diff --git a/zstd/_generate/gen.go b/zstd/_generate/gen.go
index ff0a1a9b72..23a26b21d4 100644
--- a/zstd/_generate/gen.go
+++ b/zstd/_generate/gen.go
@@ -302,28 +302,25 @@ func (o options) generateBody(name string, executeSingleTriple func(ctx *execute
 			// Update states
 			Comment("Update Offset State")
 			{
-				nBits := GP64()
+				nBits := ofState // Note: SHRXQ uses lower 6 bits of shift amount and BZHIQ lower 8 bits of count
 				lowBits := GP64()
-				MOVBQZX(ofState.As8(), nBits)
 				BZHIQ(nBits, bits, lowBits) // lowBits = bits & ((1 << nBits) - 1))
-				SHRXQ(nBits, bits, bits)    // bits >= nBits
+				SHRXQ(nBits, bits, bits)    // bits >>= nBits
 				o.nextState(name+"_ofState", ofState, lowBits, "ofTable")
 			}
 			Comment("Update Match Length State")
 			{
-				nBits := GP64()
+				nBits := mlState
 				lowBits := GP64()
-				MOVBQZX(mlState.As8(), nBits)
-				BZHIQ(nBits, bits, lowBits) // lowBits = lowBits & ((1 << nBits) - 1))
-				SHRXQ(nBits, bits, bits)    // lowBits >= nBits
+				BZHIQ(nBits, bits, lowBits) // lowBits = bits & ((1 << nBits) - 1))
+				SHRXQ(nBits, bits, bits)    // lowBits >>= nBits
 				o.nextState(name+"_mlState", mlState, lowBits, "mlTable")
 			}
 			Comment("Update Literal Length State")
 			{
-				nBits := GP64()
+				nBits := llState
 				lowBits := GP64()
-				MOVBQZX(llState.As8(), nBits)
-				BZHIQ(nBits, bits, lowBits) // lowBits = lowBits & ((1 << nBits) - 1))
+				BZHIQ(nBits, bits, lowBits) // lowBits = bits & ((1 << nBits) - 1))
 				o.nextState(name+"_llState", llState, lowBits, "llTable")
 			}
 		} else {
diff --git a/zstd/seqdec_amd64.s b/zstd/seqdec_amd64.s
index e7ad2a0ca2..212c6cac30 100644
--- a/zstd/seqdec_amd64.s
+++ b/zstd/seqdec_amd64.s
@@ -721,37 +721,34 @@ sequenceDecs_decode_bmi2_fill_2_end:
 	BZHIQ   R14, R15, R15
 
 	// Update Offset State
-	MOVBQZX R8, CX
-	BZHIQ   CX, R15, R14
-	SHRXQ   CX, R15, R15
-	MOVQ    $0x00001010, CX
-	BEXTRQ  CX, R8, CX
-	ADDQ    R14, CX
+	BZHIQ  R8, R15, CX
+	SHRXQ  R8, R15, R15
+	MOVQ   $0x00001010, R14
+	BEXTRQ R14, R8, R8
+	ADDQ   CX, R8
 
 	// Load ctx.ofTable
-	MOVQ ctx+16(FP), R8
-	MOVQ 48(R8), R8
-	MOVQ (R8)(CX*8), R8
+	MOVQ ctx+16(FP), CX
+	MOVQ 48(CX), CX
+	MOVQ (CX)(R8*8), R8
 
 	// Update Match Length State
-	MOVBQZX DI, CX
-	BZHIQ   CX, R15, R14
-	SHRXQ   CX, R15, R15
-	MOVQ    $0x00001010, CX
-	BEXTRQ  CX, DI, CX
-	ADDQ    R14, CX
+	BZHIQ  DI, R15, CX
+	SHRXQ  DI, R15, R15
+	MOVQ   $0x00001010, R14
+	BEXTRQ R14, DI, DI
+	ADDQ   CX, DI
 
