📄 zrot_sse2.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 2 页
字号:
上一页 12
	movapd	5 * SIZE(X), %xmm2	movapd	7 * SIZE(X), %xmm0	movapd	4 * SIZE(Y), %xmm1	movapd	6 * SIZE(Y), %xmm3	SHUFPD_1 %xmm2, %xmm8	SHUFPD_1 %xmm0, %xmm2	movapd	%xmm1, %xmm4	movapd	%xmm8, %xmm5	movapd	%xmm3, %xmm6	movapd	%xmm2, %xmm7	mulpd	C, %xmm8	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	mulpd	C, %xmm4	mulpd	S, %xmm5	mulpd	C, %xmm6	mulpd	S, %xmm7	addpd	%xmm1, %xmm8	addpd	%xmm3, %xmm2	subpd	%xmm5, %xmm4	subpd	%xmm7, %xmm6	movlpd	%xmm8, 4 * SIZE(X)	movhpd	%xmm8, 5 * SIZE(X)	movlpd	%xmm2, 6 * SIZE(X)	movhpd	%xmm2, 7 * SIZE(X)	movapd	%xmm4, 4 * SIZE(Y)	movapd	%xmm6, 6 * SIZE(Y)#ifdef PENTIUM4	PREFETCH	(PREFETCH_SIZE + 0) * SIZE(Y)#endif#if defined(OPTERON) || defined(BARCELONA)	PREFETCH	(PREFETCH_SIZE + 8) * SIZE(X)#endif	movapd	 9 * SIZE(X), %xmm2	movapd	11 * SIZE(X), %xmm8	movapd	 8 * SIZE(Y), %xmm1	movapd	10 * SIZE(Y), %xmm3	SHUFPD_1 %xmm2, %xmm0	SHUFPD_1 %xmm8, %xmm2	movapd	%xmm1, %xmm4	movapd	%xmm0, %xmm5	movapd	%xmm3, %xmm6	movapd	%xmm2, %xmm7	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	mulpd	C, %xmm4	mulpd	S, %xmm5	mulpd	C, %xmm6	mulpd	S, %xmm7	addpd	%xmm1, %xmm0	addpd	%xmm3, %xmm2	subpd	%xmm5, %xmm4	subpd	%xmm7, %xmm6	movlpd	%xmm0,  8 * SIZE(X)	movhpd	%xmm0,  9 * SIZE(X)	movlpd	%xmm2, 10 * SIZE(X)	movhpd	%xmm2, 11 * SIZE(X)	movapd	%xmm4,  8 * SIZE(Y)	movapd	%xmm6, 10 * SIZE(Y)#if defined(OPTERON) || defined(BARCELONA)	PREFETCH	(PREFETCH_SIZE + 8) * SIZE(Y)#endif	movapd	13 * SIZE(X), %xmm2	movapd	15 * SIZE(X), %xmm0	movapd	12 * SIZE(Y), %xmm1	movapd	14 * SIZE(Y), %xmm3	SHUFPD_1 %xmm2, %xmm8	SHUFPD_1 %xmm0, %xmm2	movapd	%xmm1, %xmm4	movapd	%xmm8, %xmm5	movapd	%xmm3, %xmm6	movapd	%xmm2, %xmm7	mulpd	C, %xmm8	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	mulpd	C, %xmm4	mulpd	S, %xmm5	mulpd	C, %xmm6	mulpd	S, %xmm7	addpd	%xmm1, %xmm8	addpd	%xmm3, %xmm2	subpd	%xmm5, %xmm4	subpd	%xmm7, %xmm6	movlpd	%xmm8, 12 * SIZE(X)	movhpd	%xmm8, 13 * SIZE(X)	movlpd	%xmm2, 14 * SIZE(X)	movhpd	%xmm2, 15 * SIZE(X)	movapd	%xmm4, 12 * SIZE(Y)	movapd	%xmm6, 14 * SIZE(Y)	addq	$16 * SIZE, Y	addq	$16 * SIZE, X	decq	%rax	jg	.L31	