viscsumcopy.s

优龙2410linux2.6.8内核源代码

			,add %dst, 96, %dst; add %len, 192 - 5*8, %len; ba,pt %icc, e3)
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,f16,
			,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, e1)
	.align		2048
vis4s:	wr		%g2, ASI_BLK_XOR, %asi	/*  LSU		Group		*/
	add		%src, 128 - 32, %src	/*  IEU0	Group		*/
	ldda		[%src-128] %asi, %f0	/*  Load	Group		*/
	ldda		[%src-64] %asi, %f16	/*  Load	Group		*/
	fmovd		%f0, %f52		/*  FPA		Group		*/
	fmovd		%f48, %f0		/*  FPA		Group		*/
	sub		%dst, 64, %dst		/*  IEU0			*/
	fpsub32		%f2, %f2, %f2		/*  FPA		Group		*/
	fpsub32		%f4, %f4, %f4		/*  FPA		Group		*/
	fpsub32		%f6, %f6, %f6		/*  FPA		Group		*/
	clr		%x4			/*  IEU0			*/
	fcmpgt32	%f32, %f8, %x5		/*  FPM		Group		*/
	faligndata	%f8, %f10, %f48		/*  FPA				*/
	fcmpgt32	%f32, %f10, %x6		/*  FPM		Group		*/
	faligndata	%f10, %f12, %f50	/*  FPA				*/
	fcmpgt32	%f32, %f12, %x7		/*  FPM		Group		*/
	faligndata	%f12, %f14, %f52	/*  FPA				*/
	fcmpgt32	%f32, %f14, %x8		/*  FPM		Group		*/
	fmovd		%f14, %f54		/*  FPA				*/
vis4:	DO_THE_TRICK(	f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,	
			,f56,f58,f60,f62,f48,f50,f52,f54,f54,
			,LDBLK(f32),	,,,,STBLK,,,,
			,bcs,pn %icc, vis4e1)
	DO_THE_TRICK(	f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
			,f56,f58,f60,f62,f48,f50,f52,f54,f54,
			,LDBLK(f0),	,,,,STBLK,,,,
			,bcs,pn %icc, vis4e2)
	DO_THE_TRICK(	f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,	
			,f56,f58,f60,f62,f48,f50,f52,f54,f54,
			,LDBLK(f16),	,,,,STBLK,,,,
			,bcc,pt %icc, vis4)
vis4e3:	DO_THE_TRICK(	f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,	
			,f56,f58,f60,f62,f48,f50,f52,f54,f32,
			,SYNC,		,,,,STBLK,ST(f48,64),ST(f50,72),ST(f52,80),
			,add %dst, 88, %dst; add %len, 192 - 4*8, %len; ba,pt %icc, e2)
vis4e1:	DO_THE_TRICK(	f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
			,f56,f58,f60,f62,f48,f50,f52,f54,f0,
			,SYNC,		,,,,STBLK,ST(f48,64),ST(f50,72),ST(f52,80),
			,add %dst, 88, %dst; add %len, 192 - 4*8, %len; ba,pt %icc, e3)
vis4e2:	DO_THE_TRICK(	f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,	
			,f56,f58,f60,f62,f48,f50,f52,f54,f16,
			,SYNC,		,,,,STBLK,ST(f48,64),ST(f50,72),ST(f52,80),
			,add %dst, 88, %dst; add %len, 192 - 4*8, %len; ba,pt %icc, e1)
	.align		2048
vis5s:	add		%src, 128 - 40, %src	/*  IEU0	Group		*/
	ldda		[%src-88] %asi, %f10	/*  Load	Group		*/
	ldda		[%src-80] %asi, %f12	/*  Load	Group		*/
