xref: /openbmc/linux/arch/mips/lib/memcpy.S (revision 1c2dd16a)
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
32#include <asm/asm.h>
33#include <asm/asm-offsets.h>
34#include <asm/export.h>
35#include <asm/regdef.h>
36
37#define dst a0
38#define src a1
39#define len a2
40
41/*
42 * Spec
43 *
44 * memcpy copies len bytes from src to dst and sets v0 to dst.
45 * It assumes that
46 *   - src and dst don't overlap
47 *   - src is readable
48 *   - dst is writable
49 * memcpy uses the standard calling convention
50 *
51 * __copy_user copies up to len bytes from src to dst and sets a2 (len) to
52 * the number of uncopied bytes due to an exception caused by a read or write.
53 * __copy_user assumes that src and dst don't overlap, and that the call is
54 * implementing one of the following:
55 *   copy_to_user
56 *     - src is readable  (no exceptions when reading src)
57 *   copy_from_user
58 *     - dst is writable  (no exceptions when writing dst)
59 * __copy_user uses a non-standard calling convention; see
60 * include/asm-mips/uaccess.h
61 *
62 * When an exception happens on a load, the handler must
63 # ensure that all of the destination buffer is overwritten to prevent
64 * leaking information to user mode programs.
65 */
66
67/*
68 * Implementation
69 */
70
71/*
72 * The exception handler for loads requires that:
73 *  1- AT contain the address of the byte just past the end of the source
74 *     of the copy,
75 *  2- src_entry <= src < AT, and
76 *  3- (dst - src) == (dst_entry - src_entry),
77 * The _entry suffix denotes values when __copy_user was called.
78 *
79 * (1) is set up up by uaccess.h and maintained by not writing AT in copy_user
80 * (2) is met by incrementing src by the number of bytes copied
81 * (3) is met by not doing loads between a pair of increments of dst and src
82 *
83 * The exception handlers for stores adjust len (if necessary) and return.
84 * These handlers do not need to overwrite any data.
85 *
86 * For __rmemcpy and memmove an exception is always a kernel bug, therefore
87 * they're not protected.
88 */
89
90/* Instruction type */
91#define LD_INSN 1
92#define ST_INSN 2
93/* Pretech type */
94#define SRC_PREFETCH 1
95#define DST_PREFETCH 2
96#define LEGACY_MODE 1
97#define EVA_MODE    2
98#define USEROP   1
99#define KERNELOP 2
100
101/*
102 * Wrapper to add an entry in the exception table
103 * in case the insn causes a memory exception.
104 * Arguments:
105 * insn    : Load/store instruction
106 * type    : Instruction type
107 * reg     : Register
108 * addr    : Address
109 * handler : Exception handler
110 */
111
112#define EXC(insn, type, reg, addr, handler)			\
113	.if \mode == LEGACY_MODE;				\
1149:		insn reg, addr;					\
115		.section __ex_table,"a";			\
116		PTR	9b, handler;				\
117		.previous;					\
118	/* This is assembled in EVA mode */			\
119	.else;							\
120		/* If loading from user or storing to user */	\
121		.if ((\from == USEROP) && (type == LD_INSN)) || \
122		    ((\to == USEROP) && (type == ST_INSN));	\
1239:			__BUILD_EVA_INSN(insn##e, reg, addr);	\
124			.section __ex_table,"a";		\
125			PTR	9b, handler;			\
126			.previous;				\
127		.else;						\
128			/*					\
129			 *  Still in EVA, but no need for	\
130			 * exception handler or EVA insn	\
131			 */					\
132			insn reg, addr;				\
133		.endif;						\
134	.endif
135
136/*
137 * Only on the 64-bit kernel we can made use of 64-bit registers.
138 */
139#ifdef CONFIG_64BIT
140#define USE_DOUBLE
141#endif
142
143#ifdef USE_DOUBLE
144
145#define LOADK ld /* No exception */
146#define LOAD(reg, addr, handler)	EXC(ld, LD_INSN, reg, addr, handler)
147#define LOADL(reg, addr, handler)	EXC(ldl, LD_INSN, reg, addr, handler)
148#define LOADR(reg, addr, handler)	EXC(ldr, LD_INSN, reg, addr, handler)
149#define STOREL(reg, addr, handler)	EXC(sdl, ST_INSN, reg, addr, handler)
150#define STORER(reg, addr, handler)	EXC(sdr, ST_INSN, reg, addr, handler)
151#define STORE(reg, addr, handler)	EXC(sd, ST_INSN, reg, addr, handler)
152#define ADD    daddu
153#define SUB    dsubu
154#define SRL    dsrl
155#define SRA    dsra
156#define SLL    dsll
157#define SLLV   dsllv
158#define SRLV   dsrlv
159#define NBYTES 8
160#define LOG_NBYTES 3
161
162/*
163 * As we are sharing code base with the mips32 tree (which use the o32 ABI
164 * register definitions). We need to redefine the register definitions from
165 * the n64 ABI register naming to the o32 ABI register naming.
