viscsumcopy.s

优龙2410linux2.6.8内核源代码

	brz,pn		%o4, 3f			/*  CTI+IEU1	Group		*/
	 sub		%g1, %o4, %g1		/*  IEU0			*/
	ldda		[%src] %asi, %f0	/*  Load			*/
	clr		%o4			/*  IEU0	Group		*/
	andcc		%dst, 8, %g0		/*  IEU1			*/
	be,pn		%icc, 1f		/*  CTI				*/
	 ldda		[%src + 8] %asi, %f2	/*  Load	Group		*/
	add		%src, 8, %src		/*  IEU0			*/
	sub		%len, 8, %len		/*  IEU1			*/
	fpadd32		%f0, %f48, %f50		/*  FPA				*/
	addcc		%dst, 8, %dst		/*  IEU1	Group		*/
	faligndata	%f0, %f2, %f16		/*  FPA				*/
	fcmpgt32	%f48, %f50, %o4		/*  FPM		Group		*/
	fmovd		%f2, %f0		/*  FPA		Group		*/
	ldda		[%src + 8] %asi, %f2	/*  Load			*/
	std		%f16, [%dst - 8]	/*  Store			*/
	fmovd		%f50, %f48		/*  FPA				*/
1:	andcc		%g1, 0x10, %g0		/*  IEU1	Group		*/
	be,pn		%icc, 1f		/*  CTI				*/
	 and		%g1, 0x20, %g1		/*  IEU0			*/
	fpadd32		%f0, %f48, %f50		/*  FPA				*/
	ldda		[%src + 16] %asi, %f4	/*  Load	Group		*/
	add		%src, 16, %src		/*  IEU0			*/
	add		%dst, 16, %dst		/*  IEU1			*/
	faligndata	%f0, %f2, %f16		/*  FPA				*/
	fcmpgt32	%f48, %f50, %g5		/*  FPM		Group		*/
	sub		%len, 16, %len		/*  IEU0			*/
	inc		%o4			/*  IEU1			*/
	std		%f16, [%dst - 16]	/*  Store	Group		*/
	fpadd32		%f2, %f50, %f48		/*  FPA				*/
	srl		%o4, 1, %o5		/*  IEU0			*/
	faligndata	%f2, %f4, %f18		/*  FPA		Group		*/
	std		%f18, [%dst - 8]	/*  Store			*/
	fcmpgt32	%f50, %f48, %o4		/*  FPM		Group		*/
	add		%o5, %sum, %sum		/*  IEU0			*/
	ldda		[%src + 8] %asi, %f2	/*  Load			*/
	fmovd		%f4, %f0		/*  FPA				*/
1:	brz,a,pn	%g1, 4f			/*  CTI+IEU1	Group		*/
	 rd		%asi, %g2		/*  LSU		Group + 4 bubbles*/
	inc		%g5			/*  IEU0			*/
	fpadd32		%f0, %f48, %f50		/*  FPA				*/
	ldda		[%src + 16] %asi, %f4	/*  Load	Group		*/
	srl		%g5, 1, %g5		/*  IEU0			*/
	add		%dst, 32, %dst		/*  IEU1			*/
	faligndata	%f0, %f2, %f16		/*  FPA				*/
	fcmpgt32	%f48, %f50, %o5		/*  FPM		Group		*/
	inc		%o4			/*  IEU0			*/
	ldda		[%src + 24] %asi, %f6	/*  Load			*/
	srl		%o4, 1, %o4		/*  IEU0	Group		*/
	add		%g5, %sum, %sum		/*  IEU1			*/
	ldda		[%src + 32] %asi, %f8	/*  Load			*/
	fpadd32		%f2, %f50, %f48		/*  FPA				*/
	faligndata	%f2, %f4, %f18		/*  FPA		Group		*/
	sub		%len, 32, %len		/*  IEU0			*/
	std		%f16, [%dst - 32]	/*  Store			*/
	fcmpgt32	%f50, %f48, %g3		/*  FPM		Group		*/
	inc		%o5			/*  IEU0			*/
	add		%o4, %sum, %sum		/*  IEU1			*/
	fpadd32		%f4, %f48, %f50		/*  FPA				*/
	faligndata	%f4, %f6, %f20		/*  FPA		Group		*/
	srl		%o5, 1, %o5		/*  IEU0			*/
	fcmpgt32	%f48, %f50, %g5		/*  FPM		Group		*/
	add		%o5, %sum, %sum		/*  IEU0			*/
	std		%f18, [%dst - 24]	/*  Store			*/
	fpadd32		%f6, %f50, %f48		/*  FPA				*/
	inc		%g3			/*  IEU0	Group		*/
