amd64/isel.c

   1 #include "all.h"
   2 #include <limits.h>
   3
   4 /* For x86_64, do the following:
   5  *
   6  * - check that constants are used only in
   7  *   places allowed
   8  * - ensure immediates always fit in 32b
   9  * - expose machine register contraints
  10  *   on instructions like division.
  11  * - implement fast locals (the streak of
  12  *   constant allocX in the first basic block)
  13  * - recognize complex addressing modes
  14  *
  15  * Invariant: the use counts that are used
  16  *            in sel() must be sound.  This
  17  *            is not so trivial, maybe the
  18  *            dce should be moved out...
  19  */
  20
  21 static int amatch(Addr *, Num *, Ref, Fn *);
  22
  23 static int
  24 noimm(Ref r, Fn *fn)
  25 {
  26         int64_t val;
  27
  28         if (rtype(r) != RCon)
  29                 return 0;
  30         switch (fn->con[r.val].type) {
  31         case CAddr:
  32                 /* we only support the 'small'
  33                  * code model of the ABI, this
  34                  * means that we can always
  35                  * address data with 32bits
  36                  */
  37                 return 0;
  38         case CBits:
  39                 val = fn->con[r.val].bits.i;
  40                 return (val < INT32_MIN || val > INT32_MAX);
  41         default:
  42                 die("invalid constant");
  43         }
  44 }
  45
  46 static int
  47 rslot(Ref r, Fn *fn)
  48 {
  49         if (rtype(r) != RTmp)
  50                 return -1;
  51         return fn->tmp[r.val].slot;
  52 }
  53
  54 static int
  55 hascon(Ref r, Con **pc, Fn *fn)
  56 {
  57         switch (rtype(r)) {
  58         case RCon:
  59                 *pc = &fn->con[r.val];
  60                 return 1;
  61         case RMem:
  62                 *pc = &fn->mem[r.val].offset;
  63                 return 1;
  64         default:
  65                 return 0;
  66         }
  67 }
  68
  69 static void
  70 fixarg(Ref *r, int k, Ins *i, Fn *fn)
  71 {
  72         char buf[32];
  73         Addr a, *m;
  74         Con cc, *c;
  75         Ref r0, r1, r2, r3;
  76         int s, n, op;
  77
  78         r1 = r0 = *r;
  79         s = rslot(r0, fn);
  80         op = i ? i->op : Ocopy;
  81         if (KBASE(k) == 1 && rtype(r0) == RCon) {
  82                 /* load floating points from memory
  83                  * slots, they can't be used as
  84                  * immediates
  85                  */
  86                 r1 = MEM(fn->nmem);
  87                 vgrow(&fn->mem, ++fn->nmem);
  88                 memset(&a, 0, sizeof a);
  89                 a.offset.type = CAddr;
  90                 n = stashbits(&fn->con[r0.val].bits, KWIDE(k) ? 8 : 4);
  91                 /* quote the name so that we do not
  92                  * add symbol prefixes on the apple
  93                  * target variant
  94                  */
  95                 sprintf(buf, "\"%sfp%d\"", T.asloc, n);
  96                 a.offset.sym.id = intern(buf);
  97                 fn->mem[fn->nmem-1] = a;
  98         }
  99         else if (op != Ocopy && k == Kl && noimm(r0, fn)) {
 100                 /* load constants that do not fit in
 101                  * a 32bit signed integer into a
 102                  * long temporary
 103                  */
 104                 r1 = newtmp("isel", Kl, fn);
 105                 emit(Ocopy, Kl, r1, r0, R);
 106         }
 107         else if (s != -1) {
