,65 @@
+//go:build amd64 && !appengine && !noasm && gc
+// +build amd64,!appengine,!noasm,gc
+
+package zstd
+
+import (
+	"fmt"
+)
+
+type buildDtableAsmContext struct {
+	// inputs
+	stateTable *uint16
+	norm       *int16
+	dt         *uint64
+
+	// outputs --- set by the procedure in the case of error;
+	// for interpretation please see the error handling part below
+	errParam1 uint64
+	errParam2 uint64
+}
+
+// buildDtable_asm is an x86 assembly implementation of fseDecoder.buildDtable.
+// Function returns non-zero exit code on error.
+//
+//go:noescape
+func buildDtable_asm(s *fseDecoder, ctx *buildDtableAsmContext) int
+
+// please keep in sync with _generate/gen_fse.go
+const (
+	errorCorruptedNormalizedCounter = 1
+	errorNewStateTooBig             = 2
+	errorNewStateNoBits             = 3
+)
+
+// buildDtable will build the decoding table.
+func (s *fseDecoder) buildDtable() error {
+	ctx := buildDtableAsmContext{
+		stateTable: &s.stateTable[0],
+		norm:       &s.norm[0],
+		dt:         (*uint64)(&s.dt[0]),
+	}
+	code := buildDtable_asm(s, &ctx)
+
+	if code != 0 {
+		switch code {
+		case errorCorruptedNormalizedCounter:
+			position := ctx.errParam1
+			return fmt.Errorf("corrupted input (position=%d, expected 0)", position)
+
+		case errorNewStateTooBig:
+			newState := decSymbol(ctx.errParam1)
+			size := ctx.errParam2
+			return fmt.Errorf("newState (%d) outside table size (%d)", newState, size)
+
+		case errorNewStateNoBits:
+			newState := decSymbol(ctx.errParam1)
+			oldState := decSymbol(ctx.errParam2)
+			return fmt.Errorf("newState (%d) == oldState (%d) and no bits", newState, oldState)
+
+		default:
+			return fmt.Errorf("buildDtable_asm returned unhandled nonzero code = %d", code)
+		}
+	}
+	return nil
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/fse_decoder_amd64.s b/vendor/github.com/klauspost/compress/zstd/fse_decoder_amd64.s
new file mode 100644
index 000000000..bcde39869
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/fse_decoder_amd64.s
@@ -0,0 +1,126 @@
+// Code generated by command: go run gen_fse.go -out ../fse_decoder_amd64.s -pkg=zstd. DO NOT EDIT.
+
+//go:build !appengine && !noasm && gc && !noasm
+
+// func buildDtable_asm(s *fseDecoder, ctx *buildDtableAsmContext) int
+TEXT ·buildDtable_asm(SB), $0-24
+	MOVQ ctx+8(FP), CX
+	MOVQ s+0(FP), DI
+
+	// Load values
+	MOVBQZX 4098(DI), DX
+	XORQ    AX, AX
+	BTSQ    DX, AX
+	MOVQ    (CX), BX
+	MOVQ    16(CX), SI
+	LEAQ    -1(AX), R8
+	MOVQ    8(CX), CX
+	MOVWQZX 4096(DI), DI
+
+	// End load values
+	// Init, lay down lowprob symbols
+	XORQ R9, R9
+	JMP  init_main_loop_condition
+
+init_main_loop:
+	MOVWQSX (CX)(R9*2), R10
+	CMPW    R10, $-1
+	JNE     do_not_update_high_threshold
+	MOVB    R9, 1(SI)(R8*8)
+	DECQ    R8
+	MOVQ    $0x0000000000000001, R10
+
+do_not_update_high_threshold:
+	MOVW R10, (BX)(R9*2)
+	INCQ R9
+
+init_main_loop_condition:
+	CMPQ R9, DI
+	JL   init_main_loop
+
+	// Spread symbols
+	// Calculate table step
+	MOVQ AX, R9
+	SHRQ $0x01, R9
+	MOVQ AX, R10
+	SHRQ $0x03, R10
+	LEAQ 3(R9)(R10*1), R9
+
+	// Fill add bits values
+	LEAQ -1(AX), R10
+	XORQ R11, R11
+	XORQ R12, R12
+	JMP  spread_main_loop_condition
+
+spread_main_loop:
+	XORQ    R13, R13
+	MOVWQSX (CX)(R12*2), R14
+	JMP     spread_inner_loop_condition
+
+spread_inner_loop:
+	MOVB R12, 1(SI)(R11*8)
+
+adjust_position:
+	ADDQ R9, R11
+	ANDQ R10, R11
+	CMPQ R11, R8
+	JG   adjust_position
+	INCQ R13
+
+spread_inner_loop_condition:
+	CMPQ R13, R14
+	JL   spread_inner_loop
+	INCQ R12
+
+spread_main_loop_condition:
+	CMPQ  R12, DI
+	JL    spread_main_loop
+	TESTQ R11, R11
+	JZ    spread_check_ok
+	MOVQ  ctx+8(FP), AX
+	MOVQ  R11, 24(AX)
+	MOVQ  $+1, ret+16(FP)
+	RET
+
+spread_check_ok:
+	// Build Decoding table
+	XORQ DI, DI
+
+build_table_main_table:
+	MOVBQZX 1(SI)(DI*8), CX
+	MOVWQZX (BX)(CX*2), R8
+	LEAQ    1(R8), R9
+	MOVW    R9, (BX)(CX*2)
+	MOVQ    R8, R9
+	BSRQ    R9, R9
+	MOVQ    DX, CX
+	SUBQ    R9, CX
+	SHLQ    CL, R8
+	SUBQ    AX, R8
+	MOVB    CL, (SI)(DI*8)
+	MOVW    R8, 2(SI)(DI*8)
+	CMPQ    R8, AX
+	JLE     build_table_check1_ok
+	MOVQ    ctx+8(FP), CX
+	MOVQ    R8, 24(CX)
+	MOVQ    AX, 32(CX)
+	MOVQ    $+2, ret+16(FP)
+	RET
+
+build_table_check1_ok:
+	TESTB CL, CL
+	JNZ   build_table_check2_ok
+	CMPW  R8, DI
+	JNE   build_table_check2_ok
+	MOVQ  ctx+8(FP), AX
+	MOVQ  R8, 24(AX)
+	MOVQ  DI, 32(AX)
+	MOVQ  $+3, ret+16(FP)
+	RET
+
+build_table_check2_ok:
+	INCQ DI
+	CMPQ DI, AX
+	JL   build_table_main_table
+	MOVQ $+0, ret+16(FP)
+	RET
diff --git a/vendor/github.com/klauspost/compress/zstd/fse_decoder_generic.go b/vendor/github.com/klauspost/compress/zstd/fse_decoder_generic.go
new file mode 100644
index 000000000..8adfebb02
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/fse_decoder_generic.go
@@ -0,0 +1,73 @@
+//go:build !amd64 || appengine || !gc || noasm
+// +build !amd64 appengine !gc noasm
+
+package zstd
+
+import (
+	"errors"
+	"fmt"
+)
+
+// 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
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/fse_encoder.go b/vendor/github.com/klauspost/compress/zstd/fse_encoder.go
new file mode 100644
index 000000000..ab26326a8
--- /dev/null
+++ b/vendor/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<<s.actualTableLog)
+	}
+	for i, v := range s.count[s.symbolLen:] {
+		if v != 0 {
+			return fmt.Errorf("warning: Found symbol out of range, %d after cut", i)
+		}
+	}
+	return nil
+}
+
+// writeCount will write the normalized histogram count to header.
+// This is read back by readNCount.
+func (s *fseEncoder) writeCount(out []byte) ([]byte, error) {
+	if s.useRLE {
+		return append(out, s.rleVal), nil
+	}
+	if s.preDefined || s.reUsed {
+		// Never write predefined.
+		return out, nil
+	}
+
+	var (
+		tableLog  = s.actualTableLog
+		tableSize = 1 << tableLog
+		previous0 bool
+		charnum   uint16
+
+		// maximum header size plus 2 extra bytes for final output if bitCount == 0.
+		maxHeaderSize = ((int(s.symbolLen) * int(tableLog)) >> 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/vendor/github.com/klauspost/compress/zstd/fse_predefined.go b/vendor/github.com/klauspost/compress/zstd/fse_predefined.go
new file mode 100644
index 000000000..474cb77d2
--- /dev/null
+++ b/vendor/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/vendor/github.com/klauspost/compress/zstd/hash.go b/vendor/github.com/klauspost/compress/zstd/hash.go
new file mode 100644
index 000000000..5d73c21eb
--- /dev/null
+++ b/vendor/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/vendor/github.com/klauspost/compress/zstd/history.go b/vendor/github.com/klauspost/compress/zstd/history.go
new file mode 100644
index 000000000..09164856d
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/history.go
@@ -0,0 +1,116 @@
+// 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 (
+	"github.com/klauspost/compress/huff0"
+)
+
+// history contains the information transferred between blocks.
+type history struct {
+	// Literal decompression
+	huffTree *huff0.Scratch
+
+	// Sequence decompression
+	decoders      sequenceDecs
+	recentOffsets [3]int
+
+	// History buffer...
+	b []byte
+
+	// ignoreBuffer is meant to ignore a number of bytes
+	// when checking for matches in history
+	ignoreBuffer int
+
+	windowSize       int
+	allocFrameBuffer int // needed?
+	error            bool
+	dict             *dict
+}
+
+// reset will reset the history to initial state of a frame.
+// The history must already have been initialized to the desired size.
+func (h *history) reset() {
+	h.b = h.b[:0]
+	h.ignoreBuffer = 0
+	h.error = false
+	h.recentOffsets = [3]int{1, 4, 8}
+	h.decoders.freeDecoders()
+	h.decoders = sequenceDecs{br: h.decoders.br}
+	h.freeHuffDecoder()
+	h.huffTree = nil
+	h.dict = nil
+	//printf("history created: %+v (l: %d, c: %d)", *h, len(h.b), cap(h.b))
+}
+
+func (h *history) freeHuffDecoder() {
+	if h.huffTree != nil {
+		if h.dict == nil || h.dict.litEnc != h.huffTree {
+			huffDecoderPool.Put(h.huffTree)
+			h.huffTree = nil
+		}
+	}
+}
+
+func (h *history) setDict(dict *dict) {
+	if dict == nil {
+		return
+	}
+	h.dict = dict
+	h.decoders.litLengths = dict.llDec
+	h.decoders.offsets = dict.ofDec
+	h.decoders.matchLengths = dict.mlDec
+	h.decoders.dict = dict.content
+	h.recentOffsets = dict.offsets
+	h.huffTree = dict.litEnc
+}
+
+// append bytes to history.
+// This function will make sure there is space for it,
+// if the buffer has been allocated with enough extra space.
+func (h *history) append(b []byte) {
+	if len(b) >= h.windowSize {
+		// Discard all history by simply overwriting
+		h.b = h.b[:h.windowSize]
+		copy(h.b, b[len(b)-h.windowSize:])
+		return
+	}
+
+	// If there is space, append it.
+	if len(b) < cap(h.b)-len(h.b) {
+		h.b = append(h.b, b...)
+		return
+	}
+
+	// Move data down so we only have window size left.
+	// We know we have less than window size in b at this point.
+	discard := len(b) + len(h.b) - h.windowSize
+	copy(h.b, h.b[discard:])
+	h.b = h.b[:h.windowSize]
+	copy(h.b[h.windowSize-len(b):], b)
+}
+
+// ensureBlock will ensure there is space for at least one block...
+func (h *history) ensureBlock() {
+	if cap(h.b) < h.allocFrameBuffer {
+		h.b = make([]byte, 0, h.allocFrameBuffer)
+		return
+	}
+
+	avail := cap(h.b) - len(h.b)
+	if avail >= h.windowSize || avail > maxCompressedBlockSize {
+		return
+	}
+	// Move data down so we only have window size left.
+	// We know we have less than window size in b at this point.
+	discard := len(h.b) - h.windowSize
+	copy(h.b, h.b[discard:])
+	h.b = h.b[:h.windowSize]
+}
+
+// append bytes to history without ever discarding anything.
+func (h *history) appendKeep(b []byte) {
+	h.b = append(h.b, b...)
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/internal/xxhash/LICENSE.txt b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/LICENSE.txt
new file mode 100644
index 000000000..24b53065f
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/LICENSE.txt
@@ -0,0 +1,22 @@
+Copyright (c) 2016 Caleb Spare
+
+MIT License
+
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of this software and associated documentation files (the
+"Software"), to deal in the Software without restriction, including
+without limitation the rights to use, copy, modify, merge, publish,
+distribute, sublicense, and/or sell copies of the Software, and to
+permit persons to whom the Software is furnished to do so, subject to
+the following conditions:
+
+The above copyright notice and this permission notice shall be
+included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
+LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
+OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
diff --git a/vendor/github.com/klauspost/compress/zstd/internal/xxhash/README.md b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/README.md
new file mode 100644
index 000000000..777290d44
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/README.md
@@ -0,0 +1,71 @@
+# xxhash
+
+VENDORED: Go to [github.com/cespare/xxhash](https://github.com/cespare/xxhash) for original package.
+
+xxhash is a Go implementation of the 64-bit [xxHash] algorithm, XXH64. This is a
+high-quality hashing algorithm that is much faster than anything in the Go
+standard library.
+
+This package provides a straightforward API:
+
+```
+func Sum64(b []byte) uint64
+func Sum64String(s string) uint64
+type Digest struct{ ... }
+    func New() *Digest
+```
+
+The `Digest` type implements hash.Hash64. Its key methods are:
+
+```
+func (*Digest) Write([]byte) (int, error)
+func (*Digest) WriteString(string) (int, error)
+func (*Digest) Sum64() uint64
+```
+
+The package is written with optimized pure Go and also contains even faster
+assembly implementations for amd64 and arm64. If desired, the `purego` build tag
+opts into using the Go code even on those architectures.
+
+[xxHash]: http://cyan4973.github.io/xxHash/
+
+## Compatibility
+
+This package is in a module and the latest code is in version 2 of the module.
+You need a version of Go with at least "minimal module compatibility" to use
+github.com/cespare/xxhash/v2:
+
+* 1.9.7+ for Go 1.9
+* 1.10.3+ for Go 1.10
+* Go 1.11 or later
+
+I recommend using the latest release of Go.
+
+## Benchmarks
+
+Here are some quick benchmarks comparing the pure-Go and assembly
+implementations of Sum64.
+
+| input size | purego    | asm       |
+| ---------- | --------- | --------- |
+| 4 B        |  1.3 GB/s |  1.2 GB/s |
+| 16 B       |  2.9 GB/s |  3.5 GB/s |
+| 100 B      |  6.9 GB/s |  8.1 GB/s |
+| 4 KB       | 11.7 GB/s | 16.7 GB/s |
+| 10 MB      | 12.0 GB/s | 17.3 GB/s |
+
+These numbers were generated on Ubuntu 20.04 with an Intel Xeon Platinum 8252C
+CPU using the following commands under Go 1.19.2:
+
+```
+benchstat <(go test -tags purego -benchtime 500ms -count 15 -bench 'Sum64$')
+benchstat <(go test -benchtime 500ms -count 15 -bench 'Sum64$')
+```
+
+## Projects using this package
+
+- [InfluxDB](https://github.com/influxdata/influxdb)
+- [Prometheus](https://github.com/prometheus/prometheus)
+- [VictoriaMetrics](https://github.com/VictoriaMetrics/VictoriaMetrics)
+- [FreeCache](https://github.com/coocood/freecache)
+- [FastCache](https://github.com/VictoriaMetrics/fastcache)
diff --git a/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash.go b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash.go
new file mode 100644
index 000000000..fc40c8200
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash.go
@@ -0,0 +1,230 @@
+// Package xxhash implements the 64-bit variant of xxHash (XXH64) as described
+// at http://cyan4973.github.io/xxHash/.
+// THIS IS VENDORED: Go to github.com/cespare/xxhash for original package.
+
+package xxhash
+
+import (
+	"encoding/binary"
+	"errors"
+	"math/bits"
+)
+
+const (
+	prime1 uint64 = 11400714785074694791
+	prime2 uint64 = 14029467366897019727
+	prime3 uint64 = 1609587929392839161
+	prime4 uint64 = 9650029242287828579
+	prime5 uint64 = 2870177450012600261
+)
+
+// Store the primes in an array as well.
+//
+// The consts are used when possible in Go code to avoid MOVs but we need a
+// contiguous array of the assembly code.
+var primes = [...]uint64{prime1, prime2, prime3, prime4, prime5}
+
+// Digest implements hash.Hash64.
+type Digest struct {
+	v1    uint64
+	v2    uint64
+	v3    uint64
+	v4    uint64
+	total uint64
+	mem   [32]byte
+	n     int // how much of mem is used
+}
+
+// New creates a new Digest that computes the 64-bit xxHash algorithm.
+func New() *Digest {
+	var d Digest
+	d.Reset()
+	return &d
+}
+
+// Reset clears the Digest's state so that it can be reused.
+func (d *Digest) Reset() {
+	d.v1 = primes[0] + prime2
+	d.v2 = prime2
+	d.v3 = 0
+	d.v4 = -primes[0]
+	d.total = 0
+	d.n = 0
+}
+
+// Size always returns 8 bytes.
+func (d *Digest) Size() int { return 8 }
+
+// BlockSize always returns 32 bytes.
+func (d *Digest) BlockSize() int { return 32 }
+
+// Write adds more data to d. It always returns len(b), nil.
+func (d *Digest) Write(b []byte) (n int, err error) {
+	n = len(b)
+	d.total += uint64(n)
+
+	memleft := d.mem[d.n&(len(d.mem)-1):]
+
+	if d.n+n < 32 {
+		// This new data doesn't even fill the current block.
+		copy(memleft, b)
+		d.n += n
+		return
+	}
+
+	if d.n > 0 {
+		// Finish off the partial block.
+		c := copy(memleft, b)
+		d.v1 = round(d.v1, u64(d.mem[0:8]))
+		d.v2 = round(d.v2, u64(d.mem[8:16]))
+		d.v3 = round(d.v3, u64(d.mem[16:24]))
+		d.v4 = round(d.v4, u64(d.mem[24:32]))
+		b = b[c:]
+		d.n = 0
+	}
+
+	if len(b) >= 32 {
+		// One or more full blocks left.
+		nw := writeBlocks(d, b)
+		b = b[nw:]
+	}
+
+	// Store any remaining partial block.
+	copy(d.mem[:], b)
+	d.n = len(b)
+
+	return
+}
+
+// Sum appends the current hash to b and returns the resulting slice.
+func (d *Digest) Sum(b []byte) []byte {
+	s := d.Sum64()
+	return append(
+		b,
+		byte(s>>56),
+		byte(s>>48),
+		byte(s>>40),
+		byte(s>>32),
+		byte(s>>24),
+		byte(s>>16),
+		byte(s>>8),
+		byte(s),
+	)
+}
+
+// Sum64 returns the current hash.
+func (d *Digest) Sum64() uint64 {
+	var h uint64
+
+	if d.total >= 32 {
+		v1, v2, v3, v4 := d.v1, d.v2, d.v3, d.v4
+		h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
+		h = mergeRound(h, v1)
+		h = mergeRound(h, v2)
+		h = mergeRound(h, v3)
+		h = mergeRound(h, v4)
+	} else {
+		h = d.v3 + prime5
+	}
+
+	h += d.total
+
+	b := d.mem[:d.n&(len(d.mem)-1)]
+	for ; len(b) >= 8; b = b[8:] {
+		k1 := round(0, u64(b[:8]))
+		h ^= k1
+		h = rol27(h)*prime1 + prime4
+	}
+	if len(b) >= 4 {
+		h ^= uint64(u32(b[:4])) * prime1
+		h = rol23(h)*prime2 + prime3
+		b = b[4:]
+	}
+	for ; len(b) > 0; b = b[1:] {
+		h ^= uint64(b[0]) * prime5
+		h = rol11(h) * prime1
+	}
+
+	h ^= h >> 33
+	h *= prime2
+	h ^= h >> 29
+	h *= prime3
+	h ^= h >> 32
+
+	return h
+}
+
+const (
+	magic         = "xxh\x06"
+	marshaledSize = len(magic) + 8*5 + 32
+)
+
+// MarshalBinary implements the encoding.BinaryMarshaler interface.
+func (d *Digest) MarshalBinary() ([]byte, error) {
+	b := make([]byte, 0, marshaledSize)
+	b = append(b, magic...)
+	b = appendUint64(b, d.v1)
+	b = appendUint64(b, d.v2)
+	b = appendUint64(b, d.v3)
+	b = appendUint64(b, d.v4)
+	b = appendUint64(b, d.total)
+	b = append(b, d.mem[:d.n]...)
+	b = b[:len(b)+len(d.mem)-d.n]
+	return b, nil
+}
+
+// UnmarshalBinary implements the encoding.BinaryUnmarshaler interface.
+func (d *Digest) UnmarshalBinary(b []byte) error {
+	if len(b) < len(magic) || string(b[:len(magic)]) != magic {
+		return errors.New("xxhash: invalid hash state identifier")
+	}
+	if len(b) != marshaledSize {
+		return errors.New("xxhash: invalid hash state size")
+	}
+	b = b[len(magic):]
+	b, d.v1 = consumeUint64(b)
+	b, d.v2 = consumeUint64(b)
+	b, d.v3 = consumeUint64(b)
+	b, d.v4 = consumeUint64(b)
+	b, d.total = consumeUint64(b)
+	copy(d.mem[:], b)
+	d.n = int(d.total % uint64(len(d.mem)))
+	return nil
+}
+
+func appendUint64(b []byte, x uint64) []byte {
+	var a [8]byte
+	binary.LittleEndian.PutUint64(a[:], x)
+	return append(b, a[:]...)
+}
+
+func consumeUint64(b []byte) ([]byte, uint64) {
+	x := u64(b)
+	return b[8:], x
+}
+
+func u64(b []byte) uint64 { return binary.LittleEndian.Uint64(b) }
+func u32(b []byte) uint32 { return binary.LittleEndian.Uint32(b) }
+
+func round(acc, input uint64) uint64 {
+	acc += input * prime2
+	acc = rol31(acc)
+	acc *= prime1
+	return acc
+}
+
+func mergeRound(acc, val uint64) uint64 {
+	val = round(0, val)
+	acc ^= val
+	acc = acc*prime1 + prime4
+	return acc
+}
+
+func rol1(x uint64) uint64  { return bits.RotateLeft64(x, 1) }
+func rol7(x uint64) uint64  { return bits.RotateLeft64(x, 7) }
+func rol11(x uint64) uint64 { return bits.RotateLeft64(x, 11) }
+func rol12(x uint64) uint64 { return bits.RotateLeft64(x, 12) }
+func rol18(x uint64) uint64 { return bits.RotateLeft64(x, 18) }
+func rol23(x uint64) uint64 { return bits.RotateLeft64(x, 23) }
+func rol27(x uint64) uint64 { return bits.RotateLeft64(x, 27) }
+func rol31(x uint64) uint64 { return bits.RotateLeft64(x, 31) }
diff --git a/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_amd64.s b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_amd64.s
new file mode 100644
index 000000000..ddb63aa91
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_amd64.s
@@ -0,0 +1,210 @@
+//go:build !appengine && gc && !purego && !noasm
+// +build !appengine
+// +build gc
+// +build !purego
+// +build !noasm
+
+#include "textflag.h"
+
+// Registers:
+#define h      AX
+#define d      AX
+#define p      SI // pointer to advance through b
+#define n      DX
+#define end    BX // loop end
+#define v1     R8
+#define v2     R9
+#define v3     R10
+#define v4     R11
+#define x      R12
+#define prime1 R13
+#define prime2 R14
+#define prime4 DI
+
+#define round(acc, x) \
+	IMULQ prime2, x   \
+	ADDQ  x, acc      \
+	ROLQ  $31, acc    \
+	IMULQ prime1, acc
+
+// round0 performs the operation x = round(0, x).
+#define round0(x) \
+	IMULQ prime2, x \
+	ROLQ  $31, x    \
+	IMULQ prime1, x
+
+// mergeRound applies a merge round on the two registers acc and x.
+// It assumes that prime1, prime2, and prime4 have been loaded.
+#define mergeRound(acc, x) \
+	round0(x)         \
+	XORQ  x, acc      \
+	IMULQ prime1, acc \
+	ADDQ  prime4, acc
+
+// blockLoop processes as many 32-byte blocks as possible,
+// updating v1, v2, v3, and v4. It assumes that there is at least one block
+// to process.
+#define blockLoop() \
+loop:  \
+	MOVQ +0(p), x  \
+	round(v1, x)   \
+	MOVQ +8(p), x  \
+	round(v2, x)   \
+	MOVQ +16(p), x \
+	round(v3, x)   \
+	MOVQ +24(p), x \
+	round(v4, x)   \
+	ADDQ $32, p    \
+	CMPQ p, end    \
+	JLE  loop
+
+// func Sum64(b []byte) uint64
+TEXT ·Sum64(SB), NOSPLIT|NOFRAME, $0-32
+	// Load fixed primes.
+	MOVQ ·primes+0(SB), prime1
+	MOVQ ·primes+8(SB), prime2
+	MOVQ ·primes+24(SB), prime4
+
+	// Load slice.
+	MOVQ b_base+0(FP), p
+	MOVQ b_len+8(FP), n
+	LEAQ (p)(n*1), end
+
+	// The first loop limit will be len(b)-32.
+	SUBQ $32, end
+
+	// Check whether we have at least one block.
+	CMPQ n, $32
+	JLT  noBlocks
+
+	// Set up initial state (v1, v2, v3, v4).
+	MOVQ prime1, v1
+	ADDQ prime2, v1
+	MOVQ prime2, v2
+	XORQ v3, v3
+	XORQ v4, v4
+	SUBQ prime1, v4
+
+	blockLoop()
+
+	MOVQ v1, h
+	ROLQ $1, h
+	MOVQ v2, x
+	ROLQ $7, x
+	ADDQ x, h
+	MOVQ v3, x
+	ROLQ $12, x
+	ADDQ x, h
+	MOVQ v4, x
+	ROLQ $18, x
+	ADDQ x, h
+
+	mergeRound(h, v1)
+	mergeRound(h, v2)
+	mergeRound(h, v3)
+	mergeRound(h, v4)
+
+	JMP afterBlocks
+
+noBlocks:
+	MOVQ ·primes+32(SB), h
+
+afterBlocks:
+	ADDQ n, h
+
+	ADDQ $24, end
+	CMPQ p, end
+	JG   try4
+
+loop8:
+	MOVQ  (p), x
+	ADDQ  $8, p
+	round0(x)
+	XORQ  x, h
+	ROLQ  $27, h
+	IMULQ prime1, h
+	ADDQ  prime4, h
+
+	CMPQ p, end
+	JLE  loop8
+
+try4:
+	ADDQ $4, end
+	CMPQ p, end
+	JG   try1
+
+	MOVL  (p), x
+	ADDQ  $4, p
+	IMULQ prime1, x
+	XORQ  x, h
+
+	ROLQ  $23, h
+	IMULQ prime2, h
+	ADDQ  ·primes+16(SB), h
+
+try1:
+	ADDQ $4, end
+	CMPQ p, end
+	JGE  finalize
+
+loop1:
+	MOVBQZX (p), x
+	ADDQ    $1, p
+	IMULQ   ·primes+32(SB), x
+	XORQ    x, h
+	ROLQ    $11, h
+	IMULQ   prime1, h
+
+	CMPQ p, end
+	JL   loop1
+
+finalize:
+	MOVQ  h, x
+	SHRQ  $33, x
+	XORQ  x, h
+	IMULQ prime2, h
+	MOVQ  h, x
+	SHRQ  $29, x
+	XORQ  x, h
+	IMULQ ·primes+16(SB), h
+	MOVQ  h, x
+	SHRQ  $32, x
+	XORQ  x, h
+
+	MOVQ h, ret+24(FP)
+	RET
+
+// func writeBlocks(d *Digest, b []byte) int
+TEXT ·writeBlocks(SB), NOSPLIT|NOFRAME, $0-40
+	// Load fixed primes needed for round.
+	MOVQ ·primes+0(SB), prime1
+	MOVQ ·primes+8(SB), prime2
+
+	// Load slice.
+	MOVQ b_base+8(FP), p
+	MOVQ b_len+16(FP), n
+	LEAQ (p)(n*1), end
+	SUBQ $32, end
+
+	// Load vN from d.
+	MOVQ s+0(FP), d
+	MOVQ 0(d), v1
+	MOVQ 8(d), v2
+	MOVQ 16(d), v3
+	MOVQ 24(d), v4
+
+	// We don't need to check the loop condition here; this function is
+	// always called with at least one block of data to process.
+	blockLoop()
+
+	// Copy vN back to d.
+	MOVQ v1, 0(d)
+	MOVQ v2, 8(d)
+	MOVQ v3, 16(d)
+	MOVQ v4, 24(d)
+
+	// The number of bytes written is p minus the old base pointer.
+	SUBQ b_base+8(FP), p
+	MOVQ p, ret+32(FP)
+
+	RET
diff --git a/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_arm64.s b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_arm64.s
new file mode 100644
index 000000000..ae7d4d329
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_arm64.s
@@ -0,0 +1,184 @@
+//go:build !appengine && gc && !purego && !noasm
+// +build !appengine
+// +build gc
+// +build !purego
+// +build !noasm
+
+#include "textflag.h"
+
+// Registers:
+#define digest	R1
+#define h	R2 // return value
+#define p	R3 // input pointer
+#define n	R4 // input length
+#define nblocks	R5 // n / 32
+#define prime1	R7
+#define prime2	R8
+#define prime3	R9
+#define prime4	R10
+#define prime5	R11
+#define v1	R12
+#define v2	R13
+#define v3	R14
+#define v4	R15
+#define x1	R20
+#define x2	R21
+#define x3	R22
+#define x4	R23
+
+#define round(acc, x) \
+	MADD prime2, acc, x, acc \
+	ROR  $64-31, acc         \
+	MUL  prime1, acc
+
+// round0 performs the operation x = round(0, x).
+#define round0(x) \
+	MUL prime2, x \
+	ROR $64-31, x \
+	MUL prime1, x
+
+#define mergeRound(acc, x) \
+	round0(x)                     \
+	EOR  x, acc                   \
+	MADD acc, prime4, prime1, acc
+
+// blockLoop processes as many 32-byte blocks as possible,
+// updating v1, v2, v3, and v4. It assumes that n >= 32.
+#define blockLoop() \
+	LSR     $5, n, nblocks  \
+	PCALIGN $16             \
+	loop:                   \
+	LDP.P   16(p), (x1, x2) \
+	LDP.P   16(p), (x3, x4) \
+	round(v1, x1)           \
+	round(v2, x2)           \
+	round(v3, x3)           \
+	round(v4, x4)           \
+	SUB     $1, nblocks     \
+	CBNZ    nblocks, loop
+
+// func Sum64(b []byte) uint64
+TEXT ·Sum64(SB), NOSPLIT|NOFRAME, $0-32
+	LDP b_base+0(FP), (p, n)
+
+	LDP  ·primes+0(SB), (prime1, prime2)
+	LDP  ·primes+16(SB), (prime3, prime4)
+	MOVD ·primes+32(SB), prime5
+
+	CMP  $32, n
+	CSEL LT, prime5, ZR, h // if n < 32 { h = prime5 } else { h = 0 }
+	BLT  afterLoop
+
+	ADD  prime1, prime2, v1
+	MOVD prime2, v2
+	MOVD $0, v3
+	NEG  prime1, v4
+
+	blockLoop()
+
+	ROR $64-1, v1, x1
+	ROR $64-7, v2, x2
+	ADD x1, x2
+	ROR $64-12, v3, x3
+	ROR $64-18, v4, x4
+	ADD x3, x4
+	ADD x2, x4, h
+
+	mergeRound(h, v1)
+	mergeRound(h, v2)
+	mergeRound(h, v3)
+	mergeRound(h, v4)
+
+afterLoop:
+	ADD n, h
+
+	TBZ   $4, n, try8
+	LDP.P 16(p), (x1, x2)
+
+	round0(x1)
+
+	// NOTE: here and below, sequencing the EOR after the ROR (using a
+	// rotated register) is worth a small but measurable speedup for small
+	// inputs.