 	// Load ctx.mlTable
-	MOVQ ctx+16(FP), DI
-	MOVQ 24(DI), DI
-	MOVQ (DI)(CX*8), DI
+	MOVQ ctx+16(FP), CX
+	MOVQ 24(CX), CX
+	MOVQ (CX)(DI*8), DI
 
 	// Update Literal Length State
-	MOVBQZX SI, CX
-	BZHIQ   CX, R15, CX
-	MOVQ    $0x00001010, R14
-	BEXTRQ  R14, SI, SI
-	ADDQ    CX, SI
+	BZHIQ  SI, R15, CX
+	MOVQ   $0x00001010, R14
+	BEXTRQ R14, SI, SI
+	ADDQ   CX, SI
 
 	// Load ctx.llTable
 	MOVQ ctx+16(FP), CX
@@ -979,37 +976,34 @@ sequenceDecs_decode_56_bmi2_fill_end:
 	BZHIQ   R14, R15, R15
 
 	// Update Offset State
-	MOVBQZX R8, CX
-	BZHIQ   CX, R15, R14
-	SHRXQ   CX, R15, R15
-	MOVQ    $0x00001010, CX
-	BEXTRQ  CX, R8, CX
-	ADDQ    R14, CX
+	BZHIQ  R8, R15, CX
+	SHRXQ  R8, R15, R15
+	MOVQ   $0x00001010, R14
+	BEXTRQ R14, R8, R8
+	ADDQ   CX, R8
 
 	// Load ctx.ofTable
-	MOVQ ctx+16(FP), R8
-	MOVQ 48(R8), R8
-	MOVQ (R8)(CX*8), R8
+	MOVQ ctx+16(FP), CX
+	MOVQ 48(CX), CX
+	MOVQ (CX)(R8*8), R8
 
 	// Update Match Length State
-	MOVBQZX DI, CX
-	BZHIQ   CX, R15, R14
-	SHRXQ   CX, R15, R15
-	MOVQ    $0x00001010, CX
-	BEXTRQ  CX, DI, CX
-	ADDQ    R14, CX
+	BZHIQ  DI, R15, CX
+	SHRXQ  DI, R15, R15
+	MOVQ   $0x00001010, R14
+	BEXTRQ R14, DI, DI
+	ADDQ   CX, DI
 
 	// Load ctx.mlTable
-	MOVQ ctx+16(FP), DI
-	MOVQ 24(DI), DI
-	MOVQ (DI)(CX*8), DI
+	MOVQ ctx+16(FP), CX
+	MOVQ 24(CX), CX
+	MOVQ (CX)(DI*8), DI
 
 	// Update Literal Length State
-	MOVBQZX SI, CX
-	BZHIQ   CX, R15, CX
-	MOVQ    $0x00001010, R14
-	BEXTRQ  R14, SI, SI
-	ADDQ    CX, SI
+	BZHIQ  SI, R15, CX
+	MOVQ   $0x00001010, R14
+	BEXTRQ R14, SI, SI
+	ADDQ   CX, SI
 
 	// Load ctx.llTable
 	MOVQ ctx+16(FP), CX
@@ -2277,37 +2271,34 @@ sequenceDecs_decodeSync_bmi2_fill_2_end:
 	BZHIQ   R13, R14, R14
 
 	// Update Offset State
-	MOVBQZX R8, CX
-	BZHIQ   CX, R14, R13
-	SHRXQ   CX, R14, R14
-	MOVQ    $0x00001010, CX
-	BEXTRQ  CX, R8, CX
-	ADDQ    R13, CX
+	BZHIQ  R8, R14, CX
+	SHRXQ  R8, R14, R14
+	MOVQ   $0x00001010, R13
+	BEXTRQ R13, R8, R8
+	ADDQ   CX, R8
 
