microblaze/lib/fastcopy.S

43 	addi	r4, r0, 4	/* n = 4 */
51 	/* n = 4 - n (yields 3, 2, 1 transfers for 1, 2, 3 addr offset) */
52 	rsubi	r4, r4, 4
81 	lwi	r10, r6, 4		/* t2 = *(s + 4) */
85 	swi	r10, r5, 4		/* *(d + 4) = t2 */
115 	lwi	r12, r8, 4	/* v = *(as + 4) */
116 	bsrli	r9, r12, 8	/* t1 = v >> 8 */
119 	bslli	r11, r12, 24	/* h = v << 24 */
120 	lwi	r12, r8, 8	/* v = *(as + 8) */
121 	bsrli	r9, r12, 8	/* t1 = v >> 8 */
123 	swi	r9, r5, 4	/* *(d + 4) = t1 */
124 	bslli	r11, r12, 24	/* h = v << 24 */
125 	lwi	r12, r8, 12	/* v = *(as + 12) */
126 	bsrli	r9, r12, 8	/* t1 = v >> 8 */
129 	bslli	r11, r12, 24	/* h = v << 24 */
130 	lwi	r12, r8, 16	/* v = *(as + 16) */
131 	bsrli	r9, r12, 8	/* t1 = v >> 8 */
134 	bslli	r11, r12, 24	/* h = v << 24 */
135 	lwi	r12, r8, 20	/* v = *(as + 20) */
136 	bsrli	r9, r12, 8	/* t1 = v >> 8 */
139 	bslli	r11, r12, 24	/* h = v << 24 */
140 	lwi	r12, r8, 24	/* v = *(as + 24) */
141 	bsrli	r9, r12, 8	/* t1 = v >> 8 */
144 	bslli	r11, r12, 24	/* h = v << 24 */
145 	lwi	r12, r8, 28	/* v = *(as + 28) */
146 	bsrli	r9, r12, 8	/* t1 = v >> 8 */
149 	bslli	r11, r12, 24	/* h = v << 24 */
150 	lwi	r12, r8, 32	/* v = *(as + 32) */
151 	bsrli	r9, r12, 8	/* t1 = v >> 8 */
154 	bslli	r11, r12, 24	/* h = v << 24 */
164 	lwi	r12, r8, 4	/* v = *(as + 4) */
165 	bsrli	r9, r12, 24	/* t1 = v >> 24 */
168 	bslli	r11, r12, 8	/* h = v << 8 */
169 	lwi	r12, r8, 8	/* v = *(as + 8) */
170 	bsrli	r9, r12, 24	/* t1 = v >> 24 */
172 	swi	r9, r5, 4	/* *(d + 4) = t1 */
173 	bslli	r11, r12, 8	/* h = v << 8 */
174 	lwi	r12, r8, 12	/* v = *(as + 12) */
175 	bsrli	r9, r12, 24	/* t1 = v >> 24 */
178 	bslli	r11, r12, 8	/* h = v << 8 */
179 	lwi	r12, r8, 16	/* v = *(as + 16) */
180 	bsrli	r9, r12, 24	/* t1 = v >> 24 */
183 	bslli	r11, r12, 8	/* h = v << 8 */
184 	lwi	r12, r8, 20	/* v = *(as + 20) */
185 	bsrli	r9, r12, 24	/* t1 = v >> 24 */
188 	bslli	r11, r12, 8	/* h = v << 8 */
189 	lwi	r12, r8, 24	/* v = *(as + 24) */
190 	bsrli	r9, r12, 24	/* t1 = v >> 24 */
193 	bslli	r11, r12, 8	/* h = v << 8 */
194 	lwi	r12, r8, 28	/* v = *(as + 28) */
195 	bsrli	r9, r12, 24	/* t1 = v >> 24 */
198 	bslli	r11, r12, 8	/* h = v << 8 */
199 	lwi	r12, r8, 32	/* v = *(as + 32) */
200 	bsrli	r9, r12, 24	/* t1 = v >> 24 */
203 	bslli	r11, r12, 8	/* h = v << 8 */
213 	lwi	r12, r8, 4	/* v = *(as + 4) */
214 	bsrli	r9, r12, 16	/* t1 = v >> 16 */
217 	bslli	r11, r12, 16	/* h = v << 16 */
218 	lwi	r12, r8, 8	/* v = *(as + 8) */
219 	bsrli	r9, r12, 16	/* t1 = v >> 16 */
221 	swi	r9, r5, 4	/* *(d + 4) = t1 */
222 	bslli	r11, r12, 16	/* h = v << 16 */
223 	lwi	r12, r8, 12	/* v = *(as + 12) */
224 	bsrli	r9, r12, 16	/* t1 = v >> 16 */
