arch/ia64/lib/memcpy_mck.S

   1 /*
   2  * Itanium 2-optimized version of memcpy and copy_user function
   3  *
   4  * Inputs:
   5  *      in0:    destination address
   6  *      in1:    source address
   7  *      in2:    number of bytes to copy
   8  * Output:
   9  *      0 if success, or number of byte NOT copied if error occurred.
  10  *
  11  * Copyright (C) 2002 Intel Corp.
  12  * Copyright (C) 2002 Ken Chen <kenneth.w.chen@intel.com>
  13  */
  14 #include <linux/config.h>
  15 #include <asm/asmmacro.h>
  16 #include <asm/page.h>
  17
  18 #define EK(y...) EX(y)
  19
  20 /* McKinley specific optimization */
  21
  22 #define retval          r8
  23 #define saved_pfs       r31
  24 #define saved_lc        r10
  25 #define saved_pr        r11
  26 #define saved_in0       r14
  27 #define saved_in1       r15
  28 #define saved_in2       r16
  29
  30 #define src0            r2
  31 #define src1            r3
  32 #define dst0            r17
  33 #define dst1            r18
  34 #define cnt             r9
  35
  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
  42 #define t1              r23
  43 #define t2              r24
  44 #define t3              r25
  45 #define t4              r26
  46 #define t5              t1      // alias!
  47 #define t6              t2      // alias!
  48 #define t7              t3      // alias!
  49 #define n8              r27
  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!
  55 #define t13             r28
  56 #define t15             r29
  57 #define tmp             r30
  58
  59 /* defines for long_copy block */
  60 #define A       0
  61 #define B       (PREFETCH_DIST)
  62 #define C       (B + PREFETCH_DIST)
  63 #define D       (C + 1)
  64 #define N       (D + 1)
  65 #define Nrot    ((N + 7) & ~7)
  66
  67 /* alias */
  68 #define in0             r32
  69 #define in1             r33
  70 #define in2             r34
  71
  72 GLOBAL_ENTRY(memcpy)
  73         and     r28=0x7,in0
  74         and     r29=0x7,in1
  75         mov     f6=f0
  76         br.cond.sptk .common_code
  77         ;;
  78 END(memcpy)
  79 GLOBAL_ENTRY(__copy_user)
  80         .prologue
  81 // check dest alignment
  82         and     r28=0x7,in0
  83         and     r29=0x7,in1
  84         mov     f6=f1
  85         mov     saved_in0=in0   // save dest pointer
  86         mov     saved_in1=in1   // save src pointer
  87         mov     saved_in2=in2   // save len
  88         ;;
  89 .common_code:
  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
  93         add     src0=0,in1
  94         sub     r30=8,r28       // for .align_dest
  95         mov     retval=r0       // initialize return value
  96         ;;
  97         add     dst0=0,in0
  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
 103         ;;
 104
 105 // both dest and src are aligned on 8-byte boundary
 106 .aligned_src:
 107         .save ar.pfs, saved_pfs
 108         alloc   saved_pfs=ar.pfs,3,Nrot-3,0,Nrot
 109         .save pr, saved_pr
 110         mov     saved_pr=pr
 111
 112         shr.u   cnt=in2,7       // this much cache line
 113         ;;
 114         cmp.lt  p6,p0=2*PREFETCH_DIST,cnt
 115         cmp.lt  p7,p8=1,cnt
 116         .save ar.lc, saved_lc
 117         mov     saved_lc=ar.lc
 118         .body
 119         add     cnt=-1,cnt
 120         add     src_pre_mem=0,in1       // prefetch src pointer
 121         add     dst_pre_mem=0,in0       // prefetch dest pointer
 122         ;;
 123 (p7)    mov     ar.lc=cnt       // prefetch count
 124 (p8)    mov     ar.lc=r0
 125 (p6)    br.cond.dpnt .long_copy
 126         ;;
 127
 128 .prefetch:
 129         lfetch.fault      [src_pre_mem], 128
 130         lfetch.fault.excl [dst_pre_mem], 128
 131         br.cloop.dptk.few .prefetch
 132         ;;
 133
 134 .medium_copy:
 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
 138         ;;
 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
 142         cmp.le  p6,p0=8,tmp
