xref: /openbmc/linux/arch/sh/lib/memcpy-sh4.S (revision 842ed298)
1/* SPDX-License-Identifier: GPL-2.0 */
2/*
3 * "memcpy" implementation of SuperH
4 *
5 * Copyright (C) 1999  Niibe Yutaka
6 * Copyright (c) 2002  STMicroelectronics Ltd
7 *   Modified from memcpy.S and micro-optimised for SH4
8 *   Stuart Menefy (stuart.menefy@st.com)
9 *
10 */
11#include <linux/linkage.h>
12
13/*
14 * void *memcpy(void *dst, const void *src, size_t n);
15 *
16 * It is assumed that there is no overlap between src and dst.
17 * If there is an overlap, then the results are undefined.
18 */
19
20	!
21	!	GHIJ KLMN OPQR -->  ...G HIJK LMNO PQR.
22	!
23
24	! Size is 16 or greater, and may have trailing bytes
25
26	.balign	32
27.Lcase1:
28	! Read a long word and write a long word at once
29	! At the start of each iteration, r7 contains last long load
30	add	#-1,r5		!  79 EX
31	mov	r4,r2		!   5 MT (0 cycles latency)
32
33	mov.l	@(r0,r5),r7	!  21 LS (2 cycles latency)
34	add	#-4,r5		!  50 EX
35
36	add	#7,r2		!  79 EX
37	!
38#ifdef CONFIG_CPU_LITTLE_ENDIAN
39	! 6 cycles, 4 bytes per iteration
403:	mov.l	@(r0,r5),r1	!  21 LS (latency=2)	! NMLK
41	mov	r7, r3		!   5 MT (latency=0)	! RQPO
42
43	cmp/hi	r2,r0		!  57 MT
44	shll16	r3		! 103 EX
45
46	mov	r1,r6		!   5 MT (latency=0)
47	shll8	r3		! 102 EX		! Oxxx
48
49	shlr8	r6		! 106 EX		! xNML
50	mov	r1, r7		!   5 MT (latency=0)
51
52	or	r6,r3		!  82 EX		! ONML
53	bt/s	3b		! 109 BR
54
55	 mov.l	r3,@-r0		!  30 LS
56#else
573:	mov.l	@(r0,r5),r1	!  21 LS (latency=2)	! KLMN
58	mov	r7,r3		!   5 MT (latency=0)	! OPQR
59
60	cmp/hi	r2,r0		!  57 MT
61	shlr16	r3		! 107 EX
62
63	shlr8	r3		! 106 EX		! xxxO
64	mov	r1,r6		!   5 MT (latency=0)
65
66	shll8	r6		! 102 EX		! LMNx
67	mov	r1,r7		!   5 MT (latency=0)
68
69	or	r6,r3		!  82 EX		! LMNO
70	bt/s	3b		! 109 BR
71
72	 mov.l	r3,@-r0		!  30 LS
73#endif
74	! Finally, copy a byte at once, if necessary
75
76	add	#4,r5		!  50 EX
77	cmp/eq	r4,r0		!  54 MT
78
79	add	#-6,r2		!  50 EX
80	bt	9f		! 109 BR
81
828:	cmp/hi	r2,r0		!  57 MT
83	mov.b	@(r0,r5),r1	!  20 LS (latency=2)
84
85	bt/s	8b		! 109 BR
86
87	 mov.b	r1,@-r0		!  29 LS
88
899:	rts
90	 nop
91
92
93	!
94	!	GHIJ KLMN OPQR -->  .GHI JKLM NOPQ R...
95	!
96
97	! Size is 16 or greater, and may have trailing bytes
98
99	.balign	32
100.Lcase3:
101	! Read a long word and write a long word at once
102	! At the start of each iteration, r7 contains last long load
103	add	#-3,r5		! 79 EX
104	mov	r4,r2		!  5 MT (0 cycles latency)
105
106	mov.l	@(r0,r5),r7	! 21 LS (2 cycles latency)
107	add	#-4,r5		! 50 EX
108
109	add	#7,r2		!  79 EX
110	!
