gemm_kernel_8x4_penryn.s

Optimized GotoBLAS libraries

	pxor	%xmm9, %xmm9

#ifndef TRMMKERNEL
	movq	K, %rax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movq	K, %rax
	subq	KK, %rax
	movq	%rax, KKK	
#else
	movq	KK, %rax
#ifdef LEFT
	addq	$2, %rax
#else
	addq	$2, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$2, %rax
	jle,pn	.L75
	ALIGN_3

.L72:	
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)

	addps	%xmm3, %xmm8
	pshufd	$0x44, %xmm0, %xmm1
	pshufd	$0x50, %xmm2, %xmm3
	mulps	%xmm1, %xmm3

	addps	%xmm3, %xmm9
	pshufd	$0xee, %xmm0, %xmm1
	movaps	-28 * SIZE(AO), %xmm0
	pshufd	$0xfa, %xmm2, %xmm3
	movaps	-28 * SIZE(BO), %xmm2
	mulps	%xmm1, %xmm3

	addps	%xmm3, %xmm8
	pshufd	$0x44, %xmm0, %xmm1
	pshufd	$0x50, %xmm2, %xmm3
	mulps	%xmm1, %xmm3

	addps	%xmm3, %xmm9
	pshufd	$0xee, %xmm0, %xmm1
	movaps	-24 * SIZE(AO), %xmm0
	pshufd	$0xfa, %xmm2, %xmm3
	movaps	-24 * SIZE(BO), %xmm2
	mulps	%xmm1, %xmm3

	subq	$-8 * SIZE, AO
	subq	$-8 * SIZE, BO

	subq	$1, %rax
	jg,pt	.L72
	ALIGN_3

.L75:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax
	je,pt	.L78
	ALIGN_3

.L76:
	addps	%xmm3, %xmm8
	pshufd	$0x44, %xmm0, %xmm1
	movsd	-30 * SIZE(AO), %xmm0
	pshufd	$0x50, %xmm2, %xmm3
	movsd	-30 * SIZE(BO), %xmm2
	mulps	%xmm1, %xmm3

	addq	$2 * SIZE, AO
	addq	$2 * SIZE, BO

	subq	$1, %rax
	jg,pt	.L76
	ALIGN_3

.L78:
	movddup	ALPHA, %xmm2

	addps	%xmm9, %xmm8
	addps	%xmm3, %xmm8

	mulps	%xmm2, %xmm8

#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movhps	0 * SIZE(CO2), %xmm0

	addps	%xmm0, %xmm8
#endif

	movsd	%xmm8,  0 * SIZE(CO1)
	movhps	%xmm8,  0 * SIZE(CO2)

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movq	K, %rax
	subq	KKK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(BO, %rax, 2), BO
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$2, KK
#endif

	addq	$2 * SIZE, CO1
	addq	$2 * SIZE, CO2
	ALIGN_4	

.L80:
	testq	$1, M
	jle,pt	.L89

#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	movq	B, BO
#else
	movq	B, BO

	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 2), BO
#endif	

	movsd	-32 * SIZE(AO), %xmm0
	pxor	%xmm8, %xmm8
	movsd	-32 * SIZE(BO), %xmm2
	pxor	%xmm9, %xmm9

#ifndef TRMMKERNEL
	movq	K, %rax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movq	K, %rax
	subq	KK, %rax
	movq	%rax, KKK	
#else
	movq	KK, %rax
#ifdef LEFT
	addq	$1, %rax
#else
	addq	$2, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$2, %rax
	jle,pn	.L85
	ALIGN_3