 108                 /* load fast locals' addresses into
 109                  * temporaries right before the
 110                  * instruction
 111                  */
 112                 r1 = newtmp("isel", Kl, fn);
 113                 emit(Oaddr, Kl, r1, SLOT(s), R);
 114         }
 115         else if (T.apple && hascon(r0, &c, fn)
 116         && c->type == CAddr && c->sym.type == SThr) {
 117                 r1 = newtmp("isel", Kl, fn);
 118                 if (c->bits.i) {
 119                         r2 = newtmp("isel", Kl, fn);
 120                         cc = (Con){.type = CBits};
 121                         cc.bits.i = c->bits.i;
 122                         r3 = newcon(&cc, fn);
 123                         emit(Oadd, Kl, r1, r2, r3);
 124                 } else
 125                         r2 = r1;
 126                 emit(Ocopy, Kl, r2, TMP(RAX), R);
 127                 r2 = newtmp("isel", Kl, fn);
 128                 r3 = newtmp("isel", Kl, fn);
 129                 emit(Ocall, 0, R, r3, CALL(17));
 130                 emit(Ocopy, Kl, TMP(RDI), r2, R);
 131                 emit(Oload, Kl, r3, r2, R);
 132                 cc = *c;
 133                 cc.bits.i = 0;
 134                 r3 = newcon(&cc, fn);
 135                 emit(Oload, Kl, r2, r3, R);
 136                 if (rtype(r0) == RMem) {
 137                         m = &fn->mem[r0.val];
 138                         m->offset.type = CUndef;
 139                         m->base = r1;
 140                         r1 = r0;
 141                 }
 142         }
 143         else if (!(isstore(op) && r == &i->arg[1])
 144         && !isload(op) && op != Ocall && rtype(r0) == RCon
 145         && fn->con[r0.val].type == CAddr) {
 146                 /* apple as does not support 32-bit
 147                  * absolute addressing, use a rip-
 148                  * relative leaq instead
 149                  */
 150                 r1 = newtmp("isel", Kl, fn);
 151                 emit(Oaddr, Kl, r1, r0, R);
 152         }
 153         else if (rtype(r0) == RMem) {
 154                 /* eliminate memory operands of
 155                  * the form $foo(%rip, ...)
 156                  */
 157                 m = &fn->mem[r0.val];
 158                 if (req(m->base, R))
 159                 if (m->offset.type == CAddr) {
 160                         r0 = newtmp("isel", Kl, fn);
 161                         emit(Oaddr, Kl, r0, newcon(&m->offset, fn), R);
 162                         m->offset.type = CUndef;
 163                         m->base = r0;
 164                 }
 165         }
 166         *r = r1;
 167 }
 168
 169 static void
 170 seladdr(Ref *r, Num *tn, Fn *fn)
 171 {
 172         Addr a;
 173         Ref r0;
 174
 175         r0 = *r;
 176         if (rtype(r0) == RTmp) {
 177                 memset(&a, 0, sizeof a);
 178                 if (!amatch(&a, tn, r0, fn))
 179                         return;
 180                 if (!req(a.base, R))
 181                 if (a.offset.type == CAddr) {
 182                         /* apple as does not support
 183                          * $foo(%r0, %r1, M); try to
 184                          * rewrite it or bail out if
 185                          * impossible
 186                          */
 187                         if (!req(a.index, R) || rtype(a.base) != RTmp)
 188                                 return;
 189                         else {
 190                                 a.index = a.base;
 191                                 a.scale = 1;