	std		%f20, [%dst - 16]	/*  Store			*/
	add		%src, 32, %src		/*  IEU1			*/
	faligndata	%f6, %f8, %f22		/*  FPA				*/
	fcmpgt32	%f50, %f48, %o4		/*  FPM		Group		*/
	srl		%g3, 1, %g3		/*  IEU0			*/
	std		%f22, [%dst - 8]	/*  Store			*/	
	add		%g3, %sum, %sum		/*  IEU0	Group		*/
3:	rd		%asi, %g2		/*  LSU		Group + 4 bubbles*/
#ifdef __KERNEL__
4:	sethi		%hi(vis0s), %g7		/*  IEU0	Group		*/
	or		%g2, ASI_BLK_OR, %g2	/*  IEU1			*/
#else
4:	rd		%pc, %g7		/*  LSU		Group + 4 bubbles*/
#endif
	inc		%g5			/*  IEU0	Group		*/
	and		%src, 0x38, %g3		/*  IEU1			*/	
	membar		#StoreLoad		/*  LSU		Group		*/
	srl		%g5, 1, %g5		/*  IEU0			*/
	inc		%o4			/*  IEU1			*/
	sll		%g3, 8, %g3		/*  IEU0	Group		*/
	sub		%len, 0xc0, %len	/*  IEU1			*/
	addcc		%g5, %sum, %sum		/*  IEU1	Group		*/
	srl		%o4, 1, %o4		/*  IEU0			*/
	add		%g7, %g3, %g7		/*  IEU0	Group		*/
	add		%o4, %sum, %sum		/*  IEU1			*/
#ifdef __KERNEL__
	jmpl		%g7 + %lo(vis0s), %g0	/*  CTI+IEU1	Group		*/
#else
	jmpl		%g7 + (vis0s - 4b), %g0	/*  CTI+IEU1	Group		*/
#endif
	 fzero		%f32			/*  FPA				*/

	.align		2048
vis0s:	wr		%g2, ASI_BLK_XOR, %asi	/*  LSU		Group		*/
	add		%src, 128, %src		/*  IEU0	Group		*/
	ldda		[%src-128] %asi, %f0	/*  Load	Group		*/
	ldda		[%src-64] %asi, %f16	/*  Load	Group		*/
	fmovd		%f48, %f62		/*  FPA		Group	f0 available*/
	faligndata	%f0, %f2, %f48		/*  FPA		Group	f2 available*/
	fcmpgt32	%f32, %f2, %x1		/*  FPM		Group	f4 available*/
	fpadd32		%f0, %f62, %f0		/*  FPA				*/
	fcmpgt32	%f32, %f4, %x2		/*  FPM		Group	f6 available*/
	faligndata	%f2, %f4, %f50		/*  FPA				*/
	fcmpgt32	%f62, %f0, %x3		/*  FPM		Group	f8 available*/
	faligndata	%f4, %f6, %f52		/*  FPA				*/
	fcmpgt32	%f32, %f6, %x4		/*  FPM		Group	f10 available*/
	inc		%x1			/*  IEU0			*/
	faligndata	%f6, %f8, %f54		/*  FPA				*/
	fcmpgt32	%f32, %f8, %x5		/*  FPM		Group	f12 available*/
	srl		%x1, 1, %x1		/*  IEU0			*/
	inc		%x2			/*  IEU1			*/
	faligndata	%f8, %f10, %f56		/*  FPA				*/
	fcmpgt32	%f32, %f10, %x6		/*  FPM		Group	f14 available*/
	srl		%x2, 1, %x2		/*  IEU0			*/
	add		%sum, %x1, %sum		/*  IEU1			*/
	faligndata	%f10, %f12, %f58	/*  FPA				*/
	fcmpgt32	%f32, %f12, %x7		/*  FPM		Group		*/
	inc		%x3			/*  IEU0			*/
	add		%sum, %x2, %sum		/*  IEU1			*/
	faligndata	%f12, %f14, %f60	/*  FPA				*/
	fcmpgt32	%f32, %f14, %x8		/*  FPM		Group		*/
	srl		%x3, 1, %x3		/*  IEU0			*/
	inc		%x4			/*  IEU1			*/
	fmovd		%f14, %f62		/*  FPA				*/
	srl		%x4, 1, %x4		/*  IEU0	Group		*/
	add		%sum, %x3, %sum		/*  IEU1			*/
vis0:	DO_THE_TRICK(	f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,
			,f48,f50,f52,f54,f56,f58,f60,f62,f62,
			,LDBLK(f32),	STBLK,,,,,,,,
