2 * Itanium 2-optimized version of memcpy and copy_user function
5 * in0: destination address
7 * in2: number of bytes to copy
9 * for memcpy: return dest
10 * for copy_user: return 0 if success,
11 * or number of byte NOT copied if error occurred.
13 * Copyright (C) 2002 Intel Corp.
14 * Copyright (C) 2002 Ken Chen <kenneth.w.chen@intel.com>
16 #include <linux/config.h>
17 #include <asm/asmmacro.h>
20 #define EK(y...) EX(y)
22 /* McKinley specific optimization */
38 /* r19-r30 are temp for each code section */
39 #define PREFETCH_DIST 8
40 #define src_pre_mem r19
41 #define dst_pre_mem r20
42 #define src_pre_l2 r21
43 #define dst_pre_l2 r22
48 #define t5 t1 // alias!
49 #define t6 t2 // alias!
50 #define t7 t3 // alias!
52 #define t9 t5 // alias!
53 #define t10 t4 // alias!
54 #define t11 t7 // alias!
55 #define t12 t6 // alias!
56 #define t14 t10 // alias!
61 /* defines for long_copy block */
63 #define B (PREFETCH_DIST)
64 #define C (B + PREFETCH_DIST)
67 #define Nrot ((N + 7) & ~7)
79 br.cond.sptk .common_code
82 GLOBAL_ENTRY(__copy_user)
84 // check dest alignment
88 mov saved_in0=in0 // save dest pointer
89 mov saved_in1=in1 // save src pointer
90 mov retval=r0 // initialize return value
93 cmp.gt p15,p0=8,in2 // check for small size
94 cmp.ne p13,p0=0,r28 // check dest alignment
95 cmp.ne p14,p0=0,r29 // check src alignment
97 sub r30=8,r28 // for .align_dest
98 mov saved_in2=in2 // save len
101 add dst1=1,in0 // dest odd index
102 cmp.le p6,p0 = 1,r30 // for .align_dest
103 (p15) br.cond.dpnt .memcpy_short
104 (p13) br.cond.dpnt .align_dest
105 (p14) br.cond.dpnt .unaligned_src
108 // both dest and src are aligned on 8-byte boundary
110 .save ar.pfs, saved_pfs
111 alloc saved_pfs=ar.pfs,3,Nrot-3,0,Nrot
115 shr.u cnt=in2,7 // this much cache line
117 cmp.lt p6,p0=2*PREFETCH_DIST,cnt
119 .save ar.lc, saved_lc
123 add src_pre_mem=0,in1 // prefetch src pointer
124 add dst_pre_mem=0,in0 // prefetch dest pointer
126 (p7) mov ar.lc=cnt // prefetch count
128 (p6) br.cond.dpnt .long_copy
132 lfetch.fault [src_pre_mem], 128
133 lfetch.fault.excl [dst_pre_mem], 128
134 br.cloop.dptk.few .prefetch
138 and tmp=31,in2 // copy length after iteration
139 shr.u r29=in2,5 // number of 32-byte iteration
140 add dst1=8,dst0 // 2nd dest pointer
142 add cnt=-1,r29 // ctop iteration adjustment
143 cmp.eq p10,p0=r29,r0 // do we really need to loop?
