diff --git a/vendor/github.com/klauspost/compress/.gitattributes b/vendor/github.com/klauspost/compress/.gitattributes
new file mode 100644
index 0000000..4024335
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/.gitattributes
@@ -0,0 +1,2 @@
+* -text
+*.bin -text -diff
diff --git a/vendor/github.com/klauspost/compress/.gitignore b/vendor/github.com/klauspost/compress/.gitignore
new file mode 100644
index 0000000..d31b378
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/.gitignore
@@ -0,0 +1,32 @@
+# Compiled Object files, Static and Dynamic libs (Shared Objects)
+*.o
+*.a
+*.so
+
+# Folders
+_obj
+_test
+
+# Architecture specific extensions/prefixes
+*.[568vq]
+[568vq].out
+
+*.cgo1.go
+*.cgo2.c
+_cgo_defun.c
+_cgo_gotypes.go
+_cgo_export.*
+
+_testmain.go
+
+*.exe
+*.test
+*.prof
+/s2/cmd/_s2sx/sfx-exe
+
+# Linux perf files
+perf.data
+perf.data.old
+
+# gdb history
+.gdb_history
diff --git a/vendor/github.com/klauspost/compress/.goreleaser.yml b/vendor/github.com/klauspost/compress/.goreleaser.yml
new file mode 100644
index 0000000..4528059
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/.goreleaser.yml
@@ -0,0 +1,123 @@
+version: 2
+
+before:
+  hooks:
+    - ./gen.sh
+
+builds:
+  -
+    id: "s2c"
+    binary: s2c
+    main: ./s2/cmd/s2c/main.go
+    flags:
+      - -trimpath
+    env:
+      - CGO_ENABLED=0
+    goos:
+      - aix
+      - linux
+      - freebsd
+      - netbsd
+      - windows
+      - darwin
+    goarch:
+      - 386
+      - amd64
+      - arm
+      - arm64
+      - ppc64
+      - ppc64le
+      - mips64
+      - mips64le
+    goarm:
+      - 7
+  -
+    id: "s2d"
+    binary: s2d
+    main: ./s2/cmd/s2d/main.go
+    flags:
+      - -trimpath
+    env:
+      - CGO_ENABLED=0
+    goos:
+      - aix
+      - linux
+      - freebsd
+      - netbsd
+      - windows
+      - darwin
+    goarch:
+      - 386
+      - amd64
+      - arm
+      - arm64
+      - ppc64
+      - ppc64le
+      - mips64
+      - mips64le
+    goarm:
+      - 7
+  -
+    id: "s2sx"
+    binary: s2sx
+    main: ./s2/cmd/_s2sx/main.go
+    flags:
+      - -modfile=s2sx.mod
+      - -trimpath
+    env:
+      - CGO_ENABLED=0
+    goos:
+      - aix
+      - linux
+      - freebsd
+      - netbsd
+      - windows
+      - darwin
+    goarch:
+      - 386
+      - amd64
+      - arm
+      - arm64
+      - ppc64
+      - ppc64le
+      - mips64
+      - mips64le
+    goarm:
+      - 7
+
+archives:
+  -
+    id: s2-binaries
+    name_template: "s2-{{ .Os }}_{{ .Arch }}{{ if .Arm }}v{{ .Arm }}{{ end }}"
+    format_overrides:
+      - goos: windows
+        format: zip
+    files:
+      - unpack/*
+      - s2/LICENSE
+      - s2/README.md
+checksum:
+  name_template: 'checksums.txt'
+snapshot:
+  version_template: "{{ .Tag }}-next"
+changelog:
+  sort: asc
+  filters:
+    exclude:
+    - '^doc:'
+    - '^docs:'
+    - '^test:'
+    - '^tests:'
+    - '^Update\sREADME.md'
+
+nfpms:
+  -
+    file_name_template: "s2_package__{{ .Os }}_{{ .Arch }}{{ if .Arm }}v{{ .Arm }}{{ end }}"
+    vendor: Klaus Post
+    homepage: https://github.com/klauspost/compress
+    maintainer: Klaus Post <klauspost@gmail.com>
+    description: S2 Compression Tool
+    license: BSD 3-Clause
+    formats:
+      - deb
+      - rpm
diff --git a/vendor/github.com/klauspost/compress/LICENSE b/vendor/github.com/klauspost/compress/LICENSE
index 1eb75ef..87d5574 100644
--- a/vendor/github.com/klauspost/compress/LICENSE
+++ b/vendor/github.com/klauspost/compress/LICENSE
@@ -26,3 +26,279 @@
 THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+------------------
+
+Files: gzhttp/*
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2016-2017 The New York Times Company
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+
+------------------
+
+Files: s2/cmd/internal/readahead/*
+
+The MIT License (MIT)
+
+Copyright (c) 2015 Klaus Post
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+
+---------------------
+Files: snappy/*
+Files: internal/snapref/*
+
+Copyright (c) 2011 The Snappy-Go Authors. All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+   * Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+   * Redistributions in binary form must reproduce the above
+copyright notice, this list of conditions and the following disclaimer
+in the documentation and/or other materials provided with the
+distribution.
+   * Neither the name of Google Inc. nor the names of its
+contributors may be used to endorse or promote products derived from
+this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+-----------------
+
+Files: s2/cmd/internal/filepathx/*
+
+Copyright 2016 The filepathx Authors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
diff --git a/vendor/github.com/klauspost/compress/README.md b/vendor/github.com/klauspost/compress/README.md
new file mode 100644
index 0000000..244ee19
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/README.md
@@ -0,0 +1,671 @@
+# compress
+
+This package provides various compression algorithms.
+
+* [zstandard](https://github.com/klauspost/compress/tree/master/zstd#zstd) compression and decompression in pure Go.
+* [S2](https://github.com/klauspost/compress/tree/master/s2#s2-compression) is a high performance replacement for Snappy.
+* Optimized [deflate](https://godoc.org/github.com/klauspost/compress/flate) packages which can be used as a dropin replacement for [gzip](https://godoc.org/github.com/klauspost/compress/gzip), [zip](https://godoc.org/github.com/klauspost/compress/zip) and [zlib](https://godoc.org/github.com/klauspost/compress/zlib).
+* [snappy](https://github.com/klauspost/compress/tree/master/snappy) is a drop-in replacement for `github.com/golang/snappy` offering better compression and concurrent streams.
+* [huff0](https://github.com/klauspost/compress/tree/master/huff0) and [FSE](https://github.com/klauspost/compress/tree/master/fse) implementations for raw entropy encoding.
+* [gzhttp](https://github.com/klauspost/compress/tree/master/gzhttp) Provides client and server wrappers for handling gzipped requests efficiently.
+* [pgzip](https://github.com/klauspost/pgzip) is a separate package that provides a very fast parallel gzip implementation.
+
+[![Go Reference](https://pkg.go.dev/badge/klauspost/compress.svg)](https://pkg.go.dev/github.com/klauspost/compress?tab=subdirectories)
+[![Go](https://github.com/klauspost/compress/actions/workflows/go.yml/badge.svg)](https://github.com/klauspost/compress/actions/workflows/go.yml)
+[![Sourcegraph Badge](https://sourcegraph.com/github.com/klauspost/compress/-/badge.svg)](https://sourcegraph.com/github.com/klauspost/compress?badge)
+
+# package usage
+
+Use `go get github.com/klauspost/compress@latest` to add it to your project.
+
+This package will support the current Go version and 2 versions back.
+
+* Use the `nounsafe` tag to disable all use of the "unsafe" package.
+* Use the `noasm` tag to disable all assembly across packages.
+
+Use the links above for more information on each.
+
+# changelog
+
+* Feb 19th, 2025 - [1.18.0](https://github.com/klauspost/compress/releases/tag/v1.18.0)
+  * Add unsafe little endian loaders https://github.com/klauspost/compress/pull/1036
+  * fix: check `r.err != nil` but return a nil value error `err` by @alingse in https://github.com/klauspost/compress/pull/1028
+  * flate: Simplify L4-6 loading https://github.com/klauspost/compress/pull/1043
+  * flate: Simplify matchlen (remove asm) https://github.com/klauspost/compress/pull/1045
+  * s2: Improve small block compression speed w/o asm https://github.com/klauspost/compress/pull/1048
+  * flate: Fix matchlen L5+L6 https://github.com/klauspost/compress/pull/1049
+  * flate: Cleanup & reduce casts https://github.com/klauspost/compress/pull/1050
+
+* Oct 11th, 2024 - [1.17.11](https://github.com/klauspost/compress/releases/tag/v1.17.11)
+  * zstd: Fix extra CRC written with multiple Close calls https://github.com/klauspost/compress/pull/1017
+  * s2: Don't use stack for index tables https://github.com/klauspost/compress/pull/1014
+  * gzhttp: No content-type on no body response code by @juliens in https://github.com/klauspost/compress/pull/1011
+  * gzhttp: Do not set the content-type when response has no body by @kevinpollet in https://github.com/klauspost/compress/pull/1013
+
+* Sep 23rd, 2024 - [1.17.10](https://github.com/klauspost/compress/releases/tag/v1.17.10)
+	* gzhttp: Add TransportAlwaysDecompress option. https://github.com/klauspost/compress/pull/978
+	* gzhttp: Add supported decompress request body by @mirecl in https://github.com/klauspost/compress/pull/1002
+	* s2: Add EncodeBuffer buffer recycling callback https://github.com/klauspost/compress/pull/982
+	* zstd: Improve memory usage on small streaming encodes https://github.com/klauspost/compress/pull/1007
+	* flate: read data written with partial flush by @vajexal in https://github.com/klauspost/compress/pull/996
+
+* Jun 12th, 2024 - [1.17.9](https://github.com/klauspost/compress/releases/tag/v1.17.9)
+	* s2: Reduce ReadFrom temporary allocations https://github.com/klauspost/compress/pull/949
+	* flate, zstd: Shave some bytes off amd64 matchLen by @greatroar in https://github.com/klauspost/compress/pull/963
+	* Upgrade zip/zlib to 1.22.4 upstream https://github.com/klauspost/compress/pull/970 https://github.com/klauspost/compress/pull/971
+	* zstd: BuildDict fails with RLE table https://github.com/klauspost/compress/pull/951
+
+* Apr 9th, 2024 - [1.17.8](https://github.com/klauspost/compress/releases/tag/v1.17.8)
+	* zstd: Reject blocks where reserved values are not 0 https://github.com/klauspost/compress/pull/885
+	* zstd: Add RLE detection+encoding https://github.com/klauspost/compress/pull/938
+
+* Feb 21st, 2024 - [1.17.7](https://github.com/klauspost/compress/releases/tag/v1.17.7)
+	* s2: Add AsyncFlush method: Complete the block without flushing by @Jille in https://github.com/klauspost/compress/pull/927
+	* s2: Fix literal+repeat exceeds dst crash https://github.com/klauspost/compress/pull/930
+  
+* Feb 5th, 2024 - [1.17.6](https://github.com/klauspost/compress/releases/tag/v1.17.6)
+	* zstd: Fix incorrect repeat coding in best mode https://github.com/klauspost/compress/pull/923
+	* s2: Fix DecodeConcurrent deadlock on errors https://github.com/klauspost/compress/pull/925
+  
+* Jan 26th, 2024 - [v1.17.5](https://github.com/klauspost/compress/releases/tag/v1.17.5)
+	* flate: Fix reset with dictionary on custom window encodes https://github.com/klauspost/compress/pull/912
+	* zstd: Add Frame header encoding and stripping https://github.com/klauspost/compress/pull/908
+	* zstd: Limit better/best default window to 8MB https://github.com/klauspost/compress/pull/913
+	* zstd: Speed improvements by @greatroar in https://github.com/klauspost/compress/pull/896 https://github.com/klauspost/compress/pull/910
+	* s2: Fix callbacks for skippable blocks and disallow 0xfe (Padding) by @Jille in https://github.com/klauspost/compress/pull/916 https://github.com/klauspost/compress/pull/917
+https://github.com/klauspost/compress/pull/919 https://github.com/klauspost/compress/pull/918
+
+* Dec 1st, 2023 - [v1.17.4](https://github.com/klauspost/compress/releases/tag/v1.17.4)
+	* huff0: Speed up symbol counting by @greatroar in https://github.com/klauspost/compress/pull/887
+	* huff0: Remove byteReader by @greatroar in https://github.com/klauspost/compress/pull/886
+	* gzhttp: Allow overriding decompression on transport https://github.com/klauspost/compress/pull/892
+	* gzhttp: Clamp compression level https://github.com/klauspost/compress/pull/890
+	* gzip: Error out if reserved bits are set https://github.com/klauspost/compress/pull/891
+
+* Nov 15th, 2023 - [v1.17.3](https://github.com/klauspost/compress/releases/tag/v1.17.3)
+	* fse: Fix max header size https://github.com/klauspost/compress/pull/881
+	* zstd: Improve better/best compression https://github.com/klauspost/compress/pull/877
+	* gzhttp: Fix missing content type on Close https://github.com/klauspost/compress/pull/883
+
+* Oct 22nd, 2023 - [v1.17.2](https://github.com/klauspost/compress/releases/tag/v1.17.2)
+	* zstd: Fix rare *CORRUPTION* output in "best" mode. See https://github.com/klauspost/compress/pull/876
+
+* Oct 14th, 2023 - [v1.17.1](https://github.com/klauspost/compress/releases/tag/v1.17.1)
+	* s2: Fix S2 "best" dictionary wrong encoding https://github.com/klauspost/compress/pull/871
+	* flate: Reduce allocations in decompressor and minor code improvements by @fakefloordiv in https://github.com/klauspost/compress/pull/869
+	* s2: Fix EstimateBlockSize on 6&7 length input https://github.com/klauspost/compress/pull/867
+
+* Sept 19th, 2023 - [v1.17.0](https://github.com/klauspost/compress/releases/tag/v1.17.0)
+	* Add experimental dictionary builder  https://github.com/klauspost/compress/pull/853
+	* Add xerial snappy read/writer https://github.com/klauspost/compress/pull/838
+	* flate: Add limited window compression https://github.com/klauspost/compress/pull/843
+	* s2: Do 2 overlapping match checks https://github.com/klauspost/compress/pull/839
+	* flate: Add amd64 assembly matchlen https://github.com/klauspost/compress/pull/837
+	* gzip: Copy bufio.Reader on Reset by @thatguystone in https://github.com/klauspost/compress/pull/860
+
+<details>
+	<summary>See changes to v1.16.x</summary>
+
+   
+* July 1st, 2023 - [v1.16.7](https://github.com/klauspost/compress/releases/tag/v1.16.7)
+	* zstd: Fix default level first dictionary encode https://github.com/klauspost/compress/pull/829
+	* s2: add GetBufferCapacity() method by @GiedriusS in https://github.com/klauspost/compress/pull/832
+
+* June 13, 2023 - [v1.16.6](https://github.com/klauspost/compress/releases/tag/v1.16.6)
+	* zstd: correctly ignore WithEncoderPadding(1) by @ianlancetaylor in https://github.com/klauspost/compress/pull/806
+	* zstd: Add amd64 match length assembly https://github.com/klauspost/compress/pull/824
+	* gzhttp: Handle informational headers by @rtribotte in https://github.com/klauspost/compress/pull/815
+	* s2: Improve Better compression slightly https://github.com/klauspost/compress/pull/663
+
+* Apr 16, 2023 - [v1.16.5](https://github.com/klauspost/compress/releases/tag/v1.16.5)
+	* zstd: readByte needs to use io.ReadFull by @jnoxon in https://github.com/klauspost/compress/pull/802
+	* gzip: Fix WriterTo after initial read https://github.com/klauspost/compress/pull/804
+
+* Apr 5, 2023 - [v1.16.4](https://github.com/klauspost/compress/releases/tag/v1.16.4)
+	* zstd: Improve zstd best efficiency by @greatroar and @klauspost in https://github.com/klauspost/compress/pull/784
+	* zstd: Respect WithAllLitEntropyCompression https://github.com/klauspost/compress/pull/792
+	* zstd: Fix amd64 not always detecting corrupt data https://github.com/klauspost/compress/pull/785
+	* zstd: Various minor improvements by @greatroar in https://github.com/klauspost/compress/pull/788 https://github.com/klauspost/compress/pull/794 https://github.com/klauspost/compress/pull/795
+	* s2: Fix huge block overflow https://github.com/klauspost/compress/pull/779
+	* s2: Allow CustomEncoder fallback https://github.com/klauspost/compress/pull/780
+	* gzhttp: Support ResponseWriter Unwrap() in gzhttp handler by @jgimenez in https://github.com/klauspost/compress/pull/799
+
+* Mar 13, 2023 - [v1.16.1](https://github.com/klauspost/compress/releases/tag/v1.16.1)
+	* zstd: Speed up + improve best encoder by @greatroar in https://github.com/klauspost/compress/pull/776
+	* gzhttp: Add optional [BREACH mitigation](https://github.com/klauspost/compress/tree/master/gzhttp#breach-mitigation). https://github.com/klauspost/compress/pull/762 https://github.com/klauspost/compress/pull/768 https://github.com/klauspost/compress/pull/769 https://github.com/klauspost/compress/pull/770 https://github.com/klauspost/compress/pull/767
+	* s2: Add Intel LZ4s converter https://github.com/klauspost/compress/pull/766
+	* zstd: Minor bug fixes https://github.com/klauspost/compress/pull/771 https://github.com/klauspost/compress/pull/772 https://github.com/klauspost/compress/pull/773
+	* huff0: Speed up compress1xDo by @greatroar in https://github.com/klauspost/compress/pull/774
+
+* Feb 26, 2023 - [v1.16.0](https://github.com/klauspost/compress/releases/tag/v1.16.0)
+	* s2: Add [Dictionary](https://github.com/klauspost/compress/tree/master/s2#dictionaries) support.  https://github.com/klauspost/compress/pull/685
+	* s2: Add Compression Size Estimate.  https://github.com/klauspost/compress/pull/752
+	* s2: Add support for custom stream encoder. https://github.com/klauspost/compress/pull/755
+	* s2: Add LZ4 block converter. https://github.com/klauspost/compress/pull/748
+	* s2: Support io.ReaderAt in ReadSeeker. https://github.com/klauspost/compress/pull/747
+	* s2c/s2sx: Use concurrent decoding. https://github.com/klauspost/compress/pull/746
+</details>
+
+<details>
+	<summary>See changes to v1.15.x</summary>
+	
+* Jan 21st, 2023 (v1.15.15)
+	* deflate: Improve level 7-9 https://github.com/klauspost/compress/pull/739
+	* zstd: Add delta encoding support by @greatroar in https://github.com/klauspost/compress/pull/728
+	* zstd: Various speed improvements by @greatroar https://github.com/klauspost/compress/pull/741 https://github.com/klauspost/compress/pull/734 https://github.com/klauspost/compress/pull/736 https://github.com/klauspost/compress/pull/744 https://github.com/klauspost/compress/pull/743 https://github.com/klauspost/compress/pull/745
+	* gzhttp: Add SuffixETag() and DropETag() options to prevent ETag collisions on compressed responses by @willbicks in https://github.com/klauspost/compress/pull/740
+
+* Jan 3rd, 2023 (v1.15.14)
+
+	* flate: Improve speed in big stateless blocks https://github.com/klauspost/compress/pull/718
+	* zstd: Minor speed tweaks by @greatroar in https://github.com/klauspost/compress/pull/716 https://github.com/klauspost/compress/pull/720
+	* export NoGzipResponseWriter for custom ResponseWriter wrappers by @harshavardhana in https://github.com/klauspost/compress/pull/722
+	* s2: Add example for indexing and existing stream https://github.com/klauspost/compress/pull/723
+
+* Dec 11, 2022 (v1.15.13)
+	* zstd: Add [MaxEncodedSize](https://pkg.go.dev/github.com/klauspost/compress@v1.15.13/zstd#Encoder.MaxEncodedSize) to encoder  https://github.com/klauspost/compress/pull/691
+	* zstd: Various tweaks and improvements https://github.com/klauspost/compress/pull/693 https://github.com/klauspost/compress/pull/695 https://github.com/klauspost/compress/pull/696 https://github.com/klauspost/compress/pull/701 https://github.com/klauspost/compress/pull/702 https://github.com/klauspost/compress/pull/703 https://github.com/klauspost/compress/pull/704 https://github.com/klauspost/compress/pull/705 https://github.com/klauspost/compress/pull/706 https://github.com/klauspost/compress/pull/707 https://github.com/klauspost/compress/pull/708
+
+* Oct 26, 2022 (v1.15.12)
+
+	* zstd: Tweak decoder allocs. https://github.com/klauspost/compress/pull/680
+	* gzhttp: Always delete `HeaderNoCompression` https://github.com/klauspost/compress/pull/683
+
+* Sept 26, 2022 (v1.15.11)
+
+	* flate: Improve level 1-3 compression  https://github.com/klauspost/compress/pull/678
+	* zstd: Improve "best" compression by @nightwolfz in https://github.com/klauspost/compress/pull/677
+	* zstd: Fix+reduce decompression allocations https://github.com/klauspost/compress/pull/668
+	* zstd: Fix non-effective noescape tag https://github.com/klauspost/compress/pull/667
+
+* Sept 16, 2022 (v1.15.10)
+
+	* zstd: Add [WithDecodeAllCapLimit](https://pkg.go.dev/github.com/klauspost/compress@v1.15.10/zstd#WithDecodeAllCapLimit) https://github.com/klauspost/compress/pull/649
+	* Add Go 1.19 - deprecate Go 1.16  https://github.com/klauspost/compress/pull/651
+	* flate: Improve level 5+6 compression https://github.com/klauspost/compress/pull/656
+	* zstd: Improve "better" compression  https://github.com/klauspost/compress/pull/657
+	* s2: Improve "best" compression https://github.com/klauspost/compress/pull/658
+	* s2: Improve "better" compression. https://github.com/klauspost/compress/pull/635
+	* s2: Slightly faster non-assembly decompression https://github.com/klauspost/compress/pull/646
+	* Use arrays for constant size copies https://github.com/klauspost/compress/pull/659
+
+* July 21, 2022 (v1.15.9)
+
+	* zstd: Fix decoder crash on amd64 (no BMI) on invalid input https://github.com/klauspost/compress/pull/645
+	* zstd: Disable decoder extended memory copies (amd64) due to possible crashes https://github.com/klauspost/compress/pull/644
+	* zstd: Allow single segments up to "max decoded size" https://github.com/klauspost/compress/pull/643
+
+* July 13, 2022 (v1.15.8)
+
+	* gzip: fix stack exhaustion bug in Reader.Read https://github.com/klauspost/compress/pull/641
+	* s2: Add Index header trim/restore https://github.com/klauspost/compress/pull/638
+	* zstd: Optimize seqdeq amd64 asm by @greatroar in https://github.com/klauspost/compress/pull/636
+	* zstd: Improve decoder memcopy https://github.com/klauspost/compress/pull/637
+	* huff0: Pass a single bitReader pointer to asm by @greatroar in https://github.com/klauspost/compress/pull/634
+	* zstd: Branchless getBits for amd64 w/o BMI2 by @greatroar in https://github.com/klauspost/compress/pull/640
+	* gzhttp: Remove header before writing https://github.com/klauspost/compress/pull/639
+
+* June 29, 2022 (v1.15.7)
+
+	* s2: Fix absolute forward seeks  https://github.com/klauspost/compress/pull/633
+	* zip: Merge upstream  https://github.com/klauspost/compress/pull/631
+	* zip: Re-add zip64 fix https://github.com/klauspost/compress/pull/624
+	* zstd: translate fseDecoder.buildDtable into asm by @WojciechMula in https://github.com/klauspost/compress/pull/598
+	* flate: Faster histograms  https://github.com/klauspost/compress/pull/620
+	* deflate: Use compound hcode  https://github.com/klauspost/compress/pull/622
+
+* June 3, 2022 (v1.15.6)
+	* s2: Improve coding for long, close matches https://github.com/klauspost/compress/pull/613
+	* s2c: Add Snappy/S2 stream recompression https://github.com/klauspost/compress/pull/611
+	* zstd: Always use configured block size https://github.com/klauspost/compress/pull/605
+	* zstd: Fix incorrect hash table placement for dict encoding in default https://github.com/klauspost/compress/pull/606
+	* zstd: Apply default config to ZipDecompressor without options https://github.com/klauspost/compress/pull/608
+	* gzhttp: Exclude more common archive formats https://github.com/klauspost/compress/pull/612
+	* s2: Add ReaderIgnoreCRC https://github.com/klauspost/compress/pull/609
+	* s2: Remove sanity load on index creation https://github.com/klauspost/compress/pull/607
+	* snappy: Use dedicated function for scoring https://github.com/klauspost/compress/pull/614
+	* s2c+s2d: Use official snappy framed extension https://github.com/klauspost/compress/pull/610
+
+* May 25, 2022 (v1.15.5)
+	* s2: Add concurrent stream decompression https://github.com/klauspost/compress/pull/602
+	* s2: Fix final emit oob read crash on amd64 https://github.com/klauspost/compress/pull/601
+	* huff0: asm implementation of Decompress1X by @WojciechMula https://github.com/klauspost/compress/pull/596
+	* zstd: Use 1 less goroutine for stream decoding https://github.com/klauspost/compress/pull/588
+	* zstd: Copy literal in 16 byte blocks when possible https://github.com/klauspost/compress/pull/592
+	* zstd: Speed up when WithDecoderLowmem(false) https://github.com/klauspost/compress/pull/599
+	* zstd: faster next state update in BMI2 version of decode by @WojciechMula in https://github.com/klauspost/compress/pull/593
+	* huff0: Do not check max size when reading table. https://github.com/klauspost/compress/pull/586
+	* flate: Inplace hashing for level 7-9 https://github.com/klauspost/compress/pull/590
+
+
+* May 11, 2022 (v1.15.4)
+	* huff0: decompress directly into output by @WojciechMula in [#577](https://github.com/klauspost/compress/pull/577)
+	* inflate: Keep dict on stack [#581](https://github.com/klauspost/compress/pull/581)
+	* zstd: Faster decoding memcopy in asm [#583](https://github.com/klauspost/compress/pull/583)
+	* zstd: Fix ignored crc [#580](https://github.com/klauspost/compress/pull/580)
+
+* May 5, 2022 (v1.15.3)
+	* zstd: Allow to ignore checksum checking by @WojciechMula [#572](https://github.com/klauspost/compress/pull/572)
+	* s2: Fix incorrect seek for io.SeekEnd in [#575](https://github.com/klauspost/compress/pull/575)
+
+* Apr 26, 2022 (v1.15.2)
+	* zstd: Add x86-64 assembly for decompression on streams and blocks. Contributed by [@WojciechMula](https://github.com/WojciechMula). Typically 2x faster.  [#528](https://github.com/klauspost/compress/pull/528) [#531](https://github.com/klauspost/compress/pull/531) [#545](https://github.com/klauspost/compress/pull/545) [#537](https://github.com/klauspost/compress/pull/537)
+	* zstd: Add options to ZipDecompressor and fixes [#539](https://github.com/klauspost/compress/pull/539)
+	* s2: Use sorted search for index [#555](https://github.com/klauspost/compress/pull/555)
+	* Minimum version is Go 1.16, added CI test on 1.18.
+
+* Mar 11, 2022 (v1.15.1)
+	* huff0: Add x86 assembly of Decode4X by @WojciechMula in [#512](https://github.com/klauspost/compress/pull/512)
+	* zstd: Reuse zip decoders in [#514](https://github.com/klauspost/compress/pull/514)
+	* zstd: Detect extra block data and report as corrupted in [#520](https://github.com/klauspost/compress/pull/520)
+	* zstd: Handle zero sized frame content size stricter in [#521](https://github.com/klauspost/compress/pull/521)
+	* zstd: Add stricter block size checks in [#523](https://github.com/klauspost/compress/pull/523)
+
+* Mar 3, 2022 (v1.15.0)
+	* zstd: Refactor decoder [#498](https://github.com/klauspost/compress/pull/498)
+	* zstd: Add stream encoding without goroutines [#505](https://github.com/klauspost/compress/pull/505)
+	* huff0: Prevent single blocks exceeding 16 bits by @klauspost in[#507](https://github.com/klauspost/compress/pull/507)
+	* flate: Inline literal emission [#509](https://github.com/klauspost/compress/pull/509)
+	* gzhttp: Add zstd to transport [#400](https://github.com/klauspost/compress/pull/400)
+	* gzhttp: Make content-type optional [#510](https://github.com/klauspost/compress/pull/510)
+
+Both compression and decompression now supports "synchronous" stream operations. This means that whenever "concurrency" is set to 1, they will operate without spawning goroutines.
+
+Stream decompression is now faster on asynchronous, since the goroutine allocation much more effectively splits the workload. On typical streams this will typically use 2 cores fully for decompression. When a stream has finished decoding no goroutines will be left over, so decoders can now safely be pooled and still be garbage collected.
+
+While the release has been extensively tested, it is recommended to testing when upgrading.
+
+</details>
+
+<details>
+	<summary>See changes to v1.14.x</summary>
+	
+* Feb 22, 2022 (v1.14.4)
+	* flate: Fix rare huffman only (-2) corruption. [#503](https://github.com/klauspost/compress/pull/503)
+	* zip: Update deprecated CreateHeaderRaw to correctly call CreateRaw by @saracen in [#502](https://github.com/klauspost/compress/pull/502)
+	* zip: don't read data descriptor early by @saracen in [#501](https://github.com/klauspost/compress/pull/501)  #501
+	* huff0: Use static decompression buffer up to 30% faster [#499](https://github.com/klauspost/compress/pull/499) [#500](https://github.com/klauspost/compress/pull/500)
+
+* Feb 17, 2022 (v1.14.3)
+	* flate: Improve fastest levels compression speed ~10% more throughput. [#482](https://github.com/klauspost/compress/pull/482) [#489](https://github.com/klauspost/compress/pull/489) [#490](https://github.com/klauspost/compress/pull/490) [#491](https://github.com/klauspost/compress/pull/491) [#494](https://github.com/klauspost/compress/pull/494)  [#478](https://github.com/klauspost/compress/pull/478)
+	* flate: Faster decompression speed, ~5-10%. [#483](https://github.com/klauspost/compress/pull/483)
+	* s2: Faster compression with Go v1.18 and amd64 microarch level 3+. [#484](https://github.com/klauspost/compress/pull/484) [#486](https://github.com/klauspost/compress/pull/486)
+
+* Jan 25, 2022 (v1.14.2)
+	* zstd: improve header decoder by @dsnet  [#476](https://github.com/klauspost/compress/pull/476)
+	* zstd: Add bigger default blocks  [#469](https://github.com/klauspost/compress/pull/469)
+	* zstd: Remove unused decompression buffer [#470](https://github.com/klauspost/compress/pull/470)
+	* zstd: Fix logically dead code by @ningmingxiao [#472](https://github.com/klauspost/compress/pull/472)
+	* flate: Improve level 7-9 [#471](https://github.com/klauspost/compress/pull/471) [#473](https://github.com/klauspost/compress/pull/473)
+	* zstd: Add noasm tag for xxhash [#475](https://github.com/klauspost/compress/pull/475)
+
+* Jan 11, 2022 (v1.14.1)
+	* s2: Add stream index in [#462](https://github.com/klauspost/compress/pull/462)
+	* flate: Speed and efficiency improvements in [#439](https://github.com/klauspost/compress/pull/439) [#461](https://github.com/klauspost/compress/pull/461) [#455](https://github.com/klauspost/compress/pull/455) [#452](https://github.com/klauspost/compress/pull/452) [#458](https://github.com/klauspost/compress/pull/458)
+	* zstd: Performance improvement in [#420]( https://github.com/klauspost/compress/pull/420) [#456](https://github.com/klauspost/compress/pull/456) [#437](https://github.com/klauspost/compress/pull/437) [#467](https://github.com/klauspost/compress/pull/467) [#468](https://github.com/klauspost/compress/pull/468)
+	* zstd: add arm64 xxhash assembly in [#464](https://github.com/klauspost/compress/pull/464)
+	* Add garbled for binaries for s2 in [#445](https://github.com/klauspost/compress/pull/445)
+</details>
+
+<details>
+	<summary>See changes to v1.13.x</summary>
+	
+* Aug 30, 2021 (v1.13.5)
+	* gz/zlib/flate: Alias stdlib errors [#425](https://github.com/klauspost/compress/pull/425)
+	* s2: Add block support to commandline tools [#413](https://github.com/klauspost/compress/pull/413)
+	* zstd: pooledZipWriter should return Writers to the same pool [#426](https://github.com/klauspost/compress/pull/426)
+	* Removed golang/snappy as external dependency for tests [#421](https://github.com/klauspost/compress/pull/421)
+
+* Aug 12, 2021 (v1.13.4)
+	* Add [snappy replacement package](https://github.com/klauspost/compress/tree/master/snappy).
+	* zstd: Fix incorrect encoding in "best" mode [#415](https://github.com/klauspost/compress/pull/415)
+
+* Aug 3, 2021 (v1.13.3) 
+	* zstd: Improve Best compression [#404](https://github.com/klauspost/compress/pull/404)
+	* zstd: Fix WriteTo error forwarding [#411](https://github.com/klauspost/compress/pull/411)
+	* gzhttp: Return http.HandlerFunc instead of http.Handler. Unlikely breaking change. [#406](https://github.com/klauspost/compress/pull/406)
+	* s2sx: Fix max size error [#399](https://github.com/klauspost/compress/pull/399)
+	* zstd: Add optional stream content size on reset [#401](https://github.com/klauspost/compress/pull/401)
+	* zstd: use SpeedBestCompression for level >= 10 [#410](https://github.com/klauspost/compress/pull/410)
+
+* Jun 14, 2021 (v1.13.1)
+	* s2: Add full Snappy output support  [#396](https://github.com/klauspost/compress/pull/396)
+	* zstd: Add configurable [Decoder window](https://pkg.go.dev/github.com/klauspost/compress/zstd#WithDecoderMaxWindow) size [#394](https://github.com/klauspost/compress/pull/394)
+	* gzhttp: Add header to skip compression  [#389](https://github.com/klauspost/compress/pull/389)
+	* s2: Improve speed with bigger output margin  [#395](https://github.com/klauspost/compress/pull/395)
+
+* Jun 3, 2021 (v1.13.0)
+	* Added [gzhttp](https://github.com/klauspost/compress/tree/master/gzhttp#gzip-handler) which allows wrapping HTTP servers and clients with GZIP compressors.
+	* zstd: Detect short invalid signatures [#382](https://github.com/klauspost/compress/pull/382)
+	* zstd: Spawn decoder goroutine only if needed. [#380](https://github.com/klauspost/compress/pull/380)
+</details>
+
+
+<details>
+	<summary>See changes to v1.12.x</summary>
+	
+* May 25, 2021 (v1.12.3)
+	* deflate: Better/faster Huffman encoding [#374](https://github.com/klauspost/compress/pull/374)
+	* deflate: Allocate less for history. [#375](https://github.com/klauspost/compress/pull/375)
+	* zstd: Forward read errors [#373](https://github.com/klauspost/compress/pull/373) 
+
+* Apr 27, 2021 (v1.12.2)
+	* zstd: Improve better/best compression [#360](https://github.com/klauspost/compress/pull/360) [#364](https://github.com/klauspost/compress/pull/364) [#365](https://github.com/klauspost/compress/pull/365)
+	* zstd: Add helpers to compress/decompress zstd inside zip files [#363](https://github.com/klauspost/compress/pull/363)
+	* deflate: Improve level 5+6 compression [#367](https://github.com/klauspost/compress/pull/367)
+	* s2: Improve better/best compression [#358](https://github.com/klauspost/compress/pull/358) [#359](https://github.com/klauspost/compress/pull/358)
+	* s2: Load after checking src limit on amd64. [#362](https://github.com/klauspost/compress/pull/362)
+	* s2sx: Limit max executable size [#368](https://github.com/klauspost/compress/pull/368) 
+
+* Apr 14, 2021 (v1.12.1)
+	* snappy package removed. Upstream added as dependency.
+	* s2: Better compression in "best" mode [#353](https://github.com/klauspost/compress/pull/353)
+	* s2sx: Add stdin input and detect pre-compressed from signature [#352](https://github.com/klauspost/compress/pull/352)
+	* s2c/s2d: Add http as possible input [#348](https://github.com/klauspost/compress/pull/348)
+	* s2c/s2d/s2sx: Always truncate when writing files [#352](https://github.com/klauspost/compress/pull/352)
+	* zstd: Reduce memory usage further when using [WithLowerEncoderMem](https://pkg.go.dev/github.com/klauspost/compress/zstd#WithLowerEncoderMem) [#346](https://github.com/klauspost/compress/pull/346)
+	* s2: Fix potential problem with amd64 assembly and profilers [#349](https://github.com/klauspost/compress/pull/349)
+</details>
+
+<details>
+	<summary>See changes to v1.11.x</summary>
+	
+* Mar 26, 2021 (v1.11.13)
+	* zstd: Big speedup on small dictionary encodes [#344](https://github.com/klauspost/compress/pull/344) [#345](https://github.com/klauspost/compress/pull/345)
+	* zstd: Add [WithLowerEncoderMem](https://pkg.go.dev/github.com/klauspost/compress/zstd#WithLowerEncoderMem) encoder option [#336](https://github.com/klauspost/compress/pull/336)
+	* deflate: Improve entropy compression [#338](https://github.com/klauspost/compress/pull/338)
+	* s2: Clean up and minor performance improvement in best [#341](https://github.com/klauspost/compress/pull/341)
+
+* Mar 5, 2021 (v1.11.12)
+	* s2: Add `s2sx` binary that creates [self extracting archives](https://github.com/klauspost/compress/tree/master/s2#s2sx-self-extracting-archives).
+	* s2: Speed up decompression on non-assembly platforms [#328](https://github.com/klauspost/compress/pull/328)
+
+* Mar 1, 2021 (v1.11.9)
+	* s2: Add ARM64 decompression assembly. Around 2x output speed. [#324](https://github.com/klauspost/compress/pull/324)
+	* s2: Improve "better" speed and efficiency. [#325](https://github.com/klauspost/compress/pull/325)
+	* s2: Fix binaries.
+
+* Feb 25, 2021 (v1.11.8)
+	* s2: Fixed occasional out-of-bounds write on amd64. Upgrade recommended.
+	* s2: Add AMD64 assembly for better mode. 25-50% faster. [#315](https://github.com/klauspost/compress/pull/315)
+	* s2: Less upfront decoder allocation. [#322](https://github.com/klauspost/compress/pull/322)
+	* zstd: Faster "compression" of incompressible data. [#314](https://github.com/klauspost/compress/pull/314)
+	* zip: Fix zip64 headers. [#313](https://github.com/klauspost/compress/pull/313)
+  
+* Jan 14, 2021 (v1.11.7)
+	* Use Bytes() interface to get bytes across packages. [#309](https://github.com/klauspost/compress/pull/309)
+	* s2: Add 'best' compression option.  [#310](https://github.com/klauspost/compress/pull/310)
+	* s2: Add ReaderMaxBlockSize, changes `s2.NewReader` signature to include varargs. [#311](https://github.com/klauspost/compress/pull/311)
+	* s2: Fix crash on small better buffers. [#308](https://github.com/klauspost/compress/pull/308)
+	* s2: Clean up decoder. [#312](https://github.com/klauspost/compress/pull/312)
+
+* Jan 7, 2021 (v1.11.6)
+	* zstd: Make decoder allocations smaller [#306](https://github.com/klauspost/compress/pull/306)
+	* zstd: Free Decoder resources when Reset is called with a nil io.Reader  [#305](https://github.com/klauspost/compress/pull/305)
+
+* Dec 20, 2020 (v1.11.4)
+	* zstd: Add Best compression mode [#304](https://github.com/klauspost/compress/pull/304)
+	* Add header decoder [#299](https://github.com/klauspost/compress/pull/299)
+	* s2: Add uncompressed stream option [#297](https://github.com/klauspost/compress/pull/297)
+	* Simplify/speed up small blocks with known max size. [#300](https://github.com/klauspost/compress/pull/300)
+	* zstd: Always reset literal dict encoder [#303](https://github.com/klauspost/compress/pull/303)
+
+* Nov 15, 2020 (v1.11.3)
+	* inflate: 10-15% faster decompression  [#293](https://github.com/klauspost/compress/pull/293)
+	* zstd: Tweak DecodeAll default allocation [#295](https://github.com/klauspost/compress/pull/295)
+
+* Oct 11, 2020 (v1.11.2)
+	* s2: Fix out of bounds read in "better" block compression [#291](https://github.com/klauspost/compress/pull/291)
+
+* Oct 1, 2020 (v1.11.1)
+	* zstd: Set allLitEntropy true in default configuration [#286](https://github.com/klauspost/compress/pull/286)
+
+* Sept 8, 2020 (v1.11.0)
+	* zstd: Add experimental compression [dictionaries](https://github.com/klauspost/compress/tree/master/zstd#dictionaries) [#281](https://github.com/klauspost/compress/pull/281)
+	* zstd: Fix mixed Write and ReadFrom calls [#282](https://github.com/klauspost/compress/pull/282)
+	* inflate/gz: Limit variable shifts, ~5% faster decompression [#274](https://github.com/klauspost/compress/pull/274)
+</details>
+
+<details>
+	<summary>See changes to v1.10.x</summary>
+ 
+* July 8, 2020 (v1.10.11) 
+	* zstd: Fix extra block when compressing with ReadFrom. [#278](https://github.com/klauspost/compress/pull/278)
+	* huff0: Also populate compression table when reading decoding table. [#275](https://github.com/klauspost/compress/pull/275)
+	
+* June 23, 2020 (v1.10.10) 
+	* zstd: Skip entropy compression in fastest mode when no matches. [#270](https://github.com/klauspost/compress/pull/270)
+	
+* June 16, 2020 (v1.10.9): 
+	* zstd: API change for specifying dictionaries. See [#268](https://github.com/klauspost/compress/pull/268)
+	* zip: update CreateHeaderRaw to handle zip64 fields. [#266](https://github.com/klauspost/compress/pull/266)
+	* Fuzzit tests removed. The service has been purchased and is no longer available.
+	
+* June 5, 2020 (v1.10.8): 
+	* 1.15x faster zstd block decompression. [#265](https://github.com/klauspost/compress/pull/265)
+	
+* June 1, 2020 (v1.10.7): 
+	* Added zstd decompression [dictionary support](https://github.com/klauspost/compress/tree/master/zstd#dictionaries)
+	* Increase zstd decompression speed up to 1.19x.  [#259](https://github.com/klauspost/compress/pull/259)
+	* Remove internal reset call in zstd compression and reduce allocations. [#263](https://github.com/klauspost/compress/pull/263)
+	
+* May 21, 2020: (v1.10.6) 
+	* zstd: Reduce allocations while decoding. [#258](https://github.com/klauspost/compress/pull/258), [#252](https://github.com/klauspost/compress/pull/252)
+	* zstd: Stricter decompression checks.
+	
+* April 12, 2020: (v1.10.5)
+	* s2-commands: Flush output when receiving SIGINT. [#239](https://github.com/klauspost/compress/pull/239)
+	
+* Apr 8, 2020: (v1.10.4) 
+	* zstd: Minor/special case optimizations. [#251](https://github.com/klauspost/compress/pull/251),  [#250](https://github.com/klauspost/compress/pull/250),  [#249](https://github.com/klauspost/compress/pull/249),  [#247](https://github.com/klauspost/compress/pull/247)
+* Mar 11, 2020: (v1.10.3) 
+	* s2: Use S2 encoder in pure Go mode for Snappy output as well. [#245](https://github.com/klauspost/compress/pull/245)
+	* s2: Fix pure Go block encoder. [#244](https://github.com/klauspost/compress/pull/244)
+	* zstd: Added "better compression" mode. [#240](https://github.com/klauspost/compress/pull/240)
+	* zstd: Improve speed of fastest compression mode by 5-10% [#241](https://github.com/klauspost/compress/pull/241)
+	* zstd: Skip creating encoders when not needed. [#238](https://github.com/klauspost/compress/pull/238)
+	
+* Feb 27, 2020: (v1.10.2) 
+	* Close to 50% speedup in inflate (gzip/zip decompression). [#236](https://github.com/klauspost/compress/pull/236) [#234](https://github.com/klauspost/compress/pull/234) [#232](https://github.com/klauspost/compress/pull/232)
+	* Reduce deflate level 1-6 memory usage up to 59%. [#227](https://github.com/klauspost/compress/pull/227)
+	
+* Feb 18, 2020: (v1.10.1)
+	* Fix zstd crash when resetting multiple times without sending data. [#226](https://github.com/klauspost/compress/pull/226)
+	* deflate: Fix dictionary use on level 1-6. [#224](https://github.com/klauspost/compress/pull/224)
+	* Remove deflate writer reference when closing. [#224](https://github.com/klauspost/compress/pull/224)
+	
+* Feb 4, 2020: (v1.10.0) 
+	* Add optional dictionary to [stateless deflate](https://pkg.go.dev/github.com/klauspost/compress/flate?tab=doc#StatelessDeflate). Breaking change, send `nil` for previous behaviour. [#216](https://github.com/klauspost/compress/pull/216)
+	* Fix buffer overflow on repeated small block deflate.  [#218](https://github.com/klauspost/compress/pull/218)
+	* Allow copying content from an existing ZIP file without decompressing+compressing. [#214](https://github.com/klauspost/compress/pull/214)
+	* Added [S2](https://github.com/klauspost/compress/tree/master/s2#s2-compression) AMD64 assembler and various optimizations. Stream speed >10GB/s.  [#186](https://github.com/klauspost/compress/pull/186)
+
+</details>
+
+<details>
+	<summary>See changes prior to v1.10.0</summary>
+
+* Jan 20,2020 (v1.9.8) Optimize gzip/deflate with better size estimates and faster table generation. [#207](https://github.com/klauspost/compress/pull/207) by [luyu6056](https://github.com/luyu6056),  [#206](https://github.com/klauspost/compress/pull/206).
+* Jan 11, 2020: S2 Encode/Decode will use provided buffer if capacity is big enough. [#204](https://github.com/klauspost/compress/pull/204) 
+* Jan 5, 2020: (v1.9.7) Fix another zstd regression in v1.9.5 - v1.9.6 removed.
+* Jan 4, 2020: (v1.9.6) Regression in v1.9.5 fixed causing corrupt zstd encodes in rare cases.
+* Jan 4, 2020: Faster IO in [s2c + s2d commandline tools](https://github.com/klauspost/compress/tree/master/s2#commandline-tools) compression/decompression. [#192](https://github.com/klauspost/compress/pull/192)
+* Dec 29, 2019: Removed v1.9.5 since fuzz tests showed a compatibility problem with the reference zstandard decoder.
+* Dec 29, 2019: (v1.9.5) zstd: 10-20% faster block compression. [#199](https://github.com/klauspost/compress/pull/199)
+* Dec 29, 2019: [zip](https://godoc.org/github.com/klauspost/compress/zip) package updated with latest Go features
+* Dec 29, 2019: zstd: Single segment flag condintions tweaked. [#197](https://github.com/klauspost/compress/pull/197)
+* Dec 18, 2019: s2: Faster compression when ReadFrom is used. [#198](https://github.com/klauspost/compress/pull/198)
+* Dec 10, 2019: s2: Fix repeat length output when just above at 16MB limit.
+* Dec 10, 2019: zstd: Add function to get decoder as io.ReadCloser. [#191](https://github.com/klauspost/compress/pull/191)
+* Dec 3, 2019: (v1.9.4) S2: limit max repeat length. [#188](https://github.com/klauspost/compress/pull/188)
+* Dec 3, 2019: Add [WithNoEntropyCompression](https://godoc.org/github.com/klauspost/compress/zstd#WithNoEntropyCompression) to zstd [#187](https://github.com/klauspost/compress/pull/187)
+* Dec 3, 2019: Reduce memory use for tests. Check for leaked goroutines.
+* Nov 28, 2019 (v1.9.3) Less allocations in stateless deflate.
+* Nov 28, 2019: 5-20% Faster huff0 decode. Impacts zstd as well. [#184](https://github.com/klauspost/compress/pull/184)
+* Nov 12, 2019 (v1.9.2) Added [Stateless Compression](#stateless-compression) for gzip/deflate.
+* Nov 12, 2019: Fixed zstd decompression of large single blocks. [#180](https://github.com/klauspost/compress/pull/180)
+* Nov 11, 2019: Set default  [s2c](https://github.com/klauspost/compress/tree/master/s2#commandline-tools) block size to 4MB.
+* Nov 11, 2019: Reduce inflate memory use by 1KB.
+* Nov 10, 2019: Less allocations in deflate bit writer.
+* Nov 10, 2019: Fix inconsistent error returned by zstd decoder.
+* Oct 28, 2019 (v1.9.1) ztsd: Fix crash when compressing blocks. [#174](https://github.com/klauspost/compress/pull/174)
+* Oct 24, 2019 (v1.9.0) zstd: Fix rare data corruption [#173](https://github.com/klauspost/compress/pull/173)
+* Oct 24, 2019 zstd: Fix huff0 out of buffer write [#171](https://github.com/klauspost/compress/pull/171) and always return errors [#172](https://github.com/klauspost/compress/pull/172) 
+* Oct 10, 2019: Big deflate rewrite, 30-40% faster with better compression [#105](https://github.com/klauspost/compress/pull/105)
+
+</details>
+
+<details>
+	<summary>See changes prior to v1.9.0</summary>
+
+* Oct 10, 2019: (v1.8.6) zstd: Allow partial reads to get flushed data. [#169](https://github.com/klauspost/compress/pull/169)
+* Oct 3, 2019: Fix inconsistent results on broken zstd streams.
+* Sep 25, 2019: Added `-rm` (remove source files) and `-q` (no output except errors) to `s2c` and `s2d` [commands](https://github.com/klauspost/compress/tree/master/s2#commandline-tools)
+* Sep 16, 2019: (v1.8.4) Add `s2c` and `s2d` [commandline tools](https://github.com/klauspost/compress/tree/master/s2#commandline-tools).
+* Sep 10, 2019: (v1.8.3) Fix s2 decoder [Skip](https://godoc.org/github.com/klauspost/compress/s2#Reader.Skip).
+* Sep 7, 2019: zstd: Added [WithWindowSize](https://godoc.org/github.com/klauspost/compress/zstd#WithWindowSize), contributed by [ianwilkes](https://github.com/ianwilkes).
+* Sep 5, 2019: (v1.8.2) Add [WithZeroFrames](https://godoc.org/github.com/klauspost/compress/zstd#WithZeroFrames) which adds full zero payload block encoding option.
+* Sep 5, 2019: Lazy initialization of zstandard predefined en/decoder tables.
+* Aug 26, 2019: (v1.8.1) S2: 1-2% compression increase in "better" compression mode.
+* Aug 26, 2019: zstd: Check maximum size of Huffman 1X compressed literals while decoding.
+* Aug 24, 2019: (v1.8.0) Added [S2 compression](https://github.com/klauspost/compress/tree/master/s2#s2-compression), a high performance replacement for Snappy. 
+* Aug 21, 2019: (v1.7.6) Fixed minor issues found by fuzzer. One could lead to zstd not decompressing.
+* Aug 18, 2019: Add [fuzzit](https://fuzzit.dev/) continuous fuzzing.
+* Aug 14, 2019: zstd: Skip incompressible data 2x faster.  [#147](https://github.com/klauspost/compress/pull/147)
+* Aug 4, 2019 (v1.7.5): Better literal compression. [#146](https://github.com/klauspost/compress/pull/146)
+* Aug 4, 2019: Faster zstd compression. [#143](https://github.com/klauspost/compress/pull/143) [#144](https://github.com/klauspost/compress/pull/144)
+* Aug 4, 2019: Faster zstd decompression. [#145](https://github.com/klauspost/compress/pull/145) [#143](https://github.com/klauspost/compress/pull/143) [#142](https://github.com/klauspost/compress/pull/142)
+* July 15, 2019 (v1.7.4): Fix double EOF block in rare cases on zstd encoder.
+* July 15, 2019 (v1.7.3): Minor speedup/compression increase in default zstd encoder.
+* July 14, 2019: zstd decoder: Fix decompression error on multiple uses with mixed content.
+* July 7, 2019 (v1.7.2): Snappy update, zstd decoder potential race fix.
+* June 17, 2019: zstd decompression bugfix.
+* June 17, 2019: fix 32 bit builds.
+* June 17, 2019: Easier use in modules (less dependencies).
+* June 9, 2019: New stronger "default" [zstd](https://github.com/klauspost/compress/tree/master/zstd#zstd) compression mode. Matches zstd default compression ratio.
+* June 5, 2019: 20-40% throughput in [zstandard](https://github.com/klauspost/compress/tree/master/zstd#zstd) compression and better compression.
+* June 5, 2019: deflate/gzip compression: Reduce memory usage of lower compression levels.
+* June 2, 2019: Added [zstandard](https://github.com/klauspost/compress/tree/master/zstd#zstd) compression!
+* May 25, 2019: deflate/gzip: 10% faster bit writer, mostly visible in lower levels.
+* Apr 22, 2019: [zstd](https://github.com/klauspost/compress/tree/master/zstd#zstd) decompression added.
+* Aug 1, 2018: Added [huff0 README](https://github.com/klauspost/compress/tree/master/huff0#huff0-entropy-compression).
+* Jul 8, 2018: Added [Performance Update 2018](#performance-update-2018) below.
+* Jun 23, 2018: Merged [Go 1.11 inflate optimizations](https://go-review.googlesource.com/c/go/+/102235). Go 1.9 is now required. Backwards compatible version tagged with [v1.3.0](https://github.com/klauspost/compress/releases/tag/v1.3.0).
+* Apr 2, 2018: Added [huff0](https://godoc.org/github.com/klauspost/compress/huff0) en/decoder. Experimental for now, API may change.
+* Mar 4, 2018: Added [FSE Entropy](https://godoc.org/github.com/klauspost/compress/fse) en/decoder. Experimental for now, API may change.
+* Nov 3, 2017: Add compression [Estimate](https://godoc.org/github.com/klauspost/compress#Estimate) function.
+* May 28, 2017: Reduce allocations when resetting decoder.
+* Apr 02, 2017: Change back to official crc32, since changes were merged in Go 1.7.
+* Jan 14, 2017: Reduce stack pressure due to array copies. See [Issue #18625](https://github.com/golang/go/issues/18625).
+* Oct 25, 2016: Level 2-4 have been rewritten and now offers significantly better performance than before.
+* Oct 20, 2016: Port zlib changes from Go 1.7 to fix zlib writer issue. Please update.
+* Oct 16, 2016: Go 1.7 changes merged. Apples to apples this package is a few percent faster, but has a significantly better balance between speed and compression per level. 
+* Mar 24, 2016: Always attempt Huffman encoding on level 4-7. This improves base 64 encoded data compression.
+* Mar 24, 2016: Small speedup for level 1-3.
+* Feb 19, 2016: Faster bit writer, level -2 is 15% faster, level 1 is 4% faster.
+* Feb 19, 2016: Handle small payloads faster in level 1-3.
+* Feb 19, 2016: Added faster level 2 + 3 compression modes.
+* Feb 19, 2016: [Rebalanced compression levels](https://blog.klauspost.com/rebalancing-deflate-compression-levels/), so there is a more even progression in terms of compression. New default level is 5.
+* Feb 14, 2016: Snappy: Merge upstream changes. 
+* Feb 14, 2016: Snappy: Fix aggressive skipping.
+* Feb 14, 2016: Snappy: Update benchmark.
+* Feb 13, 2016: Deflate: Fixed assembler problem that could lead to sub-optimal compression.
+* Feb 12, 2016: Snappy: Added AMD64 SSE 4.2 optimizations to matching, which makes easy to compress material run faster. Typical speedup is around 25%.
+* Feb 9, 2016: Added Snappy package fork. This version is 5-7% faster, much more on hard to compress content.
+* Jan 30, 2016: Optimize level 1 to 3 by not considering static dictionary or storing uncompressed. ~4-5% speedup.
+* Jan 16, 2016: Optimization on deflate level 1,2,3 compression.
+* Jan 8 2016: Merge [CL 18317](https://go-review.googlesource.com/#/c/18317): fix reading, writing of zip64 archives.
+* Dec 8 2015: Make level 1 and -2 deterministic even if write size differs.
+* Dec 8 2015: Split encoding functions, so hashing and matching can potentially be inlined. 1-3% faster on AMD64. 5% faster on other platforms.
+* Dec 8 2015: Fixed rare [one byte out-of bounds read](https://github.com/klauspost/compress/issues/20). Please update!
+* Nov 23 2015: Optimization on token writer. ~2-4% faster. Contributed by [@dsnet](https://github.com/dsnet).
+* Nov 20 2015: Small optimization to bit writer on 64 bit systems.
+* Nov 17 2015: Fixed out-of-bound errors if the underlying Writer returned an error. See [#15](https://github.com/klauspost/compress/issues/15).
+* Nov 12 2015: Added [io.WriterTo](https://golang.org/pkg/io/#WriterTo) support to gzip/inflate.
+* Nov 11 2015: Merged [CL 16669](https://go-review.googlesource.com/#/c/16669/4): archive/zip: enable overriding (de)compressors per file
+* Oct 15 2015: Added skipping on uncompressible data. Random data speed up >5x.
+
+</details>
+
+# deflate usage
+
+The packages are drop-in replacements for standard libraries. Simply replace the import path to use them:
+
+Typical speed is about 2x of the standard library packages.
+
+| old import       | new import                            | Documentation                                                           |
+|------------------|---------------------------------------|-------------------------------------------------------------------------|
+| `compress/gzip`  | `github.com/klauspost/compress/gzip`  | [gzip](https://pkg.go.dev/github.com/klauspost/compress/gzip?tab=doc)   |
+| `compress/zlib`  | `github.com/klauspost/compress/zlib`  | [zlib](https://pkg.go.dev/github.com/klauspost/compress/zlib?tab=doc)   |
+| `archive/zip`    | `github.com/klauspost/compress/zip`   | [zip](https://pkg.go.dev/github.com/klauspost/compress/zip?tab=doc)     |
+| `compress/flate` | `github.com/klauspost/compress/flate` | [flate](https://pkg.go.dev/github.com/klauspost/compress/flate?tab=doc) |
+
+* Optimized [deflate](https://godoc.org/github.com/klauspost/compress/flate) packages which can be used as a dropin replacement for [gzip](https://godoc.org/github.com/klauspost/compress/gzip), [zip](https://godoc.org/github.com/klauspost/compress/zip) and [zlib](https://godoc.org/github.com/klauspost/compress/zlib).
+
+You may also be interested in [pgzip](https://github.com/klauspost/pgzip), which is a drop in replacement for gzip, which support multithreaded compression on big files and the optimized [crc32](https://github.com/klauspost/crc32) package used by these packages.
+
+The packages contains the same as the standard library, so you can use the godoc for that: [gzip](http://golang.org/pkg/compress/gzip/), [zip](http://golang.org/pkg/archive/zip/),  [zlib](http://golang.org/pkg/compress/zlib/), [flate](http://golang.org/pkg/compress/flate/).
+
+Currently there is only minor speedup on decompression (mostly CRC32 calculation).
+
+Memory usage is typically 1MB for a Writer. stdlib is in the same range. 
+If you expect to have a lot of concurrently allocated Writers consider using 
+the stateless compress described below.
+
+For compression performance, see: [this spreadsheet](https://docs.google.com/spreadsheets/d/1nuNE2nPfuINCZJRMt6wFWhKpToF95I47XjSsc-1rbPQ/edit?usp=sharing).
+
+To disable all assembly add `-tags=noasm`. This works across all packages.
+
+# Stateless compression
+
+This package offers stateless compression as a special option for gzip/deflate. 
+It will do compression but without maintaining any state between Write calls.
+
+This means there will be no memory kept between Write calls, but compression and speed will be suboptimal.
+
+This is only relevant in cases where you expect to run many thousands of compressors concurrently, 
+but with very little activity. This is *not* intended for regular web servers serving individual requests.  
+
+Because of this, the size of actual Write calls will affect output size.
+
+In gzip, specify level `-3` / `gzip.StatelessCompression` to enable.
+
+For direct deflate use, NewStatelessWriter and StatelessDeflate are available. See [documentation](https://godoc.org/github.com/klauspost/compress/flate#NewStatelessWriter)
+
+A `bufio.Writer` can of course be used to control write sizes. For example, to use a 4KB buffer:
+
+```go
+	// replace 'ioutil.Discard' with your output.
+	gzw, err := gzip.NewWriterLevel(ioutil.Discard, gzip.StatelessCompression)
+	if err != nil {
+		return err
+	}
+	defer gzw.Close()
+
+	w := bufio.NewWriterSize(gzw, 4096)
+	defer w.Flush()
+	
+	// Write to 'w' 
+```
+
+This will only use up to 4KB in memory when the writer is idle. 
+
+Compression is almost always worse than the fastest compression level 
+and each write will allocate (a little) memory. 
+
+
+# Other packages
+
+Here are other packages of good quality and pure Go (no cgo wrappers or autoconverted code):
+
+* [github.com/pierrec/lz4](https://github.com/pierrec/lz4) - strong multithreaded LZ4 compression.
+* [github.com/cosnicolaou/pbzip2](https://github.com/cosnicolaou/pbzip2) - multithreaded bzip2 decompression.
+* [github.com/dsnet/compress](https://github.com/dsnet/compress) - brotli decompression, bzip2 writer.
+* [github.com/ronanh/intcomp](https://github.com/ronanh/intcomp) - Integer compression.
+* [github.com/spenczar/fpc](https://github.com/spenczar/fpc) - Float compression.
+* [github.com/minio/zipindex](https://github.com/minio/zipindex) - External ZIP directory index.
+* [github.com/ybirader/pzip](https://github.com/ybirader/pzip) - Fast concurrent zip archiver and extractor.
+
+# license
+
+This code is licensed under the same conditions as the original Go code. See LICENSE file.
diff --git a/vendor/github.com/klauspost/compress/SECURITY.md b/vendor/github.com/klauspost/compress/SECURITY.md
new file mode 100644
index 0000000..ca6685e
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/SECURITY.md
@@ -0,0 +1,25 @@
+# Security Policy
+
+## Supported Versions
+
+Security updates are applied only to the latest release.
+
+## Vulnerability Definition
+
+A security vulnerability is a bug that with certain input triggers a crash or an infinite loop. Most calls will have varying execution time and only in rare cases will slow operation be considered a security vulnerability.
+
+Corrupted output generally is not considered a security vulnerability, unless independent operations are able to affect each other. Note that not all functionality is re-entrant and safe to use concurrently.
+
+Out-of-memory crashes only applies if the en/decoder uses an abnormal amount of memory, with appropriate options applied, to limit maximum window size, concurrency, etc. However, if you are in doubt you are welcome to file a security issue.
+
+It is assumed that all callers are trusted, meaning internal data exposed through reflection or inspection of returned data structures is not considered a vulnerability.
+
+Vulnerabilities resulting from compiler/assembler errors should be reported upstream. Depending on the severity this package may or may not implement a workaround.
+
+## Reporting a Vulnerability
+
+If you have discovered a security vulnerability in this project, please report it privately. **Do not disclose it as a public issue.** This gives us time to work with you to fix the issue before public exposure, reducing the chance that the exploit will be used before a patch is released.
+
+Please disclose it at [security advisory](https://github.com/klauspost/compress/security/advisories/new). If possible please provide a minimal reproducer. If the issue only applies to a single platform, it would be helpful to provide access to that.
+
+This project is maintained by a team of volunteers on a reasonable-effort basis. As such, vulnerabilities will be disclosed in a best effort base.
diff --git a/vendor/github.com/klauspost/compress/compressible.go b/vendor/github.com/klauspost/compress/compressible.go
new file mode 100644
index 0000000..ea5a692
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/compressible.go
@@ -0,0 +1,85 @@
+package compress
+
+import "math"
+
+// Estimate returns a normalized compressibility estimate of block b.
+// Values close to zero are likely uncompressible.
+// Values above 0.1 are likely to be compressible.
+// Values above 0.5 are very compressible.
+// Very small lengths will return 0.
+func Estimate(b []byte) float64 {
+	if len(b) < 16 {
+		return 0
+	}
+
+	// Correctly predicted order 1
+	hits := 0
+	lastMatch := false
+	var o1 [256]byte
+	var hist [256]int
+	c1 := byte(0)
+	for _, c := range b {
+		if c == o1[c1] {
+			// We only count a hit if there was two correct predictions in a row.
+			if lastMatch {
+				hits++
+			}
+			lastMatch = true
+		} else {
+			lastMatch = false
+		}
+		o1[c1] = c
+		c1 = c
+		hist[c]++
+	}
+
+	// Use x^0.6 to give better spread
+	prediction := math.Pow(float64(hits)/float64(len(b)), 0.6)
+
+	// Calculate histogram distribution
+	variance := float64(0)
+	avg := float64(len(b)) / 256
+
+	for _, v := range hist {
+		Δ := float64(v) - avg
+		variance += Δ * Δ
+	}
+
+	stddev := math.Sqrt(float64(variance)) / float64(len(b))
+	exp := math.Sqrt(1 / float64(len(b)))
+
+	// Subtract expected stddev
+	stddev -= exp
+	if stddev < 0 {
+		stddev = 0
+	}
+	stddev *= 1 + exp
+
+	// Use x^0.4 to give better spread
+	entropy := math.Pow(stddev, 0.4)
+
+	// 50/50 weight between prediction and histogram distribution
+	return math.Pow((prediction+entropy)/2, 0.9)
+}
+
+// ShannonEntropyBits returns the number of bits minimum required to represent
+// an entropy encoding of the input bytes.
+// https://en.wiktionary.org/wiki/Shannon_entropy
+func ShannonEntropyBits(b []byte) int {
+	if len(b) == 0 {
+		return 0
+	}
+	var hist [256]int
+	for _, c := range b {
+		hist[c]++
+	}
+	shannon := float64(0)
+	invTotal := 1.0 / float64(len(b))
+	for _, v := range hist[:] {
+		if v > 0 {
+			n := float64(v)
+			shannon += math.Ceil(-math.Log2(n*invTotal) * n)
+		}
+	}
+	return int(math.Ceil(shannon))
+}
diff --git a/vendor/github.com/klauspost/compress/flate/deflate.go b/vendor/github.com/klauspost/compress/flate/deflate.go
new file mode 100644
index 0000000..af53fb8
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/deflate.go
@@ -0,0 +1,1017 @@
+// Copyright 2009 The Go Authors. All rights reserved.
+// Copyright (c) 2015 Klaus Post
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package flate
+
+import (
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+	"math"
+)
+
+const (
+	NoCompression      = 0
+	BestSpeed          = 1
+	BestCompression    = 9
+	DefaultCompression = -1
+
+	// HuffmanOnly disables Lempel-Ziv match searching and only performs Huffman
+	// entropy encoding. This mode is useful in compressing data that has
+	// already been compressed with an LZ style algorithm (e.g. Snappy or LZ4)
+	// that lacks an entropy encoder. Compression gains are achieved when
+	// certain bytes in the input stream occur more frequently than others.
+	//
+	// Note that HuffmanOnly produces a compressed output that is
+	// RFC 1951 compliant. That is, any valid DEFLATE decompressor will
+	// continue to be able to decompress this output.
+	HuffmanOnly         = -2
+	ConstantCompression = HuffmanOnly // compatibility alias.
+
+	logWindowSize    = 15
+	windowSize       = 1 << logWindowSize
+	windowMask       = windowSize - 1
+	logMaxOffsetSize = 15  // Standard DEFLATE
+	minMatchLength   = 4   // The smallest match that the compressor looks for
+	maxMatchLength   = 258 // The longest match for the compressor
+	minOffsetSize    = 1   // The shortest offset that makes any sense
+
+	// The maximum number of tokens we will encode at the time.
+	// Smaller sizes usually creates less optimal blocks.
+	// Bigger can make context switching slow.
+	// We use this for levels 7-9, so we make it big.
+	maxFlateBlockTokens = 1 << 15
+	maxStoreBlockSize   = 65535
+	hashBits            = 17 // After 17 performance degrades
+	hashSize            = 1 << hashBits
+	hashMask            = (1 << hashBits) - 1
+	hashShift           = (hashBits + minMatchLength - 1) / minMatchLength
+	maxHashOffset       = 1 << 28
+
+	skipNever = math.MaxInt32
+
+	debugDeflate = false
+)
+
+type compressionLevel struct {
+	good, lazy, nice, chain, fastSkipHashing, level int
+}
+
+// Compression levels have been rebalanced from zlib deflate defaults
+// to give a bigger spread in speed and compression.
+// See https://blog.klauspost.com/rebalancing-deflate-compression-levels/
+var levels = []compressionLevel{
+	{}, // 0
+	// Level 1-6 uses specialized algorithm - values not used
+	{0, 0, 0, 0, 0, 1},
+	{0, 0, 0, 0, 0, 2},
+	{0, 0, 0, 0, 0, 3},
+	{0, 0, 0, 0, 0, 4},
+	{0, 0, 0, 0, 0, 5},
+	{0, 0, 0, 0, 0, 6},
+	// Levels 7-9 use increasingly more lazy matching
+	// and increasingly stringent conditions for "good enough".
+	{8, 12, 16, 24, skipNever, 7},
+	{16, 30, 40, 64, skipNever, 8},
+	{32, 258, 258, 1024, skipNever, 9},
+}
+
+// advancedState contains state for the advanced levels, with bigger hash tables, etc.
+type advancedState struct {
+	// deflate state
+	length         int
+	offset         int
+	maxInsertIndex int
+	chainHead      int
+	hashOffset     int
+
+	ii uint16 // position of last match, intended to overflow to reset.
+
+	// input window: unprocessed data is window[index:windowEnd]
+	index     int
+	hashMatch [maxMatchLength + minMatchLength]uint32
+
+	// Input hash chains
+	// hashHead[hashValue] contains the largest inputIndex with the specified hash value
+	// If hashHead[hashValue] is within the current window, then
+	// hashPrev[hashHead[hashValue] & windowMask] contains the previous index
+	// with the same hash value.
+	hashHead [hashSize]uint32
+	hashPrev [windowSize]uint32
+}
+
+type compressor struct {
+	compressionLevel
+
+	h *huffmanEncoder
+	w *huffmanBitWriter
+
+	// compression algorithm
+	fill func(*compressor, []byte) int // copy data to window
+	step func(*compressor)             // process window
+
+	window     []byte
+	windowEnd  int
+	blockStart int // window index where current tokens start
+	err        error
+
+	// queued output tokens
+	tokens tokens
+	fast   fastEnc
+	state  *advancedState
+
+	sync          bool // requesting flush
+	byteAvailable bool // if true, still need to process window[index-1].
+}
+
+func (d *compressor) fillDeflate(b []byte) int {
+	s := d.state
+	if s.index >= 2*windowSize-(minMatchLength+maxMatchLength) {
+		// shift the window by windowSize
+		//copy(d.window[:], d.window[windowSize:2*windowSize])
+		*(*[windowSize]byte)(d.window) = *(*[windowSize]byte)(d.window[windowSize:])
+		s.index -= windowSize
+		d.windowEnd -= windowSize
+		if d.blockStart >= windowSize {
+			d.blockStart -= windowSize
+		} else {
+			d.blockStart = math.MaxInt32
+		}
+		s.hashOffset += windowSize
+		if s.hashOffset > maxHashOffset {
+			delta := s.hashOffset - 1
+			s.hashOffset -= delta
+			s.chainHead -= delta
+			// Iterate over slices instead of arrays to avoid copying
+			// the entire table onto the stack (Issue #18625).
+			for i, v := range s.hashPrev[:] {
+				if int(v) > delta {
+					s.hashPrev[i] = uint32(int(v) - delta)
+				} else {
+					s.hashPrev[i] = 0
+				}
+			}
+			for i, v := range s.hashHead[:] {
+				if int(v) > delta {
+					s.hashHead[i] = uint32(int(v) - delta)
+				} else {
+					s.hashHead[i] = 0
+				}
+			}
+		}
+	}
+	n := copy(d.window[d.windowEnd:], b)
+	d.windowEnd += n
+	return n
+}
+
+func (d *compressor) writeBlock(tok *tokens, index int, eof bool) error {
+	if index > 0 || eof {
+		var window []byte
+		if d.blockStart <= index {
+			window = d.window[d.blockStart:index]
+		}
+		d.blockStart = index
+		//d.w.writeBlock(tok, eof, window)
+		d.w.writeBlockDynamic(tok, eof, window, d.sync)
+		return d.w.err
+	}
+	return nil
+}
+
+// writeBlockSkip writes the current block and uses the number of tokens
+// to determine if the block should be stored on no matches, or
+// only huffman encoded.
+func (d *compressor) writeBlockSkip(tok *tokens, index int, eof bool) error {
+	if index > 0 || eof {
+		if d.blockStart <= index {
+			window := d.window[d.blockStart:index]
+			// If we removed less than a 64th of all literals
+			// we huffman compress the block.
+			if int(tok.n) > len(window)-int(tok.n>>6) {
+				d.w.writeBlockHuff(eof, window, d.sync)
+			} else {
+				// Write a dynamic huffman block.
+				d.w.writeBlockDynamic(tok, eof, window, d.sync)
+			}
+		} else {
+			d.w.writeBlock(tok, eof, nil)
+		}
+		d.blockStart = index
+		return d.w.err
+	}
+	return nil
+}
+
+// fillWindow will fill the current window with the supplied
+// dictionary and calculate all hashes.
+// This is much faster than doing a full encode.
+// Should only be used after a start/reset.
+func (d *compressor) fillWindow(b []byte) {
+	// Do not fill window if we are in store-only or huffman mode.
+	if d.level <= 0 && d.level > -MinCustomWindowSize {
+		return
+	}
+	if d.fast != nil {
+		// encode the last data, but discard the result
+		if len(b) > maxMatchOffset {
+			b = b[len(b)-maxMatchOffset:]
+		}
+		d.fast.Encode(&d.tokens, b)
+		d.tokens.Reset()
+		return
+	}
+	s := d.state
+	// If we are given too much, cut it.
+	if len(b) > windowSize {
+		b = b[len(b)-windowSize:]
+	}
+	// Add all to window.
+	n := copy(d.window[d.windowEnd:], b)
+
+	// Calculate 256 hashes at the time (more L1 cache hits)
+	loops := (n + 256 - minMatchLength) / 256
+	for j := 0; j < loops; j++ {
+		startindex := j * 256
+		end := startindex + 256 + minMatchLength - 1
+		if end > n {
+			end = n
+		}
+		tocheck := d.window[startindex:end]
+		dstSize := len(tocheck) - minMatchLength + 1
+
+		if dstSize <= 0 {
+			continue
+		}
+
+		dst := s.hashMatch[:dstSize]
+		bulkHash4(tocheck, dst)
+		var newH uint32
+		for i, val := range dst {
+			di := i + startindex
+			newH = val & hashMask
+			// Get previous value with the same hash.
+			// Our chain should point to the previous value.
+			s.hashPrev[di&windowMask] = s.hashHead[newH]
+			// Set the head of the hash chain to us.
+			s.hashHead[newH] = uint32(di + s.hashOffset)
+		}
+	}
+	// Update window information.
+	d.windowEnd += n
+	s.index = n
+}
+
+// Try to find a match starting at index whose length is greater than prevSize.
+// We only look at chainCount possibilities before giving up.
+// pos = s.index, prevHead = s.chainHead-s.hashOffset, prevLength=minMatchLength-1, lookahead
+func (d *compressor) findMatch(pos int, prevHead int, lookahead int) (length, offset int, ok bool) {
+	minMatchLook := maxMatchLength
+	if lookahead < minMatchLook {
+		minMatchLook = lookahead
+	}
+
+	win := d.window[0 : pos+minMatchLook]
+
+	// We quit when we get a match that's at least nice long
+	nice := len(win) - pos
+	if d.nice < nice {
+		nice = d.nice
+	}
+
+	// If we've got a match that's good enough, only look in 1/4 the chain.
+	tries := d.chain
+	length = minMatchLength - 1
+
+	wEnd := win[pos+length]
+	wPos := win[pos:]
+	minIndex := pos - windowSize
+	if minIndex < 0 {
+		minIndex = 0
+	}
+	offset = 0
+
+	if d.chain < 100 {
+		for i := prevHead; tries > 0; tries-- {
+			if wEnd == win[i+length] {
+				n := matchLen(win[i:i+minMatchLook], wPos)
+				if n > length {
+					length = n
+					offset = pos - i
+					ok = true
+					if n >= nice {
+						// The match is good enough that we don't try to find a better one.
+						break
+					}
+					wEnd = win[pos+n]
+				}
+			}
+			if i <= minIndex {
+				// hashPrev[i & windowMask] has already been overwritten, so stop now.
+				break
+			}
+			i = int(d.state.hashPrev[i&windowMask]) - d.state.hashOffset
+			if i < minIndex {
+				break
+			}
+		}
+		return
+	}
+
+	// Minimum gain to accept a match.
+	cGain := 4
+
+	// Some like it higher (CSV), some like it lower (JSON)
+	const baseCost = 3
+	// Base is 4 bytes at with an additional cost.
+	// Matches must be better than this.
+
+	for i := prevHead; tries > 0; tries-- {
+		if wEnd == win[i+length] {
+			n := matchLen(win[i:i+minMatchLook], wPos)
+			if n > length {
+				// Calculate gain. Estimate
+				newGain := d.h.bitLengthRaw(wPos[:n]) - int(offsetExtraBits[offsetCode(uint32(pos-i))]) - baseCost - int(lengthExtraBits[lengthCodes[(n-3)&255]])
+
+				//fmt.Println("gain:", newGain, "prev:", cGain, "raw:", d.h.bitLengthRaw(wPos[:n]), "this-len:", n, "prev-len:", length)
+				if newGain > cGain {
+					length = n
+					offset = pos - i
+					cGain = newGain
+					ok = true
+					if n >= nice {
+						// The match is good enough that we don't try to find a better one.
+						break
+					}
+					wEnd = win[pos+n]
+				}
+			}
+		}
+		if i <= minIndex {
+			// hashPrev[i & windowMask] has already been overwritten, so stop now.
+			break
+		}
+		i = int(d.state.hashPrev[i&windowMask]) - d.state.hashOffset
+		if i < minIndex {
+			break
+		}
+	}
+	return
+}
+
+func (d *compressor) writeStoredBlock(buf []byte) error {
+	if d.w.writeStoredHeader(len(buf), false); d.w.err != nil {
+		return d.w.err
+	}
+	d.w.writeBytes(buf)
+	return d.w.err
+}
+
+// hash4 returns a hash representation of the first 4 bytes
+// of the supplied slice.
+// The caller must ensure that len(b) >= 4.
+func hash4(b []byte) uint32 {
+	return hash4u(binary.LittleEndian.Uint32(b), hashBits)
+}
+
+// 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 hash4u(u uint32, h uint8) uint32 {
+	return (u * prime4bytes) >> (32 - h)
+}
+
+// bulkHash4 will compute hashes using the same
+// algorithm as hash4
+func bulkHash4(b []byte, dst []uint32) {
+	if len(b) < 4 {
+		return
+	}
+	hb := binary.LittleEndian.Uint32(b)
+
+	dst[0] = hash4u(hb, hashBits)
+	end := len(b) - 4 + 1
+	for i := 1; i < end; i++ {
+		hb = (hb >> 8) | uint32(b[i+3])<<24
+		dst[i] = hash4u(hb, hashBits)
+	}
+}
+
+func (d *compressor) initDeflate() {
+	d.window = make([]byte, 2*windowSize)
+	d.byteAvailable = false
+	d.err = nil
+	if d.state == nil {
+		return
+	}
+	s := d.state
+	s.index = 0
+	s.hashOffset = 1
