Merge pull request #2071 from baude/vendorcs

vendor in new containers/storage

38 files changed, 7674 insertions, 108 deletions

diff --git a/vendor.conf b/vendor.conf
index 1ef9ada84..6bbcb347e 100644
--- a/vendor.conf
+++ b/vendor.conf
@@ -12,7 +12,7 @@ github.com/containerd/continuity master
 github.com/containernetworking/cni v0.7.0-alpha1
 github.com/containernetworking/plugins 1562a1e60ed101aacc5e08ed9dbeba8e9f3d4ec1
 github.com/containers/image d53afe179b381fafb427e6b9cf9b1996a98c1067
-github.com/containers/storage db40f96d853dfced60c563e61fb66ba231ce7c8d
+github.com/containers/storage c044256cbfbae8a18cfc76319e11fcf4f98843b9
 github.com/containers/psgo dc0bc9fac5b715034c4310ed4d795b3182360842
 github.com/coreos/go-systemd v14
 github.com/cri-o/ocicni 2d2983e40c242322a56c22a903785e7f83eb378c
@@ -101,3 +101,6 @@ github.com/mailru/easyjson 03f2033d19d5860aef995fe360ac7d395cd8ce65
 github.com/coreos/go-iptables 25d087f3cffd9aedc0c2b7eff25f23cbf3c20fe1
 github.com/google/shlex c34317bd91bf98fab745d77b03933cf8769299fe
 github.com/pkg/profile v1.2.1
+github.com/klauspost/pgzip v1.2.1
+github.com/klauspost/compress v1.4.1
+github.com/klauspost/cpuid v1.2.0
diff --git a/vendor/github.com/containers/storage/drivers/devmapper/device_setup.go b/vendor/github.com/containers/storage/drivers/devmapper/device_setup.go
index 53aa217c9..58abca477 100644
--- a/vendor/github.com/containers/storage/drivers/devmapper/device_setup.go
+++ b/vendor/github.com/containers/storage/drivers/devmapper/device_setup.go
@@ -26,7 +26,7 @@ type directLVMConfig struct {
 var (
 	errThinpPercentMissing = errors.New("must set both `dm.thinp_percent` and `dm.thinp_metapercent` if either is specified")
 	errThinpPercentTooBig  = errors.New("combined `dm.thinp_percent` and `dm.thinp_metapercent` must not be greater than 100")
-	errMissingSetupDevice  = errors.New("must provide device path in `dm.setup_device` in order to configure direct-lvm")
+	errMissingSetupDevice  = errors.New("must provide device path in `dm.directlvm_device` in order to configure direct-lvm")
 )
 
 func validateLVMConfig(cfg directLVMConfig) error {
diff --git a/vendor/github.com/containers/storage/drivers/overlay/overlay.go b/vendor/github.com/containers/storage/drivers/overlay/overlay.go
index df736c0a9..2455b6736 100644
--- a/vendor/github.com/containers/storage/drivers/overlay/overlay.go
+++ b/vendor/github.com/containers/storage/drivers/overlay/overlay.go
@@ -14,6 +14,7 @@ import (
 	"strconv"
 	"strings"
 	"sync"
+	"syscall"
 
 	"github.com/containers/storage/drivers"
 	"github.com/containers/storage/drivers/overlayutils"
@@ -165,6 +166,10 @@ func Init(home string, options []string, uidMaps, gidMaps []idtools.IDMap) (grap
 		if err != nil {
 			os.Remove(filepath.Join(home, linkDir))
 			os.Remove(home)
+			patherr, ok := err.(*os.PathError)
+			if ok && patherr.Err == syscall.ENOSPC {
+				return nil, err
+			}
 			return nil, errors.Wrap(err, "kernel does not support overlay fs")
 		}
 	}
@@ -285,6 +290,12 @@ func supportsOverlay(home string, homeMagic graphdriver.FsMagic, rootUID, rootGI
 	exec.Command("modprobe", "overlay").Run()
 
 	layerDir, err := ioutil.TempDir(home, "compat")
+	if err != nil {
+		patherr, ok := err.(*os.PathError)
+		if ok && patherr.Err == syscall.ENOSPC {
+			return false, err
+		}
+	}
 	if err == nil {
 		// Check if reading the directory's contents populates the d_type field, which is required
 		// for proper operation of the overlay filesystem.
diff --git a/vendor/github.com/containers/storage/layers.go b/vendor/github.com/containers/storage/layers.go
index 0b532eb77..299d2f818 100644
--- a/vendor/github.com/containers/storage/layers.go
+++ b/vendor/github.com/containers/storage/layers.go
@@ -2,7 +2,6 @@ package storage
 
 import (
 	"bytes"
-	"compress/gzip"
 	"encoding/json"
 	"fmt"
 	"io"
@@ -20,6 +19,7 @@ import (
 	"github.com/containers/storage/pkg/stringid"
 	"github.com/containers/storage/pkg/system"
 	"github.com/containers/storage/pkg/truncindex"
+	"github.com/klauspost/pgzip"
 	digest "github.com/opencontainers/go-digest"
 	"github.com/opencontainers/selinux/go-selinux/label"
 	"github.com/pkg/errors"
@@ -1055,7 +1055,7 @@ func (r *layerStore) Diff(from, to string, options *DiffOptions) (io.ReadCloser,
 	}
 	defer tsfile.Close()
 
-	decompressor, err := gzip.NewReader(tsfile)
+	decompressor, err := pgzip.NewReader(tsfile)
 	if err != nil {
 		return nil, err
 	}
@@ -1116,9 +1116,9 @@ func (r *layerStore) ApplyDiff(to string, diff io.Reader) (size int64, err error
 	defragmented := io.TeeReader(io.MultiReader(bytes.NewBuffer(header[:n]), diff), compressedCounter)
 
 	tsdata := bytes.Buffer{}
-	compressor, err := gzip.NewWriterLevel(&tsdata, gzip.BestSpeed)
+	compressor, err := pgzip.NewWriterLevel(&tsdata, pgzip.BestSpeed)
 	if err != nil {
-		compressor = gzip.NewWriter(&tsdata)
+		compressor = pgzip.NewWriter(&tsdata)
 	}
 	metadata := storage.NewJSONPacker(compressor)
 	uncompressed, err := archive.DecompressStream(defragmented)
diff --git a/vendor/github.com/containers/storage/pkg/archive/archive.go b/vendor/github.com/containers/storage/pkg/archive/archive.go
index 8d6eaacf3..228d8bb82 100644
--- a/vendor/github.com/containers/storage/pkg/archive/archive.go
+++ b/vendor/github.com/containers/storage/pkg/archive/archive.go
@@ -5,7 +5,6 @@ import (
 	"bufio"
 	"bytes"
 	"compress/bzip2"
-	"compress/gzip"
 	"fmt"
 	"io"
 	"io/ioutil"
@@ -22,6 +21,7 @@ import (
 	"github.com/containers/storage/pkg/pools"
 	"github.com/containers/storage/pkg/promise"
 	"github.com/containers/storage/pkg/system"
+	gzip "github.com/klauspost/pgzip"
 	rsystem "github.com/opencontainers/runc/libcontainer/system"
 	"github.com/sirupsen/logrus"
 )
@@ -499,6 +499,8 @@ func (ta *tarAppender) addTarFile(path, name string) error {
 		hdr.Gid = ta.ChownOpts.GID
 	}
 
+	maybeTruncateHeaderModTime(hdr)
+
 	if ta.WhiteoutConverter != nil {
 		wo, err := ta.WhiteoutConverter.ConvertWrite(hdr, path, fi)
 		if err != nil {
@@ -640,11 +642,12 @@ func createTarFile(path, extractDir string, hdr *tar.Header, reader io.Reader, L
 	var errors []string
 	for key, value := range hdr.Xattrs {
 		if err := system.Lsetxattr(path, key, []byte(value), 0); err != nil {
-			if err == syscall.ENOTSUP {
+			if err == syscall.ENOTSUP || (err == syscall.EPERM && inUserns) {
 				// We ignore errors here because not all graphdrivers support
 				// xattrs *cough* old versions of AUFS *cough*. However only
 				// ENOTSUP should be emitted in that case, otherwise we still
-				// bail.
+				// bail.  We also ignore EPERM errors if we are running in a
+				// user namespace.
 				errors = append(errors, err.Error())
 				continue
 			}
diff --git a/vendor/github.com/containers/storage/pkg/archive/archive_110.go b/vendor/github.com/containers/storage/pkg/archive/archive_110.go
index 22b8b48cc..7bc44a566 100644
--- a/vendor/github.com/containers/storage/pkg/archive/archive_110.go
+++ b/vendor/github.com/containers/storage/pkg/archive/archive_110.go
@@ -4,8 +4,19 @@ package archive
 
 import (
 	"archive/tar"
+	"time"
 )
 
 func copyPassHeader(hdr *tar.Header) {
 	hdr.Format = tar.FormatPAX
 }
+
+func maybeTruncateHeaderModTime(hdr *tar.Header) {
+	if hdr.Format == tar.FormatUnknown {
+		// one of the first things archive/tar does is round this
+		// value, possibly up, if the format isn't specified, while we
+		// are much better equipped to handle truncation when scanning
+		// for changes between source and an extracted copy of this
+		hdr.ModTime = hdr.ModTime.Truncate(time.Second)
+	}
+}
diff --git a/vendor/github.com/containers/storage/pkg/archive/archive_19.go b/vendor/github.com/containers/storage/pkg/archive/archive_19.go
index d10d595fa..d19811fdb 100644
--- a/vendor/github.com/containers/storage/pkg/archive/archive_19.go
+++ b/vendor/github.com/containers/storage/pkg/archive/archive_19.go
@@ -8,3 +8,6 @@ import (
 
 func copyPassHeader(hdr *tar.Header) {
 }
+
+func maybeTruncateHeaderModTime(hdr *tar.Header) {
+}
diff --git a/vendor/github.com/containers/storage/pkg/archive/archive_ffjson.go b/vendor/github.com/containers/storage/pkg/archive/archive_ffjson.go
index 9b8103e4d..2ba44b85d 100644
--- a/vendor/github.com/containers/storage/pkg/archive/archive_ffjson.go
+++ b/vendor/github.com/containers/storage/pkg/archive/archive_ffjson.go
@@ -1,5 +1,5 @@
 // Code generated by ffjson <https://github.com/pquerna/ffjson>. DO NOT EDIT.
-// source: pkg/archive/archive.go
+// source: ./pkg/archive/archive.go
 
 package archive
 
diff --git a/vendor/github.com/containers/storage/pkg/archive/example_changes.go b/vendor/github.com/containers/storage/pkg/archive/example_changes.go
deleted file mode 100644
index 70f9c5564..000000000
--- a/vendor/github.com/containers/storage/pkg/archive/example_changes.go
+++ /dev/null
@@ -1,97 +0,0 @@
-// +build ignore
-
-// Simple tool to create an archive stream from an old and new directory
-//
-// By default it will stream the comparison of two temporary directories with junk files
-package main
-
-import (
-	"flag"
-	"fmt"
-	"io"
-	"io/ioutil"
-	"os"
-	"path"
-
-	"github.com/containers/storage/pkg/archive"
-	"github.com/sirupsen/logrus"
-)
-
-var (
-	flDebug  = flag.Bool("D", false, "debugging output")
-	flNewDir = flag.String("newdir", "", "")
-	flOldDir = flag.String("olddir", "", "")
-	log      = logrus.New()
-)
-
-func main() {
-	flag.Usage = func() {
-		fmt.Println("Produce a tar from comparing two directory paths. By default a demo tar is created of around 200 files (including hardlinks)")
-		fmt.Printf("%s [OPTIONS]\n", os.Args[0])
-		flag.PrintDefaults()
-	}
-	flag.Parse()
-	log.Out = os.Stderr
-	if (len(os.Getenv("DEBUG")) > 0) || *flDebug {
-		logrus.SetLevel(logrus.DebugLevel)
-	}
-	var newDir, oldDir string
-
-	if len(*flNewDir) == 0 {
-		var err error
-		newDir, err = ioutil.TempDir("", "storage-test-newDir")
-		if err != nil {
-			log.Fatal(err)
-		}
-		defer os.RemoveAll(newDir)
-		if _, err := prepareUntarSourceDirectory(100, newDir, true); err != nil {
-			log.Fatal(err)
-		}
-	} else {
-		newDir = *flNewDir
-	}
-
-	if len(*flOldDir) == 0 {
-		oldDir, err := ioutil.TempDir("", "storage-test-oldDir")
-		if err != nil {
-			log.Fatal(err)
-		}
-		defer os.RemoveAll(oldDir)
-	} else {
-		oldDir = *flOldDir
-	}
-
-	changes, err := archive.ChangesDirs(newDir, oldDir)
-	if err != nil {
-		log.Fatal(err)
-	}
-
-	a, err := archive.ExportChanges(newDir, changes)
-	if err != nil {
-		log.Fatal(err)
-	}
-	defer a.Close()
-
-	i, err := io.Copy(os.Stdout, a)
-	if err != nil && err != io.EOF {
-		log.Fatal(err)
-	}
-	fmt.Fprintf(os.Stderr, "wrote archive of %d bytes", i)
-}
-
-func prepareUntarSourceDirectory(numberOfFiles int, targetPath string, makeLinks bool) (int, error) {
-	fileData := []byte("fooo")
-	for n := 0; n < numberOfFiles; n++ {
-		fileName := fmt.Sprintf("file-%d", n)
-		if err := ioutil.WriteFile(path.Join(targetPath, fileName), fileData, 0700); err != nil {
-			return 0, err
-		}
-		if makeLinks {
-			if err := os.Link(path.Join(targetPath, fileName), path.Join(targetPath, fileName+"-link")); err != nil {
-				return 0, err
-			}
-		}
-	}
-	totalSize := numberOfFiles * len(fileData)
-	return totalSize, nil
-}
diff --git a/vendor/github.com/containers/storage/store.go b/vendor/github.com/containers/storage/store.go
index e0dd1b92f..a166799c6 100644
--- a/vendor/github.com/containers/storage/store.go
+++ b/vendor/github.com/containers/storage/store.go
@@ -1070,7 +1070,7 @@ func (s *store) imageTopLayerForMapping(image *Image, ristore ROImageStore, read
 		}
 		mappedLayer, _, err := rlstore.Put("", parentLayer, nil, layer.MountLabel, nil, &layerOptions, false, nil, rc)
 		if err != nil {
-			return nil, errors.Wrapf(err, "error creating ID-mapped copy of layer %q", parentLayer.ID)
+			return nil, errors.Wrapf(err, "error creating ID-mapped copy of layer %q", layer.ID)
 		}
 		if err = istore.addMappedTopLayer(image.ID, mappedLayer.ID); err != nil {
 			if err2 := rlstore.Delete(mappedLayer.ID); err2 != nil {
diff --git a/vendor/github.com/containers/storage/vendor.conf b/vendor/github.com/containers/storage/vendor.conf
index fa52584d7..936321b12 100644
--- a/vendor/github.com/containers/storage/vendor.conf
+++ b/vendor/github.com/containers/storage/vendor.conf
@@ -23,3 +23,6 @@ golang.org/x/net 7dcfb8076726a3fdd9353b6b8a1f1b6be6811bd6
 golang.org/x/sys 07c182904dbd53199946ba614a412c61d3c548f5
 gotest.tools master
 github.com/google/go-cmp master
+github.com/klauspost/pgzip v1.2.1
+github.com/klauspost/compress v1.4.1
+github.com/klauspost/cpuid v1.2.0
diff --git a/vendor/github.com/klauspost/compress/LICENSE b/vendor/github.com/klauspost/compress/LICENSE
new file mode 100644
index 000000000..744875676
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/LICENSE
@@ -0,0 +1,27 @@
+Copyright (c) 2012 The 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/README.md b/vendor/github.com/klauspost/compress/README.md
new file mode 100644
index 000000000..280d54890
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/README.md
@@ -0,0 +1,160 @@
+# compress

+

+This package is based on an optimized Deflate function, which is used by gzip/zip/zlib packages.

+

+It offers slightly better compression at lower compression settings, and up to 3x faster encoding at highest compression level.

+

+* [High Throughput Benchmark](http://blog.klauspost.com/go-gzipdeflate-benchmarks/).

+* [Small Payload/Webserver Benchmarks](http://blog.klauspost.com/gzip-performance-for-go-webservers/).

+* [Linear Time Compression](http://blog.klauspost.com/constant-time-gzipzip-compression/).

+* [Re-balancing Deflate Compression Levels](https://blog.klauspost.com/rebalancing-deflate-compression-levels/)

+

+[![Build Status](https://travis-ci.org/klauspost/compress.svg?branch=master)](https://travis-ci.org/klauspost/compress)

+[![Sourcegraph Badge](https://sourcegraph.com/github.com/klauspost/compress/-/badge.svg)](https://sourcegraph.com/github.com/klauspost/compress?badge)

+

+# changelog

+

+* 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 progresssion 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.

+

+# usage

+

+The packages are drop-in replacements for standard libraries. Simply replace the import path to use them:

+

+| old import         | new import                              |

+|--------------------|-----------------------------------------|

+| `compress/gzip`    | `github.com/klauspost/compress/gzip`    |

+| `compress/zlib`    | `github.com/klauspost/compress/zlib`    |

+| `archive/zip`      | `github.com/klauspost/compress/zip`     |

+| `compress/flate`   | `github.com/klauspost/compress/flate`   |

+

+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).

+

+# Performance Update 2018

+

+It has been a while since we have been looking at the speed of this package compared to the standard library, so I thought I would re-do my tests and give some overall recommendations based on the current state. All benchmarks have been performed with Go 1.10 on my Desktop Intel(R) Core(TM) i7-2600 CPU @3.40GHz. Since I last ran the tests, I have gotten more RAM, which means tests with big files are no longer limited by my SSD.

+

+The raw results are in my [updated spreadsheet](https://docs.google.com/spreadsheets/d/1nuNE2nPfuINCZJRMt6wFWhKpToF95I47XjSsc-1rbPQ/edit?usp=sharing). Due to cgo changes and upstream updates i could not get the cgo version of gzip to compile. Instead I included the [zstd](https://github.com/datadog/zstd) cgo implementation. If I get cgo gzip to work again, I might replace the results in the sheet.

+

+The columns to take note of are: *MB/s* - the throughput. *Reduction* - the data size reduction in percent of the original. *Rel Speed* relative speed compared to the standard libary at the same level. *Smaller* - how many percent smaller is the compressed output compared to stdlib. Negative means the output was bigger. *Loss* means the loss (or gain) in compression as a percentage difference of the input.

+

+The `gzstd` (standard library gzip) and `gzkp` (this package gzip) only uses one CPU core. [`pgzip`](https://github.com/klauspost/pgzip), [`bgzf`](https://github.com/biogo/hts/bgzf) uses all 4 cores. [`zstd`](https://github.com/DataDog/zstd) uses one core, and is a beast (but not Go, yet).

+

+

+## Overall differences.

+

+There appears to be a roughly 5-10% speed advantage over the standard library when comparing at similar compression levels.

+

+The biggest difference you will see is the result of [re-balancing](https://blog.klauspost.com/rebalancing-deflate-compression-levels/) the compression levels. I wanted by library to give a smoother transition between the compression levels than the standard library.

+

+This package attempts to provide a more smooth transition, where "1" is taking a lot of shortcuts, "5" is the reasonable trade-off and "9" is the "give me the best compression", and the values in between gives something reasonable in between. The standard library has big differences in levels 1-4, but levels 5-9 having no significant gains - often spending a lot more time than can be justified by the achieved compression.

+

+There are links to all the test data in the [spreadsheet](https://docs.google.com/spreadsheets/d/1nuNE2nPfuINCZJRMt6wFWhKpToF95I47XjSsc-1rbPQ/edit?usp=sharing) in the top left field on each tab.

+

+## Web Content

+

+This test set aims to emulate typical use in a web server. The test-set is 4GB data in 53k files, and is a mixture of (mostly) HTML, JS, CSS.

+

+Since level 1 and 9 are close to being the same code, they are quite close. But looking at the levels in-between the differences are quite big.

+

+Looking at level 6, this package is 88% faster, but will output about 6% more data. For a web server, this means you can serve 88% more data, but have to pay for 6% more bandwidth. You can draw your own conclusions on what would be the most expensive for your case.

+

+## Object files

+

+This test is for typical data files stored on a server. In this case it is a collection of Go precompiled objects. They are very compressible.

+

+The picture is similar to the web content, but with small differences since this is very compressible. Levels 2-3 offer good speed, but is sacrificing quite a bit of compression. 

+

+The standard library seems suboptimal on level 3 and 4 - offering both worse compression and speed than level 6 & 7 of this package respectively.

+

+## Highly Compressible File

+

+This is a JSON file with very high redundancy. The reduction starts at 95% on level 1, so in real life terms we are dealing with something like a highly redundant stream of data, etc.

+

+It is definitely visible that we are dealing with specialized content here, so the results are very scattered. This package does not do very well at levels 1-4, but picks up significantly at level 5 and levels 7 and 8 offering great speed for the achieved compression.

+

+So if you know you content is extremely compressible you might want to go slightly higher than the defaults. The standard library has a huge gap between levels 3 and 4 in terms of speed (2.75x slowdown), so it offers little "middle ground".

+

+## Medium-High Compressible

+

+This is a pretty common test corpus: [enwik9](http://mattmahoney.net/dc/textdata.html). It contains the first 10^9 bytes of the English Wikipedia dump on Mar. 3, 2006. This is a very good test of typical text based compression and more data heavy streams.

+

+We see a similar picture here as in "Web Content". On equal levels some compression is sacrificed for more speed. Level 5 seems to be the best trade-off between speed and size, beating stdlib level 3 in both.

+

+## Medium Compressible

+

+I will combine two test sets, one [10GB file set](http://mattmahoney.net/dc/10gb.html) and a VM disk image (~8GB). Both contain different data types and represent a typical backup scenario.

+

+The most notable thing is how quickly the standard libary drops to very low compression speeds around level 5-6 without any big gains in compression. Since this type of data is fairly common, this does not seem like good behavior.

+

+

+## Un-compressible Content

+

+This is mainly a test of how good the algorithms are at detecting un-compressible input. The standard library only offers this feature with very conservative settings at level 1. Obviously there is no reason for the algorithms to try to compress input that cannot be compressed.  The only downside is that it might skip some compressible data on false detections.

+

+

+# linear time compression (huffman only)

+

+This compression library adds a special compression level, named `HuffmanOnly`, which allows near linear time compression. This is done by completely disabling matching of previous data, and only reduce the number of bits to represent each character. 

+

+This means that often used characters, like 'e' and ' ' (space) in text use the fewest bits to represent, and rare characters like '¤' takes more bits to represent. For more information see [wikipedia](https://en.wikipedia.org/wiki/Huffman_coding) or this nice [video](https://youtu.be/ZdooBTdW5bM).

+

+Since this type of compression has much less variance, the compression speed is mostly unaffected by the input data, and is usually more than *180MB/s* for a single core.

+

+The downside is that the compression ratio is usually considerably worse than even the fastest conventional compression. The compression raio can never be better than 8:1 (12.5%). 

+

+The linear time compression can be used as a "better than nothing" mode, where you cannot risk the encoder to slow down on some content. For comparison, the size of the "Twain" text is *233460 bytes* (+29% vs. level 1) and encode speed is 144MB/s (4.5x level 1). So in this case you trade a 30% size increase for a 4 times speedup.

+

+For more information see my blog post on [Fast Linear Time Compression](http://blog.klauspost.com/constant-time-gzipzip-compression/).

+

+This is implemented on Go 1.7 as "Huffman Only" mode, though not exposed for gzip.

+

+

+# snappy package

+

+The standard snappy package has now been improved. This repo contains a copy of the snappy repo.

