diff --git a/vendor/github.com/klauspost/compress/zstd/README.md b/vendor/github.com/klauspost/compress/zstd/README.md
index e7d7eb0..c11d7fa 100644
--- a/vendor/github.com/klauspost/compress/zstd/README.md
+++ b/vendor/github.com/klauspost/compress/zstd/README.md
@@ -6,18 +6,19 @@
 
 This package provides [compression](#Compressor) to and [decompression](#Decompressor) of Zstandard content. 
 
-This package is pure Go and without use of "unsafe". 
+This package is pure Go. Use `noasm` and `nounsafe` to disable relevant features.
 
 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`.
 
-Godoc Documentation: https://godoc.org/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
 
@@ -79,6 +80,9 @@
 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.
 
@@ -105,7 +109,8 @@
 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, but each call will only run on a single goroutine.
+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`.
@@ -150,10 +155,10 @@
 
 This package:
 file    out     level   insize      outsize     millis  mb/s
-silesia.tar zskp    1   211947520   73101992    643     313.87
-silesia.tar zskp    2   211947520   67504318    969     208.38
-silesia.tar zskp    3   211947520   64595893    2007    100.68
-silesia.tar zskp    4   211947520   60995370    7691    26.28
+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
@@ -162,89 +167,99 @@
 silesia.tar zstd    9   211947520   60212393    5063    39.92
 
 gzip, stdlib/this package:
-silesia.tar gzstd   1   211947520   80007735    1654    122.21
-silesia.tar gzkp    1   211947520   80369488    1168    173.06
+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  235022249   3088    590.30
-gob-stream  zskp    2   1911399616  205669791   3786    481.34
-gob-stream  zskp    3   1911399616  175034659   9636    189.17
-gob-stream  zskp    4   1911399616  167273881   29337   62.13
+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  357382641   10251   177.82
-gob-stream  gzkp    1   1911399616  362156523   5695    320.08
+
+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  343848582   3609    264.18
-enwik9  zskp    2   1000000000  317276632   5746    165.97
-enwik9  zskp    3   1000000000  292243069   12162   78.41
-enwik9  zskp    4   1000000000  275241169   36430   26.18
+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   9604    99.30
-enwik9  gzkp    1   1000000000  383825945   6544    145.73
+
+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  699045015   10620   563.40
-github-june-2days-2019.json zskp    2   6273951764  617881763   11687   511.96
-github-june-2days-2019.json zskp    3   6273951764  524340691   34043   175.75
-github-june-2days-2019.json zskp    4   6273951764  503314661   93811   63.78
+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  1164400847  29948   199.79
-github-june-2days-2019.json gzkp    1   6273951764  1128755542  19236   311.03
+
+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  3667489370  20210   403.84
-rawstudio-mint14.tar    zskp    2   8558382592  3364592300  31873   256.07
-rawstudio-mint14.tar    zskp    3   8558382592  3158085214  77675   105.08
-rawstudio-mint14.tar    zskp    4   8558382592  3020370044  404956  20.16
+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  3926257486  57722   141.40
-rawstudio-mint14.tar    gzkp    1   8558382592  3970463184  41749   195.49
+
+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  641339945   8925    355.35
-nyc-taxi-data-10M.csv   zskp    2   3325605752  591748091   11268   281.44
-nyc-taxi-data-10M.csv   zskp    3   3325605752  530289687   25239   125.66
-nyc-taxi-data-10M.csv   zskp    4   3325605752  490907191   65939   48.10
+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  928656485   23876   132.83
-nyc-taxi-data-10M.csv   gzkp    1   3325605752  924718719   16388   193.53
+
+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
 
-Staus: STABLE - there may still be subtle bugs, but a wide variety of content has been tested.
+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/). 
@@ -274,8 +289,13 @@
 }
 ```
 
-It is important to use the "Close" function when you no longer need the Reader to stop running goroutines. 
-See "Allocation-less operation" below.
+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:
 
@@ -284,7 +304,7 @@
 
 // Create a reader that caches decompressors.
 // For this operation type we supply a nil Reader.
-var decoder, _ = zstd.NewReader(nil)
+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.
@@ -294,9 +314,12 @@
 ```
 
 Both of these cases should provide the functionality needed. 
-The decoder can be used for *concurrent* decompression of multiple buffers. 
+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.   
+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
 
@@ -348,70 +371,71 @@
 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 operates on
+The stream decoder will create goroutines that:
 
-* One goroutine reads input and splits the input to several block decoders.
-* A number of decoders will decode blocks.
-* A goroutine coordinates these blocks and sends history from one to the next.
+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 2 cores effectively.
- 
- 
+In practice this means that concurrency is often limited to utilizing about 3 cores effectively.
+  
 ### Benchmarks
 
-These are some examples of performance compared to [datadog cgo library](https://github.com/DataDog/zstd).
-
 The first two are streaming decodes and the last are smaller inputs. 
- 
+
+Running on AMD Ryzen 9 3950X 16-Core Processor. AMD64 assembly used.
+
 ```
-BenchmarkDecoderSilesia-8                          3     385000067 ns/op     550.51 MB/s        5498 B/op          8 allocs/op
-BenchmarkDecoderSilesiaCgo-8                       6     197666567 ns/op    1072.25 MB/s      270672 B/op          8 allocs/op
+BenchmarkDecoderSilesia-32    	                   5	 206878840 ns/op	1024.50 MB/s	   49808 B/op	      43 allocs/op
+BenchmarkDecoderEnwik9-32                          1	1271809000 ns/op	 786.28 MB/s	   72048 B/op	      52 allocs/op
 
-BenchmarkDecoderEnwik9-8                           1    2027001600 ns/op     493.34 MB/s       10496 B/op         18 allocs/op
-BenchmarkDecoderEnwik9Cgo-8                        2     979499200 ns/op    1020.93 MB/s      270672 B/op          8 allocs/op
+Concurrent blocks, performance:
 
-Concurrent performance:
-
-BenchmarkDecoder_DecodeAllParallel/kppkn.gtb.zst-16                28915         42469 ns/op    4340.07 MB/s         114 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallel/geo.protodata.zst-16           116505          9965 ns/op    11900.16 MB/s         16 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallel/plrabn12.txt.zst-16              8952        134272 ns/op    3588.70 MB/s         915 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallel/lcet10.txt.zst-16               11820        102538 ns/op    4161.90 MB/s         594 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallel/asyoulik.txt.zst-16             34782         34184 ns/op    3661.88 MB/s          60 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallel/alice29.txt.zst-16              27712         43447 ns/op    3500.58 MB/s          99 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallel/html_x_4.zst-16                 62826         18750 ns/op    21845.10 MB/s        104 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallel/paper-100k.pdf.zst-16          631545          1794 ns/op    57078.74 MB/s          2 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallel/fireworks.jpeg.zst-16         1690140           712 ns/op    172938.13 MB/s         1 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallel/urls.10K.zst-16                 10432        113593 ns/op    6180.73 MB/s        1143 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallel/html.zst-16                    113206         10671 ns/op    9596.27 MB/s          15 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallel/comp-data.bin.zst-16          1530615           779 ns/op    5229.49 MB/s           0 B/op          0 allocs/op
-
-BenchmarkDecoder_DecodeAllParallelCgo/kppkn.gtb.zst-16             65217         16192 ns/op    11383.34 MB/s         46 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallelCgo/geo.protodata.zst-16        292671          4039 ns/op    29363.19 MB/s          6 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallelCgo/plrabn12.txt.zst-16          26314         46021 ns/op    10470.43 MB/s        293 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallelCgo/lcet10.txt.zst-16            33897         34900 ns/op    12227.96 MB/s        205 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallelCgo/asyoulik.txt.zst-16         104348         11433 ns/op    10949.01 MB/s         20 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallelCgo/alice29.txt.zst-16           75949         15510 ns/op    9805.60 MB/s          32 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallelCgo/html_x_4.zst-16             173910          6756 ns/op    60624.29 MB/s         37 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallelCgo/paper-100k.pdf.zst-16       923076          1339 ns/op    76474.87 MB/s          1 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallelCgo/fireworks.jpeg.zst-16       922920          1351 ns/op    91102.57 MB/s          2 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallelCgo/urls.10K.zst-16              27649         43618 ns/op    16096.19 MB/s        407 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallelCgo/html.zst-16                 279073          4160 ns/op    24614.18 MB/s          6 B/op          0 allocs/op
-BenchmarkDecoder_DecodeAllParallelCgo/comp-data.bin.zst-16        749938          1579 ns/op    2581.71 MB/s           0 B/op          0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/kppkn.gtb.zst-32         	   67356	     17857 ns/op	10321.96 MB/s	        22.48 pct	     102 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/geo.protodata.zst-32     	  266656	      4421 ns/op	26823.21 MB/s	        11.89 pct	      19 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/plrabn12.txt.zst-32      	   20992	     56842 ns/op	8477.17 MB/s	        39.90 pct	     754 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/lcet10.txt.zst-32        	   27456	     43932 ns/op	9714.01 MB/s	        33.27 pct	     524 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/asyoulik.txt.zst-32      	   78432	     15047 ns/op	8319.15 MB/s	        40.34 pct	      66 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/alice29.txt.zst-32       	   65800	     18436 ns/op	8249.63 MB/s	        37.75 pct	      88 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/html_x_4.zst-32          	  102993	     11523 ns/op	35546.09 MB/s	         3.637 pct	     143 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/paper-100k.pdf.zst-32    	 1000000	      1070 ns/op	95720.98 MB/s	        80.53 pct	       3 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/fireworks.jpeg.zst-32    	  749802	      1752 ns/op	70272.35 MB/s	       100.0 pct	       5 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/urls.10K.zst-32          	   22640	     52934 ns/op	13263.37 MB/s	        26.25 pct	    1014 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/html.zst-32              	  226412	      5232 ns/op	19572.27 MB/s	        14.49 pct	      20 B/op	       0 allocs/op
+BenchmarkDecoder_DecodeAllParallel/comp-data.bin.zst-32     	  923041	      1276 ns/op	3194.71 MB/s	        31.26 pct	       0 B/op	       0 allocs/op
 ```
 
-This reflects the performance around May 2020, but this may be out of date.
+This reflects the performance around May 2022, but this may be out of date.
+
+## Zstd inside ZIP files
+
+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 sending files for reproducing errors use a service like [goobox](https://goobox.io/#/upload) or similar to share your files.
-
 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
index 8544585..d41e3e1 100644
--- a/vendor/github.com/klauspost/compress/zstd/bitreader.go
+++ b/vendor/github.com/klauspost/compress/zstd/bitreader.go
@@ -5,10 +5,12 @@
 package zstd
 
 import (
-	"encoding/binary"
 	"errors"
+	"fmt"
 	"io"
 	"math/bits"
+
+	"github.com/klauspost/compress/internal/le"
 )
 
 // bitReader reads a bitstream in reverse.
@@ -16,8 +18,8 @@
 // for aligning the input.
 type bitReader struct {
 	in       []byte
-	off      uint   // next byte to read is at in[off - 1]
 	value    uint64 // Maybe use [16]byte, but shifting is awkward.
+	cursor   int    // offset where next read should end
 	bitsRead uint8
 }
 
@@ -27,12 +29,12 @@
 		return errors.New("corrupt stream: too short")
 	}
 	b.in = in
-	b.off = uint(len(in))
 	// The highest bit of the last byte indicates where to start
 	v := in[len(in)-1]
 	if v == 0 {
 		return errors.New("corrupt stream, did not find end of stream")
 	}
+	b.cursor = len(in)
 	b.bitsRead = 64
 	b.value = 0
 	if len(in) >= 8 {
@@ -50,16 +52,16 @@
 	if n == 0 /*|| b.bitsRead >= 64 */ {
 		return 0
 	}
-	return b.getBitsFast(n)
+	return int(b.get32BitsFast(n))
 }
 
-// getBitsFast requires that at least one bit is requested every time.
+// get32BitsFast requires that at least one bit is requested every time.
 // There are no checks if the buffer is filled.
-func (b *bitReader) getBitsFast(n uint8) int {
+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 int(v)
+	return v
 }
 
 // fillFast() will make sure at least 32 bits are available.
@@ -68,21 +70,16 @@
 	if b.bitsRead < 32 {
 		return
 	}
-	// 2 bounds checks.
-	v := b.in[b.off-4:]
-	v = v[:4]
-	low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
-	b.value = (b.value << 32) | uint64(low)
+	b.cursor -= 4
+	b.value = (b.value << 32) | uint64(le.Load32(b.in, b.cursor))
 	b.bitsRead -= 32
-	b.off -= 4
 }
 
 // fillFastStart() assumes the bitreader is empty and there is at least 8 bytes to read.
 func (b *bitReader) fillFastStart() {
-	// Do single re-slice to avoid bounds checks.
-	b.value = binary.LittleEndian.Uint64(b.in[b.off-8:])
+	b.cursor -= 8
+	b.value = le.Load64(b.in, b.cursor)
 	b.bitsRead = 0
-	b.off -= 8
 }
 
 // fill() will make sure at least 32 bits are available.
@@ -90,25 +87,23 @@
 	if b.bitsRead < 32 {
 		return
 	}
-	if b.off >= 4 {
-		v := b.in[b.off-4:]
-		v = v[:4]
-		low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
-		b.value = (b.value << 32) | uint64(low)
+	if b.cursor >= 4 {
+		b.cursor -= 4
+		b.value = (b.value << 32) | uint64(le.Load32(b.in, b.cursor))
 		b.bitsRead -= 32
-		b.off -= 4
 		return
 	}
-	for b.off > 0 {
-		b.value = (b.value << 8) | uint64(b.in[b.off-1])
-		b.bitsRead -= 8
-		b.off--
+
+	b.bitsRead -= uint8(8 * b.cursor)
+	for b.cursor > 0 {
+		b.cursor -= 1
+		b.value = (b.value << 8) | uint64(b.in[b.cursor])
 	}
 }
 
 // finished returns true if all bits have been read from the bit stream.
 func (b *bitReader) finished() bool {
-	return b.off == 0 && b.bitsRead >= 64
+	return b.cursor == 0 && b.bitsRead >= 64
 }
 
 // overread returns true if more bits have been requested than is on the stream.
@@ -118,13 +113,17 @@
 
 // remain returns the number of bits remaining.
 func (b *bitReader) remain() uint {
-	return b.off*8 + 64 - uint(b.bitsRead)
+	return 8*uint(b.cursor) + 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
+	b.cursor = 0
+	if !b.finished() {
+		return fmt.Errorf("%d extra bits on block, should be 0", b.remain())
+	}
 	if b.bitsRead > 64 {
 		return io.ErrUnexpectedEOF
 	}
diff --git a/vendor/github.com/klauspost/compress/zstd/bitwriter.go b/vendor/github.com/klauspost/compress/zstd/bitwriter.go
index 303ae90..1952f17 100644
--- a/vendor/github.com/klauspost/compress/zstd/bitwriter.go
+++ b/vendor/github.com/klauspost/compress/zstd/bitwriter.go
@@ -5,8 +5,6 @@
 
 package zstd
 
-import "fmt"
-
 // bitWriter will write bits.
 // First bit will be LSB of the first byte of output.
 type bitWriter struct {
@@ -38,7 +36,7 @@
 	b.nBits += bits
 }
 
-// addBits32NC will add up to 32 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) {
@@ -46,6 +44,26 @@
 	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) {
@@ -53,80 +71,6 @@
 	b.nBits += bits
 }
 
-// flush will flush all pending full bytes.
-// There will be at least 56 bits available for writing when this has been called.
-// Using flush32 is faster, but leaves less space for writing.
-func (b *bitWriter) flush() {
-	v := b.nBits >> 3
-	switch v {
-	case 0:
-	case 1:
-		b.out = append(b.out,
-			byte(b.bitContainer),
-		)
-	case 2:
-		b.out = append(b.out,
-			byte(b.bitContainer),
-			byte(b.bitContainer>>8),
-		)
-	case 3:
-		b.out = append(b.out,
-			byte(b.bitContainer),
-			byte(b.bitContainer>>8),
-			byte(b.bitContainer>>16),
-		)
-	case 4:
-		b.out = append(b.out,
-			byte(b.bitContainer),
-			byte(b.bitContainer>>8),
-			byte(b.bitContainer>>16),
-			byte(b.bitContainer>>24),
-		)
-	case 5:
-		b.out = append(b.out,
-			byte(b.bitContainer),
-			byte(b.bitContainer>>8),
-			byte(b.bitContainer>>16),
-			byte(b.bitContainer>>24),
-			byte(b.bitContainer>>32),
-		)
-	case 6:
-		b.out = append(b.out,
-			byte(b.bitContainer),
-			byte(b.bitContainer>>8),
-			byte(b.bitContainer>>16),
-			byte(b.bitContainer>>24),
-			byte(b.bitContainer>>32),
-			byte(b.bitContainer>>40),
-		)
-	case 7:
-		b.out = append(b.out,
-			byte(b.bitContainer),
-			byte(b.bitContainer>>8),
-			byte(b.bitContainer>>16),
-			byte(b.bitContainer>>24),
-			byte(b.bitContainer>>32),
-			byte(b.bitContainer>>40),
-			byte(b.bitContainer>>48),
-		)
-	case 8:
-		b.out = append(b.out,
-			byte(b.bitContainer),
-			byte(b.bitContainer>>8),
-			byte(b.bitContainer>>16),
-			byte(b.bitContainer>>24),
-			byte(b.bitContainer>>32),
-			byte(b.bitContainer>>40),
-			byte(b.bitContainer>>48),
-			byte(b.bitContainer>>56),
-		)
-	default:
-		panic(fmt.Errorf("bits (%d) > 64", b.nBits))
-	}
-	b.bitContainer >>= v << 3
-	b.nBits &= 7
-}
-
 // flush32 will flush out, so there are at least 32 bits available for writing.
 func (b *bitWriter) flush32() {
 	if b.nBits < 32 {
@@ -153,12 +97,11 @@
 
 // close will write the alignment bit and write the final byte(s)
 // to the output.
-func (b *bitWriter) close() error {
+func (b *bitWriter) close() {
 	// End mark
 	b.addBits16Clean(1, 1)
 	// flush until next byte.
 	b.flushAlign()
-	return nil
 }
 
 // reset and continue writing by appending to out.
diff --git a/vendor/github.com/klauspost/compress/zstd/blockdec.go b/vendor/github.com/klauspost/compress/zstd/blockdec.go
index b51d922..0dd742f 100644
--- a/vendor/github.com/klauspost/compress/zstd/blockdec.go
+++ b/vendor/github.com/klauspost/compress/zstd/blockdec.go
@@ -7,6 +7,7 @@
 import (
 	"errors"
 	"fmt"
+	"hash/crc32"
 	"io"
 	"sync"
 
@@ -38,14 +39,14 @@
 	// 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
 
-	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#literals_section_header
-	maxCompressedLiteralSize = 1 << 18
-	maxRLELiteralSize        = 1 << 20
-	maxMatchLen              = 131074
-	maxSequences             = 0x7f00 + 0xffff
+	maxMatchLen  = 131074
+	maxSequences = 0x7f00 + 0xffff
 
 	// We support slightly less than the reference decoder to be able to
 	// use ints on 32 bit archs.
@@ -76,20 +77,28 @@
 	// Window size of the block.
 	WindowSize uint64
 
-	history     chan *history
-	input       chan struct{}
-	result      chan decodeOutput
-	sequenceBuf []seq
-	err         error
-	decWG       sync.WaitGroup
+	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
-	tmp     [4]byte
 
 	Type blockType
 
@@ -109,13 +118,8 @@
 
 func newBlockDec(lowMem bool) *blockDec {
 	b := blockDec{
-		lowMem:  lowMem,
-		result:  make(chan decodeOutput, 1),
-		input:   make(chan struct{}, 1),
-		history: make(chan *history, 1),
+		lowMem: lowMem,
 	}
-	b.decWG.Add(1)
-	go b.startDecoder()
 	return &b
 }
 
@@ -123,44 +127,60 @@
 // 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 := br.readSmall(3)
-	if tmp == nil {
-		if debug {
-			println("Reading block header:", io.ErrUnexpectedEOF)
-		}
-		return io.ErrUnexpectedEOF
+	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 := maxBlockSize
+	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 debug {
+		if debugDecoder {
 			println("Data size on stream:", cSize)
 		}
 		b.RLESize = 0
-		maxSize = maxCompressedBlockSize
+		maxSize = maxCompressedBlockSizeAlloc
 		if windowSize < maxCompressedBlockSize && b.lowMem {
-			maxSize = int(windowSize)
+			maxSize = int(windowSize) + compressedBlockOverAlloc
 		}
 		if cSize > maxCompressedBlockSize || uint64(cSize) > b.WindowSize {
-			if debug {
+			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
@@ -169,25 +189,25 @@
 	}
 
 	// Read block data.
-	if cap(b.dataStorage) < cSize {
-		if b.lowMem {
-			b.dataStorage = make([]byte, 0, cSize)
+	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, maxBlockSize)
+			b.dataStorage = make([]byte, 0, maxCompressedBlockSizeAlloc)
 		}
 	}
-	if cap(b.dst) <= maxSize {
-		b.dst = make([]byte, 0, maxSize+1)
-	}
-	var err error
 	b.data, err = br.readBig(cSize, b.dataStorage)
 	if err != nil {
-		if debug {
+		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
 }
 
@@ -196,85 +216,14 @@
 	b.Last = true
 	b.Type = blockTypeReserved
 	b.err = err
-	b.input <- struct{}{}
 }
 
 // Close will release resources.
 // Closed blockDec cannot be reset.
 func (b *blockDec) Close() {
-	close(b.input)
-	close(b.history)
-	close(b.result)
-	b.decWG.Wait()
 }
 
-// decodeAsync will prepare decoding the block when it receives input.
-// This will separate output and history.
-func (b *blockDec) startDecoder() {
-	defer b.decWG.Done()
-	for range b.input {
-		//println("blockDec: Got block input")
-		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, maxBlockSize)
-				}
-			}
-			o := decodeOutput{
-				d:   b,
-				b:   b.dst[:b.RLESize],
-				err: nil,
-			}
-			v := b.data[0]
-			for i := range o.b {
-				o.b[i] = v
-			}
-			hist := <-b.history
-			hist.append(o.b)
-			b.result <- o
-		case blockTypeRaw:
-			o := decodeOutput{
-				d:   b,
-				b:   b.data,
-				err: nil,
-			}
-			hist := <-b.history
-			hist.append(o.b)
-			b.result <- o
-		case blockTypeCompressed:
-			b.dst = b.dst[:0]
-			err := b.decodeCompressed(nil)
-			o := decodeOutput{
-				d:   b,
-				b:   b.dst,
-				err: err,
-			}
-			if debug {
-				println("Decompressed to", len(b.dst), "bytes, error:", err)
-			}
-			b.result <- o
-		case blockTypeReserved:
-			// Used for returning errors.
-			<-b.history
-			b.result <- decodeOutput{
-				d:   b,
-				b:   nil,
-				err: b.err,
-			}
-		default:
-			panic("Invalid block type")
-		}
-		if debug {
-			println("blockDec: Finished block")
-		}
-	}
-}
-
-// decodeAsync will prepare decoding the block when it receives the history.
-// If history is provided, it will not fetch it from the channel.
+// decodeBuf
 func (b *blockDec) decodeBuf(hist *history) error {
 	switch b.Type {
 	case blockTypeRLE:
@@ -282,7 +231,7 @@
 			if b.lowMem {
 				b.dst = make([]byte, b.RLESize)
 			} else {
-				b.dst = make([]byte, maxBlockSize)
+				b.dst = make([]byte, maxCompressedBlockSize)
 			}
 		}
 		b.dst = b.dst[:b.RLESize]
@@ -297,14 +246,23 @@
 		return nil
 	case blockTypeCompressed:
 		saved := b.dst
-		b.dst = hist.b
-		hist.b = nil
+		// 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 debug {
+		if debugDecoder {
 			println("Decompressed to total", len(b.dst), "bytes, hash:", xxhash.Sum64(b.dst), "error:", err)
 		}
-		hist.b = b.dst
-		b.dst = saved
+		if hist.ignoreBuffer == 0 {
+			hist.b = b.dst
+			b.dst = saved
+		} else {
+			hist.appendKeep(b.dst)
+		}
 		return err
 	case blockTypeReserved:
 		// Used for returning errors.
@@ -314,30 +272,18 @@
 	}
 }
 
