1/*
2 * Copyright (c) 2012-2015
3 *      MIPS Technologies, Inc., California.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 *    notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 *    notice, this list of conditions and the following disclaimer in the
12 *    documentation and/or other materials provided with the distribution.
13 * 3. Neither the name of the MIPS Technologies, Inc., nor the names of its
14 *    contributors may be used to endorse or promote products derived from
15 *    this software without specific prior written permission.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE MIPS TECHNOLOGIES, INC. ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED.  IN NO EVENT SHALL THE MIPS TECHNOLOGIES, INC. BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * SUCH DAMAGE.
28 */
29
30#ifdef __ANDROID__
31# include <private/bionic_asm.h>
32# define USE_MEMMOVE_FOR_OVERLAP
33# define PREFETCH_LOAD_HINT PREFETCH_HINT_LOAD_STREAMED
34# define PREFETCH_STORE_HINT PREFETCH_HINT_PREPAREFORSTORE
35#elif _LIBC
36# include <sysdep.h>
37# include <regdef.h>
38# include <sys/asm.h>
39# define PREFETCH_LOAD_HINT PREFETCH_HINT_LOAD_STREAMED
40# define PREFETCH_STORE_HINT PREFETCH_HINT_PREPAREFORSTORE
41#elif _COMPILING_NEWLIB
42# include "machine/asm.h"
43# include "machine/regdef.h"
44# define PREFETCH_LOAD_HINT PREFETCH_HINT_LOAD_STREAMED
45# define PREFETCH_STORE_HINT PREFETCH_HINT_PREPAREFORSTORE
46#else
47# include <regdef.h>
48# include <sys/asm.h>
49#endif
50
51/* Check to see if the MIPS architecture we are compiling for supports
52 * prefetching.
53 */
54
55#if (__mips == 4) || (__mips == 5) || (__mips == 32) || (__mips == 64)
56# ifndef DISABLE_PREFETCH
57#  define USE_PREFETCH
58# endif
59#endif
60
61#if defined(_MIPS_SIM) && ((_MIPS_SIM == _ABI64) || (_MIPS_SIM == _ABIN32))
62# ifndef DISABLE_DOUBLE
63#  define USE_DOUBLE
64# endif
65#endif
66
67
68#if __mips_isa_rev > 5
69# if (PREFETCH_STORE_HINT == PREFETCH_HINT_PREPAREFORSTORE)
70#  undef PREFETCH_STORE_HINT
71#  define PREFETCH_STORE_HINT PREFETCH_HINT_STORE_STREAMED
72# endif
73# define R6_CODE
74#endif
75
76/* Some asm.h files do not have the L macro definition.  */
77#ifndef L
78# if _MIPS_SIM == _ABIO32
79#  define L(label) $L ## label
80# else
81#  define L(label) .L ## label
82# endif
83#endif
84
85/* Some asm.h files do not have the PTR_ADDIU macro definition.  */
86#ifndef PTR_ADDIU
87# if _MIPS_SIM == _ABIO32
88#  define PTR_ADDIU	addiu
89# else
90#  define PTR_ADDIU	daddiu
91# endif
92#endif
93
94/* Some asm.h files do not have the PTR_SRA macro definition.  */
95#ifndef PTR_SRA
96# if  _MIPS_SIM == _ABIO32
97#  define PTR_SRA	sra
98# else
99#  define PTR_SRA	dsra
100# endif
101#endif
102
103/* New R6 instructions that may not be in asm.h.  */
104#ifndef PTR_LSA
105# if _MIPS_SIM == _ABIO32
106#  define PTR_LSA	lsa
107# else
108#  define PTR_LSA	dlsa
109# endif
110#endif
111
112/*
113 * Using PREFETCH_HINT_LOAD_STREAMED instead of PREFETCH_LOAD on load
114 * prefetches appears to offer a slight preformance advantage.
115 *
116 * Using PREFETCH_HINT_PREPAREFORSTORE instead of PREFETCH_STORE
117 * or PREFETCH_STORE_STREAMED offers a large performance advantage
118 * but PREPAREFORSTORE has some special restrictions to consider.
119 *
120 * Prefetch with the 'prepare for store' hint does not copy a memory
121 * location into the cache, it just allocates a cache line and zeros
122 * it out.  This means that if you do not write to the entire cache
123 * line before writing it out to memory some data will get zero'ed out
124 * when the cache line is written back to memory and data will be lost.
125 *
126 * Also if you are using this memcpy to copy overlapping buffers it may
127 * not behave correctly when using the 'prepare for store' hint.  If you
128 * use the 'prepare for store' prefetch on a memory area that is in the
129 * memcpy source (as well as the memcpy destination), then you will get
130 * some data zero'ed out before you have a chance to read it and data will
131 * be lost.
132 *
133 * If you are going to use this memcpy routine with the 'prepare for store'
134 * prefetch you may want to set USE_MEMMOVE_FOR_OVERLAP in order to avoid
135 * the problem of running memcpy on overlapping buffers.
136 *
137 * There are ifdef'ed sections of this memcpy to make sure that it does not
138 * do prefetches on cache lines that are not going to be completely written.
