diff --git a/README.md b/README.md
index 2f2a3928..b8f7ed89 100644
--- a/README.md
+++ b/README.md
@@ -65,10 +65,10 @@ After running `go generate` the [`add.s`](examples/add/add.s) file will contain
 
 // func Add(x uint64, y uint64) uint64
 TEXT ·Add(SB), NOSPLIT, $0-24
-	MOVQ	x(FP), AX
-	MOVQ	y+8(FP), CX
-	ADDQ	AX, CX
-	MOVQ	CX, ret+16(FP)
+	MOVQ x(FP), AX
+	MOVQ y+8(FP), CX
+	ADDQ AX, CX
+	MOVQ CX, ret+16(FP)
 	RET
 ```
 
@@ -137,18 +137,20 @@ The result from this code generator is:
 
 // func Sum(xs []uint64) uint64
 TEXT ·Sum(SB), NOSPLIT, $0-32
-	MOVQ	xs_base(FP), AX
-	MOVQ	xs_len+8(FP), CX
-	XORQ	DX, DX
+	MOVQ xs_base(FP), AX
+	MOVQ xs_len+8(FP), CX
+	XORQ DX, DX
+
 loop:
-	CMPQ	CX, $0x00
-	JE	done
-	ADDQ	(AX), DX
-	ADDQ	$0x08, AX
-	DECQ	CX
-	JMP	loop
+	CMPQ CX, $0x00
+	JE   done
+	ADDQ (AX), DX
+	ADDQ $0x08, AX
+	DECQ CX
+	JMP  loop
+
 done:
-	MOVQ	DX, ret+24(FP)
+	MOVQ DX, ret+24(FP)
 	RET
 ```
 
diff --git a/examples/add/README.md b/examples/add/README.md
index 108555a5..38c0d0cc 100644
--- a/examples/add/README.md
+++ b/examples/add/README.md
@@ -43,9 +43,9 @@ This produces [`add.s`](add.s) as follows:
 
 // func Add(x uint64, y uint64) uint64
 TEXT ·Add(SB), NOSPLIT, $0-24
-	MOVQ	x(FP), AX
-	MOVQ	y+8(FP), CX
-	ADDQ	AX, CX
-	MOVQ	CX, ret+16(FP)
+	MOVQ x(FP), AX
+	MOVQ y+8(FP), CX
+	ADDQ AX, CX
+	MOVQ CX, ret+16(FP)
 	RET
 ```
diff --git a/examples/add/add.s b/examples/add/add.s
index 679dfbf0..a5659b0e 100644
--- a/examples/add/add.s
+++ b/examples/add/add.s
@@ -4,8 +4,8 @@
 
 // func Add(x uint64, y uint64) uint64
 TEXT ·Add(SB), NOSPLIT, $0-24
-	MOVQ	x(FP), AX
-	MOVQ	y+8(FP), CX
-	ADDQ	AX, CX
-	MOVQ	CX, ret+16(FP)
+	MOVQ x(FP), AX
+	MOVQ y+8(FP), CX
+	ADDQ AX, CX
+	MOVQ CX, ret+16(FP)
 	RET
diff --git a/examples/args/README.md b/examples/args/README.md
index aa0e646f..8032cc7f 100644
--- a/examples/args/README.md
+++ b/examples/args/README.md
@@ -20,8 +20,8 @@ This `avo` code will generate the following assembly. Note that parameter refere
 ```s
 // func Second(x int32, y int32) int32
 TEXT ·Second(SB), NOSPLIT, $0-12
-	MOVL	y+4(FP), AX
-	MOVL	AX, ret+8(FP)
+	MOVL y+4(FP), AX
+	MOVL AX, ret+8(FP)
 	RET
 ```
 
diff --git a/examples/args/args.s b/examples/args/args.s
index cd890aec..1a99d0d1 100644
--- a/examples/args/args.s
+++ b/examples/args/args.s
@@ -4,108 +4,108 @@
 
 // func Second(x int32, y int32) int32
 TEXT ·Second(SB), NOSPLIT, $0-12
-	MOVL	y+4(FP), AX
-	MOVL	AX, ret+8(FP)
+	MOVL y+4(FP), AX
+	MOVL AX, ret+8(FP)
 	RET
 
 // func StringLen(s string) int
 TEXT ·StringLen(SB), NOSPLIT, $0-24
-	MOVQ	s_len+8(FP), AX
-	MOVQ	AX, ret+16(FP)
+	MOVQ s_len+8(FP), AX
+	MOVQ AX, ret+16(FP)
 	RET
 
 // func SliceLen(s []int) int
 TEXT ·SliceLen(SB), NOSPLIT, $0-32
-	MOVQ	s_len+8(FP), AX
-	MOVQ	AX, ret+24(FP)
+	MOVQ s_len+8(FP), AX
+	MOVQ AX, ret+24(FP)
 	RET
 
 // func SliceCap(s []int) int
 TEXT ·SliceCap(SB), NOSPLIT, $0-32
-	MOVQ	s_cap+16(FP), AX
-	MOVQ	AX, ret+24(FP)
+	MOVQ s_cap+16(FP), AX
+	MOVQ AX, ret+24(FP)
 	RET
 
 // func ArrayThree(a [7]uint64) uint64
 TEXT ·ArrayThree(SB), NOSPLIT, $0-64
-	MOVQ	a_3+24(FP), AX
-	MOVQ	AX, ret+56(FP)
+	MOVQ a_3+24(FP), AX
+	MOVQ AX, ret+56(FP)
 	RET
 
 // func FieldByte(s Struct) byte
 TEXT ·FieldByte(SB), NOSPLIT, $0-177
-	MOVB	s_Byte(FP), AL
-	MOVB	AL, ret+176(FP)
+	MOVB s_Byte(FP), AL
+	MOVB AL, ret+176(FP)
 	RET
 
 // func FieldInt8(s Struct) int8
 TEXT ·FieldInt8(SB), NOSPLIT, $0-177
-	MOVB	s_Int8+1(FP), AL
-	MOVB	AL, ret+176(FP)
+	MOVB s_Int8+1(FP), AL
+	MOVB AL, ret+176(FP)
 	RET
 
 // func FieldUint16(s Struct) uint16
 TEXT ·FieldUint16(SB), NOSPLIT, $0-178
-	MOVW	s_Uint16+2(FP), AX
-	MOVW	AX, ret+176(FP)
+	MOVW s_Uint16+2(FP), AX
+	MOVW AX, ret+176(FP)
 	RET
 
 // func FieldInt32(s Struct) int32
 TEXT ·FieldInt32(SB), NOSPLIT, $0-180
-	MOVL	s_Int32+4(FP), AX
-	MOVL	AX, ret+176(FP)
+	MOVL s_Int32+4(FP), AX
+	MOVL AX, ret+176(FP)
 	RET
 
 // func FieldUint64(s Struct) uint64
 TEXT ·FieldUint64(SB), NOSPLIT, $0-184
-	MOVQ	s_Uint64+8(FP), AX
-	MOVQ	AX, ret+176(FP)
+	MOVQ s_Uint64+8(FP), AX
+	MOVQ AX, ret+176(FP)
 	RET
 
 // func FieldFloat32(s Struct) float32
 TEXT ·FieldFloat32(SB), NOSPLIT, $0-180
-	MOVSS	s_Float32+16(FP), X0
-	MOVSS	X0, ret+176(FP)
+	MOVSS s_Float32+16(FP), X0
+	MOVSS X0, ret+176(FP)
 	RET
 
 // func FieldFloat64(s Struct) float64
 TEXT ·FieldFloat64(SB), NOSPLIT, $0-184
-	MOVSD	s_Float64+24(FP), X0
-	MOVSD	X0, ret+176(FP)
+	MOVSD s_Float64+24(FP), X0
+	MOVSD X0, ret+176(FP)
 	RET
 
 // func FieldStringLen(s Struct) int
 TEXT ·FieldStringLen(SB), NOSPLIT, $0-184
-	MOVQ	s_String_len+40(FP), AX
-	MOVQ	AX, ret+176(FP)
+	MOVQ s_String_len+40(FP), AX
+	MOVQ AX, ret+176(FP)
 	RET
 
 // func FieldSliceCap(s Struct) int
 TEXT ·FieldSliceCap(SB), NOSPLIT, $0-184
-	MOVQ	s_Slice_cap+64(FP), AX
-	MOVQ	AX, ret+176(FP)
+	MOVQ s_Slice_cap+64(FP), AX
+	MOVQ AX, ret+176(FP)
 	RET
 
 // func FieldArrayTwoBTwo(s Struct) byte
 TEXT ·FieldArrayTwoBTwo(SB), NOSPLIT, $0-177
-	MOVB	s_Array_2_B_2+114(FP), AL
-	MOVB	AL, ret+176(FP)
+	MOVB s_Array_2_B_2+114(FP), AL
+	MOVB AL, ret+176(FP)
 	RET
 
 // func FieldArrayOneC(s Struct) uint16
 TEXT ·FieldArrayOneC(SB), NOSPLIT, $0-178
-	MOVW	s_Array_1_C+100(FP), AX
-	MOVW	AX, ret+176(FP)
+	MOVW s_Array_1_C+100(FP), AX
+	MOVW AX, ret+176(FP)
 	RET
 
 // func FieldComplex64Imag(s Struct) float32
 TEXT ·FieldComplex64Imag(SB), NOSPLIT, $0-180
-	MOVSS	s_Complex64_imag+156(FP), X0
-	MOVSS	X0, ret+176(FP)
+	MOVSS s_Complex64_imag+156(FP), X0
+	MOVSS X0, ret+176(FP)
 	RET
 
 // func FieldComplex128Real(s Struct) float64
 TEXT ·FieldComplex128Real(SB), NOSPLIT, $0-184
-	MOVSD	s_Complex128_real+160(FP), X0
-	MOVSD	X0, ret+176(FP)
+	MOVSD s_Complex128_real+160(FP), X0
+	MOVSD X0, ret+176(FP)
 	RET
diff --git a/examples/backup.tar.gz b/examples/backup.tar.gz
new file mode 100644
index 00000000..48875671
Binary files /dev/null and b/examples/backup.tar.gz differ
diff --git a/examples/complex/README.md b/examples/complex/README.md
index b74b440e..8c4ce342 100644
--- a/examples/complex/README.md
+++ b/examples/complex/README.md
@@ -25,12 +25,12 @@ Generated assembly:
 ```s
 // func Norm(z complex128) float64
 TEXT ·Norm(SB), NOSPLIT, $0-24
-	MOVSD	z_real(FP), X0
-	MOVSD	z_imag+8(FP), X1
-	MULSD	X0, X0
-	MULSD	X1, X1
-	ADDSD	X1, X0
-	SQRTSD	X0, X2
-	MOVSD	X2, ret+16(FP)
+	MOVSD  z_real(FP), X0
+	MOVSD  z_imag+8(FP), X1
+	MULSD  X0, X0
+	MULSD  X1, X1
+	ADDSD  X1, X0
+	SQRTSD X0, X2
+	MOVSD  X2, ret+16(FP)
 	RET
 ```
diff --git a/examples/complex/complex.s b/examples/complex/complex.s
index aa9e4d68..ff4475cf 100644
--- a/examples/complex/complex.s
+++ b/examples/complex/complex.s
@@ -4,23 +4,23 @@
 
 // func Real(z complex128) float64
 TEXT ·Real(SB), NOSPLIT, $0-24
-	MOVSD	z_real(FP), X0
-	MOVSD	X0, ret+16(FP)
+	MOVSD z_real(FP), X0
+	MOVSD X0, ret+16(FP)
 	RET
 
 // func Imag(z complex128) float64
 TEXT ·Imag(SB), NOSPLIT, $0-24
-	MOVSD	z_imag+8(FP), X0
-	MOVSD	X0, ret+16(FP)
+	MOVSD z_imag+8(FP), X0
+	MOVSD X0, ret+16(FP)
 	RET
 
 // func Norm(z complex128) float64
 TEXT ·Norm(SB), NOSPLIT, $0-24
-	MOVSD	z_real(FP), X0
-	MOVSD	z_imag+8(FP), X1
-	MULSD	X0, X0
-	MULSD	X1, X1
-	ADDSD	X1, X0
-	SQRTSD	X0, X2
-	MOVSD	X2, ret+16(FP)
+	MOVSD  z_real(FP), X0
+	MOVSD  z_imag+8(FP), X1
+	MULSD  X0, X0
+	MULSD  X1, X1
+	ADDSD  X1, X0
+	SQRTSD X0, X2
+	MOVSD  X2, ret+16(FP)
 	RET
diff --git a/examples/data/data.s b/examples/data/data.s
index c9ce1432..6819c036 100644
--- a/examples/data/data.s
+++ b/examples/data/data.s
@@ -14,8 +14,8 @@ GLOBL bytes<>(SB), RODATA|NOPTR, $40
 
 // func DataAt(i int) byte
 TEXT ·DataAt(SB), NOSPLIT, $0-9
-	MOVQ	i(FP), AX
-	LEAQ	bytes<>(SB), CX
-	MOVB	(CX)(AX*1), AL
-	MOVB	AL, ret+8(FP)
+	MOVQ i(FP), AX
+	LEAQ bytes<>(SB), CX
+	MOVB (CX)(AX*1), AL
+	MOVB AL, ret+8(FP)
 	RET
diff --git a/examples/dot/dot.s b/examples/dot/dot.s
index 609832e9..6dc20955 100644
--- a/examples/dot/dot.s
+++ b/examples/dot/dot.s
@@ -4,55 +4,59 @@
 
 // func Dot(x []float32, y []float32) float32
 TEXT ·Dot(SB), NOSPLIT, $0-52
-	MOVQ	x_base(FP), AX
-	MOVQ	y_base+24(FP), CX
-	MOVQ	x_len+8(FP), DX
-	VXORPS	Y0, Y0, Y0
-	VXORPS	Y1, Y1, Y1
-	VXORPS	Y2, Y2, Y2
-	VXORPS	Y3, Y3, Y3
-	VXORPS	Y4, Y4, Y4
-	VXORPS	Y5, Y5, Y5
+	MOVQ   x_base(FP), AX
+	MOVQ   y_base+24(FP), CX
+	MOVQ   x_len+8(FP), DX
+	VXORPS Y0, Y0, Y0
+	VXORPS Y1, Y1, Y1
+	VXORPS Y2, Y2, Y2
+	VXORPS Y3, Y3, Y3
+	VXORPS Y4, Y4, Y4
+	VXORPS Y5, Y5, Y5
+
 blockloop:
-	CMPQ	DX, $0x00000030
-	JL	tail
-	VMOVUPS	(AX), Y6
-	VMOVUPS	32(AX), Y7
-	VMOVUPS	64(AX), Y8
-	VMOVUPS	96(AX), Y9
-	VMOVUPS	128(AX), Y10
-	VMOVUPS	160(AX), Y11
-	VFMADD231PS	(CX), Y6, Y0
-	VFMADD231PS	32(CX), Y7, Y1
-	VFMADD231PS	64(CX), Y8, Y2
-	VFMADD231PS	96(CX), Y9, Y3
-	VFMADD231PS	128(CX), Y10, Y4
-	VFMADD231PS	160(CX), Y11, Y5
-	ADDQ	$0x000000c0, AX
-	ADDQ	$0x000000c0, CX
-	SUBQ	$0x00000030, DX
-	JMP	blockloop
+	CMPQ        DX, $0x00000030
+	JL          tail
+	VMOVUPS     (AX), Y6
+	VMOVUPS     32(AX), Y7
+	VMOVUPS     64(AX), Y8
+	VMOVUPS     96(AX), Y9
+	VMOVUPS     128(AX), Y10
+	VMOVUPS     160(AX), Y11
+	VFMADD231PS (CX), Y6, Y0
+	VFMADD231PS 32(CX), Y7, Y1
+	VFMADD231PS 64(CX), Y8, Y2
+	VFMADD231PS 96(CX), Y9, Y3
+	VFMADD231PS 128(CX), Y10, Y4
+	VFMADD231PS 160(CX), Y11, Y5
+	ADDQ        $0x000000c0, AX
+	ADDQ        $0x000000c0, CX
+	SUBQ        $0x00000030, DX
+	JMP         blockloop
+
 tail:
-	VXORPS	X12, X12, X12
+	VXORPS X12, X12, X12
+
 tailloop:
-	CMPQ	DX, $0x00000000
-	JE	reduce
-	VMOVSS	(AX), X6
-	VFMADD231SS	(CX), X6, X12
-	ADDQ	$0x00000004, AX
-	ADDQ	$0x00000004, CX
-	DECQ	DX
-	JMP	tailloop
+	CMPQ        DX, $0x00000000
+	JE          reduce
+	VMOVSS      (AX), X6
+	VFMADD231SS (CX), X6, X12
+	ADDQ        $0x00000004, AX
+	ADDQ        $0x00000004, CX
+	DECQ        DX
+	JMP         tailloop
+
 reduce:
-	VADDPS	Y0, Y1, Y0
-	VADDPS	Y0, Y2, Y0
-	VADDPS	Y0, Y3, Y0
-	VADDPS	Y0, Y4, Y0
