diff --git a/s2/_generate/gen.go b/s2/_generate/gen.go
index bac31ae62a..e191847695 100644
--- a/s2/_generate/gen.go
+++ b/s2/_generate/gen.go
@@ -51,6 +51,7 @@ func main() {
 		bmi1:         false,
 		bmi2:         false,
 		snappy:       false,
+		avx2:         false,
 		outputMargin: 9,
 	}
 	o.genEncodeBlockAsm("encodeBlockAsm", 14, 6, 6, limit14B)
@@ -150,6 +151,7 @@ type options struct {
 	bmi1         bool
 	bmi2         bool
 	skipOutput   bool
+	avx2         bool
 	maxLen       int
 	maxOffset    int
 	outputMargin int // Should be at least 5.
@@ -614,6 +616,9 @@ func (o options) genEncodeBlockAsm(name string, tableBits, skipLog, hashBytes, m
 			panic(err)
 		}
 		MOVQ(U32(0), ri.Addr)
+		if o.avx2 {
+			VZEROUPPER()
+		}
 		RET()
 	}
 	Label("match_dst_size_check_" + name)
@@ -697,6 +702,9 @@ func (o options) genEncodeBlockAsm(name string, tableBits, skipLog, hashBytes, m
 					panic(err)
 				}
 				MOVQ(U32(0), ri.Addr)
+				if o.avx2 {
+					VZEROUPPER()
+				}
 				RET()
 				Label("match_nolit_dst_ok_" + name)
 			}
@@ -753,6 +761,9 @@ func (o options) genEncodeBlockAsm(name string, tableBits, skipLog, hashBytes, m
 		if err != nil {
 			panic(err)
 		}
+		if o.avx2 {
+			VZEROUPPER()
+		}
 		MOVQ(U32(0), ri.Addr)
 		RET()
 		Label("emit_remainder_ok_" + name)
@@ -801,6 +812,9 @@ func (o options) genEncodeBlockAsm(name string, tableBits, skipLog, hashBytes, m
 			JAE(ok)
 		})
 	}
+	if o.avx2 {
+		VZEROUPPER()
+	}
 	Store(length, ReturnIndex(0))
 	RET()
 }
@@ -1273,6 +1287,9 @@ func (o options) genEncodeBetterBlockAsm(name string, lTableBits, sTableBits, sk
 			panic(err)
 		}
 		MOVQ(U32(0), ri.Addr)
+		if o.avx2 {
+			VZEROUPPER()
+		}
 		RET()
 	}
 	Label("match_dst_size_check_" + name)
@@ -1385,6 +1402,9 @@ func (o options) genEncodeBetterBlockAsm(name string, lTableBits, sTableBits, sk
 				panic(err)
 			}
 			MOVQ(U32(0), ri.Addr)
+			if o.avx2 {
+				VZEROUPPER()
+			}
 			RET()
 		}
 	}
@@ -1538,6 +1558,9 @@ func (o options) genEncodeBetterBlockAsm(name string, lTableBits, sTableBits, sk
 			panic(err)
 		}
 		MOVQ(U32(0), ri.Addr)
+		if o.avx2 {
+			VZEROUPPER()
+		}
 		RET()
 		Label("emit_remainder_ok_" + name)
 	}
@@ -1579,6 +1602,9 @@ func (o options) genEncodeBetterBlockAsm(name string, lTableBits, sTableBits, sk
 		JAE(ok)
 	})
 	Store(length, ReturnIndex(0))
+	if o.avx2 {
+		VZEROUPPER()
+	}
 	RET()
 }
 
@@ -2696,6 +2722,9 @@ func (o options) genMatchLen() {
 	Load(Param("a").Len(), length)
 	l := o.matchLen("standalone", aBase, bBase, length, LabelRef("gen_match_len_end"))
 	Label("gen_match_len_end")
+	if o.avx2 {
+		VZEROUPPER()
+	}
 	Store(l.As64(), ReturnIndex(0))
 	RET()
 }
@@ -2706,11 +2735,13 @@ func (o options) genMatchLen() {
 // Uses 2 GP registers.
 func (o options) matchLen(name string, a, b, len reg.GPVirtual, end LabelRef) reg.GPVirtual {
 	Comment("matchLen")
-	if false {
-		return o.matchLenAlt(name, a, b, len, end)
-	}
 	tmp, matched := GP64(), GP32()
 	XORL(matched, matched)
+	if o.avx2 {
+		// Not faster...
+		o.matchLenAVX2(name+"Avx2", a, b, len, LabelRef("avx2_continue_"+name), end, matched)
+	}
+	Label("avx2_continue_" + name)
 
 	CMPL(len.As32(), U8(8))
 	JB(LabelRef("matchlen_match4_" + name))
@@ -2740,7 +2771,6 @@ func (o options) matchLen(name string, a, b, len reg.GPVirtual, end LabelRef) re
 	LEAL(Mem{Base: matched, Disp: 8}, matched)
 	CMPL(len.As32(), U8(8))
 	JAE(LabelRef("matchlen_loopback_" + name))
-	JZ(end)
 
 	// Less than 8 bytes left.
 	// Test 4 bytes...
@@ -2750,23 +2780,25 @@ func (o options) matchLen(name string, a, b, len reg.GPVirtual, end LabelRef) re
 	MOVL(Mem{Base: a, Index: matched, Scale: 1}, tmp.As32())
 	CMPL(Mem{Base: b, Index: matched, Scale: 1}, tmp.As32())
 	JNE(LabelRef("matchlen_match2_" + name))
-	SUBL(U8(4), len.As32())
+	LEAL(Mem{Base: len.As32(), Disp: -4}, len.As32())
 	LEAL(Mem{Base: matched, Disp: 4}, matched)
 
 	// Test 2 bytes...
 	Label("matchlen_match2_" + name)
-	CMPL(len.As32(), U8(2))
-	JB(LabelRef("matchlen_match1_" + name))
+	CMPL(len.As32(), U8(1))
+	// If we don't have 1, branch appropriately
+	JE(LabelRef("matchlen_match1_" + name))
+	JB(end)
+	// 2 or 3
 	MOVW(Mem{Base: a, Index: matched, Scale: 1}, tmp.As16())
 	CMPW(Mem{Base: b, Index: matched, Scale: 1}, tmp.As16())
 	JNE(LabelRef("matchlen_match1_" + name))
-	SUBL(U8(2), len.As32())
 	LEAL(Mem{Base: matched, Disp: 2}, matched)
+	SUBL(U8(2), len.As32())
+	JZ(end)
 
 	// Test 1 byte...
 	Label("matchlen_match1_" + name)
-	CMPL(len.As32(), U8(1))
-	JB(end)
 	MOVB(Mem{Base: a, Index: matched, Scale: 1}, tmp.As8())
 	CMPB(Mem{Base: b, Index: matched, Scale: 1}, tmp.As8())
 	JNE(end)
@@ -2780,94 +2812,47 @@ func (o options) matchLen(name string, a, b, len reg.GPVirtual, end LabelRef) re
 // Will jump to end when done and returns the length.
 // Uses 3 GP registers.
 // It is better on longer matches.
-func (o options) matchLenAlt(name string, a, b, len reg.GPVirtual, end LabelRef) reg.GPVirtual {
-	Comment("matchLenAlt")
-	tmp, tmp2, matched := GP64(), GP64(), GP32()
-	XORL(matched, matched)
-
-	CMPL(len.As32(), U8(16))
-	JB(LabelRef("matchlen_short_" + name))
+func (o options) matchLenAVX2(name string, a, b, len reg.GPVirtual, cont, end LabelRef, dst reg.GPVirtual) {
+	Comment("matchLenAVX2")
 
-	Label("matchlen_loopback_" + name)
-	MOVQ(Mem{Base: a}, tmp)
-	MOVQ(Mem{Base: a, Disp: 8}, tmp2)
-	XORQ(Mem{Base: b, Disp: 0}, tmp)
-	XORQ(Mem{Base: b, Disp: 8}, tmp2)
-	endTest := func(xored reg.GPVirtual, disp int, ok LabelRef) {
-		TESTQ(xored, xored)
-		JZ(ok)
-		// Not all match.
-		BSFQ(xored, xored)
-		SARQ(U8(3), xored)
-		LEAL(Mem{Base: matched, Index: xored, Scale: 1, Disp: disp}, matched)
-		JMP(end)
-	}
-	endTest(tmp, 0, LabelRef("matchlen_loop_tmp2_"+name))
-	Label("matchlen_loop_tmp2_" + name)
-	endTest(tmp2, 8, LabelRef("matchlen_loop_"+name))
-
-	// All 16 byte matched, update and loop.
-	Label("matchlen_loop_" + name)
-	SUBL(U8(16), len.As32())
-	ADDL(U8(16), matched)
-	ADDQ(U8(16), a)
-	ADDQ(U8(16), b)
-	CMPL(len.As32(), U8(16))
-	JAE(LabelRef("matchlen_loopback_" + name))
-
-	// Test 4 bytes at the time...
-	Label("matchlen_short_" + name)
-	lenoff := 0
-	if true {
-		lenoff = 4
-		SUBL(U8(4), len.As32())
-		JC(LabelRef("matchlen_single_resume_" + name))
-
-		Label("matchlen_four_loopback_" + name)
-		assert(func(ok LabelRef) {
-			CMPL(len.As32(), U32(math.MaxInt32))
-			JB(ok)
-		})
-
-		MOVL(Mem{Base: a}, tmp.As32())
-		XORL(Mem{Base: b}, tmp.As32())
-		{
-			JZ(LabelRef("matchlen_four_loopback_next" + name))
-			BSFL(tmp.As32(), tmp.As32())
-			SARQ(U8(3), tmp)
-			LEAL(Mem{Base: matched, Index: tmp, Scale: 1}, matched)
-			JMP(end)
-		}
-		Label("matchlen_four_loopback_next" + name)
-		ADDL(U8(4), matched)
-		ADDQ(U8(4), a)
-		ADDQ(U8(4), b)
-		SUBL(U8(4), len.As32())
-		JNC(LabelRef("matchlen_four_loopback_" + name))
+	equalMaskBits := GP64()
+	Label(name + "loop")
+	{
+		CMPQ(len, U8(32))
+		JB(cont)
+		Comment("load 32 bytes into YMM registers")
+		adata := YMM()
+		bdata := YMM()
+		equalMaskBytes := YMM()
+		VMOVDQU(Mem{Base: a}, adata)
+		VMOVDQU(Mem{Base: b}, bdata)
+		Comment("compare bytes in adata and bdata, like 'bytewise XNOR'",
+			"if the byte is the same in adata and bdata, VPCMPEQB will store 0xFF in the same position in equalMaskBytes")
+		VPCMPEQB(adata, bdata, equalMaskBytes)
+		Comment("like convert byte to bit, store equalMaskBytes into general reg")
+		VPMOVMSKB(equalMaskBytes, equalMaskBits.As32())
+		CMPL(equalMaskBits.As32(), U32(0xffffffff))
+		JNE(LabelRef(name + "cal_prefix"))
+		ADDQ(U8(32), a)
+		ADDQ(U8(32), b)
+		ADDL(U8(32), dst)
+		SUBQ(U8(32), len)
+		JZ(end)
+		JMP(LabelRef(name + "loop"))
 	}
 