+

+I would advise to use the standard package: https://github.com/golang/snappy

+

+

+# 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/flate/copy.go b/vendor/github.com/klauspost/compress/flate/copy.go
new file mode 100644
index 000000000..a3200a8f4
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/copy.go
@@ -0,0 +1,32 @@
+// Copyright 2012 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
+
+// forwardCopy is like the built-in copy function except that it always goes
+// forward from the start, even if the dst and src overlap.
+// It is equivalent to:
+//   for i := 0; i < n; i++ {
+//     mem[dst+i] = mem[src+i]
+//   }
+func forwardCopy(mem []byte, dst, src, n int) {
+	if dst <= src {
+		copy(mem[dst:dst+n], mem[src:src+n])
+		return
+	}
+	for {
+		if dst >= src+n {
+			copy(mem[dst:dst+n], mem[src:src+n])
+			return
+		}
+		// There is some forward overlap.  The destination
+		// will be filled with a repeated pattern of mem[src:src+k].
+		// We copy one instance of the pattern here, then repeat.
+		// Each time around this loop k will double.
+		k := dst - src
+		copy(mem[dst:dst+k], mem[src:src+k])
+		n -= k
+		dst += k
+	}
+}
diff --git a/vendor/github.com/klauspost/compress/flate/crc32_amd64.go b/vendor/github.com/klauspost/compress/flate/crc32_amd64.go
new file mode 100644
index 000000000..8298d309a
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/crc32_amd64.go
@@ -0,0 +1,42 @@
+//+build !noasm
+//+build !appengine
+//+build !gccgo
+
+// Copyright 2015, Klaus Post, see LICENSE for details.
+
+package flate
+
+import (
+	"github.com/klauspost/cpuid"
+)
+
+// crc32sse returns a hash for the first 4 bytes of the slice
+// len(a) must be >= 4.
+//go:noescape
+func crc32sse(a []byte) uint32
+
+// crc32sseAll calculates hashes for each 4-byte set in a.
+// dst must be east len(a) - 4 in size.
+// The size is not checked by the assembly.
+//go:noescape
+func crc32sseAll(a []byte, dst []uint32)
+
+// matchLenSSE4 returns the number of matching bytes in a and b
+// up to length 'max'. Both slices must be at least 'max'
+// bytes in size.
+//
+// TODO: drop the "SSE4" name, since it doesn't use any SSE instructions.
+//
+//go:noescape
+func matchLenSSE4(a, b []byte, max int) int
+
+// histogram accumulates a histogram of b in h.
+// h must be at least 256 entries in length,
+// and must be cleared before calling this function.
+//go:noescape
+func histogram(b []byte, h []int32)
+
+// Detect SSE 4.2 feature.
+func init() {
+	useSSE42 = cpuid.CPU.SSE42()
+}
diff --git a/vendor/github.com/klauspost/compress/flate/crc32_amd64.s b/vendor/github.com/klauspost/compress/flate/crc32_amd64.s
new file mode 100644
index 000000000..a79943727
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/crc32_amd64.s
@@ -0,0 +1,214 @@
+//+build !noasm
+//+build !appengine
+//+build !gccgo
+
+// Copyright 2015, Klaus Post, see LICENSE for details.
+
+// func crc32sse(a []byte) uint32
+TEXT ·crc32sse(SB), 4, $0
+	MOVQ a+0(FP), R10
+	XORQ BX, BX
+
+	// CRC32   dword (R10), EBX
+	BYTE $0xF2; BYTE $0x41; BYTE $0x0f
+	BYTE $0x38; BYTE $0xf1; BYTE $0x1a
+
+	MOVL BX, ret+24(FP)
+	RET
+
+// func crc32sseAll(a []byte, dst []uint32)
+TEXT ·crc32sseAll(SB), 4, $0
+	MOVQ  a+0(FP), R8      // R8: src
+	MOVQ  a_len+8(FP), R10 // input length
+	MOVQ  dst+24(FP), R9   // R9: dst
+	SUBQ  $4, R10
+	JS    end
+	JZ    one_crc
+	MOVQ  R10, R13
+	SHRQ  $2, R10          // len/4
+	ANDQ  $3, R13          // len&3
+	XORQ  BX, BX
+	ADDQ  $1, R13
+	TESTQ R10, R10
+	JZ    rem_loop
+
+crc_loop:
+	MOVQ (R8), R11
+	XORQ BX, BX
+	XORQ DX, DX
+	XORQ DI, DI
+	MOVQ R11, R12
+	SHRQ $8, R11
+	MOVQ R12, AX
+	MOVQ R11, CX
+	SHRQ $16, R12
+	SHRQ $16, R11
+	MOVQ R12, SI
+
+	// CRC32   EAX, EBX
+	BYTE $0xF2; BYTE $0x0f
+	BYTE $0x38; BYTE $0xf1; BYTE $0xd8
+
+	// CRC32   ECX, EDX
+	BYTE $0xF2; BYTE $0x0f
+	BYTE $0x38; BYTE $0xf1; BYTE $0xd1
+
+	// CRC32   ESI, EDI
+	BYTE $0xF2; BYTE $0x0f
+	BYTE $0x38; BYTE $0xf1; BYTE $0xfe
+	MOVL BX, (R9)
+	MOVL DX, 4(R9)
+	MOVL DI, 8(R9)
+
+	XORQ BX, BX
+	MOVL R11, AX
+
+	// CRC32   EAX, EBX
+	BYTE $0xF2; BYTE $0x0f
+	BYTE $0x38; BYTE $0xf1; BYTE $0xd8
+	MOVL BX, 12(R9)
+
+	ADDQ $16, R9
+	ADDQ $4, R8
+	XORQ BX, BX
+	SUBQ $1, R10
+	JNZ  crc_loop
+
+rem_loop:
+	MOVL (R8), AX
+
+	// CRC32   EAX, EBX
+	BYTE $0xF2; BYTE $0x0f
+	BYTE $0x38; BYTE $0xf1; BYTE $0xd8
+
+	MOVL BX, (R9)
+	ADDQ $4, R9
+	ADDQ $1, R8
+	XORQ BX, BX
+	SUBQ $1, R13
+	JNZ  rem_loop
+
+end:
+	RET
+
+one_crc:
+	MOVQ $1, R13
+	XORQ BX, BX
+	JMP  rem_loop
+
+// func matchLenSSE4(a, b []byte, max int) int
+TEXT ·matchLenSSE4(SB), 4, $0
+	MOVQ a_base+0(FP), SI
+	MOVQ b_base+24(FP), DI
+	MOVQ DI, DX
+	MOVQ max+48(FP), CX
+
+cmp8:
+	// As long as we are 8 or more bytes before the end of max, we can load and
+	// compare 8 bytes at a time. If those 8 bytes are equal, repeat.
+	CMPQ CX, $8
+	JLT  cmp1
+	MOVQ (SI), AX
+	MOVQ (DI), BX
+	CMPQ AX, BX
+	JNE  bsf
+	ADDQ $8, SI
+	ADDQ $8, DI
+	SUBQ $8, CX
+	JMP  cmp8
+
+bsf:
+	// If those 8 bytes were not equal, XOR the two 8 byte values, and return
+	// the index of the first byte that differs. The BSF instruction finds the
+	// least significant 1 bit, the amd64 architecture is little-endian, and
+	// the shift by 3 converts a bit index to a byte index.
+	XORQ AX, BX
+	BSFQ BX, BX
+	SHRQ $3, BX
+	ADDQ BX, DI
+
+	// Subtract off &b[0] to convert from &b[ret] to ret, and return.
+	SUBQ DX, DI
+	MOVQ DI, ret+56(FP)
+	RET
+
+cmp1:
+	// In the slices' tail, compare 1 byte at a time.
+	CMPQ CX, $0
+	JEQ  matchLenEnd
+	MOVB (SI), AX
+	MOVB (DI), BX
+	CMPB AX, BX
+	JNE  matchLenEnd
+	ADDQ $1, SI
+	ADDQ $1, DI
+	SUBQ $1, CX
+	JMP  cmp1
+
+matchLenEnd:
+	// Subtract off &b[0] to convert from &b[ret] to ret, and return.
+	SUBQ DX, DI
+	MOVQ DI, ret+56(FP)
+	RET
+
+// func histogram(b []byte, h []int32)
+TEXT ·histogram(SB), 4, $0
+	MOVQ b+0(FP), SI     // SI: &b
+	MOVQ b_len+8(FP), R9 // R9: len(b)
+	MOVQ h+24(FP), DI    // DI: Histogram
+	MOVQ R9, R8
+	SHRQ $3, R8
+	JZ   hist1
+	XORQ R11, R11
+
+loop_hist8:
+	MOVQ (SI), R10
+
+	MOVB R10, R11
+	INCL (DI)(R11*4)
+	SHRQ $8, R10
+
+	MOVB R10, R11
+	INCL (DI)(R11*4)
+	SHRQ $8, R10
+
+	MOVB R10, R11
+	INCL (DI)(R11*4)
+	SHRQ $8, R10
+
+	MOVB R10, R11
+	INCL (DI)(R11*4)
+	SHRQ $8, R10
+
+	MOVB R10, R11
+	INCL (DI)(R11*4)
+	SHRQ $8, R10
+
+	MOVB R10, R11
+	INCL (DI)(R11*4)
+	SHRQ $8, R10
+
+	MOVB R10, R11
+	INCL (DI)(R11*4)
+	SHRQ $8, R10
+
+	INCL (DI)(R10*4)
+
+	ADDQ $8, SI
+	DECQ R8
+	JNZ  loop_hist8
+
+hist1:
+	ANDQ $7, R9
+	JZ   end_hist
+	XORQ R10, R10
+
+loop_hist1:
+	MOVB (SI), R10
+	INCL (DI)(R10*4)
+	INCQ SI
+	DECQ R9
+	JNZ  loop_hist1
+
+end_hist:
+	RET
diff --git a/vendor/github.com/klauspost/compress/flate/crc32_noasm.go b/vendor/github.com/klauspost/compress/flate/crc32_noasm.go
new file mode 100644
index 000000000..dcf43bd50
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/crc32_noasm.go
@@ -0,0 +1,35 @@
+//+build !amd64 noasm appengine gccgo
+
+// Copyright 2015, Klaus Post, see LICENSE for details.
+
+package flate
+
+func init() {
+	useSSE42 = false
+}
+
+// crc32sse should never be called.
+func crc32sse(a []byte) uint32 {
+	panic("no assembler")
+}
+
+// crc32sseAll should never be called.
+func crc32sseAll(a []byte, dst []uint32) {
+	panic("no assembler")
+}
+
+// matchLenSSE4 should never be called.
+func matchLenSSE4(a, b []byte, max int) int {
+	panic("no assembler")
+	return 0
+}
+
+// histogram accumulates a histogram of b in h.
+//
+// len(h) must be >= 256, and h's elements must be all zeroes.
+func histogram(b []byte, h []int32) {
+	h = h[:256]
+	for _, t := range b {
+		h[t]++
+	}
+}
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 000000000..9e6e7ff0c
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/deflate.go
@@ -0,0 +1,1353 @@
+// 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 (
+	"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 put into a single flat block, just too
+	// stop things from getting too large.
+	maxFlateBlockTokens = 1 << 14
+	maxStoreBlockSize   = 65535
+	hashBits            = 17 // After 17 performance degrades
+	hashSize            = 1 << hashBits
+	hashMask            = (1 << hashBits) - 1
+	hashShift           = (hashBits + minMatchLength - 1) / minMatchLength
+	maxHashOffset       = 1 << 24
+
+	skipNever = math.MaxInt32
+)
+
+var useSSE42 bool
+
+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 1-4 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},
+	// For levels 5-6 we don't bother trying with lazy matches.
+	// Lazy matching is at least 30% slower, with 1.5% increase.
+	{6, 0, 12, 8, 12, 5},
+	{8, 0, 24, 16, 16, 6},
+	// Levels 7-9 use increasingly more lazy matching
+	// and increasingly stringent conditions for "good enough".
+	{8, 8, 24, 16, skipNever, 7},
+	{10, 16, 24, 64, skipNever, 8},
+	{32, 258, 258, 4096, skipNever, 9},
+}
+
+type compressor struct {
+	compressionLevel
+
+	w          *huffmanBitWriter
+	bulkHasher func([]byte, []uint32)
+
+	// compression algorithm
+	fill func(*compressor, []byte) int // copy data to window
+	step func(*compressor)             // process window
+	sync bool                          // requesting flush
+
+	// 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.
+	chainHead  int
+	hashHead   [hashSize]uint32
+	hashPrev   [windowSize]uint32
+	hashOffset int
+
+	// input window: unprocessed data is window[index:windowEnd]
+	index         int
+	window        []byte
+	windowEnd     int
+	blockStart    int  // window index where current tokens start
+	byteAvailable bool // if true, still need to process window[index-1].
+
+	// queued output tokens
+	tokens tokens
+
+	// deflate state
+	length         int
+	offset         int
+	hash           uint32
+	maxInsertIndex int
+	err            error
+	ii             uint16 // position of last match, intended to overflow to reset.
+
+	snap      snappyEnc
+	hashMatch [maxMatchLength + minMatchLength]uint32
+}
+
+func (d *compressor) fillDeflate(b []byte) int {
+	if d.index >= 2*windowSize-(minMatchLength+maxMatchLength) {
+		// shift the window by windowSize
+		copy(d.window[:], d.window[windowSize:2*windowSize])
+		d.index -= windowSize
+		d.windowEnd -= windowSize
+		if d.blockStart >= windowSize {
+			d.blockStart -= windowSize
+		} else {
+			d.blockStart = math.MaxInt32
+		}
+		d.hashOffset += windowSize
+		if d.hashOffset > maxHashOffset {
+			delta := d.hashOffset - 1
+			d.hashOffset -= delta
+			d.chainHead -= delta
+			// Iterate over slices instead of arrays to avoid copying
+			// the entire table onto the stack (Issue #18625).
+			for i, v := range d.hashPrev[:] {
+				if int(v) > delta {
+					d.hashPrev[i] = uint32(int(v) - delta)
+				} else {
+					d.hashPrev[i] = 0
+				}
+			}
+			for i, v := range d.hashHead[:] {
+				if int(v) > delta {
+					d.hashHead[i] = uint32(int(v) - delta)
+				} else {
+					d.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.tokens[:tok.n], eof, window)
+		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)
+			} else {
+				// Write a dynamic huffman block.
+				d.w.writeBlockDynamic(tok.tokens[:tok.n], eof, window)
+			}
+		} else {
+			d.w.writeBlock(tok.tokens[:tok.n], 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 mode,
+	// use constant or Snappy compression.
+	switch d.compressionLevel.level {
+	case 0, 1, 2:
+		return
+	}
+	// 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 := d.hashMatch[:dstSize]
+		d.bulkHasher(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.
+			d.hashPrev[di&windowMask] = d.hashHead[newH]
+			// Set the head of the hash chain to us.
+			d.hashHead[newH] = uint32(di + d.hashOffset)
+		}
+		d.hash = newH
+	}
+	// Update window information.
+	d.windowEnd += n
+	d.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 = d.index, prevHead = d.chainHead-d.hashOffset, prevLength=minMatchLength-1, lookahead
+func (d *compressor) findMatch(pos int, prevHead int, prevLength 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 = prevLength
+	if length >= d.good {
+		tries >>= 2
+	}
+
+	wEnd := win[pos+length]
+	wPos := win[pos:]
+	minIndex := pos - windowSize
+
+	for i := prevHead; tries > 0; tries-- {
+		if wEnd == win[i+length] {
+			n := matchLen(win[i:], wPos, minMatchLook)
+
+			if n > length && (n > minMatchLength || pos-i <= 4096) {
+				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.hashPrev[i&windowMask]) - d.hashOffset
+		if i < minIndex || i < 0 {
+			break
+		}
+	}
+	return
+}
+
+// Try to find a match starting at index whose length is greater than prevSize.
+// We only look at chainCount possibilities before giving up.
+// pos = d.index, prevHead = d.chainHead-d.hashOffset, prevLength=minMatchLength-1, lookahead
+func (d *compressor) findMatchSSE(pos int, prevHead int, prevLength 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 = prevLength
+	if length >= d.good {
+		tries >>= 2
+	}
+
+	wEnd := win[pos+length]
+	wPos := win[pos:]
+	minIndex := pos - windowSize
+
+	for i := prevHead; tries > 0; tries-- {
+		if wEnd == win[i+length] {
+			n := matchLenSSE4(win[i:], wPos, minMatchLook)
+
+			if n > length && (n > minMatchLength || pos-i <= 4096) {
+				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.hashPrev[i&windowMask]) - d.hashOffset
+		if i < minIndex || i < 0 {
+			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
+}
+
+const hashmul = 0x1e35a7bd
+
+// 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 ((uint32(b[3]) | uint32(b[2])<<8 | uint32(b[1])<<16 | uint32(b[0])<<24) * hashmul) >> (32 - hashBits)
+}
+
+// bulkHash4 will compute hashes using the same
+// algorithm as hash4
+func bulkHash4(b []byte, dst []uint32) {
+	if len(b) < minMatchLength {
+		return
+	}
+	hb := uint32(b[3]) | uint32(b[2])<<8 | uint32(b[1])<<16 | uint32(b[0])<<24
+	dst[0] = (hb * hashmul) >> (32 - hashBits)
+	end := len(b) - minMatchLength + 1
+	for i := 1; i < end; i++ {
+		hb = (hb << 8) | uint32(b[i+3])
+		dst[i] = (hb * hashmul) >> (32 - hashBits)
+	}
+}
+
+// matchLen returns the number of matching bytes in a and b
+// up to length 'max'. Both slices must be at least 'max'
+// bytes in size.
+func matchLen(a, b []byte, max int) int {
+	a = a[:max]
+	b = b[:len(a)]
+	for i, av := range a {
+		if b[i] != av {
+			return i
+		}
+	}
+	return max
+}
+
+func (d *compressor) initDeflate() {
+	d.window = make([]byte, 2*windowSize)
+	d.hashOffset = 1
+	d.length = minMatchLength - 1
+	d.offset = 0
+	d.byteAvailable = false
+	d.index = 0
+	d.hash = 0
+	d.chainHead = -1
+	d.bulkHasher = bulkHash4
+	if useSSE42 {
+		d.bulkHasher = crc32sseAll
+	}
+}
+
+// Assumes that d.fastSkipHashing != skipNever,
+// otherwise use deflateLazy
+func (d *compressor) deflate() {
+
+	// Sanity enables additional runtime tests.
+	// It's intended to be used during development
+	// to supplement the currently ad-hoc unit tests.
+	const sanity = false
+
+	if d.windowEnd-d.index < minMatchLength+maxMatchLength && !d.sync {
+		return
+	}
+
+	d.maxInsertIndex = d.windowEnd - (minMatchLength - 1)
+	if d.index < d.maxInsertIndex {
+		d.hash = hash4(d.window[d.index : d.index+minMatchLength])
+	}
+
+	for {
+		if sanity && d.index > d.windowEnd {
+			panic("index > windowEnd")
+		}
+		lookahead := d.windowEnd - d.index
+		if lookahead < minMatchLength+maxMatchLength {
+			if !d.sync {
+				return
+			}
+			if sanity && d.index > d.windowEnd {
+				panic("index > windowEnd")
+			}
+			if lookahead == 0 {
+				if d.tokens.n > 0 {
+					if d.err = d.writeBlockSkip(d.tokens, d.index, false); d.err != nil {
+						return
+					}
+					d.tokens.n = 0
+				}
+				return
+			}
+		}
+		if d.index < d.maxInsertIndex {
+			// Update the hash
+			d.hash = hash4(d.window[d.index : d.index+minMatchLength])
+			ch := d.hashHead[d.hash&hashMask]
+			d.chainHead = int(ch)
+			d.hashPrev[d.index&windowMask] = ch
+			d.hashHead[d.hash&hashMask] = uint32(d.index + d.hashOffset)
+		}
+		d.length = minMatchLength - 1
+		d.offset = 0
+		minIndex := d.index - windowSize
+		if minIndex < 0 {
+			minIndex = 0
+		}
+
+		if d.chainHead-d.hashOffset >= minIndex && lookahead > minMatchLength-1 {
+			if newLength, newOffset, ok := d.findMatch(d.index, d.chainHead-d.hashOffset, minMatchLength-1, lookahead); ok {
+				d.length = newLength
+				d.offset = newOffset
+			}
+		}
+		if d.length >= minMatchLength {
+			d.ii = 0
+			// There was a match at the previous step, and the current match is
+			// not better. Output the previous match.
+			// "d.length-3" should NOT be "d.length-minMatchLength", since the format always assume 3
+			d.tokens.tokens[d.tokens.n] = matchToken(uint32(d.length-3), uint32(d.offset-minOffsetSize))
+			d.tokens.n++
+			// 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.
+			if d.length <= d.fastSkipHashing {
+				var newIndex int
+				newIndex = d.index + d.length
+				// Calculate missing hashes
+				end := newIndex
+				if end > d.maxInsertIndex {
+					end = d.maxInsertIndex
+				}
+				end += minMatchLength - 1
+				startindex := d.index + 1
+				if startindex > d.maxInsertIndex {
+					startindex = d.maxInsertIndex
+				}
+				tocheck := d.window[startindex:end]
+				dstSize := len(tocheck) - minMatchLength + 1
+				if dstSize > 0 {
+					dst := d.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.
+						d.hashPrev[di&windowMask] = d.hashHead[newH]
+						// Set the head of the hash chain to us.
+						d.hashHead[newH] = uint32(di + d.hashOffset)
+					}
+					d.hash = newH
+				}
+				d.index = newIndex
+			} else {
+				// For matches this long, we don't bother inserting each individual
+				// item into the table.
+				d.index += d.length
+				if d.index < d.maxInsertIndex {
+					d.hash = hash4(d.window[d.index : d.index+minMatchLength])
+				}
+			}
+			if d.tokens.n == maxFlateBlockTokens {
+				// The block includes the current character
+				if d.err = d.writeBlockSkip(d.tokens, d.index, false); d.err != nil {
+					return
+				}
+				d.tokens.n = 0
+			}
+		} else {
+			d.ii++
+			end := d.index + int(d.ii>>uint(d.fastSkipHashing)) + 1
+			if end > d.windowEnd {
+				end = d.windowEnd
+			}
+			for i := d.index; i < end; i++ {
+				d.tokens.tokens[d.tokens.n] = literalToken(uint32(d.window[i]))
+				d.tokens.n++
+				if d.tokens.n == maxFlateBlockTokens {
+					if d.err = d.writeBlockSkip(d.tokens, i+1, false); d.err != nil {
+						return
+					}
+					d.tokens.n = 0
+				}
+			}
+			d.index = end
+		}
+	}
+}
+
+// deflateLazy is the same as deflate, but with d.fastSkipHashing == skipNever,
+// meaning it always has lazy matching on.
+func (d *compressor) deflateLazy() {
+	// Sanity enables additional runtime tests.
+	// It's intended to be used during development
+	// to supplement the currently ad-hoc unit tests.
+	const sanity = false
+
+	if d.windowEnd-d.index < minMatchLength+maxMatchLength && !d.sync {
+		return
+	}
+
+	d.maxInsertIndex = d.windowEnd - (minMatchLength - 1)
+	if d.index < d.maxInsertIndex {
+		d.hash = hash4(d.window[d.index : d.index+minMatchLength])
+	}
+
+	for {
+		if sanity && d.index > d.windowEnd {
+			panic("index > windowEnd")
+		}
+		lookahead := d.windowEnd - d.index
+		if lookahead < minMatchLength+maxMatchLength {
+			if !d.sync {
+				return
+			}
+			if sanity && d.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.tokens[d.tokens.n] = literalToken(uint32(d.window[d.index-1]))
+					d.tokens.n++
+					d.byteAvailable = false
+				}
+				if d.tokens.n > 0 {
+					if d.err = d.writeBlock(d.tokens, d.index, false); d.err != nil {
+						return
+					}
+					d.tokens.n = 0
+				}
+				return
+			}
+		}
+		if d.index < d.maxInsertIndex {
+			// Update the hash
+			d.hash = hash4(d.window[d.index : d.index+minMatchLength])
+			ch := d.hashHead[d.hash&hashMask]
+			d.chainHead = int(ch)
+			d.hashPrev[d.index&windowMask] = ch
+			d.hashHead[d.hash&hashMask] = uint32(d.index + d.hashOffset)
+		}
+		prevLength := d.length
+		prevOffset := d.offset
+		d.length = minMatchLength - 1
+		d.offset = 0
+		minIndex := d.index - windowSize
+		if minIndex < 0 {
+			minIndex = 0
+		}
+
+		if d.chainHead-d.hashOffset >= minIndex && lookahead > prevLength && prevLength < d.lazy {
+			if newLength, newOffset, ok := d.findMatch(d.index, d.chainHead-d.hashOffset, minMatchLength-1, lookahead); ok {
+				d.length = newLength
+				d.offset = newOffset
+			}
+		}
+		if prevLength >= minMatchLength && d.length <= prevLength {
+			// There was a match at the previous step, and the current match is
+			// not better. Output the previous match.
+			d.tokens.tokens[d.tokens.n] = matchToken(uint32(prevLength-3), uint32(prevOffset-minOffsetSize))
+			d.tokens.n++
+
+			// 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.
+			var newIndex int
+			newIndex = d.index + prevLength - 1
+			// Calculate missing hashes
+			end := newIndex
+			if end > d.maxInsertIndex {
+				end = d.maxInsertIndex
+			}
+			end += minMatchLength - 1
+			startindex := d.index + 1
+			if startindex > d.maxInsertIndex {
+				startindex = d.maxInsertIndex
+			}
+			tocheck := d.window[startindex:end]
+			dstSize := len(tocheck) - minMatchLength + 1
+			if dstSize > 0 {
+				dst := d.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.
+					d.hashPrev[di&windowMask] = d.hashHead[newH]
+					// Set the head of the hash chain to us.
+					d.hashHead[newH] = uint32(di + d.hashOffset)
+				}
+				d.hash = newH
+			}
+
+			d.index = newIndex