-// decodeCompressed will start decompressing a block.
-// If no history is supplied the decoder will decodeAsync as much as possible
-// before fetching from blockDec.history
-func (b *blockDec) decodeCompressed(hist *history) error {
-	in := b.data
-	delayedHistory := hist == nil
-
-	if delayedHistory {
-		// We must always grab history.
-		defer func() {
-			if hist == nil {
-				<-b.history
-			}
-		}()
-	}
+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 ErrBlockTooSmall
+		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 {
@@ -353,7 +299,7 @@
 			//  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 ErrBlockTooSmall
+				return in, ErrBlockTooSmall
 			}
 			litRegenSize = int(in[0]>>4) + (int(in[1]) << 4) + (int(in[2]) << 12)
 			in = in[3:]
@@ -364,7 +310,7 @@
 			// 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 ErrBlockTooSmall
+				return in, ErrBlockTooSmall
 			}
 			n := uint64(in[0]>>4) + (uint64(in[1]) << 4) + (uint64(in[2]) << 12)
 			litRegenSize = int(n & 1023)
@@ -375,7 +321,7 @@
 			fourStreams = true
 			if len(in) < 4 {
 				println("too small: litType:", litType, " sizeFormat", sizeFormat, len(in))
-				return ErrBlockTooSmall
+				return in, ErrBlockTooSmall
 			}
 			n := uint64(in[0]>>4) + (uint64(in[1]) << 4) + (uint64(in[2]) << 12) + (uint64(in[3]) << 20)
 			litRegenSize = int(n & 16383)
@@ -385,7 +331,7 @@
 			fourStreams = true
 			if len(in) < 5 {
 				println("too small: litType:", litType, " sizeFormat", sizeFormat, len(in))
-				return ErrBlockTooSmall
+				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)
@@ -393,16 +339,18 @@
 			in = in[5:]
 		}
 	}
-	if debug {
+	if debugDecoder {
 		println("literals type:", litType, "litRegenSize:", litRegenSize, "litCompSize:", litCompSize, "sizeFormat:", sizeFormat, "4X:", fourStreams)
 	}
-	var literals []byte
-	var huff *huff0.Scratch
+	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 ErrBlockTooSmall
+			return in, ErrBlockTooSmall
 		}
 		literals = in[:litRegenSize]
 		in = in[litRegenSize:]
@@ -410,19 +358,13 @@
 	case literalsBlockRLE:
 		if len(in) < 1 {
 			println("too small: litType:", litType, " sizeFormat", sizeFormat, "remain:", len(in), "want:", 1)
-			return ErrBlockTooSmall
+			return in, ErrBlockTooSmall
 		}
 		if cap(b.literalBuf) < litRegenSize {
 			if b.lowMem {
-				b.literalBuf = make([]byte, litRegenSize)
+				b.literalBuf = make([]byte, litRegenSize, litRegenSize+compressedBlockOverAlloc)
 			} else {
-				if litRegenSize > maxCompressedLiteralSize {
-					// Exceptional
-					b.literalBuf = make([]byte, litRegenSize)
-				} else {
-					b.literalBuf = make([]byte, litRegenSize, maxCompressedLiteralSize)
-
-				}
+				b.literalBuf = make([]byte, litRegenSize, maxCompressedBlockSize+compressedBlockOverAlloc)
 			}
 		}
 		literals = b.literalBuf[:litRegenSize]
@@ -431,45 +373,82 @@
 			literals[i] = v
 		}
 		in = in[1:]
-		if debug {
+		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 ErrBlockTooSmall
+			return in, ErrBlockTooSmall
 		}
 		// Store compressed literals, so we defer decoding until we get history.
 		literals = in[:litCompSize]
 		in = in[litCompSize:]
-		if debug {
+		if debugDecoder {
 			printf("Found %d compressed literals\n", litCompSize)
 		}
-	case literalsBlockCompressed:
-		if len(in) < litCompSize {
-			println("too small: litType:", litType, " sizeFormat", sizeFormat, "remain:", len(in), "want:", litCompSize)
-			return ErrBlockTooSmall
+		huff := hist.huffTree
+		if huff == nil {
+			return in, errors.New("literal block was treeless, but no history was defined")
 		}
-		literals = in[:litCompSize]
-		in = in[litCompSize:]
-		huff = huffDecoderPool.Get().(*huff0.Scratch)
-		var err error
 		// Ensure we have space to store it.
 		if cap(b.literalBuf) < litRegenSize {
 			if b.lowMem {
-				b.literalBuf = make([]byte, 0, litRegenSize)
+				b.literalBuf = make([]byte, 0, litRegenSize+compressedBlockOverAlloc)
 			} else {
-				b.literalBuf = make([]byte, 0, maxCompressedLiteralSize)
+				b.literalBuf = make([]byte, 0, maxCompressedBlockSize+compressedBlockOverAlloc)
 			}
 		}
-		if huff == nil {
-			huff = &huff0.Scratch{}
+		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 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)
@@ -478,27 +457,63 @@
 		}
 		if err != nil {
 			println("decoding compressed literals:", err)
-			return err
+			return in, err
 		}
 		// Make sure we don't leak our literals buffer
 		if len(literals) != litRegenSize {
-			return fmt.Errorf("literal output size mismatch want %d, got %d", litRegenSize, len(literals))
+			return in, fmt.Errorf("literal output size mismatch want %d, got %d", litRegenSize, len(literals))
 		}
-		if debug {
+		// 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]
-	nSeqs := 0
 	switch {
-	case seqHeader == 0:
-		in = in[1:]
 	case seqHeader < 128:
 		nSeqs = int(seqHeader)
 		in = in[1:]
@@ -515,19 +530,16 @@
 		nSeqs = 0x7f00 + int(in[1]) + (int(in[2]) << 8)
 		in = in[3:]
 	}
-	// Allocate sequences
-	if cap(b.sequenceBuf) < nSeqs {
-		if b.lowMem {
-			b.sequenceBuf = make([]seq, nSeqs)
-		} else {
-			// Allocate max
-			b.sequenceBuf = make([]seq, nSeqs, maxSequences)
+	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))
 		}
-	} else {
-		// Reuse buffer
-		b.sequenceBuf = b.sequenceBuf[:nSeqs]
+		return ErrUnexpectedBlockSize
 	}
-	var seqs = &sequenceDecs{}
+
+	var seqs = &hist.decoders
+	seqs.nSeqs = nSeqs
 	if nSeqs > 0 {
 		if len(in) < 1 {
 			return ErrBlockTooSmall
@@ -535,12 +547,15 @@
 		br := byteReader{b: in, off: 0}
 		compMode := br.Uint8()
 		br.advance(1)
-		if debug {
+		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 debug {
+			if debugDecoder {
 				println("Table", tableIndex(i), "is", mode)
 			}
 			var seq *sequenceDec
@@ -556,6 +571,9 @@
 			}
 			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 {
@@ -563,34 +581,38 @@
 				}
 				v := br.Uint8()
 				br.advance(1)
-				dec := fseDecoderPool.Get().(*fseDecoder)
+				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
 				}
-				dec.setRLE(symb)
-				seq.fse = dec
-				if debug {
-					printf("RLE set to %+v, code: %v", symb, v)
+				seq.fse.setRLE(symb)
+				if debugDecoder {
+					printf("RLE set to 0x%x, code: %v", symb, v)
 				}
 			case compModeFSE:
-				println("Reading table for", tableIndex(i))
-				dec := fseDecoderPool.Get().(*fseDecoder)
-				err := dec.readNCount(&br, uint16(maxTableSymbol[i]))
+				if debugDecoder {
+					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 = dec.transform(symbolTableX[i])
+				err = seq.fse.transform(symbolTableX[i])
 				if err != nil {
 					println("Transform table error:", err)
 					return err
 				}
-				if debug {
-					println("Read table ok", "symbolLen:", dec.symbolLen)
+				if debugDecoder {
+					println("Read table ok", "symbolLen:", seq.fse.symbolLen)
 				}
-				seq.fse = dec
 			case compModeRepeat:
 				seq.repeat = true
 			}
@@ -600,140 +622,91 @@
 		}
 		in = br.unread()
 	}
-
-	// Wait for history.
-	// All time spent after this is critical since it is strictly sequential.
-	if hist == nil {
-		hist = <-b.history
-		if hist.error {
-			return ErrDecoderClosed
-		}
-	}
-
-	// Decode treeless literal block.
-	if litType == literalsBlockTreeless {
-		// TODO: We could send the history early WITHOUT the stream history.
-		//   This would allow decoding treeless literals before the byte history is available.
-		//   Silencia stats: Treeless 4393, with: 32775, total: 37168, 11% treeless.
-		//   So not much obvious gain here.
-
-		if hist.huffTree == nil {
-			return 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)
-			} else {
-				b.literalBuf = make([]byte, 0, maxCompressedLiteralSize)
-			}
-		}
-		var err error
-		// Use our out buffer.
-		huff = hist.huffTree
-		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 err
-		}
-		if len(literals) != litRegenSize {
-			return fmt.Errorf("literal output size mismatch want %d, got %d", litRegenSize, len(literals))
-		}
-	} else {
-		if hist.huffTree != nil && huff != nil {
-			if hist.dict == nil || hist.dict.litEnc != hist.huffTree {
-				huffDecoderPool.Put(hist.huffTree)
-			}
-			hist.huffTree = nil
-		}
-	}
-	if huff != nil {
-		hist.huffTree = huff
-	}
-	if debug {
-		println("Final literals:", len(literals), "hash:", xxhash.Sum64(literals), "and", nSeqs, "sequences.")
+	if debugDecoder {
+		println("Literals:", len(seqs.literals), "hash:", xxhash.Sum64(seqs.literals), "and", seqs.nSeqs, "sequences.")
 	}
 
 	if nSeqs == 0 {
-		// Decompressed content is defined entirely as Literals Section content.
-		b.dst = append(b.dst, literals...)
-		if delayedHistory {
-			hist.append(literals)
+		if len(b.sequence) > 0 {
+			b.sequence = b.sequence[:0]
 		}
 		return nil
 	}
-
-	seqs, err := seqs.mergeHistory(&hist.decoders)
-	if err != nil {
-		return err
+	br := seqs.br
+	if br == nil {
+		br = &bitReader{}
 	}
-	if debug {
-		println("History merged ok")
-	}
-	br := &bitReader{}
 	if err := br.init(in); err != nil {
 		return err
 	}
 
-	// TODO: Investigate if sending history without decoders are faster.
-	//   This would allow the sequences to be decoded async and only have to construct stream history.
-	//   If only recent offsets were not transferred, this would be an obvious win.
-	// 	 Also, if first 3 sequences don't reference recent offsets, all sequences can be decoded.
-
-	hbytes := hist.b
-	if len(hbytes) > hist.windowSize {
-		hbytes = hbytes[len(hbytes)-hist.windowSize:]
-		// We do not need history any more.
-		if hist.dict != nil {
-			hist.dict.content = nil
-		}
-	}
-
-	if err := seqs.initialize(br, hist, literals, b.dst); err != nil {
+	if err := seqs.initialize(br, hist, b.dst); err != nil {
 		println("initializing sequences:", err)
 		return err
 	}
 
-	err = seqs.decode(nSeqs, br, hbytes)
+	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
 	}
-	if !br.finished() {
-		return fmt.Errorf("%d extra bits on block, should be 0", br.remain())
-	}
+	return b.updateHistory(hist)
+}
 
-	err = br.close()
-	if err != nil {
-		printf("Closing sequences: %v, %+v\n", err, *br)
-	}
+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 = seqs.out
-	seqs.out, seqs.literals, seqs.hist = nil, nil, nil
+	b.dst = hist.decoders.out
+	hist.recentOffsets = hist.decoders.prevOffset
 
-	if !delayedHistory {
-		// If we don't have delayed history, no need to update.
-		hist.recentOffsets = seqs.prevOffset
-		return nil
-	}
 	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.append(b.dst)
-	hist.recentOffsets = seqs.prevOffset
-	if debug {
-		println("Finished block with literals:", len(literals), "and", nSeqs, "sequences.")
-	}
+	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
index e1be092..fd35ea1 100644
--- a/vendor/github.com/klauspost/compress/zstd/blockenc.go
+++ b/vendor/github.com/klauspost/compress/zstd/blockenc.go
@@ -9,6 +9,7 @@
 	"fmt"
 	"math"
 	"math/bits"
+	"slices"
 
 	"github.com/klauspost/compress/huff0"
 )
@@ -51,7 +52,7 @@
 		if cap(b.literals) < maxCompressedBlockSize {
 			b.literals = make([]byte, 0, maxCompressedBlockSize)
 		}
-		const defSeqs = 200
+		const defSeqs = 2000
 		if cap(b.sequences) < defSeqs {
 			b.sequences = make([]seq, 0, defSeqs)
 		}
@@ -156,7 +157,7 @@
 	switch {
 	case inBits < 5:
 		lh |= (uint64(regenLen) << 3) | (1 << 60)
-		if debug {
+		if debugEncoder {
 			got := int(lh>>3) & 0xff
 			if got != regenLen {
 				panic(fmt.Sprint("litRegenSize = ", regenLen, "(want) != ", got, "(got)"))
@@ -184,7 +185,7 @@
 			lh |= 1 << 2
 		}
 		lh |= (uint64(inLen) << 4) | (uint64(compLen) << (10 + 4)) | (3 << 60)
-		if debug {
+		if debugEncoder {
 			const mmask = (1 << 24) - 1
 			n := (lh >> 4) & mmask
 			if int(n&1023) != inLen {
@@ -312,7 +313,7 @@
 	bh.setType(blockTypeRaw)
 	b.output = bh.appendTo(b.output[:0])
 	b.output = append(b.output, a...)
-	if debug {
+	if debugEncoder {
 		println("Adding RAW block, length", len(a), "last:", b.last)
 	}
 }
@@ -325,7 +326,7 @@
 	bh.setType(blockTypeRaw)
 	dst = bh.appendTo(dst)
 	dst = append(dst, src...)
-	if debug {
+	if debugEncoder {
 		println("Adding RAW block, length", len(src), "last:", b.last)
 	}
 	return dst
@@ -339,7 +340,7 @@
 
 	// Don't compress extremely small blocks
 	if len(lits) < 8 || (len(lits) < 32 && b.dictLitEnc == nil) || raw {
-		if debug {
+		if debugEncoder {
 			println("Adding RAW block, length", len(lits), "last:", b.last)
 		}
 		bh.setType(blockTypeRaw)
@@ -361,17 +362,24 @@
 	if len(lits) >= 1024 {
 		// Use 4 Streams.
 		out, reUsed, err = huff0.Compress4X(lits, b.litEnc)
-	} else if len(lits) > 32 {
+	} 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 debug {
+		if debugEncoder {
 			println("Adding RAW block, length", len(lits), "last:", b.last)
 		}
 		bh.setType(blockTypeRaw)
@@ -379,7 +387,7 @@
 		b.output = append(b.output, lits...)
 		return nil
 	case huff0.ErrUseRLE:
-		if debug {
+		if debugEncoder {
 			println("Adding RLE block, length", len(lits))
 		}
 		bh.setType(blockTypeRLE)
@@ -396,12 +404,12 @@
 	bh.setType(blockTypeCompressed)
 	var lh literalsHeader
 	if reUsed {
-		if debug {
+		if debugEncoder {
 			println("Reused tree, compressed to", len(out))
 		}
 		lh.setType(literalsBlockTreeless)
 	} else {
-		if debug {
+		if debugEncoder {
 			println("New tree, compressed to", len(out), "tree size:", len(b.litEnc.OutTable))
 		}
 		lh.setType(literalsBlockCompressed)
@@ -420,13 +428,23 @@
 	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()[:256]
+	hist := enc.Histogram()
 	maxSym := uint8(0)
 	for i, v := range data {
 		v = v & 63
@@ -440,16 +458,7 @@
 		// 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])
+	cnt := int(slices.Max(hist[:maxSym]))
 	if cnt == len(data) {
 		// RLE
 		return 0
@@ -472,8 +481,18 @@
 	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 >> 5)
+	saved := b.size - len(b.literals) - (b.size >> 6)
 	if saved < 16 {
 		if org == nil {
 			return errIncompressible
@@ -503,7 +522,7 @@
 	if len(b.literals) >= 1024 && !raw {
 		// Use 4 Streams.
 		out, reUsed, err = huff0.Compress4X(b.literals, b.litEnc)
-	} else if len(b.literals) > 32 && !raw {
+	} else if len(b.literals) > 16 && !raw {
 		// Use 1 stream
 		single = true
 		out, reUsed, err = huff0.Compress1X(b.literals, b.litEnc)
@@ -511,13 +530,24 @@
 		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 debug {
+		if debugEncoder {
 			println("Adding literals RAW, length", len(b.literals))
 		}
 	case huff0.ErrUseRLE:
@@ -525,22 +555,22 @@
 		lh.setSize(len(b.literals))
 		b.output = lh.appendTo(b.output)
 		b.output = append(b.output, b.literals[0])
-		if debug {
+		if debugEncoder {
 			println("Adding literals RLE")
 		}
 	case nil:
 		// Compressed litLen...
 		if reUsed {
-			if debug {
+			if debugEncoder {
 				println("reused tree")
 			}
 			lh.setType(literalsBlockTreeless)
 		} else {
-			if debug {
+			if debugEncoder {
 				println("new tree, size:", len(b.litEnc.OutTable))
 			}
 			lh.setType(literalsBlockCompressed)
-			if debug {
+			if debugEncoder {
 				_, _, err := huff0.ReadTable(out, nil)
 				if err != nil {
 					panic(err)
@@ -548,18 +578,18 @@
 			}
 		}
 		lh.setSizes(len(out), len(b.literals), single)
-		if debug {
+		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 debug {
+		if debugEncoder {
 			println("Adding literals compressed")
 		}
 	default:
-		if debug {
+		if debugEncoder {
 			println("Adding literals ERROR:", err)
 		}
 		return err
@@ -577,7 +607,7 @@
 		n := len(b.sequences) - 0x7f00
 		b.output = append(b.output, 255, uint8(n), uint8(n>>8))
 	}
-	if debug {
+	if debugEncoder {
 		println("Encoding", len(b.sequences), "sequences")
 	}
 	b.genCodes()
@@ -611,17 +641,17 @@
 		nSize = nSize + (nSize+2*8*16)>>4
 		switch {
 		case predefSize <= prevSize && predefSize <= nSize || forcePreDef:
-			if debug {
+			if debugEncoder {
 				println("Using predefined", predefSize>>3, "<=", nSize>>3)
 			}
 			return preDef, compModePredefined
 		case prevSize <= nSize:
-			if debug {
+			if debugEncoder {
 				println("Using previous", prevSize>>3, "<=", nSize>>3)
 			}
 			return prev, compModeRepeat
 		default:
-			if debug {
+			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])
 			}
@@ -634,7 +664,7 @@
 	if llEnc.useRLE {
 		mode |= uint8(compModeRLE) << 6
 		llEnc.setRLE(b.sequences[0].llCode)
-		if debug {
+		if debugEncoder {
 			println("llEnc.useRLE")
 		}
 	} else {
@@ -645,7 +675,7 @@
 	if ofEnc.useRLE {
 		mode |= uint8(compModeRLE) << 4
 		ofEnc.setRLE(b.sequences[0].ofCode)
-		if debug {
+		if debugEncoder {
 			println("ofEnc.useRLE")
 		}
 	} else {
@@ -657,7 +687,7 @@
 	if mlEnc.useRLE {
 		mode |= uint8(compModeRLE) << 2
 		mlEnc.setRLE(b.sequences[0].mlCode)
-		if debug {
+		if debugEncoder {
 			println("mlEnc.useRLE, code: ", b.sequences[0].mlCode, "value", b.sequences[0].matchLen)
 		}
 	} else {
@@ -666,7 +696,7 @@
 		mode |= uint8(m) << 2
 	}
 	b.output = append(b.output, mode)
-	if debug {
+	if debugEncoder {
 		printf("Compression modes: 0b%b", mode)
 	}
 	b.output, err = llEnc.writeCount(b.output)
@@ -722,71 +752,72 @@
 		println("Encoded seq", seq, s, "codes:", s.llCode, s.mlCode, s.ofCode, "states:", ll.state, ml.state, of.state, "bits:", llB, mlB, ofB)
 	}
 	seq--
-	if llEnc.maxBits+mlEnc.maxBits+ofEnc.maxBits <= 32 {
-		// No need to flush (common)
-		for seq >= 0 {
-			s = b.sequences[seq]
-			wr.flush32()
-			llB, ofB, mlB := llTT[s.llCode], ofTT[s.ofCode], mlTT[s.mlCode]
-			// tabelog max is 8 for all.
-			of.encode(ofB)
-			ml.encode(mlB)
-			ll.encode(llB)
-			wr.flush32()
+	// Store sequences in reverse...
+	for seq >= 0 {
+		s = b.sequences[seq]
 
-			// We checked that all can stay within 32 bits
-			wr.addBits32NC(s.litLen, llB.outBits)
-			wr.addBits32NC(s.matchLen, mlB.outBits)
-			wr.addBits32NC(s.offset, ofB.outBits)
+		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]
 
-			if debugSequences {
-				println("Encoded seq", seq, s)
-			}
+		// Accumulate extra bits.
+		outBits := ofB.outBits & 31
+		extraBits := uint64(s.offset & bitMask32[outBits])
+		extraBitsN := outBits
 
-			seq--
+		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)
 		}
-	} else {
-		for seq >= 0 {
-			s = b.sequences[seq]
-			wr.flush32()
-			llB, ofB, mlB := llTT[s.llCode], ofTT[s.ofCode], mlTT[s.mlCode]
-			// tabelog max is below 8 for each.
-			of.encode(ofB)
-			ml.encode(mlB)
-			ll.encode(llB)
-			wr.flush32()
 
-			// ml+ll = max 32 bits total
-			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)
-			}
-
-			seq--
-		}
+		seq--
 	}
 	ml.flush(mlEnc.actualTableLog)
 	of.flush(ofEnc.actualTableLog)
 	ll.flush(llEnc.actualTableLog)
-	err = wr.close()
-	if err != nil {
-		return err
-	}
+	wr.close()
 	b.output = wr.out
 
+	// Maybe even add a bigger margin.
 	if len(b.output)-3-bhOffset >= b.size {
-		// Maybe even add a bigger margin.
+		// Discard and encode as raw block.
+		b.output = b.encodeRawTo(b.output[:bhOffset], org)
+		b.popOffsets()
 		b.litEnc.Reuse = huff0.ReusePolicyNone
-		return errIncompressible
+		return nil
 	}
 
 	// Size is output minus block header.
 	bh.setSize(uint32(len(b.output)-bhOffset) - 3)
-	if debug {
+	if debugEncoder {
 		println("Rewriting block header", bh)
 	}
 	_ = bh.appendTo(b.output[bhOffset:bhOffset])
@@ -801,14 +832,13 @@
 		// 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()[:256]
-	ofH := b.coders.ofEnc.Histogram()[:256]
-	mlH := b.coders.mlEnc.Histogram()[:256]
+	llH := b.coders.llEnc.Histogram()
+	ofH := b.coders.ofEnc.Histogram()
+	mlH := b.coders.mlEnc.Histogram()
 	for i := range llH {
 		llH[i] = 0
 	}
@@ -820,7 +850,8 @@
 	}
 
 	var llMax, ofMax, mlMax uint8
-	for i, seq := range b.sequences {
+	for i := range b.sequences {
+		seq := &b.sequences[i]
 		v := llCode(seq.litLen)
 		seq.llCode = v
 		llH[v]++
@@ -844,16 +875,6 @@
 				panic(fmt.Errorf("mlMax > maxMatchLengthSymbol (%d), matchlen: %d", mlMax, seq.matchLen))
 			}
 		}
-		b.sequences[i] = seq
-	}
-	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))
@@ -865,7 +886,7 @@
 		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]))
+	b.coders.mlEnc.HistogramFinished(mlMax, int(slices.Max(mlH[:mlMax+1])))
+	b.coders.ofEnc.HistogramFinished(ofMax, int(slices.Max(ofH[:ofMax+1])))
+	b.coders.llEnc.HistogramFinished(llMax, int(slices.Max(llH[:llMax+1])))
 }
diff --git a/vendor/github.com/klauspost/compress/zstd/bytebuf.go b/vendor/github.com/klauspost/compress/zstd/bytebuf.go
index 658ef78..55a3885 100644
--- a/vendor/github.com/klauspost/compress/zstd/bytebuf.go
+++ b/vendor/github.com/klauspost/compress/zstd/bytebuf.go
@@ -7,13 +7,12 @@
 import (
 	"fmt"
 	"io"
-	"io/ioutil"
 )
 
 type byteBuffer interface {
 	// Read up to 8 bytes.
-	// Returns nil if no more input is available.
-	readSmall(n int) []byte
+	// Returns io.ErrUnexpectedEOF if this cannot be satisfied.
+	readSmall(n int) ([]byte, error)
 
 	// Read >8 bytes.
 	// MAY use the destination slice.
@@ -23,23 +22,23 @@
 	readByte() (byte, error)
 
 	// Skip n bytes.
-	skipN(n int) error
+	skipN(n int64) error
 }
 
 // in-memory buffer
 type byteBuf []byte
 
-func (b *byteBuf) readSmall(n int) []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
+		return nil, io.ErrUnexpectedEOF
 	}
 	r := bb[:n]
 	*b = bb[n:]
-	return r
+	return r, nil
 }
 
 func (b *byteBuf) readBig(n int, dst []byte) ([]byte, error) {
@@ -52,23 +51,22 @@
 	return r, nil
 }
 
-func (b *byteBuf) remain() []byte {
-	return *b
-}
-
 func (b *byteBuf) readByte() (byte, error) {
 	bb := *b
 	if len(bb) < 1 {
-		return 0, nil
+		return 0, io.ErrUnexpectedEOF
 	}
 	r := bb[0]
 	*b = bb[1:]
 	return r, nil
 }
 
-func (b *byteBuf) skipN(n int) error {
+func (b *byteBuf) skipN(n int64) error {
 	bb := *b
-	if len(bb) < n {
+	if n < 0 {
+		return fmt.Errorf("negative skip (%d) requested", n)
+	}
+	if int64(len(bb)) < n {
 		return io.ErrUnexpectedEOF
 	}
 	*b = bb[n:]
@@ -81,19 +79,22 @@
 	tmp [8]byte
 }
 