139 * This code is only needed and only used when PREFETCH_STORE_HINT is set to
140 * PREFETCH_HINT_PREPAREFORSTORE.  This code assumes that cache lines are
141 * 32 bytes and if the cache line is larger it will not work correctly.
142 */
143
144#ifdef USE_PREFETCH
145# define PREFETCH_HINT_LOAD		0
146# define PREFETCH_HINT_STORE		1
147# define PREFETCH_HINT_LOAD_STREAMED	4
148# define PREFETCH_HINT_STORE_STREAMED	5
149# define PREFETCH_HINT_LOAD_RETAINED	6
150# define PREFETCH_HINT_STORE_RETAINED	7
151# define PREFETCH_HINT_WRITEBACK_INVAL	25
152# define PREFETCH_HINT_PREPAREFORSTORE	30
153
154/*
155 * If we have not picked out what hints to use at this point use the
156 * standard load and store prefetch hints.
157 */
158# ifndef PREFETCH_STORE_HINT
159#  define PREFETCH_STORE_HINT PREFETCH_HINT_STORE
160# endif
161# ifndef PREFETCH_LOAD_HINT
162#  define PREFETCH_LOAD_HINT PREFETCH_HINT_LOAD
163# endif
164
165/*
166 * We double everything when USE_DOUBLE is true so we do 2 prefetches to
167 * get 64 bytes in that case.  The assumption is that each individual
168 * prefetch brings in 32 bytes.
169 */
170
171# ifdef USE_DOUBLE
172#  define PREFETCH_CHUNK 64
173#  define PREFETCH_FOR_LOAD(chunk, reg) \
174 pref PREFETCH_LOAD_HINT, (chunk)*64(reg); \
175 pref PREFETCH_LOAD_HINT, ((chunk)*64)+32(reg)
176#  define PREFETCH_FOR_STORE(chunk, reg) \
177 pref PREFETCH_STORE_HINT, (chunk)*64(reg); \
178 pref PREFETCH_STORE_HINT, ((chunk)*64)+32(reg)
179# else
180#  define PREFETCH_CHUNK 32
181#  define PREFETCH_FOR_LOAD(chunk, reg) \
182 pref PREFETCH_LOAD_HINT, (chunk)*32(reg)
183#  define PREFETCH_FOR_STORE(chunk, reg) \
184 pref PREFETCH_STORE_HINT, (chunk)*32(reg)
185# endif
186/* MAX_PREFETCH_SIZE is the maximum size of a prefetch, it must not be less
187 * than PREFETCH_CHUNK, the assumed size of each prefetch.  If the real size
188 * of a prefetch is greater than MAX_PREFETCH_SIZE and the PREPAREFORSTORE
189 * hint is used, the code will not work correctly.  If PREPAREFORSTORE is not
190 * used then MAX_PREFETCH_SIZE does not matter.  */
191# define MAX_PREFETCH_SIZE 128
192/* PREFETCH_LIMIT is set based on the fact that we never use an offset greater
193 * than 5 on a STORE prefetch and that a single prefetch can never be larger
194 * than MAX_PREFETCH_SIZE.  We add the extra 32 when USE_DOUBLE is set because
195 * we actually do two prefetches in that case, one 32 bytes after the other.  */
196# ifdef USE_DOUBLE
197#  define PREFETCH_LIMIT (5 * PREFETCH_CHUNK) + 32 + MAX_PREFETCH_SIZE
198# else
199#  define PREFETCH_LIMIT (5 * PREFETCH_CHUNK) + MAX_PREFETCH_SIZE
200# endif
201# if (PREFETCH_STORE_HINT == PREFETCH_HINT_PREPAREFORSTORE) \
202    && ((PREFETCH_CHUNK * 4) < MAX_PREFETCH_SIZE)
203/* We cannot handle this because the initial prefetches may fetch bytes that
204 * are before the buffer being copied.  We start copies with an offset
205 * of 4 so avoid this situation when using PREPAREFORSTORE.  */
206#error "PREFETCH_CHUNK is too large and/or MAX_PREFETCH_SIZE is too small."
207# endif
208#else /* USE_PREFETCH not defined */
209# define PREFETCH_FOR_LOAD(offset, reg)
210# define PREFETCH_FOR_STORE(offset, reg)
211#endif
212
213/* Allow the routine to be named something else if desired.  */
214#ifndef MEMCPY_NAME
215# define MEMCPY_NAME memcpy
216#endif
217
218/* We use these 32/64 bit registers as temporaries to do the copying.  */
219#define REG0 t0
220#define REG1 t1
221#define REG2 t2
222#define REG3 t3
223#if defined(_MIPS_SIM) && (_MIPS_SIM == _ABIO32 || _MIPS_SIM == _ABIO64)
224# define REG4 t4
225# define REG5 t5
226# define REG6 t6
227# define REG7 t7
228#else
229# define REG4 ta0
230# define REG5 ta1
231# define REG6 ta2
232# define REG7 ta3
233#endif
234
235/* We load/store 64 bits at a time when USE_DOUBLE is true.
