sparc/lib/NGmemcpy.S

42 #define LOAD(type,addr,dest)	type##a [addr] 0x80, dest
59 #define STORE_INIT(src,addr)	stx src, [addr + 0x00]
71 #ifndef XCC
72 #define XCC xcc  macro
86 	ba,pt	%xcc, __restore_asi
90 	ba,pt	%xcc, __restore_asi
95 	ba,pt	%xcc, __restore_asi
100 	ba,pt	%xcc, __restore_asi
105 	ba,pt	%xcc, __restore_asi
110 	ba,pt	%xcc, __restore_asi
115 	ba,pt	%xcc, __restore_asi
120 	ba,pt	%xcc, __restore_asi
125 	ba,pt	%xcc, __restore_asi
129 	ba,pt	%xcc, __restore_asi
134 	ba,pt	%xcc, __restore_asi
138 	ba,pt	%xcc, __restore_asi
142 	ba,pt	%xcc, __restore_asi
146 	ba,pt	%xcc, __restore_asi
151 	ba,pt	%xcc, __restore_asi
155 	ba,pt	%xcc, __restore_asi
160 	ba,pt	%xcc, __restore_asi
173 	cmp		%g2, 0
174 	tne		%xcc, 5
176 	cmp		%i2, 0
177 	be,pn		%XCC, 85f
180 	blu,a,pn	%XCC, 80f
191 	blu,pt		%XCC, 70f
192 	 andcc		%i3, 0x7, %g0
207 	be,pt		%XCC, 2f
215 	bne,pt		%XCC, 1b
235 	be,pt		%XCC, 50f
239 	be,pt		%XCC, 10f
243 	and		%i4, 0x7, GLOBAL_SPARE
253 	bg,pn	   	%XCC, 9f
268 	EX_ST(STORE_INIT(%g2, %o0 + 0x00), NG_ret_i2_plus_g1)
269 	EX_ST(STORE_INIT(%g3, %o0 + 0x08), NG_ret_i2_plus_g1_minus_8)
274 	EX_ST(STORE_INIT(%o2, %o0 + 0x10), NG_ret_i2_plus_g1_minus_16)
275 	EX_ST(STORE_INIT(%o3, %o0 + 0x18), NG_ret_i2_plus_g1_minus_24)
280 	EX_ST(STORE_INIT(%g2, %o0 + 0x20), NG_ret_i2_plus_g1_minus_32)
281 	EX_ST(STORE_INIT(%g3, %o0 + 0x28), NG_ret_i2_plus_g1_minus_40)
287 	EX_ST(STORE_INIT(%o2, %o0 + 0x30), NG_ret_i2_plus_g1_minus_48)
288 	EX_ST(STORE_INIT(%o3, %o0 + 0x38), NG_ret_i2_plus_g1_minus_56)
291 	bne,pt		%XCC, 8b
294 	ba,pt		%XCC, 60f
301 	EX_ST(STORE_INIT(%g3, %o0 + 0x00), NG_ret_i2_plus_g1)
302 	EX_ST(STORE_INIT(%o2, %o0 + 0x08), NG_ret_i2_plus_g1_minus_8)
307 	EX_ST(STORE_INIT(%o3, %o0 + 0x10), NG_ret_i2_plus_g1_minus_16)
308 	EX_ST(STORE_INIT(%g2, %o0 + 0x18), NG_ret_i2_plus_g1_minus_24)
313 	EX_ST(STORE_INIT(%g3, %o0 + 0x20), NG_ret_i2_plus_g1_minus_32)
314 	EX_ST(STORE_INIT(%o2, %o0 + 0x28), NG_ret_i2_plus_g1_minus_40)
320 	EX_ST(STORE_INIT(%o3, %o0 + 0x30), NG_ret_i2_plus_g1_minus_48)
321 	EX_ST(STORE_INIT(%g2, %o0 + 0x38), NG_ret_i2_plus_g1_minus_56)
324 	bne,pt		%XCC, 9b
