xref: /openbmc/linux/arch/mips/lib/memcpy.S (revision 3d9b448b)
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#ifndef CONFIG_CPU_MIPSR6
297	bnez	t1, .Ldst_unaligned\@
298	 nop
299	bnez	t0, .Lsrc_unaligned_dst_aligned\@
300#else
301	or	t0, t0, t1
302	bnez	t0, .Lcopy_unaligned_bytes\@
303#endif
304	/*
305	 * use delay slot for fall-through
306	 * src and dst are aligned; need to compute rem
307	 */
308.Lboth_aligned\@:
309	 SRL	t0, len, LOG_NBYTES+3	 # +3 for 8 units/iter
310	beqz	t0, .Lcleanup_both_aligned\@ # len < 8*NBYTES
311	 and	rem, len, (8*NBYTES-1)	 # rem = len % (8*NBYTES)
312	PREFS(	0, 3*32(src) )
313	PREFD(	1, 3*32(dst) )
314	.align	4
3151:
316	R10KCBARRIER(0(ra))
317	LOAD(t0, UNIT(0)(src), .Ll_exc\@)
318	LOAD(t1, UNIT(1)(src), .Ll_exc_copy\@)
319	LOAD(t2, UNIT(2)(src), .Ll_exc_copy\@)
320	LOAD(t3, UNIT(3)(src), .Ll_exc_copy\@)
321	SUB	len, len, 8*NBYTES
322	LOAD(t4, UNIT(4)(src), .Ll_exc_copy\@)
323	LOAD(t7, UNIT(5)(src), .Ll_exc_copy\@)
324	STORE(t0, UNIT(0)(dst),	.Ls_exc_p8u\@)
325	STORE(t1, UNIT(1)(dst),	.Ls_exc_p7u\@)
326	LOAD(t0, UNIT(6)(src), .Ll_exc_copy\@)
327	LOAD(t1, UNIT(7)(src), .Ll_exc_copy\@)
328	ADD	src, src, 8*NBYTES
329	ADD	dst, dst, 8*NBYTES
330	STORE(t2, UNIT(-6)(dst), .Ls_exc_p6u\@)
331	STORE(t3, UNIT(-5)(dst), .Ls_exc_p5u\@)
332	STORE(t4, UNIT(-4)(dst), .Ls_exc_p4u\@)
333	STORE(t7, UNIT(-3)(dst), .Ls_exc_p3u\@)
334	STORE(t0, UNIT(-2)(dst), .Ls_exc_p2u\@)
335	STORE(t1, UNIT(-1)(dst), .Ls_exc_p1u\@)
336	PREFS(	0, 8*32(src) )
337	PREFD(	1, 8*32(dst) )
338	bne	len, rem, 1b
339	 nop
340
341	/*
342	 * len == rem == the number of bytes left to copy < 8*NBYTES
343	 */
344.Lcleanup_both_aligned\@:
345	beqz	len, .Ldone\@
346	 sltu	t0, len, 4*NBYTES
347	bnez	t0, .Lless_than_4units\@
348	 and	rem, len, (NBYTES-1)	# rem = len % NBYTES
349	/*
350	 * len >= 4*NBYTES
351	 */
352	LOAD( t0, UNIT(0)(src),	.Ll_exc\@)
353	LOAD( t1, UNIT(1)(src),	.Ll_exc_copy\@)
354	LOAD( t2, UNIT(2)(src),	.Ll_exc_copy\@)
355	LOAD( t3, UNIT(3)(src),	.Ll_exc_copy\@)
356	SUB	len, len, 4*NBYTES
357	ADD	src, src, 4*NBYTES
358	R10KCBARRIER(0(ra))
359	STORE(t0, UNIT(0)(dst),	.Ls_exc_p4u\@)
360	STORE(t1, UNIT(1)(dst),	.Ls_exc_p3u\@)
361	STORE(t2, UNIT(2)(dst),	.Ls_exc_p2u\@)
362	STORE(t3, UNIT(3)(dst),	.Ls_exc_p1u\@)
363	.set	reorder				/* DADDI_WAR */
364	ADD	dst, dst, 4*NBYTES
365	beqz	len, .Ldone\@
366	.set	noreorder
367.Lless_than_4units\@:
368	/*
369	 * rem = len % NBYTES
370	 */
371	beq	rem, len, .Lcopy_bytes\@
372	 nop
3731:
374	R10KCBARRIER(0(ra))
375	LOAD(t0, 0(src), .Ll_exc\@)
376	ADD	src, src, NBYTES
377	SUB	len, len, NBYTES
378	STORE(t0, 0(dst), .Ls_exc_p1u\@)
379	.set	reorder				/* DADDI_WAR */
380	ADD	dst, dst, NBYTES
381	bne	rem, len, 1b
382	.set	noreorder
383
384#ifndef CONFIG_CPU_MIPSR6
385	/*
386	 * src and dst are aligned, need to copy rem bytes (rem < NBYTES)
387	 * A loop would do only a byte at a time with possible branch
388	 * mispredicts.	 Can't do an explicit LOAD dst,mask,or,STORE
389	 * because can't assume read-access to dst.  Instead, use
390	 * STREST dst, which doesn't require read access to dst.
