diff --git a/zstd/_generate/gen.go b/zstd/_generate/gen.go
index 50ce4404f0..7ef9a45aa8 100644
--- a/zstd/_generate/gen.go
+++ b/zstd/_generate/gen.go
@@ -156,8 +156,8 @@ func (o options) generateBody(name string, executeSingleTriple func(ctx *execute
 		brPointer := GP64()
 		Load(br.Field("value"), brValue)
 		Load(br.Field("bitsRead"), brBitsRead)
-		Load(br.Field("off"), brOffset)
 		Load(br.Field("in").Base(), brPointer)
+		Load(br.Field("in").Len(), brOffset)
 		ADDQ(brOffset, brPointer) // Add current offset to read pointer.
 		MOVQ(brPointer, brPointerStash)
 	}
@@ -438,7 +438,7 @@ func (o options) generateBody(name string, executeSingleTriple func(ctx *execute
 	br := Dereference(Param("br"))
 	Store(brValue, br.Field("value"))
 	Store(brBitsRead.As8(), br.Field("bitsRead"))
-	Store(brOffset, br.Field("off"))
+	Store(brOffset, br.Field("in").Len())
 
 	if !o.useSeqs {
 		Comment("Update the context")
diff --git a/zstd/bitreader.go b/zstd/bitreader.go
index 97299d499c..25ca983941 100644
--- a/zstd/bitreader.go
+++ b/zstd/bitreader.go
@@ -17,7 +17,6 @@ import (
 // for aligning the input.
 type bitReader struct {
 	in       []byte
-	off      uint   // next byte to read is at in[off - 1]
 	value    uint64 // Maybe use [16]byte, but shifting is awkward.
 	bitsRead uint8
 }
@@ -28,7 +27,6 @@ func (b *bitReader) init(in []byte) error {
 		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 {
@@ -69,21 +67,19 @@ func (b *bitReader) fillFast() {
 	if b.bitsRead < 32 {
 		return
 	}
-	// 2 bounds checks.
-	v := b.in[b.off-4:]
-	v = v[:4]
+	v := b.in[len(b.in)-4:]
+	b.in = b.in[:len(b.in)-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
 }
 
 // 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:])
+	v := b.in[len(b.in)-8:]
+	b.in = b.in[:len(b.in)-8]
+	b.value = binary.LittleEndian.Uint64(v)
 	b.bitsRead = 0
-	b.off -= 8
 }
 
 // fill() will make sure at least 32 bits are available.
@@ -91,25 +87,25 @@ func (b *bitReader) fill() {
 	if b.bitsRead < 32 {
 		return
 	}
-	if b.off >= 4 {
-		v := b.in[b.off-4:]
-		v = v[:4]
+	if len(b.in) >= 4 {
+		v := b.in[len(b.in)-4:]
+		b.in = b.in[:len(b.in)-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--
+
+	b.bitsRead -= uint8(8 * len(b.in))
+	for len(b.in) > 0 {
+		b.value = (b.value << 8) | uint64(b.in[len(b.in)-1])
+		b.in = b.in[:len(b.in)-1]
 	}
 }
 
 // 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
+	return len(b.in) == 0 && b.bitsRead >= 64
 }
 
 // overread returns true if more bits have been requested than is on the stream.
@@ -119,7 +115,7 @@ func (b *bitReader) overread() bool {
 
 // remain returns the number of bits remaining.
 func (b *bitReader) remain() uint {
-	return b.off*8 + 64 - uint(b.bitsRead)
+	return 8*uint(len(b.in)) + 64 - uint(b.bitsRead)
 }
 
 // close the bitstream and returns an error if out-of-buffer reads occurred.
diff --git a/zstd/seqdec.go b/zstd/seqdec.go
index 9405fcf101..d7fe6d82d9 100644
--- a/zstd/seqdec.go
+++ b/zstd/seqdec.go
@@ -245,7 +245,7 @@ func (s *sequenceDecs) decodeSync(hist []byte) error {
 			return io.ErrUnexpectedEOF
 		}
 		var ll, mo, ml int
-		if br.off > 4+((maxOffsetBits+16+16)>>3) {
+		if len(br.in) > 4+((maxOffsetBits+16+16)>>3) {
 			// inlined function:
 			// ll, mo, ml = s.nextFast(br, llState, mlState, ofState)
 