-func (r *readerWrapper) readSmall(n int) []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 n2 != n {
-		if debug {
+	if err != nil {
+		if err == io.EOF {
+			return nil, io.ErrUnexpectedEOF
+		}
+		if debugDecoder {
 			println("readSmall: got", n2, "want", n, "err", err)
 		}
-		return nil
+		return nil, err
 	}
-	return r.tmp[:n]
+	return r.tmp[:n], nil
 }
 
 func (r *readerWrapper) readBig(n int, dst []byte) ([]byte, error) {
@@ -108,8 +109,11 @@
 }
 
 func (r *readerWrapper) readByte() (byte, error) {
-	n2, err := r.r.Read(r.tmp[:1])
+	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 {
@@ -118,9 +122,9 @@
 	return r.tmp[0], nil
 }
 
-func (r *readerWrapper) skipN(n int) error {
-	n2, err := io.CopyN(ioutil.Discard, r.r, int64(n))
-	if n2 != int64(n) {
+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
index 2c4fca1..0e59a24 100644
--- a/vendor/github.com/klauspost/compress/zstd/bytereader.go
+++ b/vendor/github.com/klauspost/compress/zstd/bytereader.go
@@ -13,12 +13,6 @@
 	off int
 }
 
-// init will initialize the reader and set the input.
-func (b *byteReader) init(in []byte) {
-	b.b = in
-	b.off = 0
-}
-
 // advance the stream b n bytes.
 func (b *byteReader) advance(n uint) {
 	b.off += int(n)
diff --git a/vendor/github.com/klauspost/compress/zstd/decodeheader.go b/vendor/github.com/klauspost/compress/zstd/decodeheader.go
index 69736e8..6a5a298 100644
--- a/vendor/github.com/klauspost/compress/zstd/decodeheader.go
+++ b/vendor/github.com/klauspost/compress/zstd/decodeheader.go
@@ -4,7 +4,7 @@
 package zstd
 
 import (
-	"bytes"
+	"encoding/binary"
 	"errors"
 	"io"
 )
@@ -15,18 +15,50 @@
 
 // Header contains information about the first frame and block within that.
 type Header struct {
-	// Window Size the window of data to keep while decoding.
-	// Will only be set if HasFCS is false.
-	WindowSize uint64
+	// 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
 
-	// Frame content size.
-	// Expected size of the entire frame.
-	FrameContentSize uint64
+	// 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.
@@ -51,17 +83,9 @@
 		CompressedSize int
 	}
 
-	// Skippable will be true if the frame is meant to be skipped.
-	// No other information will be populated.
-	Skippable bool
-
 	// If set there is a checksum present for the block content.
+	// The checksum field at the end is always 4 bytes long.
 	HasCheckSum bool
-
-	// If this is true FrameContentSize will have a valid value
-	HasFCS bool
-
-	SingleSegment bool
 }
 
 // Decode the header from the beginning of the stream.
@@ -71,39 +95,58 @@
 // 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 {
-	if len(in) < 4 {
-		return io.ErrUnexpectedEOF
-	}
-	b, in := in[:4], in[4:]
-	if !bytes.Equal(b, frameMagic) {
-		if !bytes.Equal(b[1:4], skippableFrameMagic) || b[0]&0xf0 != 0x50 {
-			return ErrMagicMismatch
-		}
-		*h = Header{Skippable: true}
-		return nil
-	}
-	if len(in) < 1 {
-		return io.ErrUnexpectedEOF
-	}
+	_, err := h.DecodeAndStrip(in)
+	return err
+}
 
-	// Clear output
+// 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{}
-	fhd, in := in[0], in[1:]
-	h.SingleSegment = fhd&(1<<5) != 0
-	h.HasCheckSum = fhd&(1<<2) != 0
-
-	if fhd&(1<<3) != 0 {
-		return errors.New("reserved bit set on frame 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 io.ErrUnexpectedEOF
+			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)
@@ -117,13 +160,11 @@
 			size = 4
 		}
 		if len(in) < int(size) {
-			return io.ErrUnexpectedEOF
+			return nil, io.ErrUnexpectedEOF
 		}
 		b, in = in[:size], in[size:]
-		if b == nil {
-			return io.ErrUnexpectedEOF
-		}
-		switch size {
+		h.HeaderSize += int(size)
+		switch len(b) {
 		case 1:
 			h.DictionaryID = uint32(b[0])
 		case 2:
@@ -149,13 +190,11 @@
 	if fcsSize > 0 {
 		h.HasFCS = true
 		if len(in) < fcsSize {
-			return io.ErrUnexpectedEOF
+			return nil, io.ErrUnexpectedEOF
 		}
 		b, in = in[:fcsSize], in[fcsSize:]
-		if b == nil {
-			return io.ErrUnexpectedEOF
-		}
-		switch fcsSize {
+		h.HeaderSize += int(fcsSize)
+		switch len(b) {
 		case 1:
 			h.FrameContentSize = uint64(b[0])
 		case 2:
@@ -172,7 +211,7 @@
 
 	// Frame Header done, we will not fail from now on.
 	if len(in) < 3 {
-		return nil
+		return in, nil
 	}
 	tmp := in[:3]
 	bh := uint32(tmp[0]) | (uint32(tmp[1]) << 8) | (uint32(tmp[2]) << 16)
@@ -182,7 +221,7 @@
 	cSize := int(bh >> 3)
 	switch blockType {
 	case blockTypeReserved:
-		return nil
+		return in, nil
 	case blockTypeRLE:
 		h.FirstBlock.Compressed = true
 		h.FirstBlock.DecompressedSize = cSize
@@ -198,5 +237,25 @@
 	}
 
 	h.FirstBlock.OK = true
-	return nil
+	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
index f593e46..ea2a193 100644
--- a/vendor/github.com/klauspost/compress/zstd/decoder.go
+++ b/vendor/github.com/klauspost/compress/zstd/decoder.go
@@ -5,9 +5,12 @@
 package zstd
 
 import (
-	"errors"
+	"context"
+	"encoding/binary"
 	"io"
 	"sync"
+
+	"github.com/klauspost/compress/zstd/internal/xxhash"
 )
 
 // Decoder provides decoding of zstandard streams.
@@ -22,15 +25,22 @@
 	// Unreferenced decoders, ready for use.
 	decoders chan *blockDec
 
-	// Streams ready to be decoded.
-	stream chan decodeStream
-
 	// 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.
-	// Always uses copies.
-	dicts map[uint32]dict
+	dicts map[uint32]*dict
 
 	// streamWg is the waitgroup for all streams
 	streamWg sync.WaitGroup
@@ -46,7 +56,10 @@
 	output chan decodeOutput
 
 	// cancel remaining output.
-	cancel chan struct{}
+	cancel context.CancelFunc
+
+	// crc of current frame
+	crc *xxhash.Digest
 
 	flushed bool
 }
@@ -69,7 +82,7 @@
 // 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 is will considerably
+// 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()
@@ -81,7 +94,7 @@
 			return nil, err
 		}
 	}
-	d.current.output = make(chan decodeOutput, d.o.concurrent)
+	d.current.crc = xxhash.New()
 	d.current.flushed = true
 
 	if r == nil {
@@ -89,7 +102,7 @@
 	}
 
 	// Transfer option dicts.
-	d.dicts = make(map[uint32]dict, len(d.o.dicts))
+	d.dicts = make(map[uint32]*dict, len(d.o.dicts))
 	for _, dc := range d.o.dicts {
 		d.dicts[dc.id] = dc
 	}
@@ -110,12 +123,9 @@
 }
 
 // Read bytes from the decompressed stream into p.
-// Returns the number of bytes written and any error that occurred.
+// Returns the number of bytes read and any error that occurred.
 // When the stream is done, io.EOF will be returned.
 func (d *Decoder) Read(p []byte) (int, error) {
-	if d.stream == nil {
-		return 0, ErrDecoderNilInput
-	}
 	var n int
 	for {
 		if len(d.current.b) > 0 {
@@ -133,12 +143,12 @@
 				break
 			}
 			if !d.nextBlock(n == 0) {
-				return n, nil
+				return n, d.current.err
 			}
 		}
 	}
 	if len(d.current.b) > 0 {
-		if debug {
+		if debugDecoder {
 			println("returning", n, "still bytes left:", len(d.current.b))
 		}
 		// Only return error at end of block
@@ -147,7 +157,7 @@
 	if d.current.err != nil {
 		d.drainOutput()
 	}
-	if debug {
+	if debugDecoder {
 		println("returning", n, d.current.err, len(d.decoders))
 	}
 	return n, d.current.err
@@ -165,67 +175,80 @@
 
 	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 d.stream == nil {
-		d.stream = make(chan decodeStream, 1)
-		d.streamWg.Add(1)
-		go d.startStreamDecoder(d.stream)
-	}
-
-	// If bytes buffer and < 1MB, do sync decoding anyway.
-	if bb, ok := r.(byter); ok && bb.Len() < 1<<20 {
+	// 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 debug {
+		if debugDecoder {
 			println("*bytes.Buffer detected, doing sync decode, len:", bb.Len())
 		}
 		b := bb2.Bytes()
 		var dst []byte
-		if cap(d.current.b) > 0 {
-			dst = d.current.b
+		if cap(d.syncStream.dstBuf) > 0 {
+			dst = d.syncStream.dstBuf[:0]
 		}
 
-		dst, err := d.DecodeAll(b, dst[: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 debug {
+		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.cancel = make(chan struct{})
 	d.current.flushed = false
 	d.current.d = nil
+	d.syncStream.dstBuf = nil
 
-	d.stream <- decodeStream{
-		r:      r,
-		output: d.current.output,
-		cancel: d.current.cancel,
+	// 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 {
-		println("cancelling current")
-		close(d.current.cancel)
+		if debugDecoder {
+			println("cancelling current")
+		}
+		d.current.cancel()
 		d.current.cancel = nil
 	}
 	if d.current.d != nil {
-		if debug {
+		if debugDecoder {
 			printf("re-adding current decoder %p, decoders: %d", d.current.d, len(d.decoders))
 		}
 		d.decoders <- d.current.d
@@ -238,34 +261,29 @@
 	}
 	for v := range d.current.output {
 		if v.d != nil {
-			if debug {
+			if debugDecoder {
 				printf("re-adding decoder %p", v.d)
 			}
 			d.decoders <- v.d
 		}
-		if v.err == errEndOfStream {
-			println("current flushed")
-			d.current.flushed = true
-			return
-		}
 	}
+	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) {
-	if d.stream == nil {
-		return 0, ErrDecoderNilInput
-	}
 	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 {
+			if err2 != nil && (d.current.err == nil || d.current.err == io.EOF) {
 				d.current.err = err2
-				break
+			} else if n2 != len(d.current.b) {
+				d.current.err = io.ErrShortWrite
 			}
 		}
 		if d.current.err != nil {
@@ -289,19 +307,24 @@
 // DecodeAll can be used concurrently.
 // The Decoder concurrency limits will be respected.
 func (d *Decoder) DecodeAll(input, dst []byte) ([]byte, error) {
-	if d.current.err == ErrDecoderClosed {
+	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 debug {
+		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
+			frame.history.decoders.br.cursor = 0
+		}
 		d.decoders <- block
 	}()
 	frame.bBuf = input
@@ -309,34 +332,45 @@
 	for {
 		frame.history.reset()
 		err := frame.reset(&frame.bBuf)
-		if err == io.EOF {
-			if debug {
-				println("frame reset return EOF")
-			}
-			return dst, nil
-		}
-		if frame.DictionaryID != nil {
-			dict, ok := d.dicts[*frame.DictionaryID]
-			if !ok {
-				return nil, ErrUnknownDictionary
-			}
-			frame.history.setDict(&dict)
-		}
 		if err != nil {
+			if err == io.EOF {
+				if debugDecoder {
+					println("frame reset return EOF")
+				}
+				return dst, nil
+			}
 			return dst, err
 		}
-		if frame.FrameContentSize > d.o.maxDecodedSize-uint64(len(dst)) {
-			return dst, ErrDecoderSizeExceeded
+		if err = d.setDict(frame); err != nil {
+			return nil, err
 		}
-		if frame.FrameContentSize > 0 && frame.FrameContentSize < 1<<30 {
-			// Never preallocate moe than 1 GB up front.
+		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))
+				dst2 := make([]byte, len(dst), len(dst)+int(frame.FrameContentSize)+compressedBlockOverAlloc)
 				copy(dst2, dst)
 				dst = dst2
 			}
 		}
-		if cap(dst) == 0 {
+
+		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
@@ -354,8 +388,11 @@
 		if err != nil {
 			return dst, err
 		}
+		if uint64(len(dst)-initialSize) > d.o.maxDecodedSize {
+			return dst, ErrDecoderSizeExceeded
+		}
 		if len(frame.bBuf) == 0 {
-			if debug {
+			if debugDecoder {
 				println("frame dbuf empty")
 			}
 			break
@@ -370,31 +407,165 @@
 // 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 debug {
+		if debugDecoder {
 			printf("re-adding current decoder %p", d.current.d)
 		}
 		d.decoders <- d.current.d
 		d.current.d = nil
 	}
-	if d.current.err != nil {
-		// Keep error state.
-		return blocking
-	}
-
-	if blocking {
-		d.current.decodeOutput = <-d.current.output
-	} else {
-		select {
-		case d.current.decodeOutput = <-d.current.output:
-		default:
-			return false
-		}
-	}
-	if debug {
-		println("got", len(d.current.b), "bytes, error:", d.current.err)
-	}
-	return true
 }
 
 // Close will release all resources.
@@ -404,10 +575,10 @@
 		return
 	}
 	d.drainOutput()
-	if d.stream != nil {
-		close(d.stream)
+	if d.current.cancel != nil {
+		d.current.cancel()
 		d.streamWg.Wait()
-		d.stream = nil
+		d.current.cancel = nil
 	}
 	if d.decoders != nil {
 		close(d.decoders)
@@ -458,100 +629,321 @@
 	err error
 }
 
-type decodeStream struct {
-	r io.Reader
-
-	// Blocks ready to be written to output.
-	output chan decodeOutput
-
-	// cancel reading from the input
-	cancel chan struct{}
+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
 }
 
-// errEndOfStream indicates that everything from the stream was read.
-var errEndOfStream = errors.New("end-of-stream")
-
 // Create Decoder:
-// Spawn n block decoders. These accept tasks to decode a block.
-// Create goroutine that handles stream processing, this will send history to decoders as they are available.
-// Decoders update the history as they decode.
-// When a block is returned:
-// 		a) history is sent to the next decoder,
-// 		b) content written to CRC.
-// 		c) return data to WRITER.
-// 		d) wait for next block to return data.
-// Once WRITTEN, the decoders reused by the writer frame decoder for re-use.
-func (d *Decoder) startStreamDecoder(inStream chan decodeStream) {
+// 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()
-	frame := newFrameDec(d.o)
-	for stream := range inStream {
-		if debug {
-			println("got new stream")
+	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
 		}
-		br := readerWrapper{r: stream.r}
-	decodeStream:
-		for {
-			frame.history.reset()
-			err := frame.reset(&br)
-			if debug && err != nil {
-				println("Frame decoder returned", err)
+		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 err == nil && frame.DictionaryID != nil {
-				dict, ok := d.dicts[*frame.DictionaryID]
-				if !ok {
-					err = ErrUnknownDictionary
+			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 {
-					frame.history.setDict(&dict)
+					if debugDecoder {
+						println("fcs ok", block.Last, fcs, decodedFrame)
+					}
 				}
 			}
-			if err != nil {
-				stream.output <- decodeOutput{
-					err: err,
+			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
 			}
-			if debug {
-				println("starting frame decoder")
-			}
-
-			// This goroutine will forward history between frames.
-			frame.frameDone.Add(1)
-			frame.initAsync()
-
-			go frame.startDecoder(stream.output)
-		decodeFrame:
-			// Go through all blocks of the frame.
-			for {
-				dec := <-d.decoders
-				select {
-				case <-stream.cancel:
-					if !frame.sendErr(dec, io.EOF) {
-						// To not let the decoder dangle, send it back.
-						stream.output <- decodeOutput{d: dec}
-					}
-					break decodeStream
-				default:
-				}
-				err := frame.next(dec)
-				switch err {
-				case io.EOF:
-					// End of current frame, no error
-					println("EOF on next block")
-					break decodeFrame
-				case nil:
-					continue
-				default:
-					println("block decoder returned", err)
-					break decodeStream
-				}
-			}
-			// All blocks have started decoding, check if there are more frames.
-			println("waiting for done")
-			frame.frameDone.Wait()
-			println("done waiting...")
 		}
-		frame.frameDone.Wait()
-		println("Sending EOS")
-		stream.output <- decodeOutput{err: errEndOfStream}
 	}
+	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
index c0fd058..774c5f0 100644
--- a/vendor/github.com/klauspost/compress/zstd/decoder_options.go
+++ b/vendor/github.com/klauspost/compress/zstd/decoder_options.go
@@ -6,6 +6,8 @@
 
 import (
 	"errors"
+	"fmt"
+	"math/bits"
 	"runtime"
 )
 
@@ -14,19 +16,28 @@
 
 // options retains accumulated state of multiple options.
 type decoderOptions struct {
-	lowMem         bool
-	concurrent     int
-	maxDecodedSize uint64
-	dicts          []dict
+	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),
+		lowMem:          true,
+		concurrent:      runtime.GOMAXPROCS(0),
+		maxWindowSize:   MaxWindowSize,
+		decodeBufsBelow: 128 << 10,
 	}
-	o.maxDecodedSize = 1 << 63
+	if o.concurrent > 4 {
+		o.concurrent = 4
+	}
+	o.maxDecodedSize = 64 << 30
 }
 
 // WithDecoderLowmem will set whether to use a lower amount of memory,
@@ -35,16 +46,25 @@
 	return func(o *decoderOptions) error { o.lowMem = b; return nil }
 }
 
-// WithDecoderConcurrency will set the concurrency,
-// meaning the maximum number of decoders to run concurrently.
-// The value supplied must be at least 1.
-// By default this will be set to GOMAXPROCS.
+// 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 {
+		if n < 0 {
 			return errors.New("concurrency must be at least 1")
 		}
-		o.concurrent = n
+		if n == 0 {
+			o.concurrent = runtime.GOMAXPROCS(0)
+		} else {
+			o.concurrent = n
+		}
 		return nil
 	}
 }
@@ -52,8 +72,7 @@
 // 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.
-// For streaming operations, the maximum window size is capped at 1<<30 bytes.
-// Maximum and default is 1 << 63 bytes.
+// Maximum is 1 << 63 bytes. Default is 64GiB.
 func WithDecoderMaxMemory(n uint64) DOption {
 	return func(o *decoderOptions) error {
 		if n == 0 {
@@ -68,7 +87,13 @@
 }
 
 // WithDecoderDicts allows to register one or more dictionaries for the decoder.
-// If several dictionaries with the same ID is provided the last one will be used.
+//
+// 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 {
@@ -76,8 +101,69 @@
 			if err != nil {
 				return err
 			}
-			o.dicts = append(o.dicts, *d)
+			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
index fa25a18..b7b8316 100644
--- a/vendor/github.com/klauspost/compress/zstd/dict.go
+++ b/vendor/github.com/klauspost/compress/zstd/dict.go
@@ -6,6 +6,8 @@
 	"errors"
 	"fmt"
 	"io"
+	"math"
+	"sort"
 
 	"github.com/klauspost/compress/huff0"
 )
@@ -15,12 +17,14 @@
 
 	litEnc              *huff0.Scratch
 	llDec, ofDec, mlDec sequenceDec
-	//llEnc, ofEnc, mlEnc []*fseEncoder
-	offsets [3]int
-	content []byte
+	offsets             [3]int
+	content             []byte
 }
 
-var dictMagic = [4]byte{0x37, 0xa4, 0x30, 0xec}
+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 {
@@ -30,14 +34,38 @@
 	return d.id
 }
 
-// DictContentSize returns the dictionary content size or 0 if d is nil.
-func (d *dict) DictContentSize() int {
+// 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) {
@@ -50,7 +78,7 @@
 		ofDec: sequenceDec{fse: &fseDecoder{}},
 		mlDec: sequenceDec{fse: &fseDecoder{}},
 	}
-	if !bytes.Equal(b[:4], dictMagic[:]) {
+	if string(b[:4]) != dictMagic {
 		return nil, ErrMagicMismatch
 	}
 	d.id = binary.LittleEndian.Uint32(b[4:8])
@@ -62,7 +90,7 @@
 	var err error
 	d.litEnc, b, err = huff0.ReadTable(b[8:], nil)
 	if err != nil {
-		return nil, err
+		return nil, fmt.Errorf("loading literal table: %w", err)
 	}
 	d.litEnc.Reuse = huff0.ReusePolicyMust
 
@@ -82,7 +110,7 @@
 			println("Transform table error:", err)
 			return err
 		}
-		if debug {
+		if debugDecoder || debugEncoder {
 			println("Read table ok", "symbolLen:", dec.symbolLen)
 		}
 		// Set decoders as predefined so they aren't reused.
@@ -120,3 +148,418 @@
 
 	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
index 60f2986..7d250c6 100644
--- a/vendor/github.com/klauspost/compress/zstd/enc_base.go
+++ b/vendor/github.com/klauspost/compress/zstd/enc_base.go
@@ -16,6 +16,7 @@
 	cur int32
 	// maximum offset. Should be at least 2x block size.
 	maxMatchOff int32
+	bufferReset int32
 	hist        []byte
 	crc         *xxhash.Digest
 	tmp         [8]byte
@@ -38,8 +39,8 @@
 
 // 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 int) int32 {
-	if size > 0 && size < int(e.maxMatchOff) {
+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 {
@@ -56,8 +57,8 @@
 }
 
 func (e *fastBase) addBlock(src []byte) int32 {
-	if debugAsserts && e.cur > bufferReset {
-		panic(fmt.Sprintf("ecur (%d) > buffer reset (%d)", e.cur, bufferReset))
+	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) {
@@ -108,11 +109,6 @@
 	e.blk = enc
 }
 
-func (e *fastBase) matchlenNoHist(s, t int32, src []byte) int32 {
-	// Extend the match to be as long as possible.
-	return int32(matchLen(src[s:], src[t:]))
-}
-
 func (e *fastBase) matchlen(s, t int32, src []byte) int32 {
 	if debugAsserts {
 		if s < 0 {
@@ -120,7 +116,7 @@
 			panic(err)
 		}
 		if t < 0 {
-			err := fmt.Sprintf("s (%d) < 0", s)
+			err := fmt.Sprintf("t (%d) < 0", t)
 			panic(err)
 		}
 		if s-t > e.maxMatchOff {
@@ -131,8 +127,6 @@
 			panic(fmt.Sprintf("len(src)-s (%d) > maxCompressedBlockSize (%d)", len(src)-int(s), maxCompressedBlockSize))
 		}
 	}
-
-	// Extend the match to be as long as possible.
 	return int32(matchLen(src[s:], src[t:]))
 }
 
@@ -150,18 +144,19 @@
 	} else {
 		e.crc.Reset()
 	}
+	e.blk.dictLitEnc = nil
 	if d != nil {
 		low := e.lowMem
 		if singleBlock {
 			e.lowMem = true
 		}
-		e.ensureHist(d.DictContentSize() + maxCompressedBlockSize)
+		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 < bufferReset {
+	if e.cur < e.bufferReset {
 		e.cur += e.maxMatchOff + int32(len(e.hist))
 	}
 	e.hist = e.hist[:0]
diff --git a/vendor/github.com/klauspost/compress/zstd/enc_best.go b/vendor/github.com/klauspost/compress/zstd/enc_best.go
index dc1eed5..4613724 100644
--- a/vendor/github.com/klauspost/compress/zstd/enc_best.go
+++ b/vendor/github.com/klauspost/compress/zstd/enc_best.go
@@ -5,22 +5,61 @@
 package zstd
 
 import (
+	"bytes"
 	"fmt"
-	"math/bits"
+
+	"github.com/klauspost/compress"
 )
 
 const (
-	bestLongTableBits = 20                     // Bits used in the long match table
+	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 = 16                      // Bits used in the short match table
+	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.
@@ -45,14 +84,10 @@
 	)
 
 	// Protect against e.cur wraparound.
-	for e.cur >= bufferReset {
+	for e.cur >= e.bufferReset-int32(len(e.hist)) {
 		if len(e.hist) == 0 {
-			for i := range e.table[:] {
-				e.table[i] = prevEntry{}
-			}
-			for i := range e.longTable[:] {
-				e.longTable[i] = prevEntry{}
-			}
+			e.table = [bestShortTableSize]prevEntry{}
+			e.longTable = [bestLongTableSize]prevEntry{}
 			e.cur = e.maxMatchOff
 			break
 		}
@@ -100,8 +135,20 @@
 		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)]
@@ -109,6 +156,14 @@
 		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
@@ -116,7 +171,6 @@
 
 	// nextEmit is where in src the next emitLiteral should start from.
 	nextEmit := s
-	cv := load6432(src, s)
 
 	// Relative offsets
 	offset1 := int32(blk.recentOffsets[0])
@@ -130,9 +184,8 @@
 		blk.literals = append(blk.literals, src[nextEmit:until]...)
 		s.litLen = uint32(until - nextEmit)
 	}
-	_ = addLiterals
 
-	if debug {
+	if debugEncoder {
 		println("recent offsets:", blk.recentOffsets)
 	}
 
@@ -145,56 +198,104 @@
 			panic("offset0 was 0")
 		}
 
-		type match struct {
-			offset int32
-			s      int32
-			length int32
-			rep    int32
-		}
-		matchAt := func(offset int32, s int32, first uint32, rep int32) match {
-			if s-offset >= e.maxMatchOff || load3232(src, offset) != first {
-				return match{offset: offset, s: s}
-			}
-			return match{offset: offset, s: s, length: 4 + e.matchlen(s+4, offset+4, src), rep: rep}
-		}
+		const goodEnough = 250
 
