xref: /linux/arch/mips/lib/memcpy.S (revision 1c4b5ecb7ea190fa3e9f9d6891e6c90b60e04f24)
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_WD	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_WD	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#ifdef CONFIG_CPU_HAS_PREFETCH
208# define _PREF(hint, addr, type)					\
209	.if \mode == LEGACY_MODE;					\
210		kernel_pref(hint, addr);				\
211	.else;								\
212		.if ((\from == USEROP) && (type == SRC_PREFETCH)) ||	\
213		    ((\to == USEROP) && (type == DST_PREFETCH));	\
214			/*						\
215			 * PREFE has only 9 bits for the offset		\
216			 * compared to PREF which has 16, so it may	\
217			 * need to use the $at register but this	\
218			 * register should remain intact because it's	\
219			 * used later on. Therefore use $v1.		\
220			 */						\
221			.set at=v1;					\
222			user_pref(hint, addr);				\
223			.set noat;					\
224		.else;							\
225			kernel_pref(hint, addr);			\
226		.endif;							\
227	.endif
228#else
229# define _PREF(hint, addr, type)
230#endif
231
232#define PREFS(hint, addr) _PREF(hint, addr, SRC_PREFETCH)
233#define PREFD(hint, addr) _PREF(hint, addr, DST_PREFETCH)
234
235#ifdef CONFIG_CPU_LITTLE_ENDIAN
236#define LDFIRST LOADR
237#define LDREST	LOADL
238#define STFIRST STORER
239#define STREST	STOREL
240#define SHIFT_DISCARD SLLV
241#else
242#define LDFIRST LOADL
243#define LDREST	LOADR
244#define STFIRST STOREL
245#define STREST	STORER
246#define SHIFT_DISCARD SRLV
247#endif
248
249#define FIRST(unit) ((unit)*NBYTES)
250#define REST(unit)  (FIRST(unit)+NBYTES-1)
251#define UNIT(unit)  FIRST(unit)
252
253#define ADDRMASK (NBYTES-1)
254
255	.text
256	.set	noreorder
257#ifndef CONFIG_CPU_DADDI_WORKAROUNDS
258	.set	noat
259#else
260	.set	at=v1
261#endif
262
263	.align	5
264
265	/*
266	 * Macro to build the __copy_user common code
267	 * Arguments:
268	 * mode : LEGACY_MODE or EVA_MODE
269	 * from : Source operand. USEROP or KERNELOP
270	 * to   : Destination operand. USEROP or KERNELOP
271	 */
272	.macro __BUILD_COPY_USER mode, from, to
273
274	/* initialize __memcpy if this the first time we execute this macro */
275	.ifnotdef __memcpy
276	.set __memcpy, 1
277	.hidden __memcpy /* make sure it does not leak */
278	.endif
279
280	/*
281	 * Note: dst & src may be unaligned, len may be 0
282	 * Temps
283	 */
284#define rem t8
285
286	R10KCBARRIER(0(ra))
287	/*
288	 * The "issue break"s below are very approximate.
289	 * Issue delays for dcache fills will perturb the schedule, as will
290	 * load queue full replay traps, etc.
291	 *
292	 * If len < NBYTES use byte operations.
