📄 trsm_kernel_lt_8x2_sse.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 4 页
字号:
	shufps	$0x00, %xmm6, %xmm6		mulps	%xmm6, %xmm2	mulps	%xmm6, %xmm3	movss	 2 * SIZE(B), %xmm6	shufps	$0x00, %xmm6, %xmm6	movaps	%xmm6, %xmm5	mulps	%xmm2, %xmm5	mulps	%xmm3, %xmm6	subps	%xmm5, %xmm0	subps	%xmm6, %xmm1	movss	 0 * SIZE(B), %xmm6	shufps	$0x00, %xmm6, %xmm6		mulps	%xmm6, %xmm0	mulps	%xmm6, %xmm1#endif#if defined(LN) || defined(LT)	movaps	%xmm2,   0 * SIZE(B)	movaps	%xmm3,   4 * SIZE(B)	movaps	%xmm5,   8 * SIZE(B)	movaps	%xmm7,  12 * SIZE(B)	pshufd	$0x00, %xmm2, %xmm0	pshufd	$0x55, %xmm2, %xmm1	pshufd	$0xaa, %xmm2, %xmm4	pshufd	$0xff, %xmm2, %xmm6	movaps	%xmm0,   0 * SIZE(BB)	movaps	%xmm1,   4 * SIZE(BB)	movaps	%xmm4,   8 * SIZE(BB)	movaps	%xmm6,  12 * SIZE(BB)	pshufd	$0x00, %xmm3, %xmm0	pshufd	$0x55, %xmm3, %xmm1	pshufd	$0xaa, %xmm3, %xmm4	pshufd	$0xff, %xmm3, %xmm6	movaps	%xmm0,  16 * SIZE(BB)	movaps	%xmm1,  20 * SIZE(BB)	movaps	%xmm4,  24 * SIZE(BB)	movaps	%xmm6,  28 * SIZE(BB)	pshufd	$0x00, %xmm5, %xmm0	pshufd	$0x55, %xmm5, %xmm1	pshufd	$0xaa, %xmm5, %xmm4	pshufd	$0xff, %xmm5, %xmm6	movaps	%xmm0,  32 * SIZE(BB)	movaps	%xmm1,  36 * SIZE(BB)	movaps	%xmm4,  40 * SIZE(BB)	movaps	%xmm6,  44 * SIZE(BB)	pshufd	$0x00, %xmm7, %xmm0	pshufd	$0x55, %xmm7, %xmm1	pshufd	$0xaa, %xmm7, %xmm4	pshufd	$0xff, %xmm7, %xmm6	movaps	%xmm0,  48 * SIZE(BB)	movaps	%xmm1,  52 * SIZE(BB)	movaps	%xmm4,  56 * SIZE(BB)	movaps	%xmm6,  60 * SIZE(BB)#else	movaps	%xmm0,   0 * SIZE(AA)	movaps	%xmm1,   4 * SIZE(AA)	movaps	%xmm2,   8 * SIZE(AA)	movaps	%xmm3,  12 * SIZE(AA)#endif#ifdef LN	subl	$8 * SIZE, CO1#endif#if defined(LN) || defined(LT)	movaps	 %xmm2, %xmm0	shufps	 $0x88, %xmm3, %xmm2	shufps	 $0xdd, %xmm3, %xmm0	movaps	 %xmm5, %xmm4	shufps	 $0x88, %xmm7, %xmm5	shufps	 $0xdd, %xmm7, %xmm4	movlps	%xmm2, 0 * SIZE(CO1)	movhps	%xmm2, 2 * SIZE(CO1)	movlps	%xmm5, 4 * SIZE(CO1)	movhps	%xmm5, 6 * SIZE(CO1)	movlps	%xmm0, 0 * SIZE(CO1, LDC)	movhps	%xmm0, 2 * SIZE(CO1, LDC)	movlps	%xmm4, 4 * SIZE(CO1, LDC)	movhps	%xmm4, 6 * SIZE(CO1, LDC)#else	movlps	%xmm0, 0 * SIZE(CO1)	movhps	%xmm0, 2 * SIZE(CO1)	movlps	%xmm1, 4 * SIZE(CO1)	movhps	%xmm1, 6 * SIZE(CO1)	movlps	%xmm2, 0 * SIZE(CO1, LDC)	movhps	%xmm2, 2 * SIZE(CO1, LDC)	movlps	%xmm3, 4 * SIZE(CO1, LDC)	movhps	%xmm3, 6 * SIZE(CO1, LDC)#endif#ifndef LN	addl	$8 * SIZE, CO1#endif#if