.L82:	
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)

	pshufd	$0x00, %xmm0, %xmm1
	movss	-31 * SIZE(AO), %xmm0
	mulps	%xmm1, %xmm2
	addps	%xmm2, %xmm8
	movsd	-30 * SIZE(BO), %xmm2

	pshufd	$0x00, %xmm0, %xmm1
	movss	-30 * SIZE(AO), %xmm0
	mulps	%xmm1, %xmm2
	addps	%xmm2, %xmm9
	movsd	-28 * SIZE(BO), %xmm2

	pshufd	$0x00, %xmm0, %xmm1
	movss	-29 * SIZE(AO), %xmm0
	mulps	%xmm1, %xmm2
	addps	%xmm2, %xmm8
	movsd	-26 * SIZE(BO), %xmm2

	pshufd	$0x00, %xmm0, %xmm1
	movss	-28 * SIZE(AO), %xmm0
	mulps	%xmm1, %xmm2
	addps	%xmm2, %xmm9
	movsd	-24 * SIZE(BO), %xmm2

	subq	$-4 * SIZE, AO
	subq	$-8 * SIZE, BO

	subq	$1, %rax
	jg,pt	.L82
	ALIGN_3

.L85:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax
	je,pt	.L88
	ALIGN_3

.L86:
	pshufd	$0x00, %xmm0, %xmm1
	movss	-31 * SIZE(AO), %xmm0
	mulps	%xmm1, %xmm2
	addps	%xmm2, %xmm8
	movsd	-30 * SIZE(BO), %xmm2

	addq	$1 * SIZE, AO
	addq	$2 * SIZE, BO

	subq	$1, %rax
	jg,pt	.L86
	ALIGN_3

.L88:
	movddup	ALPHA, %xmm2
	addps	%xmm9, %xmm8
	mulps	%xmm2, %xmm8

	pshufd	$0x55, %xmm8, %xmm9
	pshufd	$0x00, %xmm8, %xmm8

#ifndef TRMMKERNEL
	addss	0 * SIZE(CO1), %xmm8
	addss	0 * SIZE(CO2), %xmm9
#endif

	movss	%xmm8,  0 * SIZE(CO1)
	movss	%xmm9,  0 * SIZE(CO2)

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movq	K, %rax
	subq	KKK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 2), BO
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$1, KK
#endif
	ALIGN_4

.L89:
#if defined(TRMMKERNEL) && !defined(LEFT)
	addq	$2, KK
#endif

	movq	BO, B

	leaq	(C, LDC, 2), C
	ALIGN_4

.L90:
	testq	$1, N
	jle	.L999

#if defined(TRMMKERNEL) && defined(LEFT)
        movq    OFFSET, %rax
	movq    %rax, KK
#endif

	movq	C, CO1
	movq	A, AO

	movq	M,  I
	sarq	$3, I
	jle,pn	.L100
	ALIGN_4

.L91:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	movq	B, BO
#else
	movq	B, BO

	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 8), AO
	leaq	(BO, %rax, 1), BO
#endif	

	movaps	-32 * SIZE(AO), %xmm0
	pxor	%xmm8,  %xmm8
	movaps	-28 * SIZE(AO), %xmm1
	pxor	%xmm9,  %xmm9
	movsd	-32 * SIZE(BO), %xmm2
	pxor	%xmm10, %xmm10
	prefetcht0     7 * SIZE(CO1)
	pxor	%xmm11, %xmm11

#ifndef TRMMKERNEL
	movq	K, %rax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movq	K, %rax
	subq	KK, %rax
	movq	%rax, KKK	
#else
	movq	KK, %rax
#ifdef LEFT
	addq	$8, %rax
#else
	addq	$1, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$2, %rax
	jle,pn	.L95
	ALIGN_3

.L92:	
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)

	pshufd	$0x00, %xmm2, %xmm3
	mulps	%xmm3, %xmm0
	addps	%xmm0, %xmm8
	movaps	-24 * SIZE(AO), %xmm0
	mulps	%xmm3, %xmm1
	addps	%xmm1, %xmm9
	movaps	-20 * SIZE(AO), %xmm1

	pshufd	$0x55, %xmm2, %xmm3
	movsd	-30 * SIZE(BO), %xmm2
	mulps	%xmm3, %xmm0
	addps	%xmm0, %xmm10
	movaps	-16 * SIZE(AO), %xmm0
	mulps	%xmm3, %xmm1
	addps	%xmm1, %xmm11
	movaps	-12 * SIZE(AO), %xmm1