293	 */
294	PREFS(	0, 0(src) )
295	PREFD(	1, 0(dst) )
296	sltu	t2, len, NBYTES
297	and	t1, dst, ADDRMASK
298	PREFS(	0, 1*32(src) )
299	PREFD(	1, 1*32(dst) )
300	bnez	t2, .Lcopy_bytes_checklen\@
301	 and	t0, src, ADDRMASK
302	PREFS(	0, 2*32(src) )
303	PREFD(	1, 2*32(dst) )
304#ifndef CONFIG_CPU_NO_LOAD_STORE_LR
305	bnez	t1, .Ldst_unaligned\@
306	 nop
307	bnez	t0, .Lsrc_unaligned_dst_aligned\@
308#else /* CONFIG_CPU_NO_LOAD_STORE_LR */
309	or	t0, t0, t1
310	bnez	t0, .Lcopy_unaligned_bytes\@
311#endif /* CONFIG_CPU_NO_LOAD_STORE_LR */
312	/*
313	 * use delay slot for fall-through
314	 * src and dst are aligned; need to compute rem
315	 */
316.Lboth_aligned\@:
317	 SRL	t0, len, LOG_NBYTES+3	 # +3 for 8 units/iter
318	beqz	t0, .Lcleanup_both_aligned\@ # len < 8*NBYTES
319	 and	rem, len, (8*NBYTES-1)	 # rem = len % (8*NBYTES)
320	PREFS(	0, 3*32(src) )
321	PREFD(	1, 3*32(dst) )
322	.align	4
3231:
324	R10KCBARRIER(0(ra))
325	LOAD(t0, UNIT(0)(src), .Ll_exc\@)
326	LOAD(t1, UNIT(1)(src), .Ll_exc_copy\@)
327	LOAD(t2, UNIT(2)(src), .Ll_exc_copy\@)
328	LOAD(t3, UNIT(3)(src), .Ll_exc_copy\@)
329	SUB	len, len, 8*NBYTES
330	LOAD(t4, UNIT(4)(src), .Ll_exc_copy\@)
331	LOAD(t7, UNIT(5)(src), .Ll_exc_copy\@)
332	STORE(t0, UNIT(0)(dst),	.Ls_exc_p8u\@)
333	STORE(t1, UNIT(1)(dst),	.Ls_exc_p7u\@)
334	LOAD(t0, UNIT(6)(src), .Ll_exc_copy\@)
335	LOAD(t1, UNIT(7)(src), .Ll_exc_copy\@)
336	ADD	src, src, 8*NBYTES
337	ADD	dst, dst, 8*NBYTES
338	STORE(t2, UNIT(-6)(dst), .Ls_exc_p6u\@)
339	STORE(t3, UNIT(-5)(dst), .Ls_exc_p5u\@)
340	STORE(t4, UNIT(-4)(dst), .Ls_exc_p4u\@)
341	STORE(t7, UNIT(-3)(dst), .Ls_exc_p3u\@)
342	STORE(t0, UNIT(-2)(dst), .Ls_exc_p2u\@)
343	STORE(t1, UNIT(-1)(dst), .Ls_exc_p1u\@)
344	PREFS(	0, 8*32(src) )
345	PREFD(	1, 8*32(dst) )
346	bne	len, rem, 1b
347	 nop
348
349	/*
350	 * len == rem == the number of bytes left to copy < 8*NBYTES
351	 */
352.Lcleanup_both_aligned\@:
353	beqz	len, .Ldone\@
354	 sltu	t0, len, 4*NBYTES
355	bnez	t0, .Lless_than_4units\@
356	 and	rem, len, (NBYTES-1)	# rem = len % NBYTES
357	/*
358	 * len >= 4*NBYTES
359	 */
360	LOAD( t0, UNIT(0)(src),	.Ll_exc\@)
361	LOAD( t1, UNIT(1)(src),	.Ll_exc_copy\@)
362	LOAD( t2, UNIT(2)(src),	.Ll_exc_copy\@)
363	LOAD( t3, UNIT(3)(src),	.Ll_exc_copy\@)
364	SUB	len, len, 4*NBYTES
365	ADD	src, src, 4*NBYTES
366	R10KCBARRIER(0(ra))
367	STORE(t0, UNIT(0)(dst),	.Ls_exc_p4u\@)
368	STORE(t1, UNIT(1)(dst),	.Ls_exc_p3u\@)
369	STORE(t2, UNIT(2)(dst),	.Ls_exc_p2u\@)
370	STORE(t3, UNIT(3)(dst),	.Ls_exc_p1u\@)
371	.set	reorder				/* DADDI_WAR */
372	ADD	dst, dst, 4*NBYTES
373	beqz	len, .Ldone\@
374	.set	noreorder
375.Lless_than_4units\@:
376	/*
377	 * rem = len % NBYTES
378	 */
379	beq	rem, len, .Lcopy_bytes\@
380	 nop
3811:
382	R10KCBARRIER(0(ra))
383	LOAD(t0, 0(src), .Ll_exc\@)
384	ADD	src, src, NBYTES
385	SUB	len, len, NBYTES
386	STORE(t0, 0(dst), .Ls_exc_p1u\@)
387	.set	reorder				/* DADDI_WAR */
388	ADD	dst, dst, NBYTES
389	bne	rem, len, 1b
390	.set	noreorder
391
392#ifndef CONFIG_CPU_NO_LOAD_STORE_LR
393	/*
394	 * src and dst are aligned, need to copy rem bytes (rem < NBYTES)
395	 * A loop would do only a byte at a time with possible branch
396	 * mispredicts.	 Can't do an explicit LOAD dst,mask,or,STORE
397	 * because can't assume read-access to dst.  Instead, use
398	 * STREST dst, which doesn't require read access to dst.