166 */
167#undef t0
168#undef t1
169#undef t2
170#undef t3
171#define t0	$8
172#define t1	$9
173#define t2	$10
174#define t3	$11
175#define t4	$12
176#define t5	$13
177#define t6	$14
178#define t7	$15
179
180#else
181
182#define LOADK lw /* No exception */
183#define LOAD(reg, addr, handler)	EXC(lw, LD_INSN, reg, addr, handler)
184#define LOADL(reg, addr, handler)	EXC(lwl, LD_INSN, reg, addr, handler)
185#define LOADR(reg, addr, handler)	EXC(lwr, LD_INSN, reg, addr, handler)
186#define STOREL(reg, addr, handler)	EXC(swl, ST_INSN, reg, addr, handler)
187#define STORER(reg, addr, handler)	EXC(swr, ST_INSN, reg, addr, handler)
188#define STORE(reg, addr, handler)	EXC(sw, ST_INSN, reg, addr, handler)
189#define ADD    addu
190#define SUB    subu
191#define SRL    srl
192#define SLL    sll
193#define SRA    sra
194#define SLLV   sllv
195#define SRLV   srlv
196#define NBYTES 4
197#define LOG_NBYTES 2
198
199#endif /* USE_DOUBLE */
200
201#define LOADB(reg, addr, handler)	EXC(lb, LD_INSN, reg, addr, handler)
202#define STOREB(reg, addr, handler)	EXC(sb, ST_INSN, reg, addr, handler)
203
204#define _PREF(hint, addr, type)						\
205	.if \mode == LEGACY_MODE;					\
206		PREF(hint, addr);					\
207	.else;								\
208		.if ((\from == USEROP) && (type == SRC_PREFETCH)) ||	\
209		    ((\to == USEROP) && (type == DST_PREFETCH));	\
210			/*						\
211			 * PREFE has only 9 bits for the offset		\
212			 * compared to PREF which has 16, so it may	\
213			 * need to use the $at register but this	\
214			 * register should remain intact because it's	\
215			 * used later on. Therefore use $v1.		\
216			 */						\
217			.set at=v1;					\
218			PREFE(hint, addr);				\
219			.set noat;					\
220		.else;							\
221			PREF(hint, addr);				\
222		.endif;							\
223	.endif
224
225#define PREFS(hint, addr) _PREF(hint, addr, SRC_PREFETCH)
226#define PREFD(hint, addr) _PREF(hint, addr, DST_PREFETCH)
227
228#ifdef CONFIG_CPU_LITTLE_ENDIAN
229#define LDFIRST LOADR
230#define LDREST	LOADL
231#define STFIRST STORER
232#define STREST	STOREL
233#define SHIFT_DISCARD SLLV
234#else
235#define LDFIRST LOADL
236#define LDREST	LOADR
237#define STFIRST STOREL
238#define STREST	STORER
239#define SHIFT_DISCARD SRLV
240#endif
241
242#define FIRST(unit) ((unit)*NBYTES)
243#define REST(unit)  (FIRST(unit)+NBYTES-1)
244#define UNIT(unit)  FIRST(unit)
245
246#define ADDRMASK (NBYTES-1)
247
248	.text
249	.set	noreorder
250#ifndef CONFIG_CPU_DADDI_WORKAROUNDS
251	.set	noat
252#else
253	.set	at=v1
254#endif
255
256	.align	5
257
258	/*
259	 * Macro to build the __copy_user common code
260	 * Arguments:
261	 * mode : LEGACY_MODE or EVA_MODE
262	 * from : Source operand. USEROP or KERNELOP
263	 * to   : Destination operand. USEROP or KERNELOP
264	 */
265	.macro __BUILD_COPY_USER mode, from, to
266
267	/* initialize __memcpy if this the first time we execute this macro */
268	.ifnotdef __memcpy
269	.set __memcpy, 1
270	.hidden __memcpy /* make sure it does not leak */
271	.endif
272
273	/*
274	 * Note: dst & src may be unaligned, len may be 0
275	 * Temps
276	 */
277#define rem t8
278
279	R10KCBARRIER(0(ra))
280	/*
281	 * The "issue break"s below are very approximate.