+	s.length = minMatchLength - 1
+	s.offset = 0
+	s.chainHead = -1
+}
+
+// deflateLazy is the same as deflate, but with d.fastSkipHashing == skipNever,
+// meaning it always has lazy matching on.
+func (d *compressor) deflateLazy() {
+	s := d.state
+	// Sanity enables additional runtime tests.
+	// It's intended to be used during development
+	// to supplement the currently ad-hoc unit tests.
+	const sanity = debugDeflate
+
+	if d.windowEnd-s.index < minMatchLength+maxMatchLength && !d.sync {
+		return
+	}
+	if d.windowEnd != s.index && d.chain > 100 {
+		// Get literal huffman coder.
+		if d.h == nil {
+			d.h = newHuffmanEncoder(maxFlateBlockTokens)
+		}
+		var tmp [256]uint16
+		for _, v := range d.window[s.index:d.windowEnd] {
+			tmp[v]++
+		}
+		d.h.generate(tmp[:], 15)
+	}
+
+	s.maxInsertIndex = d.windowEnd - (minMatchLength - 1)
+
+	for {
+		if sanity && s.index > d.windowEnd {
+			panic("index > windowEnd")
+		}
+		lookahead := d.windowEnd - s.index
+		if lookahead < minMatchLength+maxMatchLength {
+			if !d.sync {
+				return
+			}
+			if sanity && s.index > d.windowEnd {
+				panic("index > windowEnd")
+			}
+			if lookahead == 0 {
+				// Flush current output block if any.
+				if d.byteAvailable {
+					// There is still one pending token that needs to be flushed
+					d.tokens.AddLiteral(d.window[s.index-1])
+					d.byteAvailable = false
+				}
+				if d.tokens.n > 0 {
+					if d.err = d.writeBlock(&d.tokens, s.index, false); d.err != nil {
+						return
+					}
+					d.tokens.Reset()
+				}
+				return
+			}
+		}
+		if s.index < s.maxInsertIndex {
+			// Update the hash
+			hash := hash4(d.window[s.index:])
+			ch := s.hashHead[hash]
+			s.chainHead = int(ch)
+			s.hashPrev[s.index&windowMask] = ch
+			s.hashHead[hash] = uint32(s.index + s.hashOffset)
+		}
+		prevLength := s.length
+		prevOffset := s.offset
+		s.length = minMatchLength - 1
+		s.offset = 0
+		minIndex := s.index - windowSize
+		if minIndex < 0 {
+			minIndex = 0
+		}
+
+		if s.chainHead-s.hashOffset >= minIndex && lookahead > prevLength && prevLength < d.lazy {
+			if newLength, newOffset, ok := d.findMatch(s.index, s.chainHead-s.hashOffset, lookahead); ok {
+				s.length = newLength
+				s.offset = newOffset
+			}
+		}
+
+		if prevLength >= minMatchLength && s.length <= prevLength {
+			// No better match, but check for better match at end...
+			//
+			// Skip forward a number of bytes.
+			// Offset of 2 seems to yield best results. 3 is sometimes better.
+			const checkOff = 2
+
+			// Check all, except full length
+			if prevLength < maxMatchLength-checkOff {
+				prevIndex := s.index - 1
+				if prevIndex+prevLength < s.maxInsertIndex {
+					end := lookahead
+					if lookahead > maxMatchLength+checkOff {
+						end = maxMatchLength + checkOff
+					}
+					end += prevIndex
+
+					// Hash at match end.
+					h := hash4(d.window[prevIndex+prevLength:])
+					ch2 := int(s.hashHead[h]) - s.hashOffset - prevLength
+					if prevIndex-ch2 != prevOffset && ch2 > minIndex+checkOff {
+						length := matchLen(d.window[prevIndex+checkOff:end], d.window[ch2+checkOff:])
+						// It seems like a pure length metric is best.
+						if length > prevLength {
+							prevLength = length
+							prevOffset = prevIndex - ch2
+
+							// Extend back...
+							for i := checkOff - 1; i >= 0; i-- {
+								if prevLength >= maxMatchLength || d.window[prevIndex+i] != d.window[ch2+i] {
+									// Emit tokens we "owe"
+									for j := 0; j <= i; j++ {
+										d.tokens.AddLiteral(d.window[prevIndex+j])
+										if d.tokens.n == maxFlateBlockTokens {
+											// The block includes the current character
+											if d.err = d.writeBlock(&d.tokens, s.index, false); d.err != nil {
+												return
+											}
+											d.tokens.Reset()
+										}
+										s.index++
+										if s.index < s.maxInsertIndex {
+											h := hash4(d.window[s.index:])
+											ch := s.hashHead[h]
+											s.chainHead = int(ch)
+											s.hashPrev[s.index&windowMask] = ch
+											s.hashHead[h] = uint32(s.index + s.hashOffset)
+										}
+									}
+									break
+								} else {
+									prevLength++
+								}
+							}
+						} else if false {
+							// Check one further ahead.
+							// Only rarely better, disabled for now.
+							prevIndex++
+							h := hash4(d.window[prevIndex+prevLength:])
+							ch2 := int(s.hashHead[h]) - s.hashOffset - prevLength
+							if prevIndex-ch2 != prevOffset && ch2 > minIndex+checkOff {
+								length := matchLen(d.window[prevIndex+checkOff:end], d.window[ch2+checkOff:])
+								// It seems like a pure length metric is best.
+								if length > prevLength+checkOff {
+									prevLength = length
+									prevOffset = prevIndex - ch2
+									prevIndex--
+
+									// Extend back...
+									for i := checkOff; i >= 0; i-- {
+										if prevLength >= maxMatchLength || d.window[prevIndex+i] != d.window[ch2+i-1] {
+											// Emit tokens we "owe"
+											for j := 0; j <= i; j++ {
+												d.tokens.AddLiteral(d.window[prevIndex+j])
+												if d.tokens.n == maxFlateBlockTokens {
+													// The block includes the current character
+													if d.err = d.writeBlock(&d.tokens, s.index, false); d.err != nil {
+														return
+													}
+													d.tokens.Reset()
+												}
+												s.index++
+												if s.index < s.maxInsertIndex {
+													h := hash4(d.window[s.index:])
+													ch := s.hashHead[h]
+													s.chainHead = int(ch)
+													s.hashPrev[s.index&windowMask] = ch
+													s.hashHead[h] = uint32(s.index + s.hashOffset)
+												}
+											}
+											break
+										} else {
+											prevLength++
+										}
+									}
+								}
+							}
+						}
+					}
+				}
+			}
+			// There was a match at the previous step, and the current match is
+			// not better. Output the previous match.
+			d.tokens.AddMatch(uint32(prevLength-3), uint32(prevOffset-minOffsetSize))
+
+			// Insert in the hash table all strings up to the end of the match.
+			// index and index-1 are already inserted. If there is not enough
+			// lookahead, the last two strings are not inserted into the hash
+			// table.
+			newIndex := s.index + prevLength - 1
+			// Calculate missing hashes
+			end := newIndex
+			if end > s.maxInsertIndex {
+				end = s.maxInsertIndex
+			}
+			end += minMatchLength - 1
+			startindex := s.index + 1
+			if startindex > s.maxInsertIndex {
+				startindex = s.maxInsertIndex
+			}
+			tocheck := d.window[startindex:end]
+			dstSize := len(tocheck) - minMatchLength + 1
+			if dstSize > 0 {
+				dst := s.hashMatch[:dstSize]
+				bulkHash4(tocheck, dst)
+				var newH uint32
+				for i, val := range dst {
+					di := i + startindex
+					newH = val & hashMask
+					// Get previous value with the same hash.
+					// Our chain should point to the previous value.
+					s.hashPrev[di&windowMask] = s.hashHead[newH]
+					// Set the head of the hash chain to us.
+					s.hashHead[newH] = uint32(di + s.hashOffset)
+				}
+			}
+
+			s.index = newIndex
+			d.byteAvailable = false
+			s.length = minMatchLength - 1
+			if d.tokens.n == maxFlateBlockTokens {
+				// The block includes the current character
+				if d.err = d.writeBlock(&d.tokens, s.index, false); d.err != nil {
+					return
+				}
+				d.tokens.Reset()
+			}
+			s.ii = 0
+		} else {
+			// Reset, if we got a match this run.
+			if s.length >= minMatchLength {
+				s.ii = 0
+			}
+			// We have a byte waiting. Emit it.
+			if d.byteAvailable {
+				s.ii++
+				d.tokens.AddLiteral(d.window[s.index-1])
+				if d.tokens.n == maxFlateBlockTokens {
+					if d.err = d.writeBlock(&d.tokens, s.index, false); d.err != nil {
+						return
+					}
+					d.tokens.Reset()
+				}
+				s.index++
+
+				// If we have a long run of no matches, skip additional bytes
+				// Resets when s.ii overflows after 64KB.
+				if n := int(s.ii) - d.chain; n > 0 {
+					n = 1 + int(n>>6)
+					for j := 0; j < n; j++ {
+						if s.index >= d.windowEnd-1 {
+							break
+						}
+						d.tokens.AddLiteral(d.window[s.index-1])
+						if d.tokens.n == maxFlateBlockTokens {
+							if d.err = d.writeBlock(&d.tokens, s.index, false); d.err != nil {
+								return
+							}
+							d.tokens.Reset()
+						}
+						// Index...
+						if s.index < s.maxInsertIndex {
+							h := hash4(d.window[s.index:])
+							ch := s.hashHead[h]
+							s.chainHead = int(ch)
+							s.hashPrev[s.index&windowMask] = ch
+							s.hashHead[h] = uint32(s.index + s.hashOffset)
+						}
+						s.index++
+					}
+					// Flush last byte
+					d.tokens.AddLiteral(d.window[s.index-1])
+					d.byteAvailable = false
+					// s.length = minMatchLength - 1 // not needed, since s.ii is reset above, so it should never be > minMatchLength
+					if d.tokens.n == maxFlateBlockTokens {
+						if d.err = d.writeBlock(&d.tokens, s.index, false); d.err != nil {
+							return
+						}
+						d.tokens.Reset()
+					}
+				}
+			} else {
+				s.index++
+				d.byteAvailable = true
+			}
+		}
+	}
+}
+
+func (d *compressor) store() {
+	if d.windowEnd > 0 && (d.windowEnd == maxStoreBlockSize || d.sync) {
+		d.err = d.writeStoredBlock(d.window[:d.windowEnd])
+		d.windowEnd = 0
+	}
+}
+
+// fillWindow will fill the buffer with data for huffman-only compression.
+// The number of bytes copied is returned.
+func (d *compressor) fillBlock(b []byte) int {
+	n := copy(d.window[d.windowEnd:], b)
+	d.windowEnd += n
+	return n
+}
+
+// storeHuff will compress and store the currently added data,
+// if enough has been accumulated or we at the end of the stream.
+// Any error that occurred will be in d.err
+func (d *compressor) storeHuff() {
+	if d.windowEnd < len(d.window) && !d.sync || d.windowEnd == 0 {
+		return
+	}
+	d.w.writeBlockHuff(false, d.window[:d.windowEnd], d.sync)
+	d.err = d.w.err
+	d.windowEnd = 0
+}
+
+// storeFast will compress and store the currently added data,
+// if enough has been accumulated or we at the end of the stream.
+// Any error that occurred will be in d.err
+func (d *compressor) storeFast() {
+	// We only compress if we have maxStoreBlockSize.
+	if d.windowEnd < len(d.window) {
+		if !d.sync {
+			return
+		}
+		// Handle extremely small sizes.
+		if d.windowEnd < 128 {
+			if d.windowEnd == 0 {
+				return
+			}
+			if d.windowEnd <= 32 {
+				d.err = d.writeStoredBlock(d.window[:d.windowEnd])
+			} else {
+				d.w.writeBlockHuff(false, d.window[:d.windowEnd], true)
+				d.err = d.w.err
+			}
+			d.tokens.Reset()
+			d.windowEnd = 0
+			d.fast.Reset()
+			return
+		}
+	}
+
+	d.fast.Encode(&d.tokens, d.window[:d.windowEnd])
+	// If we made zero matches, store the block as is.
+	if d.tokens.n == 0 {
+		d.err = d.writeStoredBlock(d.window[:d.windowEnd])
+		// If we removed less than 1/16th, huffman compress the block.
+	} else if int(d.tokens.n) > d.windowEnd-(d.windowEnd>>4) {
+		d.w.writeBlockHuff(false, d.window[:d.windowEnd], d.sync)
+		d.err = d.w.err
+	} else {
+		d.w.writeBlockDynamic(&d.tokens, false, d.window[:d.windowEnd], d.sync)
+		d.err = d.w.err
+	}
+	d.tokens.Reset()
+	d.windowEnd = 0
+}
+
+// write will add input byte to the stream.
+// Unless an error occurs all bytes will be consumed.
+func (d *compressor) write(b []byte) (n int, err error) {
+	if d.err != nil {
+		return 0, d.err
+	}
+	n = len(b)
+	for len(b) > 0 {
+		if d.windowEnd == len(d.window) || d.sync {
+			d.step(d)
+		}
+		b = b[d.fill(d, b):]
+		if d.err != nil {
+			return 0, d.err
+		}
+	}
+	return n, d.err
+}
+
+func (d *compressor) syncFlush() error {
+	d.sync = true
+	if d.err != nil {
+		return d.err
+	}
+	d.step(d)
+	if d.err == nil {
+		d.w.writeStoredHeader(0, false)
+		d.w.flush()
+		d.err = d.w.err
+	}
+	d.sync = false
+	return d.err
+}
+
+func (d *compressor) init(w io.Writer, level int) (err error) {
+	d.w = newHuffmanBitWriter(w)
+
+	switch {
+	case level == NoCompression:
+		d.window = make([]byte, maxStoreBlockSize)
+		d.fill = (*compressor).fillBlock
+		d.step = (*compressor).store
+	case level == ConstantCompression:
+		d.w.logNewTablePenalty = 10
+		d.window = make([]byte, 32<<10)
+		d.fill = (*compressor).fillBlock
+		d.step = (*compressor).storeHuff
+	case level == DefaultCompression:
+		level = 5
+		fallthrough
+	case level >= 1 && level <= 6:
+		d.w.logNewTablePenalty = 7
+		d.fast = newFastEnc(level)
+		d.window = make([]byte, maxStoreBlockSize)
+		d.fill = (*compressor).fillBlock
+		d.step = (*compressor).storeFast
+	case 7 <= level && level <= 9:
+		d.w.logNewTablePenalty = 8
+		d.state = &advancedState{}
+		d.compressionLevel = levels[level]
+		d.initDeflate()
+		d.fill = (*compressor).fillDeflate
+		d.step = (*compressor).deflateLazy
+	case -level >= MinCustomWindowSize && -level <= MaxCustomWindowSize:
+		d.w.logNewTablePenalty = 7
+		d.fast = &fastEncL5Window{maxOffset: int32(-level), cur: maxStoreBlockSize}
+		d.window = make([]byte, maxStoreBlockSize)
+		d.fill = (*compressor).fillBlock
+		d.step = (*compressor).storeFast
+	default:
+		return fmt.Errorf("flate: invalid compression level %d: want value in range [-2, 9]", level)
+	}
+	d.level = level
+	return nil
+}
+
+// reset the state of the compressor.
+func (d *compressor) reset(w io.Writer) {
+	d.w.reset(w)
+	d.sync = false
+	d.err = nil
+	// We only need to reset a few things for Snappy.
+	if d.fast != nil {
+		d.fast.Reset()
+		d.windowEnd = 0
+		d.tokens.Reset()
+		return
+	}
+	switch d.compressionLevel.chain {
+	case 0:
+		// level was NoCompression or ConstantCompression.
+		d.windowEnd = 0
+	default:
+		s := d.state
+		s.chainHead = -1
+		for i := range s.hashHead {
+			s.hashHead[i] = 0
+		}
+		for i := range s.hashPrev {
+			s.hashPrev[i] = 0
+		}
+		s.hashOffset = 1
+		s.index, d.windowEnd = 0, 0
+		d.blockStart, d.byteAvailable = 0, false
+		d.tokens.Reset()
+		s.length = minMatchLength - 1
+		s.offset = 0
+		s.ii = 0
+		s.maxInsertIndex = 0
+	}
+}
+
+func (d *compressor) close() error {
+	if d.err != nil {
+		return d.err
+	}
+	d.sync = true
+	d.step(d)
+	if d.err != nil {
+		return d.err
+	}
+	if d.w.writeStoredHeader(0, true); d.w.err != nil {
+		return d.w.err
+	}
+	d.w.flush()
+	d.w.reset(nil)
+	return d.w.err
+}
+
+// NewWriter returns a new Writer compressing data at the given level.
+// Following zlib, levels range from 1 (BestSpeed) to 9 (BestCompression);
+// higher levels typically run slower but compress more.
+// Level 0 (NoCompression) does not attempt any compression; it only adds the
+// necessary DEFLATE framing.
+// Level -1 (DefaultCompression) uses the default compression level.
+// Level -2 (ConstantCompression) will use Huffman compression only, giving
+// a very fast compression for all types of input, but sacrificing considerable
+// compression efficiency.
+//
+// If level is in the range [-2, 9] then the error returned will be nil.
+// Otherwise the error returned will be non-nil.
+func NewWriter(w io.Writer, level int) (*Writer, error) {
+	var dw Writer
+	if err := dw.d.init(w, level); err != nil {
+		return nil, err
+	}
+	return &dw, nil
+}
+
+// NewWriterDict is like NewWriter but initializes the new
+// Writer with a preset dictionary.  The returned Writer behaves
+// as if the dictionary had been written to it without producing
+// any compressed output.  The compressed data written to w
+// can only be decompressed by a Reader initialized with the
+// same dictionary.
+func NewWriterDict(w io.Writer, level int, dict []byte) (*Writer, error) {
+	zw, err := NewWriter(w, level)
+	if err != nil {
+		return nil, err
+	}
+	zw.d.fillWindow(dict)
+	zw.dict = append(zw.dict, dict...) // duplicate dictionary for Reset method.
+	return zw, err
+}
+
+// MinCustomWindowSize is the minimum window size that can be sent to NewWriterWindow.
+const MinCustomWindowSize = 32
+
+// MaxCustomWindowSize is the maximum custom window that can be sent to NewWriterWindow.
+const MaxCustomWindowSize = windowSize
+
+// NewWriterWindow returns a new Writer compressing data with a custom window size.
+// windowSize must be from MinCustomWindowSize to MaxCustomWindowSize.
+func NewWriterWindow(w io.Writer, windowSize int) (*Writer, error) {
+	if windowSize < MinCustomWindowSize {
+		return nil, errors.New("flate: requested window size less than MinWindowSize")
+	}
+	if windowSize > MaxCustomWindowSize {
+		return nil, errors.New("flate: requested window size bigger than MaxCustomWindowSize")
+	}
+	var dw Writer
+	if err := dw.d.init(w, -windowSize); err != nil {
+		return nil, err
+	}
+	return &dw, nil
+}
+
+// A Writer takes data written to it and writes the compressed
+// form of that data to an underlying writer (see NewWriter).
+type Writer struct {
+	d    compressor
+	dict []byte
+}
+
+// Write writes data to w, which will eventually write the
+// compressed form of data to its underlying writer.
+func (w *Writer) Write(data []byte) (n int, err error) {
+	return w.d.write(data)
+}
+
+// Flush flushes any pending data to the underlying writer.
+// It is useful mainly in compressed network protocols, to ensure that
+// a remote reader has enough data to reconstruct a packet.
+// Flush does not return until the data has been written.
+// Calling Flush when there is no pending data still causes the Writer
+// to emit a sync marker of at least 4 bytes.
+// If the underlying writer returns an error, Flush returns that error.
+//
+// In the terminology of the zlib library, Flush is equivalent to Z_SYNC_FLUSH.
+func (w *Writer) Flush() error {
+	// For more about flushing:
+	// http://www.bolet.org/~pornin/deflate-flush.html
+	return w.d.syncFlush()
+}
+
+// Close flushes and closes the writer.
+func (w *Writer) Close() error {
+	return w.d.close()
+}
+
+// Reset discards the writer's state and makes it equivalent to
+// the result of NewWriter or NewWriterDict called with dst
+// and w's level and dictionary.
+func (w *Writer) Reset(dst io.Writer) {
+	if len(w.dict) > 0 {
+		// w was created with NewWriterDict
+		w.d.reset(dst)
+		if dst != nil {
+			w.d.fillWindow(w.dict)
+		}
+	} else {
+		// w was created with NewWriter
+		w.d.reset(dst)
+	}
+}
+
+// ResetDict discards the writer's state and makes it equivalent to
+// the result of NewWriter or NewWriterDict called with dst
+// and w's level, but sets a specific dictionary.
+func (w *Writer) ResetDict(dst io.Writer, dict []byte) {
+	w.dict = dict
+	w.d.reset(dst)
+	w.d.fillWindow(w.dict)
+}
diff --git a/vendor/github.com/klauspost/compress/flate/dict_decoder.go b/vendor/github.com/klauspost/compress/flate/dict_decoder.go
new file mode 100644
index 0000000..bb36351
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/dict_decoder.go
@@ -0,0 +1,184 @@
+// Copyright 2016 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package flate
+
+// dictDecoder implements the LZ77 sliding dictionary as used in decompression.
+// LZ77 decompresses data through sequences of two forms of commands:
+//
+//   - Literal insertions: Runs of one or more symbols are inserted into the data
+//     stream as is. This is accomplished through the writeByte method for a
+//     single symbol, or combinations of writeSlice/writeMark for multiple symbols.
+//     Any valid stream must start with a literal insertion if no preset dictionary
+//     is used.
+//
+//   - Backward copies: Runs of one or more symbols are copied from previously
+//     emitted data. Backward copies come as the tuple (dist, length) where dist
+//     determines how far back in the stream to copy from and length determines how
+//     many bytes to copy. Note that it is valid for the length to be greater than
+//     the distance. Since LZ77 uses forward copies, that situation is used to
+//     perform a form of run-length encoding on repeated runs of symbols.
+//     The writeCopy and tryWriteCopy are used to implement this command.
+//
+// For performance reasons, this implementation performs little to no sanity
+// checks about the arguments. As such, the invariants documented for each
+// method call must be respected.
+type dictDecoder struct {
+	hist []byte // Sliding window history
+
+	// Invariant: 0 <= rdPos <= wrPos <= len(hist)
+	wrPos int  // Current output position in buffer
+	rdPos int  // Have emitted hist[:rdPos] already
+	full  bool // Has a full window length been written yet?
+}
+
+// init initializes dictDecoder to have a sliding window dictionary of the given
+// size. If a preset dict is provided, it will initialize the dictionary with
+// the contents of dict.
+func (dd *dictDecoder) init(size int, dict []byte) {
+	*dd = dictDecoder{hist: dd.hist}
+
+	if cap(dd.hist) < size {
+		dd.hist = make([]byte, size)
+	}
+	dd.hist = dd.hist[:size]
+
+	if len(dict) > len(dd.hist) {
+		dict = dict[len(dict)-len(dd.hist):]
+	}
+	dd.wrPos = copy(dd.hist, dict)
+	if dd.wrPos == len(dd.hist) {
+		dd.wrPos = 0
+		dd.full = true
+	}
+	dd.rdPos = dd.wrPos
+}
+
+// histSize reports the total amount of historical data in the dictionary.
+func (dd *dictDecoder) histSize() int {
+	if dd.full {
+		return len(dd.hist)
+	}
+	return dd.wrPos
+}
+
+// availRead reports the number of bytes that can be flushed by readFlush.
+func (dd *dictDecoder) availRead() int {
+	return dd.wrPos - dd.rdPos
+}
+
+// availWrite reports the available amount of output buffer space.
+func (dd *dictDecoder) availWrite() int {
+	return len(dd.hist) - dd.wrPos
+}
+
+// writeSlice returns a slice of the available buffer to write data to.
+//
+// This invariant will be kept: len(s) <= availWrite()
+func (dd *dictDecoder) writeSlice() []byte {
+	return dd.hist[dd.wrPos:]
+}
+
+// writeMark advances the writer pointer by cnt.
+//
+// This invariant must be kept: 0 <= cnt <= availWrite()
+func (dd *dictDecoder) writeMark(cnt int) {
+	dd.wrPos += cnt
+}
+
+// writeByte writes a single byte to the dictionary.
+//
+// This invariant must be kept: 0 < availWrite()
+func (dd *dictDecoder) writeByte(c byte) {
+	dd.hist[dd.wrPos] = c
+	dd.wrPos++
+}
+
+// writeCopy copies a string at a given (dist, length) to the output.
+// This returns the number of bytes copied and may be less than the requested
+// length if the available space in the output buffer is too small.
+//
+// This invariant must be kept: 0 < dist <= histSize()
+func (dd *dictDecoder) writeCopy(dist, length int) int {
+	dstBase := dd.wrPos
+	dstPos := dstBase
+	srcPos := dstPos - dist
+	endPos := dstPos + length
+	if endPos > len(dd.hist) {
+		endPos = len(dd.hist)
+	}
+
+	// Copy non-overlapping section after destination position.
+	//
+	// This section is non-overlapping in that the copy length for this section
+	// is always less than or equal to the backwards distance. This can occur
+	// if a distance refers to data that wraps-around in the buffer.
+	// Thus, a backwards copy is performed here; that is, the exact bytes in
+	// the source prior to the copy is placed in the destination.
+	if srcPos < 0 {
+		srcPos += len(dd.hist)
+		dstPos += copy(dd.hist[dstPos:endPos], dd.hist[srcPos:])
+		srcPos = 0
+	}
+
+	// Copy possibly overlapping section before destination position.
+	//
+	// This section can overlap if the copy length for this section is larger
+	// than the backwards distance. This is allowed by LZ77 so that repeated
+	// strings can be succinctly represented using (dist, length) pairs.
+	// Thus, a forwards copy is performed here; that is, the bytes copied is
+	// possibly dependent on the resulting bytes in the destination as the copy
+	// progresses along. This is functionally equivalent to the following:
+	//
+	//	for i := 0; i < endPos-dstPos; i++ {
+	//		dd.hist[dstPos+i] = dd.hist[srcPos+i]
+	//	}
+	//	dstPos = endPos
+	//
+	for dstPos < endPos {
+		dstPos += copy(dd.hist[dstPos:endPos], dd.hist[srcPos:dstPos])
+	}
+
+	dd.wrPos = dstPos
+	return dstPos - dstBase
+}
+
+// tryWriteCopy tries to copy a string at a given (distance, length) to the
+// output. This specialized version is optimized for short distances.
+//
+// This method is designed to be inlined for performance reasons.
+//
+// This invariant must be kept: 0 < dist <= histSize()
+func (dd *dictDecoder) tryWriteCopy(dist, length int) int {
+	dstPos := dd.wrPos
+	endPos := dstPos + length
+	if dstPos < dist || endPos > len(dd.hist) {
+		return 0
+	}
+	dstBase := dstPos
+	srcPos := dstPos - dist
+
+	// Copy possibly overlapping section before destination position.
+loop:
+	dstPos += copy(dd.hist[dstPos:endPos], dd.hist[srcPos:dstPos])
+	if dstPos < endPos {
+		goto loop // Avoid for-loop so that this function can be inlined
+	}
+
+	dd.wrPos = dstPos
+	return dstPos - dstBase
+}
+
+// readFlush returns a slice of the historical buffer that is ready to be
+// emitted to the user. The data returned by readFlush must be fully consumed
+// before calling any other dictDecoder methods.
+func (dd *dictDecoder) readFlush() []byte {
+	toRead := dd.hist[dd.rdPos:dd.wrPos]
+	dd.rdPos = dd.wrPos
+	if dd.wrPos == len(dd.hist) {
+		dd.wrPos, dd.rdPos = 0, 0
+		dd.full = true
+	}
+	return toRead
+}
diff --git a/vendor/github.com/klauspost/compress/flate/fast_encoder.go b/vendor/github.com/klauspost/compress/flate/fast_encoder.go
new file mode 100644
index 0000000..0e8b163
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/fast_encoder.go
@@ -0,0 +1,232 @@
+// Copyright 2011 The Snappy-Go Authors. All rights reserved.
+// Modified for deflate by Klaus Post (c) 2015.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package flate
+
+import (
+	"fmt"
+	"math/bits"
+
+	"github.com/klauspost/compress/internal/le"
+)
+
+type fastEnc interface {
+	Encode(dst *tokens, src []byte)
+	Reset()
+}
+
+func newFastEnc(level int) fastEnc {
+	switch level {
+	case 1:
+		return &fastEncL1{fastGen: fastGen{cur: maxStoreBlockSize}}
+	case 2:
+		return &fastEncL2{fastGen: fastGen{cur: maxStoreBlockSize}}
+	case 3:
+		return &fastEncL3{fastGen: fastGen{cur: maxStoreBlockSize}}
+	case 4:
+		return &fastEncL4{fastGen: fastGen{cur: maxStoreBlockSize}}
+	case 5:
+		return &fastEncL5{fastGen: fastGen{cur: maxStoreBlockSize}}
+	case 6:
+		return &fastEncL6{fastGen: fastGen{cur: maxStoreBlockSize}}
+	default:
+		panic("invalid level specified")
+	}
+}
+
+const (
+	tableBits       = 15             // Bits used in the table
+	tableSize       = 1 << tableBits // Size of the table
+	tableShift      = 32 - tableBits // Right-shift to get the tableBits most significant bits of a uint32.
+	baseMatchOffset = 1              // The smallest match offset
+	baseMatchLength = 3              // The smallest match length per the RFC section 3.2.5
+	maxMatchOffset  = 1 << 15        // The largest match offset
+
+	bTableBits   = 17                                               // Bits used in the big tables
+	bTableSize   = 1 << bTableBits                                  // Size of the table
+	allocHistory = maxStoreBlockSize * 5                            // Size to preallocate for history.
+	bufferReset  = (1 << 31) - allocHistory - maxStoreBlockSize - 1 // Reset the buffer offset when reaching this.
+)
+
+const (
+	prime3bytes = 506832829
+	prime4bytes = 2654435761
+	prime5bytes = 889523592379
+	prime6bytes = 227718039650203
+	prime7bytes = 58295818150454627
+	prime8bytes = 0xcf1bbcdcb7a56463
+)
+
+func load3232(b []byte, i int32) uint32 {
+	return le.Load32(b, i)
+}
+
+func load6432(b []byte, i int32) uint64 {
+	return le.Load64(b, i)
+}
+
+type tableEntry struct {
+	offset int32
+}
+
+// fastGen maintains the table for matches,
+// and the previous byte block for level 2.
+// This is the generic implementation.
+type fastGen struct {
+	hist []byte
+	cur  int32
+}
+
+func (e *fastGen) addBlock(src []byte) int32 {
+	// check if we have space already
+	if len(e.hist)+len(src) > cap(e.hist) {
+		if cap(e.hist) == 0 {
+			e.hist = make([]byte, 0, allocHistory)
+		} else {
+			if cap(e.hist) < maxMatchOffset*2 {
+				panic("unexpected buffer size")
+			}
+			// Move down
+			offset := int32(len(e.hist)) - maxMatchOffset
+			// copy(e.hist[0:maxMatchOffset], e.hist[offset:])
+			*(*[maxMatchOffset]byte)(e.hist) = *(*[maxMatchOffset]byte)(e.hist[offset:])
+			e.cur += offset
+			e.hist = e.hist[:maxMatchOffset]
+		}
+	}
+	s := int32(len(e.hist))
+	e.hist = append(e.hist, src...)
+	return s
+}
+
+type tableEntryPrev struct {
+	Cur  tableEntry
+	Prev tableEntry
+}
+
+// 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) & reg8SizeMask64))
+}
+
+// hashLen returns a hash of the lowest mls bytes of with length output bits.
+// mls must be >=3 and <=8. Any other value will return hash for 4 bytes.
+// length should always be < 32.
+// Preferably length and mls should be a constant for inlining.
+func hashLen(u uint64, length, mls uint8) uint32 {
+	switch mls {
+	case 3:
+		return (uint32(u<<8) * prime3bytes) >> (32 - length)
+	case 5:
+		return uint32(((u << (64 - 40)) * prime5bytes) >> (64 - length))
+	case 6:
+		return uint32(((u << (64 - 48)) * prime6bytes) >> (64 - length))
+	case 7:
+		return uint32(((u << (64 - 56)) * prime7bytes) >> (64 - length))
+	case 8:
+		return uint32((u * prime8bytes) >> (64 - length))
+	default:
+		return (uint32(u) * prime4bytes) >> (32 - length)
+	}
+}
+
+// matchlen will return the match length between offsets and t in src.
+// The maximum length returned is maxMatchLength - 4.
+// It is assumed that s > t, that t >=0 and s < len(src).
+func (e *fastGen) matchlen(s, t int, src []byte) int32 {
+	if debugDeflate {
+		if t >= s {
+			panic(fmt.Sprint("t >=s:", t, s))
+		}
+		if int(s) >= len(src) {
+			panic(fmt.Sprint("s >= len(src):", s, len(src)))
+		}
+		if t < 0 {
+			panic(fmt.Sprint("t < 0:", t))
+		}
+		if s-t > maxMatchOffset {
+			panic(fmt.Sprint(s, "-", t, "(", s-t, ") > maxMatchLength (", maxMatchOffset, ")"))
+		}
+	}
+	s1 := min(s+maxMatchLength-4, len(src))
+	left := s1 - s
+	n := int32(0)
+	for left >= 8 {
+		diff := le.Load64(src, s) ^ le.Load64(src, t)
+		if diff != 0 {
+			return n + int32(bits.TrailingZeros64(diff)>>3)
+		}
+		s += 8
+		t += 8
+		n += 8
+		left -= 8
+	}
+
+	a := src[s:s1]
+	b := src[t:]
+	for i := range a {
+		if a[i] != b[i] {
+			break
+		}
+		n++
+	}
+	return n
+}
+
+// matchlenLong will return the match length between offsets and t in src.
+// It is assumed that s > t, that t >=0 and s < len(src).
+func (e *fastGen) matchlenLong(s, t int, src []byte) int32 {
+	if debugDeflate {
+		if t >= s {
+			panic(fmt.Sprint("t >=s:", t, s))
+		}
+		if int(s) >= len(src) {
+			panic(fmt.Sprint("s >= len(src):", s, len(src)))
+		}
+		if t < 0 {
+			panic(fmt.Sprint("t < 0:", t))
+		}
+		if s-t > maxMatchOffset {
+			panic(fmt.Sprint(s, "-", t, "(", s-t, ") > maxMatchLength (", maxMatchOffset, ")"))
+		}
+	}
+	// Extend the match to be as long as possible.
+	left := len(src) - s
+	n := int32(0)
+	for left >= 8 {
+		diff := le.Load64(src, s) ^ le.Load64(src, t)
+		if diff != 0 {
+			return n + int32(bits.TrailingZeros64(diff)>>3)
+		}
+		s += 8
+		t += 8
+		n += 8
+		left -= 8
+	}
+
+	a := src[s:]
+	b := src[t:]
+	for i := range a {
+		if a[i] != b[i] {
+			break
+		}
+		n++
+	}
+	return n
+}
+
+// Reset the encoding table.
+func (e *fastGen) Reset() {
+	if cap(e.hist) < allocHistory {
+		e.hist = make([]byte, 0, allocHistory)
+	}
+	// We offset current position so everything will be out of reach.
+	// If we are above the buffer reset it will be cleared anyway since len(hist) == 0.
+	if e.cur <= bufferReset {
+		e.cur += maxMatchOffset + int32(len(e.hist))
+	}
+	e.hist = e.hist[:0]
+}
diff --git a/vendor/github.com/klauspost/compress/flate/huffman_bit_writer.go b/vendor/github.com/klauspost/compress/flate/huffman_bit_writer.go
new file mode 100644
index 0000000..afdc8c0
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/huffman_bit_writer.go
@@ -0,0 +1,1183 @@
+// Copyright 2009 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package flate
+
+import (
+	"fmt"
+	"io"
+	"math"
+
+	"github.com/klauspost/compress/internal/le"
+)
+
+const (
+	// The largest offset code.
+	offsetCodeCount = 30
+
+	// The special code used to mark the end of a block.
+	endBlockMarker = 256
+
+	// The first length code.
+	lengthCodesStart = 257
+
+	// The number of codegen codes.
+	codegenCodeCount = 19
+	badCode          = 255
+
+	// maxPredefinedTokens is the maximum number of tokens
+	// where we check if fixed size is smaller.
+	maxPredefinedTokens = 250
+
+	// bufferFlushSize indicates the buffer size
+	// after which bytes are flushed to the writer.
+	// Should preferably be a multiple of 6, since
+	// we accumulate 6 bytes between writes to the buffer.
+	bufferFlushSize = 246
+)
+
+// Minimum length code that emits bits.
+const lengthExtraBitsMinCode = 8
+
+// The number of extra bits needed by length code X - LENGTH_CODES_START.
+var lengthExtraBits = [32]uint8{
+	/* 257 */ 0, 0, 0,
+	/* 260 */ 0, 0, 0, 0, 0, 1, 1, 1, 1, 2,
+	/* 270 */ 2, 2, 2, 3, 3, 3, 3, 4, 4, 4,
+	/* 280 */ 4, 5, 5, 5, 5, 0,
+}
+
+// The length indicated by length code X - LENGTH_CODES_START.
+var lengthBase = [32]uint8{
+	0, 1, 2, 3, 4, 5, 6, 7, 8, 10,
+	12, 14, 16, 20, 24, 28, 32, 40, 48, 56,
+	64, 80, 96, 112, 128, 160, 192, 224, 255,
+}
+
+// Minimum offset code that emits bits.
+const offsetExtraBitsMinCode = 4
+
+// offset code word extra bits.
+var offsetExtraBits = [32]int8{
+	0, 0, 0, 0, 1, 1, 2, 2, 3, 3,
+	4, 4, 5, 5, 6, 6, 7, 7, 8, 8,
+	9, 9, 10, 10, 11, 11, 12, 12, 13, 13,
+	/* extended window */
+	14, 14,
+}
+
+var offsetCombined = [32]uint32{}
+
+func init() {
+	var offsetBase = [32]uint32{
+		/* normal deflate */
+		0x000000, 0x000001, 0x000002, 0x000003, 0x000004,
+		0x000006, 0x000008, 0x00000c, 0x000010, 0x000018,
+		0x000020, 0x000030, 0x000040, 0x000060, 0x000080,
+		0x0000c0, 0x000100, 0x000180, 0x000200, 0x000300,
+		0x000400, 0x000600, 0x000800, 0x000c00, 0x001000,
+		0x001800, 0x002000, 0x003000, 0x004000, 0x006000,
+
+		/* extended window */
+		0x008000, 0x00c000,
+	}
+
+	for i := range offsetCombined[:] {
+		// Don't use extended window values...
+		if offsetExtraBits[i] == 0 || offsetBase[i] > 0x006000 {
+			continue
+		}
+		offsetCombined[i] = uint32(offsetExtraBits[i]) | (offsetBase[i] << 8)
+	}
+}
+
+// The odd order in which the codegen code sizes are written.
+var codegenOrder = []uint32{16, 17, 18, 0, 8, 7, 9, 6, 10, 5, 11, 4, 12, 3, 13, 2, 14, 1, 15}
+
+type huffmanBitWriter struct {
+	// writer is the underlying writer.
+	// Do not use it directly; use the write method, which ensures
+	// that Write errors are sticky.
+	writer io.Writer
+
+	// Data waiting to be written is bytes[0:nbytes]
+	// and then the low nbits of bits.
+	bits            uint64
+	nbits           uint8
+	nbytes          uint8
+	lastHuffMan     bool
+	literalEncoding *huffmanEncoder
+	tmpLitEncoding  *huffmanEncoder
+	offsetEncoding  *huffmanEncoder
+	codegenEncoding *huffmanEncoder
+	err             error
+	lastHeader      int
+	// Set between 0 (reused block can be up to 2x the size)
+	logNewTablePenalty uint
+	bytes              [256 + 8]byte
+	literalFreq        [lengthCodesStart + 32]uint16
+	offsetFreq         [32]uint16
+	codegenFreq        [codegenCodeCount]uint16
+
+	// codegen must have an extra space for the final symbol.
+	codegen [literalCount + offsetCodeCount + 1]uint8
+}
+
+// Huffman reuse.
+//
+// The huffmanBitWriter supports reusing huffman tables and thereby combining block sections.
+//
+// This is controlled by several variables:
+//
+// If lastHeader is non-zero the Huffman table can be reused.
+// This also indicates that a Huffman table has been generated that can output all
+// possible symbols.
+// It also indicates that an EOB has not yet been emitted, so if a new tabel is generated
+// an EOB with the previous table must be written.
+//
+// If lastHuffMan is set, a table for outputting literals has been generated and offsets are invalid.
+//
+// An incoming block estimates the output size of a new table using a 'fresh' by calculating the
+// optimal size and adding a penalty in 'logNewTablePenalty'.
+// A Huffman table is not optimal, which is why we add a penalty, and generating a new table
+// is slower both for compression and decompression.
+
+func newHuffmanBitWriter(w io.Writer) *huffmanBitWriter {
+	return &huffmanBitWriter{
+		writer:          w,
+		literalEncoding: newHuffmanEncoder(literalCount),
+		tmpLitEncoding:  newHuffmanEncoder(literalCount),
+		codegenEncoding: newHuffmanEncoder(codegenCodeCount),
+		offsetEncoding:  newHuffmanEncoder(offsetCodeCount),
+	}
+}
+
+func (w *huffmanBitWriter) reset(writer io.Writer) {
+	w.writer = writer
+	w.bits, w.nbits, w.nbytes, w.err = 0, 0, 0, nil
+	w.lastHeader = 0
+	w.lastHuffMan = false
+}
+
+func (w *huffmanBitWriter) canReuse(t *tokens) (ok bool) {
+	a := t.offHist[:offsetCodeCount]
+	b := w.offsetEncoding.codes
+	b = b[:len(a)]
+	for i, v := range a {
+		if v != 0 && b[i].zero() {
+			return false
+		}
+	}
+
+	a = t.extraHist[:literalCount-256]
+	b = w.literalEncoding.codes[256:literalCount]
+	b = b[:len(a)]
+	for i, v := range a {
+		if v != 0 && b[i].zero() {
+			return false
+		}
+	}
+
+	a = t.litHist[:256]
+	b = w.literalEncoding.codes[:len(a)]
+	for i, v := range a {
+		if v != 0 && b[i].zero() {
+			return false
+		}
+	}
+	return true
+}
+
+func (w *huffmanBitWriter) flush() {
+	if w.err != nil {
+		w.nbits = 0
+		return
+	}
+	if w.lastHeader > 0 {
+		// We owe an EOB
+		w.writeCode(w.literalEncoding.codes[endBlockMarker])
+		w.lastHeader = 0
+	}
+	n := w.nbytes
+	for w.nbits != 0 {
+		w.bytes[n] = byte(w.bits)
+		w.bits >>= 8
+		if w.nbits > 8 { // Avoid underflow
+			w.nbits -= 8
+		} else {
+			w.nbits = 0
+		}
+		n++
+	}
+	w.bits = 0
+	w.write(w.bytes[:n])
+	w.nbytes = 0
+}
+
+func (w *huffmanBitWriter) write(b []byte) {
+	if w.err != nil {
+		return
+	}
+	_, w.err = w.writer.Write(b)
+}
+
+func (w *huffmanBitWriter) writeBits(b int32, nb uint8) {
+	w.bits |= uint64(b) << (w.nbits & 63)
+	w.nbits += nb
+	if w.nbits >= 48 {
+		w.writeOutBits()
+	}
+}
+
+func (w *huffmanBitWriter) writeBytes(bytes []byte) {
+	if w.err != nil {
+		return
+	}
+	n := w.nbytes
+	if w.nbits&7 != 0 {
+		w.err = InternalError("writeBytes with unfinished bits")
+		return
+	}
+	for w.nbits != 0 {
+		w.bytes[n] = byte(w.bits)
+		w.bits >>= 8
+		w.nbits -= 8
+		n++
+	}
+	if n != 0 {
+		w.write(w.bytes[:n])
+	}
+	w.nbytes = 0
+	w.write(bytes)
+}
+
+// RFC 1951 3.2.7 specifies a special run-length encoding for specifying
+// the literal and offset lengths arrays (which are concatenated into a single
+// array).  This method generates that run-length encoding.
+//
+// The result is written into the codegen array, and the frequencies
+// of each code is written into the codegenFreq array.
+// Codes 0-15 are single byte codes. Codes 16-18 are followed by additional
+// information. Code badCode is an end marker
+//
+//	numLiterals      The number of literals in literalEncoding
+//	numOffsets       The number of offsets in offsetEncoding
+//	litenc, offenc   The literal and offset encoder to use
+func (w *huffmanBitWriter) generateCodegen(numLiterals int, numOffsets int, litEnc, offEnc *huffmanEncoder) {
+	for i := range w.codegenFreq {
+		w.codegenFreq[i] = 0
+	}
+	// Note that we are using codegen both as a temporary variable for holding
+	// a copy of the frequencies, and as the place where we put the result.
+	// This is fine because the output is always shorter than the input used
+	// so far.
+	codegen := w.codegen[:] // cache
+	// Copy the concatenated code sizes to codegen. Put a marker at the end.
+	cgnl := codegen[:numLiterals]
+	for i := range cgnl {
+		cgnl[i] = litEnc.codes[i].len()
+	}
+
+	cgnl = codegen[numLiterals : numLiterals+numOffsets]
+	for i := range cgnl {
+		cgnl[i] = offEnc.codes[i].len()
+	}
+	codegen[numLiterals+numOffsets] = badCode
+
+	size := codegen[0]
+	count := 1
+	outIndex := 0
+	for inIndex := 1; size != badCode; inIndex++ {
+		// INVARIANT: We have seen "count" copies of size that have not yet
+		// had output generated for them.
+		nextSize := codegen[inIndex]
+		if nextSize == size {
+			count++
+			continue
+		}
+		// We need to generate codegen indicating "count" of size.
+		if size != 0 {
+			codegen[outIndex] = size
+			outIndex++
+			w.codegenFreq[size]++
+			count--
+			for count >= 3 {
+				n := 6
+				if n > count {
+					n = count
+				}
+				codegen[outIndex] = 16
+				outIndex++
+				codegen[outIndex] = uint8(n - 3)
+				outIndex++
+				w.codegenFreq[16]++
+				count -= n
+			}
+		} else {
+			for count >= 11 {
+				n := 138
+				if n > count {
+					n = count
+				}
+				codegen[outIndex] = 18
+				outIndex++
+				codegen[outIndex] = uint8(n - 11)
+				outIndex++
+				w.codegenFreq[18]++
+				count -= n
+			}
+			if count >= 3 {
+				// count >= 3 && count <= 10
+				codegen[outIndex] = 17
+				outIndex++
+				codegen[outIndex] = uint8(count - 3)
+				outIndex++
+				w.codegenFreq[17]++
+				count = 0
+			}
+		}
+		count--
+		for ; count >= 0; count-- {
+			codegen[outIndex] = size
+			outIndex++
+			w.codegenFreq[size]++
+		}
+		// Set up invariant for next time through the loop.
+		size = nextSize
+		count = 1
+	}
+	// Marker indicating the end of the codegen.
+	codegen[outIndex] = badCode
+}
+
+func (w *huffmanBitWriter) codegens() int {
+	numCodegens := len(w.codegenFreq)
+	for numCodegens > 4 && w.codegenFreq[codegenOrder[numCodegens-1]] == 0 {
+		numCodegens--
+	}
+	return numCodegens
+}
+
+func (w *huffmanBitWriter) headerSize() (size, numCodegens int) {
+	numCodegens = len(w.codegenFreq)
+	for numCodegens > 4 && w.codegenFreq[codegenOrder[numCodegens-1]] == 0 {
+		numCodegens--
+	}
+	return 3 + 5 + 5 + 4 + (3 * numCodegens) +
+		w.codegenEncoding.bitLength(w.codegenFreq[:]) +
+		int(w.codegenFreq[16])*2 +
+		int(w.codegenFreq[17])*3 +
+		int(w.codegenFreq[18])*7, numCodegens
+}
+
+// dynamicSize returns the size of dynamically encoded data in bits.
+func (w *huffmanBitWriter) dynamicReuseSize(litEnc, offEnc *huffmanEncoder) (size int) {
+	size = litEnc.bitLength(w.literalFreq[:]) +
+		offEnc.bitLength(w.offsetFreq[:])
+	return size
+}
+
+// dynamicSize returns the size of dynamically encoded data in bits.
+func (w *huffmanBitWriter) dynamicSize(litEnc, offEnc *huffmanEncoder, extraBits int) (size, numCodegens int) {
+	header, numCodegens := w.headerSize()
+	size = header +
+		litEnc.bitLength(w.literalFreq[:]) +
+		offEnc.bitLength(w.offsetFreq[:]) +
+		extraBits
+	return size, numCodegens
+}
+
+// extraBitSize will return the number of bits that will be written
+// as "extra" bits on matches.
+func (w *huffmanBitWriter) extraBitSize() int {
+	total := 0
+	for i, n := range w.literalFreq[257:literalCount] {
+		total += int(n) * int(lengthExtraBits[i&31])
+	}
+	for i, n := range w.offsetFreq[:offsetCodeCount] {
+		total += int(n) * int(offsetExtraBits[i&31])
+	}
+	return total
+}
+
+// fixedSize returns the size of dynamically encoded data in bits.
+func (w *huffmanBitWriter) fixedSize(extraBits int) int {
+	return 3 +
+		fixedLiteralEncoding.bitLength(w.literalFreq[:]) +
+		fixedOffsetEncoding.bitLength(w.offsetFreq[:]) +
+		extraBits
+}
+
+// storedSize calculates the stored size, including header.
+// The function returns the size in bits and whether the block
+// fits inside a single block.
+func (w *huffmanBitWriter) storedSize(in []byte) (int, bool) {
+	if in == nil {
+		return 0, false
+	}
+	if len(in) <= maxStoreBlockSize {
+		return (len(in) + 5) * 8, true
+	}
+	return 0, false
+}
+
+func (w *huffmanBitWriter) writeCode(c hcode) {
+	// The function does not get inlined if we "& 63" the shift.
+	w.bits |= c.code64() << (w.nbits & 63)
+	w.nbits += c.len()
+	if w.nbits >= 48 {
+		w.writeOutBits()
+	}
+}
+
+// writeOutBits will write bits to the buffer.
+func (w *huffmanBitWriter) writeOutBits() {
+	bits := w.bits
+	w.bits >>= 48
+	w.nbits -= 48
+	n := w.nbytes
+
+	// We over-write, but faster...
+	le.Store64(w.bytes[n:], bits)
+	n += 6
+
+	if n >= bufferFlushSize {
+		if w.err != nil {
+			n = 0
+			return
+		}
+		w.write(w.bytes[:n])
+		n = 0
+	}
+
+	w.nbytes = n
+}
+
+// Write the header of a dynamic Huffman block to the output stream.
+//
+//	numLiterals  The number of literals specified in codegen
+//	numOffsets   The number of offsets specified in codegen
+//	numCodegens  The number of codegens used in codegen
+func (w *huffmanBitWriter) writeDynamicHeader(numLiterals int, numOffsets int, numCodegens int, isEof bool) {
+	if w.err != nil {
+		return
+	}
+	var firstBits int32 = 4
+	if isEof {
+		firstBits = 5
+	}
+	w.writeBits(firstBits, 3)
+	w.writeBits(int32(numLiterals-257), 5)
+	w.writeBits(int32(numOffsets-1), 5)
+	w.writeBits(int32(numCodegens-4), 4)
+
+	for i := 0; i < numCodegens; i++ {
+		value := uint(w.codegenEncoding.codes[codegenOrder[i]].len())
+		w.writeBits(int32(value), 3)
+	}
+
+	i := 0
+	for {
+		var codeWord = uint32(w.codegen[i])
+		i++
+		if codeWord == badCode {
+			break
+		}
+		w.writeCode(w.codegenEncoding.codes[codeWord])
+
+		switch codeWord {
+		case 16:
+			w.writeBits(int32(w.codegen[i]), 2)
+			i++
+		case 17:
+			w.writeBits(int32(w.codegen[i]), 3)
+			i++
+		case 18:
+			w.writeBits(int32(w.codegen[i]), 7)
+			i++
+		}
+	}
+}
+
+// writeStoredHeader will write a stored header.
+// If the stored block is only used for EOF,
+// it is replaced with a fixed huffman block.
+func (w *huffmanBitWriter) writeStoredHeader(length int, isEof bool) {
+	if w.err != nil {
+		return
+	}
+	if w.lastHeader > 0 {
+		// We owe an EOB
+		w.writeCode(w.literalEncoding.codes[endBlockMarker])
+		w.lastHeader = 0
+	}
+
+	// To write EOF, use a fixed encoding block. 10 bits instead of 5 bytes.
+	if length == 0 && isEof {
+		w.writeFixedHeader(isEof)
+		// EOB: 7 bits, value: 0
+		w.writeBits(0, 7)
+		w.flush()
+		return
+	}
+
+	var flag int32
+	if isEof {
+		flag = 1
+	}
+	w.writeBits(flag, 3)
+	w.flush()
+	w.writeBits(int32(length), 16)
+	w.writeBits(int32(^uint16(length)), 16)
+}
+
+func (w *huffmanBitWriter) writeFixedHeader(isEof bool) {
+	if w.err != nil {
+		return
+	}
+	if w.lastHeader > 0 {
+		// We owe an EOB
+		w.writeCode(w.literalEncoding.codes[endBlockMarker])
+		w.lastHeader = 0
+	}
+
+	// Indicate that we are a fixed Huffman block
+	var value int32 = 2
+	if isEof {
+		value = 3
+	}
+	w.writeBits(value, 3)
+}
+
+// writeBlock will write a block of tokens with the smallest encoding.
+// The original input can be supplied, and if the huffman encoded data
+// is larger than the original bytes, the data will be written as a
+// stored block.
+// If the input is nil, the tokens will always be Huffman encoded.
+func (w *huffmanBitWriter) writeBlock(tokens *tokens, eof bool, input []byte) {
+	if w.err != nil {
+		return
+	}
+
+	tokens.AddEOB()
+	if w.lastHeader > 0 {
+		// We owe an EOB
+		w.writeCode(w.literalEncoding.codes[endBlockMarker])
+		w.lastHeader = 0
+	}
+	numLiterals, numOffsets := w.indexTokens(tokens, false)
+	w.generate()
+	var extraBits int
+	storedSize, storable := w.storedSize(input)
+	if storable {
+		extraBits = w.extraBitSize()
+	}
+
+	// Figure out smallest code.
+	// Fixed Huffman baseline.
+	var literalEncoding = fixedLiteralEncoding
+	var offsetEncoding = fixedOffsetEncoding
+	var size = math.MaxInt32
+	if tokens.n < maxPredefinedTokens {
+		size = w.fixedSize(extraBits)
+	}
+
+	// Dynamic Huffman?
+	var numCodegens int
+
+	// Generate codegen and codegenFrequencies, which indicates how to encode
+	// the literalEncoding and the offsetEncoding.
+	w.generateCodegen(numLiterals, numOffsets, w.literalEncoding, w.offsetEncoding)
+	w.codegenEncoding.generate(w.codegenFreq[:], 7)
+	dynamicSize, numCodegens := w.dynamicSize(w.literalEncoding, w.offsetEncoding, extraBits)
+
+	if dynamicSize < size {
+		size = dynamicSize
+		literalEncoding = w.literalEncoding
+		offsetEncoding = w.offsetEncoding
+	}
+
+	// Stored bytes?
+	if storable && storedSize <= size {
+		w.writeStoredHeader(len(input), eof)
+		w.writeBytes(input)
+		return
+	}
+
+	// Huffman.
+	if literalEncoding == fixedLiteralEncoding {
+		w.writeFixedHeader(eof)
+	} else {
+		w.writeDynamicHeader(numLiterals, numOffsets, numCodegens, eof)
+	}
+
+	// Write the tokens.
+	w.writeTokens(tokens.Slice(), literalEncoding.codes, offsetEncoding.codes)
+}
+
+// writeBlockDynamic encodes a block using a dynamic Huffman table.
+// This should be used if the symbols used have a disproportionate
+// histogram distribution.
+// If input is supplied and the compression savings are below 1/16th of the
+// input size the block is stored.
+func (w *huffmanBitWriter) writeBlockDynamic(tokens *tokens, eof bool, input []byte, sync bool) {
+	if w.err != nil {
+		return
+	}
+
+	sync = sync || eof
+	if sync {
+		tokens.AddEOB()
+	}
+
+	// We cannot reuse pure huffman table, and must mark as EOF.
+	if (w.lastHuffMan || eof) && w.lastHeader > 0 {
+		// We will not try to reuse.
+		w.writeCode(w.literalEncoding.codes[endBlockMarker])
+		w.lastHeader = 0
+		w.lastHuffMan = false
+	}
+
+	// fillReuse enables filling of empty values.
+	// This will make encodings always reusable without testing.
+	// However, this does not appear to benefit on most cases.
+	const fillReuse = false
+
+	// Check if we can reuse...
+	if !fillReuse && w.lastHeader > 0 && !w.canReuse(tokens) {
+		w.writeCode(w.literalEncoding.codes[endBlockMarker])
+		w.lastHeader = 0
+	}
+
+	numLiterals, numOffsets := w.indexTokens(tokens, !sync)
+	extraBits := 0
+	ssize, storable := w.storedSize(input)
+
+	const usePrefs = true
+	if storable || w.lastHeader > 0 {
+		extraBits = w.extraBitSize()
+	}
+
+	var size int
+
+	// Check if we should reuse.
+	if w.lastHeader > 0 {
+		// Estimate size for using a new table.
+		// Use the previous header size as the best estimate.
+		newSize := w.lastHeader + tokens.EstimatedBits()
+		newSize += int(w.literalEncoding.codes[endBlockMarker].len()) + newSize>>w.logNewTablePenalty
+
+		// The estimated size is calculated as an optimal table.
+		// We add a penalty to make it more realistic and re-use a bit more.
+		reuseSize := w.dynamicReuseSize(w.literalEncoding, w.offsetEncoding) + extraBits
+
+		// Check if a new table is better.
+		if newSize < reuseSize {
+			// Write the EOB we owe.
+			w.writeCode(w.literalEncoding.codes[endBlockMarker])
+			size = newSize
+			w.lastHeader = 0
+		} else {
+			size = reuseSize
+		}
+
+		if tokens.n < maxPredefinedTokens {
+			if preSize := w.fixedSize(extraBits) + 7; usePrefs && preSize < size {
+				// Check if we get a reasonable size decrease.
+				if storable && ssize <= size {
+					w.writeStoredHeader(len(input), eof)
+					w.writeBytes(input)
+					return
+				}
+				w.writeFixedHeader(eof)
+				if !sync {
+					tokens.AddEOB()
+				}
+				w.writeTokens(tokens.Slice(), fixedLiteralEncoding.codes, fixedOffsetEncoding.codes)
+				return
+			}
+		}
+		// Check if we get a reasonable size decrease.
+		if storable && ssize <= size {
+			w.writeStoredHeader(len(input), eof)
+			w.writeBytes(input)
+			return
+		}
+	}
+
+	// We want a new block/table
+	if w.lastHeader == 0 {
+		if fillReuse && !sync {
+			w.fillTokens()
+			numLiterals, numOffsets = maxNumLit, maxNumDist
+		} else {
+			w.literalFreq[endBlockMarker] = 1
+		}
+
+		w.generate()
+		// Generate codegen and codegenFrequencies, which indicates how to encode
+		// the literalEncoding and the offsetEncoding.
+		w.generateCodegen(numLiterals, numOffsets, w.literalEncoding, w.offsetEncoding)
+		w.codegenEncoding.generate(w.codegenFreq[:], 7)
+
+		var numCodegens int
+		if fillReuse && !sync {
+			// Reindex for accurate size...
+			w.indexTokens(tokens, true)
+		}
+		size, numCodegens = w.dynamicSize(w.literalEncoding, w.offsetEncoding, extraBits)
+
+		// Store predefined, if we don't get a reasonable improvement.
+		if tokens.n < maxPredefinedTokens {
+			if preSize := w.fixedSize(extraBits); usePrefs && preSize <= size {
+				// Store bytes, if we don't get an improvement.
+				if storable && ssize <= preSize {
+					w.writeStoredHeader(len(input), eof)
+					w.writeBytes(input)
+					return
+				}
+				w.writeFixedHeader(eof)
+				if !sync {
+					tokens.AddEOB()
+				}
+				w.writeTokens(tokens.Slice(), fixedLiteralEncoding.codes, fixedOffsetEncoding.codes)
+				return
+			}
+		}
+
+		if storable && ssize <= size {
+			// Store bytes, if we don't get an improvement.
+			w.writeStoredHeader(len(input), eof)
+			w.writeBytes(input)
+			return
+		}
+
+		// Write Huffman table.
+		w.writeDynamicHeader(numLiterals, numOffsets, numCodegens, eof)
+		if !sync {
+			w.lastHeader, _ = w.headerSize()
+		}
+		w.lastHuffMan = false
+	}
+
+	if sync {
+		w.lastHeader = 0
+	}
+	// Write the tokens.
+	w.writeTokens(tokens.Slice(), w.literalEncoding.codes, w.offsetEncoding.codes)
+}
+
+func (w *huffmanBitWriter) fillTokens() {
+	for i, v := range w.literalFreq[:literalCount] {
+		if v == 0 {
+			w.literalFreq[i] = 1
+		}
+	}
+	for i, v := range w.offsetFreq[:offsetCodeCount] {
+		if v == 0 {
+			w.offsetFreq[i] = 1
+		}
+	}
+}
+
+// indexTokens indexes a slice of tokens, and updates
+// literalFreq and offsetFreq, and generates literalEncoding
+// and offsetEncoding.
+// The number of literal and offset tokens is returned.
+func (w *huffmanBitWriter) indexTokens(t *tokens, filled bool) (numLiterals, numOffsets int) {
+	//copy(w.literalFreq[:], t.litHist[:])
+	*(*[256]uint16)(w.literalFreq[:]) = t.litHist
+	//copy(w.literalFreq[256:], t.extraHist[:])
+	*(*[32]uint16)(w.literalFreq[256:]) = t.extraHist
+	w.offsetFreq = t.offHist
+
+	if t.n == 0 {
+		return
+	}
+	if filled {
+		return maxNumLit, maxNumDist
+	}
+	// get the number of literals
+	numLiterals = len(w.literalFreq)
+	for w.literalFreq[numLiterals-1] == 0 {
+		numLiterals--
+	}
+	// get the number of offsets
+	numOffsets = len(w.offsetFreq)
+	for numOffsets > 0 && w.offsetFreq[numOffsets-1] == 0 {
+		numOffsets--
+	}
+	if numOffsets == 0 {
+		// We haven't found a single match. If we want to go with the dynamic encoding,
+		// we should count at least one offset to be sure that the offset huffman tree could be encoded.
+		w.offsetFreq[0] = 1
+		numOffsets = 1
+	}
+	return
+}
+
+func (w *huffmanBitWriter) generate() {
+	w.literalEncoding.generate(w.literalFreq[:literalCount], 15)
+	w.offsetEncoding.generate(w.offsetFreq[:offsetCodeCount], 15)
+}
+
+// writeTokens writes a slice of tokens to the output.
+// codes for literal and offset encoding must be supplied.
+func (w *huffmanBitWriter) writeTokens(tokens []token, leCodes, oeCodes []hcode) {
+	if w.err != nil {
+		return
+	}
+	if len(tokens) == 0 {
+		return
+	}
+
+	// Only last token should be endBlockMarker.
+	var deferEOB bool
+	if tokens[len(tokens)-1] == endBlockMarker {
+		tokens = tokens[:len(tokens)-1]
+		deferEOB = true
+	}
+
+	// Create slices up to the next power of two to avoid bounds checks.
+	lits := leCodes[:256]
+	offs := oeCodes[:32]
+	lengths := leCodes[lengthCodesStart:]
+	lengths = lengths[:32]
+
+	// Go 1.16 LOVES having these on stack.
+	bits, nbits, nbytes := w.bits, w.nbits, w.nbytes
+
+	for _, t := range tokens {
+		if t < 256 {
+			//w.writeCode(lits[t.literal()])
+			c := lits[t]
+			bits |= c.code64() << (nbits & 63)
+			nbits += c.len()
+			if nbits >= 48 {
+				le.Store64(w.bytes[nbytes:], bits)
+				//*(*uint64)(unsafe.Pointer(&w.bytes[nbytes])) = bits
+				bits >>= 48
+				nbits -= 48
+				nbytes += 6
+				if nbytes >= bufferFlushSize {
+					if w.err != nil {
+						nbytes = 0
+						return
+					}
+					_, w.err = w.writer.Write(w.bytes[:nbytes])
+					nbytes = 0
+				}
+			}
+			continue
+		}
+
+		// Write the length
+		length := t.length()
+		lengthCode := lengthCode(length) & 31
+		if false {
+			w.writeCode(lengths[lengthCode])
+		} else {
+			// inlined
+			c := lengths[lengthCode]
+			bits |= c.code64() << (nbits & 63)
+			nbits += c.len()
+			if nbits >= 48 {
+				le.Store64(w.bytes[nbytes:], bits)
+				//*(*uint64)(unsafe.Pointer(&w.bytes[nbytes])) = bits
+				bits >>= 48
+				nbits -= 48
+				nbytes += 6
+				if nbytes >= bufferFlushSize {
+					if w.err != nil {
+						nbytes = 0
+						return
+					}
+					_, w.err = w.writer.Write(w.bytes[:nbytes])
+					nbytes = 0
+				}
+			}
+		}
+
+		if lengthCode >= lengthExtraBitsMinCode {
+			extraLengthBits := lengthExtraBits[lengthCode]
+			//w.writeBits(extraLength, extraLengthBits)
+			extraLength := int32(length - lengthBase[lengthCode])
+			bits |= uint64(extraLength) << (nbits & 63)
+			nbits += extraLengthBits
+			if nbits >= 48 {
+				le.Store64(w.bytes[nbytes:], bits)
+				//*(*uint64)(unsafe.Pointer(&w.bytes[nbytes])) = bits
+				bits >>= 48
+				nbits -= 48
+				nbytes += 6
+				if nbytes >= bufferFlushSize {
+					if w.err != nil {
+						nbytes = 0
+						return
+					}
+					_, w.err = w.writer.Write(w.bytes[:nbytes])
+					nbytes = 0
+				}
+			}
+		}
+		// Write the offset
+		offset := t.offset()
+		offsetCode := (offset >> 16) & 31
+		if false {
+			w.writeCode(offs[offsetCode])
+		} else {
+			// inlined
+			c := offs[offsetCode]
+			bits |= c.code64() << (nbits & 63)
+			nbits += c.len()
+			if nbits >= 48 {
+				le.Store64(w.bytes[nbytes:], bits)
+				//*(*uint64)(unsafe.Pointer(&w.bytes[nbytes])) = bits
+				bits >>= 48
+				nbits -= 48
+				nbytes += 6
+				if nbytes >= bufferFlushSize {
+					if w.err != nil {
+						nbytes = 0
+						return
+					}
+					_, w.err = w.writer.Write(w.bytes[:nbytes])
+					nbytes = 0
+				}
+			}
+		}
+
+		if offsetCode >= offsetExtraBitsMinCode {
+			offsetComb := offsetCombined[offsetCode]
+			//w.writeBits(extraOffset, extraOffsetBits)
+			bits |= uint64((offset-(offsetComb>>8))&matchOffsetOnlyMask) << (nbits & 63)
+			nbits += uint8(offsetComb)
+			if nbits >= 48 {
+				le.Store64(w.bytes[nbytes:], bits)
+				//*(*uint64)(unsafe.Pointer(&w.bytes[nbytes])) = bits
+				bits >>= 48
+				nbits -= 48
+				nbytes += 6
+				if nbytes >= bufferFlushSize {
+					if w.err != nil {
+						nbytes = 0
+						return
+					}
+					_, w.err = w.writer.Write(w.bytes[:nbytes])
+					nbytes = 0
+				}
+			}
+		}
+	}
+	// Restore...
+	w.bits, w.nbits, w.nbytes = bits, nbits, nbytes
+
+	if deferEOB {
+		w.writeCode(leCodes[endBlockMarker])
+	}
+}
+
+// huffOffset is a static offset encoder used for huffman only encoding.
+// It can be reused since we will not be encoding offset values.
+var huffOffset *huffmanEncoder
+
+func init() {
+	w := newHuffmanBitWriter(nil)
+	w.offsetFreq[0] = 1
+	huffOffset = newHuffmanEncoder(offsetCodeCount)
+	huffOffset.generate(w.offsetFreq[:offsetCodeCount], 15)
+}
+
+// writeBlockHuff encodes a block of bytes as either
+// Huffman encoded literals or uncompressed bytes if the
+// results only gains very little from compression.
+func (w *huffmanBitWriter) writeBlockHuff(eof bool, input []byte, sync bool) {
+	if w.err != nil {
+		return
+	}
+
+	// Clear histogram
+	for i := range w.literalFreq[:] {
+		w.literalFreq[i] = 0
+	}
+	if !w.lastHuffMan {
+		for i := range w.offsetFreq[:] {
+			w.offsetFreq[i] = 0
+		}
+	}
+
+	const numLiterals = endBlockMarker + 1
+	const numOffsets = 1
+
+	// Add everything as literals
+	// We have to estimate the header size.
+	// Assume header is around 70 bytes:
+	// https://stackoverflow.com/a/25454430
+	const guessHeaderSizeBits = 70 * 8
+	histogram(input, w.literalFreq[:numLiterals])
+	ssize, storable := w.storedSize(input)
+	if storable && len(input) > 1024 {
+		// Quick check for incompressible content.
+		abs := float64(0)
+		avg := float64(len(input)) / 256
+		max := float64(len(input) * 2)
+		for _, v := range w.literalFreq[:256] {
+			diff := float64(v) - avg
+			abs += diff * diff
+			if abs > max {
+				break
+			}
+		}
+		if abs < max {
+			if debugDeflate {
+				fmt.Println("stored", abs, "<", max)
+			}
+			// No chance we can compress this...
+			w.writeStoredHeader(len(input), eof)
+			w.writeBytes(input)
+			return
+		}
+	}
+	w.literalFreq[endBlockMarker] = 1
+	w.tmpLitEncoding.generate(w.literalFreq[:numLiterals], 15)
+	estBits := w.tmpLitEncoding.canReuseBits(w.literalFreq[:numLiterals])
+	if estBits < math.MaxInt32 {
+		estBits += w.lastHeader
+		if w.lastHeader == 0 {
+			estBits += guessHeaderSizeBits
+		}
+		estBits += estBits >> w.logNewTablePenalty
+	}
+
+	// Store bytes, if we don't get a reasonable improvement.
+	if storable && ssize <= estBits {
+		if debugDeflate {
+			fmt.Println("stored,", ssize, "<=", estBits)
+		}
+		w.writeStoredHeader(len(input), eof)
+		w.writeBytes(input)
+		return
+	}
+
+	if w.lastHeader > 0 {
+		reuseSize := w.literalEncoding.canReuseBits(w.literalFreq[:256])
+
+		if estBits < reuseSize {
+			if debugDeflate {
+				fmt.Println("NOT reusing, reuse:", reuseSize/8, "> new:", estBits/8, "header est:", w.lastHeader/8, "bytes")
+			}
+			// We owe an EOB
+			w.writeCode(w.literalEncoding.codes[endBlockMarker])
+			w.lastHeader = 0
+		} else if debugDeflate {
+			fmt.Println("reusing, reuse:", reuseSize/8, "> new:", estBits/8, "- header est:", w.lastHeader/8)
+		}
+	}
+
+	count := 0
+	if w.lastHeader == 0 {
+		// Use the temp encoding, so swap.
+		w.literalEncoding, w.tmpLitEncoding = w.tmpLitEncoding, w.literalEncoding
+		// Generate codegen and codegenFrequencies, which indicates how to encode
+		// the literalEncoding and the offsetEncoding.
+		w.generateCodegen(numLiterals, numOffsets, w.literalEncoding, huffOffset)
+		w.codegenEncoding.generate(w.codegenFreq[:], 7)
+		numCodegens := w.codegens()
+
+		// Huffman.
+		w.writeDynamicHeader(numLiterals, numOffsets, numCodegens, eof)
+		w.lastHuffMan = true
+		w.lastHeader, _ = w.headerSize()
+		if debugDeflate {
+			count += w.lastHeader
+			fmt.Println("header:", count/8)
+		}
+	}
+
+	encoding := w.literalEncoding.codes[:256]
+	// Go 1.16 LOVES having these on stack. At least 1.5x the speed.
+	bits, nbits, nbytes := w.bits, w.nbits, w.nbytes
+
+	if debugDeflate {
+		count -= int(nbytes)*8 + int(nbits)
+	}
+	// Unroll, write 3 codes/loop.
+	// Fastest number of unrolls.
+	for len(input) > 3 {
+		// We must have at least 48 bits free.
+		if nbits >= 8 {
+			n := nbits >> 3
+			le.Store64(w.bytes[nbytes:], bits)
+			bits >>= (n * 8) & 63
+			nbits -= n * 8
+			nbytes += n
+		}
+		if nbytes >= bufferFlushSize {
+			if w.err != nil {
+				nbytes = 0
+				return
+			}
+			if debugDeflate {
+				count += int(nbytes) * 8
+			}
+			_, w.err = w.writer.Write(w.bytes[:nbytes])
+			nbytes = 0
+		}
+		a, b := encoding[input[0]], encoding[input[1]]
+		bits |= a.code64() << (nbits & 63)
+		bits |= b.code64() << ((nbits + a.len()) & 63)
+		c := encoding[input[2]]
+		nbits += b.len() + a.len()
+		bits |= c.code64() << (nbits & 63)
+		nbits += c.len()
+		input = input[3:]
+	}
+
+	// Remaining...
+	for _, t := range input {
+		if nbits >= 48 {
+			le.Store64(w.bytes[nbytes:], bits)
+			//*(*uint64)(unsafe.Pointer(&w.bytes[nbytes])) = bits
+			bits >>= 48
+			nbits -= 48
+			nbytes += 6
+			if nbytes >= bufferFlushSize {
+				if w.err != nil {
+					nbytes = 0
+					return
+				}
+				if debugDeflate {
+					count += int(nbytes) * 8
+				}
+				_, w.err = w.writer.Write(w.bytes[:nbytes])
+				nbytes = 0
+			}
+		}
+		// Bitwriting inlined, ~30% speedup
+		c := encoding[t]
+		bits |= c.code64() << (nbits & 63)
+
+		nbits += c.len()
+		if debugDeflate {
+			count += int(c.len())
+		}
+	}
+	// Restore...
+	w.bits, w.nbits, w.nbytes = bits, nbits, nbytes
+
+	if debugDeflate {
+		nb := count + int(nbytes)*8 + int(nbits)
+		fmt.Println("wrote", nb, "bits,", nb/8, "bytes.")
+	}
+	// Flush if needed to have space.
+	if w.nbits >= 48 {
+		w.writeOutBits()
+	}
+
+	if eof || sync {
+		w.writeCode(w.literalEncoding.codes[endBlockMarker])
+		w.lastHeader = 0
+		w.lastHuffMan = false
+	}
+}
diff --git a/vendor/github.com/klauspost/compress/flate/huffman_code.go b/vendor/github.com/klauspost/compress/flate/huffman_code.go
new file mode 100644
index 0000000..be7b58b
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/huffman_code.go
@@ -0,0 +1,417 @@
+// Copyright 2009 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package flate
+
+import (
+	"math"
+	"math/bits"
+)
+
+const (
+	maxBitsLimit = 16
+	// number of valid literals
+	literalCount = 286
+)
+
+// hcode is a huffman code with a bit code and bit length.
+type hcode uint32
+
+func (h hcode) len() uint8 {
+	return uint8(h)
+}
+
+func (h hcode) code64() uint64 {
+	return uint64(h >> 8)
+}
+
+func (h hcode) zero() bool {
+	return h == 0
+}
+
+type huffmanEncoder struct {
+	codes    []hcode
+	bitCount [17]int32
+
+	// Allocate a reusable buffer with the longest possible frequency table.
+	// Possible lengths are codegenCodeCount, offsetCodeCount and literalCount.
+	// The largest of these is literalCount, so we allocate for that case.
+	freqcache [literalCount + 1]literalNode
+}
+
+type literalNode struct {
+	literal uint16
+	freq    uint16
+}
+
+// A levelInfo describes the state of the constructed tree for a given depth.
+type levelInfo struct {
+	// Our level.  for better printing
+	level int32
+
+	// The frequency of the last node at this level
+	lastFreq int32
+
+	// The frequency of the next character to add to this level
+	nextCharFreq int32
+
+	// The frequency of the next pair (from level below) to add to this level.
+	// Only valid if the "needed" value of the next lower level is 0.
+	nextPairFreq int32
+
+	// The number of chains remaining to generate for this level before moving
+	// up to the next level
+	needed int32
+}
+
+// set sets the code and length of an hcode.
+func (h *hcode) set(code uint16, length uint8) {
+	*h = hcode(length) | (hcode(code) << 8)
+}
+
+func newhcode(code uint16, length uint8) hcode {
+	return hcode(length) | (hcode(code) << 8)
+}
+
+func reverseBits(number uint16, bitLength byte) uint16 {
+	return bits.Reverse16(number << ((16 - bitLength) & 15))
+}
+
+func maxNode() literalNode { return literalNode{math.MaxUint16, math.MaxUint16} }
+
+func newHuffmanEncoder(size int) *huffmanEncoder {
+	// Make capacity to next power of two.
+	c := uint(bits.Len32(uint32(size - 1)))
+	return &huffmanEncoder{codes: make([]hcode, size, 1<<c)}
+}
+
+// Generates a HuffmanCode corresponding to the fixed literal table
+func generateFixedLiteralEncoding() *huffmanEncoder {
+	h := newHuffmanEncoder(literalCount)
+	codes := h.codes
+	var ch uint16
+	for ch = 0; ch < literalCount; ch++ {
+		var bits uint16
+		var size uint8
+		switch {
+		case ch < 144:
+			// size 8, 000110000  .. 10111111
+			bits = ch + 48
+			size = 8
+		case ch < 256:
+			// size 9, 110010000 .. 111111111
+			bits = ch + 400 - 144
+			size = 9
+		case ch < 280:
+			// size 7, 0000000 .. 0010111
+			bits = ch - 256
+			size = 7
+		default:
+			// size 8, 11000000 .. 11000111
+			bits = ch + 192 - 280
+			size = 8
+		}
+		codes[ch] = newhcode(reverseBits(bits, size), size)
+	}
+	return h
+}
+
+func generateFixedOffsetEncoding() *huffmanEncoder {
+	h := newHuffmanEncoder(30)
+	codes := h.codes
+	for ch := range codes {
+		codes[ch] = newhcode(reverseBits(uint16(ch), 5), 5)
+	}
+	return h
+}
+
+var fixedLiteralEncoding = generateFixedLiteralEncoding()
+var fixedOffsetEncoding = generateFixedOffsetEncoding()
+
+func (h *huffmanEncoder) bitLength(freq []uint16) int {
+	var total int
+	for i, f := range freq {
+		if f != 0 {
+			total += int(f) * int(h.codes[i].len())
+		}
+	}
+	return total
+}
+
+func (h *huffmanEncoder) bitLengthRaw(b []byte) int {
+	var total int
+	for _, f := range b {
+		total += int(h.codes[f].len())
+	}
+	return total
+}
+
+// canReuseBits returns the number of bits or math.MaxInt32 if the encoder cannot be reused.
+func (h *huffmanEncoder) canReuseBits(freq []uint16) int {
+	var total int
+	for i, f := range freq {
+		if f != 0 {
+			code := h.codes[i]
+			if code.zero() {
+				return math.MaxInt32
+			}
+			total += int(f) * int(code.len())
+		}
+	}
+	return total
+}
+
+// Return the number of literals assigned to each bit size in the Huffman encoding
+//
+// This method is only called when list.length >= 3
+// The cases of 0, 1, and 2 literals are handled by special case code.
+//
+// list  An array of the literals with non-zero frequencies
+//
+//	and their associated frequencies. The array is in order of increasing
+//	frequency, and has as its last element a special element with frequency
+//	MaxInt32
+//
+// maxBits     The maximum number of bits that should be used to encode any literal.
+//
+//	Must be less than 16.
+//
+// return      An integer array in which array[i] indicates the number of literals
+//
+//	that should be encoded in i bits.
+func (h *huffmanEncoder) bitCounts(list []literalNode, maxBits int32) []int32 {
+	if maxBits >= maxBitsLimit {
+		panic("flate: maxBits too large")
+	}
+	n := int32(len(list))
+	list = list[0 : n+1]
+	list[n] = maxNode()
+
+	// The tree can't have greater depth than n - 1, no matter what. This
+	// saves a little bit of work in some small cases
+	if maxBits > n-1 {
+		maxBits = n - 1
+	}
+
+	// Create information about each of the levels.
+	// A bogus "Level 0" whose sole purpose is so that
+	// level1.prev.needed==0.  This makes level1.nextPairFreq
+	// be a legitimate value that never gets chosen.
+	var levels [maxBitsLimit]levelInfo
+	// leafCounts[i] counts the number of literals at the left
+	// of ancestors of the rightmost node at level i.
+	// leafCounts[i][j] is the number of literals at the left
+	// of the level j ancestor.
+	var leafCounts [maxBitsLimit][maxBitsLimit]int32
+
+	// Descending to only have 1 bounds check.
+	l2f := int32(list[2].freq)
+	l1f := int32(list[1].freq)
+	l0f := int32(list[0].freq) + int32(list[1].freq)
+
+	for level := int32(1); level <= maxBits; level++ {
+		// For every level, the first two items are the first two characters.
+		// We initialize the levels as if we had already figured this out.
+		levels[level] = levelInfo{
+			level:        level,
+			lastFreq:     l1f,
+			nextCharFreq: l2f,
+			nextPairFreq: l0f,
+		}
+		leafCounts[level][level] = 2
+		if level == 1 {
+			levels[level].nextPairFreq = math.MaxInt32
+		}
+	}
+
+	// We need a total of 2*n - 2 items at top level and have already generated 2.
+	levels[maxBits].needed = 2*n - 4
+
+	level := uint32(maxBits)
+	for level < 16 {
+		l := &levels[level]
+		if l.nextPairFreq == math.MaxInt32 && l.nextCharFreq == math.MaxInt32 {
+			// We've run out of both leafs and pairs.
+			// End all calculations for this level.
+			// To make sure we never come back to this level or any lower level,
+			// set nextPairFreq impossibly large.
+			l.needed = 0
+			levels[level+1].nextPairFreq = math.MaxInt32
+			level++
+			continue
+		}
+
+		prevFreq := l.lastFreq
+		if l.nextCharFreq < l.nextPairFreq {
+			// The next item on this row is a leaf node.
+			n := leafCounts[level][level] + 1
+			l.lastFreq = l.nextCharFreq
+			// Lower leafCounts are the same of the previous node.
+			leafCounts[level][level] = n
+			e := list[n]
+			if e.literal < math.MaxUint16 {
+				l.nextCharFreq = int32(e.freq)
+			} else {
+				l.nextCharFreq = math.MaxInt32
+			}
+		} else {
+			// The next item on this row is a pair from the previous row.
+			// nextPairFreq isn't valid until we generate two
+			// more values in the level below
+			l.lastFreq = l.nextPairFreq
+			// Take leaf counts from the lower level, except counts[level] remains the same.
+			if true {
+				save := leafCounts[level][level]
+				leafCounts[level] = leafCounts[level-1]
+				leafCounts[level][level] = save
+			} else {
+				copy(leafCounts[level][:level], leafCounts[level-1][:level])
+			}
+			levels[l.level-1].needed = 2
+		}
+
+		if l.needed--; l.needed == 0 {
+			// We've done everything we need to do for this level.
+			// Continue calculating one level up. Fill in nextPairFreq
+			// of that level with the sum of the two nodes we've just calculated on
+			// this level.
+			if l.level == maxBits {
+				// All done!
+				break
+			}
+			levels[l.level+1].nextPairFreq = prevFreq + l.lastFreq
+			level++
+		} else {
+			// If we stole from below, move down temporarily to replenish it.
+			for levels[level-1].needed > 0 {
+				level--
+			}
+		}
+	}
+
+	// Somethings is wrong if at the end, the top level is null or hasn't used
+	// all of the leaves.
+	if leafCounts[maxBits][maxBits] != n {
+		panic("leafCounts[maxBits][maxBits] != n")
+	}
+
+	bitCount := h.bitCount[:maxBits+1]
+	bits := 1
+	counts := &leafCounts[maxBits]
+	for level := maxBits; level > 0; level-- {
+		// chain.leafCount gives the number of literals requiring at least "bits"
+		// bits to encode.
+		bitCount[bits] = counts[level] - counts[level-1]
+		bits++
+	}
+	return bitCount
+}
+
+// Look at the leaves and assign them a bit count and an encoding as specified
+// in RFC 1951 3.2.2
+func (h *huffmanEncoder) assignEncodingAndSize(bitCount []int32, list []literalNode) {
+	code := uint16(0)
+	for n, bits := range bitCount {
+		code <<= 1
+		if n == 0 || bits == 0 {
+			continue
+		}
+		// The literals list[len(list)-bits] .. list[len(list)-bits]
+		// are encoded using "bits" bits, and get the values
+		// code, code + 1, ....  The code values are
+		// assigned in literal order (not frequency order).
+		chunk := list[len(list)-int(bits):]
+
+		sortByLiteral(chunk)
+		for _, node := range chunk {
+			h.codes[node.literal] = newhcode(reverseBits(code, uint8(n)), uint8(n))
+			code++
+		}
+		list = list[0 : len(list)-int(bits)]
+	}
+}
+
+// Update this Huffman Code object to be the minimum code for the specified frequency count.
+//
+// freq  An array of frequencies, in which frequency[i] gives the frequency of literal i.
+// maxBits  The maximum number of bits to use for any literal.
+func (h *huffmanEncoder) generate(freq []uint16, maxBits int32) {
+	list := h.freqcache[:len(freq)+1]
+	codes := h.codes[:len(freq)]
+	// Number of non-zero literals
+	count := 0
+	// Set list to be the set of all non-zero literals and their frequencies
+	for i, f := range freq {
+		if f != 0 {
+			list[count] = literalNode{uint16(i), f}
+			count++
+		} else {
+			codes[i] = 0
+		}
+	}
+	list[count] = literalNode{}