+	ROR  $64-27, h
+	EOR  x1 @> 64-27, h, h
+	MADD h, prime4, prime1, h
+
+	round0(x2)
+	ROR  $64-27, h
+	EOR  x2 @> 64-27, h, h
+	MADD h, prime4, prime1, h
+
+try8:
+	TBZ    $3, n, try4
+	MOVD.P 8(p), x1
+
+	round0(x1)
+	ROR  $64-27, h
+	EOR  x1 @> 64-27, h, h
+	MADD h, prime4, prime1, h
+
+try4:
+	TBZ     $2, n, try2
+	MOVWU.P 4(p), x2
+
+	MUL  prime1, x2
+	ROR  $64-23, h
+	EOR  x2 @> 64-23, h, h
+	MADD h, prime3, prime2, h
+
+try2:
+	TBZ     $1, n, try1
+	MOVHU.P 2(p), x3
+	AND     $255, x3, x1
+	LSR     $8, x3, x2
+
+	MUL prime5, x1
+	ROR $64-11, h
+	EOR x1 @> 64-11, h, h
+	MUL prime1, h
+
+	MUL prime5, x2
+	ROR $64-11, h
+	EOR x2 @> 64-11, h, h
+	MUL prime1, h
+
+try1:
+	TBZ   $0, n, finalize
+	MOVBU (p), x4
+
+	MUL prime5, x4
+	ROR $64-11, h
+	EOR x4 @> 64-11, h, h
+	MUL prime1, h
+
+finalize:
+	EOR h >> 33, h
+	MUL prime2, h
+	EOR h >> 29, h
+	MUL prime3, h
+	EOR h >> 32, h
+
+	MOVD h, ret+24(FP)
+	RET
+
+// func writeBlocks(s *Digest, b []byte) int
+TEXT ·writeBlocks(SB), NOSPLIT|NOFRAME, $0-40
+	LDP ·primes+0(SB), (prime1, prime2)
+
+	// Load state. Assume v[1-4] are stored contiguously.
+	MOVD s+0(FP), digest
+	LDP  0(digest), (v1, v2)
+	LDP  16(digest), (v3, v4)
+
+	LDP b_base+8(FP), (p, n)
+
+	blockLoop()
+
+	// Store updated state.
+	STP (v1, v2), 0(digest)
+	STP (v3, v4), 16(digest)
+
+	BIC  $31, n
+	MOVD n, ret+32(FP)
+	RET
diff --git a/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_asm.go b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_asm.go
new file mode 100644
index 000000000..d4221edf4
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_asm.go
@@ -0,0 +1,16 @@
+//go:build (amd64 || arm64) && !appengine && gc && !purego && !noasm
+// +build amd64 arm64
+// +build !appengine
+// +build gc
+// +build !purego
+// +build !noasm
+
+package xxhash
+
+// Sum64 computes the 64-bit xxHash digest of b.
+//
+//go:noescape
+func Sum64(b []byte) uint64
+
+//go:noescape
+func writeBlocks(s *Digest, b []byte) int
diff --git a/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_other.go b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_other.go
new file mode 100644
index 000000000..0be16cefc
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_other.go
@@ -0,0 +1,76 @@
+//go:build (!amd64 && !arm64) || appengine || !gc || purego || noasm
+// +build !amd64,!arm64 appengine !gc purego noasm
+
+package xxhash
+
+// Sum64 computes the 64-bit xxHash digest of b.
+func Sum64(b []byte) uint64 {
+	// A simpler version would be
+	//   d := New()
+	//   d.Write(b)
+	//   return d.Sum64()
+	// but this is faster, particularly for small inputs.
+
+	n := len(b)
+	var h uint64
+
+	if n >= 32 {
+		v1 := primes[0] + prime2
+		v2 := prime2
+		v3 := uint64(0)
+		v4 := -primes[0]
+		for len(b) >= 32 {
+			v1 = round(v1, u64(b[0:8:len(b)]))
+			v2 = round(v2, u64(b[8:16:len(b)]))
+			v3 = round(v3, u64(b[16:24:len(b)]))
+			v4 = round(v4, u64(b[24:32:len(b)]))
+			b = b[32:len(b):len(b)]
+		}
+		h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
+		h = mergeRound(h, v1)
+		h = mergeRound(h, v2)
+		h = mergeRound(h, v3)
+		h = mergeRound(h, v4)
+	} else {
+		h = prime5
+	}
+
+	h += uint64(n)
+
+	for ; len(b) >= 8; b = b[8:] {
+		k1 := round(0, u64(b[:8]))
+		h ^= k1
+		h = rol27(h)*prime1 + prime4
+	}
+	if len(b) >= 4 {
+		h ^= uint64(u32(b[:4])) * prime1
+		h = rol23(h)*prime2 + prime3
+		b = b[4:]
+	}
+	for ; len(b) > 0; b = b[1:] {
+		h ^= uint64(b[0]) * prime5
+		h = rol11(h) * prime1
+	}
+
+	h ^= h >> 33
+	h *= prime2
+	h ^= h >> 29
+	h *= prime3
+	h ^= h >> 32
+
+	return h
+}
+
+func writeBlocks(d *Digest, b []byte) int {
+	v1, v2, v3, v4 := d.v1, d.v2, d.v3, d.v4
+	n := len(b)
+	for len(b) >= 32 {
+		v1 = round(v1, u64(b[0:8:len(b)]))
+		v2 = round(v2, u64(b[8:16:len(b)]))
+		v3 = round(v3, u64(b[16:24:len(b)]))
+		v4 = round(v4, u64(b[24:32:len(b)]))
+		b = b[32:len(b):len(b)]
+	}
+	d.v1, d.v2, d.v3, d.v4 = v1, v2, v3, v4
+	return n - len(b)
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_safe.go b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_safe.go
new file mode 100644
index 000000000..6f3b0cb10
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_safe.go
@@ -0,0 +1,11 @@
+package xxhash
+
+// Sum64String computes the 64-bit xxHash digest of s.
+func Sum64String(s string) uint64 {
+	return Sum64([]byte(s))
+}
+
+// WriteString adds more data to d. It always returns len(s), nil.
+func (d *Digest) WriteString(s string) (n int, err error) {
+	return d.Write([]byte(s))
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/matchlen_amd64.go b/vendor/github.com/klauspost/compress/zstd/matchlen_amd64.go
new file mode 100644
index 000000000..f41932b7a
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/matchlen_amd64.go
@@ -0,0 +1,16 @@
+//go:build amd64 && !appengine && !noasm && gc
+// +build amd64,!appengine,!noasm,gc
+
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+
+package zstd
+
+// matchLen returns how many bytes match in a and b
+//
+// It assumes that:
+//
+//	len(a) <= len(b) and len(a) > 0
+//
+//go:noescape
+func matchLen(a []byte, b []byte) int
diff --git a/vendor/github.com/klauspost/compress/zstd/matchlen_amd64.s b/vendor/github.com/klauspost/compress/zstd/matchlen_amd64.s
new file mode 100644
index 000000000..0782b86e3
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/matchlen_amd64.s
@@ -0,0 +1,66 @@
+// Copied from S2 implementation.
+
+//go:build !appengine && !noasm && gc && !noasm
+
+#include "textflag.h"
+
+// func matchLen(a []byte, b []byte) int
+TEXT ·matchLen(SB), NOSPLIT, $0-56
+	MOVQ a_base+0(FP), AX
+	MOVQ b_base+24(FP), CX
+	MOVQ a_len+8(FP), DX
+
+	// matchLen
+	XORL SI, SI
+	CMPL DX, $0x08
+	JB   matchlen_match4_standalone
+
+matchlen_loopback_standalone:
+	MOVQ (AX)(SI*1), BX
+	XORQ (CX)(SI*1), BX
+	JZ   matchlen_loop_standalone
+
+#ifdef GOAMD64_v3
+	TZCNTQ BX, BX
+#else
+	BSFQ BX, BX
+#endif
+	SHRL $0x03, BX
+	LEAL (SI)(BX*1), SI
+	JMP  gen_match_len_end
+
+matchlen_loop_standalone:
+	LEAL -8(DX), DX
+	LEAL 8(SI), SI
+	CMPL DX, $0x08
+	JAE  matchlen_loopback_standalone
+
+matchlen_match4_standalone:
+	CMPL DX, $0x04
+	JB   matchlen_match2_standalone
+	MOVL (AX)(SI*1), BX
+	CMPL (CX)(SI*1), BX
+	JNE  matchlen_match2_standalone
+	LEAL -4(DX), DX
+	LEAL 4(SI), SI
+
+matchlen_match2_standalone:
+	CMPL DX, $0x02
+	JB   matchlen_match1_standalone
+	MOVW (AX)(SI*1), BX
+	CMPW (CX)(SI*1), BX
+	JNE  matchlen_match1_standalone
+	LEAL -2(DX), DX
+	LEAL 2(SI), SI
+
+matchlen_match1_standalone:
+	CMPL DX, $0x01
+	JB   gen_match_len_end
+	MOVB (AX)(SI*1), BL
+	CMPB (CX)(SI*1), BL
+	JNE  gen_match_len_end
+	INCL SI
+
+gen_match_len_end:
+	MOVQ SI, ret+48(FP)
+	RET
diff --git a/vendor/github.com/klauspost/compress/zstd/matchlen_generic.go b/vendor/github.com/klauspost/compress/zstd/matchlen_generic.go
new file mode 100644
index 000000000..57b9c31c0
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/matchlen_generic.go
@@ -0,0 +1,33 @@
+//go:build !amd64 || appengine || !gc || noasm
+// +build !amd64 appengine !gc noasm
+
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+
+package zstd
+
+import (
+	"encoding/binary"
+	"math/bits"
+)
+
+// 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) {
+	for ; len(a) >= 8 && len(b) >= 8; a, b = a[8:], b[8:] {
+		diff := binary.LittleEndian.Uint64(a) ^ binary.LittleEndian.Uint64(b)
+		if diff != 0 {
+			return n + bits.TrailingZeros64(diff)>>3
+		}
+		n += 8
+	}
+
+	for i := range a {
+		if a[i] != b[i] {
+			break
+		}
+		n++
+	}
+	return n
+
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/seqdec.go b/vendor/github.com/klauspost/compress/zstd/seqdec.go
new file mode 100644
index 000000000..d7fe6d82d
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/seqdec.go
@@ -0,0 +1,503 @@
+// 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"
+	"io"
+)
+
+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 seqVals struct {
+	ll, ml, mo int
+}
+
+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 seqCompMode uint8
+
+const (
+	compModePredefined seqCompMode = iota
+	compModeRLE
+	compModeFSE
+	compModeRepeat
+)
+
+type sequenceDec struct {
+	// decoder keeps track of the current state and updates it from the bitstream.
+	fse    *fseDecoder
+	state  fseState
+	repeat bool
+}
+
+// init the state of the decoder with input from stream.
+func (s *sequenceDec) init(br *bitReader) error {
+	if s.fse == nil {
+		return errors.New("sequence decoder not defined")
+	}
+	s.state.init(br, s.fse.actualTableLog, s.fse.dt[:1<<s.fse.actualTableLog])
+	return nil
+}
+
+// sequenceDecs contains all 3 sequence decoders and their state.
+type sequenceDecs struct {
+	litLengths   sequenceDec
+	offsets      sequenceDec
+	matchLengths sequenceDec
+	prevOffset   [3]int
+	dict         []byte
+	literals     []byte
+	out          []byte
+	nSeqs        int
+	br           *bitReader
+	seqSize      int
+	windowSize   int
+	maxBits      uint8
+	maxSyncLen   uint64
+}
+
+// initialize all 3 decoders from the stream input.
+func (s *sequenceDecs) initialize(br *bitReader, hist *history, out []byte) error {
+	if err := s.litLengths.init(br); err != nil {
+		return errors.New("litLengths:" + err.Error())
+	}
+	if err := s.offsets.init(br); err != nil {
+		return errors.New("offsets:" + err.Error())
+	}
+	if err := s.matchLengths.init(br); err != nil {
+		return errors.New("matchLengths:" + err.Error())
+	}
+	s.br = br
+	s.prevOffset = hist.recentOffsets
+	s.maxBits = s.litLengths.fse.maxBits + s.offsets.fse.maxBits + s.matchLengths.fse.maxBits
+	s.windowSize = hist.windowSize
+	s.out = out
+	s.dict = nil
+	if hist.dict != nil {
+		s.dict = hist.dict.content
+	}
+	return nil
+}
+
+func (s *sequenceDecs) freeDecoders() {
+	if f := s.litLengths.fse; f != nil && !f.preDefined {
+		fseDecoderPool.Put(f)
+		s.litLengths.fse = nil
+	}
+	if f := s.offsets.fse; f != nil && !f.preDefined {
+		fseDecoderPool.Put(f)
+		s.offsets.fse = nil
+	}
+	if f := s.matchLengths.fse; f != nil && !f.preDefined {
+		fseDecoderPool.Put(f)
+		s.matchLengths.fse = nil
+	}
+}
+
+// execute will execute the decoded sequence with the provided history.
+// The sequence must be evaluated before being sent.
+func (s *sequenceDecs) execute(seqs []seqVals, hist []byte) error {
+	if len(s.dict) == 0 {
+		return s.executeSimple(seqs, hist)
+	}
+
+	// Ensure we have enough output size...
+	if len(s.out)+s.seqSize > cap(s.out) {
+		addBytes := s.seqSize + len(s.out)
+		s.out = append(s.out, make([]byte, addBytes)...)
+		s.out = s.out[:len(s.out)-addBytes]
+	}
+
+	if debugDecoder {
+		printf("Execute %d seqs with hist %d, dict %d, literals: %d into %d bytes\n", len(seqs), len(hist), len(s.dict), len(s.literals), s.seqSize)
+	}
+
+	var t = len(s.out)
+	out := s.out[:t+s.seqSize]
+
+	for _, seq := range seqs {
+		// Add literals
+		copy(out[t:], s.literals[:seq.ll])
+		t += seq.ll
+		s.literals = s.literals[seq.ll:]
+
+		// Copy from dictionary...
+		if seq.mo > t+len(hist) || seq.mo > s.windowSize {
+			if len(s.dict) == 0 {
+				return fmt.Errorf("match offset (%d) bigger than current history (%d)", seq.mo, t+len(hist))
+			}
+
+			// we may be in dictionary.
+			dictO := len(s.dict) - (seq.mo - (t + len(hist)))
+			if dictO < 0 || dictO >= len(s.dict) {
+				return fmt.Errorf("match offset (%d) bigger than current history+dict (%d)", seq.mo, t+len(hist)+len(s.dict))
+			}
+			end := dictO + seq.ml
+			if end > len(s.dict) {
+				n := len(s.dict) - dictO
+				copy(out[t:], s.dict[dictO:])
+				t += n
+				seq.ml -= n
+			} else {
+				copy(out[t:], s.dict[dictO:end])
+				t += end - dictO
+				continue
+			}
+		}
+
+		// Copy from history.
+		if v := seq.mo - t; v > 0 {
+			// v is the start position in history from end.
+			start := len(hist) - v
+			if seq.ml > v {
+				// Some goes into current block.
+				// Copy remainder of history
+				copy(out[t:], hist[start:])
+				t += v
+				seq.ml -= v
+			} else {
+				copy(out[t:], hist[start:start+seq.ml])
+				t += seq.ml
+				continue
+			}
+		}
+		// We must be in current buffer now
+		if seq.ml > 0 {
+			start := t - seq.mo
+			if seq.ml <= t-start {
+				// No overlap
+				copy(out[t:], out[start:start+seq.ml])
+				t += seq.ml
+				continue
+			} else {
+				// Overlapping copy
+				// Extend destination slice and copy one byte at the time.
+				src := out[start : start+seq.ml]
+				dst := out[t:]
+				dst = dst[:len(src)]
+				t += len(src)
+				// Destination is the space we just added.
+				for i := range src {
+					dst[i] = src[i]
+				}
+			}
+		}
+	}
+
+	// Add final literals
+	copy(out[t:], s.literals)
+	if debugDecoder {
+		t += len(s.literals)
+		if t != len(out) {
+			panic(fmt.Errorf("length mismatch, want %d, got %d, ss: %d", len(out), t, s.seqSize))
+		}
+	}
+	s.out = out
+
+	return nil
+}
+
+// decode sequences from the stream with the provided history.
+func (s *sequenceDecs) decodeSync(hist []byte) error {
+	supported, err := s.decodeSyncSimple(hist)
+	if supported {
+		return err
+	}
+
+	br := s.br
+	seqs := s.nSeqs
+	startSize := len(s.out)
+	// Grab full sizes tables, to avoid bounds checks.
+	llTable, mlTable, ofTable := s.litLengths.fse.dt[:maxTablesize], s.matchLengths.fse.dt[:maxTablesize], s.offsets.fse.dt[:maxTablesize]
+	llState, mlState, ofState := s.litLengths.state.state, s.matchLengths.state.state, s.offsets.state.state
+	out := s.out
+	maxBlockSize := maxCompressedBlockSize
+	if s.windowSize < maxBlockSize {
+		maxBlockSize = s.windowSize
+	}
+
+	if debugDecoder {
+		println("decodeSync: decoding", seqs, "sequences", br.remain(), "bits remain on stream")
+	}
+	for i := seqs - 1; i >= 0; i-- {
+		if br.overread() {
+			printf("reading sequence %d, exceeded available data. Overread by %d\n", seqs-i, -br.remain())
+			return io.ErrUnexpectedEOF
+		}
+		var ll, mo, ml int
+		if len(br.in) > 4+((maxOffsetBits+16+16)>>3) {
+			// inlined function:
+			// ll, mo, ml = s.nextFast(br, llState, mlState, ofState)
+
+			// Final will not read from stream.
+			var llB, mlB, moB uint8
+			ll, llB = llState.final()
+			ml, mlB = mlState.final()
+			mo, moB = ofState.final()
+
+			// extra bits are stored in reverse order.
+			br.fillFast()
+			mo += br.getBits(moB)
+			if s.maxBits > 32 {
+				br.fillFast()
+			}
+			ml += br.getBits(mlB)
+			ll += br.getBits(llB)
+
+			if moB > 1 {
+				s.prevOffset[2] = s.prevOffset[1]
+				s.prevOffset[1] = s.prevOffset[0]
+				s.prevOffset[0] = mo
+			} else {
+				// mo = s.adjustOffset(mo, ll, moB)
+				// Inlined for rather big speedup
+				if ll == 0 {
+					// There is an exception though, when current sequence's literals_length = 0.
+					// In this case, repeated offsets are shifted by one, so an offset_value of 1 means Repeated_Offset2,
+					// an offset_value of 2 means Repeated_Offset3, and an offset_value of 3 means Repeated_Offset1 - 1_byte.
+					mo++
+				}
+
+				if mo == 0 {
+					mo = s.prevOffset[0]
+				} else {
+					var temp int
+					if mo == 3 {
+						temp = s.prevOffset[0] - 1
+					} else {
+						temp = s.prevOffset[mo]
+					}
+
+					if temp == 0 {
+						// 0 is not valid; input is corrupted; force offset to 1
+						println("WARNING: temp was 0")
+						temp = 1
+					}
+
+					if mo != 1 {
+						s.prevOffset[2] = s.prevOffset[1]
+					}
+					s.prevOffset[1] = s.prevOffset[0]
+					s.prevOffset[0] = temp
+					mo = temp
+				}
+			}
+			br.fillFast()
+		} else {
+			ll, mo, ml = s.next(br, llState, mlState, ofState)
+			br.fill()
+		}
+
+		if debugSequences {
+			println("Seq", seqs-i-1, "Litlen:", ll, "mo:", mo, "(abs) ml:", ml)
+		}
+
+		if ll > len(s.literals) {
+			return fmt.Errorf("unexpected literal count, want %d bytes, but only %d is available", ll, len(s.literals))
+		}
+		size := ll + ml + len(out)
+		if size-startSize > maxBlockSize {
+			return fmt.Errorf("output bigger than max block size (%d)", maxBlockSize)
+		}
+		if size > cap(out) {
+			// Not enough size, which can happen under high volume block streaming conditions
+			// but could be if destination slice is too small for sync operations.
+			// over-allocating here can create a large amount of GC pressure so we try to keep
+			// it as contained as possible
+			used := len(out) - startSize
+			addBytes := 256 + ll + ml + used>>2
+			// Clamp to max block size.
+			if used+addBytes > maxBlockSize {
+				addBytes = maxBlockSize - used
+			}
+			out = append(out, make([]byte, addBytes)...)
+			out = out[:len(out)-addBytes]
+		}
+		if ml > maxMatchLen {
+			return fmt.Errorf("match len (%d) bigger than max allowed length", ml)
+		}
+
+		// Add literals
+		out = append(out, s.literals[:ll]...)
+		s.literals = s.literals[ll:]
+
+		if mo == 0 && ml > 0 {
+			return fmt.Errorf("zero matchoff and matchlen (%d) > 0", ml)
+		}
+
+		if mo > len(out)+len(hist) || mo > s.windowSize {
+			if len(s.dict) == 0 {
+				return fmt.Errorf("match offset (%d) bigger than current history (%d)", mo, len(out)+len(hist)-startSize)
+			}
+
+			// we may be in dictionary.
+			dictO := len(s.dict) - (mo - (len(out) + len(hist)))
+			if dictO < 0 || dictO >= len(s.dict) {
+				return fmt.Errorf("match offset (%d) bigger than current history (%d)", mo, len(out)+len(hist)-startSize)
+			}
+			end := dictO + ml
+			if end > len(s.dict) {
+				out = append(out, s.dict[dictO:]...)
+				ml -= len(s.dict) - dictO
+			} else {
+				out = append(out, s.dict[dictO:end]...)
+				mo = 0
+				ml = 0
+			}
+		}
+
+		// Copy from history.
+		// TODO: Blocks without history could be made to ignore this completely.
+		if v := mo - len(out); v > 0 {
+			// v is the start position in history from end.
+			start := len(hist) - v
+			if ml > v {
+				// Some goes into current block.
+				// Copy remainder of history
+				out = append(out, hist[start:]...)
+				ml -= v
+			} else {
+				out = append(out, hist[start:start+ml]...)
+				ml = 0
+			}
+		}
+		// We must be in current buffer now
+		if ml > 0 {
+			start := len(out) - mo
+			if ml <= len(out)-start {
+				// No overlap
+				out = append(out, out[start:start+ml]...)
+			} else {
+				// Overlapping copy
+				// Extend destination slice and copy one byte at the time.
+				out = out[:len(out)+ml]
+				src := out[start : start+ml]
+				// Destination is the space we just added.
+				dst := out[len(out)-ml:]
+				dst = dst[:len(src)]
+				for i := range src {
+					dst[i] = src[i]
+				}
+			}
+		}
+		if i == 0 {
+			// This is the last sequence, so we shouldn't update state.
+			break
+		}
+
+		// Manually inlined, ~ 5-20% faster
+		// Update all 3 states at once. Approx 20% faster.
+		nBits := llState.nbBits() + mlState.nbBits() + ofState.nbBits()
+		if nBits == 0 {
+			llState = llTable[llState.newState()&maxTableMask]
+			mlState = mlTable[mlState.newState()&maxTableMask]
+			ofState = ofTable[ofState.newState()&maxTableMask]
+		} else {
+			bits := br.get32BitsFast(nBits)
+
+			lowBits := uint16(bits >> ((ofState.nbBits() + mlState.nbBits()) & 31))
+			llState = llTable[(llState.newState()+lowBits)&maxTableMask]
+
+			lowBits = uint16(bits >> (ofState.nbBits() & 31))
+			lowBits &= bitMask[mlState.nbBits()&15]
+			mlState = mlTable[(mlState.newState()+lowBits)&maxTableMask]
+
+			lowBits = uint16(bits) & bitMask[ofState.nbBits()&15]
+			ofState = ofTable[(ofState.newState()+lowBits)&maxTableMask]
+		}
+	}
+
+	if size := len(s.literals) + len(out) - startSize; size > maxBlockSize {
+		return fmt.Errorf("output bigger than max block size (%d)", maxBlockSize)
+	}
+
+	// Add final literals
+	s.out = append(out, s.literals...)
+	return br.close()
+}
+
+var bitMask [16]uint16
+
+func init() {
+	for i := range bitMask[:] {
+		bitMask[i] = uint16((1 << uint(i)) - 1)
+	}
+}
+
+func (s *sequenceDecs) next(br *bitReader, llState, mlState, ofState decSymbol) (ll, mo, ml int) {
+	// Final will not read from stream.
+	ll, llB := llState.final()
+	ml, mlB := mlState.final()
+	mo, moB := ofState.final()
+
+	// extra bits are stored in reverse order.
+	br.fill()
+	mo += br.getBits(moB)
+	if s.maxBits > 32 {
+		br.fill()
+	}
+	// matchlength+literal length, max 32 bits
+	ml += br.getBits(mlB)
+	ll += br.getBits(llB)
+	mo = s.adjustOffset(mo, ll, moB)
+	return
+}
+
+func (s *sequenceDecs) adjustOffset(offset, litLen int, offsetB uint8) int {
+	if offsetB > 1 {
+		s.prevOffset[2] = s.prevOffset[1]
+		s.prevOffset[1] = s.prevOffset[0]
+		s.prevOffset[0] = offset
+		return offset
+	}
+
+	if litLen == 0 {
+		// There is an exception though, when current sequence's literals_length = 0.
+		// In this case, repeated offsets are shifted by one, so an offset_value of 1 means Repeated_Offset2,
+		// an offset_value of 2 means Repeated_Offset3, and an offset_value of 3 means Repeated_Offset1 - 1_byte.
+		offset++
+	}
+
+	if offset == 0 {
+		return s.prevOffset[0]
+	}
+	var temp int
+	if offset == 3 {
+		temp = s.prevOffset[0] - 1
+	} else {
+		temp = s.prevOffset[offset]
+	}
+
+	if temp == 0 {
+		// 0 is not valid; input is corrupted; force offset to 1
+		println("temp was 0")
+		temp = 1
+	}
+
+	if offset != 1 {
+		s.prevOffset[2] = s.prevOffset[1]
+	}
+	s.prevOffset[1] = s.prevOffset[0]
+	s.prevOffset[0] = temp
+	return temp
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/seqdec_amd64.go b/vendor/github.com/klauspost/compress/zstd/seqdec_amd64.go
new file mode 100644
index 000000000..8adabd828
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/seqdec_amd64.go
@@ -0,0 +1,394 @@
+//go:build amd64 && !appengine && !noasm && gc
+// +build amd64,!appengine,!noasm,gc
+
+package zstd
+
+import (
+	"fmt"
+	"io"
+
+	"github.com/klauspost/compress/internal/cpuinfo"
+)
+
+type decodeSyncAsmContext struct {
+	llTable     []decSymbol
+	mlTable     []decSymbol
+	ofTable     []decSymbol
+	llState     uint64
+	mlState     uint64
+	ofState     uint64
+	iteration   int
+	litRemain   int
+	out         []byte
+	outPosition int
+	literals    []byte
+	litPosition int
+	history     []byte
+	windowSize  int
+	ll          int // set on error (not for all errors, please refer to _generate/gen.go)
+	ml          int // set on error (not for all errors, please refer to _generate/gen.go)
+	mo          int // set on error (not for all errors, please refer to _generate/gen.go)
+}
+
+// sequenceDecs_decodeSync_amd64 implements the main loop of sequenceDecs.decodeSync in x86 asm.
+//
+// Please refer to seqdec_generic.go for the reference implementation.
+//
+//go:noescape
+func sequenceDecs_decodeSync_amd64(s *sequenceDecs, br *bitReader, ctx *decodeSyncAsmContext) int
+
+// sequenceDecs_decodeSync_bmi2 implements the main loop of sequenceDecs.decodeSync in x86 asm with BMI2 extensions.
+//
+//go:noescape
+func sequenceDecs_decodeSync_bmi2(s *sequenceDecs, br *bitReader, ctx *decodeSyncAsmContext) int
+
+// sequenceDecs_decodeSync_safe_amd64 does the same as above, but does not write more than output buffer.
+//
+//go:noescape
+func sequenceDecs_decodeSync_safe_amd64(s *sequenceDecs, br *bitReader, ctx *decodeSyncAsmContext) int
+
+// sequenceDecs_decodeSync_safe_bmi2 does the same as above, but does not write more than output buffer.
+//
+//go:noescape
+func sequenceDecs_decodeSync_safe_bmi2(s *sequenceDecs, br *bitReader, ctx *decodeSyncAsmContext) int
+
+// decode sequences from the stream with the provided history but without a dictionary.
+func (s *sequenceDecs) decodeSyncSimple(hist []byte) (bool, error) {
+	if len(s.dict) > 0 {
+		return false, nil
+	}
+	if s.maxSyncLen == 0 && cap(s.out)-len(s.out) < maxCompressedBlockSize {
+		return false, nil
+	}
+
+	// FIXME: Using unsafe memory copies leads to rare, random crashes
+	// with fuzz testing. It is therefore disabled for now.
+	const useSafe = true
+	/*
+		useSafe := false
+		if s.maxSyncLen == 0 && cap(s.out)-len(s.out) < maxCompressedBlockSizeAlloc {
+			useSafe = true
+		}
+		if s.maxSyncLen > 0 && cap(s.out)-len(s.out)-compressedBlockOverAlloc < int(s.maxSyncLen) {
+			useSafe = true
+		}
+		if cap(s.literals) < len(s.literals)+compressedBlockOverAlloc {
+			useSafe = true
+		}
+	*/
+
+	br := s.br
+
+	maxBlockSize := maxCompressedBlockSize
+	if s.windowSize < maxBlockSize {
+		maxBlockSize = s.windowSize
+	}
+
+	ctx := decodeSyncAsmContext{
+		llTable:     s.litLengths.fse.dt[:maxTablesize],
+		mlTable:     s.matchLengths.fse.dt[:maxTablesize],
+		ofTable:     s.offsets.fse.dt[:maxTablesize],
+		llState:     uint64(s.litLengths.state.state),
+		mlState:     uint64(s.matchLengths.state.state),
+		ofState:     uint64(s.offsets.state.state),
+		iteration:   s.nSeqs - 1,
+		litRemain:   len(s.literals),
+		out:         s.out,
+		outPosition: len(s.out),
+		literals:    s.literals,
+		windowSize:  s.windowSize,
+		history:     hist,
+	}
+
+	s.seqSize = 0
+	startSize := len(s.out)
+
+	var errCode int
+	if cpuinfo.HasBMI2() {
+		if useSafe {
+			errCode = sequenceDecs_decodeSync_safe_bmi2(s, br, &ctx)
+		} else {
+			errCode = sequenceDecs_decodeSync_bmi2(s, br, &ctx)
+		}
+	} else {
+		if useSafe {
+			errCode = sequenceDecs_decodeSync_safe_amd64(s, br, &ctx)
+		} else {
+			errCode = sequenceDecs_decodeSync_amd64(s, br, &ctx)
+		}
+	}
+	switch errCode {
+	case noError:
+		break
+
+	case errorMatchLenOfsMismatch:
+		return true, fmt.Errorf("zero matchoff and matchlen (%d) > 0", ctx.ml)
+
+	case errorMatchLenTooBig:
+		return true, fmt.Errorf("match len (%d) bigger than max allowed length", ctx.ml)
+
+	case errorMatchOffTooBig:
+		return true, fmt.Errorf("match offset (%d) bigger than current history (%d)",
+			ctx.mo, ctx.outPosition+len(hist)-startSize)
+
+	case errorNotEnoughLiterals:
+		return true, fmt.Errorf("unexpected literal count, want %d bytes, but only %d is available",
+			ctx.ll, ctx.litRemain+ctx.ll)
+
+	case errorOverread:
+		return true, io.ErrUnexpectedEOF
+
+	case errorNotEnoughSpace:
+		size := ctx.outPosition + ctx.ll + ctx.ml
+		if debugDecoder {
+			println("msl:", s.maxSyncLen, "cap", cap(s.out), "bef:", startSize, "sz:", size-startSize, "mbs:", maxBlockSize, "outsz:", cap(s.out)-startSize)
+		}
+		return true, fmt.Errorf("output bigger than max block size (%d)", maxBlockSize)
+
+	default:
+		return true, fmt.Errorf("sequenceDecs_decode returned erronous code %d", errCode)
+	}
+
+	s.seqSize += ctx.litRemain
+	if s.seqSize > maxBlockSize {
+		return true, fmt.Errorf("output bigger than max block size (%d)", maxBlockSize)
+	}
+	err := br.close()
+	if err != nil {
+		printf("Closing sequences: %v, %+v\n", err, *br)
+		return true, err
+	}
+
+	s.literals = s.literals[ctx.litPosition:]
+	t := ctx.outPosition
+	s.out = s.out[:t]
+
+	// Add final literals
+	s.out = append(s.out, s.literals...)
