trsm_kernel_ln_4x4_penryn.s

Optimized GotoBLAS libraries

#endif

#ifdef LN
	movq	OFFSET, %rax
	addq	M, %rax
	movq	%rax, KK
#endif	

	movq	K, %rax
	salq	$BASE_SHIFT + 1, %rax
	leaq	(B, %rax), BB

#ifdef LT
	movq	OFFSET, %rax
	movq	%rax, KK
#endif
	
	testq	$1, M
	jle,pt	.L60
	ALIGN_4

#ifdef LN
       movq	K, %rax
       salq	$BASE_SHIFT, %rax
       subq	%rax, AORIG
#endif

#if defined(LN) || defined(RT)
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	movq	AORIG, AO
	leaq	(AO, %rax, 1), AO
	leaq	(B,  %rax, 2), BO
#else
	movq	B, BO
#endif	

	movsd	-16 * SIZE(AO), %xmm0
	movaps	-16 * SIZE(BO), %xmm2

	pxor	%xmm8,  %xmm8
	pxor	%xmm9,  %xmm9

#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	sarq	$2, %rax
	jle,pn	.L75
	ALIGN_4

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

	shufps	$0x44, %xmm0, %xmm0
	mulpd	%xmm0, %xmm2
	movsd	-15 * SIZE(AO), %xmm0
	addpd	%xmm2, %xmm8
	movaps	-14 * SIZE(BO), %xmm2

	shufps	$0x44, %xmm0, %xmm0
	mulpd	%xmm0, %xmm2
	movsd	-14 * SIZE(AO), %xmm0
	addpd	%xmm2, %xmm9
	movaps	-12 * SIZE(BO), %xmm2

	shufps	$0x44, %xmm0, %xmm0
	mulpd	%xmm0, %xmm2
	movsd	-13 * SIZE(AO), %xmm0
	addpd	%xmm2, %xmm8
	movaps	-10 * SIZE(BO), %xmm2

	shufps	$0x44, %xmm0, %xmm0
	mulpd	%xmm0, %xmm2
	movsd	-12 * SIZE(AO), %xmm0
	addpd	%xmm2, %xmm9
	movaps	 -8 * SIZE(BO), %xmm2

	subq	$-4 * SIZE, AO
	subq	$-8 * SIZE, BO
	subq	$1, %rax
	jg,pt	.L72
	ALIGN_4

.L75:
#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je,pt	.L78
	ALIGN_4

.L76:
	shufps	$0x44, %xmm0, %xmm0
	mulpd	%xmm0, %xmm2
	movsd	-15 * SIZE(AO), %xmm0
	addpd	%xmm2, %xmm8
	movaps	-14 * SIZE(BO), %xmm2

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

	subq	$1, %rax
	jg,pt	.L76
	ALIGN_4

.L78:
#if defined(LN) || defined(RT)
	movq	KK, %rax
#ifdef LN
	subq	$1, %rax
#else
	subq	$2, %rax
#endif

	leaq	(, %rax, SIZE), %rax

	movq	AORIG, AO
	leaq	(AO, %rax, 1), AO
	leaq	(B,  %rax, 2), BO
#endif

	addpd	%xmm9, %xmm8
	movhlps %xmm8, %xmm9

#if defined(LN) || defined(LT)
	movsd	-16 * SIZE(BO), %xmm12
	movsd	-15 * SIZE(BO), %xmm13
#else
	movsd	-16 * SIZE(AO), %xmm12
	movsd	-15 * SIZE(AO), %xmm13
#endif

	subsd	%xmm8,  %xmm12
	subsd	%xmm9,  %xmm13

#ifdef LN
	movsd	-16 * SIZE(AO), %xmm8

	mulsd	 %xmm8, %xmm12
	mulsd	 %xmm8, %xmm13
#endif

#ifdef LT
	movsd	-16 * SIZE(AO), %xmm8

	mulsd	 %xmm8, %xmm12
	mulsd	 %xmm8, %xmm13
#endif

#ifdef RN
	mulsd	-16 * SIZE(BO), %xmm12
	movsd	-15 * SIZE(BO), %xmm9
	mulsd	 %xmm12, %xmm9
	subsd	 %xmm9, %xmm13