399	 *
400	 * This code should perform better than a simple loop on modern,
401	 * wide-issue mips processors because the code has fewer branches and
402	 * more instruction-level parallelism.
403	 */
404#define bits t2
405	beqz	len, .Ldone\@
406	 ADD	t1, dst, len	# t1 is just past last byte of dst
407	li	bits, 8*NBYTES
408	SLL	rem, len, 3	# rem = number of bits to keep
409	LOAD(t0, 0(src), .Ll_exc\@)
410	SUB	bits, bits, rem # bits = number of bits to discard
411	SHIFT_DISCARD t0, t0, bits
412	STREST(t0, -1(t1), .Ls_exc\@)
413	jr	ra
414	 move	len, zero
415.Ldst_unaligned\@:
416	/*
417	 * dst is unaligned
418	 * t0 = src & ADDRMASK
419	 * t1 = dst & ADDRMASK; T1 > 0
420	 * len >= NBYTES
421	 *
422	 * Copy enough bytes to align dst
423	 * Set match = (src and dst have same alignment)
424	 */
425#define match rem
426	LDFIRST(t3, FIRST(0)(src), .Ll_exc\@)
427	ADD	t2, zero, NBYTES
428	LDREST(t3, REST(0)(src), .Ll_exc_copy\@)
429	SUB	t2, t2, t1	# t2 = number of bytes copied
430	xor	match, t0, t1
431	R10KCBARRIER(0(ra))
432	STFIRST(t3, FIRST(0)(dst), .Ls_exc\@)
433	beq	len, t2, .Ldone\@
434	 SUB	len, len, t2
435	ADD	dst, dst, t2
436	beqz	match, .Lboth_aligned\@
437	 ADD	src, src, t2
438
439.Lsrc_unaligned_dst_aligned\@:
440	SRL	t0, len, LOG_NBYTES+2	 # +2 for 4 units/iter
441	PREFS(	0, 3*32(src) )
442	beqz	t0, .Lcleanup_src_unaligned\@
443	 and	rem, len, (4*NBYTES-1)	 # rem = len % 4*NBYTES
444	PREFD(	1, 3*32(dst) )
4451:
446/*
447 * Avoid consecutive LD*'s to the same register since some mips
448 * implementations can't issue them in the same cycle.
449 * It's OK to load FIRST(N+1) before REST(N) because the two addresses
450 * are to the same unit (unless src is aligned, but it's not).