	ldda		[%src-72] %asi, %f14	/*  Load	Group		*/
	wr		%g2, ASI_BLK_XOR, %asi	/*  LSU		Group		*/
	ldda		[%src-64] %asi, %f16	/*  Load	Group		*/
	fmovd		%f48, %f0		/*  FPA		Group		*/
	fmuld		%f32, %f32, %f2		/*  FPM				*/
	clr		%x4			/*  IEU0			*/
	faddd		%f32, %f32, %f4		/*  FPA		Group		*/
	fmuld		%f32, %f32, %f6		/*  FPM				*/
	clr		%x5			/*  IEU0			*/
	faddd		%f32, %f32, %f8		/*  FPA		Group		*/
	fcmpgt32	%f32, %f10, %x6		/*  FPM		Group		*/
	sub		%dst, 64, %dst		/*  IEU0			*/
	faligndata	%f10, %f12, %f48	/*  FPA				*/
	fcmpgt32	%f32, %f12, %x7		/*  FPM		Group		*/
	faligndata	%f12, %f14, %f50	/*  FPA				*/
	fcmpgt32	%f32, %f14, %x8		/*  FPM		Group		*/
	fmovd		%f14, %f52		/*  FPA				*/
vis5:	DO_THE_TRICK(	f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,	
			,f54,f56,f58,f60,f62,f48,f50,f52,f52,
			,LDBLK(f32),	,,,,,STBLK,,,
			,bcs,pn %icc, vis5e1)
	DO_THE_TRICK(	f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
			,f54,f56,f58,f60,f62,f48,f50,f52,f52,
			,LDBLK(f0),	,,,,,STBLK,,,
			,bcs,pn %icc, vis5e2)
	DO_THE_TRICK(	f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,	
			,f54,f56,f58,f60,f62,f48,f50,f52,f52,
			,LDBLK(f16),	,,,,,STBLK,,,
			,bcc,pt %icc, vis5)
vis5e3:	DO_THE_TRICK(	f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,	
			,f54,f56,f58,f60,f62,f48,f50,f52,f32,
			,SYNC,		,,,,,STBLK,ST(f48,64),ST(f50,72),
			,add %dst, 80, %dst; add %len, 192 - 3*8, %len; ba,pt %icc, e2)
vis5e1:	DO_THE_TRICK(	f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
			,f54,f56,f58,f60,f62,f48,f50,f52,f0,
			,SYNC,		,,,,,STBLK,ST(f48,64),ST(f50,72),
			,add %dst, 80, %dst; add %len, 192 - 3*8, %len; ba,pt %icc, e3)
vis5e2:	DO_THE_TRICK(	f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,	
			,f54,f56,f58,f60,f62,f48,f50,f52,f16,
			,SYNC,		,,,,,STBLK,ST(f48,64),ST(f50,72),
			,add %dst, 80, %dst; add %len, 192 - 3*8, %len; ba,pt %icc, e1)
	.align		2048
vis6s:	add		%src, 128 - 48, %src	/*  IEU0	Group		*/
	ldda		[%src-80] %asi, %f12	/*  Load	Group		*/
	ldda		[%src-72] %asi, %f14	/*  Load	Group		*/
	wr		%g2, ASI_BLK_XOR, %asi	/*  LSU		Group		*/
	ldda		[%src-64] %asi, %f16	/*  Load	Group		*/
	fmovd		%f48, %f0		/*  FPA		Group		*/
	fmuld		%f32, %f32, %f2		/*  FPM				*/
	clr		%x4			/*  IEU0			*/
	faddd		%f32, %f32, %f4		/*  FPA		Group		*/
	fmuld		%f32, %f32, %f6		/*  FPM				*/
	clr		%x5			/*  IEU0			*/
	faddd		%f32, %f32, %f8		/*  FPA		Group		*/
	fmuld		%f32, %f32, %f10	/*  FPM				*/