	mulsd	-13 * SIZE(BO), %xmm13
#endif

#ifdef RT
	mulsd	-13 * SIZE(BO), %xmm13

	movlpd	-14 * SIZE(BO), %xmm9
	mulsd	 %xmm13, %xmm9
	subsd	 %xmm9, %xmm12

	mulsd	-16 * SIZE(BO), %xmm12
#endif

#ifdef LN
	subq	$1 * SIZE, CO1
	subq	$1 * SIZE, CO2
#endif

	movsd	%xmm12,  0 * SIZE(CO1)
	movsd	%xmm13,  0 * SIZE(CO2)

#if defined(LN) || defined(LT)
	movsd	%xmm12,  -16 * SIZE(BO)
	movsd	%xmm13,  -15 * SIZE(BO)
#else
	movsd	%xmm12,  -16 * SIZE(AO)
	movsd	%xmm13,  -15 * SIZE(AO)
#endif

#ifndef LN
	addq	$1 * SIZE, CO1
	addq	$1 * SIZE, CO2
#endif

#if defined(LT) || defined(RN)
	movq	K,  %rax
	subq	KK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 2), BO
#endif

#ifdef LN
	subq	$1, KK
#endif

#ifdef LT
	addq	$1, KK
#endif

#ifdef RT
       movq	K, %rax
       salq	$BASE_SHIFT, %rax
       addq	%rax, AORIG
#endif
	ALIGN_4	

.L60:
	testq	$2, M
	jle,pt	.L70
	ALIGN_4

#ifdef LN
       movq	K, %rax
       salq	$1 + BASE_SHIFT, %rax
       subq	%rax, AORIG
#endif

#if defined(LN) || defined(RT)
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	movq	AORIG, AO
	leaq	(AO, %rax, 2), AO
	leaq	(B,  %rax, 2), BO
#else
	movq	B, BO
#endif	

	movaps	-16 * SIZE(AO), %xmm0
	pxor	%xmm8,  %xmm8
	pxor	%xmm9,  %xmm9
	movaps	-16 * SIZE(BO), %xmm2
	pxor	%xmm10, %xmm10
	pxor	%xmm11, %xmm11

#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	sarq	$2, %rax
	jle,pn	.L65
	ALIGN_4

.L62:	
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)

	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	mulpd	%xmm0, %xmm7
	movaps	-14 * SIZE(AO), %xmm0

	addpd	%xmm2, %xmm9
	addpd	%xmm7, %xmm8
	movaps	-14 * SIZE(BO), %xmm2

	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	mulpd	%xmm0, %xmm7
	movaps	-12 * SIZE(AO), %xmm0

	addpd	%xmm2, %xmm11
	addpd	%xmm7, %xmm10
	movaps	-12 * SIZE(BO), %xmm2

	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	mulpd	%xmm0, %xmm7
	movaps	-10 * SIZE(AO), %xmm0

	addpd	%xmm2, %xmm9
	addpd	%xmm7, %xmm8
	movaps	-10 * SIZE(BO), %xmm2

	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	mulpd	%xmm0, %xmm7
	movaps	 -8 * SIZE(AO), %xmm0

	addpd	%xmm2, %xmm11
	addpd	%xmm7, %xmm10
	movaps	 -8 * SIZE(BO), %xmm2

	subq	$-8 * SIZE, AO
	subq	$-8 * SIZE, BO
	subq	$1, %rax
	jg,pt	.L62
	ALIGN_4

.L65:
#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je,pt	.L68
	ALIGN_4

.L66:
	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	mulpd	%xmm0, %xmm7
	movaps	-14 * SIZE(AO), %xmm0