+
+	list = list[:count]
+	if count <= 2 {
+		// Handle the small cases here, because they are awkward for the general case code. With
+		// two or fewer literals, everything has bit length 1.
+		for i, node := range list {
+			// "list" is in order of increasing literal value.
+			h.codes[node.literal].set(uint16(i), 1)
+		}
+		return
+	}
+	sortByFreq(list)
+
+	// Get the number of literals for each bit count
+	bitCount := h.bitCounts(list, maxBits)
+	// And do the assignment
+	h.assignEncodingAndSize(bitCount, list)
+}
+
+// atLeastOne clamps the result between 1 and 15.
+func atLeastOne(v float32) float32 {
+	if v < 1 {
+		return 1
+	}
+	if v > 15 {
+		return 15
+	}
+	return v
+}
+
+func histogram(b []byte, h []uint16) {
+	if true && len(b) >= 8<<10 {
+		// Split for bigger inputs
+		histogramSplit(b, h)
+	} else {
+		h = h[:256]
+		for _, t := range b {
+			h[t]++
+		}
+	}
+}
+
+func histogramSplit(b []byte, h []uint16) {
+	// Tested, and slightly faster than 2-way.
+	// Writing to separate arrays and combining is also slightly slower.
+	h = h[:256]
+	for len(b)&3 != 0 {
+		h[b[0]]++
+		b = b[1:]
+	}
+	n := len(b) / 4
+	x, y, z, w := b[:n], b[n:], b[n+n:], b[n+n+n:]
+	y, z, w = y[:len(x)], z[:len(x)], w[:len(x)]
+	for i, t := range x {
+		v0 := &h[t]
+		v1 := &h[y[i]]
+		v3 := &h[w[i]]
+		v2 := &h[z[i]]
+		*v0++
+		*v1++
+		*v2++
+		*v3++
+	}
+}
diff --git a/vendor/github.com/klauspost/compress/flate/huffman_sortByFreq.go b/vendor/github.com/klauspost/compress/flate/huffman_sortByFreq.go
new file mode 100644
index 0000000..6c05ba8
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/huffman_sortByFreq.go
@@ -0,0 +1,159 @@
+// Copyright 2009 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package flate
+
+// Sort sorts data.
+// It makes one call to data.Len to determine n, and O(n*log(n)) calls to
+// data.Less and data.Swap. The sort is not guaranteed to be stable.
+func sortByFreq(data []literalNode) {
+	n := len(data)
+	quickSortByFreq(data, 0, n, maxDepth(n))
+}
+
+func quickSortByFreq(data []literalNode, a, b, maxDepth int) {
+	for b-a > 12 { // Use ShellSort for slices <= 12 elements
+		if maxDepth == 0 {
+			heapSort(data, a, b)
+			return
+		}
+		maxDepth--
+		mlo, mhi := doPivotByFreq(data, a, b)
+		// Avoiding recursion on the larger subproblem guarantees
+		// a stack depth of at most lg(b-a).
+		if mlo-a < b-mhi {
+			quickSortByFreq(data, a, mlo, maxDepth)
+			a = mhi // i.e., quickSortByFreq(data, mhi, b)
+		} else {
+			quickSortByFreq(data, mhi, b, maxDepth)
+			b = mlo // i.e., quickSortByFreq(data, a, mlo)
+		}
+	}
+	if b-a > 1 {
+		// Do ShellSort pass with gap 6
+		// It could be written in this simplified form cause b-a <= 12
+		for i := a + 6; i < b; i++ {
+			if data[i].freq == data[i-6].freq && data[i].literal < data[i-6].literal || data[i].freq < data[i-6].freq {
+				data[i], data[i-6] = data[i-6], data[i]
+			}
+		}
+		insertionSortByFreq(data, a, b)
+	}
+}
+
+func doPivotByFreq(data []literalNode, lo, hi int) (midlo, midhi int) {
+	m := int(uint(lo+hi) >> 1) // Written like this to avoid integer overflow.
+	if hi-lo > 40 {
+		// Tukey's ``Ninther,'' median of three medians of three.
+		s := (hi - lo) / 8
+		medianOfThreeSortByFreq(data, lo, lo+s, lo+2*s)
+		medianOfThreeSortByFreq(data, m, m-s, m+s)
+		medianOfThreeSortByFreq(data, hi-1, hi-1-s, hi-1-2*s)
+	}
+	medianOfThreeSortByFreq(data, lo, m, hi-1)
+
+	// Invariants are:
+	//	data[lo] = pivot (set up by ChoosePivot)
+	//	data[lo < i < a] < pivot
+	//	data[a <= i < b] <= pivot
+	//	data[b <= i < c] unexamined
+	//	data[c <= i < hi-1] > pivot
+	//	data[hi-1] >= pivot
+	pivot := lo
+	a, c := lo+1, hi-1
+
+	for ; a < c && (data[a].freq == data[pivot].freq && data[a].literal < data[pivot].literal || data[a].freq < data[pivot].freq); a++ {
+	}
+	b := a
+	for {
+		for ; b < c && (data[pivot].freq == data[b].freq && data[pivot].literal > data[b].literal || data[pivot].freq > data[b].freq); b++ { // data[b] <= pivot
+		}
+		for ; b < c && (data[pivot].freq == data[c-1].freq && data[pivot].literal < data[c-1].literal || data[pivot].freq < data[c-1].freq); c-- { // data[c-1] > pivot
+		}
+		if b >= c {
+			break
+		}
+		// data[b] > pivot; data[c-1] <= pivot
+		data[b], data[c-1] = data[c-1], data[b]
+		b++
+		c--
+	}
+	// If hi-c<3 then there are duplicates (by property of median of nine).
+	// Let's be a bit more conservative, and set border to 5.
+	protect := hi-c < 5
+	if !protect && hi-c < (hi-lo)/4 {
+		// Lets test some points for equality to pivot
+		dups := 0
+		if data[pivot].freq == data[hi-1].freq && data[pivot].literal > data[hi-1].literal || data[pivot].freq > data[hi-1].freq { // data[hi-1] = pivot
+			data[c], data[hi-1] = data[hi-1], data[c]
+			c++
+			dups++
+		}
+		if data[b-1].freq == data[pivot].freq && data[b-1].literal > data[pivot].literal || data[b-1].freq > data[pivot].freq { // data[b-1] = pivot
+			b--
+			dups++
+		}
+		// m-lo = (hi-lo)/2 > 6
+		// b-lo > (hi-lo)*3/4-1 > 8
+		// ==> m < b ==> data[m] <= pivot
+		if data[m].freq == data[pivot].freq && data[m].literal > data[pivot].literal || data[m].freq > data[pivot].freq { // data[m] = pivot
+			data[m], data[b-1] = data[b-1], data[m]
+			b--
+			dups++
+		}
+		// if at least 2 points are equal to pivot, assume skewed distribution
+		protect = dups > 1
+	}
+	if protect {
+		// Protect against a lot of duplicates
+		// Add invariant:
+		//	data[a <= i < b] unexamined
+		//	data[b <= i < c] = pivot
+		for {
+			for ; a < b && (data[b-1].freq == data[pivot].freq && data[b-1].literal > data[pivot].literal || data[b-1].freq > data[pivot].freq); b-- { // data[b] == pivot
+			}
+			for ; a < b && (data[a].freq == data[pivot].freq && data[a].literal < data[pivot].literal || data[a].freq < data[pivot].freq); a++ { // data[a] < pivot
+			}
+			if a >= b {
+				break
+			}
+			// data[a] == pivot; data[b-1] < pivot
+			data[a], data[b-1] = data[b-1], data[a]
+			a++
+			b--
+		}
+	}
+	// Swap pivot into middle
+	data[pivot], data[b-1] = data[b-1], data[pivot]
+	return b - 1, c
+}
+
+// Insertion sort
+func insertionSortByFreq(data []literalNode, a, b int) {
+	for i := a + 1; i < b; i++ {
+		for j := i; j > a && (data[j].freq == data[j-1].freq && data[j].literal < data[j-1].literal || data[j].freq < data[j-1].freq); j-- {
+			data[j], data[j-1] = data[j-1], data[j]
+		}
+	}
+}
+
+// quickSortByFreq, loosely following Bentley and McIlroy,
+// ``Engineering a Sort Function,'' SP&E November 1993.
+
+// medianOfThreeSortByFreq moves the median of the three values data[m0], data[m1], data[m2] into data[m1].
+func medianOfThreeSortByFreq(data []literalNode, m1, m0, m2 int) {
+	// sort 3 elements
+	if data[m1].freq == data[m0].freq && data[m1].literal < data[m0].literal || data[m1].freq < data[m0].freq {
+		data[m1], data[m0] = data[m0], data[m1]
+	}
+	// data[m0] <= data[m1]
+	if data[m2].freq == data[m1].freq && data[m2].literal < data[m1].literal || data[m2].freq < data[m1].freq {
+		data[m2], data[m1] = data[m1], data[m2]
+		// data[m0] <= data[m2] && data[m1] < data[m2]
+		if data[m1].freq == data[m0].freq && data[m1].literal < data[m0].literal || data[m1].freq < data[m0].freq {
+			data[m1], data[m0] = data[m0], data[m1]
+		}
+	}
+	// now data[m0] <= data[m1] <= data[m2]
+}
diff --git a/vendor/github.com/klauspost/compress/flate/huffman_sortByLiteral.go b/vendor/github.com/klauspost/compress/flate/huffman_sortByLiteral.go
new file mode 100644
index 0000000..93f1aea
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/huffman_sortByLiteral.go
@@ -0,0 +1,201 @@
+// Copyright 2009 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package flate
+
+// Sort sorts data.
+// It makes one call to data.Len to determine n, and O(n*log(n)) calls to
+// data.Less and data.Swap. The sort is not guaranteed to be stable.
+func sortByLiteral(data []literalNode) {
+	n := len(data)
+	quickSort(data, 0, n, maxDepth(n))
+}
+
+func quickSort(data []literalNode, a, b, maxDepth int) {
+	for b-a > 12 { // Use ShellSort for slices <= 12 elements
+		if maxDepth == 0 {
+			heapSort(data, a, b)
+			return
+		}
+		maxDepth--
+		mlo, mhi := doPivot(data, a, b)
+		// Avoiding recursion on the larger subproblem guarantees
+		// a stack depth of at most lg(b-a).
+		if mlo-a < b-mhi {
+			quickSort(data, a, mlo, maxDepth)
+			a = mhi // i.e., quickSort(data, mhi, b)
+		} else {
+			quickSort(data, mhi, b, maxDepth)
+			b = mlo // i.e., quickSort(data, a, mlo)
+		}
+	}
+	if b-a > 1 {
+		// Do ShellSort pass with gap 6
+		// It could be written in this simplified form cause b-a <= 12
+		for i := a + 6; i < b; i++ {
+			if data[i].literal < data[i-6].literal {
+				data[i], data[i-6] = data[i-6], data[i]
+			}
+		}
+		insertionSort(data, a, b)
+	}
+}
+func heapSort(data []literalNode, a, b int) {
+	first := a
+	lo := 0
+	hi := b - a
+
+	// Build heap with greatest element at top.
+	for i := (hi - 1) / 2; i >= 0; i-- {
+		siftDown(data, i, hi, first)
+	}
+
+	// Pop elements, largest first, into end of data.
+	for i := hi - 1; i >= 0; i-- {
+		data[first], data[first+i] = data[first+i], data[first]
+		siftDown(data, lo, i, first)
+	}
+}
+
+// siftDown implements the heap property on data[lo, hi).
+// first is an offset into the array where the root of the heap lies.
+func siftDown(data []literalNode, lo, hi, first int) {
+	root := lo
+	for {
+		child := 2*root + 1
+		if child >= hi {
+			break
+		}
+		if child+1 < hi && data[first+child].literal < data[first+child+1].literal {
+			child++
+		}
+		if data[first+root].literal > data[first+child].literal {
+			return
+		}
+		data[first+root], data[first+child] = data[first+child], data[first+root]
+		root = child
+	}
+}
+func doPivot(data []literalNode, lo, hi int) (midlo, midhi int) {
+	m := int(uint(lo+hi) >> 1) // Written like this to avoid integer overflow.
+	if hi-lo > 40 {
+		// Tukey's ``Ninther,'' median of three medians of three.
+		s := (hi - lo) / 8
+		medianOfThree(data, lo, lo+s, lo+2*s)
+		medianOfThree(data, m, m-s, m+s)
+		medianOfThree(data, hi-1, hi-1-s, hi-1-2*s)
+	}
+	medianOfThree(data, lo, m, hi-1)
+
+	// Invariants are:
+	//	data[lo] = pivot (set up by ChoosePivot)
+	//	data[lo < i < a] < pivot
+	//	data[a <= i < b] <= pivot
+	//	data[b <= i < c] unexamined
+	//	data[c <= i < hi-1] > pivot
+	//	data[hi-1] >= pivot
+	pivot := lo
+	a, c := lo+1, hi-1
+
+	for ; a < c && data[a].literal < data[pivot].literal; a++ {
+	}
+	b := a
+	for {
+		for ; b < c && data[pivot].literal > data[b].literal; b++ { // data[b] <= pivot
+		}
+		for ; b < c && data[pivot].literal < data[c-1].literal; c-- { // data[c-1] > pivot
+		}
+		if b >= c {
+			break
+		}
+		// data[b] > pivot; data[c-1] <= pivot
+		data[b], data[c-1] = data[c-1], data[b]
+		b++
+		c--
+	}
+	// If hi-c<3 then there are duplicates (by property of median of nine).
+	// Let's be a bit more conservative, and set border to 5.
+	protect := hi-c < 5
+	if !protect && hi-c < (hi-lo)/4 {
+		// Lets test some points for equality to pivot
+		dups := 0
+		if data[pivot].literal > data[hi-1].literal { // data[hi-1] = pivot
+			data[c], data[hi-1] = data[hi-1], data[c]
+			c++
+			dups++
+		}
+		if data[b-1].literal > data[pivot].literal { // data[b-1] = pivot
+			b--
+			dups++
+		}
+		// m-lo = (hi-lo)/2 > 6
+		// b-lo > (hi-lo)*3/4-1 > 8
+		// ==> m < b ==> data[m] <= pivot
+		if data[m].literal > data[pivot].literal { // data[m] = pivot
+			data[m], data[b-1] = data[b-1], data[m]
+			b--
+			dups++
+		}
+		// if at least 2 points are equal to pivot, assume skewed distribution
+		protect = dups > 1
+	}
+	if protect {
+		// Protect against a lot of duplicates
+		// Add invariant:
+		//	data[a <= i < b] unexamined
+		//	data[b <= i < c] = pivot
+		for {
+			for ; a < b && data[b-1].literal > data[pivot].literal; b-- { // data[b] == pivot
+			}
+			for ; a < b && data[a].literal < data[pivot].literal; a++ { // data[a] < pivot
+			}
+			if a >= b {
+				break
+			}
+			// data[a] == pivot; data[b-1] < pivot
+			data[a], data[b-1] = data[b-1], data[a]
+			a++
+			b--
+		}
+	}
+	// Swap pivot into middle
+	data[pivot], data[b-1] = data[b-1], data[pivot]
+	return b - 1, c
+}
+
+// Insertion sort
+func insertionSort(data []literalNode, a, b int) {
+	for i := a + 1; i < b; i++ {
+		for j := i; j > a && data[j].literal < data[j-1].literal; j-- {
+			data[j], data[j-1] = data[j-1], data[j]
+		}
+	}
+}
+
+// maxDepth returns a threshold at which quicksort should switch
+// to heapsort. It returns 2*ceil(lg(n+1)).
+func maxDepth(n int) int {
+	var depth int
+	for i := n; i > 0; i >>= 1 {
+		depth++
+	}
+	return depth * 2
+}
+
+// medianOfThree moves the median of the three values data[m0], data[m1], data[m2] into data[m1].
+func medianOfThree(data []literalNode, m1, m0, m2 int) {
+	// sort 3 elements
+	if data[m1].literal < data[m0].literal {
+		data[m1], data[m0] = data[m0], data[m1]
+	}
+	// data[m0] <= data[m1]
+	if data[m2].literal < data[m1].literal {
+		data[m2], data[m1] = data[m1], data[m2]
+		// data[m0] <= data[m2] && data[m1] < data[m2]
+		if data[m1].literal < data[m0].literal {
+			data[m1], data[m0] = data[m0], data[m1]
+		}
+	}
+	// now data[m0] <= data[m1] <= data[m2]
+}
diff --git a/vendor/github.com/klauspost/compress/flate/inflate.go b/vendor/github.com/klauspost/compress/flate/inflate.go
new file mode 100644
index 0000000..0d7b437
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/inflate.go
@@ -0,0 +1,865 @@
+// Copyright 2009 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package flate implements the DEFLATE compressed data format, described in
+// RFC 1951.  The gzip and zlib packages implement access to DEFLATE-based file
+// formats.
+package flate
+
+import (
+	"bufio"
+	"compress/flate"
+	"fmt"
+	"io"
+	"math/bits"
+	"sync"
+)
+
+const (
+	maxCodeLen     = 16 // max length of Huffman code
+	maxCodeLenMask = 15 // mask for max length of Huffman code
+	// The next three numbers come from the RFC section 3.2.7, with the
+	// additional proviso in section 3.2.5 which implies that distance codes
+	// 30 and 31 should never occur in compressed data.
+	maxNumLit  = 286
+	maxNumDist = 30
+	numCodes   = 19 // number of codes in Huffman meta-code
+
+	debugDecode = false
+)
+
+// Value of length - 3 and extra bits.
+type lengthExtra struct {
+	length, extra uint8
+}
+
+var decCodeToLen = [32]lengthExtra{{length: 0x0, extra: 0x0}, {length: 0x1, extra: 0x0}, {length: 0x2, extra: 0x0}, {length: 0x3, extra: 0x0}, {length: 0x4, extra: 0x0}, {length: 0x5, extra: 0x0}, {length: 0x6, extra: 0x0}, {length: 0x7, extra: 0x0}, {length: 0x8, extra: 0x1}, {length: 0xa, extra: 0x1}, {length: 0xc, extra: 0x1}, {length: 0xe, extra: 0x1}, {length: 0x10, extra: 0x2}, {length: 0x14, extra: 0x2}, {length: 0x18, extra: 0x2}, {length: 0x1c, extra: 0x2}, {length: 0x20, extra: 0x3}, {length: 0x28, extra: 0x3}, {length: 0x30, extra: 0x3}, {length: 0x38, extra: 0x3}, {length: 0x40, extra: 0x4}, {length: 0x50, extra: 0x4}, {length: 0x60, extra: 0x4}, {length: 0x70, extra: 0x4}, {length: 0x80, extra: 0x5}, {length: 0xa0, extra: 0x5}, {length: 0xc0, extra: 0x5}, {length: 0xe0, extra: 0x5}, {length: 0xff, extra: 0x0}, {length: 0x0, extra: 0x0}, {length: 0x0, extra: 0x0}, {length: 0x0, extra: 0x0}}
+
+var bitMask32 = [32]uint32{
+	0, 1, 3, 7, 0xF, 0x1F, 0x3F, 0x7F, 0xFF,
+	0x1FF, 0x3FF, 0x7FF, 0xFFF, 0x1FFF, 0x3FFF, 0x7FFF, 0xFFFF,
+	0x1ffff, 0x3ffff, 0x7FFFF, 0xfFFFF, 0x1fFFFF, 0x3fFFFF, 0x7fFFFF, 0xffFFFF,
+	0x1ffFFFF, 0x3ffFFFF, 0x7ffFFFF, 0xfffFFFF, 0x1fffFFFF, 0x3fffFFFF, 0x7fffFFFF,
+} // up to 32 bits
+
+// Initialize the fixedHuffmanDecoder only once upon first use.
+var fixedOnce sync.Once
+var fixedHuffmanDecoder huffmanDecoder
+
+// A CorruptInputError reports the presence of corrupt input at a given offset.
+type CorruptInputError = flate.CorruptInputError
+
+// An InternalError reports an error in the flate code itself.
+type InternalError string
+
+func (e InternalError) Error() string { return "flate: internal error: " + string(e) }
+
+// A ReadError reports an error encountered while reading input.
+//
+// Deprecated: No longer returned.
+type ReadError = flate.ReadError
+
+// A WriteError reports an error encountered while writing output.
+//
+// Deprecated: No longer returned.
+type WriteError = flate.WriteError
+
+// Resetter resets a ReadCloser returned by NewReader or NewReaderDict to
+// to switch to a new underlying Reader. This permits reusing a ReadCloser
+// instead of allocating a new one.
+type Resetter interface {
+	// Reset discards any buffered data and resets the Resetter as if it was
+	// newly initialized with the given reader.
+	Reset(r io.Reader, dict []byte) error
+}
+
+// The data structure for decoding Huffman tables is based on that of
+// zlib. There is a lookup table of a fixed bit width (huffmanChunkBits),
+// For codes smaller than the table width, there are multiple entries
+// (each combination of trailing bits has the same value). For codes
+// larger than the table width, the table contains a link to an overflow
+// table. The width of each entry in the link table is the maximum code
+// size minus the chunk width.
+//
+// Note that you can do a lookup in the table even without all bits
+// filled. Since the extra bits are zero, and the DEFLATE Huffman codes
+// have the property that shorter codes come before longer ones, the
+// bit length estimate in the result is a lower bound on the actual
+// number of bits.
+//
+// See the following:
+//	http://www.gzip.org/algorithm.txt
+
+// chunk & 15 is number of bits
+// chunk >> 4 is value, including table link
+
+const (
+	huffmanChunkBits  = 9
+	huffmanNumChunks  = 1 << huffmanChunkBits
+	huffmanCountMask  = 15
+	huffmanValueShift = 4
+)
+
+type huffmanDecoder struct {
+	maxRead  int                       // the maximum number of bits we can read and not overread
+	chunks   *[huffmanNumChunks]uint16 // chunks as described above
+	links    [][]uint16                // overflow links
+	linkMask uint32                    // mask the width of the link table
+}
+
+// Initialize Huffman decoding tables from array of code lengths.
+// Following this function, h is guaranteed to be initialized into a complete
+// tree (i.e., neither over-subscribed nor under-subscribed). The exception is a
+// degenerate case where the tree has only a single symbol with length 1. Empty
+// trees are permitted.
+func (h *huffmanDecoder) init(lengths []int) bool {
+	// Sanity enables additional runtime tests during Huffman
+	// table construction. It's intended to be used during
+	// development to supplement the currently ad-hoc unit tests.
+	const sanity = false
+
+	if h.chunks == nil {
+		h.chunks = new([huffmanNumChunks]uint16)
+	}
+
+	if h.maxRead != 0 {
+		*h = huffmanDecoder{chunks: h.chunks, links: h.links}
+	}
+
+	// Count number of codes of each length,
+	// compute maxRead and max length.
+	var count [maxCodeLen]int
+	var min, max int
+	for _, n := range lengths {
+		if n == 0 {
+			continue
+		}
+		if min == 0 || n < min {
+			min = n
+		}
+		if n > max {
+			max = n
+		}
+		count[n&maxCodeLenMask]++
+	}
+
+	// Empty tree. The decompressor.huffSym function will fail later if the tree
+	// is used. Technically, an empty tree is only valid for the HDIST tree and
+	// not the HCLEN and HLIT tree. However, a stream with an empty HCLEN tree
+	// is guaranteed to fail since it will attempt to use the tree to decode the
+	// codes for the HLIT and HDIST trees. Similarly, an empty HLIT tree is
+	// guaranteed to fail later since the compressed data section must be
+	// composed of at least one symbol (the end-of-block marker).
+	if max == 0 {
+		return true
+	}
+
+	code := 0
+	var nextcode [maxCodeLen]int
+	for i := min; i <= max; i++ {
+		code <<= 1
+		nextcode[i&maxCodeLenMask] = code
+		code += count[i&maxCodeLenMask]
+	}
+
+	// Check that the coding is complete (i.e., that we've
+	// assigned all 2-to-the-max possible bit sequences).
+	// Exception: To be compatible with zlib, we also need to
+	// accept degenerate single-code codings. See also
+	// TestDegenerateHuffmanCoding.
+	if code != 1<<uint(max) && !(code == 1 && max == 1) {
+		if debugDecode {
+			fmt.Println("coding failed, code, max:", code, max, code == 1<<uint(max), code == 1 && max == 1, "(one should be true)")
+		}
+		return false
+	}
+
+	h.maxRead = min
+
+	chunks := h.chunks[:]
+	for i := range chunks {
+		chunks[i] = 0
+	}
+
+	if max > huffmanChunkBits {
+		numLinks := 1 << (uint(max) - huffmanChunkBits)
+		h.linkMask = uint32(numLinks - 1)
+
+		// create link tables
+		link := nextcode[huffmanChunkBits+1] >> 1
+		if cap(h.links) < huffmanNumChunks-link {
+			h.links = make([][]uint16, huffmanNumChunks-link)
+		} else {
+			h.links = h.links[:huffmanNumChunks-link]
+		}
+		for j := uint(link); j < huffmanNumChunks; j++ {
+			reverse := int(bits.Reverse16(uint16(j)))
+			reverse >>= uint(16 - huffmanChunkBits)
+			off := j - uint(link)
+			if sanity && h.chunks[reverse] != 0 {
+				panic("impossible: overwriting existing chunk")
+			}
+			h.chunks[reverse] = uint16(off<<huffmanValueShift | (huffmanChunkBits + 1))
+			if cap(h.links[off]) < numLinks {
+				h.links[off] = make([]uint16, numLinks)
+			} else {
+				h.links[off] = h.links[off][:numLinks]
+			}
+		}
+	} else {
+		h.links = h.links[:0]
+	}
+
+	for i, n := range lengths {
+		if n == 0 {
+			continue
+		}
+		code := nextcode[n]
+		nextcode[n]++
+		chunk := uint16(i<<huffmanValueShift | n)
+		reverse := int(bits.Reverse16(uint16(code)))
+		reverse >>= uint(16 - n)
+		if n <= huffmanChunkBits {
+			for off := reverse; off < len(h.chunks); off += 1 << uint(n) {
+				// We should never need to overwrite
+				// an existing chunk. Also, 0 is
+				// never a valid chunk, because the
+				// lower 4 "count" bits should be
+				// between 1 and 15.
+				if sanity && h.chunks[off] != 0 {
+					panic("impossible: overwriting existing chunk")
+				}
+				h.chunks[off] = chunk
+			}
+		} else {
+			j := reverse & (huffmanNumChunks - 1)
+			if sanity && h.chunks[j]&huffmanCountMask != huffmanChunkBits+1 {
+				// Longer codes should have been
+				// associated with a link table above.
+				panic("impossible: not an indirect chunk")
+			}
+			value := h.chunks[j] >> huffmanValueShift
+			linktab := h.links[value]
+			reverse >>= huffmanChunkBits
+			for off := reverse; off < len(linktab); off += 1 << uint(n-huffmanChunkBits) {
+				if sanity && linktab[off] != 0 {
+					panic("impossible: overwriting existing chunk")
+				}
+				linktab[off] = chunk
+			}
+		}
+	}
+
+	if sanity {
+		// Above we've sanity checked that we never overwrote
+		// an existing entry. Here we additionally check that
+		// we filled the tables completely.
+		for i, chunk := range h.chunks {
+			if chunk == 0 {
+				// As an exception, in the degenerate
+				// single-code case, we allow odd
+				// chunks to be missing.
+				if code == 1 && i%2 == 1 {
+					continue
+				}
+				panic("impossible: missing chunk")
+			}
+		}
+		for _, linktab := range h.links {
+			for _, chunk := range linktab {
+				if chunk == 0 {
+					panic("impossible: missing chunk")
+				}
+			}
+		}
+	}
+
+	return true
+}
+
+// Reader is the actual read interface needed by NewReader.
+// If the passed in io.Reader does not also have ReadByte,
+// the NewReader will introduce its own buffering.
+type Reader interface {
+	io.Reader
+	io.ByteReader
+}
+
+type step uint8
+
+const (
+	copyData step = iota + 1
+	nextBlock
+	huffmanBytesBuffer
+	huffmanBytesReader
+	huffmanBufioReader
+	huffmanStringsReader
+	huffmanGenericReader
+)
+
+// flushMode tells decompressor when to return data
+type flushMode uint8
+
+const (
+	syncFlush    flushMode = iota // return data after sync flush block
+	partialFlush                  // return data after each block
+)
+
+// Decompress state.
+type decompressor struct {
+	// Input source.
+	r       Reader
+	roffset int64
+
+	// Huffman decoders for literal/length, distance.
+	h1, h2 huffmanDecoder
+
+	// Length arrays used to define Huffman codes.
+	bits     *[maxNumLit + maxNumDist]int
+	codebits *[numCodes]int
+
+	// Output history, buffer.
+	dict dictDecoder
+
+	// Next step in the decompression,
+	// and decompression state.
+	step      step
+	stepState int
+	err       error
+	toRead    []byte
+	hl, hd    *huffmanDecoder
+	copyLen   int
+	copyDist  int
+
+	// Temporary buffer (avoids repeated allocation).
+	buf [4]byte
+
+	// Input bits, in top of b.
+	b uint32
+
+	nb    uint
+	final bool
+
+	flushMode flushMode
+}
+
+func (f *decompressor) nextBlock() {
+	for f.nb < 1+2 {
+		if f.err = f.moreBits(); f.err != nil {
+			return
+		}
+	}
+	f.final = f.b&1 == 1
+	f.b >>= 1
+	typ := f.b & 3
+	f.b >>= 2
+	f.nb -= 1 + 2
+	switch typ {
+	case 0:
+		f.dataBlock()
+		if debugDecode {
+			fmt.Println("stored block")
+		}
+	case 1:
+		// compressed, fixed Huffman tables
+		f.hl = &fixedHuffmanDecoder
+		f.hd = nil
+		f.huffmanBlockDecoder()
+		if debugDecode {
+			fmt.Println("predefinied huffman block")
+		}
+	case 2:
+		// compressed, dynamic Huffman tables
+		if f.err = f.readHuffman(); f.err != nil {
+			break
+		}
+		f.hl = &f.h1
+		f.hd = &f.h2
+		f.huffmanBlockDecoder()
+		if debugDecode {
+			fmt.Println("dynamic huffman block")
+		}
+	default:
+		// 3 is reserved.
+		if debugDecode {
+			fmt.Println("reserved data block encountered")
+		}
+		f.err = CorruptInputError(f.roffset)
+	}
+}
+
+func (f *decompressor) Read(b []byte) (int, error) {
+	for {
+		if len(f.toRead) > 0 {
+			n := copy(b, f.toRead)
+			f.toRead = f.toRead[n:]
+			if len(f.toRead) == 0 {
+				return n, f.err
+			}
+			return n, nil
+		}
+		if f.err != nil {
+			return 0, f.err
+		}
+
+		f.doStep()
+
+		if f.err != nil && len(f.toRead) == 0 {
+			f.toRead = f.dict.readFlush() // Flush what's left in case of error
+		}
+	}
+}
+
+// WriteTo implements the io.WriteTo interface for io.Copy and friends.
+func (f *decompressor) WriteTo(w io.Writer) (int64, error) {
+	total := int64(0)
+	flushed := false
+	for {
+		if len(f.toRead) > 0 {
+			n, err := w.Write(f.toRead)
+			total += int64(n)
+			if err != nil {
+				f.err = err
+				return total, err
+			}
+			if n != len(f.toRead) {
+				return total, io.ErrShortWrite
+			}
+			f.toRead = f.toRead[:0]
+		}
+		if f.err != nil && flushed {
+			if f.err == io.EOF {
+				return total, nil
+			}
+			return total, f.err
+		}
+		if f.err == nil {
+			f.doStep()
+		}
+		if len(f.toRead) == 0 && f.err != nil && !flushed {
+			f.toRead = f.dict.readFlush() // Flush what's left in case of error
+			flushed = true
+		}
+	}
+}
+
+func (f *decompressor) Close() error {
+	if f.err == io.EOF {
+		return nil
+	}
+	return f.err
+}
+
+// RFC 1951 section 3.2.7.
+// Compression with dynamic Huffman codes
+
+var codeOrder = [...]int{16, 17, 18, 0, 8, 7, 9, 6, 10, 5, 11, 4, 12, 3, 13, 2, 14, 1, 15}
+
+func (f *decompressor) readHuffman() error {
+	// HLIT[5], HDIST[5], HCLEN[4].
+	for f.nb < 5+5+4 {
+		if err := f.moreBits(); err != nil {
+			return err
+		}
+	}
+	nlit := int(f.b&0x1F) + 257
+	if nlit > maxNumLit {
+		if debugDecode {
+			fmt.Println("nlit > maxNumLit", nlit)
+		}
+		return CorruptInputError(f.roffset)
+	}
+	f.b >>= 5
+	ndist := int(f.b&0x1F) + 1
+	if ndist > maxNumDist {
+		if debugDecode {
+			fmt.Println("ndist > maxNumDist", ndist)
+		}
+		return CorruptInputError(f.roffset)
+	}
+	f.b >>= 5
+	nclen := int(f.b&0xF) + 4
+	// numCodes is 19, so nclen is always valid.
+	f.b >>= 4
+	f.nb -= 5 + 5 + 4
+
+	// (HCLEN+4)*3 bits: code lengths in the magic codeOrder order.
+	for i := 0; i < nclen; i++ {
+		for f.nb < 3 {
+			if err := f.moreBits(); err != nil {
+				return err
+			}
+		}
+		f.codebits[codeOrder[i]] = int(f.b & 0x7)
+		f.b >>= 3
+		f.nb -= 3
+	}
+	for i := nclen; i < len(codeOrder); i++ {
+		f.codebits[codeOrder[i]] = 0
+	}
+	if !f.h1.init(f.codebits[0:]) {
+		if debugDecode {
+			fmt.Println("init codebits failed")
+		}
+		return CorruptInputError(f.roffset)
+	}
+
+	// HLIT + 257 code lengths, HDIST + 1 code lengths,
+	// using the code length Huffman code.
+	for i, n := 0, nlit+ndist; i < n; {
+		x, err := f.huffSym(&f.h1)
+		if err != nil {
+			return err
+		}
+		if x < 16 {
+			// Actual length.
+			f.bits[i] = x
+			i++
+			continue
+		}
+		// Repeat previous length or zero.
+		var rep int
+		var nb uint
+		var b int
+		switch x {
+		default:
+			return InternalError("unexpected length code")
+		case 16:
+			rep = 3
+			nb = 2
+			if i == 0 {
+				if debugDecode {
+					fmt.Println("i==0")
+				}
+				return CorruptInputError(f.roffset)
+			}
+			b = f.bits[i-1]
+		case 17:
+			rep = 3
+			nb = 3
+			b = 0
+		case 18:
+			rep = 11
+			nb = 7
+			b = 0
+		}
+		for f.nb < nb {
+			if err := f.moreBits(); err != nil {
+				if debugDecode {
+					fmt.Println("morebits:", err)
+				}
+				return err
+			}
+		}
+		rep += int(f.b & uint32(1<<(nb&regSizeMaskUint32)-1))
+		f.b >>= nb & regSizeMaskUint32
+		f.nb -= nb
+		if i+rep > n {
+			if debugDecode {
+				fmt.Println("i+rep > n", i, rep, n)
+			}
+			return CorruptInputError(f.roffset)
+		}
+		for j := 0; j < rep; j++ {
+			f.bits[i] = b
+			i++
+		}
+	}
+
+	if !f.h1.init(f.bits[0:nlit]) || !f.h2.init(f.bits[nlit:nlit+ndist]) {
+		if debugDecode {
+			fmt.Println("init2 failed")
+		}
+		return CorruptInputError(f.roffset)
+	}
+
+	// As an optimization, we can initialize the maxRead bits to read at a time
+	// for the HLIT tree to the length of the EOB marker since we know that
+	// every block must terminate with one. This preserves the property that
+	// we never read any extra bytes after the end of the DEFLATE stream.
+	if f.h1.maxRead < f.bits[endBlockMarker] {
+		f.h1.maxRead = f.bits[endBlockMarker]
+	}
+	if !f.final {
+		// If not the final block, the smallest block possible is
+		// a predefined table, BTYPE=01, with a single EOB marker.
+		// This will take up 3 + 7 bits.
+		f.h1.maxRead += 10
+	}
+
+	return nil
+}
+
+// Copy a single uncompressed data block from input to output.
+func (f *decompressor) dataBlock() {
+	// Uncompressed.
+	// Discard current half-byte.
+	left := (f.nb) & 7
+	f.nb -= left
+	f.b >>= left
+
+	offBytes := f.nb >> 3
+	// Unfilled values will be overwritten.
+	f.buf[0] = uint8(f.b)
+	f.buf[1] = uint8(f.b >> 8)
+	f.buf[2] = uint8(f.b >> 16)
+	f.buf[3] = uint8(f.b >> 24)
+
+	f.roffset += int64(offBytes)
+	f.nb, f.b = 0, 0
+
+	// Length then ones-complement of length.
+	nr, err := io.ReadFull(f.r, f.buf[offBytes:4])
+	f.roffset += int64(nr)
+	if err != nil {
+		f.err = noEOF(err)
+		return
+	}
+	n := uint16(f.buf[0]) | uint16(f.buf[1])<<8
+	nn := uint16(f.buf[2]) | uint16(f.buf[3])<<8
+	if nn != ^n {
+		if debugDecode {
+			ncomp := ^n
+			fmt.Println("uint16(nn) != uint16(^n)", nn, ncomp)
+		}
+		f.err = CorruptInputError(f.roffset)
+		return
+	}
+
+	if n == 0 {
+		if f.flushMode == syncFlush {
+			f.toRead = f.dict.readFlush()
+		}
+
+		f.finishBlock()
+		return
+	}
+
+	f.copyLen = int(n)
+	f.copyData()
+}
+
+// copyData copies f.copyLen bytes from the underlying reader into f.hist.
+// It pauses for reads when f.hist is full.
+func (f *decompressor) copyData() {
+	buf := f.dict.writeSlice()
+	if len(buf) > f.copyLen {
+		buf = buf[:f.copyLen]
+	}
+
+	cnt, err := io.ReadFull(f.r, buf)
+	f.roffset += int64(cnt)
+	f.copyLen -= cnt
+	f.dict.writeMark(cnt)
+	if err != nil {
+		f.err = noEOF(err)
+		return
+	}
+
+	if f.dict.availWrite() == 0 || f.copyLen > 0 {
+		f.toRead = f.dict.readFlush()
+		f.step = copyData
+		return
+	}
+	f.finishBlock()
+}
+
+func (f *decompressor) finishBlock() {
+	if f.final {
+		if f.dict.availRead() > 0 {
+			f.toRead = f.dict.readFlush()
+		}
+
+		f.err = io.EOF
+	} else if f.flushMode == partialFlush && f.dict.availRead() > 0 {
+		f.toRead = f.dict.readFlush()
+	}
+
+	f.step = nextBlock
+}
+
+func (f *decompressor) doStep() {
+	switch f.step {
+	case copyData:
+		f.copyData()
+	case nextBlock:
+		f.nextBlock()
+	case huffmanBytesBuffer:
+		f.huffmanBytesBuffer()
+	case huffmanBytesReader:
+		f.huffmanBytesReader()
+	case huffmanBufioReader:
+		f.huffmanBufioReader()
+	case huffmanStringsReader:
+		f.huffmanStringsReader()
+	case huffmanGenericReader:
+		f.huffmanGenericReader()
+	default:
+		panic("BUG: unexpected step state")
+	}
+}
+
+// noEOF returns err, unless err == io.EOF, in which case it returns io.ErrUnexpectedEOF.
+func noEOF(e error) error {
+	if e == io.EOF {
+		return io.ErrUnexpectedEOF
+	}
+	return e
+}
+
+func (f *decompressor) moreBits() error {
+	c, err := f.r.ReadByte()
+	if err != nil {
+		return noEOF(err)
+	}
+	f.roffset++
+	f.b |= uint32(c) << (f.nb & regSizeMaskUint32)
+	f.nb += 8
+	return nil
+}
+
+// Read the next Huffman-encoded symbol from f according to h.
+func (f *decompressor) huffSym(h *huffmanDecoder) (int, error) {
+	// Since a huffmanDecoder can be empty or be composed of a degenerate tree
+	// with single element, huffSym must error on these two edge cases. In both
+	// cases, the chunks slice will be 0 for the invalid sequence, leading it
+	// satisfy the n == 0 check below.
+	n := uint(h.maxRead)
+	// Optimization. Compiler isn't smart enough to keep f.b,f.nb in registers,
+	// but is smart enough to keep local variables in registers, so use nb and b,
+	// inline call to moreBits and reassign b,nb back to f on return.
+	nb, b := f.nb, f.b
+	for {
+		for nb < n {
+			c, err := f.r.ReadByte()
+			if err != nil {
+				f.b = b
+				f.nb = nb
+				return 0, noEOF(err)
+			}
+			f.roffset++
+			b |= uint32(c) << (nb & regSizeMaskUint32)
+			nb += 8
+		}
+		chunk := h.chunks[b&(huffmanNumChunks-1)]
+		n = uint(chunk & huffmanCountMask)
+		if n > huffmanChunkBits {
+			chunk = h.links[chunk>>huffmanValueShift][(b>>huffmanChunkBits)&h.linkMask]
+			n = uint(chunk & huffmanCountMask)
+		}
+		if n <= nb {
+			if n == 0 {
+				f.b = b
+				f.nb = nb
+				if debugDecode {
+					fmt.Println("huffsym: n==0")
+				}
+				f.err = CorruptInputError(f.roffset)
+				return 0, f.err
+			}
+			f.b = b >> (n & regSizeMaskUint32)
+			f.nb = nb - n
+			return int(chunk >> huffmanValueShift), nil
+		}
+	}
+}
+
+func makeReader(r io.Reader) Reader {
+	if rr, ok := r.(Reader); ok {
+		return rr
+	}
+	return bufio.NewReader(r)
+}
+
+func fixedHuffmanDecoderInit() {
+	fixedOnce.Do(func() {
+		// These come from the RFC section 3.2.6.
+		var bits [288]int
+		for i := 0; i < 144; i++ {
+			bits[i] = 8
+		}
+		for i := 144; i < 256; i++ {
+			bits[i] = 9
+		}
+		for i := 256; i < 280; i++ {
+			bits[i] = 7
+		}
+		for i := 280; i < 288; i++ {
+			bits[i] = 8
+		}
+		fixedHuffmanDecoder.init(bits[:])
+	})
+}
+
+func (f *decompressor) Reset(r io.Reader, dict []byte) error {
+	*f = decompressor{
+		r:        makeReader(r),
+		bits:     f.bits,
+		codebits: f.codebits,
+		h1:       f.h1,
+		h2:       f.h2,
+		dict:     f.dict,
+		step:     nextBlock,
+	}
+	f.dict.init(maxMatchOffset, dict)
+	return nil
+}
+
+type ReaderOpt func(*decompressor)
+
+// WithPartialBlock tells decompressor to return after each block,
+// so it can read data written with partial flush
+func WithPartialBlock() ReaderOpt {
+	return func(f *decompressor) {
+		f.flushMode = partialFlush
+	}
+}
+
+// WithDict initializes the reader with a preset dictionary
+func WithDict(dict []byte) ReaderOpt {
+	return func(f *decompressor) {
+		f.dict.init(maxMatchOffset, dict)
+	}
+}
+
+// NewReaderOpts returns new reader with provided options
+func NewReaderOpts(r io.Reader, opts ...ReaderOpt) io.ReadCloser {
+	fixedHuffmanDecoderInit()
+
+	var f decompressor
+	f.r = makeReader(r)
+	f.bits = new([maxNumLit + maxNumDist]int)
+	f.codebits = new([numCodes]int)
+	f.step = nextBlock
+	f.dict.init(maxMatchOffset, nil)
+
+	for _, opt := range opts {
+		opt(&f)
+	}
+
+	return &f
+}
+
+// NewReader returns a new ReadCloser that can be used
+// to read the uncompressed version of r.
+// If r does not also implement io.ByteReader,
+// the decompressor may read more data than necessary from r.
+// It is the caller's responsibility to call Close on the ReadCloser
+// when finished reading.
+//
+// The ReadCloser returned by NewReader also implements Resetter.
+func NewReader(r io.Reader) io.ReadCloser {
+	return NewReaderOpts(r)
+}
+
+// NewReaderDict is like NewReader but initializes the reader
+// with a preset dictionary. The returned Reader behaves as if
+// the uncompressed data stream started with the given dictionary,
+// which has already been read. NewReaderDict is typically used
+// to read data compressed by NewWriterDict.
+//
+// The ReadCloser returned by NewReader also implements Resetter.
+func NewReaderDict(r io.Reader, dict []byte) io.ReadCloser {
+	return NewReaderOpts(r, WithDict(dict))
+}
diff --git a/vendor/github.com/klauspost/compress/flate/inflate_gen.go b/vendor/github.com/klauspost/compress/flate/inflate_gen.go
new file mode 100644
index 0000000..2b2f993
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/inflate_gen.go
@@ -0,0 +1,1283 @@
+// Code generated by go generate gen_inflate.go. DO NOT EDIT.
+
+package flate
+
+import (
+	"bufio"
+	"bytes"
+	"fmt"
+	"math/bits"
+	"strings"
+)
+
+// Decode a single Huffman block from f.
+// hl and hd are the Huffman states for the lit/length values
+// and the distance values, respectively. If hd == nil, using the
+// fixed distance encoding associated with fixed Huffman blocks.
+func (f *decompressor) huffmanBytesBuffer() {
+	const (
+		stateInit = iota // Zero value must be stateInit
+		stateDict
+	)
+	fr := f.r.(*bytes.Buffer)
+
+	// Optimization. Compiler isn't smart enough to keep f.b,f.nb in registers,
+	// but is smart enough to keep local variables in registers, so use nb and b,
+	// inline call to moreBits and reassign b,nb back to f on return.
+	fnb, fb, dict := f.nb, f.b, &f.dict
+
+	switch f.stepState {
+	case stateInit:
+		goto readLiteral
+	case stateDict:
+		goto copyHistory
+	}
+
+readLiteral:
+	// Read literal and/or (length, distance) according to RFC section 3.2.3.
+	{
+		var v int
+		{
+			// Inlined v, err := f.huffSym(f.hl)
+			// Since a huffmanDecoder can be empty or be composed of a degenerate tree
+			// with single element, huffSym must error on these two edge cases. In both
+			// cases, the chunks slice will be 0 for the invalid sequence, leading it
+			// satisfy the n == 0 check below.
+			n := uint(f.hl.maxRead)
+			for {
+				for fnb < n {
+					c, err := fr.ReadByte()
+					if err != nil {
+						f.b, f.nb = fb, fnb
+						f.err = noEOF(err)
+						return
+					}
+					f.roffset++
+					fb |= uint32(c) << (fnb & regSizeMaskUint32)
+					fnb += 8
+				}
+				chunk := f.hl.chunks[fb&(huffmanNumChunks-1)]
+				n = uint(chunk & huffmanCountMask)
+				if n > huffmanChunkBits {
+					chunk = f.hl.links[chunk>>huffmanValueShift][(fb>>huffmanChunkBits)&f.hl.linkMask]
+					n = uint(chunk & huffmanCountMask)
+				}
+				if n <= fnb {
+					if n == 0 {
+						f.b, f.nb = fb, fnb
+						if debugDecode {
+							fmt.Println("huffsym: n==0")
+						}
+						f.err = CorruptInputError(f.roffset)
+						return
+					}
+					fb = fb >> (n & regSizeMaskUint32)
+					fnb = fnb - n
+					v = int(chunk >> huffmanValueShift)
+					break
+				}
+			}
+		}
+
+		var length int
+		switch {
+		case v < 256:
+			dict.writeByte(byte(v))
+			if dict.availWrite() == 0 {
+				f.toRead = dict.readFlush()
+				f.step = huffmanBytesBuffer
+				f.stepState = stateInit
+				f.b, f.nb = fb, fnb
+				return
+			}
+			goto readLiteral
+		case v == 256:
+			f.b, f.nb = fb, fnb
+			f.finishBlock()
+			return
+		// otherwise, reference to older data
+		case v < 265:
+			length = v - (257 - 3)
+		case v < maxNumLit:
+			val := decCodeToLen[(v - 257)]
+			length = int(val.length) + 3
+			n := uint(val.extra)
+			for fnb < n {
+				c, err := fr.ReadByte()
+				if err != nil {
+					f.b, f.nb = fb, fnb
+					if debugDecode {
+						fmt.Println("morebits n>0:", err)
+					}
+					f.err = err
+					return
+				}
+				f.roffset++
+				fb |= uint32(c) << (fnb & regSizeMaskUint32)
+				fnb += 8
+			}
+			length += int(fb & bitMask32[n])
+			fb >>= n & regSizeMaskUint32
+			fnb -= n
+		default:
+			if debugDecode {
+				fmt.Println(v, ">= maxNumLit")
+			}
+			f.err = CorruptInputError(f.roffset)
+			f.b, f.nb = fb, fnb
+			return
+		}
+
+		var dist uint32
+		if f.hd == nil {
+			for fnb < 5 {
+				c, err := fr.ReadByte()
+				if err != nil {
+					f.b, f.nb = fb, fnb
+					if debugDecode {
+						fmt.Println("morebits f.nb<5:", err)
+					}
+					f.err = err
+					return
+				}
+				f.roffset++
+				fb |= uint32(c) << (fnb & regSizeMaskUint32)
+				fnb += 8
+			}
+			dist = uint32(bits.Reverse8(uint8(fb & 0x1F << 3)))
+			fb >>= 5
+			fnb -= 5
+		} else {
+			// Since a huffmanDecoder can be empty or be composed of a degenerate tree
+			// with single element, huffSym must error on these two edge cases. In both
+			// cases, the chunks slice will be 0 for the invalid sequence, leading it
+			// satisfy the n == 0 check below.
+			n := uint(f.hd.maxRead)
+			// Optimization. Compiler isn't smart enough to keep f.b,f.nb in registers,
+			// but is smart enough to keep local variables in registers, so use nb and b,
+			// inline call to moreBits and reassign b,nb back to f on return.
+			for {
+				for fnb < n {
+					c, err := fr.ReadByte()
+					if err != nil {
+						f.b, f.nb = fb, fnb
+						f.err = noEOF(err)
+						return
+					}
+					f.roffset++
+					fb |= uint32(c) << (fnb & regSizeMaskUint32)
+					fnb += 8
+				}
+				chunk := f.hd.chunks[fb&(huffmanNumChunks-1)]
+				n = uint(chunk & huffmanCountMask)
+				if n > huffmanChunkBits {
+					chunk = f.hd.links[chunk>>huffmanValueShift][(fb>>huffmanChunkBits)&f.hd.linkMask]
+					n = uint(chunk & huffmanCountMask)
+				}
+				if n <= fnb {
+					if n == 0 {
+						f.b, f.nb = fb, fnb
+						if debugDecode {
+							fmt.Println("huffsym: n==0")
+						}
+						f.err = CorruptInputError(f.roffset)
+						return
+					}
+					fb = fb >> (n & regSizeMaskUint32)
+					fnb = fnb - n
+					dist = uint32(chunk >> huffmanValueShift)
+					break
+				}
+			}
+		}
+
+		switch {
+		case dist < 4:
+			dist++
+		case dist < maxNumDist:
+			nb := uint(dist-2) >> 1
+			// have 1 bit in bottom of dist, need nb more.
+			extra := (dist & 1) << (nb & regSizeMaskUint32)
+			for fnb < nb {
+				c, err := fr.ReadByte()
+				if err != nil {
+					f.b, f.nb = fb, fnb
+					if debugDecode {
+						fmt.Println("morebits f.nb<nb:", err)
+					}
+					f.err = err
+					return
+				}
+				f.roffset++
+				fb |= uint32(c) << (fnb & regSizeMaskUint32)
+				fnb += 8
+			}
+			extra |= fb & bitMask32[nb]
+			fb >>= nb & regSizeMaskUint32
+			fnb -= nb
+			dist = 1<<((nb+1)&regSizeMaskUint32) + 1 + extra
+			// slower: dist = bitMask32[nb+1] + 2 + extra
+		default:
+			f.b, f.nb = fb, fnb
+			if debugDecode {
+				fmt.Println("dist too big:", dist, maxNumDist)
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		// No check on length; encoding can be prescient.
+		if dist > uint32(dict.histSize()) {
+			f.b, f.nb = fb, fnb
+			if debugDecode {
+				fmt.Println("dist > dict.histSize():", dist, dict.histSize())
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		f.copyLen, f.copyDist = length, int(dist)
+		goto copyHistory
+	}
+
+copyHistory:
+	// Perform a backwards copy according to RFC section 3.2.3.
+	{
+		cnt := dict.tryWriteCopy(f.copyDist, f.copyLen)
+		if cnt == 0 {
+			cnt = dict.writeCopy(f.copyDist, f.copyLen)
+		}
+		f.copyLen -= cnt
+
+		if dict.availWrite() == 0 || f.copyLen > 0 {
+			f.toRead = dict.readFlush()
+			f.step = huffmanBytesBuffer // We need to continue this work
+			f.stepState = stateDict
+			f.b, f.nb = fb, fnb
+			return
+		}
+		goto readLiteral
+	}
+	// Not reached
+}
+
+// Decode a single Huffman block from f.
+// hl and hd are the Huffman states for the lit/length values
+// and the distance values, respectively. If hd == nil, using the
+// fixed distance encoding associated with fixed Huffman blocks.
+func (f *decompressor) huffmanBytesReader() {
+	const (
+		stateInit = iota // Zero value must be stateInit
+		stateDict
+	)
+	fr := f.r.(*bytes.Reader)
+
+	// Optimization. Compiler isn't smart enough to keep f.b,f.nb in registers,
+	// but is smart enough to keep local variables in registers, so use nb and b,
+	// inline call to moreBits and reassign b,nb back to f on return.
+	fnb, fb, dict := f.nb, f.b, &f.dict
+
+	switch f.stepState {
+	case stateInit:
+		goto readLiteral
+	case stateDict:
+		goto copyHistory
+	}
+
+readLiteral:
+	// Read literal and/or (length, distance) according to RFC section 3.2.3.
+	{
+		var v int
+		{
+			// Inlined v, err := f.huffSym(f.hl)
+			// Since a huffmanDecoder can be empty or be composed of a degenerate tree
+			// with single element, huffSym must error on these two edge cases. In both
+			// cases, the chunks slice will be 0 for the invalid sequence, leading it
+			// satisfy the n == 0 check below.
+			n := uint(f.hl.maxRead)
+			for {
+				for fnb < n {
+					c, err := fr.ReadByte()
+					if err != nil {
+						f.b, f.nb = fb, fnb
+						f.err = noEOF(err)
+						return
+					}
+					f.roffset++
+					fb |= uint32(c) << (fnb & regSizeMaskUint32)
+					fnb += 8
+				}
+				chunk := f.hl.chunks[fb&(huffmanNumChunks-1)]
+				n = uint(chunk & huffmanCountMask)
+				if n > huffmanChunkBits {
+					chunk = f.hl.links[chunk>>huffmanValueShift][(fb>>huffmanChunkBits)&f.hl.linkMask]
+					n = uint(chunk & huffmanCountMask)
+				}
+				if n <= fnb {
+					if n == 0 {
+						f.b, f.nb = fb, fnb
+						if debugDecode {
+							fmt.Println("huffsym: n==0")
+						}
+						f.err = CorruptInputError(f.roffset)
+						return
+					}
+					fb = fb >> (n & regSizeMaskUint32)
+					fnb = fnb - n
+					v = int(chunk >> huffmanValueShift)
+					break
+				}
+			}
+		}
+
+		var length int
+		switch {
+		case v < 256:
+			dict.writeByte(byte(v))
+			if dict.availWrite() == 0 {
+				f.toRead = dict.readFlush()
+				f.step = huffmanBytesReader
+				f.stepState = stateInit
+				f.b, f.nb = fb, fnb
+				return
+			}
+			goto readLiteral
+		case v == 256:
+			f.b, f.nb = fb, fnb
+			f.finishBlock()
+			return
+		// otherwise, reference to older data
+		case v < 265:
+			length = v - (257 - 3)
+		case v < maxNumLit:
+			val := decCodeToLen[(v - 257)]
+			length = int(val.length) + 3
+			n := uint(val.extra)
+			for fnb < n {
+				c, err := fr.ReadByte()
+				if err != nil {
+					f.b, f.nb = fb, fnb
+					if debugDecode {
+						fmt.Println("morebits n>0:", err)
+					}
+					f.err = err
+					return
+				}
+				f.roffset++
+				fb |= uint32(c) << (fnb & regSizeMaskUint32)
+				fnb += 8
+			}
+			length += int(fb & bitMask32[n])
+			fb >>= n & regSizeMaskUint32
+			fnb -= n
+		default:
+			if debugDecode {
+				fmt.Println(v, ">= maxNumLit")
+			}
+			f.err = CorruptInputError(f.roffset)
+			f.b, f.nb = fb, fnb
+			return
+		}
+
+		var dist uint32
+		if f.hd == nil {
+			for fnb < 5 {
+				c, err := fr.ReadByte()
+				if err != nil {
+					f.b, f.nb = fb, fnb
+					if debugDecode {
+						fmt.Println("morebits f.nb<5:", err)
+					}
+					f.err = err
+					return
+				}
+				f.roffset++
+				fb |= uint32(c) << (fnb & regSizeMaskUint32)
+				fnb += 8
+			}
+			dist = uint32(bits.Reverse8(uint8(fb & 0x1F << 3)))
+			fb >>= 5
+			fnb -= 5
+		} else {
+			// Since a huffmanDecoder can be empty or be composed of a degenerate tree
+			// with single element, huffSym must error on these two edge cases. In both
+			// cases, the chunks slice will be 0 for the invalid sequence, leading it
+			// satisfy the n == 0 check below.
+			n := uint(f.hd.maxRead)
+			// Optimization. Compiler isn't smart enough to keep f.b,f.nb in registers,
+			// but is smart enough to keep local variables in registers, so use nb and b,
+			// inline call to moreBits and reassign b,nb back to f on return.
+			for {
+				for fnb < n {
+					c, err := fr.ReadByte()
+					if err != nil {
+						f.b, f.nb = fb, fnb
+						f.err = noEOF(err)
+						return
+					}
+					f.roffset++
+					fb |= uint32(c) << (fnb & regSizeMaskUint32)
+					fnb += 8
+				}
+				chunk := f.hd.chunks[fb&(huffmanNumChunks-1)]
+				n = uint(chunk & huffmanCountMask)
+				if n > huffmanChunkBits {
+					chunk = f.hd.links[chunk>>huffmanValueShift][(fb>>huffmanChunkBits)&f.hd.linkMask]
+					n = uint(chunk & huffmanCountMask)
+				}
+				if n <= fnb {
+					if n == 0 {
+						f.b, f.nb = fb, fnb
+						if debugDecode {
+							fmt.Println("huffsym: n==0")
+						}
+						f.err = CorruptInputError(f.roffset)
+						return
+					}
+					fb = fb >> (n & regSizeMaskUint32)
+					fnb = fnb - n
+					dist = uint32(chunk >> huffmanValueShift)
+					break
+				}
+			}
+		}
+
+		switch {
+		case dist < 4:
+			dist++
+		case dist < maxNumDist:
+			nb := uint(dist-2) >> 1
+			// have 1 bit in bottom of dist, need nb more.
+			extra := (dist & 1) << (nb & regSizeMaskUint32)
+			for fnb < nb {
+				c, err := fr.ReadByte()
+				if err != nil {
+					f.b, f.nb = fb, fnb
+					if debugDecode {
+						fmt.Println("morebits f.nb<nb:", err)
+					}
+					f.err = err
+					return
+				}
+				f.roffset++
+				fb |= uint32(c) << (fnb & regSizeMaskUint32)
+				fnb += 8
+			}
+			extra |= fb & bitMask32[nb]
+			fb >>= nb & regSizeMaskUint32
+			fnb -= nb
+			dist = 1<<((nb+1)&regSizeMaskUint32) + 1 + extra
+			// slower: dist = bitMask32[nb+1] + 2 + extra
+		default:
+			f.b, f.nb = fb, fnb
+			if debugDecode {
+				fmt.Println("dist too big:", dist, maxNumDist)
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		// No check on length; encoding can be prescient.
+		if dist > uint32(dict.histSize()) {
+			f.b, f.nb = fb, fnb
+			if debugDecode {
+				fmt.Println("dist > dict.histSize():", dist, dict.histSize())
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		f.copyLen, f.copyDist = length, int(dist)
+		goto copyHistory
+	}
+
+copyHistory:
+	// Perform a backwards copy according to RFC section 3.2.3.
+	{
+		cnt := dict.tryWriteCopy(f.copyDist, f.copyLen)
+		if cnt == 0 {
+			cnt = dict.writeCopy(f.copyDist, f.copyLen)
+		}
+		f.copyLen -= cnt
+
+		if dict.availWrite() == 0 || f.copyLen > 0 {
+			f.toRead = dict.readFlush()
+			f.step = huffmanBytesReader // We need to continue this work
+			f.stepState = stateDict
+			f.b, f.nb = fb, fnb
+			return
+		}
+		goto readLiteral
+	}
+	// Not reached
+}
+
+// Decode a single Huffman block from f.
+// hl and hd are the Huffman states for the lit/length values
+// and the distance values, respectively. If hd == nil, using the
+// fixed distance encoding associated with fixed Huffman blocks.
+func (f *decompressor) huffmanBufioReader() {
+	const (
+		stateInit = iota // Zero value must be stateInit
+		stateDict
+	)
+	fr := f.r.(*bufio.Reader)
+
+	// Optimization. Compiler isn't smart enough to keep f.b,f.nb in registers,
+	// but is smart enough to keep local variables in registers, so use nb and b,
+	// inline call to moreBits and reassign b,nb back to f on return.
+	fnb, fb, dict := f.nb, f.b, &f.dict
+
+	switch f.stepState {
+	case stateInit:
+		goto readLiteral
+	case stateDict:
+		goto copyHistory
+	}
+
+readLiteral:
+	// Read literal and/or (length, distance) according to RFC section 3.2.3.
+	{
+		var v int
+		{
+			// Inlined v, err := f.huffSym(f.hl)
+			// Since a huffmanDecoder can be empty or be composed of a degenerate tree
+			// with single element, huffSym must error on these two edge cases. In both
+			// cases, the chunks slice will be 0 for the invalid sequence, leading it
+			// satisfy the n == 0 check below.
+			n := uint(f.hl.maxRead)
+			for {
+				for fnb < n {
+					c, err := fr.ReadByte()
+					if err != nil {
+						f.b, f.nb = fb, fnb
+						f.err = noEOF(err)
+						return
+					}
+					f.roffset++
+					fb |= uint32(c) << (fnb & regSizeMaskUint32)
+					fnb += 8
+				}
+				chunk := f.hl.chunks[fb&(huffmanNumChunks-1)]
+				n = uint(chunk & huffmanCountMask)
+				if n > huffmanChunkBits {
+					chunk = f.hl.links[chunk>>huffmanValueShift][(fb>>huffmanChunkBits)&f.hl.linkMask]
+					n = uint(chunk & huffmanCountMask)
+				}
+				if n <= fnb {
+					if n == 0 {
+						f.b, f.nb = fb, fnb
+						if debugDecode {
+							fmt.Println("huffsym: n==0")
+						}
+						f.err = CorruptInputError(f.roffset)
+						return
+					}
+					fb = fb >> (n & regSizeMaskUint32)
+					fnb = fnb - n
+					v = int(chunk >> huffmanValueShift)
+					break
+				}
+			}
+		}
+
+		var length int
+		switch {
+		case v < 256:
+			dict.writeByte(byte(v))
+			if dict.availWrite() == 0 {
+				f.toRead = dict.readFlush()
+				f.step = huffmanBufioReader
+				f.stepState = stateInit
+				f.b, f.nb = fb, fnb
+				return
+			}
+			goto readLiteral
+		case v == 256:
+			f.b, f.nb = fb, fnb
+			f.finishBlock()
+			return
+		// otherwise, reference to older data
+		case v < 265:
+			length = v - (257 - 3)
+		case v < maxNumLit:
+			val := decCodeToLen[(v - 257)]
+			length = int(val.length) + 3
+			n := uint(val.extra)
+			for fnb < n {
+				c, err := fr.ReadByte()
+				if err != nil {
+					f.b, f.nb = fb, fnb
+					if debugDecode {
+						fmt.Println("morebits n>0:", err)
+					}
+					f.err = err
+					return
+				}
+				f.roffset++
+				fb |= uint32(c) << (fnb & regSizeMaskUint32)
+				fnb += 8
+			}
+			length += int(fb & bitMask32[n])
+			fb >>= n & regSizeMaskUint32
+			fnb -= n
+		default:
+			if debugDecode {
+				fmt.Println(v, ">= maxNumLit")
+			}
+			f.err = CorruptInputError(f.roffset)
+			f.b, f.nb = fb, fnb
+			return
+		}
+
+		var dist uint32
+		if f.hd == nil {
+			for fnb < 5 {
+				c, err := fr.ReadByte()
+				if err != nil {
+					f.b, f.nb = fb, fnb
+					if debugDecode {
+						fmt.Println("morebits f.nb<5:", err)
+					}
+					f.err = err
+					return
+				}
+				f.roffset++
+				fb |= uint32(c) << (fnb & regSizeMaskUint32)
+				fnb += 8
+			}
+			dist = uint32(bits.Reverse8(uint8(fb & 0x1F << 3)))
+			fb >>= 5
+			fnb -= 5
+		} else {
+			// Since a huffmanDecoder can be empty or be composed of a degenerate tree
+			// with single element, huffSym must error on these two edge cases. In both
+			// cases, the chunks slice will be 0 for the invalid sequence, leading it
+			// satisfy the n == 0 check below.
+			n := uint(f.hd.maxRead)
+			// Optimization. Compiler isn't smart enough to keep f.b,f.nb in registers,
+			// but is smart enough to keep local variables in registers, so use nb and b,
+			// inline call to moreBits and reassign b,nb back to f on return.
+			for {
+				for fnb < n {
+					c, err := fr.ReadByte()
+					if err != nil {
+						f.b, f.nb = fb, fnb
+						f.err = noEOF(err)
+						return
+					}
+					f.roffset++
+					fb |= uint32(c) << (fnb & regSizeMaskUint32)
+					fnb += 8
+				}
+				chunk := f.hd.chunks[fb&(huffmanNumChunks-1)]
+				n = uint(chunk & huffmanCountMask)
+				if n > huffmanChunkBits {
+					chunk = f.hd.links[chunk>>huffmanValueShift][(fb>>huffmanChunkBits)&f.hd.linkMask]
+					n = uint(chunk & huffmanCountMask)
+				}
+				if n <= fnb {
+					if n == 0 {
+						f.b, f.nb = fb, fnb
+						if debugDecode {
+							fmt.Println("huffsym: n==0")
+						}
+						f.err = CorruptInputError(f.roffset)
+						return
+					}
+					fb = fb >> (n & regSizeMaskUint32)
+					fnb = fnb - n
+					dist = uint32(chunk >> huffmanValueShift)
+					break
+				}
+			}
+		}
+
+		switch {
+		case dist < 4:
+			dist++
+		case dist < maxNumDist:
+			nb := uint(dist-2) >> 1
+			// have 1 bit in bottom of dist, need nb more.
+			extra := (dist & 1) << (nb & regSizeMaskUint32)
+			for fnb < nb {
+				c, err := fr.ReadByte()
+				if err != nil {
+					f.b, f.nb = fb, fnb
+					if debugDecode {
+						fmt.Println("morebits f.nb<nb:", err)
+					}
+					f.err = err
+					return
+				}
+				f.roffset++
+				fb |= uint32(c) << (fnb & regSizeMaskUint32)
+				fnb += 8
+			}
+			extra |= fb & bitMask32[nb]
+			fb >>= nb & regSizeMaskUint32
+			fnb -= nb
+			dist = 1<<((nb+1)&regSizeMaskUint32) + 1 + extra
+			// slower: dist = bitMask32[nb+1] + 2 + extra
+		default:
+			f.b, f.nb = fb, fnb
+			if debugDecode {
+				fmt.Println("dist too big:", dist, maxNumDist)
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		// No check on length; encoding can be prescient.
+		if dist > uint32(dict.histSize()) {
+			f.b, f.nb = fb, fnb
+			if debugDecode {
+				fmt.Println("dist > dict.histSize():", dist, dict.histSize())
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		f.copyLen, f.copyDist = length, int(dist)
+		goto copyHistory
+	}
+
+copyHistory:
+	// Perform a backwards copy according to RFC section 3.2.3.
+	{
+		cnt := dict.tryWriteCopy(f.copyDist, f.copyLen)
+		if cnt == 0 {
+			cnt = dict.writeCopy(f.copyDist, f.copyLen)
+		}
+		f.copyLen -= cnt
+
+		if dict.availWrite() == 0 || f.copyLen > 0 {
+			f.toRead = dict.readFlush()
+			f.step = huffmanBufioReader // We need to continue this work
+			f.stepState = stateDict
+			f.b, f.nb = fb, fnb
+			return
+		}
+		goto readLiteral
+	}
+	// Not reached
+}
+
+// Decode a single Huffman block from f.
+// hl and hd are the Huffman states for the lit/length values
+// and the distance values, respectively. If hd == nil, using the
+// fixed distance encoding associated with fixed Huffman blocks.
+func (f *decompressor) huffmanStringsReader() {
+	const (
+		stateInit = iota // Zero value must be stateInit
+		stateDict
+	)
+	fr := f.r.(*strings.Reader)
+
+	// Optimization. Compiler isn't smart enough to keep f.b,f.nb in registers,
+	// but is smart enough to keep local variables in registers, so use nb and b,
+	// inline call to moreBits and reassign b,nb back to f on return.
+	fnb, fb, dict := f.nb, f.b, &f.dict
+
+	switch f.stepState {
+	case stateInit:
+		goto readLiteral
+	case stateDict:
+		goto copyHistory
+	}
+
+readLiteral:
+	// Read literal and/or (length, distance) according to RFC section 3.2.3.
+	{
+		var v int
+		{
+			// Inlined v, err := f.huffSym(f.hl)
+			// Since a huffmanDecoder can be empty or be composed of a degenerate tree
+			// with single element, huffSym must error on these two edge cases. In both
+			// cases, the chunks slice will be 0 for the invalid sequence, leading it
+			// satisfy the n == 0 check below.
+			n := uint(f.hl.maxRead)
+			for {
+				for fnb < n {
+					c, err := fr.ReadByte()
+					if err != nil {
+						f.b, f.nb = fb, fnb
+						f.err = noEOF(err)
+						return
+					}
+					f.roffset++
+					fb |= uint32(c) << (fnb & regSizeMaskUint32)
+					fnb += 8
+				}
+				chunk := f.hl.chunks[fb&(huffmanNumChunks-1)]
+				n = uint(chunk & huffmanCountMask)
+				if n > huffmanChunkBits {
+					chunk = f.hl.links[chunk>>huffmanValueShift][(fb>>huffmanChunkBits)&f.hl.linkMask]
+					n = uint(chunk & huffmanCountMask)
+				}
+				if n <= fnb {
+					if n == 0 {
+						f.b, f.nb = fb, fnb
+						if debugDecode {
+							fmt.Println("huffsym: n==0")
+						}
+						f.err = CorruptInputError(f.roffset)
+						return
+					}
+					fb = fb >> (n & regSizeMaskUint32)
+					fnb = fnb - n
+					v = int(chunk >> huffmanValueShift)
+					break
+				}
+			}
+		}
+
+		var length int
+		switch {
+		case v < 256:
+			dict.writeByte(byte(v))
+			if dict.availWrite() == 0 {
+				f.toRead = dict.readFlush()
+				f.step = huffmanStringsReader
+				f.stepState = stateInit
+				f.b, f.nb = fb, fnb
+				return
+			}
+			goto readLiteral
+		case v == 256:
+			f.b, f.nb = fb, fnb
+			f.finishBlock()
+			return
+		// otherwise, reference to older data
+		case v < 265:
+			length = v - (257 - 3)
+		case v < maxNumLit:
+			val := decCodeToLen[(v - 257)]
+			length = int(val.length) + 3
+			n := uint(val.extra)
+			for fnb < n {
+				c, err := fr.ReadByte()
+				if err != nil {
+					f.b, f.nb = fb, fnb
+					if debugDecode {
+						fmt.Println("morebits n>0:", err)
+					}
+					f.err = err
+					return
+				}
+				f.roffset++
+				fb |= uint32(c) << (fnb & regSizeMaskUint32)
+				fnb += 8
+			}
+			length += int(fb & bitMask32[n])
+			fb >>= n & regSizeMaskUint32
+			fnb -= n
+		default:
+			if debugDecode {
+				fmt.Println(v, ">= maxNumLit")
+			}
+			f.err = CorruptInputError(f.roffset)
+			f.b, f.nb = fb, fnb
+			return
+		}
+
+		var dist uint32
+		if f.hd == nil {
+			for fnb < 5 {
+				c, err := fr.ReadByte()
+				if err != nil {
+					f.b, f.nb = fb, fnb
+					if debugDecode {
+						fmt.Println("morebits f.nb<5:", err)
+					}
+					f.err = err
+					return
+				}
+				f.roffset++
+				fb |= uint32(c) << (fnb & regSizeMaskUint32)
+				fnb += 8
+			}
+			dist = uint32(bits.Reverse8(uint8(fb & 0x1F << 3)))
+			fb >>= 5
+			fnb -= 5
+		} else {
+			// Since a huffmanDecoder can be empty or be composed of a degenerate tree
+			// with single element, huffSym must error on these two edge cases. In both
+			// cases, the chunks slice will be 0 for the invalid sequence, leading it
+			// satisfy the n == 0 check below.
+			n := uint(f.hd.maxRead)
+			// Optimization. Compiler isn't smart enough to keep f.b,f.nb in registers,
+			// but is smart enough to keep local variables in registers, so use nb and b,
+			// inline call to moreBits and reassign b,nb back to f on return.
+			for {
+				for fnb < n {
+					c, err := fr.ReadByte()
+					if err != nil {
+						f.b, f.nb = fb, fnb
+						f.err = noEOF(err)
+						return
+					}
+					f.roffset++
+					fb |= uint32(c) << (fnb & regSizeMaskUint32)
+					fnb += 8
+				}
+				chunk := f.hd.chunks[fb&(huffmanNumChunks-1)]
+				n = uint(chunk & huffmanCountMask)
+				if n > huffmanChunkBits {
+					chunk = f.hd.links[chunk>>huffmanValueShift][(fb>>huffmanChunkBits)&f.hd.linkMask]
+					n = uint(chunk & huffmanCountMask)
+				}
+				if n <= fnb {
+					if n == 0 {
+						f.b, f.nb = fb, fnb
+						if debugDecode {
+							fmt.Println("huffsym: n==0")
+						}
+						f.err = CorruptInputError(f.roffset)
+						return
+					}
+					fb = fb >> (n & regSizeMaskUint32)
+					fnb = fnb - n
+					dist = uint32(chunk >> huffmanValueShift)
+					break
+				}
+			}
+		}
+
+		switch {
+		case dist < 4:
+			dist++
+		case dist < maxNumDist:
+			nb := uint(dist-2) >> 1
+			// have 1 bit in bottom of dist, need nb more.
+			extra := (dist & 1) << (nb & regSizeMaskUint32)
+			for fnb < nb {
+				c, err := fr.ReadByte()
+				if err != nil {
+					f.b, f.nb = fb, fnb
+					if debugDecode {
+						fmt.Println("morebits f.nb<nb:", err)
+					}
+					f.err = err
+					return
+				}
+				f.roffset++
+				fb |= uint32(c) << (fnb & regSizeMaskUint32)
+				fnb += 8
+			}
+			extra |= fb & bitMask32[nb]
+			fb >>= nb & regSizeMaskUint32
+			fnb -= nb
+			dist = 1<<((nb+1)&regSizeMaskUint32) + 1 + extra
+			// slower: dist = bitMask32[nb+1] + 2 + extra
+		default:
+			f.b, f.nb = fb, fnb
+			if debugDecode {
+				fmt.Println("dist too big:", dist, maxNumDist)
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		// No check on length; encoding can be prescient.
+		if dist > uint32(dict.histSize()) {
+			f.b, f.nb = fb, fnb
+			if debugDecode {
+				fmt.Println("dist > dict.histSize():", dist, dict.histSize())
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		f.copyLen, f.copyDist = length, int(dist)
+		goto copyHistory
+	}
+
+copyHistory:
+	// Perform a backwards copy according to RFC section 3.2.3.
+	{
+		cnt := dict.tryWriteCopy(f.copyDist, f.copyLen)
+		if cnt == 0 {
+			cnt = dict.writeCopy(f.copyDist, f.copyLen)
+		}
+		f.copyLen -= cnt
+
+		if dict.availWrite() == 0 || f.copyLen > 0 {
+			f.toRead = dict.readFlush()
+			f.step = huffmanStringsReader // We need to continue this work
+			f.stepState = stateDict
+			f.b, f.nb = fb, fnb
+			return
+		}
+		goto readLiteral
+	}
+	// Not reached
+}
+
+// Decode a single Huffman block from f.
+// hl and hd are the Huffman states for the lit/length values
+// and the distance values, respectively. If hd == nil, using the
+// fixed distance encoding associated with fixed Huffman blocks.
+func (f *decompressor) huffmanGenericReader() {
+	const (
+		stateInit = iota // Zero value must be stateInit
+		stateDict
+	)
+	fr := f.r.(Reader)
+
+	// Optimization. Compiler isn't smart enough to keep f.b,f.nb in registers,
+	// but is smart enough to keep local variables in registers, so use nb and b,
+	// inline call to moreBits and reassign b,nb back to f on return.
+	fnb, fb, dict := f.nb, f.b, &f.dict
+
+	switch f.stepState {
+	case stateInit:
+		goto readLiteral
+	case stateDict:
+		goto copyHistory
+	}
+
+readLiteral:
+	// Read literal and/or (length, distance) according to RFC section 3.2.3.
+	{
+		var v int
+		{
+			// Inlined v, err := f.huffSym(f.hl)
+			// Since a huffmanDecoder can be empty or be composed of a degenerate tree
+			// with single element, huffSym must error on these two edge cases. In both
+			// cases, the chunks slice will be 0 for the invalid sequence, leading it
+			// satisfy the n == 0 check below.
+			n := uint(f.hl.maxRead)
+			for {
+				for fnb < n {
+					c, err := fr.ReadByte()
+					if err != nil {
+						f.b, f.nb = fb, fnb
+						f.err = noEOF(err)
+						return
+					}
+					f.roffset++
+					fb |= uint32(c) << (fnb & regSizeMaskUint32)
+					fnb += 8
+				}
+				chunk := f.hl.chunks[fb&(huffmanNumChunks-1)]
+				n = uint(chunk & huffmanCountMask)
+				if n > huffmanChunkBits {
+					chunk = f.hl.links[chunk>>huffmanValueShift][(fb>>huffmanChunkBits)&f.hl.linkMask]
+					n = uint(chunk & huffmanCountMask)
+				}
+				if n <= fnb {
+					if n == 0 {
+						f.b, f.nb = fb, fnb
+						if debugDecode {
+							fmt.Println("huffsym: n==0")
+						}
+						f.err = CorruptInputError(f.roffset)
+						return
+					}
+					fb = fb >> (n & regSizeMaskUint32)
+					fnb = fnb - n
+					v = int(chunk >> huffmanValueShift)
+					break
+				}
+			}
+		}
+
+		var length int
+		switch {
+		case v < 256:
+			dict.writeByte(byte(v))
+			if dict.availWrite() == 0 {
+				f.toRead = dict.readFlush()
+				f.step = huffmanGenericReader
+				f.stepState = stateInit
+				f.b, f.nb = fb, fnb
+				return
+			}
+			goto readLiteral
+		case v == 256:
+			f.b, f.nb = fb, fnb
+			f.finishBlock()
+			return
+		// otherwise, reference to older data
+		case v < 265:
+			length = v - (257 - 3)
+		case v < maxNumLit:
+			val := decCodeToLen[(v - 257)]
+			length = int(val.length) + 3
+			n := uint(val.extra)
+			for fnb < n {
+				c, err := fr.ReadByte()
+				if err != nil {
+					f.b, f.nb = fb, fnb
+					if debugDecode {
+						fmt.Println("morebits n>0:", err)
+					}
+					f.err = err
+					return
+				}
+				f.roffset++
+				fb |= uint32(c) << (fnb & regSizeMaskUint32)
+				fnb += 8
+			}
+			length += int(fb & bitMask32[n])
+			fb >>= n & regSizeMaskUint32
+			fnb -= n
+		default:
+			if debugDecode {
+				fmt.Println(v, ">= maxNumLit")
+			}
+			f.err = CorruptInputError(f.roffset)
+			f.b, f.nb = fb, fnb
+			return
+		}
+
+		var dist uint32
+		if f.hd == nil {
+			for fnb < 5 {
+				c, err := fr.ReadByte()
+				if err != nil {
+					f.b, f.nb = fb, fnb
+					if debugDecode {
+						fmt.Println("morebits f.nb<5:", err)
+					}
+					f.err = err
+					return
+				}
+				f.roffset++
+				fb |= uint32(c) << (fnb & regSizeMaskUint32)
+				fnb += 8
+			}
+			dist = uint32(bits.Reverse8(uint8(fb & 0x1F << 3)))
+			fb >>= 5
+			fnb -= 5
+		} else {
+			// Since a huffmanDecoder can be empty or be composed of a degenerate tree
+			// with single element, huffSym must error on these two edge cases. In both
+			// cases, the chunks slice will be 0 for the invalid sequence, leading it
+			// satisfy the n == 0 check below.
+			n := uint(f.hd.maxRead)
+			// Optimization. Compiler isn't smart enough to keep f.b,f.nb in registers,
+			// but is smart enough to keep local variables in registers, so use nb and b,
+			// inline call to moreBits and reassign b,nb back to f on return.
+			for {
+				for fnb < n {
+					c, err := fr.ReadByte()
+					if err != nil {
+						f.b, f.nb = fb, fnb
+						f.err = noEOF(err)
+						return
+					}
+					f.roffset++
+					fb |= uint32(c) << (fnb & regSizeMaskUint32)
+					fnb += 8
+				}
+				chunk := f.hd.chunks[fb&(huffmanNumChunks-1)]
+				n = uint(chunk & huffmanCountMask)
+				if n > huffmanChunkBits {
+					chunk = f.hd.links[chunk>>huffmanValueShift][(fb>>huffmanChunkBits)&f.hd.linkMask]
+					n = uint(chunk & huffmanCountMask)
+				}
+				if n <= fnb {
+					if n == 0 {
+						f.b, f.nb = fb, fnb
+						if debugDecode {
+							fmt.Println("huffsym: n==0")
+						}
+						f.err = CorruptInputError(f.roffset)
+						return
+					}
+					fb = fb >> (n & regSizeMaskUint32)
+					fnb = fnb - n
+					dist = uint32(chunk >> huffmanValueShift)
+					break
+				}
+			}
+		}
+
+		switch {
+		case dist < 4:
+			dist++
+		case dist < maxNumDist:
+			nb := uint(dist-2) >> 1
+			// have 1 bit in bottom of dist, need nb more.
+			extra := (dist & 1) << (nb & regSizeMaskUint32)
+			for fnb < nb {
+				c, err := fr.ReadByte()
+				if err != nil {
+					f.b, f.nb = fb, fnb
+					if debugDecode {
+						fmt.Println("morebits f.nb<nb:", err)
+					}
+					f.err = err
+					return
+				}
+				f.roffset++
+				fb |= uint32(c) << (fnb & regSizeMaskUint32)
+				fnb += 8
+			}
+			extra |= fb & bitMask32[nb]
+			fb >>= nb & regSizeMaskUint32
+			fnb -= nb
+			dist = 1<<((nb+1)&regSizeMaskUint32) + 1 + extra
+			// slower: dist = bitMask32[nb+1] + 2 + extra
+		default:
+			f.b, f.nb = fb, fnb
+			if debugDecode {
+				fmt.Println("dist too big:", dist, maxNumDist)
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		// No check on length; encoding can be prescient.
+		if dist > uint32(dict.histSize()) {
+			f.b, f.nb = fb, fnb
+			if debugDecode {
+				fmt.Println("dist > dict.histSize():", dist, dict.histSize())
+			}
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		f.copyLen, f.copyDist = length, int(dist)
+		goto copyHistory
+	}
+
+copyHistory:
+	// Perform a backwards copy according to RFC section 3.2.3.
+	{
+		cnt := dict.tryWriteCopy(f.copyDist, f.copyLen)
+		if cnt == 0 {
+			cnt = dict.writeCopy(f.copyDist, f.copyLen)
+		}
+		f.copyLen -= cnt
+
+		if dict.availWrite() == 0 || f.copyLen > 0 {
+			f.toRead = dict.readFlush()
+			f.step = huffmanGenericReader // We need to continue this work
+			f.stepState = stateDict
+			f.b, f.nb = fb, fnb
+			return
+		}
+		goto readLiteral
+	}
+	// Not reached
+}
+
+func (f *decompressor) huffmanBlockDecoder() {
+	switch f.r.(type) {
+	case *bytes.Buffer:
+		f.huffmanBytesBuffer()
+	case *bytes.Reader:
+		f.huffmanBytesReader()
+	case *bufio.Reader:
+		f.huffmanBufioReader()
+	case *strings.Reader:
+		f.huffmanStringsReader()
+	case Reader:
+		f.huffmanGenericReader()
+	default:
+		f.huffmanGenericReader()
+	}
+}
diff --git a/vendor/github.com/klauspost/compress/flate/level1.go b/vendor/github.com/klauspost/compress/flate/level1.go
new file mode 100644
index 0000000..c3581a3
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/level1.go
@@ -0,0 +1,215 @@
+package flate
+
+import (
+	"fmt"
+
+	"github.com/klauspost/compress/internal/le"
+)
+
+// fastGen maintains the table for matches,
+// and the previous byte block for level 2.
+// This is the generic implementation.
+type fastEncL1 struct {
+	fastGen
+	table [tableSize]tableEntry
+}
+
+// EncodeL1 uses a similar algorithm to level 1
+func (e *fastEncL1) Encode(dst *tokens, src []byte) {
+	const (
+		inputMargin            = 12 - 1
+		minNonLiteralBlockSize = 1 + 1 + inputMargin
+		hashBytes              = 5
+	)
+	if debugDeflate && e.cur < 0 {
+		panic(fmt.Sprint("e.cur < 0: ", e.cur))
+	}
+
+	// Protect against e.cur wraparound.
+	for e.cur >= bufferReset {
+		if len(e.hist) == 0 {
+			for i := range e.table[:] {
+				e.table[i] = tableEntry{}
+			}
+			e.cur = maxMatchOffset
+			break
+		}
+		// Shift down everything in the table that isn't already too far away.
+		minOff := e.cur + int32(len(e.hist)) - maxMatchOffset
+		for i := range e.table[:] {
+			v := e.table[i].offset
+			if v <= minOff {
+				v = 0
+			} else {
+				v = v - e.cur + maxMatchOffset
+			}
+			e.table[i].offset = v
+		}
+		e.cur = maxMatchOffset
+	}
+
+	s := e.addBlock(src)
+
+	// This check isn't in the Snappy implementation, but there, the caller
+	// instead of the callee handles this case.
+	if len(src) < minNonLiteralBlockSize {
+		// We do not fill the token table.
+		// This will be picked up by caller.
+		dst.n = uint16(len(src))
+		return
+	}
+
+	// Override src
+	src = e.hist
+	nextEmit := s
+
+	// sLimit is when to stop looking for offset/length copies. The inputMargin
+	// lets us use a fast path for emitLiteral in the main loop, while we are
+	// looking for copies.
+	sLimit := int32(len(src) - inputMargin)
+
+	// nextEmit is where in src the next emitLiteral should start from.
+	cv := load6432(src, s)
+
+	for {
+		const skipLog = 5
+		const doEvery = 2
+
+		nextS := s
+		var candidate tableEntry
+		var t int32
+		for {
+			nextHash := hashLen(cv, tableBits, hashBytes)
+			candidate = e.table[nextHash]
+			nextS = s + doEvery + (s-nextEmit)>>skipLog
+			if nextS > sLimit {
+				goto emitRemainder
+			}
+
+			now := load6432(src, nextS)
+			e.table[nextHash] = tableEntry{offset: s + e.cur}
+			nextHash = hashLen(now, tableBits, hashBytes)
+			t = candidate.offset - e.cur
+			if s-t < maxMatchOffset && uint32(cv) == load3232(src, t) {
+				e.table[nextHash] = tableEntry{offset: nextS + e.cur}
+				break
+			}
+
+			// Do one right away...
+			cv = now
+			s = nextS
+			nextS++
+			candidate = e.table[nextHash]
+			now >>= 8
+			e.table[nextHash] = tableEntry{offset: s + e.cur}
+
+			t = candidate.offset - e.cur
+			if s-t < maxMatchOffset && uint32(cv) == load3232(src, t) {
+				e.table[nextHash] = tableEntry{offset: nextS + e.cur}
+				break
+			}
+			cv = now
+			s = nextS
+		}
+
+		// A 4-byte match has been found. We'll later see if more than 4 bytes
+		// match. But, prior to the match, src[nextEmit:s] are unmatched. Emit
+		// them as literal bytes.
+		for {
+			// Invariant: we have a 4-byte match at s, and no need to emit any
+			// literal bytes prior to s.
+
+			// Extend the 4-byte match as long as possible.