144 add src1=8,src0 // 2nd src pointer
148 mov ar.lc=cnt // loop setup
149 cmp.eq p16,p17 = r0,r0
151 (p10) br.dpnt.few .aligned_src_tail
155 EX(.ex_handler, (p16) ld8 r34=[src0],16)
156 EK(.ex_handler, (p16) ld8 r38=[src1],16)
157 EX(.ex_handler, (p17) st8 [dst0]=r33,16)
158 EK(.ex_handler, (p17) st8 [dst1]=r37,16)
160 EX(.ex_handler, (p16) ld8 r32=[src0],16)
161 EK(.ex_handler, (p16) ld8 r36=[src1],16)
162 EX(.ex_handler, (p16) st8 [dst0]=r34,16)
163 EK(.ex_handler, (p16) st8 [dst1]=r38,16)
168 EX(.ex_handler, (p6) ld8 t1=[src0])
171 EX(.ex_hndlr_s, (p7) ld8 t2=[src1],8)
175 EX(.ex_hndlr_s, (p8) ld8 t3=[src1])
176 EX(.ex_handler, (p6) st8 [dst0]=t1) // store byte 1
177 and in2=7,tmp // remaining length
178 EX(.ex_hndlr_d, (p7) st8 [dst1]=t2,8) // store byte 2
179 add src0=src0,r21 // setting up src pointer
180 add dst0=dst0,r21 // setting up dest pointer
182 EX(.ex_handler, (p8) st8 [dst1]=t3) // store byte 3
184 br.dptk.many .memcpy_short
187 /* code taken from copy_page_mck */
189 .rotr v[2*PREFETCH_DIST]
192 mov src_pre_mem = src0
194 mov ar.ec = 1 // special unrolled loop
196 mov dst_pre_mem = dst0
198 add src_pre_l2 = 8*8, src0
199 add dst_pre_l2 = 8*8, dst0
201 add src0 = 8, src_pre_mem // first t1 src
202 mov ar.lc = 2*PREFETCH_DIST - 1
203 shr.u cnt=in2,7 // number of lines
204 add src1 = 3*8, src_pre_mem // first t3 src
205 add dst0 = 8, dst_pre_mem // first t1 dst
206 add dst1 = 3*8, dst_pre_mem // first t3 dst
208 and tmp=127,in2 // remaining bytes after this block
209 add cnt = -(2*PREFETCH_DIST) - 1, cnt
210 // same as .line_copy loop, but with all predicated-off instructions removed:
212 EX(.ex_hndlr_lcpy_1, (p[A]) ld8 v[A] = [src_pre_mem], 128) // M0
213 EK(.ex_hndlr_lcpy_1, (p[B]) st8 [dst_pre_mem] = v[B], 128) // M2
214 br.ctop.sptk .prefetch_loop
216 cmp.eq p16, p0 = r0, r0 // reset p16 to 1
218 mov ar.ec = N // # of stages in pipeline
221 EX(.ex_handler, (p[D]) ld8 t2 = [src0], 3*8) // M0
222 EK(.ex_handler, (p[D]) ld8 t4 = [src1], 3*8) // M1
223 EX(.ex_handler_lcpy, (p[B]) st8 [dst_pre_mem] = v[B], 128) // M2 prefetch dst from memory
224 EK(.ex_handler_lcpy, (p[D]) st8 [dst_pre_l2] = n8, 128) // M3 prefetch dst from L2
226 EX(.ex_handler_lcpy, (p[A]) ld8 v[A] = [src_pre_mem], 128) // M0 prefetch src from memory
227 EK(.ex_handler_lcpy, (p[C]) ld8 n8 = [src_pre_l2], 128) // M1 prefetch src from L2
228 EX(.ex_handler, (p[D]) st8 [dst0] = t1, 8) // M2
229 EK(.ex_handler, (p[D]) st8 [dst1] = t3, 8) // M3
231 EX(.ex_handler, (p[D]) ld8 t5 = [src0], 8)
232 EK(.ex_handler, (p[D]) ld8 t7 = [src1], 3*8)
233 EX(.ex_handler, (p[D]) st8 [dst0] = t2, 3*8)
234 EK(.ex_handler, (p[D]) st8 [dst1] = t4, 3*8)
236 EX(.ex_handler, (p[D]) ld8 t6 = [src0], 3*8)
237 EK(.ex_handler, (p[D]) ld8 t10 = [src1], 8)
238 EX(.ex_handler, (p[D]) st8 [dst0] = t5, 8)
239 EK(.ex_handler, (p[D]) st8 [dst1] = t7, 3*8)
241 EX(.ex_handler, (p[D]) ld8 t9 = [src0], 3*8)
242 EK(.ex_handler, (p[D]) ld8 t11 = [src1], 3*8)
243 EX(.ex_handler, (p[D]) st8 [dst0] = t6, 3*8)
244 EK(.ex_handler, (p[D]) st8 [dst1] = t10, 8)
246 EX(.ex_handler, (p[D]) ld8 t12 = [src0], 8)
247 EK(.ex_handler, (p[D]) ld8 t14 = [src1], 8)
248 EX(.ex_handler, (p[D]) st8 [dst0] = t9, 3*8)
249 EK(.ex_handler, (p[D]) st8 [dst1] = t11, 3*8)
251 EX(.ex_handler, (p[D]) ld8 t13 = [src0], 4*8)
252 EK(.ex_handler, (p[D]) ld8 t15 = [src1], 4*8)
253 EX(.ex_handler, (p[D]) st8 [dst0] = t12, 8)
254 EK(.ex_handler, (p[D]) st8 [dst1] = t14, 8)
256 EX(.ex_handler, (p[C]) ld8 t1 = [src0], 8)
257 EK(.ex_handler, (p[C]) ld8 t3 = [src1], 8)
258 EX(.ex_handler, (p[D]) st8 [dst0] = t13, 4*8)
259 EK(.ex_handler, (p[D]) st8 [dst1] = t15, 4*8)
260 br.ctop.sptk .line_copy
267 br.sptk.many .medium_copy
270 #define BLOCK_SIZE 128*32
271 #define blocksize r23
274 // dest is on 8-byte boundary, src is not. We need to do
275 // ld8-ld8, shrp, then st8. Max 8 byte copy per cycle.
