xref: /linux/arch/mips/lib/memcpy.S (revision fd639726bf15fca8ee1a00dce8e0096d0ad9bd18)
1/*
2 * This file is subject to the terms and conditions of the GNU General Public
3 * License.  See the file "COPYING" in the main directory of this archive
4 * for more details.
5 *
6 * Unified implementation of memcpy, memmove and the __copy_user backend.
7 *
8 * Copyright (C) 1998, 99, 2000, 01, 2002 Ralf Baechle (ralf@gnu.org)
9 * Copyright (C) 1999, 2000, 01, 2002 Silicon Graphics, Inc.
10 * Copyright (C) 2002 Broadcom, Inc.
11 *   memcpy/copy_user author: Mark Vandevoorde
12 * Copyright (C) 2007  Maciej W. Rozycki
13 * Copyright (C) 2014 Imagination Technologies Ltd.
14 *
15 * Mnemonic names for arguments to memcpy/__copy_user
16 */
17
18/*
19 * Hack to resolve longstanding prefetch issue
20 *
21 * Prefetching may be fatal on some systems if we're prefetching beyond the
22 * end of memory on some systems.  It's also a seriously bad idea on non
23 * dma-coherent systems.
24 */
25#ifdef CONFIG_DMA_NONCOHERENT
26#undef CONFIG_CPU_HAS_PREFETCH
27#endif
28#ifdef CONFIG_MIPS_MALTA
29#undef CONFIG_CPU_HAS_PREFETCH
30#endif
31#ifdef CONFIG_CPU_MIPSR6
32#undef CONFIG_CPU_HAS_PREFETCH
33#endif
34
35#include <asm/asm.h>
36#include <asm/asm-offsets.h>
37#include <asm/export.h>
38#include <asm/regdef.h>
39
40#define dst a0
41#define src a1
42#define len a2
43
44/*
45 * Spec
46 *
47 * memcpy copies len bytes from src to dst and sets v0 to dst.
48 * It assumes that
49 *   - src and dst don't overlap
50 *   - src is readable
51 *   - dst is writable
52 * memcpy uses the standard calling convention
53 *
54 * __copy_user copies up to len bytes from src to dst and sets a2 (len) to
55 * the number of uncopied bytes due to an exception caused by a read or write.
56 * __copy_user assumes that src and dst don't overlap, and that the call is
57 * implementing one of the following:
58 *   copy_to_user
59 *     - src is readable  (no exceptions when reading src)
60 *   copy_from_user
61 *     - dst is writable  (no exceptions when writing dst)
62 * __copy_user uses a non-standard calling convention; see
63 * include/asm-mips/uaccess.h
64 *
65 * When an exception happens on a load, the handler must
66 # ensure that all of the destination buffer is overwritten to prevent
67 * leaking information to user mode programs.
68 */
69
70/*
71 * Implementation
72 */
73
74/*
75 * The exception handler for loads requires that:
76 *  1- AT contain the address of the byte just past the end of the source
77 *     of the copy,
78 *  2- src_entry <= src < AT, and
79 *  3- (dst - src) == (dst_entry - src_entry),
80 * The _entry suffix denotes values when __copy_user was called.
81 *
82 * (1) is set up up by uaccess.h and maintained by not writing AT in copy_user
83 * (2) is met by incrementing src by the number of bytes copied
84 * (3) is met by not doing loads between a pair of increments of dst and src
85 *
86 * The exception handlers for stores adjust len (if necessary) and return.
87 * These handlers do not need to overwrite any data.
88 *
89 * For __rmemcpy and memmove an exception is always a kernel bug, therefore
90 * they're not protected.
91 */
92
93/* Instruction type */
94#define LD_INSN 1
95#define ST_INSN 2
96/* Pretech type */
97#define SRC_PREFETCH 1
98#define DST_PREFETCH 2
99#define LEGACY_MODE 1
100#define EVA_MODE    2
101#define USEROP   1
102#define KERNELOP 2
103
104/*
105 * Wrapper to add an entry in the exception table
106 * in case the insn causes a memory exception.
