xref: /openbmc/linux/arch/ia64/lib/memcpy_mck.S (revision fb574682)
1/* SPDX-License-Identifier: GPL-2.0 */
2/*
3 * Itanium 2-optimized version of memcpy and copy_user function
4 *
5 * Inputs:
6 * 	in0:	destination address
7 *	in1:	source address
8 *	in2:	number of bytes to copy
9 * Output:
10 *	for memcpy:    return dest
11 * 	for copy_user: return 0 if success,
12 *		       or number of byte NOT copied if error occurred.
13 *
14 * Copyright (C) 2002 Intel Corp.
15 * Copyright (C) 2002 Ken Chen <kenneth.w.chen@intel.com>
16 */
17#include <asm/asmmacro.h>
18#include <asm/page.h>
19#include <asm/export.h>
20
21#define EK(y...) EX(y)
22
23/* McKinley specific optimization */
24
25#define retval		r8
26#define saved_pfs	r31
27#define saved_lc	r10
28#define saved_pr	r11
29#define saved_in0	r14
30#define saved_in1	r15
31#define saved_in2	r16
32
33#define src0		r2
34#define src1		r3
35#define dst0		r17
36#define dst1		r18
37#define cnt		r9
38
39/* r19-r30 are temp for each code section */
40#define PREFETCH_DIST	8
41#define src_pre_mem	r19
42#define dst_pre_mem	r20
43#define src_pre_l2	r21
44#define dst_pre_l2	r22
45#define t1		r23
46#define t2		r24
47#define t3		r25
48#define t4		r26
49#define t5		t1	// alias!
50#define t6		t2	// alias!
51#define t7		t3	// alias!
52#define n8		r27
53#define t9		t5	// alias!
54#define t10		t4	// alias!
55#define t11		t7	// alias!
56#define t12		t6	// alias!
57#define t14		t10	// alias!
58#define t13		r28
59#define t15		r29
60#define tmp		r30
61
62/* defines for long_copy block */
63#define	A	0
64#define B	(PREFETCH_DIST)
65#define C	(B + PREFETCH_DIST)
66#define D	(C + 1)
67#define N	(D + 1)
68#define Nrot	((N + 7) & ~7)
69
70/* alias */
71#define in0		r32
72#define in1		r33
73#define in2		r34
74
75GLOBAL_ENTRY(memcpy)
76	and	r28=0x7,in0
77	and	r29=0x7,in1
78	mov	f6=f0
79	mov	retval=in0
80	br.cond.sptk .common_code
81	;;
82END(memcpy)
83EXPORT_SYMBOL(memcpy)
84GLOBAL_ENTRY(__copy_user)
85	.prologue
86// check dest alignment
87	and	r28=0x7,in0
88	and	r29=0x7,in1
89	mov	f6=f1
90	mov	saved_in0=in0	// save dest pointer
91	mov	saved_in1=in1	// save src pointer
92	mov	retval=r0	// initialize return value
93	;;
94.common_code:
95	cmp.gt	p15,p0=8,in2	// check for small size
96	cmp.ne	p13,p0=0,r28	// check dest alignment
97	cmp.ne	p14,p0=0,r29	// check src alignment
98	add	src0=0,in1
99	sub	r30=8,r28	// for .align_dest
100	mov	saved_in2=in2	// save len
101	;;
102	add	dst0=0,in0
103	add	dst1=1,in0	// dest odd index
104	cmp.le	p6,p0 = 1,r30	// for .align_dest
105(p15)	br.cond.dpnt .memcpy_short
106(p13)	br.cond.dpnt .align_dest
107(p14)	br.cond.dpnt .unaligned_src
108	;;
109
110// both dest and src are aligned on 8-byte boundary
111.aligned_src:
112	.save ar.pfs, saved_pfs
113	alloc	saved_pfs=ar.pfs,3,Nrot-3,0,Nrot
114	.save pr, saved_pr
115	mov	saved_pr=pr
116
117	shr.u	cnt=in2,7	// this much cache line
118	;;
119	cmp.lt	p6,p0=2*PREFETCH_DIST,cnt
120	cmp.lt	p7,p8=1,cnt
121	.save ar.lc, saved_lc
122	mov	saved_lc=ar.lc
123	.body
124	add	cnt=-1,cnt
125	add	src_pre_mem=0,in1	// prefetch src pointer
126	add	dst_pre_mem=0,in0	// prefetch dest pointer
127	;;
128(p7)	mov	ar.lc=cnt	// prefetch count
129(p8)	mov	ar.lc=r0
130(p6)	br.cond.dpnt .long_copy
131	;;
132
133.prefetch:
134	lfetch.fault	  [src_pre_mem], 128
135	lfetch.fault.excl [dst_pre_mem], 128
136	br.cloop.dptk.few .prefetch
137	;;
138
139.medium_copy:
140	and	tmp=31,in2	// copy length after iteration
141	shr.u	r29=in2,5	// number of 32-byte iteration
142	add	dst1=8,dst0	// 2nd dest pointer
143	;;
144	add	cnt=-1,r29	// ctop iteration adjustment
145	cmp.eq	p10,p0=r29,r0	// do we really need to loop?