227 	bslli	r11, r12, 16	/* h = v << 16 */
228 	lwi	r12, r8, 16	/* v = *(as + 16) */
229 	bsrli	r9, r12, 16	/* t1 = v >> 16 */
232 	bslli	r11, r12, 16	/* h = v << 16 */
233 	lwi	r12, r8, 20	/* v = *(as + 20) */
234 	bsrli	r9, r12, 16	/* t1 = v >> 16 */
237 	bslli	r11, r12, 16	/* h = v << 16 */
238 	lwi	r12, r8, 24	/* v = *(as + 24) */
239 	bsrli	r9, r12, 16	/* t1 = v >> 16 */
242 	bslli	r11, r12, 16	/* h = v << 16 */
243 	lwi	r12, r8, 28	/* v = *(as + 28) */
244 	bsrli	r9, r12, 16	/* t1 = v >> 16 */
247 	bslli	r11, r12, 16	/* h = v << 16 */
248 	lwi	r12, r8, 32	/* v = *(as + 32) */
249 	bsrli	r9, r12, 16	/* t1 = v >> 16 */
252 	bslli	r11, r12, 16	/* h = v << 16 */
259 	addi	r4, r0, 4	/* n = 4 */
274 	addi	r4, r4,-4		/* n-- */
276 	addi	r10, r10, 4		/* offset++ (IN DELAY SLOT) */
283 	addi	r8, r8, 4		/* as = as + 4 */
293 	lw	r12, r8, r10	/* v = *(as + offset) */
294 	bsrli	r9, r12, 8	/* t1 = v >> 8 */
297 	bslli	r11, r12, 24	/* h = v << 24 */
298 	addi	r4, r4,-4	/* n = n - 4 */
300 	addi	r10, r10, 4	/* offset = ofset + 4 (IN DELAY SLOT) */
307 	lw	r12, r8, r10	/* v = *(as + offset) */
308 	bsrli	r9, r12, 24	/* t1 = v >> 24 */
311 	bslli	r11, r12, 8	/* h = v << 8 */
312 	addi	r4, r4,-4	/* n = n - 4 */
314 	addi	r10, r10, 4	/* offset = ofset + 4 (IN DELAY SLOT) */
321 	lw	r12, r8, r10	/* v = *(as + offset) */
322 	bsrli	r9, r12, 16	/* t1 = v >> 16 */
325 	bslli	r11, r12, 16	/* h = v << 16 */
326 	addi	r4, r4,-4	/* n = n - 4 */
328 	addi	r10, r10, 4	/* offset = ofset + 4 (IN DELAY SLOT) */
367 	addi	r4, r0, 4	/* n = 4 */
414 	lwi	r11, r6, 4		/* t3 = *(s + 4) */
418 	swi	r11, r5, 4		/* *(d + 4) = t3 */
439 	lwi	r12, r8, 28	/* v = *(as + 28) */
440 	bslli	r9, r12, 24	/* t1 = v << 24 */
443 	bsrli	r11, r12, 8	/* h = v >> 8 */
444 	lwi	r12, r8, 24	/* v = *(as + 24) */
445 	bslli	r9, r12, 24	/* t1 = v << 24 */
448 	bsrli	r11, r12, 8	/* h = v >> 8 */
449 	lwi	r12, r8, 20	/* v = *(as + 20) */
450 	bslli	r9, r12, 24	/* t1 = v << 24 */
453 	bsrli	r11, r12, 8	/* h = v >> 8 */
454 	lwi	r12, r8, 16	/* v = *(as + 16) */
455 	bslli	r9, r12, 24	/* t1 = v << 24 */
458 	bsrli	r11, r12, 8	/* h = v >> 8 */
459 	lwi	r12, r8, 12	/* v = *(as + 12) */
460 	bslli	r9, r12, 24	/* t1 = v << 24 */
463 	bsrli	r11, r12, 8	/* h = v >> 8 */
464 	lwi	r12, r8, 8	/* v = *(as + 8) */
465 	bslli	r9, r12, 24	/* t1 = v << 24 */
468 	bsrli	r11, r12, 8	/* h = v >> 8 */
469 	lwi	r12, r8, 4	/* v = *(as + 4) */
470 	bslli	r9, r12, 24	/* t1 = v << 24 */
472 	swi	r9, r5, 4	/* *(d + 4) = t1 */
473 	bsrli	r11, r12, 8	/* h = v >> 8 */
474 	lwi	r12, r8, 0	/* v = *(as + 0) */
475 	bslli	r9, r12, 24	/* t1 = v << 24 */
480 	bsrli	r11, r12, 8	/* h = v >> 8 (IN DELAY SLOT) */
