2 * Itanium 2-optimized version of memcpy and copy_user function
5 * in0: destination address
7 * in2: number of bytes to copy
9 * 0 if success, or number of byte NOT copied if error occurred.
11 * Copyright (C) 2002 Intel Corp.
12 * Copyright (C) 2002 Ken Chen <kenneth.w.chen@intel.com>
14 #include <linux/config.h>
15 #include <asm/asmmacro.h>
18 #define EK(y...) EX(y)
20 /* McKinley specific optimization */
36 /* r19-r30 are temp for each code section */
37 #define PREFETCH_DIST 8
38 #define src_pre_mem r19
39 #define dst_pre_mem r20
40 #define src_pre_l2 r21
41 #define dst_pre_l2 r22
46 #define t5 t1 // alias!
47 #define t6 t2 // alias!
48 #define t7 t3 // alias!
50 #define t9 t5 // alias!
51 #define t10 t4 // alias!
52 #define t11 t7 // alias!
53 #define t12 t6 // alias!
54 #define t14 t10 // alias!
59 /* defines for long_copy block */
61 #define B (PREFETCH_DIST)
62 #define C (B + PREFETCH_DIST)
65 #define Nrot ((N + 7) & ~7)
76 br.cond.sptk .common_code
79 GLOBAL_ENTRY(__copy_user)
81 // check dest alignment
85 mov saved_in0=in0 // save dest pointer
86 mov saved_in1=in1 // save src pointer
87 mov saved_in2=in2 // save len
90 cmp.gt p15,p0=8,in2 // check for small size
91 cmp.ne p13,p0=0,r28 // check dest alignment
92 cmp.ne p14,p0=0,r29 // check src alignment
94 sub r30=8,r28 // for .align_dest
95 mov retval=r0 // initialize return value
98 add dst1=1,in0 // dest odd index
99 cmp.le p6,p0 = 1,r30 // for .align_dest
100 (p15) br.cond.dpnt .memcpy_short
101 (p13) br.cond.dpnt .align_dest
102 (p14) br.cond.dpnt .unaligned_src
105 // both dest and src are aligned on 8-byte boundary
107 .save ar.pfs, saved_pfs
108 alloc saved_pfs=ar.pfs,3,Nrot-3,0,Nrot
112 shr.u cnt=in2,7 // this much cache line
114 cmp.lt p6,p0=2*PREFETCH_DIST,cnt
116 .save ar.lc, saved_lc
120 add src_pre_mem=0,in1 // prefetch src pointer
121 add dst_pre_mem=0,in0 // prefetch dest pointer
123 (p7) mov ar.lc=cnt // prefetch count
125 (p6) br.cond.dpnt .long_copy
129 lfetch.fault [src_pre_mem], 128
130 lfetch.fault.excl [dst_pre_mem], 128
131 br.cloop.dptk.few .prefetch
135 and tmp=31,in2 // copy length after iteration
136 shr.u r29=in2,5 // number of 32-byte iteration
137 add dst1=8,dst0 // 2nd dest pointer
139 add cnt=-1,r29 // ctop iteration adjustment
140 cmp.eq p10,p0=r29,r0 // do we really need to loop?