-		bestOf := func(a, b match) match {
-			aScore := b.s - a.s + a.length
-			bScore := a.s - b.s + b.length
-			if a.rep < 0 {
-				aScore = aScore - int32(bits.Len32(uint32(a.offset)))/8
-			}
-			if b.rep < 0 {
-				bScore = bScore - int32(bits.Len32(uint32(b.offset)))/8
-			}
-			if aScore >= bScore {
-				return a
-			}
-			return b
-		}
-		const goodEnough = 100
+		cv := load6432(src, s)
 
-		nextHashL := hash8(cv, bestLongTableBits)
-		nextHashS := hash4x64(cv, bestShortTableBits)
+		nextHashL := hashLen(cv, bestLongTableBits, bestLongLen)
+		nextHashS := hashLen(cv, bestShortTableBits, bestShortLen)
 		candidateL := e.longTable[nextHashL]
 		candidateS := e.table[nextHashS]
 
-		best := bestOf(matchAt(candidateL.offset-e.cur, s, uint32(cv), -1), matchAt(candidateL.prev-e.cur, s, uint32(cv), -1))
-		best = bestOf(best, matchAt(candidateS.offset-e.cur, s, uint32(cv), -1))
-		best = bestOf(best, matchAt(candidateS.prev-e.cur, s, uint32(cv), -1))
+		// 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 {
-			best = bestOf(best, matchAt(s-offset1+1, s+1, uint32(cv>>8), 1))
-			best = bestOf(best, matchAt(s-offset2+1, s+1, uint32(cv>>8), 2))
-			best = bestOf(best, matchAt(s-offset3+1, s+1, uint32(cv>>8), 3))
-			if best.length > 0 {
-				best = bestOf(best, matchAt(s-offset1+3, s+3, uint32(cv>>24), 1))
-				best = bestOf(best, matchAt(s-offset2+3, s+3, uint32(cv>>24), 2))
-				best = bestOf(best, matchAt(s-offset3+3, s+3, uint32(cv>>24), 3))
+			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 {
@@ -204,106 +305,117 @@
 				if s >= sLimit {
 					break encodeLoop
 				}
-				cv = load6432(src, s)
 				continue
 			}
 
-			s++
-			candidateS = e.table[hash4x64(cv>>8, bestShortTableBits)]
-			cv = load6432(src, s)
-			cv2 := load6432(src, s+1)
-			candidateL = e.longTable[hash8(cv, bestLongTableBits)]
-			candidateL2 := e.longTable[hash8(cv2, bestLongTableBits)]
+			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)]
 
-			best = bestOf(best, matchAt(candidateS.offset-e.cur, s, uint32(cv), -1))
-			best = bestOf(best, matchAt(candidateL.offset-e.cur, s, uint32(cv), -1))
-			best = bestOf(best, matchAt(candidateL.prev-e.cur, s, uint32(cv), -1))
-			best = bestOf(best, matchAt(candidateL2.offset-e.cur, s+1, uint32(cv2), -1))
-			best = bestOf(best, matchAt(candidateL2.prev-e.cur, s+1, uint32(cv2), -1))
+			// 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)
+			}
 
-			// 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 := hash8(load6432(src, sAt), bestLongTableBits)
-				candidateEnd := e.longTable[nextHashL]
-				if pos := candidateEnd.offset - e.cur - best.length; pos >= 0 {
-					bestEnd := bestOf(best, matchAt(pos, best.s, load3232(src, best.s), -1))
-					if pos := candidateEnd.prev - e.cur - best.length; pos >= 0 {
-						bestEnd = bestOf(bestEnd, matchAt(pos, best.s, load3232(src, best.s), -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)
+						}
 					}
-					best = bestEnd
 				}
 			}
 		}
 
+		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 {
-			s = best.s
 			var seq seq
 			seq.matchLen = uint32(best.length - zstdMinMatch)
+			addLiterals(&seq, best.s)
 
-			// We might be able to match backwards.
-			// Extend as long as we can.
-			start := best.s
-			// 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
-			}
-			repIndex := best.offset
-			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 = uint32(best.rep)
+			// 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:", s)
+				println("repeat sequence", seq, "next s:", best.s, "off:", best.s-best.offset)
 			}
 			blk.sequences = append(blk.sequences, seq)
 
-			// Index match start+1 (long) -> s - 1
-			index0 := s
+			// Index old s + 1 -> s - 1
 			s = best.s + best.length
-
 			nextEmit = s
-			if s >= sLimit {
-				if debug {
-					println("repeat ended", s, best.length)
 
-				}
-				break encodeLoop
-			}
 			// Index skipped...
+			end := s
+			if s > sLimit+4 {
+				end = sLimit + 4
+			}
 			off := index0 + e.cur
-			for index0 < s-1 {
+			for index0 < end {
 				cv0 := load6432(src, index0)
-				h0 := hash8(cv0, bestLongTableBits)
-				h1 := hash4x64(cv0, bestShortTableBits)
+				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:
+			case 2, 4 | 1:
 				offset1, offset2 = offset2, offset1
-			case 3:
+			case 3, 4 | 2:
 				offset1, offset2, offset3 = offset3, offset1, offset2
+			case 4 | 3:
+				offset1, offset2, offset3 = offset1-1, offset1, offset2
 			}
-			cv = load6432(src, s)
+			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.
-		s = best.s
 		t := best.offset
 		offset1, offset2, offset3 = s-t, offset1, offset2
 
@@ -311,26 +423,13 @@
 			panic(fmt.Sprintf("s (%d) <= t (%d)", s, t))
 		}
 
-		if debugAsserts && canRepeat && int(offset1) > len(src) {
+		if debugAsserts && int(offset1) > len(src) {
 			panic("invalid offset")
 		}
 
-		// Extend the n-byte match as long as possible.
-		l := best.length
-
-		// 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
+		l := best.length
 		seq.litLen = uint32(s - nextEmit)
 		seq.matchLen = uint32(l - zstdMinMatch)
 		if seq.litLen > 0 {
@@ -343,65 +442,25 @@
 		}
 		blk.sequences = append(blk.sequences, seq)
 		nextEmit = s
-		if s >= sLimit {
-			break encodeLoop
+
+		// Index old s + 1 -> s - 1 or sLimit
+		end := s
+		if s > sLimit-4 {
+			end = sLimit - 4
 		}
 
-		// Index match start+1 (long) -> s - 1
-		index0 := s - l + 1
-		// every entry
-		for index0 < s-1 {
+		off := index0 + e.cur
+		for index0 < end {
 			cv0 := load6432(src, index0)
-			h0 := hash8(cv0, bestLongTableBits)
-			h1 := hash4x64(cv0, bestShortTableBits)
-			off := index0 + e.cur
+			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++
 		}
-
-		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 := hash4x64(cv, bestShortTableBits)
-			nextHashL := hash8(cv, bestLongTableBits)
-
-			// 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] = prevEntry{offset: s + e.cur, prev: e.table[nextHashS].offset}
-			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 s >= sLimit {
+			break encodeLoop
 		}
 	}
 
@@ -412,7 +471,7 @@
 	blk.recentOffsets[0] = uint32(offset1)
 	blk.recentOffsets[1] = uint32(offset2)
 	blk.recentOffsets[2] = uint32(offset3)
-	if debug {
+	if debugEncoder {
 		println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits)
 	}
 }
@@ -425,7 +484,7 @@
 	e.Encode(blk, src)
 }
 
-// ResetDict will reset and set a dictionary if not nil
+// 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 {
@@ -441,10 +500,10 @@
 			const hashLog = bestShortTableBits
 
 			cv := load6432(d.content, i-e.maxMatchOff)
-			nextHash := hash4x64(cv, hashLog)      // 0 -> 4
-			nextHash1 := hash4x64(cv>>8, hashLog)  // 1 -> 5
-			nextHash2 := hash4x64(cv>>16, hashLog) // 2 -> 6
-			nextHash3 := hash4x64(cv>>24, hashLog) // 3 -> 7
+			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,
@@ -472,7 +531,7 @@
 		}
 		if len(d.content) >= 8 {
 			cv := load6432(d.content, 0)
-			h := hash8(cv, bestLongTableBits)
+			h := hashLen(cv, bestLongTableBits, bestLongLen)
 			e.dictLongTable[h] = prevEntry{
 				offset: e.maxMatchOff,
 				prev:   e.dictLongTable[h].offset,
@@ -482,7 +541,7 @@
 			off := 8 // First to read
 			for i := e.maxMatchOff + 1; i < end; i++ {
 				cv = cv>>8 | (uint64(d.content[off]) << 56)
-				h := hash8(cv, bestLongTableBits)
+				h := hashLen(cv, bestLongTableBits, bestLongLen)
 				e.dictLongTable[h] = prevEntry{
 					offset: i,
 					prev:   e.dictLongTable[h].offset,
diff --git a/vendor/github.com/klauspost/compress/zstd/enc_better.go b/vendor/github.com/klauspost/compress/zstd/enc_better.go
index 6049542..84a79fd 100644
--- a/vendor/github.com/klauspost/compress/zstd/enc_better.go
+++ b/vendor/github.com/klauspost/compress/zstd/enc_better.go
@@ -9,6 +9,7 @@
 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
@@ -16,6 +17,7 @@
 	// 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
@@ -60,14 +62,10 @@
 	)
 
 	// Protect against e.cur wraparound.
-	for e.cur >= bufferReset {
+	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.table = [betterShortTableSize]tableEntry{}
+			e.longTable = [betterLongTableSize]prevEntry{}
 			e.cur = e.maxMatchOff
 			break
 		}
@@ -104,9 +102,20 @@
 		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)]
@@ -138,7 +147,7 @@
 		blk.literals = append(blk.literals, src[nextEmit:until]...)
 		s.litLen = uint32(until - nextEmit)
 	}
-	if debug {
+	if debugEncoder {
 		println("recent offsets:", blk.recentOffsets)
 	}
 
@@ -147,15 +156,15 @@
 		var t int32
 		// We allow the encoder to optionally turn off repeat offsets across blocks
 		canRepeat := len(blk.sequences) > 2
-		var matched int32
+		var matched, index0 int32
 
 		for {
 			if debugAsserts && canRepeat && offset1 == 0 {
 				panic("offset0 was 0")
 			}
 
-			nextHashS := hash5(cv, betterShortTableBits)
-			nextHashL := hash8(cv, betterLongTableBits)
+			nextHashL := hashLen(cv, betterLongTableBits, betterLongLen)
+			nextHashS := hashLen(cv, betterShortTableBits, betterShortLen)
 			candidateL := e.longTable[nextHashL]
 			candidateS := e.table[nextHashS]
 
@@ -164,14 +173,15 @@
 			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)
+					length := 4 + e.matchlen(s+4+repOff, repIndex+4, src)
 
-					seq.matchLen = uint32(lenght - zstdMinMatch)
+					seq.matchLen = uint32(length - zstdMinMatch)
 
 					// We might be able to match backwards.
 					// Extend as long as we can.
@@ -200,12 +210,12 @@
 
 					// Index match start+1 (long) -> s - 1
 					index0 := s + repOff
-					s += lenght + repOff
+					s += length + repOff
 
 					nextEmit = s
 					if s >= sLimit {
-						if debug {
-							println("repeat ended", s, lenght)
+						if debugEncoder {
+							println("repeat ended", s, length)
 
 						}
 						break encodeLoop
@@ -214,10 +224,10 @@
 					for index0 < s-1 {
 						cv0 := load6432(src, index0)
 						cv1 := cv0 >> 8
-						h0 := hash8(cv0, betterLongTableBits)
+						h0 := hashLen(cv0, betterLongTableBits, betterLongLen)
 						off := index0 + e.cur
 						e.longTable[h0] = prevEntry{offset: off, prev: e.longTable[h0].offset}
-						e.table[hash5(cv1, betterShortTableBits)] = tableEntry{offset: off + 1, val: uint32(cv1)}
+						e.table[hashLen(cv1, betterShortTableBits, betterShortLen)] = tableEntry{offset: off + 1, val: uint32(cv1)}
 						index0 += 2
 					}
 					cv = load6432(src, s)
@@ -231,9 +241,9 @@
 				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)
+					length := 8 + e.matchlen(s+8+repOff2, repIndex+8, src)
 
-					seq.matchLen = uint32(lenght - zstdMinMatch)
+					seq.matchLen = uint32(length - zstdMinMatch)
 
 					// We might be able to match backwards.
 					// Extend as long as we can.
@@ -260,12 +270,11 @@
 					}
 					blk.sequences = append(blk.sequences, seq)
 
-					index0 := s + repOff2
-					s += lenght + repOff2
+					s += length + repOff2
 					nextEmit = s
 					if s >= sLimit {
-						if debug {
-							println("repeat ended", s, lenght)
+						if debugEncoder {
+							println("repeat ended", s, length)
 
 						}
 						break encodeLoop
@@ -275,10 +284,10 @@
 					for index0 < s-1 {
 						cv0 := load6432(src, index0)
 						cv1 := cv0 >> 8
-						h0 := hash8(cv0, betterLongTableBits)
+						h0 := hashLen(cv0, betterLongTableBits, betterLongLen)
 						off := index0 + e.cur
 						e.longTable[h0] = prevEntry{offset: off, prev: e.longTable[h0].offset}
-						e.table[hash5(cv1, betterShortTableBits)] = tableEntry{offset: off + 1, val: uint32(cv1)}
+						e.table[hashLen(cv1, betterShortTableBits, betterShortLen)] = tableEntry{offset: off + 1, val: uint32(cv1)}
 						index0 += 2
 					}
 					cv = load6432(src, s)
@@ -353,7 +362,7 @@
 				// See if we can find a long match at s+1
 				const checkAt = 1
 				cv := load6432(src, s+checkAt)
-				nextHashL = hash8(cv, betterLongTableBits)
+				nextHashL = hashLen(cv, betterLongTableBits, betterLongLen)
 				candidateL = e.longTable[nextHashL]
 				coffsetL = candidateL.offset - e.cur
 
@@ -413,16 +422,24 @@
 		}
 
 		// Try to find a better match by searching for a long match at the end of the current best match
-		if true && s+matched < sLimit {
-			nextHashL := hash8(load6432(src, s+matched), betterLongTableBits)
-			cv := load3232(src, s)
+		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
-			if coffsetL >= 0 && coffsetL < s && s-coffsetL < e.maxMatchOff && cv == load3232(src, coffsetL) {
+			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(s+4, coffsetL+4, src) + 4
+				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")
@@ -432,12 +449,13 @@
 
 			// Check prev long...
 			if true {
-				coffsetL = candidateL.prev - e.cur - matched
-				if coffsetL >= 0 && coffsetL < s && s-coffsetL < e.maxMatchOff && cv == load3232(src, coffsetL) {
+				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(s+4, coffsetL+4, src) + 4
+					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")
@@ -491,15 +509,15 @@
 		}
 
 		// Index match start+1 (long) -> s - 1
-		index0 := s - l + 1
+		off := index0 + e.cur
 		for index0 < s-1 {
 			cv0 := load6432(src, index0)
 			cv1 := cv0 >> 8
-			h0 := hash8(cv0, betterLongTableBits)
-			off := index0 + e.cur
+			h0 := hashLen(cv0, betterLongTableBits, betterLongLen)
 			e.longTable[h0] = prevEntry{offset: off, prev: e.longTable[h0].offset}
-			e.table[hash5(cv1, betterShortTableBits)] = tableEntry{offset: off + 1, val: uint32(cv1)}
+			e.table[hashLen(cv1, betterShortTableBits, betterShortLen)] = tableEntry{offset: off + 1, val: uint32(cv1)}
 			index0 += 2
+			off += 2
 		}
 
 		cv = load6432(src, s)
@@ -516,8 +534,8 @@
 			}
 
 			// Store this, since we have it.
-			nextHashS := hash5(cv, betterShortTableBits)
-			nextHashL := hash8(cv, betterLongTableBits)
+			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
@@ -553,7 +571,7 @@
 	}
 	blk.recentOffsets[0] = uint32(offset1)
 	blk.recentOffsets[1] = uint32(offset2)
-	if debug {
+	if debugEncoder {
 		println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits)
 	}
 }
@@ -576,7 +594,7 @@
 	)
 
 	// Protect against e.cur wraparound.
-	for e.cur >= bufferReset {
+	for e.cur >= e.bufferReset-int32(len(e.hist)) {
 		if len(e.hist) == 0 {
 			for i := range e.table[:] {
 				e.table[i] = tableEntry{}
@@ -656,7 +674,7 @@
 		blk.literals = append(blk.literals, src[nextEmit:until]...)
 		s.litLen = uint32(until - nextEmit)
 	}
-	if debug {
+	if debugEncoder {
 		println("recent offsets:", blk.recentOffsets)
 	}
 
@@ -665,15 +683,15 @@
 		var t int32
 		// We allow the encoder to optionally turn off repeat offsets across blocks
 		canRepeat := len(blk.sequences) > 2
-		var matched int32
+		var matched, index0 int32
 
 		for {
 			if debugAsserts && canRepeat && offset1 == 0 {
 				panic("offset0 was 0")
 			}
 
-			nextHashS := hash5(cv, betterShortTableBits)
-			nextHashL := hash8(cv, betterLongTableBits)
+			nextHashL := hashLen(cv, betterLongTableBits, betterLongLen)
+			nextHashS := hashLen(cv, betterShortTableBits, betterShortLen)
 			candidateL := e.longTable[nextHashL]
 			candidateS := e.table[nextHashS]
 
@@ -684,14 +702,15 @@
 			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)
+					length := 4 + e.matchlen(s+4+repOff, repIndex+4, src)
 
-					seq.matchLen = uint32(lenght - zstdMinMatch)
+					seq.matchLen = uint32(length - zstdMinMatch)
 
 					// We might be able to match backwards.
 					// Extend as long as we can.
@@ -719,13 +738,12 @@
 					blk.sequences = append(blk.sequences, seq)
 
 					// Index match start+1 (long) -> s - 1
-					index0 := s + repOff
-					s += lenght + repOff
+					s += length + repOff
 
 					nextEmit = s
 					if s >= sLimit {
-						if debug {
-							println("repeat ended", s, lenght)
+						if debugEncoder {
+							println("repeat ended", s, length)
 
 						}
 						break encodeLoop
@@ -734,11 +752,11 @@
 					for index0 < s-1 {
 						cv0 := load6432(src, index0)
 						cv1 := cv0 >> 8
-						h0 := hash8(cv0, betterLongTableBits)
+						h0 := hashLen(cv0, betterLongTableBits, betterLongLen)
 						off := index0 + e.cur
 						e.longTable[h0] = prevEntry{offset: off, prev: e.longTable[h0].offset}
 						e.markLongShardDirty(h0)
-						h1 := hash5(cv1, betterShortTableBits)
+						h1 := hashLen(cv1, betterShortTableBits, betterShortLen)
 						e.table[h1] = tableEntry{offset: off + 1, val: uint32(cv1)}
 						e.markShortShardDirty(h1)
 						index0 += 2
@@ -754,9 +772,9 @@
 				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)
+					length := 8 + e.matchlen(s+8+repOff2, repIndex+8, src)
 
-					seq.matchLen = uint32(lenght - zstdMinMatch)
+					seq.matchLen = uint32(length - zstdMinMatch)
 
 					// We might be able to match backwards.
 					// Extend as long as we can.
@@ -783,12 +801,11 @@
 					}
 					blk.sequences = append(blk.sequences, seq)
 
-					index0 := s + repOff2
-					s += lenght + repOff2
+					s += length + repOff2
 					nextEmit = s
 					if s >= sLimit {
-						if debug {
-							println("repeat ended", s, lenght)
+						if debugEncoder {
+							println("repeat ended", s, length)
 
 						}
 						break encodeLoop
@@ -798,11 +815,11 @@
 					for index0 < s-1 {
 						cv0 := load6432(src, index0)
 						cv1 := cv0 >> 8
-						h0 := hash8(cv0, betterLongTableBits)
+						h0 := hashLen(cv0, betterLongTableBits, betterLongLen)
 						off := index0 + e.cur
 						e.longTable[h0] = prevEntry{offset: off, prev: e.longTable[h0].offset}
 						e.markLongShardDirty(h0)
-						h1 := hash5(cv1, betterShortTableBits)
+						h1 := hashLen(cv1, betterShortTableBits, betterShortLen)
 						e.table[h1] = tableEntry{offset: off + 1, val: uint32(cv1)}
 						e.markShortShardDirty(h1)
 						index0 += 2
@@ -879,7 +896,7 @@
 				// See if we can find a long match at s+1
 				const checkAt = 1
 				cv := load6432(src, s+checkAt)
-				nextHashL = hash8(cv, betterLongTableBits)
+				nextHashL = hashLen(cv, betterLongTableBits, betterLongLen)
 				candidateL = e.longTable[nextHashL]
 				coffsetL = candidateL.offset - e.cur
 
@@ -940,7 +957,7 @@
 		}
 		// 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 := hash8(load6432(src, s+matched), betterLongTableBits)
+			nextHashL := hashLen(load6432(src, s+matched), betterLongTableBits, betterLongLen)
 			cv := load3232(src, s)
 			candidateL := e.longTable[nextHashL]
 			coffsetL := candidateL.offset - e.cur - matched
@@ -1017,18 +1034,18 @@
 		}
 
 		// Index match start+1 (long) -> s - 1
-		index0 := s - l + 1
+		off := index0 + e.cur
 		for index0 < s-1 {
 			cv0 := load6432(src, index0)
 			cv1 := cv0 >> 8
-			h0 := hash8(cv0, betterLongTableBits)
-			off := index0 + e.cur
+			h0 := hashLen(cv0, betterLongTableBits, betterLongLen)
 			e.longTable[h0] = prevEntry{offset: off, prev: e.longTable[h0].offset}
 			e.markLongShardDirty(h0)
-			h1 := hash5(cv1, betterShortTableBits)
+			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)
@@ -1045,8 +1062,8 @@
 			}
 
 			// Store this, since we have it.
-			nextHashS := hash5(cv, betterShortTableBits)
-			nextHashL := hash8(cv, betterLongTableBits)
+			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
@@ -1084,7 +1101,7 @@
 	}
 	blk.recentOffsets[0] = uint32(offset1)
 	blk.recentOffsets[1] = uint32(offset2)
-	if debug {
+	if debugEncoder {
 		println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits)
 	}
 }
@@ -1113,10 +1130,10 @@
 			const hashLog = betterShortTableBits
 
 			cv := load6432(d.content, i-e.maxMatchOff)
-			nextHash := hash5(cv, hashLog)      // 0 -> 4
-			nextHash1 := hash5(cv>>8, hashLog)  // 1 -> 5
-			nextHash2 := hash5(cv>>16, hashLog) // 2 -> 6
-			nextHash3 := hash5(cv>>24, hashLog) // 3 -> 7
+			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,
@@ -1145,7 +1162,7 @@
 		}
 		if len(d.content) >= 8 {
 			cv := load6432(d.content, 0)
-			h := hash8(cv, betterLongTableBits)
+			h := hashLen(cv, betterLongTableBits, betterLongLen)
 			e.dictLongTable[h] = prevEntry{
 				offset: e.maxMatchOff,
 				prev:   e.dictLongTable[h].offset,
@@ -1155,7 +1172,7 @@
 			off := 8 // First to read
 			for i := e.maxMatchOff + 1; i < end; i++ {
 				cv = cv>>8 | (uint64(d.content[off]) << 56)
-				h := hash8(cv, betterLongTableBits)
+				h := hashLen(cv, betterLongTableBits, betterLongLen)
 				e.dictLongTable[h] = prevEntry{
 					offset: i,
 					prev:   e.dictLongTable[h].offset,
diff --git a/vendor/github.com/klauspost/compress/zstd/enc_dfast.go b/vendor/github.com/klauspost/compress/zstd/enc_dfast.go
index 8629d43..d36be7b 100644
--- a/vendor/github.com/klauspost/compress/zstd/enc_dfast.go
+++ b/vendor/github.com/klauspost/compress/zstd/enc_dfast.go
@@ -10,6 +10,7 @@
 	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
@@ -17,6 +18,8 @@
 	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 {
@@ -41,14 +44,10 @@
 	)
 
 	// Protect against e.cur wraparound.
-	for e.cur >= bufferReset {
+	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.table = [dFastShortTableSize]tableEntry{}
+			e.longTable = [dFastLongTableSize]tableEntry{}
 			e.cur = e.maxMatchOff
 			break
 		}
@@ -109,7 +108,7 @@
 		blk.literals = append(blk.literals, src[nextEmit:until]...)
 		s.litLen = uint32(until - nextEmit)
 	}
-	if debug {
+	if debugEncoder {
 		println("recent offsets:", blk.recentOffsets)
 	}
 
@@ -124,8 +123,8 @@
 				panic("offset0 was 0")
 			}
 
-			nextHashS := hash5(cv, dFastShortTableBits)
-			nextHashL := hash8(cv, dFastLongTableBits)
+			nextHashL := hashLen(cv, dFastLongTableBits, dFastLongLen)
+			nextHashS := hashLen(cv, dFastShortTableBits, dFastShortLen)
 			candidateL := e.longTable[nextHashL]
 			candidateS := e.table[nextHashS]
 
@@ -139,9 +138,9 @@
 				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)
+					length := 4 + e.matchlen(s+4+repOff, repIndex+4, src)
 
-					seq.matchLen = uint32(lenght - zstdMinMatch)
+					seq.matchLen = uint32(length - zstdMinMatch)
 