-	VADDPS	Y0, Y5, Y0
-	VEXTRACTF128	$0x01, Y0, X1
-	VADDPS	X0, X1, X0
-	VADDPS	X0, X12, X0
-	VHADDPS	X0, X0, X0
-	VHADDPS	X0, X0, X0
-	MOVSS	X0, ret+48(FP)
+	VADDPS       Y0, Y1, Y0
+	VADDPS       Y0, Y2, Y0
+	VADDPS       Y0, Y3, Y0
+	VADDPS       Y0, Y4, Y0
+	VADDPS       Y0, Y5, Y0
+	VEXTRACTF128 $0x01, Y0, X1
+	VADDPS       X0, X1, X0
+	VADDPS       X0, X12, X0
+	VHADDPS      X0, X0, X0
+	VHADDPS      X0, X0, X0
+	MOVSS        X0, ret+48(FP)
 	RET
diff --git a/examples/fnv1a/fnv1a.s b/examples/fnv1a/fnv1a.s
index 9df9e01b..85bbd580 100644
--- a/examples/fnv1a/fnv1a.s
+++ b/examples/fnv1a/fnv1a.s
@@ -4,19 +4,21 @@
 
 // func Hash64(data []byte) uint64
 TEXT ·Hash64(SB), NOSPLIT, $0-32
-	MOVQ	data_base(FP), CX
-	MOVQ	data_len+8(FP), BX
-	MOVQ	$0xcbf29ce484222325, AX
-	MOVQ	$0x00000100000001b3, BP
+	MOVQ data_base(FP), CX
+	MOVQ data_len+8(FP), BX
+	MOVQ $0xcbf29ce484222325, AX
+	MOVQ $0x00000100000001b3, BP
+
 loop:
-	CMPQ	BX, $0x00
-	JE	done
-	MOVBQZX	(CX), DX
-	XORQ	DX, AX
-	MULQ	BP
-	INCQ	CX
-	DECQ	BX
-	JMP	loop
+	CMPQ    BX, $0x00
+	JE      done
+	MOVBQZX (CX), DX
+	XORQ    DX, AX
+	MULQ    BP
+	INCQ    CX
+	DECQ    BX
+	JMP     loop
+
 done:
-	MOVQ	AX, ret+24(FP)
+	MOVQ AX, ret+24(FP)
 	RET
diff --git a/examples/geohash/geohash.s b/examples/geohash/geohash.s
index c724e197..c14d6f51 100644
--- a/examples/geohash/geohash.s
+++ b/examples/geohash/geohash.s
@@ -4,21 +4,21 @@
 
 // func EncodeInt(lat float64, lng float64) uint64
 TEXT ·EncodeInt(SB), NOSPLIT, $0-24
-	MOVSD	lat(FP), X0
-	MOVSD	lng+8(FP), X1
-	MULSD	reciprocal180<>(SB), X0
-	ADDSD	onepointfive<>(SB), X0
-	MULSD	reciprocal360<>(SB), X1
-	ADDSD	onepointfive<>(SB), X1
-	MOVQ	X0, CX
-	SHRQ	$0x14, CX
-	MOVQ	X1, AX
-	SHRQ	$0x14, AX
-	PDEPQ	mask<>(SB), CX, CX
-	PDEPQ	mask<>(SB), AX, AX
-	SHLQ	$0x01, AX
-	XORQ	AX, CX
-	MOVQ	CX, ret+16(FP)
+	MOVSD lat(FP), X0
+	MOVSD lng+8(FP), X1
+	MULSD reciprocal180<>(SB), X0
+	ADDSD onepointfive<>(SB), X0
+	MULSD reciprocal360<>(SB), X1
+	ADDSD onepointfive<>(SB), X1
+	MOVQ  X0, CX
+	SHRQ  $0x14, CX
+	MOVQ  X1, AX
+	SHRQ  $0x14, AX
+	PDEPQ mask<>(SB), CX, CX
+	PDEPQ mask<>(SB), AX, AX
+	SHLQ  $0x01, AX
+	XORQ  AX, CX
+	MOVQ  CX, ret+16(FP)
 	RET
 
 DATA reciprocal180<>(SB)/8, $(0.005555555555555556)
diff --git a/examples/returns/returns.s b/examples/returns/returns.s
index 9c3b6f6d..23751f28 100644
--- a/examples/returns/returns.s
+++ b/examples/returns/returns.s
@@ -4,43 +4,43 @@
 
 // func Interval(start uint64, size uint64) (uint64, uint64)
 TEXT ·Interval(SB), NOSPLIT, $0-32
-	MOVQ	start(FP), AX
-	MOVQ	size+8(FP), CX
-	ADDQ	AX, CX
-	MOVQ	AX, ret+16(FP)
-	MOVQ	CX, ret1+24(FP)
+	MOVQ start(FP), AX
+	MOVQ size+8(FP), CX
+	ADDQ AX, CX
+	MOVQ AX, ret+16(FP)
+	MOVQ CX, ret1+24(FP)
 	RET
 
 // func Butterfly(x0 float64, x1 float64) (y0 float64, y1 float64)
 TEXT ·Butterfly(SB), NOSPLIT, $0-32
-	MOVSD	x0(FP), X0
-	MOVSD	x1+8(FP), X1
-	MOVSD	X0, X2
-	ADDSD	X1, X2
-	MOVSD	X0, X3
-	SUBSD	X1, X3
-	MOVSD	X2, y0+16(FP)
-	MOVSD	X3, y1+24(FP)
+	MOVSD x0(FP), X0
+	MOVSD x1+8(FP), X1
+	MOVSD X0, X2
+	ADDSD X1, X2
+	MOVSD X0, X3
+	SUBSD X1, X3
+	MOVSD X2, y0+16(FP)
+	MOVSD X3, y1+24(FP)
 	RET
 
 // func Septuple(byte) [7]byte
 TEXT ·Septuple(SB), NOSPLIT, $0-15
-	MOVB	arg(FP), AL
-	MOVB	AL, ret_0+8(FP)
-	MOVB	AL, ret_1+9(FP)
-	MOVB	AL, ret_2+10(FP)
-	MOVB	AL, ret_3+11(FP)
-	MOVB	AL, ret_4+12(FP)
-	MOVB	AL, ret_5+13(FP)
-	MOVB	AL, ret_6+14(FP)
+	MOVB arg(FP), AL
+	MOVB AL, ret_0+8(FP)
+	MOVB AL, ret_1+9(FP)
+	MOVB AL, ret_2+10(FP)
+	MOVB AL, ret_3+11(FP)
+	MOVB AL, ret_4+12(FP)
+	MOVB AL, ret_5+13(FP)
+	MOVB AL, ret_6+14(FP)
 	RET
 
 // func CriticalLine(t float64) complex128
 TEXT ·CriticalLine(SB), NOSPLIT, $0-24
-	MOVSD	t(FP), X0
-	MOVSD	half<>(SB), X1
-	MOVSD	X1, ret_real+8(FP)
-	MOVSD	X0, ret_imag+16(FP)
+	MOVSD t(FP), X0
+	MOVSD half<>(SB), X1
+	MOVSD X1, ret_real+8(FP)
+	MOVSD X0, ret_imag+16(FP)
 	RET
 
 DATA half<>(SB)/8, $(0.5)
@@ -48,12 +48,12 @@ GLOBL half<>(SB), RODATA|NOPTR, $8
 
 // func NewStruct(w uint16, p [2]float64, q uint64) Struct
 TEXT ·NewStruct(SB), NOSPLIT, $0-64
-	MOVW	w(FP), AX
-	MOVSD	p_0+8(FP), X0
-	MOVSD	p_1+16(FP), X1
-	MOVQ	q+24(FP), CX
-	MOVW	AX, ret_Word+32(FP)
-	MOVSD	X0, ret_Point_0+40(FP)
-	MOVSD	X1, ret_Point_1+48(FP)
-	MOVQ	CX, ret_Quad+56(FP)
+	MOVW  w(FP), AX
+	MOVSD p_0+8(FP), X0
+	MOVSD p_1+16(FP), X1
+	MOVQ  q+24(FP), CX
+	MOVW  AX, ret_Word+32(FP)
+	MOVSD X0, ret_Point_0+40(FP)
+	MOVSD X1, ret_Point_1+48(FP)
+	MOVQ  CX, ret_Quad+56(FP)
 	RET
diff --git a/examples/sha1/sha1.s b/examples/sha1/sha1.s
index 55d1abee..4f16ebac 100644
--- a/examples/sha1/sha1.s
+++ b/examples/sha1/sha1.s
@@ -2,1338 +2,1338 @@
 
 // func block(h *[5]uint32, m []byte)
 TEXT ·block(SB), $64-32
-	MOVQ	h(FP), AX
-	MOVQ	m_base+8(FP), CX
-	MOVL	(AX), DX
-	MOVL	4(AX), BX
-	MOVL	8(AX), BP
-	MOVL	12(AX), SI
-	MOVL	16(AX), DI
-	MOVL	DX, R8
-	MOVL	BX, R9
-	MOVL	BP, R10
-	MOVL	SI, R11
-	MOVL	DI, R12
-	MOVL	(CX), R13
-	BSWAPL	R13
-	MOVL	R13, (SP)
-	MOVL	R8, R14
-	ROLL	$0x05, R14
-	MOVL	R11, R15
-	XORL	R10, R15
-	ANDL	R9, R15
-	XORL	R11, R15
-	ADDL	R15, R14
-	ADDL	R12, R14
-	ADDL	$0x5a827999, R14
-	ADDL	R13, R14
-	ROLL	$0x1e, R9
-	MOVL	4(CX), R12
-	BSWAPL	R12
-	MOVL	R12, 4(SP)
-	MOVL	R14, R13
-	ROLL	$0x05, R13
-	MOVL	R10, R15
-	XORL	R9, R15
-	ANDL	R8, R15
-	XORL	R10, R15
-	ADDL	R15, R13
-	ADDL	R11, R13
-	ADDL	$0x5a827999, R13
-	ADDL	R12, R13
-	ROLL	$0x1e, R8
-	MOVL	8(CX), R11
-	BSWAPL	R11
-	MOVL	R11, 8(SP)
-	MOVL	R13, R12
-	ROLL	$0x05, R12
-	MOVL	R9, R15
-	XORL	R8, R15
-	ANDL	R14, R15
-	XORL	R9, R15
-	ADDL	R15, R12
-	ADDL	R10, R12
-	ADDL	$0x5a827999, R12
-	ADDL	R11, R12
-	ROLL	$0x1e, R14
-	MOVL	12(CX), R10
-	BSWAPL	R10
-	MOVL	R10, 12(SP)
-	MOVL	R12, R11
-	ROLL	$0x05, R11
-	MOVL	R8, R15
-	XORL	R14, R15
-	ANDL	R13, R15
-	XORL	R8, R15
-	ADDL	R15, R11
-	ADDL	R9, R11
-	ADDL	$0x5a827999, R11
-	ADDL	R10, R11
-	ROLL	$0x1e, R13
-	MOVL	16(CX), R9
-	BSWAPL	R9
-	MOVL	R9, 16(SP)
-	MOVL	R11, R10
-	ROLL	$0x05, R10
-	MOVL	R14, R15
-	XORL	R13, R15
-	ANDL	R12, R15
-	XORL	R14, R15
-	ADDL	R15, R10
-	ADDL	R8, R10
-	ADDL	$0x5a827999, R10
-	ADDL	R9, R10
-	ROLL	$0x1e, R12
-	MOVL	20(CX), R8
-	BSWAPL	R8
-	MOVL	R8, 20(SP)
-	MOVL	R10, R9
-	ROLL	$0x05, R9
-	MOVL	R13, R15
-	XORL	R12, R15
-	ANDL	R11, R15
-	XORL	R13, R15
-	ADDL	R15, R9
-	ADDL	R14, R9
-	ADDL	$0x5a827999, R9
-	ADDL	R8, R9
-	ROLL	$0x1e, R11
-	MOVL	24(CX), R8
-	BSWAPL	R8
-	MOVL	R8, 24(SP)
-	MOVL	R9, R14
-	ROLL	$0x05, R14
-	MOVL	R12, R15
-	XORL	R11, R15
-	ANDL	R10, R15
-	XORL	R12, R15
-	ADDL	R15, R14
-	ADDL	R13, R14
-	ADDL	$0x5a827999, R14
-	ADDL	R8, R14
-	ROLL	$0x1e, R10
-	MOVL	28(CX), R8
-	BSWAPL	R8
-	MOVL	R8, 28(SP)
-	MOVL	R14, R13
-	ROLL	$0x05, R13
-	MOVL	R11, R15
-	XORL	R10, R15
-	ANDL	R9, R15
-	XORL	R11, R15
-	ADDL	R15, R13
-	ADDL	R12, R13
-	ADDL	$0x5a827999, R13
-	ADDL	R8, R13
-	ROLL	$0x1e, R9
-	MOVL	32(CX), R8
-	BSWAPL	R8
-	MOVL	R8, 32(SP)
-	MOVL	R13, R12
-	ROLL	$0x05, R12
-	MOVL	R10, R15
-	XORL	R9, R15
-	ANDL	R14, R15
-	XORL	R10, R15
-	ADDL	R15, R12
-	ADDL	R11, R12
-	ADDL	$0x5a827999, R12
-	ADDL	R8, R12
-	ROLL	$0x1e, R14
-	MOVL	36(CX), R8
-	BSWAPL	R8
-	MOVL	R8, 36(SP)
-	MOVL	R12, R11
-	ROLL	$0x05, R11
-	MOVL	R9, R15
-	XORL	R14, R15
-	ANDL	R13, R15
-	XORL	R9, R15
-	ADDL	R15, R11
-	ADDL	R10, R11
-	ADDL	$0x5a827999, R11
-	ADDL	R8, R11
-	ROLL	$0x1e, R13
-	MOVL	40(CX), R8
-	BSWAPL	R8
-	MOVL	R8, 40(SP)
-	MOVL	R11, R10
-	ROLL	$0x05, R10
-	MOVL	R14, R15
-	XORL	R13, R15
-	ANDL	R12, R15
-	XORL	R14, R15
-	ADDL	R15, R10
-	ADDL	R9, R10
-	ADDL	$0x5a827999, R10
-	ADDL	R8, R10
-	ROLL	$0x1e, R12
-	MOVL	44(CX), R8
-	BSWAPL	R8
-	MOVL	R8, 44(SP)
-	MOVL	R10, R9
-	ROLL	$0x05, R9
-	MOVL	R13, R15
-	XORL	R12, R15
-	ANDL	R11, R15
-	XORL	R13, R15
-	ADDL	R15, R9
-	ADDL	R14, R9
-	ADDL	$0x5a827999, R9
-	ADDL	R8, R9
-	ROLL	$0x1e, R11
-	MOVL	48(CX), R8
-	BSWAPL	R8
-	MOVL	R8, 48(SP)
-	MOVL	R9, R14
-	ROLL	$0x05, R14
-	MOVL	R12, R15
-	XORL	R11, R15
-	ANDL	R10, R15
-	XORL	R12, R15
-	ADDL	R15, R14
-	ADDL	R13, R14
-	ADDL	$0x5a827999, R14
-	ADDL	R8, R14
-	ROLL	$0x1e, R10
-	MOVL	52(CX), R8
-	BSWAPL	R8
-	MOVL	R8, 52(SP)
-	MOVL	R14, R13
-	ROLL	$0x05, R13
-	MOVL	R11, R15
-	XORL	R10, R15
-	ANDL	R9, R15
-	XORL	R11, R15
-	ADDL	R15, R13
-	ADDL	R12, R13
-	ADDL	$0x5a827999, R13
-	ADDL	R8, R13
-	ROLL	$0x1e, R9
-	MOVL	56(CX), R8
-	BSWAPL	R8
-	MOVL	R8, 56(SP)
-	MOVL	R13, R12
-	ROLL	$0x05, R12
-	MOVL	R10, R15
-	XORL	R9, R15
-	ANDL	R14, R15
-	XORL	R10, R15
-	ADDL	R15, R12
-	ADDL	R11, R12
-	ADDL	$0x5a827999, R12
-	ADDL	R8, R12
-	ROLL	$0x1e, R14
-	MOVL	60(CX), CX
-	BSWAPL	CX
-	MOVL	CX, 60(SP)
-	MOVL	R12, R8
-	ROLL	$0x05, R8
-	MOVL	R9, R11
-	XORL	R14, R11
-	ANDL	R13, R11
-	XORL	R9, R11
-	ADDL	R11, R8
-	ADDL	R10, R8
-	ADDL	$0x5a827999, R8
-	ADDL	CX, R8
-	ROLL	$0x1e, R13
-	MOVL	52(SP), CX
-	XORL	32(SP), CX
-	XORL	8(SP), CX
-	XORL	(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, (SP)
-	MOVL	R8, R10
-	ROLL	$0x05, R10
-	MOVL	R14, R11
-	XORL	R13, R11
-	ANDL	R12, R11
-	XORL	R14, R11
-	ADDL	R11, R10
-	ADDL	R9, R10
-	ADDL	$0x5a827999, R10
-	ADDL	CX, R10
-	ROLL	$0x1e, R12
-	MOVL	56(SP), CX
-	XORL	36(SP), CX
-	XORL	12(SP), CX
-	XORL	4(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 4(SP)
-	MOVL	R10, R9
-	ROLL	$0x05, R9
-	MOVL	R13, R11
-	XORL	R12, R11
-	ANDL	R8, R11
-	XORL	R13, R11
-	ADDL	R11, R9
-	ADDL	R14, R9
-	ADDL	$0x5a827999, R9
-	ADDL	CX, R9
-	ROLL	$0x1e, R8
-	MOVL	60(SP), CX
-	XORL	40(SP), CX
-	XORL	16(SP), CX
-	XORL	8(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 8(SP)
-	MOVL	R9, R11
-	ROLL	$0x05, R11
-	MOVL	R12, R14
-	XORL	R8, R14
-	ANDL	R10, R14
-	XORL	R12, R14
-	ADDL	R14, R11
-	ADDL	R13, R11
-	ADDL	$0x5a827999, R11
-	ADDL	CX, R11
-	ROLL	$0x1e, R10
-	MOVL	(SP), CX
-	XORL	44(SP), CX
-	XORL	20(SP), CX
-	XORL	12(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 12(SP)
-	MOVL	R11, R13
-	ROLL	$0x05, R13
-	MOVL	R8, R14
-	XORL	R10, R14
-	ANDL	R9, R14
-	XORL	R8, R14
-	ADDL	R14, R13
-	ADDL	R12, R13