+	if debugDecoder {
+		t += len(s.literals)
+		if t != len(s.out) {
+			panic(fmt.Errorf("length mismatch, want %d, got %d", len(s.out), t))
+		}
+	}
+
+	return true, nil
+}
+
+// --------------------------------------------------------------------------------
+
+type decodeAsmContext struct {
+	llTable   []decSymbol
+	mlTable   []decSymbol
+	ofTable   []decSymbol
+	llState   uint64
+	mlState   uint64
+	ofState   uint64
+	iteration int
+	seqs      []seqVals
+	litRemain int
+}
+
+const noError = 0
+
+// error reported when mo == 0 && ml > 0
+const errorMatchLenOfsMismatch = 1
+
+// error reported when ml > maxMatchLen
+const errorMatchLenTooBig = 2
+
+// error reported when mo > available history or mo > s.windowSize
+const errorMatchOffTooBig = 3
+
+// error reported when the sum of literal lengths exeeceds the literal buffer size
+const errorNotEnoughLiterals = 4
+
+// error reported when capacity of `out` is too small
+const errorNotEnoughSpace = 5
+
+// error reported when bits are overread.
+const errorOverread = 6
+
+// sequenceDecs_decode implements the main loop of sequenceDecs in x86 asm.
+//
+// Please refer to seqdec_generic.go for the reference implementation.
+//
+//go:noescape
+func sequenceDecs_decode_amd64(s *sequenceDecs, br *bitReader, ctx *decodeAsmContext) int
+
+// sequenceDecs_decode implements the main loop of sequenceDecs in x86 asm.
+//
+// Please refer to seqdec_generic.go for the reference implementation.
+//
+//go:noescape
+func sequenceDecs_decode_56_amd64(s *sequenceDecs, br *bitReader, ctx *decodeAsmContext) int
+
+// sequenceDecs_decode implements the main loop of sequenceDecs in x86 asm with BMI2 extensions.
+//
+//go:noescape
+func sequenceDecs_decode_bmi2(s *sequenceDecs, br *bitReader, ctx *decodeAsmContext) int
+
+// sequenceDecs_decode implements the main loop of sequenceDecs in x86 asm with BMI2 extensions.
+//
+//go:noescape
+func sequenceDecs_decode_56_bmi2(s *sequenceDecs, br *bitReader, ctx *decodeAsmContext) int
+
+// decode sequences from the stream without the provided history.
+func (s *sequenceDecs) decode(seqs []seqVals) error {
+	br := s.br
+
+	maxBlockSize := maxCompressedBlockSize
+	if s.windowSize < maxBlockSize {
+		maxBlockSize = s.windowSize
+	}
+
+	ctx := decodeAsmContext{
+		llTable:   s.litLengths.fse.dt[:maxTablesize],
+		mlTable:   s.matchLengths.fse.dt[:maxTablesize],
+		ofTable:   s.offsets.fse.dt[:maxTablesize],
+		llState:   uint64(s.litLengths.state.state),
+		mlState:   uint64(s.matchLengths.state.state),
+		ofState:   uint64(s.offsets.state.state),
+		seqs:      seqs,
+		iteration: len(seqs) - 1,
+		litRemain: len(s.literals),
+	}
+
+	if debugDecoder {
+		println("decode: decoding", len(seqs), "sequences", br.remain(), "bits remain on stream")
+	}
+
+	s.seqSize = 0
+	lte56bits := s.maxBits+s.offsets.fse.actualTableLog+s.matchLengths.fse.actualTableLog+s.litLengths.fse.actualTableLog <= 56
+	var errCode int
+	if cpuinfo.HasBMI2() {
+		if lte56bits {
+			errCode = sequenceDecs_decode_56_bmi2(s, br, &ctx)
+		} else {
+			errCode = sequenceDecs_decode_bmi2(s, br, &ctx)
+		}
+	} else {
+		if lte56bits {
+			errCode = sequenceDecs_decode_56_amd64(s, br, &ctx)
+		} else {
+			errCode = sequenceDecs_decode_amd64(s, br, &ctx)
+		}
+	}
+	if errCode != 0 {
+		i := len(seqs) - ctx.iteration - 1
+		switch errCode {
+		case errorMatchLenOfsMismatch:
+			ml := ctx.seqs[i].ml
+			return fmt.Errorf("zero matchoff and matchlen (%d) > 0", ml)
+
+		case errorMatchLenTooBig:
+			ml := ctx.seqs[i].ml
+			return fmt.Errorf("match len (%d) bigger than max allowed length", ml)
+
+		case errorNotEnoughLiterals:
+			ll := ctx.seqs[i].ll
+			return fmt.Errorf("unexpected literal count, want %d bytes, but only %d is available", ll, ctx.litRemain+ll)
+		case errorOverread:
+			return io.ErrUnexpectedEOF
+		}
+
+		return fmt.Errorf("sequenceDecs_decode_amd64 returned erronous code %d", errCode)
+	}
+
+	if ctx.litRemain < 0 {
+		return fmt.Errorf("literal count is too big: total available %d, total requested %d",
+			len(s.literals), len(s.literals)-ctx.litRemain)
+	}
+
+	s.seqSize += ctx.litRemain
+	if s.seqSize > maxBlockSize {
+		return fmt.Errorf("output bigger than max block size (%d)", maxBlockSize)
+	}
+	if debugDecoder {
+		println("decode: ", br.remain(), "bits remain on stream. code:", errCode)
+	}
+	err := br.close()
+	if err != nil {
+		printf("Closing sequences: %v, %+v\n", err, *br)
+	}
+	return err
+}
+
+// --------------------------------------------------------------------------------
+
+type executeAsmContext struct {
+	seqs        []seqVals
+	seqIndex    int
+	out         []byte
+	history     []byte
+	literals    []byte
+	outPosition int
+	litPosition int
+	windowSize  int
+}
+
+// sequenceDecs_executeSimple_amd64 implements the main loop of sequenceDecs.executeSimple in x86 asm.
+//
+// Returns false if a match offset is too big.
+//
+// Please refer to seqdec_generic.go for the reference implementation.
+//
+//go:noescape
+func sequenceDecs_executeSimple_amd64(ctx *executeAsmContext) bool
+
+// Same as above, but with safe memcopies
+//
+//go:noescape
+func sequenceDecs_executeSimple_safe_amd64(ctx *executeAsmContext) bool
+
+// executeSimple handles cases when dictionary is not used.
+func (s *sequenceDecs) executeSimple(seqs []seqVals, hist []byte) error {
+	// Ensure we have enough output size...
+	if len(s.out)+s.seqSize+compressedBlockOverAlloc > cap(s.out) {
+		addBytes := s.seqSize + len(s.out) + compressedBlockOverAlloc
+		s.out = append(s.out, make([]byte, addBytes)...)
+		s.out = s.out[:len(s.out)-addBytes]
+	}
+
+	if debugDecoder {
+		printf("Execute %d seqs with literals: %d into %d bytes\n", len(seqs), len(s.literals), s.seqSize)
+	}
+
+	var t = len(s.out)
+	out := s.out[:t+s.seqSize]
+
+	ctx := executeAsmContext{
+		seqs:        seqs,
+		seqIndex:    0,
+		out:         out,
+		history:     hist,
+		outPosition: t,
+		litPosition: 0,
+		literals:    s.literals,
+		windowSize:  s.windowSize,
+	}
+	var ok bool
+	if cap(s.literals) < len(s.literals)+compressedBlockOverAlloc {
+		ok = sequenceDecs_executeSimple_safe_amd64(&ctx)
+	} else {
+		ok = sequenceDecs_executeSimple_amd64(&ctx)
+	}
+	if !ok {
+		return fmt.Errorf("match offset (%d) bigger than current history (%d)",
+			seqs[ctx.seqIndex].mo, ctx.outPosition+len(hist))
+	}
+	s.literals = s.literals[ctx.litPosition:]
+	t = ctx.outPosition
+
+	// Add final literals
+	copy(out[t:], s.literals)
+	if debugDecoder {
+		t += len(s.literals)
+		if t != len(out) {
+			panic(fmt.Errorf("length mismatch, want %d, got %d, ss: %d", len(out), t, s.seqSize))
+		}
+	}
+	s.out = out
+
+	return nil
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/seqdec_amd64.s b/vendor/github.com/klauspost/compress/zstd/seqdec_amd64.s
new file mode 100644
index 000000000..5b06174b8
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/seqdec_amd64.s
@@ -0,0 +1,4151 @@
+// Code generated by command: go run gen.go -out ../seqdec_amd64.s -pkg=zstd. DO NOT EDIT.
+
+//go:build !appengine && !noasm && gc && !noasm
+
+// func sequenceDecs_decode_amd64(s *sequenceDecs, br *bitReader, ctx *decodeAsmContext) int
+// Requires: CMOV
+TEXT ·sequenceDecs_decode_amd64(SB), $8-32
+	MOVQ    br+8(FP), CX
+	MOVQ    24(CX), DX
+	MOVBQZX 32(CX), BX
+	MOVQ    (CX), AX
+	MOVQ    8(CX), SI
+	ADDQ    SI, AX
+	MOVQ    AX, (SP)
+	MOVQ    ctx+16(FP), AX
+	MOVQ    72(AX), DI
+	MOVQ    80(AX), R8
+	MOVQ    88(AX), R9
+	MOVQ    104(AX), R10
+	MOVQ    s+0(FP), AX
+	MOVQ    144(AX), R11
+	MOVQ    152(AX), R12
+	MOVQ    160(AX), R13
+
+sequenceDecs_decode_amd64_main_loop:
+	MOVQ (SP), R14
+
+	// Fill bitreader to have enough for the offset and match length.
+	CMPQ SI, $0x08
+	JL   sequenceDecs_decode_amd64_fill_byte_by_byte
+	MOVQ BX, AX
+	SHRQ $0x03, AX
+	SUBQ AX, R14
+	MOVQ (R14), DX
+	SUBQ AX, SI
+	ANDQ $0x07, BX
+	JMP  sequenceDecs_decode_amd64_fill_end
+
+sequenceDecs_decode_amd64_fill_byte_by_byte:
+	CMPQ    SI, $0x00
+	JLE     sequenceDecs_decode_amd64_fill_check_overread
+	CMPQ    BX, $0x07
+	JLE     sequenceDecs_decode_amd64_fill_end
+	SHLQ    $0x08, DX
+	SUBQ    $0x01, R14
+	SUBQ    $0x01, SI
+	SUBQ    $0x08, BX
+	MOVBQZX (R14), AX
+	ORQ     AX, DX
+	JMP     sequenceDecs_decode_amd64_fill_byte_by_byte
+
+sequenceDecs_decode_amd64_fill_check_overread:
+	CMPQ BX, $0x40
+	JA   error_overread
+
+sequenceDecs_decode_amd64_fill_end:
+	// Update offset
+	MOVQ  R9, AX
+	MOVQ  BX, CX
+	MOVQ  DX, R15
+	SHLQ  CL, R15
+	MOVB  AH, CL
+	SHRQ  $0x20, AX
+	TESTQ CX, CX
+	JZ    sequenceDecs_decode_amd64_of_update_zero
+	ADDQ  CX, BX
+	CMPQ  BX, $0x40
+	JA    sequenceDecs_decode_amd64_of_update_zero
+	CMPQ  CX, $0x40
+	JAE   sequenceDecs_decode_amd64_of_update_zero
+	NEGQ  CX
+	SHRQ  CL, R15
+	ADDQ  R15, AX
+
+sequenceDecs_decode_amd64_of_update_zero:
+	MOVQ AX, 16(R10)
+
+	// Update match length
+	MOVQ  R8, AX
+	MOVQ  BX, CX
+	MOVQ  DX, R15
+	SHLQ  CL, R15
+	MOVB  AH, CL
+	SHRQ  $0x20, AX
+	TESTQ CX, CX
+	JZ    sequenceDecs_decode_amd64_ml_update_zero
+	ADDQ  CX, BX
+	CMPQ  BX, $0x40
+	JA    sequenceDecs_decode_amd64_ml_update_zero
+	CMPQ  CX, $0x40
+	JAE   sequenceDecs_decode_amd64_ml_update_zero
+	NEGQ  CX
+	SHRQ  CL, R15
+	ADDQ  R15, AX
+
+sequenceDecs_decode_amd64_ml_update_zero:
+	MOVQ AX, 8(R10)
+
+	// Fill bitreader to have enough for the remaining
+	CMPQ SI, $0x08
+	JL   sequenceDecs_decode_amd64_fill_2_byte_by_byte
+	MOVQ BX, AX
+	SHRQ $0x03, AX
+	SUBQ AX, R14
+	MOVQ (R14), DX
+	SUBQ AX, SI
+	ANDQ $0x07, BX
+	JMP  sequenceDecs_decode_amd64_fill_2_end
+
+sequenceDecs_decode_amd64_fill_2_byte_by_byte:
+	CMPQ    SI, $0x00
+	JLE     sequenceDecs_decode_amd64_fill_2_check_overread
+	CMPQ    BX, $0x07
+	JLE     sequenceDecs_decode_amd64_fill_2_end
+	SHLQ    $0x08, DX
+	SUBQ    $0x01, R14
+	SUBQ    $0x01, SI
+	SUBQ    $0x08, BX
+	MOVBQZX (R14), AX
+	ORQ     AX, DX
+	JMP     sequenceDecs_decode_amd64_fill_2_byte_by_byte
+
+sequenceDecs_decode_amd64_fill_2_check_overread:
+	CMPQ BX, $0x40
+	JA   error_overread
+
+sequenceDecs_decode_amd64_fill_2_end:
+	// Update literal length
+	MOVQ  DI, AX
+	MOVQ  BX, CX
+	MOVQ  DX, R15
+	SHLQ  CL, R15
+	MOVB  AH, CL
+	SHRQ  $0x20, AX
+	TESTQ CX, CX
+	JZ    sequenceDecs_decode_amd64_ll_update_zero
+	ADDQ  CX, BX
+	CMPQ  BX, $0x40
+	JA    sequenceDecs_decode_amd64_ll_update_zero
+	CMPQ  CX, $0x40
+	JAE   sequenceDecs_decode_amd64_ll_update_zero
+	NEGQ  CX
+	SHRQ  CL, R15
+	ADDQ  R15, AX
+
+sequenceDecs_decode_amd64_ll_update_zero:
+	MOVQ AX, (R10)
+
+	// Fill bitreader for state updates
+	MOVQ    R14, (SP)
+	MOVQ    R9, AX
+	SHRQ    $0x08, AX
+	MOVBQZX AL, AX
+	MOVQ    ctx+16(FP), CX
+	CMPQ    96(CX), $0x00
+	JZ      sequenceDecs_decode_amd64_skip_update
+
+	// Update Literal Length State
+	MOVBQZX DI, R14
+	SHRL    $0x10, DI
+	LEAQ    (BX)(R14*1), CX
+	MOVQ    DX, R15
+	MOVQ    CX, BX
+	ROLQ    CL, R15
+	MOVL    $0x00000001, BP
+	MOVB    R14, CL
+	SHLL    CL, BP
+	DECL    BP
+	ANDQ    BP, R15
+	ADDQ    R15, DI
+
+	// Load ctx.llTable
+	MOVQ ctx+16(FP), CX
+	MOVQ (CX), CX
+	MOVQ (CX)(DI*8), DI
+
+	// Update Match Length State
+	MOVBQZX R8, R14
+	SHRL    $0x10, R8
+	LEAQ    (BX)(R14*1), CX
+	MOVQ    DX, R15
+	MOVQ    CX, BX
+	ROLQ    CL, R15
+	MOVL    $0x00000001, BP
+	MOVB    R14, CL
+	SHLL    CL, BP
+	DECL    BP
+	ANDQ    BP, R15
+	ADDQ    R15, R8
+
+	// Load ctx.mlTable
+	MOVQ ctx+16(FP), CX
+	MOVQ 24(CX), CX
+	MOVQ (CX)(R8*8), R8
+
+	// Update Offset State
+	MOVBQZX R9, R14
+	SHRL    $0x10, R9
+	LEAQ    (BX)(R14*1), CX
+	MOVQ    DX, R15
+	MOVQ    CX, BX
+	ROLQ    CL, R15
+	MOVL    $0x00000001, BP
+	MOVB    R14, CL
+	SHLL    CL, BP
+	DECL    BP
+	ANDQ    BP, R15
+	ADDQ    R15, R9
+
+	// Load ctx.ofTable
+	MOVQ ctx+16(FP), CX
+	MOVQ 48(CX), CX
+	MOVQ (CX)(R9*8), R9
+
+sequenceDecs_decode_amd64_skip_update:
+	// Adjust offset
+	MOVQ 16(R10), CX
+	CMPQ AX, $0x01
+	JBE  sequenceDecs_decode_amd64_adjust_offsetB_1_or_0
+	MOVQ R12, R13
+	MOVQ R11, R12
+	MOVQ CX, R11
+	JMP  sequenceDecs_decode_amd64_after_adjust
+
+sequenceDecs_decode_amd64_adjust_offsetB_1_or_0:
+	CMPQ (R10), $0x00000000
+	JNE  sequenceDecs_decode_amd64_adjust_offset_maybezero
+	INCQ CX
+	JMP  sequenceDecs_decode_amd64_adjust_offset_nonzero
+
+sequenceDecs_decode_amd64_adjust_offset_maybezero:
+	TESTQ CX, CX
+	JNZ   sequenceDecs_decode_amd64_adjust_offset_nonzero
+	MOVQ  R11, CX
+	JMP   sequenceDecs_decode_amd64_after_adjust
+
+sequenceDecs_decode_amd64_adjust_offset_nonzero:
+	CMPQ CX, $0x01
+	JB   sequenceDecs_decode_amd64_adjust_zero
+	JEQ  sequenceDecs_decode_amd64_adjust_one
+	CMPQ CX, $0x02
+	JA   sequenceDecs_decode_amd64_adjust_three
+	JMP  sequenceDecs_decode_amd64_adjust_two
+
+sequenceDecs_decode_amd64_adjust_zero:
+	MOVQ R11, AX
+	JMP  sequenceDecs_decode_amd64_adjust_test_temp_valid
+
+sequenceDecs_decode_amd64_adjust_one:
+	MOVQ R12, AX
+	JMP  sequenceDecs_decode_amd64_adjust_test_temp_valid
+
+sequenceDecs_decode_amd64_adjust_two:
+	MOVQ R13, AX
+	JMP  sequenceDecs_decode_amd64_adjust_test_temp_valid
+
+sequenceDecs_decode_amd64_adjust_three:
+	LEAQ -1(R11), AX
+
+sequenceDecs_decode_amd64_adjust_test_temp_valid:
+	TESTQ AX, AX
+	JNZ   sequenceDecs_decode_amd64_adjust_temp_valid
+	MOVQ  $0x00000001, AX
+
+sequenceDecs_decode_amd64_adjust_temp_valid:
+	CMPQ    CX, $0x01
+	CMOVQNE R12, R13
+	MOVQ    R11, R12
+	MOVQ    AX, R11
+	MOVQ    AX, CX
+
+sequenceDecs_decode_amd64_after_adjust:
+	MOVQ CX, 16(R10)
+
+	// Check values
+	MOVQ  8(R10), AX
+	MOVQ  (R10), R14
+	LEAQ  (AX)(R14*1), R15
+	MOVQ  s+0(FP), BP
+	ADDQ  R15, 256(BP)
+	MOVQ  ctx+16(FP), R15
+	SUBQ  R14, 128(R15)
+	JS    error_not_enough_literals
+	CMPQ  AX, $0x00020002
+	JA    sequenceDecs_decode_amd64_error_match_len_too_big
+	TESTQ CX, CX
+	JNZ   sequenceDecs_decode_amd64_match_len_ofs_ok
+	TESTQ AX, AX
+	JNZ   sequenceDecs_decode_amd64_error_match_len_ofs_mismatch
+
+sequenceDecs_decode_amd64_match_len_ofs_ok:
+	ADDQ $0x18, R10
+	MOVQ ctx+16(FP), AX
+	DECQ 96(AX)
+	JNS  sequenceDecs_decode_amd64_main_loop
+	MOVQ s+0(FP), AX
+	MOVQ R11, 144(AX)
+	MOVQ R12, 152(AX)
+	MOVQ R13, 160(AX)
+	MOVQ br+8(FP), AX
+	MOVQ DX, 24(AX)
+	MOVB BL, 32(AX)
+	MOVQ SI, 8(AX)
+
+	// Return success
+	MOVQ $0x00000000, ret+24(FP)
+	RET
+
+	// Return with match length error
+sequenceDecs_decode_amd64_error_match_len_ofs_mismatch:
+	MOVQ $0x00000001, ret+24(FP)
+	RET
+
+	// Return with match too long error
+sequenceDecs_decode_amd64_error_match_len_too_big:
+	MOVQ $0x00000002, ret+24(FP)
+	RET
+
+	// Return with match offset too long error
+	MOVQ $0x00000003, ret+24(FP)
+	RET
+
+	// Return with not enough literals error
+error_not_enough_literals:
+	MOVQ $0x00000004, ret+24(FP)
+	RET
+
+	// Return with overread error
+error_overread:
+	MOVQ $0x00000006, ret+24(FP)
+	RET
+
+// func sequenceDecs_decode_56_amd64(s *sequenceDecs, br *bitReader, ctx *decodeAsmContext) int
+// Requires: CMOV
+TEXT ·sequenceDecs_decode_56_amd64(SB), $8-32
+	MOVQ    br+8(FP), CX
+	MOVQ    24(CX), DX
+	MOVBQZX 32(CX), BX
+	MOVQ    (CX), AX
+	MOVQ    8(CX), SI
+	ADDQ    SI, AX
+	MOVQ    AX, (SP)
+	MOVQ    ctx+16(FP), AX
+	MOVQ    72(AX), DI
+	MOVQ    80(AX), R8
+	MOVQ    88(AX), R9
+	MOVQ    104(AX), R10
+	MOVQ    s+0(FP), AX
+	MOVQ    144(AX), R11
+	MOVQ    152(AX), R12
+	MOVQ    160(AX), R13
+
+sequenceDecs_decode_56_amd64_main_loop:
+	MOVQ (SP), R14
+
+	// Fill bitreader to have enough for the offset and match length.
+	CMPQ SI, $0x08
+	JL   sequenceDecs_decode_56_amd64_fill_byte_by_byte
+	MOVQ BX, AX
+	SHRQ $0x03, AX
+	SUBQ AX, R14
+	MOVQ (R14), DX
+	SUBQ AX, SI
+	ANDQ $0x07, BX
+	JMP  sequenceDecs_decode_56_amd64_fill_end
+
+sequenceDecs_decode_56_amd64_fill_byte_by_byte:
+	CMPQ    SI, $0x00
+	JLE     sequenceDecs_decode_56_amd64_fill_check_overread
+	CMPQ    BX, $0x07
+	JLE     sequenceDecs_decode_56_amd64_fill_end
+	SHLQ    $0x08, DX
+	SUBQ    $0x01, R14
+	SUBQ    $0x01, SI
+	SUBQ    $0x08, BX
+	MOVBQZX (R14), AX
+	ORQ     AX, DX
+	JMP     sequenceDecs_decode_56_amd64_fill_byte_by_byte
+
+sequenceDecs_decode_56_amd64_fill_check_overread:
+	CMPQ BX, $0x40
+	JA   error_overread
+
+sequenceDecs_decode_56_amd64_fill_end:
+	// Update offset
+	MOVQ  R9, AX
+	MOVQ  BX, CX
+	MOVQ  DX, R15
+	SHLQ  CL, R15
+	MOVB  AH, CL
+	SHRQ  $0x20, AX
+	TESTQ CX, CX
+	JZ    sequenceDecs_decode_56_amd64_of_update_zero
+	ADDQ  CX, BX
+	CMPQ  BX, $0x40
+	JA    sequenceDecs_decode_56_amd64_of_update_zero
+	CMPQ  CX, $0x40
+	JAE   sequenceDecs_decode_56_amd64_of_update_zero
+	NEGQ  CX
+	SHRQ  CL, R15
+	ADDQ  R15, AX
+
+sequenceDecs_decode_56_amd64_of_update_zero:
+	MOVQ AX, 16(R10)
+
+	// Update match length
+	MOVQ  R8, AX
+	MOVQ  BX, CX
+	MOVQ  DX, R15
+	SHLQ  CL, R15
+	MOVB  AH, CL
+	SHRQ  $0x20, AX
+	TESTQ CX, CX
+	JZ    sequenceDecs_decode_56_amd64_ml_update_zero
+	ADDQ  CX, BX
+	CMPQ  BX, $0x40
+	JA    sequenceDecs_decode_56_amd64_ml_update_zero
+	CMPQ  CX, $0x40
+	JAE   sequenceDecs_decode_56_amd64_ml_update_zero
+	NEGQ  CX
+	SHRQ  CL, R15
+	ADDQ  R15, AX
+
+sequenceDecs_decode_56_amd64_ml_update_zero:
+	MOVQ AX, 8(R10)
+
+	// Update literal length
+	MOVQ  DI, AX
+	MOVQ  BX, CX
+	MOVQ  DX, R15
+	SHLQ  CL, R15
+	MOVB  AH, CL
+	SHRQ  $0x20, AX
+	TESTQ CX, CX
+	JZ    sequenceDecs_decode_56_amd64_ll_update_zero
+	ADDQ  CX, BX
+	CMPQ  BX, $0x40
+	JA    sequenceDecs_decode_56_amd64_ll_update_zero
+	CMPQ  CX, $0x40
+	JAE   sequenceDecs_decode_56_amd64_ll_update_zero
+	NEGQ  CX
+	SHRQ  CL, R15
+	ADDQ  R15, AX
+
+sequenceDecs_decode_56_amd64_ll_update_zero:
+	MOVQ AX, (R10)
+
+	// Fill bitreader for state updates
+	MOVQ    R14, (SP)
+	MOVQ    R9, AX
+	SHRQ    $0x08, AX
+	MOVBQZX AL, AX
+	MOVQ    ctx+16(FP), CX
+	CMPQ    96(CX), $0x00
+	JZ      sequenceDecs_decode_56_amd64_skip_update
+
+	// Update Literal Length State
+	MOVBQZX DI, R14
+	SHRL    $0x10, DI
+	LEAQ    (BX)(R14*1), CX
+	MOVQ    DX, R15
+	MOVQ    CX, BX
+	ROLQ    CL, R15
+	MOVL    $0x00000001, BP
+	MOVB    R14, CL
+	SHLL    CL, BP
+	DECL    BP
+	ANDQ    BP, R15
+	ADDQ    R15, DI
+
+	// Load ctx.llTable
+	MOVQ ctx+16(FP), CX
+	MOVQ (CX), CX
+	MOVQ (CX)(DI*8), DI
+
+	// Update Match Length State
+	MOVBQZX R8, R14
+	SHRL    $0x10, R8
+	LEAQ    (BX)(R14*1), CX
+	MOVQ    DX, R15
+	MOVQ    CX, BX
+	ROLQ    CL, R15
+	MOVL    $0x00000001, BP
+	MOVB    R14, CL
+	SHLL    CL, BP
+	DECL    BP
+	ANDQ    BP, R15
+	ADDQ    R15, R8
+
+	// Load ctx.mlTable
+	MOVQ ctx+16(FP), CX
+	MOVQ 24(CX), CX
+	MOVQ (CX)(R8*8), R8
+
+	// Update Offset State
+	MOVBQZX R9, R14
+	SHRL    $0x10, R9
+	LEAQ    (BX)(R14*1), CX
+	MOVQ    DX, R15
+	MOVQ    CX, BX
+	ROLQ    CL, R15
+	MOVL    $0x00000001, BP
+	MOVB    R14, CL
+	SHLL    CL, BP
+	DECL    BP
+	ANDQ    BP, R15
+	ADDQ    R15, R9
+
+	// Load ctx.ofTable
+	MOVQ ctx+16(FP), CX
+	MOVQ 48(CX), CX
+	MOVQ (CX)(R9*8), R9
+
+sequenceDecs_decode_56_amd64_skip_update:
+	// Adjust offset
+	MOVQ 16(R10), CX
+	CMPQ AX, $0x01
+	JBE  sequenceDecs_decode_56_amd64_adjust_offsetB_1_or_0
+	MOVQ R12, R13
+	MOVQ R11, R12
+	MOVQ CX, R11
+	JMP  sequenceDecs_decode_56_amd64_after_adjust
+
+sequenceDecs_decode_56_amd64_adjust_offsetB_1_or_0:
+	CMPQ (R10), $0x00000000
+	JNE  sequenceDecs_decode_56_amd64_adjust_offset_maybezero
+	INCQ CX
+	JMP  sequenceDecs_decode_56_amd64_adjust_offset_nonzero
+
+sequenceDecs_decode_56_amd64_adjust_offset_maybezero:
+	TESTQ CX, CX
+	JNZ   sequenceDecs_decode_56_amd64_adjust_offset_nonzero
+	MOVQ  R11, CX
+	JMP   sequenceDecs_decode_56_amd64_after_adjust
+
+sequenceDecs_decode_56_amd64_adjust_offset_nonzero:
+	CMPQ CX, $0x01
+	JB   sequenceDecs_decode_56_amd64_adjust_zero
+	JEQ  sequenceDecs_decode_56_amd64_adjust_one
+	CMPQ CX, $0x02
+	JA   sequenceDecs_decode_56_amd64_adjust_three
+	JMP  sequenceDecs_decode_56_amd64_adjust_two
+
+sequenceDecs_decode_56_amd64_adjust_zero:
+	MOVQ R11, AX
+	JMP  sequenceDecs_decode_56_amd64_adjust_test_temp_valid
+
+sequenceDecs_decode_56_amd64_adjust_one:
+	MOVQ R12, AX
+	JMP  sequenceDecs_decode_56_amd64_adjust_test_temp_valid
+
+sequenceDecs_decode_56_amd64_adjust_two:
+	MOVQ R13, AX
+	JMP  sequenceDecs_decode_56_amd64_adjust_test_temp_valid
+
+sequenceDecs_decode_56_amd64_adjust_three:
+	LEAQ -1(R11), AX
+
+sequenceDecs_decode_56_amd64_adjust_test_temp_valid:
+	TESTQ AX, AX
+	JNZ   sequenceDecs_decode_56_amd64_adjust_temp_valid
+	MOVQ  $0x00000001, AX
+
+sequenceDecs_decode_56_amd64_adjust_temp_valid:
+	CMPQ    CX, $0x01
+	CMOVQNE R12, R13
+	MOVQ    R11, R12
+	MOVQ    AX, R11
+	MOVQ    AX, CX
+
+sequenceDecs_decode_56_amd64_after_adjust:
+	MOVQ CX, 16(R10)
+
+	// Check values
+	MOVQ  8(R10), AX
+	MOVQ  (R10), R14
+	LEAQ  (AX)(R14*1), R15
+	MOVQ  s+0(FP), BP
+	ADDQ  R15, 256(BP)
+	MOVQ  ctx+16(FP), R15
+	SUBQ  R14, 128(R15)
+	JS    error_not_enough_literals
+	CMPQ  AX, $0x00020002
+	JA    sequenceDecs_decode_56_amd64_error_match_len_too_big
+	TESTQ CX, CX
+	JNZ   sequenceDecs_decode_56_amd64_match_len_ofs_ok
+	TESTQ AX, AX
+	JNZ   sequenceDecs_decode_56_amd64_error_match_len_ofs_mismatch
+
+sequenceDecs_decode_56_amd64_match_len_ofs_ok:
+	ADDQ $0x18, R10
+	MOVQ ctx+16(FP), AX
+	DECQ 96(AX)
+	JNS  sequenceDecs_decode_56_amd64_main_loop
+	MOVQ s+0(FP), AX
+	MOVQ R11, 144(AX)
+	MOVQ R12, 152(AX)
+	MOVQ R13, 160(AX)
+	MOVQ br+8(FP), AX
+	MOVQ DX, 24(AX)
+	MOVB BL, 32(AX)
+	MOVQ SI, 8(AX)
+
+	// Return success
+	MOVQ $0x00000000, ret+24(FP)
+	RET
+
+	// Return with match length error
+sequenceDecs_decode_56_amd64_error_match_len_ofs_mismatch:
+	MOVQ $0x00000001, ret+24(FP)
+	RET
+
+	// Return with match too long error
+sequenceDecs_decode_56_amd64_error_match_len_too_big:
+	MOVQ $0x00000002, ret+24(FP)
+	RET
+
+	// Return with match offset too long error
+	MOVQ $0x00000003, ret+24(FP)
+	RET
+
+	// Return with not enough literals error
+error_not_enough_literals:
+	MOVQ $0x00000004, ret+24(FP)
+	RET
+
+	// Return with overread error
+error_overread:
+	MOVQ $0x00000006, ret+24(FP)
+	RET
+
+// func sequenceDecs_decode_bmi2(s *sequenceDecs, br *bitReader, ctx *decodeAsmContext) int
+// Requires: BMI, BMI2, CMOV
+TEXT ·sequenceDecs_decode_bmi2(SB), $8-32
+	MOVQ    br+8(FP), BX
+	MOVQ    24(BX), AX
+	MOVBQZX 32(BX), DX
+	MOVQ    (BX), CX
+	MOVQ    8(BX), BX
+	ADDQ    BX, CX
+	MOVQ    CX, (SP)
+	MOVQ    ctx+16(FP), CX
+	MOVQ    72(CX), SI
+	MOVQ    80(CX), DI
+	MOVQ    88(CX), R8
+	MOVQ    104(CX), R9
+	MOVQ    s+0(FP), CX
+	MOVQ    144(CX), R10
+	MOVQ    152(CX), R11
+	MOVQ    160(CX), R12
+
+sequenceDecs_decode_bmi2_main_loop:
+	MOVQ (SP), R13
+
+	// Fill bitreader to have enough for the offset and match length.