141 add src1=8,src0 // 2nd src pointer
145 mov ar.lc=cnt // loop setup
146 cmp.eq p16,p17 = r0,r0
148 (p10) br.dpnt.few .aligned_src_tail
152 EX(.ex_handler, (p16) ld8 r34=[src0],16)
153 EK(.ex_handler, (p16) ld8 r38=[src1],16)
154 EX(.ex_handler, (p17) st8 [dst0]=r33,16)
155 EK(.ex_handler, (p17) st8 [dst1]=r37,16)
157 EX(.ex_handler, (p16) ld8 r32=[src0],16)
158 EK(.ex_handler, (p16) ld8 r36=[src1],16)
159 EX(.ex_handler, (p16) st8 [dst0]=r34,16)
160 EK(.ex_handler, (p16) st8 [dst1]=r38,16)
165 EX(.ex_handler, (p6) ld8 t1=[src0])
168 EX(.ex_hndlr_s, (p7) ld8 t2=[src1],8)
172 EX(.ex_hndlr_s, (p8) ld8 t3=[src1])
173 EX(.ex_handler, (p6) st8 [dst0]=t1) // store byte 1
174 and in2=7,tmp // remaining length
175 EX(.ex_hndlr_d, (p7) st8 [dst1]=t2,8) // store byte 2
176 add src0=src0,r21 // setting up src pointer
177 add dst0=dst0,r21 // setting up dest pointer
179 EX(.ex_handler, (p8) st8 [dst1]=t3) // store byte 3
181 br.dptk.many .memcpy_short
184 /* code taken from copy_page_mck */
186 .rotr v[2*PREFETCH_DIST]
189 mov src_pre_mem = src0
191 mov ar.ec = 1 // special unrolled loop
193 mov dst_pre_mem = dst0
195 add src_pre_l2 = 8*8, src0
196 add dst_pre_l2 = 8*8, dst0
198 add src0 = 8, src_pre_mem // first t1 src
199 mov ar.lc = 2*PREFETCH_DIST - 1
200 shr.u cnt=in2,7 // number of lines
201 add src1 = 3*8, src_pre_mem // first t3 src
202 add dst0 = 8, dst_pre_mem // first t1 dst
203 add dst1 = 3*8, dst_pre_mem // first t3 dst
205 and tmp=127,in2 // remaining bytes after this block
206 add cnt = -(2*PREFETCH_DIST) - 1, cnt
207 // same as .line_copy loop, but with all predicated-off instructions removed:
209 EX(.ex_hndlr_lcpy_1, (p[A]) ld8 v[A] = [src_pre_mem], 128) // M0
210 EK(.ex_hndlr_lcpy_1, (p[B]) st8 [dst_pre_mem] = v[B], 128) // M2
211 br.ctop.sptk .prefetch_loop
213 cmp.eq p16, p0 = r0, r0 // reset p16 to 1
215 mov ar.ec = N // # of stages in pipeline
218 EX(.ex_handler, (p[D]) ld8 t2 = [src0], 3*8) // M0
219 EK(.ex_handler, (p[D]) ld8 t4 = [src1], 3*8) // M1
220 EX(.ex_handler_lcpy, (p[B]) st8 [dst_pre_mem] = v[B], 128) // M2 prefetch dst from memory
221 EK(.ex_handler_lcpy, (p[D]) st8 [dst_pre_l2] = n8, 128) // M3 prefetch dst from L2
223 EX(.ex_handler_lcpy, (p[A]) ld8 v[A] = [src_pre_mem], 128) // M0 prefetch src from memory
224 EK(.ex_handler_lcpy, (p[C]) ld8 n8 = [src_pre_l2], 128) // M1 prefetch src from L2
225 EX(.ex_handler, (p[D]) st8 [dst0] = t1, 8) // M2
226 EK(.ex_handler, (p[D]) st8 [dst1] = t3, 8) // M3
228 EX(.ex_handler, (p[D]) ld8 t5 = [src0], 8)
229 EK(.ex_handler, (p[D]) ld8 t7 = [src1], 3*8)
230 EX(.ex_handler, (p[D]) st8 [dst0] = t2, 3*8)
231 EK(.ex_handler, (p[D]) st8 [dst1] = t4, 3*8)
233 EX(.ex_handler, (p[D]) ld8 t6 = [src0], 3*8)
234 EK(.ex_handler, (p[D]) ld8 t10 = [src1], 8)
235 EX(.ex_handler, (p[D]) st8 [dst0] = t5, 8)
236 EK(.ex_handler, (p[D]) st8 [dst1] = t7, 3*8)
238 EX(.ex_handler, (p[D]) ld8 t9 = [src0], 3*8)
239 EK(.ex_handler, (p[D]) ld8 t11 = [src1], 3*8)
240 EX(.ex_handler, (p[D]) st8 [dst0] = t6, 3*8)
241 EK(.ex_handler, (p[D]) st8 [dst1] = t10, 8)
243 EX(.ex_handler, (p[D]) ld8 t12 = [src0], 8)
244 EK(.ex_handler, (p[D]) ld8 t14 = [src1], 8)
245 EX(.ex_handler, (p[D]) st8 [dst0] = t9, 3*8)
246 EK(.ex_handler, (p[D]) st8 [dst1] = t11, 3*8)
248 EX(.ex_handler, (p[D]) ld8 t13 = [src0], 4*8)
249 EK(.ex_handler, (p[D]) ld8 t15 = [src1], 4*8)
250 EX(.ex_handler, (p[D]) st8 [dst0] = t12, 8)
251 EK(.ex_handler, (p[D]) st8 [dst1] = t14, 8)
253 EX(.ex_handler, (p[C]) ld8 t1 = [src0], 8)
254 EK(.ex_handler, (p[C]) ld8 t3 = [src1], 8)
255 EX(.ex_handler, (p[D]) st8 [dst0] = t13, 4*8)
256 EK(.ex_handler, (p[D]) st8 [dst1] = t15, 4*8)
257 br.ctop.sptk .line_copy
264 br.sptk.many .medium_copy
267 #define BLOCK_SIZE 128*32
268 #define blocksize r23
271 // dest is on 8-byte boundary, src is not. We need to do
272 // ld8-ld8, shrp, then st8. Max 8 byte copy per cycle.