-	ADDL	$0x5a827999, R13
-	ADDL	CX, R13
-	ROLL	$0x1e, R9
-	MOVL	4(SP), CX
-	XORL	48(SP), CX
-	XORL	24(SP), CX
-	XORL	16(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 16(SP)
-	MOVL	R13, R12
-	ROLL	$0x05, R12
-	MOVL	R11, R14
-	XORL	R9, R14
-	XORL	R10, R14
-	ADDL	R14, R12
-	ADDL	R8, R12
-	ADDL	$0x6ed9eba1, R12
-	ADDL	CX, R12
-	ROLL	$0x1e, R11
-	MOVL	8(SP), CX
-	XORL	52(SP), CX
-	XORL	28(SP), CX
-	XORL	20(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 20(SP)
-	MOVL	R12, R8
-	ROLL	$0x05, R8
-	MOVL	R13, R14
-	XORL	R11, R14
-	XORL	R9, R14
-	ADDL	R14, R8
-	ADDL	R10, R8
-	ADDL	$0x6ed9eba1, R8
-	ADDL	CX, R8
-	ROLL	$0x1e, R13
-	MOVL	12(SP), CX
-	XORL	56(SP), CX
-	XORL	32(SP), CX
-	XORL	24(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 24(SP)
-	MOVL	R8, R10
-	ROLL	$0x05, R10
-	MOVL	R12, R14
-	XORL	R13, R14
-	XORL	R11, R14
-	ADDL	R14, R10
-	ADDL	R9, R10
-	ADDL	$0x6ed9eba1, R10
-	ADDL	CX, R10
-	ROLL	$0x1e, R12
-	MOVL	16(SP), CX
-	XORL	60(SP), CX
-	XORL	36(SP), CX
-	XORL	28(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 28(SP)
-	MOVL	R10, R9
-	ROLL	$0x05, R9
-	MOVL	R8, R14
-	XORL	R12, R14
-	XORL	R13, R14
-	ADDL	R14, R9
-	ADDL	R11, R9
-	ADDL	$0x6ed9eba1, R9
-	ADDL	CX, R9
-	ROLL	$0x1e, R8
-	MOVL	20(SP), CX
-	XORL	(SP), CX
-	XORL	40(SP), CX
-	XORL	32(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 32(SP)
-	MOVL	R9, R11
-	ROLL	$0x05, R11
-	MOVL	R10, R14
-	XORL	R8, R14
-	XORL	R12, R14
-	ADDL	R14, R11
-	ADDL	R13, R11
-	ADDL	$0x6ed9eba1, R11
-	ADDL	CX, R11
-	ROLL	$0x1e, R10
-	MOVL	24(SP), CX
-	XORL	4(SP), CX
-	XORL	44(SP), CX
-	XORL	36(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 36(SP)
-	MOVL	R11, R13
-	ROLL	$0x05, R13
-	MOVL	R9, R14
-	XORL	R10, R14
-	XORL	R8, R14
-	ADDL	R14, R13
-	ADDL	R12, R13
-	ADDL	$0x6ed9eba1, R13
-	ADDL	CX, R13
-	ROLL	$0x1e, R9
-	MOVL	28(SP), CX
-	XORL	8(SP), CX
-	XORL	48(SP), CX
-	XORL	40(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 40(SP)
-	MOVL	R13, R12
-	ROLL	$0x05, R12
-	MOVL	R11, R14
-	XORL	R9, R14
-	XORL	R10, R14
-	ADDL	R14, R12
-	ADDL	R8, R12
-	ADDL	$0x6ed9eba1, R12
-	ADDL	CX, R12
-	ROLL	$0x1e, R11
-	MOVL	32(SP), CX
-	XORL	12(SP), CX
-	XORL	52(SP), CX
-	XORL	44(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 44(SP)
-	MOVL	R12, R8
-	ROLL	$0x05, R8
-	MOVL	R13, R14
-	XORL	R11, R14
-	XORL	R9, R14
-	ADDL	R14, R8
-	ADDL	R10, R8
-	ADDL	$0x6ed9eba1, R8
-	ADDL	CX, R8
-	ROLL	$0x1e, R13
-	MOVL	36(SP), CX
-	XORL	16(SP), CX
-	XORL	56(SP), CX
-	XORL	48(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 48(SP)
-	MOVL	R8, R10
-	ROLL	$0x05, R10
-	MOVL	R12, R14
-	XORL	R13, R14
-	XORL	R11, R14
-	ADDL	R14, R10
-	ADDL	R9, R10
-	ADDL	$0x6ed9eba1, R10
-	ADDL	CX, R10
-	ROLL	$0x1e, R12
-	MOVL	40(SP), CX
-	XORL	20(SP), CX
-	XORL	60(SP), CX
-	XORL	52(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 52(SP)
-	MOVL	R10, R9
-	ROLL	$0x05, R9
-	MOVL	R8, R14
-	XORL	R12, R14
-	XORL	R13, R14
-	ADDL	R14, R9
-	ADDL	R11, R9
-	ADDL	$0x6ed9eba1, R9
-	ADDL	CX, R9
-	ROLL	$0x1e, R8
-	MOVL	44(SP), CX
-	XORL	24(SP), CX
-	XORL	(SP), CX
-	XORL	56(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 56(SP)
-	MOVL	R9, R11
-	ROLL	$0x05, R11
-	MOVL	R10, R14
-	XORL	R8, R14
-	XORL	R12, R14
-	ADDL	R14, R11
-	ADDL	R13, R11
-	ADDL	$0x6ed9eba1, R11
-	ADDL	CX, R11
-	ROLL	$0x1e, R10
-	MOVL	48(SP), CX
-	XORL	28(SP), CX
-	XORL	4(SP), CX
-	XORL	60(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 60(SP)
-	MOVL	R11, R13
-	ROLL	$0x05, R13
-	MOVL	R9, R14
-	XORL	R10, R14
-	XORL	R8, R14
-	ADDL	R14, R13
-	ADDL	R12, R13
-	ADDL	$0x6ed9eba1, R13
-	ADDL	CX, R13
-	ROLL	$0x1e, R9
-	MOVL	52(SP), CX
-	XORL	32(SP), CX
-	XORL	8(SP), CX
-	XORL	(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, (SP)
-	MOVL	R13, R12
-	ROLL	$0x05, R12
-	MOVL	R11, R14
-	XORL	R9, R14
-	XORL	R10, R14
-	ADDL	R14, R12
-	ADDL	R8, R12
-	ADDL	$0x6ed9eba1, R12
-	ADDL	CX, R12
-	ROLL	$0x1e, R11
-	MOVL	56(SP), CX
-	XORL	36(SP), CX
-	XORL	12(SP), CX
-	XORL	4(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 4(SP)
-	MOVL	R12, R8
-	ROLL	$0x05, R8
-	MOVL	R13, R14
-	XORL	R11, R14
-	XORL	R9, R14
-	ADDL	R14, R8
-	ADDL	R10, R8
-	ADDL	$0x6ed9eba1, R8
-	ADDL	CX, R8
-	ROLL	$0x1e, R13
-	MOVL	60(SP), CX
-	XORL	40(SP), CX
-	XORL	16(SP), CX
-	XORL	8(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 8(SP)
-	MOVL	R8, R10
-	ROLL	$0x05, R10
-	MOVL	R12, R14
-	XORL	R13, R14
-	XORL	R11, R14
-	ADDL	R14, R10
-	ADDL	R9, R10
-	ADDL	$0x6ed9eba1, R10
-	ADDL	CX, R10
-	ROLL	$0x1e, R12
-	MOVL	(SP), CX
-	XORL	44(SP), CX
-	XORL	20(SP), CX
-	XORL	12(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 12(SP)
-	MOVL	R10, R9
-	ROLL	$0x05, R9
-	MOVL	R8, R14
-	XORL	R12, R14
-	XORL	R13, R14
-	ADDL	R14, R9
-	ADDL	R11, R9
-	ADDL	$0x6ed9eba1, R9
-	ADDL	CX, R9
-	ROLL	$0x1e, R8
-	MOVL	4(SP), CX
-	XORL	48(SP), CX
-	XORL	24(SP), CX
-	XORL	16(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 16(SP)
-	MOVL	R9, R11
-	ROLL	$0x05, R11
-	MOVL	R10, R14
-	XORL	R8, R14
-	XORL	R12, R14
-	ADDL	R14, R11
-	ADDL	R13, R11
-	ADDL	$0x6ed9eba1, R11
-	ADDL	CX, R11
-	ROLL	$0x1e, R10
-	MOVL	8(SP), CX
-	XORL	52(SP), CX
-	XORL	28(SP), CX
-	XORL	20(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 20(SP)
-	MOVL	R11, R13
-	ROLL	$0x05, R13
-	MOVL	R9, R14
-	XORL	R10, R14
-	XORL	R8, R14
-	ADDL	R14, R13
-	ADDL	R12, R13
-	ADDL	$0x6ed9eba1, R13
-	ADDL	CX, R13
-	ROLL	$0x1e, R9
-	MOVL	12(SP), CX
-	XORL	56(SP), CX
-	XORL	32(SP), CX
-	XORL	24(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 24(SP)
-	MOVL	R13, R12
-	ROLL	$0x05, R12
-	MOVL	R11, R14
-	XORL	R9, R14
-	XORL	R10, R14
-	ADDL	R14, R12
-	ADDL	R8, R12
-	ADDL	$0x6ed9eba1, R12
-	ADDL	CX, R12
-	ROLL	$0x1e, R11
-	MOVL	16(SP), CX
-	XORL	60(SP), CX
-	XORL	36(SP), CX
-	XORL	28(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 28(SP)
-	MOVL	R12, R8
-	ROLL	$0x05, R8
-	MOVL	R13, R14
-	XORL	R11, R14
-	XORL	R9, R14
-	ADDL	R14, R8
-	ADDL	R10, R8
-	ADDL	$0x6ed9eba1, R8
-	ADDL	CX, R8
-	ROLL	$0x1e, R13
-	MOVL	20(SP), CX
-	XORL	(SP), CX
-	XORL	40(SP), CX
-	XORL	32(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 32(SP)
-	MOVL	R8, R10
-	ROLL	$0x05, R10
-	MOVL	R12, R14
-	ORL	R13, R14
-	ANDL	R11, R14
-	MOVL	R12, R15
-	ANDL	R13, R15
-	ORL	R14, R15
-	ADDL	R15, R10
-	ADDL	R9, R10
-	ADDL	$0x8f1bbcdc, R10
-	ADDL	CX, R10
-	ROLL	$0x1e, R12
-	MOVL	24(SP), CX
-	XORL	4(SP), CX
-	XORL	44(SP), CX
-	XORL	36(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 36(SP)
-	MOVL	R10, R9
-	ROLL	$0x05, R9
-	MOVL	R8, R14
-	ORL	R12, R14
-	ANDL	R13, R14
-	MOVL	R8, R15
-	ANDL	R12, R15
-	ORL	R14, R15
-	ADDL	R15, R9
-	ADDL	R11, R9
-	ADDL	$0x8f1bbcdc, R9
-	ADDL	CX, R9
-	ROLL	$0x1e, R8
-	MOVL	28(SP), CX
-	XORL	8(SP), CX
-	XORL	48(SP), CX
-	XORL	40(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 40(SP)
-	MOVL	R9, R11
-	ROLL	$0x05, R11
-	MOVL	R10, R14
-	ORL	R8, R14
-	ANDL	R12, R14
-	MOVL	R10, R15
-	ANDL	R8, R15
-	ORL	R14, R15
-	ADDL	R15, R11
-	ADDL	R13, R11
-	ADDL	$0x8f1bbcdc, R11
-	ADDL	CX, R11
-	ROLL	$0x1e, R10
-	MOVL	32(SP), CX
-	XORL	12(SP), CX
-	XORL	52(SP), CX
-	XORL	44(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 44(SP)
-	MOVL	R11, R13
-	ROLL	$0x05, R13
-	MOVL	R9, R14
-	ORL	R10, R14
-	ANDL	R8, R14
-	MOVL	R9, R15
-	ANDL	R10, R15
-	ORL	R14, R15
-	ADDL	R15, R13
-	ADDL	R12, R13
-	ADDL	$0x8f1bbcdc, R13
-	ADDL	CX, R13
-	ROLL	$0x1e, R9
-	MOVL	36(SP), CX
-	XORL	16(SP), CX
-	XORL	56(SP), CX
-	XORL	48(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 48(SP)
-	MOVL	R13, R12
-	ROLL	$0x05, R12
-	MOVL	R11, R14
-	ORL	R9, R14
-	ANDL	R10, R14
-	MOVL	R11, R15
-	ANDL	R9, R15
-	ORL	R14, R15
-	ADDL	R15, R12
-	ADDL	R8, R12
-	ADDL	$0x8f1bbcdc, R12
-	ADDL	CX, R12
-	ROLL	$0x1e, R11
-	MOVL	40(SP), CX
-	XORL	20(SP), CX
-	XORL	60(SP), CX
-	XORL	52(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 52(SP)
-	MOVL	R12, R8
-	ROLL	$0x05, R8
-	MOVL	R13, R14
-	ORL	R11, R14
-	ANDL	R9, R14
-	MOVL	R13, R15
-	ANDL	R11, R15
-	ORL	R14, R15
-	ADDL	R15, R8
-	ADDL	R10, R8
-	ADDL	$0x8f1bbcdc, R8
-	ADDL	CX, R8
-	ROLL	$0x1e, R13
-	MOVL	44(SP), CX
-	XORL	24(SP), CX
-	XORL	(SP), CX
-	XORL	56(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 56(SP)
-	MOVL	R8, R10
-	ROLL	$0x05, R10
-	MOVL	R12, R14
-	ORL	R13, R14
-	ANDL	R11, R14
-	MOVL	R12, R15
-	ANDL	R13, R15
-	ORL	R14, R15
-	ADDL	R15, R10
-	ADDL	R9, R10
-	ADDL	$0x8f1bbcdc, R10
-	ADDL	CX, R10
-	ROLL	$0x1e, R12
-	MOVL	48(SP), CX
-	XORL	28(SP), CX
-	XORL	4(SP), CX
-	XORL	60(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 60(SP)
-	MOVL	R10, R9
-	ROLL	$0x05, R9
-	MOVL	R8, R14
-	ORL	R12, R14
-	ANDL	R13, R14
-	MOVL	R8, R15
-	ANDL	R12, R15
-	ORL	R14, R15
-	ADDL	R15, R9
-	ADDL	R11, R9
-	ADDL	$0x8f1bbcdc, R9
-	ADDL	CX, R9
-	ROLL	$0x1e, R8
-	MOVL	52(SP), CX
-	XORL	32(SP), CX
-	XORL	8(SP), CX
-	XORL	(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, (SP)
-	MOVL	R9, R11
-	ROLL	$0x05, R11
-	MOVL	R10, R14
-	ORL	R8, R14
-	ANDL	R12, R14
-	MOVL	R10, R15
-	ANDL	R8, R15
-	ORL	R14, R15
-	ADDL	R15, R11
-	ADDL	R13, R11
-	ADDL	$0x8f1bbcdc, R11
-	ADDL	CX, R11
-	ROLL	$0x1e, R10
-	MOVL	56(SP), CX
-	XORL	36(SP), CX
-	XORL	12(SP), CX
-	XORL	4(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 4(SP)
-	MOVL	R11, R13
-	ROLL	$0x05, R13
-	MOVL	R9, R14
-	ORL	R10, R14
-	ANDL	R8, R14
-	MOVL	R9, R15
-	ANDL	R10, R15
-	ORL	R14, R15
-	ADDL	R15, R13
-	ADDL	R12, R13
-	ADDL	$0x8f1bbcdc, R13
-	ADDL	CX, R13
-	ROLL	$0x1e, R9
-	MOVL	60(SP), CX
-	XORL	40(SP), CX
-	XORL	16(SP), CX
-	XORL	8(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 8(SP)
-	MOVL	R13, R12
-	ROLL	$0x05, R12
-	MOVL	R11, R14
-	ORL	R9, R14
-	ANDL	R10, R14
-	MOVL	R11, R15
-	ANDL	R9, R15
-	ORL	R14, R15
-	ADDL	R15, R12
-	ADDL	R8, R12
-	ADDL	$0x8f1bbcdc, R12
-	ADDL	CX, R12
-	ROLL	$0x1e, R11
-	MOVL	(SP), CX
-	XORL	44(SP), CX
-	XORL	20(SP), CX
-	XORL	12(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 12(SP)
-	MOVL	R12, R8
-	ROLL	$0x05, R8
-	MOVL	R13, R14
-	ORL	R11, R14
-	ANDL	R9, R14
-	MOVL	R13, R15