236 * The C_ prefix stands for CHUNK and is used to avoid macro name
237 * conflicts with system header files.  */
238
239#ifdef USE_DOUBLE
240# define C_ST	sd
241# define C_LD	ld
242# if __MIPSEB
243#  define C_LDHI	ldl	/* high part is left in big-endian	*/
244#  define C_STHI	sdl	/* high part is left in big-endian	*/
245#  define C_LDLO	ldr	/* low part is right in big-endian	*/
246#  define C_STLO	sdr	/* low part is right in big-endian	*/
247# else
248#  define C_LDHI	ldr	/* high part is right in little-endian	*/
249#  define C_STHI	sdr	/* high part is right in little-endian	*/
250#  define C_LDLO	ldl	/* low part is left in little-endian	*/
251#  define C_STLO	sdl	/* low part is left in little-endian	*/
252# endif
253# define C_ALIGN	dalign	/* r6 align instruction			*/
254#else
255# define C_ST	sw
256# define C_LD	lw
257# if __MIPSEB
258#  define C_LDHI	lwl	/* high part is left in big-endian	*/
259#  define C_STHI	swl	/* high part is left in big-endian	*/
260#  define C_LDLO	lwr	/* low part is right in big-endian	*/
261#  define C_STLO	swr	/* low part is right in big-endian	*/
262# else
263#  define C_LDHI	lwr	/* high part is right in little-endian	*/
264#  define C_STHI	swr	/* high part is right in little-endian	*/
265#  define C_LDLO	lwl	/* low part is left in little-endian	*/
266#  define C_STLO	swl	/* low part is left in little-endian	*/
267# endif
268# define C_ALIGN	align	/* r6 align instruction			*/
269#endif
270
271/* Bookkeeping values for 32 vs. 64 bit mode.  */
272#ifdef USE_DOUBLE
273# define NSIZE 8
274# define NSIZEMASK 0x3f
275# define NSIZEDMASK 0x7f
276#else
277# define NSIZE 4
278# define NSIZEMASK 0x1f
279# define NSIZEDMASK 0x3f
280#endif
281#define UNIT(unit) ((unit)*NSIZE)
282#define UNITM1(unit) (((unit)*NSIZE)-1)
283
284#ifdef __ANDROID__
285LEAF(MEMCPY_NAME, 0)
286#else
287LEAF(MEMCPY_NAME)
288#endif
289	.set	nomips16
290	.set	noreorder
291/*
292 * Below we handle the case where memcpy is called with overlapping src and dst.
293 * Although memcpy is not required to handle this case, some parts of Android
294 * like Skia rely on such usage. We call memmove to handle such cases.
295 */
296#ifdef USE_MEMMOVE_FOR_OVERLAP
297	PTR_SUBU t0,a0,a1
298	PTR_SRA	t2,t0,31
299	xor	t1,t0,t2
300	PTR_SUBU t0,t1,t2
301	sltu	t2,t0,a2
302	beq	t2,zero,L(memcpy)
303	nop
304#if defined(__LP64__)
305	daddiu	sp,sp,-8
306	SETUP_GP64(0,MEMCPY_NAME)
307	LA	t9,memmove
308	RESTORE_GP64
309	jr	t9
310	daddiu	sp,sp,8
311#else
312	LA	t9,memmove
313	jr	t9
314	nop
315#endif
316L(memcpy):
317#endif
318/*
319 * If the size is less than 2*NSIZE (8 or 16), go to L(lastb).  Regardless of
320 * size, copy dst pointer to v0 for the return value.
321 */
322	slti	t2,a2,(2 * NSIZE)
323	bne	t2,zero,L(lastb)
324#if defined(RETURN_FIRST_PREFETCH) || defined(RETURN_LAST_PREFETCH)
325	move	v0,zero
326#else
327	move	v0,a0
328#endif
329
330#ifndef R6_CODE
331
332/*
333 * If src and dst have different alignments, go to L(unaligned), if they
334 * have the same alignment (but are not actually aligned) do a partial
335 * load/store to make them aligned.  If they are both already aligned
336 * we can start copying at L(aligned).
337 */
338	xor	t8,a1,a0
339	andi	t8,t8,(NSIZE-1)		/* t8 is a0/a1 word-displacement */
340	bne	t8,zero,L(unaligned)
341	PTR_SUBU a3, zero, a0
342
343	andi	a3,a3,(NSIZE-1)		/* copy a3 bytes to align a0/a1	  */
344	beq	a3,zero,L(aligned)	/* if a3=0, it is already aligned */
345	PTR_SUBU a2,a2,a3		/* a2 is the remining bytes count */
346
347	C_LDHI	t8,0(a1)
348	PTR_ADDU a1,a1,a3
349	C_STHI	t8,0(a0)
350	PTR_ADDU a0,a0,a3
351
352#else /* R6_CODE */
353
354/*
355 * Align the destination and hope that the source gets aligned too.  If it
356 * doesn't we jump to L(r6_unaligned*) to do unaligned copies using the r6
357 * align instruction.