 					// We might be able to match backwards.
 					// Extend as long as we can.
@@ -167,11 +166,11 @@
 						println("repeat sequence", seq, "next s:", s)
 					}
 					blk.sequences = append(blk.sequences, seq)
-					s += lenght + repOff
+					s += length + repOff
 					nextEmit = s
 					if s >= sLimit {
-						if debug {
-							println("repeat ended", s, lenght)
+						if debugEncoder {
+							println("repeat ended", s, length)
 
 						}
 						break encodeLoop
@@ -208,7 +207,7 @@
 				// See if we can find a long match at s+1
 				const checkAt = 1
 				cv := load6432(src, s+checkAt)
-				nextHashL = hash8(cv, dFastLongTableBits)
+				nextHashL = hashLen(cv, dFastLongTableBits, dFastLongLen)
 				candidateL = e.longTable[nextHashL]
 				coffsetL = s - (candidateL.offset - e.cur) + checkAt
 
@@ -304,16 +303,16 @@
 		cv1 := load6432(src, index1)
 		te0 := tableEntry{offset: index0 + e.cur, val: uint32(cv0)}
 		te1 := tableEntry{offset: index1 + e.cur, val: uint32(cv1)}
-		e.longTable[hash8(cv0, dFastLongTableBits)] = te0
-		e.longTable[hash8(cv1, dFastLongTableBits)] = te1
+		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[hash5(cv0, dFastShortTableBits)] = te0
-		e.table[hash5(cv1, dFastShortTableBits)] = te1
+		e.table[hashLen(cv0, dFastShortTableBits, dFastShortLen)] = te0
+		e.table[hashLen(cv1, dFastShortTableBits, dFastShortLen)] = te1
 
 		cv = load6432(src, s)
 
@@ -330,8 +329,8 @@
 			}
 
 			// Store this, since we have it.
-			nextHashS := hash5(cv, dFastShortTableBits)
-			nextHashL := hash8(cv, dFastLongTableBits)
+			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
@@ -368,7 +367,7 @@
 	}
 	blk.recentOffsets[0] = uint32(offset1)
 	blk.recentOffsets[1] = uint32(offset2)
-	if debug {
+	if debugEncoder {
 		println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits)
 	}
 }
@@ -385,7 +384,7 @@
 	)
 
 	// Protect against e.cur wraparound.
-	if e.cur >= bufferReset {
+	if e.cur >= e.bufferReset {
 		for i := range e.table[:] {
 			e.table[i] = tableEntry{}
 		}
@@ -427,7 +426,7 @@
 		blk.literals = append(blk.literals, src[nextEmit:until]...)
 		s.litLen = uint32(until - nextEmit)
 	}
-	if debug {
+	if debugEncoder {
 		println("recent offsets:", blk.recentOffsets)
 	}
 
@@ -436,8 +435,8 @@
 		var t int32
 		for {
 
-			nextHashS := hash5(cv, dFastShortTableBits)
-			nextHashL := hash8(cv, dFastLongTableBits)
+			nextHashL := hashLen(cv, dFastLongTableBits, dFastLongLen)
+			nextHashS := hashLen(cv, dFastShortTableBits, dFastShortLen)
 			candidateL := e.longTable[nextHashL]
 			candidateS := e.table[nextHashS]
 
@@ -483,7 +482,7 @@
 					s += length + repOff
 					nextEmit = s
 					if s >= sLimit {
-						if debug {
+						if debugEncoder {
 							println("repeat ended", s, length)
 
 						}
@@ -521,7 +520,7 @@
 				// See if we can find a long match at s+1
 				const checkAt = 1
 				cv := load6432(src, s+checkAt)
-				nextHashL = hash8(cv, dFastLongTableBits)
+				nextHashL = hashLen(cv, dFastLongTableBits, dFastLongLen)
 				candidateL = e.longTable[nextHashL]
 				coffsetL = s - (candidateL.offset - e.cur) + checkAt
 
@@ -614,16 +613,16 @@
 		cv1 := load6432(src, index1)
 		te0 := tableEntry{offset: index0 + e.cur, val: uint32(cv0)}
 		te1 := tableEntry{offset: index1 + e.cur, val: uint32(cv1)}
-		e.longTable[hash8(cv0, dFastLongTableBits)] = te0
-		e.longTable[hash8(cv1, dFastLongTableBits)] = te1
+		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[hash5(cv0, dFastShortTableBits)] = te0
-		e.table[hash5(cv1, dFastShortTableBits)] = te1
+		e.table[hashLen(cv0, dFastShortTableBits, dFastShortLen)] = te0
+		e.table[hashLen(cv1, dFastShortTableBits, dFastShortLen)] = te1
 
 		cv = load6432(src, s)
 
@@ -640,8 +639,8 @@
 			}
 
 			// Store this, since we have it.
-			nextHashS := hash5(cv1>>8, dFastShortTableBits)
-			nextHashL := hash8(cv, dFastLongTableBits)
+			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
@@ -677,12 +676,12 @@
 		blk.literals = append(blk.literals, src[nextEmit:]...)
 		blk.extraLits = len(src) - int(nextEmit)
 	}
-	if debug {
+	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 < bufferReset {
+	if e.cur < e.bufferReset {
 		e.cur += int32(len(src))
 	}
 }
@@ -697,7 +696,7 @@
 	)
 
 	// Protect against e.cur wraparound.
-	for e.cur >= bufferReset {
+	for e.cur >= e.bufferReset-int32(len(e.hist)) {
 		if len(e.hist) == 0 {
 			for i := range e.table[:] {
 				e.table[i] = tableEntry{}
@@ -767,7 +766,7 @@
 		blk.literals = append(blk.literals, src[nextEmit:until]...)
 		s.litLen = uint32(until - nextEmit)
 	}
-	if debug {
+	if debugEncoder {
 		println("recent offsets:", blk.recentOffsets)
 	}
 
@@ -782,8 +781,8 @@
 				panic("offset0 was 0")
 			}
 
-			nextHashS := hash5(cv, dFastShortTableBits)
-			nextHashL := hash8(cv, dFastLongTableBits)
+			nextHashL := hashLen(cv, dFastLongTableBits, dFastLongLen)
+			nextHashS := hashLen(cv, dFastShortTableBits, dFastShortLen)
 			candidateL := e.longTable[nextHashL]
 			candidateS := e.table[nextHashS]
 
@@ -799,9 +798,9 @@
 				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)
+					length := 4 + e.matchlen(s+4+repOff, repIndex+4, src)
 
-					seq.matchLen = uint32(lenght - zstdMinMatch)
+					seq.matchLen = uint32(length - zstdMinMatch)
 
 					// We might be able to match backwards.
 					// Extend as long as we can.
@@ -827,11 +826,11 @@
 						println("repeat sequence", seq, "next s:", s)
 					}
 					blk.sequences = append(blk.sequences, seq)
-					s += lenght + repOff
+					s += length + repOff
 					nextEmit = s
 					if s >= sLimit {
-						if debug {
-							println("repeat ended", s, lenght)
+						if debugEncoder {
+							println("repeat ended", s, length)
 
 						}
 						break encodeLoop
@@ -868,7 +867,7 @@
 				// See if we can find a long match at s+1
 				const checkAt = 1
 				cv := load6432(src, s+checkAt)
-				nextHashL = hash8(cv, dFastLongTableBits)
+				nextHashL = hashLen(cv, dFastLongTableBits, dFastLongLen)
 				candidateL = e.longTable[nextHashL]
 				coffsetL = s - (candidateL.offset - e.cur) + checkAt
 
@@ -965,8 +964,8 @@
 		cv1 := load6432(src, index1)
 		te0 := tableEntry{offset: index0 + e.cur, val: uint32(cv0)}
 		te1 := tableEntry{offset: index1 + e.cur, val: uint32(cv1)}
-		longHash1 := hash8(cv0, dFastLongTableBits)
-		longHash2 := hash8(cv0, dFastLongTableBits)
+		longHash1 := hashLen(cv0, dFastLongTableBits, dFastLongLen)
+		longHash2 := hashLen(cv1, dFastLongTableBits, dFastLongLen)
 		e.longTable[longHash1] = te0
 		e.longTable[longHash2] = te1
 		e.markLongShardDirty(longHash1)
@@ -977,8 +976,8 @@
 		te1.offset++
 		te0.val = uint32(cv0)
 		te1.val = uint32(cv1)
-		hashVal1 := hash5(cv0, dFastShortTableBits)
-		hashVal2 := hash5(cv1, dFastShortTableBits)
+		hashVal1 := hashLen(cv0, dFastShortTableBits, dFastShortLen)
+		hashVal2 := hashLen(cv1, dFastShortTableBits, dFastShortLen)
 		e.table[hashVal1] = te0
 		e.markShardDirty(hashVal1)
 		e.table[hashVal2] = te1
@@ -999,8 +998,8 @@
 			}
 
 			// Store this, since we have it.
-			nextHashS := hash5(cv, dFastShortTableBits)
-			nextHashL := hash8(cv, dFastLongTableBits)
+			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
@@ -1039,7 +1038,7 @@
 	}
 	blk.recentOffsets[0] = uint32(offset1)
 	blk.recentOffsets[1] = uint32(offset2)
-	if debug {
+	if debugEncoder {
 		println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits)
 	}
 	// If we encoded more than 64K mark all dirty.
@@ -1071,21 +1070,21 @@
 		}
 		if len(d.content) >= 8 {
 			cv := load6432(d.content, 0)
-			e.dictLongTable[hash8(cv, dFastLongTableBits)] = tableEntry{
+			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[hash8(cv, dFastLongTableBits)] = tableEntry{
+				e.dictLongTable[hashLen(cv, dFastLongTableBits, dFastLongLen)] = tableEntry{
 					val:    uint32(cv),
 					offset: i,
 				}
 			}
 		}
 		e.lastDictID = d.id
-		e.allDirty = true
+		allDirty = true
 	}
 	// Reset table to initial state
 	e.cur = e.maxMatchOff
@@ -1100,7 +1099,8 @@
 	}
 
 	if allDirty || dirtyShardCnt > dLongTableShardCnt/2 {
-		copy(e.longTable[:], e.dictLongTable)
+		//copy(e.longTable[:], e.dictLongTable)
+		e.longTable = *(*[dFastLongTableSize]tableEntry)(e.dictLongTable)
 		for i := range e.longTableShardDirty {
 			e.longTableShardDirty[i] = false
 		}
@@ -1111,7 +1111,9 @@
 			continue
 		}
 
-		copy(e.longTable[i*dLongTableShardSize:(i+1)*dLongTableShardSize], e.dictLongTable[i*dLongTableShardSize:(i+1)*dLongTableShardSize])
+		// 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
 	}
 }
diff --git a/vendor/github.com/klauspost/compress/zstd/enc_fast.go b/vendor/github.com/klauspost/compress/zstd/enc_fast.go
index ba4a17e..f45a3da 100644
--- a/vendor/github.com/klauspost/compress/zstd/enc_fast.go
+++ b/vendor/github.com/klauspost/compress/zstd/enc_fast.go
@@ -6,17 +6,16 @@
 
 import (
 	"fmt"
-	"math"
-	"math/bits"
 )
 
 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
-	tableMask      = tableSize - 1                    // Mask for table indices. Redundant, but can eliminate bounds checks.
-	maxMatchLength = 131074
+	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 {
@@ -44,7 +43,7 @@
 	)
 
 	// Protect against e.cur wraparound.
-	for e.cur >= bufferReset {
+	for e.cur >= e.bufferReset-int32(len(e.hist)) {
 		if len(e.hist) == 0 {
 			for i := range e.table[:] {
 				e.table[i] = tableEntry{}
@@ -86,7 +85,7 @@
 	// TEMPLATE
 	const hashLog = tableBits
 	// seems global, but would be nice to tweak.
-	const kSearchStrength = 7
+	const kSearchStrength = 6
 
 	// nextEmit is where in src the next emitLiteral should start from.
 	nextEmit := s
@@ -103,7 +102,7 @@
 		blk.literals = append(blk.literals, src[nextEmit:until]...)
 		s.litLen = uint32(until - nextEmit)
 	}
-	if debug {
+	if debugEncoder {
 		println("recent offsets:", blk.recentOffsets)
 	}
 
@@ -122,8 +121,8 @@
 				panic("offset0 was 0")
 			}
 
-			nextHash := hash6(cv, hashLog)
-			nextHash2 := hash6(cv>>8, hashLog)
+			nextHash := hashLen(cv, hashLog, tableFastHashLen)
+			nextHash2 := hashLen(cv>>8, hashLog, tableFastHashLen)
 			candidate := e.table[nextHash]
 			candidate2 := e.table[nextHash2]
 			repIndex := s - offset1 + 2
@@ -134,21 +133,7 @@
 			if canRepeat && repIndex >= 0 && load3232(src, repIndex) == uint32(cv>>16) {
 				// Consider history as well.
 				var seq seq
-				var length int32
-				// length = 4 + e.matchlen(s+6, repIndex+4, src)
-				{
-					a := src[s+6:]
-					b := src[repIndex+4:]
-					endI := len(a) & (math.MaxInt32 - 7)
-					length = int32(endI) + 4
-					for i := 0; i < endI; i += 8 {
-						if diff := load64(a, i) ^ load64(b, i); diff != 0 {
-							length = int32(i+bits.TrailingZeros64(diff)>>3) + 4
-							break
-						}
-					}
-				}
-
+				length := 4 + e.matchlen(s+6, repIndex+4, src)
 				seq.matchLen = uint32(length - zstdMinMatch)
 
 				// We might be able to match backwards.
@@ -178,7 +163,7 @@
 				s += length + 2
 				nextEmit = s
 				if s >= sLimit {
-					if debug {
+					if debugEncoder {
 						println("repeat ended", s, length)
 
 					}
@@ -235,20 +220,7 @@
 		}
 
 		// Extend the 4-byte match as long as possible.
-		//l := e.matchlen(s+4, t+4, src) + 4
-		var l int32
-		{
-			a := src[s+4:]
-			b := src[t+4:]
-			endI := len(a) & (math.MaxInt32 - 7)
-			l = int32(endI) + 4
-			for i := 0; i < endI; i += 8 {
-				if diff := load64(a, i) ^ load64(b, i); diff != 0 {
-					l = int32(i+bits.TrailingZeros64(diff)>>3) + 4
-					break
-				}
-			}
-		}
+		l := e.matchlen(s+4, t+4, src) + 4
 
 		// Extend backwards
 		tMin := s - e.maxMatchOff
@@ -285,23 +257,10 @@
 		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)
-			var l int32
-			{
-				a := src[s+4:]
-				b := src[o2+4:]
-				endI := len(a) & (math.MaxInt32 - 7)
-				l = int32(endI) + 4
-				for i := 0; i < endI; i += 8 {
-					if diff := load64(a, i) ^ load64(b, i); diff != 0 {
-						l = int32(i+bits.TrailingZeros64(diff)>>3) + 4
-						break
-					}
-				}
-			}
+			l := 4 + e.matchlen(s+4, o2+4, src)
 
 			// Store this, since we have it.
-			nextHash := hash6(cv, hashLog)
+			nextHash := hashLen(cv, hashLog, tableFastHashLen)
 			e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)}
 			seq.matchLen = uint32(l) - zstdMinMatch
 			seq.litLen = 0
@@ -330,7 +289,7 @@
 	}
 	blk.recentOffsets[0] = uint32(offset1)
 	blk.recentOffsets[1] = uint32(offset2)
-	if debug {
+	if debugEncoder {
 		println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits)
 	}
 }
@@ -343,14 +302,14 @@
 		inputMargin            = 8
 		minNonLiteralBlockSize = 1 + 1 + inputMargin
 	)
-	if debug {
-		if len(src) > maxBlockSize {
+	if debugEncoder {
+		if len(src) > maxCompressedBlockSize {
 			panic("src too big")
 		}
 	}
 
 	// Protect against e.cur wraparound.
-	if e.cur >= bufferReset {
+	if e.cur >= e.bufferReset {
 		for i := range e.table[:] {
 			e.table[i] = tableEntry{}
 		}
@@ -374,7 +333,7 @@
 	// TEMPLATE
 	const hashLog = tableBits
 	// seems global, but would be nice to tweak.
-	const kSearchStrength = 8
+	const kSearchStrength = 6
 
 	// nextEmit is where in src the next emitLiteral should start from.
 	nextEmit := s
@@ -391,7 +350,7 @@
 		blk.literals = append(blk.literals, src[nextEmit:until]...)
 		s.litLen = uint32(until - nextEmit)
 	}
-	if debug {
+	if debugEncoder {
 		println("recent offsets:", blk.recentOffsets)
 	}
 
@@ -405,8 +364,8 @@
 		// By not using them for the first 3 matches
 
 		for {
-			nextHash := hash6(cv, hashLog)
-			nextHash2 := hash6(cv>>8, hashLog)
+			nextHash := hashLen(cv, hashLog, tableFastHashLen)
+			nextHash2 := hashLen(cv>>8, hashLog, tableFastHashLen)
 			candidate := e.table[nextHash]
 			candidate2 := e.table[nextHash2]
 			repIndex := s - offset1 + 2
@@ -417,21 +376,7 @@
 			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)
-				// length := 4 + int32(matchLen(src[s+6:], src[repIndex+4:]))
-				var length int32
-				{
-					a := src[s+6:]
-					b := src[repIndex+4:]
-					endI := len(a) & (math.MaxInt32 - 7)
-					length = int32(endI) + 4
-					for i := 0; i < endI; i += 8 {
-						if diff := load64(a, i) ^ load64(b, i); diff != 0 {
-							length = int32(i+bits.TrailingZeros64(diff)>>3) + 4
-							break
-						}
-					}
-				}
+				length := 4 + e.matchlen(s+6, repIndex+4, src)
 
 				seq.matchLen = uint32(length - zstdMinMatch)
 
@@ -462,7 +407,7 @@
 				s += length + 2
 				nextEmit = s
 				if s >= sLimit {
-					if debug {
+					if debugEncoder {
 						println("repeat ended", s, length)
 
 					}
@@ -521,21 +466,7 @@
 			panic(fmt.Sprintf("t (%d) < 0 ", t))
 		}
 		// Extend the 4-byte match as long as possible.
-		//l := e.matchlenNoHist(s+4, t+4, src) + 4
-		// l := int32(matchLen(src[s+4:], src[t+4:])) + 4
-		var l int32
-		{
-			a := src[s+4:]
-			b := src[t+4:]
-			endI := len(a) & (math.MaxInt32 - 7)
-			l = int32(endI) + 4
-			for i := 0; i < endI; i += 8 {
-				if diff := load64(a, i) ^ load64(b, i); diff != 0 {
-					l = int32(i+bits.TrailingZeros64(diff)>>3) + 4
-					break
-				}
-			}
-		}
+		l := e.matchlen(s+4, t+4, src) + 4
 
 		// Extend backwards
 		tMin := s - e.maxMatchOff
@@ -572,24 +503,10 @@
 		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.matchlenNoHist(s+4, o2+4, src)
-			// l := 4 + int32(matchLen(src[s+4:], src[o2+4:]))
-			var l int32
-			{
-				a := src[s+4:]
-				b := src[o2+4:]
-				endI := len(a) & (math.MaxInt32 - 7)
-				l = int32(endI) + 4
-				for i := 0; i < endI; i += 8 {
-					if diff := load64(a, i) ^ load64(b, i); diff != 0 {
-						l = int32(i+bits.TrailingZeros64(diff)>>3) + 4
-						break
-					}
-				}
-			}
+			l := 4 + e.matchlen(s+4, o2+4, src)
 
 			// Store this, since we have it.
-			nextHash := hash6(cv, hashLog)
+			nextHash := hashLen(cv, hashLog, tableFastHashLen)
 			e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)}
 			seq.matchLen = uint32(l) - zstdMinMatch
 			seq.litLen = 0
@@ -616,11 +533,11 @@
 		blk.literals = append(blk.literals, src[nextEmit:]...)
 		blk.extraLits = len(src) - int(nextEmit)
 	}
-	if debug {
+	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 < bufferReset {
+	if e.cur < e.bufferReset {
 		e.cur += int32(len(src))
 	}
 }
@@ -637,11 +554,9 @@
 		return
 	}
 	// Protect against e.cur wraparound.
-	for e.cur >= bufferReset {
+	for e.cur >= e.bufferReset-int32(len(e.hist)) {
 		if len(e.hist) == 0 {
-			for i := range e.table[:] {
-				e.table[i] = tableEntry{}
-			}
+			e.table = [tableSize]tableEntry{}
 			e.cur = e.maxMatchOff
 			break
 		}
@@ -696,7 +611,7 @@
 		blk.literals = append(blk.literals, src[nextEmit:until]...)
 		s.litLen = uint32(until - nextEmit)
 	}
-	if debug {
+	if debugEncoder {
 		println("recent offsets:", blk.recentOffsets)
 	}
 
@@ -715,8 +630,8 @@
 				panic("offset0 was 0")
 			}
 
-			nextHash := hash6(cv, hashLog)
-			nextHash2 := hash6(cv>>8, hashLog)
+			nextHash := hashLen(cv, hashLog, tableFastHashLen)
+			nextHash2 := hashLen(cv>>8, hashLog, tableFastHashLen)
 			candidate := e.table[nextHash]
 			candidate2 := e.table[nextHash2]
 			repIndex := s - offset1 + 2
@@ -729,20 +644,7 @@
 			if canRepeat && repIndex >= 0 && load3232(src, repIndex) == uint32(cv>>16) {
 				// Consider history as well.
 				var seq seq
-				var length int32
-				// length = 4 + e.matchlen(s+6, repIndex+4, src)
-				{
-					a := src[s+6:]
-					b := src[repIndex+4:]
-					endI := len(a) & (math.MaxInt32 - 7)
-					length = int32(endI) + 4
-					for i := 0; i < endI; i += 8 {
-						if diff := load64(a, i) ^ load64(b, i); diff != 0 {
-							length = int32(i+bits.TrailingZeros64(diff)>>3) + 4
-							break
-						}
-					}
-				}
+				length := 4 + e.matchlen(s+6, repIndex+4, src)
 
 				seq.matchLen = uint32(length - zstdMinMatch)
 
@@ -773,7 +675,7 @@
 				s += length + 2
 				nextEmit = s
 				if s >= sLimit {
-					if debug {
+					if debugEncoder {
 						println("repeat ended", s, length)
 
 					}
@@ -830,20 +732,7 @@
 		}
 
 		// Extend the 4-byte match as long as possible.
-		//l := e.matchlen(s+4, t+4, src) + 4
-		var l int32
-		{
-			a := src[s+4:]
-			b := src[t+4:]
-			endI := len(a) & (math.MaxInt32 - 7)
-			l = int32(endI) + 4
-			for i := 0; i < endI; i += 8 {
-				if diff := load64(a, i) ^ load64(b, i); diff != 0 {
-					l = int32(i+bits.TrailingZeros64(diff)>>3) + 4
-					break
-				}
-			}
-		}
+		l := e.matchlen(s+4, t+4, src) + 4
 
 		// Extend backwards
 		tMin := s - e.maxMatchOff
@@ -880,23 +769,10 @@
 		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)
-			var l int32
-			{
-				a := src[s+4:]
-				b := src[o2+4:]
-				endI := len(a) & (math.MaxInt32 - 7)
-				l = int32(endI) + 4
-				for i := 0; i < endI; i += 8 {
-					if diff := load64(a, i) ^ load64(b, i); diff != 0 {
-						l = int32(i+bits.TrailingZeros64(diff)>>3) + 4
-						break
-					}
-				}
-			}
+			l := 4 + e.matchlen(s+4, o2+4, src)
 
 			// Store this, since we have it.
-			nextHash := hash6(cv, hashLog)
+			nextHash := hashLen(cv, hashLog, tableFastHashLen)
 			e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)}
 			e.markShardDirty(nextHash)
 			seq.matchLen = uint32(l) - zstdMinMatch
@@ -926,7 +802,7 @@
 	}
 	blk.recentOffsets[0] = uint32(offset1)
 	blk.recentOffsets[1] = uint32(offset2)
-	if debug {
+	if debugEncoder {
 		println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits)
 	}
 }
@@ -953,13 +829,12 @@
 		}
 		if true {
 			end := e.maxMatchOff + int32(len(d.content)) - 8
-			for i := e.maxMatchOff; i < end; i += 3 {
+			for i := e.maxMatchOff; i < end; i += 2 {
 				const hashLog = tableBits
 
 				cv := load6432(d.content, i-e.maxMatchOff)
-				nextHash := hash6(cv, hashLog)      // 0 -> 5
-				nextHash1 := hash6(cv>>8, hashLog)  // 1 -> 6
-				nextHash2 := hash6(cv>>16, hashLog) // 2 -> 7
+				nextHash := hashLen(cv, hashLog, tableFastHashLen)     // 0 -> 6
+				nextHash1 := hashLen(cv>>8, hashLog, tableFastHashLen) // 1 -> 7
 				e.dictTable[nextHash] = tableEntry{
 					val:    uint32(cv),
 					offset: i,
@@ -968,10 +843,6 @@
 					val:    uint32(cv >> 8),
 					offset: i + 1,
 				}
-				e.dictTable[nextHash2] = tableEntry{
-					val:    uint32(cv >> 16),
-					offset: i + 2,
-				}
 			}
 		}
 		e.lastDictID = d.id
@@ -991,7 +862,8 @@
 	const shardCnt = tableShardCnt
 	const shardSize = tableShardSize
 	if e.allDirty || dirtyShardCnt > shardCnt*4/6 {
-		copy(e.table[:], e.dictTable)
+		//copy(e.table[:], e.dictTable)
+		e.table = *(*[tableSize]tableEntry)(e.dictTable)
 		for i := range e.tableShardDirty {
 			e.tableShardDirty[i] = false
 		}
@@ -1003,7 +875,8 @@
 			continue
 		}
 