 192                                 a.base = R;
 193                         }
 194                 }
 195                 chuse(r0, -1, fn);
 196                 vgrow(&fn->mem, ++fn->nmem);
 197                 fn->mem[fn->nmem-1] = a;
 198                 chuse(a.base, +1, fn);
 199                 chuse(a.index, +1, fn);
 200                 *r = MEM(fn->nmem-1);
 201         }
 202 }
 203
 204 static int
 205 cmpswap(Ref arg[2], int op)
 206 {
 207         switch (op) {
 208         case NCmpI+Cflt:
 209         case NCmpI+Cfle:
 210                 return 1;
 211         case NCmpI+Cfgt:
 212         case NCmpI+Cfge:
 213                 return 0;
 214         }
 215         return rtype(arg[0]) == RCon;
 216 }
 217
 218 static void
 219 selcmp(Ref arg[2], int k, int swap, Fn *fn)
 220 {
 221         Ref r;
 222         Ins *icmp;
 223
 224         if (swap) {
 225                 r = arg[1];
 226                 arg[1] = arg[0];
 227                 arg[0] = r;
 228         }
 229         emit(Oxcmp, k, R, arg[1], arg[0]);
 230         icmp = curi;
 231         if (rtype(arg[0]) == RCon) {
 232                 assert(k != Kw);
 233                 icmp->arg[1] = newtmp("isel", k, fn);
 234                 emit(Ocopy, k, icmp->arg[1], arg[0], R);
 235                 fixarg(&curi->arg[0], k, curi, fn);
 236         }
 237         fixarg(&icmp->arg[0], k, icmp, fn);
 238         fixarg(&icmp->arg[1], k, icmp, fn);
 239 }
 240
 241 static void
 242 sel(Ins i, Num *tn, Fn *fn)
 243 {
 244         Ref r0, r1, tmp[7];
 245         int x, j, k, kc, sh, swap;
 246         Ins *i0, *i1;
 247
 248         if (rtype(i.to) == RTmp)
 249         if (!isreg(i.to) && !isreg(i.arg[0]) && !isreg(i.arg[1]))
 250         if (fn->tmp[i.to.val].nuse == 0) {
 251                 chuse(i.arg[0], -1, fn);
 252                 chuse(i.arg[1], -1, fn);
 253                 return;
 254         }
 255         i0 = curi;
 256         k = i.cls;
 257         switch (i.op) {
 258         case Odiv:
 259         case Orem:
 260         case Oudiv:
 261         case Ourem:
 262                 if (KBASE(k) == 1)
 263                         goto Emit;
 264                 if (i.op == Odiv || i.op == Oudiv)
 265                         r0 = TMP(RAX), r1 = TMP(RDX);
 266                 else
 267                         r0 = TMP(RDX), r1 = TMP(RAX);
 268                 emit(Ocopy, k, i.to, r0, R);
 269                 emit(Ocopy, k, R, r1, R);
 270                 if (rtype(i.arg[1]) == RCon) {
 271                         /* immediates not allowed for
 272                          * divisions in x86
 273                          */
 274                         r0 = newtmp("isel", k, fn);
 275                 } else
 276                         r0 = i.arg[1];
 277                 if (fn->tmp[r0.val].slot != -1)
 278                         err("unlikely argument %%%s in %s",
 279                                 fn->tmp[r0.val].name, optab[i.op].name);
 280                 if (i.op == Odiv || i.op == Orem) {
 281                         emit(Oxidiv, k, R, r0, R);
 282                         emit(Osign, k, TMP(RDX), TMP(RAX), R);
 283                 } else {
 284                         emit(Oxdiv, k, R, r0, R);
 285                         emit(Ocopy, k, TMP(RDX), CON_Z, R);
 286                 }
 287                 emit(Ocopy, k, TMP(RAX), i.arg[0], R);
 288                 fixarg(&curi->arg[0], k, curi, fn);
 289                 if (rtype(i.arg[1]) == RCon)
 290                         emit(Ocopy, k, r0, i.arg[1], R);