107 * Arguments:
108 * insn    : Load/store instruction
109 * type    : Instruction type
110 * reg     : Register
111 * addr    : Address
112 * handler : Exception handler
113 */
114
115#define EXC(insn, type, reg, addr, handler)			\
116	.if \mode == LEGACY_MODE;				\
1179:		insn reg, addr;					\
118		.section __ex_table,"a";			\
119		PTR	9b, handler;				\
120		.previous;					\
121	/* This is assembled in EVA mode */			\
122	.else;							\
123		/* If loading from user or storing to user */	\
124		.if ((\from == USEROP) && (type == LD_INSN)) || \
125		    ((\to == USEROP) && (type == ST_INSN));	\
1269:			__BUILD_EVA_INSN(insn##e, reg, addr);	\
127			.section __ex_table,"a";		\
128			PTR	9b, handler;			\
129			.previous;				\
130		.else;						\
131			/*					\
132			 *  Still in EVA, but no need for	\
133			 * exception handler or EVA insn	\
134			 */					\
135			insn reg, addr;				\
136		.endif;						\
137	.endif
138
139/*
140 * Only on the 64-bit kernel we can made use of 64-bit registers.
141 */
142#ifdef CONFIG_64BIT
143#define USE_DOUBLE
144#endif
145
146#ifdef USE_DOUBLE
147
148#define LOADK ld /* No exception */
149#define LOAD(reg, addr, handler)	EXC(ld, LD_INSN, reg, addr, handler)
150#define LOADL(reg, addr, handler)	EXC(ldl, LD_INSN, reg, addr, handler)
151#define LOADR(reg, addr, handler)	EXC(ldr, LD_INSN, reg, addr, handler)
152#define STOREL(reg, addr, handler)	EXC(sdl, ST_INSN, reg, addr, handler)
153#define STORER(reg, addr, handler)	EXC(sdr, ST_INSN, reg, addr, handler)
154#define STORE(reg, addr, handler)	EXC(sd, ST_INSN, reg, addr, handler)
155#define ADD    daddu
156#define SUB    dsubu
157#define SRL    dsrl
158#define SRA    dsra
159#define SLL    dsll
160#define SLLV   dsllv
161#define SRLV   dsrlv
162#define NBYTES 8
163#define LOG_NBYTES 3
164
165/*
166 * As we are sharing code base with the mips32 tree (which use the o32 ABI
167 * register definitions). We need to redefine the register definitions from
168 * the n64 ABI register naming to the o32 ABI register naming.
169 */
170#undef t0
171#undef t1
172#undef t2
173#undef t3
174#define t0	$8
175#define t1	$9
176#define t2	$10
177#define t3	$11
178#define t4	$12
179#define t5	$13
180#define t6	$14
181#define t7	$15
182
183#else
184
185#define LOADK lw /* No exception */
186#define LOAD(reg, addr, handler)	EXC(lw, LD_INSN, reg, addr, handler)
187#define LOADL(reg, addr, handler)	EXC(lwl, LD_INSN, reg, addr, handler)
188#define LOADR(reg, addr, handler)	EXC(lwr, LD_INSN, reg, addr, handler)
189#define STOREL(reg, addr, handler)	EXC(swl, ST_INSN, reg, addr, handler)
190#define STORER(reg, addr, handler)	EXC(swr, ST_INSN, reg, addr, handler)
191#define STORE(reg, addr, handler)	EXC(sw, ST_INSN, reg, addr, handler)
192#define ADD    addu
193#define SUB    subu
194#define SRL    srl
195#define SLL    sll
196#define SRA    sra
197#define SLLV   sllv