+			l := e.matchlenLong(int(s+4), int(t+4), src) + 4
+
+			// Extend backwards
+			for t > 0 && s > nextEmit && le.Load8(src, t-1) == le.Load8(src, s-1) {
+				s--
+				t--
+				l++
+			}
+			if nextEmit < s {
+				if false {
+					emitLiteral(dst, src[nextEmit:s])
+				} else {
+					for _, v := range src[nextEmit:s] {
+						dst.tokens[dst.n] = token(v)
+						dst.litHist[v]++
+						dst.n++
+					}
+				}
+			}
+
+			// Save the match found
+			if false {
+				dst.AddMatchLong(l, uint32(s-t-baseMatchOffset))
+			} else {
+				// Inlined...
+				xoffset := uint32(s - t - baseMatchOffset)
+				xlength := l
+				oc := offsetCode(xoffset)
+				xoffset |= oc << 16
+				for xlength > 0 {
+					xl := xlength
+					if xl > 258 {
+						if xl > 258+baseMatchLength {
+							xl = 258
+						} else {
+							xl = 258 - baseMatchLength
+						}
+					}
+					xlength -= xl
+					xl -= baseMatchLength
+					dst.extraHist[lengthCodes1[uint8(xl)]]++
+					dst.offHist[oc]++
+					dst.tokens[dst.n] = token(matchType | uint32(xl)<<lengthShift | xoffset)
+					dst.n++
+				}
+			}
+			s += l
+			nextEmit = s
+			if nextS >= s {
+				s = nextS + 1
+			}
+			if s >= sLimit {
+				// Index first pair after match end.
+				if int(s+l+8) < len(src) {
+					cv := load6432(src, s)
+					e.table[hashLen(cv, tableBits, hashBytes)] = tableEntry{offset: s + e.cur}
+				}
+				goto emitRemainder
+			}
+
+			// We could immediately start working at s now, but to improve
+			// compression we first update the hash table at s-2 and at s. If
+			// another emitCopy is not our next move, also calculate nextHash
+			// at s+1. At least on GOARCH=amd64, these three hash calculations
+			// are faster as one load64 call (with some shifts) instead of
+			// three load32 calls.
+			x := load6432(src, s-2)
+			o := e.cur + s - 2
+			prevHash := hashLen(x, tableBits, hashBytes)
+			e.table[prevHash] = tableEntry{offset: o}
+			x >>= 16
+			currHash := hashLen(x, tableBits, hashBytes)
+			candidate = e.table[currHash]
+			e.table[currHash] = tableEntry{offset: o + 2}
+
+			t = candidate.offset - e.cur
+			if s-t > maxMatchOffset || uint32(x) != load3232(src, t) {
+				cv = x >> 8
+				s++
+				break
+			}
+		}
+	}
+
+emitRemainder:
+	if int(nextEmit) < len(src) {
+		// If nothing was added, don't encode literals.
+		if dst.n == 0 {
+			return
+		}
+		emitLiteral(dst, src[nextEmit:])
+	}
+}
diff --git a/vendor/github.com/klauspost/compress/flate/level2.go b/vendor/github.com/klauspost/compress/flate/level2.go
new file mode 100644
index 0000000..c8d047f
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/level2.go
@@ -0,0 +1,214 @@
+package flate
+
+import "fmt"
+
+// fastGen maintains the table for matches,
+// and the previous byte block for level 2.
+// This is the generic implementation.
+type fastEncL2 struct {
+	fastGen
+	table [bTableSize]tableEntry
+}
+
+// EncodeL2 uses a similar algorithm to level 1, but is capable
+// of matching across blocks giving better compression at a small slowdown.
+func (e *fastEncL2) Encode(dst *tokens, src []byte) {
+	const (
+		inputMargin            = 12 - 1
+		minNonLiteralBlockSize = 1 + 1 + inputMargin
+		hashBytes              = 5
+	)
+
+	if debugDeflate && e.cur < 0 {
+		panic(fmt.Sprint("e.cur < 0: ", e.cur))
+	}
+
+	// Protect against e.cur wraparound.
+	for e.cur >= bufferReset {
+		if len(e.hist) == 0 {
+			for i := range e.table[:] {
+				e.table[i] = tableEntry{}
+			}
+			e.cur = maxMatchOffset
+			break
+		}
+		// Shift down everything in the table that isn't already too far away.
+		minOff := e.cur + int32(len(e.hist)) - maxMatchOffset
+		for i := range e.table[:] {
+			v := e.table[i].offset
+			if v <= minOff {
+				v = 0
+			} else {
+				v = v - e.cur + maxMatchOffset
+			}
+			e.table[i].offset = v
+		}
+		e.cur = maxMatchOffset
+	}
+
+	s := e.addBlock(src)
+
+	// This check isn't in the Snappy implementation, but there, the caller
+	// instead of the callee handles this case.
+	if len(src) < minNonLiteralBlockSize {
+		// We do not fill the token table.
+		// This will be picked up by caller.
+		dst.n = uint16(len(src))
+		return
+	}
+
+	// Override src
+	src = e.hist
+	nextEmit := s
+
+	// sLimit is when to stop looking for offset/length copies. The inputMargin
+	// lets us use a fast path for emitLiteral in the main loop, while we are
+	// looking for copies.
+	sLimit := int32(len(src) - inputMargin)
+
+	// nextEmit is where in src the next emitLiteral should start from.
+	cv := load6432(src, s)
+	for {
+		// When should we start skipping if we haven't found matches in a long while.
+		const skipLog = 5
+		const doEvery = 2
+
+		nextS := s
+		var candidate tableEntry
+		for {
+			nextHash := hashLen(cv, bTableBits, hashBytes)
+			s = nextS
+			nextS = s + doEvery + (s-nextEmit)>>skipLog
+			if nextS > sLimit {
+				goto emitRemainder
+			}
+			candidate = e.table[nextHash]
+			now := load6432(src, nextS)
+			e.table[nextHash] = tableEntry{offset: s + e.cur}
+			nextHash = hashLen(now, bTableBits, hashBytes)
+
+			offset := s - (candidate.offset - e.cur)
+			if offset < maxMatchOffset && uint32(cv) == load3232(src, candidate.offset-e.cur) {
+				e.table[nextHash] = tableEntry{offset: nextS + e.cur}
+				break
+			}
+
+			// Do one right away...
+			cv = now
+			s = nextS
+			nextS++
+			candidate = e.table[nextHash]
+			now >>= 8
+			e.table[nextHash] = tableEntry{offset: s + e.cur}
+
+			offset = s - (candidate.offset - e.cur)
+			if offset < maxMatchOffset && uint32(cv) == load3232(src, candidate.offset-e.cur) {
+				break
+			}
+			cv = now
+		}
+
+		// A 4-byte match has been found. We'll later see if more than 4 bytes
+		// match. But, prior to the match, src[nextEmit:s] are unmatched. Emit
+		// them as literal bytes.
+
+		// Call emitCopy, and then see if another emitCopy could be our next
+		// move. Repeat until we find no match for the input immediately after
+		// what was consumed by the last emitCopy call.
+		//
+		// If we exit this loop normally then we need to call emitLiteral next,
+		// though we don't yet know how big the literal will be. We handle that
+		// by proceeding to the next iteration of the main loop. We also can
+		// exit this loop via goto if we get close to exhausting the input.
+		for {
+			// Invariant: we have a 4-byte match at s, and no need to emit any
+			// literal bytes prior to s.
+
+			// Extend the 4-byte match as long as possible.
+			t := candidate.offset - e.cur
+			l := e.matchlenLong(int(s+4), int(t+4), src) + 4
+
+			// Extend backwards
+			for t > 0 && s > nextEmit && src[t-1] == src[s-1] {
+				s--
+				t--
+				l++
+			}
+			if nextEmit < s {
+				if false {
+					emitLiteral(dst, src[nextEmit:s])
+				} else {
+					for _, v := range src[nextEmit:s] {
+						dst.tokens[dst.n] = token(v)
+						dst.litHist[v]++
+						dst.n++
+					}
+				}
+			}
+
+			dst.AddMatchLong(l, uint32(s-t-baseMatchOffset))
+			s += l
+			nextEmit = s
+			if nextS >= s {
+				s = nextS + 1
+			}
+
+			if s >= sLimit {
+				// Index first pair after match end.
+				if int(s+l+8) < len(src) {
+					cv := load6432(src, s)
+					e.table[hashLen(cv, bTableBits, hashBytes)] = tableEntry{offset: s + e.cur}
+				}
+				goto emitRemainder
+			}
+
+			// Store every second hash in-between, but offset by 1.
+			for i := s - l + 2; i < s-5; i += 7 {
+				x := load6432(src, i)
+				nextHash := hashLen(x, bTableBits, hashBytes)
+				e.table[nextHash] = tableEntry{offset: e.cur + i}
+				// Skip one
+				x >>= 16
+				nextHash = hashLen(x, bTableBits, hashBytes)
+				e.table[nextHash] = tableEntry{offset: e.cur + i + 2}
+				// Skip one
+				x >>= 16
+				nextHash = hashLen(x, bTableBits, hashBytes)
+				e.table[nextHash] = tableEntry{offset: e.cur + i + 4}
+			}
+
+			// We could immediately start working at s now, but to improve
+			// compression we first update the hash table at s-2 to s. If
+			// another emitCopy is not our next move, also calculate nextHash
+			// at s+1. At least on GOARCH=amd64, these three hash calculations
+			// are faster as one load64 call (with some shifts) instead of
+			// three load32 calls.
+			x := load6432(src, s-2)
+			o := e.cur + s - 2
+			prevHash := hashLen(x, bTableBits, hashBytes)
+			prevHash2 := hashLen(x>>8, bTableBits, hashBytes)
+			e.table[prevHash] = tableEntry{offset: o}
+			e.table[prevHash2] = tableEntry{offset: o + 1}
+			currHash := hashLen(x>>16, bTableBits, hashBytes)
+			candidate = e.table[currHash]
+			e.table[currHash] = tableEntry{offset: o + 2}
+
+			offset := s - (candidate.offset - e.cur)
+			if offset > maxMatchOffset || uint32(x>>16) != load3232(src, candidate.offset-e.cur) {
+				cv = x >> 24
+				s++
+				break
+			}
+		}
+	}
+
+emitRemainder:
+	if int(nextEmit) < len(src) {
+		// If nothing was added, don't encode literals.
+		if dst.n == 0 {
+			return
+		}
+
+		emitLiteral(dst, src[nextEmit:])
+	}
+}
diff --git a/vendor/github.com/klauspost/compress/flate/level3.go b/vendor/github.com/klauspost/compress/flate/level3.go
new file mode 100644
index 0000000..33f9fb1
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/level3.go
@@ -0,0 +1,241 @@
+package flate
+
+import "fmt"
+
+// fastEncL3
+type fastEncL3 struct {
+	fastGen
+	table [1 << 16]tableEntryPrev
+}
+
+// Encode uses a similar algorithm to level 2, will check up to two candidates.
+func (e *fastEncL3) Encode(dst *tokens, src []byte) {
+	const (
+		inputMargin            = 12 - 1
+		minNonLiteralBlockSize = 1 + 1 + inputMargin
+		tableBits              = 16
+		tableSize              = 1 << tableBits
+		hashBytes              = 5
+	)
+
+	if debugDeflate && e.cur < 0 {
+		panic(fmt.Sprint("e.cur < 0: ", e.cur))
+	}
+
+	// Protect against e.cur wraparound.
+	for e.cur >= bufferReset {
+		if len(e.hist) == 0 {
+			for i := range e.table[:] {
+				e.table[i] = tableEntryPrev{}
+			}
+			e.cur = maxMatchOffset
+			break
+		}
+		// Shift down everything in the table that isn't already too far away.
+		minOff := e.cur + int32(len(e.hist)) - maxMatchOffset
+		for i := range e.table[:] {
+			v := e.table[i]
+			if v.Cur.offset <= minOff {
+				v.Cur.offset = 0
+			} else {
+				v.Cur.offset = v.Cur.offset - e.cur + maxMatchOffset
+			}
+			if v.Prev.offset <= minOff {
+				v.Prev.offset = 0
+			} else {
+				v.Prev.offset = v.Prev.offset - e.cur + maxMatchOffset
+			}
+			e.table[i] = v
+		}
+		e.cur = maxMatchOffset
+	}
+
+	s := e.addBlock(src)
+
+	// Skip if too small.
+	if len(src) < minNonLiteralBlockSize {
+		// We do not fill the token table.
+		// This will be picked up by caller.
+		dst.n = uint16(len(src))
+		return
+	}
+
+	// Override src
+	src = e.hist
+	nextEmit := s
+
+	// sLimit is when to stop looking for offset/length copies. The inputMargin
+	// lets us use a fast path for emitLiteral in the main loop, while we are
+	// looking for copies.
+	sLimit := int32(len(src) - inputMargin)
+
+	// nextEmit is where in src the next emitLiteral should start from.
+	cv := load6432(src, s)
+	for {
+		const skipLog = 7
+		nextS := s
+		var candidate tableEntry
+		for {
+			nextHash := hashLen(cv, tableBits, hashBytes)
+			s = nextS
+			nextS = s + 1 + (s-nextEmit)>>skipLog
+			if nextS > sLimit {
+				goto emitRemainder
+			}
+			candidates := e.table[nextHash]
+			now := load6432(src, nextS)
+
+			// Safe offset distance until s + 4...
+			minOffset := e.cur + s - (maxMatchOffset - 4)
+			e.table[nextHash] = tableEntryPrev{Prev: candidates.Cur, Cur: tableEntry{offset: s + e.cur}}
+
+			// Check both candidates
+			candidate = candidates.Cur
+			if candidate.offset < minOffset {
+				cv = now
+				// Previous will also be invalid, we have nothing.
+				continue
+			}
+
+			if uint32(cv) == load3232(src, candidate.offset-e.cur) {
+				if candidates.Prev.offset < minOffset || uint32(cv) != load3232(src, candidates.Prev.offset-e.cur) {
+					break
+				}
+				// Both match and are valid, pick longest.
+				offset := s - (candidate.offset - e.cur)
+				o2 := s - (candidates.Prev.offset - e.cur)
+				l1, l2 := matchLen(src[s+4:], src[s-offset+4:]), matchLen(src[s+4:], src[s-o2+4:])
+				if l2 > l1 {
+					candidate = candidates.Prev
+				}
+				break
+			} else {
+				// We only check if value mismatches.
+				// Offset will always be invalid in other cases.
+				candidate = candidates.Prev
+				if candidate.offset > minOffset && uint32(cv) == load3232(src, candidate.offset-e.cur) {
+					break
+				}
+			}
+			cv = now
+		}
+
+		// Call emitCopy, and then see if another emitCopy could be our next
+		// move. Repeat until we find no match for the input immediately after
+		// what was consumed by the last emitCopy call.
+		//
+		// If we exit this loop normally then we need to call emitLiteral next,
+		// though we don't yet know how big the literal will be. We handle that
+		// by proceeding to the next iteration of the main loop. We also can
+		// exit this loop via goto if we get close to exhausting the input.
+		for {
+			// Invariant: we have a 4-byte match at s, and no need to emit any
+			// literal bytes prior to s.
+
+			// Extend the 4-byte match as long as possible.
+			//
+			t := candidate.offset - e.cur
+			l := e.matchlenLong(int(s+4), int(t+4), src) + 4
+
+			// Extend backwards
+			for t > 0 && s > nextEmit && src[t-1] == src[s-1] {
+				s--
+				t--
+				l++
+			}
+			if nextEmit < s {
+				if false {
+					emitLiteral(dst, src[nextEmit:s])
+				} else {
+					for _, v := range src[nextEmit:s] {
+						dst.tokens[dst.n] = token(v)
+						dst.litHist[v]++
+						dst.n++
+					}
+				}
+			}
+
+			dst.AddMatchLong(l, uint32(s-t-baseMatchOffset))
+			s += l
+			nextEmit = s
+			if nextS >= s {
+				s = nextS + 1
+			}
+
+			if s >= sLimit {
+				t += l
+				// Index first pair after match end.
+				if int(t+8) < len(src) && t > 0 {
+					cv = load6432(src, t)
+					nextHash := hashLen(cv, tableBits, hashBytes)
+					e.table[nextHash] = tableEntryPrev{
+						Prev: e.table[nextHash].Cur,
+						Cur:  tableEntry{offset: e.cur + t},
+					}
+				}
+				goto emitRemainder
+			}
+
+			// Store every 5th hash in-between.
+			for i := s - l + 2; i < s-5; i += 6 {
+				nextHash := hashLen(load6432(src, i), tableBits, hashBytes)
+				e.table[nextHash] = tableEntryPrev{
+					Prev: e.table[nextHash].Cur,
+					Cur:  tableEntry{offset: e.cur + i}}
+			}
+			// We could immediately start working at s now, but to improve
+			// compression we first update the hash table at s-2 to s.
+			x := load6432(src, s-2)
+			prevHash := hashLen(x, tableBits, hashBytes)
+
+			e.table[prevHash] = tableEntryPrev{
+				Prev: e.table[prevHash].Cur,
+				Cur:  tableEntry{offset: e.cur + s - 2},
+			}
+			x >>= 8
+			prevHash = hashLen(x, tableBits, hashBytes)
+
+			e.table[prevHash] = tableEntryPrev{
+				Prev: e.table[prevHash].Cur,
+				Cur:  tableEntry{offset: e.cur + s - 1},
+			}
+			x >>= 8
+			currHash := hashLen(x, tableBits, hashBytes)
+			candidates := e.table[currHash]
+			cv = x
+			e.table[currHash] = tableEntryPrev{
+				Prev: candidates.Cur,
+				Cur:  tableEntry{offset: s + e.cur},
+			}
+
+			// Check both candidates
+			candidate = candidates.Cur
+			minOffset := e.cur + s - (maxMatchOffset - 4)
+
+			if candidate.offset > minOffset {
+				if uint32(cv) == load3232(src, candidate.offset-e.cur) {
+					// Found a match...
+					continue
+				}
+				candidate = candidates.Prev
+				if candidate.offset > minOffset && uint32(cv) == load3232(src, candidate.offset-e.cur) {
+					// Match at prev...
+					continue
+				}
+			}
+			cv = x >> 8
+			s++
+			break
+		}
+	}
+
+emitRemainder:
+	if int(nextEmit) < len(src) {
+		// If nothing was added, don't encode literals.
+		if dst.n == 0 {
+			return
+		}
+
+		emitLiteral(dst, src[nextEmit:])
+	}
+}
diff --git a/vendor/github.com/klauspost/compress/flate/level4.go b/vendor/github.com/klauspost/compress/flate/level4.go
new file mode 100644
index 0000000..88509e1
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/level4.go
@@ -0,0 +1,221 @@
+package flate
+
+import "fmt"
+
+type fastEncL4 struct {
+	fastGen
+	table  [tableSize]tableEntry
+	bTable [tableSize]tableEntry
+}
+
+func (e *fastEncL4) Encode(dst *tokens, src []byte) {
+	const (
+		inputMargin            = 12 - 1
+		minNonLiteralBlockSize = 1 + 1 + inputMargin
+		hashShortBytes         = 4
+	)
+	if debugDeflate && e.cur < 0 {
+		panic(fmt.Sprint("e.cur < 0: ", e.cur))
+	}
+	// Protect against e.cur wraparound.
+	for e.cur >= bufferReset {
+		if len(e.hist) == 0 {
+			for i := range e.table[:] {
+				e.table[i] = tableEntry{}
+			}
+			for i := range e.bTable[:] {
+				e.bTable[i] = tableEntry{}
+			}
+			e.cur = maxMatchOffset
+			break
+		}
+		// Shift down everything in the table that isn't already too far away.
+		minOff := e.cur + int32(len(e.hist)) - maxMatchOffset
+		for i := range e.table[:] {
+			v := e.table[i].offset
+			if v <= minOff {
+				v = 0
+			} else {
+				v = v - e.cur + maxMatchOffset
+			}
+			e.table[i].offset = v
+		}
+		for i := range e.bTable[:] {
+			v := e.bTable[i].offset
+			if v <= minOff {
+				v = 0
+			} else {
+				v = v - e.cur + maxMatchOffset
+			}
+			e.bTable[i].offset = v
+		}
+		e.cur = maxMatchOffset
+	}
+
+	s := e.addBlock(src)
+
+	// This check isn't in the Snappy implementation, but there, the caller
+	// instead of the callee handles this case.
+	if len(src) < minNonLiteralBlockSize {
+		// We do not fill the token table.
+		// This will be picked up by caller.
+		dst.n = uint16(len(src))
+		return
+	}
+
+	// Override src
+	src = e.hist
+	nextEmit := s
+
+	// sLimit is when to stop looking for offset/length copies. The inputMargin
+	// lets us use a fast path for emitLiteral in the main loop, while we are
+	// looking for copies.
+	sLimit := int32(len(src) - inputMargin)
+
+	// nextEmit is where in src the next emitLiteral should start from.
+	cv := load6432(src, s)
+	for {
+		const skipLog = 6
+		const doEvery = 1
+
+		nextS := s
+		var t int32
+		for {
+			nextHashS := hashLen(cv, tableBits, hashShortBytes)
+			nextHashL := hash7(cv, tableBits)
+
+			s = nextS
+			nextS = s + doEvery + (s-nextEmit)>>skipLog
+			if nextS > sLimit {
+				goto emitRemainder
+			}
+			// Fetch a short+long candidate
+			sCandidate := e.table[nextHashS]
+			lCandidate := e.bTable[nextHashL]
+			next := load6432(src, nextS)
+			entry := tableEntry{offset: s + e.cur}
+			e.table[nextHashS] = entry
+			e.bTable[nextHashL] = entry
+
+			t = lCandidate.offset - e.cur
+			if s-t < maxMatchOffset && uint32(cv) == load3232(src, t) {
+				// We got a long match. Use that.
+				break
+			}
+
+			t = sCandidate.offset - e.cur
+			if s-t < maxMatchOffset && uint32(cv) == load3232(src, t) {
+				// Found a 4 match...
+				lCandidate = e.bTable[hash7(next, tableBits)]
+
+				// If the next long is a candidate, check if we should use that instead...
+				lOff := lCandidate.offset - e.cur
+				if nextS-lOff < maxMatchOffset && load3232(src, lOff) == uint32(next) {
+					l1, l2 := matchLen(src[s+4:], src[t+4:]), matchLen(src[nextS+4:], src[nextS-lOff+4:])
+					if l2 > l1 {
+						s = nextS
+						t = lCandidate.offset - e.cur
+					}
+				}
+				break
+			}
+			cv = next
+		}
+
+		// A 4-byte match has been found. We'll later see if more than 4 bytes
+		// match. But, prior to the match, src[nextEmit:s] are unmatched. Emit
+		// them as literal bytes.
+
+		// Extend the 4-byte match as long as possible.
+		l := e.matchlenLong(int(s+4), int(t+4), src) + 4
+
+		// Extend backwards
+		for t > 0 && s > nextEmit && src[t-1] == src[s-1] {
+			s--
+			t--
+			l++
+		}
+		if nextEmit < s {
+			if false {
+				emitLiteral(dst, src[nextEmit:s])
+			} else {
+				for _, v := range src[nextEmit:s] {
+					dst.tokens[dst.n] = token(v)
+					dst.litHist[v]++
+					dst.n++
+				}
+			}
+		}
+		if debugDeflate {
+			if t >= s {
+				panic("s-t")
+			}
+			if (s - t) > maxMatchOffset {
+				panic(fmt.Sprintln("mmo", t))
+			}
+			if l < baseMatchLength {
+				panic("bml")
+			}
+		}
+
+		dst.AddMatchLong(l, uint32(s-t-baseMatchOffset))
+		s += l
+		nextEmit = s
+		if nextS >= s {
+			s = nextS + 1
+		}
+
+		if s >= sLimit {
+			// Index first pair after match end.
+			if int(s+8) < len(src) {
+				cv := load6432(src, s)
+				e.table[hashLen(cv, tableBits, hashShortBytes)] = tableEntry{offset: s + e.cur}
+				e.bTable[hash7(cv, tableBits)] = tableEntry{offset: s + e.cur}
+			}
+			goto emitRemainder
+		}
+
+		// Store every 3rd hash in-between
+		if true {
+			i := nextS
+			if i < s-1 {
+				cv := load6432(src, i)
+				t := tableEntry{offset: i + e.cur}
+				t2 := tableEntry{offset: t.offset + 1}
+				e.bTable[hash7(cv, tableBits)] = t
+				e.bTable[hash7(cv>>8, tableBits)] = t2
+				e.table[hashLen(cv>>8, tableBits, hashShortBytes)] = t2
+
+				i += 3
+				for ; i < s-1; i += 3 {
+					cv := load6432(src, i)
+					t := tableEntry{offset: i + e.cur}
+					t2 := tableEntry{offset: t.offset + 1}
+					e.bTable[hash7(cv, tableBits)] = t
+					e.bTable[hash7(cv>>8, tableBits)] = t2
+					e.table[hashLen(cv>>8, tableBits, hashShortBytes)] = t2
+				}
+			}
+		}
+
+		// We could immediately start working at s now, but to improve
+		// compression we first update the hash table at s-1 and at s.
+		x := load6432(src, s-1)
+		o := e.cur + s - 1
+		prevHashS := hashLen(x, tableBits, hashShortBytes)
+		prevHashL := hash7(x, tableBits)
+		e.table[prevHashS] = tableEntry{offset: o}
+		e.bTable[prevHashL] = tableEntry{offset: o}
+		cv = x >> 8
+	}
+
+emitRemainder:
+	if int(nextEmit) < len(src) {
+		// If nothing was added, don't encode literals.
+		if dst.n == 0 {
+			return
+		}
+
+		emitLiteral(dst, src[nextEmit:])
+	}
+}
diff --git a/vendor/github.com/klauspost/compress/flate/level5.go b/vendor/github.com/klauspost/compress/flate/level5.go
new file mode 100644
index 0000000..6e5c215
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/level5.go
@@ -0,0 +1,708 @@
+package flate
+
+import "fmt"
+
+type fastEncL5 struct {
+	fastGen
+	table  [tableSize]tableEntry
+	bTable [tableSize]tableEntryPrev
+}
+
+func (e *fastEncL5) Encode(dst *tokens, src []byte) {
+	const (
+		inputMargin            = 12 - 1
+		minNonLiteralBlockSize = 1 + 1 + inputMargin
+		hashShortBytes         = 4
+	)
+	if debugDeflate && e.cur < 0 {
+		panic(fmt.Sprint("e.cur < 0: ", e.cur))
+	}
+
+	// Protect against e.cur wraparound.
+	for e.cur >= bufferReset {
+		if len(e.hist) == 0 {
+			for i := range e.table[:] {
+				e.table[i] = tableEntry{}
+			}
+			for i := range e.bTable[:] {
+				e.bTable[i] = tableEntryPrev{}
+			}
+			e.cur = maxMatchOffset
+			break
+		}
+		// Shift down everything in the table that isn't already too far away.
+		minOff := e.cur + int32(len(e.hist)) - maxMatchOffset
+		for i := range e.table[:] {
+			v := e.table[i].offset
+			if v <= minOff {
+				v = 0
+			} else {
+				v = v - e.cur + maxMatchOffset
+			}
+			e.table[i].offset = v
+		}
+		for i := range e.bTable[:] {
+			v := e.bTable[i]
+			if v.Cur.offset <= minOff {
+				v.Cur.offset = 0
+				v.Prev.offset = 0
+			} else {
+				v.Cur.offset = v.Cur.offset - e.cur + maxMatchOffset
+				if v.Prev.offset <= minOff {
+					v.Prev.offset = 0
+				} else {
+					v.Prev.offset = v.Prev.offset - e.cur + maxMatchOffset
+				}
+			}
+			e.bTable[i] = v
+		}
+		e.cur = maxMatchOffset
+	}
+
+	s := e.addBlock(src)
+
+	// This check isn't in the Snappy implementation, but there, the caller
+	// instead of the callee handles this case.
+	if len(src) < minNonLiteralBlockSize {
+		// We do not fill the token table.
+		// This will be picked up by caller.
+		dst.n = uint16(len(src))
+		return
+	}
+
+	// Override src
+	src = e.hist
+	nextEmit := s
+
+	// sLimit is when to stop looking for offset/length copies. The inputMargin
+	// lets us use a fast path for emitLiteral in the main loop, while we are
+	// looking for copies.
+	sLimit := int32(len(src) - inputMargin)
+
+	// nextEmit is where in src the next emitLiteral should start from.
+	cv := load6432(src, s)
+	for {
+		const skipLog = 6
+		const doEvery = 1
+
+		nextS := s
+		var l int32
+		var t int32
+		for {
+			nextHashS := hashLen(cv, tableBits, hashShortBytes)
+			nextHashL := hash7(cv, tableBits)
+
+			s = nextS
+			nextS = s + doEvery + (s-nextEmit)>>skipLog
+			if nextS > sLimit {
+				goto emitRemainder
+			}
+			// Fetch a short+long candidate
+			sCandidate := e.table[nextHashS]
+			lCandidate := e.bTable[nextHashL]
+			next := load6432(src, nextS)
+			entry := tableEntry{offset: s + e.cur}
+			e.table[nextHashS] = entry
+			eLong := &e.bTable[nextHashL]
+			eLong.Cur, eLong.Prev = entry, eLong.Cur
+
+			nextHashS = hashLen(next, tableBits, hashShortBytes)
+			nextHashL = hash7(next, tableBits)
+
+			t = lCandidate.Cur.offset - e.cur
+			if s-t < maxMatchOffset {
+				if uint32(cv) == load3232(src, t) {
+					// Store the next match
+					e.table[nextHashS] = tableEntry{offset: nextS + e.cur}
+					eLong := &e.bTable[nextHashL]
+					eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur}, eLong.Cur
+
+					t2 := lCandidate.Prev.offset - e.cur
+					if s-t2 < maxMatchOffset && uint32(cv) == load3232(src, t2) {
+						l = e.matchlen(int(s+4), int(t+4), src) + 4
+						ml1 := e.matchlen(int(s+4), int(t2+4), src) + 4
+						if ml1 > l {
+							t = t2
+							l = ml1
+							break
+						}
+					}
+					break
+				}
+				t = lCandidate.Prev.offset - e.cur
+				if s-t < maxMatchOffset && uint32(cv) == load3232(src, t) {
+					// Store the next match
+					e.table[nextHashS] = tableEntry{offset: nextS + e.cur}
+					eLong := &e.bTable[nextHashL]
+					eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur}, eLong.Cur
+					break
+				}
+			}
+
+			t = sCandidate.offset - e.cur
+			if s-t < maxMatchOffset && uint32(cv) == load3232(src, t) {
+				// Found a 4 match...
+				l = e.matchlen(int(s+4), int(t+4), src) + 4
+				lCandidate = e.bTable[nextHashL]
+				// Store the next match
+
+				e.table[nextHashS] = tableEntry{offset: nextS + e.cur}
+				eLong := &e.bTable[nextHashL]
+				eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur}, eLong.Cur
+
+				// If the next long is a candidate, use that...
+				t2 := lCandidate.Cur.offset - e.cur
+				if nextS-t2 < maxMatchOffset {
+					if load3232(src, t2) == uint32(next) {
+						ml := e.matchlen(int(nextS+4), int(t2+4), src) + 4
+						if ml > l {
+							t = t2
+							s = nextS
+							l = ml
+							break
+						}
+					}
+					// If the previous long is a candidate, use that...
+					t2 = lCandidate.Prev.offset - e.cur
+					if nextS-t2 < maxMatchOffset && load3232(src, t2) == uint32(next) {
+						ml := e.matchlen(int(nextS+4), int(t2+4), src) + 4
+						if ml > l {
+							t = t2
+							s = nextS
+							l = ml
+							break
+						}
+					}
+				}
+				break
+			}
+			cv = next
+		}
+
+		// A 4-byte match has been found. We'll later see if more than 4 bytes
+		// match. But, prior to the match, src[nextEmit:s] are unmatched. Emit
+		// them as literal bytes.
+
+		if l == 0 {
+			// Extend the 4-byte match as long as possible.
+			l = e.matchlenLong(int(s+4), int(t+4), src) + 4
+		} else if l == maxMatchLength {
+			l += e.matchlenLong(int(s+l), int(t+l), src)
+		}
+
+		// Try to locate a better match by checking the end of best match...
+		if sAt := s + l; l < 30 && sAt < sLimit {
+			// Allow some bytes at the beginning to mismatch.
+			// Sweet spot is 2/3 bytes depending on input.
+			// 3 is only a little better when it is but sometimes a lot worse.
+			// The skipped bytes are tested in Extend backwards,
+			// and still picked up as part of the match if they do.
+			const skipBeginning = 2
+			eLong := e.bTable[hash7(load6432(src, sAt), tableBits)].Cur.offset
+			t2 := eLong - e.cur - l + skipBeginning
+			s2 := s + skipBeginning
+			off := s2 - t2
+			if t2 >= 0 && off < maxMatchOffset && off > 0 {
+				if l2 := e.matchlenLong(int(s2), int(t2), src); l2 > l {
+					t = t2
+					l = l2
+					s = s2
+				}
+			}
+		}
+
+		// Extend backwards
+		for t > 0 && s > nextEmit && src[t-1] == src[s-1] {
+			s--
+			t--
+			l++
+		}
+		if nextEmit < s {
+			if false {
+				emitLiteral(dst, src[nextEmit:s])
+			} else {
+				for _, v := range src[nextEmit:s] {
+					dst.tokens[dst.n] = token(v)
+					dst.litHist[v]++
+					dst.n++
+				}
+			}
+		}
+		if debugDeflate {
+			if t >= s {
+				panic(fmt.Sprintln("s-t", s, t))
+			}
+			if (s - t) > maxMatchOffset {
+				panic(fmt.Sprintln("mmo", s-t))
+			}
+			if l < baseMatchLength {
+				panic("bml")
+			}
+		}
+
+		dst.AddMatchLong(l, uint32(s-t-baseMatchOffset))
+		s += l
+		nextEmit = s
+		if nextS >= s {
+			s = nextS + 1
+		}
+
+		if s >= sLimit {
+			goto emitRemainder
+		}
+
+		// Store every 3rd hash in-between.
+		if true {
+			const hashEvery = 3
+			i := s - l + 1
+			if i < s-1 {
+				cv := load6432(src, i)
+				t := tableEntry{offset: i + e.cur}
+				e.table[hashLen(cv, tableBits, hashShortBytes)] = t
+				eLong := &e.bTable[hash7(cv, tableBits)]
+				eLong.Cur, eLong.Prev = t, eLong.Cur
+
+				// Do an long at i+1
+				cv >>= 8
+				t = tableEntry{offset: t.offset + 1}
+				eLong = &e.bTable[hash7(cv, tableBits)]
+				eLong.Cur, eLong.Prev = t, eLong.Cur
+
+				// We only have enough bits for a short entry at i+2
+				cv >>= 8
+				t = tableEntry{offset: t.offset + 1}
+				e.table[hashLen(cv, tableBits, hashShortBytes)] = t
+
+				// Skip one - otherwise we risk hitting 's'
+				i += 4
+				for ; i < s-1; i += hashEvery {
+					cv := load6432(src, i)
+					t := tableEntry{offset: i + e.cur}
+					t2 := tableEntry{offset: t.offset + 1}
+					eLong := &e.bTable[hash7(cv, tableBits)]
+					eLong.Cur, eLong.Prev = t, eLong.Cur
+					e.table[hashLen(cv>>8, tableBits, hashShortBytes)] = t2
+				}
+			}
+		}
+
+		// We could immediately start working at s now, but to improve
+		// compression we first update the hash table at s-1 and at s.
+		x := load6432(src, s-1)
+		o := e.cur + s - 1
+		prevHashS := hashLen(x, tableBits, hashShortBytes)
+		prevHashL := hash7(x, tableBits)
+		e.table[prevHashS] = tableEntry{offset: o}
+		eLong := &e.bTable[prevHashL]
+		eLong.Cur, eLong.Prev = tableEntry{offset: o}, eLong.Cur
+		cv = x >> 8
+	}
+
+emitRemainder:
+	if int(nextEmit) < len(src) {
+		// If nothing was added, don't encode literals.
+		if dst.n == 0 {
+			return
+		}
+
+		emitLiteral(dst, src[nextEmit:])
+	}
+}
+
+// fastEncL5Window is a level 5 encoder,
+// but with a custom window size.
+type fastEncL5Window struct {
+	hist      []byte
+	cur       int32
+	maxOffset int32
+	table     [tableSize]tableEntry
+	bTable    [tableSize]tableEntryPrev
+}
+
+func (e *fastEncL5Window) Encode(dst *tokens, src []byte) {
+	const (
+		inputMargin            = 12 - 1
+		minNonLiteralBlockSize = 1 + 1 + inputMargin
+		hashShortBytes         = 4
+	)
+	maxMatchOffset := e.maxOffset
+	if debugDeflate && e.cur < 0 {
+		panic(fmt.Sprint("e.cur < 0: ", e.cur))
+	}
+
+	// Protect against e.cur wraparound.
+	for e.cur >= bufferReset {
+		if len(e.hist) == 0 {
+			for i := range e.table[:] {
+				e.table[i] = tableEntry{}
+			}
+			for i := range e.bTable[:] {
+				e.bTable[i] = tableEntryPrev{}
+			}
+			e.cur = maxMatchOffset
+			break
+		}
+		// Shift down everything in the table that isn't already too far away.
+		minOff := e.cur + int32(len(e.hist)) - maxMatchOffset
+		for i := range e.table[:] {
+			v := e.table[i].offset
+			if v <= minOff {
+				v = 0
+			} else {
+				v = v - e.cur + maxMatchOffset
+			}
+			e.table[i].offset = v
+		}
+		for i := range e.bTable[:] {
+			v := e.bTable[i]
+			if v.Cur.offset <= minOff {
+				v.Cur.offset = 0
+				v.Prev.offset = 0
+			} else {
+				v.Cur.offset = v.Cur.offset - e.cur + maxMatchOffset
+				if v.Prev.offset <= minOff {
+					v.Prev.offset = 0
+				} else {
+					v.Prev.offset = v.Prev.offset - e.cur + maxMatchOffset
+				}
+			}
+			e.bTable[i] = v
+		}
+		e.cur = maxMatchOffset
+	}
+
+	s := e.addBlock(src)
+
+	// This check isn't in the Snappy implementation, but there, the caller
+	// instead of the callee handles this case.
+	if len(src) < minNonLiteralBlockSize {
+		// We do not fill the token table.
+		// This will be picked up by caller.
+		dst.n = uint16(len(src))
+		return
+	}
+
+	// Override src
+	src = e.hist
+	nextEmit := s
+
+	// sLimit is when to stop looking for offset/length copies. The inputMargin
+	// lets us use a fast path for emitLiteral in the main loop, while we are
+	// looking for copies.
+	sLimit := int32(len(src) - inputMargin)
+
+	// nextEmit is where in src the next emitLiteral should start from.
+	cv := load6432(src, s)
+	for {
+		const skipLog = 6
+		const doEvery = 1
+
+		nextS := s
+		var l int32
+		var t int32
+		for {
+			nextHashS := hashLen(cv, tableBits, hashShortBytes)
+			nextHashL := hash7(cv, tableBits)
+
+			s = nextS
+			nextS = s + doEvery + (s-nextEmit)>>skipLog
+			if nextS > sLimit {
+				goto emitRemainder
+			}
+			// Fetch a short+long candidate
+			sCandidate := e.table[nextHashS]
+			lCandidate := e.bTable[nextHashL]
+			next := load6432(src, nextS)
+			entry := tableEntry{offset: s + e.cur}
+			e.table[nextHashS] = entry
+			eLong := &e.bTable[nextHashL]
+			eLong.Cur, eLong.Prev = entry, eLong.Cur
+
+			nextHashS = hashLen(next, tableBits, hashShortBytes)
+			nextHashL = hash7(next, tableBits)
+
+			t = lCandidate.Cur.offset - e.cur
+			if s-t < maxMatchOffset {
+				if uint32(cv) == load3232(src, t) {
+					// Store the next match
+					e.table[nextHashS] = tableEntry{offset: nextS + e.cur}
+					eLong := &e.bTable[nextHashL]
+					eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur}, eLong.Cur
+
+					t2 := lCandidate.Prev.offset - e.cur
+					if s-t2 < maxMatchOffset && uint32(cv) == load3232(src, t2) {
+						l = e.matchlen(s+4, t+4, src) + 4
+						ml1 := e.matchlen(s+4, t2+4, src) + 4
+						if ml1 > l {
+							t = t2
+							l = ml1
+							break
+						}
+					}
+					break
+				}
+				t = lCandidate.Prev.offset - e.cur
+				if s-t < maxMatchOffset && uint32(cv) == load3232(src, t) {
+					// Store the next match
+					e.table[nextHashS] = tableEntry{offset: nextS + e.cur}
+					eLong := &e.bTable[nextHashL]
+					eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur}, eLong.Cur
+					break
+				}
+			}
+
+			t = sCandidate.offset - e.cur
+			if s-t < maxMatchOffset && uint32(cv) == load3232(src, t) {
+				// Found a 4 match...
+				l = e.matchlen(s+4, t+4, src) + 4
+				lCandidate = e.bTable[nextHashL]
+				// Store the next match
+
+				e.table[nextHashS] = tableEntry{offset: nextS + e.cur}
+				eLong := &e.bTable[nextHashL]
+				eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur}, eLong.Cur
+
+				// If the next long is a candidate, use that...
+				t2 := lCandidate.Cur.offset - e.cur
+				if nextS-t2 < maxMatchOffset {
+					if load3232(src, t2) == uint32(next) {
+						ml := e.matchlen(nextS+4, t2+4, src) + 4
+						if ml > l {
+							t = t2
+							s = nextS
+							l = ml
+							break
+						}
+					}
+					// If the previous long is a candidate, use that...
+					t2 = lCandidate.Prev.offset - e.cur
+					if nextS-t2 < maxMatchOffset && load3232(src, t2) == uint32(next) {
+						ml := e.matchlen(nextS+4, t2+4, src) + 4
+						if ml > l {
+							t = t2
+							s = nextS
+							l = ml
+							break
+						}
+					}
+				}
+				break
+			}
+			cv = next
+		}
+
+		// A 4-byte match has been found. We'll later see if more than 4 bytes
+		// match. But, prior to the match, src[nextEmit:s] are unmatched. Emit
+		// them as literal bytes.
+
+		if l == 0 {
+			// Extend the 4-byte match as long as possible.
+			l = e.matchlenLong(s+4, t+4, src) + 4
+		} else if l == maxMatchLength {
+			l += e.matchlenLong(s+l, t+l, src)
+		}
+
+		// Try to locate a better match by checking the end of best match...
+		if sAt := s + l; l < 30 && sAt < sLimit {
+			// Allow some bytes at the beginning to mismatch.
+			// Sweet spot is 2/3 bytes depending on input.
+			// 3 is only a little better when it is but sometimes a lot worse.
+			// The skipped bytes are tested in Extend backwards,
+			// and still picked up as part of the match if they do.
+			const skipBeginning = 2
+			eLong := e.bTable[hash7(load6432(src, sAt), tableBits)].Cur.offset
+			t2 := eLong - e.cur - l + skipBeginning
+			s2 := s + skipBeginning
+			off := s2 - t2
+			if t2 >= 0 && off < maxMatchOffset && off > 0 {
+				if l2 := e.matchlenLong(s2, t2, src); l2 > l {
+					t = t2
+					l = l2
+					s = s2
+				}
+			}
+		}
+
+		// Extend backwards
+		for t > 0 && s > nextEmit && src[t-1] == src[s-1] {
+			s--
+			t--
+			l++
+		}
+		if nextEmit < s {
+			if false {
+				emitLiteral(dst, src[nextEmit:s])
+			} else {
+				for _, v := range src[nextEmit:s] {
+					dst.tokens[dst.n] = token(v)
+					dst.litHist[v]++
+					dst.n++
+				}
+			}
+		}
+		if debugDeflate {
+			if t >= s {
+				panic(fmt.Sprintln("s-t", s, t))
+			}
+			if (s - t) > maxMatchOffset {
+				panic(fmt.Sprintln("mmo", s-t))
+			}
+			if l < baseMatchLength {
+				panic("bml")
+			}
+		}
+
+		dst.AddMatchLong(l, uint32(s-t-baseMatchOffset))
+		s += l
+		nextEmit = s
+		if nextS >= s {
+			s = nextS + 1
+		}
+
+		if s >= sLimit {
+			goto emitRemainder
+		}
+
+		// Store every 3rd hash in-between.
+		if true {
+			const hashEvery = 3
+			i := s - l + 1
+			if i < s-1 {
+				cv := load6432(src, i)
+				t := tableEntry{offset: i + e.cur}
+				e.table[hashLen(cv, tableBits, hashShortBytes)] = t
+				eLong := &e.bTable[hash7(cv, tableBits)]
+				eLong.Cur, eLong.Prev = t, eLong.Cur
+
+				// Do an long at i+1
+				cv >>= 8
+				t = tableEntry{offset: t.offset + 1}
+				eLong = &e.bTable[hash7(cv, tableBits)]
+				eLong.Cur, eLong.Prev = t, eLong.Cur
+
+				// We only have enough bits for a short entry at i+2
+				cv >>= 8
+				t = tableEntry{offset: t.offset + 1}
+				e.table[hashLen(cv, tableBits, hashShortBytes)] = t
+
+				// Skip one - otherwise we risk hitting 's'
+				i += 4
+				for ; i < s-1; i += hashEvery {
+					cv := load6432(src, i)
+					t := tableEntry{offset: i + e.cur}
+					t2 := tableEntry{offset: t.offset + 1}
+					eLong := &e.bTable[hash7(cv, tableBits)]
+					eLong.Cur, eLong.Prev = t, eLong.Cur
+					e.table[hashLen(cv>>8, tableBits, hashShortBytes)] = t2
+				}
+			}
+		}
+
+		// We could immediately start working at s now, but to improve
+		// compression we first update the hash table at s-1 and at s.
+		x := load6432(src, s-1)
+		o := e.cur + s - 1
+		prevHashS := hashLen(x, tableBits, hashShortBytes)
+		prevHashL := hash7(x, tableBits)
+		e.table[prevHashS] = tableEntry{offset: o}
+		eLong := &e.bTable[prevHashL]
+		eLong.Cur, eLong.Prev = tableEntry{offset: o}, eLong.Cur
+		cv = x >> 8
+	}
+
+emitRemainder:
+	if int(nextEmit) < len(src) {
+		// If nothing was added, don't encode literals.
+		if dst.n == 0 {
+			return
+		}
+
+		emitLiteral(dst, src[nextEmit:])
+	}
+}
+
+// Reset the encoding table.
+func (e *fastEncL5Window) Reset() {
+	// We keep the same allocs, since we are compressing the same block sizes.
+	if cap(e.hist) < allocHistory {
+		e.hist = make([]byte, 0, allocHistory)
+	}
+
+	// We offset current position so everything will be out of reach.
+	// If we are above the buffer reset it will be cleared anyway since len(hist) == 0.
+	if e.cur <= int32(bufferReset) {
+		e.cur += e.maxOffset + int32(len(e.hist))
+	}
+	e.hist = e.hist[:0]
+}
+
+func (e *fastEncL5Window) addBlock(src []byte) int32 {
+	// check if we have space already
+	maxMatchOffset := e.maxOffset
+
+	if len(e.hist)+len(src) > cap(e.hist) {
+		if cap(e.hist) == 0 {
+			e.hist = make([]byte, 0, allocHistory)
+		} else {
+			if cap(e.hist) < int(maxMatchOffset*2) {
+				panic("unexpected buffer size")
+			}
+			// Move down
+			offset := int32(len(e.hist)) - maxMatchOffset
+			copy(e.hist[0:maxMatchOffset], e.hist[offset:])
+			e.cur += offset
+			e.hist = e.hist[:maxMatchOffset]
+		}
+	}
+	s := int32(len(e.hist))
+	e.hist = append(e.hist, src...)
+	return s
+}
+
+// matchlen will return the match length between offsets and t in src.
+// The maximum length returned is maxMatchLength - 4.
+// It is assumed that s > t, that t >=0 and s < len(src).
+func (e *fastEncL5Window) matchlen(s, t int32, src []byte) int32 {
+	if debugDecode {
+		if t >= s {
+			panic(fmt.Sprint("t >=s:", t, s))
+		}
+		if int(s) >= len(src) {
+			panic(fmt.Sprint("s >= len(src):", s, len(src)))
+		}
+		if t < 0 {
+			panic(fmt.Sprint("t < 0:", t))
+		}
+		if s-t > e.maxOffset {
+			panic(fmt.Sprint(s, "-", t, "(", s-t, ") > maxMatchLength (", maxMatchOffset, ")"))
+		}
+	}
+	s1 := int(s) + maxMatchLength - 4
+	if s1 > len(src) {
+		s1 = len(src)
+	}
+
+	// Extend the match to be as long as possible.
+	return int32(matchLen(src[s:s1], src[t:]))
+}
+
+// matchlenLong will return the match length between offsets and t in src.
+// It is assumed that s > t, that t >=0 and s < len(src).
+func (e *fastEncL5Window) matchlenLong(s, t int32, src []byte) int32 {
+	if debugDeflate {
+		if t >= s {
+			panic(fmt.Sprint("t >=s:", t, s))
+		}
+		if int(s) >= len(src) {
+			panic(fmt.Sprint("s >= len(src):", s, len(src)))
+		}
+		if t < 0 {
+			panic(fmt.Sprint("t < 0:", t))
+		}
+		if s-t > e.maxOffset {
+			panic(fmt.Sprint(s, "-", t, "(", s-t, ") > maxMatchLength (", maxMatchOffset, ")"))
+		}
+	}
+	// Extend the match to be as long as possible.
+	return int32(matchLen(src[s:], src[t:]))
+}
diff --git a/vendor/github.com/klauspost/compress/flate/level6.go b/vendor/github.com/klauspost/compress/flate/level6.go
new file mode 100644
index 0000000..96f5bb4
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/level6.go
@@ -0,0 +1,325 @@
+package flate
+
+import "fmt"
+
+type fastEncL6 struct {
+	fastGen
+	table  [tableSize]tableEntry
+	bTable [tableSize]tableEntryPrev
+}
+
+func (e *fastEncL6) Encode(dst *tokens, src []byte) {
+	const (
+		inputMargin            = 12 - 1
+		minNonLiteralBlockSize = 1 + 1 + inputMargin
+		hashShortBytes         = 4
+	)
+	if debugDeflate && e.cur < 0 {
+		panic(fmt.Sprint("e.cur < 0: ", e.cur))
+	}
+
+	// Protect against e.cur wraparound.
+	for e.cur >= bufferReset {
+		if len(e.hist) == 0 {
+			for i := range e.table[:] {
+				e.table[i] = tableEntry{}
+			}
+			for i := range e.bTable[:] {
+				e.bTable[i] = tableEntryPrev{}
+			}
+			e.cur = maxMatchOffset
+			break
+		}
+		// Shift down everything in the table that isn't already too far away.
+		minOff := e.cur + int32(len(e.hist)) - maxMatchOffset
+		for i := range e.table[:] {
+			v := e.table[i].offset
+			if v <= minOff {
+				v = 0
+			} else {
+				v = v - e.cur + maxMatchOffset
+			}
+			e.table[i].offset = v
+		}
+		for i := range e.bTable[:] {
+			v := e.bTable[i]
+			if v.Cur.offset <= minOff {
+				v.Cur.offset = 0
+				v.Prev.offset = 0
+			} else {
+				v.Cur.offset = v.Cur.offset - e.cur + maxMatchOffset
+				if v.Prev.offset <= minOff {
+					v.Prev.offset = 0
+				} else {
+					v.Prev.offset = v.Prev.offset - e.cur + maxMatchOffset
+				}
+			}
+			e.bTable[i] = v
+		}
+		e.cur = maxMatchOffset
+	}
+
+	s := e.addBlock(src)
+
+	// This check isn't in the Snappy implementation, but there, the caller
+	// instead of the callee handles this case.
+	if len(src) < minNonLiteralBlockSize {
+		// We do not fill the token table.
+		// This will be picked up by caller.
+		dst.n = uint16(len(src))
+		return
+	}
+
+	// Override src
+	src = e.hist
+	nextEmit := s
+
+	// sLimit is when to stop looking for offset/length copies. The inputMargin
+	// lets us use a fast path for emitLiteral in the main loop, while we are
+	// looking for copies.
+	sLimit := int32(len(src) - inputMargin)
+
+	// nextEmit is where in src the next emitLiteral should start from.
+	cv := load6432(src, s)
+	// Repeat MUST be > 1 and within range
+	repeat := int32(1)
+	for {
+		const skipLog = 7
+		const doEvery = 1
+
+		nextS := s
+		var l int32
+		var t int32
+		for {
+			nextHashS := hashLen(cv, tableBits, hashShortBytes)
+			nextHashL := hash7(cv, tableBits)
+			s = nextS
+			nextS = s + doEvery + (s-nextEmit)>>skipLog
+			if nextS > sLimit {
+				goto emitRemainder
+			}
+			// Fetch a short+long candidate
+			sCandidate := e.table[nextHashS]
+			lCandidate := e.bTable[nextHashL]
+			next := load6432(src, nextS)
+			entry := tableEntry{offset: s + e.cur}
+			e.table[nextHashS] = entry
+			eLong := &e.bTable[nextHashL]
+			eLong.Cur, eLong.Prev = entry, eLong.Cur
+
+			// Calculate hashes of 'next'
+			nextHashS = hashLen(next, tableBits, hashShortBytes)
+			nextHashL = hash7(next, tableBits)
+
+			t = lCandidate.Cur.offset - e.cur
+			if s-t < maxMatchOffset {
+				if uint32(cv) == load3232(src, t) {
+					// Long candidate matches at least 4 bytes.
+
+					// Store the next match
+					e.table[nextHashS] = tableEntry{offset: nextS + e.cur}
+					eLong := &e.bTable[nextHashL]
+					eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur}, eLong.Cur
+
+					// Check the previous long candidate as well.
+					t2 := lCandidate.Prev.offset - e.cur
+					if s-t2 < maxMatchOffset && uint32(cv) == load3232(src, t2) {
+						l = e.matchlen(int(s+4), int(t+4), src) + 4
+						ml1 := e.matchlen(int(s+4), int(t2+4), src) + 4
+						if ml1 > l {
+							t = t2
+							l = ml1
+							break
+						}
+					}
+					break
+				}
+				// Current value did not match, but check if previous long value does.
+				t = lCandidate.Prev.offset - e.cur
+				if s-t < maxMatchOffset && uint32(cv) == load3232(src, t) {
+					// Store the next match
+					e.table[nextHashS] = tableEntry{offset: nextS + e.cur}
+					eLong := &e.bTable[nextHashL]
+					eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur}, eLong.Cur
+					break
+				}
+			}
+
+			t = sCandidate.offset - e.cur
+			if s-t < maxMatchOffset && uint32(cv) == load3232(src, t) {
+				// Found a 4 match...
+				l = e.matchlen(int(s+4), int(t+4), src) + 4
+
+				// Look up next long candidate (at nextS)
+				lCandidate = e.bTable[nextHashL]
+
+				// Store the next match
+				e.table[nextHashS] = tableEntry{offset: nextS + e.cur}
+				eLong := &e.bTable[nextHashL]
+				eLong.Cur, eLong.Prev = tableEntry{offset: nextS + e.cur}, eLong.Cur
+
+				// Check repeat at s + repOff
+				const repOff = 1
+				t2 := s - repeat + repOff
+				if load3232(src, t2) == uint32(cv>>(8*repOff)) {
+					ml := e.matchlen(int(s+4+repOff), int(t2+4), src) + 4
+					if ml > l {
+						t = t2
+						l = ml
+						s += repOff
+						// Not worth checking more.
+						break
+					}
+				}
+
+				// If the next long is a candidate, use that...
+				t2 = lCandidate.Cur.offset - e.cur
+				if nextS-t2 < maxMatchOffset {
+					if load3232(src, t2) == uint32(next) {
+						ml := e.matchlen(int(nextS+4), int(t2+4), src) + 4
+						if ml > l {
+							t = t2
+							s = nextS
+							l = ml
+							// This is ok, but check previous as well.
+						}
+					}
+					// If the previous long is a candidate, use that...
+					t2 = lCandidate.Prev.offset - e.cur
+					if nextS-t2 < maxMatchOffset && load3232(src, t2) == uint32(next) {
+						ml := e.matchlen(int(nextS+4), int(t2+4), src) + 4
+						if ml > l {
+							t = t2
+							s = nextS
+							l = ml
+							break
+						}
+					}
+				}
+				break
+			}
+			cv = next
+		}
+
+		// A 4-byte match has been found. We'll later see if more than 4 bytes
+		// match. But, prior to the match, src[nextEmit:s] are unmatched. Emit
+		// them as literal bytes.
+
+		// Extend the 4-byte match as long as possible.
+		if l == 0 {
+			l = e.matchlenLong(int(s+4), int(t+4), src) + 4
+		} else if l == maxMatchLength {
+			l += e.matchlenLong(int(s+l), int(t+l), src)
+		}
+
+		// Try to locate a better match by checking the end-of-match...
+		if sAt := s + l; sAt < sLimit {
+			// Allow some bytes at the beginning to mismatch.
+			// Sweet spot is 2/3 bytes depending on input.
+			// 3 is only a little better when it is but sometimes a lot worse.
+			// The skipped bytes are tested in Extend backwards,
+			// and still picked up as part of the match if they do.
+			const skipBeginning = 2
+			eLong := &e.bTable[hash7(load6432(src, sAt), tableBits)]
+			// Test current
+			t2 := eLong.Cur.offset - e.cur - l + skipBeginning
+			s2 := s + skipBeginning
+			off := s2 - t2
+			if off < maxMatchOffset {
+				if off > 0 && t2 >= 0 {
+					if l2 := e.matchlenLong(int(s2), int(t2), src); l2 > l {
+						t = t2
+						l = l2
+						s = s2
+					}
+				}
+				// Test next:
+				t2 = eLong.Prev.offset - e.cur - l + skipBeginning
+				off := s2 - t2
+				if off > 0 && off < maxMatchOffset && t2 >= 0 {
+					if l2 := e.matchlenLong(int(s2), int(t2), src); l2 > l {
+						t = t2
+						l = l2
+						s = s2
+					}
+				}
+			}
+		}
+
+		// Extend backwards
+		for t > 0 && s > nextEmit && src[t-1] == src[s-1] {
+			s--
+			t--
+			l++
+		}
+		if nextEmit < s {
+			if false {
+				emitLiteral(dst, src[nextEmit:s])
+			} else {
+				for _, v := range src[nextEmit:s] {
+					dst.tokens[dst.n] = token(v)
+					dst.litHist[v]++
+					dst.n++
+				}
+			}
+		}
+		if false {
+			if t >= s {
+				panic(fmt.Sprintln("s-t", s, t))
+			}
+			if (s - t) > maxMatchOffset {
+				panic(fmt.Sprintln("mmo", s-t))
+			}
+			if l < baseMatchLength {
+				panic("bml")
+			}
+		}
+
+		dst.AddMatchLong(l, uint32(s-t-baseMatchOffset))
+		repeat = s - t
+		s += l
+		nextEmit = s
+		if nextS >= s {
+			s = nextS + 1
+		}
+
+		if s >= sLimit {
+			// Index after match end.
+			for i := nextS + 1; i < int32(len(src))-8; i += 2 {
+				cv := load6432(src, i)
+				e.table[hashLen(cv, tableBits, hashShortBytes)] = tableEntry{offset: i + e.cur}
+				eLong := &e.bTable[hash7(cv, tableBits)]
+				eLong.Cur, eLong.Prev = tableEntry{offset: i + e.cur}, eLong.Cur
+			}
+			goto emitRemainder
+		}
+
+		// Store every long hash in-between and every second short.
+		if true {
+			for i := nextS + 1; i < s-1; i += 2 {
+				cv := load6432(src, i)
+				t := tableEntry{offset: i + e.cur}
+				t2 := tableEntry{offset: t.offset + 1}
+				eLong := &e.bTable[hash7(cv, tableBits)]
+				eLong2 := &e.bTable[hash7(cv>>8, tableBits)]
+				e.table[hashLen(cv, tableBits, hashShortBytes)] = t
+				eLong.Cur, eLong.Prev = t, eLong.Cur
+				eLong2.Cur, eLong2.Prev = t2, eLong2.Cur
+			}
+		}
+
+		// We could immediately start working at s now, but to improve
+		// compression we first update the hash table at s-1 and at s.
+		cv = load6432(src, s)
+	}
+
+emitRemainder:
+	if int(nextEmit) < len(src) {
+		// If nothing was added, don't encode literals.
+		if dst.n == 0 {
+			return
+		}
+
+		emitLiteral(dst, src[nextEmit:])
+	}
+}
diff --git a/vendor/github.com/klauspost/compress/flate/matchlen_generic.go b/vendor/github.com/klauspost/compress/flate/matchlen_generic.go
new file mode 100644
index 0000000..6149384
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/matchlen_generic.go
@@ -0,0 +1,34 @@
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+
+package flate
+
+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/flate/regmask_amd64.go b/vendor/github.com/klauspost/compress/flate/regmask_amd64.go
new file mode 100644
index 0000000..6ed2806
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/regmask_amd64.go
@@ -0,0 +1,37 @@
+package flate
+
+const (
+	// Masks for shifts with register sizes of the shift value.
+	// This can be used to work around the x86 design of shifting by mod register size.
+	// It can be used when a variable shift is always smaller than the register size.
+
+	// reg8SizeMaskX - shift value is 8 bits, shifted is X
+	reg8SizeMask8  = 7
+	reg8SizeMask16 = 15
+	reg8SizeMask32 = 31
+	reg8SizeMask64 = 63
+
+	// reg16SizeMaskX - shift value is 16 bits, shifted is X
+	reg16SizeMask8  = reg8SizeMask8
+	reg16SizeMask16 = reg8SizeMask16
+	reg16SizeMask32 = reg8SizeMask32
+	reg16SizeMask64 = reg8SizeMask64
+
+	// reg32SizeMaskX - shift value is 32 bits, shifted is X
+	reg32SizeMask8  = reg8SizeMask8
+	reg32SizeMask16 = reg8SizeMask16
+	reg32SizeMask32 = reg8SizeMask32
+	reg32SizeMask64 = reg8SizeMask64
+
+	// reg64SizeMaskX - shift value is 64 bits, shifted is X
+	reg64SizeMask8  = reg8SizeMask8
+	reg64SizeMask16 = reg8SizeMask16
+	reg64SizeMask32 = reg8SizeMask32
+	reg64SizeMask64 = reg8SizeMask64
+
+	// regSizeMaskUintX - shift value is uint, shifted is X
+	regSizeMaskUint8  = reg8SizeMask8
+	regSizeMaskUint16 = reg8SizeMask16
+	regSizeMaskUint32 = reg8SizeMask32
+	regSizeMaskUint64 = reg8SizeMask64
+)
diff --git a/vendor/github.com/klauspost/compress/flate/regmask_other.go b/vendor/github.com/klauspost/compress/flate/regmask_other.go
new file mode 100644
index 0000000..1b7a2cb
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/regmask_other.go
@@ -0,0 +1,40 @@
+//go:build !amd64
+// +build !amd64
+
+package flate
+
+const (
+	// Masks for shifts with register sizes of the shift value.
+	// This can be used to work around the x86 design of shifting by mod register size.
+	// It can be used when a variable shift is always smaller than the register size.
+
+	// reg8SizeMaskX - shift value is 8 bits, shifted is X
+	reg8SizeMask8  = 0xff
+	reg8SizeMask16 = 0xff
+	reg8SizeMask32 = 0xff
+	reg8SizeMask64 = 0xff
+
+	// reg16SizeMaskX - shift value is 16 bits, shifted is X
+	reg16SizeMask8  = 0xffff
+	reg16SizeMask16 = 0xffff
+	reg16SizeMask32 = 0xffff
+	reg16SizeMask64 = 0xffff
+
+	// reg32SizeMaskX - shift value is 32 bits, shifted is X
+	reg32SizeMask8  = 0xffffffff
+	reg32SizeMask16 = 0xffffffff
+	reg32SizeMask32 = 0xffffffff
+	reg32SizeMask64 = 0xffffffff
+
+	// reg64SizeMaskX - shift value is 64 bits, shifted is X
+	reg64SizeMask8  = 0xffffffffffffffff
+	reg64SizeMask16 = 0xffffffffffffffff
+	reg64SizeMask32 = 0xffffffffffffffff
+	reg64SizeMask64 = 0xffffffffffffffff
+
+	// regSizeMaskUintX - shift value is uint, shifted is X
+	regSizeMaskUint8  = ^uint(0)
+	regSizeMaskUint16 = ^uint(0)
+	regSizeMaskUint32 = ^uint(0)
+	regSizeMaskUint64 = ^uint(0)
+)
diff --git a/vendor/github.com/klauspost/compress/flate/stateless.go b/vendor/github.com/klauspost/compress/flate/stateless.go
new file mode 100644
index 0000000..13b9b10
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/stateless.go
@@ -0,0 +1,313 @@
+package flate
+
+import (
+	"io"
+	"math"
+	"sync"
+
+	"github.com/klauspost/compress/internal/le"
+)
+
+const (
+	maxStatelessBlock = math.MaxInt16
+	// dictionary will be taken from maxStatelessBlock, so limit it.
+	maxStatelessDict = 8 << 10
+
+	slTableBits  = 13
+	slTableSize  = 1 << slTableBits
+	slTableShift = 32 - slTableBits
+)
+
+type statelessWriter struct {
+	dst    io.Writer
+	closed bool
+}
+
+func (s *statelessWriter) Close() error {
+	if s.closed {
+		return nil
+	}
+	s.closed = true
+	// Emit EOF block
+	return StatelessDeflate(s.dst, nil, true, nil)
+}
+
+func (s *statelessWriter) Write(p []byte) (n int, err error) {
+	err = StatelessDeflate(s.dst, p, false, nil)
+	if err != nil {
+		return 0, err
+	}
+	return len(p), nil
+}
+
+func (s *statelessWriter) Reset(w io.Writer) {
+	s.dst = w
+	s.closed = false
+}
+
+// NewStatelessWriter will do compression but without maintaining any state
+// between Write calls.
+// There will be no memory kept between Write calls,
+// but compression and speed will be suboptimal.
+// Because of this, the size of actual Write calls will affect output size.
+func NewStatelessWriter(dst io.Writer) io.WriteCloser {
+	return &statelessWriter{dst: dst}
+}
+
+// bitWriterPool contains bit writers that can be reused.
+var bitWriterPool = sync.Pool{
+	New: func() interface{} {
+		return newHuffmanBitWriter(nil)
+	},
+}
+
+// StatelessDeflate allows compressing directly to a Writer without retaining state.
+// When returning everything will be flushed.
+// Up to 8KB of an optional dictionary can be given which is presumed to precede the block.
+// Longer dictionaries will be truncated and will still produce valid output.
+// Sending nil dictionary is perfectly fine.
+func StatelessDeflate(out io.Writer, in []byte, eof bool, dict []byte) error {
+	var dst tokens
+	bw := bitWriterPool.Get().(*huffmanBitWriter)
+	bw.reset(out)
+	defer func() {
+		// don't keep a reference to our output
+		bw.reset(nil)
+		bitWriterPool.Put(bw)
+	}()
+	if eof && len(in) == 0 {
+		// Just write an EOF block.
+		// Could be faster...
+		bw.writeStoredHeader(0, true)
+		bw.flush()
+		return bw.err
+	}
+
+	// Truncate dict
+	if len(dict) > maxStatelessDict {
+		dict = dict[len(dict)-maxStatelessDict:]
+	}
+
+	// For subsequent loops, keep shallow dict reference to avoid alloc+copy.
+	var inDict []byte
+
+	for len(in) > 0 {
+		todo := in
+		if len(inDict) > 0 {
+			if len(todo) > maxStatelessBlock-maxStatelessDict {
+				todo = todo[:maxStatelessBlock-maxStatelessDict]
+			}
+		} else if len(todo) > maxStatelessBlock-len(dict) {
+			todo = todo[:maxStatelessBlock-len(dict)]
+		}
+		inOrg := in
+		in = in[len(todo):]
+		uncompressed := todo
+		if len(dict) > 0 {
+			// combine dict and source
+			bufLen := len(todo) + len(dict)
+			combined := make([]byte, bufLen)
+			copy(combined, dict)
+			copy(combined[len(dict):], todo)
+			todo = combined
+		}
+		// Compress
+		if len(inDict) == 0 {
+			statelessEnc(&dst, todo, int16(len(dict)))
+		} else {
+			statelessEnc(&dst, inDict[:maxStatelessDict+len(todo)], maxStatelessDict)
+		}
+		isEof := eof && len(in) == 0
+
+		if dst.n == 0 {
+			bw.writeStoredHeader(len(uncompressed), isEof)
+			if bw.err != nil {
+				return bw.err
+			}
+			bw.writeBytes(uncompressed)
+		} else if int(dst.n) > len(uncompressed)-len(uncompressed)>>4 {
+			// If we removed less than 1/16th, huffman compress the block.
+			bw.writeBlockHuff(isEof, uncompressed, len(in) == 0)
+		} else {
+			bw.writeBlockDynamic(&dst, isEof, uncompressed, len(in) == 0)
+		}
+		if len(in) > 0 {
+			// Retain a dict if we have more
+			inDict = inOrg[len(uncompressed)-maxStatelessDict:]
+			dict = nil
+			dst.Reset()
+		}
+		if bw.err != nil {
+			return bw.err
+		}
+	}
+	if !eof {
+		// Align, only a stored block can do that.
+		bw.writeStoredHeader(0, false)
+	}
+	bw.flush()
+	return bw.err
+}
+
+func hashSL(u uint32) uint32 {
+	return (u * 0x1e35a7bd) >> slTableShift
+}
+
+func load3216(b []byte, i int16) uint32 {
+	return le.Load32(b, i)
+}
+
+func load6416(b []byte, i int16) uint64 {
+	return le.Load64(b, i)
+}
+
+func statelessEnc(dst *tokens, src []byte, startAt int16) {
+	const (
+		inputMargin            = 12 - 1
+		minNonLiteralBlockSize = 1 + 1 + inputMargin
+	)
+
+	type tableEntry struct {
+		offset int16
+	}
+
+	var table [slTableSize]tableEntry
+
+	// This check isn't in the Snappy implementation, but there, the caller
+	// instead of the callee handles this case.
+	if len(src)-int(startAt) < minNonLiteralBlockSize {
+		// We do not fill the token table.
+		// This will be picked up by caller.
+		dst.n = 0
+		return
+	}
+	// Index until startAt
+	if startAt > 0 {
+		cv := load3232(src, 0)
+		for i := int16(0); i < startAt; i++ {
+			table[hashSL(cv)] = tableEntry{offset: i}
+			cv = (cv >> 8) | (uint32(src[i+4]) << 24)
+		}
+	}
+
+	s := startAt + 1
+	nextEmit := startAt
+	// sLimit is when to stop looking for offset/length copies. The inputMargin
+	// lets us use a fast path for emitLiteral in the main loop, while we are
+	// looking for copies.
+	sLimit := int16(len(src) - inputMargin)
+
+	// nextEmit is where in src the next emitLiteral should start from.
+	cv := load3216(src, s)
+
+	for {
+		const skipLog = 5
+		const doEvery = 2
+
+		nextS := s
+		var candidate tableEntry
+		for {
+			nextHash := hashSL(cv)
+			candidate = table[nextHash]
+			nextS = s + doEvery + (s-nextEmit)>>skipLog
+			if nextS > sLimit || nextS <= 0 {
+				goto emitRemainder
+			}
+
+			now := load6416(src, nextS)
+			table[nextHash] = tableEntry{offset: s}
+			nextHash = hashSL(uint32(now))
+
+			if cv == load3216(src, candidate.offset) {
+				table[nextHash] = tableEntry{offset: nextS}
+				break
+			}
+
+			// Do one right away...
+			cv = uint32(now)
+			s = nextS
+			nextS++
+			candidate = table[nextHash]
+			now >>= 8
+			table[nextHash] = tableEntry{offset: s}
+
+			if cv == load3216(src, candidate.offset) {
+				table[nextHash] = tableEntry{offset: nextS}
+				break
+			}
+			cv = uint32(now)
+			s = nextS
+		}
+
+		// A 4-byte match has been found. We'll later see if more than 4 bytes
+		// match. But, prior to the match, src[nextEmit:s] are unmatched. Emit
+		// them as literal bytes.
+		for {
+			// Invariant: we have a 4-byte match at s, and no need to emit any
+			// literal bytes prior to s.
+
+			// Extend the 4-byte match as long as possible.
+			t := candidate.offset
+			l := int16(matchLen(src[s+4:], src[t+4:]) + 4)
+
+			// Extend backwards
+			for t > 0 && s > nextEmit && src[t-1] == src[s-1] {
+				s--
+				t--
+				l++
+			}
+			if nextEmit < s {
+				if false {
+					emitLiteral(dst, src[nextEmit:s])
+				} else {
+					for _, v := range src[nextEmit:s] {
+						dst.tokens[dst.n] = token(v)
+						dst.litHist[v]++
+						dst.n++
+					}
+				}
+			}
+
+			// Save the match found
+			dst.AddMatchLong(int32(l), uint32(s-t-baseMatchOffset))
+			s += l
+			nextEmit = s
+			if nextS >= s {
+				s = nextS + 1
+			}
+			if s >= sLimit {
+				goto emitRemainder
+			}
+
+			// We could immediately start working at s now, but to improve
+			// compression we first update the hash table at s-2 and at s. If
+			// another emitCopy is not our next move, also calculate nextHash
+			// at s+1. At least on GOARCH=amd64, these three hash calculations
+			// are faster as one load64 call (with some shifts) instead of
+			// three load32 calls.
+			x := load6416(src, s-2)
+			o := s - 2
+			prevHash := hashSL(uint32(x))
+			table[prevHash] = tableEntry{offset: o}
+			x >>= 16
+			currHash := hashSL(uint32(x))
+			candidate = table[currHash]
+			table[currHash] = tableEntry{offset: o + 2}
+
+			if uint32(x) != load3216(src, candidate.offset) {
+				cv = uint32(x >> 8)
+				s++
+				break
+			}
+		}
+	}
+
+emitRemainder:
+	if int(nextEmit) < len(src) {
+		// If nothing was added, don't encode literals.
+		if dst.n == 0 {
+			return
+		}
+		emitLiteral(dst, src[nextEmit:])
+	}
+}
diff --git a/vendor/github.com/klauspost/compress/flate/token.go b/vendor/github.com/klauspost/compress/flate/token.go
new file mode 100644
index 0000000..d818790
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/token.go
@@ -0,0 +1,379 @@
+// Copyright 2009 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package flate
+
+import (
+	"bytes"
+	"encoding/binary"
+	"fmt"
+	"io"
+	"math"
+)
+
+const (
+	// bits 0-16  	xoffset = offset - MIN_OFFSET_SIZE, or literal - 16 bits
+	// bits 16-22	offsetcode - 5 bits
+	// bits 22-30   xlength = length - MIN_MATCH_LENGTH - 8 bits
+	// bits 30-32   type   0 = literal  1=EOF  2=Match   3=Unused - 2 bits
+	lengthShift         = 22
+	offsetMask          = 1<<lengthShift - 1
+	typeMask            = 3 << 30
+	literalType         = 0 << 30
+	matchType           = 1 << 30
+	matchOffsetOnlyMask = 0xffff
+)
+
+// The length code for length X (MIN_MATCH_LENGTH <= X <= MAX_MATCH_LENGTH)
+// is lengthCodes[length - MIN_MATCH_LENGTH]
+var lengthCodes = [256]uint8{
+	0, 1, 2, 3, 4, 5, 6, 7, 8, 8,
+	9, 9, 10, 10, 11, 11, 12, 12, 12, 12,
+	13, 13, 13, 13, 14, 14, 14, 14, 15, 15,
+	15, 15, 16, 16, 16, 16, 16, 16, 16, 16,
+	17, 17, 17, 17, 17, 17, 17, 17, 18, 18,
+	18, 18, 18, 18, 18, 18, 19, 19, 19, 19,
+	19, 19, 19, 19, 20, 20, 20, 20, 20, 20,
+	20, 20, 20, 20, 20, 20, 20, 20, 20, 20,
+	21, 21, 21, 21, 21, 21, 21, 21, 21, 21,
+	21, 21, 21, 21, 21, 21, 22, 22, 22, 22,
+	22, 22, 22, 22, 22, 22, 22, 22, 22, 22,
+	22, 22, 23, 23, 23, 23, 23, 23, 23, 23,
+	23, 23, 23, 23, 23, 23, 23, 23, 24, 24,
+	24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
+	24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
+	24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
+	25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
+	25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
+	25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
+	25, 25, 26, 26, 26, 26, 26, 26, 26, 26,
+	26, 26, 26, 26, 26, 26, 26, 26, 26, 26,
+	26, 26, 26, 26, 26, 26, 26, 26, 26, 26,
+	26, 26, 26, 26, 27, 27, 27, 27, 27, 27,
+	27, 27, 27, 27, 27, 27, 27, 27, 27, 27,
+	27, 27, 27, 27, 27, 27, 27, 27, 27, 27,
+	27, 27, 27, 27, 27, 28,
+}
+
+// lengthCodes1 is length codes, but starting at 1.
+var lengthCodes1 = [256]uint8{
+	1, 2, 3, 4, 5, 6, 7, 8, 9, 9,
+	10, 10, 11, 11, 12, 12, 13, 13, 13, 13,
+	14, 14, 14, 14, 15, 15, 15, 15, 16, 16,
+	16, 16, 17, 17, 17, 17, 17, 17, 17, 17,
+	18, 18, 18, 18, 18, 18, 18, 18, 19, 19,
+	19, 19, 19, 19, 19, 19, 20, 20, 20, 20,
+	20, 20, 20, 20, 21, 21, 21, 21, 21, 21,
+	21, 21, 21, 21, 21, 21, 21, 21, 21, 21,
+	22, 22, 22, 22, 22, 22, 22, 22, 22, 22,
+	22, 22, 22, 22, 22, 22, 23, 23, 23, 23,
+	23, 23, 23, 23, 23, 23, 23, 23, 23, 23,
+	23, 23, 24, 24, 24, 24, 24, 24, 24, 24,
+	24, 24, 24, 24, 24, 24, 24, 24, 25, 25,
+	25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
+	25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
+	25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
+	26, 26, 26, 26, 26, 26, 26, 26, 26, 26,
+	26, 26, 26, 26, 26, 26, 26, 26, 26, 26,
+	26, 26, 26, 26, 26, 26, 26, 26, 26, 26,
+	26, 26, 27, 27, 27, 27, 27, 27, 27, 27,
+	27, 27, 27, 27, 27, 27, 27, 27, 27, 27,
+	27, 27, 27, 27, 27, 27, 27, 27, 27, 27,
+	27, 27, 27, 27, 28, 28, 28, 28, 28, 28,
+	28, 28, 28, 28, 28, 28, 28, 28, 28, 28,
+	28, 28, 28, 28, 28, 28, 28, 28, 28, 28,
+	28, 28, 28, 28, 28, 29,
+}
+
+var offsetCodes = [256]uint32{
+	0, 1, 2, 3, 4, 4, 5, 5, 6, 6, 6, 6, 7, 7, 7, 7,
+	8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9,
+	10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+	11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+	12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12,
+	12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12,
+	13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13,
+	13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13,
+	14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14,
+	14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14,
+	14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14,
+	14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14,
+	15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15,
+	15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15,
+	15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15,
+	15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15,
+}
+
+// offsetCodes14 are offsetCodes, but with 14 added.
+var offsetCodes14 = [256]uint32{
+	14, 15, 16, 17, 18, 18, 19, 19, 20, 20, 20, 20, 21, 21, 21, 21,
+	22, 22, 22, 22, 22, 22, 22, 22, 23, 23, 23, 23, 23, 23, 23, 23,
+	24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
+	25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
+	26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26,
+	26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26,
+	27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27,
+	27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27,
+	28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28,
+	28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28,
+	28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28,
+	28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28,
+	29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29,
+	29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29,
+	29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29,
+	29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29,
+}
+
+type token uint32
+
+type tokens struct {
+	extraHist [32]uint16  // codes 256->maxnumlit
+	offHist   [32]uint16  // offset codes
+	litHist   [256]uint16 // codes 0->255
+	nFilled   int
+	n         uint16 // Must be able to contain maxStoreBlockSize
+	tokens    [maxStoreBlockSize + 1]token
+}
+
+func (t *tokens) Reset() {
+	if t.n == 0 {
+		return
+	}
+	t.n = 0
+	t.nFilled = 0
+	for i := range t.litHist[:] {
+		t.litHist[i] = 0
+	}
+	for i := range t.extraHist[:] {
+		t.extraHist[i] = 0
+	}
+	for i := range t.offHist[:] {
+		t.offHist[i] = 0
+	}
+}
+
+func (t *tokens) Fill() {
+	if t.n == 0 {
+		return
+	}
+	for i, v := range t.litHist[:] {
+		if v == 0 {
+			t.litHist[i] = 1
+			t.nFilled++
+		}
+	}
+	for i, v := range t.extraHist[:literalCount-256] {
+		if v == 0 {
+			t.nFilled++
+			t.extraHist[i] = 1
+		}
+	}
+	for i, v := range t.offHist[:offsetCodeCount] {
+		if v == 0 {
+			t.offHist[i] = 1
+		}
+	}
+}
+
+func indexTokens(in []token) tokens {
+	var t tokens
+	t.indexTokens(in)
+	return t
+}
+
+func (t *tokens) indexTokens(in []token) {
+	t.Reset()
+	for _, tok := range in {
+		if tok < matchType {
+			t.AddLiteral(tok.literal())
+			continue
+		}
+		t.AddMatch(uint32(tok.length()), tok.offset()&matchOffsetOnlyMask)
+	}
+}
+
+// emitLiteral writes a literal chunk and returns the number of bytes written.
+func emitLiteral(dst *tokens, lit []byte) {
+	for _, v := range lit {
+		dst.tokens[dst.n] = token(v)
+		dst.litHist[v]++
+		dst.n++
+	}
+}
+
+func (t *tokens) AddLiteral(lit byte) {
+	t.tokens[t.n] = token(lit)
+	t.litHist[lit]++
+	t.n++
+}
+
+// from https://stackoverflow.com/a/28730362
+func mFastLog2(val float32) float32 {
+	ux := int32(math.Float32bits(val))
+	log2 := (float32)(((ux >> 23) & 255) - 128)
+	ux &= -0x7f800001
+	ux += 127 << 23
+	uval := math.Float32frombits(uint32(ux))
+	log2 += ((-0.34484843)*uval+2.02466578)*uval - 0.67487759
+	return log2
+}
+
+// EstimatedBits will return an minimum size estimated by an *optimal*
+// compression of the block.
+// The size of the block
+func (t *tokens) EstimatedBits() int {
+	shannon := float32(0)
+	bits := int(0)
+	nMatches := 0
+	total := int(t.n) + t.nFilled
+	if total > 0 {
+		invTotal := 1.0 / float32(total)
+		for _, v := range t.litHist[:] {
+			if v > 0 {
+				n := float32(v)
+				shannon += atLeastOne(-mFastLog2(n*invTotal)) * n
+			}
+		}
+		// Just add 15 for EOB
+		shannon += 15
+		for i, v := range t.extraHist[1 : literalCount-256] {
+			if v > 0 {
+				n := float32(v)
+				shannon += atLeastOne(-mFastLog2(n*invTotal)) * n
+				bits += int(lengthExtraBits[i&31]) * int(v)
+				nMatches += int(v)
+			}
+		}
+	}
+	if nMatches > 0 {
+		invTotal := 1.0 / float32(nMatches)
+		for i, v := range t.offHist[:offsetCodeCount] {
+			if v > 0 {
+				n := float32(v)
+				shannon += atLeastOne(-mFastLog2(n*invTotal)) * n
+				bits += int(offsetExtraBits[i&31]) * int(v)
+			}
+		}
+	}
+	return int(shannon) + bits
+}
+
+// AddMatch adds a match to the tokens.
+// This function is very sensitive to inlining and right on the border.
+func (t *tokens) AddMatch(xlength uint32, xoffset uint32) {
+	if debugDeflate {
+		if xlength >= maxMatchLength+baseMatchLength {
+			panic(fmt.Errorf("invalid length: %v", xlength))
+		}
+		if xoffset >= maxMatchOffset+baseMatchOffset {
+			panic(fmt.Errorf("invalid offset: %v", xoffset))
+		}
+	}
+	oCode := offsetCode(xoffset)
+	xoffset |= oCode << 16
+
+	t.extraHist[lengthCodes1[uint8(xlength)]]++
+	t.offHist[oCode&31]++
+	t.tokens[t.n] = token(matchType | xlength<<lengthShift | xoffset)
+	t.n++
+}
+
+// AddMatchLong adds a match to the tokens, potentially longer than max match length.
+// Length should NOT have the base subtracted, only offset should.
+func (t *tokens) AddMatchLong(xlength int32, xoffset uint32) {
+	if debugDeflate {
+		if xoffset >= maxMatchOffset+baseMatchOffset {
+			panic(fmt.Errorf("invalid offset: %v", xoffset))
+		}
+	}
+	oc := offsetCode(xoffset)
+	xoffset |= oc << 16
+	for xlength > 0 {
+		xl := xlength
+		if xl > 258 {
+			// We need to have at least baseMatchLength left over for next loop.
+			if xl > 258+baseMatchLength {
+				xl = 258
+			} else {
+				xl = 258 - baseMatchLength
+			}
+		}
+		xlength -= xl
+		xl -= baseMatchLength
+		t.extraHist[lengthCodes1[uint8(xl)]]++
+		t.offHist[oc&31]++
+		t.tokens[t.n] = token(matchType | uint32(xl)<<lengthShift | xoffset)
+		t.n++
+	}
+}
+
+func (t *tokens) AddEOB() {
+	t.tokens[t.n] = token(endBlockMarker)
+	t.extraHist[0]++
+	t.n++
+}
+
+func (t *tokens) Slice() []token {
+	return t.tokens[:t.n]
+}
+
+// VarInt returns the tokens as varint encoded bytes.
+func (t *tokens) VarInt() []byte {
+	var b = make([]byte, binary.MaxVarintLen32*int(t.n))
+	var off int
+	for _, v := range t.tokens[:t.n] {
+		off += binary.PutUvarint(b[off:], uint64(v))
+	}
+	return b[:off]
+}
+
+// FromVarInt restores t to the varint encoded tokens provided.
+// Any data in t is removed.
+func (t *tokens) FromVarInt(b []byte) error {
+	var buf = bytes.NewReader(b)
+	var toks []token
+	for {
+		r, err := binary.ReadUvarint(buf)
+		if err == io.EOF {
+			break
+		}
+		if err != nil {
+			return err
+		}
+		toks = append(toks, token(r))
+	}
+	t.indexTokens(toks)
+	return nil
+}
+
+// Returns the type of a token
+func (t token) typ() uint32 { return uint32(t) & typeMask }
+
+// Returns the literal of a literal token
+func (t token) literal() uint8 { return uint8(t) }
+
+// Returns the extra offset of a match token
+func (t token) offset() uint32 { return uint32(t) & offsetMask }
+
+func (t token) length() uint8 { return uint8(t >> lengthShift) }
+
+// Convert length to code.
+func lengthCode(len uint8) uint8 { return lengthCodes[len] }
+
+// Returns the offset code corresponding to a specific offset
+func offsetCode(off uint32) uint32 {
+	if false {
+		if off < uint32(len(offsetCodes)) {
+			return offsetCodes[off&255]
+		} else if off>>7 < uint32(len(offsetCodes)) {
+			return offsetCodes[(off>>7)&255] + 14
+		} else {
+			return offsetCodes[(off>>14)&255] + 28
+		}
+	}
+	if off < uint32(len(offsetCodes)) {
+		return offsetCodes[uint8(off)]
+	}
+	return offsetCodes14[uint8(off>>7)]
+}
diff --git a/vendor/github.com/klauspost/compress/fse/bitwriter.go b/vendor/github.com/klauspost/compress/fse/bitwriter.go
index 43e4636..e82fa3b 100644
--- a/vendor/github.com/klauspost/compress/fse/bitwriter.go
+++ b/vendor/github.com/klauspost/compress/fse/bitwriter.go
@@ -152,12 +152,11 @@
 