	clr		%x6			/*  IEU0			*/
	fcmpgt32	%f32, %f12, %x7		/*  FPM		Group		*/
	sub		%dst, 64, %dst		/*  IEU0			*/
	fcmpgt32	%f32, %f14, %x8		/*  FPM		Group		*/
	faligndata	%f12, %f14, %f48	/*  FPA				*/
	fmovd		%f14, %f50		/*  FPA		Group		*/
vis6:	DO_THE_TRICK(	f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,	
			,f52,f54,f56,f58,f60,f62,f48,f50,f50,
			,LDBLK(f32),	,,,,,,STBLK,,
			,bcs,pn %icc, vis6e1)
	DO_THE_TRICK(	f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
			,f52,f54,f56,f58,f60,f62,f48,f50,f50,
			,LDBLK(f0),	,,,,,,STBLK,,
			,bcs,pn %icc, vis6e2)
	DO_THE_TRICK(	f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,	
			,f52,f54,f56,f58,f60,f62,f48,f50,f50,
			,LDBLK(f16),	,,,,,,STBLK,,
			,bcc,pt %icc, vis6)
vis6e3:	DO_THE_TRICK(	f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,	
			,f52,f54,f56,f58,f60,f62,f48,f50,f32,
			,SYNC,		,,,,,,STBLK,ST(f48,64),
			,add %dst, 72, %dst; add %len, 192 - 2*8, %len; ba,pt %icc, e2)
vis6e1:	DO_THE_TRICK(	f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
			,f52,f54,f56,f58,f60,f62,f48,f50,f0,
			,SYNC,		,,,,,,STBLK,ST(f48,64),
			,add %dst, 72, %dst; add %len, 192 - 2*8, %len; ba,pt %icc, e3)
vis6e2:	DO_THE_TRICK(	f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,	
			,f52,f54,f56,f58,f60,f62,f48,f50,f16,
			,SYNC,		,,,,,,STBLK,ST(f48,64),
			,add %dst, 72, %dst; add %len, 192 - 2*8, %len; ba,pt %icc, e1)
	.align		2048
vis7s:	add		%src, 128 - 56, %src	/*  IEU0	Group		*/
	ldda		[%src-72] %asi, %f14	/*  Load	Group		*/
	wr		%g2, ASI_BLK_XOR, %asi	/*  LSU		Group		*/
	ldda		[%src-64] %asi, %f16	/*  Load	Group		*/
	fmovd		%f48, %f0		/*  FPA		Group		*/
	fmuld		%f32, %f32, %f2		/*  FPM				*/
	clr		%x4			/*  IEU0			*/
	faddd		%f32, %f32, %f4		/*  FPA		Group		*/
	fmuld		%f32, %f32, %f6		/*  FPM				*/
	clr		%x5			/*  IEU0			*/
	faddd		%f32, %f32, %f8		/*  FPA		Group		*/
	fmuld		%f32, %f32, %f10	/*  FPM				*/
	clr		%x6			/*  IEU0			*/
	faddd		%f32, %f32, %f12	/*  FPA		Group		*/
	clr		%x7			/*  IEU0			*/
	fcmpgt32	%f32, %f14, %x8		/*  FPM		Group		*/
	sub		%dst, 64, %dst		/*  IEU0			*/
	fmovd		%f14, %f48		/*  FPA				*/
vis7:	DO_THE_TRICK(	f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,	
			,f50,f52,f54,f56,f58,f60,f62,f48,f48,
			,LDBLK(f32),	,,,,,,,STBLK,
			,bcs,pn %icc, vis7e1)
	DO_THE_TRICK(	f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
			,f50,f52,f54,f56,f58,f60,f62,f48,f48,
			,LDBLK(f0),	,,,,,,,STBLK,
