sparc/lib/U1memcpy.S

16 #define ASI_BLK_P 0xf0
17 #define FPRS_FEF  0x04
21 #define VISExit and %o5, FPRS_FEF, %o5; wr %o5, 0x0, %fprs
24 #define VISExit and %o5, FPRS_FEF, %o5; wr %o5, 0x0, %fprs
66 #ifndef XCC
67 #define XCC xcc  macro
83 	add			%src, 0x40, %src;			\
84 	subcc			%GLOBAL_SPARE, 0x40, %GLOBAL_SPARE;	\
85 	be,pn			%xcc, jmptgt;				\
86 	 add			%dest, 0x40, %dest;			\
98 	add			%dest, 0x40, %dest;	\
103 	add			%dest, 0x40, %dest;	\
104 	ba,pt			%xcc, target;		\
109 	bl,pn			%xcc, 95f;		\
116 	bl,pn			%xcc, 95f;	\
121 	ba,a,pt			%xcc, 93f;
155 	add		%GLOBAL_SPARE, 0x80, %GLOBAL_SPARE
162 	add		%GLOBAL_SPARE, 0x40, %GLOBAL_SPARE
197 	 add		%GLOBAL_SPARE, 0x8, %o0
203 	 add		%GLOBAL_SPARE, 0x10, %o0
249 	cmp		%g2, 0
250 	tne		%xcc, 5
253 	cmp		%o2, 0
254 	be,pn		%XCC, 85f
257 	blu,a,pn	%XCC, 80f
261 	blu,pt		%XCC, 70f
262 	 andcc		%o3, 0x7, %g0
264 	/* Clobbers o5/g1/g2/g3/g7/icc/xcc.  */
268 	andcc		%o0, 0x3f, %g2
269 	be,pt		%XCC, 2f
271 	/* Compute abs((dst & 0x3f) - 0x40) into %g2.  This is the number
276 	sub		%g2, 0x40, %g2
279 	andcc		%g2, 0x7, %g1
281 	 and		%g2, 0x38, %g2
283 1:	subcc		%g1, 0x1, %g1
284 	EX_LD_FP(LOAD(ldub, %o1 + 0x00, %o3), U1_g1_1_fp)
286 	bgu,pt		%XCC, 1b
287 	 add		%o1, 0x1, %o1
291 2:	cmp		%g2, 0x0
292 	and		%o1, 0x7, %g1
297 1:	EX_LD_FP(LOAD(ldd, %o1 + 0x8, %f6), U1_g2_0_fp)
298 	add		%o1, 0x8, %o1
299 	subcc		%g2, 0x8, %g2
303 	 add		%o0, 0x8, %o0
305 	EX_LD_FP(LOAD(ldd, %o1 + 0x8, %f4), U1_g2_0_fp)
306 	add		%o1, 0x8, %o1
307 	subcc		%g2, 0x8, %g2
311 	 add		%o0, 0x8, %o0
317 	subcc		%o2, 0x40, %GLOBAL_SPARE
319 	andncc		%GLOBAL_SPARE, (0x40 - 1), %GLOBAL_SPARE
322 	andn		%o1, (0x40 - 1), %o1
324 	andncc		%g3, 0x7, %g3
326 	sub		%g3, 0x8, %g3
333 	add		%o1, 0x40, %o1
336 	add		%o1, 0x40, %o1
337 	sub		%GLOBAL_SPARE, 0x80, %GLOBAL_SPARE
339 	add		%o1, 0x40, %o1
363 	ba,pt		%xcc, 1b+4
384 	ba,pt		%xcc, 1b+4
405 	ba,pt		%xcc, 1b+4
426 	ba,pt		%xcc, 1b+4
447 	ba,pt		%xcc, 1b+4
468 	ba,pt		%xcc, 1b+4
489 	ba,pt		%xcc, 1b+4
510 	ba,pt		%xcc, 1b+4
555 	bl,pn		%xcc, 95f
562 	bge,pt		%xcc, 93b
572 	bne,pt		%xcc, 1b
582 	bne,pn		%XCC, 75f
585 72:	andn		%o2, 0xf, %GLOBAL_SPARE
586 	and		%o2, 0xf, %o2
587 1:	EX_LD(LOAD(ldx, %o1 + 0x00, %o5), U1_gs_0)
588 	EX_LD(LOAD(ldx, %o1 + 0x08, %g1), U1_gs_0)
589 	subcc		%GLOBAL_SPARE, 0x10, %GLOBAL_SPARE
591 	add		%o1, 0x8, %o1
593 	bgu,pt		%XCC, 1b
594 	 add		%o1, 0x8, %o1
595 73:	andcc		%o2, 0x8, %g0
596 	be,pt		%XCC, 1f
599 	sub		%o2, 0x8, %o2
601 	add		%o1, 0x8, %o1
602 1:	andcc		%o2, 0x4, %g0
603 	be,pt		%XCC, 1f
606 	sub		%o2, 0x4, %o2
608 	add		%o1, 0x4, %o1
609 1:	cmp		%o2, 0
610 	be,pt		%XCC, 85f
612 	ba,pt		%xcc, 90f
615 75:	andcc		%o0, 0x7, %g1
616 	sub		%g1, 0x8, %g1
628 	andcc		%o1, 0x7, %g1
635 	ba,a,pt		%xcc, 73b
638 	andn		%o1, 0x7, %o1
641 	andn		%o2, 0x7, %GLOBAL_SPARE
643 1:	EX_LD(LOAD(ldx, %o1 + 0x8, %g3), U1_gs_0_o2_adj)
644 	subcc		%GLOBAL_SPARE, 0x8, %GLOBAL_SPARE
645 	add		%o1, 0x8, %o1
649 	add		%o0, 0x8, %o0
654 	andcc		%o2, 0x7, %o2
657 	ba,pt		%xcc, 90f
661 80:	/* 0 < len <= 16 */
662 	andcc		%o3, 0x3, %g0
663 	bne,pn		%XCC, 90f
669 	bgu,pt		%XCC, 1b
679 	bgu,pt		%XCC, 90b