ALIGN_3.L34:	testq	$7, N	jle	.L999	testq	$4, N	jle	.L35	movapd	1 * SIZE(X), %xmm2	movapd	3 * SIZE(X), %xmm8	movapd	0 * SIZE(Y), %xmm1	movapd	2 * SIZE(Y), %xmm3	SHUFPD_1 %xmm2, %xmm0	SHUFPD_1 %xmm8, %xmm2	movapd	%xmm1, %xmm4	movapd	%xmm0, %xmm5	movapd	%xmm3, %xmm6	movapd	%xmm2, %xmm7	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	mulpd	C, %xmm4	mulpd	S, %xmm5	mulpd	C, %xmm6	mulpd	S, %xmm7	addpd	%xmm1, %xmm0	addpd	%xmm3, %xmm2	subpd	%xmm5, %xmm4	subpd	%xmm7, %xmm6	movlpd	%xmm0, 0 * SIZE(X)	movhpd	%xmm0, 1 * SIZE(X)	movlpd	%xmm2, 2 * SIZE(X)	movhpd	%xmm2, 3 * SIZE(X)	movapd	%xmm4, 0 * SIZE(Y)	movapd	%xmm6, 2 * SIZE(Y)	movapd	5 * SIZE(X), %xmm2	movapd	7 * SIZE(X), %xmm0	movapd	4 * SIZE(Y), %xmm1	movapd	6 * SIZE(Y), %xmm3	SHUFPD_1 %xmm2, %xmm8	SHUFPD_1 %xmm0, %xmm2	movapd	%xmm1, %xmm4	movapd	%xmm8, %xmm5	movapd	%xmm3, %xmm6	movapd	%xmm2, %xmm7	mulpd	C, %xmm8	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	mulpd	C, %xmm4	mulpd	S, %xmm5	mulpd	C, %xmm6	mulpd	S, %xmm7	addpd	%xmm1, %xmm8	addpd	%xmm3, %xmm2	subpd	%xmm5, %xmm4	subpd	%xmm7, %xmm6	movlpd	%xmm8, 4 * SIZE(X)	movhpd	%xmm8, 5 * SIZE(X)	movlpd	%xmm2, 6 * SIZE(X)	movhpd	%xmm2, 7 * SIZE(X)	movapd	%xmm4, 4 * SIZE(Y)	movapd	%xmm6, 6 * SIZE(Y)	addq	$8 * SIZE, Y	addq	$8 * SIZE, X	ALIGN_3.L35:	testq	$2, N	jle	.L36	movapd	1 * SIZE(X), %xmm2	movapd	3 * SIZE(X), %xmm8	movapd	0 * SIZE(Y), %xmm1	movapd	2 * SIZE(Y), %xmm3	SHUFPD_1 %xmm2, %xmm0	SHUFPD_1 %xmm8, %xmm2	movapd	%xmm1, %xmm4	movapd	%xmm0, %xmm5	movapd	%xmm3, %xmm6	movapd	%xmm2, %xmm7	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	mulpd	C, %xmm4	mulpd	S, %xmm5	mulpd	C, %xmm6	mulpd	S, %xmm7	addpd	%xmm1, %xmm0	addpd	%xmm3, %xmm2	subpd	%xmm5, %xmm4	subpd	%xmm7, %xmm6	movlpd	%xmm0, 0 * SIZE(X)	movhpd	%xmm0, 1 * SIZE(X)	movlpd	%xmm2, 2 * SIZE(X)	movhpd	%xmm2, 3 * SIZE(X)	movapd	%xmm4, 0 * SIZE(Y)	movapd	%xmm6, 2 * SIZE(Y)	movapd	%xmm8, %xmm0	addq	$4 * SIZE, Y	addq	$4 * SIZE, X	ALIGN_3.L36:	testq	$1, N	jle	.L999	movapd	 1 * SIZE(X), %xmm4	movapd	 0 * SIZE(Y), %xmm1	SHUFPD_1 %xmm4, %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movlpd	%xmm0, 0 * SIZE(X)	movhpd	%xmm0, 1 * SIZE(X)	movapd	%xmm2, 0 * SIZE(Y)	jmp	.L999	ALIGN_3.L40:	movsd	0 * SIZE(Y), %xmm1	movsd	0 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulsd	C, %xmm0	mulsd	S, %xmm1	mulsd	C, %xmm2	mulsd	S, %xmm3	addsd	%xmm1, %xmm0	subsd	%xmm3, %xmm2	movsd	%xmm0, 0 * SIZE(X)	movsd	%xmm2, 0 * SIZE(Y)	addq	$1 * SIZE, Y	addq	$1 * SIZE, X	decq	N	jle	.L47	movq	N,  %rax	sarq	$3, %rax	jle	.L44	movapd	0 * SIZE(Y), %xmm1	movapd	2 * SIZE(Y), %xmm3	movapd	4 * SIZE(Y), %xmm9	movapd	6 * SIZE(Y), %xmm11	movapd	0 * SIZE(X), %xmm0	movapd	2 * SIZE(X), %xmm2	movapd	4 * SIZE(X), %xmm8	movapd	6 * SIZE(X), %xmm10		decq	%rax	jle	.L42	ALIGN_3.L41:#ifdef PENTIUM4	PREFETCH	(PREFETCH_SIZE + 0) * SIZE(X)#endif#if defined(OPTERON) || defined(BARCELONA)	PREFETCH	(PREFETCH_SIZE + 0) * SIZE(X)#endif	movapd	%xmm1, %xmm4	mulpd	S, %xmm1	movapd	%xmm3, %xmm6	mulpd	S, %xmm3	movapd	%xmm0, %xmm5	mulpd	C, %xmm0	movapd	%xmm2, %xmm7	mulpd	C, %xmm2	mulpd	C, %xmm4	mulpd	S, %xmm5	mulpd	C, %xmm6	mulpd	S, %xmm7	addpd	%xmm1, %xmm0	movapd	 8 * SIZE(Y), %xmm1	addpd	%xmm3, %xmm2	movapd	10 * SIZE(Y), %xmm3	subpd	%xmm5, %xmm4	subpd	%xmm7, %xmm6#if defined(OPTERON) || defined(BARCELONA)	PREFETCH	(PREFETCH_SIZE + 0) * SIZE(Y)#endif	movapd	%xmm0, 0 * SIZE(X)	movapd	 8 * SIZE(X), %xmm0	movapd	%xmm2, 2 * SIZE(X)	movapd	10 * SIZE(X), %xmm2	movapd	%xmm4, 0 * SIZE(Y)	movapd	%xmm6, 2 * SIZE(Y)	movapd	%xmm9, %xmm4	mulpd	S, %xmm9	movapd	%xmm8, %xmm5	mulpd	C, %xmm8	movapd	%xmm11, %xmm6	mulpd	S, %xmm11	movapd	%xmm10, %xmm7	mulpd	C, %xmm10	mulpd	C, %xmm4	mulpd	S, %xmm5	mulpd	C, %xmm6	mulpd	S, %xmm7	addpd	%xmm9,  %xmm8	movapd	12 * SIZE(Y), %xmm9	addpd	%xmm11, %xmm10	movapd	14 * SIZE(Y), %xmm11	subpd	%xmm5,  %xmm4	subpd	%xmm7,  %xmm6	movapd	%xmm8, 4 * SIZE(X)	movapd	12 * SIZE(X), %xmm8	movapd	%xmm10,6 * SIZE(X)	movapd	14 * SIZE(X), %xmm10	movapd	%xmm4, 4 * SIZE(Y)	movapd	%xmm6, 6 * SIZE(Y)#ifdef PENTIUM4	PREFETCH	(PREFETCH_SIZE + 0) * SIZE(Y)#endif#if defined(OPTERON) || defined(BARCELONA)	PREFETCH	(PREFETCH_SIZE + 8) * SIZE(X)#endif	movapd	%xmm1, %xmm4	mulpd	S, %xmm1	movapd	%xmm3, %xmm6	mulpd	S, %xmm3	movapd	%xmm0, %xmm5	mulpd	C, %xmm0	movapd	%xmm2, %xmm7	mulpd	C, %xmm2	mulpd	C, %xmm4	mulpd	S, %xmm5	mulpd	