+			d.byteAvailable = false
+			d.length = minMatchLength - 1
+			if d.tokens.n == maxFlateBlockTokens {
+				// The block includes the current character
+				if d.err = d.writeBlock(d.tokens, d.index, false); d.err != nil {
+					return
+				}
+				d.tokens.n = 0
+			}
+		} else {
+			// Reset, if we got a match this run.
+			if d.length >= minMatchLength {
+				d.ii = 0
+			}
+			// We have a byte waiting. Emit it.
+			if d.byteAvailable {
+				d.ii++
+				d.tokens.tokens[d.tokens.n] = literalToken(uint32(d.window[d.index-1]))
+				d.tokens.n++
+				if d.tokens.n == maxFlateBlockTokens {
+					if d.err = d.writeBlock(d.tokens, d.index, false); d.err != nil {
+						return
+					}
+					d.tokens.n = 0
+				}
+				d.index++
+
+				// If we have a long run of no matches, skip additional bytes
+				// Resets when d.ii overflows after 64KB.
+				if d.ii > 31 {
+					n := int(d.ii >> 5)
+					for j := 0; j < n; j++ {
+						if d.index >= d.windowEnd-1 {
+							break
+						}
+
+						d.tokens.tokens[d.tokens.n] = literalToken(uint32(d.window[d.index-1]))
+						d.tokens.n++
+						if d.tokens.n == maxFlateBlockTokens {
+							if d.err = d.writeBlock(d.tokens, d.index, false); d.err != nil {
+								return
+							}
+							d.tokens.n = 0
+						}
+						d.index++
+					}
+					// Flush last byte
+					d.tokens.tokens[d.tokens.n] = literalToken(uint32(d.window[d.index-1]))
+					d.tokens.n++
+					d.byteAvailable = false
+					// d.length = minMatchLength - 1 // not needed, since d.ii is reset above, so it should never be > minMatchLength
+					if d.tokens.n == maxFlateBlockTokens {
+						if d.err = d.writeBlock(d.tokens, d.index, false); d.err != nil {
+							return
+						}
+						d.tokens.n = 0
+					}
+				}
+			} else {
+				d.index++
+				d.byteAvailable = true
+			}
+		}
+	}
+}
+
+// Assumes that d.fastSkipHashing != skipNever,
+// otherwise use deflateLazySSE
+func (d *compressor) deflateSSE() {
+
+	// Sanity enables additional runtime tests.
+	// It's intended to be used during development
+	// to supplement the currently ad-hoc unit tests.
+	const sanity = false
+
+	if d.windowEnd-d.index < minMatchLength+maxMatchLength && !d.sync {
+		return
+	}
+
+	d.maxInsertIndex = d.windowEnd - (minMatchLength - 1)
+	if d.index < d.maxInsertIndex {
+		d.hash = crc32sse(d.window[d.index:d.index+minMatchLength]) & hashMask
+	}
+
+	for {
+		if sanity && d.index > d.windowEnd {
+			panic("index > windowEnd")
+		}
+		lookahead := d.windowEnd - d.index
+		if lookahead < minMatchLength+maxMatchLength {
+			if !d.sync {
+				return
+			}
+			if sanity && d.index > d.windowEnd {
+				panic("index > windowEnd")
+			}
+			if lookahead == 0 {
+				if d.tokens.n > 0 {
+					if d.err = d.writeBlockSkip(d.tokens, d.index, false); d.err != nil {
+						return
+					}
+					d.tokens.n = 0
+				}
+				return
+			}
+		}
+		if d.index < d.maxInsertIndex {
+			// Update the hash
+			d.hash = crc32sse(d.window[d.index:d.index+minMatchLength]) & hashMask
+			ch := d.hashHead[d.hash]
+			d.chainHead = int(ch)
+			d.hashPrev[d.index&windowMask] = ch
+			d.hashHead[d.hash] = uint32(d.index + d.hashOffset)
+		}
+		d.length = minMatchLength - 1
+		d.offset = 0
+		minIndex := d.index - windowSize
+		if minIndex < 0 {
+			minIndex = 0
+		}
+
+		if d.chainHead-d.hashOffset >= minIndex && lookahead > minMatchLength-1 {
+			if newLength, newOffset, ok := d.findMatchSSE(d.index, d.chainHead-d.hashOffset, minMatchLength-1, lookahead); ok {
+				d.length = newLength
+				d.offset = newOffset
+			}
+		}
+		if d.length >= minMatchLength {
+			d.ii = 0
+			// There was a match at the previous step, and the current match is
+			// not better. Output the previous match.
+			// "d.length-3" should NOT be "d.length-minMatchLength", since the format always assume 3
+			d.tokens.tokens[d.tokens.n] = matchToken(uint32(d.length-3), uint32(d.offset-minOffsetSize))
+			d.tokens.n++
+			// 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.
+			if d.length <= d.fastSkipHashing {
+				var newIndex int
+				newIndex = d.index + d.length
+				// Calculate missing hashes
+				end := newIndex
+				if end > d.maxInsertIndex {
+					end = d.maxInsertIndex
+				}
+				end += minMatchLength - 1
+				startindex := d.index + 1
+				if startindex > d.maxInsertIndex {
+					startindex = d.maxInsertIndex
+				}
+				tocheck := d.window[startindex:end]
+				dstSize := len(tocheck) - minMatchLength + 1
+				if dstSize > 0 {
+					dst := d.hashMatch[:dstSize]
+
+					crc32sseAll(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.
+						d.hashPrev[di&windowMask] = d.hashHead[newH]
+						// Set the head of the hash chain to us.
+						d.hashHead[newH] = uint32(di + d.hashOffset)
+					}
+					d.hash = newH
+				}
+				d.index = newIndex
+			} else {
+				// For matches this long, we don't bother inserting each individual
+				// item into the table.
+				d.index += d.length
+				if d.index < d.maxInsertIndex {
+					d.hash = crc32sse(d.window[d.index:d.index+minMatchLength]) & hashMask
+				}
+			}
+			if d.tokens.n == maxFlateBlockTokens {
+				// The block includes the current character
+				if d.err = d.writeBlockSkip(d.tokens, d.index, false); d.err != nil {
+					return
+				}
+				d.tokens.n = 0
+			}
+		} else {
+			d.ii++
+			end := d.index + int(d.ii>>5) + 1
+			if end > d.windowEnd {
+				end = d.windowEnd
+			}
+			for i := d.index; i < end; i++ {
+				d.tokens.tokens[d.tokens.n] = literalToken(uint32(d.window[i]))
+				d.tokens.n++
+				if d.tokens.n == maxFlateBlockTokens {
+					if d.err = d.writeBlockSkip(d.tokens, i+1, false); d.err != nil {
+						return
+					}
+					d.tokens.n = 0
+				}
+			}
+			d.index = end
+		}
+	}
+}
+
+// deflateLazy is the same as deflate, but with d.fastSkipHashing == skipNever,
+// meaning it always has lazy matching on.
+func (d *compressor) deflateLazySSE() {
+	// Sanity enables additional runtime tests.
+	// It's intended to be used during development
+	// to supplement the currently ad-hoc unit tests.
+	const sanity = false
+
+	if d.windowEnd-d.index < minMatchLength+maxMatchLength && !d.sync {
+		return
+	}
+
+	d.maxInsertIndex = d.windowEnd - (minMatchLength - 1)
+	if d.index < d.maxInsertIndex {
+		d.hash = crc32sse(d.window[d.index:d.index+minMatchLength]) & hashMask
+	}
+
+	for {
+		if sanity && d.index > d.windowEnd {
+			panic("index > windowEnd")
+		}
+		lookahead := d.windowEnd - d.index
+		if lookahead < minMatchLength+maxMatchLength {
+			if !d.sync {
+				return
+			}
+			if sanity && d.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.tokens[d.tokens.n] = literalToken(uint32(d.window[d.index-1]))
+					d.tokens.n++
+					d.byteAvailable = false
+				}
+				if d.tokens.n > 0 {
+					if d.err = d.writeBlock(d.tokens, d.index, false); d.err != nil {
+						return
+					}
+					d.tokens.n = 0
+				}
+				return
+			}
+		}
+		if d.index < d.maxInsertIndex {
+			// Update the hash
+			d.hash = crc32sse(d.window[d.index:d.index+minMatchLength]) & hashMask
+			ch := d.hashHead[d.hash]
+			d.chainHead = int(ch)
+			d.hashPrev[d.index&windowMask] = ch
+			d.hashHead[d.hash] = uint32(d.index + d.hashOffset)
+		}
+		prevLength := d.length
+		prevOffset := d.offset
+		d.length = minMatchLength - 1
+		d.offset = 0
+		minIndex := d.index - windowSize
+		if minIndex < 0 {
+			minIndex = 0
+		}
+
+		if d.chainHead-d.hashOffset >= minIndex && lookahead > prevLength && prevLength < d.lazy {
+			if newLength, newOffset, ok := d.findMatchSSE(d.index, d.chainHead-d.hashOffset, minMatchLength-1, lookahead); ok {
+				d.length = newLength
+				d.offset = newOffset
+			}
+		}
+		if prevLength >= minMatchLength && d.length <= prevLength {
+			// There was a match at the previous step, and the current match is
+			// not better. Output the previous match.
+			d.tokens.tokens[d.tokens.n] = matchToken(uint32(prevLength-3), uint32(prevOffset-minOffsetSize))
+			d.tokens.n++
+
+			// 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.
+			var newIndex int
+			newIndex = d.index + prevLength - 1
+			// Calculate missing hashes
+			end := newIndex
+			if end > d.maxInsertIndex {
+				end = d.maxInsertIndex
+			}
+			end += minMatchLength - 1
+			startindex := d.index + 1
+			if startindex > d.maxInsertIndex {
+				startindex = d.maxInsertIndex
+			}
+			tocheck := d.window[startindex:end]
+			dstSize := len(tocheck) - minMatchLength + 1
+			if dstSize > 0 {
+				dst := d.hashMatch[:dstSize]
+				crc32sseAll(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.
+					d.hashPrev[di&windowMask] = d.hashHead[newH]
+					// Set the head of the hash chain to us.
+					d.hashHead[newH] = uint32(di + d.hashOffset)
+				}
+				d.hash = newH
+			}
+
+			d.index = newIndex
+			d.byteAvailable = false
+			d.length = minMatchLength - 1
+			if d.tokens.n == maxFlateBlockTokens {
+				// The block includes the current character
+				if d.err = d.writeBlock(d.tokens, d.index, false); d.err != nil {
+					return
+				}
+				d.tokens.n = 0
+			}
+		} else {
+			// Reset, if we got a match this run.
+			if d.length >= minMatchLength {
+				d.ii = 0
+			}
+			// We have a byte waiting. Emit it.
+			if d.byteAvailable {
+				d.ii++
+				d.tokens.tokens[d.tokens.n] = literalToken(uint32(d.window[d.index-1]))
+				d.tokens.n++
+				if d.tokens.n == maxFlateBlockTokens {
+					if d.err = d.writeBlock(d.tokens, d.index, false); d.err != nil {
+						return
+					}
+					d.tokens.n = 0
+				}
+				d.index++
+
+				// If we have a long run of no matches, skip additional bytes
+				// Resets when d.ii overflows after 64KB.
+				if d.ii > 31 {
+					n := int(d.ii >> 6)
+					for j := 0; j < n; j++ {
+						if d.index >= d.windowEnd-1 {
+							break
+						}
+
+						d.tokens.tokens[d.tokens.n] = literalToken(uint32(d.window[d.index-1]))
+						d.tokens.n++
+						if d.tokens.n == maxFlateBlockTokens {
+							if d.err = d.writeBlock(d.tokens, d.index, false); d.err != nil {
+								return
+							}
+							d.tokens.n = 0
+						}
+						d.index++
+					}
+					// Flush last byte
+					d.tokens.tokens[d.tokens.n] = literalToken(uint32(d.window[d.index-1]))
+					d.tokens.n++
+					d.byteAvailable = false
+					// d.length = minMatchLength - 1 // not needed, since d.ii is reset above, so it should never be > minMatchLength
+					if d.tokens.n == maxFlateBlockTokens {
+						if d.err = d.writeBlock(d.tokens, d.index, false); d.err != nil {
+							return
+						}
+						d.tokens.n = 0
+					}
+				}
+			} else {
+				d.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.err = d.w.err
+	d.windowEnd = 0
+}
+
+// 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) storeSnappy() {
+	// We only compress if we have maxStoreBlockSize.
+	if d.windowEnd < maxStoreBlockSize {
+		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])
+				d.tokens.n = 0
+				d.windowEnd = 0
+			} else {
+				d.w.writeBlockHuff(false, d.window[:d.windowEnd])
+				d.err = d.w.err
+			}
+			d.tokens.n = 0
+			d.windowEnd = 0
+			d.snap.Reset()
+			return
+		}
+	}
+
+	d.snap.Encode(&d.tokens, d.window[:d.windowEnd])
+	// If we made zero matches, store the block as is.
+	if int(d.tokens.n) == d.windowEnd {
+		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.err = d.w.err
+	} else {
+		d.w.writeBlockDynamic(d.tokens.tokens[:d.tokens.n], false, d.window[:d.windowEnd])
+		d.err = d.w.err
+	}
+	d.tokens.n = 0
+	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 {
+		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.window = make([]byte, maxStoreBlockSize)
+		d.fill = (*compressor).fillBlock
+		d.step = (*compressor).storeHuff
+	case level >= 1 && level <= 4:
+		d.snap = newSnappy(level)
+		d.window = make([]byte, maxStoreBlockSize)
+		d.fill = (*compressor).fillBlock
+		d.step = (*compressor).storeSnappy
+	case level == DefaultCompression:
+		level = 5
+		fallthrough
+	case 5 <= level && level <= 9:
+		d.compressionLevel = levels[level]
+		d.initDeflate()
+		d.fill = (*compressor).fillDeflate
+		if d.fastSkipHashing == skipNever {
+			if useSSE42 {
+				d.step = (*compressor).deflateLazySSE
+			} else {
+				d.step = (*compressor).deflateLazy
+			}
+		} else {
+			if useSSE42 {
+				d.step = (*compressor).deflateSSE
+			} else {
+				d.step = (*compressor).deflate
+
+			}
+		}
+	default:
+		return fmt.Errorf("flate: invalid compression level %d: want value in range [-2, 9]", 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.snap != nil {
+		d.snap.Reset()
+		d.windowEnd = 0
+		d.tokens.n = 0
+		return
+	}
+	switch d.compressionLevel.chain {
+	case 0:
+		// level was NoCompression or ConstantCompresssion.
+		d.windowEnd = 0
+	default:
+		d.chainHead = -1
+		for i := range d.hashHead {
+			d.hashHead[i] = 0
+		}
+		for i := range d.hashPrev {
+			d.hashPrev[i] = 0
+		}
+		d.hashOffset = 1
+		d.index, d.windowEnd = 0, 0
+		d.blockStart, d.byteAvailable = 0, false
+		d.tokens.n = 0
+		d.length = minMatchLength - 1
+		d.offset = 0
+		d.hash = 0
+		d.ii = 0
+		d.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()
+	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) {
+	dw := &dictWriter{w}
+	zw, err := NewWriter(dw, 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
+}
+
+type dictWriter struct {
+	w io.Writer
+}
+
+func (w *dictWriter) Write(b []byte) (n int, err error) {
+	return w.w.Write(b)
+}
+
+// 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 dw, ok := w.d.w.writer.(*dictWriter); ok {
+		// w was created with NewWriterDict
+		dw.w = dst
+		w.d.reset(dw)
+		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 000000000..71c75a065
--- /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/huffman_bit_writer.go b/vendor/github.com/klauspost/compress/flate/huffman_bit_writer.go
new file mode 100644
index 000000000..f9b2a699a
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/huffman_bit_writer.go
@@ -0,0 +1,701 @@
+// 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 (
+	"io"
+)
+
+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
+
+	// 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 = 240
+
+	// bufferSize is the actual output byte buffer size.
+	// It must have additional headroom for a flush
+	// which can contain up to 8 bytes.
+	bufferSize = bufferFlushSize + 8
+)
+
+// The number of extra bits needed by length code X - LENGTH_CODES_START.
+var lengthExtraBits = []int8{
+	/* 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 = []uint32{
+	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,
+}
+
+// offset code word extra bits.
+var offsetExtraBits = []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, 15, 15, 16, 16, 17, 17, 18, 18, 19, 19, 20, 20,
+}
+
+var offsetBase = []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, 0x010000, 0x018000, 0x020000,
+	0x030000, 0x040000, 0x060000, 0x080000, 0x0c0000,
+	0x100000, 0x180000, 0x200000, 0x300000,
+}
+
+// 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           uint
+	bytes           [bufferSize]byte
+	codegenFreq     [codegenCodeCount]int32
+	nbytes          int
+	literalFreq     []int32
+	offsetFreq      []int32
+	codegen         []uint8
+	literalEncoding *huffmanEncoder
+	offsetEncoding  *huffmanEncoder
+	codegenEncoding *huffmanEncoder
+	err             error
+}
+
+func newHuffmanBitWriter(w io.Writer) *huffmanBitWriter {
+	return &huffmanBitWriter{
+		writer:          w,
+		literalFreq:     make([]int32, maxNumLit),
+		offsetFreq:      make([]int32, offsetCodeCount),
+		codegen:         make([]uint8, maxNumLit+offsetCodeCount+1),
+		literalEncoding: newHuffmanEncoder(maxNumLit),
+		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.bytes = [bufferSize]byte{}
+}
+
+func (w *huffmanBitWriter) flush() {
+	if w.err != nil {
+		w.nbits = 0
+		return
+	}
+	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 uint) {
+	if w.err != nil {
+		return
+	}
+	w.bits |= uint64(b) << w.nbits
+	w.nbits += nb
+	if w.nbits >= 48 {
+		bits := w.bits
+		w.bits >>= 48
+		w.nbits -= 48
+		n := w.nbytes
+		bytes := w.bytes[n : n+6]
+		bytes[0] = byte(bits)
+		bytes[1] = byte(bits >> 8)
+		bytes[2] = byte(bits >> 16)
+		bytes[3] = byte(bits >> 24)
+		bytes[4] = byte(bits >> 32)
+		bytes[5] = byte(bits >> 40)
+		n += 6
+		if n >= bufferFlushSize {
+			w.write(w.bytes[:n])
+			n = 0
+		}
+		w.nbytes = n
+	}
+}
+
+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] = uint8(litEnc.codes[i].len)
+	}
+
+	cgnl = codegen[numLiterals : numLiterals+numOffsets]
+	for i := range cgnl {
+		cgnl[i] = uint8(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
+}
+
+// dynamicSize returns the size of dynamically encoded data in bits.
+func (w *huffmanBitWriter) dynamicSize(litEnc, offEnc *huffmanEncoder, extraBits int) (size, numCodegens int) {
+	numCodegens = len(w.codegenFreq)
+	for numCodegens > 4 && w.codegenFreq[codegenOrder[numCodegens-1]] == 0 {
+		numCodegens--
+	}
+	header := 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
+	size = header +
+		litEnc.bitLength(w.literalFreq) +
+		offEnc.bitLength(w.offsetFreq) +
+		extraBits
+
+	return size, numCodegens
+}
+
+// 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) {
+	if w.err != nil {
+		return
+	}
+	w.bits |= uint64(c.code) << w.nbits
+	w.nbits += uint(c.len)
+	if w.nbits >= 48 {
+		bits := w.bits
+		w.bits >>= 48
+		w.nbits -= 48
+		n := w.nbytes
+		bytes := w.bytes[n : n+6]
+		bytes[0] = byte(bits)
+		bytes[1] = byte(bits >> 8)
+		bytes[2] = byte(bits >> 16)
+		bytes[3] = byte(bits >> 24)
+		bytes[4] = byte(bits >> 32)
+		bytes[5] = byte(bits >> 40)
+		n += 6
+		if n >= bufferFlushSize {
+			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 int = int(w.codegen[i])
+		i++
+		if codeWord == badCode {
+			break
+		}
+		w.writeCode(w.codegenEncoding.codes[uint32(codeWord)])
+
+		switch codeWord {
+		case 16:
+			w.writeBits(int32(w.codegen[i]), 2)
+			i++
+			break
+		case 17:
+			w.writeBits(int32(w.codegen[i]), 3)
+			i++
+			break
+		case 18:
+			w.writeBits(int32(w.codegen[i]), 7)
+			i++
+			break
+		}
+	}
+}
+
+func (w *huffmanBitWriter) writeStoredHeader(length int, isEof bool) {
+	if w.err != nil {
+		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
+	}
+	// 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 []token, eof bool, input []byte) {
+	if w.err != nil {
+		return
+	}
+
+	tokens = append(tokens, endBlockMarker)
+	numLiterals, numOffsets := w.indexTokens(tokens)
+
+	var extraBits int
+	storedSize, storable := w.storedSize(input)
+	if storable {
+		// We only bother calculating the costs of the extra bits required by
+		// the length of offset fields (which will be the same for both fixed
+		// and dynamic encoding), if we need to compare those two encodings
+		// against stored encoding.
+		for lengthCode := lengthCodesStart + 8; lengthCode < numLiterals; lengthCode++ {
+			// First eight length codes have extra size = 0.
+			extraBits += int(w.literalFreq[lengthCode]) * int(lengthExtraBits[lengthCode-lengthCodesStart])
+		}
+		for offsetCode := 4; offsetCode < numOffsets; offsetCode++ {
+			// First four offset codes have extra size = 0.
+			extraBits += int(w.offsetFreq[offsetCode]) * int(offsetExtraBits[offsetCode])
+		}
+	}
+
+	// Figure out smallest code.
+	// Fixed Huffman baseline.
+	var literalEncoding = fixedLiteralEncoding
+	var offsetEncoding = fixedOffsetEncoding
+	var 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, 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 []token, eof bool, input []byte) {
+	if w.err != nil {
+		return
+	}
+
+	tokens = append(tokens, endBlockMarker)
+	numLiterals, numOffsets := w.indexTokens(tokens)
+
+	// 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)
+	size, numCodegens := w.dynamicSize(w.literalEncoding, w.offsetEncoding, 0)
+
+	// Store bytes, if we don't get a reasonable improvement.
+	if ssize, storable := w.storedSize(input); storable && ssize < (size+size>>4) {
+		w.writeStoredHeader(len(input), eof)
+		w.writeBytes(input)
+		return
+	}
+
+	// Write Huffman table.
+	w.writeDynamicHeader(numLiterals, numOffsets, numCodegens, eof)
+
+	// Write the tokens.
+	w.writeTokens(tokens, w.literalEncoding.codes, w.offsetEncoding.codes)
+}
+
+// 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(tokens []token) (numLiterals, numOffsets int) {
+	for i := range w.literalFreq {
+		w.literalFreq[i] = 0
+	}
+	for i := range w.offsetFreq {
+		w.offsetFreq[i] = 0
+	}
+
+	for _, t := range tokens {
+		if t < matchType {
+			w.literalFreq[t.literal()]++
+			continue
+		}
+		length := t.length()
+		offset := t.offset()
+		w.literalFreq[lengthCodesStart+lengthCode(length)]++
+		w.offsetFreq[offsetCode(offset)]++
+	}
+
+	// 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
+	}
+	w.literalEncoding.generate(w.literalFreq, 15)
+	w.offsetEncoding.generate(w.offsetFreq, 15)
+	return
+}
+
+// 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
+	}
+	for _, t := range tokens {
+		if t < matchType {
+			w.writeCode(leCodes[t.literal()])
+			continue
+		}
+		// Write the length
+		length := t.length()
+		lengthCode := lengthCode(length)
+		w.writeCode(leCodes[lengthCode+lengthCodesStart])
+		extraLengthBits := uint(lengthExtraBits[lengthCode])
+		if extraLengthBits > 0 {
+			extraLength := int32(length - lengthBase[lengthCode])
+			w.writeBits(extraLength, extraLengthBits)
+		}
+		// Write the offset
+		offset := t.offset()
+		offsetCode := offsetCode(offset)