146	add	src1=8,src0	// 2nd src pointer
147	cmp.le	p6,p0=8,tmp
148	;;
149	cmp.le	p7,p0=16,tmp
150	mov	ar.lc=cnt	// loop setup
151	cmp.eq	p16,p17 = r0,r0
152	mov	ar.ec=2
153(p10)	br.dpnt.few .aligned_src_tail
154	;;
155	TEXT_ALIGN(32)
1561:
157EX(.ex_handler, (p16)	ld8	r34=[src0],16)
158EK(.ex_handler, (p16)	ld8	r38=[src1],16)
159EX(.ex_handler, (p17)	st8	[dst0]=r33,16)
160EK(.ex_handler, (p17)	st8	[dst1]=r37,16)
161	;;
162EX(.ex_handler, (p16)	ld8	r32=[src0],16)
163EK(.ex_handler, (p16)	ld8	r36=[src1],16)
164EX(.ex_handler, (p16)	st8	[dst0]=r34,16)
165EK(.ex_handler, (p16)	st8	[dst1]=r38,16)
166	br.ctop.dptk.few 1b
167	;;
168
169.aligned_src_tail:
170EX(.ex_handler, (p6)	ld8	t1=[src0])
171	mov	ar.lc=saved_lc
172	mov	ar.pfs=saved_pfs
173EX(.ex_hndlr_s, (p7)	ld8	t2=[src1],8)
174	cmp.le	p8,p0=24,tmp
175	and	r21=-8,tmp
176	;;
177EX(.ex_hndlr_s, (p8)	ld8	t3=[src1])
178EX(.ex_handler, (p6)	st8	[dst0]=t1)	// store byte 1
179	and	in2=7,tmp	// remaining length
180EX(.ex_hndlr_d, (p7)	st8	[dst1]=t2,8)	// store byte 2
181	add	src0=src0,r21	// setting up src pointer
182	add	dst0=dst0,r21	// setting up dest pointer
183	;;
184EX(.ex_handler, (p8)	st8	[dst1]=t3)	// store byte 3
185	mov	pr=saved_pr,-1
186	br.dptk.many .memcpy_short
187	;;
188
189/* code taken from copy_page_mck */
190.long_copy:
191	.rotr v[2*PREFETCH_DIST]
192	.rotp p[N]
193
194	mov src_pre_mem = src0
195	mov pr.rot = 0x10000
196	mov ar.ec = 1				// special unrolled loop
197
198	mov dst_pre_mem = dst0
199
200	add src_pre_l2 = 8*8, src0
201	add dst_pre_l2 = 8*8, dst0
202	;;
203	add src0 = 8, src_pre_mem		// first t1 src
204	mov ar.lc = 2*PREFETCH_DIST - 1
205	shr.u cnt=in2,7				// number of lines
206	add src1 = 3*8, src_pre_mem		// first t3 src
207	add dst0 = 8, dst_pre_mem		// first t1 dst
208	add dst1 = 3*8, dst_pre_mem		// first t3 dst
209	;;
210	and tmp=127,in2				// remaining bytes after this block
211	add cnt = -(2*PREFETCH_DIST) - 1, cnt
212	// same as .line_copy loop, but with all predicated-off instructions removed:
213.prefetch_loop:
214EX(.ex_hndlr_lcpy_1, (p[A])	ld8 v[A] = [src_pre_mem], 128)		// M0
215EK(.ex_hndlr_lcpy_1, (p[B])	st8 [dst_pre_mem] = v[B], 128)		// M2
216	br.ctop.sptk .prefetch_loop
217	;;
218	cmp.eq p16, p0 = r0, r0			// reset p16 to 1