 143         ;;
 144         cmp.le  p7,p0=16,tmp
 145         mov     ar.lc=cnt       // loop setup
 146         cmp.eq  p16,p17 = r0,r0
 147         mov     ar.ec=2
 148 (p10)   br.dpnt.few .aligned_src_tail
 149         ;;
 150         TEXT_ALIGN(32)
 151 1:
 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)
 156         ;;
 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)
 161         br.ctop.dptk.few 1b
 162         ;;
 163
 164 .aligned_src_tail:
 165 EX(.ex_handler, (p6)    ld8     t1=[src0])
 166         mov     ar.lc=saved_lc
 167         mov     ar.pfs=saved_pfs
 168 EX(.ex_hndlr_s, (p7)    ld8     t2=[src1],8)
 169         cmp.le  p8,p0=24,tmp
 170         and     r21=-8,tmp
 171         ;;
 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
 178         ;;
 179 EX(.ex_handler, (p8)    st8     [dst1]=t3)      // store byte 3
 180         mov     pr=saved_pr,-1
 181         br.dptk.many .memcpy_short
 182         ;;
 183
 184 /* code taken from copy_page_mck */
 185 .long_copy:
 186         .rotr v[2*PREFETCH_DIST]
 187         .rotp p[N]
 188
 189         mov src_pre_mem = src0
 190         mov pr.rot = 0x10000
 191         mov ar.ec = 1                           // special unrolled loop
 192
 193         mov dst_pre_mem = dst0
 194
 195         add src_pre_l2 = 8*8, src0
 196         add dst_pre_l2 = 8*8, dst0
 197         ;;
 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
 204         ;;
 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:
 208 .prefetch_loop:
 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
 212         ;;
 213         cmp.eq p16, p0 = r0, r0                 // reset p16 to 1
 214         mov ar.lc = cnt
 215         mov ar.ec = N                           // # of stages in pipeline
 216         ;;
 217 .line_copy:
 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
 222         ;;
 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
 227         ;;
 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)
 232         ;;
 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)
 237         ;;
 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)
 242         ;;
 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)
 247         ;;
 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)
 252         ;;
 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
 258         ;;
 259
 260         add dst0=-8,dst0
 261         add src0=-8,src0
 262         mov in2=tmp
 263         .restore sp
 264         br.sptk.many .medium_copy
 265         ;;
 266
 267 #define BLOCK_SIZE      128*32
 268 #define blocksize       r23
 269 #define curlen          r24
 270
 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.
 273 .unaligned_src:
 274         .prologue
 275         .save ar.pfs, saved_pfs
 276         alloc   saved_pfs=ar.pfs,3,5,0,8
 277         .save ar.lc, saved_lc
 278         mov     saved_lc=ar.lc
 279         .save pr, saved_pr
 280         mov     saved_pr=pr
 281         .body
 282 .4k_block:
 283         mov     saved_in0=dst0  // need to save all input arguments
 284         mov     saved_in2=in2
 285         mov     blocksize=BLOCK_SIZE
 286         ;;
 287         cmp.lt  p6,p7=blocksize,in2
 288         mov     saved_in1=src0
 289         ;;
 290 (p6)    mov     in2=blocksize
 291         ;;
 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
 296         ;;
 297         cmp.lt  p7,p8=1,r21
 298         add     cnt=-1,r21
 299         ;;
 300
 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
 304 (p7)    mov     ar.lc = cnt
 305 (p8)    mov     ar.lc = r0
 306         ;;
 307         TEXT_ALIGN(32)
 308 1:      lfetch.fault      [src_pre_mem], 128
 309         lfetch.fault.excl [dst_pre_mem], 128
 310         br.cloop.dptk.few 1b
 311         ;;
 312
 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
 318         ;;
 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
 322         ;;
 323
 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.