+		w.writeCode(oeCodes[offsetCode])
+		extraOffsetBits := uint(offsetExtraBits[offsetCode])
+		if extraOffsetBits > 0 {
+			extraOffset := int32(offset - offsetBase[offsetCode])
+			w.writeBits(extraOffset, extraOffsetBits)
+		}
+	}
+}
+
+// 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, 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) {
+	if w.err != nil {
+		return
+	}
+
+	// Clear histogram
+	for i := range w.literalFreq {
+		w.literalFreq[i] = 0
+	}
+
+	// Add everything as literals
+	histogram(input, w.literalFreq)
+
+	w.literalFreq[endBlockMarker] = 1
+
+	const numLiterals = endBlockMarker + 1
+	const numOffsets = 1
+
+	w.literalEncoding.generate(w.literalFreq, 15)
+
+	// Figure out smallest code.
+	// Always use dynamic Huffman or Store
+	var numCodegens int
+
+	// 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)
+	size, numCodegens := w.dynamicSize(w.literalEncoding, huffOffset, 0)
+
+	// Store bytes, if we don't get a reasonable improvement.
+	if ssize, storable := w.storedSize(input); storable && ssize < (size+size>>4) {
+		w.writeStoredHeader(len(input), eof)
+		w.writeBytes(input)
+		return
+	}
+
+	// Huffman.
+	w.writeDynamicHeader(numLiterals, numOffsets, numCodegens, eof)
+	encoding := w.literalEncoding.codes[:257]
+	n := w.nbytes
+	for _, t := range input {
+		// Bitwriting inlined, ~30% speedup
+		c := encoding[t]
+		w.bits |= uint64(c.code) << w.nbits
+		w.nbits += uint(c.len)
+		if w.nbits < 48 {
+			continue
+		}
+		// Store 6 bytes
+		bits := w.bits
+		w.bits >>= 48
+		w.nbits -= 48
+		bytes := w.bytes[n : n+6]
+		bytes[0] = byte(bits)
+		bytes[1] = byte(bits >> 8)
+		bytes[2] = byte(bits >> 16)
+		bytes[3] = byte(bits >> 24)
+		bytes[4] = byte(bits >> 32)
+		bytes[5] = byte(bits >> 40)
+		n += 6
+		if n < bufferFlushSize {
+			continue
+		}
+		w.write(w.bytes[:n])
+		if w.err != nil {
+			return // Return early in the event of write failures
+		}
+		n = 0
+	}
+	w.nbytes = n
+	w.writeCode(encoding[endBlockMarker])
+}
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 000000000..bdcbd823b
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/huffman_code.go
@@ -0,0 +1,344 @@
+// 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"
+	"sort"
+)
+
+// hcode is a huffman code with a bit code and bit length.
+type hcode struct {
+	code, len uint16
+}
+
+type huffmanEncoder struct {
+	codes     []hcode
+	freqcache []literalNode
+	bitCount  [17]int32
+	lns       byLiteral // stored to avoid repeated allocation in generate
+	lfs       byFreq    // stored to avoid repeated allocation in generate
+}
+
+type literalNode struct {
+	literal uint16
+	freq    int32
+}
+
+// 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 uint16) {
+	h.len = length
+	h.code = code
+}
+
+func maxNode() literalNode { return literalNode{math.MaxUint16, math.MaxInt32} }
+
+func newHuffmanEncoder(size int) *huffmanEncoder {
+	return &huffmanEncoder{codes: make([]hcode, size)}
+}
+
+// Generates a HuffmanCode corresponding to the fixed literal table
+func generateFixedLiteralEncoding() *huffmanEncoder {
+	h := newHuffmanEncoder(maxNumLit)
+	codes := h.codes
+	var ch uint16
+	for ch = 0; ch < maxNumLit; ch++ {
+		var bits uint16
+		var size uint16
+		switch {
+		case ch < 144:
+			// size 8, 000110000  .. 10111111
+			bits = ch + 48
+			size = 8
+			break
+		case ch < 256:
+			// size 9, 110010000 .. 111111111
+			bits = ch + 400 - 144
+			size = 9
+			break
+		case ch < 280:
+			// size 7, 0000000 .. 0010111
+			bits = ch - 256
+			size = 7
+			break
+		default:
+			// size 8, 11000000 .. 11000111
+			bits = ch + 192 - 280
+			size = 8
+		}
+		codes[ch] = hcode{code: reverseBits(bits, byte(size)), len: size}
+	}
+	return h
+}
+
+func generateFixedOffsetEncoding() *huffmanEncoder {
+	h := newHuffmanEncoder(30)
+	codes := h.codes
+	for ch := range codes {
+		codes[ch] = hcode{code: reverseBits(uint16(ch), 5), len: 5}
+	}
+	return h
+}
+
+var fixedLiteralEncoding *huffmanEncoder = generateFixedLiteralEncoding()
+var fixedOffsetEncoding *huffmanEncoder = generateFixedOffsetEncoding()
+
+func (h *huffmanEncoder) bitLength(freq []int32) int {
+	var total int
+	for i, f := range freq {
+		if f != 0 {
+			total += int(f) * int(h.codes[i].len)
+		}
+	}
+	return total
+}
+
+const maxBitsLimit = 16
+
+// 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
+
+	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:     list[1].freq,
+			nextCharFreq: list[2].freq,
+			nextPairFreq: list[0].freq + list[1].freq,
+		}
+		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 := maxBits
+	for {
+		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
+			l.nextCharFreq = list[n].freq
+		} 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.
+			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):]
+
+		h.lns.sort(chunk)
+		for _, node := range chunk {
+			h.codes[node.literal] = hcode{code: reverseBits(code, uint8(n)), len: uint16(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 []int32, maxBits int32) {
+	if h.freqcache == nil {
+		// Allocate a reusable buffer with the longest possible frequency table.
+		// Possible lengths are codegenCodeCount, offsetCodeCount and maxNumLit.
+		// The largest of these is maxNumLit, so we allocate for that case.
+		h.freqcache = make([]literalNode, maxNumLit+1)
+	}
+	list := h.freqcache[:len(freq)+1]
+	// 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 {
+			list[count] = literalNode{}
+			h.codes[i].len = 0
+		}
+	}
+	list[len(freq)] = 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
+	}
+	h.lfs.sort(list)
+
+	// Get the number of literals for each bit count
+	bitCount := h.bitCounts(list, maxBits)
+	// And do the assignment
+	h.assignEncodingAndSize(bitCount, list)
+}
+
+type byLiteral []literalNode
+
+func (s *byLiteral) sort(a []literalNode) {
+	*s = byLiteral(a)
+	sort.Sort(s)
+}
+
+func (s byLiteral) Len() int { return len(s) }
+
+func (s byLiteral) Less(i, j int) bool {
+	return s[i].literal < s[j].literal
+}
+
+func (s byLiteral) Swap(i, j int) { s[i], s[j] = s[j], s[i] }
+
+type byFreq []literalNode
+
+func (s *byFreq) sort(a []literalNode) {
+	*s = byFreq(a)
+	sort.Sort(s)
+}
+
+func (s byFreq) Len() int { return len(s) }
+
+func (s byFreq) Less(i, j int) bool {
+	if s[i].freq == s[j].freq {
+		return s[i].literal < s[j].literal
+	}
+	return s[i].freq < s[j].freq
+}
+
+func (s byFreq) Swap(i, j int) { s[i], s[j] = s[j], s[i] }
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 000000000..800d0ce9e
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/inflate.go
@@ -0,0 +1,880 @@
+// 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"
+	"io"
+	"math/bits"
+	"strconv"
+	"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
+)
+
+// 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 int64
+
+func (e CorruptInputError) Error() string {
+	return "flate: corrupt input before offset " + strconv.FormatInt(int64(e), 10)
+}
+
+// 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 struct {
+	Offset int64 // byte offset where error occurred
+	Err    error // error returned by underlying Read
+}
+
+func (e *ReadError) Error() string {
+	return "flate: read error at offset " + strconv.FormatInt(e.Offset, 10) + ": " + e.Err.Error()
+}
+
+// A WriteError reports an error encountered while writing output.
+//
+// Deprecated: No longer returned.
+type WriteError struct {
+	Offset int64 // byte offset where error occurred
+	Err    error // error returned by underlying Write
+}
+
+func (e *WriteError) Error() string {
+	return "flate: write error at offset " + strconv.FormatInt(e.Offset, 10) + ": " + e.Err.Error()
+}
+
+// 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 {
+	min      int                       // the minimum code length
+	chunks   *[huffmanNumChunks]uint32 // chunks as described above
+	links    [][]uint32                // 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 = &[huffmanNumChunks]uint32{}
+	}
+	if h.min != 0 {
+		*h = huffmanDecoder{chunks: h.chunks, links: h.links}
+	}
+
+	// Count number of codes of each length,
+	// compute min 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) {
+		return false
+	}
+
+	h.min = 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([][]uint32, 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] = uint32(off<<huffmanValueShift | (huffmanChunkBits + 1))
+			if cap(h.links[off]) < numLinks {
+				h.links[off] = make([]uint32, numLinks)
+			} else {
+				links := h.links[off][:0]
+				h.links[off] = links[:numLinks]
+			}
+		}
+	} else {
+		h.links = h.links[:0]
+	}
+
+	for i, n := range lengths {
+		if n == 0 {
+			continue
+		}
+		code := nextcode[n]
+		nextcode[n]++
+		chunk := uint32(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
+}
+
+// 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
+}
+
+// Decompress state.
+type decompressor struct {
+	// Input source.
+	r       Reader
+	roffset int64
+
+	// Input bits, in top of b.
+	b  uint32
+	nb uint
+
+	// 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
+
+	// Temporary buffer (avoids repeated allocation).
+	buf [4]byte
+
+	// Next step in the decompression,
+	// and decompression state.
+	step      func(*decompressor)
+	stepState int
+	final     bool
+	err       error
+	toRead    []byte
+	hl, hd    *huffmanDecoder
+	copyLen   int
+	copyDist  int
+}
+
+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()
+	case 1:
+		// compressed, fixed Huffman tables
+		f.hl = &fixedHuffmanDecoder
+		f.hd = nil
+		f.huffmanBlock()
+	case 2:
+		// compressed, dynamic Huffman tables
+		if f.err = f.readHuffman(); f.err != nil {
+			break
+		}
+		f.hl = &f.h1
+		f.hd = &f.h2
+		f.huffmanBlock()
+	default:
+		// 3 is reserved.
+		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.step(f)
+		if f.err != nil && len(f.toRead) == 0 {
+			f.toRead = f.dict.readFlush() // Flush what's left in case of error
+		}
+	}
+}
+
+// Support 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.step(f)
+		}
+		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 {
+		return CorruptInputError(f.roffset)
+	}
+	f.b >>= 5
+	ndist := int(f.b&0x1F) + 1
+	if ndist > maxNumDist {
+		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:]) {
+		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 {
+				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 {
+				return err
+			}
+		}
+		rep += int(f.b & uint32(1<<nb-1))
+		f.b >>= nb
+		f.nb -= nb
+		if 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]) {
+		return CorruptInputError(f.roffset)
+	}
+
+	// As an optimization, we can initialize the min 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.min < f.bits[endBlockMarker] {
+		f.h1.min = f.bits[endBlockMarker]
+	}
+
+	return nil
+}
+
+// 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) huffmanBlock() {
+	const (
+		stateInit = iota // Zero value must be stateInit
+		stateDict
+	)
+
+	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.
+	{
+		v, err := f.huffSym(f.hl)
+		if err != nil {
+			f.err = err
+			return
+		}
+		var n uint // number of bits extra
+		var length int
+		switch {
+		case v < 256:
+			f.dict.writeByte(byte(v))
+			if f.dict.availWrite() == 0 {
+				f.toRead = f.dict.readFlush()
+				f.step = (*decompressor).huffmanBlock
+				f.stepState = stateInit
+				return
+			}
+			goto readLiteral
+		case v == 256:
+			f.finishBlock()
+			return
+		// otherwise, reference to older data
+		case v < 265:
+			length = v - (257 - 3)
+			n = 0
+		case v < 269:
+			length = v*2 - (265*2 - 11)
+			n = 1
+		case v < 273:
+			length = v*4 - (269*4 - 19)
+			n = 2
+		case v < 277:
+			length = v*8 - (273*8 - 35)
+			n = 3
+		case v < 281:
+			length = v*16 - (277*16 - 67)
+			n = 4
+		case v < 285:
+			length = v*32 - (281*32 - 131)
+			n = 5
+		case v < maxNumLit:
+			length = 258
+			n = 0
+		default:
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+		if n > 0 {
+			for f.nb < n {
+				if err = f.moreBits(); err != nil {
+					f.err = err
+					return
+				}
+			}
+			length += int(f.b & uint32(1<<n-1))
+			f.b >>= n
+			f.nb -= n
+		}
+
+		var dist int
+		if f.hd == nil {
+			for f.nb < 5 {
+				if err = f.moreBits(); err != nil {
+					f.err = err
+					return
+				}
+			}
+			dist = int(bits.Reverse8(uint8(f.b & 0x1F << 3)))
+			f.b >>= 5
+			f.nb -= 5
+		} else {
+			if dist, err = f.huffSym(f.hd); err != nil {
+				f.err = err
+				return
+			}
+		}
+
+		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
+			for f.nb < nb {
+				if err = f.moreBits(); err != nil {
+					f.err = err
+					return
+				}
+			}
+			extra |= int(f.b & uint32(1<<nb-1))
+			f.b >>= nb
+			f.nb -= nb
+			dist = 1<<(nb+1) + 1 + extra
+		default:
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		// No check on length; encoding can be prescient.
+		if dist > f.dict.histSize() {
+			f.err = CorruptInputError(f.roffset)
+			return
+		}
+
+		f.copyLen, f.copyDist = length, dist
+		goto copyHistory
+	}
+
+copyHistory:
+	// Perform a backwards copy according to RFC section 3.2.3.
+	{
+		cnt := f.dict.tryWriteCopy(f.copyDist, f.copyLen)
+		if cnt == 0 {
+			cnt = f.dict.writeCopy(f.copyDist, f.copyLen)
+		}
+		f.copyLen -= cnt
+
+		if f.dict.availWrite() == 0 || f.copyLen > 0 {
+			f.toRead = f.dict.readFlush()
+			f.step = (*decompressor).huffmanBlock // We need to continue this work
+			f.stepState = stateDict
+			return
+		}
+		goto readLiteral
+	}
+}
+
+// Copy a single uncompressed data block from input to output.
+func (f *decompressor) dataBlock() {
+	// Uncompressed.
+	// Discard current half-byte.
+	f.nb = 0
+	f.b = 0
+
+	// Length then ones-complement of length.
+	nr, err := io.ReadFull(f.r, f.buf[0:4])
+	f.roffset += int64(nr)
+	if err != nil {
+		f.err = noEOF(err)
+		return
+	}
+	n := int(f.buf[0]) | int(f.buf[1])<<8
+	nn := int(f.buf[2]) | int(f.buf[3])<<8
+	if uint16(nn) != uint16(^n) {
+		f.err = CorruptInputError(f.roffset)
+		return
+	}
+
+	if n == 0 {
+		f.toRead = f.dict.readFlush()
+		f.finishBlock()
+		return
+	}
+
+	f.copyLen = 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 = (*decompressor).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
+	}
+	f.step = (*decompressor).nextBlock
+}
+
+// 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
+	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.min)
+	// 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 & 31)
+			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
+				f.err = CorruptInputError(f.roffset)
+				return 0, f.err
+			}
+			f.b = b >> (n & 31)
+			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:     (*decompressor).nextBlock,
+	}
+	f.dict.init(maxMatchOffset, dict)
+	return nil
+}
+
+// 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 {
+	fixedHuffmanDecoderInit()
+
+	var f decompressor
+	f.r = makeReader(r)
+	f.bits = new([maxNumLit + maxNumDist]int)
+	f.codebits = new([numCodes]int)
+	f.step = (*decompressor).nextBlock
+	f.dict.init(maxMatchOffset, nil)
+	return &f
+}
+
+// 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 {
+	fixedHuffmanDecoderInit()
+
+	var f decompressor
+	f.r = makeReader(r)
+	f.bits = new([maxNumLit + maxNumDist]int)
+	f.codebits = new([numCodes]int)
+	f.step = (*decompressor).nextBlock
+	f.dict.init(maxMatchOffset, dict)
+	return &f
+}
diff --git a/vendor/github.com/klauspost/compress/flate/reverse_bits.go b/vendor/github.com/klauspost/compress/flate/reverse_bits.go
new file mode 100644
index 000000000..c1a02720d
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/reverse_bits.go
@@ -0,0 +1,48 @@
+// 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
+
+var reverseByte = [256]byte{
+	0x00, 0x80, 0x40, 0xc0, 0x20, 0xa0, 0x60, 0xe0,
+	0x10, 0x90, 0x50, 0xd0, 0x30, 0xb0, 0x70, 0xf0,
+	0x08, 0x88, 0x48, 0xc8, 0x28, 0xa8, 0x68, 0xe8,
+	0x18, 0x98, 0x58, 0xd8, 0x38, 0xb8, 0x78, 0xf8,
+	0x04, 0x84, 0x44, 0xc4, 0x24, 0xa4, 0x64, 0xe4,
+	0x14, 0x94, 0x54, 0xd4, 0x34, 0xb4, 0x74, 0xf4,
+	0x0c, 0x8c, 0x4c, 0xcc, 0x2c, 0xac, 0x6c, 0xec,
+	0x1c, 0x9c, 0x5c, 0xdc, 0x3c, 0xbc, 0x7c, 0xfc,
+	0x02, 0x82, 0x42, 0xc2, 0x22, 0xa2, 0x62, 0xe2,
+	0x12, 0x92, 0x52, 0xd2, 0x32, 0xb2, 0x72, 0xf2,
+	0x0a, 0x8a, 0x4a, 0xca, 0x2a, 0xaa, 0x6a, 0xea,
+	0x1a, 0x9a, 0x5a, 0xda, 0x3a, 0xba, 0x7a, 0xfa,
+	0x06, 0x86, 0x46, 0xc6, 0x26, 0xa6, 0x66, 0xe6,
+	0x16, 0x96, 0x56, 0xd6, 0x36, 0xb6, 0x76, 0xf6,
+	0x0e, 0x8e, 0x4e, 0xce, 0x2e, 0xae, 0x6e, 0xee,
+	0x1e, 0x9e, 0x5e, 0xde, 0x3e, 0xbe, 0x7e, 0xfe,
+	0x01, 0x81, 0x41, 0xc1, 0x21, 0xa1, 0x61, 0xe1,
+	0x11, 0x91, 0x51, 0xd1, 0x31, 0xb1, 0x71, 0xf1,
+	0x09, 0x89, 0x49, 0xc9, 0x29, 0xa9, 0x69, 0xe9,
+	0x19, 0x99, 0x59, 0xd9, 0x39, 0xb9, 0x79, 0xf9,
+	0x05, 0x85, 0x45, 0xc5, 0x25, 0xa5, 0x65, 0xe5,
+	0x15, 0x95, 0x55, 0xd5, 0x35, 0xb5, 0x75, 0xf5,
+	0x0d, 0x8d, 0x4d, 0xcd, 0x2d, 0xad, 0x6d, 0xed,
+	0x1d, 0x9d, 0x5d, 0xdd, 0x3d, 0xbd, 0x7d, 0xfd,
+	0x03, 0x83, 0x43, 0xc3, 0x23, 0xa3, 0x63, 0xe3,
+	0x13, 0x93, 0x53, 0xd3, 0x33, 0xb3, 0x73, 0xf3,
+	0x0b, 0x8b, 0x4b, 0xcb, 0x2b, 0xab, 0x6b, 0xeb,
+	0x1b, 0x9b, 0x5b, 0xdb, 0x3b, 0xbb, 0x7b, 0xfb,
+	0x07, 0x87, 0x47, 0xc7, 0x27, 0xa7, 0x67, 0xe7,
+	0x17, 0x97, 0x57, 0xd7, 0x37, 0xb7, 0x77, 0xf7,
+	0x0f, 0x8f, 0x4f, 0xcf, 0x2f, 0xaf, 0x6f, 0xef,
+	0x1f, 0x9f, 0x5f, 0xdf, 0x3f, 0xbf, 0x7f, 0xff,
+}
+
+func reverseUint16(v uint16) uint16 {
+	return uint16(reverseByte[v>>8]) | uint16(reverseByte[v&0xFF])<<8
+}
+
+func reverseBits(number uint16, bitLength byte) uint16 {
+	return reverseUint16(number << uint8(16-bitLength))
+}
diff --git a/vendor/github.com/klauspost/compress/flate/snappy.go b/vendor/github.com/klauspost/compress/flate/snappy.go
new file mode 100644
index 000000000..d853320a7
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/snappy.go
@@ -0,0 +1,900 @@
+// 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
+
+// emitLiteral writes a literal chunk and returns the number of bytes written.
+func emitLiteral(dst *tokens, lit []byte) {
+	ol := int(dst.n)
+	for i, v := range lit {
+		dst.tokens[(i+ol)&maxStoreBlockSize] = token(v)
+	}
+	dst.n += uint16(len(lit))
+}
+
+// emitCopy writes a copy chunk and returns the number of bytes written.
+func emitCopy(dst *tokens, offset, length int) {
+	dst.tokens[dst.n] = matchToken(uint32(length-3), uint32(offset-minOffsetSize))
+	dst.n++
+}
+
+type snappyEnc interface {
+	Encode(dst *tokens, src []byte)
+	Reset()
+}
+
+func newSnappy(level int) snappyEnc {
+	switch level {
+	case 1:
+		return &snappyL1{}
+	case 2:
+		return &snappyL2{snappyGen: snappyGen{cur: maxStoreBlockSize, prev: make([]byte, 0, maxStoreBlockSize)}}
+	case 3:
+		return &snappyL3{snappyGen: snappyGen{cur: maxStoreBlockSize, prev: make([]byte, 0, maxStoreBlockSize)}}
+	case 4:
+		return &snappyL4{snappyL3{snappyGen: snappyGen{cur: maxStoreBlockSize, prev: make([]byte, 0, maxStoreBlockSize)}}}
+	default:
+		panic("invalid level specified")
+	}
+}
+
+const (
+	tableBits       = 14             // Bits used in the table
+	tableSize       = 1 << tableBits // Size of the table
+	tableMask       = tableSize - 1  // Mask for table indices. Redundant, but can eliminate bounds checks.
+	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
+)
+
+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
+}
+
+func hash(u uint32) uint32 {
+	return (u * 0x1e35a7bd) >> tableShift
+}
+
+// snappyL1 encapsulates level 1 compression
+type snappyL1 struct{}
+
+func (e *snappyL1) Reset() {}
+
+func (e *snappyL1) Encode(dst *tokens, src []byte) {
+	const (
+		inputMargin            = 16 - 1
+		minNonLiteralBlockSize = 1 + 1 + inputMargin
+	)
+
+	// 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
+	}
+
+	// Initialize the hash table.
+	//
+	// The table element type is uint16, as s < sLimit and sLimit < len(src)
+	// and len(src) <= maxStoreBlockSize and maxStoreBlockSize == 65535.
+	var table [tableSize]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))
+
+	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))
+			if s-candidate <= maxMatchOffset && 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.
+		emitLiteral(dst, 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 Snappy's:
+			//	s = extendMatch(src, candidate+4, s+4)
+			s += 4
+			s1 := base + maxMatchLength
+			if s1 > len(src) {
+				s1 = len(src)
+			}
+			a := src[s:s1]
+			b := src[candidate+4:]
+			b = b[:len(a)]
+			l := len(a)
+			for i := range a {
+				if a[i] != b[i] {
+					l = i
+					break
+				}
+			}
+			s += l
+
+			// matchToken is flate's equivalent of Snappy's emitCopy.
+			dst.tokens[dst.n] = matchToken(uint32(s-base-baseMatchLength), uint32(base-candidate-baseMatchOffset))
+			dst.n++
+			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))
+			table[prevHash&tableMask] = uint16(s - 1)
+			currHash := hash(uint32(x >> 8))
+			candidate = int(table[currHash&tableMask])
+			table[currHash&tableMask] = uint16(s)
+			if s-candidate > maxMatchOffset || uint32(x>>8) != load32(src, candidate) {
+				nextHash = hash(uint32(x >> 16))
+				s++
+				break
+			}
+		}
+	}
+
+emitRemainder:
+	if nextEmit < len(src) {
+		emitLiteral(dst, src[nextEmit:])
+	}
+}
+
+type tableEntry struct {
+	val    uint32
+	offset int32
+}
+
+func load3232(b []byte, i int32) 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 load6432(b []byte, i int32) 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
+}
+
+// snappyGen maintains the table for matches,