-	// Test one at the time
-	Label("matchlen_single_resume_" + name)
-	if true {
-		// Less than 16 bytes left.
-		if lenoff > 0 {
-			ADDL(U8(lenoff), len.As32())
+	Label(name + "cal_prefix")
+	{
+		NOTQ(equalMaskBits)
+		if o.bmi1 {
+			TZCNTQ(equalMaskBits, equalMaskBits)
+		} else {
+			BSFQ(equalMaskBits, equalMaskBits)
 		}
-		TESTL(len.As32(), len.As32())
-		JZ(end)
-
-		Label("matchlen_single_loopback_" + name)
-		MOVB(Mem{Base: a}, tmp.As8())
-		CMPB(Mem{Base: b}, tmp.As8())
-		JNE(end)
-		INCL(matched)
-		INCQ(a)
-		INCQ(b)
-		DECL(len.As32())
-		JNZ(LabelRef("matchlen_single_loopback_" + name))
+		ADDL(equalMaskBits.As32(), dst)
 	}
 	JMP(end)
-	return matched
+	return
 }
 
 func (o options) cvtLZ4BlockAsm(lz4s bool) {
diff --git a/s2/encodeblock_amd64.s b/s2/encodeblock_amd64.s
index 63456c0f5f..54031aa313 100644
--- a/s2/encodeblock_amd64.s
+++ b/s2/encodeblock_amd64.s
@@ -274,7 +274,6 @@ matchlen_loop_repeat_extend_encodeBlockAsm:
 	LEAL 8(R11), R11
 	CMPL R8, $0x08
 	JAE  matchlen_loopback_repeat_extend_encodeBlockAsm
-	JZ   repeat_extend_forward_end_encodeBlockAsm
 
 matchlen_match4_repeat_extend_encodeBlockAsm:
 	CMPL R8, $0x04
@@ -282,21 +281,21 @@ matchlen_match4_repeat_extend_encodeBlockAsm:
 	MOVL (R9)(R11*1), R10
 	CMPL (BX)(R11*1), R10
 	JNE  matchlen_match2_repeat_extend_encodeBlockAsm
-	SUBL $0x04, R8
+	LEAL -4(R8), R8
 	LEAL 4(R11), R11
 
 matchlen_match2_repeat_extend_encodeBlockAsm:
-	CMPL R8, $0x02
-	JB   matchlen_match1_repeat_extend_encodeBlockAsm
+	CMPL R8, $0x01
+	JE   matchlen_match1_repeat_extend_encodeBlockAsm
+	JB   repeat_extend_forward_end_encodeBlockAsm
 	MOVW (R9)(R11*1), R10
 	CMPW (BX)(R11*1), R10
 	JNE  matchlen_match1_repeat_extend_encodeBlockAsm
-	SUBL $0x02, R8
 	LEAL 2(R11), R11
+	SUBL $0x02, R8
+	JZ   repeat_extend_forward_end_encodeBlockAsm
 
 matchlen_match1_repeat_extend_encodeBlockAsm:
-	CMPL R8, $0x01
-	JB   repeat_extend_forward_end_encodeBlockAsm
 	MOVB (R9)(R11*1), R10
 	CMPB (BX)(R11*1), R10
 	JNE  repeat_extend_forward_end_encodeBlockAsm
@@ -877,7 +876,6 @@ matchlen_loop_match_nolit_encodeBlockAsm:
 	LEAL 8(R9), R9
 	CMPL SI, $0x08
 	JAE  matchlen_loopback_match_nolit_encodeBlockAsm
-	JZ   match_nolit_end_encodeBlockAsm
 
 matchlen_match4_match_nolit_encodeBlockAsm:
 	CMPL SI, $0x04
@@ -885,21 +883,21 @@ matchlen_match4_match_nolit_encodeBlockAsm:
 	MOVL (DI)(R9*1), R8
 	CMPL (BX)(R9*1), R8
 	JNE  matchlen_match2_match_nolit_encodeBlockAsm
-	SUBL $0x04, SI
+	LEAL -4(SI), SI
 	LEAL 4(R9), R9
 
 matchlen_match2_match_nolit_encodeBlockAsm:
-	CMPL SI, $0x02
-	JB   matchlen_match1_match_nolit_encodeBlockAsm
+	CMPL SI, $0x01
+	JE   matchlen_match1_match_nolit_encodeBlockAsm
+	JB   match_nolit_end_encodeBlockAsm
 	MOVW (DI)(R9*1), R8
 	CMPW (BX)(R9*1), R8
 	JNE  matchlen_match1_match_nolit_encodeBlockAsm
-	SUBL $0x02, SI
 	LEAL 2(R9), R9
+	SUBL $0x02, SI
+	JZ   match_nolit_end_encodeBlockAsm
 
 matchlen_match1_match_nolit_encodeBlockAsm:
-	CMPL SI, $0x01
-	JB   match_nolit_end_encodeBlockAsm
 	MOVB (DI)(R9*1), R8
 	CMPB (BX)(R9*1), R8
 	JNE  match_nolit_end_encodeBlockAsm
@@ -1637,7 +1635,6 @@ matchlen_loop_repeat_extend_encodeBlockAsm4MB:
 	LEAL 8(R11), R11
 	CMPL R8, $0x08
 	JAE  matchlen_loopback_repeat_extend_encodeBlockAsm4MB
-	JZ   repeat_extend_forward_end_encodeBlockAsm4MB
 
 matchlen_match4_repeat_extend_encodeBlockAsm4MB:
 	CMPL R8, $0x04
@@ -1645,21 +1642,21 @@ matchlen_match4_repeat_extend_encodeBlockAsm4MB:
 	MOVL (R9)(R11*1), R10
 	CMPL (BX)(R11*1), R10
 	JNE  matchlen_match2_repeat_extend_encodeBlockAsm4MB
-	SUBL $0x04, R8
+	LEAL -4(R8), R8
 	LEAL 4(R11), R11
 
 matchlen_match2_repeat_extend_encodeBlockAsm4MB:
-	CMPL R8, $0x02
-	JB   matchlen_match1_repeat_extend_encodeBlockAsm4MB
+	CMPL R8, $0x01
+	JE   matchlen_match1_repeat_extend_encodeBlockAsm4MB
+	JB   repeat_extend_forward_end_encodeBlockAsm4MB
 	MOVW (R9)(R11*1), R10
 	CMPW (BX)(R11*1), R10
 	JNE  matchlen_match1_repeat_extend_encodeBlockAsm4MB
-	SUBL $0x02, R8
 	LEAL 2(R11), R11
+	SUBL $0x02, R8
+	JZ   repeat_extend_forward_end_encodeBlockAsm4MB
 
 matchlen_match1_repeat_extend_encodeBlockAsm4MB:
-	CMPL R8, $0x01
-	JB   repeat_extend_forward_end_encodeBlockAsm4MB
 	MOVB (R9)(R11*1), R10
 	CMPB (BX)(R11*1), R10
 	JNE  repeat_extend_forward_end_encodeBlockAsm4MB
@@ -2190,7 +2187,6 @@ matchlen_loop_match_nolit_encodeBlockAsm4MB:
 	LEAL 8(R9), R9
 	CMPL SI, $0x08
 	JAE  matchlen_loopback_match_nolit_encodeBlockAsm4MB
-	JZ   match_nolit_end_encodeBlockAsm4MB
 
 matchlen_match4_match_nolit_encodeBlockAsm4MB:
 	CMPL SI, $0x04
@@ -2198,21 +2194,21 @@ matchlen_match4_match_nolit_encodeBlockAsm4MB:
 	MOVL (DI)(R9*1), R8
 	CMPL (BX)(R9*1), R8
 	JNE  matchlen_match2_match_nolit_encodeBlockAsm4MB
-	SUBL $0x04, SI
+	LEAL -4(SI), SI
 	LEAL 4(R9), R9
 
 matchlen_match2_match_nolit_encodeBlockAsm4MB:
-	CMPL SI, $0x02
-	JB   matchlen_match1_match_nolit_encodeBlockAsm4MB
+	CMPL SI, $0x01
+	JE   matchlen_match1_match_nolit_encodeBlockAsm4MB
+	JB   match_nolit_end_encodeBlockAsm4MB
 	MOVW (DI)(R9*1), R8
 	CMPW (BX)(R9*1), R8
 	JNE  matchlen_match1_match_nolit_encodeBlockAsm4MB
-	SUBL $0x02, SI
 	LEAL 2(R9), R9
+	SUBL $0x02, SI
+	JZ   match_nolit_end_encodeBlockAsm4MB
 
 matchlen_match1_match_nolit_encodeBlockAsm4MB:
-	CMPL SI, $0x01
-	JB   match_nolit_end_encodeBlockAsm4MB
 	MOVB (DI)(R9*1), R8
 	CMPB (BX)(R9*1), R8
 	JNE  match_nolit_end_encodeBlockAsm4MB
@@ -2902,7 +2898,6 @@ matchlen_loop_repeat_extend_encodeBlockAsm12B:
 	LEAL 8(R11), R11
 	CMPL R8, $0x08
 	JAE  matchlen_loopback_repeat_extend_encodeBlockAsm12B
-	JZ   repeat_extend_forward_end_encodeBlockAsm12B
 