219	mov ar.lc = cnt
220	mov ar.ec = N				// # of stages in pipeline
221	;;
222.line_copy:
223EX(.ex_handler,	(p[D])	ld8 t2 = [src0], 3*8)			// M0
224EK(.ex_handler,	(p[D])	ld8 t4 = [src1], 3*8)			// M1
225EX(.ex_handler_lcpy,	(p[B])	st8 [dst_pre_mem] = v[B], 128)		// M2 prefetch dst from memory
226EK(.ex_handler_lcpy,	(p[D])	st8 [dst_pre_l2] = n8, 128)		// M3 prefetch dst from L2
227	;;
228EX(.ex_handler_lcpy,	(p[A])	ld8 v[A] = [src_pre_mem], 128)		// M0 prefetch src from memory
229EK(.ex_handler_lcpy,	(p[C])	ld8 n8 = [src_pre_l2], 128)		// M1 prefetch src from L2
230EX(.ex_handler,	(p[D])	st8 [dst0] =  t1, 8)			// M2
231EK(.ex_handler,	(p[D])	st8 [dst1] =  t3, 8)			// M3
232	;;
233EX(.ex_handler,	(p[D])	ld8  t5 = [src0], 8)
234EK(.ex_handler,	(p[D])	ld8  t7 = [src1], 3*8)
235EX(.ex_handler,	(p[D])	st8 [dst0] =  t2, 3*8)
236EK(.ex_handler,	(p[D])	st8 [dst1] =  t4, 3*8)
237	;;
238EX(.ex_handler,	(p[D])	ld8  t6 = [src0], 3*8)
239EK(.ex_handler,	(p[D])	ld8 t10 = [src1], 8)
240EX(.ex_handler,	(p[D])	st8 [dst0] =  t5, 8)
241EK(.ex_handler,	(p[D])	st8 [dst1] =  t7, 3*8)
242	;;
243EX(.ex_handler,	(p[D])	ld8  t9 = [src0], 3*8)
244EK(.ex_handler,	(p[D])	ld8 t11 = [src1], 3*8)
245EX(.ex_handler,	(p[D])	st8 [dst0] =  t6, 3*8)
246EK(.ex_handler,	(p[D])	st8 [dst1] = t10, 8)
247	;;
248EX(.ex_handler,	(p[D])	ld8 t12 = [src0], 8)
249EK(.ex_handler,	(p[D])	ld8 t14 = [src1], 8)
250EX(.ex_handler,	(p[D])	st8 [dst0] =  t9, 3*8)
251EK(.ex_handler,	(p[D])	st8 [dst1] = t11, 3*8)
252	;;
253EX(.ex_handler,	(p[D])	ld8 t13 = [src0], 4*8)
254EK(.ex_handler,	(p[D])	ld8 t15 = [src1], 4*8)
255EX(.ex_handler,	(p[D])	st8 [dst0] = t12, 8)
256EK(.ex_handler,	(p[D])	st8 [dst1] = t14, 8)
257	;;
258EX(.ex_handler,	(p[C])	ld8  t1 = [src0], 8)
259EK(.ex_handler,	(p[C])	ld8  t3 = [src1], 8)
260EX(.ex_handler,	(p[D])	st8 [dst0] = t13, 4*8)
261EK(.ex_handler,	(p[D])	st8 [dst1] = t15, 4*8)
262	br.ctop.sptk .line_copy
263	;;
264
265	add dst0=-8,dst0
266	add src0=-8,src0
267	mov in2=tmp
268	.restore sp
269	br.sptk.many .medium_copy
270	;;
271
272#define BLOCK_SIZE	128*32
273#define blocksize	r23
274#define curlen		r24
275
276// dest is on 8-byte boundary, src is not. We need to do
277// ld8-ld8, shrp, then st8.  Max 8 byte copy per cycle.