 291                 break;
 292         case Osar:
 293         case Oshr:
 294         case Oshl:
 295                 r0 = i.arg[1];
 296                 if (rtype(r0) == RCon)
 297                         goto Emit;
 298                 if (fn->tmp[r0.val].slot != -1)
 299                         err("unlikely argument %%%s in %s",
 300                                 fn->tmp[r0.val].name, optab[i.op].name);
 301                 i.arg[1] = TMP(RCX);
 302                 emit(Ocopy, Kw, R, TMP(RCX), R);
 303                 emiti(i);
 304                 i1 = curi;
 305                 emit(Ocopy, Kw, TMP(RCX), r0, R);
 306                 fixarg(&i1->arg[0], argcls(&i, 0), i1, fn);
 307                 break;
 308         case Ouwtof:
 309                 r0 = newtmp("utof", Kl, fn);
 310                 emit(Osltof, k, i.to, r0, R);
 311                 emit(Oextuw, Kl, r0, i.arg[0], R);
 312                 fixarg(&curi->arg[0], k, curi, fn);
 313                 break;
 314         case Oultof:
 315                 /* %mask =l and %arg.0, 1
 316                  * %isbig =l shr %arg.0, 63
 317                  * %divided =l shr %arg.0, %isbig
 318                  * %or =l or %mask, %divided
 319                  * %float =d sltof %or
 320                  * %cast =l cast %float
 321                  * %addend =l shl %isbig, 52
 322                  * %sum =l add %cast, %addend
 323                  * %result =d cast %sum
 324                  */
 325                 r0 = newtmp("utof", k, fn);
 326                 if (k == Ks)
 327                         kc = Kw, sh = 23;
 328                 else
 329                         kc = Kl, sh = 52;
 330                 for (j=0; j<4; j++)
 331                         tmp[j] = newtmp("utof", Kl, fn);
 332                 for (; j<7; j++)
 333                         tmp[j] = newtmp("utof", kc, fn);
 334                 emit(Ocast, k, i.to, tmp[6], R);
 335                 emit(Oadd, kc, tmp[6], tmp[4], tmp[5]);
 336                 emit(Oshl, kc, tmp[5], tmp[1], getcon(sh, fn));
 337                 emit(Ocast, kc, tmp[4], r0, R);
 338                 emit(Osltof, k, r0, tmp[3], R);
 339                 emit(Oor, Kl, tmp[3], tmp[0], tmp[2]);
 340                 emit(Oshr, Kl, tmp[2], i.arg[0], tmp[1]);
 341                 sel(*curi++, 0, fn);
 342                 emit(Oshr, Kl, tmp[1], i.arg[0], getcon(63, fn));
 343                 fixarg(&curi->arg[0], Kl, curi, fn);
 344                 emit(Oand, Kl, tmp[0], i.arg[0], getcon(1, fn));
 345                 fixarg(&curi->arg[0], Kl, curi, fn);
 346                 break;
 347         case Ostoui:
 348                 i.op = Ostosi;
 349                 kc = Ks;
 350                 tmp[4] = getcon(0xdf000000, fn);
 351                 goto Oftoui;
 352         case Odtoui:
 353                 i.op = Odtosi;
 354                 kc = Kd;
 355                 tmp[4] = getcon(0xc3e0000000000000, fn);
 356         Oftoui:
 357                 if (k == Kw) {
 358                         r0 = newtmp("ftou", Kl, fn);
 359                         emit(Ocopy, Kw, i.to, r0, R);
 360                         i.cls = Kl;
 361                         i.to = r0;
 362                         goto Emit;
 363                 }
 364                 /* %try0 =l {s,d}tosi %fp
 365                  * %mask =l sar %try0, 63
 366                  *
 367                  *    mask is all ones if the first
 368                  *    try was oob, all zeroes o.w.