198#define SRLV   srlv
199#define NBYTES 4
200#define LOG_NBYTES 2
201
202#endif /* USE_DOUBLE */
203
204#define LOADB(reg, addr, handler)	EXC(lb, LD_INSN, reg, addr, handler)
205#define STOREB(reg, addr, handler)	EXC(sb, ST_INSN, reg, addr, handler)
206
207#define _PREF(hint, addr, type)						\
208	.if \mode == LEGACY_MODE;					\
209		PREF(hint, addr);					\
210	.else;								\
211		.if ((\from == USEROP) && (type == SRC_PREFETCH)) ||	\
212		    ((\to == USEROP) && (type == DST_PREFETCH));	\
213			/*						\
214			 * PREFE has only 9 bits for the offset		\
215			 * compared to PREF which has 16, so it may	\
216			 * need to use the $at register but this	\
217			 * register should remain intact because it's	\
218			 * used later on. Therefore use $v1.		\
219			 */						\
220			.set at=v1;					\
221			PREFE(hint, addr);				\
222			.set noat;					\
223		.else;							\
224			PREF(hint, addr);				\
225		.endif;							\
226	.endif
227
228#define PREFS(hint, addr) _PREF(hint, addr, SRC_PREFETCH)
229#define PREFD(hint, addr) _PREF(hint, addr, DST_PREFETCH)
230
231#ifdef CONFIG_CPU_LITTLE_ENDIAN
232#define LDFIRST LOADR
233#define LDREST	LOADL
234#define STFIRST STORER
235#define STREST	STOREL
236#define SHIFT_DISCARD SLLV
237#else
238#define LDFIRST LOADL
239#define LDREST	LOADR
240#define STFIRST STOREL
241#define STREST	STORER
242#define SHIFT_DISCARD SRLV
243#endif
244
245#define FIRST(unit) ((unit)*NBYTES)
246#define REST(unit)  (FIRST(unit)+NBYTES-1)
247#define UNIT(unit)  FIRST(unit)
248
249#define ADDRMASK (NBYTES-1)
250
251	.text
252	.set	noreorder
253#ifndef CONFIG_CPU_DADDI_WORKAROUNDS
254	.set	noat
255#else
256	.set	at=v1
257#endif
258
259	.align	5
260
261	/*
262	 * Macro to build the __copy_user common code
263	 * Arguments:
264	 * mode : LEGACY_MODE or EVA_MODE
265	 * from : Source operand. USEROP or KERNELOP
266	 * to   : Destination operand. USEROP or KERNELOP
267	 */
268	.macro __BUILD_COPY_USER mode, from, to
269
270	/* initialize __memcpy if this the first time we execute this macro */
271	.ifnotdef __memcpy
272	.set __memcpy, 1
273	.hidden __memcpy /* make sure it does not leak */
274	.endif
275
276	/*
277	 * Note: dst & src may be unaligned, len may be 0
278	 * Temps
279	 */
280#define rem t8
281
282	R10KCBARRIER(0(ra))
283	/*
284	 * The "issue break"s below are very approximate.
285	 * Issue delays for dcache fills will perturb the schedule, as will
286	 * load queue full replay traps, etc.
287	 *
288	 * If len < NBYTES use byte operations.
289	 */
290	PREFS(	0, 0(src) )
291	PREFD(	1, 0(dst) )
292	sltu	t2, len, NBYTES
293	and	t1, dst, ADDRMASK
294	PREFS(	0, 1*32(src) )
295	PREFD(	1, 1*32(dst) )
296	bnez	t2, .Lcopy_bytes_checklen\@
297	 and	t0, src, ADDRMASK
298	PREFS(	0, 2*32(src) )
299	PREFD(	1, 2*32(dst) )
300#ifndef CONFIG_CPU_MIPSR6
301	bnez	t1, .Ldst_unaligned\@
302	 nop
303	bnez	t0, .Lsrc_unaligned_dst_aligned\@