111#ifdef CONFIG_CPU_LITTLE_ENDIAN
112	! 6 cycles, 4 bytes per iteration
1133:	mov.l	@(r0,r5),r1	!  21 LS (latency=2)	! NMLK
114	mov	r7, r3		!   5 MT (latency=0)	! RQPO
115
116	cmp/hi	r2,r0		!  57 MT
117	shll8	r3		! 102 EX		! QPOx
118
119	mov	r1,r6		!   5 MT (latency=0)
120	shlr16	r6		! 107 EX
121
122	shlr8	r6		! 106 EX		! xxxN
123	mov	r1, r7		!   5 MT (latency=0)
124
125	or	r6,r3		!  82 EX		! QPON
126	bt/s	3b		! 109 BR
127
128	 mov.l	r3,@-r0		!  30 LS
129#else
1303:	mov	r7,r3		! OPQR
131	shlr8	r3		! xOPQ
132	mov.l	@(r0,r5),r7	! KLMN
133	mov	r7,r6
134	shll16	r6
135	shll8	r6		! Nxxx
136	or	r6,r3		! NOPQ
137	cmp/hi	r2,r0
138	bt/s	3b
139	 mov.l	r3,@-r0
140#endif
141
142	! Finally, copy a byte at once, if necessary
143
144	add	#6,r5		!  50 EX
145	cmp/eq	r4,r0		!  54 MT
146
147	add	#-6,r2		!  50 EX
148	bt	9f		! 109 BR
149
1508:	cmp/hi	r2,r0		!  57 MT
151	mov.b	@(r0,r5),r1	!  20 LS (latency=2)
152
153	bt/s	8b		! 109 BR
154
155	 mov.b	r1,@-r0		!  29 LS
156
1579:	rts
158	 nop
159
160ENTRY(memcpy)
161
162	! Calculate the invariants which will be used in the remainder
163	! of the code:
164	!
165	!      r4   -->  [ ...  ] DST             [ ...  ] SRC
166	!	         [ ...  ]                 [ ...  ]
167	!	           :                        :
168	!      r0   -->  [ ...  ]       r0+r5 --> [ ...  ]
169	!
170	!
171
172	! Short circuit the common case of src, dst and len being 32 bit aligned
173	! and test for zero length move
174
175	mov	r6, r0		!   5 MT (0 cycle latency)
176	or	r4, r0		!  82 EX
177
178	or	r5, r0		!  82 EX
179	tst	r6, r6		!  86 MT
180
181	bt/s	99f		! 111 BR		(zero len)
182	 tst	#3, r0		!  87 MT
183
184	mov	r4, r0		!   5 MT (0 cycle latency)
185	add	r6, r0		!  49 EX
186
187	mov	#16, r1		!   6 EX
188	bt/s	.Lcase00	! 111 BR		(aligned)
189
190	 sub	r4, r5		!  75 EX
191
192	! Arguments are not nicely long word aligned or zero len.
193	! Check for small copies, and if so do a simple byte at a time copy.
194	!
195	! Deciding on an exact value of 'small' is not easy, as the point at which
196	! using the optimised routines become worthwhile varies (these are the
197	! cycle counts for differnet sizes using byte-at-a-time vs. optimised):
198	!	size	byte-at-time	long	word	byte
199	!	16	42		39-40	46-50	50-55
200	!	24	58		43-44	54-58	62-67
201	!	36	82		49-50	66-70	80-85
202	! However the penalty for getting it 'wrong' is much higher for long word
203	! aligned data (and this is more common), so use a value of 16.
204
205	cmp/gt	r6,r1		!  56 MT
206
207	add	#-1,r5		!  50 EX
208	bf/s	6f		! 108 BR		(not small)
209
210	 mov	r5, r3		!   5 MT (latency=0)
211	shlr	r6		! 104 EX
212
213	mov.b	@(r0,r5),r1	!  20 LS (latency=2)
214	bf/s	4f		! 111 BR
215
216	 add	#-1,r3		!  50 EX
217	tst	r6, r6		!  86 MT
218
219	bt/s	98f		! 110 BR
220	 mov.b	r1,@-r0		!  29 LS
221
222	! 4 cycles, 2 bytes per iteration
2233:	mov.b	@(r0,r5),r1	!  20 LS (latency=2)
224
2254:	mov.b	@(r0,r3),r2	!  20 LS (latency=2)
226	dt	r6		!  67 EX
227
228	mov.b	r1,@-r0		!  29 LS
229	bf/s	3b		! 111 BR
230
231	 mov.b	r2,@-r0		!  29 LS
23298:
233	rts
234	 nop
235
23699:	rts
237	 mov	r4, r0
238
239	! Size is not small, so its worthwhile looking for optimisations.