 369                  *
 370                  * %fps ={s,d} sub %fp, (1<<63)
 371                  * %try1 =l {s,d}tosi %fps
 372                  *
 373                  * %tmp =l and %mask, %try1
 374                  * %res =l or %tmp, %try0
 375                  */
 376                 r0 = newtmp("ftou", kc, fn);
 377                 for (j=0; j<4; j++)
 378                         tmp[j] = newtmp("ftou", Kl, fn);
 379                 emit(Oor, Kl, i.to, tmp[0], tmp[3]);
 380                 emit(Oand, Kl, tmp[3], tmp[2], tmp[1]);
 381                 emit(i.op, Kl, tmp[2], r0, R);
 382                 emit(Oadd, kc, r0, tmp[4], i.arg[0]);
 383                 i1 = curi; /* fixarg() can change curi */
 384                 fixarg(&i1->arg[0], kc, i1, fn);
 385                 fixarg(&i1->arg[1], kc, i1, fn);
 386                 emit(Osar, Kl, tmp[1], tmp[0], getcon(63, fn));
 387                 emit(i.op, Kl, tmp[0], i.arg[0], R);
 388                 fixarg(&curi->arg[0], Kl, curi, fn);
 389                 break;
 390         case Onop:
 391                 break;
 392         case Ostored:
 393         case Ostores:
 394         case Ostorel:
 395         case Ostorew:
 396         case Ostoreh:
 397         case Ostoreb:
 398                 if (rtype(i.arg[0]) == RCon) {
 399                         if (i.op == Ostored)
 400                                 i.op = Ostorel;
 401                         if (i.op == Ostores)
 402                                 i.op = Ostorew;
 403                 }
 404                 seladdr(&i.arg[1], tn, fn);
 405                 goto Emit;
 406         case_Oload:
 407                 seladdr(&i.arg[0], tn, fn);
 408                 goto Emit;
 409         case Odbgloc:
 410         case Ocall:
 411         case Osalloc:
 412         case Ocopy:
 413         case Oadd:
 414         case Osub:
 415         case Oneg:
 416         case Omul:
 417         case Oand:
 418         case Oor:
 419         case Oxor:
 420         case Oxtest:
 421         case Ostosi:
 422         case Odtosi:
 423         case Oswtof:
 424         case Osltof:
 425         case Oexts:
 426         case Otruncd:
 427         case Ocast:
 428         case_OExt:
 429 Emit:
 430                 emiti(i);
 431                 i1 = curi; /* fixarg() can change curi */
 432                 fixarg(&i1->arg[0], argcls(&i, 0), i1, fn);
 433                 fixarg(&i1->arg[1], argcls(&i, 1), i1, fn);
 434                 break;
 435         case Oalloc4:
 436         case Oalloc8:
 437         case Oalloc16:
 438                 salloc(i.to, i.arg[0], fn);
 439                 break;
 440         default:
 441                 if (isext(i.op))
 442                         goto case_OExt;
 443                 if (isload(i.op))
 444                         goto case_Oload;
 445                 if (iscmp(i.op, &kc, &x)) {
 446                         switch (x) {
 447                         case NCmpI+Cfeq:
 448                                 /* zf is set when operands are
 449                                  * unordered, so we may have to
 450                                  * check pf
 451                                  */
 452                                 r0 = newtmp("isel", Kw, fn);
 453                                 r1 = newtmp("isel", Kw, fn);
 454                                 emit(Oand, Kw, i.to, r0, r1);
 455                                 emit(Oflagfo, k, r1, R, R);
 456                                 i.to = r0;
 457                                 break;
 458                         case NCmpI+Cfne:
 459                                 r0 = newtmp("isel", Kw, fn);
 460                                 r1 = newtmp("isel", Kw, fn);
 461                                 emit(Oor, Kw, i.to, r0, r1);
 462                                 emit(Oflagfuo, k, r1, R, R);
 463                                 i.to = r0;
 464                                 break;
 465                         }
 466                         swap = cmpswap(i.arg, x);
 467                         if (swap)