-		copy(e.table[i*shardSize:(i+1)*shardSize], e.dictTable[i*shardSize:(i+1)*shardSize])
+		//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
diff --git a/vendor/github.com/klauspost/compress/zstd/encoder.go b/vendor/github.com/klauspost/compress/zstd/encoder.go
index 4871dd0..8f8223c 100644
--- a/vendor/github.com/klauspost/compress/zstd/encoder.go
+++ b/vendor/github.com/klauspost/compress/zstd/encoder.go
@@ -6,8 +6,10 @@
 
 import (
 	"crypto/rand"
+	"errors"
 	"fmt"
 	"io"
+	"math"
 	rdebug "runtime/debug"
 	"sync"
 
@@ -33,7 +35,7 @@
 	Block() *blockEnc
 	CRC() *xxhash.Digest
 	AppendCRC([]byte) []byte
-	WindowSize(size int) int32
+	WindowSize(size int64) int32
 	UseBlock(*blockEnc)
 	Reset(d *dict, singleBlock bool)
 }
@@ -48,6 +50,8 @@
 	err              error
 	writeErr         error
 	nWritten         int64
+	nInput           int64
+	frameContentSize int64
 	headerWritten    bool
 	eofWritten       bool
 	fullFrameWritten bool
@@ -96,23 +100,25 @@
 	if cap(s.filling) == 0 {
 		s.filling = make([]byte, 0, e.o.blockSize)
 	}
-	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)
+	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()
 	}
-	if s.writing == nil {
-		s.writing = &blockEnc{lowMem: e.o.lowMem}
-		s.writing.init()
-	}
-	s.writing.initNewEncode()
 	s.filling = s.filling[:0]
-	s.current = s.current[:0]
-	s.previous = s.previous[:0]
 	s.encoder.Reset(e.o.dict, false)
 	s.headerWritten = false
 	s.eofWritten = false
@@ -120,7 +126,21 @@
 	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.
@@ -130,6 +150,9 @@
 // and write CRC if requested.
 func (e *Encoder) Write(p []byte) (n int, err error) {
 	s := &e.state
+	if s.eofWritten {
+		return 0, ErrEncoderClosed
+	}
 	for len(p) > 0 {
 		if len(p)+len(s.filling) < e.o.blockSize {
 			if e.o.crc {
@@ -183,13 +206,14 @@
 			return nil
 		}
 		if final && len(s.filling) > 0 {
-			s.current = e.EncodeAll(s.filling, s.current[:0])
+			s.current = e.encodeAll(s.encoder, s.filling, s.current[:0])
 			var n2 int
 			n2, s.err = s.w.Write(s.current)
 			if s.err != nil {
 				return s.err
 			}
 			s.nWritten += int64(n2)
+			s.nInput += int64(len(s.filling))
 			s.current = s.current[:0]
 			s.filling = s.filling[:0]
 			s.headerWritten = true
@@ -200,17 +224,14 @@
 
 		var tmp [maxHeaderSize]byte
 		fh := frameHeader{
-			ContentSize:   0,
-			WindowSize:    uint32(s.encoder.WindowSize(0)),
+			ContentSize:   uint64(s.frameContentSize),
+			WindowSize:    uint32(s.encoder.WindowSize(s.frameContentSize)),
 			SingleSegment: false,
 			Checksum:      e.o.crc,
 			DictID:        e.o.dict.ID(),
 		}
 
-		dst, err := fh.appendTo(tmp[:0])
-		if err != nil {
-			return err
-		}
+		dst := fh.appendTo(tmp[:0])
 		s.headerWritten = true
 		s.wWg.Wait()
 		var n2 int
@@ -241,11 +262,41 @@
 		return s.err
 	}
 
+	// SYNC:
+	if e.o.concurrent == 1 {
+		src := s.filling
+		s.nInput += int64(len(s.filling))
+		if debugEncoder {
+			println("Adding sync block,", len(src), "bytes, final:", final)
+		}
+		enc := s.encoder
+		blk := enc.Block()
+		blk.reset(nil)
+		enc.Encode(blk, src)
+		blk.last = final
+		if final {
+			s.eofWritten = true
+		}
+
+		s.err = blk.encode(src, e.o.noEntropy, !e.o.allLitEntropy)
+		if s.err != nil {
+			return s.err
+		}
+		_, s.err = s.w.Write(blk.output)
+		s.nWritten += int64(len(blk.output))
+		s.filling = s.filling[:0]
+		return s.err
+	}
+
 	// Move blocks forward.
 	s.filling, s.current, s.previous = s.previous[:0], s.filling, s.current
+	s.nInput += int64(len(s.current))
 	s.wg.Add(1)
+	if final {
+		s.eofWritten = true
+	}
 	go func(src []byte) {
-		if debug {
+		if debugEncoder {
 			println("Adding block,", len(src), "bytes, final:", final)
 		}
 		defer func() {
@@ -259,9 +310,6 @@
 		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 {
@@ -282,22 +330,8 @@
 				}
 				s.wWg.Done()
 			}()
-			err := errIncompressible
-			// If we got the exact same number of literals as input,
-			// assume the literals cannot be compressed.
-			if len(src) != len(blk.literals) || len(src) != e.o.blockSize {
-				err = blk.encode(src, e.o.noEntropy, !e.o.allLitEntropy)
-			}
-			switch err {
-			case errIncompressible:
-				if debug {
-					println("Storing incompressible block as raw")
-				}
-				blk.encodeRaw(src)
-				// In fast mode, we do not transfer offsets, so we don't have to deal with changing the.
-			case nil:
-			default:
-				s.writeErr = err
+			s.writeErr = blk.encode(src, e.o.noEntropy, !e.o.allLitEntropy)
+			if s.writeErr != nil {
 				return
 			}
 			_, s.writeErr = s.w.Write(blk.output)
@@ -313,7 +347,7 @@
 //
 // The Copy function uses ReaderFrom if available.
 func (e *Encoder) ReadFrom(r io.Reader) (n int64, err error) {
-	if debug {
+	if debugEncoder {
 		println("Using ReadFrom")
 	}
 
@@ -336,20 +370,20 @@
 		switch err {
 		case io.EOF:
 			e.state.filling = e.state.filling[:len(e.state.filling)-len(src)]
-			if debug {
+			if debugEncoder {
 				println("ReadFrom: got EOF final block:", len(e.state.filling))
 			}
 			return n, nil
 		case nil:
 		default:
-			if debug {
+			if debugEncoder {
 				println("ReadFrom: got error:", err)
 			}
 			e.state.err = err
 			return n, err
 		}
 		if len(src) > 0 {
-			if debug {
+			if debugEncoder {
 				println("ReadFrom: got space left in source:", len(src))
 			}
 			continue
@@ -371,12 +405,20 @@
 	if len(s.filling) > 0 {
 		err := e.nextBlock(false)
 		if err != nil {
+			// Ignore Flush after Close.
+			if errors.Is(s.err, ErrEncoderClosed) {
+				return nil
+			}
 			return err
 		}
 	}
 	s.wg.Wait()
 	s.wWg.Wait()
 	if s.err != nil {
+		// Ignore Flush after Close.
+		if errors.Is(s.err, ErrEncoderClosed) {
+			return nil
+		}
 		return s.err
 	}
 	return s.writeErr
@@ -392,8 +434,16 @@
 	}
 	err := e.nextBlock(true)
 	if err != nil {
+		if errors.Is(s.err, ErrEncoderClosed) {
+			return nil
+		}
 		return err
 	}
+	if s.frameContentSize > 0 {
+		if s.nInput != s.frameContentSize {
+			return fmt.Errorf("frame content size %d given, but %d bytes was written", s.frameContentSize, s.nInput)
+		}
+	}
 	if e.state.fullFrameWritten {
 		return s.err
 	}
@@ -424,6 +474,11 @@
 		}
 		_, s.err = s.w.Write(frame)
 	}
+	if s.err == nil {
+		s.err = ErrEncoderClosed
+		return nil
+	}
+
 	return s.err
 }
 
@@ -434,6 +489,15 @@
 // Data compressed with EncodeAll can be decoded with the Decoder,
 // using either a stream or DecodeAll.
 func (e *Encoder) EncodeAll(src, dst []byte) []byte {
+	e.init.Do(e.initialize)
+	enc := <-e.encoders
+	defer func() {
+		e.encoders <- enc
+	}()
+	return e.encodeAll(enc, src, dst)
+}
+
+func (e *Encoder) encodeAll(enc encoder, src, dst []byte) []byte {
 	if len(src) == 0 {
 		if e.o.fullZero {
 			// Add frame header.
@@ -445,7 +509,7 @@
 				Checksum: false,
 				DictID:   0,
 			}
-			dst, _ = fh.appendTo(dst)
+			dst = fh.appendTo(dst)
 
 			// Write raw block as last one only.
 			var blk blockHeader
@@ -456,21 +520,15 @@
 		}
 		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 1MB.
-	single := len(src) < 1<<20 && len(src) > MinWindowSize
+
+	// 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(len(src))),
+		WindowSize:    uint32(enc.WindowSize(int64(len(src)))),
 		SingleSegment: single,
 		Checksum:      e.o.crc,
 		DictID:        e.o.dict.ID(),
@@ -480,13 +538,10 @@
 	if len(dst) == 0 && cap(dst) == 0 && len(src) < 1<<20 && !e.o.lowMem {
 		dst = make([]byte, 0, len(src))
 	}
-	dst, err := fh.appendTo(dst)
-	if err != nil {
-		panic(err)
-	}
+	dst = fh.appendTo(dst)
 
 	// If we can do everything in one block, prefer that.
-	if len(src) <= maxCompressedBlockSize {
+	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 {
@@ -502,25 +557,15 @@
 
 		// If we got the exact same number of literals as input,
 		// assume the literals cannot be compressed.
-		err := errIncompressible
 		oldout := blk.output
-		if len(blk.literals) != len(src) || len(src) != e.o.blockSize {
-			// Output directly to dst
-			blk.output = dst
-			err = blk.encode(src, e.o.noEntropy, !e.o.allLitEntropy)
-		}
+		// Output directly to dst
+		blk.output = dst
 
-		switch err {
-		case errIncompressible:
-			if debug {
-				println("Storing incompressible block as raw")
-			}
-			dst = blk.encodeRawTo(dst, src)
-		case nil:
-			dst = blk.output
-		default:
+		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)
@@ -539,25 +584,11 @@
 			if len(src) == 0 {
 				blk.last = true
 			}
-			err := errIncompressible
-			// If we got the exact same number of literals as input,
-			// assume the literals cannot be compressed.
-			if len(blk.literals) != len(todo) || len(todo) != e.o.blockSize {
-				err = blk.encode(todo, e.o.noEntropy, !e.o.allLitEntropy)
-			}
-
-			switch err {
-			case errIncompressible:
-				if debug {
-					println("Storing incompressible block as raw")
-				}
-				dst = blk.encodeRawTo(dst, todo)
-				blk.popOffsets()
-			case nil:
-				dst = append(dst, blk.output...)
-			default:
+			err := blk.encode(todo, e.o.noEntropy, !e.o.allLitEntropy)
+			if err != nil {
 				panic(err)
 			}
+			dst = append(dst, blk.output...)
 			blk.reset(nil)
 		}
 	}
@@ -567,6 +598,7 @@
 	// 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)
@@ -574,3 +606,37 @@
 	}
 	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
index 16d4ab6..20671dc 100644
--- a/vendor/github.com/klauspost/compress/zstd/encoder_options.go
+++ b/vendor/github.com/klauspost/compress/zstd/encoder_options.go
@@ -3,6 +3,8 @@
 import (
 	"errors"
 	"fmt"
+	"math"
+	"math/bits"
 	"runtime"
 	"strings"
 )
@@ -24,6 +26,7 @@
 	allLitEntropy   bool
 	customWindow    bool
 	customALEntropy bool
+	customBlockSize bool
 	lowMem          bool
 	dict            *dict
 }
@@ -33,10 +36,10 @@
 		concurrent:    runtime.GOMAXPROCS(0),
 		crc:           true,
 		single:        nil,
-		blockSize:     1 << 16,
+		blockSize:     maxCompressedBlockSize,
 		windowSize:    8 << 20,
 		level:         SpeedDefault,
-		allLitEntropy: true,
+		allLitEntropy: false,
 		lowMem:        false,
 	}
 }
@@ -46,22 +49,22 @@
 	switch o.level {
 	case SpeedFastest:
 		if o.dict != nil {
-			return &fastEncoderDict{fastEncoder: fastEncoder{fastBase: fastBase{maxMatchOff: int32(o.windowSize), lowMem: o.lowMem}}}
+			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), 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), lowMem: o.lowMem}}}}
+			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), 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), lowMem: o.lowMem}}}
+			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), 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), lowMem: o.lowMem}}
+		return &bestFastEncoder{fastBase: fastBase{maxMatchOff: int32(o.windowSize), bufferReset: math.MaxInt32 - int32(o.windowSize*2), lowMem: o.lowMem}}
 	}
 	panic("unknown compression level")
 }
@@ -75,6 +78,7 @@
 // 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 {
@@ -90,7 +94,7 @@
 // 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.
+// The default value is determined by the compression level and max 8MB.
 func WithWindowSize(n int) EOption {
 	return func(o *encoderOptions) error {
 		switch {
@@ -106,6 +110,7 @@
 		o.customWindow = true
 		if o.blockSize > o.windowSize {
 			o.blockSize = o.windowSize
+			o.customBlockSize = true
 		}
 		return nil
 	}
@@ -124,7 +129,7 @@
 		}
 		// No need to waste our time.
 		if n == 1 {
-			o.pad = 0
+			n = 0
 		}
 		if n > 1<<30 {
 			return fmt.Errorf("padding must less than 1GB (1<<30 bytes) ")
@@ -188,10 +193,9 @@
 		return SpeedDefault
 	case level >= 6 && level < 10:
 		return SpeedBetterCompression
-	case level >= 10:
-		return SpeedBetterCompression
+	default:
+		return SpeedBestCompression
 	}
-	return SpeedDefault
 }
 
 // String provides a string representation of the compression level.
@@ -222,16 +226,19 @@
 			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 = 16 << 20
+				o.windowSize = 8 << 20
 			case SpeedBestCompression:
-				o.windowSize = 32 << 20
+				o.windowSize = 8 << 20
 			}
 		}
 		if !o.customALEntropy {
-			o.allLitEntropy = l > SpeedFastest
+			o.allLitEntropy = l > SpeedDefault
 		}
 
 		return nil
@@ -278,7 +285,7 @@
 // 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.
+// 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 {
@@ -299,7 +306,13 @@
 }
 
 // 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)
@@ -310,3 +323,17 @@
 		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
index 693c5f0..e47af66 100644
--- a/vendor/github.com/klauspost/compress/zstd/framedec.go
+++ b/vendor/github.com/klauspost/compress/zstd/framedec.go
@@ -5,30 +5,20 @@
 package zstd
 
 import (
-	"bytes"
+	"encoding/binary"
 	"encoding/hex"
 	"errors"
-	"hash"
 	"io"
-	"sync"
 
 	"github.com/klauspost/compress/zstd/internal/xxhash"
 )
 
 type frameDec struct {
-	o      decoderOptions
-	crc    hash.Hash64
-	offset int64
+	o   decoderOptions
+	crc *xxhash.Digest
 
 	WindowSize uint64
 
-	// maxWindowSize is the maximum windows size to support.
-	// should never be bigger than max-int.
-	maxWindowSize uint64
-
-	// In order queue of blocks being decoded.
-	decoding chan *blockDec
-
 	// Frame history passed between blocks
 	history history
 
@@ -38,35 +28,32 @@
 	bBuf byteBuf
 
 	FrameContentSize uint64
-	frameDone        sync.WaitGroup
 
-	DictionaryID  *uint32
+	DictionaryID  uint32
 	HasCheckSum   bool
 	SingleSegment bool
-
-	// asyncRunning indicates whether the async routine processes input on 'decoding'.
-	asyncRunningMu sync.Mutex
-	asyncRunning   bool
 }
 
 const (
-	// The minimum Window_Size is 1 KB.
+	// 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
 )
 
-var (
-	frameMagic          = []byte{0x28, 0xb5, 0x2f, 0xfd}
-	skippableFrameMagic = []byte{0x2a, 0x4d, 0x18}
+const (
+	frameMagic          = "\x28\xb5\x2f\xfd"
+	skippableFrameMagic = "\x2a\x4d\x18"
 )
 
 func newFrameDec(o decoderOptions) *frameDec {
-	d := frameDec{
-		o:             o,
-		maxWindowSize: MaxWindowSize,
+	if o.maxWindowSize > o.maxDecodedSize {
+		o.maxWindowSize = o.maxDecodedSize
 	}
-	if d.maxWindowSize > o.maxDecodedSize {
-		d.maxWindowSize = o.maxDecodedSize
+	d := frameDec{
+		o: o,
 	}
 	return &d
 }
@@ -78,44 +65,68 @@
 func (d *frameDec) reset(br byteBuffer) error {
 	d.HasCheckSum = false
 	d.WindowSize = 0
-	var b []byte
+	var signature [4]byte
 	for {
-		b = br.readSmall(4)
-		if b == nil {
+		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
 		}
-		if !bytes.Equal(b[1:4], skippableFrameMagic) || b[0]&0xf0 != 0x50 {
-			if debug {
-				println("Not skippable", hex.EncodeToString(b), hex.EncodeToString(skippableFrameMagic))
+		// 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 = br.readSmall(4)
-		if b == nil {
-			println("Reading Frame Size EOF")
-			return io.ErrUnexpectedEOF
+		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(int(n))
+		err = br.skipN(int64(n))
 		if err != nil {
-			if debug {
+			if debugDecoder {
 				println("Reading discarded frame", err)
 			}
 			return err
 		}
 	}
-	if !bytes.Equal(b, frameMagic) {
-		println("Got magic numbers: ", b, "want:", frameMagic)
+	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 {
-		println("Reading Frame_Header_Descriptor", err)
+		if debugDecoder {
+			println("Reading Frame_Header_Descriptor", err)
+		}
 		return err
 	}
 	d.SingleSegment = fhd&(1<<5) != 0
@@ -130,10 +141,14 @@
 	if !d.SingleSegment {
 		wd, err := br.readByte()
 		if err != nil {
-			println("Reading Window_Descriptor", err)
+			if debugDecoder {
+				println("Reading Window_Descriptor", err)
+			}
 			return err
 		}
-		printf("raw: %x, mantissa: %d, exponent: %d\n", wd, wd&7, wd>>3)
+		if debugDecoder {
+			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)
@@ -142,20 +157,19 @@
 
 	// Read Dictionary_ID
 	// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#dictionary_id
-	d.DictionaryID = nil
+	d.DictionaryID = 0
 	if size := fhd & 3; size != 0 {
 		if size == 3 {
 			size = 4
 		}
-		b = br.readSmall(int(size))
-		if b == nil {
-			if debug {
-				println("Reading Dictionary_ID", io.ErrUnexpectedEOF)
-			}
-			return io.ErrUnexpectedEOF
+
+		b, err := br.readSmall(int(size))
+		if err != nil {
+			println("Reading Dictionary_ID", err)
+			return err
 		}
 		var id uint32
-		switch size {
+		switch len(b) {
 		case 1:
 			id = uint32(b[0])
 		case 2:
@@ -163,14 +177,10 @@
 		case 4:
 			id = uint32(b[0]) | (uint32(b[1]) << 8) | (uint32(b[2]) << 16) | (uint32(b[3]) << 24)
 		}
-		if debug {
+		if debugDecoder {
 			println("Dict size", size, "ID:", id)
 		}
-		if id > 0 {
-			// ID 0 means "sorry, no dictionary anyway".
-			// https://github.com/facebook/zstd/blob/dev/doc/zstd_compression_format.md#dictionary-format
-			d.DictionaryID = &id
-		}
+		d.DictionaryID = id
 	}
 
 	// Read Frame_Content_Size
@@ -185,14 +195,14 @@
 	default:
 		fcsSize = 1 << v
 	}
-	d.FrameContentSize = 0
+	d.FrameContentSize = fcsUnknown
 	if fcsSize > 0 {
-		b := br.readSmall(fcsSize)
-		if b == nil {
-			println("Reading Frame content", io.ErrUnexpectedEOF)
-			return io.ErrUnexpectedEOF
+		b, err := br.readSmall(fcsSize)
+		if err != nil {
+			println("Reading Frame content", err)
+			return err
 		}
-		switch fcsSize {
+		switch len(b) {
 		case 1:
 			d.FrameContentSize = uint64(b[0])
 		case 2:
@@ -205,10 +215,11 @@
 			d2 := uint32(b[4]) | (uint32(b[5]) << 8) | (uint32(b[6]) << 16) | (uint32(b[7]) << 24)
 			d.FrameContentSize = uint64(d1) | (uint64(d2) << 32)
 		}
-		if debug {
-			println("field size bits:", v, "fcsSize:", fcsSize, "FrameContentSize:", d.FrameContentSize, hex.EncodeToString(b[:fcsSize]), "singleseg:", d.SingleSegment, "window:", d.WindowSize)
+		if debugDecoder {
+			println("Read FCS:", d.FrameContentSize)
 		}
 	}
+
 	// Move this to shared.
 	d.HasCheckSum = fhd&(1<<2) != 0
 	if d.HasCheckSum {
@@ -218,29 +229,52 @@
 		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
+		}
 	}
 
-	if d.WindowSize > d.maxWindowSize {
-		printf("window size %d > max %d\n", d.WindowSize, d.maxWindowSize)
-		return ErrWindowSizeExceeded
-	}
 	// The minimum Window_Size is 1 KB.
 	if d.WindowSize < MinWindowSize {
-		println("got window size: ", d.WindowSize)
+		if debugDecoder {
+			println("got window size: ", d.WindowSize)
+		}
 		return ErrWindowSizeTooSmall
 	}
 	d.history.windowSize = int(d.WindowSize)
-	if d.o.lowMem && d.history.windowSize < maxBlockSize {
-		d.history.maxSize = d.history.windowSize * 2
+	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 {
-		d.history.maxSize = d.history.windowSize + maxBlockSize
+		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
@@ -248,244 +282,131 @@
 
 // next will start decoding the next block from stream.
 func (d *frameDec) next(block *blockDec) error {
-	if debug {
-		printf("decoding new block %p:%p", block, block.data)
+	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.
-		d.sendErr(block, err)
+		block.sendErr(err)
 		return err
 	}
-	block.input <- struct{}{}
-	if debug {
-		println("next block:", block)
-	}
-	d.asyncRunningMu.Lock()
-	defer d.asyncRunningMu.Unlock()
-	if !d.asyncRunning {
-		return nil
-	}
-	if block.Last {
-		// We indicate the frame is done by sending io.EOF
-		d.decoding <- block
-		return io.EOF
-	}
-	d.decoding <- block
 	return nil
 }
 
-// sendEOF will queue an error block on the frame.
-// This will cause the frame decoder to return when it encounters the block.
-// Returns true if the decoder was added.
-func (d *frameDec) sendErr(block *blockDec, err error) bool {
-	d.asyncRunningMu.Lock()
-	defer d.asyncRunningMu.Unlock()
-	if !d.asyncRunning {
-		return false
-	}
-
-	println("sending error", err.Error())
-	block.sendErr(err)
-	d.decoding <- block
-	return true
-}
-
-// checkCRC will check the checksum if the frame has one.
+// checkCRC will check the checksum, assuming the frame has one.
 // Will return ErrCRCMismatch if crc check failed, otherwise nil.
 func (d *frameDec) checkCRC() error {
-	if !d.HasCheckSum {
-		return nil
-	}
-	var tmp [4]byte
-	got := d.crc.Sum64()
-	// Flip to match file order.
-	tmp[0] = byte(got >> 0)
-	tmp[1] = byte(got >> 8)
-	tmp[2] = byte(got >> 16)
-	tmp[3] = byte(got >> 24)
-
 	// We can overwrite upper tmp now
-	want := d.rawInput.readSmall(4)
-	if want == nil {
-		println("CRC missing?")
-		return io.ErrUnexpectedEOF
+	buf, err := d.rawInput.readSmall(4)
+	if err != nil {
+		println("CRC missing?", err)
+		return err
 	}
 
-	if !bytes.Equal(tmp[:], want) {
-		if debug {
-			println("CRC Check Failed:", tmp[:], "!=", want)
+	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 debug {
-		println("CRC ok", tmp[:])
+	if debugDecoder {
+		printf("CRC ok %08x\n", got)
 	}
 	return nil
 }
 
-func (d *frameDec) initAsync() {
-	if !d.o.lowMem && !d.SingleSegment {
-		// set max extra size history to 10MB.
-		d.history.maxSize = d.history.windowSize + maxBlockSize*5
+// 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)
 	}
-	// re-alloc if more than one extra block size.
-	if d.o.lowMem && cap(d.history.b) > d.history.maxSize+maxBlockSize {
-		d.history.b = make([]byte, 0, d.history.maxSize)
-	}
-	if cap(d.history.b) < d.history.maxSize {
-		d.history.b = make([]byte, 0, d.history.maxSize)
-	}
-	if cap(d.decoding) < d.o.concurrent {
-		d.decoding = make(chan *blockDec, d.o.concurrent)
-	}
-	if debug {
-		h := d.history
-		printf("history init. len: %d, cap: %d", len(h.b), cap(h.b))
-	}
-	d.asyncRunningMu.Lock()
-	d.asyncRunning = true
-	d.asyncRunningMu.Unlock()
+	return err
 }
 