282	 * Issue delays for dcache fills will perturb the schedule, as will
283	 * load queue full replay traps, etc.
284	 *
285	 * If len < NBYTES use byte operations.
286	 */
287	PREFS(	0, 0(src) )
288	PREFD(	1, 0(dst) )
289	sltu	t2, len, NBYTES
290	and	t1, dst, ADDRMASK
291	PREFS(	0, 1*32(src) )
292	PREFD(	1, 1*32(dst) )
293	bnez	t2, .Lcopy_bytes_checklen\@
294	 and	t0, src, ADDRMASK
295	PREFS(	0, 2*32(src) )
296	PREFD(	1, 2*32(dst) )
297#ifndef CONFIG_CPU_MIPSR6
298	bnez	t1, .Ldst_unaligned\@
299	 nop
300	bnez	t0, .Lsrc_unaligned_dst_aligned\@
301#else
302	or	t0, t0, t1
303	bnez	t0, .Lcopy_unaligned_bytes\@
304#endif
305	/*
306	 * use delay slot for fall-through
307	 * src and dst are aligned; need to compute rem
308	 */
309.Lboth_aligned\@:
310	 SRL	t0, len, LOG_NBYTES+3	 # +3 for 8 units/iter
311	beqz	t0, .Lcleanup_both_aligned\@ # len < 8*NBYTES
312	 and	rem, len, (8*NBYTES-1)	 # rem = len % (8*NBYTES)
313	PREFS(	0, 3*32(src) )
314	PREFD(	1, 3*32(dst) )
315	.align	4
3161:
317	R10KCBARRIER(0(ra))
318	LOAD(t0, UNIT(0)(src), .Ll_exc\@)
319	LOAD(t1, UNIT(1)(src), .Ll_exc_copy\@)
320	LOAD(t2, UNIT(2)(src), .Ll_exc_copy\@)
321	LOAD(t3, UNIT(3)(src), .Ll_exc_copy\@)
322	SUB	len, len, 8*NBYTES
323	LOAD(t4, UNIT(4)(src), .Ll_exc_copy\@)
324	LOAD(t7, UNIT(5)(src), .Ll_exc_copy\@)
325	STORE(t0, UNIT(0)(dst),	.Ls_exc_p8u\@)
326	STORE(t1, UNIT(1)(dst),	.Ls_exc_p7u\@)
327	LOAD(t0, UNIT(6)(src), .Ll_exc_copy\@)
328	LOAD(t1, UNIT(7)(src), .Ll_exc_copy\@)
329	ADD	src, src, 8*NBYTES
330	ADD	dst, dst, 8*NBYTES
331	STORE(t2, UNIT(-6)(dst), .Ls_exc_p6u\@)
332	STORE(t3, UNIT(-5)(dst), .Ls_exc_p5u\@)
333	STORE(t4, UNIT(-4)(dst), .Ls_exc_p4u\@)
334	STORE(t7, UNIT(-3)(dst), .Ls_exc_p3u\@)
335	STORE(t0, UNIT(-2)(dst), .Ls_exc_p2u\@)
336	STORE(t1, UNIT(-1)(dst), .Ls_exc_p1u\@)
337	PREFS(	0, 8*32(src) )
338	PREFD(	1, 8*32(dst) )
339	bne	len, rem, 1b
340	 nop
341
342	/*
343	 * len == rem == the number of bytes left to copy < 8*NBYTES
344	 */
345.Lcleanup_both_aligned\@:
346	beqz	len, .Ldone\@
347	 sltu	t0, len, 4*NBYTES
348	bnez	t0, .Lless_than_4units\@
349	 and	rem, len, (NBYTES-1)	# rem = len % NBYTES
350	/*
351	 * len >= 4*NBYTES
352	 */
353	LOAD( t0, UNIT(0)(src),	.Ll_exc\@)
354	LOAD( t1, UNIT(1)(src),	.Ll_exc_copy\@)
355	LOAD( t2, UNIT(2)(src),	.Ll_exc_copy\@)
356	LOAD( t3, UNIT(3)(src),	.Ll_exc_copy\@)
357	SUB	len, len, 4*NBYTES
358	ADD	src, src, 4*NBYTES
359	R10KCBARRIER(0(ra))
360	STORE(t0, UNIT(0)(dst),	.Ls_exc_p4u\@)
361	STORE(t1, UNIT(1)(dst),	.Ls_exc_p3u\@)
362	STORE(t2, UNIT(2)(dst),	.Ls_exc_p2u\@)
363	STORE(t3, UNIT(3)(dst),	.Ls_exc_p1u\@)