C, %xmm6	mulpd	S, %xmm7	addpd	%xmm1, %xmm0	movapd	16 * SIZE(Y), %xmm1	addpd	%xmm3, %xmm2	movapd	18 * SIZE(Y), %xmm3	subpd	%xmm5, %xmm4	subpd	%xmm7, %xmm6	movapd	%xmm0,  8 * SIZE(X)	movapd	16 * SIZE(X), %xmm0	movapd	%xmm2, 10 * SIZE(X)	movapd	18 * SIZE(X), %xmm2	movapd	%xmm4,  8 * SIZE(Y)	movapd	%xmm6, 10 * SIZE(Y)#if defined(OPTERON) || defined(BARCELONA)	PREFETCH	(PREFETCH_SIZE + 8) * SIZE(Y)#endif	movapd	%xmm9, %xmm4	mulpd	S, %xmm9	movapd	%xmm8, %xmm5	mulpd	C, %xmm8	movapd	%xmm11, %xmm6	mulpd	S, %xmm11	movapd	%xmm10, %xmm7	mulpd	C, %xmm10	mulpd	C, %xmm4	mulpd	S, %xmm5	mulpd	C, %xmm6	mulpd	S, %xmm7	addpd	%xmm9,  %xmm8	movapd	20 * SIZE(Y), %xmm9	addpd	%xmm11, %xmm10	movapd	22 * SIZE(Y), %xmm11	subpd	%xmm5,  %xmm4	subpd	%xmm7,  %xmm6	movapd	%xmm8,  12 * SIZE(X)	movapd	20 * SIZE(X), %xmm8	movapd	%xmm10, 14 * SIZE(X)	movapd	22 * SIZE(X), %xmm10	movapd	%xmm4,  12 * SIZE(Y)	movapd	%xmm6,  14 * SIZE(Y)	addq	$16 * SIZE, X	addq	$16 * SIZE, Y	decq	%rax	jg	.L41	ALIGN_3.L42:	movapd	%xmm1, %xmm4	mulpd	S, %xmm1	movapd	%xmm3, %xmm6	mulpd	S, %xmm3	movapd	%xmm0, %xmm5	mulpd	C, %xmm0	movapd	%xmm2, %xmm7	mulpd	C, %xmm2	mulpd	C, %xmm4	mulpd	S, %xmm5	mulpd	C, %xmm6	mulpd	S, %xmm7	addpd	%xmm1, %xmm0	movapd	 8 * SIZE(Y), %xmm1	addpd	%xmm3, %xmm2	movapd	10 * SIZE(Y), %xmm3	subpd	%xmm5, %xmm4	subpd	%xmm7, %xmm6	movapd	%xmm0, 0 * SIZE(X)	movapd	 8 * SIZE(X), %xmm0	movapd	%xmm2, 2 * SIZE(X)	movapd	10 * SIZE(X), %xmm2	movapd	%xmm4, 0 * SIZE(Y)	movapd	%xmm6, 2 * SIZE(Y)	movapd	%xmm9, %xmm4	mulpd	S, %xmm9	movapd	%xmm8, %xmm5	mulpd	C, %xmm8	movapd	%xmm11, %xmm6	mulpd	S, %xmm11	movapd	%xmm10, %xmm7	mulpd	C, %xmm10	mulpd	C, %xmm4	mulpd	S, %xmm5	mulpd	C, %xmm6	mulpd	S, %xmm7	addpd	%xmm9,  %xmm8	movapd	12 * SIZE(Y), %xmm9	addpd	%xmm11, %xmm10	movapd	14 * SIZE(Y), %xmm11	subpd	%xmm5,  %xmm4	subpd	%xmm7,  %xmm6	movapd	%xmm8, 4 * SIZE(X)	movapd	12 * SIZE(X), %xmm8	movapd	%xmm10,6 * SIZE(X)	movapd	14 * SIZE(X), %xmm10	movapd	%xmm4, 4 * SIZE(Y)	movapd	%xmm6, 6 * SIZE(Y)	movapd	%xmm1, %xmm4	mulpd	S, %xmm1	movapd	%xmm3, %xmm6	mulpd	S, %xmm3	movapd	%xmm0, %xmm5	mulpd	C, %xmm0	movapd	%xmm2, %xmm7	mulpd	C, %xmm2	mulpd	C, %xmm4	mulpd	S, %xmm5	mulpd	C, %xmm6	