📄 zrot_sse.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 2 页
字号:
上一页 12
	movaps	%xmm11, %xmm6	mulps	S, %xmm11	movaps	%xmm10, %xmm7	mulps	C, %xmm10	mulps	C, %xmm4	mulps	S, %xmm5	mulps	C, %xmm6	mulps	S, %xmm7	addps	%xmm9,  %xmm8	addps	%xmm11, %xmm10	subps	%xmm5,  %xmm4	subps	%xmm7,  %xmm6	movaps	%xmm8,  24 * SIZE(X)	movaps	%xmm10, 28 * SIZE(X)	movlps	%xmm4,  24 * SIZE(Y)	movhps	%xmm4,  26 * SIZE(Y)	movlps	%xmm6,  28 * SIZE(Y)	movhps	%xmm6,  30 * SIZE(Y)	addq	$32 * SIZE, X	addq	$32 * SIZE, Y	ALIGN_3.L24:	testq	$15, N	jle	.L999	testq	$8, N	jle	.L25	MOVSD	0 * SIZE(Y), %xmm1	movhps	2 * SIZE(Y), %xmm1	movaps	0 * SIZE(X), %xmm0	MOVSD	4 * SIZE(Y), %xmm3	movhps	6 * SIZE(Y), %xmm3	movaps	4 * SIZE(X), %xmm2	movaps	%xmm1, %xmm4	movaps	%xmm0, %xmm5	movaps	%xmm3, %xmm6	movaps	%xmm2, %xmm7	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	mulps	C, %xmm4	mulps	S, %xmm5	mulps	C, %xmm6	mulps	S, %xmm7	addps	%xmm1, %xmm0	addps	%xmm3, %xmm2	subps	%xmm5, %xmm4	subps	%xmm7, %xmm6	movaps	%xmm0, 0 * SIZE(X)	movaps	%xmm2, 4 * SIZE(X)	movlps	%xmm4, 0 * SIZE(Y)	movhps	%xmm4, 2 * SIZE(Y)	movlps	%xmm6, 4 * SIZE(Y)	movhps	%xmm6, 6 * SIZE(Y)	MOVSD	 8 * SIZE(Y), %xmm1	movhps	10 * SIZE(Y), %xmm1	movaps	 8 * SIZE(X), %xmm0	MOVSD	12 * SIZE(Y), %xmm3	movhps	14 * SIZE(Y), %xmm3	movaps	12 * SIZE(X), %xmm2	movaps	%xmm1, %xmm4	movaps	%xmm0, %xmm5	movaps	%xmm3, %xmm6	movaps	%xmm2, %xmm7	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	mulps	C, %xmm4	mulps	S, %xmm5	mulps	C, %xmm6	mulps	S, %xmm7	addps	%xmm1, %xmm0	addps	%xmm3, %xmm2	subps	%xmm5, %xmm4	subps	%xmm7, %xmm6	movaps	%xmm0,  8 * SIZE(X)	movaps	%xmm2, 12 * SIZE(X)	movlps	%xmm4,  8 * SIZE(Y)	movhps	%xmm4, 10 * SIZE(Y)	movlps	%xmm6, 12 * SIZE(Y)	movhps	%xmm6, 14 * SIZE(Y)	addq	$16 * SIZE, X	addq	$16 * SIZE, Y	ALIGN_3.L25:	testq	$4, N	jle	.L26	MOVSD	0 * SIZE(Y), %xmm1	movhps	2 * SIZE(Y), %xmm1	movaps	0 * SIZE(X), %xmm0	MOVSD	4 * SIZE(Y), %xmm3	movhps	6 * SIZE(Y), %xmm3	movaps	4 * SIZE(X), %xmm2	movaps	%xmm1, %xmm4	movaps	%xmm0, %xmm5	movaps	%xmm3, %xmm6	movaps	%xmm2, %xmm7	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	mulps	C, %xmm4	mulps	S, %xmm5	