+// and the previous byte block for level 2.
+// This is the generic implementation.
+type snappyGen struct {
+	prev []byte
+	cur  int32
+}
+
+// snappyGen maintains the table for matches,
+// and the previous byte block for level 2.
+// This is the generic implementation.
+type snappyL2 struct {
+	snappyGen
+	table [tableSize]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 *snappyL2) Encode(dst *tokens, src []byte) {
+	const (
+		inputMargin            = 8 - 1
+		minNonLiteralBlockSize = 1 + 1 + inputMargin
+	)
+
+	// Protect against e.cur wraparound.
+	if e.cur > 1<<30 {
+		for i := range e.table[:] {
+			e.table[i] = tableEntry{}
+		}
+		e.cur = maxStoreBlockSize
+	}
+
+	// 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))
+		e.cur += maxStoreBlockSize
+		e.prev = e.prev[:0]
+		return
+	}
+
+	// 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.
+	nextEmit := int32(0)
+	s := int32(0)
+	cv := load3232(src, s)
+	nextHash := hash(cv)
+
+	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 := int32(32)
+
+		nextS := s
+		var candidate tableEntry
+		for {
+			s = nextS
+			bytesBetweenHashLookups := skip >> 5
+			nextS = s + bytesBetweenHashLookups
+			skip += bytesBetweenHashLookups
+			if nextS > sLimit {
+				goto emitRemainder
+			}
+			candidate = e.table[nextHash&tableMask]
+			now := load3232(src, nextS)
+			e.table[nextHash&tableMask] = tableEntry{offset: s + e.cur, val: cv}
+			nextHash = hash(now)
+
+			offset := s - (candidate.offset - e.cur)
+			if offset > maxMatchOffset || cv != candidate.val {
+				// Out of range or not matched.
+				cv = now
+				continue
+			}
+			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.
+		emitLiteral(dst, 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.
+
+			// Extend the 4-byte match as long as possible.
+			//
+			s += 4
+			t := candidate.offset - e.cur + 4
+			l := e.matchlen(s, t, src)
+
+			// matchToken is flate's equivalent of Snappy's emitCopy. (length,offset)
+			dst.tokens[dst.n] = matchToken(uint32(l+4-baseMatchLength), uint32(s-t-baseMatchOffset))
+			dst.n++
+			s += l
+			nextEmit = s
+			if s >= sLimit {
+				t += l
+				// Index first pair after match end.
+				if int(t+4) < len(src) && t > 0 {
+					cv := load3232(src, t)
+					e.table[hash(cv)&tableMask] = tableEntry{offset: t + e.cur, val: cv}
+				}
+				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 := load6432(src, s-1)
+			prevHash := hash(uint32(x))
+			e.table[prevHash&tableMask] = tableEntry{offset: e.cur + s - 1, val: uint32(x)}
+			x >>= 8
+			currHash := hash(uint32(x))
+			candidate = e.table[currHash&tableMask]
+			e.table[currHash&tableMask] = tableEntry{offset: e.cur + s, val: uint32(x)}
+
+			offset := s - (candidate.offset - e.cur)
+			if offset > maxMatchOffset || uint32(x) != candidate.val {
+				cv = uint32(x >> 8)
+				nextHash = hash(cv)
+				s++
+				break
+			}
+		}
+	}
+
+emitRemainder:
+	if int(nextEmit) < len(src) {
+		emitLiteral(dst, src[nextEmit:])
+	}
+	e.cur += int32(len(src))
+	e.prev = e.prev[:len(src)]
+	copy(e.prev, src)
+}
+
+type tableEntryPrev struct {
+	Cur  tableEntry
+	Prev tableEntry
+}
+
+// snappyL3
+type snappyL3 struct {
+	snappyGen
+	table [tableSize]tableEntryPrev
+}
+
+// Encode uses a similar algorithm to level 2, will check up to two candidates.
+func (e *snappyL3) Encode(dst *tokens, src []byte) {
+	const (
+		inputMargin            = 8 - 1
+		minNonLiteralBlockSize = 1 + 1 + inputMargin
+	)
+
+	// Protect against e.cur wraparound.
+	if e.cur > 1<<30 {
+		for i := range e.table[:] {
+			e.table[i] = tableEntryPrev{}
+		}
+		e.snappyGen = snappyGen{cur: maxStoreBlockSize, prev: e.prev[:0]}
+	}
+
+	// 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))
+		e.cur += maxStoreBlockSize
+		e.prev = e.prev[:0]
+		return
+	}
+
+	// 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.
+	nextEmit := int32(0)
+	s := int32(0)
+	cv := load3232(src, s)
+	nextHash := hash(cv)
+
+	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 := int32(32)
+
+		nextS := s
+		var candidate tableEntry
+		for {
+			s = nextS
+			bytesBetweenHashLookups := skip >> 5
+			nextS = s + bytesBetweenHashLookups
+			skip += bytesBetweenHashLookups
+			if nextS > sLimit {
+				goto emitRemainder
+			}
+			candidates := e.table[nextHash&tableMask]
+			now := load3232(src, nextS)
+			e.table[nextHash&tableMask] = tableEntryPrev{Prev: candidates.Cur, Cur: tableEntry{offset: s + e.cur, val: cv}}
+			nextHash = hash(now)
+
+			// Check both candidates
+			candidate = candidates.Cur
+			if cv == candidate.val {
+				offset := s - (candidate.offset - e.cur)
+				if offset <= maxMatchOffset {
+					break
+				}
+			} else {
+				// We only check if value mismatches.
+				// Offset will always be invalid in other cases.
+				candidate = candidates.Prev
+				if cv == candidate.val {
+					offset := s - (candidate.offset - e.cur)
+					if offset <= maxMatchOffset {
+						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.
+		emitLiteral(dst, 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.
+
+			// Extend the 4-byte match as long as possible.
+			//
+			s += 4
+			t := candidate.offset - e.cur + 4
+			l := e.matchlen(s, t, src)
+
+			// matchToken is flate's equivalent of Snappy's emitCopy. (length,offset)
+			dst.tokens[dst.n] = matchToken(uint32(l+4-baseMatchLength), uint32(s-t-baseMatchOffset))
+			dst.n++
+			s += l
+			nextEmit = s
+			if s >= sLimit {
+				t += l
+				// Index first pair after match end.
+				if int(t+4) < len(src) && t > 0 {
+					cv := load3232(src, t)
+					nextHash = hash(cv)
+					e.table[nextHash&tableMask] = tableEntryPrev{
+						Prev: e.table[nextHash&tableMask].Cur,
+						Cur:  tableEntry{offset: e.cur + t, val: cv},
+					}
+				}
+				goto emitRemainder
+			}
+
+			// We could immediately start working at s now, but to improve
+			// compression we first update the hash table at s-3 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-3)
+			prevHash := hash(uint32(x))
+			e.table[prevHash&tableMask] = tableEntryPrev{
+				Prev: e.table[prevHash&tableMask].Cur,
+				Cur:  tableEntry{offset: e.cur + s - 3, val: uint32(x)},
+			}
+			x >>= 8
+			prevHash = hash(uint32(x))
+
+			e.table[prevHash&tableMask] = tableEntryPrev{
+				Prev: e.table[prevHash&tableMask].Cur,
+				Cur:  tableEntry{offset: e.cur + s - 2, val: uint32(x)},
+			}
+			x >>= 8
+			prevHash = hash(uint32(x))
+
+			e.table[prevHash&tableMask] = tableEntryPrev{
+				Prev: e.table[prevHash&tableMask].Cur,
+				Cur:  tableEntry{offset: e.cur + s - 1, val: uint32(x)},
+			}
+			x >>= 8
+			currHash := hash(uint32(x))
+			candidates := e.table[currHash&tableMask]
+			cv = uint32(x)
+			e.table[currHash&tableMask] = tableEntryPrev{
+				Prev: candidates.Cur,
+				Cur:  tableEntry{offset: s + e.cur, val: cv},
+			}
+
+			// Check both candidates
+			candidate = candidates.Cur
+			if cv == candidate.val {
+				offset := s - (candidate.offset - e.cur)
+				if offset <= maxMatchOffset {
+					continue
+				}
+			} else {
+				// We only check if value mismatches.
+				// Offset will always be invalid in other cases.
+				candidate = candidates.Prev
+				if cv == candidate.val {
+					offset := s - (candidate.offset - e.cur)
+					if offset <= maxMatchOffset {
+						continue
+					}
+				}
+			}
+			cv = uint32(x >> 8)
+			nextHash = hash(cv)
+			s++
+			break
+		}
+	}
+
+emitRemainder:
+	if int(nextEmit) < len(src) {
+		emitLiteral(dst, src[nextEmit:])
+	}
+	e.cur += int32(len(src))
+	e.prev = e.prev[:len(src)]
+	copy(e.prev, src)
+}
+
+// snappyL4
+type snappyL4 struct {
+	snappyL3
+}
+
+// Encode uses a similar algorithm to level 3,
+// but will check up to two candidates if first isn't long enough.
+func (e *snappyL4) Encode(dst *tokens, src []byte) {
+	const (
+		inputMargin            = 8 - 3
+		minNonLiteralBlockSize = 1 + 1 + inputMargin
+		matchLenGood           = 12
+	)
+
+	// Protect against e.cur wraparound.
+	if e.cur > 1<<30 {
+		for i := range e.table[:] {
+			e.table[i] = tableEntryPrev{}
+		}
+		e.snappyGen = snappyGen{cur: maxStoreBlockSize, prev: e.prev[:0]}
+	}
+
+	// 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))
+		e.cur += maxStoreBlockSize
+		e.prev = e.prev[:0]
+		return
+	}
+
+	// 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.
+	nextEmit := int32(0)
+	s := int32(0)
+	cv := load3232(src, s)
+	nextHash := hash(cv)
+
+	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 := int32(32)
+
+		nextS := s
+		var candidate tableEntry
+		var candidateAlt tableEntry
+		for {
+			s = nextS
+			bytesBetweenHashLookups := skip >> 5
+			nextS = s + bytesBetweenHashLookups
+			skip += bytesBetweenHashLookups
+			if nextS > sLimit {
+				goto emitRemainder
+			}
+			candidates := e.table[nextHash&tableMask]
+			now := load3232(src, nextS)
+			e.table[nextHash&tableMask] = tableEntryPrev{Prev: candidates.Cur, Cur: tableEntry{offset: s + e.cur, val: cv}}
+			nextHash = hash(now)
+
+			// Check both candidates
+			candidate = candidates.Cur
+			if cv == candidate.val {
+				offset := s - (candidate.offset - e.cur)
+				if offset < maxMatchOffset {
+					offset = s - (candidates.Prev.offset - e.cur)
+					if cv == candidates.Prev.val && offset < maxMatchOffset {
+						candidateAlt = candidates.Prev
+					}
+					break
+				}
+			} else {
+				// We only check if value mismatches.
+				// Offset will always be invalid in other cases.
+				candidate = candidates.Prev
+				if cv == candidate.val {
+					offset := s - (candidate.offset - e.cur)
+					if offset < maxMatchOffset {
+						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.
+		emitLiteral(dst, 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.
+
+			// Extend the 4-byte match as long as possible.
+			//
+			s += 4
+			t := candidate.offset - e.cur + 4
+			l := e.matchlen(s, t, src)
+			// Try alternative candidate if match length < matchLenGood.
+			if l < matchLenGood-4 && candidateAlt.offset != 0 {
+				t2 := candidateAlt.offset - e.cur + 4
+				l2 := e.matchlen(s, t2, src)
+				if l2 > l {
+					l = l2
+					t = t2
+				}
+			}
+			// matchToken is flate's equivalent of Snappy's emitCopy. (length,offset)
+			dst.tokens[dst.n] = matchToken(uint32(l+4-baseMatchLength), uint32(s-t-baseMatchOffset))
+			dst.n++
+			s += l
+			nextEmit = s
+			if s >= sLimit {
+				t += l
+				// Index first pair after match end.
+				if int(t+4) < len(src) && t > 0 {
+					cv := load3232(src, t)
+					nextHash = hash(cv)
+					e.table[nextHash&tableMask] = tableEntryPrev{
+						Prev: e.table[nextHash&tableMask].Cur,
+						Cur:  tableEntry{offset: e.cur + t, val: cv},
+					}
+				}
+				goto emitRemainder
+			}
+
+			// We could immediately start working at s now, but to improve
+			// compression we first update the hash table at s-3 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-3)
+			prevHash := hash(uint32(x))
+			e.table[prevHash&tableMask] = tableEntryPrev{
+				Prev: e.table[prevHash&tableMask].Cur,
+				Cur:  tableEntry{offset: e.cur + s - 3, val: uint32(x)},
+			}
+			x >>= 8
+			prevHash = hash(uint32(x))
+
+			e.table[prevHash&tableMask] = tableEntryPrev{
+				Prev: e.table[prevHash&tableMask].Cur,
+				Cur:  tableEntry{offset: e.cur + s - 2, val: uint32(x)},
+			}
+			x >>= 8
+			prevHash = hash(uint32(x))
+
+			e.table[prevHash&tableMask] = tableEntryPrev{
+				Prev: e.table[prevHash&tableMask].Cur,
+				Cur:  tableEntry{offset: e.cur + s - 1, val: uint32(x)},
+			}
+			x >>= 8
+			currHash := hash(uint32(x))
+			candidates := e.table[currHash&tableMask]
+			cv = uint32(x)
+			e.table[currHash&tableMask] = tableEntryPrev{
+				Prev: candidates.Cur,
+				Cur:  tableEntry{offset: s + e.cur, val: cv},
+			}
+
+			// Check both candidates
+			candidate = candidates.Cur
+			candidateAlt = tableEntry{}
+			if cv == candidate.val {
+				offset := s - (candidate.offset - e.cur)
+				if offset <= maxMatchOffset {
+					offset = s - (candidates.Prev.offset - e.cur)
+					if cv == candidates.Prev.val && offset <= maxMatchOffset {
+						candidateAlt = candidates.Prev
+					}
+					continue
+				}
+			} else {
+				// We only check if value mismatches.
+				// Offset will always be invalid in other cases.
+				candidate = candidates.Prev
+				if cv == candidate.val {
+					offset := s - (candidate.offset - e.cur)
+					if offset <= maxMatchOffset {
+						continue
+					}
+				}
+			}
+			cv = uint32(x >> 8)
+			nextHash = hash(cv)
+			s++
+			break
+		}
+	}
+
+emitRemainder:
+	if int(nextEmit) < len(src) {
+		emitLiteral(dst, src[nextEmit:])
+	}
+	e.cur += int32(len(src))
+	e.prev = e.prev[:len(src)]
+	copy(e.prev, src)
+}
+
+func (e *snappyGen) matchlen(s, t int32, src []byte) int32 {
+	s1 := int(s) + maxMatchLength - 4
+	if s1 > len(src) {
+		s1 = len(src)
+	}
+
+	// If we are inside the current block
+	if t >= 0 {
+		b := src[t:]
+		a := src[s:s1]
+		b = b[:len(a)]
+		// Extend the match to be as long as possible.
+		for i := range a {
+			if a[i] != b[i] {
+				return int32(i)
+			}
+		}
+		return int32(len(a))
+	}
+
+	// We found a match in the previous block.
+	tp := int32(len(e.prev)) + t
+	if tp < 0 {
+		return 0
+	}
+
+	// Extend the match to be as long as possible.
+	a := src[s:s1]
+	b := e.prev[tp:]
+	if len(b) > len(a) {
+		b = b[:len(a)]
+	}
+	a = a[:len(b)]
+	for i := range b {
+		if a[i] != b[i] {
+			return int32(i)
+		}
+	}
+
+	// If we reached our limit, we matched everything we are
+	// allowed to in the previous block and we return.
+	n := int32(len(b))
+	if int(s+n) == s1 {
+		return n
+	}
+
+	// Continue looking for more matches in the current block.
+	a = src[s+n : s1]
+	b = src[:len(a)]
+	for i := range a {
+		if a[i] != b[i] {
+			return int32(i) + n
+		}
+	}
+	return int32(len(a)) + n
+}
+
+// Reset the encoding table.
+func (e *snappyGen) Reset() {
+	e.prev = e.prev[:0]
+	e.cur += maxMatchOffset
+}
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 000000000..4f275ea61
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/token.go
@@ -0,0 +1,115 @@
+// 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"
+
+const (
+	// 2 bits:   type   0 = literal  1=EOF  2=Match   3=Unused
+	// 8 bits:   xlength = length - MIN_MATCH_LENGTH
+	// 22 bits   xoffset = offset - MIN_OFFSET_SIZE, or literal
+	lengthShift = 22
+	offsetMask  = 1<<lengthShift - 1
+	typeMask    = 3 << 30
+	literalType = 0 << 30
+	matchType   = 1 << 30
+)
+
+// The length code for length X (MIN_MATCH_LENGTH <= X <= MAX_MATCH_LENGTH)
+// is lengthCodes[length - MIN_MATCH_LENGTH]
+var lengthCodes = [...]uint32{
+	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,
+}
+
+var offsetCodes = [...]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,
+}
+
+type token uint32
+
+type tokens struct {
+	tokens [maxStoreBlockSize + 1]token
+	n      uint16 // Must be able to contain maxStoreBlockSize
+}
+
+// Convert a literal into a literal token.
+func literalToken(literal uint32) token { return token(literalType + literal) }
+
+// Convert a < xlength, xoffset > pair into a match token.
+func matchToken(xlength uint32, xoffset uint32) token {
+	return token(matchType + xlength<<lengthShift + xoffset)
+}
+
+func matchTokend(xlength uint32, xoffset uint32) token {
+	if xlength > maxMatchLength || xoffset > maxMatchOffset {
+		panic(fmt.Sprintf("Invalid match: len: %d, offset: %d\n", xlength, xoffset))
+		return token(matchType)
+	}
+	return token(matchType + xlength<<lengthShift + xoffset)
+}
+
+// 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() uint32 { return uint32(t - literalType) }
+
+// Returns the extra offset of a match token
+func (t token) offset() uint32 { return uint32(t) & offsetMask }
+
+func (t token) length() uint32 { return uint32((t - matchType) >> lengthShift) }
+
+func lengthCode(len uint32) uint32 { return lengthCodes[len] }
+
+// Returns the offset code corresponding to a specific offset
+func offsetCode(off uint32) uint32 {
+	if off < uint32(len(offsetCodes)) {
+		return offsetCodes[off]
+	} else if off>>7 < uint32(len(offsetCodes)) {
+		return offsetCodes[off>>7] + 14
+	} else {
+		return offsetCodes[off>>14] + 28
+	}
+}
diff --git a/vendor/github.com/klauspost/cpuid/LICENSE b/vendor/github.com/klauspost/cpuid/LICENSE
new file mode 100644
index 000000000..5cec7ee94
--- /dev/null
+++ b/vendor/github.com/klauspost/cpuid/LICENSE
@@ -0,0 +1,22 @@
+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.
+
diff --git a/vendor/github.com/klauspost/cpuid/README.md b/vendor/github.com/klauspost/cpuid/README.md
new file mode 100644
index 000000000..b2b6bee87
--- /dev/null
+++ b/vendor/github.com/klauspost/cpuid/README.md
@@ -0,0 +1,145 @@
+# cpuid
+Package cpuid provides information about the CPU running the current program.
+
+CPU features are detected on startup, and kept for fast access through the life of the application.
+Currently x86 / x64 (AMD64) is supported, and no external C (cgo) code is used, which should make the library very easy to use.
+
+You can access the CPU information by accessing the shared CPU variable of the cpuid library.
+
+Package home: https://github.com/klauspost/cpuid
+
+[![GoDoc][1]][2] [![Build Status][3]][4]
+
+[1]: https://godoc.org/github.com/klauspost/cpuid?status.svg
+[2]: https://godoc.org/github.com/klauspost/cpuid
+[3]: https://travis-ci.org/klauspost/cpuid.svg
+[4]: https://travis-ci.org/klauspost/cpuid
+
+# features
+## CPU Instructions
+*  **CMOV** (i686 CMOV)
+*  **NX** (NX (No-Execute) bit)
+*  **AMD3DNOW** (AMD 3DNOW)
+*  **AMD3DNOWEXT** (AMD 3DNowExt)
+*  **MMX** (standard MMX)
+*  **MMXEXT** (SSE integer functions or AMD MMX ext)
+*  **SSE** (SSE functions)
+*  **SSE2** (P4 SSE functions)
+*  **SSE3** (Prescott SSE3 functions)
+*  **SSSE3** (Conroe SSSE3 functions)
+*  **SSE4** (Penryn SSE4.1 functions)
+*  **SSE4A** (AMD Barcelona microarchitecture SSE4a instructions)
+*  **SSE42** (Nehalem SSE4.2 functions)
+*  **AVX** (AVX functions)
+*  **AVX2** (AVX2 functions)
+*  **FMA3** (Intel FMA 3)
+*  **FMA4** (Bulldozer FMA4 functions)
+*  **XOP** (Bulldozer XOP functions)
+*  **F16C** (Half-precision floating-point conversion)
+*  **BMI1** (Bit Manipulation Instruction Set 1)
+*  **BMI2** (Bit Manipulation Instruction Set 2)
+*  **TBM** (AMD Trailing Bit Manipulation)
+*  **LZCNT** (LZCNT instruction)
+*  **POPCNT** (POPCNT instruction)
+*  **AESNI** (Advanced Encryption Standard New Instructions)
+*  **CLMUL** (Carry-less Multiplication)
+*  **HTT** (Hyperthreading (enabled))
+*  **HLE** (Hardware Lock Elision)
+*  **RTM** (Restricted Transactional Memory)
+*  **RDRAND** (RDRAND instruction is available)
+*  **RDSEED** (RDSEED instruction is available)
+*  **ADX** (Intel ADX (Multi-Precision Add-Carry Instruction Extensions))
+*  **SHA** (Intel SHA Extensions)
+*  **AVX512F** (AVX-512 Foundation)
+*  **AVX512DQ** (AVX-512 Doubleword and Quadword Instructions)
+*  **AVX512IFMA** (AVX-512 Integer Fused Multiply-Add Instructions)
+*  **AVX512PF** (AVX-512 Prefetch Instructions)
+*  **AVX512ER** (AVX-512 Exponential and Reciprocal Instructions)
+*  **AVX512CD** (AVX-512 Conflict Detection Instructions)
+*  **AVX512BW** (AVX-512 Byte and Word Instructions)
+*  **AVX512VL** (AVX-512 Vector Length Extensions)
+*  **AVX512VBMI** (AVX-512 Vector Bit Manipulation Instructions)
+*  **MPX** (Intel MPX (Memory Protection Extensions))
+*  **ERMS** (Enhanced REP MOVSB/STOSB)
+*  **RDTSCP** (RDTSCP Instruction)
+*  **CX16** (CMPXCHG16B Instruction)
+*  **SGX** (Software Guard Extensions, with activation details)
+
+## Performance
+*  **RDTSCP()** Returns current cycle count. Can be used for benchmarking.
+*  **SSE2SLOW** (SSE2 is supported, but usually not faster)
+*  **SSE3SLOW** (SSE3 is supported, but usually not faster)
+*  **ATOM** (Atom processor, some SSSE3 instructions are slower)
+*  **Cache line** (Probable size of a cache line).
+*  **L1, L2, L3 Cache size** on newer Intel/AMD CPUs.
+
+## Cpu Vendor/VM
+* **Intel**
+* **AMD**
+* **VIA**
+* **Transmeta**
+* **NSC**
+* **KVM**  (Kernel-based Virtual Machine)
+* **MSVM** (Microsoft Hyper-V or Windows Virtual PC)
+* **VMware**
+* **XenHVM**
+
+# installing
+
+```go get github.com/klauspost/cpuid```
+
+# example
+
+```Go
+package main
+
+import (
+	"fmt"
+	"github.com/klauspost/cpuid"
+)
+
+func main() {
+	// Print basic CPU information:
+	fmt.Println("Name:", cpuid.CPU.BrandName)
+	fmt.Println("PhysicalCores:", cpuid.CPU.PhysicalCores)
+	fmt.Println("ThreadsPerCore:", cpuid.CPU.ThreadsPerCore)
+	fmt.Println("LogicalCores:", cpuid.CPU.LogicalCores)
+	fmt.Println("Family", cpuid.CPU.Family, "Model:", cpuid.CPU.Model)
+	fmt.Println("Features:", cpuid.CPU.Features)
+	fmt.Println("Cacheline bytes:", cpuid.CPU.CacheLine)
+	fmt.Println("L1 Data Cache:", cpuid.CPU.Cache.L1D, "bytes")
+	fmt.Println("L1 Instruction Cache:", cpuid.CPU.Cache.L1D, "bytes")
+	fmt.Println("L2 Cache:", cpuid.CPU.Cache.L2, "bytes")
+	fmt.Println("L3 Cache:", cpuid.CPU.Cache.L3, "bytes")
+
+	// Test if we have a specific feature:
+	if cpuid.CPU.SSE() {
+		fmt.Println("We have Streaming SIMD Extensions")
+	}
+}
+```
+
+Sample output:
+```
+>go run main.go
+Name: Intel(R) Core(TM) i5-2540M CPU @ 2.60GHz
+PhysicalCores: 2
+ThreadsPerCore: 2
+LogicalCores: 4
+Family 6 Model: 42
+Features: CMOV,MMX,MMXEXT,SSE,SSE2,SSE3,SSSE3,SSE4.1,SSE4.2,AVX,AESNI,CLMUL
+Cacheline bytes: 64
+We have Streaming SIMD Extensions