	addpd	%xmm2, %xmm9
	addpd	%xmm7, %xmm8
	movaps	-14 * SIZE(BO), %xmm2

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

	subq	$1, %rax
	jg,pt	.L66
	ALIGN_4

.L68:
#if defined(LN) || defined(RT)
	movq	KK, %rax
#ifdef LN
	subq	$2, %rax
#else
	subq	$2, %rax
#endif

	leaq	(, %rax, SIZE), %rax

	movq	AORIG, AO
	leaq	(AO, %rax, 2), AO
	leaq	(B,  %rax, 2), BO
#endif

	addpd	%xmm10, %xmm8
	addpd	%xmm11, %xmm9

	movapd	 %xmm8, %xmm0
	movsd    %xmm9, %xmm8
	movsd    %xmm0, %xmm9


#if defined(LN) || defined(LT)
	movapd	%xmm8, %xmm0
	unpcklpd %xmm9, %xmm8
	unpckhpd %xmm9, %xmm0

	movapd	 -16 * SIZE(BO), %xmm9
	movapd	 -14 * SIZE(BO), %xmm13

	subpd	%xmm8,  %xmm9
	subpd	%xmm0,  %xmm13
#else
	movapd	-16 * SIZE(AO), %xmm0
	movapd	-14 * SIZE(AO), %xmm2

	subpd	%xmm8, %xmm0
	subpd	%xmm9, %xmm2
#endif

#ifdef LN
	movddup	-13 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm13

	movddup	-14 * SIZE(AO), %xmm10
	mulpd	 %xmm13, %xmm10
	subpd	 %xmm10, %xmm9

	movddup	-16 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm9
#endif

#ifdef LT
	movddup	-16 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm9

	movddup	-15 * SIZE(AO), %xmm10
	mulpd	 %xmm9, %xmm10
	subpd	 %xmm10, %xmm13

	movddup	-13 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm13
#endif

#ifdef RN
	movddup	-16 * SIZE(BO), %xmm8
	mulpd	 %xmm8, %xmm0

	movddup	-15 * SIZE(BO), %xmm9
	mulpd	 %xmm0, %xmm9
	subpd	 %xmm9, %xmm2

	movddup	-13 * SIZE(BO), %xmm8
	mulpd	 %xmm8, %xmm2
#endif

#ifdef RT
	movddup	-13 * SIZE(BO), %xmm8
	mulpd	 %xmm8, %xmm2

	movddup	-14 * SIZE(BO), %xmm9
	mulpd	 %xmm2, %xmm9
	subpd	 %xmm9, %xmm0

	movddup	-16 * SIZE(BO), %xmm8
	mulpd	 %xmm8, %xmm0
#endif

#ifdef LN
	subq	$2 * SIZE, CO1
	subq	$2 * SIZE, CO2
#endif

#if defined(LN) || defined(LT)
	movsd	%xmm9,  0 * SIZE(CO1)
	movsd	%xmm13, 1 * SIZE(CO1)

	movhpd	%xmm9,  0 * SIZE(CO2)
	movhpd	%xmm13, 1 * SIZE(CO2)
#else
	movsd	%xmm0,  0 * SIZE(CO1)
	movhpd	%xmm0,  1 * SIZE(CO1)

	movsd	%xmm2,  0 * SIZE(CO2)
	movhpd	%xmm2,  1 * SIZE(CO2)
#endif

#if defined(LN) || defined(LT)
	movapd	%xmm9,  -16 * SIZE(BO)
	movapd	%xmm13, -14 * SIZE(BO)
#else
	movapd	%xmm0,  -16 * SIZE(AO)
	movapd	%xmm2,  -14 * SIZE(AO)
#endif

#ifndef LN
	addq	$2 * SIZE, CO1
	addq	$2 * SIZE, CO2
#endif

#if defined(LT) || defined(RN)
	movq	K,  %rax
	subq	KK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(BO, %rax, 2), BO
#endif

#ifdef LN
	subq	$2, KK
#endif

#ifdef LT
	addq	$2, KK
#endif

#ifdef RT
       movq	K, %rax
       salq	$1 + BASE_SHIFT, %rax
       addq	%rax, AORIG
#endif
	ALIGN_4	

