xref: /linux/arch/mips/lib/memcpy.S (revision 93d90ad708b8da6efc0e487b66111aa9db7f70c7)
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/regdef.h>
35
36#define dst a0
37#define src a1
38#define len a2
39
40/*
41 * Spec
42 *
43 * memcpy copies len bytes from src to dst and sets v0 to dst.
44 * It assumes that
45 *   - src and dst don't overlap
46 *   - src is readable
47 *   - dst is writable
48 * memcpy uses the standard calling convention
49 *
50 * __copy_user copies up to len bytes from src to dst and sets a2 (len) to
51 * the number of uncopied bytes due to an exception caused by a read or write.
52 * __copy_user assumes that src and dst don't overlap, and that the call is
53 * implementing one of the following:
54 *   copy_to_user
55 *     - src is readable  (no exceptions when reading src)
56 *   copy_from_user
57 *     - dst is writable  (no exceptions when writing dst)
58 * __copy_user uses a non-standard calling convention; see
59 * include/asm-mips/uaccess.h
60 *
61 * When an exception happens on a load, the handler must
62 # ensure that all of the destination buffer is overwritten to prevent
63 * leaking information to user mode programs.
64 */
65
66/*
67 * Implementation
68 */
69
70/*
71 * The exception handler for loads requires that:
72 *  1- AT contain the address of the byte just past the end of the source
73 *     of the copy,
74 *  2- src_entry <= src < AT, and
75 *  3- (dst - src) == (dst_entry - src_entry),
76 * The _entry suffix denotes values when __copy_user was called.
77 *
78 * (1) is set up up by uaccess.h and maintained by not writing AT in copy_user
79 * (2) is met by incrementing src by the number of bytes copied
80 * (3) is met by not doing loads between a pair of increments of dst and src
81 *
82 * The exception handlers for stores adjust len (if necessary) and return.
83 * These handlers do not need to overwrite any data.
84 *
85 * For __rmemcpy and memmove an exception is always a kernel bug, therefore
86 * they're not protected.
87 */
88
89/* Instruction type */
90#define LD_INSN 1
91#define ST_INSN 2
92/* Pretech type */
93#define SRC_PREFETCH 1
94#define DST_PREFETCH 2
95#define LEGACY_MODE 1
96#define EVA_MODE    2
97#define USEROP   1
98#define KERNELOP 2
99
100/*
101 * Wrapper to add an entry in the exception table
102 * in case the insn causes a memory exception.
103 * Arguments:
104 * insn    : Load/store instruction
105 * type    : Instruction type
106 * reg     : Register
107 * addr    : Address
108 * handler : Exception handler
109 */
110
111#define EXC(insn, type, reg, addr, handler)			\
112	.if \mode == LEGACY_MODE;				\
1139:		insn reg, addr;					\
114		.section __ex_table,"a";			\
115		PTR	9b, handler;				\
116		.previous;					\
117	/* This is assembled in EVA mode */			\
118	.else;							\
119		/* If loading from user or storing to user */	\
120		.if ((\from == USEROP) && (type == LD_INSN)) || \
121		    ((\to == USEROP) && (type == ST_INSN));	\
1229:			__BUILD_EVA_INSN(insn##e, reg, addr);	\
123			.section __ex_table,"a";		\
124			PTR	9b, handler;			\
125			.previous;				\
126		.else;						\
127			/*					\
128			 *  Still in EVA, but no need for	\
129			 * exception handler or EVA insn	\
130			 */					\
131			insn reg, addr;				\
132		.endif;						\
133	.endif
134
135/*
136 * Only on the 64-bit kernel we can made use of 64-bit registers.