 matchlen_match4_repeat_extend_encodeBlockAsm12B:
 	CMPL R8, $0x04
@@ -2910,21 +2905,21 @@ matchlen_match4_repeat_extend_encodeBlockAsm12B:
 	MOVL (R9)(R11*1), R10
 	CMPL (BX)(R11*1), R10
 	JNE  matchlen_match2_repeat_extend_encodeBlockAsm12B
-	SUBL $0x04, R8
+	LEAL -4(R8), R8
 	LEAL 4(R11), R11
 
 matchlen_match2_repeat_extend_encodeBlockAsm12B:
-	CMPL R8, $0x02
-	JB   matchlen_match1_repeat_extend_encodeBlockAsm12B
+	CMPL R8, $0x01
+	JE   matchlen_match1_repeat_extend_encodeBlockAsm12B
+	JB   repeat_extend_forward_end_encodeBlockAsm12B
 	MOVW (R9)(R11*1), R10
 	CMPW (BX)(R11*1), R10
 	JNE  matchlen_match1_repeat_extend_encodeBlockAsm12B
-	SUBL $0x02, R8
 	LEAL 2(R11), R11
+	SUBL $0x02, R8
+	JZ   repeat_extend_forward_end_encodeBlockAsm12B
 
 matchlen_match1_repeat_extend_encodeBlockAsm12B:
-	CMPL R8, $0x01
-	JB   repeat_extend_forward_end_encodeBlockAsm12B
 	MOVB (R9)(R11*1), R10
 	CMPB (BX)(R11*1), R10
 	JNE  repeat_extend_forward_end_encodeBlockAsm12B
@@ -3333,7 +3328,6 @@ matchlen_loop_match_nolit_encodeBlockAsm12B:
 	LEAL 8(R9), R9
 	CMPL SI, $0x08
 	JAE  matchlen_loopback_match_nolit_encodeBlockAsm12B
-	JZ   match_nolit_end_encodeBlockAsm12B
 
 matchlen_match4_match_nolit_encodeBlockAsm12B:
 	CMPL SI, $0x04
@@ -3341,21 +3335,21 @@ matchlen_match4_match_nolit_encodeBlockAsm12B:
 	MOVL (DI)(R9*1), R8
 	CMPL (BX)(R9*1), R8
 	JNE  matchlen_match2_match_nolit_encodeBlockAsm12B
-	SUBL $0x04, SI
+	LEAL -4(SI), SI
 	LEAL 4(R9), R9
 
 matchlen_match2_match_nolit_encodeBlockAsm12B:
-	CMPL SI, $0x02
-	JB   matchlen_match1_match_nolit_encodeBlockAsm12B
+	CMPL SI, $0x01
+	JE   matchlen_match1_match_nolit_encodeBlockAsm12B
+	JB   match_nolit_end_encodeBlockAsm12B
 	MOVW (DI)(R9*1), R8
 	CMPW (BX)(R9*1), R8
 	JNE  matchlen_match1_match_nolit_encodeBlockAsm12B
-	SUBL $0x02, SI
 	LEAL 2(R9), R9
+	SUBL $0x02, SI
+	JZ   match_nolit_end_encodeBlockAsm12B
 
 matchlen_match1_match_nolit_encodeBlockAsm12B:
-	CMPL SI, $0x01
-	JB   match_nolit_end_encodeBlockAsm12B
 	MOVB (DI)(R9*1), R8
 	CMPB (BX)(R9*1), R8
 	JNE  match_nolit_end_encodeBlockAsm12B
@@ -3935,7 +3929,6 @@ matchlen_loop_repeat_extend_encodeBlockAsm10B:
 	LEAL 8(R11), R11
 	CMPL R8, $0x08
 	JAE  matchlen_loopback_repeat_extend_encodeBlockAsm10B
-	JZ   repeat_extend_forward_end_encodeBlockAsm10B
 
 matchlen_match4_repeat_extend_encodeBlockAsm10B:
 	CMPL R8, $0x04
@@ -3943,21 +3936,21 @@ matchlen_match4_repeat_extend_encodeBlockAsm10B:
 	MOVL (R9)(R11*1), R10
 	CMPL (BX)(R11*1), R10
 	JNE  matchlen_match2_repeat_extend_encodeBlockAsm10B
-	SUBL $0x04, R8
+	LEAL -4(R8), R8
 	LEAL 4(R11), R11
 
 matchlen_match2_repeat_extend_encodeBlockAsm10B:
-	CMPL R8, $0x02
-	JB   matchlen_match1_repeat_extend_encodeBlockAsm10B
+	CMPL R8, $0x01
+	JE   matchlen_match1_repeat_extend_encodeBlockAsm10B
+	JB   repeat_extend_forward_end_encodeBlockAsm10B
 	MOVW (R9)(R11*1), R10
 	CMPW (BX)(R11*1), R10
 	JNE  matchlen_match1_repeat_extend_encodeBlockAsm10B
-	SUBL $0x02, R8
 	LEAL 2(R11), R11
+	SUBL $0x02, R8
+	JZ   repeat_extend_forward_end_encodeBlockAsm10B
 
 matchlen_match1_repeat_extend_encodeBlockAsm10B:
-	CMPL R8, $0x01
-	JB   repeat_extend_forward_end_encodeBlockAsm10B
 	MOVB (R9)(R11*1), R10
 	CMPB (BX)(R11*1), R10
 	JNE  repeat_extend_forward_end_encodeBlockAsm10B
@@ -4366,7 +4359,6 @@ matchlen_loop_match_nolit_encodeBlockAsm10B:
 	LEAL 8(R9), R9
 	CMPL SI, $0x08
 	JAE  matchlen_loopback_match_nolit_encodeBlockAsm10B
-	JZ   match_nolit_end_encodeBlockAsm10B
 
 matchlen_match4_match_nolit_encodeBlockAsm10B:
 	CMPL SI, $0x04
@@ -4374,21 +4366,21 @@ matchlen_match4_match_nolit_encodeBlockAsm10B:
 	MOVL (DI)(R9*1), R8
 	CMPL (BX)(R9*1), R8
 	JNE  matchlen_match2_match_nolit_encodeBlockAsm10B
-	SUBL $0x04, SI
+	LEAL -4(SI), SI
 	LEAL 4(R9), R9
 
 matchlen_match2_match_nolit_encodeBlockAsm10B:
-	CMPL SI, $0x02
-	JB   matchlen_match1_match_nolit_encodeBlockAsm10B
+	CMPL SI, $0x01
+	JE   matchlen_match1_match_nolit_encodeBlockAsm10B
+	JB   match_nolit_end_encodeBlockAsm10B
 	MOVW (DI)(R9*1), R8
 	CMPW (BX)(R9*1), R8
 	JNE  matchlen_match1_match_nolit_encodeBlockAsm10B
-	SUBL $0x02, SI
 	LEAL 2(R9), R9
+	SUBL $0x02, SI
+	JZ   match_nolit_end_encodeBlockAsm10B
 
 matchlen_match1_match_nolit_encodeBlockAsm10B:
-	CMPL SI, $0x01
-	JB   match_nolit_end_encodeBlockAsm10B
 	MOVB (DI)(R9*1), R8
 	CMPB (BX)(R9*1), R8
 	JNE  match_nolit_end_encodeBlockAsm10B
@@ -4968,7 +4960,6 @@ matchlen_loop_repeat_extend_encodeBlockAsm8B:
 	LEAL 8(R11), R11
 	CMPL R8, $0x08
 	JAE  matchlen_loopback_repeat_extend_encodeBlockAsm8B
-	JZ   repeat_extend_forward_end_encodeBlockAsm8B
 
 matchlen_match4_repeat_extend_encodeBlockAsm8B:
 	CMPL R8, $0x04
@@ -4976,21 +4967,21 @@ matchlen_match4_repeat_extend_encodeBlockAsm8B:
 	MOVL (R9)(R11*1), R10
 	CMPL (BX)(R11*1), R10
 	JNE  matchlen_match2_repeat_extend_encodeBlockAsm8B
-	SUBL $0x04, R8
+	LEAL -4(R8), R8
 	LEAL 4(R11), R11
 
 matchlen_match2_repeat_extend_encodeBlockAsm8B:
-	CMPL R8, $0x02
-	JB   matchlen_match1_repeat_extend_encodeBlockAsm8B
+	CMPL R8, $0x01
+	JE   matchlen_match1_repeat_extend_encodeBlockAsm8B
+	JB   repeat_extend_forward_end_encodeBlockAsm8B
 	MOVW (R9)(R11*1), R10
 	CMPW (BX)(R11*1), R10
 	JNE  matchlen_match1_repeat_extend_encodeBlockAsm8B
-	SUBL $0x02, R8
 	LEAL 2(R11), R11
+	SUBL $0x02, R8
+	JZ   repeat_extend_forward_end_encodeBlockAsm8B
 
 matchlen_match1_repeat_extend_encodeBlockAsm8B:
-	CMPL R8, $0x01
-	JB   repeat_extend_forward_end_encodeBlockAsm8B
 	MOVB (R9)(R11*1), R10
 	CMPB (BX)(R11*1), R10
 	JNE  repeat_extend_forward_end_encodeBlockAsm8B
@@ -5385,7 +5376,6 @@ matchlen_loop_match_nolit_encodeBlockAsm8B:
 	LEAL 8(R9), R9
 	CMPL SI, $0x08
 	JAE  matchlen_loopback_match_nolit_encodeBlockAsm8B
-	JZ   match_nolit_end_encodeBlockAsm8B
 
 matchlen_match4_match_nolit_encodeBlockAsm8B:
 	CMPL SI, $0x04
@@ -5393,21 +5383,21 @@ matchlen_match4_match_nolit_encodeBlockAsm8B:
 	MOVL (DI)(R9*1), R8
 	CMPL (BX)(R9*1), R8
 	JNE  matchlen_match2_match_nolit_encodeBlockAsm8B
-	SUBL $0x04, SI
+	LEAL -4(SI), SI
 	LEAL 4(R9), R9
 