240	! First align destination to a long word boundary.
241	!
242	! r5 = normal value -1
243
2446:	tst	#3, r0		!  87 MT
245        mov	#3, r3		!   6 EX
246
247	bt/s	2f		! 111 BR
248	 and	r0,r3		!  78 EX
249
250	! 3 cycles, 1 byte per iteration
2511:	dt	r3		!  67 EX
252	mov.b	@(r0,r5),r1	!  19 LS (latency=2)
253
254	add	#-1, r6		!  79 EX
255	bf/s	1b		! 109 BR
256
257	 mov.b	r1,@-r0		!  28 LS
258
2592:	add	#1, r5		!  79 EX
260
261	! Now select the appropriate bulk transfer code based on relative
262	! alignment of src and dst.
263
264	mov	r0, r3		!   5 MT (latency=0)
265
266	mov	r5, r0		!   5 MT (latency=0)
267	tst	#1, r0		!  87 MT
268
269	bf/s	1f		! 111 BR
270	 mov	#64, r7		!   6 EX
271
272	! bit 0 clear
273
274	cmp/ge	r7, r6		!  55 MT
275
276	bt/s	2f		! 111 BR
277	 tst	#2, r0		!  87 MT
278
279	! small
280	bt/s	.Lcase0
281	 mov	r3, r0
282
283	bra	.Lcase2
284	 nop
285
286	! big
2872:	bt/s	.Lcase0b
288	 mov	r3, r0
289
290	bra	.Lcase2b
291	 nop
292
293	! bit 0 set
2941:	tst	#2, r0		! 87 MT
295
296	bt/s	.Lcase1
297	 mov	r3, r0
298
299	bra	.Lcase3
300	 nop
301
302
303	!
304	!	GHIJ KLMN OPQR -->  GHIJ KLMN OPQR
305	!
306
307	! src, dst and size are all long word aligned
308	! size is non-zero
309
310	.balign	32
311.Lcase00:
312	mov	#64, r1		!   6 EX
313	mov	r5, r3		!   5 MT (latency=0)
314
315	cmp/gt	r6, r1		!  56 MT
316	add	#-4, r5		!  50 EX
317
318	bf	.Lcase00b	! 108 BR		(big loop)
319	shlr2	r6		! 105 EX
320
321	shlr	r6		! 104 EX
322	mov.l	@(r0, r5), r1	!  21 LS (latency=2)
323
324	bf/s	4f		! 111 BR
325	 add	#-8, r3		!  50 EX
326
327	tst	r6, r6		!  86 MT
328	bt/s	5f		! 110 BR
329
330	 mov.l	r1,@-r0		!  30 LS
331
332	! 4 cycles, 2 long words per iteration
3333:	mov.l	@(r0, r5), r1	!  21 LS (latency=2)
334
3354:	mov.l	@(r0, r3), r2	!  21 LS (latency=2)
336	dt	r6		!  67 EX
337
338	mov.l	r1, @-r0	!  30 LS
339	bf/s	3b		! 109 BR
340
341	 mov.l	r2, @-r0	!  30 LS
342
3435:	rts
344	 nop
345
346
347	! Size is 16 or greater and less than 64, but may have trailing bytes
348
349	.balign	32
350.Lcase0:
351	add	#-4, r5		!  50 EX
352	mov	r4, r7		!   5 MT (latency=0)
353
354	mov.l	@(r0, r5), r1	!  21 LS (latency=2)
355	mov	#4, r2		!   6 EX
356
357	add	#11, r7		!  50 EX
358	tst	r2, r6		!  86 MT
359
360	mov	r5, r3		!   5 MT (latency=0)
361	bt/s	4f		! 111 BR
362
363	 add	#-4, r3		!  50 EX
364	mov.l	r1,@-r0		!  30 LS
365
366	! 4 cycles, 2 long words per iteration
3673:	mov.l	@(r0, r5), r1	!  21 LS (latency=2)
368
3694:	mov.l	@(r0, r3), r2	!  21 LS (latency=2)
370	cmp/hi	r7, r0
371
372	mov.l	r1, @-r0	!  30 LS
373	bt/s	3b		! 109 BR
374
375	 mov.l	r2, @-r0	!  30 LS
376
377	! Copy the final 0-3 bytes
378
379	add	#3,r5		!  50 EX
380
381	cmp/eq	r0, r4		!  54 MT
382	add	#-10, r7	!  50 EX
383
384	bt	9f		! 110 BR
385
386	! 3 cycles, 1 byte per iteration
3871:	mov.b	@(r0,r5),r1	!  19 LS
388	cmp/hi	r7,r0		!  57 MT
389
390	bt/s	1b		! 111 BR
391	 mov.b	r1,@-r0		!  28 LS
392
3939:	rts
394	 nop
395
396	! Size is at least 64 bytes, so will be going round the big loop at least once.