mulpd	S, %xmm7	addpd	%xmm1, %xmm0	addpd	%xmm3, %xmm2	subpd	%xmm5, %xmm4	subpd	%xmm7, %xmm6	movapd	%xmm0,  8 * SIZE(X)	movapd	%xmm2, 10 * SIZE(X)	movapd	%xmm4,  8 * SIZE(Y)	movapd	%xmm6, 10 * SIZE(Y)	movapd	%xmm9, %xmm4	mulpd	S, %xmm9	movapd	%xmm8, %xmm5	mulpd	C, %xmm8	movapd	%xmm11, %xmm6	mulpd	S, %xmm11	movapd	%xmm10, %xmm7	mulpd	C, %xmm10	mulpd	C, %xmm4	mulpd	S, %xmm5	mulpd	C, %xmm6	mulpd	S, %xmm7	addpd	%xmm9,  %xmm8	addpd	%xmm11, %xmm10	subpd	%xmm5,  %xmm4	subpd	%xmm7,  %xmm6	movapd	%xmm8,  12 * SIZE(X)	movapd	%xmm10, 14 * SIZE(X)	movapd	%xmm4,  12 * SIZE(Y)	movapd	%xmm6,  14 * SIZE(Y)	addq	$16 * SIZE, X	addq	$16 * SIZE, Y	ALIGN_3.L44:	testq	$4, N	jle	.L45	movapd	0 * SIZE(Y), %xmm1	movapd	0 * SIZE(X), %xmm0	movapd	2 * SIZE(Y), %xmm3	movapd	2 * SIZE(X), %xmm2	movapd	%xmm1, %xmm4	movapd	%xmm0, %xmm5	movapd	%xmm3, %xmm6	movapd	%xmm2, %xmm7	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	mulpd	C, %xmm4	mulpd	S, %xmm5	mulpd	C, %xmm6	mulpd	S, %xmm7	addpd	%xmm1, %xmm0	addpd	%xmm3, %xmm2	subpd	%xmm5, %xmm4	subpd	%xmm7, %xmm6	movapd	%xmm0, 0 * SIZE(X)	movapd	%xmm2, 2 * SIZE(X)	movapd	%xmm4, 0 * SIZE(Y)	movapd	%xmm6, 2 * SIZE(Y)	movapd	4 * SIZE(Y), %xmm1	movapd	4 * SIZE(X), %xmm0	movapd	6 * SIZE(Y), %xmm3	movapd	6 * SIZE(X), %xmm2	movapd	%xmm1, %xmm4	movapd	%xmm0, %xmm5	movapd	%xmm3, %xmm6	movapd	%xmm2, %xmm7	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	mulpd	C, %xmm4	mulpd	S, %xmm5	mulpd	C, %xmm6	mulpd	S, %xmm7	addpd	%xmm1, %xmm0	addpd	%xmm3, %xmm2	subpd	%xmm5, %xmm4	subpd	%xmm7, %xmm6	movapd	%xmm0, 4 * SIZE(X)	movapd	%xmm2, 6 * SIZE(X)	movapd	%xmm4, 4 * SIZE(Y)	movapd	%xmm6, 6 * SIZE(Y)	addq	$8 * SIZE, X	addq	$8 * SIZE, Y	ALIGN_3.L45:	testq	$2, N	jle	.L46	movapd	0 * SIZE(Y), %xmm1	movapd	0 * SIZE(X), %xmm0	movapd	2 * SIZE(Y), %xmm3	movapd	2 * SIZE(X), %xmm2	movapd	%xmm1, %xmm4	movapd	%xmm0, %xmm5	movapd	%xmm3, %xmm6	movapd	%xmm2, %xmm7	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	mulpd	C, %xmm4	mulpd	S, %xmm5	mulpd	C, %xmm6	mulpd	S, %xmm7	addpd	%xmm1, %xmm0	addpd	%xmm3, %xmm2	subpd	%xmm5, %xmm4	subpd	%xmm7, %xmm6	movapd	%xmm0, 0 * SIZE(X)	movapd	%xmm2, 2 * SIZE(X)	movapd	%xmm4, 0 * SIZE(Y)	