 matchlen_match2_match_nolit_encodeBlockAsm8B:
-	CMPL SI, $0x02
-	JB   matchlen_match1_match_nolit_encodeBlockAsm8B
+	CMPL SI, $0x01
+	JE   matchlen_match1_match_nolit_encodeBlockAsm8B
+	JB   match_nolit_end_encodeBlockAsm8B
 	MOVW (DI)(R9*1), R8
 	CMPW (BX)(R9*1), R8
 	JNE  matchlen_match1_match_nolit_encodeBlockAsm8B
-	SUBL $0x02, SI
 	LEAL 2(R9), R9
+	SUBL $0x02, SI
+	JZ   match_nolit_end_encodeBlockAsm8B
 
 matchlen_match1_match_nolit_encodeBlockAsm8B:
-	CMPL SI, $0x01
-	JB   match_nolit_end_encodeBlockAsm8B
 	MOVB (DI)(R9*1), R8
 	CMPB (BX)(R9*1), R8
 	JNE  match_nolit_end_encodeBlockAsm8B
@@ -5889,7 +5879,6 @@ matchlen_loop_match_nolit_encodeBetterBlockAsm:
 	LEAL 8(R11), R11
 	CMPL DI, $0x08
 	JAE  matchlen_loopback_match_nolit_encodeBetterBlockAsm
-	JZ   match_nolit_end_encodeBetterBlockAsm
 
 matchlen_match4_match_nolit_encodeBetterBlockAsm:
 	CMPL DI, $0x04
@@ -5897,21 +5886,21 @@ matchlen_match4_match_nolit_encodeBetterBlockAsm:
 	MOVL (R8)(R11*1), R10
 	CMPL (R9)(R11*1), R10
 	JNE  matchlen_match2_match_nolit_encodeBetterBlockAsm
-	SUBL $0x04, DI
+	LEAL -4(DI), DI
 	LEAL 4(R11), R11
 
 matchlen_match2_match_nolit_encodeBetterBlockAsm:
-	CMPL DI, $0x02
-	JB   matchlen_match1_match_nolit_encodeBetterBlockAsm
+	CMPL DI, $0x01
+	JE   matchlen_match1_match_nolit_encodeBetterBlockAsm
+	JB   match_nolit_end_encodeBetterBlockAsm
 	MOVW (R8)(R11*1), R10
 	CMPW (R9)(R11*1), R10
 	JNE  matchlen_match1_match_nolit_encodeBetterBlockAsm
-	SUBL $0x02, DI
 	LEAL 2(R11), R11
+	SUBL $0x02, DI
+	JZ   match_nolit_end_encodeBetterBlockAsm
 
 matchlen_match1_match_nolit_encodeBetterBlockAsm:
-	CMPL DI, $0x01
-	JB   match_nolit_end_encodeBetterBlockAsm
 	MOVB (R8)(R11*1), R10
 	CMPB (R9)(R11*1), R10
 	JNE  match_nolit_end_encodeBetterBlockAsm
@@ -6962,7 +6951,6 @@ matchlen_loop_match_nolit_encodeBetterBlockAsm4MB:
 	LEAL 8(R11), R11
 	CMPL DI, $0x08
 	JAE  matchlen_loopback_match_nolit_encodeBetterBlockAsm4MB
-	JZ   match_nolit_end_encodeBetterBlockAsm4MB
 
 matchlen_match4_match_nolit_encodeBetterBlockAsm4MB:
 	CMPL DI, $0x04
@@ -6970,21 +6958,21 @@ matchlen_match4_match_nolit_encodeBetterBlockAsm4MB:
 	MOVL (R8)(R11*1), R10
 	CMPL (R9)(R11*1), R10
 	JNE  matchlen_match2_match_nolit_encodeBetterBlockAsm4MB
-	SUBL $0x04, DI
+	LEAL -4(DI), DI
 	LEAL 4(R11), R11
 
 matchlen_match2_match_nolit_encodeBetterBlockAsm4MB:
-	CMPL DI, $0x02
-	JB   matchlen_match1_match_nolit_encodeBetterBlockAsm4MB
+	CMPL DI, $0x01
+	JE   matchlen_match1_match_nolit_encodeBetterBlockAsm4MB
+	JB   match_nolit_end_encodeBetterBlockAsm4MB
 	MOVW (R8)(R11*1), R10
 	CMPW (R9)(R11*1), R10
 	JNE  matchlen_match1_match_nolit_encodeBetterBlockAsm4MB
-	SUBL $0x02, DI
 	LEAL 2(R11), R11
+	SUBL $0x02, DI
+	JZ   match_nolit_end_encodeBetterBlockAsm4MB
 
 matchlen_match1_match_nolit_encodeBetterBlockAsm4MB:
-	CMPL DI, $0x01
-	JB   match_nolit_end_encodeBetterBlockAsm4MB
 	MOVB (R8)(R11*1), R10
 	CMPB (R9)(R11*1), R10
 	JNE  match_nolit_end_encodeBetterBlockAsm4MB
@@ -7961,7 +7949,6 @@ matchlen_loop_match_nolit_encodeBetterBlockAsm12B:
 	LEAL 8(R11), R11
 	CMPL DI, $0x08
 	JAE  matchlen_loopback_match_nolit_encodeBetterBlockAsm12B
-	JZ   match_nolit_end_encodeBetterBlockAsm12B
 
 matchlen_match4_match_nolit_encodeBetterBlockAsm12B:
 	CMPL DI, $0x04
@@ -7969,21 +7956,21 @@ matchlen_match4_match_nolit_encodeBetterBlockAsm12B:
 	MOVL (R8)(R11*1), R10
 	CMPL (R9)(R11*1), R10
 	JNE  matchlen_match2_match_nolit_encodeBetterBlockAsm12B
-	SUBL $0x04, DI
+	LEAL -4(DI), DI
 	LEAL 4(R11), R11
 
 matchlen_match2_match_nolit_encodeBetterBlockAsm12B:
-	CMPL DI, $0x02
-	JB   matchlen_match1_match_nolit_encodeBetterBlockAsm12B
+	CMPL DI, $0x01
+	JE   matchlen_match1_match_nolit_encodeBetterBlockAsm12B
+	JB   match_nolit_end_encodeBetterBlockAsm12B
 	MOVW (R8)(R11*1), R10
 	CMPW (R9)(R11*1), R10
 	JNE  matchlen_match1_match_nolit_encodeBetterBlockAsm12B
-	SUBL $0x02, DI
 	LEAL 2(R11), R11
+	SUBL $0x02, DI
+	JZ   match_nolit_end_encodeBetterBlockAsm12B
 
 matchlen_match1_match_nolit_encodeBetterBlockAsm12B:
-	CMPL DI, $0x01
-	JB   match_nolit_end_encodeBetterBlockAsm12B
 	MOVB (R8)(R11*1), R10
 	CMPB (R9)(R11*1), R10
 	JNE  match_nolit_end_encodeBetterBlockAsm12B
@@ -8813,7 +8800,6 @@ matchlen_loop_match_nolit_encodeBetterBlockAsm10B:
 	LEAL 8(R11), R11
 	CMPL DI, $0x08
 	JAE  matchlen_loopback_match_nolit_encodeBetterBlockAsm10B
-	JZ   match_nolit_end_encodeBetterBlockAsm10B
 
 matchlen_match4_match_nolit_encodeBetterBlockAsm10B:
 	CMPL DI, $0x04
@@ -8821,21 +8807,21 @@ matchlen_match4_match_nolit_encodeBetterBlockAsm10B:
 	MOVL (R8)(R11*1), R10
 	CMPL (R9)(R11*1), R10
 	JNE  matchlen_match2_match_nolit_encodeBetterBlockAsm10B
-	SUBL $0x04, DI
+	LEAL -4(DI), DI
 	LEAL 4(R11), R11
 
 matchlen_match2_match_nolit_encodeBetterBlockAsm10B:
-	CMPL DI, $0x02
-	JB   matchlen_match1_match_nolit_encodeBetterBlockAsm10B
+	CMPL DI, $0x01
+	JE   matchlen_match1_match_nolit_encodeBetterBlockAsm10B
+	JB   match_nolit_end_encodeBetterBlockAsm10B
 	MOVW (R8)(R11*1), R10
 	CMPW (R9)(R11*1), R10
 	JNE  matchlen_match1_match_nolit_encodeBetterBlockAsm10B
-	SUBL $0x02, DI
 	LEAL 2(R11), R11
+	SUBL $0x02, DI
+	JZ   match_nolit_end_encodeBetterBlockAsm10B
 
 matchlen_match1_match_nolit_encodeBetterBlockAsm10B:
-	CMPL DI, $0x01
-	JB   match_nolit_end_encodeBetterBlockAsm10B
 	MOVB (R8)(R11*1), R10
 	CMPB (R9)(R11*1), R10
 	JNE  match_nolit_end_encodeBetterBlockAsm10B
@@ -9665,7 +9651,6 @@ matchlen_loop_match_nolit_encodeBetterBlockAsm8B:
 	LEAL 8(R11), R11
 	CMPL DI, $0x08
 	JAE  matchlen_loopback_match_nolit_encodeBetterBlockAsm8B
-	JZ   match_nolit_end_encodeBetterBlockAsm8B
 
 matchlen_match4_match_nolit_encodeBetterBlockAsm8B:
 	CMPL DI, $0x04
@@ -9673,21 +9658,21 @@ matchlen_match4_match_nolit_encodeBetterBlockAsm8B:
 	MOVL (R8)(R11*1), R10
 	CMPL (R9)(R11*1), R10
 	JNE  matchlen_match2_match_nolit_encodeBetterBlockAsm8B
-	SUBL $0x04, DI
+	LEAL -4(DI), DI
 	LEAL 4(R11), R11
 
