Added all required dependencies

36 files changed, 8870 insertions, 0 deletions

diff --git a/vendor/github.com/klauspost/compress/LICENSE b/vendor/github.com/klauspost/compress/LICENSE
new file mode 100644
index 0000000000..7448756763
--- /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/flate/copy.go b/vendor/github.com/klauspost/compress/flate/copy.go
new file mode 100644
index 0000000000..a3200a8f49
--- /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 0000000000..70a6095e60
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/crc32_amd64.go
@@ -0,0 +1,41 @@
+//+build !noasm
+//+build !appengine
+
+// 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 0000000000..2fb2079b9d
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/crc32_amd64.s
@@ -0,0 +1,213 @@
+//+build !noasm
+//+build !appengine
+
+// 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 0000000000..bd98bd598f
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/crc32_noasm.go
@@ -0,0 +1,35 @@
+//+build !amd64 noasm appengine
+
+// 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 0000000000..76e9682f7e
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/deflate.go
@@ -0,0 +1,1351 @@
+// 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
+			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 0000000000..71c75a065e
--- /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/gen.go b/vendor/github.com/klauspost/compress/flate/gen.go
new file mode 100644
index 0000000000..154c89a488
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/gen.go
@@ -0,0 +1,265 @@
+// 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.
+
+// +build ignore
+
+// This program generates fixedhuff.go
+// Invoke as
+//
+//	go run gen.go -output fixedhuff.go
+
+package main
+
+import (
+	"bytes"
+	"flag"
+	"fmt"
+	"go/format"
+	"io/ioutil"
+	"log"
+)
+
+var filename = flag.String("output", "fixedhuff.go", "output file name")
+
+const maxCodeLen = 16
+
+// Note: the definition of the huffmanDecoder struct is copied from
+// inflate.go, as it is private to the implementation.
+
+// 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(bits []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.min != 0 {
+		*h = huffmanDecoder{}
+	}
+
+	// Count number of codes of each length,
+	// compute min and max length.
+	var count [maxCodeLen]int
+	var min, max int
+	for _, n := range bits {
+		if n == 0 {
+			continue
+		}
+		if min == 0 || n < min {
+			min = n
+		}
+		if n > max {
+			max = n
+		}
+		count[n]++
+	}
+
+	// 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] = code
+		code += count[i]
+	}
+
+	// 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
+	if max > huffmanChunkBits {
+		numLinks := 1 << (uint(max) - huffmanChunkBits)
+		h.linkMask = uint32(numLinks - 1)
+
+		// create link tables
+		link := nextcode[huffmanChunkBits+1] >> 1
+		h.links = make([][]uint32, huffmanNumChunks-link)
+		for j := uint(link); j < huffmanNumChunks; j++ {
+			reverse := int(reverseByte[j>>8]) | int(reverseByte[j&0xff])<<8
+			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))
+			h.links[off] = make([]uint32, numLinks)
+		}
+	}
+
+	for i, n := range bits {
+		if n == 0 {
+			continue
+		}
+		code := nextcode[n]
+		nextcode[n]++
+		chunk := uint32(i<<huffmanValueShift | n)
+		reverse := int(reverseByte[code>>8]) | int(reverseByte[code&0xff])<<8
+		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
+}
+
+func main() {
+	flag.Parse()
+
+	var h huffmanDecoder
+	var bits [288]int
+	initReverseByte()
+	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
+	}
+	h.init(bits[:])
+	if h.links != nil {
+		log.Fatal("Unexpected links table in fixed Huffman decoder")
+	}
+
+	var buf bytes.Buffer
+
+	fmt.Fprintf(&buf, `// Copyright 2013 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.`+"\n\n")
+
+	fmt.Fprintln(&buf, "package flate")
+	fmt.Fprintln(&buf)
+	fmt.Fprintln(&buf, "// autogenerated by go run gen.go -output fixedhuff.go, DO NOT EDIT")
+	fmt.Fprintln(&buf)
+	fmt.Fprintln(&buf, "var fixedHuffmanDecoder = huffmanDecoder{")
+	fmt.Fprintf(&buf, "\t%d,\n", h.min)
+	fmt.Fprintln(&buf, "\t[huffmanNumChunks]uint32{")
+	for i := 0; i < huffmanNumChunks; i++ {
+		if i&7 == 0 {
+			fmt.Fprintf(&buf, "\t\t")
+		} else {
+			fmt.Fprintf(&buf, " ")
+		}
+		fmt.Fprintf(&buf, "0x%04x,", h.chunks[i])
+		if i&7 == 7 {
+			fmt.Fprintln(&buf)
+		}
+	}
+	fmt.Fprintln(&buf, "\t},")
+	fmt.Fprintln(&buf, "\tnil, 0,")
+	fmt.Fprintln(&buf, "}")
+
+	data, err := format.Source(buf.Bytes())
+	if err != nil {
+		log.Fatal(err)
+	}
+	err = ioutil.WriteFile(*filename, data, 0644)
+	if err != nil {
+		log.Fatal(err)
+	}
+}
+
+var reverseByte [256]byte
+
+func initReverseByte() {
+	for x := 0; x < 256; x++ {
+		var result byte
+		for i := uint(0); i < 8; i++ {
+			result |= byte(((x >> i) & 1) << (7 - i))
+		}
+		reverseByte[x] = result
+	}
+}
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 0000000000..f9b2a699a3
--- /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 0000000000..bdcbd823b0
--- /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 0000000000..53b63d9a0b
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/inflate.go
@@ -0,0 +1,846 @@
+// 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"
+	"strconv"
+	"sync"
+)
+
+const (
+	maxCodeLen = 16 // 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(bits []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.min != 0 {
+		*h = huffmanDecoder{}
+	}
+
+	// Count number of codes of each length,
+	// compute min and max length.
+	var count [maxCodeLen]int
+	var min, max int
+	for _, n := range bits {
+		if n == 0 {
+			continue
+		}
+		if min == 0 || n < min {
+			min = n
+		}
+		if n > max {
+			max = n
+		}
+		count[n]++
+	}
+
+	// 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] = code
+		code += count[i]
+	}
+
+	// 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
+	if max > huffmanChunkBits {
+		numLinks := 1 << (uint(max) - huffmanChunkBits)
+		h.linkMask = uint32(numLinks - 1)
+
+		// create link tables
+		link := nextcode[huffmanChunkBits+1] >> 1
+		h.links = make([][]uint32, huffmanNumChunks-link)
+		for j := uint(link); j < huffmanNumChunks; j++ {
+			reverse := int(reverseByte[j>>8]) | int(reverseByte[j&0xff])<<8
+			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))
+			h.links[off] = make([]uint32, numLinks)
+		}
+	}
+
+	for i, n := range bits {
+		if n == 0 {
+			continue
+		}
+		code := nextcode[n]
+		nextcode[n]++
+		chunk := uint32(i<<huffmanValueShift | n)
+		reverse := int(reverseByte[code>>8]) | int(reverseByte[code&0xff])<<8
+		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(reverseByte[(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 {
+		if err == io.EOF {
+			err = io.ErrUnexpectedEOF
+		}
+		f.err = 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 {
+		if err == io.EOF {
+			err = io.ErrUnexpectedEOF
+		}
+		f.err = 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
+}
+
+func (f *decompressor) moreBits() error {
+	c, err := f.r.ReadByte()
+	if err != nil {
+		if err == io.EOF {
+			err = io.ErrUnexpectedEOF
+		}
+		return 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)
+	for {
+		for f.nb < n {
+			if err := f.moreBits(); err != nil {
+				return 0, err
+			}
+		}
+		chunk := h.chunks[f.b&(huffmanNumChunks-1)]
+		n = uint(chunk & huffmanCountMask)
+		if n > huffmanChunkBits {
+			chunk = h.links[chunk>>huffmanValueShift][(f.b>>huffmanChunkBits)&h.linkMask]
+			n = uint(chunk & huffmanCountMask)
+		}
+		if n <= f.nb {
+			if n == 0 {
+				f.err = CorruptInputError(f.roffset)
+				return 0, f.err
+			}
+			f.b >>= n
+			f.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,
+		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 0000000000..c1a02720d1
--- /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 0000000000..0bbd946c01
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/flate/snappy.go
@@ -0,0 +1,856 @@
+// 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))
+			// TODO: < should be <=, and add a test for that.
+			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)
+			// TODO: >= should be >, and add a test for that.
+			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            = 16 - 1
+		minNonLiteralBlockSize = 1 + 1 + inputMargin
+	)
+
+	// Ensure that e.cur doesn't wrap, mainly an issue on 32 bits.
+	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 {
+				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            = 16 - 1
+		minNonLiteralBlockSize = 1 + 1 + inputMargin
+	)
+
+	// Ensure that e.cur doesn't wrap, mainly an issue on 32 bits.
+	if e.cur > 1<<30 {
+		for i := range e.table {
+			e.table[i] = tableEntryPrev{}
+		}
+		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
+			}
+			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 {
+				goto emitRemainder
+			}
+
+			// We could immediately start working at s now, but to improve
+			// compression we first update the hash table at s-2, 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-2)
+			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            = 16 - 1
+		minNonLiteralBlockSize = 1 + 1 + inputMargin
+		matchLenGood           = 12
+	)
+
+	// Ensure that e.cur doesn't wrap, mainly an issue on 32 bits.
+	if e.cur > 1<<30 {
+		for i := range e.table {
+			e.table[i] = tableEntryPrev{}
+		}
+		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
+		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 {
+				goto emitRemainder
+			}
+
+			// We could immediately start working at s now, but to improve
+			// compression we first update the hash table at s-2, 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-2)
+			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)
+		}
+	}
+	n := int32(len(b))
+	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 + 1
+}
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 0000000000..4f275ea61d
--- /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/compress/gzip/gunzip.go b/vendor/github.com/klauspost/compress/gzip/gunzip.go
new file mode 100644
index 0000000000..e73fab3f0f
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/gzip/gunzip.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 gzip implements reading and writing of gzip format compressed files,
+// as specified in RFC 1952.
+package gzip
+
+import (
+	"bufio"
+	"encoding/binary"
+	"errors"
+	"io"
+	"time"
+
+	"github.com/klauspost/compress/flate"
+	"github.com/klauspost/crc32"
+)
+
+const (
+	gzipID1     = 0x1f
+	gzipID2     = 0x8b
+	gzipDeflate = 8
+	flagText    = 1 << 0
+	flagHdrCrc  = 1 << 1
+	flagExtra   = 1 << 2
+	flagName    = 1 << 3
+	flagComment = 1 << 4
+)
+
+var (
+	// ErrChecksum is returned when reading GZIP data that has an invalid checksum.
+	ErrChecksum = errors.New("gzip: invalid checksum")
+	// ErrHeader is returned when reading GZIP data that has an invalid header.
+	ErrHeader = errors.New("gzip: invalid header")
+)
+
+var le = binary.LittleEndian
+
+// noEOF converts io.EOF to io.ErrUnexpectedEOF.
+func noEOF(err error) error {
+	if err == io.EOF {
+		return io.ErrUnexpectedEOF
+	}
+	return err
+}
+
+// The gzip file stores a header giving metadata about the compressed file.
+// That header is exposed as the fields of the Writer and Reader structs.
+//
+// Strings must be UTF-8 encoded and may only contain Unicode code points
+// U+0001 through U+00FF, due to limitations of the GZIP file format.
+type Header struct {
+	Comment string    // comment
+	Extra   []byte    // "extra data"
+	ModTime time.Time // modification time
+	Name    string    // file name
+	OS      byte      // operating system type
+}
+
+// A Reader is an io.Reader that can be read to retrieve
+// uncompressed data from a gzip-format compressed file.
+//
+// In general, a gzip file can be a concatenation of gzip files,
+// each with its own header. Reads from the Reader
+// return the concatenation of the uncompressed data of each.
+// Only the first header is recorded in the Reader fields.
+//
+// Gzip files store a length and checksum of the uncompressed data.
+// The Reader will return a ErrChecksum when Read
+// reaches the end of the uncompressed data if it does not
+// have the expected length or checksum. Clients should treat data
+// returned by Read as tentative until they receive the io.EOF
+// marking the end of the data.
+type Reader struct {
+	Header       // valid after NewReader or Reader.Reset
+	r            flate.Reader
+	decompressor io.ReadCloser
+	digest       uint32 // CRC-32, IEEE polynomial (section 8)
+	size         uint32 // Uncompressed size (section 2.3.1)
+	buf          [512]byte
+	err          error
+	multistream  bool
+}
+
+// NewReader creates a new Reader reading the given reader.
+// If r does not also implement io.ByteReader,
+// the decompressor may read more data than necessary from r.
+//
+// It is the caller's responsibility to call Close on the Reader when done.
+//
+// The Reader.Header fields will be valid in the Reader returned.