364	.set	reorder				/* DADDI_WAR */
365	ADD	dst, dst, 4*NBYTES
366	beqz	len, .Ldone\@
367	.set	noreorder
368.Lless_than_4units\@:
369	/*
370	 * rem = len % NBYTES
371	 */
372	beq	rem, len, .Lcopy_bytes\@
373	 nop
3741:
375	R10KCBARRIER(0(ra))
376	LOAD(t0, 0(src), .Ll_exc\@)
377	ADD	src, src, NBYTES
378	SUB	len, len, NBYTES
379	STORE(t0, 0(dst), .Ls_exc_p1u\@)
380	.set	reorder				/* DADDI_WAR */
381	ADD	dst, dst, NBYTES
382	bne	rem, len, 1b
383	.set	noreorder
384
385#ifndef CONFIG_CPU_MIPSR6
386	/*
387	 * src and dst are aligned, need to copy rem bytes (rem < NBYTES)
388	 * A loop would do only a byte at a time with possible branch
389	 * mispredicts.	 Can't do an explicit LOAD dst,mask,or,STORE
390	 * because can't assume read-access to dst.  Instead, use
391	 * STREST dst, which doesn't require read access to dst.
392	 *
393	 * This code should perform better than a simple loop on modern,
394	 * wide-issue mips processors because the code has fewer branches and
395	 * more instruction-level parallelism.
396	 */
397#define bits t2
398	beqz	len, .Ldone\@
399	 ADD	t1, dst, len	# t1 is just past last byte of dst
400	li	bits, 8*NBYTES
401	SLL	rem, len, 3	# rem = number of bits to keep
402	LOAD(t0, 0(src), .Ll_exc\@)
403	SUB	bits, bits, rem # bits = number of bits to discard
404	SHIFT_DISCARD t0, t0, bits
405	STREST(t0, -1(t1), .Ls_exc\@)
406	jr	ra
407	 move	len, zero
408.Ldst_unaligned\@:
409	/*
410	 * dst is unaligned
411	 * t0 = src & ADDRMASK
412	 * t1 = dst & ADDRMASK; T1 > 0
413	 * len >= NBYTES
414	 *
415	 * Copy enough bytes to align dst
416	 * Set match = (src and dst have same alignment)
417	 */
418#define match rem
419	LDFIRST(t3, FIRST(0)(src), .Ll_exc\@)
420	ADD	t2, zero, NBYTES
421	LDREST(t3, REST(0)(src), .Ll_exc_copy\@)
422	SUB	t2, t2, t1	# t2 = number of bytes copied
423	xor	match, t0, t1
424	R10KCBARRIER(0(ra))
425	STFIRST(t3, FIRST(0)(dst), .Ls_exc\@)
426	beq	len, t2, .Ldone\@
427	 SUB	len, len, t2
428	ADD	dst, dst, t2
429	beqz	match, .Lboth_aligned\@
430	 ADD	src, src, t2
431
432.Lsrc_unaligned_dst_aligned\@:
433	SRL	t0, len, LOG_NBYTES+2	 # +2 for 4 units/iter
434	PREFS(	0, 3*32(src) )
435	beqz	t0, .Lcleanup_src_unaligned\@
436	 and	rem, len, (4*NBYTES-1)	 # rem = len % 4*NBYTES
437	PREFD(	1, 3*32(dst) )
4381:
439/*
440 * Avoid consecutive LD*'s to the same register since some mips
441 * implementations can't issue them in the same cycle.
442 * It's OK to load FIRST(N+1) before REST(N) because the two addresses
443 * are to the same unit (unless src is aligned, but it's not).