 // 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/fse/compress.go b/vendor/github.com/klauspost/compress/fse/compress.go
index 6f34191..074018d 100644
--- a/vendor/github.com/klauspost/compress/fse/compress.go
+++ b/vendor/github.com/klauspost/compress/fse/compress.go
@@ -146,54 +146,51 @@
 		c1.encodeZero(tt[src[ip-2]])
 		ip -= 2
 	}
+	src = src[:ip]
 
 	// Main compression loop.
 	switch {
 	case !s.zeroBits && s.actualTableLog <= 8:
 		// We can encode 4 symbols without requiring a flush.
 		// We do not need to check if any output is 0 bits.
-		for ip >= 4 {
+		for ; len(src) >= 4; src = src[:len(src)-4] {
 			s.bw.flush32()
-			v3, v2, v1, v0 := src[ip-4], src[ip-3], src[ip-2], src[ip-1]
+			v3, v2, v1, v0 := src[len(src)-4], src[len(src)-3], src[len(src)-2], src[len(src)-1]
 			c2.encode(tt[v0])
 			c1.encode(tt[v1])
 			c2.encode(tt[v2])
 			c1.encode(tt[v3])
-			ip -= 4
 		}
 	case !s.zeroBits:
 		// We do not need to check if any output is 0 bits.
-		for ip >= 4 {
+		for ; len(src) >= 4; src = src[:len(src)-4] {
 			s.bw.flush32()
-			v3, v2, v1, v0 := src[ip-4], src[ip-3], src[ip-2], src[ip-1]
+			v3, v2, v1, v0 := src[len(src)-4], src[len(src)-3], src[len(src)-2], src[len(src)-1]
 			c2.encode(tt[v0])
 			c1.encode(tt[v1])
 			s.bw.flush32()
 			c2.encode(tt[v2])
 			c1.encode(tt[v3])
-			ip -= 4
 		}
 	case s.actualTableLog <= 8:
 		// We can encode 4 symbols without requiring a flush
-		for ip >= 4 {
+		for ; len(src) >= 4; src = src[:len(src)-4] {
 			s.bw.flush32()
-			v3, v2, v1, v0 := src[ip-4], src[ip-3], src[ip-2], src[ip-1]
+			v3, v2, v1, v0 := src[len(src)-4], src[len(src)-3], src[len(src)-2], src[len(src)-1]
 			c2.encodeZero(tt[v0])
 			c1.encodeZero(tt[v1])
 			c2.encodeZero(tt[v2])
 			c1.encodeZero(tt[v3])
-			ip -= 4
 		}
 	default:
-		for ip >= 4 {
+		for ; len(src) >= 4; src = src[:len(src)-4] {
 			s.bw.flush32()
-			v3, v2, v1, v0 := src[ip-4], src[ip-3], src[ip-2], src[ip-1]
+			v3, v2, v1, v0 := src[len(src)-4], src[len(src)-3], src[len(src)-2], src[len(src)-1]
 			c2.encodeZero(tt[v0])
 			c1.encodeZero(tt[v1])
 			s.bw.flush32()
 			c2.encodeZero(tt[v2])
 			c1.encodeZero(tt[v3])
-			ip -= 4
 		}
 	}
 