mulps	C, %xmm6	mulps	S, %xmm7	addps	%xmm1, %xmm0	addps	%xmm3, %xmm2	subps	%xmm5, %xmm4	subps	%xmm7, %xmm6	movaps	%xmm0, 0 * SIZE(X)	movaps	%xmm2, 4 * SIZE(X)	movlps	%xmm4, 0 * SIZE(Y)	movhps	%xmm4, 2 * SIZE(Y)	movlps	%xmm6, 4 * SIZE(Y)	movhps	%xmm6, 6 * SIZE(Y)	addq	$8 * SIZE, X	addq	$8 * SIZE, Y	ALIGN_3.L26:	testq	$2, N	jle	.L27	MOVSD	0 * SIZE(Y), %xmm1	movhps	2 * SIZE(Y), %xmm1	movaps	0 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movaps	%xmm0, 0 * SIZE(X)	movlps	%xmm2, 0 * SIZE(Y)	movhps	%xmm2, 2 * SIZE(Y)	addq	$4 * SIZE, X	addq	$4 * SIZE, Y	ALIGN_3.L27:	testq	$1, N	jle	.L999	movsd	0 * SIZE(Y), %xmm1	movsd	0 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, 0 * SIZE(X)	movlps	%xmm2, 0 * SIZE(Y)	jmp	.L999	ALIGN_3.L30:	movq	N,  %rax	sarq	$4, %rax	jle	.L34	MOVSD	 0 * SIZE(Y), %xmm1	movhps	 2 * SIZE(Y), %xmm1	MOVSD	 4 * SIZE(Y), %xmm3	movhps	 6 * SIZE(Y), %xmm3	MOVSD	 8 * SIZE(Y), %xmm9	movhps	10 * SIZE(Y), %xmm9	MOVSD	12 * SIZE(Y), %xmm11	movhps	14 * SIZE(Y), %xmm11	MOVSD	 0 * SIZE(X), %xmm0	movhps	 2 * SIZE(X), %xmm0	MOVSD	 4 * SIZE(X), %xmm2	movhps	 6 * SIZE(X), %xmm2	MOVSD	 8 * SIZE(X), %xmm8	movhps	10 * SIZE(X), %xmm8	MOVSD	12 * SIZE(X), %xmm10	movhps	14 * SIZE(X), %xmm10		decq	%rax	jle	.L32	ALIGN_3.L31:#ifdef PENTIUM4	PREFETCH	(PREFETCH_SIZE + 0) * SIZE(X)#endif#if defined(OPTERON) || defined(BARCELONA)	PREFETCH	(PREFETCH_SIZE + 0) * SIZE(X)#endif	movaps	%xmm1, %xmm4	mulps	S, %xmm1	movaps	%xmm3, %xmm6	mulps	S, %xmm3	movaps	%xmm0, %xmm5	mulps	C, %xmm0	movaps	%xmm2, %xmm7	mulps	C, %xmm2	mulps	C, %xmm4	mulps	S, %xmm5	mulps	C, %xmm6	mulps	S, %xmm7	addps	%xmm1, %xmm0	MOVSD	16 * SIZE(Y), %xmm1	movhps	18 * SIZE(Y), %xmm1	addps	%xmm3, %xmm2	MOVSD	20 * SIZE(Y), %xmm3	movhps	22 * SIZE(Y), %xmm3	subps	%xmm5, %xmm4	subps	%xmm7, %xmm6#if defined(OPTERON) || defined(BARCELONA)	PREFETCH	(PREFETCH_SIZE + 0) * SIZE(Y)#endif	movlps	%xmm0,  0 * SIZE(X)	movhps	%xmm0,  2 * SIZE(X)	MOVSD	16 * SIZE(X), %xmm0	movhps	18 * SIZE(X), %xmm0	movlps	%xmm2,  4 * SIZE(X)	movhps	%xmm2,  6 * SIZE(X)	MOVSD	20 * SIZE(X), %xmm2	movhps	22 * SIZE(X), %xmm2	movlps	%xmm4,  0 * SIZE(Y)	movhps	%xmm4,  2 * SIZE(Y)	movlps	%xmm6,  4 * SIZE(Y)	movhps	%xmm6,  6 * SIZE(Y)	movaps	%xmm9, %xmm4	mulps	S, %xmm9	movaps	%xmm8, %xmm5	mulps	C, %xmm8	movaps	%xmm11, %xmm6	mulps	S, %xmm11	movaps	%xmm10, %xmm7	mulps	C, %xmm10	mulps	C, %xmm4	mulps	S, %xmm5	mulps	C, %xmm6	mulps	S, %xmm7	addps	%xmm9,  %xmm8	MOVSD	24 * SIZE(Y), %xmm9	movhps	26 * SIZE(Y), %xmm9	addps	%xmm11, %xmm10	MOVSD	28 * SIZE(Y), %xmm11	movhps	30 * SIZE(Y), %xmm11	subps	%xmm5,  %xmm4	subps	%xmm7,  %xmm6	movlps	%xmm8,   8 * SIZE(X)	movhps	%xmm8,  10 * SIZE(X)	MOVSD	24 * SIZE(X), %xmm8	movhps	26 * SIZE(X), %xmm8	movlps	%xmm10, 12 * SIZE(X)	movhps	%xmm10, 14 * SIZE(X)	MOVSD	28 * SIZE(X), %xmm10	movhps	30 * SIZE(X), %xmm10	movlps	%xmm4,   8 * SIZE(Y)	movhps	%xmm4,  10 * SIZE(Y)	movlps	%xmm6,  12 * SIZE(Y)	movhps	%xmm6,  14 * SIZE(Y)#ifdef PENTIUM4	PREFETCH	(PREFETCH_SIZE + 0) * SIZE(Y)#endif#if defined(OPTERON) || defined(BARCELONA)	PREFETCH	(PREFETCH_SIZE + 16) * SIZE(X)#endif	movaps	%xmm1, %xmm4	mulps	S, %xmm1	movaps	%xmm3, %xmm6	mulps	S, %xmm3	movaps	%xmm0, %xmm5	mulps	C, %xmm0	movaps	%xmm2, %xmm7	mulps	C, %xmm2	mulps	C, %xmm4	mulps	S, %xmm5	mulps	C, %xmm6	mulps	S, %xmm7	addps	%xmm1, %xmm0	MOVSD	32 * SIZE(Y), %xmm1	movhps	34 * SIZE(Y), %xmm1	addps	%xmm3, %xmm2	MOVSD	36 * SIZE(Y), %xmm3	movhps	38 * SIZE(Y), %xmm3	subps	%xmm5, %xmm4	subps	%xmm7, %xmm6	movlps	%xmm0, 16 * SIZE(X)	movhps	%xmm0, 18 * SIZE(X)	MOVSD	32 * SIZE(X), %xmm0	movhps	34 * SIZE(X), %xmm0	movlps	%xmm2, 20 * SIZE(X)	movhps	%xmm2, 22 * SIZE(X)	MOVSD	36 * SIZE(X), %xmm2	movhps	38 * SIZE(X), %xmm2	movlps	%xmm4, 16 * SIZE(Y)	movhps	%xmm4, 18 * SIZE(Y)	movlps	%xmm6, 20 * SIZE(Y)	movhps	%xmm6, 22 * SIZE(Y)#if defined(OPTERON) || defined(BARCELONA)	PREFETCH	(PREFETCH_SIZE + 16) * SIZE(Y)#endif	movaps	%xmm9, %xmm4	mulps	S, %xmm9	movaps	%xmm8, %xmm5	mulps	C, %xmm8	movaps	%xmm11, %xmm6	mulps	S, %xmm11	movaps	%xmm10, %xmm7	mulps	C, %xmm10	mulps	C, %xmm4	mulps	S, %xmm5	mulps	C, %xmm6	mulps	S, %xmm7	addps	%xmm9,  %xmm8	MOVSD	40 * SIZE(Y), %xmm9	movhps	42 * SIZE(Y), %xmm9	addps	%xmm11, %xmm10	MOVSD	44 * SIZE(Y), %xmm11	movhps	46 * SIZE(Y), %xmm11	subps	%xmm5,  %xmm4	subps	%xmm7,  %xmm6	