391	 *
392	 * This code should perform better than a simple loop on modern,
393	 * wide-issue mips processors because the code has fewer branches and
394	 * more instruction-level parallelism.
395	 */
396#define bits t2
397	beqz	len, .Ldone\@
398	 ADD	t1, dst, len	# t1 is just past last byte of dst
399	li	bits, 8*NBYTES
400	SLL	rem, len, 3	# rem = number of bits to keep
401	LOAD(t0, 0(src), .Ll_exc\@)
402	SUB	bits, bits, rem # bits = number of bits to discard
403	SHIFT_DISCARD t0, t0, bits
404	STREST(t0, -1(t1), .Ls_exc\@)
405	jr	ra
406	 move	len, zero
407.Ldst_unaligned\@:
408	/*
409	 * dst is unaligned
410	 * t0 = src & ADDRMASK
411	 * t1 = dst & ADDRMASK; T1 > 0
412	 * len >= NBYTES
413	 *
414	 * Copy enough bytes to align dst
415	 * Set match = (src and dst have same alignment)
416	 */
417#define match rem
418	LDFIRST(t3, FIRST(0)(src), .Ll_exc\@)
419	ADD	t2, zero, NBYTES
420	LDREST(t3, REST(0)(src), .Ll_exc_copy\@)
421	SUB	t2, t2, t1	# t2 = number of bytes copied
422	xor	match, t0, t1
423	R10KCBARRIER(0(ra))
424	STFIRST(t3, FIRST(0)(dst), .Ls_exc\@)
425	beq	len, t2, .Ldone\@
426	 SUB	len, len, t2
427	ADD	dst, dst, t2
428	beqz	match, .Lboth_aligned\@
429	 ADD	src, src, t2
430
431.Lsrc_unaligned_dst_aligned\@:
432	SRL	t0, len, LOG_NBYTES+2	 # +2 for 4 units/iter
433	PREFS(	0, 3*32(src) )
434	beqz	t0, .Lcleanup_src_unaligned\@
435	 and	rem, len, (4*NBYTES-1)	 # rem = len % 4*NBYTES
436	PREFD(	1, 3*32(dst) )
4371:
438/*
439 * Avoid consecutive LD*'s to the same register since some mips
440 * implementations can't issue them in the same cycle.
441 * It's OK to load FIRST(N+1) before REST(N) because the two addresses
442 * are to the same unit (unless src is aligned, but it's not).
443 */
444	R10KCBARRIER(0(ra))
445	LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
446	LDFIRST(t1, FIRST(1)(src), .Ll_exc_copy\@)
447	SUB	len, len, 4*NBYTES
448	LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
449	LDREST(t1, REST(1)(src), .Ll_exc_copy\@)
450	LDFIRST(t2, FIRST(2)(src), .Ll_exc_copy\@)
451	LDFIRST(t3, FIRST(3)(src), .Ll_exc_copy\@)
452	LDREST(t2, REST(2)(src), .Ll_exc_copy\@)
453	LDREST(t3, REST(3)(src), .Ll_exc_copy\@)
454	PREFS(	0, 9*32(src) )		# 0 is PREF_LOAD  (not streamed)
455	ADD	src, src, 4*NBYTES
456#ifdef CONFIG_CPU_SB1
457	nop				# improves slotting
458#endif
459	STORE(t0, UNIT(0)(dst),	.Ls_exc_p4u\@)
460	STORE(t1, UNIT(1)(dst),	.Ls_exc_p3u\@)
461	STORE(t2, UNIT(2)(dst),	.Ls_exc_p2u\@)
462	STORE(t3, UNIT(3)(dst),	.Ls_exc_p1u\@)
463	PREFD(	1, 9*32(dst) )		# 1 is PREF_STORE (not streamed)
464	.set	reorder				/* DADDI_WAR */
465	ADD	dst, dst, 4*NBYTES
466	bne	len, rem, 1b
467	.set	noreorder
468
469.Lcleanup_src_unaligned\@:
470	beqz	len, .Ldone\@
471	 and	rem, len, NBYTES-1  # rem = len % NBYTES
472	beq	rem, len, .Lcopy_bytes\@
473	 nop
4741:
475	R10KCBARRIER(0(ra))
476	LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
477	LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
478	ADD	src, src, NBYTES
479	SUB	len, len, NBYTES
480	STORE(t0, 0(dst), .Ls_exc_p1u\@)
481	.set	reorder				/* DADDI_WAR */
482	ADD	dst, dst, NBYTES
483	bne	len, rem, 1b
484	.set	noreorder
485
486#endif /* !CONFIG_CPU_MIPSR6 */
487.Lcopy_bytes_checklen\@:
488	beqz	len, .Ldone\@
489	 nop
490.Lcopy_bytes\@:
491	/* 0 < len < NBYTES  */
492	R10KCBARRIER(0(ra))
493#define COPY_BYTE(N)			\
494	LOADB(t0, N(src), .Ll_exc\@);	\
495	SUB	len, len, 1;		\
496	beqz	len, .Ldone\@;		\
497	STOREB(t0, N(dst), .Ls_exc_p1\@)
498
499	COPY_BYTE(0)
500	COPY_BYTE(1)
501#ifdef USE_DOUBLE
502	COPY_BYTE(2)
503	COPY_BYTE(3)
504	COPY_BYTE(4)
505	COPY_BYTE(5)
506#endif
507	LOADB(t0, NBYTES-2(src), .Ll_exc\@)
508	SUB	len, len, 1
509	jr	ra
510	STOREB(t0, NBYTES-2(dst), .Ls_exc_p1\@)
511.Ldone\@:
512	jr	ra
513	 nop
514
515#ifdef CONFIG_CPU_MIPSR6
516.Lcopy_unaligned_bytes\@:
5171:
518	COPY_BYTE(0)
519	COPY_BYTE(1)
520	COPY_BYTE(2)
521	COPY_BYTE(3)
522	COPY_BYTE(4)
523	COPY_BYTE(5)
524	COPY_BYTE(6)
525	COPY_BYTE(7)
526	ADD	src, src, 8
527	b	1b
528	 ADD	dst, dst, 8
529#endif /* CONFIG_CPU_MIPSR6 */
530	.if __memcpy == 1
531	END(memcpy)
532	.set __memcpy, 0
533	.hidden __memcpy
534	.endif
535
536.Ll_exc_copy\@:
537	/*
538	 * Copy bytes from src until faulting load address (or until a
539	 * lb faults)
540	 *
541	 * When reached by a faulting LDFIRST/LDREST, THREAD_BUADDR($28)
542	 * may be more than a byte beyond the last address.
543	 * Hence, the lb below may get an exception.
544	 *
545	 * Assumes src < THREAD_BUADDR($28)
546	 */
547	LOADK	t0, TI_TASK($28)
548	 nop
549	LOADK	t0, THREAD_BUADDR(t0)
5501:
551	LOADB(t1, 0(src), .Ll_exc\@)
552	ADD	src, src, 1
553	sb	t1, 0(dst)	# can't fault -- we're copy_from_user
554	.set	reorder				/* DADDI_WAR */
555	ADD	dst, dst, 1
556	bne	src, t0, 1b
557	.set	noreorder
558.Ll_exc\@:
559	LOADK	t0, TI_TASK($28)
560	 nop
561	LOADK	t0, THREAD_BUADDR(t0)	# t0 is just past last good address
562	 nop
563	SUB	len, AT, t0		# len number of uncopied bytes
564	bnez	t6, .Ldone\@	/* Skip the zeroing part if inatomic */
565	/*
566	 * Here's where we rely on src and dst being incremented in tandem,
567	 *   See (3) above.
568	 * dst += (fault addr - src) to put dst at first byte to clear
569	 */
570	ADD	dst, t0			# compute start address in a1
571	SUB	dst, src
572	/*
573	 * Clear len bytes starting at dst.  Can't call __bzero because it
574	 * might modify len.  An inefficient loop for these rare times...