+	CMPQ BX, $0x08
+	JL   sequenceDecs_decode_bmi2_fill_byte_by_byte
+	MOVQ DX, CX
+	SHRQ $0x03, CX
+	SUBQ CX, R13
+	MOVQ (R13), AX
+	SUBQ CX, BX
+	ANDQ $0x07, DX
+	JMP  sequenceDecs_decode_bmi2_fill_end
+
+sequenceDecs_decode_bmi2_fill_byte_by_byte:
+	CMPQ    BX, $0x00
+	JLE     sequenceDecs_decode_bmi2_fill_check_overread
+	CMPQ    DX, $0x07
+	JLE     sequenceDecs_decode_bmi2_fill_end
+	SHLQ    $0x08, AX
+	SUBQ    $0x01, R13
+	SUBQ    $0x01, BX
+	SUBQ    $0x08, DX
+	MOVBQZX (R13), CX
+	ORQ     CX, AX
+	JMP     sequenceDecs_decode_bmi2_fill_byte_by_byte
+
+sequenceDecs_decode_bmi2_fill_check_overread:
+	CMPQ DX, $0x40
+	JA   error_overread
+
+sequenceDecs_decode_bmi2_fill_end:
+	// Update offset
+	MOVQ   $0x00000808, CX
+	BEXTRQ CX, R8, R14
+	MOVQ   AX, R15
+	LEAQ   (DX)(R14*1), CX
+	ROLQ   CL, R15
+	BZHIQ  R14, R15, R15
+	MOVQ   CX, DX
+	MOVQ   R8, CX
+	SHRQ   $0x20, CX
+	ADDQ   R15, CX
+	MOVQ   CX, 16(R9)
+
+	// Update match length
+	MOVQ   $0x00000808, CX
+	BEXTRQ CX, DI, R14
+	MOVQ   AX, R15
+	LEAQ   (DX)(R14*1), CX
+	ROLQ   CL, R15
+	BZHIQ  R14, R15, R15
+	MOVQ   CX, DX
+	MOVQ   DI, CX
+	SHRQ   $0x20, CX
+	ADDQ   R15, CX
+	MOVQ   CX, 8(R9)
+
+	// Fill bitreader to have enough for the remaining
+	CMPQ BX, $0x08
+	JL   sequenceDecs_decode_bmi2_fill_2_byte_by_byte
+	MOVQ DX, CX
+	SHRQ $0x03, CX
+	SUBQ CX, R13
+	MOVQ (R13), AX
+	SUBQ CX, BX
+	ANDQ $0x07, DX
+	JMP  sequenceDecs_decode_bmi2_fill_2_end
+
+sequenceDecs_decode_bmi2_fill_2_byte_by_byte:
+	CMPQ    BX, $0x00
+	JLE     sequenceDecs_decode_bmi2_fill_2_check_overread
+	CMPQ    DX, $0x07
+	JLE     sequenceDecs_decode_bmi2_fill_2_end
+	SHLQ    $0x08, AX
+	SUBQ    $0x01, R13
+	SUBQ    $0x01, BX
+	SUBQ    $0x08, DX
+	MOVBQZX (R13), CX
+	ORQ     CX, AX
+	JMP     sequenceDecs_decode_bmi2_fill_2_byte_by_byte
+
+sequenceDecs_decode_bmi2_fill_2_check_overread:
+	CMPQ DX, $0x40
+	JA   error_overread
+
+sequenceDecs_decode_bmi2_fill_2_end:
+	// Update literal length
+	MOVQ   $0x00000808, CX
+	BEXTRQ CX, SI, R14
+	MOVQ   AX, R15
+	LEAQ   (DX)(R14*1), CX
+	ROLQ   CL, R15
+	BZHIQ  R14, R15, R15
+	MOVQ   CX, DX
+	MOVQ   SI, CX
+	SHRQ   $0x20, CX
+	ADDQ   R15, CX
+	MOVQ   CX, (R9)
+
+	// Fill bitreader for state updates
+	MOVQ    R13, (SP)
+	MOVQ    $0x00000808, CX
+	BEXTRQ  CX, R8, R13
+	MOVQ    ctx+16(FP), CX
+	CMPQ    96(CX), $0x00
+	JZ      sequenceDecs_decode_bmi2_skip_update
+	LEAQ    (SI)(DI*1), R14
+	ADDQ    R8, R14
+	MOVBQZX R14, R14
+	LEAQ    (DX)(R14*1), CX
+	MOVQ    AX, R15
+	MOVQ    CX, DX
+	ROLQ    CL, R15
+	BZHIQ   R14, R15, R15
+
+	// Update Offset State
+	BZHIQ R8, R15, CX
+	SHRXQ R8, R15, R15
+	SHRL  $0x10, R8
+	ADDQ  CX, R8
+
+	// Load ctx.ofTable
+	MOVQ ctx+16(FP), CX
+	MOVQ 48(CX), CX
+	MOVQ (CX)(R8*8), R8
+
+	// Update Match Length State
+	BZHIQ DI, R15, CX
+	SHRXQ DI, R15, R15
+	SHRL  $0x10, DI
+	ADDQ  CX, DI
+
+	// Load ctx.mlTable
+	MOVQ ctx+16(FP), CX
+	MOVQ 24(CX), CX
+	MOVQ (CX)(DI*8), DI
+
+	// Update Literal Length State
+	BZHIQ SI, R15, CX
+	SHRL  $0x10, SI
+	ADDQ  CX, SI
+
+	// Load ctx.llTable
+	MOVQ ctx+16(FP), CX
+	MOVQ (CX), CX
+	MOVQ (CX)(SI*8), SI
+
+sequenceDecs_decode_bmi2_skip_update:
+	// Adjust offset
+	MOVQ 16(R9), CX
+	CMPQ R13, $0x01
+	JBE  sequenceDecs_decode_bmi2_adjust_offsetB_1_or_0
+	MOVQ R11, R12
+	MOVQ R10, R11
+	MOVQ CX, R10
+	JMP  sequenceDecs_decode_bmi2_after_adjust
+
+sequenceDecs_decode_bmi2_adjust_offsetB_1_or_0:
+	CMPQ (R9), $0x00000000
+	JNE  sequenceDecs_decode_bmi2_adjust_offset_maybezero
+	INCQ CX
+	JMP  sequenceDecs_decode_bmi2_adjust_offset_nonzero
+
+sequenceDecs_decode_bmi2_adjust_offset_maybezero:
+	TESTQ CX, CX
+	JNZ   sequenceDecs_decode_bmi2_adjust_offset_nonzero
+	MOVQ  R10, CX
+	JMP   sequenceDecs_decode_bmi2_after_adjust
+
+sequenceDecs_decode_bmi2_adjust_offset_nonzero:
+	CMPQ CX, $0x01
+	JB   sequenceDecs_decode_bmi2_adjust_zero
+	JEQ  sequenceDecs_decode_bmi2_adjust_one
+	CMPQ CX, $0x02
+	JA   sequenceDecs_decode_bmi2_adjust_three
+	JMP  sequenceDecs_decode_bmi2_adjust_two
+
+sequenceDecs_decode_bmi2_adjust_zero:
+	MOVQ R10, R13
+	JMP  sequenceDecs_decode_bmi2_adjust_test_temp_valid
+
+sequenceDecs_decode_bmi2_adjust_one:
+	MOVQ R11, R13
+	JMP  sequenceDecs_decode_bmi2_adjust_test_temp_valid
+
+sequenceDecs_decode_bmi2_adjust_two:
+	MOVQ R12, R13
+	JMP  sequenceDecs_decode_bmi2_adjust_test_temp_valid
+
+sequenceDecs_decode_bmi2_adjust_three:
+	LEAQ -1(R10), R13
+
+sequenceDecs_decode_bmi2_adjust_test_temp_valid:
+	TESTQ R13, R13
+	JNZ   sequenceDecs_decode_bmi2_adjust_temp_valid
+	MOVQ  $0x00000001, R13
+
+sequenceDecs_decode_bmi2_adjust_temp_valid:
+	CMPQ    CX, $0x01
+	CMOVQNE R11, R12
+	MOVQ    R10, R11
+	MOVQ    R13, R10
+	MOVQ    R13, CX
+
+sequenceDecs_decode_bmi2_after_adjust:
+	MOVQ CX, 16(R9)
+
+	// Check values
+	MOVQ  8(R9), R13
+	MOVQ  (R9), R14
+	LEAQ  (R13)(R14*1), R15
+	MOVQ  s+0(FP), BP
+	ADDQ  R15, 256(BP)
+	MOVQ  ctx+16(FP), R15
+	SUBQ  R14, 128(R15)
+	JS    error_not_enough_literals
+	CMPQ  R13, $0x00020002
+	JA    sequenceDecs_decode_bmi2_error_match_len_too_big
+	TESTQ CX, CX
+	JNZ   sequenceDecs_decode_bmi2_match_len_ofs_ok
+	TESTQ R13, R13
+	JNZ   sequenceDecs_decode_bmi2_error_match_len_ofs_mismatch
+
+sequenceDecs_decode_bmi2_match_len_ofs_ok:
+	ADDQ $0x18, R9
+	MOVQ ctx+16(FP), CX
+	DECQ 96(CX)
+	JNS  sequenceDecs_decode_bmi2_main_loop
+	MOVQ s+0(FP), CX
+	MOVQ R10, 144(CX)
+	MOVQ R11, 152(CX)
+	MOVQ R12, 160(CX)
+	MOVQ br+8(FP), CX
+	MOVQ AX, 24(CX)
+	MOVB DL, 32(CX)
+	MOVQ BX, 8(CX)
+
+	// Return success
+	MOVQ $0x00000000, ret+24(FP)
+	RET
+
+	// Return with match length error
+sequenceDecs_decode_bmi2_error_match_len_ofs_mismatch:
+	MOVQ $0x00000001, ret+24(FP)
+	RET
+
+	// Return with match too long error
+sequenceDecs_decode_bmi2_error_match_len_too_big:
+	MOVQ $0x00000002, ret+24(FP)
+	RET
+
+	// Return with match offset too long error
+	MOVQ $0x00000003, ret+24(FP)
+	RET
+
+	// Return with not enough literals error
+error_not_enough_literals:
+	MOVQ $0x00000004, ret+24(FP)
+	RET
+
+	// Return with overread error
+error_overread:
+	MOVQ $0x00000006, ret+24(FP)
+	RET
+
+// func sequenceDecs_decode_56_bmi2(s *sequenceDecs, br *bitReader, ctx *decodeAsmContext) int
+// Requires: BMI, BMI2, CMOV
+TEXT ·sequenceDecs_decode_56_bmi2(SB), $8-32
+	MOVQ    br+8(FP), BX
+	MOVQ    24(BX), AX
+	MOVBQZX 32(BX), DX
+	MOVQ    (BX), CX
+	MOVQ    8(BX), BX
+	ADDQ    BX, CX
+	MOVQ    CX, (SP)
+	MOVQ    ctx+16(FP), CX
+	MOVQ    72(CX), SI
+	MOVQ    80(CX), DI
+	MOVQ    88(CX), R8
+	MOVQ    104(CX), R9
+	MOVQ    s+0(FP), CX
+	MOVQ    144(CX), R10
+	MOVQ    152(CX), R11
+	MOVQ    160(CX), R12
+
+sequenceDecs_decode_56_bmi2_main_loop:
+	MOVQ (SP), R13
+
+	// Fill bitreader to have enough for the offset and match length.
+	CMPQ BX, $0x08
+	JL   sequenceDecs_decode_56_bmi2_fill_byte_by_byte
+	MOVQ DX, CX
+	SHRQ $0x03, CX
+	SUBQ CX, R13
+	MOVQ (R13), AX
+	SUBQ CX, BX
+	ANDQ $0x07, DX
+	JMP  sequenceDecs_decode_56_bmi2_fill_end
+
+sequenceDecs_decode_56_bmi2_fill_byte_by_byte:
+	CMPQ    BX, $0x00
+	JLE     sequenceDecs_decode_56_bmi2_fill_check_overread
+	CMPQ    DX, $0x07
+	JLE     sequenceDecs_decode_56_bmi2_fill_end
+	SHLQ    $0x08, AX
+	SUBQ    $0x01, R13
+	SUBQ    $0x01, BX
+	SUBQ    $0x08, DX
+	MOVBQZX (R13), CX
+	ORQ     CX, AX
+	JMP     sequenceDecs_decode_56_bmi2_fill_byte_by_byte
+
+sequenceDecs_decode_56_bmi2_fill_check_overread:
+	CMPQ DX, $0x40
+	JA   error_overread
+
+sequenceDecs_decode_56_bmi2_fill_end:
+	// Update offset
+	MOVQ   $0x00000808, CX
+	BEXTRQ CX, R8, R14
+	MOVQ   AX, R15
+	LEAQ   (DX)(R14*1), CX
+	ROLQ   CL, R15
+	BZHIQ  R14, R15, R15
+	MOVQ   CX, DX
+	MOVQ   R8, CX
+	SHRQ   $0x20, CX
+	ADDQ   R15, CX
+	MOVQ   CX, 16(R9)
+
+	// Update match length
+	MOVQ   $0x00000808, CX
+	BEXTRQ CX, DI, R14
+	MOVQ   AX, R15
+	LEAQ   (DX)(R14*1), CX
+	ROLQ   CL, R15
+	BZHIQ  R14, R15, R15
+	MOVQ   CX, DX
+	MOVQ   DI, CX
+	SHRQ   $0x20, CX
+	ADDQ   R15, CX
+	MOVQ   CX, 8(R9)
+
+	// Update literal length
+	MOVQ   $0x00000808, CX
+	BEXTRQ CX, SI, R14
+	MOVQ   AX, R15
+	LEAQ   (DX)(R14*1), CX
+	ROLQ   CL, R15
+	BZHIQ  R14, R15, R15
+	MOVQ   CX, DX
+	MOVQ   SI, CX
+	SHRQ   $0x20, CX
+	ADDQ   R15, CX
+	MOVQ   CX, (R9)
+
+	// Fill bitreader for state updates
+	MOVQ    R13, (SP)
+	MOVQ    $0x00000808, CX
+	BEXTRQ  CX, R8, R13
+	MOVQ    ctx+16(FP), CX
+	CMPQ    96(CX), $0x00
+	JZ      sequenceDecs_decode_56_bmi2_skip_update
+	LEAQ    (SI)(DI*1), R14
+	ADDQ    R8, R14
+	MOVBQZX R14, R14
+	LEAQ    (DX)(R14*1), CX
+	MOVQ    AX, R15
+	MOVQ    CX, DX
+	ROLQ    CL, R15
+	BZHIQ   R14, R15, R15
+
+	// Update Offset State
+	BZHIQ R8, R15, CX
+	SHRXQ R8, R15, R15
+	SHRL  $0x10, R8
+	ADDQ  CX, R8
+
+	// Load ctx.ofTable
+	MOVQ ctx+16(FP), CX
+	MOVQ 48(CX), CX
+	MOVQ (CX)(R8*8), R8
+
+	// Update Match Length State
+	BZHIQ DI, R15, CX
+	SHRXQ DI, R15, R15
+	SHRL  $0x10, DI
+	ADDQ  CX, DI
+
+	// Load ctx.mlTable
+	MOVQ ctx+16(FP), CX
+	MOVQ 24(CX), CX
+	MOVQ (CX)(DI*8), DI
+
+	// Update Literal Length State
+	BZHIQ SI, R15, CX
+	SHRL  $0x10, SI
+	ADDQ  CX, SI
+
+	// Load ctx.llTable
+	MOVQ ctx+16(FP), CX
+	MOVQ (CX), CX
+	MOVQ (CX)(SI*8), SI
+
+sequenceDecs_decode_56_bmi2_skip_update:
+	// Adjust offset
+	MOVQ 16(R9), CX
+	CMPQ R13, $0x01
+	JBE  sequenceDecs_decode_56_bmi2_adjust_offsetB_1_or_0
+	MOVQ R11, R12
+	MOVQ R10, R11
+	MOVQ CX, R10
+	JMP  sequenceDecs_decode_56_bmi2_after_adjust
+
+sequenceDecs_decode_56_bmi2_adjust_offsetB_1_or_0:
+	CMPQ (R9), $0x00000000
+	JNE  sequenceDecs_decode_56_bmi2_adjust_offset_maybezero
+	INCQ CX
+	JMP  sequenceDecs_decode_56_bmi2_adjust_offset_nonzero
+
+sequenceDecs_decode_56_bmi2_adjust_offset_maybezero:
+	TESTQ CX, CX
+	JNZ   sequenceDecs_decode_56_bmi2_adjust_offset_nonzero
+	MOVQ  R10, CX
+	JMP   sequenceDecs_decode_56_bmi2_after_adjust
+
+sequenceDecs_decode_56_bmi2_adjust_offset_nonzero:
+	CMPQ CX, $0x01
+	JB   sequenceDecs_decode_56_bmi2_adjust_zero
+	JEQ  sequenceDecs_decode_56_bmi2_adjust_one
+	CMPQ CX, $0x02
+	JA   sequenceDecs_decode_56_bmi2_adjust_three
+	JMP  sequenceDecs_decode_56_bmi2_adjust_two
+
+sequenceDecs_decode_56_bmi2_adjust_zero:
+	MOVQ R10, R13
+	JMP  sequenceDecs_decode_56_bmi2_adjust_test_temp_valid
+
+sequenceDecs_decode_56_bmi2_adjust_one:
+	MOVQ R11, R13
+	JMP  sequenceDecs_decode_56_bmi2_adjust_test_temp_valid
+
+sequenceDecs_decode_56_bmi2_adjust_two:
+	MOVQ R12, R13
+	JMP  sequenceDecs_decode_56_bmi2_adjust_test_temp_valid
+
+sequenceDecs_decode_56_bmi2_adjust_three:
+	LEAQ -1(R10), R13
+
+sequenceDecs_decode_56_bmi2_adjust_test_temp_valid:
+	TESTQ R13, R13
+	JNZ   sequenceDecs_decode_56_bmi2_adjust_temp_valid
+	MOVQ  $0x00000001, R13
+
+sequenceDecs_decode_56_bmi2_adjust_temp_valid:
+	CMPQ    CX, $0x01
+	CMOVQNE R11, R12
+	MOVQ    R10, R11
+	MOVQ    R13, R10
+	MOVQ    R13, CX
+
+sequenceDecs_decode_56_bmi2_after_adjust:
+	MOVQ CX, 16(R9)
+
+	// Check values
+	MOVQ  8(R9), R13
+	MOVQ  (R9), R14
+	LEAQ  (R13)(R14*1), R15
+	MOVQ  s+0(FP), BP
+	ADDQ  R15, 256(BP)
+	MOVQ  ctx+16(FP), R15
+	SUBQ  R14, 128(R15)
+	JS    error_not_enough_literals
+	CMPQ  R13, $0x00020002
+	JA    sequenceDecs_decode_56_bmi2_error_match_len_too_big
+	TESTQ CX, CX
+	JNZ   sequenceDecs_decode_56_bmi2_match_len_ofs_ok
+	TESTQ R13, R13
+	JNZ   sequenceDecs_decode_56_bmi2_error_match_len_ofs_mismatch
+
+sequenceDecs_decode_56_bmi2_match_len_ofs_ok:
+	ADDQ $0x18, R9
+	MOVQ ctx+16(FP), CX
+	DECQ 96(CX)
+	JNS  sequenceDecs_decode_56_bmi2_main_loop
+	MOVQ s+0(FP), CX
+	MOVQ R10, 144(CX)
+	MOVQ R11, 152(CX)
+	MOVQ R12, 160(CX)
+	MOVQ br+8(FP), CX
+	MOVQ AX, 24(CX)
+	MOVB DL, 32(CX)
+	MOVQ BX, 8(CX)
+
+	// Return success
+	MOVQ $0x00000000, ret+24(FP)
+	RET
+
+	// Return with match length error
+sequenceDecs_decode_56_bmi2_error_match_len_ofs_mismatch:
+	MOVQ $0x00000001, ret+24(FP)
+	RET
+
+	// Return with match too long error
+sequenceDecs_decode_56_bmi2_error_match_len_too_big:
+	MOVQ $0x00000002, ret+24(FP)
+	RET
+
+	// Return with match offset too long error
+	MOVQ $0x00000003, ret+24(FP)
+	RET
+
+	// Return with not enough literals error
+error_not_enough_literals:
+	MOVQ $0x00000004, ret+24(FP)
+	RET
+
+	// Return with overread error
+error_overread:
+	MOVQ $0x00000006, ret+24(FP)
+	RET
+
+// func sequenceDecs_executeSimple_amd64(ctx *executeAsmContext) bool
+// Requires: SSE
+TEXT ·sequenceDecs_executeSimple_amd64(SB), $8-9
+	MOVQ  ctx+0(FP), R10
+	MOVQ  8(R10), CX
+	TESTQ CX, CX
+	JZ    empty_seqs
+	MOVQ  (R10), AX
+	MOVQ  24(R10), DX
+	MOVQ  32(R10), BX
+	MOVQ  80(R10), SI
+	MOVQ  104(R10), DI
+	MOVQ  120(R10), R8
+	MOVQ  56(R10), R9
+	MOVQ  64(R10), R10
+	ADDQ  R10, R9
+
+	// seqsBase += 24 * seqIndex
+	LEAQ (DX)(DX*2), R11
+	SHLQ $0x03, R11
+	ADDQ R11, AX
+
+	// outBase += outPosition
+	ADDQ DI, BX
+
+main_loop:
+	MOVQ (AX), R11
+	MOVQ 16(AX), R12
+	MOVQ 8(AX), R13
+
+	// Copy literals
+	TESTQ R11, R11
+	JZ    check_offset
+	XORQ  R14, R14
+
+copy_1:
+	MOVUPS (SI)(R14*1), X0
+	MOVUPS X0, (BX)(R14*1)
+	ADDQ   $0x10, R14
+	CMPQ   R14, R11
+	JB     copy_1
+	ADDQ   R11, SI
+	ADDQ   R11, BX
+	ADDQ   R11, DI
+
+	// Malformed input if seq.mo > t+len(hist) || seq.mo > s.windowSize)
+check_offset:
+	LEAQ (DI)(R10*1), R11
+	CMPQ R12, R11
+	JG   error_match_off_too_big
+	CMPQ R12, R8
+	JG   error_match_off_too_big
+
+	// Copy match from history
+	MOVQ R12, R11
+	SUBQ DI, R11
+	JLS  copy_match
+	MOVQ R9, R14
+	SUBQ R11, R14
+	CMPQ R13, R11
+	JG   copy_all_from_history
+	MOVQ R13, R11
+	SUBQ $0x10, R11
+	JB   copy_4_small
+
+copy_4_loop:
+	MOVUPS (R14), X0
+	MOVUPS X0, (BX)
+	ADDQ   $0x10, R14
+	ADDQ   $0x10, BX
+	SUBQ   $0x10, R11
+	JAE    copy_4_loop
+	LEAQ   16(R14)(R11*1), R14
+	LEAQ   16(BX)(R11*1), BX
+	MOVUPS -16(R14), X0
+	MOVUPS X0, -16(BX)
+	JMP    copy_4_end
+
+copy_4_small:
+	CMPQ R13, $0x03
+	JE   copy_4_move_3
+	CMPQ R13, $0x08
+	JB   copy_4_move_4through7
+	JMP  copy_4_move_8through16
+
+copy_4_move_3:
+	MOVW (R14), R11
+	MOVB 2(R14), R12
+	MOVW R11, (BX)
+	MOVB R12, 2(BX)
+	ADDQ R13, R14
+	ADDQ R13, BX
+	JMP  copy_4_end
+
+copy_4_move_4through7:
+	MOVL (R14), R11
+	MOVL -4(R14)(R13*1), R12
+	MOVL R11, (BX)
+	MOVL R12, -4(BX)(R13*1)
+	ADDQ R13, R14
+	ADDQ R13, BX
+	JMP  copy_4_end
+
+copy_4_move_8through16:
+	MOVQ (R14), R11
+	MOVQ -8(R14)(R13*1), R12
+	MOVQ R11, (BX)
+	MOVQ R12, -8(BX)(R13*1)
+	ADDQ R13, R14
+	ADDQ R13, BX
+
+copy_4_end:
+	ADDQ R13, DI
+	ADDQ $0x18, AX
+	INCQ DX
+	CMPQ DX, CX
+	JB   main_loop
+	JMP  loop_finished
+
+copy_all_from_history:
+	MOVQ R11, R15
+	SUBQ $0x10, R15
+	JB   copy_5_small
+
+copy_5_loop:
+	MOVUPS (R14), X0
+	MOVUPS X0, (BX)
+	ADDQ   $0x10, R14
+	ADDQ   $0x10, BX
+	SUBQ   $0x10, R15
+	JAE    copy_5_loop
+	LEAQ   16(R14)(R15*1), R14
+	LEAQ   16(BX)(R15*1), BX
+	MOVUPS -16(R14), X0
+	MOVUPS X0, -16(BX)
+	JMP    copy_5_end
+
+copy_5_small:
+	CMPQ R11, $0x03
+	JE   copy_5_move_3
+	JB   copy_5_move_1or2
+	CMPQ R11, $0x08
+	JB   copy_5_move_4through7
+	JMP  copy_5_move_8through16
+
+copy_5_move_1or2:
+	MOVB (R14), R15
+	MOVB -1(R14)(R11*1), BP
+	MOVB R15, (BX)
+	MOVB BP, -1(BX)(R11*1)
+	ADDQ R11, R14
+	ADDQ R11, BX
+	JMP  copy_5_end
+
+copy_5_move_3:
+	MOVW (R14), R15
+	MOVB 2(R14), BP
+	MOVW R15, (BX)
+	MOVB BP, 2(BX)
+	ADDQ R11, R14
+	ADDQ R11, BX
+	JMP  copy_5_end
+
+copy_5_move_4through7:
+	MOVL (R14), R15
+	MOVL -4(R14)(R11*1), BP
+	MOVL R15, (BX)
+	MOVL BP, -4(BX)(R11*1)
+	ADDQ R11, R14
+	ADDQ R11, BX
+	JMP  copy_5_end
+
+copy_5_move_8through16:
+	MOVQ (R14), R15
+	MOVQ -8(R14)(R11*1), BP
+	MOVQ R15, (BX)
+	MOVQ BP, -8(BX)(R11*1)
+	ADDQ R11, R14
+	ADDQ R11, BX
+
+copy_5_end:
+	ADDQ R11, DI
+	SUBQ R11, R13
+
+	// Copy match from the current buffer
+copy_match:
+	MOVQ BX, R11
+	SUBQ R12, R11
+
+	// ml <= mo
+	CMPQ R13, R12
+	JA   copy_overlapping_match
+
+	// Copy non-overlapping match
+	ADDQ R13, DI
+	MOVQ BX, R12
+	ADDQ R13, BX
+
+copy_2:
+	MOVUPS (R11), X0
+	MOVUPS X0, (R12)
+	ADDQ   $0x10, R11
+	ADDQ   $0x10, R12
+	SUBQ   $0x10, R13
+	JHI    copy_2
+	JMP    handle_loop
+
+	// Copy overlapping match
+copy_overlapping_match:
+	ADDQ R13, DI
+
+copy_slow_3:
+	MOVB (R11), R12
+	MOVB R12, (BX)
+	INCQ R11
+	INCQ BX
+	DECQ R13
+	JNZ  copy_slow_3
+
+handle_loop:
+	ADDQ $0x18, AX
+	INCQ DX
+	CMPQ DX, CX
+	JB   main_loop
+
+loop_finished:
+	// Return value
+	MOVB $0x01, ret+8(FP)
+
+	// Update the context
+	MOVQ ctx+0(FP), AX
+	MOVQ DX, 24(AX)
+	MOVQ DI, 104(AX)
+	SUBQ 80(AX), SI
+	MOVQ SI, 112(AX)
+	RET
+
+error_match_off_too_big:
+	// Return value
+	MOVB $0x00, ret+8(FP)
+
+	// Update the context
+	MOVQ ctx+0(FP), AX
+	MOVQ DX, 24(AX)
+	MOVQ DI, 104(AX)
+	SUBQ 80(AX), SI
+	MOVQ SI, 112(AX)
+	RET
+
+empty_seqs:
+	// Return value
+	MOVB $0x01, ret+8(FP)
+	RET
+
+// func sequenceDecs_executeSimple_safe_amd64(ctx *executeAsmContext) bool
+// Requires: SSE
+TEXT ·sequenceDecs_executeSimple_safe_amd64(SB), $8-9
+	MOVQ  ctx+0(FP), R10
+	MOVQ  8(R10), CX
+	TESTQ CX, CX
+	JZ    empty_seqs
+	MOVQ  (R10), AX
+	MOVQ  24(R10), DX
+	MOVQ  32(R10), BX
+	MOVQ  80(R10), SI
+	MOVQ  104(R10), DI
+	MOVQ  120(R10), R8
+	MOVQ  56(R10), R9
+	MOVQ  64(R10), R10
+	ADDQ  R10, R9
+
+	// seqsBase += 24 * seqIndex
+	LEAQ (DX)(DX*2), R11
+	SHLQ $0x03, R11
+	ADDQ R11, AX
+
+	// outBase += outPosition
+	ADDQ DI, BX
+
+main_loop:
+	MOVQ (AX), R11
+	MOVQ 16(AX), R12
+	MOVQ 8(AX), R13
+
+	// Copy literals
+	TESTQ R11, R11
+	JZ    check_offset
+	MOVQ  R11, R14
+	SUBQ  $0x10, R14
+	JB    copy_1_small
+
+copy_1_loop:
+	MOVUPS (SI), X0
+	MOVUPS X0, (BX)
+	ADDQ   $0x10, SI
+	ADDQ   $0x10, BX
+	SUBQ   $0x10, R14
+	JAE    copy_1_loop
+	LEAQ   16(SI)(R14*1), SI
+	LEAQ   16(BX)(R14*1), BX
+	MOVUPS -16(SI), X0
+	MOVUPS X0, -16(BX)
+	JMP    copy_1_end
+
+copy_1_small:
+	CMPQ R11, $0x03
+	JE   copy_1_move_3
+	JB   copy_1_move_1or2
+	CMPQ R11, $0x08
+	JB   copy_1_move_4through7
+	JMP  copy_1_move_8through16
+
+copy_1_move_1or2:
+	MOVB (SI), R14
+	MOVB -1(SI)(R11*1), R15
+	MOVB R14, (BX)
+	MOVB R15, -1(BX)(R11*1)
+	ADDQ R11, SI
+	ADDQ R11, BX
+	JMP  copy_1_end
+
+copy_1_move_3:
+	MOVW (SI), R14
+	MOVB 2(SI), R15