 330 #define LOOP_SIZE 6
 331 1:
 332         mov     r29=ip          // jmp_table thread
 333         mov     ar.lc=cnt
 334         ;;
 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
 338         ;;
 339         add     r29=r29,r28             // jmp_table thread
 340         cmp.eq  p16,p17=r0,r0
 341         ;;
 342         mov     b6=r29                  // jmp_table thread
 343         ;;
 344         br.cond.sptk.few b6
 345
 346 // for 8-15 byte case
 347 // We will skip the loop, but need to replicate the side effect
 348 // that the loop produces.
 349 .noloop:
 350 EX(.ex_handler, (p6)    ld8     r37=[src1],8)
 351         add     src0=8,src0
 352 (p6)    shl     r25=r30,3
 353         ;;
 354 EX(.ex_handler, (p6)    ld8     r27=[src1])
 355 (p6)    shr.u   r28=r37,r25
 356 (p6)    sub     r26=64,r25
 357         ;;
 358 (p6)    shl     r27=r27,r26
 359         ;;
 360 (p6)    or      r21=r28,r27
 361
 362 .unaligned_src_tail:
 363 /* check if we have more than blocksize to copy, if so go back */
 364         cmp.gt  p8,p0=saved_in2,blocksize
 365         ;;
 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
 370         ;;
 371
 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.
 375  * src side:
 376  *
 377  *   xxxxxx xx                   <----- r21 has xxxxxxxx already
 378  * -------- -------- --------
 379  * 0        8        16
 380  *          ^
 381  *          |
 382  *          src1
 383  *
 384  * dst
 385  * -------- -------- --------
 386  * ^
 387  * |
 388  * dst1
 389  */
 390 EX(.ex_handler, (p6)    st8     [dst1]=r21,8)   // more than 8 byte to copy
 391 (p6)    add     curlen=-8,curlen        // update length
 392         mov     ar.pfs=saved_pfs
 393         ;;
 394         mov     ar.lc=saved_lc
 395         mov     pr=saved_pr,-1
 396         mov     in2=curlen      // remaining length
 397         mov     dst0=dst1       // dest pointer
 398         add     src0=src1,r30   // forward by src alignment
 399         ;;
 400
 401 // 7 byte or smaller.
 402 .memcpy_short:
 403         cmp.le  p8,p9   = 1,in2
 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
 409         ;;
 410
 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
 414         ;;
 415
 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
 419
 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
 423         ;;
 424
 425         cmp.le  p6,p7   = 5,in2
 426         cmp.le  p8,p9   = 6,in2
 427         cmp.le  p10,p11 = 7,in2
 428
 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
 432         ;;
 433
 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
 437         ;;
 438
 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
 442
 443 EX(.ex_handler_short, (p10)     ld1     t7=[src0],2)
 444 (p11)   br.ret.dptk rp          // 6 byte copy
 445         ;;
 446
 447 EX(.ex_handler_short, (p10)     st1     [dst0]=t7,2)
 448         br.ret.dptk rp          // done all cases
 449
 450
 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
 455
 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
 461  */
 462
 463 .align_dest:
 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
 470         ;;
 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
 476         ;;
 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)
 481         ;;
 482 EX(.ex_handler_short, (p11)     ld1     t6=[src1],2)
 483 EK(.ex_handler_short, (p12)     ld1     t7=[src0],2)
 484         cmp.eq  p6,p7=r28,r29
 485 EX(.ex_handler_short, (p9)      st1     [dst1] = t4,2)
 486 EK(.ex_handler_short, (p10)     st1     [dst0] = t5,2)
 487         sub     in2=in2,r30
 488         ;;
 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
 492         add     src0=in1,r30
 493 (p6)    br.cond.dptk .aligned_src
 494 (p7)    br.cond.dpnt .unaligned_src
 495         ;;
 496
 497 /* main loop body in jump table format */
 498 #define COPYU(shift)                                                                    \
 499 1:                                                                                      \
 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 */    \
 504                  nop.m  0;                                                              \
 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 */         \
 512                  ;;
 513         TEXT_ALIGN(32)
 514 .jump_table:
 515         COPYU(8)        // unaligned cases
 516 .jmp1:
 517         COPYU(16)
 518         COPYU(24)
 519         COPYU(32)
 520         COPYU(40)
 521         COPYU(48)
 522         COPYU(56)
 523
 524 #undef A
 525 #undef B
 526 #undef C
 527 #undef D
 528
 529 /*
 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.