+func NewReader(r io.Reader) (*Reader, error) {
+	z := new(Reader)
+	if err := z.Reset(r); err != nil {
+		return nil, err
+	}
+	return z, nil
+}
+
+// Reset discards the Reader z's state and makes it equivalent to the
+// result of its original state from NewReader, but reading from r instead.
+// This permits reusing a Reader rather than allocating a new one.
+func (z *Reader) Reset(r io.Reader) error {
+	*z = Reader{
+		decompressor: z.decompressor,
+		multistream:  true,
+	}
+	if rr, ok := r.(flate.Reader); ok {
+		z.r = rr
+	} else {
+		z.r = bufio.NewReader(r)
+	}
+	z.Header, z.err = z.readHeader()
+	return z.err
+}
+
+// Multistream controls whether the reader supports multistream files.
+//
+// If enabled (the default), the Reader expects the input to be a sequence
+// of individually gzipped data streams, each with its own header and
+// trailer, ending at EOF. The effect is that the concatenation of a sequence
+// of gzipped files is treated as equivalent to the gzip of the concatenation
+// of the sequence. This is standard behavior for gzip readers.
+//
+// Calling Multistream(false) disables this behavior; disabling the behavior
+// can be useful when reading file formats that distinguish individual gzip
+// data streams or mix gzip data streams with other data streams.
+// In this mode, when the Reader reaches the end of the data stream,
+// Read returns io.EOF. If the underlying reader implements io.ByteReader,
+// it will be left positioned just after the gzip stream.
+// To start the next stream, call z.Reset(r) followed by z.Multistream(false).
+// If there is no next stream, z.Reset(r) will return io.EOF.
+func (z *Reader) Multistream(ok bool) {
+	z.multistream = ok
+}
+
+// readString reads a NUL-terminated string from z.r.
+// It treats the bytes read as being encoded as ISO 8859-1 (Latin-1) and
+// will output a string encoded using UTF-8.
+// This method always updates z.digest with the data read.
+func (z *Reader) readString() (string, error) {
+	var err error
+	needConv := false
+	for i := 0; ; i++ {
+		if i >= len(z.buf) {
+			return "", ErrHeader
+		}
+		z.buf[i], err = z.r.ReadByte()
+		if err != nil {
+			return "", err
+		}
+		if z.buf[i] > 0x7f {
+			needConv = true
+		}
+		if z.buf[i] == 0 {
+			// Digest covers the NUL terminator.
+			z.digest = crc32.Update(z.digest, crc32.IEEETable, z.buf[:i+1])
+
+			// Strings are ISO 8859-1, Latin-1 (RFC 1952, section 2.3.1).
+			if needConv {
+				s := make([]rune, 0, i)
+				for _, v := range z.buf[:i] {
+					s = append(s, rune(v))
+				}
+				return string(s), nil
+			}
+			return string(z.buf[:i]), nil
+		}
+	}
+}
+
+// readHeader reads the GZIP header according to section 2.3.1.
+// This method does not set z.err.
+func (z *Reader) readHeader() (hdr Header, err error) {
+	if _, err = io.ReadFull(z.r, z.buf[:10]); err != nil {
+		// RFC 1952, section 2.2, says the following:
+		//	A gzip file consists of a series of "members" (compressed data sets).
+		//
+		// Other than this, the specification does not clarify whether a
+		// "series" is defined as "one or more" or "zero or more". To err on the
+		// side of caution, Go interprets this to mean "zero or more".
+		// Thus, it is okay to return io.EOF here.
+		return hdr, err
+	}
+	if z.buf[0] != gzipID1 || z.buf[1] != gzipID2 || z.buf[2] != gzipDeflate {
+		return hdr, ErrHeader
+	}
+	flg := z.buf[3]
+	hdr.ModTime = time.Unix(int64(le.Uint32(z.buf[4:8])), 0)
+	// z.buf[8] is XFL and is currently ignored.
+	hdr.OS = z.buf[9]
+	z.digest = crc32.ChecksumIEEE(z.buf[:10])
+
+	if flg&flagExtra != 0 {
+		if _, err = io.ReadFull(z.r, z.buf[:2]); err != nil {
+			return hdr, noEOF(err)
+		}
+		z.digest = crc32.Update(z.digest, crc32.IEEETable, z.buf[:2])
+		data := make([]byte, le.Uint16(z.buf[:2]))
+		if _, err = io.ReadFull(z.r, data); err != nil {
+			return hdr, noEOF(err)
+		}
+		z.digest = crc32.Update(z.digest, crc32.IEEETable, data)
+		hdr.Extra = data
+	}
+
+	var s string
+	if flg&flagName != 0 {
+		if s, err = z.readString(); err != nil {
+			return hdr, err
+		}
+		hdr.Name = s
+	}
+
+	if flg&flagComment != 0 {
+		if s, err = z.readString(); err != nil {
+			return hdr, err
+		}
+		hdr.Comment = s
+	}
+
+	if flg&flagHdrCrc != 0 {
+		if _, err = io.ReadFull(z.r, z.buf[:2]); err != nil {
+			return hdr, noEOF(err)
+		}
+		digest := le.Uint16(z.buf[:2])
+		if digest != uint16(z.digest) {
+			return hdr, ErrHeader
+		}
+	}
+
+	z.digest = 0
+	if z.decompressor == nil {
+		z.decompressor = flate.NewReader(z.r)
+	} else {
+		z.decompressor.(flate.Resetter).Reset(z.r, nil)
+	}
+	return hdr, nil
+}
+
+// Read implements io.Reader, reading uncompressed bytes from its underlying Reader.
+func (z *Reader) Read(p []byte) (n int, err error) {
+	if z.err != nil {
+		return 0, z.err
+	}
+
+	n, z.err = z.decompressor.Read(p)
+	z.digest = crc32.Update(z.digest, crc32.IEEETable, p[:n])
+	z.size += uint32(n)
+	if z.err != io.EOF {
+		// In the normal case we return here.
+		return n, z.err
+	}
+
+	// Finished file; check checksum and size.
+	if _, err := io.ReadFull(z.r, z.buf[:8]); err != nil {
+		z.err = noEOF(err)
+		return n, z.err
+	}
+	digest := le.Uint32(z.buf[:4])
+	size := le.Uint32(z.buf[4:8])
+	if digest != z.digest || size != z.size {
+		z.err = ErrChecksum
+		return n, z.err
+	}
+	z.digest, z.size = 0, 0
+
+	// File is ok; check if there is another.
+	if !z.multistream {
+		return n, io.EOF
+	}
+	z.err = nil // Remove io.EOF
+
+	if _, z.err = z.readHeader(); z.err != nil {
+		return n, z.err
+	}
+
+	// Read from next file, if necessary.
+	if n > 0 {
+		return n, nil
+	}
+	return z.Read(p)
+}
+
+// Support the io.WriteTo interface for io.Copy and friends.
+func (z *Reader) WriteTo(w io.Writer) (int64, error) {
+	total := int64(0)
+	crcWriter := crc32.NewIEEE()
+	for {
+		if z.err != nil {
+			if z.err == io.EOF {
+				return total, nil
+			}
+			return total, z.err
+		}
+
+		// We write both to output and digest.
+		mw := io.MultiWriter(w, crcWriter)
+		n, err := z.decompressor.(io.WriterTo).WriteTo(mw)
+		total += n
+		z.size += uint32(n)
+		if err != nil {
+			z.err = err
+			return total, z.err
+		}
+
+		// Finished file; check checksum + size.
+		if _, err := io.ReadFull(z.r, z.buf[0:8]); err != nil {
+			if err == io.EOF {
+				err = io.ErrUnexpectedEOF
+			}
+			z.err = err
+			return total, err
+		}
+		z.digest = crcWriter.Sum32()
+		digest := le.Uint32(z.buf[:4])
+		size := le.Uint32(z.buf[4:8])
+		if digest != z.digest || size != z.size {
+			z.err = ErrChecksum
+			return total, z.err
+		}
+		z.digest, z.size = 0, 0
+
+		// File is ok; check if there is another.
+		if !z.multistream {
+			return total, nil
+		}
+		crcWriter.Reset()
+		z.err = nil // Remove io.EOF
+
+		if _, z.err = z.readHeader(); z.err != nil {
+			if z.err == io.EOF {
+				return total, nil
+			}
+			return total, z.err
+		}
+	}
+}
+
+// Close closes the Reader. It does not close the underlying io.Reader.
+// In order for the GZIP checksum to be verified, the reader must be
+// fully consumed until the io.EOF.
+func (z *Reader) Close() error { return z.decompressor.Close() }
diff --git a/vendor/github.com/klauspost/compress/gzip/gzip.go b/vendor/github.com/klauspost/compress/gzip/gzip.go
new file mode 100644
index 0000000000..a0f3ed0fcf
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/gzip/gzip.go
@@ -0,0 +1,251 @@
+// Copyright 2010 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package gzip
+
+import (
+	"errors"
+	"fmt"
+	"io"
+
+	"github.com/klauspost/compress/flate"
+	"github.com/klauspost/crc32"
+)
+
+// 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      // written at first call to Write, Flush, or Close
+	w           io.Writer
+	level       int
+	wroteHeader bool
+	compressor  *flate.Writer
+	digest      uint32 // CRC-32, IEEE polynomial (section 8)
+	size        uint32 // Uncompressed size (section 2.3.1)
+	closed      bool
+	buf         [10]byte
+	err         error
+}
+
+// NewWriter returns a new Writer.
+// Writes to the returned writer are compressed and written to w.
+//
+// It is the caller's responsibility to call Close on the WriteCloser when done.
+// Writes may be buffered and not flushed until Close.
+//
+// Callers that wish to set the fields in Writer.Header must do so before
+// the first call to Write, Flush, or Close.
+func NewWriter(w io.Writer) *Writer {
+	z, _ := NewWriterLevel(w, DefaultCompression)
+	return z
+}
+
+// NewWriterLevel is like NewWriter but specifies the compression level instead
+// of assuming DefaultCompression.
+//
+// The compression level can be DefaultCompression, NoCompression, or any
+// integer value between BestSpeed and BestCompression inclusive. The error
+// returned will be nil if the level is valid.
+func NewWriterLevel(w io.Writer, level int) (*Writer, error) {
+	if level < HuffmanOnly || level > BestCompression {
+		return nil, fmt.Errorf("gzip: invalid compression level: %d", level)
+	}
+	z := new(Writer)
+	z.init(w, level)
+	return z, nil
+}
+
+func (z *Writer) init(w io.Writer, level int) {
+	compressor := z.compressor
+	if compressor != nil {
+		compressor.Reset(w)
+	}
+	*z = Writer{
+		Header: Header{
+			OS: 255, // unknown
+		},
+		w:          w,
+		level:      level,
+		compressor: compressor,
+	}
+}
+
+// Reset discards the Writer z's state and makes it equivalent to the
+// result of its original state from NewWriter or NewWriterLevel, but
+// writing to w instead. This permits reusing a Writer rather than
+// allocating a new one.
+func (z *Writer) Reset(w io.Writer) {
+	z.init(w, z.level)
+}
+
+// writeBytes writes a length-prefixed byte slice to z.w.
+func (z *Writer) writeBytes(b []byte) error {
+	if len(b) > 0xffff {
+		return errors.New("gzip.Write: Extra data is too large")
+	}
+	le.PutUint16(z.buf[:2], uint16(len(b)))
+	_, err := z.w.Write(z.buf[:2])
+	if err != nil {
+		return err
+	}
+	_, err = z.w.Write(b)
+	return err
+}
+
+// writeString writes a UTF-8 string s in GZIP's format to z.w.
+// GZIP (RFC 1952) specifies that strings are NUL-terminated ISO 8859-1 (Latin-1).
+func (z *Writer) writeString(s string) (err error) {
+	// GZIP stores Latin-1 strings; error if non-Latin-1; convert if non-ASCII.
+	needconv := false
+	for _, v := range s {
+		if v == 0 || v > 0xff {
+			return errors.New("gzip.Write: non-Latin-1 header string")
+		}
+		if v > 0x7f {
+			needconv = true
+		}
+	}
+	if needconv {
+		b := make([]byte, 0, len(s))
+		for _, v := range s {
+			b = append(b, byte(v))
+		}
+		_, err = z.w.Write(b)
+	} else {
+		_, err = io.WriteString(z.w, s)
+	}
+	if err != nil {
+		return err
+	}
+	// GZIP strings are NUL-terminated.
+	z.buf[0] = 0
+	_, err = z.w.Write(z.buf[:1])
+	return err
+}
+
+// Write writes a compressed form of p to the underlying io.Writer. The
+// compressed bytes are not necessarily flushed until the Writer is closed.
+func (z *Writer) Write(p []byte) (int, error) {
+	if z.err != nil {
+		return 0, z.err
+	}
+	var n int
+	// Write the GZIP header lazily.