+	MOVW R14, (BX)
+	MOVB R15, 2(BX)
+	ADDQ R11, SI
+	ADDQ R11, BX
+	JMP  copy_1_end
+
+copy_1_move_4through7:
+	MOVL (SI), R14
+	MOVL -4(SI)(R11*1), R15
+	MOVL R14, (BX)
+	MOVL R15, -4(BX)(R11*1)
+	ADDQ R11, SI
+	ADDQ R11, BX
+	JMP  copy_1_end
+
+copy_1_move_8through16:
+	MOVQ (SI), R14
+	MOVQ -8(SI)(R11*1), R15
+	MOVQ R14, (BX)
+	MOVQ R15, -8(BX)(R11*1)
+	ADDQ R11, SI
+	ADDQ R11, BX
+
+copy_1_end:
+	ADDQ R11, DI
+
+	// Malformed input if seq.mo > t+len(hist) || seq.mo > s.windowSize)
+check_offset:
+	LEAQ (DI)(R10*1), R11
+	CMPQ R12, R11
+	JG   error_match_off_too_big
+	CMPQ R12, R8
+	JG   error_match_off_too_big
+
+	// Copy match from history
+	MOVQ R12, R11
+	SUBQ DI, R11
+	JLS  copy_match
+	MOVQ R9, R14
+	SUBQ R11, R14
+	CMPQ R13, R11
+	JG   copy_all_from_history
+	MOVQ R13, R11
+	SUBQ $0x10, R11
+	JB   copy_4_small
+
+copy_4_loop:
+	MOVUPS (R14), X0
+	MOVUPS X0, (BX)
+	ADDQ   $0x10, R14
+	ADDQ   $0x10, BX
+	SUBQ   $0x10, R11
+	JAE    copy_4_loop
+	LEAQ   16(R14)(R11*1), R14
+	LEAQ   16(BX)(R11*1), BX
+	MOVUPS -16(R14), X0
+	MOVUPS X0, -16(BX)
+	JMP    copy_4_end
+
+copy_4_small:
+	CMPQ R13, $0x03
+	JE   copy_4_move_3
+	CMPQ R13, $0x08
+	JB   copy_4_move_4through7
+	JMP  copy_4_move_8through16
+
+copy_4_move_3:
+	MOVW (R14), R11
+	MOVB 2(R14), R12
+	MOVW R11, (BX)
+	MOVB R12, 2(BX)
+	ADDQ R13, R14
+	ADDQ R13, BX
+	JMP  copy_4_end
+
+copy_4_move_4through7:
+	MOVL (R14), R11
+	MOVL -4(R14)(R13*1), R12
+	MOVL R11, (BX)
+	MOVL R12, -4(BX)(R13*1)
+	ADDQ R13, R14
+	ADDQ R13, BX
+	JMP  copy_4_end
+
+copy_4_move_8through16:
+	MOVQ (R14), R11
+	MOVQ -8(R14)(R13*1), R12
+	MOVQ R11, (BX)
+	MOVQ R12, -8(BX)(R13*1)
+	ADDQ R13, R14
+	ADDQ R13, BX
+
+copy_4_end:
+	ADDQ R13, DI
+	ADDQ $0x18, AX
+	INCQ DX
+	CMPQ DX, CX
+	JB   main_loop
+	JMP  loop_finished
+
+copy_all_from_history:
+	MOVQ R11, R15
+	SUBQ $0x10, R15
+	JB   copy_5_small
+
+copy_5_loop:
+	MOVUPS (R14), X0
+	MOVUPS X0, (BX)
+	ADDQ   $0x10, R14
+	ADDQ   $0x10, BX
+	SUBQ   $0x10, R15
+	JAE    copy_5_loop
+	LEAQ   16(R14)(R15*1), R14
+	LEAQ   16(BX)(R15*1), BX
+	MOVUPS -16(R14), X0
+	MOVUPS X0, -16(BX)
+	JMP    copy_5_end
+
+copy_5_small:
+	CMPQ R11, $0x03
+	JE   copy_5_move_3
+	JB   copy_5_move_1or2
+	CMPQ R11, $0x08
+	JB   copy_5_move_4through7
+	JMP  copy_5_move_8through16
+
+copy_5_move_1or2:
+	MOVB (R14), R15
+	MOVB -1(R14)(R11*1), BP
+	MOVB R15, (BX)
+	MOVB BP, -1(BX)(R11*1)
+	ADDQ R11, R14
+	ADDQ R11, BX
+	JMP  copy_5_end
+
+copy_5_move_3:
+	MOVW (R14), R15
+	MOVB 2(R14), BP
+	MOVW R15, (BX)
+	MOVB BP, 2(BX)
+	ADDQ R11, R14
+	ADDQ R11, BX
+	JMP  copy_5_end
+
+copy_5_move_4through7:
+	MOVL (R14), R15
+	MOVL -4(R14)(R11*1), BP
+	MOVL R15, (BX)
+	MOVL BP, -4(BX)(R11*1)
+	ADDQ R11, R14
+	ADDQ R11, BX
+	JMP  copy_5_end
+
+copy_5_move_8through16:
+	MOVQ (R14), R15
+	MOVQ -8(R14)(R11*1), BP
+	MOVQ R15, (BX)
+	MOVQ BP, -8(BX)(R11*1)
+	ADDQ R11, R14
+	ADDQ R11, BX
+
+copy_5_end:
+	ADDQ R11, DI
+	SUBQ R11, R13
+
+	// Copy match from the current buffer
+copy_match:
+	MOVQ BX, R11
+	SUBQ R12, R11
+
+	// ml <= mo
+	CMPQ R13, R12
+	JA   copy_overlapping_match
+
+	// Copy non-overlapping match
+	ADDQ R13, DI
+	MOVQ R13, R12
+	SUBQ $0x10, R12
+	JB   copy_2_small
+
+copy_2_loop:
+	MOVUPS (R11), X0
+	MOVUPS X0, (BX)
+	ADDQ   $0x10, R11
+	ADDQ   $0x10, BX
+	SUBQ   $0x10, R12
+	JAE    copy_2_loop
+	LEAQ   16(R11)(R12*1), R11
+	LEAQ   16(BX)(R12*1), BX
+	MOVUPS -16(R11), X0
+	MOVUPS X0, -16(BX)
+	JMP    copy_2_end
+
+copy_2_small:
+	CMPQ R13, $0x03
+	JE   copy_2_move_3
+	JB   copy_2_move_1or2
+	CMPQ R13, $0x08
+	JB   copy_2_move_4through7
+	JMP  copy_2_move_8through16
+
+copy_2_move_1or2:
+	MOVB (R11), R12
+	MOVB -1(R11)(R13*1), R14
+	MOVB R12, (BX)
+	MOVB R14, -1(BX)(R13*1)
+	ADDQ R13, R11
+	ADDQ R13, BX
+	JMP  copy_2_end
+
+copy_2_move_3:
+	MOVW (R11), R12
+	MOVB 2(R11), R14
+	MOVW R12, (BX)
+	MOVB R14, 2(BX)
+	ADDQ R13, R11
+	ADDQ R13, BX
+	JMP  copy_2_end
+
+copy_2_move_4through7:
+	MOVL (R11), R12
+	MOVL -4(R11)(R13*1), R14
+	MOVL R12, (BX)
+	MOVL R14, -4(BX)(R13*1)
+	ADDQ R13, R11
+	ADDQ R13, BX
+	JMP  copy_2_end
+
+copy_2_move_8through16:
+	MOVQ (R11), R12
+	MOVQ -8(R11)(R13*1), R14
+	MOVQ R12, (BX)
+	MOVQ R14, -8(BX)(R13*1)
+	ADDQ R13, R11
+	ADDQ R13, BX
+
+copy_2_end:
+	JMP handle_loop
+
+	// Copy overlapping match
+copy_overlapping_match:
+	ADDQ R13, DI
+
+copy_slow_3:
+	MOVB (R11), R12
+	MOVB R12, (BX)
+	INCQ R11
+	INCQ BX
+	DECQ R13
+	JNZ  copy_slow_3
+
+handle_loop:
+	ADDQ $0x18, AX
+	INCQ DX
+	CMPQ DX, CX
+	JB   main_loop
+
+loop_finished:
+	// Return value
+	MOVB $0x01, ret+8(FP)
+
+	// Update the context
+	MOVQ ctx+0(FP), AX
+	MOVQ DX, 24(AX)
+	MOVQ DI, 104(AX)
+	SUBQ 80(AX), SI
+	MOVQ SI, 112(AX)
+	RET
+
+error_match_off_too_big:
+	// Return value
+	MOVB $0x00, ret+8(FP)
+
+	// Update the context
+	MOVQ ctx+0(FP), AX
+	MOVQ DX, 24(AX)
+	MOVQ DI, 104(AX)
+	SUBQ 80(AX), SI
+	MOVQ SI, 112(AX)
+	RET
+
+empty_seqs:
+	// Return value
+	MOVB $0x01, ret+8(FP)
+	RET
+
+// func sequenceDecs_decodeSync_amd64(s *sequenceDecs, br *bitReader, ctx *decodeSyncAsmContext) int
+// Requires: CMOV, SSE
+TEXT ·sequenceDecs_decodeSync_amd64(SB), $64-32
+	MOVQ    br+8(FP), CX
+	MOVQ    24(CX), DX
+	MOVBQZX 32(CX), BX
+	MOVQ    (CX), AX
+	MOVQ    8(CX), SI
+	ADDQ    SI, AX
+	MOVQ    AX, (SP)
+	MOVQ    ctx+16(FP), AX
+	MOVQ    72(AX), DI
+	MOVQ    80(AX), R8
+	MOVQ    88(AX), R9
+	XORQ    CX, CX
+	MOVQ    CX, 8(SP)
+	MOVQ    CX, 16(SP)
+	MOVQ    CX, 24(SP)
+	MOVQ    112(AX), R10
+	MOVQ    128(AX), CX
+	MOVQ    CX, 32(SP)
+	MOVQ    144(AX), R11
+	MOVQ    136(AX), R12
+	MOVQ    200(AX), CX
+	MOVQ    CX, 56(SP)
+	MOVQ    176(AX), CX
+	MOVQ    CX, 48(SP)
+	MOVQ    184(AX), AX
+	MOVQ    AX, 40(SP)
+	MOVQ    40(SP), AX
+	ADDQ    AX, 48(SP)
+
+	// Calculate poiter to s.out[cap(s.out)] (a past-end pointer)
+	ADDQ R10, 32(SP)
+
+	// outBase += outPosition
+	ADDQ R12, R10
+
+sequenceDecs_decodeSync_amd64_main_loop:
+	MOVQ (SP), R13
+
+	// Fill bitreader to have enough for the offset and match length.
+	CMPQ SI, $0x08
+	JL   sequenceDecs_decodeSync_amd64_fill_byte_by_byte
+	MOVQ BX, AX
+	SHRQ $0x03, AX
+	SUBQ AX, R13
+	MOVQ (R13), DX
+	SUBQ AX, SI
+	ANDQ $0x07, BX
+	JMP  sequenceDecs_decodeSync_amd64_fill_end
+
+sequenceDecs_decodeSync_amd64_fill_byte_by_byte:
+	CMPQ    SI, $0x00
+	JLE     sequenceDecs_decodeSync_amd64_fill_check_overread
+	CMPQ    BX, $0x07
+	JLE     sequenceDecs_decodeSync_amd64_fill_end
+	SHLQ    $0x08, DX
+	SUBQ    $0x01, R13
+	SUBQ    $0x01, SI
+	SUBQ    $0x08, BX
+	MOVBQZX (R13), AX
+	ORQ     AX, DX
+	JMP     sequenceDecs_decodeSync_amd64_fill_byte_by_byte
+
+sequenceDecs_decodeSync_amd64_fill_check_overread:
+	CMPQ BX, $0x40
+	JA   error_overread
+
+sequenceDecs_decodeSync_amd64_fill_end:
+	// Update offset
+	MOVQ  R9, AX
+	MOVQ  BX, CX
+	MOVQ  DX, R14
+	SHLQ  CL, R14
+	MOVB  AH, CL
+	SHRQ  $0x20, AX
+	TESTQ CX, CX
+	JZ    sequenceDecs_decodeSync_amd64_of_update_zero
+	ADDQ  CX, BX
+	CMPQ  BX, $0x40
+	JA    sequenceDecs_decodeSync_amd64_of_update_zero
+	CMPQ  CX, $0x40
+	JAE   sequenceDecs_decodeSync_amd64_of_update_zero
+	NEGQ  CX
+	SHRQ  CL, R14
+	ADDQ  R14, AX
+
+sequenceDecs_decodeSync_amd64_of_update_zero:
+	MOVQ AX, 8(SP)
+
+	// Update match length
+	MOVQ  R8, AX
+	MOVQ  BX, CX
+	MOVQ  DX, R14
+	SHLQ  CL, R14
+	MOVB  AH, CL
+	SHRQ  $0x20, AX
+	TESTQ CX, CX
+	JZ    sequenceDecs_decodeSync_amd64_ml_update_zero
+	ADDQ  CX, BX
+	CMPQ  BX, $0x40
+	JA    sequenceDecs_decodeSync_amd64_ml_update_zero
+	CMPQ  CX, $0x40
+	JAE   sequenceDecs_decodeSync_amd64_ml_update_zero
+	NEGQ  CX
+	SHRQ  CL, R14
+	ADDQ  R14, AX
+
+sequenceDecs_decodeSync_amd64_ml_update_zero:
+	MOVQ AX, 16(SP)
+
+	// Fill bitreader to have enough for the remaining
+	CMPQ SI, $0x08
+	JL   sequenceDecs_decodeSync_amd64_fill_2_byte_by_byte
+	MOVQ BX, AX
+	SHRQ $0x03, AX
+	SUBQ AX, R13
+	MOVQ (R13), DX
+	SUBQ AX, SI
+	ANDQ $0x07, BX
+	JMP  sequenceDecs_decodeSync_amd64_fill_2_end
+
+sequenceDecs_decodeSync_amd64_fill_2_byte_by_byte:
+	CMPQ    SI, $0x00
+	JLE     sequenceDecs_decodeSync_amd64_fill_2_check_overread
+	CMPQ    BX, $0x07
+	JLE     sequenceDecs_decodeSync_amd64_fill_2_end
+	SHLQ    $0x08, DX
+	SUBQ    $0x01, R13
+	SUBQ    $0x01, SI
+	SUBQ    $0x08, BX
+	MOVBQZX (R13), AX
+	ORQ     AX, DX
+	JMP     sequenceDecs_decodeSync_amd64_fill_2_byte_by_byte
+
+sequenceDecs_decodeSync_amd64_fill_2_check_overread:
+	CMPQ BX, $0x40
+	JA   error_overread
+
+sequenceDecs_decodeSync_amd64_fill_2_end:
+	// Update literal length
+	MOVQ  DI, AX
+	MOVQ  BX, CX
+	MOVQ  DX, R14
+	SHLQ  CL, R14
+	MOVB  AH, CL
+	SHRQ  $0x20, AX
+	TESTQ CX, CX
+	JZ    sequenceDecs_decodeSync_amd64_ll_update_zero
+	ADDQ  CX, BX
+	CMPQ  BX, $0x40
+	JA    sequenceDecs_decodeSync_amd64_ll_update_zero
+	CMPQ  CX, $0x40
+	JAE   sequenceDecs_decodeSync_amd64_ll_update_zero
+	NEGQ  CX
+	SHRQ  CL, R14
+	ADDQ  R14, AX
+
+sequenceDecs_decodeSync_amd64_ll_update_zero:
+	MOVQ AX, 24(SP)
+
+	// Fill bitreader for state updates
+	MOVQ    R13, (SP)
+	MOVQ    R9, AX
+	SHRQ    $0x08, AX
+	MOVBQZX AL, AX
+	MOVQ    ctx+16(FP), CX
+	CMPQ    96(CX), $0x00
+	JZ      sequenceDecs_decodeSync_amd64_skip_update
+
+	// Update Literal Length State
+	MOVBQZX DI, R13
+	SHRL    $0x10, DI
+	LEAQ    (BX)(R13*1), CX
+	MOVQ    DX, R14
+	MOVQ    CX, BX
+	ROLQ    CL, R14
+	MOVL    $0x00000001, R15
+	MOVB    R13, CL
+	SHLL    CL, R15
+	DECL    R15
+	ANDQ    R15, R14
+	ADDQ    R14, DI
+
+	// Load ctx.llTable
+	MOVQ ctx+16(FP), CX
+	MOVQ (CX), CX
+	MOVQ (CX)(DI*8), DI
+
+	// Update Match Length State
+	MOVBQZX R8, R13
+	SHRL    $0x10, R8
+	LEAQ    (BX)(R13*1), CX
+	MOVQ    DX, R14
+	MOVQ    CX, BX
+	ROLQ    CL, R14
+	MOVL    $0x00000001, R15
+	MOVB    R13, CL
+	SHLL    CL, R15
+	DECL    R15
+	ANDQ    R15, R14
+	ADDQ    R14, R8
+
+	// Load ctx.mlTable
+	MOVQ ctx+16(FP), CX
+	MOVQ 24(CX), CX
+	MOVQ (CX)(R8*8), R8
+
+	// Update Offset State
+	MOVBQZX R9, R13
+	SHRL    $0x10, R9
+	LEAQ    (BX)(R13*1), CX
+	MOVQ    DX, R14
+	MOVQ    CX, BX
+	ROLQ    CL, R14
+	MOVL    $0x00000001, R15
+	MOVB    R13, CL
+	SHLL    CL, R15
+	DECL    R15
+	ANDQ    R15, R14
+	ADDQ    R14, R9
+
+	// Load ctx.ofTable
+	MOVQ ctx+16(FP), CX
+	MOVQ 48(CX), CX
+	MOVQ (CX)(R9*8), R9
+
+sequenceDecs_decodeSync_amd64_skip_update:
+	// Adjust offset
+	MOVQ   s+0(FP), CX
+	MOVQ   8(SP), R13
+	CMPQ   AX, $0x01
+	JBE    sequenceDecs_decodeSync_amd64_adjust_offsetB_1_or_0
+	MOVUPS 144(CX), X0
+	MOVQ   R13, 144(CX)
+	MOVUPS X0, 152(CX)
+	JMP    sequenceDecs_decodeSync_amd64_after_adjust
+
+sequenceDecs_decodeSync_amd64_adjust_offsetB_1_or_0:
+	CMPQ 24(SP), $0x00000000
+	JNE  sequenceDecs_decodeSync_amd64_adjust_offset_maybezero
+	INCQ R13
+	JMP  sequenceDecs_decodeSync_amd64_adjust_offset_nonzero
+
+sequenceDecs_decodeSync_amd64_adjust_offset_maybezero:
+	TESTQ R13, R13
+	JNZ   sequenceDecs_decodeSync_amd64_adjust_offset_nonzero
+	MOVQ  144(CX), R13
+	JMP   sequenceDecs_decodeSync_amd64_after_adjust
+
+sequenceDecs_decodeSync_amd64_adjust_offset_nonzero:
+	MOVQ    R13, AX
+	XORQ    R14, R14
+	MOVQ    $-1, R15
+	CMPQ    R13, $0x03
+	CMOVQEQ R14, AX
+	CMOVQEQ R15, R14
+	ADDQ    144(CX)(AX*8), R14
+	JNZ     sequenceDecs_decodeSync_amd64_adjust_temp_valid
+	MOVQ    $0x00000001, R14
+
+sequenceDecs_decodeSync_amd64_adjust_temp_valid:
+	CMPQ R13, $0x01
+	JZ   sequenceDecs_decodeSync_amd64_adjust_skip
+	MOVQ 152(CX), AX
+	MOVQ AX, 160(CX)
+
+sequenceDecs_decodeSync_amd64_adjust_skip:
+	MOVQ 144(CX), AX
+	MOVQ AX, 152(CX)
+	MOVQ R14, 144(CX)
+	MOVQ R14, R13
+
+sequenceDecs_decodeSync_amd64_after_adjust:
+	MOVQ R13, 8(SP)
+
+	// Check values
+	MOVQ  16(SP), AX
+	MOVQ  24(SP), CX
+	LEAQ  (AX)(CX*1), R14
+	MOVQ  s+0(FP), R15
+	ADDQ  R14, 256(R15)
+	MOVQ  ctx+16(FP), R14
+	SUBQ  CX, 104(R14)
+	JS    error_not_enough_literals
+	CMPQ  AX, $0x00020002
+	JA    sequenceDecs_decodeSync_amd64_error_match_len_too_big
+	TESTQ R13, R13
+	JNZ   sequenceDecs_decodeSync_amd64_match_len_ofs_ok
+	TESTQ AX, AX
+	JNZ   sequenceDecs_decodeSync_amd64_error_match_len_ofs_mismatch
+
+sequenceDecs_decodeSync_amd64_match_len_ofs_ok:
+	MOVQ 24(SP), AX
+	MOVQ 8(SP), CX
+	MOVQ 16(SP), R13
+
+	// Check if we have enough space in s.out
+	LEAQ (AX)(R13*1), R14
+	ADDQ R10, R14
+	CMPQ R14, 32(SP)
+	JA   error_not_enough_space
+
+	// Copy literals
+	TESTQ AX, AX
+	JZ    check_offset
+	XORQ  R14, R14
+
+copy_1:
+	MOVUPS (R11)(R14*1), X0
+	MOVUPS X0, (R10)(R14*1)
+	ADDQ   $0x10, R14
+	CMPQ   R14, AX
+	JB     copy_1
+	ADDQ   AX, R11
+	ADDQ   AX, R10
+	ADDQ   AX, R12
+
+	// Malformed input if seq.mo > t+len(hist) || seq.mo > s.windowSize)
+check_offset:
+	MOVQ R12, AX
+	ADDQ 40(SP), AX
+	CMPQ CX, AX
+	JG   error_match_off_too_big
+	CMPQ CX, 56(SP)
+	JG   error_match_off_too_big
+
+	// Copy match from history
+	MOVQ CX, AX
+	SUBQ R12, AX
+	JLS  copy_match
+	MOVQ 48(SP), R14
+	SUBQ AX, R14
+	CMPQ R13, AX
+	JG   copy_all_from_history
+	MOVQ R13, AX
+	SUBQ $0x10, AX
+	JB   copy_4_small
+
+copy_4_loop:
+	MOVUPS (R14), X0
+	MOVUPS X0, (R10)
+	ADDQ   $0x10, R14
+	ADDQ   $0x10, R10
+	SUBQ   $0x10, AX
+	JAE    copy_4_loop
+	LEAQ   16(R14)(AX*1), R14
+	LEAQ   16(R10)(AX*1), R10
+	MOVUPS -16(R14), X0
+	MOVUPS X0, -16(R10)
+	JMP    copy_4_end
+
+copy_4_small:
+	CMPQ R13, $0x03
+	JE   copy_4_move_3
+	CMPQ R13, $0x08
+	JB   copy_4_move_4through7
+	JMP  copy_4_move_8through16
+
+copy_4_move_3:
+	MOVW (R14), AX
+	MOVB 2(R14), CL
+	MOVW AX, (R10)
+	MOVB CL, 2(R10)
+	ADDQ R13, R14
+	ADDQ R13, R10
+	JMP  copy_4_end
+
+copy_4_move_4through7:
+	MOVL (R14), AX
+	MOVL -4(R14)(R13*1), CX
+	MOVL AX, (R10)
+	MOVL CX, -4(R10)(R13*1)
+	ADDQ R13, R14
+	ADDQ R13, R10
+	JMP  copy_4_end
+
+copy_4_move_8through16:
+	MOVQ (R14), AX
+	MOVQ -8(R14)(R13*1), CX
+	MOVQ AX, (R10)
+	MOVQ CX, -8(R10)(R13*1)
+	ADDQ R13, R14
+	ADDQ R13, R10
+
+copy_4_end:
+	ADDQ R13, R12
+	JMP  handle_loop
+	JMP loop_finished
+
+copy_all_from_history:
+	MOVQ AX, R15
+	SUBQ $0x10, R15
+	JB   copy_5_small
+
+copy_5_loop:
+	MOVUPS (R14), X0
+	MOVUPS X0, (R10)
+	ADDQ   $0x10, R14
+	ADDQ   $0x10, R10
+	SUBQ   $0x10, R15
+	JAE    copy_5_loop
+	LEAQ   16(R14)(R15*1), R14
+	LEAQ   16(R10)(R15*1), R10
+	MOVUPS -16(R14), X0
+	MOVUPS X0, -16(R10)
+	JMP    copy_5_end
+
+copy_5_small:
+	CMPQ AX, $0x03
+	JE   copy_5_move_3
+	JB   copy_5_move_1or2
+	CMPQ AX, $0x08
+	JB   copy_5_move_4through7
+	JMP  copy_5_move_8through16
+
+copy_5_move_1or2:
+	MOVB (R14), R15
+	MOVB -1(R14)(AX*1), BP
+	MOVB R15, (R10)
+	MOVB BP, -1(R10)(AX*1)
+	ADDQ AX, R14
+	ADDQ AX, R10
+	JMP  copy_5_end
+
+copy_5_move_3:
+	MOVW (R14), R15
+	MOVB 2(R14), BP
+	MOVW R15, (R10)
+	MOVB BP, 2(R10)
+	ADDQ AX, R14
+	ADDQ AX, R10
+	JMP  copy_5_end
+
+copy_5_move_4through7:
+	MOVL (R14), R15
+	MOVL -4(R14)(AX*1), BP
+	MOVL R15, (R10)
+	MOVL BP, -4(R10)(AX*1)
+	ADDQ AX, R14
+	ADDQ AX, R10
+	JMP  copy_5_end
+
+copy_5_move_8through16:
+	MOVQ (R14), R15
+	MOVQ -8(R14)(AX*1), BP
+	MOVQ R15, (R10)
+	MOVQ BP, -8(R10)(AX*1)
+	ADDQ AX, R14
+	ADDQ AX, R10
+
+copy_5_end:
+	ADDQ AX, R12
+	SUBQ AX, R13
+
+	// Copy match from the current buffer
+copy_match:
+	MOVQ R10, AX
+	SUBQ CX, AX
+
+	// ml <= mo
+	CMPQ R13, CX
+	JA   copy_overlapping_match
+
+	// Copy non-overlapping match
+	ADDQ R13, R12
+	MOVQ R10, CX
+	ADDQ R13, R10
+
+copy_2:
+	MOVUPS (AX), X0
+	MOVUPS X0, (CX)
+	ADDQ   $0x10, AX
+	ADDQ   $0x10, CX
+	SUBQ   $0x10, R13
+	JHI    copy_2
+	JMP    handle_loop
+
+	// Copy overlapping match
+copy_overlapping_match:
+	ADDQ R13, R12
+
+copy_slow_3:
+	MOVB (AX), CL
+	MOVB CL, (R10)
+	INCQ AX
+	INCQ R10
+	DECQ R13
+	JNZ  copy_slow_3
+
+handle_loop:
+	MOVQ ctx+16(FP), AX
+	DECQ 96(AX)
+	JNS  sequenceDecs_decodeSync_amd64_main_loop
+
+loop_finished:
+	MOVQ br+8(FP), AX
+	MOVQ DX, 24(AX)
+	MOVB BL, 32(AX)
+	MOVQ SI, 8(AX)
+
+	// Update the context
+	MOVQ ctx+16(FP), AX
+	MOVQ R12, 136(AX)
+	MOVQ 144(AX), CX
+	SUBQ CX, R11
+	MOVQ R11, 168(AX)
+
+	// Return success
+	MOVQ $0x00000000, ret+24(FP)
+	RET
+
+	// Return with match length error
+sequenceDecs_decodeSync_amd64_error_match_len_ofs_mismatch:
+	MOVQ 16(SP), AX
+	MOVQ ctx+16(FP), CX
+	MOVQ AX, 216(CX)
+	MOVQ $0x00000001, ret+24(FP)
+	RET
+
+	// Return with match too long error
+sequenceDecs_decodeSync_amd64_error_match_len_too_big:
+	MOVQ ctx+16(FP), AX
+	MOVQ 16(SP), CX
+	MOVQ CX, 216(AX)
+	MOVQ $0x00000002, ret+24(FP)
+	RET
+
+	// Return with match offset too long error
+error_match_off_too_big:
+	MOVQ ctx+16(FP), AX
+	MOVQ 8(SP), CX
+	MOVQ CX, 224(AX)
+	MOVQ R12, 136(AX)
+	MOVQ $0x00000003, ret+24(FP)
+	RET
+
+	// Return with not enough literals error
+error_not_enough_literals:
+	MOVQ ctx+16(FP), AX
+	MOVQ 24(SP), CX
+	MOVQ CX, 208(AX)
+	MOVQ $0x00000004, ret+24(FP)
+	RET
+
+	// Return with overread error
+error_overread:
+	MOVQ $0x00000006, ret+24(FP)
+	RET
+
+	// Return with not enough output space error
+error_not_enough_space:
+	MOVQ ctx+16(FP), AX
+	MOVQ 24(SP), CX
+	MOVQ CX, 208(AX)
+	MOVQ 16(SP), CX
+	MOVQ CX, 216(AX)
+	MOVQ R12, 136(AX)
+	MOVQ $0x00000005, ret+24(FP)
+	RET
+
+// func sequenceDecs_decodeSync_bmi2(s *sequenceDecs, br *bitReader, ctx *decodeSyncAsmContext) int
+// Requires: BMI, BMI2, CMOV, SSE
+TEXT ·sequenceDecs_decodeSync_bmi2(SB), $64-32
+	MOVQ    br+8(FP), BX
+	MOVQ    24(BX), AX
+	MOVBQZX 32(BX), DX
+	MOVQ    (BX), CX
+	MOVQ    8(BX), BX
+	ADDQ    BX, CX
+	MOVQ    CX, (SP)
+	MOVQ    ctx+16(FP), CX
+	MOVQ    72(CX), SI
+	MOVQ    80(CX), DI
+	MOVQ    88(CX), R8
+	XORQ    R9, R9
+	MOVQ    R9, 8(SP)
+	MOVQ    R9, 16(SP)
+	MOVQ    R9, 24(SP)
+	MOVQ    112(CX), R9
+	MOVQ    128(CX), R10
+	MOVQ    R10, 32(SP)
+	MOVQ    144(CX), R10
+	MOVQ    136(CX), R11
+	MOVQ    200(CX), R12
+	MOVQ    R12, 56(SP)
+	MOVQ    176(CX), R12
+	MOVQ    R12, 48(SP)
+	MOVQ    184(CX), CX
+	MOVQ    CX, 40(SP)
+	MOVQ    40(SP), CX
+	ADDQ    CX, 48(SP)
+
+	// Calculate poiter to s.out[cap(s.out)] (a past-end pointer)
+	ADDQ R9, 32(SP)
+
+	// outBase += outPosition
+	ADDQ R11, R9
+
+sequenceDecs_decodeSync_bmi2_main_loop:
+	MOVQ (SP), R12
+
+	// Fill bitreader to have enough for the offset and match length.
