diff --git a/huff0/decompress.go b/huff0/decompress.go
index 9b7cc8e97b..f3db4a3b55 100644
--- a/huff0/decompress.go
+++ b/huff0/decompress.go
@@ -20,7 +20,7 @@ type dEntrySingle struct {
 
 // double-symbols decoding
 type dEntryDouble struct {
-	seq   uint16
+	seq   [4]byte
 	nBits uint8
 	len   uint8
 }
@@ -914,7 +914,7 @@ func (d *Decoder) decompress4X8bit(dst, src []byte) ([]byte, error) {
 	out := dst
 	dstEvery := (dstSize + 3) / 4
 
-	shift := (8 - d.actualTableLog) & 7
+	shift := (56 + (8 - d.actualTableLog)) & 63
 
 	const tlSize = 1 << 8
 	single := d.dt.single[:tlSize]
@@ -935,79 +935,91 @@ func (d *Decoder) decompress4X8bit(dst, src []byte) ([]byte, error) {
 			// Interleave 2 decodes.
 			const stream = 0
 			const stream2 = 1
-			br[stream].fillFast()
-			br[stream2].fillFast()
-
-			v := single[br[stream].peekByteFast()>>shift].entry
+			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[off+bufoff*stream] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
-
-			v2 := single[br[stream2].peekByteFast()>>shift].entry
 			buf[off+bufoff*stream2] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
 
-			v = single[br[stream].peekByteFast()>>shift].entry
+			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[off+bufoff*stream+1] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
-
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
 			buf[off+bufoff*stream2+1] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
 
-			v = single[br[stream].peekByteFast()>>shift].entry
+			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[off+bufoff*stream+2] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
-
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
 			buf[off+bufoff*stream2+2] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
 
-			v = single[br[stream].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream+3] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
-
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
+			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[off+bufoff*stream2+3] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
+			buf[off+bufoff*stream+3] = uint8(v >> 8)
 		}
 
 		{
 			const stream = 2
 			const stream2 = 3
-			br[stream].fillFast()
-			br[stream2].fillFast()
-
-			v := single[br[stream].peekByteFast()>>shift].entry
+			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[off+bufoff*stream] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
-
-			v2 := single[br[stream2].peekByteFast()>>shift].entry
 			buf[off+bufoff*stream2] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
 
-			v = single[br[stream].peekByteFast()>>shift].entry
+			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[off+bufoff*stream+1] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
-
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
 			buf[off+bufoff*stream2+1] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
 
-			v = single[br[stream].peekByteFast()>>shift].entry
+			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[off+bufoff*stream+2] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
-
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
 			buf[off+bufoff*stream2+2] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
-
-			v = single[br[stream].peekByteFast()>>shift].entry
-			buf[off+bufoff*stream+3] = uint8(v >> 8)
-			br[stream].advance(uint8(v))
 
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
+			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[off+bufoff*stream2+3] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
+			buf[off+bufoff*stream+3] = uint8(v >> 8)
 		}
 
 		off += 4
@@ -1073,7 +1085,7 @@ func (d *Decoder) decompress4X8bit(dst, src []byte) ([]byte, error) {
 			}
 
 			// Read value and increment offset.
-			v := single[br.peekByteFast()>>shift].entry
+			v := single[uint8(br.value>>shift)].entry
 			nBits := uint8(v)
 			br.advance(nBits)
 			bitsLeft -= int(nBits)
@@ -1121,7 +1133,7 @@ func (d *Decoder) decompress4X8bitExactly(dst, src []byte) ([]byte, error) {
 	out := dst
 	dstEvery := (dstSize + 3) / 4
 
-	const shift = 0
+	const shift = 56
 	const tlSize = 1 << 8
 	const tlMask = tlSize - 1
 	single := d.dt.single[:tlSize]
@@ -1145,37 +1157,41 @@ func (d *Decoder) decompress4X8bitExactly(dst, src []byte) ([]byte, error) {
 			br[stream].fillFast()
 			br[stream2].fillFast()
 
-			v := single[br[stream].peekByteFast()>>shift].entry
+			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)
-			br[stream].advance(uint8(v))
-
-			v2 := single[br[stream2].peekByteFast()>>shift].entry
 			buf[off+bufoff*stream2] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
 
-			v = single[br[stream].peekByteFast()>>shift].entry
+			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)
-			br[stream].advance(uint8(v))
-
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
 			buf[off+bufoff*stream2+1] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
 
-			v = single[br[stream].peekByteFast()>>shift].entry
+			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)
-			br[stream].advance(uint8(v))
-
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
 			buf[off+bufoff*stream2+2] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
 
-			v = single[br[stream].peekByteFast()>>shift].entry
+			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)
-			br[stream].advance(uint8(v))
-
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
 			buf[off+bufoff*stream2+3] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
 		}
 