defined(LT) || defined(RN)	movl	K,  %eax	subl	KK, %eax	leal	(,%eax, SIZE), %eax	leal	(AA, %eax, 8), AA#ifdef LT	addl	$16 * SIZE, B#endif#endif#ifdef LN	subl	$8, KK	movl	BORIG, B#endif#ifdef LT	addl	$8, KK#endif#ifdef RT	movl	K, %eax	movl	BORIG, B	sall	$3 + BASE_SHIFT, %eax	addl	%eax, AORIG#endif	decl	%ebx			# i --	jg	.L10	ALIGN_2	.L30:	testl	$4, M	jle	.L50#ifdef LN       movl	K, %eax       sall	$2 + BASE_SHIFT, %eax       subl	%eax, AORIG#endif#if defined(LN) || defined(RT)	movl	KK, %eax	movl	AORIG, AA	sall	$2 + BASE_SHIFT, %eax	addl	%eax, AA#endif	leal	BUFFER, BB#if defined(LN) || defined(RT)	movl	KK, %eax	sall	$1 + BASE_SHIFT, %eax	leal	(BB, %eax, 4), BB#endif		movaps	 0 * SIZE(BB), %xmm2	xorps	%xmm4, %xmm4	movaps	 0 * SIZE(AA), %xmm0	xorps	%xmm5, %xmm5	movaps	 16 * SIZE(BB), %xmm3	xorps	%xmm6, %xmm6	movaps	 16 * SIZE(AA), %xmm1	xorps	%xmm7, %xmm7#if defined(LT) || defined(RN)	movl	KK, %eax#else	movl	K, %eax	subl	KK, %eax#endif	sarl	$3, %eax	je	.L32	ALIGN_2.L31:	mulps	%xmm0, %xmm2	mulps	 4 * SIZE(BB), %xmm0	addps	%xmm2, %xmm4	movaps	 8 * SIZE(BB), %xmm2	addps	%xmm0, %xmm5	movaps	 4 * SIZE(AA), %xmm0	mulps	%xmm0, %xmm2	mulps	12 * SIZE(BB), %xmm0	addps	%xmm2, %xmm6	movaps	32 * SIZE(BB), %xmm2	addps	%xmm0, %xmm7	movaps	 8 * SIZE(AA), %xmm0	mulps	%xmm0, %xmm3	mulps	20 * SIZE(BB), %xmm0	addps	%xmm3, %xmm4	movaps	24 * SIZE(BB), %xmm3	addps	%xmm0, %xmm5	movaps	12 * SIZE(AA), %xmm0	mulps	%xmm0, %xmm3	mulps	28 * SIZE(BB), %xmm0	addps	%xmm3, %xmm6	movaps	48 * SIZE(BB), %xmm3	addps	%xmm0, %xmm7	movaps	32 * SIZE(AA), %xmm0	mulps	%xmm1, %xmm2	mulps	36 * SIZE(BB), %xmm1	addps	%xmm2, %xmm4	movaps	40 * SIZE(BB), %xmm2	addps	%xmm1, %xmm5	movaps	20 * SIZE(AA), %xmm1	mulps	%xmm1, %xmm2	mulps	44 * SIZE(BB), %xmm1	addps	%xmm2, %xmm6	movaps	64 * SIZE(BB), %xmm2	addps	%xmm1, %xmm7	movaps	24 * SIZE(AA), %xmm1	mulps	%xmm1, %xmm3	mulps	52 * SIZE(BB), %xmm1	addps	%xmm3, %xmm4	movaps	56 * SIZE(BB), %xmm3	addps	%xmm1, %xmm5	movaps	28 * SIZE(AA), %xmm1	mulps	%xmm1, %xmm3	mulps	60 * SIZE(BB), %xmm1	addps	%xmm3, %xmm6	movaps	80 * SIZE(BB), %xmm3	addps	%xmm1, %xmm7	movaps	48 * SIZE(AA), %xmm1	addl   $32 * SIZE, AA	addl   $64 * SIZE, BB	decl   %eax	jne    .L31	ALIGN_2.L32:#if defined(LT) || defined(RN)	movl	KK, %eax#else	movl	K, %eax	subl	KK, %eax#endif	andl	$7, %eax		# if (k & 1)	BRANCH	je .L34.L33:	mulps	%xmm0, %xmm2	mulps	 4 * SIZE(BB), %xmm0	addps	%xmm2, %xmm4	movaps	 8 * SIZE(BB), %xmm2	addps	%xmm0, %xmm5	movaps	 4 * SIZE(AA), %xmm0	addl	$4 * SIZE, AA	addl	$8 * SIZE, BB	decl	%eax	jg	.L33	ALIGN_4.L34:	addps	%xmm6, %xmm4	addps	%xmm7, %xmm5#if defined(LN) || defined(RT)	movl	KK, %eax#ifdef LN	subl	$4, %eax#else	subl	$2, %eax#endif	movl	AORIG, AA	movl	BORIG, B	leal	BUFFER, BB	sall	$BASE_SHIFT, %eax	leal	(AA, %eax, 4), AA	leal	(B,  %eax, 2), B	leal	(BB, %eax, 8), BB#endif#if defined(LN) || defined(LT)	movaps	 %xmm4, %xmm0	unpcklps %xmm5, %xmm4	unpckhps %xmm5, %xmm0	movaps	 0 * SIZE(B), %xmm2	movaps	 4 * SIZE(B), %xmm3	subps	%xmm4,  %xmm2	subps	%xmm0,  %xmm3#else	movaps	 0 * SIZE(AA), %xmm0	movaps	 4 * SIZE(AA), %xmm2	subps	%xmm4, %xmm0	subps	%xmm5, %xmm2#endif#if defined(LN) || defined(LT)	movaps	MASK, %xmm6#endif#ifdef LN	movss	15 * SIZE(AA), %xmm0	movaps	 %xmm6, %xmm1	shufps	$0x00, %xmm0,  %xmm1	mulps	%xmm1, %xmm3	pshufd	$0xee, %xmm3, %xmm1	movss	14 * SIZE(AA), %xmm0	pshufd	$0x50, %xmm0, %xmm0	mulps	%xmm1, %xmm0	subps	%xmm0, %xmm3	movsd	12 * SIZE(AA), %xmm0	pshufd	$0x50, %xmm0, %xmm0	mulps	%xmm1, %xmm0	subps	%xmm0, %xmm2	movss	10 * SIZE(AA), %xmm0	shufps	$0x00, %xmm6,  %xmm0	mulps	%xmm0, %xmm3	pshufd	$0x44, %xmm3, %xmm1	movsd	 8 * SIZE(AA), %xmm0	pshufd	$0x50, %xmm0, %xmm0	mulps	%xmm1, %xmm0	subps	%xmm0, %xmm2	movss	 5 * SIZE(AA), %xmm0	movaps	 %xmm6, %xmm1	shufps	$0x00, %xmm0,  %xmm1	mulps	%xmm1, %xmm2	pshufd	$0xee, %xmm2, %xmm1	movss	 4 * SIZE(AA), %xmm0	pshufd	$0x50, %xmm0, %xmm0	mulps	%xmm1, %xmm0	subps	%xmm0, %xmm2	movss	 0 * SIZE(AA), %xmm0	shufps	$0x00, %xmm6,  %xmm0	mulps	%xmm0, %xmm2#endif#ifdef LT	movss	 0 * SIZE(AA), %xmm0	shufps	$0x00, %xmm6,  %xmm0	mulps	%xmm0, %xmm2	pshufd	$0x44, %xmm2, %xmm1	movss	 1 * SIZE(AA), %xmm0	pshufd	$0x05, %xmm0, %xmm0	mulps	%xmm1, %xmm0	subps	%xmm0, %xmm2	movsd	 2 * SIZE(AA), %xmm0	pshufd	$0x50, %xmm0, %xmm0	mulps	%xmm1, %xmm0	subps	%xmm0, %xmm3	movss	 5 * SIZE(AA), %xmm0	movaps	 %xmm6, %xmm1	shufps	$0x00, %xmm0,  %xmm1	mulps	%xmm1, %xmm2	pshufd	$0xee, %xmm2, %xmm1	movsd	 6 * SIZE(AA), %xmm0	pshufd	$0x50, %xmm0, %xmm0	mulps	%xmm1, %xmm0	subps	%xmm0, %xmm3	movss	10 * SIZE(AA), %xmm0	shufps	$0x00, %xmm6,  %xmm0	mulps	%xmm0, %xmm3	pshufd	$0x44, %xmm3, %xmm1	movss	11 * SIZE(AA), %xmm0	pshufd	$0x05, %xmm0, %xmm0	mulps	%xmm1, %xmm0	subps	%xmm0, %xmm3	movss	15 * SIZE(AA), %xmm0	movaps	 %xmm6, %xmm1	shufps	$0x00, %xmm0,  %xmm1	mulps	%xmm1, %xmm3#endif#ifdef RN	movss	 0 * SIZE(B), %xmm6	shufps	$0x00, %xmm6, %xmm6		mulps	%xmm6, %xmm0	movss	 1 * SIZE(B), %xmm6	shufps	$0x00, %xmm6, %xmm6	movaps	%xmm6, %xmm5	mulps	%xmm0, %xmm5	subps	%xmm5, %xmm2	movss	 3 * SIZE(B), %xmm6	shufps	$0x00, %xmm6, %xmm6		mulps	%xmm6, %xmm2#endif#ifdef RT	movss	 3 * SIZE(B), %xmm6	shufps	$0x00, %xmm6, %xmm6		mulps	%xmm6, %xmm2	movss	 2 * SIZE(B), %xmm6	shufps	$0x00, %xmm6, %xmm6	movaps	%xmm6, %xmm5	mulps	%xmm2, %xmm5	subps	%xmm5, %xmm0	movss	 0 * SIZE(B), %xmm6	shufps	$0x00, %xmm6, %xmm6		mulps	%xmm6, %xmm0#endif#if defined(LN) || defined(LT)	movaps	%xmm2,   0 * SIZE(B)	movaps	%xmm3,   4 * SIZE(B)	pshufd	$0x00, %xmm2, %xmm0	pshufd	$0x55, %xmm2, %xmm1	pshufd	$0xaa, %xmm2, %xmm4	pshufd	$0xff, %xmm2, %xmm6	movaps	%xmm0,   0 * SIZE(BB)	movaps	%xmm1,   4 * SIZE(BB)	movaps	%xmm4,   8 * SIZE(BB)	movaps	%xmm6,  12 * SIZE(BB)	pshufd	$0x00, %xmm3, %xmm0	pshufd	$0x55, %xmm3, %xmm1	pshufd	$0xaa, %xmm3, %xmm4	pshufd	$0xff, %xmm3, %xmm6	movaps	%xmm0,  16 * SIZE(BB)	movaps	%xmm1,  20 * SIZE(BB)	movaps	%xmm4,  24 * SIZE(BB)	movaps	%xmm6,  28 * SIZE(BB)#else	movaps	%xmm0,   0 * SIZE(AA)	movaps	%xmm2,   4 * SIZE(AA)#endif#ifdef LN	subl	$4 * SIZE, CO1#endif#if defined(LN) || defined(LT)	movaps	 %xmm2, %xmm0	shufps	 $0x88, %xmm3, %xmm2	shufps	 $0xdd, %xmm3, %xmm0	movlps	%xmm2, 0 * SIZE(CO1)	movhps	%xmm2, 2 * SIZE(CO1)	movlps	%xmm0, 0 * SIZE(CO1, LDC)	movhps	%xmm0, 2 * SIZE(CO1, LDC)#else	movlps	%xmm0, 0 * SIZE(CO1)	movhps	%xmm0, 2 * SIZE(CO1)	movlps	%xmm2, 0 * SIZE(CO1, LDC)	movhps	%xmm2, 2 * SIZE(CO1, LDC)#endif#ifndef LN	addl	$4 * SIZE, CO1#endif#if defined(LT) || defined(RN)	movl	K,  %eax	subl	KK, %eax	leal	(,%eax, SIZE), %eax	leal	(AA, %eax, 4), AA#ifdef LT	addl	$8 * SIZE, B#endif#endif#ifdef LN	subl	$4, KK	movl	BORIG, B#endif#ifdef LT	addl	$4, KK#endif#ifdef RT	movl	K, %eax	movl	BORIG, B	sall	$2 + BASE_SHIFT, %eax	addl	%eax, AORIG#endif	ALIGN_2	.L50:	testl	$2, M	jle	.L70#ifdef LN       movl	K, %eax       sall	$1 + BASE_SHIFT, %eax       subl	%eax, AORIG#endif#if defined(LN) || defined(RT)	movl	KK, %eax	movl	AORIG, AA	sall	$1 + BASE_SHIFT, %eax	addl	%eax, AA#endif	leal	BUFFER, BB#if defined(LN) || defined(RT)	movl	KK, %eax	sall	$1 + BASE_SHIFT, %eax	leal	(BB, %eax, 4), BB#endif		movlps	 0 * SIZE(BB), %xmm2	xorps	