@@ -452,18 +452,13 @@ func (s *sequenceDecs) next(br *bitReader, llState, mlState, ofState decSymbol)
 
 	// extra bits are stored in reverse order.
 	br.fill()
-	if s.maxBits <= 32 {
-		mo += br.getBits(moB)
-		ml += br.getBits(mlB)
-		ll += br.getBits(llB)
-	} else {
-		mo += br.getBits(moB)
+	mo += br.getBits(moB)
+	if s.maxBits > 32 {
 		br.fill()
-		// matchlength+literal length, max 32 bits
-		ml += br.getBits(mlB)
-		ll += br.getBits(llB)
-
 	}
+	// matchlength+literal length, max 32 bits
+	ml += br.getBits(mlB)
+	ll += br.getBits(llB)
 	mo = s.adjustOffset(mo, ll, moB)
 	return
 }
diff --git a/zstd/seqdec_amd64.s b/zstd/seqdec_amd64.s
index b6f4ba6fc5..974b99725f 100644
--- a/zstd/seqdec_amd64.s
+++ b/zstd/seqdec_amd64.s
@@ -5,11 +5,11 @@
 // func sequenceDecs_decode_amd64(s *sequenceDecs, br *bitReader, ctx *decodeAsmContext) int
 // Requires: CMOV
 TEXT ·sequenceDecs_decode_amd64(SB), $8-32
-	MOVQ    br+8(FP), AX
-	MOVQ    32(AX), DX
-	MOVBQZX 40(AX), BX
-	MOVQ    24(AX), SI
-	MOVQ    (AX), AX
+	MOVQ    br+8(FP), CX
+	MOVQ    24(CX), DX
+	MOVBQZX 32(CX), BX
+	MOVQ    (CX), AX
+	MOVQ    8(CX), SI
 	ADDQ    SI, AX
 	MOVQ    AX, (SP)
 	MOVQ    ctx+16(FP), AX
@@ -301,9 +301,9 @@ sequenceDecs_decode_amd64_match_len_ofs_ok:
 	MOVQ R12, 152(AX)
 	MOVQ R13, 160(AX)
 	MOVQ br+8(FP), AX
-	MOVQ DX, 32(AX)
-	MOVB BL, 40(AX)
-	MOVQ SI, 24(AX)
+	MOVQ DX, 24(AX)
+	MOVB BL, 32(AX)
+	MOVQ SI, 8(AX)
 
 	// Return success
 	MOVQ $0x00000000, ret+24(FP)
@@ -336,11 +336,11 @@ error_overread:
 // func sequenceDecs_decode_56_amd64(s *sequenceDecs, br *bitReader, ctx *decodeAsmContext) int
 // Requires: CMOV
 TEXT ·sequenceDecs_decode_56_amd64(SB), $8-32
-	MOVQ    br+8(FP), AX
-	MOVQ    32(AX), DX
-	MOVBQZX 40(AX), BX
-	MOVQ    24(AX), SI
-	MOVQ    (AX), AX
+	MOVQ    br+8(FP), CX
+	MOVQ    24(CX), DX
+	MOVBQZX 32(CX), BX
+	MOVQ    (CX), AX
+	MOVQ    8(CX), SI
 	ADDQ    SI, AX
 	MOVQ    AX, (SP)
 	MOVQ    ctx+16(FP), AX
@@ -603,9 +603,9 @@ sequenceDecs_decode_56_amd64_match_len_ofs_ok:
 	MOVQ R12, 152(AX)
 	MOVQ R13, 160(AX)
 	MOVQ br+8(FP), AX
-	MOVQ DX, 32(AX)
-	MOVB BL, 40(AX)
-	MOVQ SI, 24(AX)
+	MOVQ DX, 24(AX)
+	MOVB BL, 32(AX)
+	MOVQ SI, 8(AX)
 