278 .save ar.pfs, saved_pfs
279 alloc saved_pfs=ar.pfs,3,5,0,8
280 .save ar.lc, saved_lc
286 mov saved_in0=dst0 // need to save all input arguments
288 mov blocksize=BLOCK_SIZE
290 cmp.lt p6,p7=blocksize,in2
293 (p6) mov in2=blocksize
295 shr.u r21=in2,7 // this much cache line
296 shr.u r22=in2,4 // number of 16-byte iteration
297 and curlen=15,in2 // copy length after iteration
298 and r30=7,src0 // source alignment
304 add src_pre_mem=0,src0 // prefetch src pointer
305 add dst_pre_mem=0,dst0 // prefetch dest pointer
306 and src0=-8,src0 // 1st src pointer
311 1: lfetch.fault [src_pre_mem], 128
312 lfetch.fault.excl [dst_pre_mem], 128
316 shladd dst1=r22,3,dst0 // 2nd dest pointer
317 shladd src1=r22,3,src0 // 2nd src pointer
318 cmp.eq p8,p9=r22,r0 // do we really need to loop?
319 cmp.le p6,p7=8,curlen; // have at least 8 byte remaining?
320 add cnt=-1,r22 // ctop iteration adjustment
322 EX(.ex_handler, (p9) ld8 r33=[src0],8) // loop primer
323 EK(.ex_handler, (p9) ld8 r37=[src1],8)
324 (p8) br.dpnt.few .noloop
327 // The jump address is calculated based on src alignment. The COPYU
328 // macro below need to confine its size to power of two, so an entry
329 // can be caulated using shl instead of an expensive multiply. The
330 // size is then hard coded by the following #define to match the
331 // actual size. This make it somewhat tedious when COPYU macro gets
332 // changed and this need to be adjusted to match.
335 mov r29=ip // jmp_table thread
338 add r29=.jump_table - 1b - (.jmp1-.jump_table), r29
339 shl r28=r30, LOOP_SIZE // jmp_table thread
340 mov ar.ec=2 // loop setup
342 add r29=r29,r28 // jmp_table thread
345 mov b6=r29 // jmp_table thread
349 // for 8-15 byte case
350 // We will skip the loop, but need to replicate the side effect
351 // that the loop produces.
353 EX(.ex_handler, (p6) ld8 r37=[src1],8)
357 EX(.ex_handler, (p6) ld8 r27=[src1])
358 (p6) shr.u r28=r37,r25
366 /* check if we have more than blocksize to copy, if so go back */
367 cmp.gt p8,p0=saved_in2,blocksize
369 (p8) add dst0=saved_in0,blocksize
370 (p8) add src0=saved_in1,blocksize
371 (p8) sub in2=saved_in2,blocksize
372 (p8) br.dpnt .4k_block
375 /* we have up to 15 byte to copy in the tail.
376 * part of work is already done in the jump table code
377 * we are at the following state.
380 * xxxxxx xx <----- r21 has xxxxxxxx already
381 * -------- -------- --------
388 * -------- -------- --------
393 EX(.ex_handler, (p6) st8 [dst1]=r21,8) // more than 8 byte to copy
394 (p6) add curlen=-8,curlen // update length
399 mov in2=curlen // remaining length
400 mov dst0=dst1 // dest pointer
401 add src0=src1,r30 // forward by src alignment
404 // 7 byte or smaller.