+	if !z.wroteHeader {
+		z.wroteHeader = true
+		z.buf[0] = gzipID1
+		z.buf[1] = gzipID2
+		z.buf[2] = gzipDeflate
+		z.buf[3] = 0
+		if z.Extra != nil {
+			z.buf[3] |= 0x04
+		}
+		if z.Name != "" {
+			z.buf[3] |= 0x08
+		}
+		if z.Comment != "" {
+			z.buf[3] |= 0x10
+		}
+		le.PutUint32(z.buf[4:8], uint32(z.ModTime.Unix()))
+		if z.level == BestCompression {
+			z.buf[8] = 2
+		} else if z.level == BestSpeed {
+			z.buf[8] = 4
+		} else {
+			z.buf[8] = 0
+		}
+		z.buf[9] = z.OS
+		n, z.err = z.w.Write(z.buf[:10])
+		if z.err != nil {
+			return n, z.err
+		}
+		if z.Extra != nil {
+			z.err = z.writeBytes(z.Extra)
+			if z.err != nil {
+				return n, z.err
+			}
+		}
+		if z.Name != "" {
+			z.err = z.writeString(z.Name)
+			if z.err != nil {
+				return n, z.err
+			}
+		}
+		if z.Comment != "" {
+			z.err = z.writeString(z.Comment)
+			if z.err != nil {
+				return n, z.err
+			}
+		}
+		if z.compressor == nil {
+			z.compressor, _ = flate.NewWriter(z.w, z.level)
+		}
+	}
+	z.size += uint32(len(p))
+	z.digest = crc32.Update(z.digest, crc32.IEEETable, p)
+	n, z.err = z.compressor.Write(p)
+	return n, z.err
+}
+
+// Flush flushes any pending compressed data to the underlying writer.
+//
+// It is useful mainly in compressed network protocols, to ensure that
+// a remote reader has enough data to reconstruct a packet. Flush does
+// not return until the data has been written. If the underlying
+// writer returns an error, Flush returns that error.
+//
+// In the terminology of the zlib library, Flush is equivalent to Z_SYNC_FLUSH.
+func (z *Writer) Flush() error {
+	if z.err != nil {
+		return z.err
+	}
+	if z.closed {
+		return nil
+	}
+	if !z.wroteHeader {
+		z.Write(nil)
+		if z.err != nil {
+			return z.err
+		}
+	}
+	z.err = z.compressor.Flush()
+	return z.err
+}
+
+// Close closes the Writer, flushing any unwritten data to the underlying
+// io.Writer, but does not close the underlying io.Writer.
+func (z *Writer) Close() error {
+	if z.err != nil {
+		return z.err
+	}
+	if z.closed {
+		return nil
+	}
+	z.closed = true
+	if !z.wroteHeader {
+		z.Write(nil)
+		if z.err != nil {
+			return z.err
+		}
+	}
+	z.err = z.compressor.Close()
+	if z.err != nil {
+		return z.err
+	}
+	le.PutUint32(z.buf[:4], z.digest)
+	le.PutUint32(z.buf[4:8], z.size)
+	_, z.err = z.w.Write(z.buf[:8])
+	return z.err
+}
diff --git a/vendor/github.com/klauspost/cpuid/LICENSE b/vendor/github.com/klauspost/cpuid/LICENSE
new file mode 100644
index 0000000000..5cec7ee949
--- /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 0000000000..b2b6bee879
--- /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 0000000000..9230ca5628
--- /dev/null
+++ b/vendor/github.com/klauspost/cpuid/cpuid.go
@@ -0,0 +1,1022 @@
+// 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
+
+	// 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
+
+	// 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 = sgx(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
+}
+
+func (c CPUInfo) RDTSCP() bool {
+	return c.Features&RDTSCP != 0
+}
+
+func (c CPUInfo) CX16() bool {
+	return c.Features&CX16 != 0
+}
+
+// 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 sgx(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, _ := 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
+		}
+
+		// 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 0000000000..4d731711e4
--- /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 0000000000..3c1d60e422
--- /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 0000000000..a5f04dd6d0
--- /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 0000000000..909c5d9a7a
--- /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 0000000000..c060b8165e
--- /dev/null
+++ b/vendor/github.com/klauspost/cpuid/generate.go
@@ -0,0 +1,3 @@
+package cpuid
+
+//go:generate go run private-gen.go
diff --git a/vendor/github.com/klauspost/cpuid/private-gen.go b/vendor/github.com/klauspost/cpuid/private-gen.go
new file mode 100644
index 0000000000..437333d292
--- /dev/null
+++ b/vendor/github.com/klauspost/cpuid/private-gen.go
@@ -0,0 +1,476 @@
+// +build ignore
+
+package main
+
+import (
+	"bytes"
+	"fmt"
+	"go/ast"
+	"go/parser"
+	"go/printer"
+	"go/token"
+	"io"
+	"io/ioutil"
+	"log"
+	"os"
+	"reflect"
+	"strings"
+	"unicode"
+	"unicode/utf8"
+)
+
+var inFiles = []string{"cpuid.go", "cpuid_test.go"}
+var copyFiles = []string{"cpuid_amd64.s", "cpuid_386.s", "detect_ref.go", "detect_intel.go"}
+var fileSet = token.NewFileSet()
+var reWrites = []rewrite{
+	initRewrite("CPUInfo -> cpuInfo"),
+	initRewrite("Vendor -> vendor"),
+	initRewrite("Flags -> flags"),
+	initRewrite("Detect -> detect"),
+	initRewrite("CPU -> cpu"),
+}
+var excludeNames = map[string]bool{"string": true, "join": true, "trim": true,
+	// cpuid_test.go
+	"t": true, "println": true, "logf": true, "log": true, "fatalf": true, "fatal": true,
+}
+
+var excludePrefixes = []string{"test", "benchmark"}
+
+func main() {
+	Package := "private"
+	parserMode := parser.ParseComments
+	exported := make(map[string]rewrite)
+	for _, file := range inFiles {
+		in, err := os.Open(file)
+		if err != nil {
+			log.Fatalf("opening input", err)
+		}
+
+		src, err := ioutil.ReadAll(in)
+		if err != nil {
+			log.Fatalf("reading input", err)
+		}
+
+		astfile, err := parser.ParseFile(fileSet, file, src, parserMode)
+		if err != nil {
+			log.Fatalf("parsing input", err)
+		}
+
+		for _, rw := range reWrites {
+			astfile = rw(astfile)
+		}
+
+		// Inspect the AST and print all identifiers and literals.
+		var startDecl token.Pos
+		var endDecl token.Pos
+		ast.Inspect(astfile, func(n ast.Node) bool {
+			var s string
+			switch x := n.(type) {
+			case *ast.Ident:
+				if x.IsExported() {
+					t := strings.ToLower(x.Name)
+					for _, pre := range excludePrefixes {
+						if strings.HasPrefix(t, pre) {
+							return true
+						}
+					}
+					if excludeNames[t] != true {
+						//if x.Pos() > startDecl && x.Pos() < endDecl {
+						exported[x.Name] = initRewrite(x.Name + " -> " + t)
+					}
+				}
+
+			case *ast.GenDecl:
+				if x.Tok == token.CONST && x.Lparen > 0 {
+					startDecl = x.Lparen
+					endDecl = x.Rparen
+					// fmt.Printf("Decl:%s -> %s\n", fileSet.Position(startDecl), fileSet.Position(endDecl))
+				}
+			}
+			if s != "" {
+				fmt.Printf("%s:\t%s\n", fileSet.Position(n.Pos()), s)
+			}
+			return true
+		})
+
+		for _, rw := range exported {
+			astfile = rw(astfile)
+		}
+
+		var buf bytes.Buffer
+
+		printer.Fprint(&buf, fileSet, astfile)
+
+		// Remove package documentation and insert information
+		s := buf.String()
+		ind := strings.Index(buf.String(), "\npackage cpuid")
+		s = s[ind:]
+		s = "// Generated, DO NOT EDIT,\n" +
+			"// but copy it to your own project and rename the package.\n" +
+			"// See more at http://github.com/klauspost/cpuid\n" +
+			s
+
+		outputName := Package + string(os.PathSeparator) + file
+
+		err = ioutil.WriteFile(outputName, []byte(s), 0644)
+		if err != nil {
+			log.Fatalf("writing output: %s", err)
+		}
+		log.Println("Generated", outputName)
+	}
+
+	for _, file := range copyFiles {
+		dst := ""
+		if strings.HasPrefix(file, "cpuid") {
+			dst = Package + string(os.PathSeparator) + file
+		} else {
+			dst = Package + string(os.PathSeparator) + "cpuid_" + file
+		}
+		err := copyFile(file, dst)
+		if err != nil {
+			log.Fatalf("copying file: %s", err)
+		}
+		log.Println("Copied", dst)
+	}
+}
+
+// CopyFile copies a file from src to dst. If src and dst files exist, and are
+// the same, then return success. Copy the file contents from src to dst.
+func copyFile(src, dst string) (err error) {
+	sfi, err := os.Stat(src)
+	if err != nil {
+		return
+	}
+	if !sfi.Mode().IsRegular() {
+		// cannot copy non-regular files (e.g., directories,
+		// symlinks, devices, etc.)
+		return fmt.Errorf("CopyFile: non-regular source file %s (%q)", sfi.Name(), sfi.Mode().String())
+	}
+	dfi, err := os.Stat(dst)
+	if err != nil {
+		if !os.IsNotExist(err) {
+			return
+		}
+	} else {
+		if !(dfi.Mode().IsRegular()) {
+			return fmt.Errorf("CopyFile: non-regular destination file %s (%q)", dfi.Name(), dfi.Mode().String())
+		}
+		if os.SameFile(sfi, dfi) {
+			return
+		}
+	}
+	err = copyFileContents(src, dst)
+	return
+}
+
+// copyFileContents copies the contents of the file named src to the file named
+// by dst. The file will be created if it does not already exist. If the
+// destination file exists, all it's contents will be replaced by the contents
+// of the source file.
+func copyFileContents(src, dst string) (err error) {
+	in, err := os.Open(src)
+	if err != nil {
+		return
+	}
+	defer in.Close()
+	out, err := os.Create(dst)
+	if err != nil {
+		return
+	}
+	defer func() {
+		cerr := out.Close()
+		if err == nil {
+			err = cerr
+		}
+	}()
+	if _, err = io.Copy(out, in); err != nil {
+		return
+	}
+	err = out.Sync()
+	return
+}
+
+type rewrite func(*ast.File) *ast.File
+
+// Mostly copied from gofmt
+func initRewrite(rewriteRule string) rewrite {
+	f := strings.Split(rewriteRule, "->")
+	if len(f) != 2 {
+		fmt.Fprintf(os.Stderr, "rewrite rule must be of the form 'pattern -> replacement'\n")
+		os.Exit(2)
+	}
+	pattern := parseExpr(f[0], "pattern")
+	replace := parseExpr(f[1], "replacement")
+	return func(p *ast.File) *ast.File { return rewriteFile(pattern, replace, p) }
+}
+
+// parseExpr parses s as an expression.
+// It might make sense to expand this to allow statement patterns,
+// but there are problems with preserving formatting and also
+// with what a wildcard for a statement looks like.
+func parseExpr(s, what string) ast.Expr {
+	x, err := parser.ParseExpr(s)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "parsing %s %s at %s\n", what, s, err)
+		os.Exit(2)
+	}
+	return x
+}
+
+// Keep this function for debugging.
+/*
+func dump(msg string, val reflect.Value) {
+	fmt.Printf("%s:\n", msg)
+	ast.Print(fileSet, val.Interface())
+	fmt.Println()
+}
+*/
+
+// rewriteFile applies the rewrite rule 'pattern -> replace' to an entire file.
+func rewriteFile(pattern, replace ast.Expr, p *ast.File) *ast.File {
+	cmap := ast.NewCommentMap(fileSet, p, p.Comments)
+	m := make(map[string]reflect.Value)
+	pat := reflect.ValueOf(pattern)
+	repl := reflect.ValueOf(replace)
+
+	var rewriteVal func(val reflect.Value) reflect.Value
+	rewriteVal = func(val reflect.Value) reflect.Value {
+		// don't bother if val is invalid to start with
+		if !val.IsValid() {
+			return reflect.Value{}
+		}
+		for k := range m {
+			delete(m, k)
+		}
+		val = apply(rewriteVal, val)
+		if match(m, pat, val) {
+			val = subst(m, repl, reflect.ValueOf(val.Interface().(ast.Node).Pos()))
+		}
+		return val
+	}
+
+	r := apply(rewriteVal, reflect.ValueOf(p)).Interface().(*ast.File)
+	r.Comments = cmap.Filter(r).Comments() // recreate comments list
+	return r
+}
+
+// set is a wrapper for x.Set(y); it protects the caller from panics if x cannot be changed to y.