.L70:
	movq	M,  I
	sarq	$2, I	# i = (m >> 2)
	jle,pn	.L79
	ALIGN_4

.L51:
#ifdef LN
       movq	K, %rax
       salq	$2 + BASE_SHIFT, %rax
       subq	%rax, AORIG
#endif

#if defined(LN) || defined(RT)
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	movq	AORIG, AO
	leaq	(AO, %rax, 4), AO
	leaq	(B,  %rax, 2), BO
#else
	movq	B, BO
#endif	

	prefetcht2	 -16 * SIZE(BB)
	subq		 $-4 * SIZE, BB


	movaps	-16 * SIZE(AO), %xmm0
	movaps	-14 * SIZE(AO), %xmm1
	movaps	-16 * SIZE(BO), %xmm2

#ifdef LN
	prefetcht0    -4 * SIZE(CO1)
	pxor	%xmm8,  %xmm8
	pxor	%xmm9,  %xmm9
	prefetcht0    -4 * SIZE(CO2)
	pxor	%xmm12, %xmm12
	pxor	%xmm13, %xmm13
#else
	prefetcht0     3 * SIZE(CO1)
	pxor	%xmm8,  %xmm8
	pxor	%xmm9,  %xmm9
	prefetcht0     3 * SIZE(CO2)
	pxor	%xmm12, %xmm12
	pxor	%xmm13, %xmm13
#endif

#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	sarq	$2, %rax
	jle,pn	.L55
	ALIGN_4

.L52:	
	movaps	%xmm2, %xmm4
	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)
	mulpd	%xmm1, %xmm4

	movaps	%xmm7, %xmm6
	mulpd	%xmm0, %xmm7
	movaps	-12 * SIZE(AO), %xmm0
	mulpd	%xmm1, %xmm6
	movaps	-10 * SIZE(AO), %xmm1

	addpd	%xmm2, %xmm9
	movaps	-14 * SIZE(BO), %xmm2
	addpd	%xmm4, %xmm13
	addpd	%xmm7, %xmm8
	addpd	%xmm6, %xmm12

	movaps	%xmm2, %xmm4
	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	PREFETCH (PREFETCHSIZE +  8) * SIZE(AO)
	mulpd	%xmm1, %xmm4

	movaps	%xmm7, %xmm6
	mulpd	%xmm0, %xmm7
	movaps	 -8 * SIZE(AO), %xmm0
	mulpd	%xmm1, %xmm6
	movaps	 -6 * SIZE(AO), %xmm1

	addpd	%xmm2, %xmm9
	movaps	-12 * SIZE(BO), %xmm2
	addpd	%xmm4, %xmm13
	addpd	%xmm7, %xmm8
	addpd	%xmm6, %xmm12

	movaps	%xmm2, %xmm4
	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	mulpd	%xmm1, %xmm4

	movaps	%xmm7, %xmm6
	mulpd	%xmm0, %xmm7
	movaps	 -4 * SIZE(AO), %xmm0
	mulpd	%xmm1, %xmm6
	movaps	 -2 * SIZE(AO), %xmm1

	addpd	%xmm2, %xmm9
	movaps	-10 * SIZE(BO), %xmm2
	addpd	%xmm4, %xmm13
	addpd	%xmm7, %xmm8
	addpd	%xmm6, %xmm12

	movaps	%xmm2, %xmm4
	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	mulpd	%xmm1, %xmm4

	movaps	%xmm7, %xmm6
	mulpd	%xmm0, %xmm7
	movaps	  0 * SIZE(AO), %xmm0
	mulpd	%xmm1, %xmm6
	movaps	  2 * SIZE(AO), %xmm1

	addpd	%xmm2, %xmm9
	movaps	 -8 * SIZE(BO), %xmm2
	addpd	%xmm4, %xmm13
	addpd	%xmm7, %xmm8
	addpd	%xmm6, %xmm12