@@ -202,7 +199,8 @@
 	c2.flush(s.actualTableLog)
 	c1.flush(s.actualTableLog)
 
-	return s.bw.close()
+	s.bw.close()
+	return nil
 }
 
 // writeCount will write the normalized histogram count to header.
@@ -214,7 +212,7 @@
 		previous0 bool
 		charnum   uint16
 
-		maxHeaderSize = ((int(s.symbolLen) * int(tableLog)) >> 3) + 3
+		maxHeaderSize = ((int(s.symbolLen)*int(tableLog) + 4 + 2) >> 3) + 3
 
 		// Write Table Size
 		bitStream = uint32(tableLog - minTablelog)
@@ -459,15 +457,17 @@
 	for _, v := range in {
 		s.count[v]++
 	}
-	m := uint32(0)
+	m, symlen := uint32(0), s.symbolLen
 	for i, v := range s.count[:] {
+		if v == 0 {
+			continue
+		}
 		if v > m {
 			m = v
 		}
-		if v > 0 {
-			s.symbolLen = uint16(i) + 1
-		}
+		symlen = uint16(i) + 1
 	}
+	s.symbolLen = symlen
 	return int(m)
 }
 
diff --git a/vendor/github.com/klauspost/compress/fse/decompress.go b/vendor/github.com/klauspost/compress/fse/decompress.go
index 926f5f1..0c7dd4f 100644
--- a/vendor/github.com/klauspost/compress/fse/decompress.go
+++ b/vendor/github.com/klauspost/compress/fse/decompress.go
@@ -15,7 +15,7 @@
 // It is possible, but by no way guaranteed that corrupt data will
 // return an error.
 // It is up to the caller to verify integrity of the returned data.
-// Use a predefined Scrach to set maximum acceptable output size.
+// Use a predefined Scratch to set maximum acceptable output size.
 func Decompress(b []byte, s *Scratch) ([]byte, error) {
 	s, err := s.prepare(b)
 	if err != nil {
@@ -260,7 +260,9 @@
 // If the buffer is over-read an error is returned.
 func (s *Scratch) decompress() error {
 	br := &s.bits
-	br.init(s.br.unread())
+	if err := br.init(s.br.unread()); err != nil {
+		return err
+	}
 
 	var s1, s2 decoder
 	// Initialize and decode first state and symbol.
diff --git a/vendor/github.com/klauspost/compress/gen.sh b/vendor/github.com/klauspost/compress/gen.sh
new file mode 100644
index 0000000..aff9422
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/gen.sh
@@ -0,0 +1,4 @@
+#!/bin/sh
+
+cd s2/cmd/_s2sx/ || exit 1
+go generate .
diff --git a/vendor/github.com/klauspost/compress/gzip/gunzip.go b/vendor/github.com/klauspost/compress/gzip/gunzip.go
new file mode 100644
index 0000000..00a0a2c
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/gzip/gunzip.go
@@ -0,0 +1,380 @@
+// Copyright 2009 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package gzip implements reading and writing of gzip format compressed files,
+// as specified in RFC 1952.
+package gzip
+
+import (
+	"bufio"
+	"compress/gzip"
+	"encoding/binary"
+	"hash/crc32"
+	"io"
+	"time"
+
+	"github.com/klauspost/compress/flate"
+)
+
+const (
+	gzipID1     = 0x1f
+	gzipID2     = 0x8b
+	gzipDeflate = 8
+	flagText    = 1 << 0
+	flagHdrCrc  = 1 << 1
+	flagExtra   = 1 << 2
+	flagName    = 1 << 3
+	flagComment = 1 << 4
+)
+
+var (
+	// ErrChecksum is returned when reading GZIP data that has an invalid checksum.
+	ErrChecksum = gzip.ErrChecksum
+	// ErrHeader is returned when reading GZIP data that has an invalid header.
+	ErrHeader = gzip.ErrHeader
+)
+
+var le = binary.LittleEndian
+
+// noEOF converts io.EOF to io.ErrUnexpectedEOF.
+func noEOF(err error) error {
+	if err == io.EOF {
+		return io.ErrUnexpectedEOF
+	}
+	return err
+}
+
+// The gzip file stores a header giving metadata about the compressed file.
+// That header is exposed as the fields of the Writer and Reader structs.
+//
+// Strings must be UTF-8 encoded and may only contain Unicode code points
+// U+0001 through U+00FF, due to limitations of the GZIP file format.
+type Header struct {
+	Comment string    // comment
+	Extra   []byte    // "extra data"
+	ModTime time.Time // modification time
+	Name    string    // file name
+	OS      byte      // operating system type
+}
+
+// A Reader is an io.Reader that can be read to retrieve
+// uncompressed data from a gzip-format compressed file.
+//
+// In general, a gzip file can be a concatenation of gzip files,
+// each with its own header. Reads from the Reader
+// return the concatenation of the uncompressed data of each.
+// Only the first header is recorded in the Reader fields.
+//
+// Gzip files store a length and checksum of the uncompressed data.
+// The Reader will return a ErrChecksum when Read
+// reaches the end of the uncompressed data if it does not
+// have the expected length or checksum. Clients should treat data
+// returned by Read as tentative until they receive the io.EOF
+// marking the end of the data.
+type Reader struct {
+	Header       // valid after NewReader or Reader.Reset
+	r            flate.Reader
+	br           *bufio.Reader
+	decompressor io.ReadCloser
+	digest       uint32 // CRC-32, IEEE polynomial (section 8)
+	size         uint32 // Uncompressed size (section 2.3.1)
+	buf          [512]byte
+	err          error
+	multistream  bool
+}
+
+// NewReader creates a new Reader reading the given reader.
+// If r does not also implement io.ByteReader,
+// the decompressor may read more data than necessary from r.
+//
+// It is the caller's responsibility to call Close on the Reader when done.
+//
+// The Reader.Header fields will be valid in the Reader returned.
+func NewReader(r io.Reader) (*Reader, error) {
+	z := new(Reader)
+	if err := z.Reset(r); err != nil {
+		return nil, err
+	}
+	return z, nil
+}
+
+// Reset discards the Reader z's state and makes it equivalent to the
+// result of its original state from NewReader, but reading from r instead.
+// This permits reusing a Reader rather than allocating a new one.
+func (z *Reader) Reset(r io.Reader) error {
+	*z = Reader{
+		decompressor: z.decompressor,
+		multistream:  true,
+		br:           z.br,
+	}
+	if rr, ok := r.(flate.Reader); ok {
+		z.r = rr
+	} else {
+		// Reuse if we can.
+		if z.br != nil {
+			z.br.Reset(r)
+		} else {
+			z.br = bufio.NewReader(r)
+		}
+		z.r = z.br
+	}
+	z.Header, z.err = z.readHeader()
+	return z.err
+}
+
+// Multistream controls whether the reader supports multistream files.
+//
+// If enabled (the default), the Reader expects the input to be a sequence
+// of individually gzipped data streams, each with its own header and
+// trailer, ending at EOF. The effect is that the concatenation of a sequence
+// of gzipped files is treated as equivalent to the gzip of the concatenation
+// of the sequence. This is standard behavior for gzip readers.
+//
+// Calling Multistream(false) disables this behavior; disabling the behavior
+// can be useful when reading file formats that distinguish individual gzip
+// data streams or mix gzip data streams with other data streams.
+// In this mode, when the Reader reaches the end of the data stream,
+// Read returns io.EOF. If the underlying reader implements io.ByteReader,
+// it will be left positioned just after the gzip stream.
+// To start the next stream, call z.Reset(r) followed by z.Multistream(false).
+// If there is no next stream, z.Reset(r) will return io.EOF.
+func (z *Reader) Multistream(ok bool) {
+	z.multistream = ok
+}
+
+// readString reads a NUL-terminated string from z.r.
+// It treats the bytes read as being encoded as ISO 8859-1 (Latin-1) and
+// will output a string encoded using UTF-8.
+// This method always updates z.digest with the data read.
+func (z *Reader) readString() (string, error) {
+	var err error
+	needConv := false
+	for i := 0; ; i++ {
+		if i >= len(z.buf) {
+			return "", ErrHeader
+		}
+		z.buf[i], err = z.r.ReadByte()
+		if err != nil {
+			return "", err
+		}
+		if z.buf[i] > 0x7f {
+			needConv = true
+		}
+		if z.buf[i] == 0 {
+			// Digest covers the NUL terminator.
+			z.digest = crc32.Update(z.digest, crc32.IEEETable, z.buf[:i+1])
+
+			// Strings are ISO 8859-1, Latin-1 (RFC 1952, section 2.3.1).
+			if needConv {
+				s := make([]rune, 0, i)
+				for _, v := range z.buf[:i] {
+					s = append(s, rune(v))
+				}
+				return string(s), nil
+			}
+			return string(z.buf[:i]), nil
+		}
+	}
+}
+
+// readHeader reads the GZIP header according to section 2.3.1.
+// This method does not set z.err.
+func (z *Reader) readHeader() (hdr Header, err error) {
+	if _, err = io.ReadFull(z.r, z.buf[:10]); err != nil {
+		// RFC 1952, section 2.2, says the following:
+		//	A gzip file consists of a series of "members" (compressed data sets).
+		//
+		// Other than this, the specification does not clarify whether a
+		// "series" is defined as "one or more" or "zero or more". To err on the
+		// side of caution, Go interprets this to mean "zero or more".
+		// Thus, it is okay to return io.EOF here.
+		return hdr, err
+	}
+	if z.buf[0] != gzipID1 || z.buf[1] != gzipID2 || z.buf[2] != gzipDeflate {
+		return hdr, ErrHeader
+	}
+	flg := z.buf[3]
+	hdr.ModTime = time.Unix(int64(le.Uint32(z.buf[4:8])), 0)
+	// z.buf[8] is XFL and is currently ignored.
+	hdr.OS = z.buf[9]
+	z.digest = crc32.ChecksumIEEE(z.buf[:10])
+
+	if flg&flagExtra != 0 {
+		if _, err = io.ReadFull(z.r, z.buf[:2]); err != nil {
+			return hdr, noEOF(err)
+		}
+		z.digest = crc32.Update(z.digest, crc32.IEEETable, z.buf[:2])
+		data := make([]byte, le.Uint16(z.buf[:2]))
+		if _, err = io.ReadFull(z.r, data); err != nil {
+			return hdr, noEOF(err)
+		}
+		z.digest = crc32.Update(z.digest, crc32.IEEETable, data)
+		hdr.Extra = data
+	}
+
+	var s string
+	if flg&flagName != 0 {
+		if s, err = z.readString(); err != nil {
+			return hdr, err
+		}
+		hdr.Name = s
+	}
+
+	if flg&flagComment != 0 {
+		if s, err = z.readString(); err != nil {
+			return hdr, err
+		}
+		hdr.Comment = s
+	}
+
+	if flg&flagHdrCrc != 0 {
+		if _, err = io.ReadFull(z.r, z.buf[:2]); err != nil {
+			return hdr, noEOF(err)
+		}
+		digest := le.Uint16(z.buf[:2])
+		if digest != uint16(z.digest) {
+			return hdr, ErrHeader
+		}
+	}
+
+	// Reserved FLG bits must be zero.
+	if flg>>5 != 0 {
+		return hdr, ErrHeader
+	}
+
+	z.digest = 0
+	if z.decompressor == nil {
+		z.decompressor = flate.NewReader(z.r)
+	} else {
+		z.decompressor.(flate.Resetter).Reset(z.r, nil)
+	}
+	return hdr, nil
+}
+
+// Read implements io.Reader, reading uncompressed bytes from its underlying Reader.
+func (z *Reader) Read(p []byte) (n int, err error) {
+	if z.err != nil {
+		return 0, z.err
+	}
+
+	for n == 0 {
+		n, z.err = z.decompressor.Read(p)
+		z.digest = crc32.Update(z.digest, crc32.IEEETable, p[:n])
+		z.size += uint32(n)
+		if z.err != io.EOF {
+			// In the normal case we return here.
+			return n, z.err
+		}
+
+		// Finished file; check checksum and size.
+		if _, err := io.ReadFull(z.r, z.buf[:8]); err != nil {
+			z.err = noEOF(err)
+			return n, z.err
+		}
+		digest := le.Uint32(z.buf[:4])
+		size := le.Uint32(z.buf[4:8])
+		if digest != z.digest || size != z.size {
+			z.err = ErrChecksum
+			return n, z.err
+		}
+		z.digest, z.size = 0, 0
+
+		// File is ok; check if there is another.
+		if !z.multistream {
+			return n, io.EOF
+		}
+		z.err = nil // Remove io.EOF
+
+		if _, z.err = z.readHeader(); z.err != nil {
+			return n, z.err
+		}
+	}
+
+	return n, nil
+}
+
+type crcer interface {
+	io.Writer
+	Sum32() uint32
+	Reset()
+}
+type crcUpdater struct {
+	z *Reader
+}
+
+func (c *crcUpdater) Write(p []byte) (int, error) {
+	c.z.digest = crc32.Update(c.z.digest, crc32.IEEETable, p)
+	return len(p), nil
+}
+
+func (c *crcUpdater) Sum32() uint32 {
+	return c.z.digest
+}
+
+func (c *crcUpdater) Reset() {
+	c.z.digest = 0
+}
+
+// WriteTo support the io.WriteTo interface for io.Copy and friends.
+func (z *Reader) WriteTo(w io.Writer) (int64, error) {
+	total := int64(0)
+	crcWriter := crcer(crc32.NewIEEE())
+	if z.digest != 0 {
+		crcWriter = &crcUpdater{z: z}
+	}
+	for {
+		if z.err != nil {
+			if z.err == io.EOF {
+				return total, nil
+			}
+			return total, z.err
+		}
+
+		// We write both to output and digest.
+		mw := io.MultiWriter(w, crcWriter)
+		n, err := z.decompressor.(io.WriterTo).WriteTo(mw)
+		total += n
+		z.size += uint32(n)
+		if err != nil {
+			z.err = err
+			return total, z.err
+		}
+
+		// Finished file; check checksum + size.
+		if _, err := io.ReadFull(z.r, z.buf[0:8]); err != nil {
+			if err == io.EOF {
+				err = io.ErrUnexpectedEOF
+			}
+			z.err = err
+			return total, err
+		}
+		z.digest = crcWriter.Sum32()
+		digest := le.Uint32(z.buf[:4])
+		size := le.Uint32(z.buf[4:8])
+		if digest != z.digest || size != z.size {
+			z.err = ErrChecksum
+			return total, z.err
+		}
+		z.digest, z.size = 0, 0
+
+		// File is ok; check if there is another.
+		if !z.multistream {
+			return total, nil
+		}
+		crcWriter.Reset()
+		z.err = nil // Remove io.EOF
+
+		if _, z.err = z.readHeader(); z.err != nil {
+			if z.err == io.EOF {
+				return total, nil
+			}
+			return total, z.err
+		}
+	}
+}
+
+// Close closes the Reader. It does not close the underlying io.Reader.
+// In order for the GZIP checksum to be verified, the reader must be
+// fully consumed until the io.EOF.
+func (z *Reader) Close() error { return z.decompressor.Close() }
diff --git a/vendor/github.com/klauspost/compress/gzip/gzip.go b/vendor/github.com/klauspost/compress/gzip/gzip.go
new file mode 100644
index 0000000..5bc7205
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/gzip/gzip.go
@@ -0,0 +1,290 @@
+// Copyright 2010 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package gzip
+
+import (
+	"errors"
+	"fmt"
+	"hash/crc32"
+	"io"
+
+	"github.com/klauspost/compress/flate"
+)
+
+// These constants are copied from the flate package, so that code that imports
+// "compress/gzip" does not also have to import "compress/flate".
+const (
+	NoCompression       = flate.NoCompression
+	BestSpeed           = flate.BestSpeed
+	BestCompression     = flate.BestCompression
+	DefaultCompression  = flate.DefaultCompression
+	ConstantCompression = flate.ConstantCompression
+	HuffmanOnly         = flate.HuffmanOnly
+
+	// StatelessCompression will do compression but without maintaining any state
+	// between Write calls.
+	// There will be no memory kept between Write calls,
+	// but compression and speed will be suboptimal.
+	// Because of this, the size of actual Write calls will affect output size.
+	StatelessCompression = -3
+)
+
+// A Writer is an io.WriteCloser.
+// Writes to a Writer are compressed and written to w.
+type Writer struct {
+	Header      // written at first call to Write, Flush, or Close
+	w           io.Writer
+	level       int
+	err         error
+	compressor  *flate.Writer
+	digest      uint32 // CRC-32, IEEE polynomial (section 8)
+	size        uint32 // Uncompressed size (section 2.3.1)
+	wroteHeader bool
+	closed      bool
+	buf         [10]byte
+}
+
+// NewWriter returns a new Writer.
+// Writes to the returned writer are compressed and written to w.
+//
+// It is the caller's responsibility to call Close on the WriteCloser when done.
+// Writes may be buffered and not flushed until Close.
+//
+// Callers that wish to set the fields in Writer.Header must do so before
+// the first call to Write, Flush, or Close.
+func NewWriter(w io.Writer) *Writer {
+	z, _ := NewWriterLevel(w, DefaultCompression)
+	return z
+}
+
+// NewWriterLevel is like NewWriter but specifies the compression level instead
+// of assuming DefaultCompression.
+//
+// The compression level can be DefaultCompression, NoCompression, or any
+// integer value between BestSpeed and BestCompression inclusive. The error
+// returned will be nil if the level is valid.
+func NewWriterLevel(w io.Writer, level int) (*Writer, error) {
+	if level < StatelessCompression || level > BestCompression {
+		return nil, fmt.Errorf("gzip: invalid compression level: %d", level)
+	}
+	z := new(Writer)
+	z.init(w, level)
+	return z, nil
+}
+
+// MinCustomWindowSize is the minimum window size that can be sent to NewWriterWindow.
+const MinCustomWindowSize = flate.MinCustomWindowSize
+
+// MaxCustomWindowSize is the maximum custom window that can be sent to NewWriterWindow.
+const MaxCustomWindowSize = flate.MaxCustomWindowSize
+
+// NewWriterWindow returns a new Writer compressing data with a custom window size.
+// windowSize must be from MinCustomWindowSize to MaxCustomWindowSize.
+func NewWriterWindow(w io.Writer, windowSize int) (*Writer, error) {
+	if windowSize < MinCustomWindowSize {
+		return nil, errors.New("gzip: requested window size less than MinWindowSize")
+	}
+	if windowSize > MaxCustomWindowSize {
+		return nil, errors.New("gzip: requested window size bigger than MaxCustomWindowSize")
+	}
+
+	z := new(Writer)
+	z.init(w, -windowSize)
+	return z, nil
+}
+
+func (z *Writer) init(w io.Writer, level int) {
+	compressor := z.compressor
+	if level != StatelessCompression {
+		if compressor != nil {
+			compressor.Reset(w)
+		}
+	}
+
+	*z = Writer{
+		Header: Header{
+			OS: 255, // unknown
+		},
+		w:          w,
+		level:      level,
+		compressor: compressor,
+	}
+}
+
+// Reset discards the Writer z's state and makes it equivalent to the
+// result of its original state from NewWriter or NewWriterLevel, but
+// writing to w instead. This permits reusing a Writer rather than
+// allocating a new one.
+func (z *Writer) Reset(w io.Writer) {
+	z.init(w, z.level)
+}
+
+// writeBytes writes a length-prefixed byte slice to z.w.
+func (z *Writer) writeBytes(b []byte) error {
+	if len(b) > 0xffff {
+		return errors.New("gzip.Write: Extra data is too large")
+	}
+	le.PutUint16(z.buf[:2], uint16(len(b)))
+	_, err := z.w.Write(z.buf[:2])
+	if err != nil {
+		return err
+	}
+	_, err = z.w.Write(b)
+	return err
+}
+
+// writeString writes a UTF-8 string s in GZIP's format to z.w.
+// GZIP (RFC 1952) specifies that strings are NUL-terminated ISO 8859-1 (Latin-1).
+func (z *Writer) writeString(s string) (err error) {
+	// GZIP stores Latin-1 strings; error if non-Latin-1; convert if non-ASCII.
+	needconv := false
+	for _, v := range s {
+		if v == 0 || v > 0xff {
+			return errors.New("gzip.Write: non-Latin-1 header string")
+		}
+		if v > 0x7f {
+			needconv = true
+		}
+	}
+	if needconv {
+		b := make([]byte, 0, len(s))
+		for _, v := range s {
+			b = append(b, byte(v))
+		}
+		_, err = z.w.Write(b)
+	} else {
+		_, err = io.WriteString(z.w, s)
+	}
+	if err != nil {
+		return err
+	}
+	// GZIP strings are NUL-terminated.
+	z.buf[0] = 0
+	_, err = z.w.Write(z.buf[:1])
+	return err
+}
+
+// Write writes a compressed form of p to the underlying io.Writer. The
+// compressed bytes are not necessarily flushed until the Writer is closed.
+func (z *Writer) Write(p []byte) (int, error) {
+	if z.err != nil {
+		return 0, z.err
+	}
+	var n int
+	// Write the GZIP header lazily.
+	if !z.wroteHeader {
+		z.wroteHeader = true
+		z.buf[0] = gzipID1
+		z.buf[1] = gzipID2
+		z.buf[2] = gzipDeflate
+		z.buf[3] = 0
+		if z.Extra != nil {
+			z.buf[3] |= 0x04
+		}
+		if z.Name != "" {
+			z.buf[3] |= 0x08
+		}
+		if z.Comment != "" {
+			z.buf[3] |= 0x10
+		}
+		le.PutUint32(z.buf[4:8], uint32(z.ModTime.Unix()))
+		if z.level == BestCompression {
+			z.buf[8] = 2
+		} else if z.level == BestSpeed {
+			z.buf[8] = 4
+		} else {
+			z.buf[8] = 0
+		}
+		z.buf[9] = z.OS
+		n, z.err = z.w.Write(z.buf[:10])
+		if z.err != nil {
+			return n, z.err
+		}
+		if z.Extra != nil {
+			z.err = z.writeBytes(z.Extra)
+			if z.err != nil {
+				return n, z.err
+			}
+		}
+		if z.Name != "" {
+			z.err = z.writeString(z.Name)
+			if z.err != nil {
+				return n, z.err
+			}
+		}
+		if z.Comment != "" {
+			z.err = z.writeString(z.Comment)
+			if z.err != nil {
+				return n, z.err
+			}
+		}
+
+		if z.compressor == nil && z.level != StatelessCompression {
+			z.compressor, _ = flate.NewWriter(z.w, z.level)
+		}
+	}
+	z.size += uint32(len(p))
+	z.digest = crc32.Update(z.digest, crc32.IEEETable, p)
+	if z.level == StatelessCompression {
+		return len(p), flate.StatelessDeflate(z.w, p, false, nil)
+	}
+	n, z.err = z.compressor.Write(p)
+	return n, z.err
+}
+
+// Flush flushes any pending compressed data to the underlying writer.
+//
+// It is useful mainly in compressed network protocols, to ensure that
+// a remote reader has enough data to reconstruct a packet. Flush does
+// not return until the data has been written. If the underlying
+// writer returns an error, Flush returns that error.
+//
+// In the terminology of the zlib library, Flush is equivalent to Z_SYNC_FLUSH.
+func (z *Writer) Flush() error {
+	if z.err != nil {
+		return z.err
+	}
+	if z.closed || z.level == StatelessCompression {
+		return nil
+	}
+	if !z.wroteHeader {
+		z.Write(nil)
+		if z.err != nil {
+			return z.err
+		}
+	}
+	z.err = z.compressor.Flush()
+	return z.err
+}
+
+// Close closes the Writer, flushing any unwritten data to the underlying
+// io.Writer, but does not close the underlying io.Writer.
+func (z *Writer) Close() error {
+	if z.err != nil {
+		return z.err
+	}
+	if z.closed {
+		return nil
+	}
+	z.closed = true
+	if !z.wroteHeader {
+		z.Write(nil)
+		if z.err != nil {
+			return z.err
+		}
+	}
+	if z.level == StatelessCompression {
+		z.err = flate.StatelessDeflate(z.w, nil, true, nil)
+	} else {
+		z.err = z.compressor.Close()
+	}
+	if z.err != nil {
+		return z.err
+	}
+	le.PutUint32(z.buf[:4], z.digest)
+	le.PutUint32(z.buf[4:8], z.size)
+	_, z.err = z.w.Write(z.buf[:8])
+	return z.err
+}
diff --git a/vendor/github.com/klauspost/compress/huff0/bitreader.go b/vendor/github.com/klauspost/compress/huff0/bitreader.go
index a4979e8..bfc7a52 100644
--- a/vendor/github.com/klauspost/compress/huff0/bitreader.go
+++ b/vendor/github.com/klauspost/compress/huff0/bitreader.go
@@ -6,120 +6,16 @@
 package huff0
 
 import (
-	"encoding/binary"
 	"errors"
+	"fmt"
 	"io"
+
+	"github.com/klauspost/compress/internal/le"
 )
 
 // bitReader reads a bitstream in reverse.
 // The last set bit indicates the start of the stream and is used
 // for aligning the input.
-type bitReader struct {
-	in       []byte
-	off      uint // next byte to read is at in[off - 1]
-	value    uint64
-	bitsRead uint8
-}
-
-// init initializes and resets the bit reader.
-func (b *bitReader) init(in []byte) error {
-	if len(in) < 1 {
-		return errors.New("corrupt stream: too short")
-	}
-	b.in = in
-	b.off = uint(len(in))
-	// The highest bit of the last byte indicates where to start
-	v := in[len(in)-1]
-	if v == 0 {
-		return errors.New("corrupt stream, did not find end of stream")
-	}
-	b.bitsRead = 64
-	b.value = 0
-	if len(in) >= 8 {
-		b.fillFastStart()
-	} else {
-		b.fill()
-		b.fill()
-	}
-	b.bitsRead += 8 - uint8(highBit32(uint32(v)))
-	return nil
-}
-
-// peekBitsFast requires that at least one bit is requested every time.
-// There are no checks if the buffer is filled.
-func (b *bitReader) peekBitsFast(n uint8) uint16 {
-	const regMask = 64 - 1
-	v := uint16((b.value << (b.bitsRead & regMask)) >> ((regMask + 1 - n) & regMask))
-	return v
-}
-
-// fillFast() will make sure at least 32 bits are available.
-// There must be at least 4 bytes available.
-func (b *bitReader) fillFast() {
-	if b.bitsRead < 32 {
-		return
-	}
-
-	// 2 bounds checks.
-	v := b.in[b.off-4 : b.off]
-	v = v[:4]
-	low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
-	b.value = (b.value << 32) | uint64(low)
-	b.bitsRead -= 32
-	b.off -= 4
-}
-
-func (b *bitReader) advance(n uint8) {
-	b.bitsRead += n
-}
-
-// fillFastStart() assumes the bitreader is empty and there is at least 8 bytes to read.
-func (b *bitReader) fillFastStart() {
-	// Do single re-slice to avoid bounds checks.
-	b.value = binary.LittleEndian.Uint64(b.in[b.off-8:])
-	b.bitsRead = 0
-	b.off -= 8
-}
-
-// fill() will make sure at least 32 bits are available.
-func (b *bitReader) fill() {
-	if b.bitsRead < 32 {
-		return
-	}
-	if b.off > 4 {
-		v := b.in[b.off-4:]
-		v = v[:4]
-		low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
-		b.value = (b.value << 32) | uint64(low)
-		b.bitsRead -= 32
-		b.off -= 4
-		return
-	}
-	for b.off > 0 {
-		b.value = (b.value << 8) | uint64(b.in[b.off-1])
-		b.bitsRead -= 8
-		b.off--
-	}
-}
-
-// 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
-}
-
-// close the bitstream and returns an error if out-of-buffer reads occurred.
-func (b *bitReader) close() error {
-	// Release reference.
-	b.in = nil
-	if b.bitsRead > 64 {
-		return io.ErrUnexpectedEOF
-	}
-	return nil
-}
-
-// bitReader reads a bitstream in reverse.
-// The last set bit indicates the start of the stream and is used
-// for aligning the input.
 type bitReaderBytes struct {
 	in       []byte
 	off      uint // next byte to read is at in[off - 1]
@@ -151,7 +47,7 @@
 	return nil
 }
 
-// peekBitsFast requires that at least one bit is requested every time.
+// peekByteFast requires that at least one byte is requested every time.
 // There are no checks if the buffer is filled.
 func (b *bitReaderBytes) peekByteFast() uint8 {
 	got := uint8(b.value >> 56)
@@ -171,9 +67,7 @@
 	}
 
 	// 2 bounds checks.
-	v := b.in[b.off-4 : b.off]
-	v = v[:4]
-	low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	low := le.Load32(b.in, b.off-4)
 	b.value |= uint64(low) << (b.bitsRead - 32)
 	b.bitsRead -= 32
 	b.off -= 4
@@ -182,7 +76,7 @@
 // fillFastStart() assumes the bitReaderBytes is empty and there is at least 8 bytes to read.
 func (b *bitReaderBytes) fillFastStart() {
 	// Do single re-slice to avoid bounds checks.
-	b.value = binary.LittleEndian.Uint64(b.in[b.off-8:])
+	b.value = le.Load64(b.in, b.off-8)
 	b.bitsRead = 0
 	b.off -= 8
 }
@@ -192,10 +86,8 @@
 	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)
+	if b.off >= 4 {
+		low := le.Load32(b.in, b.off-4)
 		b.value |= uint64(low) << (b.bitsRead - 32)
 		b.bitsRead -= 32
 		b.off -= 4
@@ -213,10 +105,17 @@
 	return b.off == 0 && b.bitsRead >= 64
 }
 
+func (b *bitReaderBytes) remaining() uint {
+	return b.off*8 + uint(64-b.bitsRead)
+}
+
 // close the bitstream and returns an error if out-of-buffer reads occurred.
 func (b *bitReaderBytes) close() error {
 	// Release reference.
 	b.in = nil
+	if b.remaining() > 0 {
+		return fmt.Errorf("corrupt input: %d bits remain on stream", b.remaining())
+	}
 	if b.bitsRead > 64 {
 		return io.ErrUnexpectedEOF
 	}
@@ -275,10 +174,7 @@
 		return
 	}
 
-	// 2 bounds checks.
-	v := b.in[b.off-4 : b.off]
-	v = v[:4]
-	low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	low := le.Load32(b.in, b.off-4)
 	b.value |= uint64(low) << ((b.bitsRead - 32) & 63)
 	b.bitsRead -= 32
 	b.off -= 4
@@ -286,8 +182,7 @@
 
 // fillFastStart() assumes the bitReaderShifted is empty and there is at least 8 bytes to read.
 func (b *bitReaderShifted) fillFastStart() {
-	// Do single re-slice to avoid bounds checks.
-	b.value = binary.LittleEndian.Uint64(b.in[b.off-8:])
+	b.value = le.Load64(b.in, b.off-8)
 	b.bitsRead = 0
 	b.off -= 8
 }
@@ -298,9 +193,7 @@
 		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)
+		low := le.Load32(b.in, b.off-4)
 		b.value |= uint64(low) << ((b.bitsRead - 32) & 63)
 		b.bitsRead -= 32
 		b.off -= 4
@@ -313,15 +206,17 @@
 	}
 }
 
-// finished returns true if all bits have been read from the bit stream.
-func (b *bitReaderShifted) finished() bool {
-	return b.off == 0 && b.bitsRead >= 64
+func (b *bitReaderShifted) remaining() uint {
+	return b.off*8 + uint(64-b.bitsRead)
 }
 
 // close the bitstream and returns an error if out-of-buffer reads occurred.
 func (b *bitReaderShifted) close() error {
 	// Release reference.
 	b.in = nil
+	if b.remaining() > 0 {
+		return fmt.Errorf("corrupt input: %d bits remain on stream", b.remaining())
+	}
 	if b.bitsRead > 64 {
 		return io.ErrUnexpectedEOF
 	}
diff --git a/vendor/github.com/klauspost/compress/huff0/bitwriter.go b/vendor/github.com/klauspost/compress/huff0/bitwriter.go
index 6bce4e8..0ebc9aa 100644
--- a/vendor/github.com/klauspost/compress/huff0/bitwriter.go
+++ b/vendor/github.com/klauspost/compress/huff0/bitwriter.go
@@ -5,8 +5,6 @@
 
 package huff0
 
-import "fmt"
-
 // bitWriter will write bits.
 // First bit will be LSB of the first byte of output.
 type bitWriter struct {
@@ -15,22 +13,6 @@
 	out          []byte
 }
 
-// bitMask16 is bitmasks. Has extra to avoid bounds check.
-var bitMask16 = [32]uint16{
-	0, 1, 3, 7, 0xF, 0x1F,
-	0x3F, 0x7F, 0xFF, 0x1FF, 0x3FF, 0x7FF,
-	0xFFF, 0x1FFF, 0x3FFF, 0x7FFF, 0xFFFF, 0xFFFF,
-	0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
-	0xFFFF, 0xFFFF} /* up to 16 bits */
-
-// addBits16NC will add up to 16 bits.
-// It will not check if there is space for them,
-// so the caller must ensure that it has flushed recently.
-func (b *bitWriter) addBits16NC(value uint16, bits uint8) {
-	b.bitContainer |= uint64(value&bitMask16[bits&31]) << (b.nBits & 63)
-	b.nBits += bits
-}
-
 // addBits16Clean will add up to 16 bits. value may not contain more set bits than indicated.
 // It will not check if there is space for them, so the caller must ensure that it has flushed recently.
 func (b *bitWriter) addBits16Clean(value uint16, bits uint8) {
@@ -70,102 +52,20 @@
 	b.nBits += encA.nBits + encB.nBits
 }
 
-// addBits16ZeroNC will add up to 16 bits.
+// encFourSymbols adds up to 32 bits from four symbols.
 // It will not check if there is space for them,
-// so the caller must ensure that it has flushed recently.
-// This is fastest if bits can be zero.
-func (b *bitWriter) addBits16ZeroNC(value uint16, bits uint8) {
-	if bits == 0 {
-		return
-	}
-	value <<= (16 - bits) & 15
-	value >>= (16 - bits) & 15
-	b.bitContainer |= uint64(value) << (b.nBits & 63)
-	b.nBits += bits
-}
-
-// flush will flush all pending full bytes.
-// There will be at least 56 bits available for writing when this has been called.
-// Using flush32 is faster, but leaves less space for writing.
-func (b *bitWriter) flush() {
-	v := b.nBits >> 3
-	switch v {
-	case 0:
-		return
-	case 1:
-		b.out = append(b.out,
-			byte(b.bitContainer),
-		)
-		b.bitContainer >>= 1 << 3
-	case 2:
-		b.out = append(b.out,
-			byte(b.bitContainer),
-			byte(b.bitContainer>>8),
-		)
-		b.bitContainer >>= 2 << 3
-	case 3:
-		b.out = append(b.out,
-			byte(b.bitContainer),
-			byte(b.bitContainer>>8),
-			byte(b.bitContainer>>16),
-		)
-		b.bitContainer >>= 3 << 3
-	case 4:
-		b.out = append(b.out,
-			byte(b.bitContainer),
-			byte(b.bitContainer>>8),
-			byte(b.bitContainer>>16),
-			byte(b.bitContainer>>24),
-		)
-		b.bitContainer >>= 4 << 3
-	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),
-		)
-		b.bitContainer >>= 5 << 3
-	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),
-		)
-		b.bitContainer >>= 6 << 3
-	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),
-		)
-		b.bitContainer >>= 7 << 3
-	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),
-		)
-		b.bitContainer = 0
-		b.nBits = 0
-		return
-	default:
-		panic(fmt.Errorf("bits (%d) > 64", b.nBits))
-	}
-	b.nBits &= 7
+// so the caller must ensure that b has been flushed recently.
+func (b *bitWriter) encFourSymbols(encA, encB, encC, encD cTableEntry) {
+	bitsA := encA.nBits
+	bitsB := bitsA + encB.nBits
+	bitsC := bitsB + encC.nBits
+	bitsD := bitsC + encD.nBits
+	combined := uint64(encA.val) |
+		(uint64(encB.val) << (bitsA & 63)) |
+		(uint64(encC.val) << (bitsB & 63)) |
+		(uint64(encD.val) << (bitsC & 63))
+	b.bitContainer |= combined << (b.nBits & 63)
+	b.nBits += bitsD
 }
 
 // flush32 will flush out, so there are at least 32 bits available for writing.
@@ -194,17 +94,9 @@
 
 // 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.
-func (b *bitWriter) reset(out []byte) {
-	b.bitContainer = 0
-	b.nBits = 0
-	b.out = out
 }
diff --git a/vendor/github.com/klauspost/compress/huff0/bytereader.go b/vendor/github.com/klauspost/compress/huff0/bytereader.go
deleted file mode 100644
index 50bcdf6..0000000
--- a/vendor/github.com/klauspost/compress/huff0/bytereader.go
+++ /dev/null
@@ -1,54 +0,0 @@
-// Copyright 2018 Klaus Post. All rights reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-// Based on work Copyright (c) 2013, Yann Collet, released under BSD License.
-
-package huff0
-
-// byteReader provides a byte reader that reads
-// little endian values from a byte stream.
-// The input stream is manually advanced.
-// The reader performs no bounds checks.
-type byteReader struct {
-	b   []byte
-	off int
-}
-
-// init will initialize the reader and set the input.
-func (b *byteReader) init(in []byte) {
-	b.b = in
-	b.off = 0
-}
-
-// advance the stream b n bytes.
-func (b *byteReader) advance(n uint) {
-	b.off += int(n)
-}
-
-// Int32 returns a little endian int32 starting at current offset.
-func (b byteReader) Int32() int32 {
-	v3 := int32(b.b[b.off+3])
-	v2 := int32(b.b[b.off+2])
-	v1 := int32(b.b[b.off+1])
-	v0 := int32(b.b[b.off])
-	return (v3 << 24) | (v2 << 16) | (v1 << 8) | v0
-}
-
-// Uint32 returns a little endian uint32 starting at current offset.
-func (b byteReader) Uint32() uint32 {
-	v3 := uint32(b.b[b.off+3])
-	v2 := uint32(b.b[b.off+2])
-	v1 := uint32(b.b[b.off+1])
-	v0 := uint32(b.b[b.off])
-	return (v3 << 24) | (v2 << 16) | (v1 << 8) | v0
-}
-
-// unread returns the unread portion of the input.
-func (b byteReader) unread() []byte {
-	return b.b[b.off:]
-}
-
-// remain will return the number of bytes remaining.
-func (b byteReader) remain() int {
-	return len(b.b) - b.off
-}
diff --git a/vendor/github.com/klauspost/compress/huff0/compress.go b/vendor/github.com/klauspost/compress/huff0/compress.go
index 0823c92..84aa3d1 100644
--- a/vendor/github.com/klauspost/compress/huff0/compress.go
+++ b/vendor/github.com/klauspost/compress/huff0/compress.go
@@ -2,6 +2,7 @@
 
 import (
 	"fmt"
+	"math"
 	"runtime"
 	"sync"
 )
@@ -161,11 +162,75 @@
 	return s.Out, false, nil
 }
 
-func (s *Scratch) compress1X(src []byte) ([]byte, error) {
-	return s.compress1xDo(s.Out, src)
+// EstimateSizes will estimate the data sizes
+func EstimateSizes(in []byte, s *Scratch) (tableSz, dataSz, reuseSz int, err error) {
+	s, err = s.prepare(in)
+	if err != nil {
+		return 0, 0, 0, err
+	}
+
+	// Create histogram, if none was provided.
+	tableSz, dataSz, reuseSz = -1, -1, -1
+	maxCount := s.maxCount
+	var canReuse = false
+	if maxCount == 0 {
+		maxCount, canReuse = s.countSimple(in)
+	} else {
+		canReuse = s.canUseTable(s.prevTable)
+	}
+
+	// We want the output size to be less than this:
+	wantSize := len(in)
+	if s.WantLogLess > 0 {
+		wantSize -= wantSize >> s.WantLogLess
+	}
+
+	// Reset for next run.
+	s.clearCount = true
+	s.maxCount = 0
+	if maxCount >= len(in) {
+		if maxCount > len(in) {
+			return 0, 0, 0, fmt.Errorf("maxCount (%d) > length (%d)", maxCount, len(in))
+		}
+		if len(in) == 1 {
+			return 0, 0, 0, ErrIncompressible
+		}
+		// One symbol, use RLE
+		return 0, 0, 0, ErrUseRLE
+	}
+	if maxCount == 1 || maxCount < (len(in)>>7) {
+		// Each symbol present maximum once or too well distributed.
+		return 0, 0, 0, ErrIncompressible
+	}
+
+	// Calculate new table.
+	err = s.buildCTable()
+	if err != nil {
+		return 0, 0, 0, err
+	}
+
+	if false && !s.canUseTable(s.cTable) {
+		panic("invalid table generated")
+	}
+
+	tableSz, err = s.cTable.estTableSize(s)
+	if err != nil {
+		return 0, 0, 0, err
+	}
+	if canReuse {
+		reuseSz = s.prevTable.estimateSize(s.count[:s.symbolLen])
+	}
+	dataSz = s.cTable.estimateSize(s.count[:s.symbolLen])
+
+	// Restore
+	return tableSz, dataSz, reuseSz, nil
 }
 
-func (s *Scratch) compress1xDo(dst, src []byte) ([]byte, error) {
+func (s *Scratch) compress1X(src []byte) ([]byte, error) {
+	return s.compress1xDo(s.Out, src), nil
+}
+
+func (s *Scratch) compress1xDo(dst, src []byte) []byte {
 	var bw = bitWriter{out: dst}
 
 	// N is length divisible by 4.
@@ -183,8 +248,7 @@
 			tmp := src[n : n+4]
 			// tmp should be len 4
 			bw.flush32()
-			bw.encTwoSymbols(cTable, tmp[3], tmp[2])
-			bw.encTwoSymbols(cTable, tmp[1], tmp[0])
+			bw.encFourSymbols(cTable[tmp[3]], cTable[tmp[2]], cTable[tmp[1]], cTable[tmp[0]])
 		}
 	} else {
 		for ; n >= 0; n -= 4 {
@@ -196,8 +260,8 @@
 			bw.encTwoSymbols(cTable, tmp[1], tmp[0])
 		}
 	}
-	err := bw.close()
-	return bw.out, err
+	bw.close()
+	return bw.out
 }
 
 var sixZeros [6]byte
@@ -219,11 +283,11 @@
 		}
 		src = src[len(toDo):]
 
-		var err error
 		idx := len(s.Out)
-		s.Out, err = s.compress1xDo(s.Out, toDo)
-		if err != nil {
-			return nil, err
+		s.Out = s.compress1xDo(s.Out, toDo)
+		if len(s.Out)-idx > math.MaxUint16 {
+			// We cannot store the size in the jump table
+			return nil, ErrIncompressible
 		}
 		// Write compressed length as little endian before block.
 		if i < 3 {
@@ -247,7 +311,6 @@
 
 	segmentSize := (len(src) + 3) / 4
 	var wg sync.WaitGroup
-	var errs [4]error
 	wg.Add(4)
 	for i := 0; i < 4; i++ {
 		toDo := src
@@ -258,16 +321,17 @@
 
 		// Separate goroutine for each block.
 		go func(i int) {
-			s.tmpOut[i], errs[i] = s.compress1xDo(s.tmpOut[i][:0], toDo)
+			s.tmpOut[i] = s.compress1xDo(s.tmpOut[i][:0], toDo)
 			wg.Done()
 		}(i)
 	}
 	wg.Wait()
 	for i := 0; i < 4; i++ {
-		if errs[i] != nil {
-			return nil, errs[i]
-		}
 		o := s.tmpOut[i]
+		if len(o) > math.MaxUint16 {
+			// We cannot store the size in the jump table
+			return nil, ErrIncompressible
+		}
 		// Write compressed length as little endian before block.
 		if i < 3 {
 			// Last length is not written.
@@ -286,35 +350,36 @@
 // Does not update s.clearCount.
 func (s *Scratch) countSimple(in []byte) (max int, reuse bool) {
 	reuse = true
+	_ = s.count // Assert that s != nil to speed up the following loop.
 	for _, v := range in {
 		s.count[v]++
 	}
 	m := uint32(0)
 	if len(s.prevTable) > 0 {
 		for i, v := range s.count[:] {
+			if v == 0 {
+				continue
+			}
 			if v > m {
 				m = v
 			}
-			if v > 0 {
-				s.symbolLen = uint16(i) + 1
-				if i >= len(s.prevTable) {
-					reuse = false
-				} else {
-					if s.prevTable[i].nBits == 0 {
-						reuse = false
-					}
-				}
+			s.symbolLen = uint16(i) + 1
+			if i >= len(s.prevTable) {
+				reuse = false
+			} else if s.prevTable[i].nBits == 0 {
+				reuse = false
 			}
 		}
 		return int(m), reuse
 	}
 	for i, v := range s.count[:] {
+		if v == 0 {
+			continue
+		}
 		if v > m {
 			m = v
 		}
-		if v > 0 {
-			s.symbolLen = uint16(i) + 1
-		}
+		s.symbolLen = uint16(i) + 1
 	}
 	return int(m), false
 }
@@ -331,6 +396,7 @@
 	return true
 }
 
+//lint:ignore U1000 used for debugging
 func (s *Scratch) validateTable(c cTable) bool {
 	if len(c) < int(s.symbolLen) {
 		return false
@@ -350,7 +416,7 @@
 
 // minTableLog provides the minimum logSize to safely represent a distribution.
 func (s *Scratch) minTableLog() uint8 {
-	minBitsSrc := highBit32(uint32(s.br.remain())) + 1
+	minBitsSrc := highBit32(uint32(s.srcLen)) + 1
 	minBitsSymbols := highBit32(uint32(s.symbolLen-1)) + 2
 	if minBitsSrc < minBitsSymbols {
 		return uint8(minBitsSrc)
@@ -362,7 +428,7 @@
 func (s *Scratch) optimalTableLog() {
 	tableLog := s.TableLog
 	minBits := s.minTableLog()
-	maxBitsSrc := uint8(highBit32(uint32(s.br.remain()-1))) - 1
+	maxBitsSrc := uint8(highBit32(uint32(s.srcLen-1))) - 1
 	if maxBitsSrc < tableLog {
 		// Accuracy can be reduced
 		tableLog = maxBitsSrc
@@ -410,34 +476,35 @@
 	// Different from reference implementation.
 	huffNode0 := s.nodes[0 : huffNodesLen+1]
 
-	for huffNode[nonNullRank].count == 0 {
+	for huffNode[nonNullRank].count() == 0 {
 		nonNullRank--
 	}
 
 	lowS := int16(nonNullRank)
 	nodeRoot := nodeNb + lowS - 1
 	lowN := nodeNb
-	huffNode[nodeNb].count = huffNode[lowS].count + huffNode[lowS-1].count
-	huffNode[lowS].parent, huffNode[lowS-1].parent = uint16(nodeNb), uint16(nodeNb)
+	huffNode[nodeNb].setCount(huffNode[lowS].count() + huffNode[lowS-1].count())
+	huffNode[lowS].setParent(nodeNb)
+	huffNode[lowS-1].setParent(nodeNb)
 	nodeNb++
 	lowS -= 2
 	for n := nodeNb; n <= nodeRoot; n++ {
-		huffNode[n].count = 1 << 30
+		huffNode[n].setCount(1 << 30)
 	}
 	// fake entry, strong barrier
-	huffNode0[0].count = 1 << 31
+	huffNode0[0].setCount(1 << 31)
 
 	// create parents
 	for nodeNb <= nodeRoot {
 		var n1, n2 int16
-		if huffNode0[lowS+1].count < huffNode0[lowN+1].count {
+		if huffNode0[lowS+1].count() < huffNode0[lowN+1].count() {
 			n1 = lowS
 			lowS--
 		} else {
 			n1 = lowN
 			lowN++
 		}
-		if huffNode0[lowS+1].count < huffNode0[lowN+1].count {
+		if huffNode0[lowS+1].count() < huffNode0[lowN+1].count() {
 			n2 = lowS
 			lowS--
 		} else {
@@ -445,18 +512,19 @@
 			lowN++
 		}
 
-		huffNode[nodeNb].count = huffNode0[n1+1].count + huffNode0[n2+1].count
-		huffNode0[n1+1].parent, huffNode0[n2+1].parent = uint16(nodeNb), uint16(nodeNb)
+		huffNode[nodeNb].setCount(huffNode0[n1+1].count() + huffNode0[n2+1].count())
+		huffNode0[n1+1].setParent(nodeNb)
+		huffNode0[n2+1].setParent(nodeNb)
 		nodeNb++
 	}
 
 	// distribute weights (unlimited tree height)
-	huffNode[nodeRoot].nbBits = 0
+	huffNode[nodeRoot].setNbBits(0)
 	for n := nodeRoot - 1; n >= startNode; n-- {
-		huffNode[n].nbBits = huffNode[huffNode[n].parent].nbBits + 1
+		huffNode[n].setNbBits(huffNode[huffNode[n].parent()].nbBits() + 1)
 	}
 	for n := uint16(0); n <= nonNullRank; n++ {
-		huffNode[n].nbBits = huffNode[huffNode[n].parent].nbBits + 1
+		huffNode[n].setNbBits(huffNode[huffNode[n].parent()].nbBits() + 1)
 	}
 	s.actualTableLog = s.setMaxHeight(int(nonNullRank))
 	maxNbBits := s.actualTableLog
@@ -468,7 +536,7 @@
 	var nbPerRank [tableLogMax + 1]uint16
 	var valPerRank [16]uint16
 	for _, v := range huffNode[:nonNullRank+1] {
-		nbPerRank[v.nbBits]++
+		nbPerRank[v.nbBits()]++
 	}
 	// determine stating value per rank
 	{
@@ -483,7 +551,7 @@
 
 	// push nbBits per symbol, symbol order
 	for _, v := range huffNode[:nonNullRank+1] {
-		s.cTable[v.symbol].nBits = v.nbBits
+		s.cTable[v.symbol()].nBits = v.nbBits()
 	}
 
 	// assign value within rank, symbol order
@@ -529,12 +597,12 @@
 		pos := rank[r].current
 		rank[r].current++
 		prev := nodes[(pos-1)&huffNodesMask]
-		for pos > rank[r].base && c > prev.count {
+		for pos > rank[r].base && c > prev.count() {
 			nodes[pos&huffNodesMask] = prev
 			pos--
 			prev = nodes[(pos-1)&huffNodesMask]
 		}
-		nodes[pos&huffNodesMask] = nodeElt{count: c, symbol: byte(n)}
+		nodes[pos&huffNodesMask] = makeNodeElt(c, byte(n))
 	}
 }
 
@@ -543,7 +611,7 @@
 	huffNode := s.nodes[1 : huffNodesLen+1]
 	//huffNode = huffNode[: huffNodesLen]
 
-	largestBits := huffNode[lastNonNull].nbBits
+	largestBits := huffNode[lastNonNull].nbBits()
 
 	// early exit : no elt > maxNbBits
 	if largestBits <= maxNbBits {
@@ -553,14 +621,14 @@
 	baseCost := int(1) << (largestBits - maxNbBits)
 	n := uint32(lastNonNull)
 
-	for huffNode[n].nbBits > maxNbBits {
-		totalCost += baseCost - (1 << (largestBits - huffNode[n].nbBits))
-		huffNode[n].nbBits = maxNbBits
+	for huffNode[n].nbBits() > maxNbBits {
+		totalCost += baseCost - (1 << (largestBits - huffNode[n].nbBits()))
+		huffNode[n].setNbBits(maxNbBits)
 		n--
 	}
 	// n stops at huffNode[n].nbBits <= maxNbBits
 