			,bcs,pn %icc, vis7e2)
	DO_THE_TRICK(	f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,	
			,f50,f52,f54,f56,f58,f60,f62,f48,f48,
			,LDBLK(f16),	,,,,,,,STBLK,
			,bcc,pt %icc, vis7)
vis7e3:	DO_THE_TRICK(	f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,	
			,f50,f52,f54,f56,f58,f60,f62,f48,f32,
			,SYNC,		,,,,,,,STBLK,
			,add %dst, 64, %dst; add %len, 192 - 1*8, %len; ba,pt %icc, e2)
vis7e1:	DO_THE_TRICK(	f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
			,f50,f52,f54,f56,f58,f60,f62,f48,f0,
			,SYNC,		,,,,,,,STBLK,
			,add %dst, 64, %dst; add %len, 192 - 1*8, %len; ba,pt %icc, e3)
vis7e2:	DO_THE_TRICK(	f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,	
			,f50,f52,f54,f56,f58,f60,f62,f48,f16,
			,SYNC,		,,,,,,,STBLK,
			,add %dst, 64, %dst; add %len, 192 - 1*8, %len; ba,pt %icc, e1)
e1:	END_THE_TRICK1(	f0,f2,f4,f6,f8,f10,f12,f14,f16,f6)
e2:	END_THE_TRICK1(	f16,f18,f20,f22,f24,f26,f28,f30,f32,f6)
e3:	END_THE_TRICK1(	f32,f34,f36,f38,f40,f42,f44,f46,f0,f6)
ett:	rd		%asi, %x4		/*  LSU		Group+4bubbles	*/
	rd		%gsr, %x3		/*  LSU		Group+4bubbles	*/
#ifdef __KERNEL__
	srl		%x4, 3, %x5		/*  IEU0	Group		*/
	xor		%x4, ASI_BLK_XOR1, %x4	/*  IEU1			*/
	wr		%x4, %x5, %asi		/*  LSU		Group+4bubbles	*/
#else
	wr		%x4, ASI_BLK_XOR, %asi	/*  LSU		Group+4bubbles	*/
#endif
	andcc		%x3, 7, %x3		/*  IEU1	Group		*/
	add		%dst, 8, %dst		/*  IEU0			*/
	bne,pn		%icc, 1f		/*  CTI				*/
	 fzero		%f10			/*  FPA				*/
	brz,a,pn	%len, 2f		/*  CTI+IEU1	Group		*/
	 std		%f6, [%dst - 8]		/*  Store			*/
1:	cmp		%len, 8			/*  IEU1			*/
	blu,pn		%icc, 3f		/*  CTI				*/
	 sub		%src, 64, %src		/*  IEU0	Group		*/
1:	ldda		[%src] %asi, %f2	/*  Load	Group		*/
	fpadd32		%f10, %f2, %f12		/*  FPA		Group+load stall*/
	add		%src, 8, %src		/*  IEU0			*/
	add		%dst, 8, %dst		/*  IEU1			*/
	faligndata	%f6, %f2, %f14		/*  FPA		Group		*/
	fcmpgt32	%f10, %f12, %x5		/*  FPM		Group		*/
	std		%f14, [%dst - 16]	/*  Store			*/
	fmovd		%f2, %f6		/*  FPA				*/
	fmovd		%f12, %f10		/*  FPA		Group		*/
	sub		%len, 8, %len		/*  IEU1			*/
	fzero		%f16			/*  FPA		Group - FPU nop	*/
	fzero		%f18			/*  FPA		Group - FPU nop	*/
	inc		%x5			/*  IEU0			*/
	srl		%x5, 1, %x5		/*  IEU0	Group (regdep)	*/
	cmp		%len, 8			/*  IEU1			*/
	bgeu,pt		%icc, 1b		/*  CTI				*/
	 add		%x5, %sum, %sum		/*  IEU0	Group		*/
3:	brz,a,pt	%x3, 2f			/*  CTI+IEU1			*/
	 std		%f6, [%dst - 8]		/*  Store	Group		*/
	st		%f7, [%dst - 8]		/*  Store	Group		*/