304#else
305	or	t0, t0, t1
306	bnez	t0, .Lcopy_unaligned_bytes\@
307#endif
308	/*
309	 * use delay slot for fall-through
310	 * src and dst are aligned; need to compute rem
311	 */
312.Lboth_aligned\@:
313	 SRL	t0, len, LOG_NBYTES+3	 # +3 for 8 units/iter
314	beqz	t0, .Lcleanup_both_aligned\@ # len < 8*NBYTES
315	 and	rem, len, (8*NBYTES-1)	 # rem = len % (8*NBYTES)
316	PREFS(	0, 3*32(src) )
317	PREFD(	1, 3*32(dst) )
318	.align	4
3191:
320	R10KCBARRIER(0(ra))
321	LOAD(t0, UNIT(0)(src), .Ll_exc\@)
322	LOAD(t1, UNIT(1)(src), .Ll_exc_copy\@)
323	LOAD(t2, UNIT(2)(src), .Ll_exc_copy\@)
324	LOAD(t3, UNIT(3)(src), .Ll_exc_copy\@)
325	SUB	len, len, 8*NBYTES
326	LOAD(t4, UNIT(4)(src), .Ll_exc_copy\@)
327	LOAD(t7, UNIT(5)(src), .Ll_exc_copy\@)
328	STORE(t0, UNIT(0)(dst),	.Ls_exc_p8u\@)
329	STORE(t1, UNIT(1)(dst),	.Ls_exc_p7u\@)
330	LOAD(t0, UNIT(6)(src), .Ll_exc_copy\@)
331	LOAD(t1, UNIT(7)(src), .Ll_exc_copy\@)
332	ADD	src, src, 8*NBYTES
333	ADD	dst, dst, 8*NBYTES
334	STORE(t2, UNIT(-6)(dst), .Ls_exc_p6u\@)
335	STORE(t3, UNIT(-5)(dst), .Ls_exc_p5u\@)
336	STORE(t4, UNIT(-4)(dst), .Ls_exc_p4u\@)
337	STORE(t7, UNIT(-3)(dst), .Ls_exc_p3u\@)
338	STORE(t0, UNIT(-2)(dst), .Ls_exc_p2u\@)
339	STORE(t1, UNIT(-1)(dst), .Ls_exc_p1u\@)
340	PREFS(	0, 8*32(src) )
341	PREFD(	1, 8*32(dst) )
342	bne	len, rem, 1b
343	 nop
344
345	/*
346	 * len == rem == the number of bytes left to copy < 8*NBYTES
347	 */
348.Lcleanup_both_aligned\@:
349	beqz	len, .Ldone\@
350	 sltu	t0, len, 4*NBYTES
351	bnez	t0, .Lless_than_4units\@
352	 and	rem, len, (NBYTES-1)	# rem = len % NBYTES
353	/*
354	 * len >= 4*NBYTES
355	 */
356	LOAD( t0, UNIT(0)(src),	.Ll_exc\@)
357	LOAD( t1, UNIT(1)(src),	.Ll_exc_copy\@)
358	LOAD( t2, UNIT(2)(src),	.Ll_exc_copy\@)
359	LOAD( t3, UNIT(3)(src),	.Ll_exc_copy\@)
360	SUB	len, len, 4*NBYTES
361	ADD	src, src, 4*NBYTES
362	R10KCBARRIER(0(ra))
363	STORE(t0, UNIT(0)(dst),	.Ls_exc_p4u\@)
364	STORE(t1, UNIT(1)(dst),	.Ls_exc_p3u\@)
365	STORE(t2, UNIT(2)(dst),	.Ls_exc_p2u\@)
366	STORE(t3, UNIT(3)(dst),	.Ls_exc_p1u\@)
367	.set	reorder				/* DADDI_WAR */
368	ADD	dst, dst, 4*NBYTES
369	beqz	len, .Ldone\@
370	.set	noreorder
371.Lless_than_4units\@:
372	/*
373	 * rem = len % NBYTES
374	 */
375	beq	rem, len, .Lcopy_bytes\@
376	 nop
3771:
378	R10KCBARRIER(0(ra))
379	LOAD(t0, 0(src), .Ll_exc\@)
380	ADD	src, src, NBYTES
381	SUB	len, len, NBYTES
382	STORE(t0, 0(dst), .Ls_exc_p1u\@)
383	.set	reorder				/* DADDI_WAR */
384	ADD	dst, dst, NBYTES
385	bne	rem, len, 1b
386	.set	noreorder
387
388#ifndef CONFIG_CPU_MIPSR6
389	/*
390	 * src and dst are aligned, need to copy rem bytes (rem < NBYTES)
391	 * A loop would do only a byte at a time with possible branch
392	 * mispredicts.	 Can't do an explicit LOAD dst,mask,or,STORE
393	 * because can't assume read-access to dst.  Instead, use
394	 * STREST dst, which doesn't require read access to dst.