 matchlen_match2_match_nolit_encodeBetterBlockAsm8B:
-	CMPL DI, $0x02
-	JB   matchlen_match1_match_nolit_encodeBetterBlockAsm8B
+	CMPL DI, $0x01
+	JE   matchlen_match1_match_nolit_encodeBetterBlockAsm8B
+	JB   match_nolit_end_encodeBetterBlockAsm8B
 	MOVW (R8)(R11*1), R10
 	CMPW (R9)(R11*1), R10
 	JNE  matchlen_match1_match_nolit_encodeBetterBlockAsm8B
-	SUBL $0x02, DI
 	LEAL 2(R11), R11
+	SUBL $0x02, DI
+	JZ   match_nolit_end_encodeBetterBlockAsm8B
 
 matchlen_match1_match_nolit_encodeBetterBlockAsm8B:
-	CMPL DI, $0x01
-	JB   match_nolit_end_encodeBetterBlockAsm8B
 	MOVB (R8)(R11*1), R10
 	CMPB (R9)(R11*1), R10
 	JNE  match_nolit_end_encodeBetterBlockAsm8B
@@ -10615,7 +10600,6 @@ matchlen_loop_repeat_extend_encodeSnappyBlockAsm:
 	LEAL 8(R10), R10
 	CMPL DI, $0x08
 	JAE  matchlen_loopback_repeat_extend_encodeSnappyBlockAsm
-	JZ   repeat_extend_forward_end_encodeSnappyBlockAsm
 
 matchlen_match4_repeat_extend_encodeSnappyBlockAsm:
 	CMPL DI, $0x04
@@ -10623,21 +10607,21 @@ matchlen_match4_repeat_extend_encodeSnappyBlockAsm:
 	MOVL (R8)(R10*1), R9
 	CMPL (BX)(R10*1), R9
 	JNE  matchlen_match2_repeat_extend_encodeSnappyBlockAsm
-	SUBL $0x04, DI
+	LEAL -4(DI), DI
 	LEAL 4(R10), R10
 
 matchlen_match2_repeat_extend_encodeSnappyBlockAsm:
-	CMPL DI, $0x02
-	JB   matchlen_match1_repeat_extend_encodeSnappyBlockAsm
+	CMPL DI, $0x01
+	JE   matchlen_match1_repeat_extend_encodeSnappyBlockAsm
+	JB   repeat_extend_forward_end_encodeSnappyBlockAsm
 	MOVW (R8)(R10*1), R9
 	CMPW (BX)(R10*1), R9
 	JNE  matchlen_match1_repeat_extend_encodeSnappyBlockAsm
-	SUBL $0x02, DI
 	LEAL 2(R10), R10
+	SUBL $0x02, DI
+	JZ   repeat_extend_forward_end_encodeSnappyBlockAsm
 
 matchlen_match1_repeat_extend_encodeSnappyBlockAsm:
-	CMPL DI, $0x01
-	JB   repeat_extend_forward_end_encodeSnappyBlockAsm
 	MOVB (R8)(R10*1), R9
 	CMPB (BX)(R10*1), R9
 	JNE  repeat_extend_forward_end_encodeSnappyBlockAsm
@@ -10938,7 +10922,6 @@ matchlen_loop_match_nolit_encodeSnappyBlockAsm:
 	LEAL 8(R9), R9
 	CMPL SI, $0x08
 	JAE  matchlen_loopback_match_nolit_encodeSnappyBlockAsm
-	JZ   match_nolit_end_encodeSnappyBlockAsm
 
 matchlen_match4_match_nolit_encodeSnappyBlockAsm:
 	CMPL SI, $0x04
@@ -10946,21 +10929,21 @@ matchlen_match4_match_nolit_encodeSnappyBlockAsm:
 	MOVL (DI)(R9*1), R8
 	CMPL (BX)(R9*1), R8
 	JNE  matchlen_match2_match_nolit_encodeSnappyBlockAsm
-	SUBL $0x04, SI
+	LEAL -4(SI), SI
 	LEAL 4(R9), R9
 
 matchlen_match2_match_nolit_encodeSnappyBlockAsm:
-	CMPL SI, $0x02
-	JB   matchlen_match1_match_nolit_encodeSnappyBlockAsm
+	CMPL SI, $0x01
+	JE   matchlen_match1_match_nolit_encodeSnappyBlockAsm
+	JB   match_nolit_end_encodeSnappyBlockAsm
 	MOVW (DI)(R9*1), R8
 	CMPW (BX)(R9*1), R8
 	JNE  matchlen_match1_match_nolit_encodeSnappyBlockAsm
-	SUBL $0x02, SI
 	LEAL 2(R9), R9
+	SUBL $0x02, SI
+	JZ   match_nolit_end_encodeSnappyBlockAsm
 
 matchlen_match1_match_nolit_encodeSnappyBlockAsm:
-	CMPL SI, $0x01
-	JB   match_nolit_end_encodeSnappyBlockAsm
 	MOVB (DI)(R9*1), R8
 	CMPB (BX)(R9*1), R8
 	JNE  match_nolit_end_encodeSnappyBlockAsm
@@ -11479,7 +11462,6 @@ matchlen_loop_repeat_extend_encodeSnappyBlockAsm64K:
 	LEAL 8(R10), R10
 	CMPL DI, $0x08
 	JAE  matchlen_loopback_repeat_extend_encodeSnappyBlockAsm64K
-	JZ   repeat_extend_forward_end_encodeSnappyBlockAsm64K
 
 matchlen_match4_repeat_extend_encodeSnappyBlockAsm64K:
 	CMPL DI, $0x04
@@ -11487,21 +11469,21 @@ matchlen_match4_repeat_extend_encodeSnappyBlockAsm64K:
 	MOVL (R8)(R10*1), R9
 	CMPL (BX)(R10*1), R9
 	JNE  matchlen_match2_repeat_extend_encodeSnappyBlockAsm64K
-	SUBL $0x04, DI
+	LEAL -4(DI), DI
 	LEAL 4(R10), R10
 
 matchlen_match2_repeat_extend_encodeSnappyBlockAsm64K:
-	CMPL DI, $0x02
-	JB   matchlen_match1_repeat_extend_encodeSnappyBlockAsm64K
+	CMPL DI, $0x01
+	JE   matchlen_match1_repeat_extend_encodeSnappyBlockAsm64K
+	JB   repeat_extend_forward_end_encodeSnappyBlockAsm64K
 	MOVW (R8)(R10*1), R9
 	CMPW (BX)(R10*1), R9
 	JNE  matchlen_match1_repeat_extend_encodeSnappyBlockAsm64K
-	SUBL $0x02, DI
 	LEAL 2(R10), R10
+	SUBL $0x02, DI
+	JZ   repeat_extend_forward_end_encodeSnappyBlockAsm64K
 
 matchlen_match1_repeat_extend_encodeSnappyBlockAsm64K:
-	CMPL DI, $0x01
-	JB   repeat_extend_forward_end_encodeSnappyBlockAsm64K
 	MOVB (R8)(R10*1), R9
 	CMPB (BX)(R10*1), R9
 	JNE  repeat_extend_forward_end_encodeSnappyBlockAsm64K
@@ -11762,7 +11744,6 @@ matchlen_loop_match_nolit_encodeSnappyBlockAsm64K:
 	LEAL 8(R9), R9
 	CMPL SI, $0x08
 	JAE  matchlen_loopback_match_nolit_encodeSnappyBlockAsm64K
-	JZ   match_nolit_end_encodeSnappyBlockAsm64K
 
 matchlen_match4_match_nolit_encodeSnappyBlockAsm64K:
 	CMPL SI, $0x04
@@ -11770,21 +11751,21 @@ matchlen_match4_match_nolit_encodeSnappyBlockAsm64K:
 	MOVL (DI)(R9*1), R8
 	CMPL (BX)(R9*1), R8
 	JNE  matchlen_match2_match_nolit_encodeSnappyBlockAsm64K
-	SUBL $0x04, SI
+	LEAL -4(SI), SI
 	LEAL 4(R9), R9
 
 matchlen_match2_match_nolit_encodeSnappyBlockAsm64K:
-	CMPL SI, $0x02
-	JB   matchlen_match1_match_nolit_encodeSnappyBlockAsm64K
+	CMPL SI, $0x01
+	JE   matchlen_match1_match_nolit_encodeSnappyBlockAsm64K
+	JB   match_nolit_end_encodeSnappyBlockAsm64K
 	MOVW (DI)(R9*1), R8
 	CMPW (BX)(R9*1), R8
 	JNE  matchlen_match1_match_nolit_encodeSnappyBlockAsm64K
-	SUBL $0x02, SI
 	LEAL 2(R9), R9
+	SUBL $0x02, SI
+	JZ   match_nolit_end_encodeSnappyBlockAsm64K
 
 matchlen_match1_match_nolit_encodeSnappyBlockAsm64K:
-	CMPL SI, $0x01
-	JB   match_nolit_end_encodeSnappyBlockAsm64K
 	MOVB (DI)(R9*1), R8
 	CMPB (BX)(R9*1), R8
 	JNE  match_nolit_end_encodeSnappyBlockAsm64K
@@ -12263,7 +12244,6 @@ matchlen_loop_repeat_extend_encodeSnappyBlockAsm12B:
 	LEAL 8(R10), R10
 	CMPL DI, $0x08
 	JAE  matchlen_loopback_repeat_extend_encodeSnappyBlockAsm12B
-	JZ   repeat_extend_forward_end_encodeSnappyBlockAsm12B
 
 matchlen_match4_repeat_extend_encodeSnappyBlockAsm12B:
 	CMPL DI, $0x04
@@ -12271,21 +12251,21 @@ matchlen_match4_repeat_extend_encodeSnappyBlockAsm12B:
 	MOVL (R8)(R10*1), R9
 	CMPL (BX)(R10*1), R9
 	JNE  matchlen_match2_repeat_extend_encodeSnappyBlockAsm12B
-	SUBL $0x04, DI
+	LEAL -4(DI), DI
 	LEAL 4(R10), R10
 