397	!
398	!   r2 = rounded up r4
399	!   r3 = rounded down r0
400
401	.balign	32
402.Lcase0b:
403	add	#-4, r5		!  50 EX
404
405.Lcase00b:
406	mov	r0, r3		!   5 MT (latency=0)
407	mov	#(~0x1f), r1	!   6 EX
408
409	and	r1, r3		!  78 EX
410	mov	r4, r2		!   5 MT (latency=0)
411
412	cmp/eq	r3, r0		!  54 MT
413	add	#0x1f, r2	!  50 EX
414
415	bt/s	1f		! 110 BR
416	 and	r1, r2		!  78 EX
417
418	! copy initial words until cache line aligned
419
420	mov.l	@(r0, r5), r1	!  21 LS (latency=2)
421	tst	#4, r0		!  87 MT
422
423	mov	r5, r6		!   5 MT (latency=0)
424	add	#-4, r6		!  50 EX
425
426	bt/s	4f		! 111 BR
427	 add	#8, r3		!  50 EX
428
429	tst	#0x18, r0	!  87 MT
430
431	bt/s	1f		! 109 BR
432	 mov.l	r1,@-r0		!  30 LS
433
434	! 4 cycles, 2 long words per iteration
4353:	mov.l	@(r0, r5), r1	!  21 LS (latency=2)
436
4374:	mov.l	@(r0, r6), r7	!  21 LS (latency=2)
438	cmp/eq	r3, r0		!  54 MT
439
440	mov.l	r1, @-r0	!  30 LS
441	bf/s	3b		! 109 BR
442
443	 mov.l	r7, @-r0	!  30 LS
444
445	! Copy the cache line aligned blocks
446	!
447	! In use: r0, r2, r4, r5
448	! Scratch: r1, r3, r6, r7
449	!
450	! We could do this with the four scratch registers, but if src
451	! and dest hit the same cache line, this will thrash, so make
452	! use of additional registers.
453	!
454	! We also need r0 as a temporary (for movca), so 'undo' the invariant:
455	!   r5:	 src (was r0+r5)
456	!   r1:	 dest (was r0)
457	! this can be reversed at the end, so we don't need to save any extra
458	! state.
459	!