-	ANDL	R11, R15
-	ORL	R14, R15
-	ADDL	R15, R8
-	ADDL	R10, R8
-	ADDL	$0x8f1bbcdc, R8
-	ADDL	CX, R8
-	ROLL	$0x1e, R13
-	MOVL	4(SP), CX
-	XORL	48(SP), CX
-	XORL	24(SP), CX
-	XORL	16(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 16(SP)
-	MOVL	R8, R10
-	ROLL	$0x05, R10
-	MOVL	R12, R14
-	ORL	R13, R14
-	ANDL	R11, R14
-	MOVL	R12, R15
-	ANDL	R13, R15
-	ORL	R14, R15
-	ADDL	R15, R10
-	ADDL	R9, R10
-	ADDL	$0x8f1bbcdc, R10
-	ADDL	CX, R10
-	ROLL	$0x1e, R12
-	MOVL	8(SP), CX
-	XORL	52(SP), CX
-	XORL	28(SP), CX
-	XORL	20(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 20(SP)
-	MOVL	R10, R9
-	ROLL	$0x05, R9
-	MOVL	R8, R14
-	ORL	R12, R14
-	ANDL	R13, R14
-	MOVL	R8, R15
-	ANDL	R12, R15
-	ORL	R14, R15
-	ADDL	R15, R9
-	ADDL	R11, R9
-	ADDL	$0x8f1bbcdc, R9
-	ADDL	CX, R9
-	ROLL	$0x1e, R8
-	MOVL	12(SP), CX
-	XORL	56(SP), CX
-	XORL	32(SP), CX
-	XORL	24(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 24(SP)
-	MOVL	R9, R11
-	ROLL	$0x05, R11
-	MOVL	R10, R14
-	ORL	R8, R14
-	ANDL	R12, R14
-	MOVL	R10, R15
-	ANDL	R8, R15
-	ORL	R14, R15
-	ADDL	R15, R11
-	ADDL	R13, R11
-	ADDL	$0x8f1bbcdc, R11
-	ADDL	CX, R11
-	ROLL	$0x1e, R10
-	MOVL	16(SP), CX
-	XORL	60(SP), CX
-	XORL	36(SP), CX
-	XORL	28(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 28(SP)
-	MOVL	R11, R13
-	ROLL	$0x05, R13
-	MOVL	R9, R14
-	ORL	R10, R14
-	ANDL	R8, R14
-	MOVL	R9, R15
-	ANDL	R10, R15
-	ORL	R14, R15
-	ADDL	R15, R13
-	ADDL	R12, R13
-	ADDL	$0x8f1bbcdc, R13
-	ADDL	CX, R13
-	ROLL	$0x1e, R9
-	MOVL	20(SP), CX
-	XORL	(SP), CX
-	XORL	40(SP), CX
-	XORL	32(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 32(SP)
-	MOVL	R13, R12
-	ROLL	$0x05, R12
-	MOVL	R11, R14
-	ORL	R9, R14
-	ANDL	R10, R14
-	MOVL	R11, R15
-	ANDL	R9, R15
-	ORL	R14, R15
-	ADDL	R15, R12
-	ADDL	R8, R12
-	ADDL	$0x8f1bbcdc, R12
-	ADDL	CX, R12
-	ROLL	$0x1e, R11
-	MOVL	24(SP), CX
-	XORL	4(SP), CX
-	XORL	44(SP), CX
-	XORL	36(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 36(SP)
-	MOVL	R12, R8
-	ROLL	$0x05, R8
-	MOVL	R13, R14
-	ORL	R11, R14
-	ANDL	R9, R14
-	MOVL	R13, R15
-	ANDL	R11, R15
-	ORL	R14, R15
-	ADDL	R15, R8
-	ADDL	R10, R8
-	ADDL	$0x8f1bbcdc, R8
-	ADDL	CX, R8
-	ROLL	$0x1e, R13
-	MOVL	28(SP), CX
-	XORL	8(SP), CX
-	XORL	48(SP), CX
-	XORL	40(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 40(SP)
-	MOVL	R8, R10
-	ROLL	$0x05, R10
-	MOVL	R12, R14
-	ORL	R13, R14
-	ANDL	R11, R14
-	MOVL	R12, R15
-	ANDL	R13, R15
-	ORL	R14, R15
-	ADDL	R15, R10
-	ADDL	R9, R10
-	ADDL	$0x8f1bbcdc, R10
-	ADDL	CX, R10
-	ROLL	$0x1e, R12
-	MOVL	32(SP), CX
-	XORL	12(SP), CX
-	XORL	52(SP), CX
-	XORL	44(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 44(SP)
-	MOVL	R10, R9
-	ROLL	$0x05, R9
-	MOVL	R8, R14
-	ORL	R12, R14
-	ANDL	R13, R14
-	MOVL	R8, R15
-	ANDL	R12, R15
-	ORL	R14, R15
-	ADDL	R15, R9
-	ADDL	R11, R9
-	ADDL	$0x8f1bbcdc, R9
-	ADDL	CX, R9
-	ROLL	$0x1e, R8
-	MOVL	36(SP), CX
-	XORL	16(SP), CX
-	XORL	56(SP), CX
-	XORL	48(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 48(SP)
-	MOVL	R9, R11
-	ROLL	$0x05, R11
-	MOVL	R10, R14
-	XORL	R8, R14
-	XORL	R12, R14
-	ADDL	R14, R11
-	ADDL	R13, R11
-	ADDL	$0xca62c1d6, R11
-	ADDL	CX, R11
-	ROLL	$0x1e, R10
-	MOVL	40(SP), CX
-	XORL	20(SP), CX
-	XORL	60(SP), CX
-	XORL	52(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 52(SP)
-	MOVL	R11, R13
-	ROLL	$0x05, R13
-	MOVL	R9, R14
-	XORL	R10, R14
-	XORL	R8, R14
-	ADDL	R14, R13
-	ADDL	R12, R13
-	ADDL	$0xca62c1d6, R13
-	ADDL	CX, R13
-	ROLL	$0x1e, R9
-	MOVL	44(SP), CX
-	XORL	24(SP), CX
-	XORL	(SP), CX
-	XORL	56(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 56(SP)
-	MOVL	R13, R12
-	ROLL	$0x05, R12
-	MOVL	R11, R14
-	XORL	R9, R14
-	XORL	R10, R14
-	ADDL	R14, R12
-	ADDL	R8, R12
-	ADDL	$0xca62c1d6, R12
-	ADDL	CX, R12
-	ROLL	$0x1e, R11
-	MOVL	48(SP), CX
-	XORL	28(SP), CX
-	XORL	4(SP), CX
-	XORL	60(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 60(SP)
-	MOVL	R12, R8
-	ROLL	$0x05, R8
-	MOVL	R13, R14
-	XORL	R11, R14
-	XORL	R9, R14
-	ADDL	R14, R8
-	ADDL	R10, R8
-	ADDL	$0xca62c1d6, R8
-	ADDL	CX, R8
-	ROLL	$0x1e, R13
-	MOVL	52(SP), CX
-	XORL	32(SP), CX
-	XORL	8(SP), CX
-	XORL	(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, (SP)
-	MOVL	R8, R10
-	ROLL	$0x05, R10
-	MOVL	R12, R14
-	XORL	R13, R14
-	XORL	R11, R14
-	ADDL	R14, R10
-	ADDL	R9, R10
-	ADDL	$0xca62c1d6, R10
-	ADDL	CX, R10
-	ROLL	$0x1e, R12
-	MOVL	56(SP), CX
-	XORL	36(SP), CX
-	XORL	12(SP), CX
-	XORL	4(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 4(SP)
-	MOVL	R10, R9
-	ROLL	$0x05, R9
-	MOVL	R8, R14
-	XORL	R12, R14
-	XORL	R13, R14
-	ADDL	R14, R9
-	ADDL	R11, R9
-	ADDL	$0xca62c1d6, R9
-	ADDL	CX, R9
-	ROLL	$0x1e, R8
-	MOVL	60(SP), CX
-	XORL	40(SP), CX
-	XORL	16(SP), CX
-	XORL	8(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 8(SP)
-	MOVL	R9, R11
-	ROLL	$0x05, R11
-	MOVL	R10, R14
-	XORL	R8, R14
-	XORL	R12, R14
-	ADDL	R14, R11
-	ADDL	R13, R11
-	ADDL	$0xca62c1d6, R11
-	ADDL	CX, R11
-	ROLL	$0x1e, R10
-	MOVL	(SP), CX
-	XORL	44(SP), CX
-	XORL	20(SP), CX
-	XORL	12(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 12(SP)
-	MOVL	R11, R13
-	ROLL	$0x05, R13
-	MOVL	R9, R14
-	XORL	R10, R14
-	XORL	R8, R14
-	ADDL	R14, R13
-	ADDL	R12, R13
-	ADDL	$0xca62c1d6, R13
-	ADDL	CX, R13
-	ROLL	$0x1e, R9
-	MOVL	4(SP), CX
-	XORL	48(SP), CX
-	XORL	24(SP), CX
-	XORL	16(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 16(SP)
-	MOVL	R13, R12
-	ROLL	$0x05, R12
-	MOVL	R11, R14
-	XORL	R9, R14
-	XORL	R10, R14
-	ADDL	R14, R12
-	ADDL	R8, R12
-	ADDL	$0xca62c1d6, R12
-	ADDL	CX, R12
-	ROLL	$0x1e, R11
-	MOVL	8(SP), CX
-	XORL	52(SP), CX
-	XORL	28(SP), CX
-	XORL	20(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 20(SP)
-	MOVL	R12, R8
-	ROLL	$0x05, R8
-	MOVL	R13, R14
-	XORL	R11, R14
-	XORL	R9, R14
-	ADDL	R14, R8
-	ADDL	R10, R8
-	ADDL	$0xca62c1d6, R8
-	ADDL	CX, R8
-	ROLL	$0x1e, R13
-	MOVL	12(SP), CX
-	XORL	56(SP), CX
-	XORL	32(SP), CX
-	XORL	24(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 24(SP)
-	MOVL	R8, R10
-	ROLL	$0x05, R10
-	MOVL	R12, R14
-	XORL	R13, R14
-	XORL	R11, R14
-	ADDL	R14, R10
-	ADDL	R9, R10
-	ADDL	$0xca62c1d6, R10
-	ADDL	CX, R10
-	ROLL	$0x1e, R12
-	MOVL	16(SP), CX
-	XORL	60(SP), CX
-	XORL	36(SP), CX
-	XORL	28(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 28(SP)
-	MOVL	R10, R9
-	ROLL	$0x05, R9
-	MOVL	R8, R14
-	XORL	R12, R14
-	XORL	R13, R14
-	ADDL	R14, R9
-	ADDL	R11, R9
-	ADDL	$0xca62c1d6, R9
-	ADDL	CX, R9
-	ROLL	$0x1e, R8
-	MOVL	20(SP), CX
-	XORL	(SP), CX
-	XORL	40(SP), CX
-	XORL	32(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 32(SP)
-	MOVL	R9, R11
-	ROLL	$0x05, R11
-	MOVL	R10, R14
-	XORL	R8, R14
-	XORL	R12, R14
-	ADDL	R14, R11
-	ADDL	R13, R11
-	ADDL	$0xca62c1d6, R11
-	ADDL	CX, R11
-	ROLL	$0x1e, R10
-	MOVL	24(SP), CX
-	XORL	4(SP), CX
-	XORL	44(SP), CX
-	XORL	36(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 36(SP)
-	MOVL	R11, R13
-	ROLL	$0x05, R13
-	MOVL	R9, R14
-	XORL	R10, R14
-	XORL	R8, R14
-	ADDL	R14, R13
-	ADDL	R12, R13
-	ADDL	$0xca62c1d6, R13
-	ADDL	CX, R13
-	ROLL	$0x1e, R9
-	MOVL	28(SP), CX
-	XORL	8(SP), CX
-	XORL	48(SP), CX
-	XORL	40(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 40(SP)
-	MOVL	R13, R12
-	ROLL	$0x05, R12
-	MOVL	R11, R14
-	XORL	R9, R14
-	XORL	R10, R14
-	ADDL	R14, R12
-	ADDL	R8, R12
-	ADDL	$0xca62c1d6, R12
-	ADDL	CX, R12
-	ROLL	$0x1e, R11
-	MOVL	32(SP), CX
-	XORL	12(SP), CX
-	XORL	52(SP), CX
-	XORL	44(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 44(SP)
-	MOVL	R12, R8
-	ROLL	$0x05, R8
-	MOVL	R13, R14
-	XORL	R11, R14
-	XORL	R9, R14
-	ADDL	R14, R8
-	ADDL	R10, R8
-	ADDL	$0xca62c1d6, R8
-	ADDL	CX, R8
-	ROLL	$0x1e, R13
-	MOVL	36(SP), CX
-	XORL	16(SP), CX
-	XORL	56(SP), CX
-	XORL	48(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 48(SP)
-	MOVL	R8, R10
-	ROLL	$0x05, R10
-	MOVL	R12, R14
-	XORL	R13, R14
-	XORL	R11, R14
-	ADDL	R14, R10
-	ADDL	R9, R10
-	ADDL	$0xca62c1d6, R10
-	ADDL	CX, R10
-	ROLL	$0x1e, R12
-	MOVL	40(SP), CX
-	XORL	20(SP), CX
-	XORL	60(SP), CX
-	XORL	52(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 52(SP)
-	MOVL	R10, R9
-	ROLL	$0x05, R9
-	MOVL	R8, R14
-	XORL	R12, R14
-	XORL	R13, R14
-	ADDL	R14, R9
-	ADDL	R11, R9
-	ADDL	$0xca62c1d6, R9
-	ADDL	CX, R9
-	ROLL	$0x1e, R8
-	MOVL	44(SP), CX
-	XORL	24(SP), CX
-	XORL	(SP), CX
-	XORL	56(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 56(SP)
-	MOVL	R9, R11
-	ROLL	$0x05, R11
-	MOVL	R10, R14
-	XORL	R8, R14
-	XORL	R12, R14
-	ADDL	R14, R11
-	ADDL	R13, R11
-	ADDL	$0xca62c1d6, R11
-	ADDL	CX, R11
-	ROLL	$0x1e, R10
-	MOVL	48(SP), CX
-	XORL	28(SP), CX
-	XORL	4(SP), CX
-	XORL	60(SP), CX
-	ROLL	$0x01, CX
-	MOVL	CX, 60(SP)
-	MOVL	R11, R13
-	ROLL	$0x05, R13
-	MOVL	R9, R14
-	XORL	R10, R14
-	XORL	R8, R14
-	ADDL	R14, R13
-	ADDL	R12, R13
-	ADDL	$0xca62c1d6, R13
-	ADDL	CX, R13
-	ROLL	$0x1e, R9
-	ADDL	R13, DX
-	ADDL	R11, BX
-	ADDL	R9, BP
-	ADDL	R10, SI
-	ADDL	R8, DI
-	MOVL	DX, (AX)
-	MOVL	BX, 4(AX)
-	MOVL	BP, 8(AX)
-	MOVL	SI, 12(AX)
-	MOVL	DI, 16(AX)
+	MOVQ   h(FP), AX
+	MOVQ   m_base+8(FP), CX
+	MOVL   (AX), DX
+	MOVL   4(AX), BX
+	MOVL   8(AX), BP
+	MOVL   12(AX), SI
+	MOVL   16(AX), DI
+	MOVL   DX, R8
+	MOVL   BX, R9
+	MOVL   BP, R10
+	MOVL   SI, R11
+	MOVL   DI, R12
+	MOVL   (CX), R13
+	BSWAPL R13
+	MOVL   R13, (SP)
+	MOVL   R8, R14
+	ROLL   $0x05, R14
+	MOVL   R11, R15
+	XORL   R10, R15
+	ANDL   R9, R15
+	XORL   R11, R15
+	ADDL   R15, R14
+	ADDL   R12, R14
+	ADDL   $0x5a827999, R14
+	ADDL   R13, R14
+	ROLL   $0x1e, R9
+	MOVL   4(CX), R12
+	BSWAPL R12
+	MOVL   R12, 4(SP)
+	MOVL   R14, R13
+	ROLL   $0x05, R13
+	MOVL   R10, R15
+	XORL   R9, R15
+	ANDL   R8, R15
+	XORL   R10, R15
+	ADDL   R15, R13
+	ADDL   R11, R13
+	ADDL   $0x5a827999, R13
+	ADDL   R12, R13
+	ROLL   $0x1e, R8
+	MOVL   8(CX), R11
+	BSWAPL R11
+	MOVL   R11, 8(SP)
+	MOVL   R13, R12
+	ROLL   $0x05, R12
+	MOVL   R9, R15
+	XORL   R8, R15
+	ANDL   R14, R15
+	XORL   R9, R15
+	ADDL   R15, R12
+	ADDL   R10, R12
+	ADDL   $0x5a827999, R12
+	ADDL   R11, R12
+	ROLL   $0x1e, R14