-// startDecoder will start decoding blocks and write them to the writer.
-// The decoder will stop as soon as an error occurs or at end of frame.
-// When the frame has finished decoding the *bufio.Reader
-// containing the remaining input will be sent on frameDec.frameDone.
-func (d *frameDec) startDecoder(output chan decodeOutput) {
-	written := int64(0)
-
-	defer func() {
-		d.asyncRunningMu.Lock()
-		d.asyncRunning = false
-		d.asyncRunningMu.Unlock()
-
-		// Drain the currently decoding.
-		d.history.error = true
-	flushdone:
-		for {
-			select {
-			case b := <-d.decoding:
-				b.history <- &d.history
-				output <- <-b.result
-			default:
-				break flushdone
-			}
-		}
-		println("frame decoder done, signalling done")
-		d.frameDone.Done()
-	}()
-	// Get decoder for first block.
-	block := <-d.decoding
-	block.history <- &d.history
-	for {
-		var next *blockDec
-		// Get result
-		r := <-block.result
-		if r.err != nil {
-			println("Result contained error", r.err)
-			output <- r
-			return
-		}
-		if debug {
-			println("got result, from ", d.offset, "to", d.offset+int64(len(r.b)))
-			d.offset += int64(len(r.b))
-		}
-		if !block.Last {
-			// Send history to next block
-			select {
-			case next = <-d.decoding:
-				if debug {
-					println("Sending ", len(d.history.b), "bytes as history")
-				}
-				next.history <- &d.history
-			default:
-				// Wait until we have sent the block, so
-				// other decoders can potentially get the decoder.
-				next = nil
-			}
-		}
-
-		// Add checksum, async to decoding.
-		if d.HasCheckSum {
-			n, err := d.crc.Write(r.b)
-			if err != nil {
-				r.err = err
-				if n != len(r.b) {
-					r.err = io.ErrShortWrite
-				}
-				output <- r
-				return
-			}
-		}
-		written += int64(len(r.b))
-		if d.SingleSegment && uint64(written) > d.FrameContentSize {
-			println("runDecoder: single segment and", uint64(written), ">", d.FrameContentSize)
-			r.err = ErrFrameSizeExceeded
-			output <- r
-			return
-		}
-		if block.Last {
-			r.err = d.checkCRC()
-			output <- r
-			return
-		}
-		output <- r
-		if next == nil {
-			// There was no decoder available, we wait for one now that we have sent to the writer.
-			if debug {
-				println("Sending ", len(d.history.b), " bytes as history")
-			}
-			next = <-d.decoding
-			next.history <- &d.history
-		}
-		block = next
-	}
-}
-
-// runDecoder will create a sync decoder that will decode a block of data.
+// 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 debug {
+		if debugDecoder {
 			println("next block:", dec)
 		}
 		err = dec.decodeBuf(&d.history)
-		if err != nil || dec.Last {
+		if err != nil {
 			break
 		}
-		if uint64(len(d.history.b)) > d.o.maxDecodedSize {
+		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.SingleSegment && uint64(len(d.history.b)) > d.o.maxDecodedSize {
-			println("runDecoder: single segment and", uint64(len(d.history.b)), ">", d.o.maxDecodedSize)
+		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.HasCheckSum {
-			var n int
-			n, err = d.crc.Write(dst[crcStart:])
-			if err == nil {
-				if n != len(dst)-crcStart {
-					err = io.ErrShortWrite
-				} else {
-					err = d.checkCRC()
-				}
+		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()
 			}
 		}
 	}
diff --git a/vendor/github.com/klauspost/compress/zstd/frameenc.go b/vendor/github.com/klauspost/compress/zstd/frameenc.go
index 4ef7f5a..667ca06 100644
--- a/vendor/github.com/klauspost/compress/zstd/frameenc.go
+++ b/vendor/github.com/klauspost/compress/zstd/frameenc.go
@@ -22,7 +22,7 @@
 
 const maxHeaderSize = 14
 
-func (f frameHeader) appendTo(dst []byte) ([]byte, error) {
+func (f frameHeader) appendTo(dst []byte) []byte {
 	dst = append(dst, frameMagic...)
 	var fhd uint8
 	if f.Checksum {
@@ -76,7 +76,7 @@
 		if f.SingleSegment {
 			dst = append(dst, uint8(f.ContentSize))
 		}
-		// Unless SingleSegment is set, framessizes < 256 are nto stored.
+		// Unless SingleSegment is set, framessizes < 256 are not stored.
 	case 1:
 		f.ContentSize -= 256
 		dst = append(dst, uint8(f.ContentSize), uint8(f.ContentSize>>8))
@@ -88,7 +88,7 @@
 	default:
 		panic("invalid fcs")
 	}
-	return dst, nil
+	return dst
 }
 
 const skippableFrameHeader = 4 + 4
diff --git a/vendor/github.com/klauspost/compress/zstd/fse_decoder.go b/vendor/github.com/klauspost/compress/zstd/fse_decoder.go
index e6d3d49..2f8860a 100644
--- a/vendor/github.com/klauspost/compress/zstd/fse_decoder.go
+++ b/vendor/github.com/klauspost/compress/zstd/fse_decoder.go
@@ -5,8 +5,10 @@
 package zstd
 
 import (
+	"encoding/binary"
 	"errors"
 	"fmt"
+	"io"
 )
 
 const (
@@ -178,10 +180,32 @@
 		return fmt.Errorf("corruption detected (total %d != %d)", gotTotal, 1<<s.actualTableLog)
 	}
 	b.advance((bitCount + 7) >> 3)
-	// println(s.norm[:s.symbolLen], s.symbolLen)
 	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.
@@ -204,18 +228,10 @@
 	return uint16(d >> 16)
 }
 
-func (d decSymbol) baseline() uint32 {
-	return uint32(d >> 32)
-}
-
 func (d decSymbol) baselineInt() int {
 	return int(d >> 32)
 }
 
-func (d *decSymbol) set(nbits, addBits uint8, newState uint16, baseline uint32) {
-	*d = decSymbol(nbits) | (decSymbol(addBits) << 8) | (decSymbol(newState) << 16) | (decSymbol(baseline) << 32)
-}
-
 func (d *decSymbol) setNBits(nBits uint8) {
 	const mask = 0xffffffffffffff00
 	*d = (*d & mask) | decSymbol(nBits)
@@ -231,11 +247,6 @@
 	*d = (*d & mask) | decSymbol(state)<<16
 }
 
-func (d *decSymbol) setBaseline(baseline uint32) {
-	const mask = 0xffffffff
-	*d = (*d & mask) | decSymbol(baseline)<<32
-}
-
 func (d *decSymbol) setExt(addBits uint8, baseline uint32) {
 	const mask = 0xffff00ff
 	*d = (*d & mask) | (decSymbol(addBits) << 8) | (decSymbol(baseline) << 32)
@@ -257,68 +268,6 @@
 	s.dt[0] = symbol
 }
 
-// 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--
-				symbolNext[i] = 1
-			} else {
-				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))
-				position = (position + step) & tableMask
-				for position > highThreshold {
-					// lowprob area
-					position = (position + step) & tableMask
-				}
-			}
-		}
-		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
-}
-
 // 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.
@@ -352,34 +301,7 @@
 	s.state = dt[br.getBits(tableLog)]
 }
 
-// next returns the current symbol and sets the next state.
-// At least tablelog bits must be available in the bit reader.
-func (s *fseState) next(br *bitReader) {
-	lowBits := uint16(br.getBits(s.state.nbBits()))
-	s.state = s.dt[s.state.newState()+lowBits]
-}
-
-// finished returns true if all bits have been read from the bitstream
-// and the next state would require reading bits from the input.
-func (s *fseState) finished(br *bitReader) bool {
-	return br.finished() && s.state.nbBits() > 0
-}
-
-// final returns the current state symbol without decoding the next.
-func (s *fseState) final() (int, uint8) {
-	return s.state.baselineInt(), s.state.addBits()
-}
-
 // final returns the current state symbol without decoding the next.
 func (s decSymbol) final() (int, uint8) {
 	return s.baselineInt(), s.addBits()
 }
-
-// nextFast returns the next symbol and sets the next state.
-// This can only be used if no symbols are 0 bits.
-// At least tablelog bits must be available in the bit reader.
-func (s *fseState) nextFast(br *bitReader) (uint32, uint8) {
-	lowBits := uint16(br.getBitsFast(s.state.nbBits()))
-	s.state = s.dt[s.state.newState()+lowBits]
-	return s.state.baseline(), s.state.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 0000000..d04a829
--- /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 0000000..bcde398
--- /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 0000000..8adfebb
--- /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
index c74681b..ab26326 100644
--- a/vendor/github.com/klauspost/compress/zstd/fse_encoder.go
+++ b/vendor/github.com/klauspost/compress/zstd/fse_encoder.go
@@ -62,9 +62,8 @@
 // To indicate that you have populated the histogram call HistogramFinished
 // with the value of the highest populated symbol, as well as the number of entries
 // in the most populated entry. These are accepted at face value.
-// The returned slice will always be length 256.
-func (s *fseEncoder) Histogram() []uint32 {
-	return s.count[:]
+func (s *fseEncoder) Histogram() *[256]uint32 {
+	return &s.count
 }
 
 // HistogramFinished can be called to indicate that the histogram has been populated.
@@ -77,21 +76,6 @@
 	s.clearCount = maxCount != 0
 }
 
-// prepare will prepare and allocate scratch tables used for both compression and decompression.
-func (s *fseEncoder) prepare() (*fseEncoder, error) {
-	if s == nil {
-		s = &fseEncoder{}
-	}
-	s.useRLE = false
-	if s.clearCount && s.maxCount == 0 {
-		for i := range s.count {
-			s.count[i] = 0
-		}
-		s.clearCount = false
-	}
-	return s, nil
-}
-
 // allocCtable will allocate tables needed for compression.
 // If existing tables a re big enough, they are simply re-used.
 func (s *fseEncoder) allocCtable() {
@@ -229,7 +213,7 @@
 		deltaFindState: 0,
 		deltaNbBits:    0,
 	}
-	if debug {
+	if debugEncoder {
 		println("setRLE: val", val, "symbolTT", s.ct.symbolTT[val])
 	}
 	s.rleVal = val
@@ -710,14 +694,6 @@
 	c.state = c.stateTable[lu]
 }
 
-// encode the output symbol provided and write it to the bitstream.
-func (c *cState) encode(symbolTT symbolTransform) {
-	nbBitsOut := (uint32(c.state) + symbolTT.deltaNbBits) >> 16
-	dstState := int32(c.state>>(nbBitsOut&15)) + int32(symbolTT.deltaFindState)
-	c.bw.addBits16NC(c.state, uint8(nbBitsOut))
-	c.state = c.stateTable[dstState]
-}
-
 // flush will write the tablelog to the output and flush the remaining full bytes.
 func (c *cState) flush(tableLog uint8) {
 	c.bw.flush32()
diff --git a/vendor/github.com/klauspost/compress/zstd/hash.go b/vendor/github.com/klauspost/compress/zstd/hash.go
index 4a75206..5d73c21 100644
--- a/vendor/github.com/klauspost/compress/zstd/hash.go
+++ b/vendor/github.com/klauspost/compress/zstd/hash.go
@@ -13,65 +13,23 @@
 	prime8bytes = 0xcf1bbcdcb7a56463
 )
 
-// hashLen returns a hash of the lowest l bytes of u for a size size of h bytes.
-// l must be >=4 and <=8. Any other value will return hash for 4 bytes.
-// h should always be <32.
-// Preferably h and l should be a constant.
-// FIXME: This does NOT get resolved, if 'mls' is constant,
-//  so this cannot be used.
-func hashLen(u uint64, hashLog, mls uint8) uint32 {
+// 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 hash5(u, hashLog)
+		return uint32(((u << (64 - 40)) * prime5bytes) >> (64 - length))
 	case 6:
-		return hash6(u, hashLog)
+		return uint32(((u << (64 - 48)) * prime6bytes) >> (64 - length))
 	case 7:
-		return hash7(u, hashLog)
+		return uint32(((u << (64 - 56)) * prime7bytes) >> (64 - length))
 	case 8:
-		return hash8(u, hashLog)
+		return uint32((u * prime8bytes) >> (64 - length))
 	default:
-		return hash4x64(u, hashLog)
+		return (uint32(u) * prime4bytes) >> (32 - length)
 	}
 }
-
-// hash3 returns the hash of the lower 3 bytes of u to fit in a hash table with h bits.
-// Preferably h should be a constant and should always be <32.
-func hash3(u uint32, h uint8) uint32 {
-	return ((u << (32 - 24)) * prime3bytes) >> ((32 - h) & 31)
-}
-
-// hash4 returns the hash of u to fit in a hash table with h bits.
-// Preferably h should be a constant and should always be <32.
-func hash4(u uint32, h uint8) uint32 {
-	return (u * prime4bytes) >> ((32 - h) & 31)
-}
-
-// hash4x64 returns the hash of the lowest 4 bytes of u to fit in a hash table with h bits.
-// Preferably h should be a constant and should always be <32.
-func hash4x64(u uint64, h uint8) uint32 {
-	return (uint32(u) * prime4bytes) >> ((32 - h) & 31)
-}
-
-// hash5 returns the hash of the lowest 5 bytes of u to fit in a hash table with h bits.
-// Preferably h should be a constant and should always be <64.
-func hash5(u uint64, h uint8) uint32 {
-	return uint32(((u << (64 - 40)) * prime5bytes) >> ((64 - h) & 63))
-}
-
-// hash6 returns the hash of the lowest 6 bytes of u to fit in a hash table with h bits.
-// Preferably h should be a constant and should always be <64.
-func hash6(u uint64, h uint8) uint32 {
-	return uint32(((u << (64 - 48)) * prime6bytes) >> ((64 - h) & 63))
-}
-
-// hash7 returns the hash of the lowest 7 bytes of u to fit in a hash table with h bits.
-// Preferably h should be a constant and should always be <64.
-func hash7(u uint64, h uint8) uint32 {
-	return uint32(((u << (64 - 56)) * prime7bytes) >> ((64 - h) & 63))
-}
-
-// hash8 returns the hash of u to fit in a hash table with h bits.
-// Preferably h should be a constant and should always be <64.
-func hash8(u uint64, h uint8) uint32 {
-	return uint32((u * prime8bytes) >> ((64 - h) & 63))
-}
diff --git a/vendor/github.com/klauspost/compress/zstd/history.go b/vendor/github.com/klauspost/compress/zstd/history.go
index f783e32..0916485 100644
--- a/vendor/github.com/klauspost/compress/zstd/history.go
+++ b/vendor/github.com/klauspost/compress/zstd/history.go
@@ -10,42 +10,50 @@
 
 // history contains the information transferred between blocks.
 type history struct {
-	b             []byte
-	huffTree      *huff0.Scratch
-	recentOffsets [3]int
+	// Literal decompression
+	huffTree *huff0.Scratch
+
+	// Sequence decompression
 	decoders      sequenceDecs
-	windowSize    int
-	maxSize       int
-	error         bool
-	dict          *dict
+	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}
-	if f := h.decoders.litLengths.fse; f != nil && !f.preDefined {
-		fseDecoderPool.Put(f)
-	}
-	if f := h.decoders.offsets.fse; f != nil && !f.preDefined {
-		fseDecoderPool.Put(f)
-	}
-	if f := h.decoders.matchLengths.fse; f != nil && !f.preDefined {
-		fseDecoderPool.Put(f)
-	}
-	h.decoders = sequenceDecs{}
-	if h.huffTree != nil {
-		if h.dict == nil || h.dict.litEnc != h.huffTree {
-			huffDecoderPool.Put(h.huffTree)
-		}
-	}
+	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
@@ -54,6 +62,7 @@
 	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
 }
@@ -83,6 +92,24 @@
 	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/README.md b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/README.md
index 69aa3bb..777290d 100644
--- a/vendor/github.com/klauspost/compress/zstd/internal/xxhash/README.md
+++ b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/README.md
@@ -2,12 +2,7 @@
 
 VENDORED: Go to [github.com/cespare/xxhash](https://github.com/cespare/xxhash) for original package.
 
-
-[![GoDoc](https://godoc.org/github.com/cespare/xxhash?status.svg)](https://godoc.org/github.com/cespare/xxhash)
-[![Build Status](https://travis-ci.org/cespare/xxhash.svg?branch=master)](https://travis-ci.org/cespare/xxhash)
-
-xxhash is a Go implementation of the 64-bit
-[xxHash](http://cyan4973.github.io/xxHash/) algorithm, XXH64. This is a
+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.
 
@@ -28,31 +23,49 @@
 func (*Digest) Sum64() uint64
 ```
 
-This implementation provides a fast pure-Go implementation and an even faster
-assembly implementation for amd64.
+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 |
-| --- | --- | --- |
-| 5 B   |  979.66 MB/s |  1291.17 MB/s  |
-| 100 B | 7475.26 MB/s | 7973.40 MB/s  |
-| 4 KB  | 17573.46 MB/s | 17602.65 MB/s |
-| 10 MB | 17131.46 MB/s | 17142.16 MB/s |
+| 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 18.04 with an Intel i7-8700K CPU using
-the following commands under Go 1.11.2:
+These numbers were generated on Ubuntu 20.04 with an Intel Xeon Platinum 8252C
+CPU using the following commands under Go 1.19.2:
 
 ```
-$ go test -tags purego -benchtime 10s -bench '/xxhash,direct,bytes'
-$ go test -benchtime 10s -bench '/xxhash,direct,bytes'
+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
index 426b9ca..fc40c82 100644
--- a/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash.go
+++ b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash.go
@@ -18,19 +18,11 @@
 	prime5 uint64 = 2870177450012600261
 )
 
-// NOTE(caleb): I'm using both consts and vars of the primes. Using consts where
-// possible in the Go code is worth a small (but measurable) performance boost
-// by avoiding some MOVQs. Vars are needed for the asm and also are useful for
-// convenience in the Go code in a few places where we need to intentionally
-// avoid constant arithmetic (e.g., v1 := prime1 + prime2 fails because the
-// result overflows a uint64).
-var (
-	prime1v = prime1
-	prime2v = prime2
-	prime3v = prime3
-	prime4v = prime4
-	prime5v = prime5
-)
+// 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 {
@@ -52,10 +44,10 @@
 
 // Reset clears the Digest's state so that it can be reused.
 func (d *Digest) Reset() {
-	d.v1 = prime1v + prime2
+	d.v1 = primes[0] + prime2
 	d.v2 = prime2
 	d.v3 = 0
-	d.v4 = -prime1v
+	d.v4 = -primes[0]
 	d.total = 0
 	d.n = 0
 }
@@ -71,21 +63,23 @@
 	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(d.mem[d.n:], b)
+		copy(memleft, b)
 		d.n += n
 		return
 	}
 
 	if d.n > 0 {
 		// Finish off the partial block.
-		copy(d.mem[d.n:], b)
+		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[32-d.n:]
+		b = b[c:]
 		d.n = 0
 	}
 
@@ -135,21 +129,20 @@
 
 	h += d.total
 
-	i, end := 0, d.n
-	for ; i+8 <= end; i += 8 {
-		k1 := round(0, u64(d.mem[i:i+8]))
+	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 i+4 <= end {
-		h ^= uint64(u32(d.mem[i:i+4])) * prime1
+	if len(b) >= 4 {
+		h ^= uint64(u32(b[:4])) * prime1
 		h = rol23(h)*prime2 + prime3
-		i += 4
+		b = b[4:]
 	}
-	for i < end {
-		h ^= uint64(d.mem[i]) * prime5
+	for ; len(b) > 0; b = b[1:] {
+		h ^= uint64(b[0]) * prime5
 		h = rol11(h) * prime1
-		i++
 	}
 
 	h ^= h >> 33
@@ -195,7 +188,6 @@
 	b, d.v4 = consumeUint64(b)
 	b, d.total = consumeUint64(b)
 	copy(d.mem[:], b)
-	b = b[len(d.mem):]
 	d.n = int(d.total % uint64(len(d.mem)))
 	return nil
 }
diff --git a/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_amd64.go b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_amd64.go
deleted file mode 100644
index 35318d7..0000000
--- a/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_amd64.go
+++ /dev/null
@@ -1,13 +0,0 @@
-// +build !appengine
-// +build gc
-// +build !purego
-
-package xxhash
-
-// Sum64 computes the 64-bit xxHash digest of b.
-//
-//go:noescape
-func Sum64(b []byte) uint64
-
-//go:noescape
-func writeBlocks(*Digest, []byte) int
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
index 2c9c535..ddb63aa 100644
--- a/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_amd64.s
+++ b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_amd64.s
@@ -1,215 +1,210 @@
+//go:build !appengine && gc && !purego && !noasm
 // +build !appengine
 // +build gc
 // +build !purego
+// +build !noasm
 
 #include "textflag.h"
 
-// Register allocation:
-// AX	h
-// CX	pointer to advance through b
-// DX	n
-// BX	loop end
-// R8	v1, k1
-// R9	v2
-// R10	v3
-// R11	v4
-// R12	tmp
-// R13	prime1v
-// R14	prime2v
-// R15	prime4v
+// 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
 
-// round reads from and advances the buffer pointer in CX.
-// It assumes that R13 has prime1v and R14 has prime2v.
-#define round(r) \
-	MOVQ  (CX), R12 \
-	ADDQ  $8, CX    \
-	IMULQ R14, R12  \
-	ADDQ  R12, r    \
-	ROLQ  $31, r    \
-	IMULQ R13, r
+#define round(acc, x) \
+	IMULQ prime2, x   \
+	ADDQ  x, acc      \
+	ROLQ  $31, acc    \
+	IMULQ prime1, acc
 
-// mergeRound applies a merge round on the two registers acc and val.
-// It assumes that R13 has prime1v, R14 has prime2v, and R15 has prime4v.
-#define mergeRound(acc, val) \
-	IMULQ R14, val \
-	ROLQ  $31, val \
-	IMULQ R13, val \
-	XORQ  val, acc \
-	IMULQ R13, acc \
-	ADDQ  R15, 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, $0-32
+TEXT ·Sum64(SB), NOSPLIT|NOFRAME, $0-32
 	// Load fixed primes.
-	MOVQ ·prime1v(SB), R13
-	MOVQ ·prime2v(SB), R14
-	MOVQ ·prime4v(SB), R15
+	MOVQ ·primes+0(SB), prime1
+	MOVQ ·primes+8(SB), prime2
+	MOVQ ·primes+24(SB), prime4
 
 	// Load slice.
-	MOVQ b_base+0(FP), CX
-	MOVQ b_len+8(FP), DX
-	LEAQ (CX)(DX*1), BX
+	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, BX
+	SUBQ $32, end
 
 	// Check whether we have at least one block.
-	CMPQ DX, $32
+	CMPQ n, $32
 	JLT  noBlocks
 
 	// Set up initial state (v1, v2, v3, v4).
-	MOVQ R13, R8
-	ADDQ R14, R8
-	MOVQ R14, R9
-	XORQ R10, R10
-	XORQ R11, R11
-	SUBQ R13, R11
+	MOVQ prime1, v1
+	ADDQ prime2, v1
+	MOVQ prime2, v2
+	XORQ v3, v3
+	XORQ v4, v4
+	SUBQ prime1, v4
 
-	// Loop until CX > BX.
-blockLoop:
-	round(R8)
-	round(R9)
-	round(R10)
-	round(R11)
+	blockLoop()
 
-	CMPQ CX, BX
-	JLE  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
 
-	MOVQ R8, AX
-	ROLQ $1, AX
-	MOVQ R9, R12
-	ROLQ $7, R12
-	ADDQ R12, AX
-	MOVQ R10, R12
-	ROLQ $12, R12
-	ADDQ R12, AX
-	MOVQ R11, R12
-	ROLQ $18, R12
-	ADDQ R12, AX
-
-	mergeRound(AX, R8)
-	mergeRound(AX, R9)
-	mergeRound(AX, R10)
-	mergeRound(AX, R11)
+	mergeRound(h, v1)
+	mergeRound(h, v2)
+	mergeRound(h, v3)
+	mergeRound(h, v4)
 
 	JMP afterBlocks
 
 noBlocks:
-	MOVQ ·prime5v(SB), AX
+	MOVQ ·primes+32(SB), h
 
 afterBlocks:
-	ADDQ DX, AX
+	ADDQ n, h
 
-	// Right now BX has len(b)-32, and we want to loop until CX > len(b)-8.
-	ADDQ $24, BX
+	ADDQ $24, end
+	CMPQ p, end
+	JG   try4
 
-	CMPQ CX, BX
-	JG   fourByte
+loop8:
+	MOVQ  (p), x
+	ADDQ  $8, p
+	round0(x)
+	XORQ  x, h
+	ROLQ  $27, h
+	IMULQ prime1, h
+	ADDQ  prime4, h
 
-wordLoop:
-	// Calculate k1.
-	MOVQ  (CX), R8
-	ADDQ  $8, CX
-	IMULQ R14, R8
-	ROLQ  $31, R8
-	IMULQ R13, R8
+	CMPQ p, end
+	JLE  loop8
 
-	XORQ  R8, AX
-	ROLQ  $27, AX
-	IMULQ R13, AX
-	ADDQ  R15, AX
+try4:
+	ADDQ $4, end
+	CMPQ p, end
+	JG   try1
 
-	CMPQ CX, BX
-	JLE  wordLoop
+	MOVL  (p), x
+	ADDQ  $4, p
+	IMULQ prime1, x
+	XORQ  x, h
 