4601:	mov.l	r8, @-r15	!  30 LS
461	add	r0, r5		!  49 EX
462
463	mov.l	r9, @-r15	!  30 LS
464	mov	r0, r1		!   5 MT (latency=0)
465
466	mov.l	r10, @-r15	!  30 LS
467	add	#-0x1c, r5	!  50 EX
468
469	mov.l	r11, @-r15	!  30 LS
470
471	! 16 cycles, 32 bytes per iteration
4722:	mov.l	@(0x00,r5),r0	! 18 LS (latency=2)
473	add	#-0x20, r1	! 50 EX
474	mov.l	@(0x04,r5),r3	! 18 LS (latency=2)
475	mov.l	@(0x08,r5),r6	! 18 LS (latency=2)
476	mov.l	@(0x0c,r5),r7	! 18 LS (latency=2)
477	mov.l	@(0x10,r5),r8	! 18 LS (latency=2)
478	mov.l	@(0x14,r5),r9	! 18 LS (latency=2)
479	mov.l	@(0x18,r5),r10	! 18 LS (latency=2)
480	mov.l	@(0x1c,r5),r11	! 18 LS (latency=2)
481	movca.l	r0,@r1		! 40 LS (latency=3-7)
482	mov.l	r3,@(0x04,r1)	! 33 LS
483	mov.l	r6,@(0x08,r1)	! 33 LS
484	mov.l	r7,@(0x0c,r1)	! 33 LS
485
486	mov.l	r8,@(0x10,r1)	! 33 LS
487	add	#-0x20, r5	! 50 EX
488
489	mov.l	r9,@(0x14,r1)	! 33 LS
490	cmp/eq	r2,r1		! 54 MT
491
492	mov.l	r10,@(0x18,r1)	!  33 LS
493	bf/s	2b		! 109 BR
494
495	 mov.l	r11,@(0x1c,r1)	!  33 LS
496
497	mov	r1, r0		!   5 MT (latency=0)
498
499	mov.l	@r15+, r11	!  15 LS
500	sub	r1, r5		!  75 EX
501
502	mov.l	@r15+, r10	!  15 LS
503	cmp/eq	r4, r0		!  54 MT
504
505	bf/s	1f		! 109 BR
506	 mov.l	 @r15+, r9	!  15 LS
507
508	rts
5091:	 mov.l	@r15+, r8	!  15 LS
510	sub	r4, r1		!  75 EX		(len remaining)
511
512	! number of trailing bytes is non-zero
513	!
514	! invariants restored (r5 already decremented by 4)
515	! also r1=num bytes remaining
516
517	mov	#4, r2		!   6 EX
518	mov	r4, r7		!   5 MT (latency=0)
519
520	add	#0x1c, r5	!  50 EX		(back to -4)
521	cmp/hs	r2, r1		!  58 MT
522
523	bf/s	5f		! 108 BR
524	 add	 #11, r7	!  50 EX
525
526	mov.l	@(r0, r5), r6	!  21 LS (latency=2)
527	tst	r2, r1		!  86 MT
528
529	mov	r5, r3		!   5 MT (latency=0)
530	bt/s	4f		! 111 BR
531
532	 add	#-4, r3		!  50 EX
533	cmp/hs	r2, r1		!  58 MT
534
535	bt/s	5f		! 111 BR
536	 mov.l	r6,@-r0		!  30 LS
537
538	! 4 cycles, 2 long words per iteration
5393:	mov.l	@(r0, r5), r6	!  21 LS (latency=2)
540
5414:	mov.l	@(r0, r3), r2	!  21 LS (latency=2)
542	cmp/hi	r7, r0
543
544	mov.l	r6, @-r0	!  30 LS
545	bt/s	3b		! 109 BR
546
547	 mov.l	r2, @-r0	!  30 LS
548
549	! Copy the final 0-3 bytes
550
5515:	cmp/eq	r0, r4		!  54 MT
552	add	#-10, r7	!  50 EX
553
554	bt	9f		! 110 BR
555	add	#3,r5		!  50 EX
556
557	! 3 cycles, 1 byte per iteration
5581:	mov.b	@(r0,r5),r1	!  19 LS
559	cmp/hi	r7,r0		!  57 MT
560
561	bt/s	1b		! 111 BR
562	 mov.b	r1,@-r0		!  28 LS
563
5649:	rts
565	 nop
566
567	!
568	!	GHIJ KLMN OPQR -->  ..GH IJKL MNOP QR..
569	!