278.unaligned_src:
279	.prologue
280	.save ar.pfs, saved_pfs
281	alloc	saved_pfs=ar.pfs,3,5,0,8
282	.save ar.lc, saved_lc
283	mov	saved_lc=ar.lc
284	.save pr, saved_pr
285	mov	saved_pr=pr
286	.body
287.4k_block:
288	mov	saved_in0=dst0	// need to save all input arguments
289	mov	saved_in2=in2
290	mov	blocksize=BLOCK_SIZE
291	;;
292	cmp.lt	p6,p7=blocksize,in2
293	mov	saved_in1=src0
294	;;
295(p6)	mov	in2=blocksize
296	;;
297	shr.u	r21=in2,7	// this much cache line
298	shr.u	r22=in2,4	// number of 16-byte iteration
299	and	curlen=15,in2	// copy length after iteration
300	and	r30=7,src0	// source alignment
301	;;
302	cmp.lt	p7,p8=1,r21
303	add	cnt=-1,r21
304	;;
305
306	add	src_pre_mem=0,src0	// prefetch src pointer
307	add	dst_pre_mem=0,dst0	// prefetch dest pointer
308	and	src0=-8,src0		// 1st src pointer
309(p7)	mov	ar.lc = cnt
310(p8)	mov	ar.lc = r0
311	;;
312	TEXT_ALIGN(32)
3131:	lfetch.fault	  [src_pre_mem], 128
314	lfetch.fault.excl [dst_pre_mem], 128
315	br.cloop.dptk.few 1b
316	;;
317
318	shladd	dst1=r22,3,dst0	// 2nd dest pointer
319	shladd	src1=r22,3,src0	// 2nd src pointer
320	cmp.eq	p8,p9=r22,r0	// do we really need to loop?
321	cmp.le	p6,p7=8,curlen;	// have at least 8 byte remaining?
322	add	cnt=-1,r22	// ctop iteration adjustment
323	;;
324EX(.ex_handler, (p9)	ld8	r33=[src0],8)	// loop primer
325EK(.ex_handler, (p9)	ld8	r37=[src1],8)
326(p8)	br.dpnt.few .noloop
327	;;
328
329// The jump address is calculated based on src alignment. The COPYU
330// macro below need to confine its size to power of two, so an entry
331// can be caulated using shl instead of an expensive multiply. The
332// size is then hard coded by the following #define to match the
333// actual size.  This make it somewhat tedious when COPYU macro gets
334// changed and this need to be adjusted to match.
335#define LOOP_SIZE 6
3361:
337	mov	r29=ip		// jmp_table thread
338	mov	ar.lc=cnt
339	;;
340	add	r29=.jump_table - 1b - (.jmp1-.jump_table), r29
341	shl	r28=r30, LOOP_SIZE	// jmp_table thread
342	mov	ar.ec=2		// loop setup
343	;;
344	add	r29=r29,r28		// jmp_table thread
345	cmp.eq	p16,p17=r0,r0
346	;;
347	mov	b6=r29			// jmp_table thread
348	;;
349	br.cond.sptk.few b6
350
351// for 8-15 byte case
352// We will skip the loop, but need to replicate the side effect
353// that the loop produces.
354.noloop:
355EX(.ex_handler, (p6)	ld8	r37=[src1],8)
356	add	src0=8,src0
357(p6)	shl	r25=r30,3
358	;;
359EX(.ex_handler, (p6)	ld8	r27=[src1])
360(p6)	shr.u	r28=r37,r25
361(p6)	sub	r26=64,r25
362	;;
363(p6)	shl	r27=r27,r26
364	;;
365(p6)	or	r21=r28,r27
366
367.unaligned_src_tail:
368/* check if we have more than blocksize to copy, if so go back */
369	cmp.gt	p8,p0=saved_in2,blocksize
370	;;
371(p8)	add	dst0=saved_in0,blocksize
372(p8)	add	src0=saved_in1,blocksize
373(p8)	sub	in2=saved_in2,blocksize
374(p8)	br.dpnt	.4k_block
375	;;
376
377/* we have up to 15 byte to copy in the tail.