-fourByte:
-	ADDQ $4, BX
-	CMPQ CX, BX
-	JG   singles
+	ROLQ  $23, h
+	IMULQ prime2, h
+	ADDQ  ·primes+16(SB), h
 
-	MOVL  (CX), R8
-	ADDQ  $4, CX
-	IMULQ R13, R8
-	XORQ  R8, AX
-
-	ROLQ  $23, AX
-	IMULQ R14, AX
-	ADDQ  ·prime3v(SB), AX
-
-singles:
-	ADDQ $4, BX
-	CMPQ CX, BX
+try1:
+	ADDQ $4, end
+	CMPQ p, end
 	JGE  finalize
 
-singlesLoop:
-	MOVBQZX (CX), R12
-	ADDQ    $1, CX
-	IMULQ   ·prime5v(SB), R12
-	XORQ    R12, AX
+loop1:
+	MOVBQZX (p), x
+	ADDQ    $1, p
+	IMULQ   ·primes+32(SB), x
+	XORQ    x, h
+	ROLQ    $11, h
+	IMULQ   prime1, h
 
-	ROLQ  $11, AX
-	IMULQ R13, AX
-
-	CMPQ CX, BX
-	JL   singlesLoop
+	CMPQ p, end
+	JL   loop1
 
 finalize:
-	MOVQ  AX, R12
-	SHRQ  $33, R12
-	XORQ  R12, AX
-	IMULQ R14, AX
-	MOVQ  AX, R12
-	SHRQ  $29, R12
-	XORQ  R12, AX
-	IMULQ ·prime3v(SB), AX
-	MOVQ  AX, R12
-	SHRQ  $32, R12
-	XORQ  R12, AX
+	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 AX, ret+24(FP)
+	MOVQ h, ret+24(FP)
 	RET
 
-// writeBlocks uses the same registers as above except that it uses AX to store
-// the d pointer.
-
 // func writeBlocks(d *Digest, b []byte) int
-TEXT ·writeBlocks(SB), NOSPLIT, $0-40
+TEXT ·writeBlocks(SB), NOSPLIT|NOFRAME, $0-40
 	// Load fixed primes needed for round.
-	MOVQ ·prime1v(SB), R13
-	MOVQ ·prime2v(SB), R14
+	MOVQ ·primes+0(SB), prime1
+	MOVQ ·primes+8(SB), prime2
 
 	// Load slice.
-	MOVQ arg1_base+8(FP), CX
-	MOVQ arg1_len+16(FP), DX
-	LEAQ (CX)(DX*1), BX
-	SUBQ $32, BX
+	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 arg+0(FP), AX
-	MOVQ 0(AX), R8   // v1
-	MOVQ 8(AX), R9   // v2
-	MOVQ 16(AX), R10 // v3
-	MOVQ 24(AX), R11 // v4
+	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:
-	round(R8)
-	round(R9)
-	round(R10)
-	round(R11)
-
-	CMPQ CX, BX
-	JLE  blockLoop
+	blockLoop()
 
 	// Copy vN back to d.
-	MOVQ R8, 0(AX)
-	MOVQ R9, 8(AX)
-	MOVQ R10, 16(AX)
-	MOVQ R11, 24(AX)
+	MOVQ v1, 0(d)
+	MOVQ v2, 8(d)
+	MOVQ v3, 16(d)
+	MOVQ v4, 24(d)
 
-	// The number of bytes written is CX minus the old base pointer.
-	SUBQ arg1_base+8(FP), CX
-	MOVQ CX, ret+32(FP)
+	// 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 0000000..ae7d4d3
--- /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 0000000..d4221ed
--- /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
index 4a5a821..0be16ce 100644
--- a/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_other.go
+++ b/vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_other.go
@@ -1,4 +1,5 @@
-// +build !amd64 appengine !gc purego
+//go:build (!amd64 && !arm64) || appengine || !gc || purego || noasm
+// +build !amd64,!arm64 appengine !gc purego noasm
 
 package xxhash
 
@@ -14,10 +15,10 @@
 	var h uint64
 
 	if n >= 32 {
-		v1 := prime1v + prime2
+		v1 := primes[0] + prime2
 		v2 := prime2
 		v3 := uint64(0)
-		v4 := -prime1v
+		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)]))
@@ -36,19 +37,18 @@
 
 	h += uint64(n)
 
-	i, end := 0, len(b)
-	for ; i+8 <= end; i += 8 {
-		k1 := round(0, u64(b[i:i+8:len(b)]))
+	for ; len(b) >= 8; b = b[8:] {
+		k1 := round(0, u64(b[:8]))
 		h ^= k1
 		h = rol27(h)*prime1 + prime4
 	}
-	if i+4 <= end {
-		h ^= uint64(u32(b[i:i+4:len(b)])) * prime1
+	if len(b) >= 4 {
+		h ^= uint64(u32(b[:4])) * prime1
 		h = rol23(h)*prime2 + prime3
-		i += 4
+		b = b[4:]
 	}
-	for ; i < end; i++ {
-		h ^= uint64(b[i]) * prime5
+	for ; len(b) > 0; b = b[1:] {
+		h ^= uint64(b[0]) * prime5
 		h = rol11(h) * prime1
 	}
 
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 0000000..f41932b
--- /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 0000000..0782b86
--- /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 0000000..bea1779
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/zstd/matchlen_generic.go
@@ -0,0 +1,38 @@
+//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 (
+	"math/bits"
+
+	"github.com/klauspost/compress/internal/le"
+)
+
+// matchLen returns the maximum common prefix length of a and b.
+// a must be the shortest of the two.
+func matchLen(a, b []byte) (n int) {
+	left := len(a)
+	for left >= 8 {
+		diff := le.Load64(a, n) ^ le.Load64(b, n)
+		if diff != 0 {
+			return n + bits.TrailingZeros64(diff)>>3
+		}
+		n += 8
+		left -= 8
+	}
+	a = a[n:]
+	b = b[n:]
+
+	for i := range a {
+		if a[i] != b[i] {
+			break
+		}
+		n++
+	}
+	return n
+
+}
diff --git a/vendor/github.com/klauspost/compress/zstd/seqdec.go b/vendor/github.com/klauspost/compress/zstd/seqdec.go
index 1dd39e6..9a7de82 100644
--- a/vendor/github.com/klauspost/compress/zstd/seqdec.go
+++ b/vendor/github.com/klauspost/compress/zstd/seqdec.go
@@ -20,6 +20,10 @@
 	llCode, mlCode, ofCode uint8
 }
 
+type seqVals struct {
+	ll, ml, mo int
+}
+
 func (s seq) String() string {
 	if s.offset <= 3 {
 		if s.offset == 0 {
@@ -61,16 +65,19 @@
 	offsets      sequenceDec
 	matchLengths sequenceDec
 	prevOffset   [3]int
-	hist         []byte
 	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, literals, out []byte) error {
+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())
 	}
@@ -80,8 +87,7 @@
 	if err := s.matchLengths.init(br); err != nil {
 		return errors.New("matchLengths:" + err.Error())
 	}
-	s.literals = literals
-	s.hist = hist.b
+	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
@@ -93,20 +99,153 @@
 	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) decode(seqs int, br *bitReader, hist []byte) error {
+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\n", seqs-i)
+			printf("reading sequence %d, exceeded available data. Overread by %d\n", seqs-i, -br.remain())
 			return io.ErrUnexpectedEOF
 		}
 		var ll, mo, ml int
-		if br.off > 4+((maxOffsetBits+16+16)>>3) {
+		if br.cursor > 4+((maxOffsetBits+16+16)>>3) {
 			// inlined function:
 			// ll, mo, ml = s.nextFast(br, llState, mlState, ofState)
 
@@ -151,7 +290,7 @@
 
 					if temp == 0 {
 						// 0 is not valid; input is corrupted; force offset to 1
-						println("temp was 0")
+						println("WARNING: temp was 0")
 						temp = 1
 					}
 
@@ -176,51 +315,49 @@
 		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(s.out)
+		size := ll + ml + len(out)
 		if size-startSize > maxBlockSize {
-			return fmt.Errorf("output (%d) bigger than max block size", size)
+			return fmt.Errorf("output bigger than max block size (%d)", maxBlockSize)
 		}
-		if size > cap(s.out) {
+		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(s.out) - startSize
+			used := len(out) - startSize
 			addBytes := 256 + ll + ml + used>>2
 			// Clamp to max block size.
 			if used+addBytes > maxBlockSize {
 				addBytes = maxBlockSize - used
 			}
-			s.out = append(s.out, make([]byte, addBytes)...)
-			s.out = s.out[:len(s.out)-addBytes]
+			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
-		s.out = append(s.out, s.literals[:ll]...)
+		out = append(out, s.literals[:ll]...)
 		s.literals = s.literals[ll:]
-		out := s.out
 
 		if mo == 0 && ml > 0 {
 			return fmt.Errorf("zero matchoff and matchlen (%d) > 0", ml)
 		}
 
-		if mo > len(s.out)+len(hist) || mo > s.windowSize {
+		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(s.out)+len(hist))
+				return fmt.Errorf("match offset (%d) bigger than current history (%d)", mo, len(out)+len(hist)-startSize)
 			}
 
 			// we may be in dictionary.
-			dictO := len(s.dict) - (mo - (len(s.out) + len(hist)))
+			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(s.out)+len(hist))
+				return fmt.Errorf("match offset (%d) bigger than current history (%d)", mo, len(out)+len(hist)-startSize)
 			}
 			end := dictO + ml
 			if end > len(s.dict) {
 				out = append(out, s.dict[dictO:]...)
-				mo -= len(s.dict) - dictO
 				ml -= len(s.dict) - dictO
 			} else {
 				out = append(out, s.dict[dictO:end]...)
@@ -231,26 +368,25 @@
 
 		// Copy from history.
 		// TODO: Blocks without history could be made to ignore this completely.
-		if v := mo - len(s.out); v > 0 {
+		if v := mo - len(out); v > 0 {
 			// v is the start position in history from end.
-			start := len(s.hist) - v
+			start := len(hist) - v
 			if ml > v {
 				// Some goes into current block.
 				// Copy remainder of history
-				out = append(out, s.hist[start:]...)
-				mo -= v
+				out = append(out, hist[start:]...)
 				ml -= v
 			} else {
-				out = append(out, s.hist[start:start+ml]...)
+				out = append(out, hist[start:start+ml]...)
 				ml = 0
 			}
 		}
 		// We must be in current buffer now
 		if ml > 0 {
-			start := len(s.out) - mo
-			if ml <= len(s.out)-start {
+			start := len(out) - mo
+			if ml <= len(out)-start {
 				// No overlap
-				out = append(out, s.out[start:start+ml]...)
+				out = append(out, out[start:start+ml]...)
 			} else {
 				// Overlapping copy
 				// Extend destination slice and copy one byte at the time.
@@ -264,7 +400,6 @@
 				}
 			}
 		}
-		s.out = out
 		if i == 0 {
 			// This is the last sequence, so we shouldn't update state.
 			break
@@ -278,7 +413,8 @@
 			mlState = mlTable[mlState.newState()&maxTableMask]
 			ofState = ofTable[ofState.newState()&maxTableMask]
 		} else {
-			bits := br.getBitsFast(nBits)
+			bits := br.get32BitsFast(nBits)
+
 			lowBits := uint16(bits >> ((ofState.nbBits() + mlState.nbBits()) & 31))
 			llState = llTable[(llState.newState()+lowBits)&maxTableMask]
 
@@ -291,19 +427,13 @@
 		}
 	}
 
-	// Add final literals
-	s.out = append(s.out, s.literals...)
-	return nil
-}
+	if size := len(s.literals) + len(out) - startSize; size > maxBlockSize {
+		return fmt.Errorf("output bigger than max block size (%d)", maxBlockSize)
+	}
 
-// update states, at least 27 bits must be available.
-func (s *sequenceDecs) update(br *bitReader) {
-	// Max 8 bits
-	s.litLengths.state.next(br)
-	// Max 9 bits
-	s.matchLengths.state.next(br)
-	// Max 8 bits
-	s.offsets.state.next(br)
+	// Add final literals
+	s.out = append(out, s.literals...)
+	return br.close()
 }
 
 var bitMask [16]uint16
@@ -314,87 +444,6 @@
 	}
 }
 
-// update states, at least 27 bits must be available.
-func (s *sequenceDecs) updateAlt(br *bitReader) {
-	// Update all 3 states at once. Approx 20% faster.
-	a, b, c := s.litLengths.state.state, s.matchLengths.state.state, s.offsets.state.state
-
-	nBits := a.nbBits() + b.nbBits() + c.nbBits()
-	if nBits == 0 {
-		s.litLengths.state.state = s.litLengths.state.dt[a.newState()]
-		s.matchLengths.state.state = s.matchLengths.state.dt[b.newState()]
-		s.offsets.state.state = s.offsets.state.dt[c.newState()]
-		return
-	}
-	bits := br.getBitsFast(nBits)
-	lowBits := uint16(bits >> ((c.nbBits() + b.nbBits()) & 31))
-	s.litLengths.state.state = s.litLengths.state.dt[a.newState()+lowBits]
-
-	lowBits = uint16(bits >> (c.nbBits() & 31))
-	lowBits &= bitMask[b.nbBits()&15]
-	s.matchLengths.state.state = s.matchLengths.state.dt[b.newState()+lowBits]
-
-	lowBits = uint16(bits) & bitMask[c.nbBits()&15]
-	s.offsets.state.state = s.offsets.state.dt[c.newState()+lowBits]
-}
-
-// nextFast will return new states when there are at least 4 unused bytes left on the stream when done.
-func (s *sequenceDecs) nextFast(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.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
-		return
-	}
-	// 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]
-		return
-	}
-	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("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
-	return
-}
-
 func (s *sequenceDecs) next(br *bitReader, llState, mlState, ofState decSymbol) (ll, mo, ml int) {
 	// Final will not read from stream.
 	ll, llB := llState.final()
@@ -403,18 +452,13 @@
 
 	// extra bits are stored in reverse order.
 	br.fill()
-	if s.maxBits <= 32 {
-		mo += br.getBits(moB)
-		ml += br.getBits(mlB)
-		ll += br.getBits(llB)
-	} else {
-		mo += br.getBits(moB)
+	mo += br.getBits(moB)
+	if s.maxBits > 32 {
 		br.fill()
-		// matchlength+literal length, max 32 bits
-		ml += br.getBits(mlB)
-		ll += br.getBits(llB)
-
 	}
+	// matchlength+literal length, max 32 bits
+	ml += br.getBits(mlB)
+	ll += br.getBits(llB)
 	mo = s.adjustOffset(mo, ll, moB)
 	return
 }
@@ -457,36 +501,3 @@
 	s.prevOffset[0] = temp
 	return temp
 }
-
-// mergeHistory will merge history.
-func (s *sequenceDecs) mergeHistory(hist *sequenceDecs) (*sequenceDecs, error) {
-	for i := uint(0); i < 3; i++ {
-		var sNew, sHist *sequenceDec
-		switch i {
-		default:
-			// same as "case 0":
-			sNew = &s.litLengths
-			sHist = &hist.litLengths
-		case 1:
-			sNew = &s.offsets
-			sHist = &hist.offsets
-		case 2:
-			sNew = &s.matchLengths
-			sHist = &hist.matchLengths
-		}
-		if sNew.repeat {
-			if sHist.fse == nil {
-				return nil, fmt.Errorf("sequence stream %d, repeat requested, but no history", i)
-			}
-			continue
-		}
-		if sNew.fse == nil {
-			return nil, fmt.Errorf("sequence stream %d, no fse found", i)
-		}
-		if sHist.fse != nil && !sHist.fse.preDefined {
-			fseDecoderPool.Put(sHist.fse)
-		}
-		sHist.fse = sNew.fse
-	}
-	return hist, nil
-}
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 0000000..c59f17e
--- /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 erroneous 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 erroneous 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 0000000..a708ca6
--- /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 40(CX), BX
+	MOVQ    (CX), AX
+	MOVQ    32(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, 40(AX)
+	MOVQ SI, 32(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 40(CX), BX
+	MOVQ    (CX), AX
+	MOVQ    32(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, 40(AX)
+	MOVQ SI, 32(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 40(BX), DX
+	MOVQ    (BX), CX
+	MOVQ    32(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, 40(CX)
+	MOVQ BX, 32(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 40(BX), DX
+	MOVQ    (BX), CX
+	MOVQ    32(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, 40(CX)
+	MOVQ BX, 32(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 40(CX), BX
+	MOVQ    (CX), AX
+	MOVQ    32(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 pointer 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, 40(AX)
+	MOVQ SI, 32(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 40(BX), DX
+	MOVQ    (BX), CX
+	MOVQ    32(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 pointer 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, 40(CX)
+	MOVQ BX, 32(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 40(CX), BX
+	MOVQ    (CX), AX
+	MOVQ    32(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 pointer 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, 40(AX)
+	MOVQ SI, 32(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 40(BX), DX
+	MOVQ    (BX), CX
+	MOVQ    32(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 pointer 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, 40(CX)
+	MOVQ BX, 32(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 0000000..7cec219
--- /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 br.cursor > 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
index 8014174..65045ea 100644
--- a/vendor/github.com/klauspost/compress/zstd/seqenc.go
+++ b/vendor/github.com/klauspost/compress/zstd/seqenc.go
@@ -69,7 +69,6 @@
 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
@@ -102,7 +101,6 @@
 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
diff --git a/vendor/github.com/klauspost/compress/zstd/snappy.go b/vendor/github.com/klauspost/compress/zstd/snappy.go
index 9d9d1d5..a17381b 100644
--- a/vendor/github.com/klauspost/compress/zstd/snappy.go
+++ b/vendor/github.com/klauspost/compress/zstd/snappy.go
@@ -10,8 +10,8 @@
 	"hash/crc32"
 	"io"
 
-	"github.com/golang/snappy"
 	"github.com/klauspost/compress/huff0"
+	snappy "github.com/klauspost/compress/internal/snapref"
 )
 
 const (
@@ -95,10 +95,9 @@
 	var written int64
 	var readHeader bool
 	{
-		var header []byte
-		var n int
-		header, r.err = frameHeader{WindowSize: snappyMaxBlockSize}.appendTo(r.buf[:0])
+		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
@@ -198,12 +197,12 @@
 
 			n, r.err = w.Write(r.block.output)
 			if r.err != nil {
-				return written, err
+				return written, r.err
 			}
 			written += int64(n)
 			continue
 		case chunkTypeUncompressedData:
-			if debug {
+			if debugEncoder {
 				println("Uncompressed, chunklen", chunkLen)
 			}
 			// Section 4.3. Uncompressed data (chunk type 0x01).
@@ -240,13 +239,13 @@
 			}
 			n, r.err = w.Write(r.block.output)
 			if r.err != nil {
-				return written, err
+				return written, r.err
 			}
 			written += int64(n)
 			continue
 
 		case chunkTypeStreamIdentifier:
-			if debug {
+			if debugEncoder {
 				println("stream id", chunkLen, len(snappyMagicBody))
 			}
 			// Section 4.1. Stream identifier (chunk type 0xff).
diff --git a/vendor/github.com/klauspost/compress/zstd/zip.go b/vendor/github.com/klauspost/compress/zstd/zip.go
index e35a0a2..29c15c8 100644
--- a/vendor/github.com/klauspost/compress/zstd/zip.go
+++ b/vendor/github.com/klauspost/compress/zstd/zip.go
@@ -13,40 +13,63 @@
 // See https://www.winzip.com/win/en/comp_info.html
 const ZipMethodWinZip = 93
 
-// ZipMethodPKWare is the method number used by PKWARE to indicate Zstandard compression.
-// See https://pkware.cachefly.net/webdocs/APPNOTE/APPNOTE-6.3.7.TXT
+// 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
 
-var zipReaderPool sync.Pool
-
-// newZipReader cannot be used since we would leak goroutines...
-func newZipReader(r io.Reader) io.ReadCloser {
-	dec, ok := zipReaderPool.Get().(*Decoder)
-	if ok {
-		dec.Reset(r)
-	} else {
-		d, err := NewReader(r, WithDecoderConcurrency(1), WithDecoderLowmem(true))
-		if err != nil {
-			panic(err)
-		}
-		dec = d
+// 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 &pooledZipReader{dec: dec}
+	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
-	dec *Decoder
+	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")
+		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
 }
 
@@ -56,15 +79,16 @@
 	var err error
 	if r.dec != nil {
 		err = r.dec.Reset(nil)
-		zipReaderPool.Put(r.dec)
+		r.pool.Put(r.dec)
 		r.dec = nil
 	}
 	return err
 }
 
 type pooledZipWriter struct {
-	mu  sync.Mutex // guards Close and Read
-	enc *Encoder
+	mu   sync.Mutex // guards Close and Read
+	enc  *Encoder
+	pool *sync.Pool
 }
 
 func (w *pooledZipWriter) Write(p []byte) (n int, err error) {
@@ -82,7 +106,7 @@
 	var err error
 	if w.enc != nil {
 		err = w.enc.Close()
-		zipReaderPool.Put(w.enc)
+		w.pool.Put(w.enc)
 		w.enc = nil
 	}
 	return err
@@ -103,18 +127,15 @@
 				return nil, err
 			}
 		}
-		return &pooledZipWriter{enc: enc}, nil
+		return &pooledZipWriter{enc: enc, pool: &pool}, nil
 	}
 }
 
 // ZipDecompressor returns a decompressor that can be registered with zip libraries.
 // See ZipCompressor for example.
-func ZipDecompressor() func(r io.Reader) io.ReadCloser {
-	return func(r io.Reader) io.ReadCloser {
-		d, err := NewReader(r, WithDecoderConcurrency(1), WithDecoderLowmem(true))
-		if err != nil {
-			panic(err)
-		}
-		return d.IOReadCloser()
-	}
+// 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
index 1ba308c..6252b46 100644
--- a/vendor/github.com/klauspost/compress/zstd/zstd.go
+++ b/vendor/github.com/klauspost/compress/zstd/zstd.go
@@ -5,16 +5,22 @@
 
 import (
 	"bytes"
-	"encoding/binary"
 	"errors"
 	"log"
 	"math"
-	"math/bits"
+
+	"github.com/klauspost/compress/internal/le"
 )
 
 // enable debug printing
 const debug = false
 
+// enable encoding debug printing
+const debugEncoder = debug
+
+// enable decoding debug printing
+const debugDecoder = debug
+
 // Enable extra assertions.
 const debugAsserts = debug || false
 
@@ -30,8 +36,8 @@
 // zstdMinMatch is the minimum zstd match length.
 const zstdMinMatch = 3
 
-// Reset the buffer offset when reaching this.
-const bufferReset = math.MaxInt32 - MaxWindowSize
+// fcsUnknown is used for unknown frame content size.
+const fcsUnknown = math.MaxUint64
 
 var (
 	// ErrReservedBlockType is returned when a reserved block type is found.
@@ -46,6 +52,10 @@
 	// 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")
@@ -62,13 +72,16 @@
 	ErrDecoderSizeExceeded = errors.New("decompressed size exceeds configured limit")
 
 	// ErrUnknownDictionary is returned if the dictionary ID is unknown.
-	// For the time being dictionaries are not supported.
 	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")
 
@@ -76,66 +89,33 @@
 	// Close has been called.
 	ErrDecoderClosed = errors.New("decoder used after Close")
 
+	// ErrEncoderClosed will be returned if the Encoder was used after
+	// Close has been called.
+	ErrEncoderClosed = errors.New("encoder used after Close")
+
 	// ErrDecoderNilInput is returned when a nil Reader was provided
 	// and an operation other than Reset/DecodeAll/Close was attempted.
 	ErrDecoderNilInput = errors.New("nil input provided as reader")
 )
 
 func println(a ...interface{}) {
-	if debug {
+	if debug || debugDecoder || debugEncoder {
 		log.Println(a...)
 	}
 }
 
 func printf(format string, a ...interface{}) {
-	if debug {
+	if debug || debugDecoder || debugEncoder {
 		log.Printf(format, a...)
 	}
 }
 
-// matchLenFast does matching, but will not match the last up to 7 bytes.
-func matchLenFast(a, b []byte) int {
-	endI := len(a) & (math.MaxInt32 - 7)
-	for i := 0; i < endI; i += 8 {
-		if diff := load64(a, i) ^ load64(b, i); diff != 0 {
-			return i + bits.TrailingZeros64(diff)>>3
-		}
-	}
-	return endI
-}
-
-// matchLen returns the maximum length.
-// a must be the shortest of the two.
-// The function also returns whether all bytes matched.
-func matchLen(a, b []byte) int {
-	b = b[:len(a)]
-	for i := 0; i < len(a)-7; i += 8 {
-		if diff := load64(a, i) ^ load64(b, i); diff != 0 {
-			return i + (bits.TrailingZeros64(diff) >> 3)
-		}
-	}
-
-	checked := (len(a) >> 3) << 3
-	a = a[checked:]
-	b = b[checked:]
-	for i := range a {
-		if a[i] != b[i] {
-			return i + checked
-		}
-	}
-	return len(a) + checked
-}
-
 func load3232(b []byte, i int32) uint32 {
-	return binary.LittleEndian.Uint32(b[i:])
+	return le.Load32(b, i)
 }
 
 func load6432(b []byte, i int32) uint64 {
-	return binary.LittleEndian.Uint64(b[i:])
-}
-
-func load64(b []byte, i int) uint64 {
-	return binary.LittleEndian.Uint64(b[i:])
+	return le.Load64(b, i)
 }
 
 type byter interface {