movapd	%xmm6, 2 * SIZE(Y)	addq	$4 * SIZE, X	addq	$4 * SIZE, Y	ALIGN_3.L46:	testq	$1, N	jle	.L47	movapd	0 * SIZE(Y), %xmm1	movapd	0 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movapd	%xmm0, 0 * SIZE(X)	movapd	%xmm2, 0 * SIZE(Y)	addq	$2 * SIZE, Y	addq	$2 * SIZE, X	ALIGN_3.L47:	movsd	0 * SIZE(Y), %xmm1	movsd	0 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulsd	C, %xmm0	mulsd	S, %xmm1	mulsd	C, %xmm2	mulsd	S, %xmm3	addsd	%xmm1, %xmm0	subsd	%xmm3, %xmm2	movsd	%xmm0, 0 * SIZE(X)	movsd	%xmm2, 0 * SIZE(Y)	jmp	.L999	ALIGN_3.L50:	movq	N,  %rax	sarq	$2, %rax	jle	.L55	ALIGN_3.L53:	MOVSD	0 * SIZE(Y), %xmm1	movhpd	1 * SIZE(Y), %xmm1	MOVSD	0 * SIZE(X), %xmm0	movhpd	1 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movlpd	%xmm0, 0 * SIZE(X)	movhpd	%xmm0, 1 * SIZE(X)	movlpd	%xmm2, 0 * SIZE(Y)	movhpd	%xmm2, 1 * SIZE(Y)	addq	INCX, X	addq	INCY, Y	MOVSD	0 * SIZE(Y), %xmm1	movhpd	1 * SIZE(Y), %xmm1	MOVSD	0 * SIZE(X), %xmm0	movhpd	1 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movlpd	%xmm0, 0 * SIZE(X)	movhpd	%xmm0, 1 * SIZE(X)	movlpd	%xmm2, 0 * SIZE(Y)	movhpd	%xmm2, 1 * SIZE(Y)	addq	INCX, X	addq	INCY, Y	MOVSD	0 * SIZE(Y), %xmm1	movhpd	1 * SIZE(Y), %xmm1	MOVSD	0 * SIZE(X), %xmm0	movhpd	1 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movlpd	%xmm0, 0 * SIZE(X)	movhpd	%xmm0, 1 * SIZE(X)	movlpd	%xmm2, 0 * SIZE(Y)	movhpd	%xmm2, 1 * SIZE(Y)	addq	INCX, X	addq	INCY, Y	MOVSD	0 * SIZE(Y), %xmm1	movhpd	1 * SIZE(Y), %xmm1	MOVSD	0 * SIZE(X), %xmm0	movhpd	1 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movlpd	%xmm0, 0 * SIZE(X)	movhpd	%xmm0, 1 * SIZE(X)	movlpd	%xmm2, 0 * SIZE(Y)	movhpd	%xmm2, 1 * SIZE(Y)	addq	INCX, X	addq	INCY, Y	decq	%rax	jg	.L53	ALIGN_3.L55:	movq	N,  %rax	andq	$3, %rax	jle	.L999	ALIGN_3.L56:	MOVSD	0 * SIZE(Y), %xmm1	movhpd	1 * SIZE(Y), %xmm1	MOVSD	0 * SIZE(X), %xmm0	movhpd	1 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movlpd	%xmm0, 0 * SIZE(X)	movhpd	%xmm0, 1 * SIZE(X)	movlpd	%xmm2, 0 * SIZE(Y)	movhpd	%xmm2, 1 * SIZE(Y)	addq	INCX, X	addq	INCY, Y	decq	%rax	jg	.L56	ALIGN_3.L999:	RESTOREREGISTERS	ret	EPILOGUE
上一页 12
⌨️ 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 + -