575	 */
576	.set	reorder				/* DADDI_WAR */
577	SUB	src, len, 1
578	beqz	len, .Ldone\@
579	.set	noreorder
5801:	sb	zero, 0(dst)
581	ADD	dst, dst, 1
582#ifndef CONFIG_CPU_DADDI_WORKAROUNDS
583	bnez	src, 1b
584	 SUB	src, src, 1
585#else
586	.set	push
587	.set	noat
588	li	v1, 1
589	bnez	src, 1b
590	 SUB	src, src, v1
591	.set	pop
592#endif
593	jr	ra
594	 nop
595
596
597#define SEXC(n)							\
598	.set	reorder;			/* DADDI_WAR */ \
599.Ls_exc_p ## n ## u\@:						\
600	ADD	len, len, n*NBYTES;				\
601	jr	ra;						\
602	.set	noreorder
603
604SEXC(8)
605SEXC(7)
606SEXC(6)
607SEXC(5)
608SEXC(4)
609SEXC(3)
610SEXC(2)
611SEXC(1)
612
613.Ls_exc_p1\@:
614	.set	reorder				/* DADDI_WAR */
615	ADD	len, len, 1
616	jr	ra
617	.set	noreorder
618.Ls_exc\@:
619	jr	ra
620	 nop
621	.endm
622
623	.align	5
624LEAF(memmove)
625	ADD	t0, a0, a2
626	ADD	t1, a1, a2
627	sltu	t0, a1, t0			# dst + len <= src -> memcpy
628	sltu	t1, a0, t1			# dst >= src + len -> memcpy
629	and	t0, t1
630	beqz	t0, .L__memcpy
631	 move	v0, a0				/* return value */
632	beqz	a2, .Lr_out
633	END(memmove)
634
635	/* fall through to __rmemcpy */
636LEAF(__rmemcpy)					/* a0=dst a1=src a2=len */
637	 sltu	t0, a1, a0
638	beqz	t0, .Lr_end_bytes_up		# src >= dst
639	 nop
640	ADD	a0, a2				# dst = dst + len
641	ADD	a1, a2				# src = src + len
642
643.Lr_end_bytes:
644	R10KCBARRIER(0(ra))
645	lb	t0, -1(a1)
646	SUB	a2, a2, 0x1
647	sb	t0, -1(a0)
648	SUB	a1, a1, 0x1
649	.set	reorder				/* DADDI_WAR */
650	SUB	a0, a0, 0x1
651	bnez	a2, .Lr_end_bytes
652	.set	noreorder
653
654.Lr_out:
655	jr	ra
656	 move	a2, zero
657
658.Lr_end_bytes_up:
659	R10KCBARRIER(0(ra))
660	lb	t0, (a1)
661	SUB	a2, a2, 0x1
662	sb	t0, (a0)
663	ADD	a1, a1, 0x1
664	.set	reorder				/* DADDI_WAR */
665	ADD	a0, a0, 0x1
666	bnez	a2, .Lr_end_bytes_up
667	.set	noreorder
668
669	jr	ra
670	 move	a2, zero
671	END(__rmemcpy)
672
673/*
674 * t6 is used as a flag to note inatomic mode.
675 */
676LEAF(__copy_user_inatomic)
677	b	__copy_user_common
678	li	t6, 1
679	END(__copy_user_inatomic)
680
681/*
682 * A combined memcpy/__copy_user
683 * __copy_user sets len to 0 for success; else to an upper bound of
684 * the number of uncopied bytes.
685 * memcpy sets v0 to dst.
686 */
687	.align	5
688LEAF(memcpy)					/* a0=dst a1=src a2=len */
689	move	v0, dst				/* return value */
690.L__memcpy:
691FEXPORT(__copy_user)
692	li	t6, 0	/* not inatomic */
693__copy_user_common:
694	/* Legacy Mode, user <-> user */
695	__BUILD_COPY_USER LEGACY_MODE USEROP USEROP
696
697#ifdef CONFIG_EVA
698
699/*
700 * For EVA we need distinct symbols for reading and writing to user space.
701 * This is because we need to use specific EVA instructions to perform the
702 * virtual <-> physical translation when a virtual address is actually in user
703 * space
704 */
705
706LEAF(__copy_user_inatomic_eva)
707	b       __copy_from_user_common
708	li	t6, 1
709	END(__copy_user_inatomic_eva)
710
711/*
712 * __copy_from_user (EVA)
713 */
714
715LEAF(__copy_from_user_eva)
716	li	t6, 0	/* not inatomic */
717__copy_from_user_common:
718	__BUILD_COPY_USER EVA_MODE USEROP KERNELOP
719END(__copy_from_user_eva)
720
721
722
723/*
724 * __copy_to_user (EVA)
725 */
726
727LEAF(__copy_to_user_eva)
728__BUILD_COPY_USER EVA_MODE KERNELOP USEROP
729END(__copy_to_user_eva)
730
731/*
732 * __copy_in_user (EVA)
733 */
734
735LEAF(__copy_in_user_eva)
736__BUILD_COPY_USER EVA_MODE USEROP USEROP
737END(__copy_in_user_eva)
738
739#endif
740