 matchlen_match2_repeat_extend_encodeSnappyBlockAsm12B:
-	CMPL DI, $0x02
-	JB   matchlen_match1_repeat_extend_encodeSnappyBlockAsm12B
+	CMPL DI, $0x01
+	JE   matchlen_match1_repeat_extend_encodeSnappyBlockAsm12B
+	JB   repeat_extend_forward_end_encodeSnappyBlockAsm12B
 	MOVW (R8)(R10*1), R9
 	CMPW (BX)(R10*1), R9
 	JNE  matchlen_match1_repeat_extend_encodeSnappyBlockAsm12B
-	SUBL $0x02, DI
 	LEAL 2(R10), R10
+	SUBL $0x02, DI
+	JZ   repeat_extend_forward_end_encodeSnappyBlockAsm12B
 
 matchlen_match1_repeat_extend_encodeSnappyBlockAsm12B:
-	CMPL DI, $0x01
-	JB   repeat_extend_forward_end_encodeSnappyBlockAsm12B
 	MOVB (R8)(R10*1), R9
 	CMPB (BX)(R10*1), R9
 	JNE  repeat_extend_forward_end_encodeSnappyBlockAsm12B
@@ -12546,7 +12526,6 @@ matchlen_loop_match_nolit_encodeSnappyBlockAsm12B:
 	LEAL 8(R9), R9
 	CMPL SI, $0x08
 	JAE  matchlen_loopback_match_nolit_encodeSnappyBlockAsm12B
-	JZ   match_nolit_end_encodeSnappyBlockAsm12B
 
 matchlen_match4_match_nolit_encodeSnappyBlockAsm12B:
 	CMPL SI, $0x04
@@ -12554,21 +12533,21 @@ matchlen_match4_match_nolit_encodeSnappyBlockAsm12B:
 	MOVL (DI)(R9*1), R8
 	CMPL (BX)(R9*1), R8
 	JNE  matchlen_match2_match_nolit_encodeSnappyBlockAsm12B
-	SUBL $0x04, SI
+	LEAL -4(SI), SI
 	LEAL 4(R9), R9
 
 matchlen_match2_match_nolit_encodeSnappyBlockAsm12B:
-	CMPL SI, $0x02
-	JB   matchlen_match1_match_nolit_encodeSnappyBlockAsm12B
+	CMPL SI, $0x01
+	JE   matchlen_match1_match_nolit_encodeSnappyBlockAsm12B
+	JB   match_nolit_end_encodeSnappyBlockAsm12B
 	MOVW (DI)(R9*1), R8
 	CMPW (BX)(R9*1), R8
 	JNE  matchlen_match1_match_nolit_encodeSnappyBlockAsm12B
-	SUBL $0x02, SI
 	LEAL 2(R9), R9
+	SUBL $0x02, SI
+	JZ   match_nolit_end_encodeSnappyBlockAsm12B
 
 matchlen_match1_match_nolit_encodeSnappyBlockAsm12B:
-	CMPL SI, $0x01
-	JB   match_nolit_end_encodeSnappyBlockAsm12B
 	MOVB (DI)(R9*1), R8
 	CMPB (BX)(R9*1), R8
 	JNE  match_nolit_end_encodeSnappyBlockAsm12B
@@ -13047,7 +13026,6 @@ matchlen_loop_repeat_extend_encodeSnappyBlockAsm10B:
 	LEAL 8(R10), R10
 	CMPL DI, $0x08
 	JAE  matchlen_loopback_repeat_extend_encodeSnappyBlockAsm10B
-	JZ   repeat_extend_forward_end_encodeSnappyBlockAsm10B
 
 matchlen_match4_repeat_extend_encodeSnappyBlockAsm10B:
 	CMPL DI, $0x04
@@ -13055,21 +13033,21 @@ matchlen_match4_repeat_extend_encodeSnappyBlockAsm10B:
 	MOVL (R8)(R10*1), R9
 	CMPL (BX)(R10*1), R9
 	JNE  matchlen_match2_repeat_extend_encodeSnappyBlockAsm10B
-	SUBL $0x04, DI
+	LEAL -4(DI), DI
 	LEAL 4(R10), R10
 
 matchlen_match2_repeat_extend_encodeSnappyBlockAsm10B:
-	CMPL DI, $0x02
-	JB   matchlen_match1_repeat_extend_encodeSnappyBlockAsm10B
+	CMPL DI, $0x01
+	JE   matchlen_match1_repeat_extend_encodeSnappyBlockAsm10B
+	JB   repeat_extend_forward_end_encodeSnappyBlockAsm10B
 	MOVW (R8)(R10*1), R9
 	CMPW (BX)(R10*1), R9
 	JNE  matchlen_match1_repeat_extend_encodeSnappyBlockAsm10B
-	SUBL $0x02, DI
 	LEAL 2(R10), R10
+	SUBL $0x02, DI
+	JZ   repeat_extend_forward_end_encodeSnappyBlockAsm10B
 
 matchlen_match1_repeat_extend_encodeSnappyBlockAsm10B:
-	CMPL DI, $0x01
-	JB   repeat_extend_forward_end_encodeSnappyBlockAsm10B
 	MOVB (R8)(R10*1), R9
 	CMPB (BX)(R10*1), R9
 	JNE  repeat_extend_forward_end_encodeSnappyBlockAsm10B
@@ -13330,7 +13308,6 @@ matchlen_loop_match_nolit_encodeSnappyBlockAsm10B:
 	LEAL 8(R9), R9
 	CMPL SI, $0x08
 	JAE  matchlen_loopback_match_nolit_encodeSnappyBlockAsm10B
-	JZ   match_nolit_end_encodeSnappyBlockAsm10B
 
 matchlen_match4_match_nolit_encodeSnappyBlockAsm10B:
 	CMPL SI, $0x04
@@ -13338,21 +13315,21 @@ matchlen_match4_match_nolit_encodeSnappyBlockAsm10B:
 	MOVL (DI)(R9*1), R8
 	CMPL (BX)(R9*1), R8
 	JNE  matchlen_match2_match_nolit_encodeSnappyBlockAsm10B
-	SUBL $0x04, SI
+	LEAL -4(SI), SI
 	LEAL 4(R9), R9
 
 matchlen_match2_match_nolit_encodeSnappyBlockAsm10B:
-	CMPL SI, $0x02
-	JB   matchlen_match1_match_nolit_encodeSnappyBlockAsm10B
+	CMPL SI, $0x01
+	JE   matchlen_match1_match_nolit_encodeSnappyBlockAsm10B
+	JB   match_nolit_end_encodeSnappyBlockAsm10B
 	MOVW (DI)(R9*1), R8
 	CMPW (BX)(R9*1), R8
 	JNE  matchlen_match1_match_nolit_encodeSnappyBlockAsm10B
-	SUBL $0x02, SI
 	LEAL 2(R9), R9
+	SUBL $0x02, SI
+	JZ   match_nolit_end_encodeSnappyBlockAsm10B
 
 matchlen_match1_match_nolit_encodeSnappyBlockAsm10B:
-	CMPL SI, $0x01
-	JB   match_nolit_end_encodeSnappyBlockAsm10B
 	MOVB (DI)(R9*1), R8
 	CMPB (BX)(R9*1), R8
 	JNE  match_nolit_end_encodeSnappyBlockAsm10B
@@ -13831,7 +13808,6 @@ matchlen_loop_repeat_extend_encodeSnappyBlockAsm8B:
 	LEAL 8(R10), R10
 	CMPL DI, $0x08
 	JAE  matchlen_loopback_repeat_extend_encodeSnappyBlockAsm8B
-	JZ   repeat_extend_forward_end_encodeSnappyBlockAsm8B
 
 matchlen_match4_repeat_extend_encodeSnappyBlockAsm8B:
 	CMPL DI, $0x04
@@ -13839,21 +13815,21 @@ matchlen_match4_repeat_extend_encodeSnappyBlockAsm8B:
 	MOVL (R8)(R10*1), R9
 	CMPL (BX)(R10*1), R9
 	JNE  matchlen_match2_repeat_extend_encodeSnappyBlockAsm8B
-	SUBL $0x04, DI
+	LEAL -4(DI), DI
 	LEAL 4(R10), R10
 
 matchlen_match2_repeat_extend_encodeSnappyBlockAsm8B:
-	CMPL DI, $0x02
-	JB   matchlen_match1_repeat_extend_encodeSnappyBlockAsm8B
+	CMPL DI, $0x01
+	JE   matchlen_match1_repeat_extend_encodeSnappyBlockAsm8B
+	JB   repeat_extend_forward_end_encodeSnappyBlockAsm8B
 	MOVW (R8)(R10*1), R9
 	CMPW (BX)(R10*1), R9
 	JNE  matchlen_match1_repeat_extend_encodeSnappyBlockAsm8B
-	SUBL $0x02, DI
 	LEAL 2(R10), R10
+	SUBL $0x02, DI
+	JZ   repeat_extend_forward_end_encodeSnappyBlockAsm8B
 
 matchlen_match1_repeat_extend_encodeSnappyBlockAsm8B:
-	CMPL DI, $0x01
-	JB   repeat_extend_forward_end_encodeSnappyBlockAsm8B
 	MOVB (R8)(R10*1), R9
 	CMPB (BX)(R10*1), R9
 	JNE  repeat_extend_forward_end_encodeSnappyBlockAsm8B
@@ -14112,7 +14088,6 @@ matchlen_loop_match_nolit_encodeSnappyBlockAsm8B:
 	LEAL 8(R9), R9
 	CMPL SI, $0x08
 	JAE  matchlen_loopback_match_nolit_encodeSnappyBlockAsm8B
-	JZ   match_nolit_end_encodeSnappyBlockAsm8B
 
 matchlen_match4_match_nolit_encodeSnappyBlockAsm8B:
 	CMPL SI, $0x04
@@ -14120,21 +14095,21 @@ matchlen_match4_match_nolit_encodeSnappyBlockAsm8B:
 	MOVL (DI)(R9*1), R8
 	CMPL (BX)(R9*1), R8
 	JNE  matchlen_match2_match_nolit_encodeSnappyBlockAsm8B
-	SUBL $0x04, SI
+	LEAL -4(SI), SI
 	LEAL 4(R9), R9
 
