Update dependencies (#1800)

11 files changed, 1992 insertions, 419 deletions

diff --git a/vendor/github.com/klauspost/compress/huff0/autogen.go b/vendor/github.com/klauspost/compress/huff0/autogen.go
new file mode 100644
index 00000000..ff2c69d6
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/huff0/autogen.go
@@ -0,0 +1,5 @@
+package huff0
+
+//go:generate go run generate.go
+//go:generate asmfmt -w decompress_amd64.s
+//go:generate asmfmt -w decompress_8b_amd64.s
diff --git a/vendor/github.com/klauspost/compress/huff0/bitreader.go b/vendor/github.com/klauspost/compress/huff0/bitreader.go
index a4979e88..451160ed 100644
--- a/vendor/github.com/klauspost/compress/huff0/bitreader.go
+++ b/vendor/github.com/klauspost/compress/huff0/bitreader.go
@@ -8,118 +8,13 @@ package huff0
 import (
 	"encoding/binary"
 	"errors"
+	"fmt"
 	"io"
 )
 
 // bitReader reads a bitstream in reverse.
 // The last set bit indicates the start of the stream and is used
 // for aligning the input.
-type bitReader struct {
-	in       []byte
-	off      uint // next byte to read is at in[off - 1]
-	value    uint64
-	bitsRead uint8
-}
-
-// init initializes and resets the bit reader.
-func (b *bitReader) init(in []byte) error {
-	if len(in) < 1 {
-		return errors.New("corrupt stream: too short")
-	}
-	b.in = in
-	b.off = uint(len(in))
-	// The highest bit of the last byte indicates where to start
-	v := in[len(in)-1]
-	if v == 0 {
-		return errors.New("corrupt stream, did not find end of stream")
-	}
-	b.bitsRead = 64
-	b.value = 0
-	if len(in) >= 8 {
-		b.fillFastStart()
-	} else {
-		b.fill()
-		b.fill()
-	}
-	b.bitsRead += 8 - uint8(highBit32(uint32(v)))
-	return nil
-}
-
-// peekBitsFast requires that at least one bit is requested every time.
-// There are no checks if the buffer is filled.
-func (b *bitReader) peekBitsFast(n uint8) uint16 {
-	const regMask = 64 - 1
-	v := uint16((b.value << (b.bitsRead & regMask)) >> ((regMask + 1 - n) & regMask))
-	return v
-}
-
-// fillFast() will make sure at least 32 bits are available.
-// There must be at least 4 bytes available.
-func (b *bitReader) fillFast() {
-	if b.bitsRead < 32 {
-		return
-	}
-
-	// 2 bounds checks.
-	v := b.in[b.off-4 : b.off]
-	v = v[:4]
-	low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
-	b.value = (b.value << 32) | uint64(low)
-	b.bitsRead -= 32
-	b.off -= 4
-}
-
-func (b *bitReader) advance(n uint8) {
-	b.bitsRead += n
-}
-
-// fillFastStart() assumes the bitreader is empty and there is at least 8 bytes to read.
-func (b *bitReader) fillFastStart() {
-	// Do single re-slice to avoid bounds checks.
-	b.value = binary.LittleEndian.Uint64(b.in[b.off-8:])
-	b.bitsRead = 0
-	b.off -= 8
-}
-
-// fill() will make sure at least 32 bits are available.
-func (b *bitReader) fill() {
-	if b.bitsRead < 32 {
-		return
-	}
-	if b.off > 4 {
-		v := b.in[b.off-4:]
-		v = v[:4]
-		low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
-		b.value = (b.value << 32) | uint64(low)
-		b.bitsRead -= 32
-		b.off -= 4
-		return
-	}
-	for b.off > 0 {
-		b.value = (b.value << 8) | uint64(b.in[b.off-1])
-		b.bitsRead -= 8
-		b.off--
-	}
-}
-
-// finished returns true if all bits have been read from the bit stream.
-func (b *bitReader) finished() bool {
-	return b.off == 0 && b.bitsRead >= 64
-}
-
-// close the bitstream and returns an error if out-of-buffer reads occurred.
-func (b *bitReader) close() error {
-	// Release reference.
-	b.in = nil
-	if b.bitsRead > 64 {
-		return io.ErrUnexpectedEOF
-	}
-	return nil
-}
-
-// bitReader reads a bitstream in reverse.
-// The last set bit indicates the start of the stream and is used
-// for aligning the input.
 type bitReaderBytes struct {
 	in       []byte
 	off      uint // next byte to read is at in[off - 1]
@@ -213,10 +108,17 @@ func (b *bitReaderBytes) finished() bool {
 	return b.off == 0 && b.bitsRead >= 64
 }
 
+func (b *bitReaderBytes) remaining() uint {
+	return b.off*8 + uint(64-b.bitsRead)
+}
+
 // close the bitstream and returns an error if out-of-buffer reads occurred.
 func (b *bitReaderBytes) close() error {
 	// Release reference.
 	b.in = nil
+	if b.remaining() > 0 {
+		return fmt.Errorf("corrupt input: %d bits remain on stream", b.remaining())
+	}
 	if b.bitsRead > 64 {
 		return io.ErrUnexpectedEOF
 	}
@@ -263,6 +165,11 @@ func (b *bitReaderShifted) peekBitsFast(n uint8) uint16 {
 	return uint16(b.value >> ((64 - n) & 63))
 }
 
+// peekTopBits(n) is equvialent to peekBitFast(64 - n)
+func (b *bitReaderShifted) peekTopBits(n uint8) uint16 {
+	return uint16(b.value >> n)
+}
+
 func (b *bitReaderShifted) advance(n uint8) {
 	b.bitsRead += n
 	b.value <<= n & 63
@@ -318,10 +225,17 @@ func (b *bitReaderShifted) finished() bool {
 	return b.off == 0 && b.bitsRead >= 64
 }
 
+func (b *bitReaderShifted) remaining() uint {
+	return b.off*8 + uint(64-b.bitsRead)
+}
+
 // close the bitstream and returns an error if out-of-buffer reads occurred.
 func (b *bitReaderShifted) close() error {
 	// Release reference.
 	b.in = nil
+	if b.remaining() > 0 {
+		return fmt.Errorf("corrupt input: %d bits remain on stream", b.remaining())
+	}
 	if b.bitsRead > 64 {
 		return io.ErrUnexpectedEOF
 	}
diff --git a/vendor/github.com/klauspost/compress/huff0/compress.go b/vendor/github.com/klauspost/compress/huff0/compress.go
index 8323dc05..bc95ac62 100644
--- a/vendor/github.com/klauspost/compress/huff0/compress.go
+++ b/vendor/github.com/klauspost/compress/huff0/compress.go
@@ -2,6 +2,7 @@ package huff0
 
 import (
 	"fmt"
+	"math"
 	"runtime"
 	"sync"
 )
@@ -289,6 +290,10 @@ func (s *Scratch) compress4X(src []byte) ([]byte, error) {
 		if err != nil {
 			return nil, err
 		}
+		if len(s.Out)-idx > math.MaxUint16 {
+			// We cannot store the size in the jump table
+			return nil, ErrIncompressible
+		}
 		// Write compressed length as little endian before block.
 		if i < 3 {
 			// Last length is not written.
@@ -332,6 +337,10 @@ func (s *Scratch) compress4Xp(src []byte) ([]byte, error) {
 			return nil, errs[i]
 		}
 		o := s.tmpOut[i]
+		if len(o) > math.MaxUint16 {
+			// We cannot store the size in the jump table
+			return nil, ErrIncompressible
+		}
 		// Write compressed length as little endian before block.
 		if i < 3 {
 			// Last length is not written.
diff --git a/vendor/github.com/klauspost/compress/huff0/decompress.go b/vendor/github.com/klauspost/compress/huff0/decompress.go
index 2a06bd1a..04f65299 100644
--- a/vendor/github.com/klauspost/compress/huff0/decompress.go
+++ b/vendor/github.com/klauspost/compress/huff0/decompress.go
@@ -4,6 +4,7 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"sync"
 
 	"github.com/klauspost/compress/fse"
 )
@@ -216,6 +217,7 @@ func (s *Scratch) Decoder() *Decoder {
 	return &Decoder{
 		dt:             s.dt,
 		actualTableLog: s.actualTableLog,
+		bufs:           &s.decPool,
 	}
 }
 
@@ -223,6 +225,15 @@ func (s *Scratch) Decoder() *Decoder {
 type Decoder struct {
 	dt             dTable
 	actualTableLog uint8
+	bufs           *sync.Pool
+}
+
+func (d *Decoder) buffer() *[4][256]byte {
+	buf, ok := d.bufs.Get().(*[4][256]byte)
+	if ok {
+		return buf
+	}
+	return &[4][256]byte{}
 }
 
 // Decompress1X will decompress a 1X encoded stream.
@@ -249,7 +260,8 @@ func (d *Decoder) Decompress1X(dst, src []byte) ([]byte, error) {
 	dt := d.dt.single[:tlSize]
 
 	// Use temp table to avoid bound checks/append penalty.
-	var buf [256]byte
+	bufs := d.buffer()
+	buf := &bufs[0]
 	var off uint8
 
 	for br.off >= 8 {
@@ -277,6 +289,7 @@ func (d *Decoder) Decompress1X(dst, src []byte) ([]byte, error) {
 		if off == 0 {
 			if len(dst)+256 > maxDecodedSize {
 				br.close()
+				d.bufs.Put(bufs)
 				return nil, ErrMaxDecodedSizeExceeded
 			}
 			dst = append(dst, buf[:]...)
@@ -284,6 +297,7 @@ func (d *Decoder) Decompress1X(dst, src []byte) ([]byte, error) {
 	}
 
 	if len(dst)+int(off) > maxDecodedSize {
+		d.bufs.Put(bufs)
 		br.close()
 		return nil, ErrMaxDecodedSizeExceeded
 	}
@@ -310,6 +324,7 @@ func (d *Decoder) Decompress1X(dst, src []byte) ([]byte, error) {
 			}
 		}
 		if len(dst) >= maxDecodedSize {
+			d.bufs.Put(bufs)
 			br.close()
 			return nil, ErrMaxDecodedSizeExceeded
 		}
@@ -319,6 +334,7 @@ func (d *Decoder) Decompress1X(dst, src []byte) ([]byte, error) {
 		bitsLeft -= nBits
 		dst = append(dst, uint8(v.entry>>8))
 	}
+	d.bufs.Put(bufs)
 	return dst, br.close()
 }
 
@@ -341,7 +357,8 @@ func (d *Decoder) decompress1X8Bit(dst, src []byte) ([]byte, error) {
 	dt := d.dt.single[:256]
 