%xmm4, %xmm4	movlps	 0 * SIZE(AA), %xmm0	xorps	%xmm5, %xmm5	movlps	 16 * SIZE(BB), %xmm3	xorps	%xmm6, %xmm6	movlps	 8 * SIZE(AA), %xmm1	xorps	%xmm7, %xmm7#if defined(LT) || defined(RN)	movl	KK, %eax#else	movl	K, %eax	subl	KK, %eax#endif	sarl	$3, %eax	je	.L52	ALIGN_2.L51:	mulps	%xmm0, %xmm2	addps	%xmm2, %xmm4	movlps	 4 * SIZE(BB), %xmm2	mulps	%xmm0, %xmm2	movlps	 2 * SIZE(AA), %xmm0	addps	%xmm2, %xmm5	movlps	 8 * SIZE(BB), %xmm2	mulps	%xmm0, %xmm2	addps	%xmm2, %xmm6	movlps	12 * SIZE(BB), %xmm2	mulps	%xmm0, %xmm2	movlps	 4 * SIZE(AA), %xmm0	addps	%xmm2, %xmm7	movlps	32 * SIZE(BB), %xmm2	mulps	%xmm0, %xmm3	addps	%xmm3, %xmm4	movlps	20 * SIZE(BB), %xmm3	mulps	%xmm0, %xmm3	movlps	 6 * SIZE(AA), %xmm0	addps	%xmm3, %xmm5	movlps	24 * SIZE(BB), %xmm3	mulps	%xmm0, %xmm3	addps	%xmm3, %xmm6	movlps	28 * SIZE(BB), %xmm3	mulps	%xmm0, %xmm3	movlps	16 * SIZE(AA), %xmm0	addps	%xmm3, %xmm7	movlps	48 * SIZE(BB), %xmm3	mulps	%xmm1, %xmm2	addps	%xmm2, %xmm4	movlps	36 * SIZE(BB), %xmm2	mulps	%xmm1, %xmm2	movlps	10 * SIZE(AA), %xmm1	addps	%xmm2, %xmm5	movlps	40 * SIZE(BB), %xmm2	mulps	%xmm1, %xmm2	addps	%xmm2, %xmm6	movlps	44 * SIZE(BB), %xmm2	mulps	%xmm1, %xmm2	movlps	12 * SIZE(AA), %xmm1	addps	%xmm2, %xmm7	movlps	64 * SIZE(BB), %xmm2	mulps	%xmm1, %xmm3	addps	%xmm3, %xmm4	movlps	52 * SIZE(BB), %xmm3	mulps	%xmm1, %xmm3	movlps	14 * SIZE(AA), %xmm1	addps	%xmm3, %xmm5	movlps	56 * SIZE(BB), %xmm3	mulps	%xmm1, %xmm3	addps	%xmm3, %xmm6	movlps	60 * SIZE(BB), %xmm3	mulps	%xmm1, %xmm3	movlps	24 * SIZE(AA), %xmm1	addps	%xmm3, %xmm7	movlps	80 * SIZE(BB), %xmm3	addl   $16 * SIZE, AA	addl   $64 * SIZE, BB	decl   %eax	jne    .L51	ALIGN_2.L52:#if defined(LT) || defined(RN)	movl	KK, %eax#else	movl	K, %eax	subl	KK, %eax#endif	andl	$7, %eax		# if (k & 1)	BRANCH	je .L54.L53:	mulps	%xmm0, %xmm2	addps	%xmm2, %xmm4	movlps	 4 * SIZE(BB), %xmm2	mulps	%xmm0, %xmm2	movlps	 2 * SIZE(AA), %xmm0	addps	%xmm2, %xmm5	movlps	 8 * SIZE(BB), %xmm2	addl	$2 * SIZE, AA	addl	$8 * SIZE, BB	decl	%eax	jg	.L53	ALIGN_4.L54:	addps	%xmm6, %xmm4	addps	%xmm7, %xmm5#if defined(LN) || defined(RT)	movl	KK, %eax#ifdef LN	subl	$2, %eax#else	subl	$2, %eax#endif	movl	AORIG, AA	movl	BORIG, B	leal	BUFFER, BB	sall	$BASE_SHIFT, %eax	leal	(AA, %eax, 2), AA	leal	(B,  %eax, 2), B	leal	(BB, %eax, 8), BB#endif#if defined(LN) || defined(LT)	unpcklps %xmm5, %xmm4	movaps	 0 * SIZE(B), %xmm2	