	PREFETCH (PREFETCHSIZE + 16) * SIZE(AO)

	pshufd	$0x00, %xmm2, %xmm3
	mulps	%xmm3, %xmm0
	addps	%xmm0, %xmm8
	movaps	 -8 * SIZE(AO), %xmm0
	mulps	%xmm3, %xmm1
	addps	%xmm1, %xmm9
	movaps	 -4 * SIZE(AO), %xmm1

	pshufd	$0x55, %xmm2, %xmm3
	movsd	-28 * SIZE(BO), %xmm2
	mulps	%xmm3, %xmm0
	addps	%xmm0, %xmm10
	movaps	  0 * SIZE(AO), %xmm0
	mulps	%xmm3, %xmm1
	addps	%xmm1, %xmm11
	movaps	  4 * SIZE(AO), %xmm1

	subq	$-32 * SIZE, AO
	subq	$ -4 * SIZE, BO

	subq	$1, %rax
	jg,pt	.L92
	ALIGN_3

.L95:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je,pt	.L98
	ALIGN_3

.L96:
	pshufd	$0x00, %xmm2, %xmm3
	movss	-31 * SIZE(BO), %xmm2
	mulps	%xmm3, %xmm0
	addps	%xmm0, %xmm8
	movaps	-24 * SIZE(AO), %xmm0
	mulps	%xmm3, %xmm1
	addps	%xmm1, %xmm9
	movaps	-20 * SIZE(AO), %xmm1

	addq	$8 * SIZE, AO
	addq	$1 * SIZE, BO

	subq	$1, %rax
	jg,pt	.L96
	ALIGN_3

.L98:
	movddup	ALPHA, %xmm7

	addps	%xmm10, %xmm8
	addps	%xmm11, %xmm9

	mulps	%xmm7, %xmm8
	mulps	%xmm7, %xmm9

#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movhps	2 * SIZE(CO1), %xmm0
	movsd	4 * SIZE(CO1), %xmm1
	movhps	6 * SIZE(CO1), %xmm1

	addps	%xmm0, %xmm8
	addps	%xmm1, %xmm9
#endif

	movsd	%xmm8,  0 * SIZE(CO1)
	movhps	%xmm8,  2 * SIZE(CO1)
	movsd	%xmm9,  4 * SIZE(CO1)
	movhps	%xmm9,  6 * SIZE(CO1)

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movq	K, %rax
	subq	KKK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 8), AO
	leaq	(BO, %rax, 1), BO
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$8, KK
#endif

	addq	$8 * SIZE, CO1
	decq	I
	jg,pt	.L91
	ALIGN_4	

.L100:
	testq	$4, M
	jle,pt	.L110

#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	movq	B, BO
#else
	movq	B, BO

	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 1), BO
#endif	

	movaps	-32 * SIZE(AO), %xmm0
	pxor	%xmm8, %xmm8
	movsd	-32 * SIZE(BO), %xmm2
	pxor	%xmm9, %xmm9

#ifndef TRMMKERNEL
	movq	K, %rax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movq	K, %rax
	subq	KK, %rax
	movq	%rax, KKK	
#else
	movq	KK, %rax
#ifdef LEFT
	addq	$4, %rax
#else
	addq	$1, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$2, %rax
	jle,pn	.L105
	ALIGN_3