 	// Use temp table to avoid bound checks/append penalty.
-	var buf [256]byte
+	bufs := d.buffer()
+	buf := &bufs[0]
 	var off uint8
 
 	switch d.actualTableLog {
@@ -369,6 +386,7 @@ func (d *Decoder) decompress1X8Bit(dst, src []byte) ([]byte, error) {
 			if off == 0 {
 				if len(dst)+256 > maxDecodedSize {
 					br.close()
+					d.bufs.Put(bufs)
 					return nil, ErrMaxDecodedSizeExceeded
 				}
 				dst = append(dst, buf[:]...)
@@ -398,6 +416,7 @@ func (d *Decoder) decompress1X8Bit(dst, src []byte) ([]byte, error) {
 			if off == 0 {
 				if len(dst)+256 > maxDecodedSize {
 					br.close()
+					d.bufs.Put(bufs)
 					return nil, ErrMaxDecodedSizeExceeded
 				}
 				dst = append(dst, buf[:]...)
@@ -426,6 +445,7 @@ func (d *Decoder) decompress1X8Bit(dst, src []byte) ([]byte, error) {
 			off += 4
 			if off == 0 {
 				if len(dst)+256 > maxDecodedSize {
+					d.bufs.Put(bufs)
 					br.close()
 					return nil, ErrMaxDecodedSizeExceeded
 				}
@@ -455,6 +475,7 @@ func (d *Decoder) decompress1X8Bit(dst, src []byte) ([]byte, error) {
 			off += 4
 			if off == 0 {
 				if len(dst)+256 > maxDecodedSize {
+					d.bufs.Put(bufs)
 					br.close()
 					return nil, ErrMaxDecodedSizeExceeded
 				}
@@ -484,6 +505,7 @@ func (d *Decoder) decompress1X8Bit(dst, src []byte) ([]byte, error) {
 			off += 4
 			if off == 0 {
 				if len(dst)+256 > maxDecodedSize {
+					d.bufs.Put(bufs)
 					br.close()
 					return nil, ErrMaxDecodedSizeExceeded
 				}
@@ -513,6 +535,7 @@ func (d *Decoder) decompress1X8Bit(dst, src []byte) ([]byte, error) {
 			off += 4
 			if off == 0 {
 				if len(dst)+256 > maxDecodedSize {
+					d.bufs.Put(bufs)
 					br.close()
 					return nil, ErrMaxDecodedSizeExceeded
 				}
@@ -542,6 +565,7 @@ func (d *Decoder) decompress1X8Bit(dst, src []byte) ([]byte, error) {
 			off += 4
 			if off == 0 {
 				if len(dst)+256 > maxDecodedSize {
+					d.bufs.Put(bufs)
 					br.close()
 					return nil, ErrMaxDecodedSizeExceeded
 				}
@@ -571,6 +595,7 @@ func (d *Decoder) decompress1X8Bit(dst, src []byte) ([]byte, error) {
 			off += 4
 			if off == 0 {
 				if len(dst)+256 > maxDecodedSize {
+					d.bufs.Put(bufs)
 					br.close()
 					return nil, ErrMaxDecodedSizeExceeded
 				}
@@ -578,10 +603,12 @@ func (d *Decoder) decompress1X8Bit(dst, src []byte) ([]byte, error) {
 			}
 		}
 	default:
+		d.bufs.Put(bufs)
 		return nil, fmt.Errorf("invalid tablelog: %d", d.actualTableLog)
 	}
 
 	if len(dst)+int(off) > maxDecodedSize {
+		d.bufs.Put(bufs)
 		br.close()
 		return nil, ErrMaxDecodedSizeExceeded
 	}
@@ -601,6 +628,7 @@ func (d *Decoder) decompress1X8Bit(dst, src []byte) ([]byte, error) {
 		}
 		if len(dst) >= maxDecodedSize {
 			br.close()
+			d.bufs.Put(bufs)
 			return nil, ErrMaxDecodedSizeExceeded
 		}
 		v := dt[br.peekByteFast()>>shift]
@@ -609,6 +637,7 @@ func (d *Decoder) decompress1X8Bit(dst, src []byte) ([]byte, error) {
 		bitsLeft -= int8(nBits)
 		dst = append(dst, uint8(v.entry>>8))
 	}
+	d.bufs.Put(bufs)
 	return dst, br.close()
 }
 
@@ -628,7 +657,8 @@ func (d *Decoder) decompress1X8BitExactly(dst, src []byte) ([]byte, error) {
 	dt := d.dt.single[:256]
 
 	// Use temp table to avoid bound checks/append penalty.
-	var buf [256]byte
+	bufs := d.buffer()
+	buf := &bufs[0]
 	var off uint8
 
 	const shift = 56
@@ -655,6 +685,7 @@ func (d *Decoder) decompress1X8BitExactly(dst, src []byte) ([]byte, error) {
 		off += 4
 		if off == 0 {
 			if len(dst)+256 > maxDecodedSize {
+				d.bufs.Put(bufs)
 				br.close()
 				return nil, ErrMaxDecodedSizeExceeded
 			}
@@ -663,6 +694,7 @@ func (d *Decoder) decompress1X8BitExactly(dst, src []byte) ([]byte, error) {
 	}
 
 	if len(dst)+int(off) > maxDecodedSize {
+		d.bufs.Put(bufs)
 		br.close()
 		return nil, ErrMaxDecodedSizeExceeded
 	}
@@ -679,6 +711,7 @@ func (d *Decoder) decompress1X8BitExactly(dst, src []byte) ([]byte, error) {
 			}
 		}
 		if len(dst) >= maxDecodedSize {
+			d.bufs.Put(bufs)
 			br.close()
 			return nil, ErrMaxDecodedSizeExceeded
 		}
@@ -688,6 +721,7 @@ func (d *Decoder) decompress1X8BitExactly(dst, src []byte) ([]byte, error) {
 		bitsLeft -= int8(nBits)
 		dst = append(dst, uint8(v.entry>>8))
 	}
+	d.bufs.Put(bufs)
 	return dst, br.close()
 }
 
@@ -695,192 +729,6 @@ func (d *Decoder) decompress1X8BitExactly(dst, src []byte) ([]byte, error) {
 // The length of the supplied input must match the end of a block exactly.
 // The *capacity* of the dst slice must match the destination size of
 // the uncompressed data exactly.
-func (d *Decoder) Decompress4X(dst, src []byte) ([]byte, error) {
-	if len(d.dt.single) == 0 {
-		return nil, errors.New("no table loaded")
-	}
-	if len(src) < 6+(4*1) {
-		return nil, errors.New("input too small")
-	}
-	if use8BitTables && d.actualTableLog <= 8 {
-		return d.decompress4X8bit(dst, src)
-	}
-
-	var br [4]bitReaderShifted
-	start := 6
-	for i := 0; i < 3; i++ {
-		length := int(src[i*2]) | (int(src[i*2+1]) << 8)
-		if start+length >= len(src) {
-			return nil, errors.New("truncated input (or invalid offset)")
-		}
-		err := br[i].init(src[start : start+length])
-		if err != nil {
-			return nil, err
-		}
-		start += length
-	}
-	err := br[3].init(src[start:])
-	if err != nil {
-		return nil, err
-	}
-
-	// destination, offset to match first output
-	dstSize := cap(dst)
-	dst = dst[:dstSize]
-	out := dst
-	dstEvery := (dstSize + 3) / 4
-
-	const tlSize = 1 << tableLogMax
-	const tlMask = tlSize - 1
-	single := d.dt.single[:tlSize]
-
-	// Use temp table to avoid bound checks/append penalty.
-	var buf [256]byte
-	var off uint8
-	var decoded int
-
-	// Decode 2 values from each decoder/loop.
-	const bufoff = 256 / 4
-	for {
-		if br[0].off < 4 || br[1].off < 4 || br[2].off < 4 || br[3].off < 4 {
-			break
-		}
-
-		{
-			const stream = 0
-			const stream2 = 1
-			br[stream].fillFast()
-			br[stream2].fillFast()
-
-			val := br[stream].peekBitsFast(d.actualTableLog)
-			val2 := br[stream2].peekBitsFast(d.actualTableLog)
-			v := single[val&tlMask]
-			v2 := single[val2&tlMask]
-			br[stream].advance(uint8(v.entry))
-			br[stream2].advance(uint8(v2.entry))
-			buf[off+bufoff*stream] = uint8(v.entry >> 8)
-			buf[off+bufoff*stream2] = uint8(v2.entry >> 8)
-
-			val = br[stream].peekBitsFast(d.actualTableLog)
-			val2 = br[stream2].peekBitsFast(d.actualTableLog)
-			v = single[val&tlMask]
-			v2 = single[val2&tlMask]
-			br[stream].advance(uint8(v.entry))
-			br[stream2].advance(uint8(v2.entry))
-			buf[off+bufoff*stream+1] = uint8(v.entry >> 8)
-			buf[off+bufoff*stream2+1] = uint8(v2.entry >> 8)
-		}
-
-		{
-			const stream = 2
-			const stream2 = 3
-			br[stream].fillFast()
-			br[stream2].fillFast()
-
-			val := br[stream].peekBitsFast(d.actualTableLog)
-			val2 := br[stream2].peekBitsFast(d.actualTableLog)
-			v := single[val&tlMask]
-			v2 := single[val2&tlMask]
-			br[stream].advance(uint8(v.entry))
-			br[stream2].advance(uint8(v2.entry))
-			buf[off+bufoff*stream] = uint8(v.entry >> 8)
-			buf[off+bufoff*stream2] = uint8(v2.entry >> 8)
-
-			val = br[stream].peekBitsFast(d.actualTableLog)
-			val2 = br[stream2].peekBitsFast(d.actualTableLog)
-			v = single[val&tlMask]
-			v2 = single[val2&tlMask]
-			br[stream].advance(uint8(v.entry))
-			br[stream2].advance(uint8(v2.entry))
-			buf[off+bufoff*stream+1] = uint8(v.entry >> 8)
-			buf[off+bufoff*stream2+1] = uint8(v2.entry >> 8)
-		}
-
-		off += 2
-
-		if off == bufoff {
-			if bufoff > dstEvery {
-				return nil, errors.New("corruption detected: stream overrun 1")
-			}
-			copy(out, buf[:bufoff])
-			copy(out[dstEvery:], buf[bufoff:bufoff*2])
-			copy(out[dstEvery*2:], buf[bufoff*2:bufoff*3])
-			copy(out[dstEvery*3:], buf[bufoff*3:bufoff*4])
-			off = 0
-			out = out[bufoff:]
-			decoded += 256
-			// There must at least be 3 buffers left.
-			if len(out) < dstEvery*3 {
-				return nil, errors.New("corruption detected: stream overrun 2")
-			}
-		}
-	}
-	if off > 0 {
-		ioff := int(off)
-		if len(out) < dstEvery*3+ioff {
-			return nil, errors.New("corruption detected: stream overrun 3")
-		}
-		copy(out, buf[:off])
-		copy(out[dstEvery:dstEvery+ioff], buf[bufoff:bufoff*2])
-		copy(out[dstEvery*2:dstEvery*2+ioff], buf[bufoff*2:bufoff*3])
-		copy(out[dstEvery*3:dstEvery*3+ioff], buf[bufoff*3:bufoff*4])
-		decoded += int(off) * 4
-		out = out[off:]
-	}
-
-	// Decode remaining.
-	for i := range br {
-		offset := dstEvery * i
-		br := &br[i]
-		bitsLeft := br.off*8 + uint(64-br.bitsRead)
-		for bitsLeft > 0 {
-			br.fill()
-			if false && br.bitsRead >= 32 {
-				if br.off >= 4 {
-					v := br.in[br.off-4:]
-					v = v[:4]
-					low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
-					br.value = (br.value << 32) | uint64(low)
-					br.bitsRead -= 32
-					br.off -= 4
-				} else {
-					for br.off > 0 {
-						br.value = (br.value << 8) | uint64(br.in[br.off-1])
-						br.bitsRead -= 8
-						br.off--
-					}
-				}
-			}
-			// end inline...
-			if offset >= len(out) {
-				return nil, errors.New("corruption detected: stream overrun 4")
-			}
-
-			// Read value and increment offset.
-			val := br.peekBitsFast(d.actualTableLog)
-			v := single[val&tlMask].entry
-			nBits := uint8(v)
-			br.advance(nBits)
-			bitsLeft -= uint(nBits)
-			out[offset] = uint8(v >> 8)
-			offset++
-		}
-		decoded += offset - dstEvery*i
-		err = br.close()
-		if err != nil {
-			return nil, err
-		}
-	}
-	if dstSize != decoded {
-		return nil, errors.New("corruption detected: short output block")
-	}
-	return dst, nil
-}
-
-// Decompress4X will decompress a 4X encoded stream.
-// The length of the supplied input must match the end of a block exactly.
-// The *capacity* of the dst slice must match the destination size of
-// the uncompressed data exactly.
 func (d *Decoder) decompress4X8bit(dst, src []byte) ([]byte, error) {
 	if d.actualTableLog == 8 {
 		return d.decompress4X8bitExactly(dst, src)
@@ -916,12 +764,12 @@ func (d *Decoder) decompress4X8bit(dst, src []byte) ([]byte, error) {
 	single := d.dt.single[:tlSize]
 
 	// Use temp table to avoid bound checks/append penalty.
-	var buf [256]byte
+	buf := d.buffer()
 	var off uint8
 	var decoded int
 