			,bcs,pn %icc, vis0e1)
	DO_THE_TRICK(	f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
			,f48,f50,f52,f54,f56,f58,f60,f62,f62,
			,LDBLK(f0),	STBLK,,,,,,,,
			,bcs,pn %icc, vis0e2)
	DO_THE_TRICK(	f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,
			,f48,f50,f52,f54,f56,f58,f60,f62,f62,
			,LDBLK(f16),	STBLK,,,,,,,,
			,bcc,pt %icc, vis0)
vis0e3:	DO_THE_TRICK(	f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,
			,f48,f50,f52,f54,f56,f58,f60,f62,f32,
			,SYNC,		STBLK,ST(f48,64),ST(f50,8),ST(f52,16),ST(f54,24),ST(f56,32),ST(f58,40),ST(f60,48),
			,add %dst, 56, %dst; add %len, 192 - 8*8, %len; ba,pt %icc, e2)
vis0e1:	DO_THE_TRICK(	f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
			,f48,f50,f52,f54,f56,f58,f60,f62,f0,
			,SYNC,		STBLK,ST(f48,64),ST(f50,8),ST(f52,16),ST(f54,24),ST(f56,32),ST(f58,40),ST(f60,48),
			,add %dst, 56, %dst; add %len, 192 - 8*8, %len; ba,pt %icc, e3)
vis0e2:	DO_THE_TRICK(	f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,
			,f48,f50,f52,f54,f56,f58,f60,f62,f16,
			,SYNC,		STBLK,ST(f48,64),ST(f50,8),ST(f52,16),ST(f54,24),ST(f56,32),ST(f58,40),ST(f60,48),
			,add %dst, 56, %dst; add %len, 192 - 8*8, %len; ba,pt %icc, e1)
	.align		2048
vis1s:	wr		%g2, ASI_BLK_XOR, %asi	/*  LSU		Group		*/
	add		%src, 128 - 8, %src	/*  IEU0	Group		*/
	ldda		[%src-128] %asi, %f0	/*  Load	Group		*/
	ldda		[%src-64] %asi, %f16	/*  Load	Group		*/
	fmovd		%f0, %f58		/*  FPA		Group		*/
	fmovd		%f48, %f0		/*  FPA		Group		*/
	fcmpgt32	%f32, %f2, %x2		/*  FPM		Group		*/
	faligndata	%f2, %f4, %f48		/*  FPA				*/
	fcmpgt32	%f32, %f4, %x3		/*  FPM		Group		*/
	faligndata	%f4, %f6, %f50		/*  FPA				*/
	fcmpgt32	%f32, %f6, %x4		/*  FPM		Group		*/
	faligndata	%f6, %f8, %f52		/*  FPA				*/
	fcmpgt32	%f32, %f8, %x5		/*  FPM		Group		*/
	inc		%x2			/*  IEU1			*/
	faligndata	%f8, %f10, %f54		/*  FPA				*/
	fcmpgt32	%f32, %f10, %x6		/*  FPM		Group		*/
	srl		%x2, 1, %x2		/*  IEU0			*/
	faligndata	%f10, %f12, %f56	/*  FPA				*/
	fcmpgt32	%f32, %f12, %x7		/*  FPM		Group		*/
	inc		%x3			/*  IEU0			*/
	add		%sum, %x2, %sum		/*  IEU1			*/
	faligndata	%f12, %f14, %f58	/*  FPA				*/
	fcmpgt32	%f32, %f14, %x8		/*  FPM		Group		*/
	srl		%x3, 1, %x3		/*  IEU0			*/
	inc		%x4			/*  IEU1			*/
	fmovd		%f14, %f60		/*  FPA				*/
	srl		%x4, 1, %x4		/*  IEU0	Group		*/
	add		%sum, %x3, %sum		/*  IEU1			*/
vis1:	DO_THE_TRICK(	f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,	
			,f62,f48,f50,f52,f54,f56,f58,f60,f60,
			,LDBLK(f32),	,STBLK,,,,,,,
			,bcs,pn %icc, vis1e1)
	DO_THE_TRICK(	f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