 matchlen_match2_match_nolit_encodeSnappyBlockAsm8B:
-	CMPL SI, $0x02
-	JB   matchlen_match1_match_nolit_encodeSnappyBlockAsm8B
+	CMPL SI, $0x01
+	JE   matchlen_match1_match_nolit_encodeSnappyBlockAsm8B
+	JB   match_nolit_end_encodeSnappyBlockAsm8B
 	MOVW (DI)(R9*1), R8
 	CMPW (BX)(R9*1), R8
 	JNE  matchlen_match1_match_nolit_encodeSnappyBlockAsm8B
-	SUBL $0x02, SI
 	LEAL 2(R9), R9
+	SUBL $0x02, SI
+	JZ   match_nolit_end_encodeSnappyBlockAsm8B
 
 matchlen_match1_match_nolit_encodeSnappyBlockAsm8B:
-	CMPL SI, $0x01
-	JB   match_nolit_end_encodeSnappyBlockAsm8B
 	MOVB (DI)(R9*1), R8
 	CMPB (BX)(R9*1), R8
 	JNE  match_nolit_end_encodeSnappyBlockAsm8B
@@ -14523,7 +14498,6 @@ matchlen_loop_match_nolit_encodeSnappyBetterBlockAsm:
 	LEAL 8(R11), R11
 	CMPL DI, $0x08
 	JAE  matchlen_loopback_match_nolit_encodeSnappyBetterBlockAsm
-	JZ   match_nolit_end_encodeSnappyBetterBlockAsm
 
 matchlen_match4_match_nolit_encodeSnappyBetterBlockAsm:
 	CMPL DI, $0x04
@@ -14531,21 +14505,21 @@ matchlen_match4_match_nolit_encodeSnappyBetterBlockAsm:
 	MOVL (R8)(R11*1), R10
 	CMPL (R9)(R11*1), R10
 	JNE  matchlen_match2_match_nolit_encodeSnappyBetterBlockAsm
-	SUBL $0x04, DI
+	LEAL -4(DI), DI
 	LEAL 4(R11), R11
 
 matchlen_match2_match_nolit_encodeSnappyBetterBlockAsm:
-	CMPL DI, $0x02
-	JB   matchlen_match1_match_nolit_encodeSnappyBetterBlockAsm
+	CMPL DI, $0x01
+	JE   matchlen_match1_match_nolit_encodeSnappyBetterBlockAsm
+	JB   match_nolit_end_encodeSnappyBetterBlockAsm
 	MOVW (R8)(R11*1), R10
 	CMPW (R9)(R11*1), R10
 	JNE  matchlen_match1_match_nolit_encodeSnappyBetterBlockAsm
-	SUBL $0x02, DI
 	LEAL 2(R11), R11
+	SUBL $0x02, DI
+	JZ   match_nolit_end_encodeSnappyBetterBlockAsm
 
 matchlen_match1_match_nolit_encodeSnappyBetterBlockAsm:
-	CMPL DI, $0x01
-	JB   match_nolit_end_encodeSnappyBetterBlockAsm
 	MOVB (R8)(R11*1), R10
 	CMPB (R9)(R11*1), R10
 	JNE  match_nolit_end_encodeSnappyBetterBlockAsm
@@ -15147,7 +15121,6 @@ matchlen_loop_match_nolit_encodeSnappyBetterBlockAsm64K:
 	LEAL 8(R11), R11
 	CMPL DI, $0x08
 	JAE  matchlen_loopback_match_nolit_encodeSnappyBetterBlockAsm64K
-	JZ   match_nolit_end_encodeSnappyBetterBlockAsm64K
 
 matchlen_match4_match_nolit_encodeSnappyBetterBlockAsm64K:
 	CMPL DI, $0x04
@@ -15155,21 +15128,21 @@ matchlen_match4_match_nolit_encodeSnappyBetterBlockAsm64K:
 	MOVL (R8)(R11*1), R10
 	CMPL (R9)(R11*1), R10
 	JNE  matchlen_match2_match_nolit_encodeSnappyBetterBlockAsm64K
-	SUBL $0x04, DI
+	LEAL -4(DI), DI
 	LEAL 4(R11), R11
 
 matchlen_match2_match_nolit_encodeSnappyBetterBlockAsm64K:
-	CMPL DI, $0x02
-	JB   matchlen_match1_match_nolit_encodeSnappyBetterBlockAsm64K
+	CMPL DI, $0x01
+	JE   matchlen_match1_match_nolit_encodeSnappyBetterBlockAsm64K
+	JB   match_nolit_end_encodeSnappyBetterBlockAsm64K
 	MOVW (R8)(R11*1), R10
 	CMPW (R9)(R11*1), R10
 	JNE  matchlen_match1_match_nolit_encodeSnappyBetterBlockAsm64K
-	SUBL $0x02, DI
 	LEAL 2(R11), R11
+	SUBL $0x02, DI
+	JZ   match_nolit_end_encodeSnappyBetterBlockAsm64K
 
 matchlen_match1_match_nolit_encodeSnappyBetterBlockAsm64K:
-	CMPL DI, $0x01
-	JB   match_nolit_end_encodeSnappyBetterBlockAsm64K
 	MOVB (R8)(R11*1), R10
 	CMPB (R9)(R11*1), R10
 	JNE  match_nolit_end_encodeSnappyBetterBlockAsm64K
@@ -15706,7 +15679,6 @@ matchlen_loop_match_nolit_encodeSnappyBetterBlockAsm12B:
 	LEAL 8(R11), R11
 	CMPL DI, $0x08
 	JAE  matchlen_loopback_match_nolit_encodeSnappyBetterBlockAsm12B
-	JZ   match_nolit_end_encodeSnappyBetterBlockAsm12B
 
 matchlen_match4_match_nolit_encodeSnappyBetterBlockAsm12B:
 	CMPL DI, $0x04
@@ -15714,21 +15686,21 @@ matchlen_match4_match_nolit_encodeSnappyBetterBlockAsm12B:
 	MOVL (R8)(R11*1), R10
 	CMPL (R9)(R11*1), R10
 	JNE  matchlen_match2_match_nolit_encodeSnappyBetterBlockAsm12B
-	SUBL $0x04, DI
+	LEAL -4(DI), DI
 	LEAL 4(R11), R11
 
 matchlen_match2_match_nolit_encodeSnappyBetterBlockAsm12B:
-	CMPL DI, $0x02
-	JB   matchlen_match1_match_nolit_encodeSnappyBetterBlockAsm12B
+	CMPL DI, $0x01
+	JE   matchlen_match1_match_nolit_encodeSnappyBetterBlockAsm12B
+	JB   match_nolit_end_encodeSnappyBetterBlockAsm12B
 	MOVW (R8)(R11*1), R10
 	CMPW (R9)(R11*1), R10
 	JNE  matchlen_match1_match_nolit_encodeSnappyBetterBlockAsm12B
-	SUBL $0x02, DI
 	LEAL 2(R11), R11
+	SUBL $0x02, DI
+	JZ   match_nolit_end_encodeSnappyBetterBlockAsm12B
 
 matchlen_match1_match_nolit_encodeSnappyBetterBlockAsm12B:
-	CMPL DI, $0x01
-	JB   match_nolit_end_encodeSnappyBetterBlockAsm12B
 	MOVB (R8)(R11*1), R10
 	CMPB (R9)(R11*1), R10
 	JNE  match_nolit_end_encodeSnappyBetterBlockAsm12B
@@ -16265,7 +16237,6 @@ matchlen_loop_match_nolit_encodeSnappyBetterBlockAsm10B:
 	LEAL 8(R11), R11
 	CMPL DI, $0x08
 	JAE  matchlen_loopback_match_nolit_encodeSnappyBetterBlockAsm10B
-	JZ   match_nolit_end_encodeSnappyBetterBlockAsm10B
 
 matchlen_match4_match_nolit_encodeSnappyBetterBlockAsm10B:
 	CMPL DI, $0x04
@@ -16273,21 +16244,21 @@ matchlen_match4_match_nolit_encodeSnappyBetterBlockAsm10B:
 	MOVL (R8)(R11*1), R10
 	CMPL (R9)(R11*1), R10
 	JNE  matchlen_match2_match_nolit_encodeSnappyBetterBlockAsm10B
-	SUBL $0x04, DI
+	LEAL -4(DI), DI
 	LEAL 4(R11), R11
 
 matchlen_match2_match_nolit_encodeSnappyBetterBlockAsm10B:
-	CMPL DI, $0x02
-	JB   matchlen_match1_match_nolit_encodeSnappyBetterBlockAsm10B
+	CMPL DI, $0x01
+	JE   matchlen_match1_match_nolit_encodeSnappyBetterBlockAsm10B
+	JB   match_nolit_end_encodeSnappyBetterBlockAsm10B
 	MOVW (R8)(R11*1), R10
 	CMPW (R9)(R11*1), R10
 	JNE  matchlen_match1_match_nolit_encodeSnappyBetterBlockAsm10B
-	SUBL $0x02, DI
 	LEAL 2(R11), R11
+	SUBL $0x02, DI
+	JZ   match_nolit_end_encodeSnappyBetterBlockAsm10B
 
 matchlen_match1_match_nolit_encodeSnappyBetterBlockAsm10B:
-	CMPL DI, $0x01
-	JB   match_nolit_end_encodeSnappyBetterBlockAsm10B
 	MOVB (R8)(R11*1), R10
 	CMPB (R9)(R11*1), R10
 	JNE  match_nolit_end_encodeSnappyBetterBlockAsm10B
@@ -16824,7 +16795,6 @@ matchlen_loop_match_nolit_encodeSnappyBetterBlockAsm8B:
 	LEAL 8(R11), R11
 	CMPL DI, $0x08
 	JAE  matchlen_loopback_match_nolit_encodeSnappyBetterBlockAsm8B
-	JZ   match_nolit_end_encodeSnappyBetterBlockAsm8B
 
 matchlen_match4_match_nolit_encodeSnappyBetterBlockAsm8B:
 	CMPL DI, $0x04
@@ -16832,21 +16802,21 @@ matchlen_match4_match_nolit_encodeSnappyBetterBlockAsm8B:
 	MOVL (R8)(R11*1), R10
 	CMPL (R9)(R11*1), R10
 	JNE  matchlen_match2_match_nolit_encodeSnappyBetterBlockAsm8B
-	SUBL $0x04, DI
+	LEAL -4(DI), DI
 	LEAL 4(R11), R11
 