movlps	%xmm8,  24 * SIZE(X)	movhps	%xmm8,  26 * SIZE(X)	MOVSD	40 * SIZE(X), %xmm8	movhps	42 * SIZE(X), %xmm8	movlps	%xmm10, 28 * SIZE(X)	movhps	%xmm10, 30 * SIZE(X)	MOVSD	44 * SIZE(X), %xmm10	movhps	46 * SIZE(X), %xmm10	movlps	%xmm4,  24 * SIZE(Y)	movhps	%xmm4,  26 * SIZE(Y)	movlps	%xmm6,  28 * SIZE(Y)	movhps	%xmm6,  30 * SIZE(Y)	addq	$32 * SIZE, X	addq	$32 * SIZE, Y	decq	%rax	jg	.L31	ALIGN_3.L32:	movaps	%xmm1, %xmm4	mulps	S, %xmm1	movaps	%xmm3, %xmm6	mulps	S, %xmm3	movaps	%xmm0, %xmm5	mulps	C, %xmm0	movaps	%xmm2, %xmm7	mulps	C, %xmm2	mulps	C, %xmm4	mulps	S, %xmm5	mulps	C, %xmm6	mulps	S, %xmm7	addps	%xmm1, %xmm0	MOVSD	16 * SIZE(Y), %xmm1	movhps	18 * SIZE(Y), %xmm1	addps	%xmm3, %xmm2	MOVSD	20 * SIZE(Y), %xmm3	movhps	22 * SIZE(Y), %xmm3	subps	%xmm5, %xmm4	subps	%xmm7, %xmm6	movlps	%xmm0,  0 * SIZE(X)	movhps	%xmm0,  2 * SIZE(X)	MOVSD	16 * SIZE(X), %xmm0	movhps	18 * SIZE(X), %xmm0	movlps	%xmm2,  4 * SIZE(X)	movhps	%xmm2,  6 * SIZE(X)	MOVSD	20 * SIZE(X), %xmm2	movhps	22 * SIZE(X), %xmm2	MOVSD	%xmm4,  0 * SIZE(Y)	movhps	%xmm4,  2 * SIZE(Y)	MOVSD	%xmm6,  4 * SIZE(Y)	movhps	%xmm6,  6 * SIZE(Y)	movaps	%xmm9, %xmm4	mulps	S, %xmm9	movaps	%xmm8, %xmm5	mulps	C, %xmm8	movaps	%xmm11, %xmm6	mulps	S, %xmm11	movaps	%xmm10, %xmm7	mulps	C, %xmm10	mulps	C, %xmm4	mulps	S, %xmm5	mulps	C, %xmm6	mulps	S, %xmm7	addps	%xmm9,  %xmm8	MOVSD	24 * SIZE(Y), %xmm9	movhps	26 * SIZE(Y), %xmm9	addps	%xmm11, %xmm10	MOVSD	28 * SIZE(Y), %xmm11	movhps	30 * SIZE(Y), %xmm11	subps	%xmm5,  %xmm4	subps	%xmm7,  %xmm6	movlps	%xmm8,   8 * SIZE(X)	movhps	%xmm8,  10 * SIZE(X)	MOVSD	24 * SIZE(X), %xmm8	movhps	26 * SIZE(X), %xmm8	movlps	%xmm10, 12 * SIZE(X)	movhps	%xmm10, 14 * SIZE(X)	MOVSD	28 * SIZE(X), %xmm10	movhps	30 * SIZE(X), %xmm10	movlps	%xmm4,  8 * SIZE(Y)	movhps	%xmm4, 10 * SIZE(Y)	movlps	%xmm6, 12 * SIZE(Y)	movhps	%xmm6, 14 * SIZE(Y)	movaps	%xmm1, %xmm4	mulps	S, %xmm1	movaps	%xmm3, %xmm6	mulps	S, %xmm3	movaps	%xmm0, %xmm5	mulps	C, %xmm0	movaps	%xmm2, %xmm7	mulps	C, %xmm2	mulps	C, %xmm4	mulps	S, %xmm5	mulps	C, %xmm6	mulps	S, %xmm7	addps	%xmm1, %xmm0	addps	%xmm3, %xmm2	subps	%xmm5, %xmm4	subps	%xmm7, %xmm6	movlps	%xmm0, 16 * SIZE(X)	movhps	%xmm0, 18 * SIZE(X)	movlps	%xmm2, 20 * SIZE(X)	movhps	%xmm2, 22 * SIZE(X)	movlps	%xmm4, 16 * SIZE(Y)	movhps	%xmm4, 18 * SIZE(Y)	movlps	%xmm6, 20 * SIZE(Y)	movhps	%xmm6, 22 * SIZE(Y)	movaps	%xmm9, %xmm4	mulps	S, %xmm9	movaps	%xmm8, %xmm5	mulps	C, %xmm8	movaps	%xmm11, %xmm6	mulps	S, %xmm11	movaps	%xmm10, %xmm7	mulps	C, %xmm10	mulps	C, %xmm4	mulps	S, %xmm5	mulps	C, %xmm6	mulps	S, %xmm7	addps	%xmm9,  %xmm8	addps	%xmm11, %xmm10	subps	%xmm5,  %xmm4	subps	%xmm7,  %xmm6	movlps	%xmm8,  24 * SIZE(X)	movhps	%xmm8,  26 * SIZE(X)	movlps	%xmm10, 28 * SIZE(X)	movhps	%xmm10, 30 * SIZE(X)	movlps	%xmm4,  24 * SIZE(Y)	movhps	%xmm4,  26 * SIZE(Y)	movlps	%xmm6,  28 * SIZE(Y)	movhps	%xmm6,  30 * SIZE(Y)	addq	$32 * SIZE, X	addq	$32 * SIZE, Y	ALIGN_3.L34:	testq	$15, N	jle	.L999	testq	$8, N	jle	.L35	MOVSD	0 * SIZE(Y), %xmm1	movhps	2 * SIZE(Y), %xmm1	MOVSD	0 * SIZE(X), %xmm0	movhps	2 * SIZE(X), %xmm0	MOVSD	4 * SIZE(Y), %xmm3	movhps	6 * SIZE(Y), %xmm3	MOVSD	4 * SIZE(X), %xmm2	movhps	6 * SIZE(X), %xmm2	movaps	%xmm1, %xmm4	movaps	%xmm0, %xmm5	movaps	%xmm3, %xmm6	movaps	%xmm2, %xmm7	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	mulps	C, %xmm4	mulps	S, %xmm5	mulps	C, %xmm6	mulps	S, %xmm7	addps	%xmm1, %xmm0	addps	%xmm3, %xmm2	subps	%xmm5, %xmm4	subps	%xmm7, %xmm6	movlps	%xmm0, 0 * SIZE(X)	movhps	%xmm0, 2 * SIZE(X)	movlps	%xmm2, 4 * SIZE(X)	movhps	%xmm2, 6 * SIZE(X)	movlps	%xmm4, 0 * SIZE(Y)	movhps	%xmm4, 2 * SIZE(Y)	movlps	%xmm6, 4 * SIZE(Y)	movhps	%xmm6, 6 * SIZE(Y)	MOVSD	 8 * SIZE(Y), %xmm1	movhps	10 * SIZE(Y), %xmm1	MOVSD	 8 * SIZE(X), %xmm0	movhps	10 * SIZE(X), %xmm0	MOVSD	12 * SIZE(Y), %xmm3	movhps	14 * SIZE(Y), %xmm3	MOVSD	12 * SIZE(X), %xmm2	movhps	14 * SIZE(X), %xmm2	movaps	%xmm1, %xmm4	movaps	%xmm0, %xmm5	movaps	%xmm3, %xmm6	movaps	%xmm2, %xmm7	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	mulps	C, %xmm4	mulps	S, %xmm5	mulps	C, %xmm6	mulps	S, %xmm7	addps	%xmm1, %xmm0	addps	%xmm3, %xmm2	subps	%xmm5, %xmm4	subps	%xmm7, %xmm6	movlps	%xmm0,  8 * SIZE(X)	movhps	%xmm0, 10 * SIZE(X)	movlps	%xmm2, 12 * SIZE(X)	movhps	%xmm2, 14 * SIZE(X)	movlps	%xmm4,  8 * SIZE(Y)	movhps	%xmm4, 