137 */
138#ifdef CONFIG_64BIT
139#define USE_DOUBLE
140#endif
141
142#ifdef USE_DOUBLE
143
144#define LOADK ld /* No exception */
145#define LOAD(reg, addr, handler)	EXC(ld, LD_INSN, reg, addr, handler)
146#define LOADL(reg, addr, handler)	EXC(ldl, LD_INSN, reg, addr, handler)
147#define LOADR(reg, addr, handler)	EXC(ldr, LD_INSN, reg, addr, handler)
148#define STOREL(reg, addr, handler)	EXC(sdl, ST_INSN, reg, addr, handler)
149#define STORER(reg, addr, handler)	EXC(sdr, ST_INSN, reg, addr, handler)
150#define STORE(reg, addr, handler)	EXC(sd, ST_INSN, reg, addr, handler)
151#define ADD    daddu
152#define SUB    dsubu
153#define SRL    dsrl
154#define SRA    dsra
155#define SLL    dsll
156#define SLLV   dsllv
157#define SRLV   dsrlv
158#define NBYTES 8
159#define LOG_NBYTES 3
160
161/*
162 * As we are sharing code base with the mips32 tree (which use the o32 ABI
163 * register definitions). We need to redefine the register definitions from
164 * the n64 ABI register naming to the o32 ABI register naming.
165 */
166#undef t0
167#undef t1
168#undef t2
169#undef t3
170#define t0	$8
171#define t1	$9
172#define t2	$10
173#define t3	$11
174#define t4	$12
175#define t5	$13
176#define t6	$14
177#define t7	$15
178
179#else
180
181#define LOADK lw /* No exception */
182#define LOAD(reg, addr, handler)	EXC(lw, LD_INSN, reg, addr, handler)
183#define LOADL(reg, addr, handler)	EXC(lwl, LD_INSN, reg, addr, handler)
184#define LOADR(reg, addr, handler)	EXC(lwr, LD_INSN, reg, addr, handler)
185#define STOREL(reg, addr, handler)	EXC(swl, ST_INSN, reg, addr, handler)
186#define STORER(reg, addr, handler)	EXC(swr, ST_INSN, reg, addr, handler)
187#define STORE(reg, addr, handler)	EXC(sw, ST_INSN, reg, addr, handler)
188#define ADD    addu
189#define SUB    subu
190#define SRL    srl
191#define SLL    sll
192#define SRA    sra
193#define SLLV   sllv
194#define SRLV   srlv
195#define NBYTES 4
196#define LOG_NBYTES 2
197
198#endif /* USE_DOUBLE */
199
200#define LOADB(reg, addr, handler)	EXC(lb, LD_INSN, reg, addr, handler)
201#define STOREB(reg, addr, handler)	EXC(sb, ST_INSN, reg, addr, handler)
202
203#define _PREF(hint, addr, type)						\
204	.if \mode == LEGACY_MODE;					\
205		PREF(hint, addr);					\
206	.else;								\
207		.if ((\from == USEROP) && (type == SRC_PREFETCH)) ||	\
208		    ((\to == USEROP) && (type == DST_PREFETCH));	\
209			/*						\
210			 * PREFE has only 9 bits for the offset		\
211			 * compared to PREF which has 16, so it may	\
212			 * need to use the $at register but this	\