	subq	$-16 * SIZE, AO
	subq	$ -8 * SIZE, BO
	subq	$1, %rax
	jg,pt	.L52
	ALIGN_4

.L55:
#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je,pt	.L58
	ALIGN_4

.L56:
	movaps	%xmm2, %xmm4
	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	mulpd	%xmm1, %xmm4

	movaps	%xmm7, %xmm6
	mulpd	%xmm0, %xmm7
	movaps	-12 * SIZE(AO), %xmm0
	mulpd	%xmm1, %xmm6
	movaps	-10 * SIZE(AO), %xmm1

	addpd	%xmm2, %xmm9
	movaps	-14 * SIZE(BO), %xmm2
	addpd	%xmm4, %xmm13
	addpd	%xmm7, %xmm8
	addpd	%xmm6, %xmm12

	addq	$4 * SIZE, AO
	addq	$2 * SIZE, BO

	subq	$1, %rax
	jg,pt	.L56
	ALIGN_4

.L58:
#if defined(LN) || defined(RT)
	movq	KK, %rax
#ifdef LN
	subq	$4, %rax
#else
	subq	$2, %rax
#endif

	leaq	(, %rax, SIZE), %rax

	movq	AORIG, AO
	leaq	(AO, %rax, 4), AO
	leaq	(B,  %rax, 2), BO
#endif

	movapd	 %xmm8, %xmm0
	movsd    %xmm9, %xmm8
	movsd    %xmm0, %xmm9

	movapd	 %xmm12, %xmm0
	movsd    %xmm13, %xmm12
	movsd    %xmm0,  %xmm13

#if defined(LN) || defined(LT)
	movapd	%xmm8,  %xmm0
	unpcklpd %xmm9, %xmm8
	unpckhpd %xmm9, %xmm0

	movapd	%xmm12, %xmm4
	unpcklpd %xmm13, %xmm12
	unpckhpd %xmm13, %xmm4

	movapd	 -16 * SIZE(BO), %xmm9
	movapd	 -14 * SIZE(BO), %xmm13
	movapd	 -12 * SIZE(BO), %xmm1
	movapd	 -10 * SIZE(BO), %xmm5

	subpd	%xmm8,  %xmm9
	subpd	%xmm0,  %xmm13
	subpd	%xmm12, %xmm1
	subpd	%xmm4,  %xmm5
#else
	movapd	 -16 * SIZE(AO), %xmm0
	movapd	 -14 * SIZE(AO), %xmm1
	movapd	 -12 * SIZE(AO), %xmm2
	movapd	 -10 * SIZE(AO), %xmm3

	subpd	%xmm8,  %xmm0
	subpd	%xmm12, %xmm1
	subpd	%xmm9,  %xmm2
	subpd	%xmm13, %xmm3
#endif

#ifdef LN
	movddup	 -1 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm5
	movddup	 -2 * SIZE(AO), %xmm10
	mulpd	 %xmm5, %xmm10
	subpd	 %xmm10, %xmm1
	movddup	 -3 * SIZE(AO), %xmm12
	mulpd	 %xmm5, %xmm12
	subpd	 %xmm12, %xmm13
	movddup	 -4 * SIZE(AO), %xmm14
	mulpd	 %xmm5, %xmm14
	subpd	 %xmm14, %xmm9

	movddup	 -6 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm1
	movddup	 -7 * SIZE(AO), %xmm10
	mulpd	 %xmm1, %xmm10
	subpd	 %xmm10, %xmm13
	movddup	 -8 * SIZE(AO), %xmm12
	mulpd	 %xmm1, %xmm12
	subpd	 %xmm12, %xmm9

	movddup	-11 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm13
	movddup	-12 * SIZE(AO), %xmm10
	mulpd	 %xmm13, %xmm10
	subpd	 %xmm10, %xmm9