407 cmp.le p10,p11 = 2,in2
408 cmp.le p12,p13 = 3,in2
409 cmp.le p14,p15 = 4,in2
410 add src1=1,src0 // second src pointer
411 add dst1=1,dst0 // second dest pointer
414 EX(.ex_handler_short, (p8) ld1 t1=[src0],2)
415 EK(.ex_handler_short, (p10) ld1 t2=[src1],2)
416 (p9) br.ret.dpnt rp // 0 byte copy
419 EX(.ex_handler_short, (p8) st1 [dst0]=t1,2)
420 EK(.ex_handler_short, (p10) st1 [dst1]=t2,2)
421 (p11) br.ret.dpnt rp // 1 byte copy
423 EX(.ex_handler_short, (p12) ld1 t3=[src0],2)
424 EK(.ex_handler_short, (p14) ld1 t4=[src1],2)
425 (p13) br.ret.dpnt rp // 2 byte copy
430 cmp.le p10,p11 = 7,in2
432 EX(.ex_handler_short, (p12) st1 [dst0]=t3,2)
433 EK(.ex_handler_short, (p14) st1 [dst1]=t4,2)
434 (p15) br.ret.dpnt rp // 3 byte copy
437 EX(.ex_handler_short, (p6) ld1 t5=[src0],2)
438 EK(.ex_handler_short, (p8) ld1 t6=[src1],2)
439 (p7) br.ret.dpnt rp // 4 byte copy
442 EX(.ex_handler_short, (p6) st1 [dst0]=t5,2)
443 EK(.ex_handler_short, (p8) st1 [dst1]=t6,2)
444 (p9) br.ret.dptk rp // 5 byte copy
446 EX(.ex_handler_short, (p10) ld1 t7=[src0],2)
447 (p11) br.ret.dptk rp // 6 byte copy
450 EX(.ex_handler_short, (p10) st1 [dst0]=t7,2)
451 br.ret.dptk rp // done all cases
454 /* Align dest to nearest 8-byte boundary. We know we have at
455 * least 7 bytes to copy, enough to crawl to 8-byte boundary.
456 * Actual number of byte to crawl depend on the dest alignment.
457 * 7 byte or less is taken care at .memcpy_short
459 * src0 - source even index
460 * src1 - source odd index
461 * dst0 - dest even index
462 * dst1 - dest odd index
463 * r30 - distance to 8-byte boundary
467 add src1=1,in1 // source odd index
468 cmp.le p7,p0 = 2,r30 // for .align_dest
469 cmp.le p8,p0 = 3,r30 // for .align_dest
470 EX(.ex_handler_short, (p6) ld1 t1=[src0],2)
471 cmp.le p9,p0 = 4,r30 // for .align_dest
472 cmp.le p10,p0 = 5,r30
474 EX(.ex_handler_short, (p7) ld1 t2=[src1],2)
475 EK(.ex_handler_short, (p8) ld1 t3=[src0],2)
476 cmp.le p11,p0 = 6,r30
477 EX(.ex_handler_short, (p6) st1 [dst0] = t1,2)
478 cmp.le p12,p0 = 7,r30
480 EX(.ex_handler_short, (p9) ld1 t4=[src1],2)
481 EK(.ex_handler_short, (p10) ld1 t5=[src0],2)
482 EX(.ex_handler_short, (p7) st1 [dst1] = t2,2)
483 EK(.ex_handler_short, (p8) st1 [dst0] = t3,2)
485 EX(.ex_handler_short, (p11) ld1 t6=[src1],2)
486 EK(.ex_handler_short, (p12) ld1 t7=[src0],2)
488 EX(.ex_handler_short, (p9) st1 [dst1] = t4,2)
489 EK(.ex_handler_short, (p10) st1 [dst0] = t5,2)
492 EX(.ex_handler_short, (p11) st1 [dst1] = t6,2)
493 EK(.ex_handler_short, (p12) st1 [dst0] = t7)
494 add dst0=in0,r30 // setup arguments
496 (p6) br.cond.dptk .aligned_src
497 (p7) br.cond.dpnt .unaligned_src
500 /* main loop body in jump table format */
501 #define COPYU(shift) \
503 EX(.ex_handler, (p16) ld8 r32=[src0],8); /* 1 */ \
504 EK(.ex_handler, (p16) ld8 r36=[src1],8); \
505 (p17) shrp r35=r33,r34,shift;; /* 1 */ \
506 EX(.ex_handler, (p6) ld8 r22=[src1]); /* common, prime for tail section */ \
508 (p16) shrp r38=r36,r37,shift; \
509 EX(.ex_handler, (p17) st8 [dst0]=r35,8); /* 1 */ \
510 EK(.ex_handler, (p17) st8 [dst1]=r39,8); \
511 br.ctop.dptk.few 1b;; \
512 (p7) add src1=-8,src1; /* back out for <8 byte case */ \
513 shrp r21=r22,r38,shift; /* speculative work */ \
514 br.sptk.few .unaligned_src_tail /* branch out of jump table */ \
518 COPYU(8) // unaligned cases
533 * Due to lack of local tag support in gcc 2.x assembler, it is not clear which
534 * instruction failed in the bundle. The exception algorithm is that we
535 * first figure out the faulting address, then detect if there is any
536 * progress made on the copy, if so, redo the copy from last known copied
537 * location up to the faulting address (exclusive). In the copy_from_user
538 * case, remaining byte in kernel buffer will be zeroed.