213			 * register should remain intact because it's	\
214			 * used later on. Therefore use $v1.		\
215			 */						\
216			.set at=v1;					\
217			PREFE(hint, addr);				\
218			.set noat;					\
219		.else;							\
220			PREF(hint, addr);				\
221		.endif;							\
222	.endif
223
224#define PREFS(hint, addr) _PREF(hint, addr, SRC_PREFETCH)
225#define PREFD(hint, addr) _PREF(hint, addr, DST_PREFETCH)
226
227#ifdef CONFIG_CPU_LITTLE_ENDIAN
228#define LDFIRST LOADR
229#define LDREST	LOADL
230#define STFIRST STORER
231#define STREST	STOREL
232#define SHIFT_DISCARD SLLV
233#else
234#define LDFIRST LOADL
235#define LDREST	LOADR
236#define STFIRST STOREL
237#define STREST	STORER
238#define SHIFT_DISCARD SRLV
239#endif
240
241#define FIRST(unit) ((unit)*NBYTES)
242#define REST(unit)  (FIRST(unit)+NBYTES-1)
243#define UNIT(unit)  FIRST(unit)
244
245#define ADDRMASK (NBYTES-1)
246
247	.text
248	.set	noreorder
249#ifndef CONFIG_CPU_DADDI_WORKAROUNDS
250	.set	noat
251#else
252	.set	at=v1
253#endif
254
255	.align	5
256
257	/*
258	 * Macro to build the __copy_user common code
259	 * Arguements:
260	 * mode : LEGACY_MODE or EVA_MODE
261	 * from : Source operand. USEROP or KERNELOP
262	 * to   : Destination operand. USEROP or KERNELOP
263	 */
264	.macro __BUILD_COPY_USER mode, from, to
265
266	/* initialize __memcpy if this the first time we execute this macro */
267	.ifnotdef __memcpy
268	.set __memcpy, 1
269	.hidden __memcpy /* make sure it does not leak */
270	.endif
271
272	/*
273	 * Note: dst & src may be unaligned, len may be 0
274	 * Temps
275	 */
276#define rem t8
277
278	R10KCBARRIER(0(ra))
279	/*
280	 * The "issue break"s below are very approximate.
281	 * Issue delays for dcache fills will perturb the schedule, as will
282	 * load queue full replay traps, etc.
283	 *
284	 * If len < NBYTES use byte operations.
285	 */
286	PREFS(	0, 0(src) )
287	PREFD(	1, 0(dst) )
288	sltu	t2, len, NBYTES
289	and	t1, dst, ADDRMASK
290	PREFS(	0, 1*32(src) )
291	PREFD(	1, 1*32(dst) )
292	bnez	t2, .Lcopy_bytes_checklen\@
293	 and	t0, src, ADDRMASK
294	PREFS(	0, 2*32(src) )
295	PREFD(	1, 2*32(dst) )
296	bnez	t1, .Ldst_unaligned\@
297	 nop
298	bnez	t0, .Lsrc_unaligned_dst_aligned\@
299	/*
300	 * use delay slot for fall-through
301	 * src and dst are aligned; need to compute rem
302	 */
303.Lboth_aligned\@:
304	 SRL	t0, len, LOG_NBYTES+3	 # +3 for 8 units/iter
305	beqz	t0, .Lcleanup_both_aligned\@ # len < 8*NBYTES
306	 and	rem, len, (8*NBYTES-1)	 # rem = len % (8*NBYTES)
307	PREFS(	0, 3*32(src) )
308	PREFD(	1, 3*32(dst) )
309	.align	4
3101:
311	R10KCBARRIER(0(ra))