395	 *
396	 * This code should perform better than a simple loop on modern,
397	 * wide-issue mips processors because the code has fewer branches and
398	 * more instruction-level parallelism.
399	 */
400#define bits t2
401	beqz	len, .Ldone\@
402	 ADD	t1, dst, len	# t1 is just past last byte of dst
403	li	bits, 8*NBYTES
404	SLL	rem, len, 3	# rem = number of bits to keep
405	LOAD(t0, 0(src), .Ll_exc\@)
406	SUB	bits, bits, rem # bits = number of bits to discard
407	SHIFT_DISCARD t0, t0, bits
408	STREST(t0, -1(t1), .Ls_exc\@)
409	jr	ra
410	 move	len, zero
411.Ldst_unaligned\@:
412	/*
413	 * dst is unaligned
414	 * t0 = src & ADDRMASK
415	 * t1 = dst & ADDRMASK; T1 > 0
416	 * len >= NBYTES
417	 *
418	 * Copy enough bytes to align dst
419	 * Set match = (src and dst have same alignment)
420	 */
421#define match rem
422	LDFIRST(t3, FIRST(0)(src), .Ll_exc\@)
423	ADD	t2, zero, NBYTES
424	LDREST(t3, REST(0)(src), .Ll_exc_copy\@)
425	SUB	t2, t2, t1	# t2 = number of bytes copied
426	xor	match, t0, t1
427	R10KCBARRIER(0(ra))
428	STFIRST(t3, FIRST(0)(dst), .Ls_exc\@)
429	beq	len, t2, .Ldone\@
430	 SUB	len, len, t2
431	ADD	dst, dst, t2
432	beqz	match, .Lboth_aligned\@
433	 ADD	src, src, t2
434
435.Lsrc_unaligned_dst_aligned\@:
436	SRL	t0, len, LOG_NBYTES+2	 # +2 for 4 units/iter
437	PREFS(	0, 3*32(src) )
438	beqz	t0, .Lcleanup_src_unaligned\@
439	 and	rem, len, (4*NBYTES-1)	 # rem = len % 4*NBYTES
440	PREFD(	1, 3*32(dst) )
4411:
442/*
443 * Avoid consecutive LD*'s to the same register since some mips
444 * implementations can't issue them in the same cycle.
445 * It's OK to load FIRST(N+1) before REST(N) because the two addresses
446 * are to the same unit (unless src is aligned, but it's not).
447 */
448	R10KCBARRIER(0(ra))
449	LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
450	LDFIRST(t1, FIRST(1)(src), .Ll_exc_copy\@)
451	SUB	len, len, 4*NBYTES
452	LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
453	LDREST(t1, REST(1)(src), .Ll_exc_copy\@)
454	LDFIRST(t2, FIRST(2)(src), .Ll_exc_copy\@)
455	LDFIRST(t3, FIRST(3)(src), .Ll_exc_copy\@)
456	LDREST(t2, REST(2)(src), .Ll_exc_copy\@)
457	LDREST(t3, REST(3)(src), .Ll_exc_copy\@)
458	PREFS(	0, 9*32(src) )		# 0 is PREF_LOAD  (not streamed)