451 */
452	R10KCBARRIER(0(ra))
453	LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
454	LDFIRST(t1, FIRST(1)(src), .Ll_exc_copy\@)
455	SUB	len, len, 4*NBYTES
456	LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
457	LDREST(t1, REST(1)(src), .Ll_exc_copy\@)
458	LDFIRST(t2, FIRST(2)(src), .Ll_exc_copy\@)
459	LDFIRST(t3, FIRST(3)(src), .Ll_exc_copy\@)
460	LDREST(t2, REST(2)(src), .Ll_exc_copy\@)
461	LDREST(t3, REST(3)(src), .Ll_exc_copy\@)
462	PREFS(	0, 9*32(src) )		# 0 is PREF_LOAD  (not streamed)
463	ADD	src, src, 4*NBYTES
464#ifdef CONFIG_CPU_SB1
465	nop				# improves slotting
466#endif
467	STORE(t0, UNIT(0)(dst),	.Ls_exc_p4u\@)
468	STORE(t1, UNIT(1)(dst),	.Ls_exc_p3u\@)
469	STORE(t2, UNIT(2)(dst),	.Ls_exc_p2u\@)
470	STORE(t3, UNIT(3)(dst),	.Ls_exc_p1u\@)
471	PREFD(	1, 9*32(dst) )		# 1 is PREF_STORE (not streamed)
472	.set	reorder				/* DADDI_WAR */
473	ADD	dst, dst, 4*NBYTES
474	bne	len, rem, 1b
475	.set	noreorder
476
477.Lcleanup_src_unaligned\@:
478	beqz	len, .Ldone\@
479	 and	rem, len, NBYTES-1  # rem = len % NBYTES
480	beq	rem, len, .Lcopy_bytes\@
481	 nop
4821:
483	R10KCBARRIER(0(ra))
484	LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
485	LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
486	ADD	src, src, NBYTES
487	SUB	len, len, NBYTES
488	STORE(t0, 0(dst), .Ls_exc_p1u\@)
489	.set	reorder				/* DADDI_WAR */
490	ADD	dst, dst, NBYTES
491	bne	len, rem, 1b
492	.set	noreorder
493
494#endif /* !CONFIG_CPU_NO_LOAD_STORE_LR */
495.Lcopy_bytes_checklen\@:
496	beqz	len, .Ldone\@
497	 nop
498.Lcopy_bytes\@:
499	/* 0 < len < NBYTES  */
500	R10KCBARRIER(0(ra))
501#define COPY_BYTE(N)			\
502	LOADB(t0, N(src), .Ll_exc\@);	\
503	SUB	len, len, 1;		\
504	beqz	len, .Ldone\@;		\
505	STOREB(t0, N(dst), .Ls_exc_p1\@)
506
507	COPY_BYTE(0)
508	COPY_BYTE(1)
509#ifdef USE_DOUBLE
510	COPY_BYTE(2)
511	COPY_BYTE(3)
512	COPY_BYTE(4)
513	COPY_BYTE(5)
514#endif
515	LOADB(t0, NBYTES-2(src), .Ll_exc\@)
516	SUB	len, len, 1
517	jr	ra
518	STOREB(t0, NBYTES-2(dst), .Ls_exc_p1\@)
519.Ldone\@:
520	jr	ra
521	 nop
522
523#ifdef CONFIG_CPU_NO_LOAD_STORE_LR
524.Lcopy_unaligned_bytes\@:
5251:
526	COPY_BYTE(0)
527	COPY_BYTE(1)
528	COPY_BYTE(2)
529	COPY_BYTE(3)
530	COPY_BYTE(4)
531	COPY_BYTE(5)
532	COPY_BYTE(6)
533	COPY_BYTE(7)
534	ADD	src, src, 8
535	b	1b
536	 ADD	dst, dst, 8
537#endif /* CONFIG_CPU_NO_LOAD_STORE_LR */
538	.if __memcpy == 1
539	END(memcpy)
540	.set __memcpy, 0
541	.hidden __memcpy
542	.endif
543
544.Ll_exc_copy\@:
545	/*
546	 * Copy bytes from src until faulting load address (or until a
547	 * lb faults)
548	 *
549	 * When reached by a faulting LDFIRST/LDREST, THREAD_BUADDR($28)
550	 * may be more than a byte beyond the last address.
551	 * Hence, the lb below may get an exception.