+	MOVL   12(CX), R10
+	BSWAPL R10
+	MOVL   R10, 12(SP)
+	MOVL   R12, R11
+	ROLL   $0x05, R11
+	MOVL   R8, R15
+	XORL   R14, R15
+	ANDL   R13, R15
+	XORL   R8, R15
+	ADDL   R15, R11
+	ADDL   R9, R11
+	ADDL   $0x5a827999, R11
+	ADDL   R10, R11
+	ROLL   $0x1e, R13
+	MOVL   16(CX), R9
+	BSWAPL R9
+	MOVL   R9, 16(SP)
+	MOVL   R11, R10
+	ROLL   $0x05, R10
+	MOVL   R14, R15
+	XORL   R13, R15
+	ANDL   R12, R15
+	XORL   R14, R15
+	ADDL   R15, R10
+	ADDL   R8, R10
+	ADDL   $0x5a827999, R10
+	ADDL   R9, R10
+	ROLL   $0x1e, R12
+	MOVL   20(CX), R8
+	BSWAPL R8
+	MOVL   R8, 20(SP)
+	MOVL   R10, R9
+	ROLL   $0x05, R9
+	MOVL   R13, R15
+	XORL   R12, R15
+	ANDL   R11, R15
+	XORL   R13, R15
+	ADDL   R15, R9
+	ADDL   R14, R9
+	ADDL   $0x5a827999, R9
+	ADDL   R8, R9
+	ROLL   $0x1e, R11
+	MOVL   24(CX), R8
+	BSWAPL R8
+	MOVL   R8, 24(SP)
+	MOVL   R9, R14
+	ROLL   $0x05, R14
+	MOVL   R12, R15
+	XORL   R11, R15
+	ANDL   R10, R15
+	XORL   R12, R15
+	ADDL   R15, R14
+	ADDL   R13, R14
+	ADDL   $0x5a827999, R14
+	ADDL   R8, R14
+	ROLL   $0x1e, R10
+	MOVL   28(CX), R8
+	BSWAPL R8
+	MOVL   R8, 28(SP)
+	MOVL   R14, R13
+	ROLL   $0x05, R13
+	MOVL   R11, R15
+	XORL   R10, R15
+	ANDL   R9, R15
+	XORL   R11, R15
+	ADDL   R15, R13
+	ADDL   R12, R13
+	ADDL   $0x5a827999, R13
+	ADDL   R8, R13
+	ROLL   $0x1e, R9
+	MOVL   32(CX), R8
+	BSWAPL R8
+	MOVL   R8, 32(SP)
+	MOVL   R13, R12
+	ROLL   $0x05, R12
+	MOVL   R10, R15
+	XORL   R9, R15
+	ANDL   R14, R15
+	XORL   R10, R15
+	ADDL   R15, R12
+	ADDL   R11, R12
+	ADDL   $0x5a827999, R12
+	ADDL   R8, R12
+	ROLL   $0x1e, R14
+	MOVL   36(CX), R8
+	BSWAPL R8
+	MOVL   R8, 36(SP)
+	MOVL   R12, R11
+	ROLL   $0x05, R11
+	MOVL   R9, R15
+	XORL   R14, R15
+	ANDL   R13, R15
+	XORL   R9, R15
+	ADDL   R15, R11
+	ADDL   R10, R11
+	ADDL   $0x5a827999, R11
+	ADDL   R8, R11
+	ROLL   $0x1e, R13
+	MOVL   40(CX), R8
+	BSWAPL R8
+	MOVL   R8, 40(SP)
+	MOVL   R11, R10
+	ROLL   $0x05, R10
+	MOVL   R14, R15
+	XORL   R13, R15
+	ANDL   R12, R15
+	XORL   R14, R15
+	ADDL   R15, R10
+	ADDL   R9, R10
+	ADDL   $0x5a827999, R10
+	ADDL   R8, R10
+	ROLL   $0x1e, R12
+	MOVL   44(CX), R8
+	BSWAPL R8
+	MOVL   R8, 44(SP)
+	MOVL   R10, R9
+	ROLL   $0x05, R9
+	MOVL   R13, R15
+	XORL   R12, R15
+	ANDL   R11, R15
+	XORL   R13, R15
+	ADDL   R15, R9
+	ADDL   R14, R9
+	ADDL   $0x5a827999, R9
+	ADDL   R8, R9
+	ROLL   $0x1e, R11
+	MOVL   48(CX), R8
+	BSWAPL R8
+	MOVL   R8, 48(SP)
+	MOVL   R9, R14
+	ROLL   $0x05, R14
+	MOVL   R12, R15
+	XORL   R11, R15
+	ANDL   R10, R15
+	XORL   R12, R15
+	ADDL   R15, R14
+	ADDL   R13, R14
+	ADDL   $0x5a827999, R14
+	ADDL   R8, R14
+	ROLL   $0x1e, R10
+	MOVL   52(CX), R8
+	BSWAPL R8
+	MOVL   R8, 52(SP)
+	MOVL   R14, R13
+	ROLL   $0x05, R13
+	MOVL   R11, R15
+	XORL   R10, R15
+	ANDL   R9, R15
+	XORL   R11, R15
+	ADDL   R15, R13
+	ADDL   R12, R13
+	ADDL   $0x5a827999, R13
+	ADDL   R8, R13
+	ROLL   $0x1e, R9
+	MOVL   56(CX), R8
+	BSWAPL R8
+	MOVL   R8, 56(SP)
+	MOVL   R13, R12
+	ROLL   $0x05, R12
+	MOVL   R10, R15
+	XORL   R9, R15
+	ANDL   R14, R15
+	XORL   R10, R15
+	ADDL   R15, R12
+	ADDL   R11, R12
+	ADDL   $0x5a827999, R12
+	ADDL   R8, R12
+	ROLL   $0x1e, R14
+	MOVL   60(CX), CX
+	BSWAPL CX
+	MOVL   CX, 60(SP)
+	MOVL   R12, R8
+	ROLL   $0x05, R8
+	MOVL   R9, R11
+	XORL   R14, R11
+	ANDL   R13, R11
+	XORL   R9, R11
+	ADDL   R11, R8
+	ADDL   R10, R8
+	ADDL   $0x5a827999, R8
+	ADDL   CX, R8
+	ROLL   $0x1e, R13
+	MOVL   52(SP), CX
+	XORL   32(SP), CX
+	XORL   8(SP), CX
+	XORL   (SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, (SP)
+	MOVL   R8, R10
+	ROLL   $0x05, R10
+	MOVL   R14, R11
+	XORL   R13, R11
+	ANDL   R12, R11
+	XORL   R14, R11
+	ADDL   R11, R10
+	ADDL   R9, R10
+	ADDL   $0x5a827999, R10
+	ADDL   CX, R10
+	ROLL   $0x1e, R12
+	MOVL   56(SP), CX
+	XORL   36(SP), CX
+	XORL   12(SP), CX
+	XORL   4(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 4(SP)
+	MOVL   R10, R9
+	ROLL   $0x05, R9
+	MOVL   R13, R11
+	XORL   R12, R11
+	ANDL   R8, R11
+	XORL   R13, R11
+	ADDL   R11, R9
+	ADDL   R14, R9
+	ADDL   $0x5a827999, R9
+	ADDL   CX, R9
+	ROLL   $0x1e, R8
+	MOVL   60(SP), CX
+	XORL   40(SP), CX
+	XORL   16(SP), CX
+	XORL   8(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 8(SP)
+	MOVL   R9, R11
+	ROLL   $0x05, R11
+	MOVL   R12, R14
+	XORL   R8, R14
+	ANDL   R10, R14
+	XORL   R12, R14
+	ADDL   R14, R11
+	ADDL   R13, R11
+	ADDL   $0x5a827999, R11
+	ADDL   CX, R11
+	ROLL   $0x1e, R10
+	MOVL   (SP), CX
+	XORL   44(SP), CX
+	XORL   20(SP), CX
+	XORL   12(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 12(SP)
+	MOVL   R11, R13
+	ROLL   $0x05, R13
+	MOVL   R8, R14
+	XORL   R10, R14
+	ANDL   R9, R14
+	XORL   R8, R14
+	ADDL   R14, R13
+	ADDL   R12, R13
+	ADDL   $0x5a827999, R13
+	ADDL   CX, R13
+	ROLL   $0x1e, R9
+	MOVL   4(SP), CX
+	XORL   48(SP), CX
+	XORL   24(SP), CX
+	XORL   16(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 16(SP)
+	MOVL   R13, R12
+	ROLL   $0x05, R12
+	MOVL   R11, R14
+	XORL   R9, R14
+	XORL   R10, R14
+	ADDL   R14, R12
+	ADDL   R8, R12
+	ADDL   $0x6ed9eba1, R12
+	ADDL   CX, R12
+	ROLL   $0x1e, R11
+	MOVL   8(SP), CX
+	XORL   52(SP), CX
+	XORL   28(SP), CX
+	XORL   20(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 20(SP)
+	MOVL   R12, R8
+	ROLL   $0x05, R8
+	MOVL   R13, R14
+	XORL   R11, R14
+	XORL   R9, R14
+	ADDL   R14, R8
+	ADDL   R10, R8
+	ADDL   $0x6ed9eba1, R8
+	ADDL   CX, R8
+	ROLL   $0x1e, R13
+	MOVL   12(SP), CX
+	XORL   56(SP), CX
+	XORL   32(SP), CX
+	XORL   24(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 24(SP)
+	MOVL   R8, R10
+	ROLL   $0x05, R10
+	MOVL   R12, R14
+	XORL   R13, R14
+	XORL   R11, R14
+	ADDL   R14, R10
+	ADDL   R9, R10
+	ADDL   $0x6ed9eba1, R10
+	ADDL   CX, R10
+	ROLL   $0x1e, R12
+	MOVL   16(SP), CX
+	XORL   60(SP), CX
+	XORL   36(SP), CX
+	XORL   28(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 28(SP)
+	MOVL   R10, R9
+	ROLL   $0x05, R9
+	MOVL   R8, R14
+	XORL   R12, R14
+	XORL   R13, R14
+	ADDL   R14, R9
+	ADDL   R11, R9
+	ADDL   $0x6ed9eba1, R9
+	ADDL   CX, R9
+	ROLL   $0x1e, R8
+	MOVL   20(SP), CX
+	XORL   (SP), CX
+	XORL   40(SP), CX
+	XORL   32(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 32(SP)
+	MOVL   R9, R11
+	ROLL   $0x05, R11
+	MOVL   R10, R14
+	XORL   R8, R14
+	XORL   R12, R14
+	ADDL   R14, R11
+	ADDL   R13, R11
+	ADDL   $0x6ed9eba1, R11
+	ADDL   CX, R11
+	ROLL   $0x1e, R10
+	MOVL   24(SP), CX
+	XORL   4(SP), CX
+	XORL   44(SP), CX
+	XORL   36(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 36(SP)
+	MOVL   R11, R13
+	ROLL   $0x05, R13
+	MOVL   R9, R14
+	XORL   R10, R14
+	XORL   R8, R14
+	ADDL   R14, R13
+	ADDL   R12, R13
+	ADDL   $0x6ed9eba1, R13
+	ADDL   CX, R13
+	ROLL   $0x1e, R9
+	MOVL   28(SP), CX
+	XORL   8(SP), CX
+	XORL   48(SP), CX
+	XORL   40(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 40(SP)
+	MOVL   R13, R12
+	ROLL   $0x05, R12
+	MOVL   R11, R14
+	XORL   R9, R14
+	XORL   R10, R14
+	ADDL   R14, R12
+	ADDL   R8, R12
+	ADDL   $0x6ed9eba1, R12
+	ADDL   CX, R12
+	ROLL   $0x1e, R11
+	MOVL   32(SP), CX
+	XORL   12(SP), CX
+	XORL   52(SP), CX
+	XORL   44(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 44(SP)
+	MOVL   R12, R8
+	ROLL   $0x05, R8
+	MOVL   R13, R14
+	XORL   R11, R14
+	XORL   R9, R14
+	ADDL   R14, R8
+	ADDL   R10, R8
+	ADDL   $0x6ed9eba1, R8
+	ADDL   CX, R8
+	ROLL   $0x1e, R13
+	MOVL   36(SP), CX
+	XORL   16(SP), CX
+	XORL   56(SP), CX
+	XORL   48(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 48(SP)
+	MOVL   R8, R10
+	ROLL   $0x05, R10
+	MOVL   R12, R14
+	XORL   R13, R14
+	XORL   R11, R14
+	ADDL   R14, R10
+	ADDL   R9, R10
+	ADDL   $0x6ed9eba1, R10
+	ADDL   CX, R10
+	ROLL   $0x1e, R12
+	MOVL   40(SP), CX
+	XORL   20(SP), CX
+	XORL   60(SP), CX
+	XORL   52(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 52(SP)
+	MOVL   R10, R9
+	ROLL   $0x05, R9
+	MOVL   R8, R14
+	XORL   R12, R14
+	XORL   R13, R14
+	ADDL   R14, R9
+	ADDL   R11, R9
+	ADDL   $0x6ed9eba1, R9
+	ADDL   CX, R9
+	ROLL   $0x1e, R8
+	MOVL   44(SP), CX
+	XORL   24(SP), CX
+	XORL   (SP), CX
+	XORL   56(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 56(SP)
+	MOVL   R9, R11
+	ROLL   $0x05, R11
+	MOVL   R10, R14
+	XORL   R8, R14
+	XORL   R12, R14
+	ADDL   R14, R11
+	ADDL   R13, R11
+	ADDL   $0x6ed9eba1, R11
+	ADDL   CX, R11
+	ROLL   $0x1e, R10
+	MOVL   48(SP), CX
+	XORL   28(SP), CX
+	XORL   4(SP), CX
+	XORL   60(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 60(SP)
+	MOVL   R11, R13
+	ROLL   $0x05, R13
+	MOVL   R9, R14
+	XORL   R10, R14
+	XORL   R8, R14
+	ADDL   R14, R13
+	ADDL   R12, R13
+	ADDL   $0x6ed9eba1, R13
+	ADDL   CX, R13
+	ROLL   $0x1e, R9
+	MOVL   52(SP), CX
+	XORL   32(SP), CX
+	XORL   8(SP), CX
+	XORL   (SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, (SP)
+	MOVL   R13, R12
+	ROLL   $0x05, R12
+	MOVL   R11, R14
+	XORL   R9, R14
+	XORL   R10, R14
+	ADDL   R14, R12
+	ADDL   R8, R12
+	ADDL   $0x6ed9eba1, R12
+	ADDL   CX, R12
+	ROLL   $0x1e, R11
+	MOVL   56(SP), CX
+	XORL   36(SP), CX
+	XORL   12(SP), CX
+	XORL   4(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 4(SP)
+	MOVL   R12, R8
+	ROLL   $0x05, R8
+	MOVL   R13, R14
+	XORL   R11, R14
+	XORL   R9, R14
+	ADDL   R14, R8
+	ADDL   R10, R8
+	ADDL   $0x6ed9eba1, R8
+	ADDL   CX, R8
+	ROLL   $0x1e, R13
+	MOVL   60(SP), CX
+	XORL   40(SP), CX
+	XORL   16(SP), CX
+	XORL   8(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 8(SP)
+	MOVL   R8, R10
+	ROLL   $0x05, R10
+	MOVL   R12, R14
+	XORL   R13, R14
+	XORL   R11, R14
+	ADDL   R14, R10
+	ADDL   R9, R10
+	ADDL   $0x6ed9eba1, R10
+	ADDL   CX, R10
+	ROLL   $0x1e, R12
+	MOVL   (SP), CX
+	XORL   44(SP), CX
+	XORL   20(SP), CX
+	XORL   12(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 12(SP)
+	MOVL   R10, R9
+	ROLL   $0x05, R9
+	MOVL   R8, R14
+	XORL   R12, R14
+	XORL   R13, R14
+	ADDL   R14, R9
+	ADDL   R11, R9
+	ADDL   $0x6ed9eba1, R9
+	ADDL   CX, R9