 	// Decode 4 values from each decoder/loop.
-	const bufoff = 256 / 4
+	const bufoff = 256
 	for {
 		if br[0].off < 4 || br[1].off < 4 || br[2].off < 4 || br[3].off < 4 {
 			break
@@ -942,8 +790,8 @@ func (d *Decoder) decompress4X8bit(dst, src []byte) ([]byte, error) {
 			br1.value <<= v & 63
 			br2.bitsRead += uint8(v2)
 			br2.value <<= v2 & 63
-			buf[off+bufoff*stream] = uint8(v >> 8)
-			buf[off+bufoff*stream2] = uint8(v2 >> 8)
+			buf[stream][off] = uint8(v >> 8)
+			buf[stream2][off] = uint8(v2 >> 8)
 
 			v = single[uint8(br1.value>>shift)].entry
 			v2 = single[uint8(br2.value>>shift)].entry
@@ -951,8 +799,8 @@ func (d *Decoder) decompress4X8bit(dst, src []byte) ([]byte, error) {
 			br1.value <<= v & 63
 			br2.bitsRead += uint8(v2)
 			br2.value <<= v2 & 63
-			buf[off+bufoff*stream+1] = uint8(v >> 8)
-			buf[off+bufoff*stream2+1] = uint8(v2 >> 8)
+			buf[stream][off+1] = uint8(v >> 8)
+			buf[stream2][off+1] = uint8(v2 >> 8)
 
 			v = single[uint8(br1.value>>shift)].entry
 			v2 = single[uint8(br2.value>>shift)].entry
@@ -960,8 +808,8 @@ func (d *Decoder) decompress4X8bit(dst, src []byte) ([]byte, error) {
 			br1.value <<= v & 63
 			br2.bitsRead += uint8(v2)
 			br2.value <<= v2 & 63
-			buf[off+bufoff*stream+2] = uint8(v >> 8)
-			buf[off+bufoff*stream2+2] = uint8(v2 >> 8)
+			buf[stream][off+2] = uint8(v >> 8)
+			buf[stream2][off+2] = uint8(v2 >> 8)
 
 			v = single[uint8(br1.value>>shift)].entry
 			v2 = single[uint8(br2.value>>shift)].entry
@@ -969,8 +817,8 @@ func (d *Decoder) decompress4X8bit(dst, src []byte) ([]byte, error) {
 			br1.value <<= v & 63
 			br2.bitsRead += uint8(v2)
 			br2.value <<= v2 & 63
-			buf[off+bufoff*stream2+3] = uint8(v2 >> 8)
-			buf[off+bufoff*stream+3] = uint8(v >> 8)
+			buf[stream][off+3] = uint8(v >> 8)
+			buf[stream2][off+3] = uint8(v2 >> 8)
 		}
 
 		{
@@ -987,8 +835,8 @@ func (d *Decoder) decompress4X8bit(dst, src []byte) ([]byte, error) {
 			br1.value <<= v & 63
 			br2.bitsRead += uint8(v2)
 			br2.value <<= v2 & 63
-			buf[off+bufoff*stream] = uint8(v >> 8)
-			buf[off+bufoff*stream2] = uint8(v2 >> 8)
+			buf[stream][off] = uint8(v >> 8)
+			buf[stream2][off] = uint8(v2 >> 8)
 
 			v = single[uint8(br1.value>>shift)].entry
 			v2 = single[uint8(br2.value>>shift)].entry
@@ -996,8 +844,8 @@ func (d *Decoder) decompress4X8bit(dst, src []byte) ([]byte, error) {
 			br1.value <<= v & 63
 			br2.bitsRead += uint8(v2)
 			br2.value <<= v2 & 63
-			buf[off+bufoff*stream+1] = uint8(v >> 8)
-			buf[off+bufoff*stream2+1] = uint8(v2 >> 8)
+			buf[stream][off+1] = uint8(v >> 8)
+			buf[stream2][off+1] = uint8(v2 >> 8)
 
 			v = single[uint8(br1.value>>shift)].entry
 			v2 = single[uint8(br2.value>>shift)].entry
@@ -1005,8 +853,8 @@ func (d *Decoder) decompress4X8bit(dst, src []byte) ([]byte, error) {
 			br1.value <<= v & 63
 			br2.bitsRead += uint8(v2)
 			br2.value <<= v2 & 63
-			buf[off+bufoff*stream+2] = uint8(v >> 8)
-			buf[off+bufoff*stream2+2] = uint8(v2 >> 8)
+			buf[stream][off+2] = uint8(v >> 8)
+			buf[stream2][off+2] = uint8(v2 >> 8)
 
 			v = single[uint8(br1.value>>shift)].entry
 			v2 = single[uint8(br2.value>>shift)].entry
@@ -1014,25 +862,26 @@ func (d *Decoder) decompress4X8bit(dst, src []byte) ([]byte, error) {
 			br1.value <<= v & 63
 			br2.bitsRead += uint8(v2)
 			br2.value <<= v2 & 63
-			buf[off+bufoff*stream2+3] = uint8(v2 >> 8)
-			buf[off+bufoff*stream+3] = uint8(v >> 8)
+			buf[stream][off+3] = uint8(v >> 8)
+			buf[stream2][off+3] = uint8(v2 >> 8)
 		}
 
 		off += 4
 
-		if off == bufoff {
+		if off == 0 {
 			if bufoff > dstEvery {
+				d.bufs.Put(buf)
 				return nil, errors.New("corruption detected: stream overrun 1")
 			}
-			copy(out, buf[:bufoff])
-			copy(out[dstEvery:], buf[bufoff:bufoff*2])
-			copy(out[dstEvery*2:], buf[bufoff*2:bufoff*3])
-			copy(out[dstEvery*3:], buf[bufoff*3:bufoff*4])
-			off = 0
+			copy(out, buf[0][:])
+			copy(out[dstEvery:], buf[1][:])
+			copy(out[dstEvery*2:], buf[2][:])
+			copy(out[dstEvery*3:], buf[3][:])
 			out = out[bufoff:]
-			decoded += 256
+			decoded += bufoff * 4
 			// There must at least be 3 buffers left.
 			if len(out) < dstEvery*3 {
+				d.bufs.Put(buf)
 				return nil, errors.New("corruption detected: stream overrun 2")
 			}
 		}
@@ -1040,23 +889,31 @@ func (d *Decoder) decompress4X8bit(dst, src []byte) ([]byte, error) {
 	if off > 0 {
 		ioff := int(off)
 		if len(out) < dstEvery*3+ioff {
+			d.bufs.Put(buf)
 			return nil, errors.New("corruption detected: stream overrun 3")
 		}
-		copy(out, buf[:off])
-		copy(out[dstEvery:dstEvery+ioff], buf[bufoff:bufoff*2])
-		copy(out[dstEvery*2:dstEvery*2+ioff], buf[bufoff*2:bufoff*3])
-		copy(out[dstEvery*3:dstEvery*3+ioff], buf[bufoff*3:bufoff*4])
+		copy(out, buf[0][:off])
+		copy(out[dstEvery:], buf[1][:off])
+		copy(out[dstEvery*2:], buf[2][:off])
+		copy(out[dstEvery*3:], buf[3][:off])
 		decoded += int(off) * 4
 		out = out[off:]
 	}
 
 	// Decode remaining.
+	// Decode remaining.
+	remainBytes := dstEvery - (decoded / 4)
 	for i := range br {
 		offset := dstEvery * i
+		endsAt := offset + remainBytes
+		if endsAt > len(out) {
+			endsAt = len(out)
+		}
 		br := &br[i]
-		bitsLeft := int(br.off*8) + int(64-br.bitsRead)
+		bitsLeft := br.remaining()
 		for bitsLeft > 0 {
 			if br.finished() {
+				d.bufs.Put(buf)
 				return nil, io.ErrUnexpectedEOF
 			}
 			if br.bitsRead >= 56 {
@@ -1076,7 +933,8 @@ func (d *Decoder) decompress4X8bit(dst, src []byte) ([]byte, error) {
 				}
 			}
 			// end inline...
-			if offset >= len(out) {
+			if offset >= endsAt {
+				d.bufs.Put(buf)
 				return nil, errors.New("corruption detected: stream overrun 4")
 			}
 
@@ -1084,16 +942,22 @@ func (d *Decoder) decompress4X8bit(dst, src []byte) ([]byte, error) {
 			v := single[uint8(br.value>>shift)].entry
 			nBits := uint8(v)
 			br.advance(nBits)
-			bitsLeft -= int(nBits)
+			bitsLeft -= uint(nBits)
 			out[offset] = uint8(v >> 8)
 			offset++
 		}
+		if offset != endsAt {
+			d.bufs.Put(buf)
+			return nil, fmt.Errorf("corruption detected: short output block %d, end %d != %d", i, offset, endsAt)
+		}
 		decoded += offset - dstEvery*i
 		err = br.close()
 		if err != nil {
+			d.bufs.Put(buf)
 			return nil, err
 		}
 	}
+	d.bufs.Put(buf)
 	if dstSize != decoded {
 		return nil, errors.New("corruption detected: short output block")
 	}
@@ -1135,12 +999,12 @@ func (d *Decoder) decompress4X8bitExactly(dst, src []byte) ([]byte, error) {
 	single := d.dt.single[:tlSize]
 
 	// Use temp table to avoid bound checks/append penalty.
-	var buf [256]byte
+	buf := d.buffer()
 	var off uint8
 	var decoded int
 
 	// Decode 4 values from each decoder/loop.
-	const bufoff = 256 / 4
+	const bufoff = 256
 	for {
 		if br[0].off < 4 || br[1].off < 4 || br[2].off < 4 || br[3].off < 4 {
 			break
@@ -1150,104 +1014,109 @@ func (d *Decoder) decompress4X8bitExactly(dst, src []byte) ([]byte, error) {
 			// Interleave 2 decodes.
 			const stream = 0
 			const stream2 = 1
-			br[stream].fillFast()
-			br[stream2].fillFast()
-
-			v := single[uint8(br[stream].value>>shift)].entry
-			v2 := single[uint8(br[stream2].value>>shift)].entry
-			br[stream].bitsRead += uint8(v)
-			br[stream].value <<= v & 63
-			br[stream2].bitsRead += uint8(v2)
-			br[stream2].value <<= v2 & 63
-			buf[off+bufoff*stream] = uint8(v >> 8)
-			buf[off+bufoff*stream2] = uint8(v2 >> 8)
-
-			v = single[uint8(br[stream].value>>shift)].entry
-			v2 = single[uint8(br[stream2].value>>shift)].entry
-			br[stream].bitsRead += uint8(v)
-			br[stream].value <<= v & 63
-			br[stream2].bitsRead += uint8(v2)
-			br[stream2].value <<= v2 & 63
-			buf[off+bufoff*stream+1] = uint8(v >> 8)
-			buf[off+bufoff*stream2+1] = uint8(v2 >> 8)
-
-			v = single[uint8(br[stream].value>>shift)].entry
-			v2 = single[uint8(br[stream2].value>>shift)].entry
-			br[stream].bitsRead += uint8(v)
-			br[stream].value <<= v & 63
-			br[stream2].bitsRead += uint8(v2)
-			br[stream2].value <<= v2 & 63
-			buf[off+bufoff*stream+2] = uint8(v >> 8)
-			buf[off+bufoff*stream2+2] = uint8(v2 >> 8)
-
-			v = single[uint8(br[stream].value>>shift)].entry
-			v2 = single[uint8(br[stream2].value>>shift)].entry
-			br[stream].bitsRead += uint8(v)
-			br[stream].value <<= v & 63
-			br[stream2].bitsRead += uint8(v2)
-			br[stream2].value <<= v2 & 63
-			buf[off+bufoff*stream+3] = uint8(v >> 8)
-			buf[off+bufoff*stream2+3] = uint8(v2 >> 8)
+			br1 := &br[stream]
+			br2 := &br[stream2]
+			br1.fillFast()
+			br2.fillFast()
+
+			v := single[uint8(br1.value>>shift)].entry
+			v2 := single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off] = uint8(v >> 8)
+			buf[stream2][off] = uint8(v2 >> 8)
+
+			v = single[uint8(br1.value>>shift)].entry
+			v2 = single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off+1] = uint8(v >> 8)
+			buf[stream2][off+1] = uint8(v2 >> 8)
+
+			v = single[uint8(br1.value>>shift)].entry
+			v2 = single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off+2] = uint8(v >> 8)
+			buf[stream2][off+2] = uint8(v2 >> 8)
+
+			v = single[uint8(br1.value>>shift)].entry
+			v2 = single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off+3] = uint8(v >> 8)
+			buf[stream2][off+3] = uint8(v2 >> 8)
 		}
 