444 */
445	R10KCBARRIER(0(ra))
446	LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
447	LDFIRST(t1, FIRST(1)(src), .Ll_exc_copy\@)
448	SUB	len, len, 4*NBYTES
449	LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
450	LDREST(t1, REST(1)(src), .Ll_exc_copy\@)
451	LDFIRST(t2, FIRST(2)(src), .Ll_exc_copy\@)
452	LDFIRST(t3, FIRST(3)(src), .Ll_exc_copy\@)
453	LDREST(t2, REST(2)(src), .Ll_exc_copy\@)
454	LDREST(t3, REST(3)(src), .Ll_exc_copy\@)
455	PREFS(	0, 9*32(src) )		# 0 is PREF_LOAD  (not streamed)
456	ADD	src, src, 4*NBYTES
457#ifdef CONFIG_CPU_SB1
458	nop				# improves slotting
459#endif
460	STORE(t0, UNIT(0)(dst),	.Ls_exc_p4u\@)
461	STORE(t1, UNIT(1)(dst),	.Ls_exc_p3u\@)
462	STORE(t2, UNIT(2)(dst),	.Ls_exc_p2u\@)
463	STORE(t3, UNIT(3)(dst),	.Ls_exc_p1u\@)
464	PREFD(	1, 9*32(dst) )		# 1 is PREF_STORE (not streamed)
465	.set	reorder				/* DADDI_WAR */
466	ADD	dst, dst, 4*NBYTES
467	bne	len, rem, 1b
468	.set	noreorder
469
470.Lcleanup_src_unaligned\@:
471	beqz	len, .Ldone\@
472	 and	rem, len, NBYTES-1  # rem = len % NBYTES
473	beq	rem, len, .Lcopy_bytes\@
474	 nop
4751:
476	R10KCBARRIER(0(ra))
477	LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
478	LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
479	ADD	src, src, NBYTES
480	SUB	len, len, NBYTES
481	STORE(t0, 0(dst), .Ls_exc_p1u\@)
482	.set	reorder				/* DADDI_WAR */
483	ADD	dst, dst, NBYTES
484	bne	len, rem, 1b
485	.set	noreorder
486
487#endif /* !CONFIG_CPU_MIPSR6 */
488.Lcopy_bytes_checklen\@:
489	beqz	len, .Ldone\@
490	 nop
491.Lcopy_bytes\@:
492	/* 0 < len < NBYTES  */
493	R10KCBARRIER(0(ra))
494#define COPY_BYTE(N)			\
495	LOADB(t0, N(src), .Ll_exc\@);	\
496	SUB	len, len, 1;		\
497	beqz	len, .Ldone\@;		\
498	STOREB(t0, N(dst), .Ls_exc_p1\@)
499
500	COPY_BYTE(0)
501	COPY_BYTE(1)
502#ifdef USE_DOUBLE
503	COPY_BYTE(2)
504	COPY_BYTE(3)
505	COPY_BYTE(4)
506	COPY_BYTE(5)
507#endif
508	LOADB(t0, NBYTES-2(src), .Ll_exc\@)
509	SUB	len, len, 1
510	jr	ra
511	STOREB(t0, NBYTES-2(dst), .Ls_exc_p1\@)
512.Ldone\@:
513	jr	ra
514	 nop
515
516#ifdef CONFIG_CPU_MIPSR6
517.Lcopy_unaligned_bytes\@:
5181:
519	COPY_BYTE(0)
520	COPY_BYTE(1)
521	COPY_BYTE(2)
522	COPY_BYTE(3)
523	COPY_BYTE(4)
524	COPY_BYTE(5)
525	COPY_BYTE(6)
526	COPY_BYTE(7)
527	ADD	src, src, 8
528	b	1b
529	 ADD	dst, dst, 8
530#endif /* CONFIG_CPU_MIPSR6 */
531	.if __memcpy == 1
532	END(memcpy)
533	.set __memcpy, 0
534	.hidden __memcpy
535	.endif
536
537.Ll_exc_copy\@:
538	/*
539	 * Copy bytes from src until faulting load address (or until a
540	 * lb faults)
541	 *
542	 * When reached by a faulting LDFIRST/LDREST, THREAD_BUADDR($28)
543	 * may be more than a byte beyond the last address.
544	 * Hence, the lb below may get an exception.