+	ROLL   $0x1e, R8
+	MOVL   4(SP), CX
+	XORL   48(SP), CX
+	XORL   24(SP), CX
+	XORL   16(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 16(SP)
+	MOVL   R9, R11
+	ROLL   $0x05, R11
+	MOVL   R10, R14
+	XORL   R8, R14
+	XORL   R12, R14
+	ADDL   R14, R11
+	ADDL   R13, R11
+	ADDL   $0x6ed9eba1, R11
+	ADDL   CX, R11
+	ROLL   $0x1e, R10
+	MOVL   8(SP), CX
+	XORL   52(SP), CX
+	XORL   28(SP), CX
+	XORL   20(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 20(SP)
+	MOVL   R11, R13
+	ROLL   $0x05, R13
+	MOVL   R9, R14
+	XORL   R10, R14
+	XORL   R8, R14
+	ADDL   R14, R13
+	ADDL   R12, R13
+	ADDL   $0x6ed9eba1, R13
+	ADDL   CX, R13
+	ROLL   $0x1e, R9
+	MOVL   12(SP), CX
+	XORL   56(SP), CX
+	XORL   32(SP), CX
+	XORL   24(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 24(SP)
+	MOVL   R13, R12
+	ROLL   $0x05, R12
+	MOVL   R11, R14
+	XORL   R9, R14
+	XORL   R10, R14
+	ADDL   R14, R12
+	ADDL   R8, R12
+	ADDL   $0x6ed9eba1, R12
+	ADDL   CX, R12
+	ROLL   $0x1e, R11
+	MOVL   16(SP), CX
+	XORL   60(SP), CX
+	XORL   36(SP), CX
+	XORL   28(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 28(SP)
+	MOVL   R12, R8
+	ROLL   $0x05, R8
+	MOVL   R13, R14
+	XORL   R11, R14
+	XORL   R9, R14
+	ADDL   R14, R8
+	ADDL   R10, R8
+	ADDL   $0x6ed9eba1, R8
+	ADDL   CX, R8
+	ROLL   $0x1e, R13
+	MOVL   20(SP), CX
+	XORL   (SP), CX
+	XORL   40(SP), CX
+	XORL   32(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 32(SP)
+	MOVL   R8, R10
+	ROLL   $0x05, R10
+	MOVL   R12, R14
+	ORL    R13, R14
+	ANDL   R11, R14
+	MOVL   R12, R15
+	ANDL   R13, R15
+	ORL    R14, R15
+	ADDL   R15, R10
+	ADDL   R9, R10
+	ADDL   $0x8f1bbcdc, R10
+	ADDL   CX, R10
+	ROLL   $0x1e, R12
+	MOVL   24(SP), CX
+	XORL   4(SP), CX
+	XORL   44(SP), CX
+	XORL   36(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 36(SP)
+	MOVL   R10, R9
+	ROLL   $0x05, R9
+	MOVL   R8, R14
+	ORL    R12, R14
+	ANDL   R13, R14
+	MOVL   R8, R15
+	ANDL   R12, R15
+	ORL    R14, R15
+	ADDL   R15, R9
+	ADDL   R11, R9
+	ADDL   $0x8f1bbcdc, R9
+	ADDL   CX, R9
+	ROLL   $0x1e, R8
+	MOVL   28(SP), CX
+	XORL   8(SP), CX
+	XORL   48(SP), CX
+	XORL   40(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 40(SP)
+	MOVL   R9, R11
+	ROLL   $0x05, R11
+	MOVL   R10, R14
+	ORL    R8, R14
+	ANDL   R12, R14
+	MOVL   R10, R15
+	ANDL   R8, R15
+	ORL    R14, R15
+	ADDL   R15, R11
+	ADDL   R13, R11
+	ADDL   $0x8f1bbcdc, R11
+	ADDL   CX, R11
+	ROLL   $0x1e, R10
+	MOVL   32(SP), CX
+	XORL   12(SP), CX
+	XORL   52(SP), CX
+	XORL   44(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 44(SP)
+	MOVL   R11, R13
+	ROLL   $0x05, R13
+	MOVL   R9, R14
+	ORL    R10, R14
+	ANDL   R8, R14
+	MOVL   R9, R15
+	ANDL   R10, R15
+	ORL    R14, R15
+	ADDL   R15, R13
+	ADDL   R12, R13
+	ADDL   $0x8f1bbcdc, R13
+	ADDL   CX, R13
+	ROLL   $0x1e, R9
+	MOVL   36(SP), CX
+	XORL   16(SP), CX
+	XORL   56(SP), CX
+	XORL   48(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 48(SP)
+	MOVL   R13, R12
+	ROLL   $0x05, R12
+	MOVL   R11, R14
+	ORL    R9, R14
+	ANDL   R10, R14
+	MOVL   R11, R15
+	ANDL   R9, R15
+	ORL    R14, R15
+	ADDL   R15, R12
+	ADDL   R8, R12
+	ADDL   $0x8f1bbcdc, R12
+	ADDL   CX, R12
+	ROLL   $0x1e, R11
+	MOVL   40(SP), CX
+	XORL   20(SP), CX
+	XORL   60(SP), CX
+	XORL   52(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 52(SP)
+	MOVL   R12, R8
+	ROLL   $0x05, R8
+	MOVL   R13, R14
+	ORL    R11, R14
+	ANDL   R9, R14
+	MOVL   R13, R15
+	ANDL   R11, R15
+	ORL    R14, R15
+	ADDL   R15, R8
+	ADDL   R10, R8
+	ADDL   $0x8f1bbcdc, R8
+	ADDL   CX, R8
+	ROLL   $0x1e, R13
+	MOVL   44(SP), CX
+	XORL   24(SP), CX
+	XORL   (SP), CX
+	XORL   56(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 56(SP)
+	MOVL   R8, R10
+	ROLL   $0x05, R10
+	MOVL   R12, R14
+	ORL    R13, R14
+	ANDL   R11, R14
+	MOVL   R12, R15
+	ANDL   R13, R15
+	ORL    R14, R15
+	ADDL   R15, R10
+	ADDL   R9, R10
+	ADDL   $0x8f1bbcdc, R10
+	ADDL   CX, R10
+	ROLL   $0x1e, R12
+	MOVL   48(SP), CX
+	XORL   28(SP), CX
+	XORL   4(SP), CX
+	XORL   60(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 60(SP)
+	MOVL   R10, R9
+	ROLL   $0x05, R9
+	MOVL   R8, R14
+	ORL    R12, R14
+	ANDL   R13, R14
+	MOVL   R8, R15
+	ANDL   R12, R15
+	ORL    R14, R15
+	ADDL   R15, R9
+	ADDL   R11, R9
+	ADDL   $0x8f1bbcdc, R9
+	ADDL   CX, R9
+	ROLL   $0x1e, R8
+	MOVL   52(SP), CX
+	XORL   32(SP), CX
+	XORL   8(SP), CX
+	XORL   (SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, (SP)
+	MOVL   R9, R11
+	ROLL   $0x05, R11
+	MOVL   R10, R14
+	ORL    R8, R14
+	ANDL   R12, R14
+	MOVL   R10, R15
+	ANDL   R8, R15
+	ORL    R14, R15
+	ADDL   R15, R11
+	ADDL   R13, R11
+	ADDL   $0x8f1bbcdc, R11
+	ADDL   CX, R11
+	ROLL   $0x1e, R10
+	MOVL   56(SP), CX
+	XORL   36(SP), CX
+	XORL   12(SP), CX
+	XORL   4(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 4(SP)
+	MOVL   R11, R13
+	ROLL   $0x05, R13
+	MOVL   R9, R14
+	ORL    R10, R14
+	ANDL   R8, R14
+	MOVL   R9, R15
+	ANDL   R10, R15
+	ORL    R14, R15
+	ADDL   R15, R13
+	ADDL   R12, R13
+	ADDL   $0x8f1bbcdc, R13
+	ADDL   CX, R13
+	ROLL   $0x1e, R9
+	MOVL   60(SP), CX
+	XORL   40(SP), CX
+	XORL   16(SP), CX
+	XORL   8(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 8(SP)
+	MOVL   R13, R12
+	ROLL   $0x05, R12
+	MOVL   R11, R14
+	ORL    R9, R14
+	ANDL   R10, R14
+	MOVL   R11, R15
+	ANDL   R9, R15
+	ORL    R14, R15
+	ADDL   R15, R12
+	ADDL   R8, R12
+	ADDL   $0x8f1bbcdc, R12
+	ADDL   CX, R12
+	ROLL   $0x1e, R11
+	MOVL   (SP), CX
+	XORL   44(SP), CX
+	XORL   20(SP), CX
+	XORL   12(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 12(SP)
+	MOVL   R12, R8
+	ROLL   $0x05, R8
+	MOVL   R13, R14
+	ORL    R11, R14
+	ANDL   R9, R14
+	MOVL   R13, R15
+	ANDL   R11, R15
+	ORL    R14, R15
+	ADDL   R15, R8
+	ADDL   R10, R8
+	ADDL   $0x8f1bbcdc, R8
+	ADDL   CX, R8
+	ROLL   $0x1e, R13
+	MOVL   4(SP), CX
+	XORL   48(SP), CX
+	XORL   24(SP), CX
+	XORL   16(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 16(SP)
+	MOVL   R8, R10
+	ROLL   $0x05, R10
+	MOVL   R12, R14
+	ORL    R13, R14
+	ANDL   R11, R14
+	MOVL   R12, R15
+	ANDL   R13, R15
+	ORL    R14, R15
+	ADDL   R15, R10
+	ADDL   R9, R10
+	ADDL   $0x8f1bbcdc, R10
+	ADDL   CX, R10
+	ROLL   $0x1e, R12
+	MOVL   8(SP), CX
+	XORL   52(SP), CX
+	XORL   28(SP), CX
+	XORL   20(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 20(SP)
+	MOVL   R10, R9
+	ROLL   $0x05, R9
+	MOVL   R8, R14
+	ORL    R12, R14
+	ANDL   R13, R14
+	MOVL   R8, R15
+	ANDL   R12, R15
+	ORL    R14, R15
+	ADDL   R15, R9
+	ADDL   R11, R9
+	ADDL   $0x8f1bbcdc, R9
+	ADDL   CX, R9
+	ROLL   $0x1e, R8
+	MOVL   12(SP), CX
+	XORL   56(SP), CX
+	XORL   32(SP), CX
+	XORL   24(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 24(SP)
+	MOVL   R9, R11
+	ROLL   $0x05, R11
+	MOVL   R10, R14
+	ORL    R8, R14
+	ANDL   R12, R14
+	MOVL   R10, R15
+	ANDL   R8, R15
+	ORL    R14, R15
+	ADDL   R15, R11
+	ADDL   R13, R11
+	ADDL   $0x8f1bbcdc, R11
+	ADDL   CX, R11
+	ROLL   $0x1e, R10
+	MOVL   16(SP), CX
+	XORL   60(SP), CX
+	XORL   36(SP), CX
+	XORL   28(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 28(SP)
+	MOVL   R11, R13
+	ROLL   $0x05, R13
+	MOVL   R9, R14
+	ORL    R10, R14
+	ANDL   R8, R14
+	MOVL   R9, R15
+	ANDL   R10, R15
+	ORL    R14, R15
+	ADDL   R15, R13
+	ADDL   R12, R13
+	ADDL   $0x8f1bbcdc, R13
+	ADDL   CX, R13
+	ROLL   $0x1e, R9
+	MOVL   20(SP), CX
+	XORL   (SP), CX
+	XORL   40(SP), CX
+	XORL   32(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 32(SP)
+	MOVL   R13, R12
+	ROLL   $0x05, R12
+	MOVL   R11, R14
+	ORL    R9, R14
+	ANDL   R10, R14
+	MOVL   R11, R15
+	ANDL   R9, R15
+	ORL    R14, R15
+	ADDL   R15, R12
+	ADDL   R8, R12
+	ADDL   $0x8f1bbcdc, R12
+	ADDL   CX, R12
+	ROLL   $0x1e, R11
+	MOVL   24(SP), CX
+	XORL   4(SP), CX
+	XORL   44(SP), CX
+	XORL   36(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 36(SP)
+	MOVL   R12, R8
+	ROLL   $0x05, R8
+	MOVL   R13, R14
+	ORL    R11, R14
+	ANDL   R9, R14
+	MOVL   R13, R15
+	ANDL   R11, R15
+	ORL    R14, R15
+	ADDL   R15, R8
+	ADDL   R10, R8
+	ADDL   $0x8f1bbcdc, R8
+	ADDL   CX, R8
+	ROLL   $0x1e, R13
+	MOVL   28(SP), CX
+	XORL   8(SP), CX
+	XORL   48(SP), CX
+	XORL   40(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 40(SP)
+	MOVL   R8, R10
+	ROLL   $0x05, R10
+	MOVL   R12, R14
+	ORL    R13, R14
+	ANDL   R11, R14
+	MOVL   R12, R15
+	ANDL   R13, R15
+	ORL    R14, R15
+	ADDL   R15, R10
+	ADDL   R9, R10
+	ADDL   $0x8f1bbcdc, R10
+	ADDL   CX, R10
+	ROLL   $0x1e, R12
+	MOVL   32(SP), CX
+	XORL   12(SP), CX
+	XORL   52(SP), CX
+	XORL   44(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 44(SP)
+	MOVL   R10, R9
+	ROLL   $0x05, R9
+	MOVL   R8, R14
+	ORL    R12, R14
+	ANDL   R13, R14
+	MOVL   R8, R15
+	ANDL   R12, R15
+	ORL    R14, R15
+	ADDL   R15, R9
+	ADDL   R11, R9
+	ADDL   $0x8f1bbcdc, R9
+	ADDL   CX, R9
+	ROLL   $0x1e, R8
+	MOVL   36(SP), CX
+	XORL   16(SP), CX
+	XORL   56(SP), CX
+	XORL   48(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 48(SP)
+	MOVL   R9, R11
+	ROLL   $0x05, R11
+	MOVL   R10, R14
+	XORL   R8, R14
+	XORL   R12, R14
+	ADDL   R14, R11
+	ADDL   R13, R11
+	ADDL   $0xca62c1d6, R11
+	ADDL   CX, R11
+	ROLL   $0x1e, R10
+	MOVL   40(SP), CX
+	XORL   20(SP), CX
+	XORL   60(SP), CX
+	XORL   52(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 52(SP)
+	MOVL   R11, R13
+	ROLL   $0x05, R13
+	MOVL   R9, R14
+	XORL   R10, R14
+	XORL   R8, R14
+	ADDL   R14, R13
+	ADDL   R12, R13
+	ADDL   $0xca62c1d6, R13
+	ADDL   CX, R13
+	ROLL   $0x1e, R9
+	MOVL   44(SP), CX
+	XORL   24(SP), CX
+	XORL   (SP), CX
+	XORL   56(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 56(SP)
+	MOVL   R13, R12
+	ROLL   $0x05, R12
+	MOVL   R11, R14
+	XORL   R9, R14
+	XORL   R10, R14
+	ADDL   R14, R12
+	ADDL   R8, R12
+	ADDL   $0xca62c1d6, R12
+	ADDL   CX, R12
+	ROLL   $0x1e, R11
+	MOVL   48(SP), CX
+	XORL   28(SP), CX
+	XORL   4(SP), CX