.L102:	
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)

	pshufd	$0x00, %xmm2, %xmm3
	mulps	%xmm0, %xmm3
	movaps	-28 * SIZE(AO), %xmm0
	addps	%xmm3, %xmm8

	pshufd	$0x55, %xmm2, %xmm3
	movsd	-30 * SIZE(BO), %xmm2
	mulps	%xmm0, %xmm3
	movaps	-24 * SIZE(AO), %xmm0
	addps	%xmm3, %xmm9

	pshufd	$0x00, %xmm2, %xmm3
	mulps	%xmm0, %xmm3
	movaps	-20 * SIZE(AO), %xmm0
	addps	%xmm3, %xmm8

	pshufd	$0x55, %xmm2, %xmm3
	movsd	-28 * SIZE(BO), %xmm2
	mulps	%xmm0, %xmm3
	movaps	-16 * SIZE(AO), %xmm0
	addps	%xmm3, %xmm9

	subq	$-16 * SIZE, AO
	subq	$ -4 * SIZE, BO

	subq	$1, %rax
	jg,pt	.L102
	ALIGN_3

.L105:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax
	je,pt	.L108
	ALIGN_3

.L106:
	pshufd	$0x00, %xmm2, %xmm3
	movss	-31 * SIZE(BO), %xmm2
	mulps	%xmm0, %xmm3
	movaps	-28 * SIZE(AO), %xmm0
	addps	%xmm3, %xmm8

	addq	$4 * SIZE, AO
	addq	$1 * SIZE, BO

	subq	$1, %rax
	jg,pt	.L106
	ALIGN_3

.L108:
	movddup	ALPHA, %xmm7

	addps	%xmm9, %xmm8
	mulps	%xmm7, %xmm8

#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movhps	2 * SIZE(CO1), %xmm0

	addps	%xmm0, %xmm8
#endif

	movsd	%xmm8,  0 * SIZE(CO1)
	movhps	%xmm8,  2 * SIZE(CO1)

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movq	K, %rax
	subq	KKK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 1), BO
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$4, KK
#endif

	addq	$4 * SIZE, CO1
	ALIGN_4	

.L110:
	testq	$2, M
	jle,pt	.L120

#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	movq	B, BO
#else
	movq	B, BO

	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(BO, %rax, 1), BO
#endif	

	movaps	-32 * SIZE(AO), %xmm0
	pxor	%xmm3, %xmm3
	movsd	-32 * SIZE(BO), %xmm2
	pxor	%xmm8, %xmm8
	pxor	%xmm9, %xmm9

#ifndef TRMMKERNEL
	movq	K, %rax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movq	K, %rax
	subq	KK, %rax
	movq	%rax, KKK	
#else
	movq	KK, %rax
#ifdef LEFT
	addq	$2, %rax
#else
	addq	$1, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$2, %rax
	jle,pn	.L115
	ALIGN_3

.L112:	
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)

	pshufd	$0x00, %xmm2, %xmm3
	mulps	%xmm0, %xmm3
	movsd	-30 * SIZE(AO), %xmm0
	addps	%xmm3, %xmm8

	pshufd	$0x55, %xmm2, %xmm3
	movsd	-30 * SIZE(BO), %xmm2
	mulps	%xmm0, %xmm3
	movsd	-28 * SIZE(AO), %xmm0
	addps	%xmm3, %xmm8

	pshufd	$0x00, %xmm2, %xmm3
	mulps	%xmm0, %xmm3
	movsd	-26 * SIZE(AO), %xmm0
	addps	%xmm3, %xmm8

	pshufd	$0x55, %xmm2, %xmm3
	movsd	-28 * SIZE(BO), %xmm2
	mulps	%xmm0, %xmm3
	movsd	-24 * SIZE(AO), %xmm0
	addps	%xmm3, %xmm8

	subq	$-8 * SIZE, AO
	subq	$-4 * SIZE, BO

	subq	$1, %rax
	jg,pt	.L112
	ALIGN_3

.L115:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax
	je,pt	.L118
	ALIGN_3