378 * part of work is already done in the jump table code
379 * we are at the following state.
380 * src side:
381 *
382 *   xxxxxx xx                   <----- r21 has xxxxxxxx already
383 * -------- -------- --------
384 * 0        8        16
385 *          ^
386 *          |
387 *          src1
388 *
389 * dst
390 * -------- -------- --------
391 * ^
392 * |
393 * dst1
394 */
395EX(.ex_handler, (p6)	st8	[dst1]=r21,8)	// more than 8 byte to copy
396(p6)	add	curlen=-8,curlen	// update length
397	mov	ar.pfs=saved_pfs
398	;;
399	mov	ar.lc=saved_lc
400	mov	pr=saved_pr,-1
401	mov	in2=curlen	// remaining length
402	mov	dst0=dst1	// dest pointer
403	add	src0=src1,r30	// forward by src alignment
404	;;
405
406// 7 byte or smaller.
407.memcpy_short:
408	cmp.le	p8,p9   = 1,in2
409	cmp.le	p10,p11 = 2,in2
410	cmp.le	p12,p13 = 3,in2
411	cmp.le	p14,p15 = 4,in2
412	add	src1=1,src0	// second src pointer
413	add	dst1=1,dst0	// second dest pointer
414	;;
415
416EX(.ex_handler_short, (p8)	ld1	t1=[src0],2)
417EK(.ex_handler_short, (p10)	ld1	t2=[src1],2)
418(p9)	br.ret.dpnt rp		// 0 byte copy
419	;;
420
421EX(.ex_handler_short, (p8)	st1	[dst0]=t1,2)
422EK(.ex_handler_short, (p10)	st1	[dst1]=t2,2)
423(p11)	br.ret.dpnt rp		// 1 byte copy
424
425EX(.ex_handler_short, (p12)	ld1	t3=[src0],2)
426EK(.ex_handler_short, (p14)	ld1	t4=[src1],2)
427(p13)	br.ret.dpnt rp		// 2 byte copy
428	;;
429
430	cmp.le	p6,p7   = 5,in2
431	cmp.le	p8,p9   = 6,in2
432	cmp.le	p10,p11 = 7,in2
433
434EX(.ex_handler_short, (p12)	st1	[dst0]=t3,2)
435EK(.ex_handler_short, (p14)	st1	[dst1]=t4,2)
436(p15)	br.ret.dpnt rp		// 3 byte copy
437	;;
438
439EX(.ex_handler_short, (p6)	ld1	t5=[src0],2)
440EK(.ex_handler_short, (p8)	ld1	t6=[src1],2)
441(p7)	br.ret.dpnt rp		// 4 byte copy
442	;;
443
444EX(.ex_handler_short, (p6)	st1	[dst0]=t5,2)
445EK(.ex_handler_short, (p8)	st1	[dst1]=t6,2)
446(p9)	br.ret.dptk rp		// 5 byte copy
447
448EX(.ex_handler_short, (p10)	ld1	t7=[src0],2)
449(p11)	br.ret.dptk rp		// 6 byte copy
450	;;
451
452EX(.ex_handler_short, (p10)	st1	[dst0]=t7,2)
453	br.ret.dptk rp		// done all cases
454
455
456/* Align dest to nearest 8-byte boundary. We know we have at
457 * least 7 bytes to copy, enough to crawl to 8-byte boundary.
458 * Actual number of byte to crawl depend on the dest alignment.