 		{
 			const stream = 2
 			const stream2 = 3
-			br[stream].fillFast()
-			br[stream2].fillFast()
-
-			v := single[uint8(br[stream].value>>shift)].entry
-			v2 := single[uint8(br[stream2].value>>shift)].entry
-			br[stream].bitsRead += uint8(v)
-			br[stream].value <<= v & 63
-			br[stream2].bitsRead += uint8(v2)
-			br[stream2].value <<= v2 & 63
-			buf[off+bufoff*stream] = uint8(v >> 8)
-			buf[off+bufoff*stream2] = uint8(v2 >> 8)
-
-			v = single[uint8(br[stream].value>>shift)].entry
-			v2 = single[uint8(br[stream2].value>>shift)].entry
-			br[stream].bitsRead += uint8(v)
-			br[stream].value <<= v & 63
-			br[stream2].bitsRead += uint8(v2)
-			br[stream2].value <<= v2 & 63
-			buf[off+bufoff*stream+1] = uint8(v >> 8)
-			buf[off+bufoff*stream2+1] = uint8(v2 >> 8)
-
-			v = single[uint8(br[stream].value>>shift)].entry
-			v2 = single[uint8(br[stream2].value>>shift)].entry
-			br[stream].bitsRead += uint8(v)
-			br[stream].value <<= v & 63
-			br[stream2].bitsRead += uint8(v2)
-			br[stream2].value <<= v2 & 63
-			buf[off+bufoff*stream+2] = uint8(v >> 8)
-			buf[off+bufoff*stream2+2] = uint8(v2 >> 8)
-
-			v = single[uint8(br[stream].value>>shift)].entry
-			v2 = single[uint8(br[stream2].value>>shift)].entry
-			br[stream].bitsRead += uint8(v)
-			br[stream].value <<= v & 63
-			br[stream2].bitsRead += uint8(v2)
-			br[stream2].value <<= v2 & 63
-			buf[off+bufoff*stream+3] = uint8(v >> 8)
-			buf[off+bufoff*stream2+3] = uint8(v2 >> 8)
+			br1 := &br[stream]
+			br2 := &br[stream2]
+			br1.fillFast()
+			br2.fillFast()
+
+			v := single[uint8(br1.value>>shift)].entry
+			v2 := single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off] = uint8(v >> 8)
+			buf[stream2][off] = uint8(v2 >> 8)
+
+			v = single[uint8(br1.value>>shift)].entry
+			v2 = single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off+1] = uint8(v >> 8)
+			buf[stream2][off+1] = uint8(v2 >> 8)
+
+			v = single[uint8(br1.value>>shift)].entry
+			v2 = single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off+2] = uint8(v >> 8)
+			buf[stream2][off+2] = uint8(v2 >> 8)
+
+			v = single[uint8(br1.value>>shift)].entry
+			v2 = single[uint8(br2.value>>shift)].entry
+			br1.bitsRead += uint8(v)
+			br1.value <<= v & 63
+			br2.bitsRead += uint8(v2)
+			br2.value <<= v2 & 63
+			buf[stream][off+3] = uint8(v >> 8)
+			buf[stream2][off+3] = uint8(v2 >> 8)
 		}
 
 		off += 4
 
-		if off == bufoff {
+		if off == 0 {
 			if bufoff > dstEvery {
+				d.bufs.Put(buf)
 				return nil, errors.New("corruption detected: stream overrun 1")
 			}
-			copy(out, buf[:bufoff])
-			copy(out[dstEvery:], buf[bufoff:bufoff*2])
-			copy(out[dstEvery*2:], buf[bufoff*2:bufoff*3])
-			copy(out[dstEvery*3:], buf[bufoff*3:bufoff*4])
-			off = 0
+			copy(out, buf[0][:])
+			copy(out[dstEvery:], buf[1][:])
+			copy(out[dstEvery*2:], buf[2][:])
+			copy(out[dstEvery*3:], buf[3][:])
 			out = out[bufoff:]
-			decoded += 256
+			decoded += bufoff * 4
 			// There must at least be 3 buffers left.
 			if len(out) < dstEvery*3 {
+				d.bufs.Put(buf)
 				return nil, errors.New("corruption detected: stream overrun 2")
 			}
 		}
@@ -1257,21 +1126,27 @@ func (d *Decoder) decompress4X8bitExactly(dst, src []byte) ([]byte, error) {
 		if len(out) < dstEvery*3+ioff {
 			return nil, errors.New("corruption detected: stream overrun 3")
 		}
-		copy(out, buf[:off])
-		copy(out[dstEvery:dstEvery+ioff], buf[bufoff:bufoff*2])
-		copy(out[dstEvery*2:dstEvery*2+ioff], buf[bufoff*2:bufoff*3])
-		copy(out[dstEvery*3:dstEvery*3+ioff], buf[bufoff*3:bufoff*4])
+		copy(out, buf[0][:off])
+		copy(out[dstEvery:], buf[1][:off])
+		copy(out[dstEvery*2:], buf[2][:off])
+		copy(out[dstEvery*3:], buf[3][:off])
 		decoded += int(off) * 4
 		out = out[off:]
 	}
 
 	// Decode remaining.
+	remainBytes := dstEvery - (decoded / 4)
 	for i := range br {
 		offset := dstEvery * i
+		endsAt := offset + remainBytes
+		if endsAt > len(out) {
+			endsAt = len(out)
+		}
 		br := &br[i]
-		bitsLeft := int(br.off*8) + int(64-br.bitsRead)
+		bitsLeft := br.remaining()
 		for bitsLeft > 0 {
 			if br.finished() {
+				d.bufs.Put(buf)
 				return nil, io.ErrUnexpectedEOF
 			}
 			if br.bitsRead >= 56 {
@@ -1291,7 +1166,8 @@ func (d *Decoder) decompress4X8bitExactly(dst, src []byte) ([]byte, error) {
 				}
 			}
 			// end inline...
-			if offset >= len(out) {
+			if offset >= endsAt {
+				d.bufs.Put(buf)
 				return nil, errors.New("corruption detected: stream overrun 4")
 			}
 