+func set(x, y reflect.Value) {
+	// don't bother if x cannot be set or y is invalid
+	if !x.CanSet() || !y.IsValid() {
+		return
+	}
+	defer func() {
+		if x := recover(); x != nil {
+			if s, ok := x.(string); ok &&
+				(strings.Contains(s, "type mismatch") || strings.Contains(s, "not assignable")) {
+				// x cannot be set to y - ignore this rewrite
+				return
+			}
+			panic(x)
+		}
+	}()
+	x.Set(y)
+}
+
+// Values/types for special cases.
+var (
+	objectPtrNil = reflect.ValueOf((*ast.Object)(nil))
+	scopePtrNil  = reflect.ValueOf((*ast.Scope)(nil))
+
+	identType     = reflect.TypeOf((*ast.Ident)(nil))
+	objectPtrType = reflect.TypeOf((*ast.Object)(nil))
+	positionType  = reflect.TypeOf(token.NoPos)
+	callExprType  = reflect.TypeOf((*ast.CallExpr)(nil))
+	scopePtrType  = reflect.TypeOf((*ast.Scope)(nil))
+)
+
+// apply replaces each AST field x in val with f(x), returning val.
+// To avoid extra conversions, f operates on the reflect.Value form.
+func apply(f func(reflect.Value) reflect.Value, val reflect.Value) reflect.Value {
+	if !val.IsValid() {
+		return reflect.Value{}
+	}
+
+	// *ast.Objects introduce cycles and are likely incorrect after
+	// rewrite; don't follow them but replace with nil instead
+	if val.Type() == objectPtrType {
+		return objectPtrNil
+	}
+
+	// similarly for scopes: they are likely incorrect after a rewrite;
+	// replace them with nil
+	if val.Type() == scopePtrType {
+		return scopePtrNil
+	}
+
+	switch v := reflect.Indirect(val); v.Kind() {
+	case reflect.Slice:
+		for i := 0; i < v.Len(); i++ {
+			e := v.Index(i)
+			set(e, f(e))
+		}
+	case reflect.Struct:
+		for i := 0; i < v.NumField(); i++ {
+			e := v.Field(i)
+			set(e, f(e))
+		}
+	case reflect.Interface:
+		e := v.Elem()
+		set(v, f(e))
+	}
+	return val
+}
+
+func isWildcard(s string) bool {
+	rune, size := utf8.DecodeRuneInString(s)
+	return size == len(s) && unicode.IsLower(rune)
+}
+
+// match returns true if pattern matches val,
+// recording wildcard submatches in m.
+// If m == nil, match checks whether pattern == val.
+func match(m map[string]reflect.Value, pattern, val reflect.Value) bool {
+	// Wildcard matches any expression.  If it appears multiple
+	// times in the pattern, it must match the same expression
+	// each time.
+	if m != nil && pattern.IsValid() && pattern.Type() == identType {
+		name := pattern.Interface().(*ast.Ident).Name
+		if isWildcard(name) && val.IsValid() {
+			// wildcards only match valid (non-nil) expressions.
+			if _, ok := val.Interface().(ast.Expr); ok && !val.IsNil() {
+				if old, ok := m[name]; ok {
+					return match(nil, old, val)
+				}
+				m[name] = val
+				return true
+			}
+		}
+	}
+
+	// Otherwise, pattern and val must match recursively.
+	if !pattern.IsValid() || !val.IsValid() {
+		return !pattern.IsValid() && !val.IsValid()
+	}
+	if pattern.Type() != val.Type() {
+		return false
+	}
+
+	// Special cases.
+	switch pattern.Type() {
+	case identType:
+		// For identifiers, only the names need to match
+		// (and none of the other *ast.Object information).
+		// This is a common case, handle it all here instead
+		// of recursing down any further via reflection.
+		p := pattern.Interface().(*ast.Ident)
+		v := val.Interface().(*ast.Ident)
+		return p == nil && v == nil || p != nil && v != nil && p.Name == v.Name
+	case objectPtrType, positionType:
+		// object pointers and token positions always match
+		return true
+	case callExprType:
+		// For calls, the Ellipsis fields (token.Position) must
+		// match since that is how f(x) and f(x...) are different.
+		// Check them here but fall through for the remaining fields.
+		p := pattern.Interface().(*ast.CallExpr)
+		v := val.Interface().(*ast.CallExpr)
+		if p.Ellipsis.IsValid() != v.Ellipsis.IsValid() {
+			return false
+		}
+	}
+
+	p := reflect.Indirect(pattern)
+	v := reflect.Indirect(val)
+	if !p.IsValid() || !v.IsValid() {
+		return !p.IsValid() && !v.IsValid()
+	}
+
+	switch p.Kind() {
+	case reflect.Slice:
+		if p.Len() != v.Len() {
+			return false
+		}
+		for i := 0; i < p.Len(); i++ {
+			if !match(m, p.Index(i), v.Index(i)) {
+				return false
+			}
+		}
+		return true
+
+	case reflect.Struct:
+		for i := 0; i < p.NumField(); i++ {
+			if !match(m, p.Field(i), v.Field(i)) {
+				return false
+			}
+		}
+		return true
+
+	case reflect.Interface:
+		return match(m, p.Elem(), v.Elem())
+	}
+
+	// Handle token integers, etc.
+	return p.Interface() == v.Interface()
+}
+
+// subst returns a copy of pattern with values from m substituted in place
+// of wildcards and pos used as the position of tokens from the pattern.
+// if m == nil, subst returns a copy of pattern and doesn't change the line
+// number information.
+func subst(m map[string]reflect.Value, pattern reflect.Value, pos reflect.Value) reflect.Value {
+	if !pattern.IsValid() {
+		return reflect.Value{}
+	}
+
+	// Wildcard gets replaced with map value.
+	if m != nil && pattern.Type() == identType {
+		name := pattern.Interface().(*ast.Ident).Name
+		if isWildcard(name) {
+			if old, ok := m[name]; ok {
+				return subst(nil, old, reflect.Value{})
+			}
+		}
+	}
+
+	if pos.IsValid() && pattern.Type() == positionType {
+		// use new position only if old position was valid in the first place
+		if old := pattern.Interface().(token.Pos); !old.IsValid() {
+			return pattern
+		}
+		return pos
+	}
+
+	// Otherwise copy.
+	switch p := pattern; p.Kind() {
+	case reflect.Slice:
+		v := reflect.MakeSlice(p.Type(), p.Len(), p.Len())
+		for i := 0; i < p.Len(); i++ {
+			v.Index(i).Set(subst(m, p.Index(i), pos))
+		}
+		return v
+
+	case reflect.Struct:
+		v := reflect.New(p.Type()).Elem()
+		for i := 0; i < p.NumField(); i++ {
+			v.Field(i).Set(subst(m, p.Field(i), pos))
+		}
+		return v
+
+	case reflect.Ptr:
+		v := reflect.New(p.Type()).Elem()
+		if elem := p.Elem(); elem.IsValid() {
+			v.Set(subst(m, elem, pos).Addr())
+		}
+		return v
+
+	case reflect.Interface:
+		v := reflect.New(p.Type()).Elem()
+		if elem := p.Elem(); elem.IsValid() {
+			v.Set(subst(m, elem, pos))
+		}
+		return v
+	}
+
+	return pattern
+}
diff --git a/vendor/github.com/klauspost/crc32/LICENSE b/vendor/github.com/klauspost/crc32/LICENSE
new file mode 100644
index 0000000000..4fd5963e39
--- /dev/null
+++ b/vendor/github.com/klauspost/crc32/LICENSE
@@ -0,0 +1,28 @@
+Copyright (c) 2012 The Go Authors. All rights reserved.
+Copyright (c) 2015 Klaus Post
+
+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/crc32/README.md b/vendor/github.com/klauspost/crc32/README.md
new file mode 100644
index 0000000000..029625d360
--- /dev/null
+++ b/vendor/github.com/klauspost/crc32/README.md
@@ -0,0 +1,87 @@
+# crc32
+CRC32 hash with x64 optimizations
+
+This package is a drop-in replacement for the standard library `hash/crc32` package, that features SSE 4.2 optimizations on x64 platforms, for a 10x speedup.
+
+[![Build Status](https://travis-ci.org/klauspost/crc32.svg?branch=master)](https://travis-ci.org/klauspost/crc32)
+
+# usage
+
+Install using `go get github.com/klauspost/crc32`. This library is based on Go 1.5 code and requires Go 1.3 or newer.
+
+Replace `import "hash/crc32"` with `import "github.com/klauspost/crc32"` and you are good to go.
+
+# changes
+* Oct 20, 2016: Changes have been merged to upstream Go. Package updated to match.
+* Dec 4, 2015: Uses the "slice-by-8" trick more extensively, which gives a 1.5 to 2.5x speedup if assembler is unavailable.
+
+
+# performance
+
+For *Go 1.7* performance is equivalent to the standard library. So if you use this package for Go 1.7 you can switch back.
+
+
+For IEEE tables (the most common), there is approximately a factor 10 speedup with "CLMUL" (Carryless multiplication) instruction:
+```
+benchmark            old ns/op     new ns/op     delta
+BenchmarkCrc32KB     99955         10258         -89.74%
+
+benchmark            old MB/s     new MB/s     speedup
+BenchmarkCrc32KB     327.83       3194.20      9.74x
+```
+
+For other tables and "CLMUL"  capable machines the performance is the same as the standard library.
+
+Here are some detailed benchmarks, comparing to go 1.5 standard library with and without assembler enabled.
+
+```
+Std:   Standard Go 1.5 library
+Crc:   Indicates IEEE type CRC.
+40B:   Size of each slice encoded.
+NoAsm: Assembler was disabled (ie. not an AMD64 or SSE 4.2+ capable machine).
+Castagnoli: Castagnoli CRC type.
+
+BenchmarkStdCrc40B-4            10000000               158 ns/op         252.88 MB/s
+BenchmarkCrc40BNoAsm-4          20000000               105 ns/op         377.38 MB/s (slice8)
+BenchmarkCrc40B-4               20000000               105 ns/op         378.77 MB/s (slice8)
+
+BenchmarkStdCrc1KB-4              500000              3604 ns/op         284.10 MB/s
+BenchmarkCrc1KBNoAsm-4           1000000              1463 ns/op         699.79 MB/s (slice8)
+BenchmarkCrc1KB-4                3000000               396 ns/op        2583.69 MB/s (asm)
+
+BenchmarkStdCrc8KB-4              200000             11417 ns/op         717.48 MB/s (slice8)
+BenchmarkCrc8KBNoAsm-4            200000             11317 ns/op         723.85 MB/s (slice8)
+BenchmarkCrc8KB-4                 500000              2919 ns/op        2805.73 MB/s (asm)
+
+BenchmarkStdCrc32KB-4              30000             45749 ns/op         716.24 MB/s (slice8)
+BenchmarkCrc32KBNoAsm-4            30000             45109 ns/op         726.42 MB/s (slice8)
+BenchmarkCrc32KB-4                100000             11497 ns/op        2850.09 MB/s (asm)
+
+BenchmarkStdNoAsmCastagnol40B-4 10000000               161 ns/op         246.94 MB/s
+BenchmarkStdCastagnoli40B-4     50000000              28.4 ns/op        1410.69 MB/s (asm)
+BenchmarkCastagnoli40BNoAsm-4   20000000               100 ns/op         398.01 MB/s (slice8)
+BenchmarkCastagnoli40B-4        50000000              28.2 ns/op        1419.54 MB/s (asm)
+
+BenchmarkStdNoAsmCastagnoli1KB-4  500000              3622 ns/op        282.67 MB/s
+BenchmarkStdCastagnoli1KB-4     10000000               144 ns/op        7099.78 MB/s (asm)
+BenchmarkCastagnoli1KBNoAsm-4    1000000              1475 ns/op         694.14 MB/s (slice8)
+BenchmarkCastagnoli1KB-4        10000000               146 ns/op        6993.35 MB/s (asm)
+
+BenchmarkStdNoAsmCastagnoli8KB-4  50000              28781 ns/op         284.63 MB/s
+BenchmarkStdCastagnoli8KB-4      1000000              1029 ns/op        7957.89 MB/s (asm)
+BenchmarkCastagnoli8KBNoAsm-4     200000             11410 ns/op         717.94 MB/s (slice8)
+BenchmarkCastagnoli8KB-4         1000000              1000 ns/op        8188.71 MB/s (asm)
+
+BenchmarkStdNoAsmCastagnoli32KB-4  10000            115426 ns/op         283.89 MB/s
+BenchmarkStdCastagnoli32KB-4      300000              4065 ns/op        8059.13 MB/s (asm)
+BenchmarkCastagnoli32KBNoAsm-4     30000             45171 ns/op         725.41 MB/s (slice8)
+BenchmarkCastagnoli32KB-4         500000              4077 ns/op        8035.89 MB/s (asm)
+```
+
+The IEEE assembler optimizations has been submitted and will be part of the Go 1.6 standard library.
+
+However, the improved use of slice-by-8 has not, but will probably be submitted for Go 1.7.