.L116:
	pshufd	$0x00, %xmm2, %xmm3
	movss	-31 * SIZE(BO), %xmm2
	mulps	%xmm0, %xmm3
	movsd	-30 * SIZE(AO), %xmm0
	addps	%xmm3, %xmm8

	addq	$2 * SIZE, AO
	addq	$1 * SIZE, BO

	subq	$1, %rax
	jg,pt	.L116
	ALIGN_3

.L118:
	movddup	ALPHA, %xmm2

	mulps	%xmm2, %xmm8

#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0

	addps	%xmm0, %xmm8
#endif

	movsd	%xmm8,  0 * SIZE(CO1)

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movq	K, %rax
	subq	KKK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(BO, %rax, 1), BO
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$2, KK
#endif

	addq	$2 * SIZE, CO1
	ALIGN_4	

.L120:
	testq	$1, M
	jle,pt	.L999

#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	movq	B, BO
#else
	movq	B, BO

	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 1), BO
#endif	

	movss	-32 * SIZE(AO), %xmm0
	pxor	%xmm8, %xmm8
	movss	-32 * SIZE(BO), %xmm2
	pxor	%xmm9, %xmm9

#ifndef TRMMKERNEL
	movq	K, %rax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movq	K, %rax
	subq	KK, %rax
	movq	%rax, KKK	
#else
	movq	KK, %rax
#ifdef LEFT
	addq	$1, %rax
#else
	addq	$1, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$2, %rax
	jle,pn	.L125
	ALIGN_3

.L122:	
	mulss	%xmm0, %xmm2
	movss	-31 * SIZE(AO), %xmm0
	addss	%xmm2, %xmm8
	movss	-31 * SIZE(BO), %xmm2

	mulss	%xmm0, %xmm2
	movss	-30 * SIZE(AO), %xmm0
	addss	%xmm2, %xmm9
	movss	-30 * SIZE(BO), %xmm2

	mulss	%xmm0, %xmm2
	movss	-29 * SIZE(AO), %xmm0
	addss	%xmm2, %xmm8
	movss	-29 * SIZE(BO), %xmm2

	mulss	%xmm0, %xmm2
	movss	-28 * SIZE(AO), %xmm0
	addss	%xmm2, %xmm9
	movss	-28 * SIZE(BO), %xmm2

	subq	$-4 * SIZE, AO
	subq	$-4 * SIZE, BO

	subq	$1, %rax
	jg,pt	.L122
	ALIGN_3

.L125:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax
	je,pt	.L128
	ALIGN_3

.L126:
	mulss	%xmm0, %xmm2
	movss	-31 * SIZE(AO), %xmm0
	addss	%xmm2, %xmm8
	movss	-31 * SIZE(BO), %xmm2

	addq	$1 * SIZE, AO
	addq	$1 * SIZE, BO

	subq	$1, %rax
	jg,pt	.L126
	ALIGN_3

.L128:
	movss	ALPHA, %xmm2
	addss	%xmm9, %xmm8
	mulss	%xmm2, %xmm8

#ifndef TRMMKERNEL
	addss	0 * SIZE(CO1), %xmm8
#endif

	movss	%xmm8,  0 * SIZE(CO1)
	ALIGN_4

.L999:
	movq	  0(%rsp), %rbx
	movq	  8(%rsp), %rbp
	movq	 16(%rsp), %r12
	movq	 24(%rsp), %r13
	movq	 32(%rsp), %r14
	movq	 40(%rsp), %r15

#ifdef WINDOWS_ABI
	movq	 48(%rsp), %rdi
	movq	 56(%rsp), %rsi
	movups	 64(%rsp), %xmm6
	movups	 80(%rsp), %xmm7
	movups	 96(%rsp), %xmm8
	movups	112(%rsp), %xmm9
	movups	128(%rsp), %xmm10
	movups	144(%rsp), %xmm11
	movups	160(%rsp), %xmm12
	movups	176(%rsp), %xmm13
	movups	192(%rsp), %xmm14
	movups	208(%rsp), %xmm15
#endif

	addq	$STACKSIZE, %rsp
	ret

	EPILOGUE