459	ADD	src, src, 4*NBYTES
460#ifdef CONFIG_CPU_SB1
461	nop				# improves slotting
462#endif
463	STORE(t0, UNIT(0)(dst),	.Ls_exc_p4u\@)
464	STORE(t1, UNIT(1)(dst),	.Ls_exc_p3u\@)
465	STORE(t2, UNIT(2)(dst),	.Ls_exc_p2u\@)
466	STORE(t3, UNIT(3)(dst),	.Ls_exc_p1u\@)
467	PREFD(	1, 9*32(dst) )		# 1 is PREF_STORE (not streamed)
468	.set	reorder				/* DADDI_WAR */
469	ADD	dst, dst, 4*NBYTES
470	bne	len, rem, 1b
471	.set	noreorder
472
473.Lcleanup_src_unaligned\@:
474	beqz	len, .Ldone\@
475	 and	rem, len, NBYTES-1  # rem = len % NBYTES
476	beq	rem, len, .Lcopy_bytes\@
477	 nop
4781:
479	R10KCBARRIER(0(ra))
480	LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
481	LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
482	ADD	src, src, NBYTES
483	SUB	len, len, NBYTES
484	STORE(t0, 0(dst), .Ls_exc_p1u\@)
485	.set	reorder				/* DADDI_WAR */
486	ADD	dst, dst, NBYTES
487	bne	len, rem, 1b
488	.set	noreorder
489
490#endif /* !CONFIG_CPU_MIPSR6 */
491.Lcopy_bytes_checklen\@:
492	beqz	len, .Ldone\@
493	 nop
494.Lcopy_bytes\@:
495	/* 0 < len < NBYTES  */
496	R10KCBARRIER(0(ra))
497#define COPY_BYTE(N)			\
498	LOADB(t0, N(src), .Ll_exc\@);	\
499	SUB	len, len, 1;		\
500	beqz	len, .Ldone\@;		\
501	STOREB(t0, N(dst), .Ls_exc_p1\@)
502
503	COPY_BYTE(0)
504	COPY_BYTE(1)
505#ifdef USE_DOUBLE
506	COPY_BYTE(2)
507	COPY_BYTE(3)
508	COPY_BYTE(4)
509	COPY_BYTE(5)
510#endif
511	LOADB(t0, NBYTES-2(src), .Ll_exc\@)
512	SUB	len, len, 1
513	jr	ra
514	STOREB(t0, NBYTES-2(dst), .Ls_exc_p1\@)
515.Ldone\@:
516	jr	ra
517	 nop
518
519#ifdef CONFIG_CPU_MIPSR6
520.Lcopy_unaligned_bytes\@:
5211:
522	COPY_BYTE(0)
523	COPY_BYTE(1)
524	COPY_BYTE(2)
525	COPY_BYTE(3)
526	COPY_BYTE(4)
527	COPY_BYTE(5)
528	COPY_BYTE(6)
529	COPY_BYTE(7)
530	ADD	src, src, 8
531	b	1b
532	 ADD	dst, dst, 8
533#endif /* CONFIG_CPU_MIPSR6 */
534	.if __memcpy == 1
535	END(memcpy)
536	.set __memcpy, 0
537	.hidden __memcpy
538	.endif
539
540.Ll_exc_copy\@:
541	/*
542	 * Copy bytes from src until faulting load address (or until a
543	 * lb faults)
544	 *
545	 * When reached by a faulting LDFIRST/LDREST, THREAD_BUADDR($28)
546	 * may be more than a byte beyond the last address.
547	 * Hence, the lb below may get an exception.