@@ -1299,16 +1175,23 @@ func (d *Decoder) decompress4X8bitExactly(dst, src []byte) ([]byte, error) {
 			v := single[br.peekByteFast()].entry
 			nBits := uint8(v)
 			br.advance(nBits)
-			bitsLeft -= int(nBits)
+			bitsLeft -= uint(nBits)
 			out[offset] = uint8(v >> 8)
 			offset++
 		}
+		if offset != endsAt {
+			d.bufs.Put(buf)
+			return nil, fmt.Errorf("corruption detected: short output block %d, end %d != %d", i, offset, endsAt)
+		}
+
 		decoded += offset - dstEvery*i
 		err = br.close()
 		if err != nil {
+			d.bufs.Put(buf)
 			return nil, err
 		}
 	}
+	d.bufs.Put(buf)
 	if dstSize != decoded {
 		return nil, errors.New("corruption detected: short output block")
 	}
diff --git a/vendor/github.com/klauspost/compress/huff0/decompress_8b_amd64.s b/vendor/github.com/klauspost/compress/huff0/decompress_8b_amd64.s
new file mode 100644
index 00000000..0d6cb1a9
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/huff0/decompress_8b_amd64.s
@@ -0,0 +1,488 @@
+// +build !appengine
+// +build gc
+// +build !noasm
+
+#include "textflag.h"
+#include "funcdata.h"
+#include "go_asm.h"
+
+#define bufoff      256 // see decompress.go, we're using [4][256]byte table
+
+// func decompress4x_main_loop_x86(pbr0, pbr1, pbr2, pbr3 *bitReaderShifted,
+//	peekBits uint8, buf *byte, tbl *dEntrySingle) (int, bool)
+TEXT ·decompress4x_8b_loop_x86(SB), NOSPLIT, $8
+#define off             R8
+#define buffer          DI
+#define table           SI
+
+#define br_bits_read    R9
+#define br_value        R10
+#define br_offset       R11
+#define peek_bits       R12
+#define exhausted       DX
+
+#define br0             R13
+#define br1             R14
+#define br2             R15
+#define br3             BP
+
+	MOVQ BP, 0(SP)
+
+	XORQ exhausted, exhausted // exhausted = false
+	XORQ off, off             // off = 0
+
+	MOVBQZX peekBits+32(FP), peek_bits
+	MOVQ    buf+40(FP), buffer
+	MOVQ    tbl+48(FP), table
+
+	MOVQ pbr0+0(FP), br0
+	MOVQ pbr1+8(FP), br1
+	MOVQ pbr2+16(FP), br2
+	MOVQ pbr3+24(FP), br3
+
+main_loop:
+
+	// const stream = 0
+	// br0.fillFast()
+	MOVBQZX bitReaderShifted_bitsRead(br0), br_bits_read
+	MOVQ    bitReaderShifted_value(br0), br_value
+	MOVQ    bitReaderShifted_off(br0), br_offset
+
+	// if b.bitsRead >= 32 {
+	CMPQ br_bits_read, $32
+	JB   skip_fill0
+
+	SUBQ $32, br_bits_read // b.bitsRead -= 32
+	SUBQ $4, br_offset     // b.off -= 4
+
+	// v := b.in[b.off-4 : b.off]
+	// v = v[:4]
+	// low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	MOVQ bitReaderShifted_in(br0), AX
+	MOVL 0(br_offset)(AX*1), AX       // AX = uint32(b.in[b.off:b.off+4])
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+	MOVQ br_bits_read, CX
+	SHLQ CL, AX
+	ORQ  AX, br_value
+
+	// exhausted = exhausted || (br0.off < 4)
+	CMPQ  br_offset, $4
+	SETLT DL
+	ORB   DL, DH
+
+	// }
+skip_fill0:
+
+	// val0 := br0.peekTopBits(peekBits)
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v0 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br0.advance(uint8(v0.entry))
+	MOVB    AH, BL           // BL = uint8(v0.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// val1 := br0.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v1 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br0.advance(uint8(v1.entry))
+	MOVB    AH, BH           // BH = uint8(v1.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CX, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off] = uint8(v0.entry >> 8)
+	// buf[stream][off+1] = uint8(v1.entry >> 8)
+	MOVW BX, 0(buffer)(off*1)
+
+	// SECOND PART:
+	// val2 := br0.peekTopBits(peekBits)
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v2 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br0.advance(uint8(v0.entry))
+	MOVB    AH, BL           // BL = uint8(v0.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// val3 := br0.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v3 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br0.advance(uint8(v1.entry))
+	MOVB    AH, BH           // BH = uint8(v1.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CX, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off+2] = uint8(v2.entry >> 8)
+	// buf[stream][off+3] = uint8(v3.entry >> 8)
+	MOVW BX, 0+2(buffer)(off*1)
+
+	// update the bitrader reader structure
+	MOVB br_bits_read, bitReaderShifted_bitsRead(br0)
+	MOVQ br_value, bitReaderShifted_value(br0)
+	MOVQ br_offset, bitReaderShifted_off(br0)
+
+	// const stream = 1
+	// br1.fillFast()
+	MOVBQZX bitReaderShifted_bitsRead(br1), br_bits_read
+	MOVQ    bitReaderShifted_value(br1), br_value
+	MOVQ    bitReaderShifted_off(br1), br_offset
+
+	// if b.bitsRead >= 32 {
+	CMPQ br_bits_read, $32
+	JB   skip_fill1
+
+	SUBQ $32, br_bits_read // b.bitsRead -= 32
+	SUBQ $4, br_offset     // b.off -= 4
+
+	// v := b.in[b.off-4 : b.off]
+	// v = v[:4]
+	// low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	MOVQ bitReaderShifted_in(br1), AX
+	MOVL 0(br_offset)(AX*1), AX       // AX = uint32(b.in[b.off:b.off+4])
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+	MOVQ br_bits_read, CX
+	SHLQ CL, AX
+	ORQ  AX, br_value
+
+	// exhausted = exhausted || (br1.off < 4)
+	CMPQ  br_offset, $4
+	SETLT DL
+	ORB   DL, DH
+
+	// }
+skip_fill1:
+
+	// val0 := br1.peekTopBits(peekBits)
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v0 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br1.advance(uint8(v0.entry))
+	MOVB    AH, BL           // BL = uint8(v0.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// val1 := br1.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v1 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br1.advance(uint8(v1.entry))
+	MOVB    AH, BH           // BH = uint8(v1.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CX, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off] = uint8(v0.entry >> 8)
+	// buf[stream][off+1] = uint8(v1.entry >> 8)
+	MOVW BX, 256(buffer)(off*1)
+
+	// SECOND PART:
+	// val2 := br1.peekTopBits(peekBits)
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v2 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br1.advance(uint8(v0.entry))
+	MOVB    AH, BL           // BL = uint8(v0.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// val3 := br1.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v3 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br1.advance(uint8(v1.entry))
+	MOVB    AH, BH           // BH = uint8(v1.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CX, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off+2] = uint8(v2.entry >> 8)
+	// buf[stream][off+3] = uint8(v3.entry >> 8)
+	MOVW BX, 256+2(buffer)(off*1)
+
+	// update the bitrader reader structure
+	MOVB br_bits_read, bitReaderShifted_bitsRead(br1)
+	MOVQ br_value, bitReaderShifted_value(br1)
+	MOVQ br_offset, bitReaderShifted_off(br1)
+
+	// const stream = 2
+	// br2.fillFast()
+	MOVBQZX bitReaderShifted_bitsRead(br2), br_bits_read
+	MOVQ    bitReaderShifted_value(br2), br_value
+	MOVQ    bitReaderShifted_off(br2), br_offset
+
+	// if b.bitsRead >= 32 {
+	CMPQ br_bits_read, $32
+	JB   skip_fill2
+
+	SUBQ $32, br_bits_read // b.bitsRead -= 32
+	SUBQ $4, br_offset     // b.off -= 4
+
+	// v := b.in[b.off-4 : b.off]
+	// v = v[:4]
+	// low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	MOVQ bitReaderShifted_in(br2), AX
+	MOVL 0(br_offset)(AX*1), AX       // AX = uint32(b.in[b.off:b.off+4])
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+	MOVQ br_bits_read, CX
+	SHLQ CL, AX
+	ORQ  AX, br_value
+
+	// exhausted = exhausted || (br2.off < 4)
+	CMPQ  br_offset, $4
+	SETLT DL
+	ORB   DL, DH
+
+	// }
+skip_fill2:
+
+	// val0 := br2.peekTopBits(peekBits)
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v0 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br2.advance(uint8(v0.entry))
+	MOVB    AH, BL           // BL = uint8(v0.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// val1 := br2.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v1 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br2.advance(uint8(v1.entry))
+	MOVB    AH, BH           // BH = uint8(v1.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CX, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off] = uint8(v0.entry >> 8)
+	// buf[stream][off+1] = uint8(v1.entry >> 8)
+	MOVW BX, 512(buffer)(off*1)
+
+	// SECOND PART:
+	// val2 := br2.peekTopBits(peekBits)
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v2 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br2.advance(uint8(v0.entry))
+	MOVB    AH, BL           // BL = uint8(v0.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// val3 := br2.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v3 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br2.advance(uint8(v1.entry))
+	MOVB    AH, BH           // BH = uint8(v1.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CX, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off+2] = uint8(v2.entry >> 8)
+	// buf[stream][off+3] = uint8(v3.entry >> 8)
+	MOVW BX, 512+2(buffer)(off*1)
+
+	// update the bitrader reader structure
+	MOVB br_bits_read, bitReaderShifted_bitsRead(br2)
+	MOVQ br_value, bitReaderShifted_value(br2)
+	MOVQ br_offset, bitReaderShifted_off(br2)
+
+	// const stream = 3
+	// br3.fillFast()
+	MOVBQZX bitReaderShifted_bitsRead(br3), br_bits_read
+	MOVQ    bitReaderShifted_value(br3), br_value
+	MOVQ    bitReaderShifted_off(br3), br_offset
+
+	// if b.bitsRead >= 32 {
+	CMPQ br_bits_read, $32
+	JB   skip_fill3
+
+	SUBQ $32, br_bits_read // b.bitsRead -= 32
+	SUBQ $4, br_offset     // b.off -= 4
+
+	// v := b.in[b.off-4 : b.off]
+	// v = v[:4]
+	// low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	MOVQ bitReaderShifted_in(br3), AX
+	MOVL 0(br_offset)(AX*1), AX       // AX = uint32(b.in[b.off:b.off+4])
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+	MOVQ br_bits_read, CX
+	SHLQ CL, AX
+	ORQ  AX, br_value
+
+	// exhausted = exhausted || (br3.off < 4)
+	CMPQ  br_offset, $4
+	SETLT DL
+	ORB   DL, DH
+
+	// }
+skip_fill3:
+
+	// val0 := br3.peekTopBits(peekBits)
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v0 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br3.advance(uint8(v0.entry))
+	MOVB    AH, BL           // BL = uint8(v0.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// val1 := br3.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v1 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br3.advance(uint8(v1.entry))
+	MOVB    AH, BH           // BH = uint8(v1.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CX, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off] = uint8(v0.entry >> 8)
+	// buf[stream][off+1] = uint8(v1.entry >> 8)
+	MOVW BX, 768(buffer)(off*1)
+
+	// SECOND PART:
+	// val2 := br3.peekTopBits(peekBits)
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v2 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br3.advance(uint8(v0.entry))
+	MOVB    AH, BL           // BL = uint8(v0.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// val3 := br3.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+	// v3 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br3.advance(uint8(v1.entry))
+	MOVB    AH, BH           // BH = uint8(v1.entry >> 8)
+	MOVBQZX AL, CX
+	SHLQ    CX, br_value     // value <<= n
+	ADDQ    CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off+2] = uint8(v2.entry >> 8)
+	// buf[stream][off+3] = uint8(v3.entry >> 8)
+	MOVW BX, 768+2(buffer)(off*1)
+
+	// update the bitrader reader structure
+	MOVB br_bits_read, bitReaderShifted_bitsRead(br3)
+	MOVQ br_value, bitReaderShifted_value(br3)
+	MOVQ br_offset, bitReaderShifted_off(br3)