358 */
359	andi	t8,a0,7
360	lapc	t9,L(atable)
361	PTR_LSA	t9,t8,t9,2
362	jrc	t9
363L(atable):
364	bc	L(lb0)
365	bc	L(lb7)
366	bc	L(lb6)
367	bc	L(lb5)
368	bc	L(lb4)
369	bc	L(lb3)
370	bc	L(lb2)
371	bc	L(lb1)
372L(lb7):
373	lb	a3, 6(a1)
374	sb	a3, 6(a0)
375L(lb6):
376	lb	a3, 5(a1)
377	sb	a3, 5(a0)
378L(lb5):
379	lb	a3, 4(a1)
380	sb	a3, 4(a0)
381L(lb4):
382	lb	a3, 3(a1)
383	sb	a3, 3(a0)
384L(lb3):
385	lb	a3, 2(a1)
386	sb	a3, 2(a0)
387L(lb2):
388	lb	a3, 1(a1)
389	sb	a3, 1(a0)
390L(lb1):
391	lb	a3, 0(a1)
392	sb	a3, 0(a0)
393
394	li	t9,8
395	subu	t8,t9,t8
396	PTR_SUBU a2,a2,t8
397	PTR_ADDU a0,a0,t8
398	PTR_ADDU a1,a1,t8
399L(lb0):
400
401	andi	t8,a1,(NSIZE-1)
402	lapc	t9,L(jtable)
403	PTR_LSA	t9,t8,t9,2
404	jrc	t9
405L(jtable):
406	bc	L(aligned)
407	bc	L(r6_unaligned1)
408	bc	L(r6_unaligned2)
409	bc	L(r6_unaligned3)
410# ifdef USE_DOUBLE
411	bc	L(r6_unaligned4)
412	bc	L(r6_unaligned5)
413	bc	L(r6_unaligned6)
414	bc	L(r6_unaligned7)
415# endif
416#endif /* R6_CODE */
417
418L(aligned):
419
420/*
421 * Now dst/src are both aligned to (word or double word) aligned addresses
422 * Set a2 to count how many bytes we have to copy after all the 64/128 byte
423 * chunks are copied and a3 to the dst pointer after all the 64/128 byte
424 * chunks have been copied.  We will loop, incrementing a0 and a1 until a0
425 * equals a3.
426 */
427
428	andi	t8,a2,NSIZEDMASK /* any whole 64-byte/128-byte chunks? */
429	beq	a2,t8,L(chkw)	 /* if a2==t8, no 64-byte/128-byte chunks */
430	PTR_SUBU a3,a2,t8	 /* subtract from a2 the reminder */
431	PTR_ADDU a3,a0,a3	 /* Now a3 is the final dst after loop */
432
433/* When in the loop we may prefetch with the 'prepare to store' hint,
434 * in this case the a0+x should not be past the "t0-32" address.  This
435 * means: for x=128 the last "safe" a0 address is "t0-160".  Alternatively,
436 * for x=64 the last "safe" a0 address is "t0-96" In the current version we
437 * will use "prefetch hint,128(a0)", so "t0-160" is the limit.
438 */
439#if defined(USE_PREFETCH) && (PREFETCH_STORE_HINT == PREFETCH_HINT_PREPAREFORSTORE)
440	PTR_ADDU t0,a0,a2		/* t0 is the "past the end" address */
441	PTR_SUBU t9,t0,PREFETCH_LIMIT	/* t9 is the "last safe pref" address */
442#endif
443	PREFETCH_FOR_LOAD  (0, a1)
444	PREFETCH_FOR_LOAD  (1, a1)
445	PREFETCH_FOR_LOAD  (2, a1)
446	PREFETCH_FOR_LOAD  (3, a1)
447#if defined(USE_PREFETCH) && (PREFETCH_STORE_HINT != PREFETCH_HINT_PREPAREFORSTORE)
448	PREFETCH_FOR_STORE (1, a0)
449	PREFETCH_FOR_STORE (2, a0)
450	PREFETCH_FOR_STORE (3, a0)
451#endif
452#if defined(RETURN_FIRST_PREFETCH) && defined(USE_PREFETCH)
453# if PREFETCH_STORE_HINT == PREFETCH_HINT_PREPAREFORSTORE
454	sltu    v1,t9,a0
455	bgtz    v1,L(skip_set)
456	nop
457	PTR_ADDIU v0,a0,(PREFETCH_CHUNK*4)
458L(skip_set):
459# else
460	PTR_ADDIU v0,a0,(PREFETCH_CHUNK*1)
461# endif
462#endif
463#if defined(RETURN_LAST_PREFETCH) && defined(USE_PREFETCH) \
464    && (PREFETCH_STORE_HINT != PREFETCH_HINT_PREPAREFORSTORE)
465	PTR_ADDIU v0,a0,(PREFETCH_CHUNK*3)
466# ifdef USE_DOUBLE