subps	%xmm4,  %xmm2#else	movsd	 0 * SIZE(AA), %xmm0	movsd	 2 * SIZE(AA), %xmm2	subps	%xmm4, %xmm0	subps	%xmm5, %xmm2#endif#if defined(LN) || defined(LT)	movaps	MASK, %xmm6#endif#ifdef LN	movss	 3 * SIZE(AA), %xmm0	movaps	 %xmm6, %xmm1	shufps	$0x00, %xmm0,  %xmm1	mulps	%xmm1, %xmm2	pshufd	$0xee, %xmm2, %xmm1	movss	 2 * SIZE(AA), %xmm0	pshufd	$0x50, %xmm0, %xmm0	mulps	%xmm1, %xmm0	subps	%xmm0, %xmm2	movss	 0 * SIZE(AA), %xmm0	shufps	$0x00, %xmm6,  %xmm0	mulps	%xmm0, %xmm2#endif#ifdef LT	movss	 0 * SIZE(AA), %xmm0	shufps	$0x00, %xmm6,  %xmm0	mulps	%xmm0, %xmm2	pshufd	$0x44, %xmm2, %xmm1	movss	 1 * SIZE(AA), %xmm0	pshufd	$0x05, %xmm0, %xmm0	mulps	%xmm1, %xmm0	subps	%xmm0, %xmm2	movss	 3 * SIZE(AA), %xmm0	movaps	 %xmm6, %xmm1	shufps	$0x00, %xmm0,  %xmm1	mulps	%xmm1, %xmm2#endif#ifdef RN	movss	 0 * SIZE(B), %xmm6	shufps	$0x00, %xmm6, %xmm6		mulps	%xmm6, %xmm0	movss	 1 * SIZE(B), %xmm6	shufps	$0x00, %xmm6, %xmm6	movaps	%xmm6, %xmm5	mulps	%xmm0, %xmm5	subps	%xmm5, %xmm2	movss	 3 * SIZE(B), %xmm6	shufps	$0x00, %xmm6, %xmm6		mulps	%xmm6, %xmm2#endif#ifdef RT	movss	 3 * SIZE(B), %xmm6	shufps	$0x00, %xmm6, %xmm6		mulps	%xmm6, %xmm2	movss	 2 * SIZE(B), %xmm6	shufps	$0x00, %xmm6, %xmm6	movaps	%xmm6, %xmm5	mulps	%xmm2, %xmm5	subps	%xmm5, %xmm0	movss	 0 * SIZE(B), %xmm6	shufps	$0x00, %xmm6, %xmm6		mulps	%xmm6, %xmm0#endif#if defined(LN) || defined(LT)	movaps	%xmm2,   0 * SIZE(B)	pshufd	$0x00, %xmm2, %xmm0	pshufd	$0x55, %xmm2, %xmm1	pshufd	$0xaa, %xmm2, %xmm4	pshufd	$0xff, %xmm2, %xmm6	movaps	%xmm0,   0 * SIZE(BB)	movaps	%xmm1,   4 * SIZE(BB)	movaps	%xmm4,   8 * SIZE(BB)	movaps	%xmm6,  12 * SIZE(BB)#else	movlps	%xmm0,   0 * SIZE(AA)	movlps	%xmm2,   2 * SIZE(AA)#endif#ifdef LN	subl	$2 * SIZE, CO1#endif#if defined(LN) || defined(LT)	movaps	 %xmm2, %xmm0	shufps	 $0x88, %xmm3, %xmm2	shufps	 $0xdd, %xmm3, %xmm0	movlps	%xmm2, 0 * SIZE(CO1)	movlps	%xmm0, 0 * SIZE(CO1, LDC)#else	movlps	%xmm0, 0 * SIZE(CO1)	movlps	%xmm2, 0 * SIZE(CO1, LDC)#endif#ifndef LN	addl	$2 * SIZE, CO1#endif#if defined(LT) || defined(RN)	movl	K,  %eax	subl	KK, %eax	leal	(,%eax, SIZE), %eax	leal	(AA, %eax, 2), AA#ifdef LT	addl	$4 * SIZE, B#endif#endif#ifdef LN	subl	$2, KK	movl	BORIG, B#endif#ifdef LT	addl	$2, KK#endif#ifdef RT	movl	K, %eax
⌨️ 快捷键说明

复制代码 Ctrl + C
搜索代码 Ctrl + F
全屏模式 F11
切换主题 Ctrl + Shift + D
显示快捷键 ?
增大字号 Ctrl + =
减小字号 Ctrl + -