 	// Return success
 	MOVQ $0x00000000, ret+24(FP)
@@ -638,11 +638,11 @@ error_overread:
 // func sequenceDecs_decode_bmi2(s *sequenceDecs, br *bitReader, ctx *decodeAsmContext) int
 // Requires: BMI, BMI2, CMOV
 TEXT ·sequenceDecs_decode_bmi2(SB), $8-32
-	MOVQ    br+8(FP), CX
-	MOVQ    32(CX), AX
-	MOVBQZX 40(CX), DX
-	MOVQ    24(CX), BX
-	MOVQ    (CX), CX
+	MOVQ    br+8(FP), BX
+	MOVQ    24(BX), AX
+	MOVBQZX 32(BX), DX
+	MOVQ    (BX), CX
+	MOVQ    8(BX), BX
 	ADDQ    BX, CX
 	MOVQ    CX, (SP)
 	MOVQ    ctx+16(FP), CX
@@ -892,9 +892,9 @@ sequenceDecs_decode_bmi2_match_len_ofs_ok:
 	MOVQ R11, 152(CX)
 	MOVQ R12, 160(CX)
 	MOVQ br+8(FP), CX
-	MOVQ AX, 32(CX)
-	MOVB DL, 40(CX)
-	MOVQ BX, 24(CX)
+	MOVQ AX, 24(CX)
+	MOVB DL, 32(CX)
+	MOVQ BX, 8(CX)
 
 	// Return success
 	MOVQ $0x00000000, ret+24(FP)
@@ -927,11 +927,11 @@ error_overread:
 // func sequenceDecs_decode_56_bmi2(s *sequenceDecs, br *bitReader, ctx *decodeAsmContext) int
 // Requires: BMI, BMI2, CMOV
 TEXT ·sequenceDecs_decode_56_bmi2(SB), $8-32
-	MOVQ    br+8(FP), CX
-	MOVQ    32(CX), AX
-	MOVBQZX 40(CX), DX
-	MOVQ    24(CX), BX
-	MOVQ    (CX), CX
+	MOVQ    br+8(FP), BX
+	MOVQ    24(BX), AX
+	MOVBQZX 32(BX), DX
+	MOVQ    (BX), CX
+	MOVQ    8(BX), BX
 	ADDQ    BX, CX
 	MOVQ    CX, (SP)
 	MOVQ    ctx+16(FP), CX
@@ -1152,9 +1152,9 @@ sequenceDecs_decode_56_bmi2_match_len_ofs_ok:
 	MOVQ R11, 152(CX)
 	MOVQ R12, 160(CX)
 	MOVQ br+8(FP), CX
-	MOVQ AX, 32(CX)
-	MOVB DL, 40(CX)
-	MOVQ BX, 24(CX)
+	MOVQ AX, 24(CX)
+	MOVB DL, 32(CX)
+	MOVQ BX, 8(CX)
 
 	// Return success
 	MOVQ $0x00000000, ret+24(FP)
@@ -1797,11 +1797,11 @@ empty_seqs:
 // func sequenceDecs_decodeSync_amd64(s *sequenceDecs, br *bitReader, ctx *decodeSyncAsmContext) int
 // Requires: CMOV, SSE
 TEXT ·sequenceDecs_decodeSync_amd64(SB), $64-32
-	MOVQ    br+8(FP), AX
-	MOVQ    32(AX), DX
-	MOVBQZX 40(AX), BX
-	MOVQ    24(AX), SI
-	MOVQ    (AX), AX
+	MOVQ    br+8(FP), CX
+	MOVQ    24(CX), DX
+	MOVBQZX 32(CX), BX
+	MOVQ    (CX), AX
+	MOVQ    8(CX), SI
 	ADDQ    SI, AX
 	MOVQ    AX, (SP)
 	MOVQ    ctx+16(FP), AX
@@ -2295,9 +2295,9 @@ handle_loop:
 
 loop_finished:
 	MOVQ br+8(FP), AX
-	MOVQ DX, 32(AX)
-	MOVB BL, 40(AX)
-	MOVQ SI, 24(AX)
+	MOVQ DX, 24(AX)
+	MOVB BL, 32(AX)
+	MOVQ SI, 8(AX)
 