312	LOAD(t0, UNIT(0)(src), .Ll_exc\@)
313	LOAD(t1, UNIT(1)(src), .Ll_exc_copy\@)
314	LOAD(t2, UNIT(2)(src), .Ll_exc_copy\@)
315	LOAD(t3, UNIT(3)(src), .Ll_exc_copy\@)
316	SUB	len, len, 8*NBYTES
317	LOAD(t4, UNIT(4)(src), .Ll_exc_copy\@)
318	LOAD(t7, UNIT(5)(src), .Ll_exc_copy\@)
319	STORE(t0, UNIT(0)(dst),	.Ls_exc_p8u\@)
320	STORE(t1, UNIT(1)(dst),	.Ls_exc_p7u\@)
321	LOAD(t0, UNIT(6)(src), .Ll_exc_copy\@)
322	LOAD(t1, UNIT(7)(src), .Ll_exc_copy\@)
323	ADD	src, src, 8*NBYTES
324	ADD	dst, dst, 8*NBYTES
325	STORE(t2, UNIT(-6)(dst), .Ls_exc_p6u\@)
326	STORE(t3, UNIT(-5)(dst), .Ls_exc_p5u\@)
327	STORE(t4, UNIT(-4)(dst), .Ls_exc_p4u\@)
328	STORE(t7, UNIT(-3)(dst), .Ls_exc_p3u\@)
329	STORE(t0, UNIT(-2)(dst), .Ls_exc_p2u\@)
330	STORE(t1, UNIT(-1)(dst), .Ls_exc_p1u\@)
331	PREFS(	0, 8*32(src) )
332	PREFD(	1, 8*32(dst) )
333	bne	len, rem, 1b
334	 nop
335
336	/*
337	 * len == rem == the number of bytes left to copy < 8*NBYTES
338	 */
339.Lcleanup_both_aligned\@:
340	beqz	len, .Ldone\@
341	 sltu	t0, len, 4*NBYTES
342	bnez	t0, .Lless_than_4units\@
343	 and	rem, len, (NBYTES-1)	# rem = len % NBYTES
344	/*
345	 * len >= 4*NBYTES
346	 */
347	LOAD( t0, UNIT(0)(src),	.Ll_exc\@)
348	LOAD( t1, UNIT(1)(src),	.Ll_exc_copy\@)
349	LOAD( t2, UNIT(2)(src),	.Ll_exc_copy\@)
350	LOAD( t3, UNIT(3)(src),	.Ll_exc_copy\@)
351	SUB	len, len, 4*NBYTES
352	ADD	src, src, 4*NBYTES
353	R10KCBARRIER(0(ra))
354	STORE(t0, UNIT(0)(dst),	.Ls_exc_p4u\@)
355	STORE(t1, UNIT(1)(dst),	.Ls_exc_p3u\@)
356	STORE(t2, UNIT(2)(dst),	.Ls_exc_p2u\@)
357	STORE(t3, UNIT(3)(dst),	.Ls_exc_p1u\@)
358	.set	reorder				/* DADDI_WAR */
359	ADD	dst, dst, 4*NBYTES
360	beqz	len, .Ldone\@
361	.set	noreorder
362.Lless_than_4units\@:
363	/*
364	 * rem = len % NBYTES
365	 */
366	beq	rem, len, .Lcopy_bytes\@
367	 nop
3681:
369	R10KCBARRIER(0(ra))
370	LOAD(t0, 0(src), .Ll_exc\@)
371	ADD	src, src, NBYTES
372	SUB	len, len, NBYTES
373	STORE(t0, 0(dst), .Ls_exc_p1u\@)
374	.set	reorder				/* DADDI_WAR */
375	ADD	dst, dst, NBYTES
376	bne	rem, len, 1b
377	.set	noreorder
378
379	/*
380	 * src and dst are aligned, need to copy rem bytes (rem < NBYTES)
381	 * A loop would do only a byte at a time with possible branch
382	 * mispredicts.	 Can't do an explicit LOAD dst,mask,or,STORE
383	 * because can't assume read-access to dst.  Instead, use
384	 * STREST dst, which doesn't require read access to dst.
385	 *
386	 * This code should perform better than a simple loop on modern,
387	 * wide-issue mips processors because the code has fewer branches and
388	 * more instruction-level parallelism.