275 .save ar.pfs, saved_pfs
276 alloc saved_pfs=ar.pfs,3,5,0,8
277 .save ar.lc, saved_lc
283 mov saved_in0=dst0 // need to save all input arguments
285 mov blocksize=BLOCK_SIZE
287 cmp.lt p6,p7=blocksize,in2
290 (p6) mov in2=blocksize
292 shr.u r21=in2,7 // this much cache line
293 shr.u r22=in2,4 // number of 16-byte iteration
294 and curlen=15,in2 // copy length after iteration
295 and r30=7,src0 // source alignment
301 add src_pre_mem=0,src0 // prefetch src pointer
302 add dst_pre_mem=0,dst0 // prefetch dest pointer
303 and src0=-8,src0 // 1st src pointer
308 1: lfetch.fault [src_pre_mem], 128
309 lfetch.fault.excl [dst_pre_mem], 128
313 shladd dst1=r22,3,dst0 // 2nd dest pointer
314 shladd src1=r22,3,src0 // 2nd src pointer
315 cmp.eq p8,p9=r22,r0 // do we really need to loop?
316 cmp.le p6,p7=8,curlen; // have at least 8 byte remaining?
317 add cnt=-1,r22 // ctop iteration adjustment
319 EX(.ex_handler, (p9) ld8 r33=[src0],8) // loop primer
320 EK(.ex_handler, (p9) ld8 r37=[src1],8)
321 (p8) br.dpnt.few .noloop
324 // The jump address is calculated based on src alignment. The COPYU
325 // macro below need to confine its size to power of two, so an entry
326 // can be caulated using shl instead of an expensive multiply. The
327 // size is then hard coded by the following #define to match the
328 // actual size. This make it somewhat tedious when COPYU macro gets
329 // changed and this need to be adjusted to match.
332 mov r29=ip // jmp_table thread
335 add r29=.jump_table - 1b - (.jmp1-.jump_table), r29
336 shl r28=r30, LOOP_SIZE // jmp_table thread
337 mov ar.ec=2 // loop setup
339 add r29=r29,r28 // jmp_table thread
342 mov b6=r29 // jmp_table thread
346 // for 8-15 byte case
347 // We will skip the loop, but need to replicate the side effect
348 // that the loop produces.
350 EX(.ex_handler, (p6) ld8 r37=[src1],8)
354 EX(.ex_handler, (p6) ld8 r27=[src1])
355 (p6) shr.u r28=r37,r25
363 /* check if we have more than blocksize to copy, if so go back */
364 cmp.gt p8,p0=saved_in2,blocksize
366 (p8) add dst0=saved_in0,blocksize
367 (p8) add src0=saved_in1,blocksize
368 (p8) sub in2=saved_in2,blocksize
369 (p8) br.dpnt .4k_block
372 /* we have up to 15 byte to copy in the tail.
373 * part of work is already done in the jump table code
374 * we are at the following state.
377 * xxxxxx xx <----- r21 has xxxxxxxx already
378 * -------- -------- --------
385 * -------- -------- --------
390 EX(.ex_handler, (p6) st8 [dst1]=r21,8) // more than 8 byte to copy
391 (p6) add curlen=-8,curlen // update length
396 mov in2=curlen // remaining length
397 mov dst0=dst1 // dest pointer
398 add src0=src1,r30 // forward by src alignment
401 // 7 byte or smaller.