+
+# license
+
+Standard Go license. Changes are Copyright (c) 2015 Klaus Post under same conditions.
diff --git a/vendor/github.com/klauspost/crc32/crc32.go b/vendor/github.com/klauspost/crc32/crc32.go
new file mode 100644
index 0000000000..8aa91b17e9
--- /dev/null
+++ b/vendor/github.com/klauspost/crc32/crc32.go
@@ -0,0 +1,207 @@
+// 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 crc32 implements the 32-bit cyclic redundancy check, or CRC-32,
+// checksum. See http://en.wikipedia.org/wiki/Cyclic_redundancy_check for
+// information.
+//
+// Polynomials are represented in LSB-first form also known as reversed representation.
+//
+// See http://en.wikipedia.org/wiki/Mathematics_of_cyclic_redundancy_checks#Reversed_representations_and_reciprocal_polynomials
+// for information.
+package crc32
+
+import (
+	"hash"
+	"sync"
+)
+
+// The size of a CRC-32 checksum in bytes.
+const Size = 4
+
+// Predefined polynomials.
+const (
+	// IEEE is by far and away the most common CRC-32 polynomial.
+	// Used by ethernet (IEEE 802.3), v.42, fddi, gzip, zip, png, ...
+	IEEE = 0xedb88320
+
+	// Castagnoli's polynomial, used in iSCSI.
+	// Has better error detection characteristics than IEEE.
+	// http://dx.doi.org/10.1109/26.231911
+	Castagnoli = 0x82f63b78
+
+	// Koopman's polynomial.
+	// Also has better error detection characteristics than IEEE.
+	// http://dx.doi.org/10.1109/DSN.2002.1028931
+	Koopman = 0xeb31d82e
+)
+
+// Table is a 256-word table representing the polynomial for efficient processing.
+type Table [256]uint32
+
+// This file makes use of functions implemented in architecture-specific files.
+// The interface that they implement is as follows:
+//
+//    // archAvailableIEEE reports whether an architecture-specific CRC32-IEEE
+//    // algorithm is available.
+//    archAvailableIEEE() bool
+//
+//    // archInitIEEE initializes the architecture-specific CRC3-IEEE algorithm.
+//    // It can only be called if archAvailableIEEE() returns true.
+//    archInitIEEE()
+//
+//    // archUpdateIEEE updates the given CRC32-IEEE. It can only be called if
+//    // archInitIEEE() was previously called.
+//    archUpdateIEEE(crc uint32, p []byte) uint32
+//
+//    // archAvailableCastagnoli reports whether an architecture-specific
+//    // CRC32-C algorithm is available.
+//    archAvailableCastagnoli() bool
+//
+//    // archInitCastagnoli initializes the architecture-specific CRC32-C
+//    // algorithm. It can only be called if archAvailableCastagnoli() returns
+//    // true.
+//    archInitCastagnoli()
+//
+//    // archUpdateCastagnoli updates the given CRC32-C. It can only be called
+//    // if archInitCastagnoli() was previously called.
+//    archUpdateCastagnoli(crc uint32, p []byte) uint32
+
+// castagnoliTable points to a lazily initialized Table for the Castagnoli
+// polynomial. MakeTable will always return this value when asked to make a
+// Castagnoli table so we can compare against it to find when the caller is
+// using this polynomial.
+var castagnoliTable *Table
+var castagnoliTable8 *slicing8Table
+var castagnoliArchImpl bool
+var updateCastagnoli func(crc uint32, p []byte) uint32
+var castagnoliOnce sync.Once
+
+func castagnoliInit() {
+	castagnoliTable = simpleMakeTable(Castagnoli)
+	castagnoliArchImpl = archAvailableCastagnoli()
+
+	if castagnoliArchImpl {
+		archInitCastagnoli()
+		updateCastagnoli = archUpdateCastagnoli
+	} else {
+		// Initialize the slicing-by-8 table.
+		castagnoliTable8 = slicingMakeTable(Castagnoli)
+		updateCastagnoli = func(crc uint32, p []byte) uint32 {
+			return slicingUpdate(crc, castagnoliTable8, p)
+		}
+	}
+}
+
+// IEEETable is the table for the IEEE polynomial.
+var IEEETable = simpleMakeTable(IEEE)
+
+// ieeeTable8 is the slicing8Table for IEEE
+var ieeeTable8 *slicing8Table
+var ieeeArchImpl bool
+var updateIEEE func(crc uint32, p []byte) uint32
+var ieeeOnce sync.Once
+
+func ieeeInit() {
+	ieeeArchImpl = archAvailableIEEE()
+
+	if ieeeArchImpl {
+		archInitIEEE()
+		updateIEEE = archUpdateIEEE
+	} else {
+		// Initialize the slicing-by-8 table.
+		ieeeTable8 = slicingMakeTable(IEEE)
+		updateIEEE = func(crc uint32, p []byte) uint32 {
+			return slicingUpdate(crc, ieeeTable8, p)
+		}
+	}
+}
+
+// MakeTable returns a Table constructed from the specified polynomial.
+// The contents of this Table must not be modified.
+func MakeTable(poly uint32) *Table {
+	switch poly {
+	case IEEE:
+		ieeeOnce.Do(ieeeInit)
+		return IEEETable
+	case Castagnoli:
+		castagnoliOnce.Do(castagnoliInit)
+		return castagnoliTable
+	}
+	return simpleMakeTable(poly)
+}
+
+// digest represents the partial evaluation of a checksum.
+type digest struct {
+	crc uint32
+	tab *Table
+}
+
+// New creates a new hash.Hash32 computing the CRC-32 checksum
+// using the polynomial represented by the Table.
+// Its Sum method will lay the value out in big-endian byte order.
+func New(tab *Table) hash.Hash32 {
+	if tab == IEEETable {
+		ieeeOnce.Do(ieeeInit)
+	}
+	return &digest{0, tab}
+}
+
+// NewIEEE creates a new hash.Hash32 computing the CRC-32 checksum
+// using the IEEE polynomial.
+// Its Sum method will lay the value out in big-endian byte order.
+func NewIEEE() hash.Hash32 { return New(IEEETable) }
+
+func (d *digest) Size() int { return Size }
+
+func (d *digest) BlockSize() int { return 1 }
+
+func (d *digest) Reset() { d.crc = 0 }
+
+// Update returns the result of adding the bytes in p to the crc.
+func Update(crc uint32, tab *Table, p []byte) uint32 {
+	switch tab {
+	case castagnoliTable:
+		return updateCastagnoli(crc, p)
+	case IEEETable:
+		// Unfortunately, because IEEETable is exported, IEEE may be used without a
+		// call to MakeTable. We have to make sure it gets initialized in that case.
+		ieeeOnce.Do(ieeeInit)
+		return updateIEEE(crc, p)
+	default:
+		return simpleUpdate(crc, tab, p)
+	}
+}
+
+func (d *digest) Write(p []byte) (n int, err error) {
+	switch d.tab {
+	case castagnoliTable:
+		d.crc = updateCastagnoli(d.crc, p)
+	case IEEETable:
+		// We only create digest objects through New() which takes care of
+		// initialization in this case.
+		d.crc = updateIEEE(d.crc, p)
+	default:
+		d.crc = simpleUpdate(d.crc, d.tab, p)
+	}
+	return len(p), nil
+}
+
+func (d *digest) Sum32() uint32 { return d.crc }
+
+func (d *digest) Sum(in []byte) []byte {
+	s := d.Sum32()
+	return append(in, byte(s>>24), byte(s>>16), byte(s>>8), byte(s))
+}
+
+// Checksum returns the CRC-32 checksum of data
+// using the polynomial represented by the Table.
+func Checksum(data []byte, tab *Table) uint32 { return Update(0, tab, data) }
+
+// ChecksumIEEE returns the CRC-32 checksum of data
+// using the IEEE polynomial.
+func ChecksumIEEE(data []byte) uint32 {
+	ieeeOnce.Do(ieeeInit)
+	return updateIEEE(0, data)
+}
diff --git a/vendor/github.com/klauspost/crc32/crc32_amd64.go b/vendor/github.com/klauspost/crc32/crc32_amd64.go
new file mode 100644
index 0000000000..af2a0b844b
--- /dev/null
+++ b/vendor/github.com/klauspost/crc32/crc32_amd64.go
@@ -0,0 +1,230 @@
+// Copyright 2011 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.
+
+// +build !appengine,!gccgo
+
+// AMD64-specific hardware-assisted CRC32 algorithms. See crc32.go for a
+// description of the interface that each architecture-specific file
+// implements.
+
+package crc32
+
+import "unsafe"
+
+// This file contains the code to call the SSE 4.2 version of the Castagnoli
+// and IEEE CRC.
+
+// haveSSE41/haveSSE42/haveCLMUL are defined in crc_amd64.s and use
+// CPUID to test for SSE 4.1, 4.2 and CLMUL support.
+func haveSSE41() bool
+func haveSSE42() bool
+func haveCLMUL() bool
+
+// castagnoliSSE42 is defined in crc32_amd64.s and uses the SSE4.2 CRC32
+// instruction.
+//go:noescape
+func castagnoliSSE42(crc uint32, p []byte) uint32
+
+// castagnoliSSE42Triple is defined in crc32_amd64.s and uses the SSE4.2 CRC32
+// instruction.
+//go:noescape
+func castagnoliSSE42Triple(
+	crcA, crcB, crcC uint32,
+	a, b, c []byte,
+	rounds uint32,
+) (retA uint32, retB uint32, retC uint32)
+
+// ieeeCLMUL is defined in crc_amd64.s and uses the PCLMULQDQ
+// instruction as well as SSE 4.1.
+//go:noescape
+func ieeeCLMUL(crc uint32, p []byte) uint32
+
+var sse42 = haveSSE42()
+var useFastIEEE = haveCLMUL() && haveSSE41()
+
+const castagnoliK1 = 168
+const castagnoliK2 = 1344
+
+type sse42Table [4]Table
+
+var castagnoliSSE42TableK1 *sse42Table
+var castagnoliSSE42TableK2 *sse42Table
+
+func archAvailableCastagnoli() bool {
+	return sse42
+}
+
+func archInitCastagnoli() {
+	if !sse42 {
+		panic("arch-specific Castagnoli not available")
+	}
+	castagnoliSSE42TableK1 = new(sse42Table)
+	castagnoliSSE42TableK2 = new(sse42Table)
+	// See description in updateCastagnoli.
+	//    t[0][i] = CRC(i000, O)
+	//    t[1][i] = CRC(0i00, O)
+	//    t[2][i] = CRC(00i0, O)
+	//    t[3][i] = CRC(000i, O)
+	// where O is a sequence of K zeros.
+	var tmp [castagnoliK2]byte
+	for b := 0; b < 4; b++ {
+		for i := 0; i < 256; i++ {
+			val := uint32(i) << uint32(b*8)
+			castagnoliSSE42TableK1[b][i] = castagnoliSSE42(val, tmp[:castagnoliK1])
+			castagnoliSSE42TableK2[b][i] = castagnoliSSE42(val, tmp[:])
+		}
+	}
+}
+
+// castagnoliShift computes the CRC32-C of K1 or K2 zeroes (depending on the
+// table given) with the given initial crc value. This corresponds to
+// CRC(crc, O) in the description in updateCastagnoli.
+func castagnoliShift(table *sse42Table, crc uint32) uint32 {
+	return table[3][crc>>24] ^
+		table[2][(crc>>16)&0xFF] ^
+		table[1][(crc>>8)&0xFF] ^
+		table[0][crc&0xFF]
+}
+
+func archUpdateCastagnoli(crc uint32, p []byte) uint32 {
+	if !sse42 {
+		panic("not available")
+	}
+
+	// This method is inspired from the algorithm in Intel's white paper:
+	//    "Fast CRC Computation for iSCSI Polynomial Using CRC32 Instruction"
+	// The same strategy of splitting the buffer in three is used but the
+	// combining calculation is different; the complete derivation is explained
+	// below.
+	//
+	// -- The basic idea --
+	//
+	// The CRC32 instruction (available in SSE4.2) can process 8 bytes at a
+	// time. In recent Intel architectures the instruction takes 3 cycles;
+	// however the processor can pipeline up to three instructions if they
+	// don't depend on each other.
+	//
+	// Roughly this means that we can process three buffers in about the same
+	// time we can process one buffer.
+	//
+	// The idea is then to split the buffer in three, CRC the three pieces
+	// separately and then combine the results.