 	// Update the context
 	MOVQ ctx+16(FP), AX
@@ -2362,11 +2362,11 @@ error_not_enough_space:
 // func sequenceDecs_decodeSync_bmi2(s *sequenceDecs, br *bitReader, ctx *decodeSyncAsmContext) int
 // Requires: BMI, BMI2, CMOV, SSE
 TEXT ·sequenceDecs_decodeSync_bmi2(SB), $64-32
-	MOVQ    br+8(FP), CX
-	MOVQ    32(CX), AX
-	MOVBQZX 40(CX), DX
-	MOVQ    24(CX), BX
-	MOVQ    (CX), CX
+	MOVQ    br+8(FP), BX
+	MOVQ    24(BX), AX
+	MOVBQZX 32(BX), DX
+	MOVQ    (BX), CX
+	MOVQ    8(BX), BX
 	ADDQ    BX, CX
 	MOVQ    CX, (SP)
 	MOVQ    ctx+16(FP), CX
@@ -2818,9 +2818,9 @@ handle_loop:
 
 loop_finished:
 	MOVQ br+8(FP), CX
-	MOVQ AX, 32(CX)
-	MOVB DL, 40(CX)
-	MOVQ BX, 24(CX)
+	MOVQ AX, 24(CX)
+	MOVB DL, 32(CX)
+	MOVQ BX, 8(CX)
 
 	// Update the context
 	MOVQ ctx+16(FP), AX
@@ -2885,11 +2885,11 @@ error_not_enough_space:
 // func sequenceDecs_decodeSync_safe_amd64(s *sequenceDecs, br *bitReader, ctx *decodeSyncAsmContext) int
 // Requires: CMOV, SSE
 TEXT ·sequenceDecs_decodeSync_safe_amd64(SB), $64-32
-	MOVQ    br+8(FP), AX
-	MOVQ    32(AX), DX
-	MOVBQZX 40(AX), BX
-	MOVQ    24(AX), SI
-	MOVQ    (AX), AX
+	MOVQ    br+8(FP), CX
+	MOVQ    24(CX), DX
+	MOVBQZX 32(CX), BX
+	MOVQ    (CX), AX
+	MOVQ    8(CX), SI
 	ADDQ    SI, AX
 	MOVQ    AX, (SP)
 	MOVQ    ctx+16(FP), AX
@@ -3485,9 +3485,9 @@ handle_loop:
 
 loop_finished:
 	MOVQ br+8(FP), AX
-	MOVQ DX, 32(AX)
-	MOVB BL, 40(AX)
-	MOVQ SI, 24(AX)
+	MOVQ DX, 24(AX)
+	MOVB BL, 32(AX)
+	MOVQ SI, 8(AX)
 
 	// Update the context
 	MOVQ ctx+16(FP), AX
@@ -3552,11 +3552,11 @@ error_not_enough_space:
 // func sequenceDecs_decodeSync_safe_bmi2(s *sequenceDecs, br *bitReader, ctx *decodeSyncAsmContext) int
 // Requires: BMI, BMI2, CMOV, SSE
 TEXT ·sequenceDecs_decodeSync_safe_bmi2(SB), $64-32
-	MOVQ    br+8(FP), CX
-	MOVQ    32(CX), AX
-	MOVBQZX 40(CX), DX
-	MOVQ    24(CX), BX
-	MOVQ    (CX), CX
+	MOVQ    br+8(FP), BX
+	MOVQ    24(BX), AX
+	MOVBQZX 32(BX), DX
+	MOVQ    (BX), CX
+	MOVQ    8(BX), BX
 	ADDQ    BX, CX
 	MOVQ    CX, (SP)
 	MOVQ    ctx+16(FP), CX
@@ -4110,9 +4110,9 @@ handle_loop:
 
 loop_finished:
 	MOVQ br+8(FP), CX
-	MOVQ AX, 32(CX)
-	MOVB DL, 40(CX)
-	MOVQ BX, 24(CX)
+	MOVQ AX, 24(CX)
+	MOVB DL, 32(CX)
+	MOVQ BX, 8(CX)
 
 	// Update the context
 	MOVQ ctx+16(FP), AX
diff --git a/zstd/seqdec_generic.go b/zstd/seqdec_generic.go
index ac2a80d291..2fb35b788c 100644
--- a/zstd/seqdec_generic.go
+++ b/zstd/seqdec_generic.go
@@ -29,7 +29,7 @@ func (s *sequenceDecs) decode(seqs []seqVals) error {
 	}
 	for i := range seqs {
 		var ll, mo, ml int
-		if br.off > 4+((maxOffsetBits+16+16)>>3) {
+		if len(br.in) > 4+((maxOffsetBits+16+16)>>3) {
 			// inlined function:
 			// ll, mo, ml = s.nextFast(br, llState, mlState, ofState)