			,f62,f48,f50,f52,f54,f56,f58,f60,f60,
			,LDBLK(f0),	,STBLK,,,,,,,
			,bcs,pn %icc, vis1e2)
	DO_THE_TRICK(	f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,	
			,f62,f48,f50,f52,f54,f56,f58,f60,f60,
			,LDBLK(f16),	,STBLK,,,,,,,
			,bcc,pt %icc, vis1)
vis1e3:	DO_THE_TRICK(	f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,	
			,f62,f48,f50,f52,f54,f56,f58,f60,f32,
			,SYNC,		,STBLK,ST(f48,0),ST(f50,8),ST(f52,16),ST(f54,24),ST(f56,32),ST(f58,40),
			,add %dst, 48, %dst; add %len, 192 - 7*8, %len; ba,pt %icc, e2)
vis1e1:	DO_THE_TRICK(	f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
			,f62,f48,f50,f52,f54,f56,f58,f60,f0,
			,SYNC,		,STBLK,ST(f48,0),ST(f50,8),ST(f52,16),ST(f54,24),ST(f56,32),ST(f58,40),
			,add %dst, 48, %dst; add %len, 192 - 7*8, %len; ba,pt %icc, e3)
vis1e2:	DO_THE_TRICK(	f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,	
			,f62,f48,f50,f52,f54,f56,f58,f60,f16,
			,SYNC,		,STBLK,ST(f48,0),ST(f50,8),ST(f52,16),ST(f54,24),ST(f56,32),ST(f58,40),
			,add %dst, 48, %dst; add %len, 192 - 7*8, %len; ba,pt %icc, e1)
	.align		2048
vis2s:	wr		%g2, ASI_BLK_XOR, %asi	/*  LSU		Group		*/
	add		%src, 128 - 16, %src	/*  IEU0	Group		*/
	ldda		[%src-128] %asi, %f0	/*  Load	Group		*/
	ldda		[%src-64] %asi, %f16	/*  Load	Group		*/
	fmovd		%f0, %f56		/*  FPA		Group		*/
	fmovd		%f48, %f0		/*  FPA		Group		*/	
	sub		%dst, 64, %dst		/*  IEU0			*/
	fpsub32		%f2, %f2, %f2		/*  FPA		Group		*/
	fcmpgt32	%f32, %f4, %x3		/*  FPM		Group		*/
	faligndata	%f4, %f6, %f48		/*  FPA				*/
	fcmpgt32	%f32, %f6, %x4		/*  FPM		Group		*/
	faligndata	%f6, %f8, %f50		/*  FPA				*/
	fcmpgt32	%f32, %f8, %x5		/*  FPM		Group		*/
	faligndata	%f8, %f10, %f52		/*  FPA				*/
	fcmpgt32	%f32, %f10, %x6		/*  FPM		Group		*/
	faligndata	%f10, %f12, %f54	/*  FPA				*/
	fcmpgt32	%f32, %f12, %x7		/*  FPM		Group		*/
	inc		%x3			/*  IEU0			*/
	faligndata	%f12, %f14, %f56	/*  FPA				*/
	fcmpgt32	%f32, %f14, %x8		/*  FPM		Group		*/
	srl		%x3, 1, %x3		/*  IEU0			*/
	inc		%x4			/*  IEU1			*/
	fmovd		%f14, %f58		/*  FPA				*/
	srl		%x4, 1, %x4		/*  IEU0	Group		*/
	add		%sum, %x3, %sum		/*  IEU1			*/
vis2:	DO_THE_TRICK(	f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,	
			,f60,f62,f48,f50,f52,f54,f56,f58,f58,
			,LDBLK(f32),	,,STBLK,,,,,,
			,bcs,pn %icc, vis2e1)
	DO_THE_TRICK(	f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
			,f60,f62,f48,f50,f52,f54,f56,f58,f58,
			,LDBLK(f0),	,,STBLK,,,,,,
			,bcs,pn %icc, vis2e2)
	DO_THE_TRICK(	f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,	
			,f60,f62,f48,f50,f52,f54,f56,f58,f58,
			,LDBLK(f16),	,,STBLK,,,,,,
			,bcc,pt %icc, vis2)