404 cmp.le p10,p11 = 2,in2
405 cmp.le p12,p13 = 3,in2
406 cmp.le p14,p15 = 4,in2
407 add src1=1,src0 // second src pointer
408 add dst1=1,dst0 // second dest pointer
411 EX(.ex_handler_short, (p8) ld1 t1=[src0],2)
412 EK(.ex_handler_short, (p10) ld1 t2=[src1],2)
413 (p9) br.ret.dpnt rp // 0 byte copy
416 EX(.ex_handler_short, (p8) st1 [dst0]=t1,2)
417 EK(.ex_handler_short, (p10) st1 [dst1]=t2,2)
418 (p11) br.ret.dpnt rp // 1 byte copy
420 EX(.ex_handler_short, (p12) ld1 t3=[src0],2)
421 EK(.ex_handler_short, (p14) ld1 t4=[src1],2)
422 (p13) br.ret.dpnt rp // 2 byte copy
427 cmp.le p10,p11 = 7,in2
429 EX(.ex_handler_short, (p12) st1 [dst0]=t3,2)
430 EK(.ex_handler_short, (p14) st1 [dst1]=t4,2)
431 (p15) br.ret.dpnt rp // 3 byte copy
434 EX(.ex_handler_short, (p6) ld1 t5=[src0],2)
435 EK(.ex_handler_short, (p8) ld1 t6=[src1],2)
436 (p7) br.ret.dpnt rp // 4 byte copy
439 EX(.ex_handler_short, (p6) st1 [dst0]=t5,2)
440 EK(.ex_handler_short, (p8) st1 [dst1]=t6,2)
441 (p9) br.ret.dptk rp // 5 byte copy
443 EX(.ex_handler_short, (p10) ld1 t7=[src0],2)
444 (p11) br.ret.dptk rp // 6 byte copy
447 EX(.ex_handler_short, (p10) st1 [dst0]=t7,2)
448 br.ret.dptk rp // done all cases
451 /* Align dest to nearest 8-byte boundary. We know we have at
452 * least 7 bytes to copy, enough to crawl to 8-byte boundary.
453 * Actual number of byte to crawl depend on the dest alignment.
454 * 7 byte or less is taken care at .memcpy_short
456 * src0 - source even index
457 * src1 - source odd index
458 * dst0 - dest even index
459 * dst1 - dest odd index
460 * r30 - distance to 8-byte boundary
464 add src1=1,in1 // source odd index
465 cmp.le p7,p0 = 2,r30 // for .align_dest
466 cmp.le p8,p0 = 3,r30 // for .align_dest
467 EX(.ex_handler_short, (p6) ld1 t1=[src0],2)
468 cmp.le p9,p0 = 4,r30 // for .align_dest
469 cmp.le p10,p0 = 5,r30
471 EX(.ex_handler_short, (p7) ld1 t2=[src1],2)
472 EK(.ex_handler_short, (p8) ld1 t3=[src0],2)
473 cmp.le p11,p0 = 6,r30
474 EX(.ex_handler_short, (p6) st1 [dst0] = t1,2)
475 cmp.le p12,p0 = 7,r30
477 EX(.ex_handler_short, (p9) ld1 t4=[src1],2)
478 EK(.ex_handler_short, (p10) ld1 t5=[src0],2)
479 EX(.ex_handler_short, (p7) st1 [dst1] = t2,2)
480 EK(.ex_handler_short, (p8) st1 [dst0] = t3,2)
482 EX(.ex_handler_short, (p11) ld1 t6=[src1],2)
483 EK(.ex_handler_short, (p12) ld1 t7=[src0],2)
485 EX(.ex_handler_short, (p9) st1 [dst1] = t4,2)
486 EK(.ex_handler_short, (p10) st1 [dst0] = t5,2)
489 EX(.ex_handler_short, (p11) st1 [dst1] = t6,2)
490 EK(.ex_handler_short, (p12) st1 [dst0] = t7)
491 add dst0=in0,r30 // setup arguments
493 (p6) br.cond.dptk .aligned_src
494 (p7) br.cond.dpnt .unaligned_src
497 /* main loop body in jump table format */
498 #define COPYU(shift) \
500 EX(.ex_handler, (p16) ld8 r32=[src0],8); /* 1 */ \
501 EK(.ex_handler, (p16) ld8 r36=[src1],8); \
502 (p17) shrp r35=r33,r34,shift;; /* 1 */ \
503 EX(.ex_handler, (p6) ld8 r22=[src1]); /* common, prime for tail section */ \
505 (p16) shrp r38=r36,r37,shift; \
506 EX(.ex_handler, (p17) st8 [dst0]=r35,8); /* 1 */ \
507 EK(.ex_handler, (p17) st8 [dst1]=r39,8); \
508 br.ctop.dptk.few 1b;; \
509 (p7) add src1=-8,src1; /* back out for <8 byte case */ \
510 shrp r21=r22,r38,shift; /* speculative work */ \
511 br.sptk.few .unaligned_src_tail /* branch out of jump table */ \
515 COPYU(8) // unaligned cases
530 * Due to lack of local tag support in gcc 2.x assembler, it is not clear which
531 * instruction failed in the bundle. The exception algorithm is that we
532 * first figure out the faulting address, then detect if there is any
533 * progress made on the copy, if so, redo the copy from last known copied
534 * location up to the faulting address (exclusive). In the copy_from_user
535 * case, remaining byte in kernel buffer will be zeroed.