 		{
@@ -1184,37 +1200,41 @@ func (d *Decoder) decompress4X8bitExactly(dst, src []byte) ([]byte, error) {
 			br[stream].fillFast()
 			br[stream2].fillFast()
 
-			v := single[br[stream].peekByteFast()>>shift].entry
+			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)
-			br[stream].advance(uint8(v))
-
-			v2 := single[br[stream2].peekByteFast()>>shift].entry
 			buf[off+bufoff*stream2] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
 
-			v = single[br[stream].peekByteFast()>>shift].entry
+			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)
-			br[stream].advance(uint8(v))
-
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
 			buf[off+bufoff*stream2+1] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
 
-			v = single[br[stream].peekByteFast()>>shift].entry
+			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)
-			br[stream].advance(uint8(v))
-
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
 			buf[off+bufoff*stream2+2] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
 
-			v = single[br[stream].peekByteFast()>>shift].entry
+			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)
-			br[stream].advance(uint8(v))
-
-			v2 = single[br[stream2].peekByteFast()>>shift].entry
 			buf[off+bufoff*stream2+3] = uint8(v2 >> 8)
-			br[stream2].advance(uint8(v2))
 		}
 
 		off += 4
@@ -1280,7 +1300,7 @@ func (d *Decoder) decompress4X8bitExactly(dst, src []byte) ([]byte, error) {
 			}
 
 			// Read value and increment offset.
-			v := single[br.peekByteFast()>>shift].entry
+			v := single[br.peekByteFast()].entry
 			nBits := uint8(v)
 			br.advance(nBits)
 			bitsLeft -= int(nBits)
diff --git a/huff0/decompress_test.go b/huff0/decompress_test.go
index b95d8be0c1..a42358200b 100644
--- a/huff0/decompress_test.go
+++ b/huff0/decompress_test.go
@@ -2,6 +2,7 @@ package huff0
 
 import (
 	"bytes"
+	"fmt"
 	"testing"
 )
 
@@ -98,88 +99,94 @@ func TestDecompress1X(t *testing.T) {
 func TestDecompress4X(t *testing.T) {
 	for _, test := range testfiles {
 		t.Run(test.name, func(t *testing.T) {
-			var s = &Scratch{}
-			buf0, err := test.fn()
-			if err != nil {
-				t.Fatal(err)
-			}
-			if len(buf0) > BlockSizeMax {
-				buf0 = buf0[:BlockSizeMax]
-			}
-			b, re, err := Compress4X(buf0, s)
-			if err != test.err4X {
-				t.Errorf("want error %v (%T), got %v (%T)", test.err1X, test.err1X, err, err)
-			}
-			if err != nil {
-				t.Log(test.name, err.Error())
-				return
-			}
-			if b == nil {
-				t.Error("got no output")
-				return
-			}
-			if len(s.OutTable) == 0 {
-				t.Error("got no table definition")
-			}
-			if re {
-				t.Error("claimed to have re-used.")
-			}
-			if len(s.OutData) == 0 {
-				t.Error("got no data output")
-			}
+			for _, tl := range []uint8{0, 5, 6, 7, 8, 9, 10, 11} {
+				t.Run(fmt.Sprintf("tablelog-%d", tl), func(t *testing.T) {
+					var s = &Scratch{}
+					s.TableLog = tl
+					buf0, err := test.fn()
+					if err != nil {
+						t.Fatal(err)
+					}
+					if len(buf0) > BlockSizeMax {
+						buf0 = buf0[:BlockSizeMax]
+					}
+					b, re, err := Compress4X(buf0, s)
+					if err != test.err4X {
+						t.Errorf("want error %v (%T), got %v (%T)", test.err1X, test.err1X, err, err)
+					}
+					if err != nil {
+						t.Log(test.name, err.Error())
+						return
+					}
+					if b == nil {
+						t.Error("got no output")
+						return
+					}
+					if len(s.OutTable) == 0 {
+						t.Error("got no table definition")
+					}
+					if re {
+						t.Error("claimed to have re-used.")
+					}
+					if len(s.OutData) == 0 {
+						t.Error("got no data output")
+					}
 
-			wantRemain := len(s.OutData)
-			t.Logf("%s: %d -> %d bytes (%.2f:1) %t (table: %d bytes)", test.name, len(buf0), len(b), float64(len(buf0))/float64(len(b)), re, len(s.OutTable))
+					wantRemain := len(s.OutData)
+					t.Logf("%s: %d -> %d bytes (%.2f:1) %t (table: %d bytes)", test.name, len(buf0), len(b), float64(len(buf0))/float64(len(b)), re, len(s.OutTable))
 