467	PTR_ADDIU v0,v0,32
468# endif
469#endif
470L(loop16w):
471	C_LD	t0,UNIT(0)(a1)
472#if defined(USE_PREFETCH) && (PREFETCH_STORE_HINT == PREFETCH_HINT_PREPAREFORSTORE)
473	sltu	v1,t9,a0		/* If a0 > t9 don't use next prefetch */
474	bgtz	v1,L(skip_pref)
475#endif
476	C_LD	t1,UNIT(1)(a1)
477#ifndef R6_CODE
478	PREFETCH_FOR_STORE (4, a0)
479	PREFETCH_FOR_STORE (5, a0)
480#else
481	PREFETCH_FOR_STORE (2, a0)
482#endif
483#if defined(RETURN_LAST_PREFETCH) && defined(USE_PREFETCH)
484	PTR_ADDIU v0,a0,(PREFETCH_CHUNK*5)
485# ifdef USE_DOUBLE
486	PTR_ADDIU v0,v0,32
487# endif
488#endif
489L(skip_pref):
490	C_LD	REG2,UNIT(2)(a1)
491	C_LD	REG3,UNIT(3)(a1)
492	C_LD	REG4,UNIT(4)(a1)
493	C_LD	REG5,UNIT(5)(a1)
494	C_LD	REG6,UNIT(6)(a1)
495	C_LD	REG7,UNIT(7)(a1)
496#ifndef R6_CODE
497	PREFETCH_FOR_LOAD (4, a1)
498#else
499	PREFETCH_FOR_LOAD (3, a1)
500#endif
501	C_ST	t0,UNIT(0)(a0)
502	C_ST	t1,UNIT(1)(a0)
503	C_ST	REG2,UNIT(2)(a0)
504	C_ST	REG3,UNIT(3)(a0)
505	C_ST	REG4,UNIT(4)(a0)
506	C_ST	REG5,UNIT(5)(a0)
507	C_ST	REG6,UNIT(6)(a0)
508	C_ST	REG7,UNIT(7)(a0)
509
510	C_LD	t0,UNIT(8)(a1)
511	C_LD	t1,UNIT(9)(a1)
512	C_LD	REG2,UNIT(10)(a1)
513	C_LD	REG3,UNIT(11)(a1)
514	C_LD	REG4,UNIT(12)(a1)
515	C_LD	REG5,UNIT(13)(a1)
516	C_LD	REG6,UNIT(14)(a1)
517	C_LD	REG7,UNIT(15)(a1)
518#ifndef R6_CODE
519	PREFETCH_FOR_LOAD (5, a1)
520#endif
521	C_ST	t0,UNIT(8)(a0)
522	C_ST	t1,UNIT(9)(a0)
523	C_ST	REG2,UNIT(10)(a0)
524	C_ST	REG3,UNIT(11)(a0)
525	C_ST	REG4,UNIT(12)(a0)
526	C_ST	REG5,UNIT(13)(a0)
527	C_ST	REG6,UNIT(14)(a0)
528	C_ST	REG7,UNIT(15)(a0)
529	PTR_ADDIU a0,a0,UNIT(16)	/* adding 64/128 to dest */
530	bne	a0,a3,L(loop16w)
531	PTR_ADDIU a1,a1,UNIT(16)	/* adding 64/128 to src */
532	move	a2,t8
533
534/* Here we have src and dest word-aligned but less than 64-bytes or
535 * 128 bytes to go.  Check for a 32(64) byte chunk and copy if if there
536 * is one.  Otherwise jump down to L(chk1w) to handle the tail end of
537 * the copy.
538 */
539
540L(chkw):
541	PREFETCH_FOR_LOAD (0, a1)
542	andi	t8,a2,NSIZEMASK	/* Is there a 32-byte/64-byte chunk.  */
543				/* The t8 is the reminder count past 32-bytes */
544	beq	a2,t8,L(chk1w)	/* When a2=t8, no 32-byte chunk  */
545	nop
546	C_LD	t0,UNIT(0)(a1)
547	C_LD	t1,UNIT(1)(a1)
548	C_LD	REG2,UNIT(2)(a1)
549	C_LD	REG3,UNIT(3)(a1)
550	C_LD	REG4,UNIT(4)(a1)
551	C_LD	REG5,UNIT(5)(a1)
552	C_LD	REG6,UNIT(6)(a1)
553	C_LD	REG7,UNIT(7)(a1)
554	PTR_ADDIU a1,a1,UNIT(8)
555	C_ST	t0,UNIT(0)(a0)
556	C_ST	t1,UNIT(1)(a0)
557	C_ST	REG2,UNIT(2)(a0)
558	C_ST	REG3,UNIT(3)(a0)
559	C_ST	REG4,UNIT(4)(a0)
560	C_ST	REG5,UNIT(5)(a0)
561	C_ST	REG6,UNIT(6)(a0)
562	C_ST	REG7,UNIT(7)(a0)
563	PTR_ADDIU a0,a0,UNIT(8)
564
565/*
566 * Here we have less than 32(64) bytes to copy.  Set up for a loop to
567 * copy one word (or double word) at a time.  Set a2 to count how many
568 * bytes we have to copy after all the word (or double word) chunks are
569 * copied and a3 to the dst pointer after all the (d)word chunks have
570 * been copied.  We will loop, incrementing a0 and a1 until a0 equals a3.