537 * Take copy_from_user as an example, in the code there are multiple loads
538 * in a bundle and those multiple loads could span over two pages, the
539 * faulting address is calculated as page_round_down(max(src0, src1)).
540 * This is based on knowledge that if we can access one byte in a page, we
541 * can access any byte in that page.
543 * predicate used in the exception handler:
545 * p10-p11: src faulting addr calculation
546 * p12-p13: dst faulting addr calculation
555 #define memset_arg0 r32
556 #define memset_arg2 r33
558 #define saved_retval loc0
559 #define saved_rtlink loc1
560 #define saved_pfs_stack loc2
573 cmp.gtu p10,p11=src_pre_mem,saved_in1
574 cmp.gtu p12,p13=dst_pre_mem,saved_in0
576 (p10) add src0=8,saved_in1
577 (p11) mov src0=saved_in1
578 (p12) add dst0=8,saved_in0
579 (p13) mov dst0=saved_in0
582 // in line_copy block, the preload addresses should always ahead
583 // of the other two src/dst pointers. Furthermore, src1/dst1 should
584 // always ahead of src0/dst0.
588 mov pr=saved_pr,-1 // first restore pr, lc, and pfs
592 .ex_handler_short: // fault occurred in these sections didn't change pr, lc, pfs
593 cmp.ltu p6,p7=saved_in0, saved_in1 // get the copy direction
594 cmp.ltu p10,p11=src0,src1
595 cmp.ltu p12,p13=dst0,dst1
596 fcmp.eq p8,p0=f6,f0 // is it memcpy?
599 (p11) mov src1 = src0 // pick the larger of the two
600 (p13) mov dst0 = dst1 // make dst0 the smaller one
601 (p13) mov dst1 = tmp // and dst1 the larger one
603 (p6) dep F = r0,dst1,0,PAGE_SHIFT // usr dst round down to page boundary
604 (p7) dep F = r0,src1,0,PAGE_SHIFT // usr src round down to page boundary
606 (p6) cmp.le p14,p0=dst0,saved_in0 // no progress has been made on store
607 (p7) cmp.le p14,p0=src0,saved_in1 // no progress has been made on load
609 (p8) ld1 tmp=[src1] // force an oops for memcpy call
610 (p8) st1 [dst1]=r0 // force an oops for memcpy call
611 (p14) br.ret.sptk.many rp
614 * The remaining byte to copy is calculated as:
616 * A = (faulting_addr - orig_src) -> len to faulting ld address
618 * (faulting_addr - orig_dst) -> len to faulting st address
619 * B = (cur_dst - orig_dst) -> len copied so far
620 * C = A - B -> len need to be copied
621 * D = orig_len - A -> len need to be zeroed
623 (p6) sub A = F, saved_in0
624 (p7) sub A = F, saved_in1
627 alloc saved_pfs_stack=ar.pfs,3,3,3,0
629 sub B = dst0, saved_in0 // how many byte copied so far
631 (p8) mov A = 0; // A shouldn't be negative, cap it
636 cmp.gt p8,p0=C,r0 // more than 1 byte?
637 add memset_arg0=saved_in0, A
638 (p6) mov memset_arg2=0 // copy_to_user should not call memset
639 (p7) mov memset_arg2=D // copy_from_user need to have kbuf zeroed
642 mov saved_rtlink = b0
644 add out0=saved_in0, B
645 add out1=saved_in1, B
647 (p8) br.call.sptk.few b0=__copy_user // recursive call
650 add saved_retval=saved_retval,r8 // above might return non-zero value
651 cmp.gt p8,p0=memset_arg2,r0 // more than 1 byte?
652 mov out0=memset_arg0 // *s
654 mov out2=memset_arg2 // n
655 (p8) br.call.sptk.few b0=memset
658 mov retval=saved_retval
659 mov ar.pfs=saved_pfs_stack
663 /* end of McKinley specific optimization */