-			s.Out = nil
-			var remain []byte
-			s, remain, err = ReadTable(b, s)
-			if err != nil {
-				t.Error(err)
-				return
-			}
-			var buf bytes.Buffer
-			if s.matches(s.prevTable, &buf); buf.Len() > 0 {
-				t.Error(buf.String())
-			}
-			if len(remain) != wantRemain {
-				t.Fatalf("remain mismatch, want %d, got %d bytes", wantRemain, len(remain))
-			}
-			t.Logf("remain: %d bytes, ok", len(remain))
-			dc, err := s.Decompress4X(remain, len(buf0))
-			if err != nil {
-				t.Error(err)
-				return
-			}
-			if len(buf0) != len(dc) {
-				t.Errorf(test.name+"decompressed, want size: %d, got %d", len(buf0), len(dc))
-				if len(buf0) > len(dc) {
-					buf0 = buf0[:len(dc)]
-				} else {
-					dc = dc[:len(buf0)]
-				}
-				if !bytes.Equal(buf0, dc) {
-					if len(dc) > 1024 {
-						t.Log(string(dc[:1024]))
-						t.Errorf(test.name+"decompressed, got delta: \n(in)\t%02x !=\n(out)\t%02x\n", buf0[:1024], dc[:1024])
-					} else {
-						t.Log(string(dc))
-						t.Errorf(test.name+"decompressed, got delta: (in) %v != (out) %v\n", buf0, dc)
+					s.Out = nil
+					var remain []byte
+					s, remain, err = ReadTable(b, s)
+					if err != nil {
+						t.Error(err)
+						return
 					}
-				}
-				return
-			}
-			if !bytes.Equal(buf0, dc) {
-				if len(buf0) > 1024 {
-					t.Log(string(dc[:1024]))
-				} else {
-					t.Log(string(dc))
-				}
-				//t.Errorf(test.name+": decompressed, got delta: \n%s")
-				t.Errorf(test.name + ": decompressed, got delta")
-			}
-			if !t.Failed() {
-				t.Log("... roundtrip ok!")
+					var buf bytes.Buffer
+					if s.matches(s.prevTable, &buf); buf.Len() > 0 {
+						t.Error(buf.String())
+					}
+					if len(remain) != wantRemain {
+						t.Fatalf("remain mismatch, want %d, got %d bytes", wantRemain, len(remain))
+					}
+					t.Logf("remain: %d bytes, ok", len(remain))
+					dc, err := s.Decompress4X(remain, len(buf0))
+					if err != nil {
+						t.Error(err)
+						return
+					}
+					if len(buf0) != len(dc) {
+						t.Errorf(test.name+"decompressed, want size: %d, got %d", len(buf0), len(dc))
+						if len(buf0) > len(dc) {
+							buf0 = buf0[:len(dc)]
+						} else {
+							dc = dc[:len(buf0)]
+						}
+						if !bytes.Equal(buf0, dc) {
+							if len(dc) > 1024 {
+								t.Log(string(dc[:1024]))
+								t.Errorf(test.name+"decompressed, got delta: \n(in)\t%02x !=\n(out)\t%02x\n", buf0[:1024], dc[:1024])
+							} else {
+								t.Log(string(dc))
+								t.Errorf(test.name+"decompressed, got delta: (in) %v != (out) %v\n", buf0, dc)
+							}
+						}
+						return
+					}
+					if !bytes.Equal(buf0, dc) {
+						if len(buf0) > 1024 {
+							t.Log(string(dc[:1024]))
+						} else {
+							t.Log(string(dc))
+						}
+						//t.Errorf(test.name+": decompressed, got delta: \n%s")
+						t.Errorf(test.name + ": decompressed, got delta")
+					}
+					if !t.Failed() {
+						t.Log("... roundtrip ok!")
+					}
+
+				})
 			}
 		})
 	}
@@ -479,6 +486,43 @@ func BenchmarkDecompress4XNoTable(b *testing.B) {
 	}
 }
 
+func BenchmarkDecompress4XNoTableTableLog8(b *testing.B) {
+	for _, tt := range testfiles[:1] {
+		test := tt
+		if test.err4X != nil {
+			continue
+		}
+		b.Run(test.name, func(b *testing.B) {
+			var s = &Scratch{}
+			s.Reuse = ReusePolicyNone
+			buf0, err := test.fn()
+			if err != nil {
+				b.Fatal(err)
+			}
+			if len(buf0) > BlockSizeMax {
+				buf0 = buf0[:BlockSizeMax]
+			}
+			s.TableLog = 8
+			compressed, _, err := Compress4X(buf0, s)
+			if err != test.err1X {
+				b.Fatal("unexpected error:", err)
+			}
+			s.Out = nil
+			s, remain, _ := ReadTable(compressed, s)
+			s.Decompress4X(remain, len(buf0))
+			b.ResetTimer()
+			b.ReportAllocs()
+			b.SetBytes(int64(len(buf0)))
+			for i := 0; i < b.N; i++ {
+				_, err = s.Decompress4X(remain, len(buf0))
+				if err != nil {
+					b.Fatal(err)
+				}
+			}
+		})
+	}
+}
+
 func BenchmarkDecompress4XTable(b *testing.B) {
 	for _, tt := range testfiles {
 		test := tt