571 */
572L(chk1w):
573	andi	a2,t8,(NSIZE-1)	/* a2 is the reminder past one (d)word chunks */
574	beq	a2,t8,L(lastb)
575	PTR_SUBU a3,t8,a2	/* a3 is count of bytes in one (d)word chunks */
576	PTR_ADDU a3,a0,a3	/* a3 is the dst address after loop */
577
578/* copying in words (4-byte or 8-byte chunks) */
579L(wordCopy_loop):
580	C_LD	REG3,UNIT(0)(a1)
581	PTR_ADDIU a0,a0,UNIT(1)
582	PTR_ADDIU a1,a1,UNIT(1)
583	bne	a0,a3,L(wordCopy_loop)
584	C_ST	REG3,UNIT(-1)(a0)
585
586/* Copy the last 8 (or 16) bytes */
587L(lastb):
588	blez	a2,L(leave)
589	PTR_ADDU a3,a0,a2	/* a3 is the last dst address */
590L(lastbloop):
591	lb	v1,0(a1)
592	PTR_ADDIU a0,a0,1
593	PTR_ADDIU a1,a1,1
594	bne	a0,a3,L(lastbloop)
595	sb	v1,-1(a0)
596L(leave):
597	j	ra
598	nop
599
600#ifndef R6_CODE
601/*
602 * UNALIGNED case, got here with a3 = "negu a0"
603 * This code is nearly identical to the aligned code above
604 * but only the destination (not the source) gets aligned
605 * so we need to do partial loads of the source followed
606 * by normal stores to the destination (once we have aligned
607 * the destination).
608 */
609
610L(unaligned):
611	andi	a3,a3,(NSIZE-1)	/* copy a3 bytes to align a0/a1 */
612	beqz	a3,L(ua_chk16w) /* if a3=0, it is already aligned */
613	PTR_SUBU a2,a2,a3	/* a2 is the remining bytes count */
614
615	C_LDHI	v1,UNIT(0)(a1)
616	C_LDLO	v1,UNITM1(1)(a1)
617	PTR_ADDU a1,a1,a3
618	C_STHI	v1,UNIT(0)(a0)
619	PTR_ADDU a0,a0,a3
620
621/*
622 *  Now the destination (but not the source) is aligned
623 * Set a2 to count how many bytes we have to copy after all the 64/128 byte
624 * chunks are copied and a3 to the dst pointer after all the 64/128 byte
625 * chunks have been copied.  We will loop, incrementing a0 and a1 until a0
626 * equals a3.
627 */
628
629L(ua_chk16w):
630	andi	t8,a2,NSIZEDMASK /* any whole 64-byte/128-byte chunks? */
631	beq	a2,t8,L(ua_chkw) /* if a2==t8, no 64-byte/128-byte chunks */
632	PTR_SUBU a3,a2,t8	 /* subtract from a2 the reminder */
633	PTR_ADDU a3,a0,a3	 /* Now a3 is the final dst after loop */
634
635# if defined(USE_PREFETCH) && (PREFETCH_STORE_HINT == PREFETCH_HINT_PREPAREFORSTORE)
636	PTR_ADDU t0,a0,a2	  /* t0 is the "past the end" address */
637	PTR_SUBU t9,t0,PREFETCH_LIMIT /* t9 is the "last safe pref" address */
638# endif
639	PREFETCH_FOR_LOAD  (0, a1)
640	PREFETCH_FOR_LOAD  (1, a1)
641	PREFETCH_FOR_LOAD  (2, a1)
642# if defined(USE_PREFETCH) && (PREFETCH_STORE_HINT != PREFETCH_HINT_PREPAREFORSTORE)
643	PREFETCH_FOR_STORE (1, a0)
644	PREFETCH_FOR_STORE (2, a0)
645	PREFETCH_FOR_STORE (3, a0)
646# endif
647# if defined(RETURN_FIRST_PREFETCH) && defined(USE_PREFETCH)
648#  if (PREFETCH_STORE_HINT == PREFETCH_HINT_PREPAREFORSTORE)
649	sltu    v1,t9,a0
650	bgtz    v1,L(ua_skip_set)
651	nop
652	PTR_ADDIU v0,a0,(PREFETCH_CHUNK*4)
653L(ua_skip_set):
654#  else
655	PTR_ADDIU v0,a0,(PREFETCH_CHUNK*1)
656#  endif
657# endif
658L(ua_loop16w):
659	PREFETCH_FOR_LOAD  (3, a1)
660	C_LDHI	t0,UNIT(0)(a1)
661	C_LDHI	t1,UNIT(1)(a1)
662	C_LDHI	REG2,UNIT(2)(a1)