+
+	ADDQ $4, off // off += 2
+
+	TESTB DH, DH // any br[i].ofs < 4?
+	JNZ   end
+
+	CMPQ off, $bufoff
+	JL   main_loop
+
+end:
+	MOVQ 0(SP), BP
+
+	MOVB off, ret+56(FP)
+	RET
+
+#undef off
+#undef buffer
+#undef table
+
+#undef br_bits_read
+#undef br_value
+#undef br_offset
+#undef peek_bits
+#undef exhausted
+
+#undef br0
+#undef br1
+#undef br2
+#undef br3
diff --git a/vendor/github.com/klauspost/compress/huff0/decompress_8b_amd64.s.in b/vendor/github.com/klauspost/compress/huff0/decompress_8b_amd64.s.in
new file mode 100644
index 00000000..6d477a2c
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/huff0/decompress_8b_amd64.s.in
@@ -0,0 +1,197 @@
+// +build !appengine
+// +build gc
+// +build !noasm
+
+#include "textflag.h"
+#include "funcdata.h"
+#include "go_asm.h"
+
+
+#define bufoff      256     // see decompress.go, we're using [4][256]byte table
+
+//func decompress4x_main_loop_x86(pbr0, pbr1, pbr2, pbr3 *bitReaderShifted,
+//	peekBits uint8, buf *byte, tbl *dEntrySingle) (int, bool)
+TEXT ·decompress4x_8b_loop_x86(SB), NOSPLIT, $8
+#define off             R8
+#define buffer          DI
+#define table           SI
+
+#define br_bits_read    R9
+#define br_value        R10
+#define br_offset       R11
+#define peek_bits       R12
+#define exhausted       DX
+
+#define br0             R13
+#define br1             R14
+#define br2             R15
+#define br3             BP
+
+    MOVQ    BP, 0(SP)
+
+    XORQ    exhausted, exhausted    // exhausted = false
+    XORQ    off, off                // off = 0
+
+    MOVBQZX peekBits+32(FP), peek_bits
+    MOVQ    buf+40(FP), buffer
+    MOVQ    tbl+48(FP), table
+
+    MOVQ    pbr0+0(FP), br0
+    MOVQ    pbr1+8(FP), br1
+    MOVQ    pbr2+16(FP), br2
+    MOVQ    pbr3+24(FP), br3
+
+main_loop:
+{{ define "decode_2_values_x86" }}
+    // const stream = {{ var "id" }}
+    // br{{ var "id"}}.fillFast()
+    MOVBQZX bitReaderShifted_bitsRead(br{{ var "id" }}), br_bits_read
+    MOVQ    bitReaderShifted_value(br{{ var "id" }}), br_value
+    MOVQ    bitReaderShifted_off(br{{ var "id" }}), br_offset
+
+	// if b.bitsRead >= 32 {
+    CMPQ    br_bits_read, $32
+    JB      skip_fill{{ var "id" }}
+
+    SUBQ    $32, br_bits_read       // b.bitsRead -= 32
+    SUBQ    $4, br_offset           // b.off -= 4
+
+	// v := b.in[b.off-4 : b.off]
+	// v = v[:4]
+	// low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+    MOVQ    bitReaderShifted_in(br{{ var "id" }}), AX
+    MOVL    0(br_offset)(AX*1), AX  // AX = uint32(b.in[b.off:b.off+4])
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+    MOVQ    br_bits_read, CX
+    SHLQ    CL, AX
+    ORQ     AX, br_value
+
+    // exhausted = exhausted || (br{{ var "id"}}.off < 4)
+    CMPQ    br_offset, $4
+    SETLT   DL
+    ORB     DL, DH
+    // }
+skip_fill{{ var "id" }}:
+
+    // val0 := br{{ var "id"}}.peekTopBits(peekBits)
+    MOVQ    br_value, AX
+    MOVQ    peek_bits, CX
+    SHRQ    CL, AX                  // AX = (value >> peek_bits) & mask
+
+    // v0 := table[val0&mask]
+    MOVW    0(table)(AX*2), AX      // AX - v0
+
+    // br{{ var "id"}}.advance(uint8(v0.entry))
+    MOVB    AH, BL                  // BL = uint8(v0.entry >> 8)
+    MOVBQZX AL, CX
+    SHLQ    CL, br_value            // value <<= n
+    ADDQ    CX, br_bits_read        // bits_read += n
+
+    // val1 := br{{ var "id"}}.peekTopBits(peekBits)
+    MOVQ    peek_bits, CX
+    MOVQ    br_value, AX
+    SHRQ    CL, AX                  // AX = (value >> peek_bits) & mask
+
+    // v1 := table[val1&mask]
+    MOVW    0(table)(AX*2), AX      // AX - v1
+
+    // br{{ var "id"}}.advance(uint8(v1.entry))
+    MOVB    AH, BH                  // BH = uint8(v1.entry >> 8)
+    MOVBQZX AL, CX
+    SHLQ    CX, br_value            // value <<= n
+    ADDQ    CX, br_bits_read        // bits_read += n
+
+
+    // these two writes get coalesced
+    // buf[stream][off] = uint8(v0.entry >> 8)
+    // buf[stream][off+1] = uint8(v1.entry >> 8)
+    MOVW    BX, {{ var "bufofs" }}(buffer)(off*1)
+
+    // SECOND PART:
+    // val2 := br{{ var "id"}}.peekTopBits(peekBits)
+    MOVQ    br_value, AX
+    MOVQ    peek_bits, CX
+    SHRQ    CL, AX                  // AX = (value >> peek_bits) & mask
+
+    // v2 := table[val0&mask]
+    MOVW    0(table)(AX*2), AX      // AX - v0
+
+    // br{{ var "id"}}.advance(uint8(v0.entry))
+    MOVB    AH, BL                  // BL = uint8(v0.entry >> 8)
+    MOVBQZX AL, CX
+    SHLQ    CL, br_value            // value <<= n
+    ADDQ    CX, br_bits_read        // bits_read += n
+
+    // val3 := br{{ var "id"}}.peekTopBits(peekBits)
+    MOVQ    peek_bits, CX
+    MOVQ    br_value, AX
+    SHRQ    CL, AX                  // AX = (value >> peek_bits) & mask
+
+    // v3 := table[val1&mask]
+    MOVW    0(table)(AX*2), AX      // AX - v1
+
+    // br{{ var "id"}}.advance(uint8(v1.entry))
+    MOVB    AH, BH                  // BH = uint8(v1.entry >> 8)
+    MOVBQZX AL, CX
+    SHLQ    CX, br_value            // value <<= n
+    ADDQ    CX, br_bits_read        // bits_read += n
+
+
+    // these two writes get coalesced
+    // buf[stream][off+2] = uint8(v2.entry >> 8)
+    // buf[stream][off+3] = uint8(v3.entry >> 8)
+    MOVW    BX, {{ var "bufofs" }}+2(buffer)(off*1)
+
+    // update the bitrader reader structure
+    MOVB    br_bits_read, bitReaderShifted_bitsRead(br{{ var "id" }})
+    MOVQ    br_value, bitReaderShifted_value(br{{ var "id" }})
+    MOVQ    br_offset, bitReaderShifted_off(br{{ var "id" }})
+{{ end }}
+
+    {{ set "id" "0" }}
+    {{ set "ofs" "0" }}
+    {{ set "bufofs" "0" }} {{/* id * bufoff */}}
+    {{ template "decode_2_values_x86" . }}
+
+    {{ set "id" "1" }}
+    {{ set "ofs" "8" }}
+    {{ set "bufofs" "256" }}
+    {{ template "decode_2_values_x86" . }}
+
+    {{ set "id" "2" }}
+    {{ set "ofs" "16" }}
+    {{ set "bufofs" "512" }}
+    {{ template "decode_2_values_x86" . }}
+
+    {{ set "id" "3" }}
+    {{ set "ofs" "24" }}
+    {{ set "bufofs" "768" }}
+    {{ template "decode_2_values_x86" . }}
+
+    ADDQ    $4, off     // off += 2
+
+    TESTB   DH, DH      // any br[i].ofs < 4?
+    JNZ     end
+
+    CMPQ    off, $bufoff
+    JL      main_loop
+end:
+    MOVQ    0(SP), BP
+
+    MOVB    off, ret+56(FP)
+    RET
+#undef  off
+#undef  buffer
+#undef  table
+
+#undef  br_bits_read
+#undef  br_value
+#undef  br_offset
+#undef  peek_bits
+#undef  exhausted
+
+#undef  br0
+#undef  br1
+#undef  br2
+#undef  br3
diff --git a/vendor/github.com/klauspost/compress/huff0/decompress_amd64.go b/vendor/github.com/klauspost/compress/huff0/decompress_amd64.go
new file mode 100644
index 00000000..d47f6644
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/huff0/decompress_amd64.go
@@ -0,0 +1,181 @@
+//go:build amd64 && !appengine && !noasm && gc
+// +build amd64,!appengine,!noasm,gc
+
+// This file contains the specialisation of Decoder.Decompress4X
+// that uses an asm implementation of its main loop.
+package huff0
+
+import (
+	"errors"
+	"fmt"
+)
+
+// decompress4x_main_loop_x86 is an x86 assembler implementation
+// of Decompress4X when tablelog > 8.
+// go:noescape
+func decompress4x_main_loop_x86(pbr0, pbr1, pbr2, pbr3 *bitReaderShifted,
+	peekBits uint8, buf *byte, tbl *dEntrySingle) uint8
+
+// decompress4x_8b_loop_x86 is an x86 assembler implementation
+// of Decompress4X when tablelog <= 8 which decodes 4 entries
+// per loop.
+// go:noescape
+func decompress4x_8b_loop_x86(pbr0, pbr1, pbr2, pbr3 *bitReaderShifted,
+	peekBits uint8, buf *byte, tbl *dEntrySingle) uint8
+
+// fallback8BitSize is the size where using Go version is faster.
+const fallback8BitSize = 800
+
+// Decompress4X will decompress a 4X encoded stream.
+// The length of the supplied input must match the end of a block exactly.
+// The *capacity* of the dst slice must match the destination size of
+// the uncompressed data exactly.
+func (d *Decoder) Decompress4X(dst, src []byte) ([]byte, error) {
+	if len(d.dt.single) == 0 {
+		return nil, errors.New("no table loaded")
+	}
+	if len(src) < 6+(4*1) {
+		return nil, errors.New("input too small")
+	}
+
+	use8BitTables := d.actualTableLog <= 8
+	if cap(dst) < fallback8BitSize && use8BitTables {
+		return d.decompress4X8bit(dst, src)
+	}
+	var br [4]bitReaderShifted
+	// Decode "jump table"
+	start := 6
+	for i := 0; i < 3; i++ {
+		length := int(src[i*2]) | (int(src[i*2+1]) << 8)
+		if start+length >= len(src) {
+			return nil, errors.New("truncated input (or invalid offset)")
+		}
+		err := br[i].init(src[start : start+length])
+		if err != nil {
+			return nil, err
+		}
+		start += length
+	}
+	err := br[3].init(src[start:])
+	if err != nil {
+		return nil, err
+	}
+
+	// destination, offset to match first output
+	dstSize := cap(dst)
+	dst = dst[:dstSize]
+	out := dst
+	dstEvery := (dstSize + 3) / 4
+
+	const tlSize = 1 << tableLogMax
+	const tlMask = tlSize - 1
+	single := d.dt.single[:tlSize]
+
+	// Use temp table to avoid bound checks/append penalty.
+	buf := d.buffer()
+	var off uint8
+	var decoded int
+
+	const debug = false
+
+	// see: bitReaderShifted.peekBitsFast()
+	peekBits := uint8((64 - d.actualTableLog) & 63)
+
+	// Decode 2 values from each decoder/loop.
+	const bufoff = 256
+	for {
+		if br[0].off < 4 || br[1].off < 4 || br[2].off < 4 || br[3].off < 4 {
+			break
+		}
+
+		if use8BitTables {
+			off = decompress4x_8b_loop_x86(&br[0], &br[1], &br[2], &br[3], peekBits, &buf[0][0], &single[0])
+		} else {
+			off = decompress4x_main_loop_x86(&br[0], &br[1], &br[2], &br[3], peekBits, &buf[0][0], &single[0])
+		}
+		if debug {
+			fmt.Print("DEBUG: ")
+			fmt.Printf("off=%d,", off)
+			for i := 0; i < 4; i++ {
+				fmt.Printf(" br[%d]={bitsRead=%d, value=%x, off=%d}",
+					i, br[i].bitsRead, br[i].value, br[i].off)
+			}
+			fmt.Println("")
+		}
+
+		if off != 0 {
+			break
+		}
+
+		if bufoff > dstEvery {
+			d.bufs.Put(buf)
+			return nil, errors.New("corruption detected: stream overrun 1")
+		}
+		copy(out, buf[0][:])
+		copy(out[dstEvery:], buf[1][:])
+		copy(out[dstEvery*2:], buf[2][:])
+		copy(out[dstEvery*3:], buf[3][:])
+		out = out[bufoff:]
+		decoded += bufoff * 4
+		// There must at least be 3 buffers left.
+		if len(out) < dstEvery*3 {
+			d.bufs.Put(buf)
+			return nil, errors.New("corruption detected: stream overrun 2")
+		}
+	}
+	if off > 0 {
+		ioff := int(off)
+		if len(out) < dstEvery*3+ioff {
+			d.bufs.Put(buf)
+			return nil, errors.New("corruption detected: stream overrun 3")
+		}
+		copy(out, buf[0][:off])
+		copy(out[dstEvery:], buf[1][:off])
+		copy(out[dstEvery*2:], buf[2][:off])
+		copy(out[dstEvery*3:], buf[3][:off])
+		decoded += int(off) * 4
+		out = out[off:]
+	}
+
+	// Decode remaining.
+	remainBytes := dstEvery - (decoded / 4)
+	for i := range br {
+		offset := dstEvery * i
+		endsAt := offset + remainBytes
+		if endsAt > len(out) {
+			endsAt = len(out)
+		}
+		br := &br[i]
+		bitsLeft := br.remaining()
+		for bitsLeft > 0 {
+			br.fill()
+			if offset >= endsAt {
+				d.bufs.Put(buf)
+				return nil, errors.New("corruption detected: stream overrun 4")
+			}
+
+			// Read value and increment offset.
+			val := br.peekBitsFast(d.actualTableLog)
+			v := single[val&tlMask].entry
+			nBits := uint8(v)
+			br.advance(nBits)
+			bitsLeft -= uint(nBits)
+			out[offset] = uint8(v >> 8)
+			offset++
+		}
+		if offset != endsAt {
+			d.bufs.Put(buf)
+			return nil, fmt.Errorf("corruption detected: short output block %d, end %d != %d", i, offset, endsAt)
+		}
+		decoded += offset - dstEvery*i
+		err = br.close()
+		if err != nil {
+			return nil, err