540 * Take copy_from_user as an example, in the code there are multiple loads
541 * in a bundle and those multiple loads could span over two pages, the
542 * faulting address is calculated as page_round_down(max(src0, src1)).
543 * This is based on knowledge that if we can access one byte in a page, we
544 * can access any byte in that page.
546 * predicate used in the exception handler:
548 * p10-p11: src faulting addr calculation
549 * p12-p13: dst faulting addr calculation
558 #define memset_arg0 r32
559 #define memset_arg2 r33
561 #define saved_retval loc0
562 #define saved_rtlink loc1
563 #define saved_pfs_stack loc2
576 cmp.gtu p10,p11=src_pre_mem,saved_in1
577 cmp.gtu p12,p13=dst_pre_mem,saved_in0
579 (p10) add src0=8,saved_in1
580 (p11) mov src0=saved_in1
581 (p12) add dst0=8,saved_in0
582 (p13) mov dst0=saved_in0
585 // in line_copy block, the preload addresses should always ahead
586 // of the other two src/dst pointers. Furthermore, src1/dst1 should
587 // always ahead of src0/dst0.
591 mov pr=saved_pr,-1 // first restore pr, lc, and pfs
595 .ex_handler_short: // fault occurred in these sections didn't change pr, lc, pfs
596 cmp.ltu p6,p7=saved_in0, saved_in1 // get the copy direction
597 cmp.ltu p10,p11=src0,src1
598 cmp.ltu p12,p13=dst0,dst1
599 fcmp.eq p8,p0=f6,f0 // is it memcpy?
602 (p11) mov src1 = src0 // pick the larger of the two
603 (p13) mov dst0 = dst1 // make dst0 the smaller one
604 (p13) mov dst1 = tmp // and dst1 the larger one
606 (p6) dep F = r0,dst1,0,PAGE_SHIFT // usr dst round down to page boundary
607 (p7) dep F = r0,src1,0,PAGE_SHIFT // usr src round down to page boundary
609 (p6) cmp.le p14,p0=dst0,saved_in0 // no progress has been made on store
610 (p7) cmp.le p14,p0=src0,saved_in1 // no progress has been made on load
612 (p8) ld1 tmp=[src1] // force an oops for memcpy call
613 (p8) st1 [dst1]=r0 // force an oops for memcpy call
614 (p14) br.ret.sptk.many rp
617 * The remaining byte to copy is calculated as:
619 * A = (faulting_addr - orig_src) -> len to faulting ld address
621 * (faulting_addr - orig_dst) -> len to faulting st address
622 * B = (cur_dst - orig_dst) -> len copied so far
623 * C = A - B -> len need to be copied
624 * D = orig_len - A -> len need to be zeroed
626 (p6) sub A = F, saved_in0
627 (p7) sub A = F, saved_in1
630 alloc saved_pfs_stack=ar.pfs,3,3,3,0
632 sub B = dst0, saved_in0 // how many byte copied so far
634 (p8) mov A = 0; // A shouldn't be negative, cap it
639 cmp.gt p8,p0=C,r0 // more than 1 byte?
640 add memset_arg0=saved_in0, A
641 (p6) mov memset_arg2=0 // copy_to_user should not call memset
642 (p7) mov memset_arg2=D // copy_from_user need to have kbuf zeroed
645 mov saved_rtlink = b0
647 add out0=saved_in0, B
648 add out1=saved_in1, B
650 (p8) br.call.sptk.few b0=__copy_user // recursive call
653 add saved_retval=saved_retval,r8 // above might return non-zero value
654 cmp.gt p8,p0=memset_arg2,r0 // more than 1 byte?
655 mov out0=memset_arg0 // *s
657 mov out2=memset_arg2 // n
658 (p8) br.call.sptk.few b0=memset
661 mov retval=saved_retval
662 mov ar.pfs=saved_pfs_stack
666 /* end of McKinley specific optimization */