+	XORL   60(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 60(SP)
+	MOVL   R12, R8
+	ROLL   $0x05, R8
+	MOVL   R13, R14
+	XORL   R11, R14
+	XORL   R9, R14
+	ADDL   R14, R8
+	ADDL   R10, R8
+	ADDL   $0xca62c1d6, R8
+	ADDL   CX, R8
+	ROLL   $0x1e, R13
+	MOVL   52(SP), CX
+	XORL   32(SP), CX
+	XORL   8(SP), CX
+	XORL   (SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, (SP)
+	MOVL   R8, R10
+	ROLL   $0x05, R10
+	MOVL   R12, R14
+	XORL   R13, R14
+	XORL   R11, R14
+	ADDL   R14, R10
+	ADDL   R9, R10
+	ADDL   $0xca62c1d6, R10
+	ADDL   CX, R10
+	ROLL   $0x1e, R12
+	MOVL   56(SP), CX
+	XORL   36(SP), CX
+	XORL   12(SP), CX
+	XORL   4(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 4(SP)
+	MOVL   R10, R9
+	ROLL   $0x05, R9
+	MOVL   R8, R14
+	XORL   R12, R14
+	XORL   R13, R14
+	ADDL   R14, R9
+	ADDL   R11, R9
+	ADDL   $0xca62c1d6, R9
+	ADDL   CX, R9
+	ROLL   $0x1e, R8
+	MOVL   60(SP), CX
+	XORL   40(SP), CX
+	XORL   16(SP), CX
+	XORL   8(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 8(SP)
+	MOVL   R9, R11
+	ROLL   $0x05, R11
+	MOVL   R10, R14
+	XORL   R8, R14
+	XORL   R12, R14
+	ADDL   R14, R11
+	ADDL   R13, R11
+	ADDL   $0xca62c1d6, R11
+	ADDL   CX, R11
+	ROLL   $0x1e, R10
+	MOVL   (SP), CX
+	XORL   44(SP), CX
+	XORL   20(SP), CX
+	XORL   12(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 12(SP)
+	MOVL   R11, R13
+	ROLL   $0x05, R13
+	MOVL   R9, R14
+	XORL   R10, R14
+	XORL   R8, R14
+	ADDL   R14, R13
+	ADDL   R12, R13
+	ADDL   $0xca62c1d6, R13
+	ADDL   CX, R13
+	ROLL   $0x1e, R9
+	MOVL   4(SP), CX
+	XORL   48(SP), CX
+	XORL   24(SP), CX
+	XORL   16(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 16(SP)
+	MOVL   R13, R12
+	ROLL   $0x05, R12
+	MOVL   R11, R14
+	XORL   R9, R14
+	XORL   R10, R14
+	ADDL   R14, R12
+	ADDL   R8, R12
+	ADDL   $0xca62c1d6, R12
+	ADDL   CX, R12
+	ROLL   $0x1e, R11
+	MOVL   8(SP), CX
+	XORL   52(SP), CX
+	XORL   28(SP), CX
+	XORL   20(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 20(SP)
+	MOVL   R12, R8
+	ROLL   $0x05, R8
+	MOVL   R13, R14
+	XORL   R11, R14
+	XORL   R9, R14
+	ADDL   R14, R8
+	ADDL   R10, R8
+	ADDL   $0xca62c1d6, R8
+	ADDL   CX, R8
+	ROLL   $0x1e, R13
+	MOVL   12(SP), CX
+	XORL   56(SP), CX
+	XORL   32(SP), CX
+	XORL   24(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 24(SP)
+	MOVL   R8, R10
+	ROLL   $0x05, R10
+	MOVL   R12, R14
+	XORL   R13, R14
+	XORL   R11, R14
+	ADDL   R14, R10
+	ADDL   R9, R10
+	ADDL   $0xca62c1d6, R10
+	ADDL   CX, R10
+	ROLL   $0x1e, R12
+	MOVL   16(SP), CX
+	XORL   60(SP), CX
+	XORL   36(SP), CX
+	XORL   28(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 28(SP)
+	MOVL   R10, R9
+	ROLL   $0x05, R9
+	MOVL   R8, R14
+	XORL   R12, R14
+	XORL   R13, R14
+	ADDL   R14, R9
+	ADDL   R11, R9
+	ADDL   $0xca62c1d6, R9
+	ADDL   CX, R9
+	ROLL   $0x1e, R8
+	MOVL   20(SP), CX
+	XORL   (SP), CX
+	XORL   40(SP), CX
+	XORL   32(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 32(SP)
+	MOVL   R9, R11
+	ROLL   $0x05, R11
+	MOVL   R10, R14
+	XORL   R8, R14
+	XORL   R12, R14
+	ADDL   R14, R11
+	ADDL   R13, R11
+	ADDL   $0xca62c1d6, R11
+	ADDL   CX, R11
+	ROLL   $0x1e, R10
+	MOVL   24(SP), CX
+	XORL   4(SP), CX
+	XORL   44(SP), CX
+	XORL   36(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 36(SP)
+	MOVL   R11, R13
+	ROLL   $0x05, R13
+	MOVL   R9, R14
+	XORL   R10, R14
+	XORL   R8, R14
+	ADDL   R14, R13
+	ADDL   R12, R13
+	ADDL   $0xca62c1d6, R13
+	ADDL   CX, R13
+	ROLL   $0x1e, R9
+	MOVL   28(SP), CX
+	XORL   8(SP), CX
+	XORL   48(SP), CX
+	XORL   40(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 40(SP)
+	MOVL   R13, R12
+	ROLL   $0x05, R12
+	MOVL   R11, R14
+	XORL   R9, R14
+	XORL   R10, R14
+	ADDL   R14, R12
+	ADDL   R8, R12
+	ADDL   $0xca62c1d6, R12
+	ADDL   CX, R12
+	ROLL   $0x1e, R11
+	MOVL   32(SP), CX
+	XORL   12(SP), CX
+	XORL   52(SP), CX
+	XORL   44(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 44(SP)
+	MOVL   R12, R8
+	ROLL   $0x05, R8
+	MOVL   R13, R14
+	XORL   R11, R14
+	XORL   R9, R14
+	ADDL   R14, R8
+	ADDL   R10, R8
+	ADDL   $0xca62c1d6, R8
+	ADDL   CX, R8
+	ROLL   $0x1e, R13
+	MOVL   36(SP), CX
+	XORL   16(SP), CX
+	XORL   56(SP), CX
+	XORL   48(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 48(SP)
+	MOVL   R8, R10
+	ROLL   $0x05, R10
+	MOVL   R12, R14
+	XORL   R13, R14
+	XORL   R11, R14
+	ADDL   R14, R10
+	ADDL   R9, R10
+	ADDL   $0xca62c1d6, R10
+	ADDL   CX, R10
+	ROLL   $0x1e, R12
+	MOVL   40(SP), CX
+	XORL   20(SP), CX
+	XORL   60(SP), CX
+	XORL   52(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 52(SP)
+	MOVL   R10, R9
+	ROLL   $0x05, R9
+	MOVL   R8, R14
+	XORL   R12, R14
+	XORL   R13, R14
+	ADDL   R14, R9
+	ADDL   R11, R9
+	ADDL   $0xca62c1d6, R9
+	ADDL   CX, R9
+	ROLL   $0x1e, R8
+	MOVL   44(SP), CX
+	XORL   24(SP), CX
+	XORL   (SP), CX
+	XORL   56(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 56(SP)
+	MOVL   R9, R11
+	ROLL   $0x05, R11
+	MOVL   R10, R14
+	XORL   R8, R14
+	XORL   R12, R14
+	ADDL   R14, R11
+	ADDL   R13, R11
+	ADDL   $0xca62c1d6, R11
+	ADDL   CX, R11
+	ROLL   $0x1e, R10
+	MOVL   48(SP), CX
+	XORL   28(SP), CX
+	XORL   4(SP), CX
+	XORL   60(SP), CX
+	ROLL   $0x01, CX
+	MOVL   CX, 60(SP)
+	MOVL   R11, R13
+	ROLL   $0x05, R13
+	MOVL   R9, R14
+	XORL   R10, R14
+	XORL   R8, R14
+	ADDL   R14, R13
+	ADDL   R12, R13
+	ADDL   $0xca62c1d6, R13
+	ADDL   CX, R13
+	ROLL   $0x1e, R9
+	ADDL   R13, DX
+	ADDL   R11, BX
+	ADDL   R9, BP
+	ADDL   R10, SI
+	ADDL   R8, DI
+	MOVL   DX, (AX)
+	MOVL   BX, 4(AX)
+	MOVL   BP, 8(AX)
+	MOVL   SI, 12(AX)
+	MOVL   DI, 16(AX)
 	RET
diff --git a/examples/stadtx/stadtx.s b/examples/stadtx/stadtx.s
index 80ee9e7a..c8b542a4 100644
--- a/examples/stadtx/stadtx.s
+++ b/examples/stadtx/stadtx.s
@@ -4,307 +4,335 @@
 
 // func Hash(state *State, key []byte) uint64
 TEXT ·Hash(SB), NOSPLIT, $0-40
-	MOVQ	state(FP), AX
-	MOVQ	key_base+8(FP), CX
-	MOVQ	key_len+16(FP), DX
-	MOVQ	(AX), BX
-	MOVQ	8(AX), BP
-	MOVQ	DX, SI
-	ADDQ	$0x00000001, SI
-	MOVQ	$0xb89b0f8e1655514f, DI
-	IMULQ	DI, SI
-	XORQ	SI, BX
-	MOVQ	DX, SI
-	ADDQ	$0x00000002, SI
-	MOVQ	$0x8c6f736011bd5127, DI
-	IMULQ	DI, SI
-	XORQ	SI, BP
-	CMPQ	DX, $0x00000020
-	JGE	coreLong
-	MOVQ	DX, SI
-	SHRQ	$0x03, SI
-	CMPQ	SI, $0x00000000
-	JE	shortCore0
-	CMPQ	SI, $0x00000001
-	JE	shortCore1
-	CMPQ	SI, $0x00000002
-	JE	shortCore2
-	CMPQ	SI, $0x00000003
-	JE	shortCore3
+	MOVQ  state(FP), AX
+	MOVQ  key_base+8(FP), CX
+	MOVQ  key_len+16(FP), DX
+	MOVQ  (AX), BX
+	MOVQ  8(AX), BP
+	MOVQ  DX, SI
+	ADDQ  $0x00000001, SI
+	MOVQ  $0xb89b0f8e1655514f, DI
+	IMULQ DI, SI
+	XORQ  SI, BX
+	MOVQ  DX, SI
+	ADDQ  $0x00000002, SI
+	MOVQ  $0x8c6f736011bd5127, DI
+	IMULQ DI, SI
+	XORQ  SI, BP
+	CMPQ  DX, $0x00000020
+	JGE   coreLong
+	MOVQ  DX, SI
+	SHRQ  $0x03, SI
+	CMPQ  SI, $0x00000000
+	JE    shortCore0
+	CMPQ  SI, $0x00000001
+	JE    shortCore1
+	CMPQ  SI, $0x00000002
+	JE    shortCore2
+	CMPQ  SI, $0x00000003
+	JE    shortCore3
+
 shortCore3:
-	MOVQ	(CX), SI
-	MOVQ	$0x9c1b8e1e9628323f, DI
-	IMULQ	DI, SI
-	ADDQ	SI, BX
-	RORQ	$0x11, BX
-	XORQ	BP, BX
-	RORQ	$0x35, BP
-	ADDQ	BX, BP
-	ADDQ	$0x00000008, CX
-	SUBQ	$0x00000008, DX
+	MOVQ  (CX), SI
+	MOVQ  $0x9c1b8e1e9628323f, DI
+	IMULQ DI, SI
+	ADDQ  SI, BX
+	RORQ  $0x11, BX
+	XORQ  BP, BX
+	RORQ  $0x35, BP
+	ADDQ  BX, BP
+	ADDQ  $0x00000008, CX
+	SUBQ  $0x00000008, DX
+
 shortCore2:
-	MOVQ	(CX), SI
-	MOVQ	$0x9c1b8e1e9628323f, DI
-	IMULQ	DI, SI
-	ADDQ	SI, BX
-	RORQ	$0x11, BX
-	XORQ	BP, BX
-	RORQ	$0x35, BP
-	ADDQ	BX, BP
-	ADDQ	$0x00000008, CX
-	SUBQ	$0x00000008, DX
+	MOVQ  (CX), SI
+	MOVQ  $0x9c1b8e1e9628323f, DI
+	IMULQ DI, SI
+	ADDQ  SI, BX
+	RORQ  $0x11, BX
+	XORQ  BP, BX
+	RORQ  $0x35, BP
+	ADDQ  BX, BP
+	ADDQ  $0x00000008, CX
+	SUBQ  $0x00000008, DX
+
 shortCore1:
-	MOVQ	(CX), SI
-	MOVQ	$0x9c1b8e1e9628323f, DI
-	IMULQ	DI, SI
-	ADDQ	SI, BX
-	RORQ	$0x11, BX
-	XORQ	BP, BX
-	RORQ	$0x35, BP
-	ADDQ	BX, BP
-	ADDQ	$0x00000008, CX
-	SUBQ	$0x00000008, DX
+	MOVQ  (CX), SI
+	MOVQ  $0x9c1b8e1e9628323f, DI
+	IMULQ DI, SI
+	ADDQ  SI, BX
+	RORQ  $0x11, BX
+	XORQ  BP, BX
+	RORQ  $0x35, BP
+	ADDQ  BX, BP
+	ADDQ  $0x00000008, CX
+	SUBQ  $0x00000008, DX
+
 shortCore0:
-	CMPQ	DX, $0x00000000
-	JE	shortTail0
-	CMPQ	DX, $0x00000001
-	JE	shortTail1
-	CMPQ	DX, $0x00000002
-	JE	shortTail2
-	CMPQ	DX, $0x00000003
-	JE	shortTail3
-	CMPQ	DX, $0x00000004
-	JE	shortTail4
-	CMPQ	DX, $0x00000005
-	JE	shortTail5
-	CMPQ	DX, $0x00000006
-	JE	shortTail6
-	CMPQ	DX, $0x00000007
-	JE	shortTail7
+	CMPQ DX, $0x00000000
+	JE   shortTail0
+	CMPQ DX, $0x00000001
+	JE   shortTail1
+	CMPQ DX, $0x00000002
+	JE   shortTail2
+	CMPQ DX, $0x00000003
+	JE   shortTail3
+	CMPQ DX, $0x00000004
+	JE   shortTail4
+	CMPQ DX, $0x00000005
+	JE   shortTail5
+	CMPQ DX, $0x00000006
+	JE   shortTail6
+	CMPQ DX, $0x00000007
+	JE   shortTail7
+
 shortTail7:
-	MOVBQZX	6(CX), SI
-	SHLQ	$0x20, SI
-	ADDQ	SI, BX
+	MOVBQZX 6(CX), SI
+	SHLQ    $0x20, SI
+	ADDQ    SI, BX
+
 shortTail6:
-	MOVBQZX	5(CX), SI
-	SHLQ	$0x30, SI
-	ADDQ	SI, BP
+	MOVBQZX 5(CX), SI
+	SHLQ    $0x30, SI
+	ADDQ    SI, BP
+
 shortTail5:
-	MOVBQZX	4(CX), SI
-	SHLQ	$0x10, SI
-	ADDQ	SI, BX
+	MOVBQZX 4(CX), SI
+	SHLQ    $0x10, SI
+	ADDQ    SI, BX
+
 shortTail4:
-	MOVLQZX	(CX), SI
-	ADDQ	SI, BP
-	JMP	shortAfter
+	MOVLQZX (CX), SI
+	ADDQ    SI, BP
+	JMP     shortAfter
+
 shortTail3:
-	MOVBQZX	2(CX), SI
-	SHLQ	$0x30, SI
-	ADDQ	SI, BX
+	MOVBQZX 2(CX), SI
+	SHLQ    $0x30, SI
+	ADDQ    SI, BX
+
 shortTail2:
-	MOVWQZX	(CX), SI
-	ADDQ	SI, BP
-	JMP	shortAfter
+	MOVWQZX (CX), SI
+	ADDQ    SI, BP
+	JMP     shortAfter
+
 shortTail1:
-	MOVBQZX	(CX), SI
-	ADDQ	SI, BX
+	MOVBQZX (CX), SI
+	ADDQ    SI, BX
+
 shortTail0:
-	RORQ	$0x20, BP
-	XORQ	$0x000000ff, BP
+	RORQ $0x20, BP
+	XORQ $0x000000ff, BP
+
 shortAfter:
-	XORQ	BX, BP
-	RORQ	$0x21, BX
-	ADDQ	BP, BX
-	ROLQ	$0x11, BP
-	XORQ	BX, BP
-	ROLQ	$0x2b, BX
-	ADDQ	BP, BX
-	ROLQ	$0x1f, BP
-	SUBQ	BX, BP
-	ROLQ	$0x0d, BX
