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)
29 /* McKinley specific optimization */
45 /* r19-r30 are temp for each code section */
46 #define PREFETCH_DIST 8
47 #define src_pre_mem r19
48 #define dst_pre_mem r20
49 #define src_pre_l2 r21
50 #define dst_pre_l2 r22
55 #define t5 t1 // alias!
56 #define t6 t2 // alias!
57 #define t7 t3 // alias!
59 #define t9 t5 // alias!
60 #define t10 t4 // alias!
61 #define t11 t7 // alias!
62 #define t12 t6 // alias!
63 #define t14 t10 // alias!
68 /* defines for long_copy block */
70 #define B (PREFETCH_DIST)
71 #define C (B + PREFETCH_DIST)
74 #define Nrot ((N + 7) & ~7)
85 br.cond.sptk .common_code
87 GLOBAL_ENTRY(__copy_user)
89 // check dest alignment
93 mov saved_in0=in0 // save dest pointer
94 mov saved_in1=in1 // save src pointer
95 mov saved_in2=in2 // save len
98 cmp.gt p15,p0=8,in2 // check for small size
99 cmp.ne p13,p0=0,r28 // check dest alignment
100 cmp.ne p14,p0=0,r29 // check src alignment
102 sub r30=8,r28 // for .align_dest
103 mov retval=r0 // initialize return value
106 add dst1=1,in0 // dest odd index
107 cmp.le p6,p0 = 1,r30 // for .align_dest
108 (p15) br.cond.dpnt .memcpy_short
109 (p13) br.cond.dpnt .align_dest
110 (p14) br.cond.dpnt .unaligned_src
113 // both dest and src are aligned on 8-byte boundary
115 .save ar.pfs, saved_pfs
116 alloc saved_pfs=ar.pfs,3,Nrot-3,0,Nrot
120 shr.u cnt=in2,7 // this much cache line
122 cmp.lt p6,p0=2*PREFETCH_DIST,cnt
124 .save ar.lc, saved_lc
128 add src_pre_mem=0,in1 // prefetch src pointer
129 add dst_pre_mem=0,in0 // prefetch dest pointer
131 (p7) mov ar.lc=cnt // prefetch count
133 (p6) br.cond.dpnt .long_copy
137 lfetch.fault [src_pre_mem], 128
138 lfetch.fault.excl [dst_pre_mem], 128
139 br.cloop.dptk.few .prefetch
143 and tmp=31,in2 // copy length after iteration
144 shr.u r29=in2,5 // number of 32-byte iteration
145 add dst1=8,dst0 // 2nd dest pointer
147 add cnt=-1,r29 // ctop iteration adjustment
148 cmp.eq p10,p0=r29,r0 // do we really need to loop?
149 add src1=8,src0 // 2nd src pointer
153 mov ar.lc=cnt // loop setup
154 cmp.eq p16,p17 = r0,r0
156 (p10) br.dpnt.few .aligned_src_tail
160 EX(.ex_handler, (p16) ld8 r34=[src0],16)
161 EK(.ex_handler, (p16) ld8 r38=[src1],16)
162 EX(.ex_handler, (p17) st8 [dst0]=r33,16)
163 EK(.ex_handler, (p17) st8 [dst1]=r37,16)
165 EX(.ex_handler, (p16) ld8 r32=[src0],16)
166 EK(.ex_handler, (p16) ld8 r36=[src1],16)
167 EX(.ex_handler, (p16) st8 [dst0]=r34,16)
168 EK(.ex_handler, (p16) st8 [dst1]=r38,16)
173 EX(.ex_handler, (p6) ld8 t1=[src0])
176 EX(.ex_hndlr_s, (p7) ld8 t2=[src1],8)
180 EX(.ex_hndlr_s, (p8) ld8 t3=[src1])
181 EX(.ex_handler, (p6) st8 [dst0]=t1) // store byte 1
182 and in2=7,tmp // remaining length
183 EX(.ex_hndlr_d, (p7) st8 [dst1]=t2,8) // store byte 2
184 add src0=src0,r21 // setting up src pointer
185 add dst0=dst0,r21 // setting up dest pointer
187 EX(.ex_handler, (p8) st8 [dst1]=t3) // store byte 3