+		}
+	}
+	d.bufs.Put(buf)
+	if dstSize != decoded {
+		return nil, errors.New("corruption detected: short output block")
+	}
+	return dst, nil
+}
diff --git a/vendor/github.com/klauspost/compress/huff0/decompress_amd64.s b/vendor/github.com/klauspost/compress/huff0/decompress_amd64.s
new file mode 100644
index 00000000..2edad3ea
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/huff0/decompress_amd64.s
@@ -0,0 +1,506 @@
+// +build !appengine
+// +build gc
+// +build !noasm
+
+#include "textflag.h"
+#include "funcdata.h"
+#include "go_asm.h"
+
+#ifdef GOAMD64_v4
+#ifndef GOAMD64_v3
+#define GOAMD64_v3
+#endif
+#endif
+
+#define bufoff      256 // see decompress.go, we're using [4][256]byte table
+
+// func decompress4x_main_loop_x86(pbr0, pbr1, pbr2, pbr3 *bitReaderShifted,
+//	peekBits uint8, buf *byte, tbl *dEntrySingle) (int, bool)
+TEXT ·decompress4x_main_loop_x86(SB), NOSPLIT, $8
+#define off             R8
+#define buffer          DI
+#define table           SI
+
+#define br_bits_read    R9
+#define br_value        R10
+#define br_offset       R11
+#define peek_bits       R12
+#define exhausted       DX
+
+#define br0             R13
+#define br1             R14
+#define br2             R15
+#define br3             BP
+
+	MOVQ BP, 0(SP)
+
+	XORQ exhausted, exhausted // exhausted = false
+	XORQ off, off             // off = 0
+
+	MOVBQZX peekBits+32(FP), peek_bits
+	MOVQ    buf+40(FP), buffer
+	MOVQ    tbl+48(FP), table
+
+	MOVQ pbr0+0(FP), br0
+	MOVQ pbr1+8(FP), br1
+	MOVQ pbr2+16(FP), br2
+	MOVQ pbr3+24(FP), br3
+
+main_loop:
+
+	// const stream = 0
+	// br0.fillFast()
+	MOVBQZX bitReaderShifted_bitsRead(br0), br_bits_read
+	MOVQ    bitReaderShifted_value(br0), br_value
+	MOVQ    bitReaderShifted_off(br0), br_offset
+
+	// We must have at least 2 * max tablelog left
+	CMPQ br_bits_read, $64-22
+	JBE  skip_fill0
+
+	SUBQ $32, br_bits_read // b.bitsRead -= 32
+	SUBQ $4, br_offset     // b.off -= 4
+
+	// v := b.in[b.off-4 : b.off]
+	// v = v[:4]
+	// low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	MOVQ bitReaderShifted_in(br0), AX
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+#ifdef GOAMD64_v3
+	SHLXQ br_bits_read, 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4]) << (b.bitsRead & 63)
+
+#else
+	MOVL 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4])
+	MOVQ br_bits_read, CX
+	SHLQ CL, AX
+
+#endif
+
+	ORQ AX, br_value
+
+	// exhausted = exhausted || (br0.off < 4)
+	CMPQ  br_offset, $4
+	SETLT DL
+	ORB   DL, DH
+
+	// }
+skip_fill0:
+
+	// val0 := br0.peekTopBits(peekBits)
+#ifdef GOAMD64_v3
+	SHRXQ peek_bits, br_value, AX // AX = (value >> peek_bits) & mask
+
+#else
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+#endif
+
+	// v0 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br0.advance(uint8(v0.entry))
+	MOVB AH, BL // BL = uint8(v0.entry >> 8)
+
+#ifdef GOAMD64_v3
+	MOVBQZX AL, CX
+	SHLXQ   AX, br_value, br_value // value <<= n
+
+#else
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value // value <<= n
+
+#endif
+
+	ADDQ CX, br_bits_read // bits_read += n
+
+#ifdef GOAMD64_v3
+	SHRXQ peek_bits, br_value, AX // AX = (value >> peek_bits) & mask
+
+#else
+	// val1 := br0.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+#endif
+
+	// v1 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br0.advance(uint8(v1.entry))
+	MOVB AH, BH // BH = uint8(v1.entry >> 8)
+
+#ifdef GOAMD64_v3
+	MOVBQZX AL, CX
+	SHLXQ   AX, br_value, br_value // value <<= n
+
+#else
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value // value <<= n
+
+#endif
+
+	ADDQ CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off] = uint8(v0.entry >> 8)
+	// buf[stream][off+1] = uint8(v1.entry >> 8)
+	MOVW BX, 0(buffer)(off*1)
+
+	// update the bitrader reader structure
+	MOVB br_bits_read, bitReaderShifted_bitsRead(br0)
+	MOVQ br_value, bitReaderShifted_value(br0)
+	MOVQ br_offset, bitReaderShifted_off(br0)
+
+	// const stream = 1
+	// br1.fillFast()
+	MOVBQZX bitReaderShifted_bitsRead(br1), br_bits_read
+	MOVQ    bitReaderShifted_value(br1), br_value
+	MOVQ    bitReaderShifted_off(br1), br_offset
+
+	// We must have at least 2 * max tablelog left
+	CMPQ br_bits_read, $64-22
+	JBE  skip_fill1
+
+	SUBQ $32, br_bits_read // b.bitsRead -= 32
+	SUBQ $4, br_offset     // b.off -= 4
+
+	// v := b.in[b.off-4 : b.off]
+	// v = v[:4]
+	// low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	MOVQ bitReaderShifted_in(br1), AX
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+#ifdef GOAMD64_v3
+	SHLXQ br_bits_read, 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4]) << (b.bitsRead & 63)
+
+#else
+	MOVL 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4])
+	MOVQ br_bits_read, CX
+	SHLQ CL, AX
+
+#endif
+
+	ORQ AX, br_value
+
+	// exhausted = exhausted || (br1.off < 4)
+	CMPQ  br_offset, $4
+	SETLT DL
+	ORB   DL, DH
+
+	// }
+skip_fill1:
+
+	// val0 := br1.peekTopBits(peekBits)
+#ifdef GOAMD64_v3
+	SHRXQ peek_bits, br_value, AX // AX = (value >> peek_bits) & mask
+
+#else
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+#endif
+
+	// v0 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br1.advance(uint8(v0.entry))
+	MOVB AH, BL // BL = uint8(v0.entry >> 8)
+
+#ifdef GOAMD64_v3
+	MOVBQZX AL, CX
+	SHLXQ   AX, br_value, br_value // value <<= n
+
+#else
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value // value <<= n
+
+#endif
+
+	ADDQ CX, br_bits_read // bits_read += n
+
+#ifdef GOAMD64_v3
+	SHRXQ peek_bits, br_value, AX // AX = (value >> peek_bits) & mask
+
+#else
+	// val1 := br1.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+#endif
+
+	// v1 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br1.advance(uint8(v1.entry))
+	MOVB AH, BH // BH = uint8(v1.entry >> 8)
+
+#ifdef GOAMD64_v3
+	MOVBQZX AL, CX
+	SHLXQ   AX, br_value, br_value // value <<= n
+
+#else
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value // value <<= n
+
+#endif
+
+	ADDQ CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off] = uint8(v0.entry >> 8)
+	// buf[stream][off+1] = uint8(v1.entry >> 8)
+	MOVW BX, 256(buffer)(off*1)
+
+	// update the bitrader reader structure
+	MOVB br_bits_read, bitReaderShifted_bitsRead(br1)
+	MOVQ br_value, bitReaderShifted_value(br1)
+	MOVQ br_offset, bitReaderShifted_off(br1)
+
+	// const stream = 2
+	// br2.fillFast()
+	MOVBQZX bitReaderShifted_bitsRead(br2), br_bits_read
+	MOVQ    bitReaderShifted_value(br2), br_value
+	MOVQ    bitReaderShifted_off(br2), br_offset
+
+	// We must have at least 2 * max tablelog left
+	CMPQ br_bits_read, $64-22
+	JBE  skip_fill2
+
+	SUBQ $32, br_bits_read // b.bitsRead -= 32
+	SUBQ $4, br_offset     // b.off -= 4
+
+	// v := b.in[b.off-4 : b.off]
+	// v = v[:4]
+	// low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	MOVQ bitReaderShifted_in(br2), AX
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+#ifdef GOAMD64_v3
+	SHLXQ br_bits_read, 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4]) << (b.bitsRead & 63)
+
+#else
+	MOVL 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4])
+	MOVQ br_bits_read, CX
+	SHLQ CL, AX
+
+#endif
+
+	ORQ AX, br_value
+
+	// exhausted = exhausted || (br2.off < 4)
+	CMPQ  br_offset, $4
+	SETLT DL
+	ORB   DL, DH
+
+	// }
+skip_fill2:
+
+	// val0 := br2.peekTopBits(peekBits)
+#ifdef GOAMD64_v3
+	SHRXQ peek_bits, br_value, AX // AX = (value >> peek_bits) & mask
+
+#else
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+#endif
+
+	// v0 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br2.advance(uint8(v0.entry))
+	MOVB AH, BL // BL = uint8(v0.entry >> 8)
+
+#ifdef GOAMD64_v3
+	MOVBQZX AL, CX
+	SHLXQ   AX, br_value, br_value // value <<= n
+
+#else
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value // value <<= n
+
+#endif
+
+	ADDQ CX, br_bits_read // bits_read += n
+
+#ifdef GOAMD64_v3
+	SHRXQ peek_bits, br_value, AX // AX = (value >> peek_bits) & mask
+
+#else
+	// val1 := br2.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+#endif
+
+	// v1 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br2.advance(uint8(v1.entry))
+	MOVB AH, BH // BH = uint8(v1.entry >> 8)
+
+#ifdef GOAMD64_v3
+	MOVBQZX AL, CX
+	SHLXQ   AX, br_value, br_value // value <<= n
+
+#else
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value // value <<= n
+
+#endif
+
+	ADDQ CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off] = uint8(v0.entry >> 8)
+	// buf[stream][off+1] = uint8(v1.entry >> 8)
+	MOVW BX, 512(buffer)(off*1)
+
+	// update the bitrader reader structure
+	MOVB br_bits_read, bitReaderShifted_bitsRead(br2)
+	MOVQ br_value, bitReaderShifted_value(br2)
+	MOVQ br_offset, bitReaderShifted_off(br2)
+
+	// const stream = 3
+	// br3.fillFast()
+	MOVBQZX bitReaderShifted_bitsRead(br3), br_bits_read
+	MOVQ    bitReaderShifted_value(br3), br_value
+	MOVQ    bitReaderShifted_off(br3), br_offset
+
+	// We must have at least 2 * max tablelog left
+	CMPQ br_bits_read, $64-22
+	JBE  skip_fill3
+
+	SUBQ $32, br_bits_read // b.bitsRead -= 32
+	SUBQ $4, br_offset     // b.off -= 4
+
+	// v := b.in[b.off-4 : b.off]
+	// v = v[:4]
+	// low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+	MOVQ bitReaderShifted_in(br3), AX
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+#ifdef GOAMD64_v3
+	SHLXQ br_bits_read, 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4]) << (b.bitsRead & 63)
+
+#else
+	MOVL 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4])
+	MOVQ br_bits_read, CX
+	SHLQ CL, AX
+
+#endif
+
+	ORQ AX, br_value
+
+	// exhausted = exhausted || (br3.off < 4)
+	CMPQ  br_offset, $4
+	SETLT DL
+	ORB   DL, DH
+
+	// }
+skip_fill3:
+
+	// val0 := br3.peekTopBits(peekBits)
+#ifdef GOAMD64_v3
+	SHRXQ peek_bits, br_value, AX // AX = (value >> peek_bits) & mask
+
+#else
+	MOVQ br_value, AX
+	MOVQ peek_bits, CX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+#endif
+
+	// v0 := table[val0&mask]
+	MOVW 0(table)(AX*2), AX // AX - v0
+
+	// br3.advance(uint8(v0.entry))
+	MOVB AH, BL // BL = uint8(v0.entry >> 8)
+
+#ifdef GOAMD64_v3
+	MOVBQZX AL, CX
+	SHLXQ   AX, br_value, br_value // value <<= n
+
+#else
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value // value <<= n
+
+#endif
+
+	ADDQ CX, br_bits_read // bits_read += n
+
+#ifdef GOAMD64_v3
+	SHRXQ peek_bits, br_value, AX // AX = (value >> peek_bits) & mask
+
+#else
+	// val1 := br3.peekTopBits(peekBits)
+	MOVQ peek_bits, CX
+	MOVQ br_value, AX
+	SHRQ CL, AX        // AX = (value >> peek_bits) & mask
+
+#endif
+
+	// v1 := table[val1&mask]
+	MOVW 0(table)(AX*2), AX // AX - v1
+
+	// br3.advance(uint8(v1.entry))
+	MOVB AH, BH // BH = uint8(v1.entry >> 8)
+
+#ifdef GOAMD64_v3
+	MOVBQZX AL, CX
+	SHLXQ   AX, br_value, br_value // value <<= n
+
+#else
+	MOVBQZX AL, CX
+	SHLQ    CL, br_value // value <<= n
+
+#endif
+
+	ADDQ CX, br_bits_read // bits_read += n
+
+	// these two writes get coalesced
+	// buf[stream][off] = uint8(v0.entry >> 8)
+	// buf[stream][off+1] = uint8(v1.entry >> 8)
+	MOVW BX, 768(buffer)(off*1)
+
+	// update the bitrader reader structure
+	MOVB br_bits_read, bitReaderShifted_bitsRead(br3)
+	MOVQ br_value, bitReaderShifted_value(br3)
+	MOVQ br_offset, bitReaderShifted_off(br3)
+
+	ADDQ $2, off // off += 2
+
+	TESTB DH, DH // any br[i].ofs < 4?
+	JNZ   end
+
+	CMPQ off, $bufoff
+	JL   main_loop
+
+end:
+	MOVQ 0(SP), BP
+
+	MOVB off, ret+56(FP)
+	RET
+
+#undef off
+#undef buffer
+#undef table
+
+#undef br_bits_read
+#undef br_value
+#undef br_offset
+#undef peek_bits
+#undef exhausted
+
+#undef br0
+#undef br1
+#undef br2
+#undef br3
diff --git a/vendor/github.com/klauspost/compress/huff0/decompress_amd64.s.in b/vendor/github.com/klauspost/compress/huff0/decompress_amd64.s.in