 matchlen_match2_match_nolit_encodeSnappyBetterBlockAsm8B:
-	CMPL DI, $0x02
-	JB   matchlen_match1_match_nolit_encodeSnappyBetterBlockAsm8B
+	CMPL DI, $0x01
+	JE   matchlen_match1_match_nolit_encodeSnappyBetterBlockAsm8B
+	JB   match_nolit_end_encodeSnappyBetterBlockAsm8B
 	MOVW (R8)(R11*1), R10
 	CMPW (R9)(R11*1), R10
 	JNE  matchlen_match1_match_nolit_encodeSnappyBetterBlockAsm8B
-	SUBL $0x02, DI
 	LEAL 2(R11), R11
+	SUBL $0x02, DI
+	JZ   match_nolit_end_encodeSnappyBetterBlockAsm8B
 
 matchlen_match1_match_nolit_encodeSnappyBetterBlockAsm8B:
-	CMPL DI, $0x01
-	JB   match_nolit_end_encodeSnappyBetterBlockAsm8B
 	MOVB (R8)(R11*1), R10
 	CMPB (R9)(R11*1), R10
 	JNE  match_nolit_end_encodeSnappyBetterBlockAsm8B
@@ -17398,7 +17368,6 @@ matchlen_loop_repeat_extend_calcBlockSize:
 	LEAL 8(R10), R10
 	CMPL DI, $0x08
 	JAE  matchlen_loopback_repeat_extend_calcBlockSize
-	JZ   repeat_extend_forward_end_calcBlockSize
 
 matchlen_match4_repeat_extend_calcBlockSize:
 	CMPL DI, $0x04
@@ -17406,21 +17375,21 @@ matchlen_match4_repeat_extend_calcBlockSize:
 	MOVL (R8)(R10*1), R9
 	CMPL (BX)(R10*1), R9
 	JNE  matchlen_match2_repeat_extend_calcBlockSize
-	SUBL $0x04, DI
+	LEAL -4(DI), DI
 	LEAL 4(R10), R10
 
 matchlen_match2_repeat_extend_calcBlockSize:
-	CMPL DI, $0x02
-	JB   matchlen_match1_repeat_extend_calcBlockSize
+	CMPL DI, $0x01
+	JE   matchlen_match1_repeat_extend_calcBlockSize
+	JB   repeat_extend_forward_end_calcBlockSize
 	MOVW (R8)(R10*1), R9
 	CMPW (BX)(R10*1), R9
 	JNE  matchlen_match1_repeat_extend_calcBlockSize
-	SUBL $0x02, DI
 	LEAL 2(R10), R10
+	SUBL $0x02, DI
+	JZ   repeat_extend_forward_end_calcBlockSize
 
 matchlen_match1_repeat_extend_calcBlockSize:
-	CMPL DI, $0x01
-	JB   repeat_extend_forward_end_calcBlockSize
 	MOVB (R8)(R10*1), R9
 	CMPB (BX)(R10*1), R9
 	JNE  repeat_extend_forward_end_calcBlockSize
@@ -17610,7 +17579,6 @@ matchlen_loop_match_nolit_calcBlockSize:
 	LEAL 8(R9), R9
 	CMPL SI, $0x08
 	JAE  matchlen_loopback_match_nolit_calcBlockSize
-	JZ   match_nolit_end_calcBlockSize
 
 matchlen_match4_match_nolit_calcBlockSize:
 	CMPL SI, $0x04
@@ -17618,21 +17586,21 @@ matchlen_match4_match_nolit_calcBlockSize:
 	MOVL (DI)(R9*1), R8
 	CMPL (BX)(R9*1), R8
 	JNE  matchlen_match2_match_nolit_calcBlockSize
-	SUBL $0x04, SI
+	LEAL -4(SI), SI
 	LEAL 4(R9), R9
 
 matchlen_match2_match_nolit_calcBlockSize:
-	CMPL SI, $0x02
-	JB   matchlen_match1_match_nolit_calcBlockSize
+	CMPL SI, $0x01
+	JE   matchlen_match1_match_nolit_calcBlockSize
+	JB   match_nolit_end_calcBlockSize
 	MOVW (DI)(R9*1), R8
 	CMPW (BX)(R9*1), R8
 	JNE  matchlen_match1_match_nolit_calcBlockSize
-	SUBL $0x02, SI
 	LEAL 2(R9), R9
+	SUBL $0x02, SI
+	JZ   match_nolit_end_calcBlockSize
 
 matchlen_match1_match_nolit_calcBlockSize:
-	CMPL SI, $0x01
-	JB   match_nolit_end_calcBlockSize
 	MOVB (DI)(R9*1), R8
 	CMPB (BX)(R9*1), R8
 	JNE  match_nolit_end_calcBlockSize
@@ -17929,7 +17897,6 @@ matchlen_loop_repeat_extend_calcBlockSizeSmall:
 	LEAL 8(R10), R10
 	CMPL DI, $0x08
 	JAE  matchlen_loopback_repeat_extend_calcBlockSizeSmall
-	JZ   repeat_extend_forward_end_calcBlockSizeSmall
 
 matchlen_match4_repeat_extend_calcBlockSizeSmall:
 	CMPL DI, $0x04
@@ -17937,21 +17904,21 @@ matchlen_match4_repeat_extend_calcBlockSizeSmall:
 	MOVL (R8)(R10*1), R9
 	CMPL (BX)(R10*1), R9
 	JNE  matchlen_match2_repeat_extend_calcBlockSizeSmall
-	SUBL $0x04, DI
+	LEAL -4(DI), DI
 	LEAL 4(R10), R10
 
 matchlen_match2_repeat_extend_calcBlockSizeSmall:
-	CMPL DI, $0x02
-	JB   matchlen_match1_repeat_extend_calcBlockSizeSmall
+	CMPL DI, $0x01
+	JE   matchlen_match1_repeat_extend_calcBlockSizeSmall
+	JB   repeat_extend_forward_end_calcBlockSizeSmall
 	MOVW (R8)(R10*1), R9
 	CMPW (BX)(R10*1), R9
 	JNE  matchlen_match1_repeat_extend_calcBlockSizeSmall
-	SUBL $0x02, DI
 	LEAL 2(R10), R10
+	SUBL $0x02, DI
+	JZ   repeat_extend_forward_end_calcBlockSizeSmall
 
 matchlen_match1_repeat_extend_calcBlockSizeSmall:
-	CMPL DI, $0x01
-	JB   repeat_extend_forward_end_calcBlockSizeSmall
 	MOVB (R8)(R10*1), R9
 	CMPB (BX)(R10*1), R9
 	JNE  repeat_extend_forward_end_calcBlockSizeSmall
@@ -18111,7 +18078,6 @@ matchlen_loop_match_nolit_calcBlockSizeSmall:
 	LEAL 8(R9), R9
 	CMPL SI, $0x08
 	JAE  matchlen_loopback_match_nolit_calcBlockSizeSmall
-	JZ   match_nolit_end_calcBlockSizeSmall
 
 matchlen_match4_match_nolit_calcBlockSizeSmall:
 	CMPL SI, $0x04
@@ -18119,21 +18085,21 @@ matchlen_match4_match_nolit_calcBlockSizeSmall:
 	MOVL (DI)(R9*1), R8
 	CMPL (BX)(R9*1), R8
 	JNE  matchlen_match2_match_nolit_calcBlockSizeSmall
-	SUBL $0x04, SI
+	LEAL -4(SI), SI
 	LEAL 4(R9), R9
 
 matchlen_match2_match_nolit_calcBlockSizeSmall:
-	CMPL SI, $0x02
-	JB   matchlen_match1_match_nolit_calcBlockSizeSmall
+	CMPL SI, $0x01
+	JE   matchlen_match1_match_nolit_calcBlockSizeSmall
+	JB   match_nolit_end_calcBlockSizeSmall
 	MOVW (DI)(R9*1), R8
 	CMPW (BX)(R9*1), R8
 	JNE  matchlen_match1_match_nolit_calcBlockSizeSmall
-	SUBL $0x02, SI
 	LEAL 2(R9), R9
+	SUBL $0x02, SI
+	JZ   match_nolit_end_calcBlockSizeSmall
 
 matchlen_match1_match_nolit_calcBlockSizeSmall:
-	CMPL SI, $0x01
-	JB   match_nolit_end_calcBlockSizeSmall
 	MOVB (DI)(R9*1), R8
 	CMPB (BX)(R9*1), R8
 	JNE  match_nolit_end_calcBlockSizeSmall
@@ -18899,7 +18865,6 @@ matchlen_loop_standalone:
 	LEAL 8(SI), SI
 	CMPL DX, $0x08
 	JAE  matchlen_loopback_standalone
-	JZ   gen_match_len_end
 
 matchlen_match4_standalone:
 	CMPL DX, $0x04
@@ -18907,21 +18872,21 @@ matchlen_match4_standalone:
 	MOVL (AX)(SI*1), BX
 	CMPL (CX)(SI*1), BX
 	JNE  matchlen_match2_standalone
-	SUBL $0x04, DX
+	LEAL -4(DX), DX
 	LEAL 4(SI), SI
 
 matchlen_match2_standalone:
-	CMPL DX, $0x02
-	JB   matchlen_match1_standalone
+	CMPL DX, $0x01
+	JE   matchlen_match1_standalone
+	JB   gen_match_len_end
 	MOVW (AX)(SI*1), BX
 	CMPW (CX)(SI*1), BX
 	JNE  matchlen_match1_standalone
-	SUBL $0x02, DX
 	LEAL 2(SI), SI
+	SUBL $0x02, DX
+	JZ   gen_match_len_end
 
 matchlen_match1_standalone:
-	CMPL DX, $0x01
-	JB   gen_match_len_end
 	MOVB (AX)(SI*1), BL
 	CMPB (CX)(SI*1), BL
 	JNE  gen_match_len_end