459 * 7 byte or less is taken care at .memcpy_short
460
461 * src0 - source even index
462 * src1 - source  odd index
463 * dst0 - dest even index
464 * dst1 - dest  odd index
465 * r30  - distance to 8-byte boundary
466 */
467
468.align_dest:
469	add	src1=1,in1	// source odd index
470	cmp.le	p7,p0 = 2,r30	// for .align_dest
471	cmp.le	p8,p0 = 3,r30	// for .align_dest
472EX(.ex_handler_short, (p6)	ld1	t1=[src0],2)
473	cmp.le	p9,p0 = 4,r30	// for .align_dest
474	cmp.le	p10,p0 = 5,r30
475	;;
476EX(.ex_handler_short, (p7)	ld1	t2=[src1],2)
477EK(.ex_handler_short, (p8)	ld1	t3=[src0],2)
478	cmp.le	p11,p0 = 6,r30
479EX(.ex_handler_short, (p6)	st1	[dst0] = t1,2)
480	cmp.le	p12,p0 = 7,r30
481	;;
482EX(.ex_handler_short, (p9)	ld1	t4=[src1],2)
483EK(.ex_handler_short, (p10)	ld1	t5=[src0],2)
484EX(.ex_handler_short, (p7)	st1	[dst1] = t2,2)
485EK(.ex_handler_short, (p8)	st1	[dst0] = t3,2)
486	;;
487EX(.ex_handler_short, (p11)	ld1	t6=[src1],2)
488EK(.ex_handler_short, (p12)	ld1	t7=[src0],2)
489	cmp.eq	p6,p7=r28,r29
490EX(.ex_handler_short, (p9)	st1	[dst1] = t4,2)
491EK(.ex_handler_short, (p10)	st1	[dst0] = t5,2)
492	sub	in2=in2,r30
493	;;
494EX(.ex_handler_short, (p11)	st1	[dst1] = t6,2)
495EK(.ex_handler_short, (p12)	st1	[dst0] = t7)
496	add	dst0=in0,r30	// setup arguments
497	add	src0=in1,r30
498(p6)	br.cond.dptk .aligned_src
499(p7)	br.cond.dpnt .unaligned_src
500	;;
501
502/* main loop body in jump table format */
503#define COPYU(shift)									\
5041:											\
505EX(.ex_handler,  (p16)	ld8	r32=[src0],8);		/* 1 */				\
506EK(.ex_handler,  (p16)	ld8	r36=[src1],8);						\
507		 (p17)	shrp	r35=r33,r34,shift;;	/* 1 */				\
508EX(.ex_handler,  (p6)	ld8	r22=[src1]);	/* common, prime for tail section */	\
509		 nop.m	0;								\
510		 (p16)	shrp	r38=r36,r37,shift;					\
511EX(.ex_handler,  (p17)	st8	[dst0]=r35,8);		/* 1 */				\
512EK(.ex_handler,  (p17)	st8	[dst1]=r39,8);						\
513		 br.ctop.dptk.few 1b;;							\
514		 (p7)	add	src1=-8,src1;	/* back out for <8 byte case */		\
515		 shrp	r21=r22,r38,shift;	/* speculative work */			\
516		 br.sptk.few .unaligned_src_tail /* branch out of jump table */		\
517		 ;;
518	TEXT_ALIGN(32)
519.jump_table:
520	COPYU(8)	// unaligned cases
521.jmp1:
522	COPYU(16)
523	COPYU(24)
524	COPYU(32)
525	COPYU(40)
526	COPYU(48)
527	COPYU(56)
528
529#undef A
530#undef B
531#undef C
532#undef D
533
534/*
535 * Due to lack of local tag support in gcc 2.x assembler, it is not clear which
536 * instruction failed in the bundle.  The exception algorithm is that we
537 * first figure out the faulting address, then detect if there is any
538 * progress made on the copy, if so, redo the copy from last known copied
539 * location up to the faulting address (exclusive). In the copy_from_user
540 * case, remaining byte in kernel buffer will be zeroed.
541 *
542 * Take copy_from_user as an example, in the code there are multiple loads
543 * in a bundle and those multiple loads could span over two pages, the
544 * faulting address is calculated as page_round_down(max(src0, src1)).
545 * This is based on knowledge that if we can access one byte in a page, we
546 * can access any byte in that page.