10 * SIZE(Y)	movlps	%xmm6, 12 * SIZE(Y)	movhps	%xmm6, 14 * SIZE(Y)	addq	$16 * SIZE, X	addq	$16 * SIZE, Y	ALIGN_3.L35:	testq	$4, N	jle	.L36	MOVSD	0 * SIZE(Y), %xmm1	movhps	2 * SIZE(Y), %xmm1	MOVSD	0 * SIZE(X), %xmm0	movhps	2 * SIZE(X), %xmm0	MOVSD	4 * SIZE(Y), %xmm3	movhps	6 * SIZE(Y), %xmm3	MOVSD	4 * SIZE(X), %xmm2	movhps	6 * SIZE(X), %xmm2	movaps	%xmm1, %xmm4	movaps	%xmm0, %xmm5	movaps	%xmm3, %xmm6	movaps	%xmm2, %xmm7	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	mulps	C, %xmm4	mulps	S, %xmm5	mulps	C, %xmm6	mulps	S, %xmm7	addps	%xmm1, %xmm0	addps	%xmm3, %xmm2	subps	%xmm5, %xmm4	subps	%xmm7, %xmm6	movlps	%xmm0, 0 * SIZE(X)	movhps	%xmm0, 2 * SIZE(X)	movlps	%xmm2, 4 * SIZE(X)	movhps	%xmm2, 6 * SIZE(X)	movlps	%xmm4, 0 * SIZE(Y)	movhps	%xmm4, 2 * SIZE(Y)	movlps	%xmm6, 4 * SIZE(Y)	movhps	%xmm6, 6 * SIZE(Y)	addq	$8 * SIZE, X	addq	$8 * SIZE, Y	ALIGN_3.L36:	testq	$2, N	jle	.L37	MOVSD	0 * SIZE(Y), %xmm1	movhps	2 * SIZE(Y), %xmm1	MOVSD	0 * SIZE(X), %xmm0	movhps	2 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, 0 * SIZE(X)	movhps	%xmm0, 2 * SIZE(X)	movlps	%xmm2, 0 * SIZE(Y)	movhps	%xmm2, 2 * SIZE(Y)	addq	$4 * SIZE, X	addq	$4 * SIZE, Y	ALIGN_3.L37:	testq	$1, N	jle	.L999	movsd	0 * SIZE(Y), %xmm1	movsd	0 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, 0 * SIZE(X)	movlps	%xmm2, 0 * SIZE(Y)	jmp	.L999	ALIGN_3	ALIGN_3.L50:	movq	N,  %rax	sarq	$2, %rax	jle	.L55	ALIGN_3.L53:	MOVSD	(Y), %xmm1	movhps	(Y, INCY), %xmm1	MOVSD	(X), %xmm0	movhps	(X, INCX), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, (X)	movhps	%xmm0, (X, INCX)	movlps	%xmm2, (Y)	movhps	%xmm2, (Y, INCY)	leaq	(X, INCX, 2), X	leaq	(Y, INCY, 2), Y	MOVSD	(Y), %xmm1	movhps	(Y, INCY), %xmm1	MOVSD	(X), %xmm0	movhps	(X, INCX), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, (X)	movhps	%xmm0, (X, INCX)	movlps	%xmm2, (Y)	movhps	%xmm2, (Y, INCY)	leaq	(X, INCX, 2), X	leaq	(Y, INCY, 2), Y	decq	%rax	jg	.L53	ALIGN_3.L55:	movq	N,  %rax	andq	$3, %rax	jle	.L999	ALIGN_3.L56:	movsd	(Y), %xmm1	movsd	(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, (X)	movlps	%xmm2, (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 + -