189 br.dptk.many .memcpy_short
192 /* code taken from copy_page_mck */
194 .rotr v[2*PREFETCH_DIST]
197 mov src_pre_mem = src0
199 mov ar.ec = 1 // special unrolled loop
201 mov dst_pre_mem = dst0
203 add src_pre_l2 = 8*8, src0
204 add dst_pre_l2 = 8*8, dst0
206 add src0 = 8, src_pre_mem // first t1 src
207 mov ar.lc = 2*PREFETCH_DIST - 1
208 shr.u cnt=in2,7 // number of lines
209 add src1 = 3*8, src_pre_mem // first t3 src
210 add dst0 = 8, dst_pre_mem // first t1 dst
211 add dst1 = 3*8, dst_pre_mem // first t3 dst
213 and tmp=127,in2 // remaining bytes after this block
214 add cnt = -(2*PREFETCH_DIST) - 1, cnt
215 // same as .line_copy loop, but with all predicated-off instructions removed:
217 EX(.ex_hndlr_lcpy_1, (p[A]) ld8 v[A] = [src_pre_mem], 128) // M0
218 EK(.ex_hndlr_lcpy_1, (p[B]) st8 [dst_pre_mem] = v[B], 128) // M2
219 br.ctop.sptk .prefetch_loop
221 cmp.eq p16, p0 = r0, r0 // reset p16 to 1
223 mov ar.ec = N // # of stages in pipeline
226 EX(.ex_handler, (p[D]) ld8 t2 = [src0], 3*8) // M0
227 EK(.ex_handler, (p[D]) ld8 t4 = [src1], 3*8) // M1
228 EX(.ex_handler_lcpy, (p[B]) st8 [dst_pre_mem] = v[B], 128) // M2 prefetch dst from memory
229 EK(.ex_handler_lcpy, (p[D]) st8 [dst_pre_l2] = n8, 128) // M3 prefetch dst from L2
231 EX(.ex_handler_lcpy, (p[A]) ld8 v[A] = [src_pre_mem], 128) // M0 prefetch src from memory
232 EK(.ex_handler_lcpy, (p[C]) ld8 n8 = [src_pre_l2], 128) // M1 prefetch src from L2
233 EX(.ex_handler, (p[D]) st8 [dst0] = t1, 8) // M2
234 EK(.ex_handler, (p[D]) st8 [dst1] = t3, 8) // M3
236 EX(.ex_handler, (p[D]) ld8 t5 = [src0], 8)
237 EK(.ex_handler, (p[D]) ld8 t7 = [src1], 3*8)
238 EX(.ex_handler, (p[D]) st8 [dst0] = t2, 3*8)
239 EK(.ex_handler, (p[D]) st8 [dst1] = t4, 3*8)
241 EX(.ex_handler, (p[D]) ld8 t6 = [src0], 3*8)
242 EK(.ex_handler, (p[D]) ld8 t10 = [src1], 8)
243 EX(.ex_handler, (p[D]) st8 [dst0] = t5, 8)
244 EK(.ex_handler, (p[D]) st8 [dst1] = t7, 3*8)
246 EX(.ex_handler, (p[D]) ld8 t9 = [src0], 3*8)
247 EK(.ex_handler, (p[D]) ld8 t11 = [src1], 3*8)
248 EX(.ex_handler, (p[D]) st8 [dst0] = t6, 3*8)
249 EK(.ex_handler, (p[D]) st8 [dst1] = t10, 8)
251 EX(.ex_handler, (p[D]) ld8 t12 = [src0], 8)
252 EK(.ex_handler, (p[D]) ld8 t14 = [src1], 8)
253 EX(.ex_handler, (p[D]) st8 [dst0] = t9, 3*8)
254 EK(.ex_handler, (p[D]) st8 [dst1] = t11, 3*8)
256 EX(.ex_handler, (p[D]) ld8 t13 = [src0], 4*8)
257 EK(.ex_handler, (p[D]) ld8 t15 = [src1], 4*8)
258 EX(.ex_handler, (p[D]) st8 [dst0] = t12, 8)
259 EK(.ex_handler, (p[D]) st8 [dst1] = t14, 8)
261 EX(.ex_handler, (p[C]) ld8 t1 = [src0], 8)
262 EK(.ex_handler, (p[C]) ld8 t3 = [src1], 8)
263 EX(.ex_handler, (p[D]) st8 [dst0] = t13, 4*8)
264 EK(.ex_handler, (p[D]) st8 [dst1] = t15, 4*8)
265 br.ctop.sptk .line_copy
272 br.sptk.many .medium_copy
275 #define BLOCK_SIZE 128*32
276 #define blocksize r23
279 // dest is on 8-byte boundary, src is not. We need to do
280 // ld8-ld8, shrp, then st8. Max 8 byte copy per cycle.