+```
+
+# private package
+
+In the "private" folder you can find an autogenerated version of the library you can include in your own packages.
+
+For this purpose all exports are removed, and functions and constants are lowercased.
+
+This is not a recommended way of using the library, but provided for convenience, if it is difficult for you to use external packages.
+
+# license
+
+This code is published under an MIT license. See LICENSE file for more information.
diff --git a/vendor/github.com/klauspost/cpuid/cpuid.go b/vendor/github.com/klauspost/cpuid/cpuid.go
new file mode 100644
index 000000000..60c681bed
--- /dev/null
+++ b/vendor/github.com/klauspost/cpuid/cpuid.go
@@ -0,0 +1,1040 @@
+// Copyright (c) 2015 Klaus Post, released under MIT License. See LICENSE file.
+
+// Package cpuid provides information about the CPU running the current program.
+//
+// CPU features are detected on startup, and kept for fast access through the life of the application.
+// Currently x86 / x64 (AMD64) is supported.
+//
+// You can access the CPU information by accessing the shared CPU variable of the cpuid library.
+//
+// Package home: https://github.com/klauspost/cpuid
+package cpuid
+
+import "strings"
+
+// Vendor is a representation of a CPU vendor.
+type Vendor int
+
+const (
+	Other Vendor = iota
+	Intel
+	AMD
+	VIA
+	Transmeta
+	NSC
+	KVM  // Kernel-based Virtual Machine
+	MSVM // Microsoft Hyper-V or Windows Virtual PC
+	VMware
+	XenHVM
+)
+
+const (
+	CMOV        = 1 << iota // i686 CMOV
+	NX                      // NX (No-Execute) bit
+	AMD3DNOW                // AMD 3DNOW
+	AMD3DNOWEXT             // AMD 3DNowExt
+	MMX                     // standard MMX
+	MMXEXT                  // SSE integer functions or AMD MMX ext
+	SSE                     // SSE functions
+	SSE2                    // P4 SSE functions
+	SSE3                    // Prescott SSE3 functions
+	SSSE3                   // Conroe SSSE3 functions
+	SSE4                    // Penryn SSE4.1 functions
+	SSE4A                   // AMD Barcelona microarchitecture SSE4a instructions
+	SSE42                   // Nehalem SSE4.2 functions
+	AVX                     // AVX functions
+	AVX2                    // AVX2 functions
+	FMA3                    // Intel FMA 3
+	FMA4                    // Bulldozer FMA4 functions
+	XOP                     // Bulldozer XOP functions
+	F16C                    // Half-precision floating-point conversion
+	BMI1                    // Bit Manipulation Instruction Set 1
+	BMI2                    // Bit Manipulation Instruction Set 2
+	TBM                     // AMD Trailing Bit Manipulation
+	LZCNT                   // LZCNT instruction
+	POPCNT                  // POPCNT instruction
+	AESNI                   // Advanced Encryption Standard New Instructions
+	CLMUL                   // Carry-less Multiplication
+	HTT                     // Hyperthreading (enabled)
+	HLE                     // Hardware Lock Elision
+	RTM                     // Restricted Transactional Memory
+	RDRAND                  // RDRAND instruction is available
+	RDSEED                  // RDSEED instruction is available
+	ADX                     // Intel ADX (Multi-Precision Add-Carry Instruction Extensions)
+	SHA                     // Intel SHA Extensions
+	AVX512F                 // AVX-512 Foundation
+	AVX512DQ                // AVX-512 Doubleword and Quadword Instructions
+	AVX512IFMA              // AVX-512 Integer Fused Multiply-Add Instructions
+	AVX512PF                // AVX-512 Prefetch Instructions
+	AVX512ER                // AVX-512 Exponential and Reciprocal Instructions
+	AVX512CD                // AVX-512 Conflict Detection Instructions
+	AVX512BW                // AVX-512 Byte and Word Instructions
+	AVX512VL                // AVX-512 Vector Length Extensions
+	AVX512VBMI              // AVX-512 Vector Bit Manipulation Instructions
+	MPX                     // Intel MPX (Memory Protection Extensions)
+	ERMS                    // Enhanced REP MOVSB/STOSB
+	RDTSCP                  // RDTSCP Instruction
+	CX16                    // CMPXCHG16B Instruction
+	SGX                     // Software Guard Extensions
+	IBPB                    // Indirect Branch Restricted Speculation (IBRS) and Indirect Branch Predictor Barrier (IBPB)
+	STIBP                   // Single Thread Indirect Branch Predictors
+
+	// Performance indicators
+	SSE2SLOW // SSE2 is supported, but usually not faster
+	SSE3SLOW // SSE3 is supported, but usually not faster
+	ATOM     // Atom processor, some SSSE3 instructions are slower
+)
+
+var flagNames = map[Flags]string{
+	CMOV:        "CMOV",        // i686 CMOV
+	NX:          "NX",          // NX (No-Execute) bit
+	AMD3DNOW:    "AMD3DNOW",    // AMD 3DNOW
+	AMD3DNOWEXT: "AMD3DNOWEXT", // AMD 3DNowExt
+	MMX:         "MMX",         // Standard MMX
+	MMXEXT:      "MMXEXT",      // SSE integer functions or AMD MMX ext
+	SSE:         "SSE",         // SSE functions
+	SSE2:        "SSE2",        // P4 SSE2 functions
+	SSE3:        "SSE3",        // Prescott SSE3 functions
+	SSSE3:       "SSSE3",       // Conroe SSSE3 functions
+	SSE4:        "SSE4.1",      // Penryn SSE4.1 functions
+	SSE4A:       "SSE4A",       // AMD Barcelona microarchitecture SSE4a instructions
+	SSE42:       "SSE4.2",      // Nehalem SSE4.2 functions
+	AVX:         "AVX",         // AVX functions
+	AVX2:        "AVX2",        // AVX functions
+	FMA3:        "FMA3",        // Intel FMA 3
+	FMA4:        "FMA4",        // Bulldozer FMA4 functions
+	XOP:         "XOP",         // Bulldozer XOP functions
+	F16C:        "F16C",        // Half-precision floating-point conversion
+	BMI1:        "BMI1",        // Bit Manipulation Instruction Set 1
+	BMI2:        "BMI2",        // Bit Manipulation Instruction Set 2
+	TBM:         "TBM",         // AMD Trailing Bit Manipulation
+	LZCNT:       "LZCNT",       // LZCNT instruction
+	POPCNT:      "POPCNT",      // POPCNT instruction
+	AESNI:       "AESNI",       // Advanced Encryption Standard New Instructions
+	CLMUL:       "CLMUL",       // Carry-less Multiplication
+	HTT:         "HTT",         // Hyperthreading (enabled)
+	HLE:         "HLE",         // Hardware Lock Elision
+	RTM:         "RTM",         // Restricted Transactional Memory
+	RDRAND:      "RDRAND",      // RDRAND instruction is available
+	RDSEED:      "RDSEED",      // RDSEED instruction is available
+	ADX:         "ADX",         // Intel ADX (Multi-Precision Add-Carry Instruction Extensions)
+	SHA:         "SHA",         // Intel SHA Extensions
+	AVX512F:     "AVX512F",     // AVX-512 Foundation
+	AVX512DQ:    "AVX512DQ",    // AVX-512 Doubleword and Quadword Instructions
+	AVX512IFMA:  "AVX512IFMA",  // AVX-512 Integer Fused Multiply-Add Instructions
+	AVX512PF:    "AVX512PF",    // AVX-512 Prefetch Instructions
+	AVX512ER:    "AVX512ER",    // AVX-512 Exponential and Reciprocal Instructions
+	AVX512CD:    "AVX512CD",    // AVX-512 Conflict Detection Instructions
+	AVX512BW:    "AVX512BW",    // AVX-512 Byte and Word Instructions
+	AVX512VL:    "AVX512VL",    // AVX-512 Vector Length Extensions
+	AVX512VBMI:  "AVX512VBMI",  // AVX-512 Vector Bit Manipulation Instructions
+	MPX:         "MPX",         // Intel MPX (Memory Protection Extensions)
+	ERMS:        "ERMS",        // Enhanced REP MOVSB/STOSB
+	RDTSCP:      "RDTSCP",      // RDTSCP Instruction
+	CX16:        "CX16",        // CMPXCHG16B Instruction
+	SGX:         "SGX",         // Software Guard Extensions
+	IBPB:        "IBPB",        // Indirect Branch Restricted Speculation and Indirect Branch Predictor Barrier
+	STIBP:       "STIBP",       // Single Thread Indirect Branch Predictors
+
+	// Performance indicators
+	SSE2SLOW: "SSE2SLOW", // SSE2 supported, but usually not faster
+	SSE3SLOW: "SSE3SLOW", // SSE3 supported, but usually not faster
+	ATOM:     "ATOM",     // Atom processor, some SSSE3 instructions are slower
+
+}
+
+// CPUInfo contains information about the detected system CPU.
+type CPUInfo struct {
+	BrandName      string // Brand name reported by the CPU
+	VendorID       Vendor // Comparable CPU vendor ID
+	Features       Flags  // Features of the CPU
+	PhysicalCores  int    // Number of physical processor cores in your CPU. Will be 0 if undetectable.
+	ThreadsPerCore int    // Number of threads per physical core. Will be 1 if undetectable.
+	LogicalCores   int    // Number of physical cores times threads that can run on each core through the use of hyperthreading. Will be 0 if undetectable.
+	Family         int    // CPU family number
+	Model          int    // CPU model number
+	CacheLine      int    // Cache line size in bytes. Will be 0 if undetectable.
+	Cache          struct {
+		L1I int // L1 Instruction Cache (per core or shared). Will be -1 if undetected
+		L1D int // L1 Data Cache (per core or shared). Will be -1 if undetected
+		L2  int // L2 Cache (per core or shared). Will be -1 if undetected
+		L3  int // L3 Instruction Cache (per core or shared). Will be -1 if undetected
+	}
+	SGX       SGXSupport
+	maxFunc   uint32
+	maxExFunc uint32
+}
+
+var cpuid func(op uint32) (eax, ebx, ecx, edx uint32)
+var cpuidex func(op, op2 uint32) (eax, ebx, ecx, edx uint32)
+var xgetbv func(index uint32) (eax, edx uint32)
+var rdtscpAsm func() (eax, ebx, ecx, edx uint32)
+
+// CPU contains information about the CPU as detected on startup,
+// or when Detect last was called.
+//
+// Use this as the primary entry point to you data,
+// this way queries are
+var CPU CPUInfo
+
+func init() {
+	initCPU()
+	Detect()
+}
+
+// Detect will re-detect current CPU info.
+// This will replace the content of the exported CPU variable.
+//
+// Unless you expect the CPU to change while you are running your program
+// you should not need to call this function.
+// If you call this, you must ensure that no other goroutine is accessing the
+// exported CPU variable.
+func Detect() {
+	CPU.maxFunc = maxFunctionID()
+	CPU.maxExFunc = maxExtendedFunction()
+	CPU.BrandName = brandName()
+	CPU.CacheLine = cacheLine()
+	CPU.Family, CPU.Model = familyModel()
+	CPU.Features = support()
+	CPU.SGX = hasSGX(CPU.Features&SGX != 0)
+	CPU.ThreadsPerCore = threadsPerCore()
+	CPU.LogicalCores = logicalCores()
+	CPU.PhysicalCores = physicalCores()
+	CPU.VendorID = vendorID()
+	CPU.cacheSize()
+}
+
+// Generated here: http://play.golang.org/p/BxFH2Gdc0G
+
+// Cmov indicates support of CMOV instructions
+func (c CPUInfo) Cmov() bool {
+	return c.Features&CMOV != 0
+}
+
+// Amd3dnow indicates support of AMD 3DNOW! instructions
+func (c CPUInfo) Amd3dnow() bool {
+	return c.Features&AMD3DNOW != 0
+}
+
+// Amd3dnowExt indicates support of AMD 3DNOW! Extended instructions
+func (c CPUInfo) Amd3dnowExt() bool {
+	return c.Features&AMD3DNOWEXT != 0
+}
+
+// MMX indicates support of MMX instructions
+func (c CPUInfo) MMX() bool {
+	return c.Features&MMX != 0
+}
+
+// MMXExt indicates support of MMXEXT instructions
+// (SSE integer functions or AMD MMX ext)
+func (c CPUInfo) MMXExt() bool {
+	return c.Features&MMXEXT != 0
+}
+
+// SSE indicates support of SSE instructions
+func (c CPUInfo) SSE() bool {
+	return c.Features&SSE != 0
+}
+
+// SSE2 indicates support of SSE 2 instructions
+func (c CPUInfo) SSE2() bool {
+	return c.Features&SSE2 != 0
+}
+
+// SSE3 indicates support of SSE 3 instructions
+func (c CPUInfo) SSE3() bool {
+	return c.Features&SSE3 != 0
+}
+
+// SSSE3 indicates support of SSSE 3 instructions
+func (c CPUInfo) SSSE3() bool {
+	return c.Features&SSSE3 != 0
+}
+
+// SSE4 indicates support of SSE 4 (also called SSE 4.1) instructions
+func (c CPUInfo) SSE4() bool {
+	return c.Features&SSE4 != 0
+}
+
+// SSE42 indicates support of SSE4.2 instructions
+func (c CPUInfo) SSE42() bool {
+	return c.Features&SSE42 != 0
+}
+
+// AVX indicates support of AVX instructions
+// and operating system support of AVX instructions
+func (c CPUInfo) AVX() bool {
+	return c.Features&AVX != 0
+}
+
+// AVX2 indicates support of AVX2 instructions
+func (c CPUInfo) AVX2() bool {
+	return c.Features&AVX2 != 0
+}
+
+// FMA3 indicates support of FMA3 instructions
+func (c CPUInfo) FMA3() bool {
+	return c.Features&FMA3 != 0
+}
+
+// FMA4 indicates support of FMA4 instructions
+func (c CPUInfo) FMA4() bool {
+	return c.Features&FMA4 != 0
+}
+
+// XOP indicates support of XOP instructions
+func (c CPUInfo) XOP() bool {
+	return c.Features&XOP != 0
+}
+
+// F16C indicates support of F16C instructions
+func (c CPUInfo) F16C() bool {
+	return c.Features&F16C != 0
+}
+
+// BMI1 indicates support of BMI1 instructions
+func (c CPUInfo) BMI1() bool {
+	return c.Features&BMI1 != 0
+}
+
+// BMI2 indicates support of BMI2 instructions
+func (c CPUInfo) BMI2() bool {
+	return c.Features&BMI2 != 0
+}
+
+// TBM indicates support of TBM instructions
+// (AMD Trailing Bit Manipulation)
+func (c CPUInfo) TBM() bool {
+	return c.Features&TBM != 0
+}
+
+// Lzcnt indicates support of LZCNT instruction
+func (c CPUInfo) Lzcnt() bool {
+	return c.Features&LZCNT != 0
+}
+
+// Popcnt indicates support of POPCNT instruction
+func (c CPUInfo) Popcnt() bool {
+	return c.Features&POPCNT != 0
+}
+
+// HTT indicates the processor has Hyperthreading enabled
+func (c CPUInfo) HTT() bool {
+	return c.Features&HTT != 0
+}
+
+// SSE2Slow indicates that SSE2 may be slow on this processor
+func (c CPUInfo) SSE2Slow() bool {
+	return c.Features&SSE2SLOW != 0
+}
+
+// SSE3Slow indicates that SSE3 may be slow on this processor
+func (c CPUInfo) SSE3Slow() bool {
+	return c.Features&SSE3SLOW != 0
+}
+
+// AesNi indicates support of AES-NI instructions
+// (Advanced Encryption Standard New Instructions)
+func (c CPUInfo) AesNi() bool {
+	return c.Features&AESNI != 0
+}
+
+// Clmul indicates support of CLMUL instructions
+// (Carry-less Multiplication)
+func (c CPUInfo) Clmul() bool {
+	return c.Features&CLMUL != 0
+}
+
+// NX indicates support of NX (No-Execute) bit
+func (c CPUInfo) NX() bool {
+	return c.Features&NX != 0
+}
+
+// SSE4A indicates support of AMD Barcelona microarchitecture SSE4a instructions
+func (c CPUInfo) SSE4A() bool {
+	return c.Features&SSE4A != 0
+}
+
+// HLE indicates support of Hardware Lock Elision
+func (c CPUInfo) HLE() bool {
+	return c.Features&HLE != 0
+}
+
+// RTM indicates support of Restricted Transactional Memory
+func (c CPUInfo) RTM() bool {
+	return c.Features&RTM != 0
+}
+
+// Rdrand indicates support of RDRAND instruction is available
+func (c CPUInfo) Rdrand() bool {
+	return c.Features&RDRAND != 0
+}
+
+// Rdseed indicates support of RDSEED instruction is available
+func (c CPUInfo) Rdseed() bool {
+	return c.Features&RDSEED != 0
+}
+
+// ADX indicates support of Intel ADX (Multi-Precision Add-Carry Instruction Extensions)
+func (c CPUInfo) ADX() bool {
+	return c.Features&ADX != 0
+}
+
+// SHA indicates support of Intel SHA Extensions
+func (c CPUInfo) SHA() bool {
+	return c.Features&SHA != 0
+}
+
+// AVX512F indicates support of AVX-512 Foundation
+func (c CPUInfo) AVX512F() bool {
+	return c.Features&AVX512F != 0
+}
+
+// AVX512DQ indicates support of AVX-512 Doubleword and Quadword Instructions
+func (c CPUInfo) AVX512DQ() bool {
+	return c.Features&AVX512DQ != 0
+}
+
+// AVX512IFMA indicates support of AVX-512 Integer Fused Multiply-Add Instructions
+func (c CPUInfo) AVX512IFMA() bool {
+	return c.Features&AVX512IFMA != 0
+}
+
+// AVX512PF indicates support of AVX-512 Prefetch Instructions
+func (c CPUInfo) AVX512PF() bool {
+	return c.Features&AVX512PF != 0
+}
+
+// AVX512ER indicates support of AVX-512 Exponential and Reciprocal Instructions
+func (c CPUInfo) AVX512ER() bool {
+	return c.Features&AVX512ER != 0
+}
+
+// AVX512CD indicates support of AVX-512 Conflict Detection Instructions
+func (c CPUInfo) AVX512CD() bool {
+	return c.Features&AVX512CD != 0
+}
+
+// AVX512BW indicates support of AVX-512 Byte and Word Instructions
+func (c CPUInfo) AVX512BW() bool {
+	return c.Features&AVX512BW != 0
+}
+
+// AVX512VL indicates support of AVX-512 Vector Length Extensions
+func (c CPUInfo) AVX512VL() bool {
+	return c.Features&AVX512VL != 0
+}
+
+// AVX512VBMI indicates support of AVX-512 Vector Bit Manipulation Instructions
+func (c CPUInfo) AVX512VBMI() bool {
+	return c.Features&AVX512VBMI != 0
+}
+
+// MPX indicates support of Intel MPX (Memory Protection Extensions)
+func (c CPUInfo) MPX() bool {
+	return c.Features&MPX != 0
+}
+
+// ERMS indicates support of Enhanced REP MOVSB/STOSB
+func (c CPUInfo) ERMS() bool {
+	return c.Features&ERMS != 0
+}
+
+// RDTSCP Instruction is available.
+func (c CPUInfo) RDTSCP() bool {
+	return c.Features&RDTSCP != 0
+}
+
+// CX16 indicates if CMPXCHG16B instruction is available.
+func (c CPUInfo) CX16() bool {
+	return c.Features&CX16 != 0
+}
+
+// TSX is split into HLE (Hardware Lock Elision) and RTM (Restricted Transactional Memory) detection.
+// So TSX simply checks that.
+func (c CPUInfo) TSX() bool {
+	return c.Features&(HLE|RTM) == HLE|RTM
+}
+
+// Atom indicates an Atom processor
+func (c CPUInfo) Atom() bool {
+	return c.Features&ATOM != 0
+}
+
+// Intel returns true if vendor is recognized as Intel
+func (c CPUInfo) Intel() bool {
+	return c.VendorID == Intel
+}
+
+// AMD returns true if vendor is recognized as AMD
+func (c CPUInfo) AMD() bool {
+	return c.VendorID == AMD
+}
+
+// Transmeta returns true if vendor is recognized as Transmeta
+func (c CPUInfo) Transmeta() bool {
+	return c.VendorID == Transmeta
+}
+
+// NSC returns true if vendor is recognized as National Semiconductor
+func (c CPUInfo) NSC() bool {
+	return c.VendorID == NSC
+}
+
+// VIA returns true if vendor is recognized as VIA
+func (c CPUInfo) VIA() bool {
+	return c.VendorID == VIA
+}
+
+// RTCounter returns the 64-bit time-stamp counter
+// Uses the RDTSCP instruction. The value 0 is returned
+// if the CPU does not support the instruction.
+func (c CPUInfo) RTCounter() uint64 {
+	if !c.RDTSCP() {
+		return 0
+	}
+	a, _, _, d := rdtscpAsm()
+	return uint64(a) | (uint64(d) << 32)
+}
+
+// Ia32TscAux returns the IA32_TSC_AUX part of the RDTSCP.
+// This variable is OS dependent, but on Linux contains information
+// about the current cpu/core the code is running on.
+// If the RDTSCP instruction isn't supported on the CPU, the value 0 is returned.
+func (c CPUInfo) Ia32TscAux() uint32 {
+	if !c.RDTSCP() {
+		return 0
+	}
+	_, _, ecx, _ := rdtscpAsm()
+	return ecx
+}
+
+// LogicalCPU will return the Logical CPU the code is currently executing on.
+// This is likely to change when the OS re-schedules the running thread
+// to another CPU.
+// If the current core cannot be detected, -1 will be returned.
+func (c CPUInfo) LogicalCPU() int {
+	if c.maxFunc < 1 {
+		return -1
+	}
+	_, ebx, _, _ := cpuid(1)
+	return int(ebx >> 24)
+}
+
+// VM Will return true if the cpu id indicates we are in
+// a virtual machine. This is only a hint, and will very likely
+// have many false negatives.
+func (c CPUInfo) VM() bool {
+	switch c.VendorID {
+	case MSVM, KVM, VMware, XenHVM:
+		return true
+	}
+	return false
+}
+
+// Flags contains detected cpu features and caracteristics
+type Flags uint64
+
+// String returns a string representation of the detected
+// CPU features.
+func (f Flags) String() string {
+	return strings.Join(f.Strings(), ",")
+}
+
+// Strings returns and array of the detected features.
+func (f Flags) Strings() []string {
+	s := support()
+	r := make([]string, 0, 20)
+	for i := uint(0); i < 64; i++ {
+		key := Flags(1 << i)
+		val := flagNames[key]
+		if s&key != 0 {
+			r = append(r, val)
+		}
+	}
+	return r
+}
+
+func maxExtendedFunction() uint32 {
+	eax, _, _, _ := cpuid(0x80000000)
+	return eax
+}
+
+func maxFunctionID() uint32 {
+	a, _, _, _ := cpuid(0)
+	return a
+}
+
+func brandName() string {
+	if maxExtendedFunction() >= 0x80000004 {
+		v := make([]uint32, 0, 48)
+		for i := uint32(0); i < 3; i++ {
+			a, b, c, d := cpuid(0x80000002 + i)
+			v = append(v, a, b, c, d)
+		}
+		return strings.Trim(string(valAsString(v...)), " ")
+	}
+	return "unknown"
+}
+
+func threadsPerCore() int {
+	mfi := maxFunctionID()
+	if mfi < 0x4 || vendorID() != Intel {
+		return 1
+	}
+
+	if mfi < 0xb {
+		_, b, _, d := cpuid(1)
+		if (d & (1 << 28)) != 0 {
+			// v will contain logical core count
+			v := (b >> 16) & 255
+			if v > 1 {
+				a4, _, _, _ := cpuid(4)
+				// physical cores
+				v2 := (a4 >> 26) + 1
+				if v2 > 0 {
+					return int(v) / int(v2)
+				}
+			}
+		}
+		return 1
+	}
+	_, b, _, _ := cpuidex(0xb, 0)
+	if b&0xffff == 0 {
+		return 1
+	}
+	return int(b & 0xffff)
+}
+
+func logicalCores() int {
+	mfi := maxFunctionID()
+	switch vendorID() {
+	case Intel:
+		// Use this on old Intel processors
+		if mfi < 0xb {
+			if mfi < 1 {
+				return 0
+			}
+			// CPUID.1:EBX[23:16] represents the maximum number of addressable IDs (initial APIC ID)
+			// that can be assigned to logical processors in a physical package.
+			// The value may not be the same as the number of logical processors that are present in the hardware of a physical package.
+			_, ebx, _, _ := cpuid(1)
+			logical := (ebx >> 16) & 0xff
+			return int(logical)
+		}
+		_, b, _, _ := cpuidex(0xb, 1)
+		return int(b & 0xffff)
+	case AMD:
+		_, b, _, _ := cpuid(1)
+		return int((b >> 16) & 0xff)
+	default:
+		return 0
+	}
+}
+
+func familyModel() (int, int) {
+	if maxFunctionID() < 0x1 {
+		return 0, 0
+	}
+	eax, _, _, _ := cpuid(1)