 	// Load ctx.ofTable
-	MOVQ ctx+16(FP), R8
-	MOVQ 48(R8), R8
-	MOVQ (R8)(CX*8), R8
+	MOVQ ctx+16(FP), CX
+	MOVQ 48(CX), CX
+	MOVQ (CX)(R8*8), R8
 
 	// Update Match Length State
-	MOVBQZX DI, CX
-	BZHIQ   CX, R14, R13
-	SHRXQ   CX, R14, R14
-	MOVQ    $0x00001010, CX
-	BEXTRQ  CX, DI, CX
-	ADDQ    R13, CX
+	BZHIQ  DI, R14, CX
+	SHRXQ  DI, R14, R14
+	MOVQ   $0x00001010, R13
+	BEXTRQ R13, DI, DI
+	ADDQ   CX, DI
 
 	// Load ctx.mlTable
-	MOVQ ctx+16(FP), DI
-	MOVQ 24(DI), DI
-	MOVQ (DI)(CX*8), DI
+	MOVQ ctx+16(FP), CX
+	MOVQ 24(CX), CX
+	MOVQ (CX)(DI*8), DI
 
 	// Update Literal Length State
-	MOVBQZX SI, CX
-	BZHIQ   CX, R14, CX
-	MOVQ    $0x00001010, R13
-	BEXTRQ  R13, SI, SI
-	ADDQ    CX, SI
+	BZHIQ  SI, R14, CX
+	MOVQ   $0x00001010, R13
+	BEXTRQ R13, SI, SI
+	ADDQ   CX, SI
 
 	// Load ctx.llTable
 	MOVQ ctx+16(FP), CX
@@ -3310,37 +3301,34 @@ sequenceDecs_decodeSync_safe_bmi2_fill_2_end:
 	BZHIQ   R13, R14, R14
 
 	// Update Offset State
-	MOVBQZX R8, CX
-	BZHIQ   CX, R14, R13
-	SHRXQ   CX, R14, R14
-	MOVQ    $0x00001010, CX
-	BEXTRQ  CX, R8, CX
-	ADDQ    R13, CX
+	BZHIQ  R8, R14, CX
+	SHRXQ  R8, R14, R14
+	MOVQ   $0x00001010, R13
+	BEXTRQ R13, R8, R8
+	ADDQ   CX, R8
 
 	// Load ctx.ofTable
-	MOVQ ctx+16(FP), R8
-	MOVQ 48(R8), R8
-	MOVQ (R8)(CX*8), R8
+	MOVQ ctx+16(FP), CX
+	MOVQ 48(CX), CX
+	MOVQ (CX)(R8*8), R8
 
 	// Update Match Length State
-	MOVBQZX DI, CX
-	BZHIQ   CX, R14, R13
-	SHRXQ   CX, R14, R14
-	MOVQ    $0x00001010, CX
-	BEXTRQ  CX, DI, CX
-	ADDQ    R13, CX
+	BZHIQ  DI, R14, CX
+	SHRXQ  DI, R14, R14
+	MOVQ   $0x00001010, R13
+	BEXTRQ R13, DI, DI
+	ADDQ   CX, DI
 
 	// Load ctx.mlTable
-	MOVQ ctx+16(FP), DI
-	MOVQ 24(DI), DI
-	MOVQ (DI)(CX*8), DI
+	MOVQ ctx+16(FP), CX
+	MOVQ 24(CX), CX
+	MOVQ (CX)(DI*8), DI
 
 	// Update Literal Length State
-	MOVBQZX SI, CX
-	BZHIQ   CX, R14, CX
-	MOVQ    $0x00001010, R13
-	BEXTRQ  R13, SI, SI
-	ADDQ    CX, SI
+	BZHIQ  SI, R14, CX
+	MOVQ   $0x00001010, R13
+	BEXTRQ R13, SI, SI
+	ADDQ   CX, SI
 
 	// Load ctx.llTable
 	MOVQ ctx+16(FP), CX