389	 */
390#define bits t2
391	beqz	len, .Ldone\@
392	 ADD	t1, dst, len	# t1 is just past last byte of dst
393	li	bits, 8*NBYTES
394	SLL	rem, len, 3	# rem = number of bits to keep
395	LOAD(t0, 0(src), .Ll_exc\@)
396	SUB	bits, bits, rem # bits = number of bits to discard
397	SHIFT_DISCARD t0, t0, bits
398	STREST(t0, -1(t1), .Ls_exc\@)
399	jr	ra
400	 move	len, zero
401.Ldst_unaligned\@:
402	/*
403	 * dst is unaligned
404	 * t0 = src & ADDRMASK
405	 * t1 = dst & ADDRMASK; T1 > 0
406	 * len >= NBYTES
407	 *
408	 * Copy enough bytes to align dst
409	 * Set match = (src and dst have same alignment)
410	 */
411#define match rem
412	LDFIRST(t3, FIRST(0)(src), .Ll_exc\@)
413	ADD	t2, zero, NBYTES
414	LDREST(t3, REST(0)(src), .Ll_exc_copy\@)
415	SUB	t2, t2, t1	# t2 = number of bytes copied
416	xor	match, t0, t1
417	R10KCBARRIER(0(ra))
418	STFIRST(t3, FIRST(0)(dst), .Ls_exc\@)
419	beq	len, t2, .Ldone\@
420	 SUB	len, len, t2
421	ADD	dst, dst, t2
422	beqz	match, .Lboth_aligned\@
423	 ADD	src, src, t2
424
425.Lsrc_unaligned_dst_aligned\@:
426	SRL	t0, len, LOG_NBYTES+2	 # +2 for 4 units/iter
427	PREFS(	0, 3*32(src) )
428	beqz	t0, .Lcleanup_src_unaligned\@
429	 and	rem, len, (4*NBYTES-1)	 # rem = len % 4*NBYTES
430	PREFD(	1, 3*32(dst) )
4311:
432/*
433 * Avoid consecutive LD*'s to the same register since some mips
434 * implementations can't issue them in the same cycle.
435 * It's OK to load FIRST(N+1) before REST(N) because the two addresses
436 * are to the same unit (unless src is aligned, but it's not).
437 */
438	R10KCBARRIER(0(ra))
439	LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
440	LDFIRST(t1, FIRST(1)(src), .Ll_exc_copy\@)
441	SUB	len, len, 4*NBYTES
442	LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
443	LDREST(t1, REST(1)(src), .Ll_exc_copy\@)
444	LDFIRST(t2, FIRST(2)(src), .Ll_exc_copy\@)
445	LDFIRST(t3, FIRST(3)(src), .Ll_exc_copy\@)
446	LDREST(t2, REST(2)(src), .Ll_exc_copy\@)
447	LDREST(t3, REST(3)(src), .Ll_exc_copy\@)
448	PREFS(	0, 9*32(src) )		# 0 is PREF_LOAD  (not streamed)
449	ADD	src, src, 4*NBYTES
450#ifdef CONFIG_CPU_SB1
451	nop				# improves slotting
452#endif
453	STORE(t0, UNIT(0)(dst),	.Ls_exc_p4u\@)
454	STORE(t1, UNIT(1)(dst),	.Ls_exc_p3u\@)
455	STORE(t2, UNIT(2)(dst),	.Ls_exc_p2u\@)
456	STORE(t3, UNIT(3)(dst),	.Ls_exc_p1u\@)
457	PREFD(	1, 9*32(dst) )		# 1 is PREF_STORE (not streamed)
458	.set	reorder				/* DADDI_WAR */
459	ADD	dst, dst, 4*NBYTES
460	bne	len, rem, 1b
461	.set	noreorder
462
463.Lcleanup_src_unaligned\@:
464	beqz	len, .Ldone\@
465	 and	rem, len, NBYTES-1  # rem = len % NBYTES
466	beq	rem, len, .Lcopy_bytes\@
467	 nop
4681:
469	R10KCBARRIER(0(ra))
470	LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
471	LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