+	CMPQ BX, $0x08
+	JL   sequenceDecs_decodeSync_bmi2_fill_byte_by_byte
+	MOVQ DX, CX
+	SHRQ $0x03, CX
+	SUBQ CX, R12
+	MOVQ (R12), AX
+	SUBQ CX, BX
+	ANDQ $0x07, DX
+	JMP  sequenceDecs_decodeSync_bmi2_fill_end
+
+sequenceDecs_decodeSync_bmi2_fill_byte_by_byte:
+	CMPQ    BX, $0x00
+	JLE     sequenceDecs_decodeSync_bmi2_fill_check_overread
+	CMPQ    DX, $0x07
+	JLE     sequenceDecs_decodeSync_bmi2_fill_end
+	SHLQ    $0x08, AX
+	SUBQ    $0x01, R12
+	SUBQ    $0x01, BX
+	SUBQ    $0x08, DX
+	MOVBQZX (R12), CX
+	ORQ     CX, AX
+	JMP     sequenceDecs_decodeSync_bmi2_fill_byte_by_byte
+
+sequenceDecs_decodeSync_bmi2_fill_check_overread:
+	CMPQ DX, $0x40
+	JA   error_overread
+
+sequenceDecs_decodeSync_bmi2_fill_end:
+	// Update offset
+	MOVQ   $0x00000808, CX
+	BEXTRQ CX, R8, R13
+	MOVQ   AX, R14
+	LEAQ   (DX)(R13*1), CX
+	ROLQ   CL, R14
+	BZHIQ  R13, R14, R14
+	MOVQ   CX, DX
+	MOVQ   R8, CX
+	SHRQ   $0x20, CX
+	ADDQ   R14, CX
+	MOVQ   CX, 8(SP)
+
+	// Update match length
+	MOVQ   $0x00000808, CX
+	BEXTRQ CX, DI, R13
+	MOVQ   AX, R14
+	LEAQ   (DX)(R13*1), CX
+	ROLQ   CL, R14
+	BZHIQ  R13, R14, R14
+	MOVQ   CX, DX
+	MOVQ   DI, CX
+	SHRQ   $0x20, CX
+	ADDQ   R14, CX
+	MOVQ   CX, 16(SP)
+
+	// Fill bitreader to have enough for the remaining
+	CMPQ BX, $0x08
+	JL   sequenceDecs_decodeSync_bmi2_fill_2_byte_by_byte
+	MOVQ DX, CX
+	SHRQ $0x03, CX
+	SUBQ CX, R12
+	MOVQ (R12), AX
+	SUBQ CX, BX
+	ANDQ $0x07, DX
+	JMP  sequenceDecs_decodeSync_bmi2_fill_2_end
+
+sequenceDecs_decodeSync_bmi2_fill_2_byte_by_byte:
+	CMPQ    BX, $0x00
+	JLE     sequenceDecs_decodeSync_bmi2_fill_2_check_overread
+	CMPQ    DX, $0x07
+	JLE     sequenceDecs_decodeSync_bmi2_fill_2_end
+	SHLQ    $0x08, AX
+	SUBQ    $0x01, R12
+	SUBQ    $0x01, BX
+	SUBQ    $0x08, DX
+	MOVBQZX (R12), CX
+	ORQ     CX, AX
+	JMP     sequenceDecs_decodeSync_bmi2_fill_2_byte_by_byte
+
+sequenceDecs_decodeSync_bmi2_fill_2_check_overread:
+	CMPQ DX, $0x40
+	JA   error_overread
+
+sequenceDecs_decodeSync_bmi2_fill_2_end:
+	// Update literal length
+	MOVQ   $0x00000808, CX
+	BEXTRQ CX, SI, R13
+	MOVQ   AX, R14
+	LEAQ   (DX)(R13*1), CX
+	ROLQ   CL, R14
+	BZHIQ  R13, R14, R14
+	MOVQ   CX, DX
+	MOVQ   SI, CX
+	SHRQ   $0x20, CX
+	ADDQ   R14, CX
+	MOVQ   CX, 24(SP)
+
+	// Fill bitreader for state updates
+	MOVQ    R12, (SP)
+	MOVQ    $0x00000808, CX
+	BEXTRQ  CX, R8, R12
+	MOVQ    ctx+16(FP), CX
+	CMPQ    96(CX), $0x00
+	JZ      sequenceDecs_decodeSync_bmi2_skip_update
+	LEAQ    (SI)(DI*1), R13
+	ADDQ    R8, R13
+	MOVBQZX R13, R13
+	LEAQ    (DX)(R13*1), CX
+	MOVQ    AX, R14
+	MOVQ    CX, DX
+	ROLQ    CL, R14
+	BZHIQ   R13, R14, R14
+
+	// Update Offset State
+	BZHIQ R8, R14, CX
+	SHRXQ R8, R14, R14
+	SHRL  $0x10, R8
+	ADDQ  CX, R8
+
+	// Load ctx.ofTable
+	MOVQ ctx+16(FP), CX
+	MOVQ 48(CX), CX
+	MOVQ (CX)(R8*8), R8
+
+	// Update Match Length State
+	BZHIQ DI, R14, CX
+	SHRXQ DI, R14, R14
+	SHRL  $0x10, DI
+	ADDQ  CX, DI
+
+	// Load ctx.mlTable
+	MOVQ ctx+16(FP), CX
+	MOVQ 24(CX), CX
+	MOVQ (CX)(DI*8), DI
+
+	// Update Literal Length State
+	BZHIQ SI, R14, CX
+	SHRL  $0x10, SI
+	ADDQ  CX, SI
+
+	// Load ctx.llTable
+	MOVQ ctx+16(FP), CX
+	MOVQ (CX), CX
+	MOVQ (CX)(SI*8), SI
+
+sequenceDecs_decodeSync_bmi2_skip_update:
+	// Adjust offset
+	MOVQ   s+0(FP), CX
+	MOVQ   8(SP), R13
+	CMPQ   R12, $0x01
+	JBE    sequenceDecs_decodeSync_bmi2_adjust_offsetB_1_or_0
+	MOVUPS 144(CX), X0
+	MOVQ   R13, 144(CX)
+	MOVUPS X0, 152(CX)
+	JMP    sequenceDecs_decodeSync_bmi2_after_adjust
+
+sequenceDecs_decodeSync_bmi2_adjust_offsetB_1_or_0:
+	CMPQ 24(SP), $0x00000000
+	JNE  sequenceDecs_decodeSync_bmi2_adjust_offset_maybezero
+	INCQ R13
+	JMP  sequenceDecs_decodeSync_bmi2_adjust_offset_nonzero
+
+sequenceDecs_decodeSync_bmi2_adjust_offset_maybezero:
+	TESTQ R13, R13
+	JNZ   sequenceDecs_decodeSync_bmi2_adjust_offset_nonzero
+	MOVQ  144(CX), R13
+	JMP   sequenceDecs_decodeSync_bmi2_after_adjust
+
+sequenceDecs_decodeSync_bmi2_adjust_offset_nonzero:
+	MOVQ    R13, R12
+	XORQ    R14, R14
+	MOVQ    $-1, R15
+	CMPQ    R13, $0x03
+	CMOVQEQ R14, R12
+	CMOVQEQ R15, R14
+	ADDQ    144(CX)(R12*8), R14
+	JNZ     sequenceDecs_decodeSync_bmi2_adjust_temp_valid
+	MOVQ    $0x00000001, R14
+
+sequenceDecs_decodeSync_bmi2_adjust_temp_valid:
+	CMPQ R13, $0x01
+	JZ   sequenceDecs_decodeSync_bmi2_adjust_skip
+	MOVQ 152(CX), R12
+	MOVQ R12, 160(CX)
+
+sequenceDecs_decodeSync_bmi2_adjust_skip:
+	MOVQ 144(CX), R12
+	MOVQ R12, 152(CX)
+	MOVQ R14, 144(CX)
+	MOVQ R14, R13
+
+sequenceDecs_decodeSync_bmi2_after_adjust:
+	MOVQ R13, 8(SP)
+
+	// Check values
+	MOVQ  16(SP), CX
+	MOVQ  24(SP), R12
+	LEAQ  (CX)(R12*1), R14
+	MOVQ  s+0(FP), R15
+	ADDQ  R14, 256(R15)
+	MOVQ  ctx+16(FP), R14
+	SUBQ  R12, 104(R14)
+	JS    error_not_enough_literals
+	CMPQ  CX, $0x00020002
+	JA    sequenceDecs_decodeSync_bmi2_error_match_len_too_big
+	TESTQ R13, R13
+	JNZ   sequenceDecs_decodeSync_bmi2_match_len_ofs_ok
+	TESTQ CX, CX
+	JNZ   sequenceDecs_decodeSync_bmi2_error_match_len_ofs_mismatch
+
+sequenceDecs_decodeSync_bmi2_match_len_ofs_ok:
+	MOVQ 24(SP), CX
+	MOVQ 8(SP), R12
+	MOVQ 16(SP), R13
+
+	// Check if we have enough space in s.out
+	LEAQ (CX)(R13*1), R14
+	ADDQ R9, R14
+	CMPQ R14, 32(SP)
+	JA   error_not_enough_space
+
+	// Copy literals
+	TESTQ CX, CX
+	JZ    check_offset
+	XORQ  R14, R14
+
+copy_1:
+	MOVUPS (R10)(R14*1), X0
+	MOVUPS X0, (R9)(R14*1)
+	ADDQ   $0x10, R14
+	CMPQ   R14, CX
+	JB     copy_1
+	ADDQ   CX, R10
+	ADDQ   CX, R9
+	ADDQ   CX, R11
+
+	// Malformed input if seq.mo > t+len(hist) || seq.mo > s.windowSize)
+check_offset:
+	MOVQ R11, CX
+	ADDQ 40(SP), CX
+	CMPQ R12, CX
+	JG   error_match_off_too_big
+	CMPQ R12, 56(SP)
+	JG   error_match_off_too_big
+
+	// Copy match from history
+	MOVQ R12, CX
+	SUBQ R11, CX
+	JLS  copy_match
+	MOVQ 48(SP), R14
+	SUBQ CX, R14
+	CMPQ R13, CX
+	JG   copy_all_from_history
+	MOVQ R13, CX
+	SUBQ $0x10, CX
+	JB   copy_4_small
+
+copy_4_loop:
+	MOVUPS (R14), X0
+	MOVUPS X0, (R9)
+	ADDQ   $0x10, R14
+	ADDQ   $0x10, R9
+	SUBQ   $0x10, CX
+	JAE    copy_4_loop
+	LEAQ   16(R14)(CX*1), R14
+	LEAQ   16(R9)(CX*1), R9
+	MOVUPS -16(R14), X0
+	MOVUPS X0, -16(R9)
+	JMP    copy_4_end
+
+copy_4_small:
+	CMPQ R13, $0x03
+	JE   copy_4_move_3
+	CMPQ R13, $0x08
+	JB   copy_4_move_4through7
+	JMP  copy_4_move_8through16
+
+copy_4_move_3:
+	MOVW (R14), CX
+	MOVB 2(R14), R12
+	MOVW CX, (R9)
+	MOVB R12, 2(R9)
+	ADDQ R13, R14
+	ADDQ R13, R9
+	JMP  copy_4_end
+
+copy_4_move_4through7:
+	MOVL (R14), CX
+	MOVL -4(R14)(R13*1), R12
+	MOVL CX, (R9)
+	MOVL R12, -4(R9)(R13*1)
+	ADDQ R13, R14
+	ADDQ R13, R9
+	JMP  copy_4_end
+
+copy_4_move_8through16:
+	MOVQ (R14), CX
+	MOVQ -8(R14)(R13*1), R12
+	MOVQ CX, (R9)
+	MOVQ R12, -8(R9)(R13*1)
+	ADDQ R13, R14
+	ADDQ R13, R9
+
+copy_4_end:
+	ADDQ R13, R11
+	JMP  handle_loop
+	JMP loop_finished
+
+copy_all_from_history:
+	MOVQ CX, R15
+	SUBQ $0x10, R15
+	JB   copy_5_small
+
+copy_5_loop:
+	MOVUPS (R14), X0
+	MOVUPS X0, (R9)
+	ADDQ   $0x10, R14
+	ADDQ   $0x10, R9
+	SUBQ   $0x10, R15
+	JAE    copy_5_loop
+	LEAQ   16(R14)(R15*1), R14
+	LEAQ   16(R9)(R15*1), R9
+	MOVUPS -16(R14), X0
+	MOVUPS X0, -16(R9)
+	JMP    copy_5_end
+
+copy_5_small:
+	CMPQ CX, $0x03
+	JE   copy_5_move_3
+	JB   copy_5_move_1or2
+	CMPQ CX, $0x08
+	JB   copy_5_move_4through7
+	JMP  copy_5_move_8through16
+
+copy_5_move_1or2:
+	MOVB (R14), R15
+	MOVB -1(R14)(CX*1), BP
+	MOVB R15, (R9)
+	MOVB BP, -1(R9)(CX*1)
+	ADDQ CX, R14
+	ADDQ CX, R9
+	JMP  copy_5_end
+
+copy_5_move_3:
+	MOVW (R14), R15
+	MOVB 2(R14), BP
+	MOVW R15, (R9)
+	MOVB BP, 2(R9)
+	ADDQ CX, R14
+	ADDQ CX, R9
+	JMP  copy_5_end
+
+copy_5_move_4through7:
+	MOVL (R14), R15
+	MOVL -4(R14)(CX*1), BP
+	MOVL R15, (R9)
+	MOVL BP, -4(R9)(CX*1)
+	ADDQ CX, R14
+	ADDQ CX, R9
+	JMP  copy_5_end
+
+copy_5_move_8through16:
+	MOVQ (R14), R15
+	MOVQ -8(R14)(CX*1), BP
+	MOVQ R15, (R9)
+	MOVQ BP, -8(R9)(CX*1)
+	ADDQ CX, R14
+	ADDQ CX, R9
+
+copy_5_end:
+	ADDQ CX, R11
+	SUBQ CX, R13
+
+	// Copy match from the current buffer
+copy_match:
+	MOVQ R9, CX
+	SUBQ R12, CX
+
+	// ml <= mo
+	CMPQ R13, R12
+	JA   copy_overlapping_match
+
+	// Copy non-overlapping match
+	ADDQ R13, R11
+	MOVQ R9, R12
+	ADDQ R13, R9
+
+copy_2:
+	MOVUPS (CX), X0
+	MOVUPS X0, (R12)
+	ADDQ   $0x10, CX
+	ADDQ   $0x10, R12
+	SUBQ   $0x10, R13
+	JHI    copy_2
+	JMP    handle_loop
+
+	// Copy overlapping match
+copy_overlapping_match:
+	ADDQ R13, R11
+
+copy_slow_3:
+	MOVB (CX), R12
+	MOVB R12, (R9)
+	INCQ CX
+	INCQ R9
+	DECQ R13
+	JNZ  copy_slow_3
+
+handle_loop:
+	MOVQ ctx+16(FP), CX
+	DECQ 96(CX)
+	JNS  sequenceDecs_decodeSync_bmi2_main_loop
+
+loop_finished:
+	MOVQ br+8(FP), CX
+	MOVQ AX, 24(CX)
+	MOVB DL, 32(CX)
+	MOVQ BX, 8(CX)
+
+	// Update the context
+	MOVQ ctx+16(FP), AX
+	MOVQ R11, 136(AX)
+	MOVQ 144(AX), CX
+	SUBQ CX, R10
+	MOVQ R10, 168(AX)
+
+	// Return success
+	MOVQ $0x00000000, ret+24(FP)
+	RET
+
+	// Return with match length error
+sequenceDecs_decodeSync_bmi2_error_match_len_ofs_mismatch:
+	MOVQ 16(SP), AX
+	MOVQ ctx+16(FP), CX
+	MOVQ AX, 216(CX)
+	MOVQ $0x00000001, ret+24(FP)
+	RET
+
+	// Return with match too long error
+sequenceDecs_decodeSync_bmi2_error_match_len_too_big:
+	MOVQ ctx+16(FP), AX
+	MOVQ 16(SP), CX
+	MOVQ CX, 216(AX)
+	MOVQ $0x00000002, ret+24(FP)
+	RET
+
+	// Return with match offset too long error
+error_match_off_too_big:
+	MOVQ ctx+16(FP), AX
+	MOVQ 8(SP), CX
+	MOVQ CX, 224(AX)
+	MOVQ R11, 136(AX)
+	MOVQ $0x00000003, ret+24(FP)
+	RET
+
+	// Return with not enough literals error
+error_not_enough_literals:
+	MOVQ ctx+16(FP), AX
+	MOVQ 24(SP), CX
+	MOVQ CX, 208(AX)
+	MOVQ $0x00000004, ret+24(FP)
+	RET
+
+	// Return with overread error
+error_overread:
+	MOVQ $0x00000006, ret+24(FP)
+	RET
+
+	// Return with not enough output space error
+error_not_enough_space:
+	MOVQ ctx+16(FP), AX
+	MOVQ 24(SP), CX
+	MOVQ CX, 208(AX)
+	MOVQ 16(SP), CX
+	MOVQ CX, 216(AX)
+	MOVQ R11, 136(AX)
+	MOVQ $0x00000005, ret+24(FP)
+	RET
+
+// func sequenceDecs_decodeSync_safe_amd64(s *sequenceDecs, br *bitReader, ctx *decodeSyncAsmContext) int
+// Requires: CMOV, SSE
+TEXT ·sequenceDecs_decodeSync_safe_amd64(SB), $64-32
+	MOVQ    br+8(FP), CX
+	MOVQ    24(CX), DX
+	MOVBQZX 32(CX), BX
+	MOVQ    (CX), AX
+	MOVQ    8(CX), SI
+	ADDQ    SI, AX
+	MOVQ    AX, (SP)
+	MOVQ    ctx+16(FP), AX
+	MOVQ    72(AX), DI
+	MOVQ    80(AX), R8
+	MOVQ    88(AX), R9
+	XORQ    CX, CX
+	MOVQ    CX, 8(SP)
+	MOVQ    CX, 16(SP)
+	MOVQ    CX, 24(SP)
+	MOVQ    112(AX), R10
+	MOVQ    128(AX), CX
+	MOVQ    CX, 32(SP)
+	MOVQ    144(AX), R11
+	MOVQ    136(AX), R12
+	MOVQ    200(AX), CX
+	MOVQ    CX, 56(SP)
+	MOVQ    176(AX), CX
+	MOVQ    CX, 48(SP)
+	MOVQ    184(AX), AX
+	MOVQ    AX, 40(SP)
+	MOVQ    40(SP), AX
+	ADDQ    AX, 48(SP)
+
+	// Calculate poiter to s.out[cap(s.out)] (a past-end pointer)
+	ADDQ R10, 32(SP)
+
+	// outBase += outPosition
+	ADDQ R12, R10
+
+sequenceDecs_decodeSync_safe_amd64_main_loop:
+	MOVQ (SP), R13
+
+	// Fill bitreader to have enough for the offset and match length.
+	CMPQ SI, $0x08
+	JL   sequenceDecs_decodeSync_safe_amd64_fill_byte_by_byte
+	MOVQ BX, AX
+	SHRQ $0x03, AX
+	SUBQ AX, R13
+	MOVQ (R13), DX
+	SUBQ AX, SI
+	ANDQ $0x07, BX
+	JMP  sequenceDecs_decodeSync_safe_amd64_fill_end
+
+sequenceDecs_decodeSync_safe_amd64_fill_byte_by_byte:
+	CMPQ    SI, $0x00
+	JLE     sequenceDecs_decodeSync_safe_amd64_fill_check_overread
+	CMPQ    BX, $0x07
+	JLE     sequenceDecs_decodeSync_safe_amd64_fill_end
+	SHLQ    $0x08, DX
+	SUBQ    $0x01, R13
+	SUBQ    $0x01, SI
+	SUBQ    $0x08, BX
+	MOVBQZX (R13), AX
+	ORQ     AX, DX
+	JMP     sequenceDecs_decodeSync_safe_amd64_fill_byte_by_byte
+
+sequenceDecs_decodeSync_safe_amd64_fill_check_overread:
+	CMPQ BX, $0x40
+	JA   error_overread
+
+sequenceDecs_decodeSync_safe_amd64_fill_end:
+	// Update offset
+	MOVQ  R9, AX
+	MOVQ  BX, CX
+	MOVQ  DX, R14
+	SHLQ  CL, R14
+	MOVB  AH, CL
+	SHRQ  $0x20, AX
+	TESTQ CX, CX
+	JZ    sequenceDecs_decodeSync_safe_amd64_of_update_zero
+	ADDQ  CX, BX
+	CMPQ  BX, $0x40
+	JA    sequenceDecs_decodeSync_safe_amd64_of_update_zero
+	CMPQ  CX, $0x40
+	JAE   sequenceDecs_decodeSync_safe_amd64_of_update_zero
+	NEGQ  CX
+	SHRQ  CL, R14
+	ADDQ  R14, AX
+
+sequenceDecs_decodeSync_safe_amd64_of_update_zero:
+	MOVQ AX, 8(SP)
+
+	// Update match length
+	MOVQ  R8, AX
+	MOVQ  BX, CX
+	MOVQ  DX, R14
+	SHLQ  CL, R14
+	MOVB  AH, CL
+	SHRQ  $0x20, AX
+	TESTQ CX, CX
+	JZ    sequenceDecs_decodeSync_safe_amd64_ml_update_zero
+	ADDQ  CX, BX
+	CMPQ  BX, $0x40
+	JA    sequenceDecs_decodeSync_safe_amd64_ml_update_zero
+	CMPQ  CX, $0x40
+	JAE   sequenceDecs_decodeSync_safe_amd64_ml_update_zero
+	NEGQ  CX
+	SHRQ  CL, R14
+	ADDQ  R14, AX
+
+sequenceDecs_decodeSync_safe_amd64_ml_update_zero:
+	MOVQ AX, 16(SP)
+
+	// Fill bitreader to have enough for the remaining
+	CMPQ SI, $0x08
+	JL   sequenceDecs_decodeSync_safe_amd64_fill_2_byte_by_byte
+	MOVQ BX, AX
+	SHRQ $0x03, AX
+	SUBQ AX, R13
+	MOVQ (R13), DX
+	SUBQ AX, SI
+	ANDQ $0x07, BX
+	JMP  sequenceDecs_decodeSync_safe_amd64_fill_2_end
+
+sequenceDecs_decodeSync_safe_amd64_fill_2_byte_by_byte:
+	CMPQ    SI, $0x00
+	JLE     sequenceDecs_decodeSync_safe_amd64_fill_2_check_overread
+	CMPQ    BX, $0x07
+	JLE     sequenceDecs_decodeSync_safe_amd64_fill_2_end
+	SHLQ    $0x08, DX
+	SUBQ    $0x01, R13
+	SUBQ    $0x01, SI
+	SUBQ    $0x08, BX
+	MOVBQZX (R13), AX
+	ORQ     AX, DX
+	JMP     sequenceDecs_decodeSync_safe_amd64_fill_2_byte_by_byte
+
+sequenceDecs_decodeSync_safe_amd64_fill_2_check_overread:
+	CMPQ BX, $0x40
+	JA   error_overread
+
+sequenceDecs_decodeSync_safe_amd64_fill_2_end:
+	// Update literal length
+	MOVQ  DI, AX
+	MOVQ  BX, CX
+	MOVQ  DX, R14
+	SHLQ  CL, R14
+	MOVB  AH, CL
+	SHRQ  $0x20, AX
+	TESTQ CX, CX
+	JZ    sequenceDecs_decodeSync_safe_amd64_ll_update_zero
+	ADDQ  CX, BX
+	CMPQ  BX, $0x40
+	JA    sequenceDecs_decodeSync_safe_amd64_ll_update_zero
+	CMPQ  CX, $0x40
+	JAE   sequenceDecs_decodeSync_safe_amd64_ll_update_zero
+	NEGQ  CX
+	SHRQ  CL, R14
+	ADDQ  R14, AX
+
+sequenceDecs_decodeSync_safe_amd64_ll_update_zero:
+	MOVQ AX, 24(SP)
+
+	// Fill bitreader for state updates
+	MOVQ    R13, (SP)
+	MOVQ    R9, AX
+	SHRQ    $0x08, AX
+	MOVBQZX AL, AX
+	MOVQ    ctx+16(FP), CX
+	CMPQ    96(CX), $0x00
+	JZ      sequenceDecs_decodeSync_safe_amd64_skip_update
+
+	// Update Literal Length State
+	MOVBQZX DI, R13
+	SHRL    $0x10, DI
+	LEAQ    (BX)(R13*1), CX
+	MOVQ    DX, R14
+	MOVQ    CX, BX
+	ROLQ    CL, R14
+	MOVL    $0x00000001, R15
+	MOVB    R13, CL
+	SHLL    CL, R15
+	DECL    R15
+	ANDQ    R15, R14
+	ADDQ    R14, DI
+
+	// Load ctx.llTable
+	MOVQ ctx+16(FP), CX
+	MOVQ (CX), CX
+	MOVQ (CX)(DI*8), DI
+
+	// Update Match Length State
+	MOVBQZX R8, R13
+	SHRL    $0x10, R8
+	LEAQ    (BX)(R13*1), CX
+	MOVQ    DX, R14
+	MOVQ    CX, BX
+	ROLQ    CL, R14
+	MOVL    $0x00000001, R15
+	MOVB    R13, CL
+	SHLL    CL, R15
+	DECL    R15
+	ANDQ    R15, R14
+	ADDQ    R14, R8
+
+	// Load ctx.mlTable
+	MOVQ ctx+16(FP), CX
+	MOVQ 24(CX), CX
+	MOVQ (CX)(R8*8), R8
+
+	// Update Offset State
+	MOVBQZX R9, R13
+	SHRL    $0x10, R9
+	LEAQ    (BX)(R13*1), CX
+	MOVQ    DX, R14
+	MOVQ    CX, BX
+	ROLQ    CL, R14
+	MOVL    $0x00000001, R15
+	MOVB    R13, CL
+	SHLL    CL, R15
+	DECL    R15
+	ANDQ    R15, R14
+	ADDQ    R14, R9
+
+	// Load ctx.ofTable
+	MOVQ ctx+16(FP), CX
+	MOVQ 48(CX), CX
+	MOVQ (CX)(R9*8), R9
+
+sequenceDecs_decodeSync_safe_amd64_skip_update:
+	// Adjust offset
+	MOVQ   s+0(FP), CX
+	MOVQ   8(SP), R13
+	CMPQ   AX, $0x01
+	JBE    sequenceDecs_decodeSync_safe_amd64_adjust_offsetB_1_or_0
+	MOVUPS 144(CX), X0
+	MOVQ   R13, 144(CX)
+	MOVUPS X0, 152(CX)
+	JMP    sequenceDecs_decodeSync_safe_amd64_after_adjust
+
+sequenceDecs_decodeSync_safe_amd64_adjust_offsetB_1_or_0:
+	CMPQ 24(SP), $0x00000000
+	JNE  sequenceDecs_decodeSync_safe_amd64_adjust_offset_maybezero
+	INCQ R13
+	JMP  sequenceDecs_decodeSync_safe_amd64_adjust_offset_nonzero
+
+sequenceDecs_decodeSync_safe_amd64_adjust_offset_maybezero:
+	TESTQ R13, R13
+	JNZ   sequenceDecs_decodeSync_safe_amd64_adjust_offset_nonzero
+	MOVQ  144(CX), R13
+	JMP   sequenceDecs_decodeSync_safe_amd64_after_adjust
+
+sequenceDecs_decodeSync_safe_amd64_adjust_offset_nonzero:
+	MOVQ    R13, AX
+	XORQ    R14, R14
+	MOVQ    $-1, R15
+	CMPQ    R13, $0x03
+	CMOVQEQ R14, AX
+	CMOVQEQ R15, R14
+	ADDQ    144(CX)(AX*8), R14
+	JNZ     sequenceDecs_decodeSync_safe_amd64_adjust_temp_valid
+	MOVQ    $0x00000001, R14
+
+sequenceDecs_decodeSync_safe_amd64_adjust_temp_valid:
+	CMPQ R13, $0x01
+	JZ   sequenceDecs_decodeSync_safe_amd64_adjust_skip
+	MOVQ 152(CX), AX
+	MOVQ AX, 160(CX)
+
+sequenceDecs_decodeSync_safe_amd64_adjust_skip:
+	MOVQ 144(CX), AX
+	MOVQ AX, 152(CX)
+	MOVQ R14, 144(CX)
+	MOVQ R14, R13
+
+sequenceDecs_decodeSync_safe_amd64_after_adjust:
+	MOVQ R13, 8(SP)
+
+	// Check values
+	MOVQ  16(SP), AX
+	MOVQ  24(SP), CX
+	LEAQ  (AX)(CX*1), R14
+	MOVQ  s+0(FP), R15
+	ADDQ  R14, 256(R15)
+	MOVQ  ctx+16(FP), R14
+	SUBQ  CX, 104(R14)
+	JS    error_not_enough_literals
+	CMPQ  AX, $0x00020002
+	JA    sequenceDecs_decodeSync_safe_amd64_error_match_len_too_big
+	TESTQ R13, R13
+	JNZ   sequenceDecs_decodeSync_safe_amd64_match_len_ofs_ok
+	TESTQ AX, AX
+	JNZ   sequenceDecs_decodeSync_safe_amd64_error_match_len_ofs_mismatch
+
+sequenceDecs_decodeSync_safe_amd64_match_len_ofs_ok:
+	MOVQ 24(SP), AX
+	MOVQ 8(SP), CX
+	MOVQ 16(SP), R13
+
+	// Check if we have enough space in s.out
+	LEAQ (AX)(R13*1), R14
+	ADDQ R10, R14
+	CMPQ R14, 32(SP)
+	JA   error_not_enough_space
+
+	// Copy literals
+	TESTQ AX, AX
+	JZ    check_offset
+	MOVQ  AX, R14
+	SUBQ  $0x10, R14
+	JB    copy_1_small
+
+copy_1_loop:
+	MOVUPS (R11), X0
+	MOVUPS X0, (R10)
+	ADDQ   $0x10, R11
+	ADDQ   $0x10, R10
+	SUBQ   $0x10, R14
+	JAE    copy_1_loop
+	LEAQ   16(R11)(R14*1), R11
+	LEAQ   16(R10)(R14*1), R10
+	MOVUPS -16(R11), X0
+	MOVUPS X0, -16(R10)
+	JMP    copy_1_end
+
+copy_1_small:
+	CMPQ AX, $0x03
+	JE   copy_1_move_3
+	JB   copy_1_move_1or2
+	CMPQ AX, $0x08
+	JB   copy_1_move_4through7
+	JMP  copy_1_move_8through16
+
+copy_1_move_1or2:
+	MOVB (R11), R14
+	MOVB -1(R11)(AX*1), R15
+	MOVB R14, (R10)
+	MOVB R15, -1(R10)(AX*1)
+	ADDQ AX, R11
+	ADDQ AX, R10
+	JMP  copy_1_end
+
+copy_1_move_3:
+	MOVW (R11), R14
+	MOVB 2(R11), R15
+	MOVW R14, (R10)
+	MOVB R15, 2(R10)
+	ADDQ AX, R11
+	ADDQ AX, R10
+	JMP  copy_1_end
+
+copy_1_move_4through7:
+	MOVL (R11), R14
+	MOVL -4(R11)(AX*1), R15
+	MOVL R14, (R10)
+	MOVL R15, -4(R10)(AX*1)
+	ADDQ AX, R11
+	ADDQ AX, R10
+	JMP  copy_1_end
+
+copy_1_move_8through16:
+	MOVQ (R11), R14
+	MOVQ -8(R11)(AX*1), R15
+	MOVQ R14, (R10)
+	MOVQ R15, -8(R10)(AX*1)
+	ADDQ AX, R11
+	ADDQ AX, R10
+
+copy_1_end:
+	ADDQ AX, R12
+
+	// Malformed input if seq.mo > t+len(hist) || seq.mo > s.windowSize)
+check_offset:
+	MOVQ R12, AX
+	ADDQ 40(SP), AX
+	CMPQ CX, AX