545	 *
546	 * Assumes src < THREAD_BUADDR($28)
547	 */
548	LOADK	t0, TI_TASK($28)
549	 nop
550	LOADK	t0, THREAD_BUADDR(t0)
5511:
552	LOADB(t1, 0(src), .Ll_exc\@)
553	ADD	src, src, 1
554	sb	t1, 0(dst)	# can't fault -- we're copy_from_user
555	.set	reorder				/* DADDI_WAR */
556	ADD	dst, dst, 1
557	bne	src, t0, 1b
558	.set	noreorder
559.Ll_exc\@:
560	LOADK	t0, TI_TASK($28)
561	 nop
562	LOADK	t0, THREAD_BUADDR(t0)	# t0 is just past last good address
563	 nop
564	SUB	len, AT, t0		# len number of uncopied bytes
565	jr	ra
566	 nop
567
568#define SEXC(n)							\
569	.set	reorder;			/* DADDI_WAR */ \
570.Ls_exc_p ## n ## u\@:						\
571	ADD	len, len, n*NBYTES;				\
572	jr	ra;						\
573	.set	noreorder
574
575SEXC(8)
576SEXC(7)
577SEXC(6)
578SEXC(5)
579SEXC(4)
580SEXC(3)
581SEXC(2)
582SEXC(1)
583
584.Ls_exc_p1\@:
585	.set	reorder				/* DADDI_WAR */
586	ADD	len, len, 1
587	jr	ra
588	.set	noreorder
589.Ls_exc\@:
590	jr	ra
591	 nop
592	.endm
593
594	.align	5
595LEAF(memmove)
596EXPORT_SYMBOL(memmove)
597	ADD	t0, a0, a2
598	ADD	t1, a1, a2
599	sltu	t0, a1, t0			# dst + len <= src -> memcpy
600	sltu	t1, a0, t1			# dst >= src + len -> memcpy
601	and	t0, t1
602	beqz	t0, .L__memcpy
603	 move	v0, a0				/* return value */
604	beqz	a2, .Lr_out
605	END(memmove)
606
607	/* fall through to __rmemcpy */
608LEAF(__rmemcpy)					/* a0=dst a1=src a2=len */
609	 sltu	t0, a1, a0
610	beqz	t0, .Lr_end_bytes_up		# src >= dst
611	 nop
612	ADD	a0, a2				# dst = dst + len
613	ADD	a1, a2				# src = src + len
614
615.Lr_end_bytes:
616	R10KCBARRIER(0(ra))
617	lb	t0, -1(a1)
618	SUB	a2, a2, 0x1
619	sb	t0, -1(a0)
620	SUB	a1, a1, 0x1
621	.set	reorder				/* DADDI_WAR */
622	SUB	a0, a0, 0x1
623	bnez	a2, .Lr_end_bytes
624	.set	noreorder
625
626.Lr_out:
627	jr	ra
628	 move	a2, zero
629
630.Lr_end_bytes_up:
631	R10KCBARRIER(0(ra))
632	lb	t0, (a1)
633	SUB	a2, a2, 0x1
634	sb	t0, (a0)
635	ADD	a1, a1, 0x1
636	.set	reorder				/* DADDI_WAR */
637	ADD	a0, a0, 0x1
638	bnez	a2, .Lr_end_bytes_up
639	.set	noreorder
640
641	jr	ra
642	 move	a2, zero
643	END(__rmemcpy)
644
645/*
646 * A combined memcpy/__copy_user
647 * __copy_user sets len to 0 for success; else to an upper bound of
648 * the number of uncopied bytes.
649 * memcpy sets v0 to dst.
650 */
651	.align	5
652LEAF(memcpy)					/* a0=dst a1=src a2=len */
653EXPORT_SYMBOL(memcpy)
654	move	v0, dst				/* return value */
655.L__memcpy:
656FEXPORT(__copy_user)
657EXPORT_SYMBOL(__copy_user)
658	/* Legacy Mode, user <-> user */
659	__BUILD_COPY_USER LEGACY_MODE USEROP USEROP
660
661#ifdef CONFIG_EVA
662
663/*
664 * For EVA we need distinct symbols for reading and writing to user space.
665 * This is because we need to use specific EVA instructions to perform the
666 * virtual <-> physical translation when a virtual address is actually in user
667 * space
668 */
669
670/*
671 * __copy_from_user (EVA)
672 */
673
674LEAF(__copy_from_user_eva)
675EXPORT_SYMBOL(__copy_from_user_eva)
676	__BUILD_COPY_USER EVA_MODE USEROP KERNELOP
677END(__copy_from_user_eva)
678
679
680
681/*
682 * __copy_to_user (EVA)
683 */
684
685LEAF(__copy_to_user_eva)
686EXPORT_SYMBOL(__copy_to_user_eva)
687__BUILD_COPY_USER EVA_MODE KERNELOP USEROP
688END(__copy_to_user_eva)
689
690/*
691 * __copy_in_user (EVA)
692 */
693
694LEAF(__copy_in_user_eva)
695EXPORT_SYMBOL(__copy_in_user_eva)
696__BUILD_COPY_USER EVA_MODE USEROP USEROP
697END(__copy_in_user_eva)
698
699#endif
700