+	family := ((eax >> 8) & 0xf) + ((eax >> 20) & 0xff)
+	model := ((eax >> 4) & 0xf) + ((eax >> 12) & 0xf0)
+	return int(family), int(model)
+}
+
+func physicalCores() int {
+	switch vendorID() {
+	case Intel:
+		return logicalCores() / threadsPerCore()
+	case AMD:
+		if maxExtendedFunction() >= 0x80000008 {
+			_, _, c, _ := cpuid(0x80000008)
+			return int(c&0xff) + 1
+		}
+	}
+	return 0
+}
+
+// Except from http://en.wikipedia.org/wiki/CPUID#EAX.3D0:_Get_vendor_ID
+var vendorMapping = map[string]Vendor{
+	"AMDisbetter!": AMD,
+	"AuthenticAMD": AMD,
+	"CentaurHauls": VIA,
+	"GenuineIntel": Intel,
+	"TransmetaCPU": Transmeta,
+	"GenuineTMx86": Transmeta,
+	"Geode by NSC": NSC,
+	"VIA VIA VIA ": VIA,
+	"KVMKVMKVMKVM": KVM,
+	"Microsoft Hv": MSVM,
+	"VMwareVMware": VMware,
+	"XenVMMXenVMM": XenHVM,
+}
+
+func vendorID() Vendor {
+	_, b, c, d := cpuid(0)
+	v := valAsString(b, d, c)
+	vend, ok := vendorMapping[string(v)]
+	if !ok {
+		return Other
+	}
+	return vend
+}
+
+func cacheLine() int {
+	if maxFunctionID() < 0x1 {
+		return 0
+	}
+
+	_, ebx, _, _ := cpuid(1)
+	cache := (ebx & 0xff00) >> 5 // cflush size
+	if cache == 0 && maxExtendedFunction() >= 0x80000006 {
+		_, _, ecx, _ := cpuid(0x80000006)
+		cache = ecx & 0xff // cacheline size
+	}
+	// TODO: Read from Cache and TLB Information
+	return int(cache)
+}
+
+func (c *CPUInfo) cacheSize() {
+	c.Cache.L1D = -1
+	c.Cache.L1I = -1
+	c.Cache.L2 = -1
+	c.Cache.L3 = -1
+	vendor := vendorID()
+	switch vendor {
+	case Intel:
+		if maxFunctionID() < 4 {
+			return
+		}
+		for i := uint32(0); ; i++ {
+			eax, ebx, ecx, _ := cpuidex(4, i)
+			cacheType := eax & 15
+			if cacheType == 0 {
+				break
+			}
+			cacheLevel := (eax >> 5) & 7
+			coherency := int(ebx&0xfff) + 1
+			partitions := int((ebx>>12)&0x3ff) + 1
+			associativity := int((ebx>>22)&0x3ff) + 1
+			sets := int(ecx) + 1
+			size := associativity * partitions * coherency * sets
+			switch cacheLevel {
+			case 1:
+				if cacheType == 1 {
+					// 1 = Data Cache
+					c.Cache.L1D = size
+				} else if cacheType == 2 {
+					// 2 = Instruction Cache
+					c.Cache.L1I = size
+				} else {
+					if c.Cache.L1D < 0 {
+						c.Cache.L1I = size
+					}
+					if c.Cache.L1I < 0 {
+						c.Cache.L1I = size
+					}
+				}
+			case 2:
+				c.Cache.L2 = size
+			case 3:
+				c.Cache.L3 = size
+			}
+		}
+	case AMD:
+		// Untested.
+		if maxExtendedFunction() < 0x80000005 {
+			return
+		}
+		_, _, ecx, edx := cpuid(0x80000005)
+		c.Cache.L1D = int(((ecx >> 24) & 0xFF) * 1024)
+		c.Cache.L1I = int(((edx >> 24) & 0xFF) * 1024)
+
+		if maxExtendedFunction() < 0x80000006 {
+			return
+		}
+		_, _, ecx, _ = cpuid(0x80000006)
+		c.Cache.L2 = int(((ecx >> 16) & 0xFFFF) * 1024)
+	}
+
+	return
+}
+
+type SGXSupport struct {
+	Available           bool
+	SGX1Supported       bool
+	SGX2Supported       bool
+	MaxEnclaveSizeNot64 int64
+	MaxEnclaveSize64    int64
+}
+
+func hasSGX(available bool) (rval SGXSupport) {
+	rval.Available = available
+
+	if !available {
+		return
+	}
+
+	a, _, _, d := cpuidex(0x12, 0)
+	rval.SGX1Supported = a&0x01 != 0
+	rval.SGX2Supported = a&0x02 != 0
+	rval.MaxEnclaveSizeNot64 = 1 << (d & 0xFF)     // pow 2
+	rval.MaxEnclaveSize64 = 1 << ((d >> 8) & 0xFF) // pow 2
+
+	return
+}
+
+func support() Flags {
+	mfi := maxFunctionID()
+	vend := vendorID()
+	if mfi < 0x1 {
+		return 0
+	}
+	rval := uint64(0)
+	_, _, c, d := cpuid(1)
+	if (d & (1 << 15)) != 0 {
+		rval |= CMOV
+	}
+	if (d & (1 << 23)) != 0 {
+		rval |= MMX
+	}
+	if (d & (1 << 25)) != 0 {
+		rval |= MMXEXT
+	}
+	if (d & (1 << 25)) != 0 {
+		rval |= SSE
+	}
+	if (d & (1 << 26)) != 0 {
+		rval |= SSE2
+	}
+	if (c & 1) != 0 {
+		rval |= SSE3
+	}
+	if (c & 0x00000200) != 0 {
+		rval |= SSSE3
+	}
+	if (c & 0x00080000) != 0 {
+		rval |= SSE4
+	}
+	if (c & 0x00100000) != 0 {
+		rval |= SSE42
+	}
+	if (c & (1 << 25)) != 0 {
+		rval |= AESNI
+	}
+	if (c & (1 << 1)) != 0 {
+		rval |= CLMUL
+	}
+	if c&(1<<23) != 0 {
+		rval |= POPCNT
+	}
+	if c&(1<<30) != 0 {
+		rval |= RDRAND
+	}
+	if c&(1<<29) != 0 {
+		rval |= F16C
+	}
+	if c&(1<<13) != 0 {
+		rval |= CX16
+	}
+	if vend == Intel && (d&(1<<28)) != 0 && mfi >= 4 {
+		if threadsPerCore() > 1 {
+			rval |= HTT
+		}
+	}
+
+	// Check XGETBV, OXSAVE and AVX bits
+	if c&(1<<26) != 0 && c&(1<<27) != 0 && c&(1<<28) != 0 {
+		// Check for OS support
+		eax, _ := xgetbv(0)
+		if (eax & 0x6) == 0x6 {
+			rval |= AVX
+			if (c & 0x00001000) != 0 {
+				rval |= FMA3
+			}
+		}
+	}
+
+	// Check AVX2, AVX2 requires OS support, but BMI1/2 don't.
+	if mfi >= 7 {
+		_, ebx, ecx, edx := cpuidex(7, 0)
+		if (rval&AVX) != 0 && (ebx&0x00000020) != 0 {
+			rval |= AVX2
+		}
+		if (ebx & 0x00000008) != 0 {
+			rval |= BMI1
+			if (ebx & 0x00000100) != 0 {
+				rval |= BMI2
+			}
+		}
+		if ebx&(1<<2) != 0 {
+			rval |= SGX
+		}
+		if ebx&(1<<4) != 0 {
+			rval |= HLE
+		}
+		if ebx&(1<<9) != 0 {
+			rval |= ERMS
+		}
+		if ebx&(1<<11) != 0 {
+			rval |= RTM
+		}
+		if ebx&(1<<14) != 0 {
+			rval |= MPX
+		}
+		if ebx&(1<<18) != 0 {
+			rval |= RDSEED
+		}
+		if ebx&(1<<19) != 0 {
+			rval |= ADX
+		}
+		if ebx&(1<<29) != 0 {
+			rval |= SHA
+		}
+		if edx&(1<<26) != 0 {
+			rval |= IBPB
+		}
+		if edx&(1<<27) != 0 {
+			rval |= STIBP
+		}
+
+		// Only detect AVX-512 features if XGETBV is supported
+		if c&((1<<26)|(1<<27)) == (1<<26)|(1<<27) {
+			// Check for OS support
+			eax, _ := xgetbv(0)
+
+			// Verify that XCR0[7:5] = ‘111b’ (OPMASK state, upper 256-bit of ZMM0-ZMM15 and
+			// ZMM16-ZMM31 state are enabled by OS)
+			/// and that XCR0[2:1] = ‘11b’ (XMM state and YMM state are enabled by OS).
+			if (eax>>5)&7 == 7 && (eax>>1)&3 == 3 {
+				if ebx&(1<<16) != 0 {
+					rval |= AVX512F
+				}
+				if ebx&(1<<17) != 0 {
+					rval |= AVX512DQ
+				}
+				if ebx&(1<<21) != 0 {
+					rval |= AVX512IFMA
+				}
+				if ebx&(1<<26) != 0 {
+					rval |= AVX512PF
+				}
+				if ebx&(1<<27) != 0 {
+					rval |= AVX512ER
+				}
+				if ebx&(1<<28) != 0 {
+					rval |= AVX512CD
+				}
+				if ebx&(1<<30) != 0 {
+					rval |= AVX512BW
+				}
+				if ebx&(1<<31) != 0 {
+					rval |= AVX512VL
+				}
+				// ecx
+				if ecx&(1<<1) != 0 {
+					rval |= AVX512VBMI
+				}
+			}
+		}
+	}
+
+	if maxExtendedFunction() >= 0x80000001 {
+		_, _, c, d := cpuid(0x80000001)
+		if (c & (1 << 5)) != 0 {
+			rval |= LZCNT
+			rval |= POPCNT
+		}
+		if (d & (1 << 31)) != 0 {
+			rval |= AMD3DNOW
+		}
+		if (d & (1 << 30)) != 0 {
+			rval |= AMD3DNOWEXT
+		}
+		if (d & (1 << 23)) != 0 {
+			rval |= MMX
+		}
+		if (d & (1 << 22)) != 0 {
+			rval |= MMXEXT
+		}
+		if (c & (1 << 6)) != 0 {
+			rval |= SSE4A
+		}
+		if d&(1<<20) != 0 {
+			rval |= NX
+		}
+		if d&(1<<27) != 0 {
+			rval |= RDTSCP
+		}
+
+		/* Allow for selectively disabling SSE2 functions on AMD processors
+		   with SSE2 support but not SSE4a. This includes Athlon64, some
+		   Opteron, and some Sempron processors. MMX, SSE, or 3DNow! are faster
+		   than SSE2 often enough to utilize this special-case flag.
+		   AV_CPU_FLAG_SSE2 and AV_CPU_FLAG_SSE2SLOW are both set in this case
+		   so that SSE2 is used unless explicitly disabled by checking
+		   AV_CPU_FLAG_SSE2SLOW. */
+		if vendorID() != Intel &&
+			rval&SSE2 != 0 && (c&0x00000040) == 0 {
+			rval |= SSE2SLOW
+		}
+
+		/* XOP and FMA4 use the AVX instruction coding scheme, so they can't be
+		 * used unless the OS has AVX support. */
+		if (rval & AVX) != 0 {
+			if (c & 0x00000800) != 0 {
+				rval |= XOP
+			}
+			if (c & 0x00010000) != 0 {
+				rval |= FMA4
+			}
+		}
+
+		if vendorID() == Intel {
+			family, model := familyModel()
+			if family == 6 && (model == 9 || model == 13 || model == 14) {
+				/* 6/9 (pentium-m "banias"), 6/13 (pentium-m "dothan"), and
+				 * 6/14 (core1 "yonah") theoretically support sse2, but it's
+				 * usually slower than mmx. */
+				if (rval & SSE2) != 0 {
+					rval |= SSE2SLOW
+				}
+				if (rval & SSE3) != 0 {
+					rval |= SSE3SLOW
+				}
+			}
+			/* The Atom processor has SSSE3 support, which is useful in many cases,
+			 * but sometimes the SSSE3 version is slower than the SSE2 equivalent
+			 * on the Atom, but is generally faster on other processors supporting
+			 * SSSE3. This flag allows for selectively disabling certain SSSE3
+			 * functions on the Atom. */
+			if family == 6 && model == 28 {
+				rval |= ATOM
+			}
+		}
+	}
+	return Flags(rval)
+}
+
+func valAsString(values ...uint32) []byte {
+	r := make([]byte, 4*len(values))
+	for i, v := range values {
+		dst := r[i*4:]
+		dst[0] = byte(v & 0xff)
+		dst[1] = byte((v >> 8) & 0xff)
+		dst[2] = byte((v >> 16) & 0xff)
+		dst[3] = byte((v >> 24) & 0xff)
+		switch {
+		case dst[0] == 0:
+			return r[:i*4]
+		case dst[1] == 0:
+			return r[:i*4+1]
+		case dst[2] == 0:
+			return r[:i*4+2]
+		case dst[3] == 0:
+			return r[:i*4+3]
+		}
+	}
+	return r
+}
diff --git a/vendor/github.com/klauspost/cpuid/cpuid_386.s b/vendor/github.com/klauspost/cpuid/cpuid_386.s
new file mode 100644
index 000000000..4d731711e
--- /dev/null
+++ b/vendor/github.com/klauspost/cpuid/cpuid_386.s
@@ -0,0 +1,42 @@
+// Copyright (c) 2015 Klaus Post, released under MIT License. See LICENSE file.
+
+// +build 386,!gccgo
+
+// func asmCpuid(op uint32) (eax, ebx, ecx, edx uint32)
+TEXT ·asmCpuid(SB), 7, $0
+	XORL CX, CX
+	MOVL op+0(FP), AX
+	CPUID
+	MOVL AX, eax+4(FP)
+	MOVL BX, ebx+8(FP)
+	MOVL CX, ecx+12(FP)
+	MOVL DX, edx+16(FP)
+	RET
+
+// func asmCpuidex(op, op2 uint32) (eax, ebx, ecx, edx uint32)
+TEXT ·asmCpuidex(SB), 7, $0
+	MOVL op+0(FP), AX
+	MOVL op2+4(FP), CX
+	CPUID
+	MOVL AX, eax+8(FP)
+	MOVL BX, ebx+12(FP)
+	MOVL CX, ecx+16(FP)
+	MOVL DX, edx+20(FP)
+	RET
+
+// func xgetbv(index uint32) (eax, edx uint32)
+TEXT ·asmXgetbv(SB), 7, $0
+	MOVL index+0(FP), CX
+	BYTE $0x0f; BYTE $0x01; BYTE $0xd0 // XGETBV
+	MOVL AX, eax+4(FP)
+	MOVL DX, edx+8(FP)
+	RET
+
+// func asmRdtscpAsm() (eax, ebx, ecx, edx uint32)
+TEXT ·asmRdtscpAsm(SB), 7, $0
+	BYTE $0x0F; BYTE $0x01; BYTE $0xF9 // RDTSCP
+	MOVL AX, eax+0(FP)
+	MOVL BX, ebx+4(FP)
+	MOVL CX, ecx+8(FP)
+	MOVL DX, edx+12(FP)
+	RET
diff --git a/vendor/github.com/klauspost/cpuid/cpuid_amd64.s b/vendor/github.com/klauspost/cpuid/cpuid_amd64.s
new file mode 100644
index 000000000..3c1d60e42
--- /dev/null
+++ b/vendor/github.com/klauspost/cpuid/cpuid_amd64.s
@@ -0,0 +1,42 @@
+// Copyright (c) 2015 Klaus Post, released under MIT License. See LICENSE file.
+
+//+build amd64,!gccgo
+
+// func asmCpuid(op uint32) (eax, ebx, ecx, edx uint32)
+TEXT ·asmCpuid(SB), 7, $0
+	XORQ CX, CX
+	MOVL op+0(FP), AX
+	CPUID
+	MOVL AX, eax+8(FP)
+	MOVL BX, ebx+12(FP)
+	MOVL CX, ecx+16(FP)
+	MOVL DX, edx+20(FP)
+	RET
+
+// func asmCpuidex(op, op2 uint32) (eax, ebx, ecx, edx uint32)
+TEXT ·asmCpuidex(SB), 7, $0
+	MOVL op+0(FP), AX
+	MOVL op2+4(FP), CX
+	CPUID
+	MOVL AX, eax+8(FP)
+	MOVL BX, ebx+12(FP)
+	MOVL CX, ecx+16(FP)
+	MOVL DX, edx+20(FP)
+	RET
+
+// func asmXgetbv(index uint32) (eax, edx uint32)
+TEXT ·asmXgetbv(SB), 7, $0
+	MOVL index+0(FP), CX
+	BYTE $0x0f; BYTE $0x01; BYTE $0xd0 // XGETBV
+	MOVL AX, eax+8(FP)
+	MOVL DX, edx+12(FP)
+	RET
+
+// func asmRdtscpAsm() (eax, ebx, ecx, edx uint32)
+TEXT ·asmRdtscpAsm(SB), 7, $0
+	BYTE $0x0F; BYTE $0x01; BYTE $0xF9 // RDTSCP
+	MOVL AX, eax+0(FP)
+	MOVL BX, ebx+4(FP)
+	MOVL CX, ecx+8(FP)
+	MOVL DX, edx+12(FP)
+	RET
diff --git a/vendor/github.com/klauspost/cpuid/detect_intel.go b/vendor/github.com/klauspost/cpuid/detect_intel.go
new file mode 100644
index 000000000..a5f04dd6d
--- /dev/null
+++ b/vendor/github.com/klauspost/cpuid/detect_intel.go
@@ -0,0 +1,17 @@
+// Copyright (c) 2015 Klaus Post, released under MIT License. See LICENSE file.
+
+// +build 386,!gccgo amd64,!gccgo
+
+package cpuid
+
+func asmCpuid(op uint32) (eax, ebx, ecx, edx uint32)
+func asmCpuidex(op, op2 uint32) (eax, ebx, ecx, edx uint32)
+func asmXgetbv(index uint32) (eax, edx uint32)
+func asmRdtscpAsm() (eax, ebx, ecx, edx uint32)
+
+func initCPU() {
+	cpuid = asmCpuid
+	cpuidex = asmCpuidex
+	xgetbv = asmXgetbv
+	rdtscpAsm = asmRdtscpAsm
+}
diff --git a/vendor/github.com/klauspost/cpuid/detect_ref.go b/vendor/github.com/klauspost/cpuid/detect_ref.go
new file mode 100644
index 000000000..909c5d9a7
--- /dev/null
+++ b/vendor/github.com/klauspost/cpuid/detect_ref.go
@@ -0,0 +1,23 @@
+// Copyright (c) 2015 Klaus Post, released under MIT License. See LICENSE file.
+
+// +build !amd64,!386 gccgo
+
+package cpuid
+
+func initCPU() {
+	cpuid = func(op uint32) (eax, ebx, ecx, edx uint32) {
+		return 0, 0, 0, 0
+	}
+
+	cpuidex = func(op, op2 uint32) (eax, ebx, ecx, edx uint32) {
+		return 0, 0, 0, 0
+	}
+
+	xgetbv = func(index uint32) (eax, edx uint32) {
+		return 0, 0
+	}
+
+	rdtscpAsm = func() (eax, ebx, ecx, edx uint32) {
+		return 0, 0, 0, 0
+	}
+}
diff --git a/vendor/github.com/klauspost/cpuid/generate.go b/vendor/github.com/klauspost/cpuid/generate.go
new file mode 100644
index 000000000..90e7a98d2
--- /dev/null
+++ b/vendor/github.com/klauspost/cpuid/generate.go
@@ -0,0 +1,4 @@
+package cpuid
+
+//go:generate go run private-gen.go
+//go:generate gofmt -w ./private
diff --git a/vendor/github.com/klauspost/pgzip/GO_LICENSE b/vendor/github.com/klauspost/pgzip/GO_LICENSE
new file mode 100644
index 000000000..744875676
--- /dev/null
+++ b/vendor/github.com/klauspost/pgzip/GO_LICENSE
@@ -0,0 +1,27 @@
+Copyright (c) 2012 The 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/pgzip/LICENSE b/vendor/github.com/klauspost/pgzip/LICENSE
new file mode 100644
index 000000000..2bdc0d751
--- /dev/null
+++ b/vendor/github.com/klauspost/pgzip/LICENSE
@@ -0,0 +1,22 @@
+The MIT License (MIT)
+
+Copyright (c) 2014 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.
+
diff --git a/vendor/github.com/klauspost/pgzip/README.md b/vendor/github.com/klauspost/pgzip/README.md
new file mode 100644
index 000000000..81000996c
--- /dev/null
+++ b/vendor/github.com/klauspost/pgzip/README.md
@@ -0,0 +1,136 @@
+pgzip
+=====
+
+Go parallel gzip compression/decompression. This is a fully gzip compatible drop in replacement for "compress/gzip".
+
+This will split compression into blocks that are compressed in parallel. 
+This can be useful for compressing big amounts of data. The output is a standard gzip file.
+
+The gzip decompression is modified so it decompresses ahead of the current reader. 
+This means that reads will be non-blocking if the decompressor can keep ahead of your code reading from it. 
+CRC calculation also takes place in a separate goroutine.
+
+You should only use this if you are (de)compressing big amounts of data, 
+say **more than 1MB** at the time, otherwise you will not see any benefit, 
+and it will likely be faster to use the internal gzip library 
+or [this package](https://github.com/klauspost/compress).
+
+It is important to note that this library creates and reads *standard gzip files*. 
+You do not have to match the compressor/decompressor to get the described speedups, 
+and the gzip files are fully compatible with other gzip readers/writers.
+
+A golang variant of this is [bgzf](https://godoc.org/github.com/biogo/hts/bgzf), 
+which has the same feature, as well as seeking in the resulting file. 
+The only drawback is a slightly bigger overhead compared to this and pure gzip. 
+See a comparison below.
+
+[![GoDoc][1]][2] [![Build Status][3]][4]
+
+[1]: https://godoc.org/github.com/klauspost/pgzip?status.svg
+[2]: https://godoc.org/github.com/klauspost/pgzip
+[3]: https://travis-ci.org/klauspost/pgzip.svg
+[4]: https://travis-ci.org/klauspost/pgzip
+
+Installation
+====
+```go get github.com/klauspost/pgzip/...```
+
+You might need to get/update the dependencies:
+
+```
+go get -u github.com/klauspost/compress
+go get -u github.com/klauspost/crc32
+```
+
+Usage
+====
+[Godoc Doumentation](https://godoc.org/github.com/klauspost/pgzip)
+
+To use as a replacement for gzip, exchange 
+
+```import "compress/gzip"``` 
+with 
+```import gzip "github.com/klauspost/pgzip"```.
+
+# Changes
+
+* Oct 6, 2016: Fixed an issue if the destination writer returned an error.
+* Oct 6, 2016: Better buffer reuse, should now generate less garbage.
+* Oct 6, 2016: Output does not change based on write sizes.
+* Dec 8, 2015: Decoder now supports the io.WriterTo interface, giving a speedup and less GC pressure.
+* Oct 9, 2015: Reduced allocations by ~35 by using sync.Pool. ~15% overall speedup.
+
+Changes in [github.com/klauspost/compress](https://github.com/klauspost/compress#changelog) are also carried over, so see that for more changes.
+
+## Compression
+The simplest way to use this is to simply do the same as you would when using [compress/gzip](http://golang.org/pkg/compress/gzip). 
+
+To change the block size, use the added (*pgzip.Writer).SetConcurrency(blockSize, blocks int) function. With this you can control the approximate size of your blocks, as well as how many you want to be processing in parallel. Default values for this is SetConcurrency(250000, 16), meaning blocks are split at 250000 bytes and up to 16 blocks can be processing at once before the writer blocks.
+
+
+Example:
+```
+var b bytes.Buffer
+w := gzip.NewWriter(&b)
+w.SetConcurrency(100000, 10)
+w.Write([]byte("hello, world\n"))
+w.Close()
+```
+
+To get any performance gains, you should at least be compressing more than 1 megabyte of data at the time.
+
+You should at least have a block size of 100k and at least a number of blocks that match the number of cores your would like to utilize, but about twice the number of blocks would be the best.
+
+Another side effect of this is, that it is likely to speed up your other code, since writes to the compressor only blocks if the compressor is already compressing the number of blocks you have specified. This also means you don't have worry about buffering input to the compressor.
+
+## Decompression
+
+Decompression works similar to compression. That means that you simply call pgzip the same way as you would call [compress/gzip](http://golang.org/pkg/compress/gzip). 
+
+The only difference is that if you want to specify your own readahead, you have to use `pgzip.NewReaderN(r io.Reader, blockSize, blocks int)` to get a reader with your custom blocksizes. The `blockSize` is the size of each block decoded, and `blocks` is the maximum number of blocks that is decoded ahead.
+
+See [Example on playground](http://play.golang.org/p/uHv1B5NbDh)
+
+Performance
+====
+## Compression
+
+See my blog post in [Benchmarks of Golang Gzip](https://blog.klauspost.com/go-gzipdeflate-benchmarks/).
+
+Compression cost is usually about 0.2% with default settings with a block size of 250k.
+
+Example with GOMAXPROC set to 8 (quad core with 8 hyperthreads)
+
+Content is [Matt Mahoneys 10GB corpus](http://mattmahoney.net/dc/10gb.html). Compression level 6.
+
+Compressor  | MB/sec   | speedup | size | size overhead (lower=better)
+------------|----------|---------|------|---------
+[gzip](http://golang.org/pkg/compress/gzip) (golang) | 7.21MB/s | 1.0x | 4786608902 | 0%
+[gzip](http://github.com/klauspost/compress/gzip) (klauspost) | 10.98MB/s | 1.52x | 4781331645 | -0.11%
+[pgzip](https://github.com/klauspost/pgzip) (klauspost) | 50.76MB/s|7.04x | 4784121440 | -0.052%
+[bgzf](https://godoc.org/github.com/biogo/hts/bgzf) (biogo) | 38.65MB/s | 5.36x | 4924899484 | 2.889%
+[pargzip](https://godoc.org/github.com/golang/build/pargzip) (builder) | 32.00MB/s | 4.44x | 4791226567 | 0.096%
+
+pgzip also contains a [linear time compression](https://github.com/klauspost/compress#linear-time-compression) mode, that will allow compression at ~150MB per core per second, independent of the content.
+
+See the [complete sheet](https://docs.google.com/spreadsheets/d/1nuNE2nPfuINCZJRMt6wFWhKpToF95I47XjSsc-1rbPQ/edit?usp=sharing) for different content types and compression settings.
+
+## Decompression
+
+The decompression speedup is there because it allows you to do other work while the decompression is taking place.
+
+In the example above, the numbers are as follows on a 4 CPU machine:
+
+Decompressor | Time | Speedup
+-------------|------|--------
+[gzip](http://golang.org/pkg/compress/gzip) (golang) | 1m28.85s | 0%
+[pgzip](https://github.com/klauspost/pgzip) (golang) | 43.48s | 104%
+
+But wait, since gzip decompression is inherently singlethreaded (aside from CRC calculation) how can it be more than 100% faster?  Because pgzip due to its design also acts as a buffer. When using unbuffered gzip, you are also waiting for io when you are decompressing. If the gzip decoder can keep up, it will always have data ready for your reader, and you will not be waiting for input to the gzip decompressor to complete.
+
+This is pretty much an optimal situation for pgzip, but it reflects most common usecases for CPU intensive gzip usage.
+
+I haven't included [bgzf](https://godoc.org/github.com/biogo/hts/bgzf) in this comparison, since it only can decompress files created by a compatible encoder, and therefore cannot be considered a generic gzip decompressor. But if you are able to compress your files with a bgzf compatible program, you can expect it to scale beyond 100%.
+
+# License
+This contains large portions of code from the go repository - see GO_LICENSE for more information. The changes are released under MIT License. See LICENSE for more information.
diff --git a/vendor/github.com/klauspost/pgzip/gunzip.go b/vendor/github.com/klauspost/pgzip/gunzip.go