+	JG   error_match_off_too_big
+	CMPQ CX, 56(SP)
+	JG   error_match_off_too_big
+
+	// Copy match from history
+	MOVQ CX, AX
+	SUBQ R12, AX
+	JLS  copy_match
+	MOVQ 48(SP), R14
+	SUBQ AX, R14
+	CMPQ R13, AX
+	JG   copy_all_from_history
+	MOVQ R13, AX
+	SUBQ $0x10, AX
+	JB   copy_4_small
+
+copy_4_loop:
+	MOVUPS (R14), X0
+	MOVUPS X0, (R10)
+	ADDQ   $0x10, R14
+	ADDQ   $0x10, R10
+	SUBQ   $0x10, AX
+	JAE    copy_4_loop
+	LEAQ   16(R14)(AX*1), R14
+	LEAQ   16(R10)(AX*1), R10
+	MOVUPS -16(R14), X0
+	MOVUPS X0, -16(R10)
+	JMP    copy_4_end
+
+copy_4_small:
+	CMPQ R13, $0x03
+	JE   copy_4_move_3
+	CMPQ R13, $0x08
+	JB   copy_4_move_4through7
+	JMP  copy_4_move_8through16
+
+copy_4_move_3:
+	MOVW (R14), AX
+	MOVB 2(R14), CL
+	MOVW AX, (R10)
+	MOVB CL, 2(R10)
+	ADDQ R13, R14
+	ADDQ R13, R10
+	JMP  copy_4_end
+
+copy_4_move_4through7:
+	MOVL (R14), AX
+	MOVL -4(R14)(R13*1), CX
+	MOVL AX, (R10)
+	MOVL CX, -4(R10)(R13*1)
+	ADDQ R13, R14
+	ADDQ R13, R10
+	JMP  copy_4_end
+
+copy_4_move_8through16:
+	MOVQ (R14), AX
+	MOVQ -8(R14)(R13*1), CX
+	MOVQ AX, (R10)
+	MOVQ CX, -8(R10)(R13*1)
+	ADDQ R13, R14
+	ADDQ R13, R10
+
+copy_4_end:
+	ADDQ R13, R12
+	JMP  handle_loop
+	JMP loop_finished
+
+copy_all_from_history:
+	MOVQ AX, R15
+	SUBQ $0x10, R15
+	JB   copy_5_small
+
+copy_5_loop:
+	MOVUPS (R14), X0
+	MOVUPS X0, (R10)
+	ADDQ   $0x10, R14
+	ADDQ   $0x10, R10
+	SUBQ   $0x10, R15
+	JAE    copy_5_loop
+	LEAQ   16(R14)(R15*1), R14
+	LEAQ   16(R10)(R15*1), R10
+	MOVUPS -16(R14), X0
+	MOVUPS X0, -16(R10)
+	JMP    copy_5_end
+
+copy_5_small:
+	CMPQ AX, $0x03
+	JE   copy_5_move_3
+	JB   copy_5_move_1or2
+	CMPQ AX, $0x08
+	JB   copy_5_move_4through7
+	JMP  copy_5_move_8through16
+
+copy_5_move_1or2:
+	MOVB (R14), R15
+	MOVB -1(R14)(AX*1), BP
+	MOVB R15, (R10)
+	MOVB BP, -1(R10)(AX*1)
+	ADDQ AX, R14
+	ADDQ AX, R10
+	JMP  copy_5_end
+
+copy_5_move_3:
+	MOVW (R14), R15
+	MOVB 2(R14), BP
+	MOVW R15, (R10)
+	MOVB BP, 2(R10)
+	ADDQ AX, R14
+	ADDQ AX, R10
+	JMP  copy_5_end
+
+copy_5_move_4through7:
+	MOVL (R14), R15
+	MOVL -4(R14)(AX*1), BP
+	MOVL R15, (R10)
+	MOVL BP, -4(R10)(AX*1)
+	ADDQ AX, R14
+	ADDQ AX, R10
+	JMP  copy_5_end
+
+copy_5_move_8through16:
+	MOVQ (R14), R15
+	MOVQ -8(R14)(AX*1), BP
+	MOVQ R15, (R10)
+	MOVQ BP, -8(R10)(AX*1)
+	ADDQ AX, R14
+	ADDQ AX, R10
+
+copy_5_end:
+	ADDQ AX, R12
+	SUBQ AX, R13
+
+	// Copy match from the current buffer
+copy_match:
+	MOVQ R10, AX
+	SUBQ CX, AX
+
+	// ml <= mo
+	CMPQ R13, CX
+	JA   copy_overlapping_match
+
+	// Copy non-overlapping match
+	ADDQ R13, R12
+	MOVQ R13, CX
+	SUBQ $0x10, CX
+	JB   copy_2_small
+
+copy_2_loop:
+	MOVUPS (AX), X0
+	MOVUPS X0, (R10)
+	ADDQ   $0x10, AX
+	ADDQ   $0x10, R10
+	SUBQ   $0x10, CX
+	JAE    copy_2_loop
+	LEAQ   16(AX)(CX*1), AX
+	LEAQ   16(R10)(CX*1), R10
+	MOVUPS -16(AX), X0
+	MOVUPS X0, -16(R10)
+	JMP    copy_2_end
+
+copy_2_small:
+	CMPQ R13, $0x03
+	JE   copy_2_move_3
+	JB   copy_2_move_1or2
+	CMPQ R13, $0x08
+	JB   copy_2_move_4through7
+	JMP  copy_2_move_8through16
+
+copy_2_move_1or2:
+	MOVB (AX), CL
+	MOVB -1(AX)(R13*1), R14
+	MOVB CL, (R10)
+	MOVB R14, -1(R10)(R13*1)
+	ADDQ R13, AX
+	ADDQ R13, R10
+	JMP  copy_2_end
+
+copy_2_move_3:
+	MOVW (AX), CX
+	MOVB 2(AX), R14
+	MOVW CX, (R10)
+	MOVB R14, 2(R10)
+	ADDQ R13, AX
+	ADDQ R13, R10
+	JMP  copy_2_end
+
+copy_2_move_4through7:
+	MOVL (AX), CX
+	MOVL -4(AX)(R13*1), R14
+	MOVL CX, (R10)
+	MOVL R14, -4(R10)(R13*1)
+	ADDQ R13, AX
+	ADDQ R13, R10
+	JMP  copy_2_end
+
+copy_2_move_8through16:
+	MOVQ (AX), CX
+	MOVQ -8(AX)(R13*1), R14
+	MOVQ CX, (R10)
+	MOVQ R14, -8(R10)(R13*1)
+	ADDQ R13, AX
+	ADDQ R13, R10
+
+copy_2_end:
+	JMP handle_loop
+
+	// Copy overlapping match
+copy_overlapping_match:
+	ADDQ R13, R12
+
+copy_slow_3:
+	MOVB (AX), CL
+	MOVB CL, (R10)
+	INCQ AX
+	INCQ R10
+	DECQ R13
+	JNZ  copy_slow_3
+
+handle_loop:
+	MOVQ ctx+16(FP), AX
+	DECQ 96(AX)
+	JNS  sequenceDecs_decodeSync_safe_amd64_main_loop
+
+loop_finished:
+	MOVQ br+8(FP), AX
+	MOVQ DX, 24(AX)
+	MOVB BL, 32(AX)
+	MOVQ SI, 8(AX)
+
+	// Update the context
+	MOVQ ctx+16(FP), AX
+	MOVQ R12, 136(AX)
+	MOVQ 144(AX), CX
+	SUBQ CX, R11
+	MOVQ R11, 168(AX)
+
+	// Return success
+	MOVQ $0x00000000, ret+24(FP)
+	RET
+
+	// Return with match length error
+sequenceDecs_decodeSync_safe_amd64_error_match_len_ofs_mismatch:
+	MOVQ 16(SP), AX
+	MOVQ ctx+16(FP), CX
+	MOVQ AX, 216(CX)
+	MOVQ $0x00000001, ret+24(FP)
+	RET
+
+	// Return with match too long error
+sequenceDecs_decodeSync_safe_amd64_error_match_len_too_big:
+	MOVQ ctx+16(FP), AX
+	MOVQ 16(SP), CX
+	MOVQ CX, 216(AX)
+	MOVQ $0x00000002, ret+24(FP)
+	RET
+
+	// Return with match offset too long error
+error_match_off_too_big:
+	MOVQ ctx+16(FP), AX
+	MOVQ 8(SP), CX
+	MOVQ CX, 224(AX)
+	MOVQ R12, 136(AX)
+	MOVQ $0x00000003, ret+24(FP)
+	RET
+
+	// Return with not enough literals error
+error_not_enough_literals:
+	MOVQ ctx+16(FP), AX
+	MOVQ 24(SP), CX
+	MOVQ CX, 208(AX)
+	MOVQ $0x00000004, ret+24(FP)
+	RET
+
+	// Return with overread error
+error_overread:
+	MOVQ $0x00000006, ret+24(FP)
+	RET
+
+	// Return with not enough output space error
+error_not_enough_space:
+	MOVQ ctx+16(FP), AX
+	MOVQ 24(SP), CX
+	MOVQ CX, 208(AX)
+	MOVQ 16(SP), CX
+	MOVQ CX, 216(AX)
+	MOVQ R12, 136(AX)
+	MOVQ $0x00000005, ret+24(FP)
+	RET
+
+// func sequenceDecs_decodeSync_safe_bmi2(s *sequenceDecs, br *bitReader, ctx *decodeSyncAsmContext) int
+// Requires: BMI, BMI2, CMOV, SSE
+TEXT ·sequenceDecs_decodeSync_safe_bmi2(SB), $64-32
+	MOVQ    br+8(FP), BX
+	MOVQ    24(BX), AX
+	MOVBQZX 32(BX), DX
+	MOVQ    (BX), CX
+	MOVQ    8(BX), BX
+	ADDQ    BX, CX
+	MOVQ    CX, (SP)
+	MOVQ    ctx+16(FP), CX
+	MOVQ    72(CX), SI
+	MOVQ    80(CX), DI
+	MOVQ    88(CX), R8
+	XORQ    R9, R9
+	MOVQ    R9, 8(SP)
+	MOVQ    R9, 16(SP)
+	MOVQ    R9, 24(SP)
+	MOVQ    112(CX), R9
+	MOVQ    128(CX), R10
+	MOVQ    R10, 32(SP)
+	MOVQ    144(CX), R10
+	MOVQ    136(CX), R11
+	MOVQ    200(CX), R12
+	MOVQ    R12, 56(SP)
+	MOVQ    176(CX), R12
+	MOVQ    R12, 48(SP)
+	MOVQ    184(CX), CX
+	MOVQ    CX, 40(SP)
+	MOVQ    40(SP), CX
+	ADDQ    CX, 48(SP)
+
+	// Calculate poiter to s.out[cap(s.out)] (a past-end pointer)
+	ADDQ R9, 32(SP)
+
+	// outBase += outPosition
+	ADDQ R11, R9
+
+sequenceDecs_decodeSync_safe_bmi2_main_loop:
+	MOVQ (SP), R12
+
+	// Fill bitreader to have enough for the offset and match length.
+	CMPQ BX, $0x08
+	JL   sequenceDecs_decodeSync_safe_bmi2_fill_byte_by_byte
+	MOVQ DX, CX
+	SHRQ $0x03, CX
+	SUBQ CX, R12
+	MOVQ (R12), AX
+	SUBQ CX, BX
+	ANDQ $0x07, DX
+	JMP  sequenceDecs_decodeSync_safe_bmi2_fill_end
+
+sequenceDecs_decodeSync_safe_bmi2_fill_byte_by_byte:
+	CMPQ    BX, $0x00
+	JLE     sequenceDecs_decodeSync_safe_bmi2_fill_check_overread
+	CMPQ    DX, $0x07
+	JLE     sequenceDecs_decodeSync_safe_bmi2_fill_end
+	SHLQ    $0x08, AX
+	SUBQ    $0x01, R12
+	SUBQ    $0x01, BX
+	SUBQ    $0x08, DX
+	MOVBQZX (R12), CX
+	ORQ     CX, AX
+	JMP     sequenceDecs_decodeSync_safe_bmi2_fill_byte_by_byte
+
+sequenceDecs_decodeSync_safe_bmi2_fill_check_overread:
+	CMPQ DX, $0x40
+	JA   error_overread
+
+sequenceDecs_decodeSync_safe_bmi2_fill_end:
+	// Update offset
+	MOVQ   $0x00000808, CX
+	BEXTRQ CX, R8, R13
+	MOVQ   AX, R14
+	LEAQ   (DX)(R13*1), CX
+	ROLQ   CL, R14
+	BZHIQ  R13, R14, R14
+	MOVQ   CX, DX
+	MOVQ   R8, CX
+	SHRQ   $0x20, CX
+	ADDQ   R14, CX
+	MOVQ   CX, 8(SP)
+
+	// Update match length
+	MOVQ   $0x00000808, CX
+	BEXTRQ CX, DI, R13
+	MOVQ   AX, R14
+	LEAQ   (DX)(R13*1), CX
+	ROLQ   CL, R14
+	BZHIQ  R13, R14, R14
+	MOVQ   CX, DX
+	MOVQ   DI, CX
+	SHRQ   $0x20, CX
+	ADDQ   R14, CX
+	MOVQ   CX, 16(SP)
+
+	// Fill bitreader to have enough for the remaining
+	CMPQ BX, $0x08
+	JL   sequenceDecs_decodeSync_safe_bmi2_fill_2_byte_by_byte
+	MOVQ DX, CX
+	SHRQ $0x03, CX
+	SUBQ CX, R12
+	MOVQ (R12), AX
+	SUBQ CX, BX
+	ANDQ $0x07, DX
+	JMP  sequenceDecs_decodeSync_safe_bmi2_fill_2_end
+
+sequenceDecs_decodeSync_safe_bmi2_fill_2_byte_by_byte:
+	CMPQ    BX, $0x00
+	JLE     sequenceDecs_decodeSync_safe_bmi2_fill_2_check_overread
+	CMPQ    DX, $0x07
+	JLE     sequenceDecs_decodeSync_safe_bmi2_fill_2_end
+	SHLQ    $0x08, AX
+	SUBQ    $0x01, R12
+	SUBQ    $0x01, BX
+	SUBQ    $0x08, DX
+	MOVBQZX (R12), CX
+	ORQ     CX, AX
+	JMP     sequenceDecs_decodeSync_safe_bmi2_fill_2_byte_by_byte
+
+sequenceDecs_decodeSync_safe_bmi2_fill_2_check_overread:
+	CMPQ DX, $0x40
+	JA   error_overread
+
+sequenceDecs_decodeSync_safe_bmi2_fill_2_end:
+	// Update literal length
+	MOVQ   $0x00000808, CX
+	BEXTRQ CX, SI, R13
+	MOVQ   AX, R14
+	LEAQ   (DX)(R13*1), CX
+	ROLQ   CL, R14
+	BZHIQ  R13, R14, R14
+	MOVQ   CX, DX
+	MOVQ   SI, CX
+	SHRQ   $0x20, CX
+	ADDQ   R14, CX
+	MOVQ   CX, 24(SP)
+
+	// Fill bitreader for state updates
+	MOVQ    R12, (SP)
+	MOVQ    $0x00000808, CX
+	BEXTRQ  CX, R8, R12
+	MOVQ    ctx+16(FP), CX
+	CMPQ    96(CX), $0x00
+	JZ      sequenceDecs_decodeSync_safe_bmi2_skip_update
+	LEAQ    (SI)(DI*1), R13
+	ADDQ    R8, R13
+	MOVBQZX R13, R13
+	LEAQ    (DX)(R13*1), CX
+	MOVQ    AX, R14
+	MOVQ    CX, DX
+	ROLQ    CL, R14
+	BZHIQ   R13, R14, R14
+
+	// Update Offset State
+	BZHIQ R8, R14, CX
+	SHRXQ R8, R14, R14
+	SHRL  $0x10, R8
+	ADDQ  CX, R8
+
+	// Load ctx.ofTable
+	MOVQ ctx+16(FP), CX
+	MOVQ 48(CX), CX
+	MOVQ (CX)(R8*8), R8
+
+	// Update Match Length State
+	BZHIQ DI, R14, CX
+	SHRXQ DI, R14, R14
+	SHRL  $0x10, DI
+	ADDQ  CX, DI
+
+	// Load ctx.mlTable
+	MOVQ ctx+16(FP), CX
+	MOVQ 24(CX), CX
+	MOVQ (CX)(DI*8), DI
+
+	// Update Literal Length State
+	BZHIQ SI, R14, CX
+	SHRL  $0x10, SI
+	ADDQ  CX, SI
+
+	// Load ctx.llTable
+	MOVQ ctx+16(FP), CX
+	MOVQ (CX), CX
+	MOVQ (CX)(SI*8), SI
+
+sequenceDecs_decodeSync_safe_bmi2_skip_update:
+	// Adjust offset
+	MOVQ   s+0(FP), CX
+	MOVQ   8(SP), R13
+	CMPQ   R12, $0x01
+	JBE    sequenceDecs_decodeSync_safe_bmi2_adjust_offsetB_1_or_0
+	MOVUPS 144(CX), X0
+	MOVQ   R13, 144(CX)
+	MOVUPS X0, 152(CX)
+	JMP    sequenceDecs_decodeSync_safe_bmi2_after_adjust
+
+sequenceDecs_decodeSync_safe_bmi2_adjust_offsetB_1_or_0:
+	CMPQ 24(SP), $0x00000000
+	JNE  sequenceDecs_decodeSync_safe_bmi2_adjust_offset_maybezero
+	INCQ R13
+	JMP  sequenceDecs_decodeSync_safe_bmi2_adjust_offset_nonzero
+
+sequenceDecs_decodeSync_safe_bmi2_adjust_offset_maybezero:
+	TESTQ R13, R13
+	JNZ   sequenceDecs_decodeSync_safe_bmi2_adjust_offset_nonzero
+	MOVQ  144(CX), R13
+	JMP   sequenceDecs_decodeSync_safe_bmi2_after_adjust
+
+sequenceDecs_decodeSync_safe_bmi2_adjust_offset_nonzero:
+	MOVQ    R13, R12
+	XORQ    R14, R14
+	MOVQ    $-1, R15
+	CMPQ    R13, $0x03
+	CMOVQEQ R14, R12
+	CMOVQEQ R15, R14
+	ADDQ    144(CX)(R12*8), R14
+	JNZ     sequenceDecs_decodeSync_safe_bmi2_adjust_temp_valid
+	MOVQ    $0x00000001, R14
+
+sequenceDecs_decodeSync_safe_bmi2_adjust_temp_valid:
+	CMPQ R13, $0x01
+	JZ   sequenceDecs_decodeSync_safe_bmi2_adjust_skip
+	MOVQ 152(CX), R12
+	MOVQ R12, 160(CX)
+
+sequenceDecs_decodeSync_safe_bmi2_adjust_skip:
+	MOVQ 144(CX), R12
+	MOVQ R12, 152(CX)
+	MOVQ R14, 144(CX)
+	MOVQ R14, R13
+
+sequenceDecs_decodeSync_safe_bmi2_after_adjust:
+	MOVQ R13, 8(SP)
+
+	// Check values
+	MOVQ  16(SP), CX
+	MOVQ  24(SP), R12
+	LEAQ  (CX)(R12*1), R14
+	MOVQ  s+0(FP), R15
+	ADDQ  R14, 256(R15)
+	MOVQ  ctx+16(FP), R14
+	SUBQ  R12, 104(R14)
+	JS    error_not_enough_literals
+	CMPQ  CX, $0x00020002
+	JA    sequenceDecs_decodeSync_safe_bmi2_error_match_len_too_big
+	TESTQ R13, R13
+	JNZ   sequenceDecs_decodeSync_safe_bmi2_match_len_ofs_ok
+	TESTQ CX, CX
+	JNZ   sequenceDecs_decodeSync_safe_bmi2_error_match_len_ofs_mismatch
+
+sequenceDecs_decodeSync_safe_bmi2_match_len_ofs_ok:
+	MOVQ 24(SP), CX
+	MOVQ 8(SP), R12
+	MOVQ 16(SP), R13
+
+	// Check if we have enough space in s.out
+	LEAQ (CX)(R13*1), R14
+	ADDQ R9, R14
+	CMPQ R14, 32(SP)
+	JA   error_not_enough_space
+
+	// Copy literals
+	TESTQ CX, CX
+	JZ    check_offset
+	MOVQ  CX, R14
+	SUBQ  $0x10, R14
+	JB    copy_1_small
+
+copy_1_loop:
+	MOVUPS (R10), X0
+	MOVUPS X0, (R9)
+	ADDQ   $0x10, R10
+	ADDQ   $0x10, R9
+	SUBQ   $0x10, R14
+	JAE    copy_1_loop
+	LEAQ   16(R10)(R14*1), R10
+	LEAQ   16(R9)(R14*1), R9
+	MOVUPS -16(R10), X0
+	MOVUPS X0, -16(R9)
+	JMP    copy_1_end
+
+copy_1_small:
+	CMPQ CX, $0x03
+	JE   copy_1_move_3
+	JB   copy_1_move_1or2
+	CMPQ CX, $0x08
+	JB   copy_1_move_4through7
+	JMP  copy_1_move_8through16
+
+copy_1_move_1or2:
+	MOVB (R10), R14
+	MOVB -1(R10)(CX*1), R15
+	MOVB R14, (R9)
+	MOVB R15, -1(R9)(CX*1)
+	ADDQ CX, R10
+	ADDQ CX, R9
+	JMP  copy_1_end
+
+copy_1_move_3:
+	MOVW (R10), R14
+	MOVB 2(R10), R15
+	MOVW R14, (R9)
+	MOVB R15, 2(R9)
+	ADDQ CX, R10
+	ADDQ CX, R9
+	JMP  copy_1_end
+
+copy_1_move_4through7:
+	MOVL (R10), R14
+	MOVL -4(R10)(CX*1), R15
+	MOVL R14, (R9)
+	MOVL R15, -4(R9)(CX*1)
+	ADDQ CX, R10
+	ADDQ CX, R9
+	JMP  copy_1_end
+
+copy_1_move_8through16:
+	MOVQ (R10), R14
+	MOVQ -8(R10)(CX*1), R15
+	MOVQ R14, (R9)
+	MOVQ R15, -8(R9)(CX*1)
+	ADDQ CX, R10
+	ADDQ CX, R9
+
+copy_1_end:
+	ADDQ CX, R11
+
+	// Malformed input if seq.mo > t+len(hist) || seq.mo > s.windowSize)
+check_offset:
+	MOVQ R11, CX
+	ADDQ 40(SP), CX
+	CMPQ R12, CX
+	JG   error_match_off_too_big
+	CMPQ R12, 56(SP)
+	JG   error_match_off_too_big
+
+	// Copy match from history
+	MOVQ R12, CX
+	SUBQ R11, CX
+	JLS  copy_match
+	MOVQ 48(SP), R14
+	SUBQ CX, R14
+	CMPQ R13, CX
+	JG   copy_all_from_history
+	MOVQ R13, CX
+	SUBQ $0x10, CX
+	JB   copy_4_small
+
+copy_4_loop:
+	MOVUPS (R14), X0
+	MOVUPS X0, (R9)
+	ADDQ   $0x10, R14
+	ADDQ   $0x10, R9
+	SUBQ   $0x10, CX
+	JAE    copy_4_loop
+	LEAQ   16(R14)(CX*1), R14
+	LEAQ   16(R9)(CX*1), R9
+	MOVUPS -16(R14), X0
+	MOVUPS X0, -16(R9)
+	JMP    copy_4_end
+
+copy_4_small:
+	CMPQ R13, $0x03
+	JE   copy_4_move_3
+	CMPQ R13, $0x08
+	JB   copy_4_move_4through7
+	JMP  copy_4_move_8through16
+
+copy_4_move_3:
+	MOVW (R14), CX
+	MOVB 2(R14), R12
+	MOVW CX, (R9)
+	MOVB R12, 2(R9)
+	ADDQ R13, R14
+	ADDQ R13, R9
+	JMP  copy_4_end
+
+copy_4_move_4through7:
+	MOVL (R14), CX
+	MOVL -4(R14)(R13*1), R12
+	MOVL CX, (R9)
+	MOVL R12, -4(R9)(R13*1)
+	ADDQ R13, R14
+	ADDQ R13, R9
+	JMP  copy_4_end
+
+copy_4_move_8through16:
+	MOVQ (R14), CX
+	MOVQ -8(R14)(R13*1), R12
+	MOVQ CX, (R9)
+	MOVQ R12, -8(R9)(R13*1)
+	ADDQ R13, R14
+	ADDQ R13, R9
+
+copy_4_end:
+	ADDQ R13, R11
+	JMP  handle_loop
+	JMP loop_finished
+
+copy_all_from_history:
+	MOVQ CX, R15
+	SUBQ $0x10, R15
+	JB   copy_5_small
+
+copy_5_loop:
+	MOVUPS (R14), X0
+	MOVUPS X0, (R9)
+	ADDQ   $0x10, R14
+	ADDQ   $0x10, R9
+	SUBQ   $0x10, R15
+	JAE    copy_5_loop
+	LEAQ   16(R14)(R15*1), R14
+	LEAQ   16(R9)(R15*1), R9
+	MOVUPS -16(R14), X0
+	MOVUPS X0, -16(R9)
+	JMP    copy_5_end
+
+copy_5_small:
+	CMPQ CX, $0x03
+	JE   copy_5_move_3
+	JB   copy_5_move_1or2
+	CMPQ CX, $0x08
+	JB   copy_5_move_4through7
+	JMP  copy_5_move_8through16
+
+copy_5_move_1or2:
+	MOVB (R14), R15
+	MOVB -1(R14)(CX*1), BP
+	MOVB R15, (R9)
+	MOVB BP, -1(R9)(CX*1)
+	ADDQ CX, R14
+	ADDQ CX, R9
+	JMP  copy_5_end
+
+copy_5_move_3:
+	MOVW (R14), R15
+	MOVB 2(R14), BP
+	MOVW R15, (R9)
+	MOVB BP, 2(R9)
+	ADDQ CX, R14
+	ADDQ CX, R9
+	JMP  copy_5_end
+
+copy_5_move_4through7:
+	MOVL (R14), R15
+	MOVL -4(R14)(CX*1), BP
+	MOVL R15, (R9)
+	MOVL BP, -4(R9)(CX*1)
+	ADDQ CX, R14
+	ADDQ CX, R9
+	JMP  copy_5_end
+
+copy_5_move_8through16:
+	MOVQ (R14), R15
+	MOVQ -8(R14)(CX*1), BP
+	MOVQ R15, (R9)
+	MOVQ BP, -8(R9)(CX*1)
+	ADDQ CX, R14
+	ADDQ CX, R9
+
+copy_5_end:
+	ADDQ CX, R11
+	SUBQ CX, R13
+
+	// Copy match from the current buffer
+copy_match:
+	MOVQ R9, CX
+	SUBQ R12, CX
+
+	// ml <= mo
+	CMPQ R13, R12
+	JA   copy_overlapping_match
+
+	// Copy non-overlapping match
+	ADDQ R13, R11
+	MOVQ R13, R12
+	SUBQ $0x10, R12
+	JB   copy_2_small
+
+copy_2_loop:
+	MOVUPS (CX), X0
+	MOVUPS X0, (R9)
+	ADDQ   $0x10, CX
+	ADDQ   $0x10, R9
+	SUBQ   $0x10, R12
+	JAE    copy_2_loop
+	LEAQ   16(CX)(R12*1), CX
+	LEAQ   16(R9)(R12*1), R9
+	MOVUPS -16(CX), X0
+	MOVUPS X0, -16(R9)
+	JMP    copy_2_end
+
+copy_2_small:
+	CMPQ R13, $0x03
+	JE   copy_2_move_3
+	JB   copy_2_move_1or2
+	CMPQ R13, $0x08
+	JB   copy_2_move_4through7
+	JMP  copy_2_move_8through16
+
+copy_2_move_1or2:
+	MOVB (CX), R12
+	MOVB -1(CX)(R13*1), R14
+	MOVB R12, (R9)
+	MOVB R14, -1(R9)(R13*1)
+	ADDQ R13, CX
+	ADDQ R13, R9
+	JMP  copy_2_end
+
+copy_2_move_3:
+	MOVW (CX), R12
+	MOVB 2(CX), R14
+	MOVW R12, (R9)
+	MOVB R14, 2(R9)
+	ADDQ R13, CX
+	ADDQ R13, R9
+	JMP  copy_2_end
+
+copy_2_move_4through7:
+	MOVL (CX), R12
+	MOVL -4(CX)(R13*1), R14
+	MOVL R12, (R9)
+	MOVL R14, -4(R9)(R13*1)
+	ADDQ R13, CX
+	ADDQ R13, R9
+	JMP  copy_2_end
+
+copy_2_move_8through16:
+	MOVQ (CX), R12
+	MOVQ -8(CX)(R13*1), R14
+	MOVQ R12, (R9)
+	MOVQ R14, -8(R9)(R13*1)
+	ADDQ R13, CX
+	ADDQ R13, R9
+
+copy_2_end:
+	JMP handle_loop
+
+	// Copy overlapping match
+copy_overlapping_match:
+	ADDQ R13, R11
+
+copy_slow_3:
+	MOVB (CX), R12
+	MOVB R12, (R9)
+	INCQ CX
+	INCQ R9
+	DECQ R13
+	JNZ  copy_slow_3
+
+handle_loop:
+	MOVQ ctx+16(FP), CX
+	DECQ 96(CX)
+	JNS  sequenceDecs_decodeSync_safe_bmi2_main_loop
+
+loop_finished:
+	MOVQ br+8(FP), CX
+	MOVQ AX, 24(CX)
+	MOVB DL, 32(CX)
+	MOVQ BX, 8(CX)
+
+	// Update the context
+	MOVQ ctx+16(FP), AX
+	MOVQ R11, 136(AX)
+	MOVQ 144(AX), CX
+	SUBQ CX, R10
+	MOVQ R10, 168(AX)
+
+	// Return success
+	MOVQ $0x00000000, ret+24(FP)
+	RET
+
+	// Return with match length error
+sequenceDecs_decodeSync_safe_bmi2_error_match_len_ofs_mismatch:
+	MOVQ 16(SP), AX
+	MOVQ ctx+16(FP), CX
+	MOVQ AX, 216(CX)
+	MOVQ $0x00000001, ret+24(FP)
+	RET
+
+	// Return with match too long error
+sequenceDecs_decodeSync_safe_bmi2_error_match_len_too_big:
+	MOVQ ctx+16(FP), AX
+	MOVQ 16(SP), CX
+	MOVQ CX, 216(AX)
+	MOVQ $0x00000002, ret+24(FP)
+	RET
+
+	// Return with match offset too long error
+error_match_off_too_big:
+	MOVQ ctx+16(FP), AX
+	MOVQ 8(SP), CX
+	MOVQ CX, 224(AX)
+	MOVQ R11, 136(AX)
+	MOVQ $0x00000003, ret+24(FP)
+	RET
+
+	// Return with not enough literals error
+error_not_enough_literals:
+	MOVQ ctx+16(FP), AX
+	MOVQ 24(SP), CX
+	MOVQ CX, 208(AX)
+	MOVQ $0x00000004, ret+24(FP)
+	RET
+
+	// Return with overread error
+error_overread:
+	MOVQ $0x00000006, ret+24(FP)
+	RET
+
+	// Return with not enough output space error
+error_not_enough_space:
+	MOVQ ctx+16(FP), AX
+	MOVQ 24(SP), CX
+	MOVQ CX, 208(AX)
+	MOVQ 16(SP), CX
+	MOVQ CX, 216(AX)
+	MOVQ R11, 136(AX)
+	MOVQ $0x00000005, ret+24(FP)
+	RET
diff --git a/vendor/github.com/klauspost/compress/zstd/seqdec_generic.go b/vendor/github.com/klauspost/compress/zstd/seqdec_generic.go
new file mode 100644
index 000000000..2fb35b788
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/seqdec_generic.go
@@ -0,0 +1,237 @@
+//go:build !amd64 || appengine || !gc || noasm
+// +build !amd64 appengine !gc noasm
+
+package zstd
+
+import (
+	"fmt"
+	"io"
+)
+
+// decode sequences from the stream with the provided history but without dictionary.