570
571	.balign	32
572.Lcase2:
573	! Size is 16 or greater and less then 64, but may have trailing bytes
574
5752:	mov	r5, r6		!   5 MT (latency=0)
576	add	#-2,r5		!  50 EX
577
578	mov	r4,r2		!   5 MT (latency=0)
579	add	#-4,r6		!  50 EX
580
581	add	#7,r2		!  50 EX
5823:	mov.w	@(r0,r5),r1	!  20 LS (latency=2)
583
584	mov.w	@(r0,r6),r3	!  20 LS (latency=2)
585	cmp/hi	r2,r0		!  57 MT
586
587	mov.w	r1,@-r0		!  29 LS
588	bt/s	3b		! 111 BR
589
590	 mov.w	r3,@-r0		!  29 LS
591
592	bra	10f
593	 nop
594
595
596	.balign	32
597.Lcase2b:
598	! Size is at least 64 bytes, so will be going round the big loop at least once.
599	!
600	!   r2 = rounded up r4
601	!   r3 = rounded down r0
602
603	mov	r0, r3		!   5 MT (latency=0)
604	mov	#(~0x1f), r1	!   6 EX
605
606	and	r1, r3		!  78 EX
607	mov	r4, r2		!   5 MT (latency=0)
608
609	cmp/eq	r3, r0		!  54 MT
610	add	#0x1f, r2	!  50 EX
611
612	add	#-2, r5		!  50 EX
613	bt/s	1f		! 110 BR
614	 and	r1, r2		!  78 EX
615
616	! Copy a short word one at a time until we are cache line aligned
617	!   Normal values: r0, r2, r3, r4
618	!   Unused: r1, r6, r7
619	!   Mod: r5 (=r5-2)
620	!
621	add	#2, r3		!  50 EX
622
6232:	mov.w	@(r0,r5),r1	!  20 LS (latency=2)
624	cmp/eq	r3,r0		!  54 MT
625
626	bf/s	2b		! 111 BR
627
628	 mov.w	r1,@-r0		!  29 LS
629
630	! Copy the cache line aligned blocks
631	!
632	! In use: r0, r2, r4, r5 (=r5-2)
633	! Scratch: r1, r3, r6, r7
634	!
635	! We could do this with the four scratch registers, but if src
636	! and dest hit the same cache line, this will thrash, so make
637	! use of additional registers.
638	!
639	! We also need r0 as a temporary (for movca), so 'undo' the invariant:
640	!   r5:	 src (was r0+r5)
641	!   r1:	 dest (was r0)
642	! this can be reversed at the end, so we don't need to save any extra
643	! state.
644	!
6451:	mov.l	r8, @-r15	!  30 LS
646	add	r0, r5		!  49 EX
647
648	mov.l	r9, @-r15	!  30 LS
649	mov	r0, r1		!   5 MT (latency=0)
650
651	mov.l	r10, @-r15	!  30 LS
652	add	#-0x1e, r5	!  50 EX
653
654	mov.l	r11, @-r15	!  30 LS
655
656	mov.l	r12, @-r15	!  30 LS
657
658	! 17 cycles, 32 bytes per iteration
659#ifdef CONFIG_CPU_LITTLE_ENDIAN
6602:	mov.w	@r5+, r0	!  14 LS (latency=2)		..JI
661	add	#-0x20, r1	!  50 EX
662
663	mov.l	@r5+, r3	!  15 LS (latency=2)		NMLK
664
665	mov.l	@r5+, r6	!  15 LS (latency=2)		RQPO
666	shll16	r0		! 103 EX			JI..