283 .save ar.pfs, saved_pfs
284 alloc saved_pfs=ar.pfs,3,5,0,8
285 .save ar.lc, saved_lc
291 mov saved_in0=dst0 // need to save all input arguments
293 mov blocksize=BLOCK_SIZE
295 cmp.lt p6,p7=blocksize,in2
298 (p6) mov in2=blocksize
300 shr.u r21=in2,7 // this much cache line
301 shr.u r22=in2,4 // number of 16-byte iteration
302 and curlen=15,in2 // copy length after iteration
303 and r30=7,src0 // source alignment
309 add src_pre_mem=0,src0 // prefetch src pointer
310 add dst_pre_mem=0,dst0 // prefetch dest pointer
311 and src0=-8,src0 // 1st src pointer
316 1: lfetch.fault [src_pre_mem], 128
317 lfetch.fault.excl [dst_pre_mem], 128
321 shladd dst1=r22,3,dst0 // 2nd dest pointer
322 shladd src1=r22,3,src0 // 2nd src pointer
323 cmp.eq p8,p9=r22,r0 // do we really need to loop?
324 cmp.le p6,p7=8,curlen; // have at least 8 byte remaining?
325 add cnt=-1,r22 // ctop iteration adjustment
327 EX(.ex_handler, (p9) ld8 r33=[src0],8) // loop primer
328 EK(.ex_handler, (p9) ld8 r37=[src1],8)
329 (p8) br.dpnt.few .noloop
332 // The jump address is calculated based on src alignment. The COPYU
333 // macro below need to confine its size to power of two, so an entry
334 // can be caulated using shl instead of an expensive multiply. The
335 // size is then hard coded by the following #define to match the
336 // actual size. This make it somewhat tedious when COPYU macro gets
337 // changed and this need to be adjusted to match.
340 mov r29=ip // jmp_table thread
343 add r29=.jump_table - 1b - (.jmp1-.jump_table), r29
344 shl r28=r30, LOOP_SIZE // jmp_table thread
345 mov ar.ec=2 // loop setup
347 add r29=r29,r28 // jmp_table thread
350 mov b6=r29 // jmp_table thread
354 // for 8-15 byte case
355 // We will skip the loop, but need to replicate the side effect
356 // that the loop produces.
358 EX(.ex_handler, (p6) ld8 r37=[src1],8)
362 EX(.ex_handler, (p6) ld8 r27=[src1])
363 (p6) shr.u r28=r37,r25
371 /* check if we have more than blocksize to copy, if so go back */
372 cmp.gt p8,p0=saved_in2,blocksize
374 (p8) add dst0=saved_in0,blocksize
375 (p8) add src0=saved_in1,blocksize
376 (p8) sub in2=saved_in2,blocksize
377 (p8) br.dpnt .4k_block
380 /* we have up to 15 byte to copy in the tail.
381 * part of work is already done in the jump table code
382 * we are at the following state.
385 * xxxxxx xx <----- r21 has xxxxxxxx already
386 * -------- -------- --------
393 * -------- -------- --------
398 EX(.ex_handler, (p6) st8 [dst1]=r21,8) // more than 8 byte to copy
399 (p6) add curlen=-8,curlen // update length
404 mov in2=curlen // remaining length
405 mov dst0=dst1 // dest pointer
406 add src0=src1,r30 // forward by src alignment
409 // 7 byte or smaller.