new file mode 100644
index 00000000..330d86ae
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/huff0/decompress_amd64.s.in
@@ -0,0 +1,195 @@
+// +build !appengine
+// +build gc
+// +build !noasm
+
+#include "textflag.h"
+#include "funcdata.h"
+#include "go_asm.h"
+
+#ifdef GOAMD64_v4
+#ifndef GOAMD64_v3
+#define GOAMD64_v3
+#endif
+#endif
+
+#define bufoff      256     // see decompress.go, we're using [4][256]byte table
+
+//func decompress4x_main_loop_x86(pbr0, pbr1, pbr2, pbr3 *bitReaderShifted,
+//	peekBits uint8, buf *byte, tbl *dEntrySingle) (int, bool)
+TEXT ·decompress4x_main_loop_x86(SB), NOSPLIT, $8
+#define off             R8
+#define buffer          DI
+#define table           SI
+
+#define br_bits_read    R9
+#define br_value        R10
+#define br_offset       R11
+#define peek_bits       R12
+#define exhausted       DX
+
+#define br0             R13
+#define br1             R14
+#define br2             R15
+#define br3             BP
+
+    MOVQ    BP, 0(SP)
+
+    XORQ    exhausted, exhausted    // exhausted = false
+    XORQ    off, off                // off = 0
+
+    MOVBQZX peekBits+32(FP), peek_bits
+    MOVQ    buf+40(FP), buffer
+    MOVQ    tbl+48(FP), table
+
+    MOVQ    pbr0+0(FP), br0
+    MOVQ    pbr1+8(FP), br1
+    MOVQ    pbr2+16(FP), br2
+    MOVQ    pbr3+24(FP), br3
+
+main_loop:
+{{ define "decode_2_values_x86" }}
+    // const stream = {{ var "id" }}
+    // br{{ var "id"}}.fillFast()
+    MOVBQZX bitReaderShifted_bitsRead(br{{ var "id" }}), br_bits_read
+    MOVQ    bitReaderShifted_value(br{{ var "id" }}), br_value
+    MOVQ    bitReaderShifted_off(br{{ var "id" }}), br_offset
+
+    // We must have at least 2 * max tablelog left
+    CMPQ    br_bits_read, $64-22
+    JBE     skip_fill{{ var "id" }}
+
+    SUBQ    $32, br_bits_read       // b.bitsRead -= 32
+    SUBQ    $4, br_offset           // b.off -= 4
+
+	// v := b.in[b.off-4 : b.off]
+	// v = v[:4]
+	// low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24)
+    MOVQ    bitReaderShifted_in(br{{ var "id" }}), AX
+
+	// b.value |= uint64(low) << (b.bitsRead & 63)
+#ifdef GOAMD64_v3
+    SHLXQ   br_bits_read, 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4]) << (b.bitsRead & 63)
+#else
+    MOVL    0(br_offset)(AX*1), AX  // AX = uint32(b.in[b.off:b.off+4])
+    MOVQ    br_bits_read, CX
+    SHLQ    CL, AX
+#endif
+
+    ORQ     AX, br_value
+
+    // exhausted = exhausted || (br{{ var "id"}}.off < 4)
+    CMPQ    br_offset, $4
+    SETLT   DL
+    ORB     DL, DH
+    // }
+skip_fill{{ var "id" }}:
+
+    // val0 := br{{ var "id"}}.peekTopBits(peekBits)
+#ifdef GOAMD64_v3
+    SHRXQ   peek_bits, br_value, AX // AX = (value >> peek_bits) & mask
+#else
+    MOVQ    br_value, AX
+    MOVQ    peek_bits, CX
+    SHRQ    CL, AX                  // AX = (value >> peek_bits) & mask
+#endif
+
+    // v0 := table[val0&mask]
+    MOVW    0(table)(AX*2), AX      // AX - v0
+
+    // br{{ var "id"}}.advance(uint8(v0.entry))
+    MOVB    AH, BL                  // BL = uint8(v0.entry >> 8)
+
+#ifdef GOAMD64_v3
+    MOVBQZX AL, CX
+    SHLXQ   AX, br_value, br_value // value <<= n
+#else
+    MOVBQZX AL, CX
+    SHLQ    CL, br_value            // value <<= n
+#endif
+
+    ADDQ    CX, br_bits_read        // bits_read += n
+
+
+#ifdef GOAMD64_v3
+    SHRXQ    peek_bits, br_value, AX  // AX = (value >> peek_bits) & mask
+#else
+    // val1 := br{{ var "id"}}.peekTopBits(peekBits)
+    MOVQ    peek_bits, CX
+    MOVQ    br_value, AX
+    SHRQ    CL, AX                  // AX = (value >> peek_bits) & mask
+#endif
+
+    // v1 := table[val1&mask]
+    MOVW    0(table)(AX*2), AX      // AX - v1
+
+    // br{{ var "id"}}.advance(uint8(v1.entry))
+    MOVB    AH, BH                  // BH = uint8(v1.entry >> 8)
+
+#ifdef GOAMD64_v3
+    MOVBQZX AL, CX
+    SHLXQ   AX, br_value, br_value // value <<= n
+#else
+    MOVBQZX AL, CX
+    SHLQ    CL, br_value            // value <<= n
+#endif
+
+    ADDQ    CX, br_bits_read        // bits_read += n
+
+
+    // these two writes get coalesced
+    // buf[stream][off] = uint8(v0.entry >> 8)
+    // buf[stream][off+1] = uint8(v1.entry >> 8)
+    MOVW    BX, {{ var "bufofs" }}(buffer)(off*1)
+
+    // update the bitrader reader structure
+    MOVB    br_bits_read, bitReaderShifted_bitsRead(br{{ var "id" }})
+    MOVQ    br_value, bitReaderShifted_value(br{{ var "id" }})
+    MOVQ    br_offset, bitReaderShifted_off(br{{ var "id" }})
+{{ end }}
+
+    {{ set "id" "0" }}
+    {{ set "ofs" "0" }}
+    {{ set "bufofs" "0" }} {{/* id * bufoff */}}
+    {{ template "decode_2_values_x86" . }}
+
+    {{ set "id" "1" }}
+    {{ set "ofs" "8" }}
+    {{ set "bufofs" "256" }}
+    {{ template "decode_2_values_x86" . }}
+
+    {{ set "id" "2" }}
+    {{ set "ofs" "16" }}
+    {{ set "bufofs" "512" }}
+    {{ template "decode_2_values_x86" . }}
+
+    {{ set "id" "3" }}
+    {{ set "ofs" "24" }}
+    {{ set "bufofs" "768" }}
+    {{ template "decode_2_values_x86" . }}
+
+    ADDQ    $2, off     // off += 2
+
+    TESTB   DH, DH      // any br[i].ofs < 4?
+    JNZ     end
+
+    CMPQ    off, $bufoff
+    JL      main_loop
+end:
+    MOVQ    0(SP), BP
+
+    MOVB    off, ret+56(FP)
+    RET
+#undef  off
+#undef  buffer
+#undef  table
+
+#undef  br_bits_read
+#undef  br_value
+#undef  br_offset
+#undef  peek_bits
+#undef  exhausted
+
+#undef  br0
+#undef  br1
+#undef  br2
+#undef  br3
diff --git a/vendor/github.com/klauspost/compress/huff0/decompress_generic.go b/vendor/github.com/klauspost/compress/huff0/decompress_generic.go
new file mode 100644
index 00000000..126b4d68
--- /dev/null
+++ b/vendor/github.com/klauspost/compress/huff0/decompress_generic.go
@@ -0,0 +1,193 @@
+//go:build !amd64 || appengine || !gc || noasm
+// +build !amd64 appengine !gc noasm
+
+// This file contains a generic implementation of Decoder.Decompress4X.
+package huff0
+
+import (
+	"errors"
+	"fmt"
+)
+
+// Decompress4X will decompress a 4X encoded stream.
+// The length of the supplied input must match the end of a block exactly.
+// The *capacity* of the dst slice must match the destination size of
+// the uncompressed data exactly.
+func (d *Decoder) Decompress4X(dst, src []byte) ([]byte, error) {
+	if len(d.dt.single) == 0 {
+		return nil, errors.New("no table loaded")
+	}
+	if len(src) < 6+(4*1) {
+		return nil, errors.New("input too small")
+	}
+	if use8BitTables && d.actualTableLog <= 8 {
+		return d.decompress4X8bit(dst, src)
+	}
+
+	var br [4]bitReaderShifted
+	// Decode "jump table"
+	start := 6
+	for i := 0; i < 3; i++ {
+		length := int(src[i*2]) | (int(src[i*2+1]) << 8)
+		if start+length >= len(src) {
+			return nil, errors.New("truncated input (or invalid offset)")
+		}
+		err := br[i].init(src[start : start+length])
+		if err != nil {
+			return nil, err
+		}
+		start += length
+	}
+	err := br[3].init(src[start:])
+	if err != nil {
+		return nil, err
+	}
+
+	// destination, offset to match first output
+	dstSize := cap(dst)
+	dst = dst[:dstSize]
+	out := dst
+	dstEvery := (dstSize + 3) / 4
+
+	const tlSize = 1 << tableLogMax
+	const tlMask = tlSize - 1
+	single := d.dt.single[:tlSize]
+
+	// Use temp table to avoid bound checks/append penalty.
+	buf := d.buffer()
+	var off uint8
+	var decoded int
+
+	// Decode 2 values from each decoder/loop.
+	const bufoff = 256
+	for {
+		if br[0].off < 4 || br[1].off < 4 || br[2].off < 4 || br[3].off < 4 {
+			break
+		}
+
+		{
+			const stream = 0
+			const stream2 = 1
+			br[stream].fillFast()
+			br[stream2].fillFast()
+
+			val := br[stream].peekBitsFast(d.actualTableLog)
+			val2 := br[stream2].peekBitsFast(d.actualTableLog)
+			v := single[val&tlMask]
+			v2 := single[val2&tlMask]
+			br[stream].advance(uint8(v.entry))
+			br[stream2].advance(uint8(v2.entry))
+			buf[stream][off] = uint8(v.entry >> 8)
+			buf[stream2][off] = uint8(v2.entry >> 8)
+
+			val = br[stream].peekBitsFast(d.actualTableLog)
+			val2 = br[stream2].peekBitsFast(d.actualTableLog)
+			v = single[val&tlMask]
+			v2 = single[val2&tlMask]
+			br[stream].advance(uint8(v.entry))
+			br[stream2].advance(uint8(v2.entry))
+			buf[stream][off+1] = uint8(v.entry >> 8)
+			buf[stream2][off+1] = uint8(v2.entry >> 8)
+		}
+
+		{
+			const stream = 2
+			const stream2 = 3
+			br[stream].fillFast()
+			br[stream2].fillFast()
+
+			val := br[stream].peekBitsFast(d.actualTableLog)
+			val2 := br[stream2].peekBitsFast(d.actualTableLog)
+			v := single[val&tlMask]
+			v2 := single[val2&tlMask]
+			br[stream].advance(uint8(v.entry))
+			br[stream2].advance(uint8(v2.entry))
+			buf[stream][off] = uint8(v.entry >> 8)
+			buf[stream2][off] = uint8(v2.entry >> 8)
+
+			val = br[stream].peekBitsFast(d.actualTableLog)
+			val2 = br[stream2].peekBitsFast(d.actualTableLog)
+			v = single[val&tlMask]
+			v2 = single[val2&tlMask]
+			br[stream].advance(uint8(v.entry))
+			br[stream2].advance(uint8(v2.entry))
+			buf[stream][off+1] = uint8(v.entry >> 8)
+			buf[stream2][off+1] = uint8(v2.entry >> 8)
+		}
+
+		off += 2
+
+		if off == 0 {
+			if bufoff > dstEvery {
+				d.bufs.Put(buf)
+				return nil, errors.New("corruption detected: stream overrun 1")
+			}
+			copy(out, buf[0][:])
+			copy(out[dstEvery:], buf[1][:])
+			copy(out[dstEvery*2:], buf[2][:])
+			copy(out[dstEvery*3:], buf[3][:])
+			out = out[bufoff:]
+			decoded += bufoff * 4
+			// There must at least be 3 buffers left.
+			if len(out) < dstEvery*3 {
+				d.bufs.Put(buf)
+				return nil, errors.New("corruption detected: stream overrun 2")
+			}
+		}
+	}
+	if off > 0 {
+		ioff := int(off)
+		if len(out) < dstEvery*3+ioff {
+			d.bufs.Put(buf)
+			return nil, errors.New("corruption detected: stream overrun 3")
+		}
+		copy(out, buf[0][:off])
+		copy(out[dstEvery:], buf[1][:off])
+		copy(out[dstEvery*2:], buf[2][:off])
+		copy(out[dstEvery*3:], buf[3][:off])
+		decoded += int(off) * 4
+		out = out[off:]
+	}
+
+	// Decode remaining.
+	remainBytes := dstEvery - (decoded / 4)
+	for i := range br {
+		offset := dstEvery * i
+		endsAt := offset + remainBytes
+		if endsAt > len(out) {
+			endsAt = len(out)
+		}
+		br := &br[i]
+		bitsLeft := br.remaining()
+		for bitsLeft > 0 {
+			br.fill()
+			if offset >= endsAt {
+				d.bufs.Put(buf)
+				return nil, errors.New("corruption detected: stream overrun 4")
+			}
+
+			// Read value and increment offset.
+			val := br.peekBitsFast(d.actualTableLog)
+			v := single[val&tlMask].entry
+			nBits := uint8(v)
+			br.advance(nBits)
+			bitsLeft -= uint(nBits)
+			out[offset] = uint8(v >> 8)
+			offset++
+		}
+		if offset != endsAt {
+			d.bufs.Put(buf)
+			return nil, fmt.Errorf("corruption detected: short output block %d, end %d != %d", i, offset, endsAt)
+		}
+		decoded += offset - dstEvery*i
+		err = br.close()
+		if err != nil {
+			return nil, err
+		}
+	}
+	d.bufs.Put(buf)
+	if dstSize != decoded {
+		return nil, errors.New("corruption detected: short output block")
+	}
+	return dst, nil
+}
diff --git a/vendor/github.com/klauspost/compress/huff0/huff0.go b/vendor/github.com/klauspost/compress/huff0/huff0.go
index 3ee00ecb..e8ad17ad 100644
--- a/vendor/github.com/klauspost/compress/huff0/huff0.go
+++ b/vendor/github.com/klauspost/compress/huff0/huff0.go
@@ -8,6 +8,7 @@ import (
 	"fmt"
 	"math"
 	"math/bits"
+	"sync"
 
 	"github.com/klauspost/compress/fse"
 )
@@ -116,6 +117,7 @@ type Scratch struct {
 	nodes          []nodeElt
 	tmpOut         [4][]byte
 	fse            *fse.Scratch
+	decPool        sync.Pool // *[4][256]byte buffers.
 	huffWeight     [maxSymbolValue + 1]byte
 }