663# if defined(USE_PREFETCH) && (PREFETCH_STORE_HINT == PREFETCH_HINT_PREPAREFORSTORE)
664	sltu	v1,t9,a0
665	bgtz	v1,L(ua_skip_pref)
666# endif
667	C_LDHI	REG3,UNIT(3)(a1)
668	PREFETCH_FOR_STORE (4, a0)
669	PREFETCH_FOR_STORE (5, a0)
670L(ua_skip_pref):
671	C_LDHI	REG4,UNIT(4)(a1)
672	C_LDHI	REG5,UNIT(5)(a1)
673	C_LDHI	REG6,UNIT(6)(a1)
674	C_LDHI	REG7,UNIT(7)(a1)
675	C_LDLO	t0,UNITM1(1)(a1)
676	C_LDLO	t1,UNITM1(2)(a1)
677	C_LDLO	REG2,UNITM1(3)(a1)
678	C_LDLO	REG3,UNITM1(4)(a1)
679	C_LDLO	REG4,UNITM1(5)(a1)
680	C_LDLO	REG5,UNITM1(6)(a1)
681	C_LDLO	REG6,UNITM1(7)(a1)
682	C_LDLO	REG7,UNITM1(8)(a1)
683        PREFETCH_FOR_LOAD (4, a1)
684	C_ST	t0,UNIT(0)(a0)
685	C_ST	t1,UNIT(1)(a0)
686	C_ST	REG2,UNIT(2)(a0)
687	C_ST	REG3,UNIT(3)(a0)
688	C_ST	REG4,UNIT(4)(a0)
689	C_ST	REG5,UNIT(5)(a0)
690	C_ST	REG6,UNIT(6)(a0)
691	C_ST	REG7,UNIT(7)(a0)
692	C_LDHI	t0,UNIT(8)(a1)
693	C_LDHI	t1,UNIT(9)(a1)
694	C_LDHI	REG2,UNIT(10)(a1)
695	C_LDHI	REG3,UNIT(11)(a1)
696	C_LDHI	REG4,UNIT(12)(a1)
697	C_LDHI	REG5,UNIT(13)(a1)
698	C_LDHI	REG6,UNIT(14)(a1)
699	C_LDHI	REG7,UNIT(15)(a1)
700	C_LDLO	t0,UNITM1(9)(a1)
701	C_LDLO	t1,UNITM1(10)(a1)
702	C_LDLO	REG2,UNITM1(11)(a1)
703	C_LDLO	REG3,UNITM1(12)(a1)
704	C_LDLO	REG4,UNITM1(13)(a1)
705	C_LDLO	REG5,UNITM1(14)(a1)
706	C_LDLO	REG6,UNITM1(15)(a1)
707	C_LDLO	REG7,UNITM1(16)(a1)
708        PREFETCH_FOR_LOAD (5, a1)
709	C_ST	t0,UNIT(8)(a0)
710	C_ST	t1,UNIT(9)(a0)
711	C_ST	REG2,UNIT(10)(a0)
712	C_ST	REG3,UNIT(11)(a0)
713	C_ST	REG4,UNIT(12)(a0)
714	C_ST	REG5,UNIT(13)(a0)
715	C_ST	REG6,UNIT(14)(a0)
716	C_ST	REG7,UNIT(15)(a0)
717	PTR_ADDIU a0,a0,UNIT(16)	/* adding 64/128 to dest */
718	bne	a0,a3,L(ua_loop16w)
719	PTR_ADDIU a1,a1,UNIT(16)	/* adding 64/128 to src */
720	move	a2,t8
721
722/* Here we have src and dest word-aligned but less than 64-bytes or
723 * 128 bytes to go.  Check for a 32(64) byte chunk and copy if if there
724 * is one.  Otherwise jump down to L(ua_chk1w) to handle the tail end of
725 * the copy.  */
726
727L(ua_chkw):
728	PREFETCH_FOR_LOAD (0, a1)
729	andi	t8,a2,NSIZEMASK	  /* Is there a 32-byte/64-byte chunk.  */
730				  /* t8 is the reminder count past 32-bytes */
731	beq	a2,t8,L(ua_chk1w) /* When a2=t8, no 32-byte chunk */
732	nop
733	C_LDHI	t0,UNIT(0)(a1)
734	C_LDHI	t1,UNIT(1)(a1)
735	C_LDHI	REG2,UNIT(2)(a1)
736	C_LDHI	REG3,UNIT(3)(a1)
737	C_LDHI	REG4,UNIT(4)(a1)
738	C_LDHI	REG5,UNIT(5)(a1)
739	C_LDHI	REG6,UNIT(6)(a1)
740	C_LDHI	REG7,UNIT(7)(a1)
741	C_LDLO	t0,UNITM1(1)(a1)
742	C_LDLO	t1,UNITM1(2)(a1)
743	C_LDLO	REG2,UNITM1(3)(a1)
744	C_LDLO	REG3,UNITM1(4)(a1)
745	C_LDLO	REG4,UNITM1(5)(a1)
746	C_LDLO	REG5,UNITM1(6)(a1)
747	C_LDLO	REG6,UNITM1(7)(a1)
748	C_LDLO	REG7,UNITM1(8)(a1)
749	PTR_ADDIU a1,a1,UNIT(8)
750	C_ST	t0,UNIT(0)(a0)
751	C_ST	t1,UNIT(1)(a0)
752	C_ST	REG2,UNIT(2)(a0)
753	C_ST	REG3,UNIT(3)(a0)
754	C_ST	REG4,UNIT(4)(a0)
755	C_ST	REG5,UNIT(5)(a0)
756	C_ST	REG6,UNIT(6)(a0)
757	C_ST	REG7,UNIT(7)(a0)
758	PTR_ADDIU a0,a0,UNIT(8)
759/*
760 * Here we have less than 32(64) bytes to copy.  Set up for a loop to
761 * copy one word (or double word) at a time.