667
668	mov.l	@r5+, r7	!  15 LS (latency=2)
669	xtrct	r3, r0		!  48 EX			LKJI
670
671	mov.l	@r5+, r8	!  15 LS (latency=2)
672	xtrct	r6, r3		!  48 EX			PONM
673
674	mov.l	@r5+, r9	!  15 LS (latency=2)
675	xtrct	r7, r6		!  48 EX
676
677	mov.l	@r5+, r10	!  15 LS (latency=2)
678	xtrct	r8, r7		!  48 EX
679
680	mov.l	@r5+, r11	!  15 LS (latency=2)
681	xtrct	r9, r8		!  48 EX
682
683	mov.w	@r5+, r12	!  15 LS (latency=2)
684	xtrct	r10, r9		!  48 EX
685
686	movca.l	r0,@r1		!  40 LS (latency=3-7)
687	xtrct	r11, r10	!  48 EX
688
689	mov.l	r3, @(0x04,r1)	!  33 LS
690	xtrct	r12, r11	!  48 EX
691
692	mov.l	r6, @(0x08,r1)	!  33 LS
693
694	mov.l	r7, @(0x0c,r1)	!  33 LS
695
696	mov.l	r8, @(0x10,r1)	!  33 LS
697	add	#-0x40, r5	!  50 EX
698
699	mov.l	r9, @(0x14,r1)	!  33 LS
700	cmp/eq	r2,r1		!  54 MT
701
702	mov.l	r10, @(0x18,r1)	!  33 LS
703	bf/s	2b		! 109 BR
704
705	 mov.l	r11, @(0x1c,r1)	!  33 LS
706#else
7072:	mov.w	@(0x1e,r5), r0	!  17 LS (latency=2)
708	add	#-2, r5		!  50 EX
709
710	mov.l	@(0x1c,r5), r3	!  18 LS (latency=2)
711	add	#-4, r1		!  50 EX
712
713	mov.l	@(0x18,r5), r6	!  18 LS (latency=2)
714	shll16	r0		! 103 EX
715
716	mov.l	@(0x14,r5), r7	!  18 LS (latency=2)
717	xtrct	r3, r0		!  48 EX
718
719	mov.l	@(0x10,r5), r8	!  18 LS (latency=2)
720	xtrct	r6, r3		!  48 EX
721
722	mov.l	@(0x0c,r5), r9	!  18 LS (latency=2)
723	xtrct	r7, r6		!  48 EX
724
725	mov.l	@(0x08,r5), r10	!  18 LS (latency=2)
726	xtrct	r8, r7		!  48 EX
727
728	mov.l	@(0x04,r5), r11	!  18 LS (latency=2)
729	xtrct	r9, r8		!  48 EX
730
731	mov.l   @(0x00,r5), r12 !  18 LS (latency=2)
732    	xtrct	r10, r9		!  48 EX
733
734	movca.l	r0,@r1		!  40 LS (latency=3-7)
735	add	#-0x1c, r1	!  50 EX
736
737	mov.l	r3, @(0x18,r1)	!  33 LS
738	xtrct	r11, r10	!  48 EX
739
740	mov.l	r6, @(0x14,r1)	!  33 LS
741	xtrct	r12, r11	!  48 EX
742
743	mov.l	r7, @(0x10,r1)	!  33 LS
744
745	mov.l	r8, @(0x0c,r1)	!  33 LS
746	add	#-0x1e, r5	!  50 EX
747
748	mov.l	r9, @(0x08,r1)	!  33 LS
749	cmp/eq	r2,r1		!  54 MT
750
751	mov.l	r10, @(0x04,r1)	!  33 LS
752	bf/s	2b		! 109 BR
753
754	 mov.l	r11, @(0x00,r1)	!  33 LS
755#endif
756
757	mov.l	@r15+, r12
758	mov	r1, r0		!   5 MT (latency=0)
759
760	mov.l	@r15+, r11	!  15 LS
761	sub	r1, r5		!  75 EX
762
763	mov.l	@r15+, r10	!  15 LS
764	cmp/eq	r4, r0		!  54 MT
765
766	bf/s	1f		! 109 BR
767	 mov.l	 @r15+, r9	!  15 LS
768
769	rts
7701:	 mov.l	@r15+, r8	!  15 LS
771
772	add	#0x1e, r5	!  50 EX
773
774	! Finish off a short word at a time
775	! r5 must be invariant - 2
77610:	mov	r4,r2		!   5 MT (latency=0)
777	add	#1,r2		!  50 EX
778
779	cmp/hi	r2, r0		!  57 MT
780	bf/s	1f		! 109 BR
781
782	 add	#2, r2		!  50 EX
783
7843:	mov.w	@(r0,r5),r1	!  20 LS
785	cmp/hi	r2,r0		!  57 MT
786
787	bt/s	3b		! 109 BR
788
789	 mov.w	r1,@-r0		!  29 LS
7901:
791
792	!
793	! Finally, copy the last byte if necessary
794	cmp/eq	r4,r0		!  54 MT
795	bt/s	9b
796	 add	#1,r5
797	mov.b	@(r0,r5),r1
798	rts
799	 mov.b	r1,@-r0
800
801