	movddup	-16 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm9
#endif

#ifdef LT
	movddup	-16 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm9
	movddup	-15 * SIZE(AO), %xmm10
	mulpd	 %xmm9, %xmm10
	subpd	 %xmm10, %xmm13
	movddup	-14 * SIZE(AO), %xmm12
	mulpd	 %xmm9, %xmm12
	subpd	 %xmm12, %xmm1
	movddup	-13 * SIZE(AO), %xmm14
	mulpd	 %xmm9, %xmm14
	subpd	 %xmm14, %xmm5


	movddup	-11 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm13

	movddup	-10 * SIZE(AO), %xmm10
	mulpd	 %xmm13, %xmm10
	subpd	 %xmm10, %xmm1
	movddup	 -9 * SIZE(AO), %xmm12
	mulpd	 %xmm13, %xmm12
	subpd	 %xmm12, %xmm5

	movddup	 -6 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm1
	movddup	 -5 * SIZE(AO), %xmm10
	mulpd	 %xmm1, %xmm10
	subpd	 %xmm10, %xmm5

	movddup	 -1 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm5
#endif

#ifdef RN
	movddup	-16 * SIZE(BO), %xmm8
	mulpd	 %xmm8, %xmm0
	mulpd	 %xmm8, %xmm1

	movddup	-15 * SIZE(BO), %xmm9
	movapd	%xmm9, %xmm10
	mulpd	 %xmm0, %xmm9
	subpd	 %xmm9, %xmm2
	mulpd	 %xmm1, %xmm10
	subpd	 %xmm10, %xmm3

	movddup	-13 * SIZE(BO), %xmm8
	mulpd	 %xmm8, %xmm2
	mulpd	 %xmm8, %xmm3
#endif

#ifdef RT
	movddup	-13 * SIZE(BO), %xmm8
	mulpd	 %xmm8, %xmm2
	mulpd	 %xmm8, %xmm3

	movddup	-14 * SIZE(BO), %xmm9
	movapd	%xmm9, %xmm10
	mulpd	 %xmm2, %xmm9
	subpd	 %xmm9, %xmm0
	mulpd	 %xmm3, %xmm10
	subpd	 %xmm10, %xmm1

	movddup	-16 * SIZE(BO), %xmm8
	mulpd	 %xmm8, %xmm0
	mulpd	 %xmm8, %xmm1
#endif

#ifdef LN
	subq	$4 * SIZE, CO1
	subq	$4 * SIZE, CO2
#endif

#if defined(LN) || defined(LT)
	movsd	%xmm9,  0 * SIZE(CO1)
	movsd	%xmm13, 1 * SIZE(CO1)
	movsd	%xmm1,  2 * SIZE(CO1)
	movsd	%xmm5,  3 * SIZE(CO1)

	movhpd	%xmm9,  0 * SIZE(CO2)
	movhpd	%xmm13, 1 * SIZE(CO2)
	movhpd	%xmm1,  2 * SIZE(CO2)
	movhpd	%xmm5,  3 * SIZE(CO2)
#else
	movsd	%xmm0,  0 * SIZE(CO1)
	movhpd	%xmm0,  1 * SIZE(CO1)
	movsd	%xmm1,  2 * SIZE(CO1)
	movhpd	%xmm1,  3 * SIZE(CO1)

	movsd	%xmm2,  0 * SIZE(CO2)
	movhpd	%xmm2,  1 * SIZE(CO2)
	movsd	%xmm3,  2 * SIZE(CO2)
	movhpd	%xmm3,  3 * SIZE(CO2)
#endif

#if defined(LN) || defined(LT)
	movapd	%xmm9,  -16 * SIZE(BO)
	movapd	%xmm13, -14 * SIZE(BO)
	movapd	%xmm1,  -12 * SIZE(BO)
	movapd	%xmm5,  -10 * SIZE(BO)
#else
	movapd	%xmm0,  -16 * SIZE(AO)
	movapd	%xmm1,  -14 * SIZE(AO)
	movapd	%xmm2,  -12 * SIZE(AO)
	movapd	%xmm3,  -10 * SIZE(AO)
#endif

#ifndef LN
	addq	$4 * SIZE, CO1
	addq	$4 * SIZE, CO2