-	for huffNode[n].nbBits == maxNbBits {
+	for huffNode[n].nbBits() == maxNbBits {
 		n--
 	}
 	// n end at index of smallest symbol using < maxNbBits
@@ -581,10 +649,10 @@
 		{
 			currentNbBits := maxNbBits
 			for pos := int(n); pos >= 0; pos-- {
-				if huffNode[pos].nbBits >= currentNbBits {
+				if huffNode[pos].nbBits() >= currentNbBits {
 					continue
 				}
-				currentNbBits = huffNode[pos].nbBits // < maxNbBits
+				currentNbBits = huffNode[pos].nbBits() // < maxNbBits
 				rankLast[maxNbBits-currentNbBits] = uint32(pos)
 			}
 		}
@@ -601,8 +669,8 @@
 				if lowPos == noSymbol {
 					break
 				}
-				highTotal := huffNode[highPos].count
-				lowTotal := 2 * huffNode[lowPos].count
+				highTotal := huffNode[highPos].count()
+				lowTotal := 2 * huffNode[lowPos].count()
 				if highTotal <= lowTotal {
 					break
 				}
@@ -618,13 +686,14 @@
 				// this rank is no longer empty
 				rankLast[nBitsToDecrease-1] = rankLast[nBitsToDecrease]
 			}
-			huffNode[rankLast[nBitsToDecrease]].nbBits++
+			huffNode[rankLast[nBitsToDecrease]].setNbBits(1 +
+				huffNode[rankLast[nBitsToDecrease]].nbBits())
 			if rankLast[nBitsToDecrease] == 0 {
 				/* special case, reached largest symbol */
 				rankLast[nBitsToDecrease] = noSymbol
 			} else {
 				rankLast[nBitsToDecrease]--
-				if huffNode[rankLast[nBitsToDecrease]].nbBits != maxNbBits-nBitsToDecrease {
+				if huffNode[rankLast[nBitsToDecrease]].nbBits() != maxNbBits-nBitsToDecrease {
 					rankLast[nBitsToDecrease] = noSymbol /* this rank is now empty */
 				}
 			}
@@ -632,15 +701,15 @@
 
 		for totalCost < 0 { /* Sometimes, cost correction overshoot */
 			if rankLast[1] == noSymbol { /* special case : no rank 1 symbol (using maxNbBits-1); let's create one from largest rank 0 (using maxNbBits) */
-				for huffNode[n].nbBits == maxNbBits {
+				for huffNode[n].nbBits() == maxNbBits {
 					n--
 				}
-				huffNode[n+1].nbBits--
+				huffNode[n+1].setNbBits(huffNode[n+1].nbBits() - 1)
 				rankLast[1] = n + 1
 				totalCost++
 				continue
 			}
-			huffNode[rankLast[1]+1].nbBits--
+			huffNode[rankLast[1]+1].setNbBits(huffNode[rankLast[1]+1].nbBits() - 1)
 			rankLast[1]++
 			totalCost++
 		}
@@ -648,9 +717,26 @@
 	return maxNbBits
 }
 
-type nodeElt struct {
-	count  uint32
-	parent uint16
-	symbol byte
-	nbBits uint8
+// A nodeElt is the fields
+//
+//	count  uint32
+//	parent uint16
+//	symbol byte
+//	nbBits uint8
+//
+// in some order, all squashed into an integer so that the compiler
+// always loads and stores entire nodeElts instead of separate fields.
+type nodeElt uint64
+
+func makeNodeElt(count uint32, symbol byte) nodeElt {
+	return nodeElt(count) | nodeElt(symbol)<<48
 }
+
+func (e *nodeElt) count() uint32  { return uint32(*e) }
+func (e *nodeElt) parent() uint16 { return uint16(*e >> 32) }
+func (e *nodeElt) symbol() byte   { return byte(*e >> 48) }
+func (e *nodeElt) nbBits() uint8  { return uint8(*e >> 56) }
+
+func (e *nodeElt) setCount(c uint32) { *e = (*e)&0xffffffff00000000 | nodeElt(c) }
+func (e *nodeElt) setParent(p int16) { *e = (*e)&0xffff0000ffffffff | nodeElt(uint16(p))<<32 }
+func (e *nodeElt) setNbBits(n uint8) { *e = (*e)&0x00ffffffffffffff | nodeElt(n)<<56 }
diff --git a/vendor/github.com/klauspost/compress/huff0/decompress.go b/vendor/github.com/klauspost/compress/huff0/decompress.go
index 41703bb..0f56b02 100644
--- a/vendor/github.com/klauspost/compress/huff0/decompress.go
+++ b/vendor/github.com/klauspost/compress/huff0/decompress.go
@@ -4,13 +4,13 @@
 	"errors"
 	"fmt"
 	"io"
+	"sync"
 
 	"github.com/klauspost/compress/fse"
 )
 
 type dTable struct {
 	single []dEntrySingle
-	double []dEntryDouble
 }
 
 // single-symbols decoding
@@ -18,13 +18,6 @@
 	entry uint16
 }
 
-// double-symbols decoding
-type dEntryDouble struct {
-	seq   uint16
-	nBits uint8
-	len   uint8
-}
-
 // Uses special code for all tables that are < 8 bits.
 const use8BitTables = true
 
@@ -34,7 +27,7 @@
 // If no Scratch is provided a new one is allocated.
 // The returned Scratch can be used for encoding or decoding input using this table.
 func ReadTable(in []byte, s *Scratch) (s2 *Scratch, remain []byte, err error) {
-	s, err = s.prepare(in)
+	s, err = s.prepare(nil)
 	if err != nil {
 		return s, nil, err
 	}
@@ -68,7 +61,7 @@
 		b, err := fse.Decompress(in[:iSize], s.fse)
 		s.fse.Out = nil
 		if err != nil {
-			return s, nil, err
+			return s, nil, fmt.Errorf("fse decompress returned: %w", err)
 		}
 		if len(b) > 255 {
 			return s, nil, errors.New("corrupt input: output table too large")
@@ -216,6 +209,7 @@
 	return &Decoder{
 		dt:             s.dt,
 		actualTableLog: s.actualTableLog,
+		bufs:           &s.decPool,
 	}
 }
 
@@ -223,103 +217,15 @@
 type Decoder struct {
 	dt             dTable
 	actualTableLog uint8
+	bufs           *sync.Pool
 }
 
-// Decompress1X will decompress a 1X encoded stream.
-// The cap of the output buffer will be the maximum decompressed size.
-// The length of the supplied input must match the end of a block exactly.
-func (d *Decoder) Decompress1X(dst, src []byte) ([]byte, error) {
-	if len(d.dt.single) == 0 {
-		return nil, errors.New("no table loaded")
+func (d *Decoder) buffer() *[4][256]byte {
+	buf, ok := d.bufs.Get().(*[4][256]byte)
+	if ok {
+		return buf
 	}
-	if use8BitTables && d.actualTableLog <= 8 {
-		return d.decompress1X8Bit(dst, src)
-	}
-	var br bitReaderShifted
-	err := br.init(src)
-	if err != nil {
-		return dst, err
-	}
-	maxDecodedSize := cap(dst)
-	dst = dst[:0]
-
-	// Avoid bounds check by always having full sized table.
-	const tlSize = 1 << tableLogMax
-	const tlMask = tlSize - 1
-	dt := d.dt.single[:tlSize]
-
-	// Use temp table to avoid bound checks/append penalty.
-	var buf [256]byte
-	var off uint8
-
-	for br.off >= 8 {
-		br.fillFast()
-		v := dt[br.peekBitsFast(d.actualTableLog)&tlMask]
-		br.advance(uint8(v.entry))
-		buf[off+0] = uint8(v.entry >> 8)
-
-		v = dt[br.peekBitsFast(d.actualTableLog)&tlMask]
-		br.advance(uint8(v.entry))
-		buf[off+1] = uint8(v.entry >> 8)
-
-		// Refill
-		br.fillFast()
-
-		v = dt[br.peekBitsFast(d.actualTableLog)&tlMask]
-		br.advance(uint8(v.entry))
-		buf[off+2] = uint8(v.entry >> 8)
-
-		v = dt[br.peekBitsFast(d.actualTableLog)&tlMask]
-		br.advance(uint8(v.entry))
-		buf[off+3] = uint8(v.entry >> 8)
-
-		off += 4
-		if off == 0 {
-			if len(dst)+256 > maxDecodedSize {
-				br.close()
-				return nil, ErrMaxDecodedSizeExceeded
-			}
-			dst = append(dst, buf[:]...)
-		}
-	}
-
-	if len(dst)+int(off) > maxDecodedSize {
-		br.close()
-		return nil, ErrMaxDecodedSizeExceeded
-	}
-	dst = append(dst, buf[:off]...)
-
-	// br < 8, so uint8 is fine
-	bitsLeft := uint8(br.off)*8 + 64 - br.bitsRead
-	for bitsLeft > 0 {
-		br.fill()
-		if false && br.bitsRead >= 32 {
-			if br.off >= 4 {
-				v := br.in[br.off-4:]
-				v = v[:4]
-				low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
-				br.value = (br.value << 32) | uint64(low)
-				br.bitsRead -= 32
-				br.off -= 4
-			} else {
-				for br.off > 0 {
-					br.value = (br.value << 8) | uint64(br.in[br.off-1])
-					br.bitsRead -= 8
-					br.off--
-				}
-			}
-		}
-		if len(dst) >= maxDecodedSize {
-			br.close()
-			return nil, ErrMaxDecodedSizeExceeded
-		}
-		v := d.dt.single[br.peekBitsFast(d.actualTableLog)&tlMask]
-		nBits := uint8(v.entry)
-		br.advance(nBits)
-		bitsLeft -= nBits
-		dst = append(dst, uint8(v.entry>>8))
-	}
-	return dst, br.close()
+	return &[4][256]byte{}
 }
 
 // decompress1X8Bit will decompress a 1X encoded stream with tablelog <= 8.
@@ -341,41 +247,258 @@
 	dt := d.dt.single[:256]
 
 	// Use temp table to avoid bound checks/append penalty.
-	var buf [256]byte
+	bufs := d.buffer()
+	buf := &bufs[0]
 	var off uint8
 
-	shift := (8 - d.actualTableLog) & 7
+	switch d.actualTableLog {
+	case 8:
+		const shift = 0
+		for br.off >= 4 {
+			br.fillFast()
+			v := dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+0] = uint8(v.entry >> 8)
 
-	//fmt.Printf("mask: %b, tl:%d\n", mask, d.actualTableLog)
-	for br.off >= 4 {
-		br.fillFast()
-		v := dt[br.peekByteFast()>>shift]
-		br.advance(uint8(v.entry))
-		buf[off+0] = uint8(v.entry >> 8)
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+1] = uint8(v.entry >> 8)
 
-		v = dt[br.peekByteFast()>>shift]
-		br.advance(uint8(v.entry))
-		buf[off+1] = uint8(v.entry >> 8)
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+2] = uint8(v.entry >> 8)
 
-		v = dt[br.peekByteFast()>>shift]
-		br.advance(uint8(v.entry))
-		buf[off+2] = uint8(v.entry >> 8)
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+3] = uint8(v.entry >> 8)
 
-		v = dt[br.peekByteFast()>>shift]
-		br.advance(uint8(v.entry))
-		buf[off+3] = uint8(v.entry >> 8)
-
-		off += 4
-		if off == 0 {
-			if len(dst)+256 > maxDecodedSize {
-				br.close()
-				return nil, ErrMaxDecodedSizeExceeded
+			off += 4
+			if off == 0 {
+				if len(dst)+256 > maxDecodedSize {
+					br.close()
+					d.bufs.Put(bufs)
+					return nil, ErrMaxDecodedSizeExceeded
+				}
+				dst = append(dst, buf[:]...)
 			}
-			dst = append(dst, buf[:]...)
 		}
+	case 7:
+		const shift = 8 - 7
+		for br.off >= 4 {
+			br.fillFast()
+			v := dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+0] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+1] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+2] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+3] = uint8(v.entry >> 8)
+
+			off += 4
+			if off == 0 {
+				if len(dst)+256 > maxDecodedSize {
+					br.close()
+					d.bufs.Put(bufs)
+					return nil, ErrMaxDecodedSizeExceeded
+				}
+				dst = append(dst, buf[:]...)
+			}
+		}
+	case 6:
+		const shift = 8 - 6
+		for br.off >= 4 {
+			br.fillFast()
+			v := dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+0] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+1] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+2] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+3] = uint8(v.entry >> 8)
+
+			off += 4
+			if off == 0 {
+				if len(dst)+256 > maxDecodedSize {
+					d.bufs.Put(bufs)
+					br.close()
+					return nil, ErrMaxDecodedSizeExceeded
+				}
+				dst = append(dst, buf[:]...)
+			}
+		}
+	case 5:
+		const shift = 8 - 5
+		for br.off >= 4 {
+			br.fillFast()
+			v := dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+0] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+1] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+2] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+3] = uint8(v.entry >> 8)
+
+			off += 4
+			if off == 0 {
+				if len(dst)+256 > maxDecodedSize {
+					d.bufs.Put(bufs)
+					br.close()
+					return nil, ErrMaxDecodedSizeExceeded
+				}
+				dst = append(dst, buf[:]...)
+			}
+		}
+	case 4:
+		const shift = 8 - 4
+		for br.off >= 4 {
+			br.fillFast()
+			v := dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+0] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+1] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+2] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+3] = uint8(v.entry >> 8)
+
+			off += 4
+			if off == 0 {
+				if len(dst)+256 > maxDecodedSize {
+					d.bufs.Put(bufs)
+					br.close()
+					return nil, ErrMaxDecodedSizeExceeded
+				}
+				dst = append(dst, buf[:]...)
+			}
+		}
+	case 3:
+		const shift = 8 - 3
+		for br.off >= 4 {
+			br.fillFast()
+			v := dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+0] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+1] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+2] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+3] = uint8(v.entry >> 8)
+
+			off += 4
+			if off == 0 {
+				if len(dst)+256 > maxDecodedSize {
+					d.bufs.Put(bufs)
+					br.close()
+					return nil, ErrMaxDecodedSizeExceeded
+				}
+				dst = append(dst, buf[:]...)
+			}
+		}
+	case 2:
+		const shift = 8 - 2
+		for br.off >= 4 {
+			br.fillFast()
+			v := dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+0] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+1] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+2] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+3] = uint8(v.entry >> 8)
+
+			off += 4
+			if off == 0 {
+				if len(dst)+256 > maxDecodedSize {
+					d.bufs.Put(bufs)
+					br.close()
+					return nil, ErrMaxDecodedSizeExceeded
+				}
+				dst = append(dst, buf[:]...)
+			}
+		}
+	case 1:
+		const shift = 8 - 1
+		for br.off >= 4 {
+			br.fillFast()
+			v := dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+0] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+1] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+2] = uint8(v.entry >> 8)
+
+			v = dt[uint8(br.value>>(56+shift))]
+			br.advance(uint8(v.entry))
+			buf[off+3] = uint8(v.entry >> 8)
+
+			off += 4
+			if off == 0 {
+				if len(dst)+256 > maxDecodedSize {
+					d.bufs.Put(bufs)
+					br.close()
+					return nil, ErrMaxDecodedSizeExceeded
+				}
+				dst = append(dst, buf[:]...)
+			}
+		}
+	default:
+		d.bufs.Put(bufs)
+		return nil, fmt.Errorf("invalid tablelog: %d", d.actualTableLog)
 	}
 
 	if len(dst)+int(off) > maxDecodedSize {
+		d.bufs.Put(bufs)
 		br.close()
 		return nil, ErrMaxDecodedSizeExceeded
 	}
@@ -383,6 +506,8 @@
 
 	// br < 4, so uint8 is fine
 	bitsLeft := int8(uint8(br.off)*8 + (64 - br.bitsRead))
+	shift := (8 - d.actualTableLog) & 7
+
 	for bitsLeft > 0 {
 		if br.bitsRead >= 64-8 {
 			for br.off > 0 {
@@ -393,6 +518,7 @@
 		}
 		if len(dst) >= maxDecodedSize {
 			br.close()
+			d.bufs.Put(bufs)
 			return nil, ErrMaxDecodedSizeExceeded
 		}
 		v := dt[br.peekByteFast()>>shift]
@@ -401,6 +527,7 @@
 		bitsLeft -= int8(nBits)
 		dst = append(dst, uint8(v.entry>>8))
 	}
+	d.bufs.Put(bufs)
 	return dst, br.close()
 }
 
@@ -420,33 +547,35 @@
 	dt := d.dt.single[:256]
 
 	// Use temp table to avoid bound checks/append penalty.
-	var buf [256]byte
+	bufs := d.buffer()
+	buf := &bufs[0]
 	var off uint8
 
-	const shift = 0
+	const shift = 56
 
 	//fmt.Printf("mask: %b, tl:%d\n", mask, d.actualTableLog)
 	for br.off >= 4 {
 		br.fillFast()
-		v := dt[br.peekByteFast()>>shift]
+		v := dt[uint8(br.value>>shift)]
 		br.advance(uint8(v.entry))
 		buf[off+0] = uint8(v.entry >> 8)
 
-		v = dt[br.peekByteFast()>>shift]
+		v = dt[uint8(br.value>>shift)]
 		br.advance(uint8(v.entry))
 		buf[off+1] = uint8(v.entry >> 8)
 
-		v = dt[br.peekByteFast()>>shift]
+		v = dt[uint8(br.value>>shift)]
 		br.advance(uint8(v.entry))
 		buf[off+2] = uint8(v.entry >> 8)
 
-		v = dt[br.peekByteFast()>>shift]
+		v = dt[uint8(br.value>>shift)]
 		br.advance(uint8(v.entry))
 		buf[off+3] = uint8(v.entry >> 8)
 
 		off += 4
 		if off == 0 {
 			if len(dst)+256 > maxDecodedSize {
+				d.bufs.Put(bufs)
 				br.close()
 				return nil, ErrMaxDecodedSizeExceeded
 			}
@@ -455,6 +584,7 @@
 	}
 
 	if len(dst)+int(off) > maxDecodedSize {
+		d.bufs.Put(bufs)
 		br.close()
 		return nil, ErrMaxDecodedSizeExceeded
 	}
@@ -471,15 +601,17 @@
 			}
 		}
 		if len(dst) >= maxDecodedSize {
+			d.bufs.Put(bufs)
 			br.close()
 			return nil, ErrMaxDecodedSizeExceeded
 		}
-		v := dt[br.peekByteFast()>>shift]
+		v := dt[br.peekByteFast()]
 		nBits := uint8(v.entry)
 		br.advance(nBits)
 		bitsLeft -= int8(nBits)
 		dst = append(dst, uint8(v.entry>>8))
 	}
+	d.bufs.Put(bufs)
 	return dst, br.close()
 }
 
@@ -487,196 +619,6 @@
 // The length of the supplied input must match the end of a block exactly.
 // The *capacity* of the dst slice must match the destination size of
 // the uncompressed data exactly.
-func (d *Decoder) Decompress4X(dst, src []byte) ([]byte, error) {
-	if len(d.dt.single) == 0 {
-		return nil, errors.New("no table loaded")
-	}
-	if len(src) < 6+(4*1) {
-		return nil, errors.New("input too small")
-	}
-	if use8BitTables && d.actualTableLog <= 8 {
-		return d.decompress4X8bit(dst, src)
-	}
-
-	var br [4]bitReaderShifted
-	start := 6
-	for i := 0; i < 3; i++ {
-		length := int(src[i*2]) | (int(src[i*2+1]) << 8)
-		if start+length >= len(src) {
-			return nil, errors.New("truncated input (or invalid offset)")
-		}
-		err := br[i].init(src[start : start+length])
-		if err != nil {
-			return nil, err
-		}
-		start += length
-	}
-	err := br[3].init(src[start:])
-	if err != nil {
-		return nil, err
-	}
-
-	// destination, offset to match first output
-	dstSize := cap(dst)
-	dst = dst[:dstSize]
-	out := dst
-	dstEvery := (dstSize + 3) / 4
-
-	const tlSize = 1 << tableLogMax
-	const tlMask = tlSize - 1
-	single := d.dt.single[:tlSize]
-
-	// Use temp table to avoid bound checks/append penalty.
-	var buf [256]byte
-	var off uint8
-	var decoded int
-
-	// Decode 2 values from each decoder/loop.
-	const bufoff = 256 / 4
-	for {
-		if br[0].off < 4 || br[1].off < 4 || br[2].off < 4 || br[3].off < 4 {
-			break
-		}
-
-		{
-			const stream = 0
-			const stream2 = 1
-			br[stream].fillFast()
-			br[stream2].fillFast()
-
-			val := br[stream].peekBitsFast(d.actualTableLog)
-			v := single[val&tlMask]
-			br[stream].advance(uint8(v.entry))
-			buf[off+bufoff*stream] = uint8(v.entry >> 8)
-
-			val2 := br[stream2].peekBitsFast(d.actualTableLog)
-			v2 := single[val2&tlMask]
-			br[stream2].advance(uint8(v2.entry))
-			buf[off+bufoff*stream2] = uint8(v2.entry >> 8)
-
-			val = br[stream].peekBitsFast(d.actualTableLog)
-			v = single[val&tlMask]
-			br[stream].advance(uint8(v.entry))
-			buf[off+bufoff*stream+1] = uint8(v.entry >> 8)
-
-			val2 = br[stream2].peekBitsFast(d.actualTableLog)
-			v2 = single[val2&tlMask]
-			br[stream2].advance(uint8(v2.entry))
-			buf[off+bufoff*stream2+1] = uint8(v2.entry >> 8)
-		}
-
-		{
-			const stream = 2
-			const stream2 = 3
-			br[stream].fillFast()
-			br[stream2].fillFast()
-
-			val := br[stream].peekBitsFast(d.actualTableLog)
-			v := single[val&tlMask]
-			br[stream].advance(uint8(v.entry))
-			buf[off+bufoff*stream] = uint8(v.entry >> 8)
-
-			val2 := br[stream2].peekBitsFast(d.actualTableLog)
-			v2 := single[val2&tlMask]
-			br[stream2].advance(uint8(v2.entry))
-			buf[off+bufoff*stream2] = uint8(v2.entry >> 8)
-
-			val = br[stream].peekBitsFast(d.actualTableLog)
-			v = single[val&tlMask]
-			br[stream].advance(uint8(v.entry))
-			buf[off+bufoff*stream+1] = uint8(v.entry >> 8)
-
-			val2 = br[stream2].peekBitsFast(d.actualTableLog)
-			v2 = single[val2&tlMask]
-			br[stream2].advance(uint8(v2.entry))
-			buf[off+bufoff*stream2+1] = uint8(v2.entry >> 8)
-		}
-
-		off += 2
-
-		if off == bufoff {
-			if bufoff > dstEvery {
-				return nil, errors.New("corruption detected: stream overrun 1")
-			}
-			copy(out, buf[:bufoff])
-			copy(out[dstEvery:], buf[bufoff:bufoff*2])
-			copy(out[dstEvery*2:], buf[bufoff*2:bufoff*3])
-			copy(out[dstEvery*3:], buf[bufoff*3:bufoff*4])
-			off = 0
-			out = out[bufoff:]
-			decoded += 256
-			// There must at least be 3 buffers left.
-			if len(out) < dstEvery*3 {
-				return nil, errors.New("corruption detected: stream overrun 2")
-			}
-		}
-	}
-	if off > 0 {
-		ioff := int(off)
-		if len(out) < dstEvery*3+ioff {
-			return nil, errors.New("corruption detected: stream overrun 3")
-		}
-		copy(out, buf[:off])
-		copy(out[dstEvery:dstEvery+ioff], buf[bufoff:bufoff*2])
-		copy(out[dstEvery*2:dstEvery*2+ioff], buf[bufoff*2:bufoff*3])
-		copy(out[dstEvery*3:dstEvery*3+ioff], buf[bufoff*3:bufoff*4])
-		decoded += int(off) * 4
-		out = out[off:]
-	}
-
-	// Decode remaining.
-	for i := range br {
-		offset := dstEvery * i
-		br := &br[i]
-		bitsLeft := br.off*8 + uint(64-br.bitsRead)
-		for bitsLeft > 0 {
-			br.fill()
-			if false && br.bitsRead >= 32 {
-				if br.off >= 4 {
-					v := br.in[br.off-4:]
-					v = v[:4]
-					low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
-					br.value = (br.value << 32) | uint64(low)
-					br.bitsRead -= 32
-					br.off -= 4
-				} else {
-					for br.off > 0 {
-						br.value = (br.value << 8) | uint64(br.in[br.off-1])
-						br.bitsRead -= 8
-						br.off--
-					}
-				}
-			}
-			// end inline...
-			if offset >= len(out) {
-				return nil, errors.New("corruption detected: stream overrun 4")
-			}
-
-			// Read value and increment offset.
-			val := br.peekBitsFast(d.actualTableLog)
-			v := single[val&tlMask].entry
-			nBits := uint8(v)
-			br.advance(nBits)
-			bitsLeft -= uint(nBits)
-			out[offset] = uint8(v >> 8)
-			offset++
-		}
-		decoded += offset - dstEvery*i
-		err = br.close()
-		if err != nil {
-			return nil, err
-		}
-	}
-	if dstSize != decoded {
-		return nil, errors.New("corruption detected: short output block")
-	}
-	return dst, nil
-}
-
-// Decompress4X will decompress a 4X encoded stream.
-// The length of the supplied input must match the end of a block exactly.
-// The *capacity* of the dst slice must match the destination size of
-// the uncompressed data exactly.
 func (d *Decoder) decompress4X8bit(dst, src []byte) ([]byte, error) {
 	if d.actualTableLog == 8 {
 		return d.decompress4X8bitExactly(dst, src)
@@ -706,19 +648,18 @@
 	out := dst
 	dstEvery := (dstSize + 3) / 4
 
-	shift := (8 - d.actualTableLog) & 7
+	shift := (56 + (8 - d.actualTableLog)) & 63
 
 	const tlSize = 1 << 8
-	const tlMask = tlSize - 1
 	single := d.dt.single[:tlSize]
 
 	// Use temp table to avoid bound checks/append penalty.
-	var buf [256]byte
+	buf := d.buffer()
 	var off uint8
 	var decoded int
 
 	// Decode 4 values from each decoder/loop.
-	const bufoff = 256 / 4
+	const bufoff = 256
 	for {
 		if br[0].off < 4 || br[1].off < 4 || br[2].off < 4 || br[3].off < 4 {
 			break
@@ -728,120 +669,144 @@
 			// Interleave 2 decodes.
 			const stream = 0
 			const stream2 = 1
-			br[stream].fillFast()
-			br[stream2].fillFast()
+			br1 := &br[stream]
+			br2 := &br[stream2]
+			br1.fillFast()
+			br2.fillFast()
 
-			v := single[br[stream].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
+			v := single[uint8(br1.value>>shift)].entry
+			v2 := single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off] = uint8(v >> 8)
+			buf[stream2][off] = uint8(v2 >> 8)
 
-			v2 := single[br[stream2].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream2] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
+			v = single[uint8(br1.value>>shift)].entry
+			v2 = single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off+1] = uint8(v >> 8)
+			buf[stream2][off+1] = uint8(v2 >> 8)
 
-			v = single[br[stream].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream+1] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
+			v = single[uint8(br1.value>>shift)].entry
+			v2 = single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off+2] = uint8(v >> 8)
+			buf[stream2][off+2] = uint8(v2 >> 8)
 
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream2+1] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
-
-			v = single[br[stream].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream+2] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
-
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream2+2] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
-
-			v = single[br[stream].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream+3] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
-
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream2+3] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
+			v = single[uint8(br1.value>>shift)].entry
+			v2 = single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off+3] = uint8(v >> 8)
+			buf[stream2][off+3] = uint8(v2 >> 8)
 		}
 
 		{
 			const stream = 2
 			const stream2 = 3
-			br[stream].fillFast()
-			br[stream2].fillFast()
+			br1 := &br[stream]
+			br2 := &br[stream2]
+			br1.fillFast()
+			br2.fillFast()
 
-			v := single[br[stream].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
+			v := single[uint8(br1.value>>shift)].entry
+			v2 := single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off] = uint8(v >> 8)
+			buf[stream2][off] = uint8(v2 >> 8)
 
-			v2 := single[br[stream2].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream2] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
+			v = single[uint8(br1.value>>shift)].entry
+			v2 = single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off+1] = uint8(v >> 8)
+			buf[stream2][off+1] = uint8(v2 >> 8)
 
-			v = single[br[stream].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream+1] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
+			v = single[uint8(br1.value>>shift)].entry
+			v2 = single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off+2] = uint8(v >> 8)
+			buf[stream2][off+2] = uint8(v2 >> 8)
 
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream2+1] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
-
-			v = single[br[stream].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream+2] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
-
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream2+2] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
-
-			v = single[br[stream].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream+3] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
-
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream2+3] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
+			v = single[uint8(br1.value>>shift)].entry
+			v2 = single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off+3] = uint8(v >> 8)
+			buf[stream2][off+3] = uint8(v2 >> 8)
 		}
 
 		off += 4
 
-		if off == bufoff {
+		if off == 0 {
 			if bufoff > dstEvery {
+				d.bufs.Put(buf)
 				return nil, errors.New("corruption detected: stream overrun 1")
 			}
-			copy(out, buf[:bufoff])
-			copy(out[dstEvery:], buf[bufoff:bufoff*2])
-			copy(out[dstEvery*2:], buf[bufoff*2:bufoff*3])
-			copy(out[dstEvery*3:], buf[bufoff*3:bufoff*4])
-			off = 0
-			out = out[bufoff:]
-			decoded += 256
 			// There must at least be 3 buffers left.
-			if len(out) < dstEvery*3 {
+			if len(out)-bufoff < dstEvery*3 {
+				d.bufs.Put(buf)
 				return nil, errors.New("corruption detected: stream overrun 2")
 			}
+			//copy(out, buf[0][:])
+			//copy(out[dstEvery:], buf[1][:])
+			//copy(out[dstEvery*2:], buf[2][:])
+			*(*[bufoff]byte)(out) = buf[0]
+			*(*[bufoff]byte)(out[dstEvery:]) = buf[1]
+			*(*[bufoff]byte)(out[dstEvery*2:]) = buf[2]
+			*(*[bufoff]byte)(out[dstEvery*3:]) = buf[3]
+			out = out[bufoff:]
+			decoded += bufoff * 4
 		}
 	}
 	if off > 0 {
 		ioff := int(off)
 		if len(out) < dstEvery*3+ioff {
+			d.bufs.Put(buf)
 			return nil, errors.New("corruption detected: stream overrun 3")
 		}
-		copy(out, buf[:off])
-		copy(out[dstEvery:dstEvery+ioff], buf[bufoff:bufoff*2])
-		copy(out[dstEvery*2:dstEvery*2+ioff], buf[bufoff*2:bufoff*3])
-		copy(out[dstEvery*3:dstEvery*3+ioff], buf[bufoff*3:bufoff*4])
+		copy(out, buf[0][:off])
+		copy(out[dstEvery:], buf[1][:off])
+		copy(out[dstEvery*2:], buf[2][:off])
+		copy(out[dstEvery*3:], buf[3][:off])
 		decoded += int(off) * 4
 		out = out[off:]
 	}
 
 	// Decode remaining.
+	// Decode remaining.
+	remainBytes := dstEvery - (decoded / 4)
 	for i := range br {
 		offset := dstEvery * i
+		endsAt := offset + remainBytes
+		if endsAt > len(out) {
+			endsAt = len(out)
+		}
 		br := &br[i]
-		bitsLeft := int(br.off*8) + int(64-br.bitsRead)
+		bitsLeft := br.remaining()
 		for bitsLeft > 0 {
 			if br.finished() {
+				d.bufs.Put(buf)
 				return nil, io.ErrUnexpectedEOF
 			}
 			if br.bitsRead >= 56 {
@@ -861,24 +826,31 @@
 				}
 			}
 			// end inline...
-			if offset >= len(out) {
+			if offset >= endsAt {
+				d.bufs.Put(buf)
 				return nil, errors.New("corruption detected: stream overrun 4")
 			}
 
 			// Read value and increment offset.
-			v := single[br.peekByteFast()>>shift].entry
+			v := single[uint8(br.value>>shift)].entry
 			nBits := uint8(v)
 			br.advance(nBits)
-			bitsLeft -= int(nBits)
+			bitsLeft -= uint(nBits)
 			out[offset] = uint8(v >> 8)
 			offset++
 		}
+		if offset != endsAt {
+			d.bufs.Put(buf)
+			return nil, fmt.Errorf("corruption detected: short output block %d, end %d != %d", i, offset, endsAt)
+		}
 		decoded += offset - dstEvery*i
 		err = br.close()
 		if err != nil {
+			d.bufs.Put(buf)
 			return nil, err
 		}
 	}
+	d.bufs.Put(buf)
 	if dstSize != decoded {
 		return nil, errors.New("corruption detected: short output block")
 	}
@@ -914,18 +886,17 @@
 	out := dst
 	dstEvery := (dstSize + 3) / 4
 
-	const shift = 0
+	const shift = 56
 	const tlSize = 1 << 8
-	const tlMask = tlSize - 1
 	single := d.dt.single[:tlSize]
 
 	// Use temp table to avoid bound checks/append penalty.
-	var buf [256]byte
+	buf := d.buffer()
 	var off uint8
 	var decoded int
 
 	// Decode 4 values from each decoder/loop.
-	const bufoff = 256 / 4
+	const bufoff = 256
 	for {
 		if br[0].off < 4 || br[1].off < 4 || br[2].off < 4 || br[3].off < 4 {
 			break
@@ -935,98 +906,116 @@
 			// Interleave 2 decodes.
 			const stream = 0
 			const stream2 = 1
-			br[stream].fillFast()
-			br[stream2].fillFast()
+			br1 := &br[stream]
+			br2 := &br[stream2]
+			br1.fillFast()
+			br2.fillFast()
 
-			v := single[br[stream].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
+			v := single[uint8(br1.value>>shift)].entry
+			v2 := single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off] = uint8(v >> 8)
+			buf[stream2][off] = uint8(v2 >> 8)
 
-			v2 := single[br[stream2].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream2] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
+			v = single[uint8(br1.value>>shift)].entry
+			v2 = single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off+1] = uint8(v >> 8)
+			buf[stream2][off+1] = uint8(v2 >> 8)
 
-			v = single[br[stream].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream+1] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
+			v = single[uint8(br1.value>>shift)].entry
+			v2 = single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off+2] = uint8(v >> 8)
+			buf[stream2][off+2] = uint8(v2 >> 8)
 
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream2+1] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
-
-			v = single[br[stream].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream+2] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
-
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream2+2] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
-
-			v = single[br[stream].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream+3] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
-
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream2+3] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
+			v = single[uint8(br1.value>>shift)].entry
+			v2 = single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off+3] = uint8(v >> 8)
+			buf[stream2][off+3] = uint8(v2 >> 8)
 		}
 
 		{
 			const stream = 2
 			const stream2 = 3
-			br[stream].fillFast()
-			br[stream2].fillFast()
+			br1 := &br[stream]
+			br2 := &br[stream2]
+			br1.fillFast()
+			br2.fillFast()
 
-			v := single[br[stream].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
+			v := single[uint8(br1.value>>shift)].entry
+			v2 := single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off] = uint8(v >> 8)
+			buf[stream2][off] = uint8(v2 >> 8)
 
-			v2 := single[br[stream2].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream2] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
+			v = single[uint8(br1.value>>shift)].entry
+			v2 = single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off+1] = uint8(v >> 8)
+			buf[stream2][off+1] = uint8(v2 >> 8)
 
-			v = single[br[stream].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream+1] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
+			v = single[uint8(br1.value>>shift)].entry
+			v2 = single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off+2] = uint8(v >> 8)
+			buf[stream2][off+2] = uint8(v2 >> 8)
 
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream2+1] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
-
-			v = single[br[stream].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream+2] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
-
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream2+2] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
-
-			v = single[br[stream].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream+3] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
-
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream2+3] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
+			v = single[uint8(br1.value>>shift)].entry
+			v2 = single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off+3] = uint8(v >> 8)
+			buf[stream2][off+3] = uint8(v2 >> 8)
 		}
 
 		off += 4
 
-		if off == bufoff {
+		if off == 0 {
 			if bufoff > dstEvery {
+				d.bufs.Put(buf)
 				return nil, errors.New("corruption detected: stream overrun 1")
 			}
-			copy(out, buf[:bufoff])
-			copy(out[dstEvery:], buf[bufoff:bufoff*2])
-			copy(out[dstEvery*2:], buf[bufoff*2:bufoff*3])
-			copy(out[dstEvery*3:], buf[bufoff*3:bufoff*4])
-			off = 0
-			out = out[bufoff:]
-			decoded += 256
 			// There must at least be 3 buffers left.
-			if len(out) < dstEvery*3 {
+			if len(out)-bufoff < dstEvery*3 {
+				d.bufs.Put(buf)
 				return nil, errors.New("corruption detected: stream overrun 2")
 			}
+
+			//copy(out, buf[0][:])
+			//copy(out[dstEvery:], buf[1][:])
+			//copy(out[dstEvery*2:], buf[2][:])
+			// copy(out[dstEvery*3:], buf[3][:])
+			*(*[bufoff]byte)(out) = buf[0]
+			*(*[bufoff]byte)(out[dstEvery:]) = buf[1]
+			*(*[bufoff]byte)(out[dstEvery*2:]) = buf[2]
+			*(*[bufoff]byte)(out[dstEvery*3:]) = buf[3]
+			out = out[bufoff:]
+			decoded += bufoff * 4
 		}
 	}
 	if off > 0 {
@@ -1034,21 +1023,27 @@
 		if len(out) < dstEvery*3+ioff {
 			return nil, errors.New("corruption detected: stream overrun 3")
 		}
-		copy(out, buf[:off])
-		copy(out[dstEvery:dstEvery+ioff], buf[bufoff:bufoff*2])
-		copy(out[dstEvery*2:dstEvery*2+ioff], buf[bufoff*2:bufoff*3])
-		copy(out[dstEvery*3:dstEvery*3+ioff], buf[bufoff*3:bufoff*4])
+		copy(out, buf[0][:off])
+		copy(out[dstEvery:], buf[1][:off])
+		copy(out[dstEvery*2:], buf[2][:off])
+		copy(out[dstEvery*3:], buf[3][:off])
 		decoded += int(off) * 4
 		out = out[off:]
 	}
 
 	// Decode remaining.
+	remainBytes := dstEvery - (decoded / 4)
 	for i := range br {
 		offset := dstEvery * i
+		endsAt := offset + remainBytes
+		if endsAt > len(out) {
+			endsAt = len(out)
+		}
 		br := &br[i]
-		bitsLeft := int(br.off*8) + int(64-br.bitsRead)
+		bitsLeft := br.remaining()
 		for bitsLeft > 0 {
 			if br.finished() {
+				d.bufs.Put(buf)
 				return nil, io.ErrUnexpectedEOF
 			}
 			if br.bitsRead >= 56 {
@@ -1068,24 +1063,32 @@
 				}
 			}
 			// end inline...
-			if offset >= len(out) {
+			if offset >= endsAt {
+				d.bufs.Put(buf)
 				return nil, errors.New("corruption detected: stream overrun 4")
 			}
 
 			// Read value and increment offset.
-			v := single[br.peekByteFast()>>shift].entry
+			v := single[br.peekByteFast()].entry
 			nBits := uint8(v)
 			br.advance(nBits)
-			bitsLeft -= int(nBits)
+			bitsLeft -= uint(nBits)
 			out[offset] = uint8(v >> 8)
 			offset++
 		}
+		if offset != endsAt {
+			d.bufs.Put(buf)
+			return nil, fmt.Errorf("corruption detected: short output block %d, end %d != %d", i, offset, endsAt)
+		}
+
 		decoded += offset - dstEvery*i
 		err = br.close()
 		if err != nil {
+			d.bufs.Put(buf)
 			return nil, err
 		}
 	}
+	d.bufs.Put(buf)
 	if dstSize != decoded {
 		return nil, errors.New("corruption detected: short output block")
 	}
@@ -1133,7 +1136,7 @@
 			errs++
 		}
 		if errs > 0 {
-			fmt.Fprintf(w, "%d errros in base, stopping\n", errs)
+			fmt.Fprintf(w, "%d errors in base, stopping\n", errs)
 			continue
 		}
 		// Ensure that all combinations are covered.
@@ -1149,7 +1152,7 @@
 				errs++
 			}
 			if errs > 20 {
-				fmt.Fprintf(w, "%d errros, stopping\n", errs)
+				fmt.Fprintf(w, "%d errors, stopping\n", errs)
 				break
 			}
 		}
diff --git a/vendor/github.com/klauspost/compress/huff0/decompress_amd64.go b/vendor/github.com/klauspost/compress/huff0/decompress_amd64.go
new file mode 100644
index 0000000..ba7e8e6
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/huff0/decompress_amd64.go
@@ -0,0 +1,226 @@
+//go:build amd64 && !appengine && !noasm && gc
+// +build amd64,!appengine,!noasm,gc
+
+// This file contains the specialisation of Decoder.Decompress4X
+// and Decoder.Decompress1X that use an asm implementation of thir main loops.
+package huff0
+
+import (
+	"errors"
+	"fmt"
+
+	"github.com/klauspost/compress/internal/cpuinfo"
+)
+
+// decompress4x_main_loop_x86 is an x86 assembler implementation
+// of Decompress4X when tablelog > 8.
+//
+//go:noescape
+func decompress4x_main_loop_amd64(ctx *decompress4xContext)
+
+// decompress4x_8b_loop_x86 is an x86 assembler implementation
+// of Decompress4X when tablelog <= 8 which decodes 4 entries
+// per loop.
+//
+//go:noescape
+func decompress4x_8b_main_loop_amd64(ctx *decompress4xContext)
+
+// fallback8BitSize is the size where using Go version is faster.
+const fallback8BitSize = 800
+
+type decompress4xContext struct {
+	pbr      *[4]bitReaderShifted
+	peekBits uint8
+	out      *byte
+	dstEvery int
+	tbl      *dEntrySingle
+	decoded  int
+	limit    *byte
+}
+
+// Decompress4X will decompress a 4X encoded stream.
+// The length of the supplied input must match the end of a block exactly.
+// The *capacity* of the dst slice must match the destination size of
+// the uncompressed data exactly.
+func (d *Decoder) Decompress4X(dst, src []byte) ([]byte, error) {
+	if len(d.dt.single) == 0 {
+		return nil, errors.New("no table loaded")
+	}
+	if len(src) < 6+(4*1) {
+		return nil, errors.New("input too small")
+	}
+
+	use8BitTables := d.actualTableLog <= 8
+	if cap(dst) < fallback8BitSize && use8BitTables {
+		return d.decompress4X8bit(dst, src)
+	}
+
+	var br [4]bitReaderShifted
+	// Decode "jump table"
+	start := 6
+	for i := 0; i < 3; i++ {
+		length := int(src[i*2]) | (int(src[i*2+1]) << 8)
+		if start+length >= len(src) {
+			return nil, errors.New("truncated input (or invalid offset)")
+		}
+		err := br[i].init(src[start : start+length])
+		if err != nil {
+			return nil, err
+		}
+		start += length
+	}
+	err := br[3].init(src[start:])
+	if err != nil {
+		return nil, err
+	}
+
+	// destination, offset to match first output
+	dstSize := cap(dst)
+	dst = dst[:dstSize]
+	out := dst
+	dstEvery := (dstSize + 3) / 4
+
+	const tlSize = 1 << tableLogMax
+	const tlMask = tlSize - 1
+	single := d.dt.single[:tlSize]
+
+	var decoded int
+
+	if len(out) > 4*4 && !(br[0].off < 4 || br[1].off < 4 || br[2].off < 4 || br[3].off < 4) {
+		ctx := decompress4xContext{
+			pbr:      &br,
+			peekBits: uint8((64 - d.actualTableLog) & 63), // see: bitReaderShifted.peekBitsFast()
+			out:      &out[0],
+			dstEvery: dstEvery,
+			tbl:      &single[0],
+			limit:    &out[dstEvery-4], // Always stop decoding when first buffer gets here to avoid writing OOB on last.
+		}
+		if use8BitTables {
+			decompress4x_8b_main_loop_amd64(&ctx)
+		} else {
+			decompress4x_main_loop_amd64(&ctx)
+		}
+
+		decoded = ctx.decoded
+		out = out[decoded/4:]
+	}
+
+	// Decode remaining.
+	remainBytes := dstEvery - (decoded / 4)
+	for i := range br {
+		offset := dstEvery * i
+		endsAt := offset + remainBytes
+		if endsAt > len(out) {
+			endsAt = len(out)
+		}
+		br := &br[i]
+		bitsLeft := br.remaining()
+		for bitsLeft > 0 {
+			br.fill()
+			if offset >= endsAt {
+				return nil, errors.New("corruption detected: stream overrun 4")
+			}
+
+			// Read value and increment offset.
+			val := br.peekBitsFast(d.actualTableLog)
+			v := single[val&tlMask].entry
+			nBits := uint8(v)
+			br.advance(nBits)
+			bitsLeft -= uint(nBits)
+			out[offset] = uint8(v >> 8)
+			offset++
+		}
+		if offset != endsAt {
+			return nil, fmt.Errorf("corruption detected: short output block %d, end %d != %d", i, offset, endsAt)
+		}
+		decoded += offset - dstEvery*i
+		err = br.close()
+		if err != nil {
+			return nil, err
+		}
+	}
+	if dstSize != decoded {
+		return nil, errors.New("corruption detected: short output block")
+	}
+	return dst, nil
+}
+
+// decompress4x_main_loop_x86 is an x86 assembler implementation
+// of Decompress1X when tablelog > 8.
+//
+//go:noescape
+func decompress1x_main_loop_amd64(ctx *decompress1xContext)
+
+// decompress4x_main_loop_x86 is an x86 with BMI2 assembler implementation
+// of Decompress1X when tablelog > 8.
+//
+//go:noescape
+func decompress1x_main_loop_bmi2(ctx *decompress1xContext)
+
+type decompress1xContext struct {
+	pbr      *bitReaderShifted
+	peekBits uint8
+	out      *byte
+	outCap   int
+	tbl      *dEntrySingle
+	decoded  int
+}
+
+// Error reported by asm implementations
+const error_max_decoded_size_exeeded = -1
+
+// Decompress1X will decompress a 1X encoded stream.
+// The cap of the output buffer will be the maximum decompressed size.
+// The length of the supplied input must match the end of a block exactly.
+func (d *Decoder) Decompress1X(dst, src []byte) ([]byte, error) {
+	if len(d.dt.single) == 0 {
+		return nil, errors.New("no table loaded")
+	}
+	var br bitReaderShifted
+	err := br.init(src)
+	if err != nil {
+		return dst, err
+	}
+	maxDecodedSize := cap(dst)
+	dst = dst[:maxDecodedSize]
+
+	const tlSize = 1 << tableLogMax
+	const tlMask = tlSize - 1
+
+	if maxDecodedSize >= 4 {
+		ctx := decompress1xContext{
+			pbr:      &br,
+			out:      &dst[0],
+			outCap:   maxDecodedSize,
+			peekBits: uint8((64 - d.actualTableLog) & 63), // see: bitReaderShifted.peekBitsFast()
+			tbl:      &d.dt.single[0],
+		}
+
+		if cpuinfo.HasBMI2() {
+			decompress1x_main_loop_bmi2(&ctx)
+		} else {
+			decompress1x_main_loop_amd64(&ctx)
+		}
+		if ctx.decoded == error_max_decoded_size_exeeded {
+			return nil, ErrMaxDecodedSizeExceeded
+		}
+
+		dst = dst[:ctx.decoded]
+	}
+
+	// br < 8, so uint8 is fine
+	bitsLeft := uint8(br.off)*8 + 64 - br.bitsRead
+	for bitsLeft > 0 {
+		br.fill()
+		if len(dst) >= maxDecodedSize {
+			br.close()
+			return nil, ErrMaxDecodedSizeExceeded
+		}
+		v := d.dt.single[br.peekBitsFast(d.actualTableLog)&tlMask]
+		nBits := uint8(v.entry)
+		br.advance(nBits)
+		bitsLeft -= nBits
+		dst = append(dst, uint8(v.entry>>8))
+	}
+	return dst, br.close()
+}
diff --git a/vendor/github.com/klauspost/compress/huff0/decompress_amd64.s b/vendor/github.com/klauspost/compress/huff0/decompress_amd64.s
new file mode 100644
index 0000000..c4c7ab2
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/huff0/decompress_amd64.s
@@ -0,0 +1,830 @@
+// Code generated by command: go run gen.go -out ../decompress_amd64.s -pkg=huff0. DO NOT EDIT.
+
+//go:build amd64 && !appengine && !noasm && gc
+
+// func decompress4x_main_loop_amd64(ctx *decompress4xContext)
+TEXT ·decompress4x_main_loop_amd64(SB), $0-8
+	// Preload values
+	MOVQ    ctx+0(FP), AX
+	MOVBQZX 8(AX), DI
+	MOVQ    16(AX), BX
+	MOVQ    48(AX), SI
+	MOVQ    24(AX), R8
+	MOVQ    32(AX), R9
+	MOVQ    (AX), R10
+
+	// Main loop
+main_loop:
+	XORL  DX, DX
+	CMPQ  BX, SI
+	SETGE DL
+
+	// br0.fillFast32()
+	MOVQ    32(R10), R11
+	MOVBQZX 40(R10), R12
+	CMPQ    R12, $0x20
+	JBE     skip_fill0
+	MOVQ    24(R10), AX
+	SUBQ    $0x20, R12
+	SUBQ    $0x04, AX
+	MOVQ    (R10), R13
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+	MOVL (AX)(R13*1), R13
+	MOVQ R12, CX
+	SHLQ CL, R13
+	MOVQ AX, 24(R10)
+	ORQ  R13, R11
+
+	// exhausted += (br0.off < 4)
+	CMPQ AX, $0x04
+	ADCB $+0, DL
+
+skip_fill0:
+	// val0 := br0.peekTopBits(peekBits)
+	MOVQ R11, R13
+	MOVQ DI, CX
+	SHRQ CL, R13
+
+	// v0 := table[val0&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br0.advance(uint8(v0.entry)
+	MOVB CH, AL
+	SHLQ CL, R11
+	ADDB CL, R12
+
+	// val1 := br0.peekTopBits(peekBits)
+	MOVQ DI, CX
+	MOVQ R11, R13
+	SHRQ CL, R13
+
+	// v1 := table[val1&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br0.advance(uint8(v1.entry))
+	MOVB CH, AH
+	SHLQ CL, R11
+	ADDB CL, R12
+
+	// these two writes get coalesced
+	// out[id * dstEvery + 0] = uint8(v0.entry >> 8)
+	// out[id * dstEvery + 1] = uint8(v1.entry >> 8)
+	MOVW AX, (BX)
+
+	// update the bitreader structure
+	MOVQ R11, 32(R10)
+	MOVB R12, 40(R10)
+
+	// br1.fillFast32()
+	MOVQ    80(R10), R11
+	MOVBQZX 88(R10), R12
+	CMPQ    R12, $0x20
+	JBE     skip_fill1
+	MOVQ    72(R10), AX
+	SUBQ    $0x20, R12
+	SUBQ    $0x04, AX
+	MOVQ    48(R10), R13
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+	MOVL (AX)(R13*1), R13
+	MOVQ R12, CX
+	SHLQ CL, R13
+	MOVQ AX, 72(R10)
+	ORQ  R13, R11
+
+	// exhausted += (br1.off < 4)
+	CMPQ AX, $0x04
+	ADCB $+0, DL
+
+skip_fill1:
+	// val0 := br1.peekTopBits(peekBits)
+	MOVQ R11, R13
+	MOVQ DI, CX
+	SHRQ CL, R13
+
+	// v0 := table[val0&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br1.advance(uint8(v0.entry)
+	MOVB CH, AL
+	SHLQ CL, R11
+	ADDB CL, R12
+
+	// val1 := br1.peekTopBits(peekBits)
+	MOVQ DI, CX
+	MOVQ R11, R13
+	SHRQ CL, R13
+
+	// v1 := table[val1&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br1.advance(uint8(v1.entry))
+	MOVB CH, AH
+	SHLQ CL, R11
+	ADDB CL, R12
+
+	// these two writes get coalesced
+	// out[id * dstEvery + 0] = uint8(v0.entry >> 8)
+	// out[id * dstEvery + 1] = uint8(v1.entry >> 8)
+	MOVW AX, (BX)(R8*1)
+
+	// update the bitreader structure
+	MOVQ R11, 80(R10)
+	MOVB R12, 88(R10)
+
+	// br2.fillFast32()
+	MOVQ    128(R10), R11
+	MOVBQZX 136(R10), R12
+	CMPQ    R12, $0x20
+	JBE     skip_fill2
+	MOVQ    120(R10), AX
+	SUBQ    $0x20, R12
+	SUBQ    $0x04, AX
+	MOVQ    96(R10), R13
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+	MOVL (AX)(R13*1), R13
+	MOVQ R12, CX
+	SHLQ CL, R13
+	MOVQ AX, 120(R10)
+	ORQ  R13, R11
+
+	// exhausted += (br2.off < 4)
+	CMPQ AX, $0x04
+	ADCB $+0, DL
+
+skip_fill2:
+	// val0 := br2.peekTopBits(peekBits)
+	MOVQ R11, R13
+	MOVQ DI, CX
+	SHRQ CL, R13
+
+	// v0 := table[val0&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br2.advance(uint8(v0.entry)
+	MOVB CH, AL
+	SHLQ CL, R11
+	ADDB CL, R12
+
+	// val1 := br2.peekTopBits(peekBits)
+	MOVQ DI, CX
+	MOVQ R11, R13
+	SHRQ CL, R13
+
+	// v1 := table[val1&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br2.advance(uint8(v1.entry))
+	MOVB CH, AH
+	SHLQ CL, R11
+	ADDB CL, R12
+
+	// these two writes get coalesced
+	// out[id * dstEvery + 0] = uint8(v0.entry >> 8)
+	// out[id * dstEvery + 1] = uint8(v1.entry >> 8)
+	MOVW AX, (BX)(R8*2)
+
+	// update the bitreader structure
+	MOVQ R11, 128(R10)
+	MOVB R12, 136(R10)
+
+	// br3.fillFast32()
+	MOVQ    176(R10), R11
+	MOVBQZX 184(R10), R12
+	CMPQ    R12, $0x20
+	JBE     skip_fill3
+	MOVQ    168(R10), AX
+	SUBQ    $0x20, R12
+	SUBQ    $0x04, AX
+	MOVQ    144(R10), R13
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+	MOVL (AX)(R13*1), R13
+	MOVQ R12, CX
+	SHLQ CL, R13
+	MOVQ AX, 168(R10)
+	ORQ  R13, R11
+
+	// exhausted += (br3.off < 4)
+	CMPQ AX, $0x04
+	ADCB $+0, DL
+
+skip_fill3:
+	// val0 := br3.peekTopBits(peekBits)
+	MOVQ R11, R13
+	MOVQ DI, CX
+	SHRQ CL, R13
+
+	// v0 := table[val0&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br3.advance(uint8(v0.entry)
+	MOVB CH, AL
+	SHLQ CL, R11
+	ADDB CL, R12
+
+	// val1 := br3.peekTopBits(peekBits)
+	MOVQ DI, CX
+	MOVQ R11, R13
+	SHRQ CL, R13
+
+	// v1 := table[val1&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br3.advance(uint8(v1.entry))
+	MOVB CH, AH
+	SHLQ CL, R11
+	ADDB CL, R12
+
+	// these two writes get coalesced
+	// out[id * dstEvery + 0] = uint8(v0.entry >> 8)
+	// out[id * dstEvery + 1] = uint8(v1.entry >> 8)
+	LEAQ (R8)(R8*2), CX
+	MOVW AX, (BX)(CX*1)
+
+	// update the bitreader structure
+	MOVQ  R11, 176(R10)
+	MOVB  R12, 184(R10)
+	ADDQ  $0x02, BX
+	TESTB DL, DL
+	JZ    main_loop
+	MOVQ  ctx+0(FP), AX
+	SUBQ  16(AX), BX
+	SHLQ  $0x02, BX
+	MOVQ  BX, 40(AX)
+	RET
+
+// func decompress4x_8b_main_loop_amd64(ctx *decompress4xContext)
+TEXT ·decompress4x_8b_main_loop_amd64(SB), $0-8
+	// Preload values
+	MOVQ    ctx+0(FP), CX
+	MOVBQZX 8(CX), DI
+	MOVQ    16(CX), BX
+	MOVQ    48(CX), SI
+	MOVQ    24(CX), R8
+	MOVQ    32(CX), R9
+	MOVQ    (CX), R10
+
+	// Main loop
+main_loop:
+	XORL  DX, DX
+	CMPQ  BX, SI
+	SETGE DL
+
+	// br0.fillFast32()
+	MOVQ    32(R10), R11
+	MOVBQZX 40(R10), R12
+	CMPQ    R12, $0x20
+	JBE     skip_fill0
+	MOVQ    24(R10), R13
+	SUBQ    $0x20, R12
+	SUBQ    $0x04, R13
+	MOVQ    (R10), R14
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+	MOVL (R13)(R14*1), R14
+	MOVQ R12, CX
+	SHLQ CL, R14
+	MOVQ R13, 24(R10)
+	ORQ  R14, R11
+
+	// exhausted += (br0.off < 4)
+	CMPQ R13, $0x04
+	ADCB $+0, DL
+
+skip_fill0:
+	// val0 := br0.peekTopBits(peekBits)
+	MOVQ R11, R13
+	MOVQ DI, CX
+	SHRQ CL, R13
+
+	// v0 := table[val0&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br0.advance(uint8(v0.entry)
+	MOVB CH, AL
+	SHLQ CL, R11
+	ADDB CL, R12
+
+	// val1 := br0.peekTopBits(peekBits)
+	MOVQ R11, R13
+	MOVQ DI, CX
+	SHRQ CL, R13
+
+	// v1 := table[val0&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br0.advance(uint8(v1.entry)
+	MOVB   CH, AH
+	SHLQ   CL, R11
+	ADDB   CL, R12
+	BSWAPL AX
+
+	// val2 := br0.peekTopBits(peekBits)
+	MOVQ R11, R13
+	MOVQ DI, CX
+	SHRQ CL, R13
+
+	// v2 := table[val0&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br0.advance(uint8(v2.entry)
+	MOVB CH, AH
+	SHLQ CL, R11
+	ADDB CL, R12
+
+	// val3 := br0.peekTopBits(peekBits)
+	MOVQ R11, R13
+	MOVQ DI, CX
+	SHRQ CL, R13
+
+	// v3 := table[val0&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br0.advance(uint8(v3.entry)
+	MOVB   CH, AL
+	SHLQ   CL, R11
+	ADDB   CL, R12
+	BSWAPL AX
+
+	// these four writes get coalesced
+	// out[id * dstEvery + 0] = uint8(v0.entry >> 8)
+	// out[id * dstEvery + 1] = uint8(v1.entry >> 8)
+	// out[id * dstEvery + 3] = uint8(v2.entry >> 8)
+	// out[id * dstEvery + 4] = uint8(v3.entry >> 8)
+	MOVL AX, (BX)
+
+	// update the bitreader structure
+	MOVQ R11, 32(R10)
+	MOVB R12, 40(R10)
+
+	// br1.fillFast32()
+	MOVQ    80(R10), R11
+	MOVBQZX 88(R10), R12
+	CMPQ    R12, $0x20
+	JBE     skip_fill1
+	MOVQ    72(R10), R13
+	SUBQ    $0x20, R12
+	SUBQ    $0x04, R13
+	MOVQ    48(R10), R14
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+	MOVL (R13)(R14*1), R14
+	MOVQ R12, CX
+	SHLQ CL, R14
+	MOVQ R13, 72(R10)
+	ORQ  R14, R11
+
+	// exhausted += (br1.off < 4)
+	CMPQ R13, $0x04
+	ADCB $+0, DL
+
+skip_fill1:
+	// val0 := br1.peekTopBits(peekBits)
+	MOVQ R11, R13
+	MOVQ DI, CX
+	SHRQ CL, R13
+
+	// v0 := table[val0&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br1.advance(uint8(v0.entry)
+	MOVB CH, AL
+	SHLQ CL, R11
+	ADDB CL, R12
+
+	// val1 := br1.peekTopBits(peekBits)
+	MOVQ R11, R13
+	MOVQ DI, CX
+	SHRQ CL, R13
+
+	// v1 := table[val0&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br1.advance(uint8(v1.entry)
+	MOVB   CH, AH
+	SHLQ   CL, R11
+	ADDB   CL, R12
+	BSWAPL AX
+
+	// val2 := br1.peekTopBits(peekBits)
+	MOVQ R11, R13
+	MOVQ DI, CX
+	SHRQ CL, R13
+
+	// v2 := table[val0&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br1.advance(uint8(v2.entry)
+	MOVB CH, AH
+	SHLQ CL, R11
+	ADDB CL, R12
+
+	// val3 := br1.peekTopBits(peekBits)
+	MOVQ R11, R13
+	MOVQ DI, CX
+	SHRQ CL, R13
+
+	// v3 := table[val0&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br1.advance(uint8(v3.entry)
+	MOVB   CH, AL
+	SHLQ   CL, R11
+	ADDB   CL, R12
+	BSWAPL AX
+
+	// these four writes get coalesced
+	// out[id * dstEvery + 0] = uint8(v0.entry >> 8)
+	// out[id * dstEvery + 1] = uint8(v1.entry >> 8)
+	// out[id * dstEvery + 3] = uint8(v2.entry >> 8)
+	// out[id * dstEvery + 4] = uint8(v3.entry >> 8)
+	MOVL AX, (BX)(R8*1)
+
+	// update the bitreader structure
+	MOVQ R11, 80(R10)
+	MOVB R12, 88(R10)
+
+	// br2.fillFast32()
+	MOVQ    128(R10), R11
+	MOVBQZX 136(R10), R12
+	CMPQ    R12, $0x20
+	JBE     skip_fill2
+	MOVQ    120(R10), R13
+	SUBQ    $0x20, R12
+	SUBQ    $0x04, R13
+	MOVQ    96(R10), R14
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+	MOVL (R13)(R14*1), R14
+	MOVQ R12, CX
+	SHLQ CL, R14
+	MOVQ R13, 120(R10)
+	ORQ  R14, R11
+
+	// exhausted += (br2.off < 4)
+	CMPQ R13, $0x04
+	ADCB $+0, DL
+
+skip_fill2:
+	// val0 := br2.peekTopBits(peekBits)
+	MOVQ R11, R13
+	MOVQ DI, CX
+	SHRQ CL, R13
+
+	// v0 := table[val0&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br2.advance(uint8(v0.entry)
+	MOVB CH, AL
+	SHLQ CL, R11
+	ADDB CL, R12
+
+	// val1 := br2.peekTopBits(peekBits)
+	MOVQ R11, R13
+	MOVQ DI, CX
+	SHRQ CL, R13
+
+	// v1 := table[val0&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br2.advance(uint8(v1.entry)
+	MOVB   CH, AH
+	SHLQ   CL, R11
+	ADDB   CL, R12
+	BSWAPL AX
+
+	// val2 := br2.peekTopBits(peekBits)
+	MOVQ R11, R13
+	MOVQ DI, CX
+	SHRQ CL, R13
+
+	// v2 := table[val0&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br2.advance(uint8(v2.entry)
+	MOVB CH, AH
+	SHLQ CL, R11
+	ADDB CL, R12
+
+	// val3 := br2.peekTopBits(peekBits)
+	MOVQ R11, R13
+	MOVQ DI, CX
+	SHRQ CL, R13
+
+	// v3 := table[val0&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br2.advance(uint8(v3.entry)
+	MOVB   CH, AL
+	SHLQ   CL, R11
+	ADDB   CL, R12
+	BSWAPL AX
+
+	// these four writes get coalesced
+	// out[id * dstEvery + 0] = uint8(v0.entry >> 8)
+	// out[id * dstEvery + 1] = uint8(v1.entry >> 8)
+	// out[id * dstEvery + 3] = uint8(v2.entry >> 8)
+	// out[id * dstEvery + 4] = uint8(v3.entry >> 8)
+	MOVL AX, (BX)(R8*2)
+
+	// update the bitreader structure
+	MOVQ R11, 128(R10)
+	MOVB R12, 136(R10)
+
+	// br3.fillFast32()
+	MOVQ    176(R10), R11
+	MOVBQZX 184(R10), R12
+	CMPQ    R12, $0x20
+	JBE     skip_fill3
+	MOVQ    168(R10), R13
+	SUBQ    $0x20, R12
+	SUBQ    $0x04, R13
+	MOVQ    144(R10), R14
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+	MOVL (R13)(R14*1), R14
+	MOVQ R12, CX
+	SHLQ CL, R14
+	MOVQ R13, 168(R10)
+	ORQ  R14, R11
+
+	// exhausted += (br3.off < 4)
+	CMPQ R13, $0x04
+	ADCB $+0, DL
+
+skip_fill3:
+	// val0 := br3.peekTopBits(peekBits)
+	MOVQ R11, R13
+	MOVQ DI, CX
+	SHRQ CL, R13
+
+	// v0 := table[val0&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br3.advance(uint8(v0.entry)
+	MOVB CH, AL
+	SHLQ CL, R11
+	ADDB CL, R12
+
+	// val1 := br3.peekTopBits(peekBits)
+	MOVQ R11, R13
+	MOVQ DI, CX
+	SHRQ CL, R13
+
+	// v1 := table[val0&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br3.advance(uint8(v1.entry)
+	MOVB   CH, AH
+	SHLQ   CL, R11
+	ADDB   CL, R12
+	BSWAPL AX
+
+	// val2 := br3.peekTopBits(peekBits)
+	MOVQ R11, R13
+	MOVQ DI, CX
+	SHRQ CL, R13
+
+	// v2 := table[val0&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br3.advance(uint8(v2.entry)
+	MOVB CH, AH
+	SHLQ CL, R11
+	ADDB CL, R12
+
+	// val3 := br3.peekTopBits(peekBits)
+	MOVQ R11, R13
+	MOVQ DI, CX
+	SHRQ CL, R13
+
+	// v3 := table[val0&mask]
+	MOVW (R9)(R13*2), CX
+
+	// br3.advance(uint8(v3.entry)
+	MOVB   CH, AL
+	SHLQ   CL, R11
+	ADDB   CL, R12
+	BSWAPL AX
+
+	// these four writes get coalesced
+	// out[id * dstEvery + 0] = uint8(v0.entry >> 8)
+	// out[id * dstEvery + 1] = uint8(v1.entry >> 8)
+	// out[id * dstEvery + 3] = uint8(v2.entry >> 8)
+	// out[id * dstEvery + 4] = uint8(v3.entry >> 8)
+	LEAQ (R8)(R8*2), CX
+	MOVL AX, (BX)(CX*1)
+
+	// update the bitreader structure
+	MOVQ  R11, 176(R10)
+	MOVB  R12, 184(R10)
+	ADDQ  $0x04, BX
+	TESTB DL, DL
+	JZ    main_loop
+	MOVQ  ctx+0(FP), AX
+	SUBQ  16(AX), BX
+	SHLQ  $0x02, BX
+	MOVQ  BX, 40(AX)
+	RET
+
+// func decompress1x_main_loop_amd64(ctx *decompress1xContext)
+TEXT ·decompress1x_main_loop_amd64(SB), $0-8
+	MOVQ    ctx+0(FP), CX
+	MOVQ    16(CX), DX
+	MOVQ    24(CX), BX
+	CMPQ    BX, $0x04
+	JB      error_max_decoded_size_exceeded
+	LEAQ    (DX)(BX*1), BX
+	MOVQ    (CX), SI
+	MOVQ    (SI), R8
+	MOVQ    24(SI), R9
+	MOVQ    32(SI), R10
+	MOVBQZX 40(SI), R11
+	MOVQ    32(CX), SI
+	MOVBQZX 8(CX), DI
+	JMP     loop_condition
+
+main_loop:
+	// Check if we have room for 4 bytes in the output buffer
+	LEAQ 4(DX), CX
+	CMPQ CX, BX
+	JGE  error_max_decoded_size_exceeded
+
+	// Decode 4 values
+	CMPQ R11, $0x20
+	JL   bitReader_fillFast_1_end
+	SUBQ $0x20, R11
+	SUBQ $0x04, R9
+	MOVL (R8)(R9*1), R12
+	MOVQ R11, CX
+	SHLQ CL, R12
+	ORQ  R12, R10
+
+bitReader_fillFast_1_end:
+	MOVQ    DI, CX
+	MOVQ    R10, R12
+	SHRQ    CL, R12
+	MOVW    (SI)(R12*2), CX
+	MOVB    CH, AL
+	MOVBQZX CL, CX
+	ADDQ    CX, R11
+	SHLQ    CL, R10
+	MOVQ    DI, CX
+	MOVQ    R10, R12
+	SHRQ    CL, R12
+	MOVW    (SI)(R12*2), CX
+	MOVB    CH, AH
+	MOVBQZX CL, CX
+	ADDQ    CX, R11
+	SHLQ    CL, R10
+	BSWAPL  AX
+	CMPQ    R11, $0x20
+	JL      bitReader_fillFast_2_end
+	SUBQ    $0x20, R11
+	SUBQ    $0x04, R9
+	MOVL    (R8)(R9*1), R12
+	MOVQ    R11, CX
+	SHLQ    CL, R12
+	ORQ     R12, R10
+
+bitReader_fillFast_2_end:
+	MOVQ    DI, CX
+	MOVQ    R10, R12
+	SHRQ    CL, R12
+	MOVW    (SI)(R12*2), CX
+	MOVB    CH, AH
+	MOVBQZX CL, CX
+	ADDQ    CX, R11
+	SHLQ    CL, R10
+	MOVQ    DI, CX
+	MOVQ    R10, R12
+	SHRQ    CL, R12
+	MOVW    (SI)(R12*2), CX
+	MOVB    CH, AL
+	MOVBQZX CL, CX
+	ADDQ    CX, R11
+	SHLQ    CL, R10
+	BSWAPL  AX
+
+	// Store the decoded values
+	MOVL AX, (DX)
+	ADDQ $0x04, DX
+
+loop_condition:
+	CMPQ R9, $0x08
+	JGE  main_loop
+
+	// Update ctx structure
+	MOVQ ctx+0(FP), AX
+	SUBQ 16(AX), DX
+	MOVQ DX, 40(AX)
+	MOVQ (AX), AX
+	MOVQ R9, 24(AX)
+	MOVQ R10, 32(AX)
+	MOVB R11, 40(AX)
+	RET
+
+	// Report error
+error_max_decoded_size_exceeded:
+	MOVQ ctx+0(FP), AX
+	MOVQ $-1, CX
+	MOVQ CX, 40(AX)
+	RET
+
+// func decompress1x_main_loop_bmi2(ctx *decompress1xContext)
+// Requires: BMI2
+TEXT ·decompress1x_main_loop_bmi2(SB), $0-8
+	MOVQ    ctx+0(FP), CX
+	MOVQ    16(CX), DX
+	MOVQ    24(CX), BX
+	CMPQ    BX, $0x04
+	JB      error_max_decoded_size_exceeded
+	LEAQ    (DX)(BX*1), BX
+	MOVQ    (CX), SI
+	MOVQ    (SI), R8
+	MOVQ    24(SI), R9
+	MOVQ    32(SI), R10
+	MOVBQZX 40(SI), R11
+	MOVQ    32(CX), SI
+	MOVBQZX 8(CX), DI
+	JMP     loop_condition
+
+main_loop:
+	// Check if we have room for 4 bytes in the output buffer
+	LEAQ 4(DX), CX
+	CMPQ CX, BX
+	JGE  error_max_decoded_size_exceeded
+
+	// Decode 4 values
+	CMPQ  R11, $0x20
+	JL    bitReader_fillFast_1_end
+	SUBQ  $0x20, R11
+	SUBQ  $0x04, R9
+	MOVL  (R8)(R9*1), CX
+	SHLXQ R11, CX, CX
+	ORQ   CX, R10
+
+bitReader_fillFast_1_end:
+	SHRXQ   DI, R10, CX
+	MOVW    (SI)(CX*2), CX
+	MOVB    CH, AL
+	MOVBQZX CL, CX
+	ADDQ    CX, R11
+	SHLXQ   CX, R10, R10
+	SHRXQ   DI, R10, CX
+	MOVW    (SI)(CX*2), CX
+	MOVB    CH, AH
+	MOVBQZX CL, CX
+	ADDQ    CX, R11
+	SHLXQ   CX, R10, R10
+	BSWAPL  AX
+	CMPQ    R11, $0x20
+	JL      bitReader_fillFast_2_end
+	SUBQ    $0x20, R11
+	SUBQ    $0x04, R9
+	MOVL    (R8)(R9*1), CX
+	SHLXQ   R11, CX, CX
+	ORQ     CX, R10
+
+bitReader_fillFast_2_end:
+	SHRXQ   DI, R10, CX
+	MOVW    (SI)(CX*2), CX
+	MOVB    CH, AH
+	MOVBQZX CL, CX
+	ADDQ    CX, R11
+	SHLXQ   CX, R10, R10
+	SHRXQ   DI, R10, CX
+	MOVW    (SI)(CX*2), CX
+	MOVB    CH, AL
+	MOVBQZX CL, CX
+	ADDQ    CX, R11
+	SHLXQ   CX, R10, R10
+	BSWAPL  AX
+
+	// Store the decoded values
+	MOVL AX, (DX)
+	ADDQ $0x04, DX
+
+loop_condition:
+	CMPQ R9, $0x08
+	JGE  main_loop
+
+	// Update ctx structure
+	MOVQ ctx+0(FP), AX
+	SUBQ 16(AX), DX
+	MOVQ DX, 40(AX)
+	MOVQ (AX), AX
+	MOVQ R9, 24(AX)
+	MOVQ R10, 32(AX)
+	MOVB R11, 40(AX)
+	RET
+
+	// Report error
+error_max_decoded_size_exceeded:
+	MOVQ ctx+0(FP), AX
+	MOVQ $-1, CX
+	MOVQ CX, 40(AX)
+	RET
diff --git a/vendor/github.com/klauspost/compress/huff0/decompress_generic.go b/vendor/github.com/klauspost/compress/huff0/decompress_generic.go
new file mode 100644
index 0000000..908c17d
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/huff0/decompress_generic.go
@@ -0,0 +1,299 @@
+//go:build !amd64 || appengine || !gc || noasm
+// +build !amd64 appengine !gc noasm
+
+// This file contains a generic implementation of Decoder.Decompress4X.
+package huff0
+
+import (
+	"errors"
+	"fmt"
+)
+
+// Decompress4X will decompress a 4X encoded stream.
+// The length of the supplied input must match the end of a block exactly.
+// The *capacity* of the dst slice must match the destination size of
+// the uncompressed data exactly.
+func (d *Decoder) Decompress4X(dst, src []byte) ([]byte, error) {
+	if len(d.dt.single) == 0 {
+		return nil, errors.New("no table loaded")
+	}
+	if len(src) < 6+(4*1) {
+		return nil, errors.New("input too small")
+	}
+	if use8BitTables && d.actualTableLog <= 8 {
+		return d.decompress4X8bit(dst, src)
+	}
+
+	var br [4]bitReaderShifted
+	// Decode "jump table"
+	start := 6
+	for i := 0; i < 3; i++ {
+		length := int(src[i*2]) | (int(src[i*2+1]) << 8)
+		if start+length >= len(src) {
+			return nil, errors.New("truncated input (or invalid offset)")
+		}
+		err := br[i].init(src[start : start+length])
+		if err != nil {
+			return nil, err
+		}
+		start += length
+	}
+	err := br[3].init(src[start:])
+	if err != nil {
+		return nil, err
+	}
+
+	// destination, offset to match first output
+	dstSize := cap(dst)
+	dst = dst[:dstSize]
+	out := dst
+	dstEvery := (dstSize + 3) / 4
+
+	const tlSize = 1 << tableLogMax
+	const tlMask = tlSize - 1
+	single := d.dt.single[:tlSize]
+
+	// Use temp table to avoid bound checks/append penalty.
+	buf := d.buffer()
+	var off uint8
+	var decoded int
+
+	// Decode 2 values from each decoder/loop.
+	const bufoff = 256
+	for {
+		if br[0].off < 4 || br[1].off < 4 || br[2].off < 4 || br[3].off < 4 {
+			break
+		}
+
+		{
+			const stream = 0
+			const stream2 = 1
+			br[stream].fillFast()
+			br[stream2].fillFast()
+
+			val := br[stream].peekBitsFast(d.actualTableLog)
+			val2 := br[stream2].peekBitsFast(d.actualTableLog)
+			v := single[val&tlMask]
+			v2 := single[val2&tlMask]
+			br[stream].advance(uint8(v.entry))
+			br[stream2].advance(uint8(v2.entry))
+			buf[stream][off] = uint8(v.entry >> 8)
+			buf[stream2][off] = uint8(v2.entry >> 8)
+
+			val = br[stream].peekBitsFast(d.actualTableLog)
+			val2 = br[stream2].peekBitsFast(d.actualTableLog)
+			v = single[val&tlMask]
+			v2 = single[val2&tlMask]
+			br[stream].advance(uint8(v.entry))
+			br[stream2].advance(uint8(v2.entry))
+			buf[stream][off+1] = uint8(v.entry >> 8)
+			buf[stream2][off+1] = uint8(v2.entry >> 8)
+		}
+
+		{
+			const stream = 2
+			const stream2 = 3
+			br[stream].fillFast()
+			br[stream2].fillFast()
+
+			val := br[stream].peekBitsFast(d.actualTableLog)
+			val2 := br[stream2].peekBitsFast(d.actualTableLog)
+			v := single[val&tlMask]
+			v2 := single[val2&tlMask]
+			br[stream].advance(uint8(v.entry))
+			br[stream2].advance(uint8(v2.entry))
+			buf[stream][off] = uint8(v.entry >> 8)
+			buf[stream2][off] = uint8(v2.entry >> 8)
+
+			val = br[stream].peekBitsFast(d.actualTableLog)
+			val2 = br[stream2].peekBitsFast(d.actualTableLog)
+			v = single[val&tlMask]
+			v2 = single[val2&tlMask]
+			br[stream].advance(uint8(v.entry))
+			br[stream2].advance(uint8(v2.entry))
+			buf[stream][off+1] = uint8(v.entry >> 8)
+			buf[stream2][off+1] = uint8(v2.entry >> 8)
+		}
+
+		off += 2
+
+		if off == 0 {
+			if bufoff > dstEvery {
+				d.bufs.Put(buf)
+				return nil, errors.New("corruption detected: stream overrun 1")
+			}
+			// There must at least be 3 buffers left.
+			if len(out)-bufoff < dstEvery*3 {
+				d.bufs.Put(buf)
+				return nil, errors.New("corruption detected: stream overrun 2")
+			}
+			//copy(out, buf[0][:])
+			//copy(out[dstEvery:], buf[1][:])
+			//copy(out[dstEvery*2:], buf[2][:])
+			//copy(out[dstEvery*3:], buf[3][:])
+			*(*[bufoff]byte)(out) = buf[0]
+			*(*[bufoff]byte)(out[dstEvery:]) = buf[1]
+			*(*[bufoff]byte)(out[dstEvery*2:]) = buf[2]
+			*(*[bufoff]byte)(out[dstEvery*3:]) = buf[3]
+			out = out[bufoff:]
+			decoded += bufoff * 4
+		}
+	}
+	if off > 0 {
+		ioff := int(off)
+		if len(out) < dstEvery*3+ioff {
+			d.bufs.Put(buf)
+			return nil, errors.New("corruption detected: stream overrun 3")
+		}
+		copy(out, buf[0][:off])
+		copy(out[dstEvery:], buf[1][:off])
+		copy(out[dstEvery*2:], buf[2][:off])
+		copy(out[dstEvery*3:], buf[3][:off])
+		decoded += int(off) * 4
+		out = out[off:]
+	}
+
+	// Decode remaining.
+	remainBytes := dstEvery - (decoded / 4)
+	for i := range br {
+		offset := dstEvery * i
+		endsAt := offset + remainBytes
+		if endsAt > len(out) {
+			endsAt = len(out)
+		}
+		br := &br[i]
+		bitsLeft := br.remaining()
+		for bitsLeft > 0 {
+			br.fill()
+			if offset >= endsAt {
+				d.bufs.Put(buf)
+				return nil, errors.New("corruption detected: stream overrun 4")
+			}
+
+			// Read value and increment offset.
+			val := br.peekBitsFast(d.actualTableLog)
+			v := single[val&tlMask].entry
+			nBits := uint8(v)
+			br.advance(nBits)
+			bitsLeft -= uint(nBits)
+			out[offset] = uint8(v >> 8)
+			offset++
+		}
+		if offset != endsAt {
+			d.bufs.Put(buf)
+			return nil, fmt.Errorf("corruption detected: short output block %d, end %d != %d", i, offset, endsAt)
+		}
+		decoded += offset - dstEvery*i
+		err = br.close()
+		if err != nil {
+			return nil, err
+		}
+	}
+	d.bufs.Put(buf)
+	if dstSize != decoded {
+		return nil, errors.New("corruption detected: short output block")
+	}
+	return dst, nil
+}
+
+// Decompress1X will decompress a 1X encoded stream.
+// The cap of the output buffer will be the maximum decompressed size.
+// The length of the supplied input must match the end of a block exactly.
+func (d *Decoder) Decompress1X(dst, src []byte) ([]byte, error) {
+	if len(d.dt.single) == 0 {
+		return nil, errors.New("no table loaded")
+	}
+	if use8BitTables && d.actualTableLog <= 8 {
+		return d.decompress1X8Bit(dst, src)
+	}
+	var br bitReaderShifted
+	err := br.init(src)
+	if err != nil {
+		return dst, err
+	}
+	maxDecodedSize := cap(dst)
+	dst = dst[:0]
+
+	// Avoid bounds check by always having full sized table.
+	const tlSize = 1 << tableLogMax
+	const tlMask = tlSize - 1
+	dt := d.dt.single[:tlSize]
+
+	// Use temp table to avoid bound checks/append penalty.
+	bufs := d.buffer()
+	buf := &bufs[0]
+	var off uint8
+
+	for br.off >= 8 {
+		br.fillFast()
+		v := dt[br.peekBitsFast(d.actualTableLog)&tlMask]
+		br.advance(uint8(v.entry))
+		buf[off+0] = uint8(v.entry >> 8)
+
+		v = dt[br.peekBitsFast(d.actualTableLog)&tlMask]
+		br.advance(uint8(v.entry))
+		buf[off+1] = uint8(v.entry >> 8)
+
+		// Refill
+		br.fillFast()
+
+		v = dt[br.peekBitsFast(d.actualTableLog)&tlMask]
+		br.advance(uint8(v.entry))
+		buf[off+2] = uint8(v.entry >> 8)
+
+		v = dt[br.peekBitsFast(d.actualTableLog)&tlMask]
+		br.advance(uint8(v.entry))
+		buf[off+3] = uint8(v.entry >> 8)
+
+		off += 4
+		if off == 0 {
+			if len(dst)+256 > maxDecodedSize {
+				br.close()
+				d.bufs.Put(bufs)
+				return nil, ErrMaxDecodedSizeExceeded
+			}
+			dst = append(dst, buf[:]...)
+		}
+	}
+
+	if len(dst)+int(off) > maxDecodedSize {
+		d.bufs.Put(bufs)
+		br.close()
+		return nil, ErrMaxDecodedSizeExceeded
+	}
+	dst = append(dst, buf[:off]...)
+
+	// br < 8, so uint8 is fine
+	bitsLeft := uint8(br.off)*8 + 64 - br.bitsRead
+	for bitsLeft > 0 {
+		br.fill()
+		if false && br.bitsRead >= 32 {
+			if br.off >= 4 {
+				v := br.in[br.off-4:]
+				v = v[:4]
+				low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+				br.value = (br.value << 32) | uint64(low)
+				br.bitsRead -= 32
+				br.off -= 4
+			} else {
+				for br.off > 0 {
+					br.value = (br.value << 8) | uint64(br.in[br.off-1])
+					br.bitsRead -= 8
+					br.off--
+				}
+			}
+		}
+		if len(dst) >= maxDecodedSize {
+			d.bufs.Put(bufs)
+			br.close()
+			return nil, ErrMaxDecodedSizeExceeded
+		}
+		v := d.dt.single[br.peekBitsFast(d.actualTableLog)&tlMask]
+		nBits := uint8(v.entry)
+		br.advance(nBits)
+		bitsLeft -= nBits
+		dst = append(dst, uint8(v.entry>>8))
+	}
+	d.bufs.Put(bufs)
+	return dst, br.close()
+}
diff --git a/vendor/github.com/klauspost/compress/huff0/huff0.go b/vendor/github.com/klauspost/compress/huff0/huff0.go
index 7ec2022..77ecd68 100644
--- a/vendor/github.com/klauspost/compress/huff0/huff0.go
+++ b/vendor/github.com/klauspost/compress/huff0/huff0.go
@@ -8,6 +8,7 @@
 	"fmt"
 	"math"
 	"math/bits"
+	"sync"
 