548	 *
549	 * Assumes src < THREAD_BUADDR($28)
550	 */
551	LOADK	t0, TI_TASK($28)
552	 nop
553	LOADK	t0, THREAD_BUADDR(t0)
5541:
555	LOADB(t1, 0(src), .Ll_exc\@)
556	ADD	src, src, 1
557	sb	t1, 0(dst)	# can't fault -- we're copy_from_user
558	.set	reorder				/* DADDI_WAR */
559	ADD	dst, dst, 1
560	bne	src, t0, 1b
561	.set	noreorder
562.Ll_exc\@:
563	LOADK	t0, TI_TASK($28)
564	 nop
565	LOADK	t0, THREAD_BUADDR(t0)	# t0 is just past last good address
566	 nop
567	SUB	len, AT, t0		# len number of uncopied bytes
568	jr	ra
569	 nop
570
571#define SEXC(n)							\
572	.set	reorder;			/* DADDI_WAR */ \
573.Ls_exc_p ## n ## u\@:						\
574	ADD	len, len, n*NBYTES;				\
575	jr	ra;						\
576	.set	noreorder
577
578SEXC(8)
579SEXC(7)
580SEXC(6)
581SEXC(5)
582SEXC(4)
583SEXC(3)
584SEXC(2)
585SEXC(1)
586
587.Ls_exc_p1\@:
588	.set	reorder				/* DADDI_WAR */
589	ADD	len, len, 1
590	jr	ra
591	.set	noreorder
592.Ls_exc\@:
593	jr	ra
594	 nop
595	.endm
596
597	.align	5
598LEAF(memmove)
599EXPORT_SYMBOL(memmove)
600	ADD	t0, a0, a2
601	ADD	t1, a1, a2
602	sltu	t0, a1, t0			# dst + len <= src -> memcpy
603	sltu	t1, a0, t1			# dst >= src + len -> memcpy
604	and	t0, t1
605	beqz	t0, .L__memcpy
606	 move	v0, a0				/* return value */
607	beqz	a2, .Lr_out
608	END(memmove)
609
610	/* fall through to __rmemcpy */
611LEAF(__rmemcpy)					/* a0=dst a1=src a2=len */
612	 sltu	t0, a1, a0
613	beqz	t0, .Lr_end_bytes_up		# src >= dst
614	 nop
615	ADD	a0, a2				# dst = dst + len
616	ADD	a1, a2				# src = src + len
617
618.Lr_end_bytes:
619	R10KCBARRIER(0(ra))
620	lb	t0, -1(a1)
621	SUB	a2, a2, 0x1
622	sb	t0, -1(a0)
623	SUB	a1, a1, 0x1
624	.set	reorder				/* DADDI_WAR */
625	SUB	a0, a0, 0x1
626	bnez	a2, .Lr_end_bytes
627	.set	noreorder
628
629.Lr_out:
630	jr	ra
631	 move	a2, zero
632
633.Lr_end_bytes_up:
634	R10KCBARRIER(0(ra))
635	lb	t0, (a1)
636	SUB	a2, a2, 0x1
637	sb	t0, (a0)
638	ADD	a1, a1, 0x1
639	.set	reorder				/* DADDI_WAR */
640	ADD	a0, a0, 0x1
641	bnez	a2, .Lr_end_bytes_up
642	.set	noreorder
643
644	jr	ra
645	 move	a2, zero
646	END(__rmemcpy)
647
648/*
649 * A combined memcpy/__copy_user
650 * __copy_user sets len to 0 for success; else to an upper bound of
651 * the number of uncopied bytes.
652 * memcpy sets v0 to dst.
653 */
654	.align	5
655LEAF(memcpy)					/* a0=dst a1=src a2=len */
656EXPORT_SYMBOL(memcpy)
657	move	v0, dst				/* return value */
658.L__memcpy:
659FEXPORT(__copy_user)
660EXPORT_SYMBOL(__copy_user)
661	/* Legacy Mode, user <-> user */
662	__BUILD_COPY_USER LEGACY_MODE USEROP USEROP
663
664#ifdef CONFIG_EVA
665
666/*
667 * For EVA we need distinct symbols for reading and writing to user space.
668 * This is because we need to use specific EVA instructions to perform the
669 * virtual <-> physical translation when a virtual address is actually in user
670 * space
671 */
672
673/*
674 * __copy_from_user (EVA)
675 */
676
677LEAF(__copy_from_user_eva)
678EXPORT_SYMBOL(__copy_from_user_eva)
679	__BUILD_COPY_USER EVA_MODE USEROP KERNELOP
680END(__copy_from_user_eva)
681
682
683
684/*
685 * __copy_to_user (EVA)
686 */
687
688LEAF(__copy_to_user_eva)
689EXPORT_SYMBOL(__copy_to_user_eva)
690__BUILD_COPY_USER EVA_MODE KERNELOP USEROP
691END(__copy_to_user_eva)
692
693/*
694 * __copy_in_user (EVA)
695 */
696
697LEAF(__copy_in_user_eva)
698EXPORT_SYMBOL(__copy_in_user_eva)
699__BUILD_COPY_USER EVA_MODE USEROP USEROP
700END(__copy_in_user_eva)
701
702#endif
703