472	ADD	src, src, NBYTES
473	SUB	len, len, NBYTES
474	STORE(t0, 0(dst), .Ls_exc_p1u\@)
475	.set	reorder				/* DADDI_WAR */
476	ADD	dst, dst, NBYTES
477	bne	len, rem, 1b
478	.set	noreorder
479
480.Lcopy_bytes_checklen\@:
481	beqz	len, .Ldone\@
482	 nop
483.Lcopy_bytes\@:
484	/* 0 < len < NBYTES  */
485	R10KCBARRIER(0(ra))
486#define COPY_BYTE(N)			\
487	LOADB(t0, N(src), .Ll_exc\@);	\
488	SUB	len, len, 1;		\
489	beqz	len, .Ldone\@;		\
490	STOREB(t0, N(dst), .Ls_exc_p1\@)
491
492	COPY_BYTE(0)
493	COPY_BYTE(1)
494#ifdef USE_DOUBLE
495	COPY_BYTE(2)
496	COPY_BYTE(3)
497	COPY_BYTE(4)
498	COPY_BYTE(5)
499#endif
500	LOADB(t0, NBYTES-2(src), .Ll_exc\@)
501	SUB	len, len, 1
502	jr	ra
503	STOREB(t0, NBYTES-2(dst), .Ls_exc_p1\@)
504.Ldone\@:
505	jr	ra
506	 nop
507	.if __memcpy == 1
508	END(memcpy)
509	.set __memcpy, 0
510	.hidden __memcpy
511	.endif
512
513.Ll_exc_copy\@:
514	/*
515	 * Copy bytes from src until faulting load address (or until a
516	 * lb faults)
517	 *
518	 * When reached by a faulting LDFIRST/LDREST, THREAD_BUADDR($28)
519	 * may be more than a byte beyond the last address.
520	 * Hence, the lb below may get an exception.
521	 *
522	 * Assumes src < THREAD_BUADDR($28)
523	 */
524	LOADK	t0, TI_TASK($28)
525	 nop
526	LOADK	t0, THREAD_BUADDR(t0)
5271:
528	LOADB(t1, 0(src), .Ll_exc\@)
529	ADD	src, src, 1
530	sb	t1, 0(dst)	# can't fault -- we're copy_from_user
531	.set	reorder				/* DADDI_WAR */
532	ADD	dst, dst, 1
533	bne	src, t0, 1b
534	.set	noreorder
535.Ll_exc\@:
536	LOADK	t0, TI_TASK($28)
537	 nop
538	LOADK	t0, THREAD_BUADDR(t0)	# t0 is just past last good address
539	 nop
540	SUB	len, AT, t0		# len number of uncopied bytes
541	bnez	t6, .Ldone\@	/* Skip the zeroing part if inatomic */
542	/*
543	 * Here's where we rely on src and dst being incremented in tandem,
544	 *   See (3) above.
545	 * dst += (fault addr - src) to put dst at first byte to clear
546	 */
547	ADD	dst, t0			# compute start address in a1
548	SUB	dst, src
549	/*
550	 * Clear len bytes starting at dst.  Can't call __bzero because it
551	 * might modify len.  An inefficient loop for these rare times...
552	 */
553	.set	reorder				/* DADDI_WAR */
554	SUB	src, len, 1
555	beqz	len, .Ldone\@
556	.set	noreorder
5571:	sb	zero, 0(dst)
558	ADD	dst, dst, 1
559#ifndef CONFIG_CPU_DADDI_WORKAROUNDS
560	bnez	src, 1b
561	 SUB	src, src, 1
562#else
563	.set	push
564	.set	noat
565	li	v1, 1
566	bnez	src, 1b
567	 SUB	src, src, v1
568	.set	pop
569#endif
570	jr	ra
571	 nop
572
573
574#define SEXC(n)							\
575	.set	reorder;			/* DADDI_WAR */ \
576.Ls_exc_p ## n ## u\@:						\
577	ADD	len, len, n*NBYTES;				\
578	jr	ra;						\
579	.set	noreorder
580
581SEXC(8)
582SEXC(7)
583SEXC(6)
584SEXC(5)
585SEXC(4)
586SEXC(3)
587SEXC(2)
588SEXC(1)
589
590.Ls_exc_p1\@:
591	.set	reorder				/* DADDI_WAR */