488 	lwi	r12, r8, 28	/* v = *(as + 28) */
489 	bslli	r9, r12, 8	/* t1 = v << 8 */
492 	bsrli	r11, r12, 24	/* h = v >> 24 */
493 	lwi	r12, r8, 24	/* v = *(as + 24) */
494 	bslli	r9, r12, 8	/* t1 = v << 8 */
497 	bsrli	r11, r12, 24	/* h = v >> 24 */
498 	lwi	r12, r8, 20	/* v = *(as + 20) */
499 	bslli	r9, r12, 8	/* t1 = v << 8 */
502 	bsrli	r11, r12, 24	/* h = v >> 24 */
503 	lwi	r12, r8, 16	/* v = *(as + 16) */
504 	bslli	r9, r12, 8	/* t1 = v << 8 */
507 	bsrli	r11, r12, 24	/* h = v >> 24 */
508 	lwi	r12, r8, 12	/* v = *(as + 12) */
509 	bslli	r9, r12, 8	/* t1 = v << 8 */
512 	bsrli	r11, r12, 24	/* h = v >> 24 */
513 	lwi	r12, r8, 8	/* v = *(as + 8) */
514 	bslli	r9, r12, 8	/* t1 = v << 8 */
517 	bsrli	r11, r12, 24	/* h = v >> 24 */
518 	lwi	r12, r8, 4	/* v = *(as + 4) */
519 	bslli	r9, r12, 8	/* t1 = v << 8 */
521 	swi	r9, r5, 4	/* *(d + 4) = t1 */
522 	bsrli	r11, r12, 24	/* h = v >> 24 */
523 	lwi	r12, r8, 0	/* v = *(as + 0) */
524 	bslli	r9, r12, 8	/* t1 = v << 8 */
529 	bsrli	r11, r12, 24	/* h = v >> 24 (IN DELAY SLOT) */
537 	lwi	r12, r8, 28	/* v = *(as + 28) */
538 	bslli	r9, r12, 16	/* t1 = v << 16 */
541 	bsrli	r11, r12, 16	/* h = v >> 16 */
542 	lwi	r12, r8, 24	/* v = *(as + 24) */
543 	bslli	r9, r12, 16	/* t1 = v << 16 */
546 	bsrli	r11, r12, 16	/* h = v >> 16 */
547 	lwi	r12, r8, 20	/* v = *(as + 20) */
548 	bslli	r9, r12, 16	/* t1 = v << 16 */
551 	bsrli	r11, r12, 16	/* h = v >> 16 */
552 	lwi	r12, r8, 16	/* v = *(as + 16) */
553 	bslli	r9, r12, 16	/* t1 = v << 16 */
556 	bsrli	r11, r12, 16	/* h = v >> 16 */
557 	lwi	r12, r8, 12	/* v = *(as + 12) */
558 	bslli	r9, r12, 16	/* t1 = v << 16 */
561 	bsrli	r11, r12, 16	/* h = v >> 16 */
562 	lwi	r12, r8, 8	/* v = *(as + 8) */
563 	bslli	r9, r12, 16	/* t1 = v << 16 */
566 	bsrli	r11, r12, 16	/* h = v >> 16 */
567 	lwi	r12, r8, 4	/* v = *(as + 4) */
568 	bslli	r9, r12, 16	/* t1 = v << 16 */
570 	swi	r9, r5, 4	/* *(d + 4) = t1 */
571 	bsrli	r11, r12, 16	/* h = v >> 16 */
572 	lwi	r12, r8, 0	/* v = *(as + 0) */
573 	bslli	r9, r12, 16	/* t1 = v << 16 */
578 	bsrli	r11, r12, 16	/* h = v >> 16 (IN DELAY SLOT) */
581 	addi	r4, r0, 4	/* n = 4 */
596 	addi	r4, r4,-4		/* n-- */
615 	addi	r4, r4,-4	/* n = n - 4 */
616 	lw	r12, r8, r4	/* v = *(as + n) */
617 	bslli	r9, r12, 24	/* t1 = v << 24 */
621 	bsrli	r11, r12, 8	/* h = v >> 8 (IN DELAY SLOT) */
628 	addi	r4, r4,-4	/* n = n - 4 */
629 	lw	r12, r8, r4	/* v = *(as + n) */
630 	bslli	r9, r12, 8	/* t1 = v << 8 */
634 	bsrli	r11, r12, 24	/* h = v >> 24 (IN DELAY SLOT) */
641 	addi	r4, r4,-4	/* n = n - 4 */
642 	lw	r12, r8, r4	/* v = *(as + n) */
643 	bslli	r9, r12, 16	/* t1 = v << 16 */
647 	bsrli	r11, r12, 16	/* h = v >> 16 (IN DELAY SLOT) */