-	XORQ	BP, BX
-	SUBQ	BX, BP
-	ROLQ	$0x29, BX
-	ADDQ	BP, BX
-	ROLQ	$0x25, BP
-	XORQ	BX, BP
-	RORQ	$0x27, BX
-	ADDQ	BP, BX
-	RORQ	$0x0f, BP
-	ADDQ	BX, BP
-	ROLQ	$0x0f, BX
-	XORQ	BP, BX
-	RORQ	$0x05, BP
-	XORQ	BP, BX
-	MOVQ	BX, ret+32(FP)
+	XORQ BX, BP
+	RORQ $0x21, BX
+	ADDQ BP, BX
+	ROLQ $0x11, BP
+	XORQ BX, BP
+	ROLQ $0x2b, BX
+	ADDQ BP, BX
+	ROLQ $0x1f, BP
+	SUBQ BX, BP
+	ROLQ $0x0d, BX
+	XORQ BP, BX
+	SUBQ BX, BP
+	ROLQ $0x29, BX
+	ADDQ BP, BX
+	ROLQ $0x25, BP
+	XORQ BX, BP
+	RORQ $0x27, BX
+	ADDQ BP, BX
+	RORQ $0x0f, BP
+	ADDQ BX, BP
+	ROLQ $0x0f, BX
+	XORQ BP, BX
+	RORQ $0x05, BP
+	XORQ BP, BX
+	MOVQ BX, ret+32(FP)
 	RET
+
 coreLong:
-	MOVQ	16(AX), DI
-	MOVQ	24(AX), AX
-	MOVQ	DX, SI
-	ADDQ	$0x00000003, SI
-	MOVQ	$0x8f29bd94edce7b39, R8
-	IMULQ	R8, SI
-	XORQ	SI, DI
-	MOVQ	DX, SI
-	ADDQ	$0x00000004, SI
-	MOVQ	$0x9c1b8e1e9628323f, R8
-	IMULQ	R8, SI
-	XORQ	SI, AX
+	MOVQ  16(AX), DI
+	MOVQ  24(AX), AX
+	MOVQ  DX, SI
+	ADDQ  $0x00000003, SI
+	MOVQ  $0x8f29bd94edce7b39, R8
+	IMULQ R8, SI
+	XORQ  SI, DI
+	MOVQ  DX, SI
+	ADDQ  $0x00000004, SI
+	MOVQ  $0x9c1b8e1e9628323f, R8
+	IMULQ R8, SI
+	XORQ  SI, AX
+
 block:
-	MOVQ	(CX), SI
-	MOVQ	$0x00000000802910e3, R8
-	IMULQ	R8, SI
-	ADDQ	SI, BX
-	ROLQ	$0x39, BX
-	XORQ	AX, BX
-	MOVQ	8(CX), SI
-	MOVQ	$0x00000000819b13af, R8
-	IMULQ	R8, SI
-	ADDQ	SI, BP
-	ROLQ	$0x3f, BP
-	XORQ	DI, BP
-	MOVQ	16(CX), SI
-	MOVQ	$0x0000000091cb27e5, R8
-	IMULQ	R8, SI
-	ADDQ	SI, DI
-	RORQ	$0x2f, DI
-	ADDQ	BX, DI
-	MOVQ	24(CX), SI
-	MOVQ	$0x00000000c1a269c1, R8
-	IMULQ	R8, SI
-	ADDQ	SI, AX
-	RORQ	$0x0b, AX
-	SUBQ	BP, AX
-	ADDQ	$0x00000020, CX
-	SUBQ	$0x00000020, DX
-	CMPQ	DX, $0x00000020
-	JGE	block
-	MOVQ	DX, R8
-	MOVQ	DX, SI
-	SHRQ	$0x03, SI
-	CMPQ	SI, $0x00000000
-	JE	longCore0
-	CMPQ	SI, $0x00000001
-	JE	longCore1
-	CMPQ	SI, $0x00000002
-	JE	longCore2
-	CMPQ	SI, $0x00000003
-	JE	longCore3
+	MOVQ  (CX), SI
+	MOVQ  $0x00000000802910e3, R8
+	IMULQ R8, SI
+	ADDQ  SI, BX
+	ROLQ  $0x39, BX
+	XORQ  AX, BX
+	MOVQ  8(CX), SI
+	MOVQ  $0x00000000819b13af, R8
+	IMULQ R8, SI
+	ADDQ  SI, BP
+	ROLQ  $0x3f, BP
+	XORQ  DI, BP
+	MOVQ  16(CX), SI
+	MOVQ  $0x0000000091cb27e5, R8
+	IMULQ R8, SI
+	ADDQ  SI, DI
+	RORQ  $0x2f, DI
+	ADDQ  BX, DI
+	MOVQ  24(CX), SI
+	MOVQ  $0x00000000c1a269c1, R8
+	IMULQ R8, SI
+	ADDQ  SI, AX
+	RORQ  $0x0b, AX
+	SUBQ  BP, AX
+	ADDQ  $0x00000020, CX
+	SUBQ  $0x00000020, DX
+	CMPQ  DX, $0x00000020
+	JGE   block
+	MOVQ  DX, R8
+	MOVQ  DX, SI
+	SHRQ  $0x03, SI
+	CMPQ  SI, $0x00000000
+	JE    longCore0
+	CMPQ  SI, $0x00000001
+	JE    longCore1
+	CMPQ  SI, $0x00000002
+	JE    longCore2
+	CMPQ  SI, $0x00000003
+	JE    longCore3
+
 longCore3:
-	MOVQ	(CX), SI
-	MOVQ	$0x00000000802910e3, R9
-	IMULQ	R9, SI
-	ADDQ	SI, BX
-	ROLQ	$0x39, BX
-	XORQ	AX, BX
-	ADDQ	$0x00000008, CX
-	SUBQ	$0x00000008, DX
+	MOVQ  (CX), SI
+	MOVQ  $0x00000000802910e3, R9
+	IMULQ R9, SI
+	ADDQ  SI, BX
+	ROLQ  $0x39, BX
+	XORQ  AX, BX
+	ADDQ  $0x00000008, CX
+	SUBQ  $0x00000008, DX
+
 longCore2:
-	MOVQ	(CX), SI
-	MOVQ	$0x00000000819b13af, R9
-	IMULQ	R9, SI
-	ADDQ	SI, BP
-	ROLQ	$0x3f, BP
-	XORQ	DI, BP
-	ADDQ	$0x00000008, CX
-	SUBQ	$0x00000008, DX
+	MOVQ  (CX), SI
+	MOVQ  $0x00000000819b13af, R9
+	IMULQ R9, SI
+	ADDQ  SI, BP
+	ROLQ  $0x3f, BP
+	XORQ  DI, BP
+	ADDQ  $0x00000008, CX
+	SUBQ  $0x00000008, DX
+
 longCore1:
-	MOVQ	(CX), SI
-	MOVQ	$0x0000000091cb27e5, R9
-	IMULQ	R9, SI
-	ADDQ	SI, DI
-	RORQ	$0x2f, DI
-	ADDQ	BX, DI
-	ADDQ	$0x00000008, CX
-	SUBQ	$0x00000008, DX
+	MOVQ  (CX), SI
+	MOVQ  $0x0000000091cb27e5, R9
+	IMULQ R9, SI
+	ADDQ  SI, DI
+	RORQ  $0x2f, DI
+	ADDQ  BX, DI
+	ADDQ  $0x00000008, CX
+	SUBQ  $0x00000008, DX
+
 longCore0:
-	RORQ	$0x0b, AX
-	SUBQ	BP, AX
-	ADDQ	$0x00000001, R8
-	MOVQ	$0x9c1b8e1e9628323f, SI
-	IMULQ	SI, R8
-	XORQ	R8, BX
-	CMPQ	DX, $0x00000000
-	JE	longTail0
-	CMPQ	DX, $0x00000001
-	JE	longTail1
-	CMPQ	DX, $0x00000002
-	JE	longTail2
-	CMPQ	DX, $0x00000003
-	JE	longTail3
-	CMPQ	DX, $0x00000004
-	JE	longTail4
-	CMPQ	DX, $0x00000005
-	JE	longTail5
-	CMPQ	DX, $0x00000006
-	JE	longTail6
-	CMPQ	DX, $0x00000007
-	JE	longTail7
+	RORQ  $0x0b, AX
+	SUBQ  BP, AX
+	ADDQ  $0x00000001, R8
+	MOVQ  $0x9c1b8e1e9628323f, SI
+	IMULQ SI, R8
+	XORQ  R8, BX
+	CMPQ  DX, $0x00000000
+	JE    longTail0
+	CMPQ  DX, $0x00000001
+	JE    longTail1
+	CMPQ  DX, $0x00000002
+	JE    longTail2
+	CMPQ  DX, $0x00000003
+	JE    longTail3
+	CMPQ  DX, $0x00000004
+	JE    longTail4
+	CMPQ  DX, $0x00000005
+	JE    longTail5
+	CMPQ  DX, $0x00000006
+	JE    longTail6
+	CMPQ  DX, $0x00000007
+	JE    longTail7
+
 longTail7:
-	MOVBQZX	6(CX), SI
-	ADDQ	SI, BP
+	MOVBQZX 6(CX), SI
+	ADDQ    SI, BP
+
 longTail6:
-	MOVWQZX	4(CX), SI
-	ADDQ	SI, DI
-	MOVLQZX	(CX), SI
-	ADDQ	SI, AX
-	JMP	longAfter
+	MOVWQZX 4(CX), SI
+	ADDQ    SI, DI
+	MOVLQZX (CX), SI
+	ADDQ    SI, AX
+	JMP     longAfter
+
 longTail5:
-	MOVBQZX	4(CX), SI
-	ADDQ	SI, BP
+	MOVBQZX 4(CX), SI
+	ADDQ    SI, BP
+
 longTail4:
-	MOVLQZX	(CX), SI
-	ADDQ	SI, DI
-	JMP	longAfter
+	MOVLQZX (CX), SI
+	ADDQ    SI, DI
+	JMP     longAfter
+
 longTail3:
-	MOVBQZX	2(CX), SI
-	ADDQ	SI, AX
+	MOVBQZX 2(CX), SI
+	ADDQ    SI, AX
+
 longTail2:
-	MOVWQZX	(CX), SI
-	ADDQ	SI, BP
-	JMP	longAfter
+	MOVWQZX (CX), SI
+	ADDQ    SI, BP
+	JMP     longAfter
+
 longTail1:
-	MOVBQZX	(CX), SI
-	ADDQ	SI, DI
+	MOVBQZX (CX), SI
+	ADDQ    SI, DI
+
 longTail0:
-	ROLQ	$0x20, AX
-	XORQ	$0x000000ff, AX
+	ROLQ $0x20, AX
+	XORQ $0x000000ff, AX
+
 longAfter:
-	SUBQ	DI, BP
-	RORQ	$0x13, BX
-	SUBQ	BX, BP
-	RORQ	$0x35, BP
-	XORQ	BP, AX
-	SUBQ	AX, BX
-	ROLQ	$0x2b, AX
-	ADDQ	AX, BX
-	RORQ	$0x03, BX
-	SUBQ	BX, AX
-	RORQ	$0x2b, DI
-	SUBQ	AX, DI
-	ROLQ	$0x37, DI
-	XORQ	BX, DI
-	SUBQ	DI, BP
-	RORQ	$0x07, AX
-	SUBQ	DI, AX
-	RORQ	$0x1f, DI
-	ADDQ	DI, AX
-	SUBQ	BP, DI
-	RORQ	$0x27, AX
-	XORQ	AX, DI
-	RORQ	$0x11, AX
-	XORQ	DI, AX
-	ADDQ	AX, BP
-	RORQ	$0x09, BP
-	XORQ	BP, DI
-	ROLQ	$0x18, DI
-	XORQ	DI, AX
-	RORQ	$0x3b, AX
-	RORQ	$0x01, BX
-	SUBQ	BP, BX
-	XORQ	BP, BX
-	XORQ	AX, DI
-	XORQ	DI, BX
-	MOVQ	BX, ret+32(FP)
+	SUBQ DI, BP
+	RORQ $0x13, BX
+	SUBQ BX, BP
+	RORQ $0x35, BP
+	XORQ BP, AX
+	SUBQ AX, BX
+	ROLQ $0x2b, AX
+	ADDQ AX, BX
+	RORQ $0x03, BX
+	SUBQ BX, AX
+	RORQ $0x2b, DI
+	SUBQ AX, DI
+	ROLQ $0x37, DI
+	XORQ BX, DI
+	SUBQ DI, BP
+	RORQ $0x07, AX
+	SUBQ DI, AX
+	RORQ $0x1f, DI
+	ADDQ DI, AX
+	SUBQ BP, DI
+	RORQ $0x27, AX
+	XORQ AX, DI
+	RORQ $0x11, AX
+	XORQ DI, AX
+	ADDQ AX, BP
+	RORQ $0x09, BP
+	XORQ BP, DI
+	ROLQ $0x18, DI
+	XORQ DI, AX
+	RORQ $0x3b, AX
+	RORQ $0x01, BX
+	SUBQ BP, BX
+	XORQ BP, BX
+	XORQ AX, DI
+	XORQ DI, BX
+	MOVQ BX, ret+32(FP)
 	RET
diff --git a/examples/sum/sum.s b/examples/sum/sum.s
index 339231a5..35928d79 100644
--- a/examples/sum/sum.s
+++ b/examples/sum/sum.s
@@ -4,16 +4,18 @@
 
 // func Sum(xs []uint64) uint64
 TEXT ·Sum(SB), NOSPLIT, $0-32
-	MOVQ	xs_base(FP), AX
-	MOVQ	xs_len+8(FP), CX
-	XORQ	DX, DX
+	MOVQ xs_base(FP), AX
+	MOVQ xs_len+8(FP), CX
+	XORQ DX, DX
+
 loop:
-	CMPQ	CX, $0x00
-	JE	done
-	ADDQ	(AX), DX
-	ADDQ	$0x08, AX
-	DECQ	CX
-	JMP	loop
+	CMPQ CX, $0x00
+	JE   done
+	ADDQ (AX), DX
+	ADDQ $0x08, AX
+	DECQ CX
+	JMP  loop
+
 done:
-	MOVQ	DX, ret+24(FP)
+	MOVQ DX, ret+24(FP)
 	RET
diff --git a/internal/prnt/printer.go b/internal/prnt/printer.go
index 12cc6db9..2239946f 100644
--- a/internal/prnt/printer.go
+++ b/internal/prnt/printer.go
@@ -4,6 +4,7 @@ package prnt
 import (
 	"bytes"
 	"fmt"
+	"io"
 )
 
 // Generator provides convenience methods for code generators. In particular it
@@ -15,6 +16,11 @@ type Generator struct {
 	err error
 }
 
+// Raw provides direct access to the underlying output stream.
+func (g *Generator) Raw() io.Writer {
+	return &g.buf
+}
+
 // Printf prints to the internal buffer.
 func (g *Generator) Printf(format string, args ...interface{}) {
 	if g.err != nil {
diff --git a/printer/goasm.go b/printer/goasm.go
index b6321364..8d67162e 100644
--- a/printer/goasm.go
+++ b/printer/goasm.go
@@ -1,8 +1,10 @@
 package printer
 
 import (
+	"fmt"
 	"strconv"
 	"strings"
+	"text/tabwriter"
 
 	"github.com/mmcloughlin/avo/internal/prnt"
 	"github.com/mmcloughlin/avo/ir"
@@ -81,20 +83,29 @@ func (p *goasm) function(f *ir.Function) {
 	}
 	p.Printf(", %s\n", textsize(f))
 
+	w := p.tabwriter()
 	for _, node := range f.Nodes {
 		switch n := node.(type) {
 		case *ir.Instruction:
+			leader := []byte{tabwriter.Escape, '\t', tabwriter.Escape}
+			fmt.Fprint(w, string(leader)+n.Opcode)
 			if len(n.Operands) > 0 {
-				p.Printf("\t%s\t%s\n", n.Opcode, joinOperands(n.Operands))
-			} else {
-				p.Printf("\t%s\n", n.Opcode)
+				fmt.Fprintf(w, "\t%s", joinOperands(n.Operands))
 			}
+			fmt.Fprint(w, "\n")
 		case ir.Label:
-			p.Printf("%s:\n", n)
+			w.Flush()
+			w = p.tabwriter()
+			p.Printf("\n%s:\n", n)
 		default:
 			panic("unexpected node type")
 		}
 	}
+	w.Flush()
+}
+
+func (p *goasm) tabwriter() *tabwriter.Writer {
+	return tabwriter.NewWriter(p.Raw(), 4, 4, 1, ' ', tabwriter.StripEscape)
 }
 
 func (p *goasm) global(g *ir.Global) {
diff --git a/printer/goasm_test.go b/printer/goasm_test.go
index 408b6a35..f03c2860 100644
--- a/printer/goasm_test.go
+++ b/printer/goasm_test.go
@@ -24,10 +24,10 @@ func TestBasic(t *testing.T) {
 		"",
 		"// func add(x uint64, y uint64) uint64",
 		"TEXT ·add(SB), $0-24",
-		"\tMOVQ\tx(FP), AX",
-		"\tMOVQ\ty+8(FP), R9",
-		"\tADDQ\tAX, R9",
-		"\tMOVQ\tR9, ret+16(FP)",
+		"\tMOVQ x(FP), AX",
+		"\tMOVQ y+8(FP), R9",
+		"\tADDQ AX, R9",
+		"\tMOVQ R9, ret+16(FP)",
 		"\tRET",
 		"",
 	})
diff --git a/tests/cast/cast.s b/tests/cast/cast.s
index 7eb51a67..9af73536 100644
--- a/tests/cast/cast.s
+++ b/tests/cast/cast.s
@@ -4,9 +4,9 @@
 
 // func Split(x uint64) (q uint64, l uint32, w uint16, b uint8)
 TEXT ·Split(SB), NOSPLIT, $0-23
-	MOVQ	x(FP), AX
-	MOVQ	AX, q+8(FP)
-	MOVL	AX, l+16(FP)
-	MOVW	AX, w+20(FP)
-	MOVB	AL, b+22(FP)
+	MOVQ x(FP), AX
+	MOVQ AX, q+8(FP)
+	MOVL AX, l+16(FP)
+	MOVW AX, w+20(FP)
+	MOVB AL, b+22(FP)
 	RET