327 	ba,pt		%XCC, 60f
342 	EX_ST(STORE_INIT(%o5, %o0 + 0x00), NG_ret_i2_plus_g1)	! initializes cache line
343 	EX_ST(STORE_INIT(%o2, %o0 + 0x08), NG_ret_i2_plus_g1_minus_8)
345 	EX_ST(STORE_INIT(%o3, %o0 + 0x10), NG_ret_i2_plus_g1_minus_16)
346 	EX_ST(STORE_INIT(%o4, %o0 + 0x18), NG_ret_i2_plus_g1_minus_24)
348 	EX_ST(STORE_INIT(%o5, %o0 + 0x20), NG_ret_i2_plus_g1_minus_32)
349 	EX_ST(STORE_INIT(%o2, %o0 + 0x28), NG_ret_i2_plus_g1_minus_40)
352 	EX_ST(STORE_INIT(%o3, %o0 + 0x30), NG_ret_i2_plus_g1_minus_48)
353 	EX_ST(STORE_INIT(%o4, %o0 + 0x38), NG_ret_i2_plus_g1_minus_56)
355 	bne,pt		%XCC, 1b
358 	ba,pt		%XCC, 60f
359 	 add		%i1, 0x8, %i1
371 	EX_ST(STORE_INIT(%o4, %o0 + 0x00), NG_ret_i2_plus_g1)	! initializes cache line
372 	EX_ST(STORE_INIT(%o5, %o0 + 0x08), NG_ret_i2_plus_g1_minus_8)
374 	EX_ST(STORE_INIT(%o2, %o0 + 0x10), NG_ret_i2_plus_g1_minus_16)
375 	EX_ST(STORE_INIT(%o3, %o0 + 0x18), NG_ret_i2_plus_g1_minus_24)
378 	EX_ST(STORE_INIT(%o4, %o0 + 0x20), NG_ret_i2_plus_g1_minus_32)
379 	EX_ST(STORE_INIT(%o5, %o0 + 0x28), NG_ret_i2_plus_g1_minus_40)
380 	EX_ST(STORE_INIT(%o2, %o0 + 0x30), NG_ret_i2_plus_g1_minus_48)
381 	EX_ST(STORE_INIT(%o3, %o0 + 0x38), NG_ret_i2_plus_g1_minus_56)
383 	bne,pt	%XCC, 1b
396 	ba,a,pt		%XCC, 90f
401 	bne,pn		%XCC, 75f
405 	andn		%i2, 0xf, %i4
406 	and		%i2, 0xf, %i2
407 1:	subcc		%i4, 0x10, %i4
409 	add		%i1, 0x08, %i1
411 	sub		%i1, 0x08, %i1
413 	add		%i1, 0x8, %i1
415 	bgu,pt		%XCC, 1b
416 	 add		%i1, 0x8, %i1
417 73:	andcc		%i2, 0x8, %g0
418 	be,pt		%XCC, 1f
420 	sub		%i2, 0x8, %i2
423 	add		%i1, 0x8, %i1
424 1:	andcc		%i2, 0x4, %g0
425 	be,pt		%XCC, 1f
427 	sub		%i2, 0x4, %i2
430 	add		%i1, 0x4, %i1
431 1:	cmp		%i2, 0
432 	be,pt		%XCC, 85f
434 	ba,pt		%xcc, 90f
438 	andcc		%o0, 0x7, %g1
439 	sub		%g1, 0x8, %g1
451 	andcc		%i1, 0x7, %g1
458 	ba,a,pt		%xcc, 73b
461 	andn		%i1, 0x7, %i1
464 	andn		%i2, 0x7, %i4
466 1:	add		%i1, 0x8, %i1
468 	subcc		%i4, 0x8, %i4
472 	add		%o0, 0x8, %o0
477 	andcc		%i2, 0x7, %i2
480 	ba,pt		%xcc, 90f
484 80: /* 0 < len <= 16 */
485 	andcc		%i3, 0x3, %g0
486 	bne,pn		%XCC, 90f
493 	bgu,pt		%XCC, 1b
504 	bgu,pt		%XCC, 90b