592	ADD	len, len, 1
593	jr	ra
594	.set	noreorder
595.Ls_exc\@:
596	jr	ra
597	 nop
598	.endm
599
600	.align	5
601LEAF(memmove)
602	ADD	t0, a0, a2
603	ADD	t1, a1, a2
604	sltu	t0, a1, t0			# dst + len <= src -> memcpy
605	sltu	t1, a0, t1			# dst >= src + len -> memcpy
606	and	t0, t1
607	beqz	t0, .L__memcpy
608	 move	v0, a0				/* return value */
609	beqz	a2, .Lr_out
610	END(memmove)
611
612	/* fall through to __rmemcpy */
613LEAF(__rmemcpy)					/* a0=dst a1=src a2=len */
614	 sltu	t0, a1, a0
615	beqz	t0, .Lr_end_bytes_up		# src >= dst
616	 nop
617	ADD	a0, a2				# dst = dst + len
618	ADD	a1, a2				# src = src + len
619
620.Lr_end_bytes:
621	R10KCBARRIER(0(ra))
622	lb	t0, -1(a1)
623	SUB	a2, a2, 0x1
624	sb	t0, -1(a0)
625	SUB	a1, a1, 0x1
626	.set	reorder				/* DADDI_WAR */
627	SUB	a0, a0, 0x1
628	bnez	a2, .Lr_end_bytes
629	.set	noreorder
630
631.Lr_out:
632	jr	ra
633	 move	a2, zero
634
635.Lr_end_bytes_up:
636	R10KCBARRIER(0(ra))
637	lb	t0, (a1)
638	SUB	a2, a2, 0x1
639	sb	t0, (a0)
640	ADD	a1, a1, 0x1
641	.set	reorder				/* DADDI_WAR */
642	ADD	a0, a0, 0x1
643	bnez	a2, .Lr_end_bytes_up
644	.set	noreorder
645
646	jr	ra
647	 move	a2, zero
648	END(__rmemcpy)
649
650/*
651 * t6 is used as a flag to note inatomic mode.
652 */
653LEAF(__copy_user_inatomic)
654	b	__copy_user_common
655	li	t6, 1
656	END(__copy_user_inatomic)
657
658/*
659 * A combined memcpy/__copy_user
660 * __copy_user sets len to 0 for success; else to an upper bound of
661 * the number of uncopied bytes.
662 * memcpy sets v0 to dst.
663 */
664	.align	5
665LEAF(memcpy)					/* a0=dst a1=src a2=len */
666	move	v0, dst				/* return value */
667.L__memcpy:
668FEXPORT(__copy_user)
669	li	t6, 0	/* not inatomic */
670__copy_user_common:
671	/* Legacy Mode, user <-> user */
672	__BUILD_COPY_USER LEGACY_MODE USEROP USEROP
673
674#ifdef CONFIG_EVA
675
676/*
677 * For EVA we need distinct symbols for reading and writing to user space.
678 * This is because we need to use specific EVA instructions to perform the
679 * virtual <-> physical translation when a virtual address is actually in user
680 * space
681 */
682
683LEAF(__copy_user_inatomic_eva)
684	b       __copy_from_user_common
685	li	t6, 1
686	END(__copy_user_inatomic_eva)
687
688/*
689 * __copy_from_user (EVA)
690 */
691
692LEAF(__copy_from_user_eva)
693	li	t6, 0	/* not inatomic */
694__copy_from_user_common:
695	__BUILD_COPY_USER EVA_MODE USEROP KERNELOP
696END(__copy_from_user_eva)
697
698
699
700/*
701 * __copy_to_user (EVA)
702 */
703
704LEAF(__copy_to_user_eva)
705__BUILD_COPY_USER EVA_MODE KERNELOP USEROP
706END(__copy_to_user_eva)
707
708/*
709 * __copy_in_user (EVA)
710 */
711
712LEAF(__copy_in_user_eva)
713__BUILD_COPY_USER EVA_MODE USEROP USEROP
714END(__copy_in_user_eva)
715
716#endif
717