552	 *
553	 * Assumes src < THREAD_BUADDR($28)
554	 */
555	LOADK	t0, TI_TASK($28)
556	 nop
557	LOADK	t0, THREAD_BUADDR(t0)
5581:
559	LOADB(t1, 0(src), .Ll_exc\@)
560	ADD	src, src, 1
561	sb	t1, 0(dst)	# can't fault -- we're copy_from_user
562	.set	reorder				/* DADDI_WAR */
563	ADD	dst, dst, 1
564	bne	src, t0, 1b
565	.set	noreorder
566.Ll_exc\@:
567	LOADK	t0, TI_TASK($28)
568	 nop
569	LOADK	t0, THREAD_BUADDR(t0)	# t0 is just past last good address
570	 nop
571	SUB	len, AT, t0		# len number of uncopied bytes
572	jr	ra
573	 nop
574
575#define SEXC(n)							\
576	.set	reorder;			/* DADDI_WAR */ \
577.Ls_exc_p ## n ## u\@:						\
578	ADD	len, len, n*NBYTES;				\
579	jr	ra;						\
580	.set	noreorder
581
582SEXC(8)
583SEXC(7)
584SEXC(6)
585SEXC(5)
586SEXC(4)
587SEXC(3)
588SEXC(2)
589SEXC(1)
590
591.Ls_exc_p1\@:
592	.set	reorder				/* DADDI_WAR */
593	ADD	len, len, 1
594	jr	ra
595	.set	noreorder
596.Ls_exc\@:
597	jr	ra
598	 nop
599	.endm
600
601#ifndef CONFIG_HAVE_PLAT_MEMCPY
602	.align	5
603LEAF(memmove)
604EXPORT_SYMBOL(memmove)
605	ADD	t0, a0, a2
606	ADD	t1, a1, a2
607	sltu	t0, a1, t0			# dst + len <= src -> memcpy
608	sltu	t1, a0, t1			# dst >= src + len -> memcpy
609	and	t0, t1
610	beqz	t0, .L__memcpy
611	 move	v0, a0				/* return value */
612	beqz	a2, .Lr_out
613	END(memmove)
614
615	/* fall through to __rmemcpy */
616LEAF(__rmemcpy)					/* a0=dst a1=src a2=len */
617	 sltu	t0, a1, a0
618	beqz	t0, .Lr_end_bytes_up		# src >= dst
619	 nop
620	ADD	a0, a2				# dst = dst + len
621	ADD	a1, a2				# src = src + len
622
623.Lr_end_bytes:
624	R10KCBARRIER(0(ra))
625	lb	t0, -1(a1)
626	SUB	a2, a2, 0x1
627	sb	t0, -1(a0)
628	SUB	a1, a1, 0x1
629	.set	reorder				/* DADDI_WAR */
630	SUB	a0, a0, 0x1
631	bnez	a2, .Lr_end_bytes
632	.set	noreorder
633
634.Lr_out:
635	jr	ra
636	 move	a2, zero
637
638.Lr_end_bytes_up:
639	R10KCBARRIER(0(ra))
640	lb	t0, (a1)
641	SUB	a2, a2, 0x1
642	sb	t0, (a0)
643	ADD	a1, a1, 0x1
644	.set	reorder				/* DADDI_WAR */
645	ADD	a0, a0, 0x1
646	bnez	a2, .Lr_end_bytes_up
647	.set	noreorder
648
649	jr	ra
650	 move	a2, zero
651	END(__rmemcpy)
652
653/*
654 * A combined memcpy/__copy_user
655 * __copy_user sets len to 0 for success; else to an upper bound of
656 * the number of uncopied bytes.
657 * memcpy sets v0 to dst.
658 */
659	.align	5
660LEAF(memcpy)					/* a0=dst a1=src a2=len */
661EXPORT_SYMBOL(memcpy)
662	move	v0, dst				/* return value */
663.L__memcpy:
664#ifndef CONFIG_EVA
665FEXPORT(__raw_copy_from_user)
666EXPORT_SYMBOL(__raw_copy_from_user)
667FEXPORT(__raw_copy_to_user)
668EXPORT_SYMBOL(__raw_copy_to_user)
669#endif
670	/* Legacy Mode, user <-> user */
671	__BUILD_COPY_USER LEGACY_MODE USEROP USEROP
672
673#endif
674
675#ifdef CONFIG_EVA
676
677/*
678 * For EVA we need distinct symbols for reading and writing to user space.
679 * This is because we need to use specific EVA instructions to perform the
680 * virtual <-> physical translation when a virtual address is actually in user
681 * space
682 */
683
684/*
685 * __copy_from_user (EVA)
686 */
687
688LEAF(__raw_copy_from_user)
689EXPORT_SYMBOL(__raw_copy_from_user)
690	__BUILD_COPY_USER EVA_MODE USEROP KERNELOP
691END(__raw_copy_from_user)
692
693
694
695/*
696 * __copy_to_user (EVA)
697 */
698
699LEAF(__raw_copy_to_user)
700EXPORT_SYMBOL(__raw_copy_to_user)
701__BUILD_COPY_USER EVA_MODE KERNELOP USEROP
702END(__raw_copy_to_user)
703
704#endif
705