new file mode 100644
index 000000000..93efec714
--- /dev/null
+++ b/vendor/github.com/klauspost/pgzip/gunzip.go
@@ -0,0 +1,573 @@
+// 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 pgzip implements reading and writing of gzip format compressed files,
+// as specified in RFC 1952.
+//
+// This is a drop in replacement for "compress/gzip".
+// This will split compression into blocks that are compressed in parallel.
+// This can be useful for compressing big amounts of data.
+// The gzip decompression has not been modified, but remains in the package,
+// so you can use it as a complete replacement for "compress/gzip".
+//
+// See more at https://github.com/klauspost/pgzip
+package pgzip
+
+import (
+	"bufio"
+	"errors"
+	"hash"
+	"hash/crc32"
+	"io"
+	"sync"
+	"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
+)
+
+func makeReader(r io.Reader) flate.Reader {
+	if rr, ok := r.(flate.Reader); ok {
+		return rr
+	}
+	return bufio.NewReader(r)
+}
+
+var (
+	// ErrChecksum is returned when reading GZIP data that has an invalid checksum.
+	ErrChecksum = errors.New("gzip: invalid checksum")
+	// ErrHeader is returned when reading GZIP data that has an invalid header.
+	ErrHeader = errors.New("gzip: invalid header")
+)
+
+// 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.
+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
+	r            flate.Reader
+	decompressor io.ReadCloser
+	digest       hash.Hash32
+	size         uint32
+	flg          byte
+	buf          [512]byte
+	err          error
+	closeErr     chan error
+	multistream  bool
+
+	readAhead   chan read
+	roff        int // read offset
+	current     []byte
+	closeReader chan struct{}
+	lastBlock   bool
+	blockSize   int
+	blocks      int
+
+	activeRA bool       // Indication if readahead is active
+	mu       sync.Mutex // Lock for above
+
+	blockPool chan []byte
+}
+
+type read struct {
+	b   []byte
+	err error
+}
+
+// NewReader creates a new Reader reading the given reader.
+// The implementation buffers input and may read more data than necessary from r.
+// It is the caller's responsibility to call Close on the Reader when done.
+func NewReader(r io.Reader) (*Reader, error) {
+	z := new(Reader)
+	z.blocks = defaultBlocks
+	z.blockSize = defaultBlockSize
+	z.r = makeReader(r)
+	z.digest = crc32.NewIEEE()
+	z.multistream = true
+	z.blockPool = make(chan []byte, z.blocks)
+	for i := 0; i < z.blocks; i++ {
+		z.blockPool <- make([]byte, z.blockSize)
+	}
+	if err := z.readHeader(true); err != nil {
+		return nil, err
+	}
+	return z, nil
+}
+
+// NewReaderN creates a new Reader reading the given reader.
+// The implementation buffers input and may read more data than necessary from r.
+// It is the caller's responsibility to call Close on the Reader when done.
+//
+// With this you can control the approximate size of your blocks,
+// as well as how many blocks you want to have prefetched.
+//
+// Default values for this is blockSize = 250000, blocks = 16,
+// meaning up to 16 blocks of maximum 250000 bytes will be
+// prefetched.
+func NewReaderN(r io.Reader, blockSize, blocks int) (*Reader, error) {
+	z := new(Reader)
+	z.blocks = blocks
+	z.blockSize = blockSize
+	z.r = makeReader(r)
+	z.digest = crc32.NewIEEE()
+	z.multistream = true
+
+	// Account for too small values
+	if z.blocks <= 0 {
+		z.blocks = defaultBlocks
+	}
+	if z.blockSize <= 512 {
+		z.blockSize = defaultBlockSize
+	}
+	z.blockPool = make(chan []byte, z.blocks)
+	for i := 0; i < z.blocks; i++ {
+		z.blockPool <- make([]byte, z.blockSize)
+	}
+	if err := z.readHeader(true); 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.killReadAhead()
+	z.r = makeReader(r)
+	z.digest = crc32.NewIEEE()
+	z.size = 0
+	z.err = nil
+	z.multistream = true
+
+	// Account for uninitialized values
+	if z.blocks <= 0 {
+		z.blocks = defaultBlocks
+	}
+	if z.blockSize <= 512 {
+		z.blockSize = defaultBlockSize
+	}
+
+	if z.blockPool == nil {
+		z.blockPool = make(chan []byte, z.blocks)
+		for i := 0; i < z.blocks; i++ {
+			z.blockPool <- make([]byte, z.blockSize)
+		}
+	}
+
+	return z.readHeader(true)
+}
+
+// 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
+}
+
+// GZIP (RFC 1952) is little-endian, unlike ZLIB (RFC 1950).
+func get4(p []byte) uint32 {
+	return uint32(p[0]) | uint32(p[1])<<8 | uint32(p[2])<<16 | uint32(p[3])<<24
+}
+
+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 {
+			// GZIP (RFC 1952) specifies that strings are NUL-terminated ISO 8859-1 (Latin-1).
+			if needconv {
+				s := make([]rune, 0, i)
+				for _, v := range z.buf[0:i] {
+					s = append(s, rune(v))
+				}
+				return string(s), nil
+			}
+			return string(z.buf[0:i]), nil
+		}
+	}
+}
+
+func (z *Reader) read2() (uint32, error) {
+	_, err := io.ReadFull(z.r, z.buf[0:2])
+	if err != nil {
+		return 0, err
+	}
+	return uint32(z.buf[0]) | uint32(z.buf[1])<<8, nil
+}
+
+func (z *Reader) readHeader(save bool) error {
+	z.killReadAhead()
+
+	_, err := io.ReadFull(z.r, z.buf[0:10])
+	if err != nil {
+		return err
+	}
+	if z.buf[0] != gzipID1 || z.buf[1] != gzipID2 || z.buf[2] != gzipDeflate {
+		return ErrHeader
+	}
+	z.flg = z.buf[3]
+	if save {
+		z.ModTime = time.Unix(int64(get4(z.buf[4:8])), 0)
+		// z.buf[8] is xfl, ignored
+		z.OS = z.buf[9]
+	}
+	z.digest.Reset()
+	z.digest.Write(z.buf[0:10])
+
+	if z.flg&flagExtra != 0 {
+		n, err := z.read2()
+		if err != nil {
+			return err
+		}
+		data := make([]byte, n)
+		if _, err = io.ReadFull(z.r, data); err != nil {
+			return err
+		}
+		if save {
+			z.Extra = data
+		}
+	}
+
+	var s string
+	if z.flg&flagName != 0 {
+		if s, err = z.readString(); err != nil {
+			return err
+		}
+		if save {
+			z.Name = s
+		}
+	}
+
+	if z.flg&flagComment != 0 {
+		if s, err = z.readString(); err != nil {
+			return err
+		}
+		if save {
+			z.Comment = s
+		}
+	}
+
+	if z.flg&flagHdrCrc != 0 {
+		n, err := z.read2()
+		if err != nil {
+			return err
+		}
+		sum := z.digest.Sum32() & 0xFFFF
+		if n != sum {
+			return ErrHeader
+		}
+	}
+
+	z.digest.Reset()
+	z.decompressor = flate.NewReader(z.r)
+	z.doReadAhead()
+	return nil
+}
+
+func (z *Reader) killReadAhead() error {
+	z.mu.Lock()
+	defer z.mu.Unlock()
+	if z.activeRA {
+		if z.closeReader != nil {
+			close(z.closeReader)
+		}
+
+		// Wait for decompressor to be closed and return error, if any.
+		e, ok := <-z.closeErr
+		z.activeRA = false
+		if !ok {
+			// Channel is closed, so if there was any error it has already been returned.
+			return nil
+		}
+		return e
+	}
+	return nil
+}
+
+// Starts readahead.
+// Will return on error (including io.EOF)
+// or when z.closeReader is closed.
+func (z *Reader) doReadAhead() {
+	z.mu.Lock()
+	defer z.mu.Unlock()
+	z.activeRA = true
+
+	if z.blocks <= 0 {
+		z.blocks = defaultBlocks
+	}
+	if z.blockSize <= 512 {
+		z.blockSize = defaultBlockSize
+	}
+	ra := make(chan read, z.blocks)
+	z.readAhead = ra
+	closeReader := make(chan struct{}, 0)
+	z.closeReader = closeReader
+	z.lastBlock = false
+	closeErr := make(chan error, 1)
+	z.closeErr = closeErr
+	z.size = 0
+	z.roff = 0
+	z.current = nil
+	decomp := z.decompressor
+
+	go func() {
+		defer func() {
+			closeErr <- decomp.Close()
+			close(closeErr)
+			close(ra)
+		}()
+
+		// We hold a local reference to digest, since
+		// it way be changed by reset.
+		digest := z.digest
+		var wg sync.WaitGroup
+		for {
+			var buf []byte
+			select {
+			case buf = <-z.blockPool:
+			case <-closeReader:
+				return
+			}
+			buf = buf[0:z.blockSize]
+			// Try to fill the buffer
+			n, err := io.ReadFull(decomp, buf)
+			if err == io.ErrUnexpectedEOF {
+				if n > 0 {
+					err = nil
+				} else {
+					// If we got zero bytes, we need to establish if
+					// we reached end of stream or truncated stream.
+					_, err = decomp.Read([]byte{})
+					if err == io.EOF {
+						err = nil
+					}
+				}
+			}
+			if n < len(buf) {
+				buf = buf[0:n]
+			}
+			wg.Wait()
+			wg.Add(1)
+			go func() {
+				digest.Write(buf)
+				wg.Done()
+			}()
+			z.size += uint32(n)
+
+			// If we return any error, out digest must be ready
+			if err != nil {
+				wg.Wait()
+			}
+			select {
+			case z.readAhead <- read{b: buf, err: err}:
+			case <-closeReader:
+				// Sent on close, we don't care about the next results
+				return
+			}
+			if err != nil {
+				return
+			}
+		}
+	}()
+}
+
+func (z *Reader) Read(p []byte) (n int, err error) {
+	if z.err != nil {
+		return 0, z.err
+	}
+	if len(p) == 0 {
+		return 0, nil
+	}
+
+	for {
+		if len(z.current) == 0 && !z.lastBlock {
+			read := <-z.readAhead
+
+			if read.err != nil {
+				// If not nil, the reader will have exited
+				z.closeReader = nil
+
+				if read.err != io.EOF {
+					z.err = read.err
+					return
+				}
+				if read.err == io.EOF {
+					z.lastBlock = true
+					err = nil
+				}
+			}
+			z.current = read.b
+			z.roff = 0
+		}
+		avail := z.current[z.roff:]
+		if len(p) >= len(avail) {
+			// If len(p) >= len(current), return all content of current
+			n = copy(p, avail)
+			z.blockPool <- z.current
+			z.current = nil
+			if z.lastBlock {
+				err = io.EOF
+				break
+			}
+		} else {
+			// We copy as much as there is space for
+			n = copy(p, avail)
+			z.roff += n
+		}
+		return
+	}
+
+	// Finished file; check checksum + size.
+	if _, err := io.ReadFull(z.r, z.buf[0:8]); err != nil {
+		z.err = err
+		return 0, err
+	}
+	crc32, isize := get4(z.buf[0:4]), get4(z.buf[4:8])
+	sum := z.digest.Sum32()
+	if sum != crc32 || isize != z.size {
+		z.err = ErrChecksum
+		return 0, z.err
+	}
+
+	// File is ok; should we attempt reading one more?
+	if !z.multistream {
+		return 0, io.EOF
+	}
+
+	// Is there another?
+	if err = z.readHeader(false); err != nil {
+		z.err = err
+		return
+	}
+
+	// Yes.  Reset and read from it.
+	return z.Read(p)
+}
+
+func (z *Reader) WriteTo(w io.Writer) (n int64, err error) {
+	total := int64(0)
+	for {
+		if z.err != nil {
+			return total, z.err
+		}
+		// We write both to output and digest.
+		for {
+			// Read from input
+			read := <-z.readAhead
+			if read.err != nil {
+				// If not nil, the reader will have exited
+				z.closeReader = nil
+
+				if read.err != io.EOF {
+					z.err = read.err
+					return total, z.err
+				}
+				if read.err == io.EOF {
+					z.lastBlock = true
+					err = nil
+				}
+			}
+			// Write what we got
+			n, err := w.Write(read.b)
+			if n != len(read.b) {
+				return total, io.ErrShortWrite
+			}
+			total += int64(n)
+			if err != nil {
+				return total, err
+			}
+			// Put block back
+			z.blockPool <- read.b
+			if z.lastBlock {
+				break
+			}
+		}
+
+		// Finished file; check checksum + size.
+		if _, err := io.ReadFull(z.r, z.buf[0:8]); err != nil {
+			z.err = err
+			return total, err
+		}
+		crc32, isize := get4(z.buf[0:4]), get4(z.buf[4:8])
+		sum := z.digest.Sum32()
+		if sum != crc32 || isize != z.size {
+			z.err = ErrChecksum
+			return total, z.err
+		}
+		// File is ok; should we attempt reading one more?
+		if !z.multistream {
+			return total, nil
+		}
+
+		// Is there another?
+		err = z.readHeader(false)
+		if err == io.EOF {
+			return total, nil
+		}
+		if err != nil {
+			z.err = err
+			return total, err
+		}
+	}
+}
+
+// Close closes the Reader. It does not close the underlying io.Reader.
+func (z *Reader) Close() error {
+	return z.killReadAhead()
+}
diff --git a/vendor/github.com/klauspost/pgzip/gzip.go b/vendor/github.com/klauspost/pgzip/gzip.go
new file mode 100644
index 000000000..85d14e9cb
--- /dev/null
+++ b/vendor/github.com/klauspost/pgzip/gzip.go
@@ -0,0 +1,501 @@
+// 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 pgzip
+
+import (
+	"bytes"
+	"errors"
+	"fmt"
+	"hash"
+	"hash/crc32"
+	"io"
+	"sync"
+	"time"
+
+	"github.com/klauspost/compress/flate"
+)
+
+const (
+	defaultBlockSize = 256 << 10
+	tailSize         = 16384
+	defaultBlocks    = 16
+)
+
+// 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
+)
+
+// A Writer is an io.WriteCloser.
+// Writes to a Writer are compressed and written to w.
+type Writer struct {
+	Header
+	w             io.Writer
+	level         int
+	wroteHeader   bool
+	blockSize     int
+	blocks        int
+	currentBuffer []byte
+	prevTail      []byte
+	digest        hash.Hash32
+	size          int
+	closed        bool
+	buf           [10]byte
+	errMu         sync.RWMutex
+	err           error
+	pushedErr     chan struct{}
+	results       chan result
+	dictFlatePool sync.Pool
+	dstPool       sync.Pool
+	wg            sync.WaitGroup
+}
+
+type result struct {
+	result        chan []byte
+	notifyWritten chan struct{}
+}
+
+// Use SetConcurrency to finetune the concurrency level if needed.
+//
+// With this you can control the approximate size of your blocks,
+// as well as how many you want to be processing in parallel.
+//
+// Default values for this is SetConcurrency(250000, 16),
+// meaning blocks are split at 250000 bytes and up to 16 blocks
+// can be processing at once before the writer blocks.
+func (z *Writer) SetConcurrency(blockSize, blocks int) error {
+	if blockSize <= tailSize {
+		return fmt.Errorf("gzip: block size cannot be less than or equal to %d", tailSize)
+	}
+	if blocks <= 0 {
+		return errors.New("gzip: blocks cannot be zero or less")
+	}
+	if blockSize == z.blockSize && blocks == z.blocks {
+		return nil
+	}
+	z.blockSize = blockSize
+	z.results = make(chan result, blocks)
+	z.blocks = blocks
+	z.dstPool = sync.Pool{New: func() interface{} { return make([]byte, 0, blockSize+(blockSize)>>4) }}
+	return nil
+}
+
+// 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 or Close. The Comment and Name header fields are
+// UTF-8 strings in Go, but the underlying format requires NUL-terminated ISO
+// 8859-1 (Latin-1). NUL or non-Latin-1 runes in those strings will lead to an
+// error on Write.
+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 < ConstantCompression || level > BestCompression {
+		return nil, fmt.Errorf("gzip: invalid compression level: %d", level)
+	}
+	z := new(Writer)
+	z.SetConcurrency(defaultBlockSize, defaultBlocks)
+	z.init(w, level)
+	return z, nil
+}
+
+// This function must be used by goroutines to set an
+// error condition, since z.err access is restricted
+// to the callers goruotine.
+func (z *Writer) pushError(err error) {
+	z.errMu.Lock()
+	if z.err != nil {
+		z.errMu.Unlock()
+		return
+	}
+	z.err = err
+	close(z.pushedErr)
+	z.errMu.Unlock()
+}
+
+func (z *Writer) init(w io.Writer, level int) {
+	z.wg.Wait()
+	digest := z.digest
+	if digest != nil {
+		digest.Reset()
+	} else {
+		digest = crc32.NewIEEE()
+	}
+	z.Header = Header{OS: 255}
+	z.w = w
+	z.level = level
+	z.digest = digest
+	z.pushedErr = make(chan struct{}, 0)
+	z.results = make(chan result, z.blocks)
+	z.err = nil
+	z.closed = false
+	z.Comment = ""
+	z.Extra = nil
+	z.ModTime = time.Time{}
+	z.wroteHeader = false
+	z.currentBuffer = nil
+	z.buf = [10]byte{}
+	z.prevTail = nil
+	z.size = 0
+	if z.dictFlatePool.New == nil {
+		z.dictFlatePool.New = func() interface{} {
+			f, _ := flate.NewWriterDict(w, level, nil)
+			return f
+		}
+	}
+}
+
+// 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) {
+	if z.results != nil && !z.closed {
+		close(z.results)
+	}
+	z.SetConcurrency(defaultBlockSize, defaultBlocks)
+	z.init(w, z.level)
+}
+
+// GZIP (RFC 1952) is little-endian, unlike ZLIB (RFC 1950).
+func put2(p []byte, v uint16) {
+	p[0] = uint8(v >> 0)
+	p[1] = uint8(v >> 8)
+}
+
+func put4(p []byte, v uint32) {
+	p[0] = uint8(v >> 0)
+	p[1] = uint8(v >> 8)
+	p[2] = uint8(v >> 16)
+	p[3] = uint8(v >> 24)
+}
+
+// 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")
+	}
+	put2(z.buf[0:2], uint16(len(b)))
+	_, err := z.w.Write(z.buf[0:2])
+	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[0:1])
+	return err
+}
+
+// compressCurrent will compress the data currently buffered
+// This should only be called from the main writer/flush/closer
+func (z *Writer) compressCurrent(flush bool) {
+	r := result{}
+	r.result = make(chan []byte, 1)
+	r.notifyWritten = make(chan struct{}, 0)
+	select {
+	case z.results <- r:
+	case <-z.pushedErr:
+		return
+	}
+
+	// If block given is more than twice the block size, split it.
+	c := z.currentBuffer
+	if len(c) > z.blockSize*2 {
+		c = c[:z.blockSize]
+		z.wg.Add(1)
+		go z.compressBlock(c, z.prevTail, r, false)
+		z.prevTail = c[len(c)-tailSize:]
+		z.currentBuffer = z.currentBuffer[z.blockSize:]
+		z.compressCurrent(flush)
+		// Last one flushes if needed
+		return
+	}
+
+	z.wg.Add(1)
+	go z.compressBlock(c, z.prevTail, r, z.closed)
+	if len(c) > tailSize {
+		z.prevTail = c[len(c)-tailSize:]
+	} else {
+		z.prevTail = nil
+	}
+	z.currentBuffer = z.dstPool.Get().([]byte)
+	z.currentBuffer = z.currentBuffer[:0]
+
+	// Wait if flushing
+	if flush {
+		<-r.notifyWritten
+	}
+}
+
+// Returns an error if it has been set.
+// Cannot be used by functions that are from internal goroutines.
+func (z *Writer) checkError() error {
+	z.errMu.RLock()
+	err := z.err
+	z.errMu.RUnlock()
+	return err
+}
+
+// Write writes a compressed form of p to the underlying io.Writer. The
+// compressed bytes are not necessarily flushed to output until
+// the Writer is closed or Flush() is called.
+//
+// The function will return quickly, if there are unused buffers.
+// The sent slice (p) is copied, and the caller is free to re-use the buffer
+// when the function returns.
+//
+// Errors that occur during compression will be reported later, and a nil error
+// does not signify that the compression succeeded (since it is most likely still running)
+// That means that the call that returns an error may not be the call that caused it.
+// Only Flush and Close functions are guaranteed to return any errors up to that point.
+func (z *Writer) Write(p []byte) (int, error) {
+	if err := z.checkError(); err != nil {
+		return 0, err
+	}
+	// 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
+		}
+		put4(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
+		var n int
+		var err error
+		n, err = z.w.Write(z.buf[0:10])
+		if err != nil {
+			z.pushError(err)
+			return n, err
+		}
+		if z.Extra != nil {
+			err = z.writeBytes(z.Extra)
+			if err != nil {
+				z.pushError(err)
+				return n, err
+			}
+		}
+		if z.Name != "" {
+			err = z.writeString(z.Name)
+			if err != nil {
+				z.pushError(err)
+				return n, err
+			}
+		}
+		if z.Comment != "" {
+			err = z.writeString(z.Comment)
+			if err != nil {
+				z.pushError(err)
+				return n, err
+			}
+		}
+		// Start receiving data from compressors
+		go func() {
+			listen := z.results
+			for {
+				r, ok := <-listen
+				// If closed, we are finished.
+				if !ok {
+					return
+				}
+				buf := <-r.result
+				n, err := z.w.Write(buf)
+				if err != nil {
+					z.pushError(err)
+					close(r.notifyWritten)
+					return
+				}
+				if n != len(buf) {
+					z.pushError(fmt.Errorf("gzip: short write %d should be %d", n, len(buf)))
+					close(r.notifyWritten)
+					return
+				}
+				z.dstPool.Put(buf)
+				close(r.notifyWritten)
+			}
+		}()
+		z.currentBuffer = make([]byte, 0, z.blockSize)
+	}
+	q := p
+	for len(q) > 0 {
+		length := len(q)
+		if length+len(z.currentBuffer) > z.blockSize {
+			length = z.blockSize - len(z.currentBuffer)
+		}
+		z.digest.Write(q[:length])
+		z.currentBuffer = append(z.currentBuffer, q[:length]...)
+		if len(z.currentBuffer) >= z.blockSize {
+			z.compressCurrent(false)
+			if err := z.checkError(); err != nil {
+				return len(p) - len(q) - length, err
+			}
+		}
+		z.size += length
+		q = q[length:]
+	}
+	return len(p), z.checkError()
+}
+
+// Step 1: compresses buffer to buffer
+// Step 2: send writer to channel
+// Step 3: Close result channel to indicate we are done
+func (z *Writer) compressBlock(p, prevTail []byte, r result, closed bool) {
+	defer func() {
+		close(r.result)
+		z.wg.Done()
+	}()
+	buf := z.dstPool.Get().([]byte)
+	dest := bytes.NewBuffer(buf[:0])
+
+	compressor := z.dictFlatePool.Get().(*flate.Writer)
+	compressor.ResetDict(dest, prevTail)
+	compressor.Write(p)
+
+	err := compressor.Flush()
+	if err != nil {
+		z.pushError(err)
+		return
+	}
+	if closed {
+		err = compressor.Close()
+		if err != nil {
+			z.pushError(err)
+			return
+		}
+	}
+	z.dictFlatePool.Put(compressor)
+	// Read back buffer
+	buf = dest.Bytes()
+	r.result <- buf
+}
+
+// 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 err := z.checkError(); err != nil {
+		return err
+	}
+	if z.closed {
+		return nil
+	}
+	if !z.wroteHeader {
+		_, err := z.Write(nil)
+		if err != nil {
+			return err
+		}
+	}
+	// We send current block to compression
+	z.compressCurrent(true)
+
+	return z.checkError()
+}
+
+// UncompressedSize will return the number of bytes written.
+// pgzip only, not a function in the official gzip package.
+func (z *Writer) UncompressedSize() int {
+	return z.size
+}
+
+// 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 err := z.checkError(); err != nil {
+		return err
+	}
+	if z.closed {
+		return nil
+	}
+
+	z.closed = true
+	if !z.wroteHeader {
+		z.Write(nil)
+		if err := z.checkError(); err != nil {
+			return err
+		}
+	}
+	z.compressCurrent(true)
+	if err := z.checkError(); err != nil {
+		return err
+	}
+	close(z.results)
+	put4(z.buf[0:4], z.digest.Sum32())
+	put4(z.buf[4:8], uint32(z.size))
+	_, err := z.w.Write(z.buf[0:8])
+	if err != nil {
+		z.pushError(err)
+		return err
+	}
+	return nil
+}