+	//
+	// Combining the results requires precomputed tables, so we must choose a
+	// fixed buffer length to optimize. The longer the length, the faster; but
+	// only buffers longer than this length will use the optimization. We choose
+	// two cutoffs and compute tables for both:
+	//  - one around 512: 168*3=504
+	//  - one around 4KB: 1344*3=4032
+	//
+	// -- The nitty gritty --
+	//
+	// Let CRC(I, X) be the non-inverted CRC32-C of the sequence X (with
+	// initial non-inverted CRC I). This function has the following properties:
+	//   (a) CRC(I, AB) = CRC(CRC(I, A), B)
+	//   (b) CRC(I, A xor B) = CRC(I, A) xor CRC(0, B)
+	//
+	// Say we want to compute CRC(I, ABC) where A, B, C are three sequences of
+	// K bytes each, where K is a fixed constant. Let O be the sequence of K zero
+	// bytes.
+	//
+	// CRC(I, ABC) = CRC(I, ABO xor C)
+	//             = CRC(I, ABO) xor CRC(0, C)
+	//             = CRC(CRC(I, AB), O) xor CRC(0, C)
+	//             = CRC(CRC(I, AO xor B), O) xor CRC(0, C)
+	//             = CRC(CRC(I, AO) xor CRC(0, B), O) xor CRC(0, C)
+	//             = CRC(CRC(CRC(I, A), O) xor CRC(0, B), O) xor CRC(0, C)
+	//
+	// The castagnoliSSE42Triple function can compute CRC(I, A), CRC(0, B),
+	// and CRC(0, C) efficiently.  We just need to find a way to quickly compute
+	// CRC(uvwx, O) given a 4-byte initial value uvwx. We can precompute these
+	// values; since we can't have a 32-bit table, we break it up into four
+	// 8-bit tables:
+	//
+	//    CRC(uvwx, O) = CRC(u000, O) xor
+	//                   CRC(0v00, O) xor
+	//                   CRC(00w0, O) xor
+	//                   CRC(000x, O)
+	//
+	// We can compute tables corresponding to the four terms for all 8-bit
+	// values.
+
+	crc = ^crc
+
+	// If a buffer is long enough to use the optimization, process the first few
+	// bytes to align the buffer to an 8 byte boundary (if necessary).
+	if len(p) >= castagnoliK1*3 {
+		delta := int(uintptr(unsafe.Pointer(&p[0])) & 7)
+		if delta != 0 {
+			delta = 8 - delta
+			crc = castagnoliSSE42(crc, p[:delta])
+			p = p[delta:]
+		}
+	}
+
+	// Process 3*K2 at a time.
+	for len(p) >= castagnoliK2*3 {
+		// Compute CRC(I, A), CRC(0, B), and CRC(0, C).
+		crcA, crcB, crcC := castagnoliSSE42Triple(
+			crc, 0, 0,
+			p, p[castagnoliK2:], p[castagnoliK2*2:],
+			castagnoliK2/24)
+
+		// CRC(I, AB) = CRC(CRC(I, A), O) xor CRC(0, B)
+		crcAB := castagnoliShift(castagnoliSSE42TableK2, crcA) ^ crcB
+		// CRC(I, ABC) = CRC(CRC(I, AB), O) xor CRC(0, C)
+		crc = castagnoliShift(castagnoliSSE42TableK2, crcAB) ^ crcC
+		p = p[castagnoliK2*3:]
+	}
+
+	// Process 3*K1 at a time.
+	for len(p) >= castagnoliK1*3 {
+		// Compute CRC(I, A), CRC(0, B), and CRC(0, C).
+		crcA, crcB, crcC := castagnoliSSE42Triple(
+			crc, 0, 0,
+			p, p[castagnoliK1:], p[castagnoliK1*2:],
+			castagnoliK1/24)
+
+		// CRC(I, AB) = CRC(CRC(I, A), O) xor CRC(0, B)
+		crcAB := castagnoliShift(castagnoliSSE42TableK1, crcA) ^ crcB
+		// CRC(I, ABC) = CRC(CRC(I, AB), O) xor CRC(0, C)
+		crc = castagnoliShift(castagnoliSSE42TableK1, crcAB) ^ crcC
+		p = p[castagnoliK1*3:]
+	}
+
+	// Use the simple implementation for what's left.
+	crc = castagnoliSSE42(crc, p)
+	return ^crc
+}
+
+func archAvailableIEEE() bool {
+	return useFastIEEE
+}
+
+var archIeeeTable8 *slicing8Table
+
+func archInitIEEE() {
+	if !useFastIEEE {
+		panic("not available")
+	}
+	// We still use slicing-by-8 for small buffers.
+	archIeeeTable8 = slicingMakeTable(IEEE)
+}
+
+func archUpdateIEEE(crc uint32, p []byte) uint32 {
+	if !useFastIEEE {
+		panic("not available")
+	}
+
+	if len(p) >= 64 {
+		left := len(p) & 15
+		do := len(p) - left
+		crc = ^ieeeCLMUL(^crc, p[:do])
+		p = p[do:]
+	}
+	if len(p) == 0 {
+		return crc
+	}
+	return slicingUpdate(crc, archIeeeTable8, p)
+}
diff --git a/vendor/github.com/klauspost/crc32/crc32_amd64.s b/vendor/github.com/klauspost/crc32/crc32_amd64.s
new file mode 100644
index 0000000000..e8a7941ce7
--- /dev/null
+++ b/vendor/github.com/klauspost/crc32/crc32_amd64.s
@@ -0,0 +1,319 @@
+// Copyright 2011 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.
+
+// +build gc
+
+#define NOSPLIT 4
+#define RODATA 8
+
+// castagnoliSSE42 updates the (non-inverted) crc with the given buffer.
+//
+// func castagnoliSSE42(crc uint32, p []byte) uint32
+TEXT ·castagnoliSSE42(SB), NOSPLIT, $0
+	MOVL crc+0(FP), AX    // CRC value
+	MOVQ p+8(FP), SI      // data pointer
+	MOVQ p_len+16(FP), CX // len(p)
+
+	// If there are fewer than 8 bytes to process, skip alignment.
+	CMPQ CX, $8
+	JL   less_than_8
+
+	MOVQ SI, BX
+	ANDQ $7, BX
+	JZ   aligned
+
+	// Process the first few bytes to 8-byte align the input.
+
+	// BX = 8 - BX. We need to process this many bytes to align.
+	SUBQ $1, BX
+	XORQ $7, BX
+
+	BTQ $0, BX
+	JNC align_2
+
+	CRC32B (SI), AX
+	DECQ   CX
+	INCQ   SI
+
+align_2:
+	BTQ $1, BX
+	JNC align_4
+
+	// CRC32W (SI), AX
+	BYTE $0x66; BYTE $0xf2; BYTE $0x0f; BYTE $0x38; BYTE $0xf1; BYTE $0x06
+
+	SUBQ $2, CX
+	ADDQ $2, SI
+
+align_4:
+	BTQ $2, BX
+	JNC aligned
+
+	// CRC32L (SI), AX
+	BYTE $0xf2; BYTE $0x0f; BYTE $0x38; BYTE $0xf1; BYTE $0x06
+
+	SUBQ $4, CX
+	ADDQ $4, SI
+
+aligned:
+	// The input is now 8-byte aligned and we can process 8-byte chunks.
+	CMPQ CX, $8
+	JL   less_than_8
+
+	CRC32Q (SI), AX
+	ADDQ   $8, SI
+	SUBQ   $8, CX
+	JMP    aligned
+
+less_than_8:
+	// We may have some bytes left over; process 4 bytes, then 2, then 1.
+	BTQ $2, CX
+	JNC less_than_4
+
+	// CRC32L (SI), AX
+	BYTE $0xf2; BYTE $0x0f; BYTE $0x38; BYTE $0xf1; BYTE $0x06
+	ADDQ $4, SI
+
+less_than_4:
+	BTQ $1, CX
+	JNC less_than_2
+
+	// CRC32W (SI), AX
+	BYTE $0x66; BYTE $0xf2; BYTE $0x0f; BYTE $0x38; BYTE $0xf1; BYTE $0x06
+	ADDQ $2, SI
+
+less_than_2:
+	BTQ $0, CX
+	JNC done
+
+	CRC32B (SI), AX
+
+done:
+	MOVL AX, ret+32(FP)
+	RET
+
+// castagnoliSSE42Triple updates three (non-inverted) crcs with (24*rounds)
+// bytes from each buffer.
+//
+// func castagnoliSSE42Triple(
+//     crc1, crc2, crc3 uint32,
+//     a, b, c []byte,
+//     rounds uint32,
+// ) (retA uint32, retB uint32, retC uint32)
+TEXT ·castagnoliSSE42Triple(SB), NOSPLIT, $0
+	MOVL crcA+0(FP), AX
+	MOVL crcB+4(FP), CX
+	MOVL crcC+8(FP), DX
+
+	MOVQ a+16(FP), R8  // data pointer
+	MOVQ b+40(FP), R9  // data pointer
+	MOVQ c+64(FP), R10 // data pointer
+
+	MOVL rounds+88(FP), R11
+
+loop:
+	CRC32Q (R8), AX
+	CRC32Q (R9), CX
+	CRC32Q (R10), DX
+
+	CRC32Q 8(R8), AX
+	CRC32Q 8(R9), CX
+	CRC32Q 8(R10), DX
+
+	CRC32Q 16(R8), AX
+	CRC32Q 16(R9), CX
+	CRC32Q 16(R10), DX
+
+	ADDQ $24, R8
+	ADDQ $24, R9
+	ADDQ $24, R10
+
+	DECQ R11
+	JNZ  loop
+
+	MOVL AX, retA+96(FP)
+	MOVL CX, retB+100(FP)
+	MOVL DX, retC+104(FP)
+	RET
+
+// func haveSSE42() bool
+TEXT ·haveSSE42(SB), NOSPLIT, $0
+	XORQ AX, AX
+	INCL AX
+	CPUID
+	SHRQ $20, CX
+	ANDQ $1, CX
+	MOVB CX, ret+0(FP)
+	RET
+
+// func haveCLMUL() bool
+TEXT ·haveCLMUL(SB), NOSPLIT, $0
+	XORQ AX, AX
+	INCL AX
+	CPUID
+	SHRQ $1, CX
+	ANDQ $1, CX
+	MOVB CX, ret+0(FP)
+	RET
+
+// func haveSSE41() bool
+TEXT ·haveSSE41(SB), NOSPLIT, $0
+	XORQ AX, AX
+	INCL AX
+	CPUID
+	SHRQ $19, CX
+	ANDQ $1, CX
+	MOVB CX, ret+0(FP)
+	RET
+
+// CRC32 polynomial data
+//
+// These constants are lifted from the
+// Linux kernel, since they avoid the costly
+// PSHUFB 16 byte reversal proposed in the
+// original Intel paper.
+DATA r2r1kp<>+0(SB)/8, $0x154442bd4
+DATA r2r1kp<>+8(SB)/8, $0x1c6e41596
+DATA r4r3kp<>+0(SB)/8, $0x1751997d0
+DATA r4r3kp<>+8(SB)/8, $0x0ccaa009e
+DATA rupolykp<>+0(SB)/8, $0x1db710641
+DATA rupolykp<>+8(SB)/8, $0x1f7011641
+DATA r5kp<>+0(SB)/8, $0x163cd6124
+
+GLOBL r2r1kp<>(SB), RODATA, $16
+GLOBL r4r3kp<>(SB), RODATA, $16
+GLOBL rupolykp<>(SB), RODATA, $16
+GLOBL r5kp<>(SB), RODATA, $8
+
+// Based on http://www.intel.com/content/dam/www/public/us/en/documents/white-papers/fast-crc-computation-generic-polynomials-pclmulqdq-paper.pdf
+// len(p) must be at least 64, and must be a multiple of 16.
+
+// func ieeeCLMUL(crc uint32, p []byte) uint32
+TEXT ·ieeeCLMUL(SB), NOSPLIT, $0
+	MOVL crc+0(FP), X0    // Initial CRC value
+	MOVQ p+8(FP), SI      // data pointer
+	MOVQ p_len+16(FP), CX // len(p)
+
+	MOVOU (SI), X1
+	MOVOU 16(SI), X2
+	MOVOU 32(SI), X3
+	MOVOU 48(SI), X4
+	PXOR  X0, X1
+	ADDQ  $64, SI    // buf+=64
+	SUBQ  $64, CX    // len-=64
+	CMPQ  CX, $64    // Less than 64 bytes left
+	JB    remain64
+
+	MOVOA r2r1kp<>+0(SB), X0
+
+loopback64:
+	MOVOA X1, X5
+	MOVOA X2, X6
+	MOVOA X3, X7
+	MOVOA X4, X8
+
+	PCLMULQDQ $0, X0, X1
+	PCLMULQDQ $0, X0, X2
+	PCLMULQDQ $0, X0, X3
+	PCLMULQDQ $0, X0, X4
+
+	// Load next early
+	MOVOU (SI), X11
+	MOVOU 16(SI), X12
+	MOVOU 32(SI), X13
+	MOVOU 48(SI), X14
+
+	PCLMULQDQ $0x11, X0, X5
+	PCLMULQDQ $0x11, X0, X6
+	PCLMULQDQ $0x11, X0, X7
+	PCLMULQDQ $0x11, X0, X8
+
+	PXOR X5, X1
+	PXOR X6, X2
+	PXOR X7, X3
+	PXOR X8, X4
+
+	PXOR X11, X1
+	PXOR X12, X2
+	PXOR X13, X3
+	PXOR X14, X4
+
+	ADDQ $0x40, DI
+	ADDQ $64, SI    // buf+=64
+	SUBQ $64, CX    // len-=64
+	CMPQ CX, $64    // Less than 64 bytes left?