vis2e3:	DO_THE_TRICK(	f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,	
			,f60,f62,f48,f50,f52,f54,f56,f58,f32,
			,SYNC,		,,STBLK,ST(f48,64),ST(f50,72),ST(f52,80),ST(f54,88),ST(f56,96),
			,add %dst, 104, %dst; add %len, 192 - 6*8, %len; ba,pt %icc, e2)
vis2e1:	DO_THE_TRICK(	f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
			,f60,f62,f48,f50,f52,f54,f56,f58,f0,
			,SYNC,		,,STBLK,ST(f48,64),ST(f50,72),ST(f52,80),ST(f54,88),ST(f56,96),
			,add %dst, 104, %dst; add %len, 192 - 6*8, %len; ba,pt %icc, e3)
vis2e2:	DO_THE_TRICK(	f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,	
			,f60,f62,f48,f50,f52,f54,f56,f58,f16,
			,SYNC,		,,STBLK,ST(f48,64),ST(f50,72),ST(f52,80),ST(f54,88),ST(f56,96),
			,add %dst, 104, %dst; add %len, 192 - 6*8, %len; ba,pt %icc, e1)
	.align		2048
vis3s:	wr		%g2, ASI_BLK_XOR, %asi	/*  LSU		Group		*/
	add		%src, 128 - 24, %src	/*  IEU0	Group		*/
	ldda		[%src-128] %asi, %f0	/*  Load	Group		*/
	ldda		[%src-64] %asi, %f16	/*  Load	Group		*/
	fmovd		%f0, %f54		/*  FPA		Group		*/
	fmovd		%f48, %f0		/*  FPA		Group		*/
	sub		%dst, 64, %dst		/*  IEU0			*/
	fpsub32		%f2, %f2, %f2		/*  FPA		Group		*/
	fpsub32		%f4, %f4, %f4		/*  FPA		Group		*/
	fcmpgt32	%f32, %f6, %x4		/*  FPM		Group		*/
	faligndata	%f6, %f8, %f48		/*  FPA				*/
	fcmpgt32	%f32, %f8, %x5		/*  FPM		Group		*/
	faligndata	%f8, %f10, %f50		/*  FPA				*/
	fcmpgt32	%f32, %f10, %x6		/*  FPM		Group		*/
	faligndata	%f10, %f12, %f52	/*  FPA				*/
	fcmpgt32	%f32, %f12, %x7		/*  FPM		Group		*/
	faligndata	%f12, %f14, %f54	/*  FPA				*/
	fcmpgt32	%f32, %f14, %x8		/*  FPM		Group		*/
	fmovd		%f14, %f56		/*  FPA				*/
	inc		%x4			/*  IEU0			*/
	srl		%x4, 1, %x4		/*  IEU0	Group		*/
vis3:	DO_THE_TRICK(	f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,	
			,f58,f60,f62,f48,f50,f52,f54,f56,f56,
			,LDBLK(f32),	,,,STBLK,,,,,
			,bcs,pn %icc, vis3e1)
	DO_THE_TRICK(	f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
			,f58,f60,f62,f48,f50,f52,f54,f56,f56,
			,LDBLK(f0),	,,,STBLK,,,,,
			,bcs,pn %icc, vis3e2)
	DO_THE_TRICK(	f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,	
			,f58,f60,f62,f48,f50,f52,f54,f56,f56,
			,LDBLK(f16),	,,,STBLK,,,,,
			,bcc,pt %icc, vis3)
vis3e3:	DO_THE_TRICK(	f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,	
			,f58,f60,f62,f48,f50,f52,f54,f56,f32,
			,SYNC,		,,,STBLK,ST(f48,64),ST(f50,72),ST(f52,80),ST(f54,88),
			,add %dst, 96, %dst; add %len, 192 - 5*8, %len; ba,pt %icc, e2)
vis3e1:	DO_THE_TRICK(	f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
			,f58,f60,f62,f48,f50,f52,f54,f56,f0,
			,SYNC,		,,,STBLK,ST(f48,64),ST(f50,72),ST(f52,80),ST(f54,88),