412 cmp.le p10,p11 = 2,in2
413 cmp.le p12,p13 = 3,in2
414 cmp.le p14,p15 = 4,in2
415 add src1=1,src0 // second src pointer
416 add dst1=1,dst0 // second dest pointer
419 EX(.ex_handler_short, (p8) ld1 t1=[src0],2)
420 EK(.ex_handler_short, (p10) ld1 t2=[src1],2)
421 (p9) br.ret.dpnt rp // 0 byte copy
424 EX(.ex_handler_short, (p8) st1 [dst0]=t1,2)
425 EK(.ex_handler_short, (p10) st1 [dst1]=t2,2)
426 (p11) br.ret.dpnt rp // 1 byte copy
428 EX(.ex_handler_short, (p12) ld1 t3=[src0],2)
429 EK(.ex_handler_short, (p14) ld1 t4=[src1],2)
430 (p13) br.ret.dpnt rp // 2 byte copy
435 cmp.le p10,p11 = 7,in2
437 EX(.ex_handler_short, (p12) st1 [dst0]=t3,2)
438 EK(.ex_handler_short, (p14) st1 [dst1]=t4,2)
439 (p15) br.ret.dpnt rp // 3 byte copy
442 EX(.ex_handler_short, (p6) ld1 t5=[src0],2)
443 EK(.ex_handler_short, (p8) ld1 t6=[src1],2)
444 (p7) br.ret.dpnt rp // 4 byte copy
447 EX(.ex_handler_short, (p6) st1 [dst0]=t5,2)
448 EK(.ex_handler_short, (p8) st1 [dst1]=t6,2)
449 (p9) br.ret.dptk rp // 5 byte copy
451 EX(.ex_handler_short, (p10) ld1 t7=[src0],2)
452 (p11) br.ret.dptk rp // 6 byte copy
455 EX(.ex_handler_short, (p10) st1 [dst0]=t7,2)
456 br.ret.dptk rp // done all cases
459 /* Align dest to nearest 8-byte boundary. We know we have at
460 * least 7 bytes to copy, enough to crawl to 8-byte boundary.
461 * Actual number of byte to crawl depend on the dest alignment.
462 * 7 byte or less is taken care at .memcpy_short
464 * src0 - source even index
465 * src1 - source odd index
466 * dst0 - dest even index
467 * dst1 - dest odd index
468 * r30 - distance to 8-byte boundary
472 add src1=1,in1 // source odd index
473 cmp.le p7,p0 = 2,r30 // for .align_dest
474 cmp.le p8,p0 = 3,r30 // for .align_dest
475 EX(.ex_handler_short, (p6) ld1 t1=[src0],2)
476 cmp.le p9,p0 = 4,r30 // for .align_dest
477 cmp.le p10,p0 = 5,r30
479 EX(.ex_handler_short, (p7) ld1 t2=[src1],2)
480 EK(.ex_handler_short, (p8) ld1 t3=[src0],2)
481 cmp.le p11,p0 = 6,r30
482 EX(.ex_handler_short, (p6) st1 [dst0] = t1,2)
483 cmp.le p12,p0 = 7,r30
485 EX(.ex_handler_short, (p9) ld1 t4=[src1],2)
486 EK(.ex_handler_short, (p10) ld1 t5=[src0],2)
487 EX(.ex_handler_short, (p7) st1 [dst1] = t2,2)
488 EK(.ex_handler_short, (p8) st1 [dst0] = t3,2)
490 EX(.ex_handler_short, (p11) ld1 t6=[src1],2)
491 EK(.ex_handler_short, (p12) ld1 t7=[src0],2)
493 EX(.ex_handler_short, (p9) st1 [dst1] = t4,2)
494 EK(.ex_handler_short, (p10) st1 [dst0] = t5,2)
497 EX(.ex_handler_short, (p11) st1 [dst1] = t6,2)
498 EK(.ex_handler_short, (p12) st1 [dst0] = t7)
499 add dst0=in0,r30 // setup arguments
501 (p6) br.cond.dptk .aligned_src
502 (p7) br.cond.dpnt .unaligned_src
505 /* main loop body in jump table format */
506 #define COPYU(shift) \
508 EX(.ex_handler, (p16) ld8 r32=[src0],8); /* 1 */ \
509 EK(.ex_handler, (p16) ld8 r36=[src1],8); \
510 (p17) shrp r35=r33,r34,shift;; /* 1 */ \
511 EX(.ex_handler, (p6) ld8 r22=[src1]); /* common, prime for tail section */ \
513 (p16) shrp r38=r36,r37,shift; \
514 EX(.ex_handler, (p17) st8 [dst0]=r35,8); /* 1 */ \
515 EK(.ex_handler, (p17) st8 [dst1]=r39,8); \
516 br.ctop.dptk.few 1b;; \
517 (p7) add src1=-8,src1; /* back out for <8 byte case */ \
518 shrp r21=r22,r38,shift; /* speculative work */ \
519 br.sptk.few .unaligned_src_tail /* branch out of jump table */ \
523 COPYU(8) // unaligned cases
539 * Due to lack of local tag support in gcc 2.x assembler, it is not clear which
540 * instruction failed in the bundle. The exception algorithm is that we
541 * first figure out the faulting address, then detect if there is any
542 * progress made on the copy, if so, redo the copy from last known copied
543 * location up to the faulting address (exclusive). In the copy_from_user
544 * case, remaining byte in kernel buffer will be zeroed.