+	JGE  loopback64
+
+	// Fold result into a single register (X1)
+remain64:
+	MOVOA r4r3kp<>+0(SB), X0
+
+	MOVOA     X1, X5
+	PCLMULQDQ $0, X0, X1
+	PCLMULQDQ $0x11, X0, X5
+	PXOR      X5, X1
+	PXOR      X2, X1
+
+	MOVOA     X1, X5
+	PCLMULQDQ $0, X0, X1
+	PCLMULQDQ $0x11, X0, X5
+	PXOR      X5, X1
+	PXOR      X3, X1
+
+	MOVOA     X1, X5
+	PCLMULQDQ $0, X0, X1
+	PCLMULQDQ $0x11, X0, X5
+	PXOR      X5, X1
+	PXOR      X4, X1
+
+	// If there is less than 16 bytes left we are done
+	CMPQ CX, $16
+	JB   finish
+
+	// Encode 16 bytes
+remain16:
+	MOVOU     (SI), X10
+	MOVOA     X1, X5
+	PCLMULQDQ $0, X0, X1
+	PCLMULQDQ $0x11, X0, X5
+	PXOR      X5, X1
+	PXOR      X10, X1
+	SUBQ      $16, CX
+	ADDQ      $16, SI
+	CMPQ      CX, $16
+	JGE       remain16
+
+finish:
+	// Fold final result into 32 bits and return it
+	PCMPEQB   X3, X3
+	PCLMULQDQ $1, X1, X0
+	PSRLDQ    $8, X1
+	PXOR      X0, X1
+
+	MOVOA X1, X2
+	MOVQ  r5kp<>+0(SB), X0
+
+	// Creates 32 bit mask. Note that we don't care about upper half.
+	PSRLQ $32, X3
+
+	PSRLDQ    $4, X2
+	PAND      X3, X1
+	PCLMULQDQ $0, X0, X1
+	PXOR      X2, X1
+
+	MOVOA rupolykp<>+0(SB), X0
+
+	MOVOA     X1, X2
+	PAND      X3, X1
+	PCLMULQDQ $0x10, X0, X1
+	PAND      X3, X1
+	PCLMULQDQ $0, X0, X1
+	PXOR      X2, X1
+
+	// PEXTRD   $1, X1, AX  (SSE 4.1)
+	BYTE $0x66; BYTE $0x0f; BYTE $0x3a
+	BYTE $0x16; BYTE $0xc8; BYTE $0x01
+	MOVL AX, ret+32(FP)
+
+	RET
diff --git a/vendor/github.com/klauspost/crc32/crc32_amd64p32.go b/vendor/github.com/klauspost/crc32/crc32_amd64p32.go
new file mode 100644
index 0000000000..3222b06a5a
--- /dev/null
+++ b/vendor/github.com/klauspost/crc32/crc32_amd64p32.go
@@ -0,0 +1,43 @@
+// Copyright 2011 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.
+
+// +build !appengine,!gccgo
+
+package crc32
+
+// This file contains the code to call the SSE 4.2 version of the Castagnoli
+// CRC.
+
+// haveSSE42 is defined in crc32_amd64p32.s and uses CPUID to test for SSE 4.2
+// support.
+func haveSSE42() bool
+
+// castagnoliSSE42 is defined in crc32_amd64p32.s and uses the SSE4.2 CRC32
+// instruction.
+//go:noescape
+func castagnoliSSE42(crc uint32, p []byte) uint32
+
+var sse42 = haveSSE42()
+
+func archAvailableCastagnoli() bool {
+	return sse42
+}
+
+func archInitCastagnoli() {
+	if !sse42 {
+		panic("not available")
+	}
+	// No initialization necessary.
+}
+
+func archUpdateCastagnoli(crc uint32, p []byte) uint32 {
+	if !sse42 {
+		panic("not available")
+	}
+	return castagnoliSSE42(crc, p)
+}
+
+func archAvailableIEEE() bool                    { return false }
+func archInitIEEE()                              { panic("not available") }
+func archUpdateIEEE(crc uint32, p []byte) uint32 { panic("not available") }
diff --git a/vendor/github.com/klauspost/crc32/crc32_amd64p32.s b/vendor/github.com/klauspost/crc32/crc32_amd64p32.s
new file mode 100644
index 0000000000..a578d685cc
--- /dev/null
+++ b/vendor/github.com/klauspost/crc32/crc32_amd64p32.s
@@ -0,0 +1,67 @@
+// Copyright 2011 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.
+
+// +build gc
+
+#define NOSPLIT 4
+#define RODATA 8
+
+// func castagnoliSSE42(crc uint32, p []byte) uint32
+TEXT ·castagnoliSSE42(SB), NOSPLIT, $0
+	MOVL crc+0(FP), AX   // CRC value
+	MOVL p+4(FP), SI     // data pointer
+	MOVL p_len+8(FP), CX // len(p)
+
+	NOTL AX
+
+	// If there's less than 8 bytes to process, we do it byte-by-byte.
+	CMPQ CX, $8
+	JL   cleanup
+
+	// Process individual bytes until the input is 8-byte aligned.
+startup:
+	MOVQ SI, BX
+	ANDQ $7, BX
+	JZ   aligned
+
+	CRC32B (SI), AX
+	DECQ   CX
+	INCQ   SI
+	JMP    startup
+
+aligned:
+	// The input is now 8-byte aligned and we can process 8-byte chunks.
+	CMPQ CX, $8
+	JL   cleanup
+
+	CRC32Q (SI), AX
+	ADDQ   $8, SI
+	SUBQ   $8, CX
+	JMP    aligned
+
+cleanup:
+	// We may have some bytes left over that we process one at a time.
+	CMPQ CX, $0
+	JE   done
+
+	CRC32B (SI), AX
+	INCQ   SI
+	DECQ   CX
+	JMP    cleanup
+
+done:
+	NOTL AX
+	MOVL AX, ret+16(FP)
+	RET
+
+// func haveSSE42() bool
+TEXT ·haveSSE42(SB), NOSPLIT, $0
+	XORQ AX, AX
+	INCL AX
+	CPUID
+	SHRQ $20, CX
+	ANDQ $1, CX
+	MOVB CX, ret+0(FP)
+	RET
+
diff --git a/vendor/github.com/klauspost/crc32/crc32_generic.go b/vendor/github.com/klauspost/crc32/crc32_generic.go
new file mode 100644
index 0000000000..abacbb663d
--- /dev/null
+++ b/vendor/github.com/klauspost/crc32/crc32_generic.go
@@ -0,0 +1,89 @@
+// Copyright 2011 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.
+
+// This file contains CRC32 algorithms that are not specific to any architecture
+// and don't use hardware acceleration.
+//
+// The simple (and slow) CRC32 implementation only uses a 256*4 bytes table.
+//
+// The slicing-by-8 algorithm is a faster implementation that uses a bigger
+// table (8*256*4 bytes).
+
+package crc32
+
+// simpleMakeTable allocates and constructs a Table for the specified
+// polynomial. The table is suitable for use with the simple algorithm
+// (simpleUpdate).
+func simpleMakeTable(poly uint32) *Table {
+	t := new(Table)
+	simplePopulateTable(poly, t)
+	return t
+}
+
+// simplePopulateTable constructs a Table for the specified polynomial, suitable
+// for use with simpleUpdate.
+func simplePopulateTable(poly uint32, t *Table) {
+	for i := 0; i < 256; i++ {
+		crc := uint32(i)
+		for j := 0; j < 8; j++ {
+			if crc&1 == 1 {
+				crc = (crc >> 1) ^ poly
+			} else {
+				crc >>= 1
+			}
+		}
+		t[i] = crc
+	}
+}
+
+// simpleUpdate uses the simple algorithm to update the CRC, given a table that
+// was previously computed using simpleMakeTable.
+func simpleUpdate(crc uint32, tab *Table, p []byte) uint32 {
+	crc = ^crc
+	for _, v := range p {
+		crc = tab[byte(crc)^v] ^ (crc >> 8)
+	}
+	return ^crc
+}
+
+// Use slicing-by-8 when payload >= this value.
+const slicing8Cutoff = 16
+
+// slicing8Table is array of 8 Tables, used by the slicing-by-8 algorithm.
+type slicing8Table [8]Table
+
+// slicingMakeTable constructs a slicing8Table for the specified polynomial. The
+// table is suitable for use with the slicing-by-8 algorithm (slicingUpdate).
+func slicingMakeTable(poly uint32) *slicing8Table {
+	t := new(slicing8Table)
+	simplePopulateTable(poly, &t[0])
+	for i := 0; i < 256; i++ {
+		crc := t[0][i]
+		for j := 1; j < 8; j++ {
+			crc = t[0][crc&0xFF] ^ (crc >> 8)
+			t[j][i] = crc
+		}
+	}
+	return t
+}
+
+// slicingUpdate uses the slicing-by-8 algorithm to update the CRC, given a
+// table that was previously computed using slicingMakeTable.
+func slicingUpdate(crc uint32, tab *slicing8Table, p []byte) uint32 {
+	if len(p) >= slicing8Cutoff {
+		crc = ^crc
+		for len(p) > 8 {
+			crc ^= uint32(p[0]) | uint32(p[1])<<8 | uint32(p[2])<<16 | uint32(p[3])<<24
+			crc = tab[0][p[7]] ^ tab[1][p[6]] ^ tab[2][p[5]] ^ tab[3][p[4]] ^
+				tab[4][crc>>24] ^ tab[5][(crc>>16)&0xFF] ^
+				tab[6][(crc>>8)&0xFF] ^ tab[7][crc&0xFF]
+			p = p[8:]
+		}
+		crc = ^crc
+	}
+	if len(p) == 0 {
+		return crc
+	}
+	return simpleUpdate(crc, &tab[0], p)
+}
diff --git a/vendor/github.com/klauspost/crc32/crc32_otherarch.go b/vendor/github.com/klauspost/crc32/crc32_otherarch.go
new file mode 100644
index 0000000000..cc960764bc
--- /dev/null
+++ b/vendor/github.com/klauspost/crc32/crc32_otherarch.go
@@ -0,0 +1,15 @@
+// Copyright 2011 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.
+
+// +build !amd64,!amd64p32,!s390x
+
+package crc32
+
+func archAvailableIEEE() bool                    { return false }
+func archInitIEEE()                              { panic("not available") }
+func archUpdateIEEE(crc uint32, p []byte) uint32 { panic("not available") }
+
+func archAvailableCastagnoli() bool                    { return false }
+func archInitCastagnoli()                              { panic("not available") }
+func archUpdateCastagnoli(crc uint32, p []byte) uint32 { panic("not available") }
diff --git a/vendor/github.com/klauspost/crc32/crc32_s390x.go b/vendor/github.com/klauspost/crc32/crc32_s390x.go
new file mode 100644
index 0000000000..ce96f03281
--- /dev/null
+++ b/vendor/github.com/klauspost/crc32/crc32_s390x.go
@@ -0,0 +1,91 @@
+// 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.
+
+// +build s390x
+
+package crc32
+
+const (
+	vxMinLen    = 64
+	vxAlignMask = 15 // align to 16 bytes
+)
+
+// hasVectorFacility reports whether the machine has the z/Architecture
+// vector facility installed and enabled.
+func hasVectorFacility() bool
+
+var hasVX = hasVectorFacility()
+
+// vectorizedCastagnoli implements CRC32 using vector instructions.
+// It is defined in crc32_s390x.s.
+//go:noescape
+func vectorizedCastagnoli(crc uint32, p []byte) uint32
+
+// vectorizedIEEE implements CRC32 using vector instructions.
+// It is defined in crc32_s390x.s.
+//go:noescape
+func vectorizedIEEE(crc uint32, p []byte) uint32
+
+func archAvailableCastagnoli() bool {
+	return hasVX
+}
+
+var archCastagnoliTable8 *slicing8Table
+
+func archInitCastagnoli() {
+	if !hasVX {
+		panic("not available")
+	}
+	// We still use slicing-by-8 for small buffers.
+	archCastagnoliTable8 = slicingMakeTable(Castagnoli)
+}
+
+// archUpdateCastagnoli calculates the checksum of p using
+// vectorizedCastagnoli.
+func archUpdateCastagnoli(crc uint32, p []byte) uint32 {
+	if !hasVX {
+		panic("not available")
+	}
+	// Use vectorized function if data length is above threshold.