	sub		%dst, 4, %dst		/*  IEU0			*/
	add		%len, 4, %len		/*  IEU1			*/
2:
#ifdef __KERNEL__
	sub		%sp, 8, %sp		/*  IEU0	Group		*/
#endif
	END_THE_TRICK2(	f48,f50,f52,f54,f56,f58,f60,f10,f12,f62)
	membar		#Sync			/*  LSU		Group		*/
#ifdef __KERNEL__
	VISExit
	add		%sp, 8, %sp		/*  IEU0	Group		*/
#endif
23:	brnz,pn		%len, 26f		/*  CTI+IEU1	Group		*/
24:	 sllx		%sum, 32, %g1		/*  IEU0			*/
25:	addcc		%sum, %g1, %src		/*  IEU1	Group		*/
	srlx		%src, 32, %src		/*  IEU0	Group (regdep)	*/
	bcs,a,pn	%xcc, 1f		/*  CTI				*/
	 add		%src, 1, %src		/*  IEU1			*/
#ifndef __KERNEL__
1:	retl					/*  CTI		Group brk forced*/
	 srl		%src, 0, %src		/*  IEU0			*/
#else
1:	retl					/*  CTI		Group brk forced*/
	 ldx		[%g6 + TI_TASK], %g4	/*  Load			*/
#endif
26:	andcc		%len, 8, %g0		/*  IEU1	Group		*/
	be,pn		%icc, 1f		/*  CTI				*/
	 lduwa		[%src] %asi, %o4	/*  Load			*/
	lduwa		[%src+4] %asi, %g2	/*  Load	Group		*/
	add		%src, 8, %src		/*  IEU0			*/
	add		%dst, 8, %dst		/*  IEU1			*/
	sllx		%o4, 32, %g5		/*  IEU0	Group		*/
	stw		%o4, [%dst - 8]		/*  Store			*/
	or		%g5, %g2, %g5		/*  IEU0	Group		*/
	stw		%g2, [%dst - 4]		/*  Store			*/
	addcc		%g5, %sum, %sum		/*  IEU1	Group		*/
	bcs,a,pn	%xcc, 1f		/*  CTI				*/
	 add		%sum, 1, %sum		/*  IEU0			*/
1:	andcc		%len, 4, %g0		/*  IEU1	Group		*/
	be,a,pn		%icc, 1f		/*  CTI				*/
	 clr		%g2			/*  IEU0			*/
	lduwa		[%src] %asi, %g7	/*  Load			*/
	add		%src, 4, %src		/*  IEU0	Group		*/
	add		%dst, 4, %dst		/*  IEU1			*/
	sllx		%g7, 32, %g2		/*  IEU0	Group		*/
	stw		%g7, [%dst - 4]		/*  Store			*/
1:	andcc		%len, 2, %g0		/*  IEU1			*/
	be,a,pn		%icc, 1f		/*  CTI				*/
	 clr		%g3			/*  IEU0	Group		*/
	lduha		[%src] %asi, %g7	/*  Load			*/
	add		%src, 2, %src		/*  IEU1			*/
	add		%dst, 2, %dst		/*  IEU0	Group		*/
	sll		%g7, 16, %g3		/*  IEU0	Group		*/
	sth		%g7, [%dst - 2]		/*  Store			*/
1:	andcc		%len, 1, %g0		/*  IEU1			*/
	be,a,pn		%icc, 1f		/*  CTI				*/
	 clr		%o5			/*  IEU0	Group		*/
	lduba		[%src] %asi, %g7	/*  Load			*/
	sll		%g7, 8, %o5		/*  IEU0	Group		*/
	stb		%g7, [%dst]		/*  Store			*/
1:	or		%g2, %g3, %g3		/*  IEU1			*/
	or		%o5, %g3, %g3		/*  IEU0	Group (regdep)	*/
	addcc		%g3, %sum, %sum		/*  IEU1	Group (regdep)	*/
	bcs,a,pn	%xcc, 1f		/*  CTI				*/
	 add		%sum, 1, %sum		/*  IEU0			*/
1:	ba,pt		%xcc, 25b		/*  CTI		Group		*/
	 sllx		%sum, 32, %g1		/*  IEU0			*/

#ifdef __KERNEL__
end:

	.section	__ex_table
	.align		4
	.word		csum_partial_copy_vis, 0, end, cpc_handler
#endif