546 * Take copy_from_user as an example, in the code there are multiple loads
547 * in a bundle and those multiple loads could span over two pages, the
548 * faulting address is calculated as page_round_down(max(src0, src1)).
549 * This is based on knowledge that if we can access one byte in a page, we
550 * can access any byte in that page.
552 * predicate used in the exception handler:
554 * p10-p11: src faulting addr calculation
555 * p12-p13: dst faulting addr calculation
564 #define memset_arg0 r32
565 #define memset_arg2 r33
567 #define saved_retval loc0
568 #define saved_rtlink loc1
569 #define saved_pfs_stack loc2
582 cmp.gtu p10,p11=src_pre_mem,saved_in1
583 cmp.gtu p12,p13=dst_pre_mem,saved_in0
585 (p10) add src0=8,saved_in1
586 (p11) mov src0=saved_in1
587 (p12) add dst0=8,saved_in0
588 (p13) mov dst0=saved_in0
591 // in line_copy block, the preload addresses should always ahead
592 // of the other two src/dst pointers. Furthermore, src1/dst1 should
593 // always ahead of src0/dst0.
597 mov pr=saved_pr,-1 // first restore pr, lc, and pfs
601 .ex_handler_short: // fault occurred in these sections didn't change pr, lc, pfs
602 cmp.ltu p6,p7=saved_in0, saved_in1 // get the copy direction
603 cmp.ltu p10,p11=src0,src1
604 cmp.ltu p12,p13=dst0,dst1
605 fcmp.eq p8,p0=f6,f0 // is it memcpy?
608 (p11) mov src1 = src0 // pick the larger of the two
609 (p13) mov dst0 = dst1 // make dst0 the smaller one
610 (p13) mov dst1 = tmp // and dst1 the larger one
612 (p6) dep F = r0,dst1,0,PAGE_SHIFT // usr dst round down to page boundary
613 (p7) dep F = r0,src1,0,PAGE_SHIFT // usr src round down to page boundary
615 (p6) cmp.le p14,p0=dst0,saved_in0 // no progress has been made on store
616 (p7) cmp.le p14,p0=src0,saved_in1 // no progress has been made on load
618 (p8) ld1 tmp=[src1] // force an oops for memcpy call
619 (p8) st1 [dst1]=r0 // force an oops for memcpy call
620 (p14) br.ret.sptk.many rp
623 * The remaining byte to copy is calculated as:
625 * A = (faulting_addr - orig_src) -> len to faulting ld address
627 * (faulting_addr - orig_dst) -> len to faulting st address
628 * B = (cur_dst - orig_dst) -> len copied so far
629 * C = A - B -> len need to be copied
630 * D = orig_len - A -> len need to be zeroed
632 (p6) sub A = F, saved_in0
633 (p7) sub A = F, saved_in1
636 alloc saved_pfs_stack=ar.pfs,3,3,3,0
637 sub B = dst0, saved_in0 // how many byte copied so far
642 cmp.gt p8,p0=C,r0 // more than 1 byte?
643 add memset_arg0=saved_in0, A
644 (p6) mov memset_arg2=0 // copy_to_user should not call memset
645 (p7) mov memset_arg2=D // copy_from_user need to have kbuf zeroed
648 mov saved_rtlink = b0
650 add out0=saved_in0, B
651 add out1=saved_in1, B
653 (p8) br.call.sptk.few b0=__copy_user // recursive call
656 add saved_retval=saved_retval,r8 // above might return non-zero value
657 cmp.gt p8,p0=memset_arg2,r0 // more than 1 byte?
658 mov out0=memset_arg0 // *s
660 mov out2=memset_arg2 // n
661 (p8) br.call.sptk.few b0=memset
664 mov retval=saved_retval
665 mov ar.pfs=saved_pfs_stack
669 /* end of McKinley specific optimization */