762 */
763L(ua_chk1w):
764	andi	a2,t8,(NSIZE-1)	/* a2 is the reminder past one (d)word chunks */
765	beq	a2,t8,L(ua_smallCopy)
766	PTR_SUBU a3,t8,a2	/* a3 is count of bytes in one (d)word chunks */
767	PTR_ADDU a3,a0,a3	/* a3 is the dst address after loop */
768
769/* copying in words (4-byte or 8-byte chunks) */
770L(ua_wordCopy_loop):
771	C_LDHI	v1,UNIT(0)(a1)
772	C_LDLO	v1,UNITM1(1)(a1)
773	PTR_ADDIU a0,a0,UNIT(1)
774	PTR_ADDIU a1,a1,UNIT(1)
775	bne	a0,a3,L(ua_wordCopy_loop)
776	C_ST	v1,UNIT(-1)(a0)
777
778/* Copy the last 8 (or 16) bytes */
779L(ua_smallCopy):
780	beqz	a2,L(leave)
781	PTR_ADDU a3,a0,a2	/* a3 is the last dst address */
782L(ua_smallCopy_loop):
783	lb	v1,0(a1)
784	PTR_ADDIU a0,a0,1
785	PTR_ADDIU a1,a1,1
786	bne	a0,a3,L(ua_smallCopy_loop)
787	sb	v1,-1(a0)
788
789	j	ra
790	nop
791
792#else /* R6_CODE */
793
794# if __MIPSEB
795#  define SWAP_REGS(X,Y) X, Y
796#  define ALIGN_OFFSET(N) (N)
797# else
798#  define SWAP_REGS(X,Y) Y, X
799#  define ALIGN_OFFSET(N) (NSIZE-N)
800# endif
801# define R6_UNALIGNED_WORD_COPY(BYTEOFFSET) \
802	andi	REG7, a2, (NSIZE-1);/* REG7 is # of bytes to by bytes.     */ \
803	beq	REG7, a2, L(lastb); /* Check for bytes to copy by word	   */ \
804	PTR_SUBU a3, a2, REG7;	/* a3 is number of bytes to be copied in   */ \
805				/* (d)word chunks.			   */ \
806	move	a2, REG7;	/* a2 is # of bytes to copy byte by byte   */ \
807				/* after word loop is finished.		   */ \
808	PTR_ADDU REG6, a0, a3;	/* REG6 is the dst address after loop.	   */ \
809	PTR_SUBU REG2, a1, t8;	/* REG2 is the aligned src address.	   */ \
810	PTR_ADDU a1, a1, a3;	/* a1 is addr of source after word loop.   */ \
811	C_LD	t0, UNIT(0)(REG2);  /* Load first part of source.	   */ \
812L(r6_ua_wordcopy##BYTEOFFSET):						      \
813	C_LD	t1, UNIT(1)(REG2);  /* Load second part of source.	   */ \
814	C_ALIGN	REG3, SWAP_REGS(t1,t0), ALIGN_OFFSET(BYTEOFFSET);	      \
815	PTR_ADDIU a0, a0, UNIT(1);  /* Increment destination pointer.	   */ \
816	PTR_ADDIU REG2, REG2, UNIT(1); /* Increment aligned source pointer.*/ \
817	move	t0, t1;		/* Move second part of source to first.	   */ \
818	bne	a0, REG6,L(r6_ua_wordcopy##BYTEOFFSET);			      \
819	C_ST	REG3, UNIT(-1)(a0);					      \
820	j	L(lastb);						      \
821	nop
822
823	/* We are generating R6 code, the destination is 4 byte aligned and
824	   the source is not 4 byte aligned. t8 is 1, 2, or 3 depending on the
825           alignment of the source.  */
826
827L(r6_unaligned1):
828	R6_UNALIGNED_WORD_COPY(1)
829L(r6_unaligned2):
830	R6_UNALIGNED_WORD_COPY(2)
831L(r6_unaligned3):
832	R6_UNALIGNED_WORD_COPY(3)
833# ifdef USE_DOUBLE
834L(r6_unaligned4):
835	R6_UNALIGNED_WORD_COPY(4)
836L(r6_unaligned5):
837	R6_UNALIGNED_WORD_COPY(5)
838L(r6_unaligned6):
839	R6_UNALIGNED_WORD_COPY(6)
840L(r6_unaligned7):
841	R6_UNALIGNED_WORD_COPY(7)
842# endif
843#endif /* R6_CODE */
844
845	.set	at
846	.set	reorder
847END(MEMCPY_NAME)
848#ifndef __ANDROID__
849# ifdef _LIBC
850libc_hidden_builtin_def (MEMCPY_NAME)
851# endif
852#endif
853