547 *
548 * predicate used in the exception handler:
549 * p6-p7: direction
550 * p10-p11: src faulting addr calculation
551 * p12-p13: dst faulting addr calculation
552 */
553
554#define A	r19
555#define B	r20
556#define C	r21
557#define D	r22
558#define F	r28
559
560#define saved_retval	loc0
561#define saved_rtlink	loc1
562#define saved_pfs_stack	loc2
563
564.ex_hndlr_s:
565	add	src0=8,src0
566	br.sptk .ex_handler
567	;;
568.ex_hndlr_d:
569	add	dst0=8,dst0
570	br.sptk .ex_handler
571	;;
572.ex_hndlr_lcpy_1:
573	mov	src1=src_pre_mem
574	mov	dst1=dst_pre_mem
575	cmp.gtu	p10,p11=src_pre_mem,saved_in1
576	cmp.gtu	p12,p13=dst_pre_mem,saved_in0
577	;;
578(p10)	add	src0=8,saved_in1
579(p11)	mov	src0=saved_in1
580(p12)	add	dst0=8,saved_in0
581(p13)	mov	dst0=saved_in0
582	br.sptk	.ex_handler
583.ex_handler_lcpy:
584	// in line_copy block, the preload addresses should always ahead
585	// of the other two src/dst pointers.  Furthermore, src1/dst1 should
586	// always ahead of src0/dst0.
587	mov	src1=src_pre_mem
588	mov	dst1=dst_pre_mem
589.ex_handler:
590	mov	pr=saved_pr,-1		// first restore pr, lc, and pfs
591	mov	ar.lc=saved_lc
592	mov	ar.pfs=saved_pfs
593	;;
594.ex_handler_short: // fault occurred in these sections didn't change pr, lc, pfs
595	cmp.ltu	p6,p7=saved_in0, saved_in1	// get the copy direction
596	cmp.ltu	p10,p11=src0,src1
597	cmp.ltu	p12,p13=dst0,dst1
598	fcmp.eq	p8,p0=f6,f0		// is it memcpy?
599	mov	tmp = dst0
600	;;
601(p11)	mov	src1 = src0		// pick the larger of the two
602(p13)	mov	dst0 = dst1		// make dst0 the smaller one
603(p13)	mov	dst1 = tmp		// and dst1 the larger one
604	;;
605(p6)	dep	F = r0,dst1,0,PAGE_SHIFT // usr dst round down to page boundary
606(p7)	dep	F = r0,src1,0,PAGE_SHIFT // usr src round down to page boundary
607	;;
608(p6)	cmp.le	p14,p0=dst0,saved_in0	// no progress has been made on store
609(p7)	cmp.le	p14,p0=src0,saved_in1	// no progress has been made on load
610	mov	retval=saved_in2
611(p8)	ld1	tmp=[src1]		// force an oops for memcpy call
612(p8)	st1	[dst1]=r0		// force an oops for memcpy call
613(p14)	br.ret.sptk.many rp
614
615/*
616 * The remaining byte to copy is calculated as:
617 *
618 * A =	(faulting_addr - orig_src)	-> len to faulting ld address
619 *	or
620 * 	(faulting_addr - orig_dst)	-> len to faulting st address
621 * B =	(cur_dst - orig_dst)		-> len copied so far
622 * C =	A - B				-> len need to be copied
623 * D =	orig_len - A			-> len need to be left along
624 */
625(p6)	sub	A = F, saved_in0
626(p7)	sub	A = F, saved_in1
627	clrrrb
628	;;
629	alloc	saved_pfs_stack=ar.pfs,3,3,3,0
630	cmp.lt	p8,p0=A,r0
631	sub	B = dst0, saved_in0	// how many byte copied so far
632	;;
633(p8)	mov	A = 0;			// A shouldn't be negative, cap it
634	;;
635	sub	C = A, B
636	sub	D = saved_in2, A
637	;;
638	cmp.gt	p8,p0=C,r0		// more than 1 byte?
639	mov	r8=0
640	mov	saved_retval = D
641	mov	saved_rtlink = b0
642
643	add	out0=saved_in0, B
644	add	out1=saved_in1, B
645	mov	out2=C
646(p8)	br.call.sptk.few b0=__copy_user	// recursive call
647	;;
648
649	add	saved_retval=saved_retval,r8	// above might return non-zero value
650	;;
651
652	mov	retval=saved_retval
653	mov	ar.pfs=saved_pfs_stack
654	mov	b0=saved_rtlink
655	br.ret.sptk.many rp
656
657/* end of McKinley specific optimization */
658END(__copy_user)
659EXPORT_SYMBOL(__copy_user)
660