 536  *
 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.
 542  *
 543  * predicate used in the exception handler:
 544  * p6-p7: direction
 545  * p10-p11: src faulting addr calculation
 546  * p12-p13: dst faulting addr calculation
 547  */
 548
 549 #define A       r19
 550 #define B       r20
 551 #define C       r21
 552 #define D       r22
 553 #define F       r28
 554
 555 #define memset_arg0     r32
 556 #define memset_arg2     r33
 557
 558 #define saved_retval    loc0
 559 #define saved_rtlink    loc1
 560 #define saved_pfs_stack loc2
 561
 562 .ex_hndlr_s:
 563         add     src0=8,src0
 564         br.sptk .ex_handler
 565         ;;
 566 .ex_hndlr_d:
 567         add     dst0=8,dst0
 568         br.sptk .ex_handler
 569         ;;
 570 .ex_hndlr_lcpy_1:
 571         mov     src1=src_pre_mem
 572         mov     dst1=dst_pre_mem
 573         cmp.gtu p10,p11=src_pre_mem,saved_in1
 574         cmp.gtu p12,p13=dst_pre_mem,saved_in0
 575         ;;
 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
 580         br.sptk .ex_handler
 581 .ex_handler_lcpy:
 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.
 585         mov     src1=src_pre_mem
 586         mov     dst1=dst_pre_mem
 587 .ex_handler:
 588         mov     pr=saved_pr,-1          // first restore pr, lc, and pfs
 589         mov     ar.lc=saved_lc
 590         mov     ar.pfs=saved_pfs
 591         ;;
 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?
 597         mov     tmp = dst0
 598         ;;
 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
 602         ;;
 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
 605         ;;
 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
 608         mov     retval=saved_in2
 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
 612
 613 /*
 614  * The remaining byte to copy is calculated as:
 615  *
 616  * A =  (faulting_addr - orig_src)      -> len to faulting ld address
 617  *      or
 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
 622  */
 623 (p6)    sub     A = F, saved_in0
 624 (p7)    sub     A = F, saved_in1
 625         clrrrb
 626         ;;
 627         alloc   saved_pfs_stack=ar.pfs,3,3,3,0
 628         cmp.lt  p8,p0=A,r0
 629         sub     B = dst0, saved_in0     // how many byte copied so far
 630         ;;
 631 (p8)    mov     A = 0;                  // A shouldn't be negative, cap it
 632         ;;
 633         sub     C = A, B
 634         sub     D = saved_in2, A
 635         ;;
 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
 640         mov     r8=0
 641         mov     saved_retval = D
 642         mov     saved_rtlink = b0
 643
 644         add     out0=saved_in0, B
 645         add     out1=saved_in1, B
 646         mov     out2=C
 647 (p8)    br.call.sptk.few b0=__copy_user // recursive call
 648         ;;
 649
 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
 653         mov     out1=r0                 // c
 654         mov     out2=memset_arg2        // n
 655 (p8)    br.call.sptk.few b0=memset
 656         ;;
 657
 658         mov     retval=saved_retval
 659         mov     ar.pfs=saved_pfs_stack
 660         mov     b0=saved_rtlink
 661         br.ret.sptk.many rp
 662
 663 /* end of McKinley specific optimization */
 664 END(__copy_user)