+	if len(p) >= vxMinLen {
+		aligned := len(p) & ^vxAlignMask
+		crc = vectorizedCastagnoli(crc, p[:aligned])
+		p = p[aligned:]
+	}
+	if len(p) == 0 {
+		return crc
+	}
+	return slicingUpdate(crc, archCastagnoliTable8, p)
+}
+
+func archAvailableIEEE() bool {
+	return hasVX
+}
+
+var archIeeeTable8 *slicing8Table
+
+func archInitIEEE() {
+	if !hasVX {
+		panic("not available")
+	}
+	// We still use slicing-by-8 for small buffers.
+	archIeeeTable8 = slicingMakeTable(IEEE)
+}
+
+// archUpdateIEEE calculates the checksum of p using vectorizedIEEE.
+func archUpdateIEEE(crc uint32, p []byte) uint32 {
+	if !hasVX {
+		panic("not available")
+	}
+	// Use vectorized function if data length is above threshold.
+	if len(p) >= vxMinLen {
+		aligned := len(p) & ^vxAlignMask
+		crc = vectorizedIEEE(crc, p[:aligned])
+		p = p[aligned:]
+	}
+	if len(p) == 0 {
+		return crc
+	}
+	return slicingUpdate(crc, archIeeeTable8, p)
+}
diff --git a/vendor/github.com/klauspost/crc32/crc32_s390x.s b/vendor/github.com/klauspost/crc32/crc32_s390x.s
new file mode 100644
index 0000000000..e980ca29d6
--- /dev/null
+++ b/vendor/github.com/klauspost/crc32/crc32_s390x.s
@@ -0,0 +1,249 @@
+// 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.
+
+// +build s390x
+
+#include "textflag.h"
+
+// Vector register range containing CRC-32 constants
+
+#define CONST_PERM_LE2BE        V9
+#define CONST_R2R1              V10
+#define CONST_R4R3              V11
+#define CONST_R5                V12
+#define CONST_RU_POLY           V13
+#define CONST_CRC_POLY          V14
+
+// The CRC-32 constant block contains reduction constants to fold and
+// process particular chunks of the input data stream in parallel.
+//
+// Note that the constant definitions below are extended in order to compute
+// intermediate results with a single VECTOR GALOIS FIELD MULTIPLY instruction.
+// The rightmost doubleword can be 0 to prevent contribution to the result or
+// can be multiplied by 1 to perform an XOR without the need for a separate
+// VECTOR EXCLUSIVE OR instruction.
+//
+// The polynomials used are bit-reflected:
+//
+//            IEEE: P'(x) = 0x0edb88320
+//      Castagnoli: P'(x) = 0x082f63b78
+
+// IEEE polynomial constants
+DATA ·crcleconskp+0(SB)/8, $0x0F0E0D0C0B0A0908 // LE-to-BE mask
+DATA ·crcleconskp+8(SB)/8, $0x0706050403020100
+DATA ·crcleconskp+16(SB)/8, $0x00000001c6e41596 // R2
+DATA ·crcleconskp+24(SB)/8, $0x0000000154442bd4 // R1
+DATA ·crcleconskp+32(SB)/8, $0x00000000ccaa009e // R4
+DATA ·crcleconskp+40(SB)/8, $0x00000001751997d0 // R3
+DATA ·crcleconskp+48(SB)/8, $0x0000000000000000
+DATA ·crcleconskp+56(SB)/8, $0x0000000163cd6124 // R5
+DATA ·crcleconskp+64(SB)/8, $0x0000000000000000
+DATA ·crcleconskp+72(SB)/8, $0x00000001F7011641 // u'
+DATA ·crcleconskp+80(SB)/8, $0x0000000000000000
+DATA ·crcleconskp+88(SB)/8, $0x00000001DB710641 // P'(x) << 1
+
+GLOBL ·crcleconskp(SB), RODATA, $144
+
+// Castagonli Polynomial constants
+DATA ·crccleconskp+0(SB)/8, $0x0F0E0D0C0B0A0908 // LE-to-BE mask
+DATA ·crccleconskp+8(SB)/8, $0x0706050403020100
+DATA ·crccleconskp+16(SB)/8, $0x000000009e4addf8 // R2
+DATA ·crccleconskp+24(SB)/8, $0x00000000740eef02 // R1
+DATA ·crccleconskp+32(SB)/8, $0x000000014cd00bd6 // R4
+DATA ·crccleconskp+40(SB)/8, $0x00000000f20c0dfe // R3
+DATA ·crccleconskp+48(SB)/8, $0x0000000000000000
+DATA ·crccleconskp+56(SB)/8, $0x00000000dd45aab8 // R5
+DATA ·crccleconskp+64(SB)/8, $0x0000000000000000
+DATA ·crccleconskp+72(SB)/8, $0x00000000dea713f1 // u'
+DATA ·crccleconskp+80(SB)/8, $0x0000000000000000
+DATA ·crccleconskp+88(SB)/8, $0x0000000105ec76f0 // P'(x) << 1
+
+GLOBL ·crccleconskp(SB), RODATA, $144
+
+// func hasVectorFacility() bool
+TEXT ·hasVectorFacility(SB), NOSPLIT, $24-1
+	MOVD  $x-24(SP), R1
+	XC    $24, 0(R1), 0(R1) // clear the storage
+	MOVD  $2, R0            // R0 is the number of double words stored -1
+	WORD  $0xB2B01000       // STFLE 0(R1)
+	XOR   R0, R0            // reset the value of R0
+	MOVBZ z-8(SP), R1
+	AND   $0x40, R1
+	BEQ   novector
+
+vectorinstalled:
+	// check if the vector instruction has been enabled
+	VLEIB  $0, $0xF, V16
+	VLGVB  $0, V16, R1
+	CMPBNE R1, $0xF, novector
+	MOVB   $1, ret+0(FP)      // have vx
+	RET
+
+novector:
+	MOVB $0, ret+0(FP) // no vx
+	RET
+
+// The CRC-32 function(s) use these calling conventions:
+//
+// Parameters:
+//
+//      R2:    Initial CRC value, typically ~0; and final CRC (return) value.
+//      R3:    Input buffer pointer, performance might be improved if the
+//             buffer is on a doubleword boundary.
+//      R4:    Length of the buffer, must be 64 bytes or greater.
+//
+// Register usage:
+//
+//      R5:     CRC-32 constant pool base pointer.
+//      V0:     Initial CRC value and intermediate constants and results.
+//      V1..V4: Data for CRC computation.
+//      V5..V8: Next data chunks that are fetched from the input buffer.
+//
+//      V9..V14: CRC-32 constants.
+
+// func vectorizedIEEE(crc uint32, p []byte) uint32
+TEXT ·vectorizedIEEE(SB), NOSPLIT, $0
+	MOVWZ crc+0(FP), R2    // R2 stores the CRC value
+	MOVD  p+8(FP), R3      // data pointer
+	MOVD  p_len+16(FP), R4 // len(p)
+
+	MOVD $·crcleconskp(SB), R5
+	BR   vectorizedBody<>(SB)
+
+// func vectorizedCastagnoli(crc uint32, p []byte) uint32
+TEXT ·vectorizedCastagnoli(SB), NOSPLIT, $0
+	MOVWZ crc+0(FP), R2    // R2 stores the CRC value
+	MOVD  p+8(FP), R3      // data pointer
+	MOVD  p_len+16(FP), R4 // len(p)
+
+	// R5: crc-32 constant pool base pointer, constant is used to reduce crc
+	MOVD $·crccleconskp(SB), R5
+	BR   vectorizedBody<>(SB)
+
+TEXT vectorizedBody<>(SB), NOSPLIT, $0
+	XOR $0xffffffff, R2                         // NOTW R2
+	VLM 0(R5), CONST_PERM_LE2BE, CONST_CRC_POLY
+
+	// Load the initial CRC value into the rightmost word of V0
+	VZERO V0
+	VLVGF $3, R2, V0
+
+	// Crash if the input size is less than 64-bytes.
+	CMP R4, $64
+	BLT crash
+
+	// Load a 64-byte data chunk and XOR with CRC
+	VLM 0(R3), V1, V4 // 64-bytes into V1..V4
+
+	// Reflect the data if the CRC operation is in the bit-reflected domain
+	VPERM V1, V1, CONST_PERM_LE2BE, V1
+	VPERM V2, V2, CONST_PERM_LE2BE, V2
+	VPERM V3, V3, CONST_PERM_LE2BE, V3
+	VPERM V4, V4, CONST_PERM_LE2BE, V4
+
+	VX  V0, V1, V1 // V1 ^= CRC
+	ADD $64, R3    // BUF = BUF + 64
+	ADD $(-64), R4
+
+	// Check remaining buffer size and jump to proper folding method
+	CMP R4, $64
+	BLT less_than_64bytes
+
+fold_64bytes_loop:
+	// Load the next 64-byte data chunk into V5 to V8
+	VLM   0(R3), V5, V8
+	VPERM V5, V5, CONST_PERM_LE2BE, V5
+	VPERM V6, V6, CONST_PERM_LE2BE, V6
+	VPERM V7, V7, CONST_PERM_LE2BE, V7
+	VPERM V8, V8, CONST_PERM_LE2BE, V8
+
+	// Perform a GF(2) multiplication of the doublewords in V1 with
+	// the reduction constants in V0.  The intermediate result is
+	// then folded (accumulated) with the next data chunk in V5 and
+	// stored in V1.  Repeat this step for the register contents
+	// in V2, V3, and V4 respectively.
+
+	VGFMAG CONST_R2R1, V1, V5, V1
+	VGFMAG CONST_R2R1, V2, V6, V2
+	VGFMAG CONST_R2R1, V3, V7, V3
+	VGFMAG CONST_R2R1, V4, V8, V4
+
+	// Adjust buffer pointer and length for next loop
+	ADD $64, R3    // BUF = BUF + 64
+	ADD $(-64), R4 // LEN = LEN - 64
+
+	CMP R4, $64
+	BGE fold_64bytes_loop
+
+less_than_64bytes:
+	// Fold V1 to V4 into a single 128-bit value in V1
+	VGFMAG CONST_R4R3, V1, V2, V1
+	VGFMAG CONST_R4R3, V1, V3, V1
+	VGFMAG CONST_R4R3, V1, V4, V1
+
+	// Check whether to continue with 64-bit folding
+	CMP R4, $16
+	BLT final_fold
+
+fold_16bytes_loop:
+	VL    0(R3), V2                    // Load next data chunk
+	VPERM V2, V2, CONST_PERM_LE2BE, V2
+
+	VGFMAG CONST_R4R3, V1, V2, V1 // Fold next data chunk
+
+	// Adjust buffer pointer and size for folding next data chunk
+	ADD $16, R3
+	ADD $-16, R4
+
+	// Process remaining data chunks
+	CMP R4, $16
+	BGE fold_16bytes_loop
+
+final_fold:
+	VLEIB $7, $0x40, V9
+	VSRLB V9, CONST_R4R3, V0
+	VLEIG $0, $1, V0
+
+	VGFMG V0, V1, V1
+
+	VLEIB  $7, $0x20, V9        // Shift by words
+	VSRLB  V9, V1, V2           // Store remaining bits in V2
+	VUPLLF V1, V1               // Split rightmost doubleword
+	VGFMAG CONST_R5, V1, V2, V1 // V1 = (V1 * R5) XOR V2
+
+	// The input values to the Barret reduction are the degree-63 polynomial
+	// in V1 (R(x)), degree-32 generator polynomial, and the reduction
+	// constant u.  The Barret reduction result is the CRC value of R(x) mod
+	// P(x).
+	//
+	// The Barret reduction algorithm is defined as:
+	//
+	//    1. T1(x) = floor( R(x) / x^32 ) GF2MUL u
+	//    2. T2(x) = floor( T1(x) / x^32 ) GF2MUL P(x)
+	//    3. C(x)  = R(x) XOR T2(x) mod x^32
+	//
+	// Note: To compensate the division by x^32, use the vector unpack
+	// instruction to move the leftmost word into the leftmost doubleword
+	// of the vector register.  The rightmost doubleword is multiplied
+	// with zero to not contribute to the intermedate results.
+
+	// T1(x) = floor( R(x) / x^32 ) GF2MUL u
+	VUPLLF V1, V2
+	VGFMG  CONST_RU_POLY, V2, V2
+
+	// Compute the GF(2) product of the CRC polynomial in VO with T1(x) in
+	// V2 and XOR the intermediate result, T2(x),  with the value in V1.
+	// The final result is in the rightmost word of V2.
+
+	VUPLLF V2, V2
+	VGFMAG CONST_CRC_POLY, V2, V1, V2
+
+done:
+	VLGVF $2, V2, R2
+	XOR   $0xffffffff, R2  // NOTW R2
+	MOVWZ R2, ret + 32(FP)
+	RET
+
+crash:
+	MOVD $0, (R0) // input size is less than 64-bytes