 	"github.com/klauspost/compress/fse"
 )
@@ -87,7 +88,7 @@
 	// Decoders will return ErrMaxDecodedSizeExceeded is this limit is exceeded.
 	MaxDecodedSize int
 
-	br byteReader
+	srcLen int
 
 	// MaxSymbolValue will override the maximum symbol value of the next block.
 	MaxSymbolValue uint8
@@ -116,6 +117,7 @@
 	nodes          []nodeElt
 	tmpOut         [4][]byte
 	fse            *fse.Scratch
+	decPool        sync.Pool // *[4][256]byte buffers.
 	huffWeight     [maxSymbolValue + 1]byte
 }
 
@@ -168,7 +170,7 @@
 	if s.fse == nil {
 		s.fse = &fse.Scratch{}
 	}
-	s.br.init(in)
+	s.srcLen = len(in)
 
 	return s, nil
 }
@@ -245,6 +247,68 @@
 	return nil
 }
 
+func (c cTable) estTableSize(s *Scratch) (sz int, err error) {
+	var (
+		// precomputed conversion table
+		bitsToWeight [tableLogMax + 1]byte
+		huffLog      = s.actualTableLog
+		// last weight is not saved.
+		maxSymbolValue = uint8(s.symbolLen - 1)
+		huffWeight     = s.huffWeight[:256]
+	)
+	const (
+		maxFSETableLog = 6
+	)
+	// convert to weight
+	bitsToWeight[0] = 0
+	for n := uint8(1); n < huffLog+1; n++ {
+		bitsToWeight[n] = huffLog + 1 - n
+	}
+
+	// Acquire histogram for FSE.
+	hist := s.fse.Histogram()
+	hist = hist[:256]
+	for i := range hist[:16] {
+		hist[i] = 0
+	}
+	for n := uint8(0); n < maxSymbolValue; n++ {
+		v := bitsToWeight[c[n].nBits] & 15
+		huffWeight[n] = v
+		hist[v]++
+	}
+
+	// FSE compress if feasible.
+	if maxSymbolValue >= 2 {
+		huffMaxCnt := uint32(0)
+		huffMax := uint8(0)
+		for i, v := range hist[:16] {
+			if v == 0 {
+				continue
+			}
+			huffMax = byte(i)
+			if v > huffMaxCnt {
+				huffMaxCnt = v
+			}
+		}
+		s.fse.HistogramFinished(huffMax, int(huffMaxCnt))
+		s.fse.TableLog = maxFSETableLog
+		b, err := fse.Compress(huffWeight[:maxSymbolValue], s.fse)
+		if err == nil && len(b) < int(s.symbolLen>>1) {
+			sz += 1 + len(b)
+			return sz, nil
+		}
+		// Unable to compress (RLE/uncompressible)
+	}
+	// write raw values as 4-bits (max : 15)
+	if maxSymbolValue > (256 - 128) {
+		// should not happen : likely means source cannot be compressed
+		return 0, ErrIncompressible
+	}
+	// special case, pack weights 4 bits/weight.
+	sz += 1 + int(maxSymbolValue/2)
+	return sz, nil
+}
+
 // estimateSize returns the estimated size in bytes of the input represented in the
 // histogram supplied.
 func (c cTable) estimateSize(hist []uint32) int {
diff --git a/vendor/github.com/klauspost/compress/internal/cpuinfo/cpuinfo.go b/vendor/github.com/klauspost/compress/internal/cpuinfo/cpuinfo.go
new file mode 100644
index 0000000..3954c51
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/internal/cpuinfo/cpuinfo.go
@@ -0,0 +1,34 @@
+// Package cpuinfo gives runtime info about the current CPU.
+//
+// This is a very limited module meant for use internally
+// in this project. For more versatile solution check
+// https://github.com/klauspost/cpuid.
+package cpuinfo
+
+// HasBMI1 checks whether an x86 CPU supports the BMI1 extension.
+func HasBMI1() bool {
+	return hasBMI1
+}
+
+// HasBMI2 checks whether an x86 CPU supports the BMI2 extension.
+func HasBMI2() bool {
+	return hasBMI2
+}
+
+// DisableBMI2 will disable BMI2, for testing purposes.
+// Call returned function to restore previous state.
+func DisableBMI2() func() {
+	old := hasBMI2
+	hasBMI2 = false
+	return func() {
+		hasBMI2 = old
+	}
+}
+
+// HasBMI checks whether an x86 CPU supports both BMI1 and BMI2 extensions.
+func HasBMI() bool {
+	return HasBMI1() && HasBMI2()
+}
+
+var hasBMI1 bool
+var hasBMI2 bool
diff --git a/vendor/github.com/klauspost/compress/internal/cpuinfo/cpuinfo_amd64.go b/vendor/github.com/klauspost/compress/internal/cpuinfo/cpuinfo_amd64.go
new file mode 100644
index 0000000..e802579
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/internal/cpuinfo/cpuinfo_amd64.go
@@ -0,0 +1,11 @@
+//go:build amd64 && !appengine && !noasm && gc
+// +build amd64,!appengine,!noasm,gc
+
+package cpuinfo
+
+// go:noescape
+func x86extensions() (bmi1, bmi2 bool)
+
+func init() {
+	hasBMI1, hasBMI2 = x86extensions()
+}
diff --git a/vendor/github.com/klauspost/compress/internal/cpuinfo/cpuinfo_amd64.s b/vendor/github.com/klauspost/compress/internal/cpuinfo/cpuinfo_amd64.s
new file mode 100644
index 0000000..4465fbe
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/internal/cpuinfo/cpuinfo_amd64.s
@@ -0,0 +1,36 @@
+// +build !appengine
+// +build gc
+// +build !noasm
+
+#include "textflag.h"
+#include "funcdata.h"
+#include "go_asm.h"
+
+TEXT ·x86extensions(SB), NOSPLIT, $0
+	// 1. determine max EAX value
+	XORQ AX, AX
+	CPUID
+
+	CMPQ AX, $7
+	JB   unsupported
+
+	// 2. EAX = 7, ECX = 0 --- see Table 3-8 "Information Returned by CPUID Instruction"
+	MOVQ $7, AX
+	MOVQ $0, CX
+	CPUID
+
+	BTQ   $3, BX // bit 3 = BMI1
+	SETCS AL
+
+	BTQ   $8, BX // bit 8 = BMI2
+	SETCS AH
+
+	MOVB AL, bmi1+0(FP)
+	MOVB AH, bmi2+1(FP)
+	RET
+
+unsupported:
+	XORQ AX, AX
+	MOVB AL, bmi1+0(FP)
+	MOVB AL, bmi2+1(FP)
+	RET
diff --git a/vendor/github.com/klauspost/compress/internal/le/le.go b/vendor/github.com/klauspost/compress/internal/le/le.go
new file mode 100644
index 0000000..e54909e
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/internal/le/le.go
@@ -0,0 +1,5 @@
+package le
+
+type Indexer interface {
+	int | int8 | int16 | int32 | int64 | uint | uint8 | uint16 | uint32 | uint64
+}
diff --git a/vendor/github.com/klauspost/compress/internal/le/unsafe_disabled.go b/vendor/github.com/klauspost/compress/internal/le/unsafe_disabled.go
new file mode 100644
index 0000000..0cfb5c0
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/internal/le/unsafe_disabled.go
@@ -0,0 +1,42 @@
+//go:build !(amd64 || arm64 || ppc64le || riscv64) || nounsafe || purego || appengine
+
+package le
+
+import (
+	"encoding/binary"
+)
+
+// Load8 will load from b at index i.
+func Load8[I Indexer](b []byte, i I) byte {
+	return b[i]
+}
+
+// Load16 will load from b at index i.
+func Load16[I Indexer](b []byte, i I) uint16 {
+	return binary.LittleEndian.Uint16(b[i:])
+}
+
+// Load32 will load from b at index i.
+func Load32[I Indexer](b []byte, i I) uint32 {
+	return binary.LittleEndian.Uint32(b[i:])
+}
+
+// Load64 will load from b at index i.
+func Load64[I Indexer](b []byte, i I) uint64 {
+	return binary.LittleEndian.Uint64(b[i:])
+}
+
+// Store16 will store v at b.
+func Store16(b []byte, v uint16) {
+	binary.LittleEndian.PutUint16(b, v)
+}
+
+// Store32 will store v at b.
+func Store32(b []byte, v uint32) {
+	binary.LittleEndian.PutUint32(b, v)
+}
+
+// Store64 will store v at b.
+func Store64(b []byte, v uint64) {
+	binary.LittleEndian.PutUint64(b, v)
+}
diff --git a/vendor/github.com/klauspost/compress/internal/le/unsafe_enabled.go b/vendor/github.com/klauspost/compress/internal/le/unsafe_enabled.go
new file mode 100644
index 0000000..ada45cd
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/internal/le/unsafe_enabled.go
@@ -0,0 +1,55 @@
+// We enable 64 bit LE platforms:
+
+//go:build (amd64 || arm64 || ppc64le || riscv64) && !nounsafe && !purego && !appengine
+
+package le
+
+import (
+	"unsafe"
+)
+
+// Load8 will load from b at index i.
+func Load8[I Indexer](b []byte, i I) byte {
+	//return binary.LittleEndian.Uint16(b[i:])
+	//return *(*uint16)(unsafe.Pointer(&b[i]))
+	return *(*byte)(unsafe.Add(unsafe.Pointer(unsafe.SliceData(b)), i))
+}
+
+// Load16 will load from b at index i.
+func Load16[I Indexer](b []byte, i I) uint16 {
+	//return binary.LittleEndian.Uint16(b[i:])
+	//return *(*uint16)(unsafe.Pointer(&b[i]))
+	return *(*uint16)(unsafe.Add(unsafe.Pointer(unsafe.SliceData(b)), i))
+}
+
+// Load32 will load from b at index i.
+func Load32[I Indexer](b []byte, i I) uint32 {
+	//return binary.LittleEndian.Uint32(b[i:])
+	//return *(*uint32)(unsafe.Pointer(&b[i]))
+	return *(*uint32)(unsafe.Add(unsafe.Pointer(unsafe.SliceData(b)), i))
+}
+
+// Load64 will load from b at index i.
+func Load64[I Indexer](b []byte, i I) uint64 {
+	//return binary.LittleEndian.Uint64(b[i:])
+	//return *(*uint64)(unsafe.Pointer(&b[i]))
+	return *(*uint64)(unsafe.Add(unsafe.Pointer(unsafe.SliceData(b)), i))
+}
+
+// Store16 will store v at b.
+func Store16(b []byte, v uint16) {
+	//binary.LittleEndian.PutUint16(b, v)
+	*(*uint16)(unsafe.Pointer(unsafe.SliceData(b))) = v
+}
+
+// Store32 will store v at b.
+func Store32(b []byte, v uint32) {
+	//binary.LittleEndian.PutUint32(b, v)
+	*(*uint32)(unsafe.Pointer(unsafe.SliceData(b))) = v
+}
+
+// Store64 will store v at b.
+func Store64(b []byte, v uint64) {
+	//binary.LittleEndian.PutUint64(b, v)
+	*(*uint64)(unsafe.Pointer(unsafe.SliceData(b))) = v
+}
diff --git a/vendor/github.com/klauspost/compress/internal/snapref/LICENSE b/vendor/github.com/klauspost/compress/internal/snapref/LICENSE
new file mode 100644
index 0000000..6050c10
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/internal/snapref/LICENSE
@@ -0,0 +1,27 @@
+Copyright (c) 2011 The Snappy-Go Authors. All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+   * Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+   * Redistributions in binary form must reproduce the above
+copyright notice, this list of conditions and the following disclaimer
+in the documentation and/or other materials provided with the
+distribution.
+   * Neither the name of Google Inc. nor the names of its
+contributors may be used to endorse or promote products derived from
+this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
diff --git a/vendor/github.com/klauspost/compress/internal/snapref/decode.go b/vendor/github.com/klauspost/compress/internal/snapref/decode.go
new file mode 100644
index 0000000..40796a4
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/internal/snapref/decode.go
@@ -0,0 +1,264 @@
+// Copyright 2011 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package snapref
+
+import (
+	"encoding/binary"
+	"errors"
+	"io"
+)
+
+var (
+	// ErrCorrupt reports that the input is invalid.
+	ErrCorrupt = errors.New("snappy: corrupt input")
+	// ErrTooLarge reports that the uncompressed length is too large.
+	ErrTooLarge = errors.New("snappy: decoded block is too large")
+	// ErrUnsupported reports that the input isn't supported.
+	ErrUnsupported = errors.New("snappy: unsupported input")
+
+	errUnsupportedLiteralLength = errors.New("snappy: unsupported literal length")
+)
+
+// DecodedLen returns the length of the decoded block.
+func DecodedLen(src []byte) (int, error) {
+	v, _, err := decodedLen(src)
+	return v, err
+}
+
+// decodedLen returns the length of the decoded block and the number of bytes
+// that the length header occupied.
+func decodedLen(src []byte) (blockLen, headerLen int, err error) {
+	v, n := binary.Uvarint(src)
+	if n <= 0 || v > 0xffffffff {
+		return 0, 0, ErrCorrupt
+	}
+
+	const wordSize = 32 << (^uint(0) >> 32 & 1)
+	if wordSize == 32 && v > 0x7fffffff {
+		return 0, 0, ErrTooLarge
+	}
+	return int(v), n, nil
+}
+
+const (
+	decodeErrCodeCorrupt                  = 1
+	decodeErrCodeUnsupportedLiteralLength = 2
+)
+
+// Decode returns the decoded form of src. The returned slice may be a sub-
+// slice of dst if dst was large enough to hold the entire decoded block.
+// Otherwise, a newly allocated slice will be returned.
+//
+// The dst and src must not overlap. It is valid to pass a nil dst.
+//
+// Decode handles the Snappy block format, not the Snappy stream format.
+func Decode(dst, src []byte) ([]byte, error) {
+	dLen, s, err := decodedLen(src)
+	if err != nil {
+		return nil, err
+	}
+	if dLen <= len(dst) {
+		dst = dst[:dLen]
+	} else {
+		dst = make([]byte, dLen)
+	}
+	switch decode(dst, src[s:]) {
+	case 0:
+		return dst, nil
+	case decodeErrCodeUnsupportedLiteralLength:
+		return nil, errUnsupportedLiteralLength
+	}
+	return nil, ErrCorrupt
+}
+
+// NewReader returns a new Reader that decompresses from r, using the framing
+// format described at
+// https://github.com/google/snappy/blob/master/framing_format.txt
+func NewReader(r io.Reader) *Reader {
+	return &Reader{
+		r:       r,
+		decoded: make([]byte, maxBlockSize),
+		buf:     make([]byte, maxEncodedLenOfMaxBlockSize+checksumSize),
+	}
+}
+
+// Reader is an io.Reader that can read Snappy-compressed bytes.
+//
+// Reader handles the Snappy stream format, not the Snappy block format.
+type Reader struct {
+	r       io.Reader
+	err     error
+	decoded []byte
+	buf     []byte
+	// decoded[i:j] contains decoded bytes that have not yet been passed on.
+	i, j       int
+	readHeader bool
+}
+
+// Reset discards any buffered data, resets all state, and switches the Snappy
+// reader to read from r. This permits reusing a Reader rather than allocating
+// a new one.
+func (r *Reader) Reset(reader io.Reader) {
+	r.r = reader
+	r.err = nil
+	r.i = 0
+	r.j = 0
+	r.readHeader = false
+}
+
+func (r *Reader) readFull(p []byte, allowEOF bool) (ok bool) {
+	if _, r.err = io.ReadFull(r.r, p); r.err != nil {
+		if r.err == io.ErrUnexpectedEOF || (r.err == io.EOF && !allowEOF) {
+			r.err = ErrCorrupt
+		}
+		return false
+	}
+	return true
+}
+
+func (r *Reader) fill() error {
+	for r.i >= r.j {
+		if !r.readFull(r.buf[:4], true) {
+			return r.err
+		}
+		chunkType := r.buf[0]
+		if !r.readHeader {
+			if chunkType != chunkTypeStreamIdentifier {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			r.readHeader = true
+		}
+		chunkLen := int(r.buf[1]) | int(r.buf[2])<<8 | int(r.buf[3])<<16
+		if chunkLen > len(r.buf) {
+			r.err = ErrUnsupported
+			return r.err
+		}
+
+		// The chunk types are specified at
+		// https://github.com/google/snappy/blob/master/framing_format.txt
+		switch chunkType {
+		case chunkTypeCompressedData:
+			// Section 4.2. Compressed data (chunk type 0x00).
+			if chunkLen < checksumSize {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			buf := r.buf[:chunkLen]
+			if !r.readFull(buf, false) {
+				return r.err
+			}
+			checksum := uint32(buf[0]) | uint32(buf[1])<<8 | uint32(buf[2])<<16 | uint32(buf[3])<<24
+			buf = buf[checksumSize:]
+
+			n, err := DecodedLen(buf)
+			if err != nil {
+				r.err = err
+				return r.err
+			}
+			if n > len(r.decoded) {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			if _, err := Decode(r.decoded, buf); err != nil {
+				r.err = err
+				return r.err
+			}
+			if crc(r.decoded[:n]) != checksum {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			r.i, r.j = 0, n
+			continue
+
+		case chunkTypeUncompressedData:
+			// Section 4.3. Uncompressed data (chunk type 0x01).
+			if chunkLen < checksumSize {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			buf := r.buf[:checksumSize]
+			if !r.readFull(buf, false) {
+				return r.err
+			}
+			checksum := uint32(buf[0]) | uint32(buf[1])<<8 | uint32(buf[2])<<16 | uint32(buf[3])<<24
+			// Read directly into r.decoded instead of via r.buf.
+			n := chunkLen - checksumSize
+			if n > len(r.decoded) {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			if !r.readFull(r.decoded[:n], false) {
+				return r.err
+			}
+			if crc(r.decoded[:n]) != checksum {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			r.i, r.j = 0, n
+			continue
+
+		case chunkTypeStreamIdentifier:
+			// Section 4.1. Stream identifier (chunk type 0xff).
+			if chunkLen != len(magicBody) {
+				r.err = ErrCorrupt
+				return r.err
+			}
+			if !r.readFull(r.buf[:len(magicBody)], false) {
+				return r.err
+			}
+			for i := 0; i < len(magicBody); i++ {
+				if r.buf[i] != magicBody[i] {
+					r.err = ErrCorrupt
+					return r.err
+				}
+			}
+			continue
+		}
+
+		if chunkType <= 0x7f {
+			// Section 4.5. Reserved unskippable chunks (chunk types 0x02-0x7f).
+			r.err = ErrUnsupported
+			return r.err
+		}
+		// Section 4.4 Padding (chunk type 0xfe).
+		// Section 4.6. Reserved skippable chunks (chunk types 0x80-0xfd).
+		if !r.readFull(r.buf[:chunkLen], false) {
+			return r.err
+		}
+	}
+
+	return nil
+}
+
+// Read satisfies the io.Reader interface.
+func (r *Reader) Read(p []byte) (int, error) {
+	if r.err != nil {
+		return 0, r.err
+	}
+
+	if err := r.fill(); err != nil {
+		return 0, err
+	}
+
+	n := copy(p, r.decoded[r.i:r.j])
+	r.i += n
+	return n, nil
+}
+
+// ReadByte satisfies the io.ByteReader interface.
+func (r *Reader) ReadByte() (byte, error) {
+	if r.err != nil {
+		return 0, r.err
+	}
+
+	if err := r.fill(); err != nil {
+		return 0, err
+	}
+
+	c := r.decoded[r.i]
+	r.i++
+	return c, nil
+}
diff --git a/vendor/github.com/klauspost/compress/internal/snapref/decode_other.go b/vendor/github.com/klauspost/compress/internal/snapref/decode_other.go
new file mode 100644
index 0000000..77395a6
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/internal/snapref/decode_other.go
@@ -0,0 +1,113 @@
+// Copyright 2016 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package snapref
+
+// decode writes the decoding of src to dst. It assumes that the varint-encoded
+// length of the decompressed bytes has already been read, and that len(dst)
+// equals that length.
+//
+// It returns 0 on success or a decodeErrCodeXxx error code on failure.
+func decode(dst, src []byte) int {
+	var d, s, offset, length int
+	for s < len(src) {
+		switch src[s] & 0x03 {
+		case tagLiteral:
+			x := uint32(src[s] >> 2)
+			switch {
+			case x < 60:
+				s++
+			case x == 60:
+				s += 2
+				if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+					return decodeErrCodeCorrupt
+				}
+				x = uint32(src[s-1])
+			case x == 61:
+				s += 3
+				if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+					return decodeErrCodeCorrupt
+				}
+				x = uint32(src[s-2]) | uint32(src[s-1])<<8
+			case x == 62:
+				s += 4
+				if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+					return decodeErrCodeCorrupt
+				}
+				x = uint32(src[s-3]) | uint32(src[s-2])<<8 | uint32(src[s-1])<<16
+			case x == 63:
+				s += 5
+				if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+					return decodeErrCodeCorrupt
+				}
+				x = uint32(src[s-4]) | uint32(src[s-3])<<8 | uint32(src[s-2])<<16 | uint32(src[s-1])<<24
+			}
+			length = int(x) + 1
+			if length <= 0 {
+				return decodeErrCodeUnsupportedLiteralLength
+			}
+			if length > len(dst)-d || length > len(src)-s {
+				return decodeErrCodeCorrupt
+			}
+			copy(dst[d:], src[s:s+length])
+			d += length
+			s += length
+			continue
+
+		case tagCopy1:
+			s += 2
+			if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+				return decodeErrCodeCorrupt
+			}
+			length = 4 + int(src[s-2])>>2&0x7
+			offset = int(uint32(src[s-2])&0xe0<<3 | uint32(src[s-1]))
+
+		case tagCopy2:
+			s += 3
+			if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+				return decodeErrCodeCorrupt
+			}
+			length = 1 + int(src[s-3])>>2
+			offset = int(uint32(src[s-2]) | uint32(src[s-1])<<8)
+
+		case tagCopy4:
+			s += 5
+			if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line.
+				return decodeErrCodeCorrupt
+			}
+			length = 1 + int(src[s-5])>>2
+			offset = int(uint32(src[s-4]) | uint32(src[s-3])<<8 | uint32(src[s-2])<<16 | uint32(src[s-1])<<24)
+		}
+
+		if offset <= 0 || d < offset || length > len(dst)-d {
+			return decodeErrCodeCorrupt
+		}
+		// Copy from an earlier sub-slice of dst to a later sub-slice.
+		// If no overlap, use the built-in copy:
+		if offset >= length {
+			copy(dst[d:d+length], dst[d-offset:])
+			d += length
+			continue
+		}
+
+		// Unlike the built-in copy function, this byte-by-byte copy always runs
+		// forwards, even if the slices overlap. Conceptually, this is:
+		//
+		// d += forwardCopy(dst[d:d+length], dst[d-offset:])
+		//
+		// We align the slices into a and b and show the compiler they are the same size.
+		// This allows the loop to run without bounds checks.
+		a := dst[d : d+length]
+		b := dst[d-offset:]
+		b = b[:len(a)]
+		for i := range a {
+			a[i] = b[i]
+		}
+		d += length
+	}
+	if d != len(dst) {
+		return decodeErrCodeCorrupt
+	}
+	return 0
+}
diff --git a/vendor/github.com/klauspost/compress/internal/snapref/encode.go b/vendor/github.com/klauspost/compress/internal/snapref/encode.go
new file mode 100644
index 0000000..13c6040
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/internal/snapref/encode.go
@@ -0,0 +1,289 @@
+// Copyright 2011 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package snapref
+
+import (
+	"encoding/binary"
+	"errors"
+	"io"
+)
+
+// Encode returns the encoded form of src. The returned slice may be a sub-
+// slice of dst if dst was large enough to hold the entire encoded block.
+// Otherwise, a newly allocated slice will be returned.
+//
+// The dst and src must not overlap. It is valid to pass a nil dst.
+//
+// Encode handles the Snappy block format, not the Snappy stream format.
+func Encode(dst, src []byte) []byte {
+	if n := MaxEncodedLen(len(src)); n < 0 {
+		panic(ErrTooLarge)
+	} else if len(dst) < n {
+		dst = make([]byte, n)
+	}
+
+	// The block starts with the varint-encoded length of the decompressed bytes.
+	d := binary.PutUvarint(dst, uint64(len(src)))
+
+	for len(src) > 0 {
+		p := src
+		src = nil
+		if len(p) > maxBlockSize {
+			p, src = p[:maxBlockSize], p[maxBlockSize:]
+		}
+		if len(p) < minNonLiteralBlockSize {
+			d += emitLiteral(dst[d:], p)
+		} else {
+			d += encodeBlock(dst[d:], p)
+		}
+	}
+	return dst[:d]
+}
+
+// inputMargin is the minimum number of extra input bytes to keep, inside
+// encodeBlock's inner loop. On some architectures, this margin lets us
+// implement a fast path for emitLiteral, where the copy of short (<= 16 byte)
+// literals can be implemented as a single load to and store from a 16-byte
+// register. That literal's actual length can be as short as 1 byte, so this
+// can copy up to 15 bytes too much, but that's OK as subsequent iterations of
+// the encoding loop will fix up the copy overrun, and this inputMargin ensures
+// that we don't overrun the dst and src buffers.
+const inputMargin = 16 - 1
+
+// minNonLiteralBlockSize is the minimum size of the input to encodeBlock that
+// could be encoded with a copy tag. This is the minimum with respect to the
+// algorithm used by encodeBlock, not a minimum enforced by the file format.
+//
+// The encoded output must start with at least a 1 byte literal, as there are
+// no previous bytes to copy. A minimal (1 byte) copy after that, generated
+// from an emitCopy call in encodeBlock's main loop, would require at least
+// another inputMargin bytes, for the reason above: we want any emitLiteral
+// calls inside encodeBlock's main loop to use the fast path if possible, which
+// requires being able to overrun by inputMargin bytes. Thus,
+// minNonLiteralBlockSize equals 1 + 1 + inputMargin.
+//
+// The C++ code doesn't use this exact threshold, but it could, as discussed at
+// https://groups.google.com/d/topic/snappy-compression/oGbhsdIJSJ8/discussion
+// The difference between Go (2+inputMargin) and C++ (inputMargin) is purely an
+// optimization. It should not affect the encoded form. This is tested by
+// TestSameEncodingAsCppShortCopies.
+const minNonLiteralBlockSize = 1 + 1 + inputMargin
+
+// MaxEncodedLen returns the maximum length of a snappy block, given its
+// uncompressed length.
+//
+// It will return a negative value if srcLen is too large to encode.
+func MaxEncodedLen(srcLen int) int {
+	n := uint64(srcLen)
+	if n > 0xffffffff {
+		return -1
+	}
+	// Compressed data can be defined as:
+	//    compressed := item* literal*
+	//    item       := literal* copy
+	//
+	// The trailing literal sequence has a space blowup of at most 62/60
+	// since a literal of length 60 needs one tag byte + one extra byte
+	// for length information.
+	//
+	// Item blowup is trickier to measure. Suppose the "copy" op copies
+	// 4 bytes of data. Because of a special check in the encoding code,
+	// we produce a 4-byte copy only if the offset is < 65536. Therefore
+	// the copy op takes 3 bytes to encode, and this type of item leads
+	// to at most the 62/60 blowup for representing literals.
+	//
+	// Suppose the "copy" op copies 5 bytes of data. If the offset is big
+	// enough, it will take 5 bytes to encode the copy op. Therefore the
+	// worst case here is a one-byte literal followed by a five-byte copy.
+	// That is, 6 bytes of input turn into 7 bytes of "compressed" data.
+	//
+	// This last factor dominates the blowup, so the final estimate is:
+	n = 32 + n + n/6
+	if n > 0xffffffff {
+		return -1
+	}
+	return int(n)
+}
+
+var errClosed = errors.New("snappy: Writer is closed")
+
+// NewWriter returns a new Writer that compresses to w.
+//
+// The Writer returned does not buffer writes. There is no need to Flush or
+// Close such a Writer.
+//
+// Deprecated: the Writer returned is not suitable for many small writes, only
+// for few large writes. Use NewBufferedWriter instead, which is efficient
+// regardless of the frequency and shape of the writes, and remember to Close
+// that Writer when done.
+func NewWriter(w io.Writer) *Writer {
+	return &Writer{
+		w:    w,
+		obuf: make([]byte, obufLen),
+	}
+}
+
+// NewBufferedWriter returns a new Writer that compresses to w, using the
+// framing format described at
+// https://github.com/google/snappy/blob/master/framing_format.txt
+//
+// The Writer returned buffers writes. Users must call Close to guarantee all
+// data has been forwarded to the underlying io.Writer. They may also call
+// Flush zero or more times before calling Close.
+func NewBufferedWriter(w io.Writer) *Writer {
+	return &Writer{
+		w:    w,
+		ibuf: make([]byte, 0, maxBlockSize),
+		obuf: make([]byte, obufLen),
+	}
+}
+
+// Writer is an io.Writer that can write Snappy-compressed bytes.
+//
+// Writer handles the Snappy stream format, not the Snappy block format.
+type Writer struct {
+	w   io.Writer
+	err error
+
+	// ibuf is a buffer for the incoming (uncompressed) bytes.
+	//
+	// Its use is optional. For backwards compatibility, Writers created by the
+	// NewWriter function have ibuf == nil, do not buffer incoming bytes, and
+	// therefore do not need to be Flush'ed or Close'd.
+	ibuf []byte
+
+	// obuf is a buffer for the outgoing (compressed) bytes.
+	obuf []byte
+
+	// wroteStreamHeader is whether we have written the stream header.
+	wroteStreamHeader bool
+}
+
+// Reset discards the writer's state and switches the Snappy writer to write to
+// w. This permits reusing a Writer rather than allocating a new one.
+func (w *Writer) Reset(writer io.Writer) {
+	w.w = writer
+	w.err = nil
+	if w.ibuf != nil {
+		w.ibuf = w.ibuf[:0]
+	}
+	w.wroteStreamHeader = false
+}
+
+// Write satisfies the io.Writer interface.
+func (w *Writer) Write(p []byte) (nRet int, errRet error) {
+	if w.ibuf == nil {
+		// Do not buffer incoming bytes. This does not perform or compress well
+		// if the caller of Writer.Write writes many small slices. This
+		// behavior is therefore deprecated, but still supported for backwards
+		// compatibility with code that doesn't explicitly Flush or Close.
+		return w.write(p)
+	}
+
+	// The remainder of this method is based on bufio.Writer.Write from the
+	// standard library.
+
+	for len(p) > (cap(w.ibuf)-len(w.ibuf)) && w.err == nil {
+		var n int
+		if len(w.ibuf) == 0 {
+			// Large write, empty buffer.
+			// Write directly from p to avoid copy.
+			n, _ = w.write(p)
+		} else {
+			n = copy(w.ibuf[len(w.ibuf):cap(w.ibuf)], p)
+			w.ibuf = w.ibuf[:len(w.ibuf)+n]
+			w.Flush()
+		}
+		nRet += n
+		p = p[n:]
+	}
+	if w.err != nil {
+		return nRet, w.err
+	}
+	n := copy(w.ibuf[len(w.ibuf):cap(w.ibuf)], p)
+	w.ibuf = w.ibuf[:len(w.ibuf)+n]
+	nRet += n
+	return nRet, nil
+}
+
+func (w *Writer) write(p []byte) (nRet int, errRet error) {
+	if w.err != nil {
+		return 0, w.err
+	}
+	for len(p) > 0 {
+		obufStart := len(magicChunk)
+		if !w.wroteStreamHeader {
+			w.wroteStreamHeader = true
+			copy(w.obuf, magicChunk)
+			obufStart = 0
+		}
+
+		var uncompressed []byte
+		if len(p) > maxBlockSize {
+			uncompressed, p = p[:maxBlockSize], p[maxBlockSize:]
+		} else {
+			uncompressed, p = p, nil
+		}
+		checksum := crc(uncompressed)
+
+		// Compress the buffer, discarding the result if the improvement
+		// isn't at least 12.5%.
+		compressed := Encode(w.obuf[obufHeaderLen:], uncompressed)
+		chunkType := uint8(chunkTypeCompressedData)
+		chunkLen := 4 + len(compressed)
+		obufEnd := obufHeaderLen + len(compressed)
+		if len(compressed) >= len(uncompressed)-len(uncompressed)/8 {
+			chunkType = chunkTypeUncompressedData
+			chunkLen = 4 + len(uncompressed)
+			obufEnd = obufHeaderLen
+		}
+
+		// Fill in the per-chunk header that comes before the body.
+		w.obuf[len(magicChunk)+0] = chunkType
+		w.obuf[len(magicChunk)+1] = uint8(chunkLen >> 0)
+		w.obuf[len(magicChunk)+2] = uint8(chunkLen >> 8)
+		w.obuf[len(magicChunk)+3] = uint8(chunkLen >> 16)
+		w.obuf[len(magicChunk)+4] = uint8(checksum >> 0)
+		w.obuf[len(magicChunk)+5] = uint8(checksum >> 8)
+		w.obuf[len(magicChunk)+6] = uint8(checksum >> 16)
+		w.obuf[len(magicChunk)+7] = uint8(checksum >> 24)
+
+		if _, err := w.w.Write(w.obuf[obufStart:obufEnd]); err != nil {
+			w.err = err
+			return nRet, err
+		}
+		if chunkType == chunkTypeUncompressedData {
+			if _, err := w.w.Write(uncompressed); err != nil {
+				w.err = err
+				return nRet, err
+			}
+		}
+		nRet += len(uncompressed)
+	}
+	return nRet, nil
+}
+
+// Flush flushes the Writer to its underlying io.Writer.
+func (w *Writer) Flush() error {
+	if w.err != nil {
+		return w.err
+	}
+	if len(w.ibuf) == 0 {
+		return nil
+	}
+	w.write(w.ibuf)
+	w.ibuf = w.ibuf[:0]
+	return w.err
+}
+
+// Close calls Flush and then closes the Writer.
+func (w *Writer) Close() error {
+	w.Flush()
+	ret := w.err
+	if w.err == nil {
+		w.err = errClosed
+	}
+	return ret
+}
diff --git a/vendor/github.com/klauspost/compress/internal/snapref/encode_other.go b/vendor/github.com/klauspost/compress/internal/snapref/encode_other.go
new file mode 100644
index 0000000..2754bac
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/internal/snapref/encode_other.go
@@ -0,0 +1,250 @@
+// Copyright 2016 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package snapref
+
+func load32(b []byte, i int) uint32 {
+	b = b[i : i+4 : len(b)] // Help the compiler eliminate bounds checks on the next line.
+	return uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24
+}
+
+func load64(b []byte, i int) uint64 {
+	b = b[i : i+8 : len(b)] // Help the compiler eliminate bounds checks on the next line.
+	return uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 |
+		uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56
+}
+
+// emitLiteral writes a literal chunk and returns the number of bytes written.
+//
+// It assumes that:
+//
+//	dst is long enough to hold the encoded bytes
+//	1 <= len(lit) && len(lit) <= 65536
+func emitLiteral(dst, lit []byte) int {
+	i, n := 0, uint(len(lit)-1)
+	switch {
+	case n < 60:
+		dst[0] = uint8(n)<<2 | tagLiteral
+		i = 1
+	case n < 1<<8:
+		dst[0] = 60<<2 | tagLiteral
+		dst[1] = uint8(n)
+		i = 2
+	default:
+		dst[0] = 61<<2 | tagLiteral
+		dst[1] = uint8(n)
+		dst[2] = uint8(n >> 8)
+		i = 3
+	}
+	return i + copy(dst[i:], lit)
+}
+
+// emitCopy writes a copy chunk and returns the number of bytes written.
+//
+// It assumes that:
+//
+//	dst is long enough to hold the encoded bytes
+//	1 <= offset && offset <= 65535
+//	4 <= length && length <= 65535
+func emitCopy(dst []byte, offset, length int) int {
+	i := 0
+	// The maximum length for a single tagCopy1 or tagCopy2 op is 64 bytes. The
+	// threshold for this loop is a little higher (at 68 = 64 + 4), and the
+	// length emitted down below is a little lower (at 60 = 64 - 4), because
+	// it's shorter to encode a length 67 copy as a length 60 tagCopy2 followed
+	// by a length 7 tagCopy1 (which encodes as 3+2 bytes) than to encode it as
+	// a length 64 tagCopy2 followed by a length 3 tagCopy2 (which encodes as
+	// 3+3 bytes). The magic 4 in the 64±4 is because the minimum length for a
+	// tagCopy1 op is 4 bytes, which is why a length 3 copy has to be an
+	// encodes-as-3-bytes tagCopy2 instead of an encodes-as-2-bytes tagCopy1.
+	for length >= 68 {
+		// Emit a length 64 copy, encoded as 3 bytes.
+		dst[i+0] = 63<<2 | tagCopy2
+		dst[i+1] = uint8(offset)
+		dst[i+2] = uint8(offset >> 8)
+		i += 3
+		length -= 64
+	}
+	if length > 64 {
+		// Emit a length 60 copy, encoded as 3 bytes.
+		dst[i+0] = 59<<2 | tagCopy2
+		dst[i+1] = uint8(offset)
+		dst[i+2] = uint8(offset >> 8)
+		i += 3
+		length -= 60
+	}
+	if length >= 12 || offset >= 2048 {
+		// Emit the remaining copy, encoded as 3 bytes.
+		dst[i+0] = uint8(length-1)<<2 | tagCopy2
+		dst[i+1] = uint8(offset)
+		dst[i+2] = uint8(offset >> 8)
+		return i + 3
+	}
+	// Emit the remaining copy, encoded as 2 bytes.
+	dst[i+0] = uint8(offset>>8)<<5 | uint8(length-4)<<2 | tagCopy1
+	dst[i+1] = uint8(offset)
+	return i + 2
+}
+
+func hash(u, shift uint32) uint32 {
+	return (u * 0x1e35a7bd) >> shift
+}
+
+// EncodeBlockInto exposes encodeBlock but checks dst size.
+func EncodeBlockInto(dst, src []byte) (d int) {
+	if MaxEncodedLen(len(src)) > len(dst) {
+		return 0
+	}
+
+	// encodeBlock breaks on too big blocks, so split.
+	for len(src) > 0 {
+		p := src
+		src = nil
+		if len(p) > maxBlockSize {
+			p, src = p[:maxBlockSize], p[maxBlockSize:]
+		}
+		if len(p) < minNonLiteralBlockSize {
+			d += emitLiteral(dst[d:], p)
+		} else {
+			d += encodeBlock(dst[d:], p)
+		}
+	}
+	return d
+}
+
+// encodeBlock encodes a non-empty src to a guaranteed-large-enough dst. It
+// assumes that the varint-encoded length of the decompressed bytes has already
+// been written.
+//
+// It also assumes that:
+//
+//	len(dst) >= MaxEncodedLen(len(src)) &&
+//	minNonLiteralBlockSize <= len(src) && len(src) <= maxBlockSize
+func encodeBlock(dst, src []byte) (d int) {
+	// Initialize the hash table. Its size ranges from 1<<8 to 1<<14 inclusive.
+	// The table element type is uint16, as s < sLimit and sLimit < len(src)
+	// and len(src) <= maxBlockSize and maxBlockSize == 65536.
+	const (
+		maxTableSize = 1 << 14
+		// tableMask is redundant, but helps the compiler eliminate bounds
+		// checks.
+		tableMask = maxTableSize - 1
+	)
+	shift := uint32(32 - 8)
+	for tableSize := 1 << 8; tableSize < maxTableSize && tableSize < len(src); tableSize *= 2 {
+		shift--
+	}
+	// In Go, all array elements are zero-initialized, so there is no advantage
+	// to a smaller tableSize per se. However, it matches the C++ algorithm,
+	// and in the asm versions of this code, we can get away with zeroing only
+	// the first tableSize elements.
+	var table [maxTableSize]uint16
+
+	// sLimit is when to stop looking for offset/length copies. The inputMargin
+	// lets us use a fast path for emitLiteral in the main loop, while we are
+	// looking for copies.
+	sLimit := len(src) - inputMargin
+
+	// nextEmit is where in src the next emitLiteral should start from.
+	nextEmit := 0
+
+	// The encoded form must start with a literal, as there are no previous
+	// bytes to copy, so we start looking for hash matches at s == 1.
+	s := 1
+	nextHash := hash(load32(src, s), shift)
+
+	for {
+		// Copied from the C++ snappy implementation:
+		//
+		// Heuristic match skipping: If 32 bytes are scanned with no matches
+		// found, start looking only at every other byte. If 32 more bytes are
+		// scanned (or skipped), look at every third byte, etc.. When a match
+		// is found, immediately go back to looking at every byte. This is a
+		// small loss (~5% performance, ~0.1% density) for compressible data
+		// due to more bookkeeping, but for non-compressible data (such as
+		// JPEG) it's a huge win since the compressor quickly "realizes" the
+		// data is incompressible and doesn't bother looking for matches
+		// everywhere.
+		//
+		// The "skip" variable keeps track of how many bytes there are since
+		// the last match; dividing it by 32 (ie. right-shifting by five) gives
+		// the number of bytes to move ahead for each iteration.
+		skip := 32
+
+		nextS := s
+		candidate := 0
+		for {
+			s = nextS
+			bytesBetweenHashLookups := skip >> 5
+			nextS = s + bytesBetweenHashLookups
+			skip += bytesBetweenHashLookups
+			if nextS > sLimit {
+				goto emitRemainder
+			}
+			candidate = int(table[nextHash&tableMask])
+			table[nextHash&tableMask] = uint16(s)
+			nextHash = hash(load32(src, nextS), shift)
+			if load32(src, s) == load32(src, candidate) {
+				break
+			}
+		}
+
+		// A 4-byte match has been found. We'll later see if more than 4 bytes
+		// match. But, prior to the match, src[nextEmit:s] are unmatched. Emit
+		// them as literal bytes.
+		d += emitLiteral(dst[d:], src[nextEmit:s])
+
+		// Call emitCopy, and then see if another emitCopy could be our next
+		// move. Repeat until we find no match for the input immediately after
+		// what was consumed by the last emitCopy call.
+		//
+		// If we exit this loop normally then we need to call emitLiteral next,
+		// though we don't yet know how big the literal will be. We handle that
+		// by proceeding to the next iteration of the main loop. We also can
+		// exit this loop via goto if we get close to exhausting the input.
+		for {
+			// Invariant: we have a 4-byte match at s, and no need to emit any
+			// literal bytes prior to s.
+			base := s
+
+			// Extend the 4-byte match as long as possible.
+			//
+			// This is an inlined version of:
+			//	s = extendMatch(src, candidate+4, s+4)
+			s += 4
+			for i := candidate + 4; s < len(src) && src[i] == src[s]; i, s = i+1, s+1 {
+			}
+
+			d += emitCopy(dst[d:], base-candidate, s-base)
+			nextEmit = s
+			if s >= sLimit {
+				goto emitRemainder
+			}
+
+			// We could immediately start working at s now, but to improve
+			// compression we first update the hash table at s-1 and at s. If
+			// another emitCopy is not our next move, also calculate nextHash
+			// at s+1. At least on GOARCH=amd64, these three hash calculations
+			// are faster as one load64 call (with some shifts) instead of
+			// three load32 calls.
+			x := load64(src, s-1)
+			prevHash := hash(uint32(x>>0), shift)
+			table[prevHash&tableMask] = uint16(s - 1)
+			currHash := hash(uint32(x>>8), shift)
+			candidate = int(table[currHash&tableMask])
+			table[currHash&tableMask] = uint16(s)
+			if uint32(x>>8) != load32(src, candidate) {
+				nextHash = hash(uint32(x>>16), shift)
+				s++
+				break
+			}
+		}
+	}
+
+emitRemainder:
+	if nextEmit < len(src) {
+		d += emitLiteral(dst[d:], src[nextEmit:])
+	}
+	return d
+}
diff --git a/vendor/github.com/klauspost/compress/internal/snapref/snappy.go b/vendor/github.com/klauspost/compress/internal/snapref/snappy.go
new file mode 100644
index 0000000..34d01f4
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/internal/snapref/snappy.go
@@ -0,0 +1,98 @@
+// Copyright 2011 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package snapref implements the Snappy compression format. It aims for very
+// high speeds and reasonable compression.
+//
+// There are actually two Snappy formats: block and stream. They are related,
+// but different: trying to decompress block-compressed data as a Snappy stream
+// will fail, and vice versa. The block format is the Decode and Encode
+// functions and the stream format is the Reader and Writer types.
+//
+// The block format, the more common case, is used when the complete size (the
+// number of bytes) of the original data is known upfront, at the time
+// compression starts. The stream format, also known as the framing format, is
+// for when that isn't always true.
+//
+// The canonical, C++ implementation is at https://github.com/google/snappy and
+// it only implements the block format.
+package snapref
+
+import (
+	"hash/crc32"
+)
+
+/*
+Each encoded block begins with the varint-encoded length of the decoded data,
+followed by a sequence of chunks. Chunks begin and end on byte boundaries. The
+first byte of each chunk is broken into its 2 least and 6 most significant bits
+called l and m: l ranges in [0, 4) and m ranges in [0, 64). l is the chunk tag.
+Zero means a literal tag. All other values mean a copy tag.
+
+For literal tags:
+  - If m < 60, the next 1 + m bytes are literal bytes.
+  - Otherwise, let n be the little-endian unsigned integer denoted by the next
+    m - 59 bytes. The next 1 + n bytes after that are literal bytes.
+
+For copy tags, length bytes are copied from offset bytes ago, in the style of
+Lempel-Ziv compression algorithms. In particular:
+  - For l == 1, the offset ranges in [0, 1<<11) and the length in [4, 12).
+    The length is 4 + the low 3 bits of m. The high 3 bits of m form bits 8-10
+    of the offset. The next byte is bits 0-7 of the offset.
+  - For l == 2, the offset ranges in [0, 1<<16) and the length in [1, 65).
+    The length is 1 + m. The offset is the little-endian unsigned integer
+    denoted by the next 2 bytes.
+  - For l == 3, this tag is a legacy format that is no longer issued by most
+    encoders. Nonetheless, the offset ranges in [0, 1<<32) and the length in
+    [1, 65). The length is 1 + m. The offset is the little-endian unsigned
+    integer denoted by the next 4 bytes.
+*/
+const (
+	tagLiteral = 0x00
+	tagCopy1   = 0x01
+	tagCopy2   = 0x02
+	tagCopy4   = 0x03
+)
+
+const (
+	checksumSize    = 4
+	chunkHeaderSize = 4
+	magicChunk      = "\xff\x06\x00\x00" + magicBody
+	magicBody       = "sNaPpY"
+
+	// maxBlockSize is the maximum size of the input to encodeBlock. It is not
+	// part of the wire format per se, but some parts of the encoder assume
+	// that an offset fits into a uint16.
+	//
+	// Also, for the framing format (Writer type instead of Encode function),
+	// https://github.com/google/snappy/blob/master/framing_format.txt says
+	// that "the uncompressed data in a chunk must be no longer than 65536
+	// bytes".
+	maxBlockSize = 65536
+
+	// maxEncodedLenOfMaxBlockSize equals MaxEncodedLen(maxBlockSize), but is
+	// hard coded to be a const instead of a variable, so that obufLen can also
+	// be a const. Their equivalence is confirmed by
+	// TestMaxEncodedLenOfMaxBlockSize.
+	maxEncodedLenOfMaxBlockSize = 76490
+
+	obufHeaderLen = len(magicChunk) + checksumSize + chunkHeaderSize
+	obufLen       = obufHeaderLen + maxEncodedLenOfMaxBlockSize
+)
+
+const (
+	chunkTypeCompressedData   = 0x00
+	chunkTypeUncompressedData = 0x01
+	chunkTypePadding          = 0xfe
+	chunkTypeStreamIdentifier = 0xff
+)
+
+var crcTable = crc32.MakeTable(crc32.Castagnoli)
+
+// crc implements the checksum specified in section 3 of
+// https://github.com/google/snappy/blob/master/framing_format.txt
+func crc(b []byte) uint32 {
+	c := crc32.Update(0, crcTable, b)
+	return uint32(c>>15|c<<17) + 0xa282ead8
+}
diff --git a/vendor/github.com/klauspost/compress/s2sx.mod b/vendor/github.com/klauspost/compress/s2sx.mod
new file mode 100644
index 0000000..81bda5e
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/s2sx.mod
@@ -0,0 +1,3 @@
+module github.com/klauspost/compress
+
+go 1.22
diff --git a/vendor/github.com/klauspost/compress/s2sx.sum b/vendor/github.com/klauspost/compress/s2sx.sum
new file mode 100644
index 0000000..e69de29
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/s2sx.sum
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 {