+func (s *sequenceDecs) decodeSyncSimple(hist []byte) (bool, error) {
+	return false, nil
+}
+
+// decode sequences from the stream without the provided history.
+func (s *sequenceDecs) decode(seqs []seqVals) error {
+	br := s.br
+
+	// Grab full sizes tables, to avoid bounds checks.
+	llTable, mlTable, ofTable := s.litLengths.fse.dt[:maxTablesize], s.matchLengths.fse.dt[:maxTablesize], s.offsets.fse.dt[:maxTablesize]
+	llState, mlState, ofState := s.litLengths.state.state, s.matchLengths.state.state, s.offsets.state.state
+	s.seqSize = 0
+	litRemain := len(s.literals)
+
+	maxBlockSize := maxCompressedBlockSize
+	if s.windowSize < maxBlockSize {
+		maxBlockSize = s.windowSize
+	}
+	for i := range seqs {
+		var ll, mo, ml int
+		if len(br.in) > 4+((maxOffsetBits+16+16)>>3) {
+			// inlined function:
+			// ll, mo, ml = s.nextFast(br, llState, mlState, ofState)
+
+			// Final will not read from stream.
+			var llB, mlB, moB uint8
+			ll, llB = llState.final()
+			ml, mlB = mlState.final()
+			mo, moB = ofState.final()
+
+			// extra bits are stored in reverse order.
+			br.fillFast()
+			mo += br.getBits(moB)
+			if s.maxBits > 32 {
+				br.fillFast()
+			}
+			ml += br.getBits(mlB)
+			ll += br.getBits(llB)
+
+			if moB > 1 {
+				s.prevOffset[2] = s.prevOffset[1]
+				s.prevOffset[1] = s.prevOffset[0]
+				s.prevOffset[0] = mo
+			} else {
+				// mo = s.adjustOffset(mo, ll, moB)
+				// Inlined for rather big speedup
+				if ll == 0 {
+					// There is an exception though, when current sequence's literals_length = 0.
+					// In this case, repeated offsets are shifted by one, so an offset_value of 1 means Repeated_Offset2,
+					// an offset_value of 2 means Repeated_Offset3, and an offset_value of 3 means Repeated_Offset1 - 1_byte.
+					mo++
+				}
+
+				if mo == 0 {
+					mo = s.prevOffset[0]
+				} else {
+					var temp int
+					if mo == 3 {
+						temp = s.prevOffset[0] - 1
+					} else {
+						temp = s.prevOffset[mo]
+					}
+
+					if temp == 0 {
+						// 0 is not valid; input is corrupted; force offset to 1
+						println("WARNING: temp was 0")
+						temp = 1
+					}
+
+					if mo != 1 {
+						s.prevOffset[2] = s.prevOffset[1]
+					}
+					s.prevOffset[1] = s.prevOffset[0]
+					s.prevOffset[0] = temp
+					mo = temp
+				}
+			}
+			br.fillFast()
+		} else {
+			if br.overread() {
+				if debugDecoder {
+					printf("reading sequence %d, exceeded available data\n", i)
+				}
+				return io.ErrUnexpectedEOF
+			}
+			ll, mo, ml = s.next(br, llState, mlState, ofState)
+			br.fill()
+		}
+
+		if debugSequences {
+			println("Seq", i, "Litlen:", ll, "mo:", mo, "(abs) ml:", ml)
+		}
+		// Evaluate.
+		// We might be doing this async, so do it early.
+		if mo == 0 && ml > 0 {
+			return fmt.Errorf("zero matchoff and matchlen (%d) > 0", ml)
+		}
+		if ml > maxMatchLen {
+			return fmt.Errorf("match len (%d) bigger than max allowed length", ml)
+		}
+		s.seqSize += ll + ml
+		if s.seqSize > maxBlockSize {
+			return fmt.Errorf("output bigger than max block size (%d)", maxBlockSize)
+		}
+		litRemain -= ll
+		if litRemain < 0 {
+			return fmt.Errorf("unexpected literal count, want %d bytes, but only %d is available", ll, litRemain+ll)
+		}
+		seqs[i] = seqVals{
+			ll: ll,
+			ml: ml,
+			mo: mo,
+		}
+		if i == len(seqs)-1 {
+			// This is the last sequence, so we shouldn't update state.
+			break
+		}
+
+		// Manually inlined, ~ 5-20% faster
+		// Update all 3 states at once. Approx 20% faster.
+		nBits := llState.nbBits() + mlState.nbBits() + ofState.nbBits()
+		if nBits == 0 {
+			llState = llTable[llState.newState()&maxTableMask]
+			mlState = mlTable[mlState.newState()&maxTableMask]
+			ofState = ofTable[ofState.newState()&maxTableMask]
+		} else {
+			bits := br.get32BitsFast(nBits)
+			lowBits := uint16(bits >> ((ofState.nbBits() + mlState.nbBits()) & 31))
+			llState = llTable[(llState.newState()+lowBits)&maxTableMask]
+
+			lowBits = uint16(bits >> (ofState.nbBits() & 31))
+			lowBits &= bitMask[mlState.nbBits()&15]
+			mlState = mlTable[(mlState.newState()+lowBits)&maxTableMask]
+
+			lowBits = uint16(bits) & bitMask[ofState.nbBits()&15]
+			ofState = ofTable[(ofState.newState()+lowBits)&maxTableMask]
+		}
+	}
+	s.seqSize += litRemain
+	if s.seqSize > maxBlockSize {
+		return fmt.Errorf("output bigger than max block size (%d)", maxBlockSize)
+	}
+	err := br.close()
+	if err != nil {
+		printf("Closing sequences: %v, %+v\n", err, *br)
+	}
+	return err
+}
+
+// executeSimple handles cases when a dictionary is not used.
+func (s *sequenceDecs) executeSimple(seqs []seqVals, hist []byte) error {
+	// Ensure we have enough output size...
+	if len(s.out)+s.seqSize > cap(s.out) {
+		addBytes := s.seqSize + len(s.out)
+		s.out = append(s.out, make([]byte, addBytes)...)
+		s.out = s.out[:len(s.out)-addBytes]
+	}
+
+	if debugDecoder {
+		printf("Execute %d seqs with literals: %d into %d bytes\n", len(seqs), len(s.literals), s.seqSize)
+	}
+
+	var t = len(s.out)
+	out := s.out[:t+s.seqSize]
+
+	for _, seq := range seqs {
+		// Add literals
+		copy(out[t:], s.literals[:seq.ll])
+		t += seq.ll
+		s.literals = s.literals[seq.ll:]
+
+		// Malformed input
+		if seq.mo > t+len(hist) || seq.mo > s.windowSize {
+			return fmt.Errorf("match offset (%d) bigger than current history (%d)", seq.mo, t+len(hist))
+		}
+
+		// Copy from history.
+		if v := seq.mo - t; v > 0 {
+			// v is the start position in history from end.
+			start := len(hist) - v
+			if seq.ml > v {
+				// Some goes into the current block.
+				// Copy remainder of history
+				copy(out[t:], hist[start:])
+				t += v
+				seq.ml -= v
+			} else {
+				copy(out[t:], hist[start:start+seq.ml])
+				t += seq.ml
+				continue
+			}
+		}
+
+		// We must be in the current buffer now
+		if seq.ml > 0 {
+			start := t - seq.mo
+			if seq.ml <= t-start {
+				// No overlap
+				copy(out[t:], out[start:start+seq.ml])
+				t += seq.ml
+			} else {
+				// Overlapping copy
+				// Extend destination slice and copy one byte at the time.
+				src := out[start : start+seq.ml]
+				dst := out[t:]
+				dst = dst[:len(src)]
+				t += len(src)
+				// Destination is the space we just added.
+				for i := range src {
+					dst[i] = src[i]
+				}
+			}
+		}
+	}
+	// Add final literals
+	copy(out[t:], s.literals)
+	if debugDecoder {
+		t += len(s.literals)
+		if t != len(out) {
+			panic(fmt.Errorf("length mismatch, want %d, got %d, ss: %d", len(out), t, s.seqSize))
+		}
+	}
+	s.out = out
+
+	return nil
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/seqenc.go b/vendor/github.com/klauspost/compress/zstd/seqenc.go
new file mode 100644
index 000000000..8014174a7
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/seqenc.go
@@ -0,0 +1,114 @@
+// 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 "math/bits"
+
+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 {
+		// Compiler insists on bounds check (Go 1.12)
+		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 {
+		// Compiler insists on bounds check (Go 1.12)
+		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/vendor/github.com/klauspost/compress/zstd/snappy.go b/vendor/github.com/klauspost/compress/zstd/snappy.go
new file mode 100644
index 000000000..ec13594e8
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/snappy.go
@@ -0,0 +1,434 @@
+// 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"
+	"hash/crc32"
+	"io"
+
+	"github.com/klauspost/compress/huff0"
+	snappy "github.com/klauspost/compress/internal/snapref"
+)
+
+const (
+	snappyTagLiteral = 0x00
+	snappyTagCopy1   = 0x01
+	snappyTagCopy2   = 0x02
+	snappyTagCopy4   = 0x03
+)
+
+const (
+	snappyChecksumSize = 4
+	snappyMagicBody    = "sNaPpY"
+
+	// snappyMaxBlockSize is the maximum size of the input to encodeBlock. It is not
+	// part of the wire format per se, but some parts of the encoder assume
+	// that an offset fits into a uint16.
+	//
+	// Also, for the framing format (Writer type instead of Encode function),
+	// https://github.com/google/snappy/blob/master/framing_format.txt says
+	// that "the uncompressed data in a chunk must be no longer than 65536
+	// bytes".
+	snappyMaxBlockSize = 65536
+
+	// snappyMaxEncodedLenOfMaxBlockSize equals MaxEncodedLen(snappyMaxBlockSize), but is
+	// hard coded to be a const instead of a variable, so that obufLen can also
+	// be a const. Their equivalence is confirmed by
+	// TestMaxEncodedLenOfMaxBlockSize.
+	snappyMaxEncodedLenOfMaxBlockSize = 76490
+)
+
+const (
+	chunkTypeCompressedData   = 0x00
+	chunkTypeUncompressedData = 0x01
+	chunkTypePadding          = 0xfe
+	chunkTypeStreamIdentifier = 0xff
+)
+
+var (
+	// ErrSnappyCorrupt reports that the input is invalid.
+	ErrSnappyCorrupt = errors.New("snappy: corrupt input")
+	// ErrSnappyTooLarge reports that the uncompressed length is too large.
+	ErrSnappyTooLarge = errors.New("snappy: decoded block is too large")
+	// ErrSnappyUnsupported reports that the input isn't supported.
+	ErrSnappyUnsupported = errors.New("snappy: unsupported input")
+
+	errUnsupportedLiteralLength = errors.New("snappy: unsupported literal length")
+)
+
+// SnappyConverter can read SnappyConverter-compressed streams and convert them to zstd.
+// Conversion is done by converting the stream directly from Snappy without intermediate
+// full decoding.
+// Therefore the compression ratio is much less than what can be done by a full decompression
+// and compression, and a faulty Snappy stream may lead to a faulty Zstandard stream without
+// any errors being generated.
+// No CRC value is being generated and not all CRC values of the Snappy stream are checked.
+// However, it provides really fast recompression of Snappy streams.
+// The converter can be reused to avoid allocations, even after errors.
+type SnappyConverter struct {
+	r     io.Reader
+	err   error
+	buf   []byte
+	block *blockEnc
+}
+
+// Convert the Snappy stream supplied in 'in' and write the zStandard stream to 'w'.
+// If any error is detected on the Snappy stream it is returned.
+// The number of bytes written is returned.
+func (r *SnappyConverter) Convert(in io.Reader, w io.Writer) (int64, error) {
+	initPredefined()
+	r.err = nil
+	r.r = in
+	if r.block == nil {
+		r.block = &blockEnc{}
+		r.block.init()
+	}
+	r.block.initNewEncode()
+	if len(r.buf) != snappyMaxEncodedLenOfMaxBlockSize+snappyChecksumSize {
+		r.buf = make([]byte, snappyMaxEncodedLenOfMaxBlockSize+snappyChecksumSize)
+	}
+	r.block.litEnc.Reuse = huff0.ReusePolicyNone
+	var written int64
+	var readHeader bool
+	{
+		header := frameHeader{WindowSize: snappyMaxBlockSize}.appendTo(r.buf[:0])
+
+		var n int
+		n, r.err = w.Write(header)
+		if r.err != nil {
+			return written, r.err
+		}
+		written += int64(n)
+	}
+
+	for {
+		if !r.readFull(r.buf[:4], true) {
+			// Add empty last block
+			r.block.reset(nil)
+			r.block.last = true
+			err := r.block.encodeLits(r.block.literals, false)
+			if err != nil {
+				return written, err
+			}
+			n, err := w.Write(r.block.output)
+			if err != nil {
+				return written, err
+			}
+			written += int64(n)
+
+			return written, r.err
+		}
+		chunkType := r.buf[0]
+		if !readHeader {
+			if chunkType != chunkTypeStreamIdentifier {
+				println("chunkType != chunkTypeStreamIdentifier", chunkType)
+				r.err = ErrSnappyCorrupt
+				return written, r.err
+			}
+			readHeader = true
+		}
+		chunkLen := int(r.buf[1]) | int(r.buf[2])<<8 | int(r.buf[3])<<16
+		if chunkLen > len(r.buf) {
+			println("chunkLen > len(r.buf)", chunkType)
+			r.err = ErrSnappyUnsupported
+			return written, r.err
+		}
+
+		// The chunk types are specified at
+		// https://github.com/google/snappy/blob/master/framing_format.txt
+		switch chunkType {
+		case chunkTypeCompressedData:
+			// Section 4.2. Compressed data (chunk type 0x00).
+			if chunkLen < snappyChecksumSize {
+				println("chunkLen < snappyChecksumSize", chunkLen, snappyChecksumSize)
+				r.err = ErrSnappyCorrupt
+				return written, r.err
+			}
+			buf := r.buf[:chunkLen]
+			if !r.readFull(buf, false) {
+				return written, r.err
+			}
+			//checksum := uint32(buf[0]) | uint32(buf[1])<<8 | uint32(buf[2])<<16 | uint32(buf[3])<<24
+			buf = buf[snappyChecksumSize:]
+
+			n, hdr, err := snappyDecodedLen(buf)
+			if err != nil {
+				r.err = err
+				return written, r.err
+			}
+			buf = buf[hdr:]
+			if n > snappyMaxBlockSize {
+				println("n > snappyMaxBlockSize", n, snappyMaxBlockSize)
+				r.err = ErrSnappyCorrupt
+				return written, r.err
+			}
+			r.block.reset(nil)
+			r.block.pushOffsets()
+			if err := decodeSnappy(r.block, buf); err != nil {
+				r.err = err
+				return written, r.err
+			}
+			if r.block.size+r.block.extraLits != n {
+				printf("invalid size, want %d, got %d\n", n, r.block.size+r.block.extraLits)
+				r.err = ErrSnappyCorrupt
+				return written, r.err
+			}
+			err = r.block.encode(nil, false, false)
+			switch err {
+			case errIncompressible:
+				r.block.popOffsets()
+				r.block.reset(nil)
+				r.block.literals, err = snappy.Decode(r.block.literals[:n], r.buf[snappyChecksumSize:chunkLen])
+				if err != nil {
+					return written, err
+				}
+				err = r.block.encodeLits(r.block.literals, false)
+				if err != nil {
+					return written, err
+				}
+			case nil:
+			default:
+				return written, err
+			}
+
+			n, r.err = w.Write(r.block.output)
+			if r.err != nil {
+				return written, err
+			}
+			written += int64(n)
+			continue
+		case chunkTypeUncompressedData:
+			if debugEncoder {
+				println("Uncompressed, chunklen", chunkLen)
+			}
+			// Section 4.3. Uncompressed data (chunk type 0x01).
+			if chunkLen < snappyChecksumSize {
+				println("chunkLen < snappyChecksumSize", chunkLen, snappyChecksumSize)
+				r.err = ErrSnappyCorrupt
+				return written, r.err
+			}
+			r.block.reset(nil)
+			buf := r.buf[:snappyChecksumSize]
+			if !r.readFull(buf, false) {
+				return written, r.err
+			}
+			checksum := uint32(buf[0]) | uint32(buf[1])<<8 | uint32(buf[2])<<16 | uint32(buf[3])<<24
+			// Read directly into r.decoded instead of via r.buf.
+			n := chunkLen - snappyChecksumSize
+			if n > snappyMaxBlockSize {
+				println("n > snappyMaxBlockSize", n, snappyMaxBlockSize)
+				r.err = ErrSnappyCorrupt
+				return written, r.err
+			}
+			r.block.literals = r.block.literals[:n]
+			if !r.readFull(r.block.literals, false) {
+				return written, r.err
+			}
+			if snappyCRC(r.block.literals) != checksum {
+				println("literals crc mismatch")
+				r.err = ErrSnappyCorrupt
+				return written, r.err
+			}
+			err := r.block.encodeLits(r.block.literals, false)
+			if err != nil {
+				return written, err
+			}
+			n, r.err = w.Write(r.block.output)
+			if r.err != nil {
+				return written, err
+			}
+			written += int64(n)
+			continue
+
+		case chunkTypeStreamIdentifier:
+			if debugEncoder {
+				println("stream id", chunkLen, len(snappyMagicBody))
+			}
+			// Section 4.1. Stream identifier (chunk type 0xff).
+			if chunkLen != len(snappyMagicBody) {
+				println("chunkLen != len(snappyMagicBody)", chunkLen, len(snappyMagicBody))
+				r.err = ErrSnappyCorrupt
+				return written, r.err
+			}
+			if !r.readFull(r.buf[:len(snappyMagicBody)], false) {
+				return written, r.err
+			}
+			for i := 0; i < len(snappyMagicBody); i++ {
+				if r.buf[i] != snappyMagicBody[i] {
+					println("r.buf[i] != snappyMagicBody[i]", r.buf[i], snappyMagicBody[i], i)
+					r.err = ErrSnappyCorrupt
+					return written, r.err
+				}
+			}
+			continue
+		}
+
+		if chunkType <= 0x7f {
+			// Section 4.5. Reserved unskippable chunks (chunk types 0x02-0x7f).
+			println("chunkType <= 0x7f")
+			r.err = ErrSnappyUnsupported
+			return written, r.err
+		}
+		// Section 4.4 Padding (chunk type 0xfe).
+		// Section 4.6. Reserved skippable chunks (chunk types 0x80-0xfd).
+		if !r.readFull(r.buf[:chunkLen], false) {
+			return written, r.err
+		}
+	}
+}
+
+// decodeSnappy writes the decoding of src to dst. It assumes that the varint-encoded
+// length of the decompressed bytes has already been read.
+func decodeSnappy(blk *blockEnc, src []byte) error {
+	//decodeRef(make([]byte, snappyMaxBlockSize), src)
+	var s, length int
+	lits := blk.extraLits
+	var offset uint32
+	for s < len(src) {
+		switch src[s] & 0x03 {
+		case snappyTagLiteral:
+			x := uint32(src[s] >> 2)
+			switch {
+			case x < 60:
+				s++
+			case x == 60:
+				s += 2
+				if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+					println("uint(s) > uint(len(src)", s, src)
+					return ErrSnappyCorrupt
+				}
+				x = uint32(src[s-1])
+			case x == 61:
+				s += 3
+				if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+					println("uint(s) > uint(len(src)", s, src)
+					return ErrSnappyCorrupt
+				}
+				x = uint32(src[s-2]) | uint32(src[s-1])<<8
+			case x == 62:
+				s += 4
+				if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+					println("uint(s) > uint(len(src)", s, src)
+					return ErrSnappyCorrupt
+				}
+				x = uint32(src[s-3]) | uint32(src[s-2])<<8 | uint32(src[s-1])<<16
+			case x == 63:
+				s += 5
+				if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+					println("uint(s) > uint(len(src)", s, src)
+					return ErrSnappyCorrupt
+				}
+				x = uint32(src[s-4]) | uint32(src[s-3])<<8 | uint32(src[s-2])<<16 | uint32(src[s-1])<<24
+			}
+			if x > snappyMaxBlockSize {
+				println("x > snappyMaxBlockSize", x, snappyMaxBlockSize)
+				return ErrSnappyCorrupt
+			}
+			length = int(x) + 1
+			if length <= 0 {
+				println("length <= 0 ", length)
+
+				return errUnsupportedLiteralLength
+			}
+			//if length > snappyMaxBlockSize-d || uint32(length) > len(src)-s {
+			//	return ErrSnappyCorrupt
+			//}
+
+			blk.literals = append(blk.literals, src[s:s+length]...)
+			//println(length, "litLen")
+			lits += length
+			s += length
+			continue
+
+		case snappyTagCopy1:
+			s += 2
+			if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+				println("uint(s) > uint(len(src)", s, len(src))
+				return ErrSnappyCorrupt
+			}
+			length = 4 + int(src[s-2])>>2&0x7
+			offset = uint32(src[s-2])&0xe0<<3 | uint32(src[s-1])
+
+		case snappyTagCopy2:
+			s += 3
+			if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+				println("uint(s) > uint(len(src)", s, len(src))
+				return ErrSnappyCorrupt
+			}
+			length = 1 + int(src[s-3])>>2
+			offset = uint32(src[s-2]) | uint32(src[s-1])<<8
+
+		case snappyTagCopy4:
+			s += 5
+			if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+				println("uint(s) > uint(len(src)", s, len(src))
+				return ErrSnappyCorrupt
+			}
+			length = 1 + int(src[s-5])>>2
+			offset = uint32(src[s-4]) | uint32(src[s-3])<<8 | uint32(src[s-2])<<16 | uint32(src[s-1])<<24
+		}
+
+		if offset <= 0 || blk.size+lits < int(offset) /*|| length > len(blk)-d */ {
+			println("offset <= 0 || blk.size+lits < int(offset)", offset, blk.size+lits, int(offset), blk.size, lits)
+
+			return ErrSnappyCorrupt
+		}
+
+		// Check if offset is one of the recent offsets.
+		// Adjusts the output offset accordingly.
+		// Gives a tiny bit of compression, typically around 1%.
+		if false {
+			offset = blk.matchOffset(offset, uint32(lits))
+		} else {
+			offset += 3
+		}
+
+		blk.sequences = append(blk.sequences, seq{
+			litLen:   uint32(lits),
+			offset:   offset,
+			matchLen: uint32(length) - zstdMinMatch,
+		})
+		blk.size += length + lits
+		lits = 0
+	}
+	blk.extraLits = lits
+	return nil
+}
+
+func (r *SnappyConverter) readFull(p []byte, allowEOF bool) (ok bool) {
+	if _, r.err = io.ReadFull(r.r, p); r.err != nil {
+		if r.err == io.ErrUnexpectedEOF || (r.err == io.EOF && !allowEOF) {
+			r.err = ErrSnappyCorrupt
+		}
+		return false
+	}
+	return true
+}
+
+var crcTable = crc32.MakeTable(crc32.Castagnoli)
+
+// crc implements the checksum specified in section 3 of
+// https://github.com/google/snappy/blob/master/framing_format.txt
+func snappyCRC(b []byte) uint32 {
+	c := crc32.Update(0, crcTable, b)
+	return c>>15 | c<<17 + 0xa282ead8
+}
+
+// snappyDecodedLen returns the length of the decoded block and the number of bytes
+// that the length header occupied.
+func snappyDecodedLen(src []byte) (blockLen, headerLen int, err error) {
+	v, n := binary.Uvarint(src)
+	if n <= 0 || v > 0xffffffff {
+		return 0, 0, ErrSnappyCorrupt
+	}
+
+	const wordSize = 32 << (^uint(0) >> 32 & 1)
+	if wordSize == 32 && v > 0x7fffffff {
+		return 0, 0, ErrSnappyTooLarge
+	}
+	return int(v), n, nil
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/zip.go b/vendor/github.com/klauspost/compress/zstd/zip.go
new file mode 100644
index 000000000..29c15c8c4
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/zip.go
@@ -0,0 +1,141 @@
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+
+package zstd
+
+import (
+	"errors"
+	"io"
+	"sync"
+)
+
+// ZipMethodWinZip is the method for Zstandard compressed data inside Zip files for WinZip.
+// See https://www.winzip.com/win/en/comp_info.html
+const ZipMethodWinZip = 93
+
+// ZipMethodPKWare is the original method number used by PKWARE to indicate Zstandard compression.
+// Deprecated: This has been deprecated by PKWARE, use ZipMethodWinZip instead for compression.
+// See https://pkware.cachefly.net/webdocs/APPNOTE/APPNOTE-6.3.9.TXT
+const ZipMethodPKWare = 20
+
+// zipReaderPool is the default reader pool.
+var zipReaderPool = sync.Pool{New: func() interface{} {
+	z, err := NewReader(nil, WithDecoderLowmem(true), WithDecoderMaxWindow(128<<20), WithDecoderConcurrency(1))
+	if err != nil {
+		panic(err)
+	}
+	return z
+}}
+
+// newZipReader creates a pooled zip decompressor.
+func newZipReader(opts ...DOption) func(r io.Reader) io.ReadCloser {
+	pool := &zipReaderPool
+	if len(opts) > 0 {
+		opts = append([]DOption{WithDecoderLowmem(true), WithDecoderMaxWindow(128 << 20)}, opts...)
+		// Force concurrency 1
+		opts = append(opts, WithDecoderConcurrency(1))
+		// Create our own pool
+		pool = &sync.Pool{}
+	}
+	return func(r io.Reader) io.ReadCloser {
+		dec, ok := pool.Get().(*Decoder)
+		if ok {
+			dec.Reset(r)
+		} else {
+			d, err := NewReader(r, opts...)
+			if err != nil {
+				panic(err)
+			}
+			dec = d
+		}
+		return &pooledZipReader{dec: dec, pool: pool}
+	}
+}
+
+type pooledZipReader struct {
+	mu   sync.Mutex // guards Close and Read
+	pool *sync.Pool
+	dec  *Decoder
+}
+
+func (r *pooledZipReader) Read(p []byte) (n int, err error) {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	if r.dec == nil {
+		return 0, errors.New("read after close or EOF")
+	}
+	dec, err := r.dec.Read(p)
+	if err == io.EOF {
+		r.dec.Reset(nil)
+		r.pool.Put(r.dec)
+		r.dec = nil
+	}
+	return dec, err
+}
+
+func (r *pooledZipReader) Close() error {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	var err error
+	if r.dec != nil {
+		err = r.dec.Reset(nil)
+		r.pool.Put(r.dec)
+		r.dec = nil
+	}
+	return err
+}
+
+type pooledZipWriter struct {
+	mu   sync.Mutex // guards Close and Read
+	enc  *Encoder
+	pool *sync.Pool
+}
+
+func (w *pooledZipWriter) Write(p []byte) (n int, err error) {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	if w.enc == nil {
+		return 0, errors.New("Write after Close")
+	}
+	return w.enc.Write(p)
+}
+
+func (w *pooledZipWriter) Close() error {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	var err error
+	if w.enc != nil {
+		err = w.enc.Close()
+		w.pool.Put(w.enc)
+		w.enc = nil
+	}
+	return err
+}
+
+// ZipCompressor returns a compressor that can be registered with zip libraries.
+// The provided encoder options will be used on all encodes.
+func ZipCompressor(opts ...EOption) func(w io.Writer) (io.WriteCloser, error) {
+	var pool sync.Pool
+	return func(w io.Writer) (io.WriteCloser, error) {
+		enc, ok := pool.Get().(*Encoder)
+		if ok {
+			enc.Reset(w)
+		} else {
+			var err error
+			enc, err = NewWriter(w, opts...)
+			if err != nil {
+				return nil, err
+			}
+		}
+		return &pooledZipWriter{enc: enc, pool: &pool}, nil
+	}
+}
+
+// ZipDecompressor returns a decompressor that can be registered with zip libraries.
+// See ZipCompressor for example.
+// Options can be specified. WithDecoderConcurrency(1) is forced,
+// and by default a 128MB maximum decompression window is specified.
+// The window size can be overridden if required.
+func ZipDecompressor(opts ...DOption) func(r io.Reader) io.ReadCloser {
+	return newZipReader(opts...)
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/zstd.go b/vendor/github.com/klauspost/compress/zstd/zstd.go
new file mode 100644
index 000000000..4be7cc736
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/zstd.go
@@ -0,0 +1,121 @@
+// 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"
+	"encoding/binary"
+	"errors"
+	"log"
+	"math"
+)
+
+// 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")
+
+	// 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 binary.LittleEndian.Uint32(b[:len(b):len(b)][i:])
+}
+
+func load6432(b []byte, i int32) uint64 {
+	return binary.LittleEndian.Uint64(b[:len(b):len(b)][i:])
+}
+
+type byter interface {
+	Bytes() []byte
+	Len() int
+}
+
+var _ byter = &bytes.Buffer{}