 468                                 x = cmpop(x);
 469                         emit(Oflag+x, k, i.to, R, R);
 470                         selcmp(i.arg, kc, swap, fn);
 471                         break;
 472                 }
 473                 die("unknown instruction %s", optab[i.op].name);
 474         }
 475
 476         while (i0>curi && --i0) {
 477                 assert(rslot(i0->arg[0], fn) == -1);
 478                 assert(rslot(i0->arg[1], fn) == -1);
 479         }
 480 }
 481
 482 static Ins *
 483 flagi(Ins *i0, Ins *i)
 484 {
 485         while (i>i0) {
 486                 i--;
 487                 if (amd64_op[i->op].zflag)
 488                         return i;
 489                 if (amd64_op[i->op].lflag)
 490                         continue;
 491                 return 0;
 492         }
 493         return 0;
 494 }
 495
 496 static void
 497 seljmp(Blk *b, Fn *fn)
 498 {
 499         Ref r;
 500         int c, k, swap;
 501         Ins *fi;
 502         Tmp *t;
 503
 504         if (b->jmp.type == Jret0
 505         || b->jmp.type == Jjmp
 506         || b->jmp.type == Jhlt)
 507                 return;
 508         assert(b->jmp.type == Jjnz);
 509         r = b->jmp.arg;
 510         t = &fn->tmp[r.val];
 511         b->jmp.arg = R;
 512         assert(rtype(r) == RTmp);
 513         if (b->s1 == b->s2) {
 514                 chuse(r, -1, fn);
 515                 b->jmp.type = Jjmp;
 516                 b->s2 = 0;
 517                 return;
 518         }
 519         fi = flagi(b->ins, &b->ins[b->nins]);
 520         if (!fi || !req(fi->to, r)) {
 521                 selcmp((Ref[2]){r, CON_Z}, Kw, 0, fn);
 522                 b->jmp.type = Jjf + Cine;
 523         }
 524         else if (iscmp(fi->op, &k, &c)
 525              && c != NCmpI+Cfeq /* see sel() */
 526              && c != NCmpI+Cfne) {
 527                 swap = cmpswap(fi->arg, c);
 528                 if (swap)
 529                         c = cmpop(c);
 530                 if (t->nuse == 1) {
 531                         selcmp(fi->arg, k, swap, fn);
 532                         *fi = (Ins){.op = Onop};
 533                 }
 534                 b->jmp.type = Jjf + c;
 535         }
 536         else if (fi->op == Oand && t->nuse == 1
 537              && (rtype(fi->arg[0]) == RTmp ||
 538                  rtype(fi->arg[1]) == RTmp)) {
 539                 fi->op = Oxtest;
 540                 fi->to = R;
 541                 b->jmp.type = Jjf + Cine;
 542                 if (rtype(fi->arg[1]) == RCon) {
 543                         r = fi->arg[1];
 544                         fi->arg[1] = fi->arg[0];
 545                         fi->arg[0] = r;
 546                 }
 547         }
 548         else {
 549                 /* since flags are not tracked in liveness,
 550                  * the result of the flag-setting instruction
 551                  * has to be marked as live
 552                  */
 553                 if (t->nuse == 1)
 554                         emit(Ocopy, Kw, R, r, R);
 555                 b->jmp.type = Jjf + Cine;
 556         }
 557 }
 558
 559 enum {
 560         Pob,
 561         Pbis,
 562         Pois,
 563         Pobis,
 564         Pbi1,
 565         Pobi1,
 566 };
 567
 568 /* mgen generated code
 569  *
 570  * (with-vars (o b i s)
 571  *   (patterns
 572  *     (ob   (add (con o) (tmp b)))
 573  *     (bis  (add (tmp b) (mul (tmp i) (con s 1 2 4 8))))
 574  *     (ois  (add (con o) (mul (tmp i) (con s 1 2 4 8))))
 575  *     (obis (add (con o) (tmp b) (mul (tmp i) (con s 1 2 4 8))))
 576  *     (bi1  (add (tmp b) (tmp i)))
 577  *     (obi1 (add (con o) (tmp b) (tmp i)))
 578  * ))
 579  */
 580
 581 static int
 582 opn(int op, int l, int r)
 583 {
 584         static uchar Oaddtbl[91] = {
 585                 2,
 586                 2,2,
 587                 4,4,5,
 588                 6,6,8,8,
 589                 4,4,9,10,9,
 590                 7,7,5,8,9,5,
 591                 4,4,12,10,12,12,12,
 592                 4,4,9,10,9,9,12,9,
 593                 11,11,5,8,9,5,12,9,5,
 594                 7,7,5,8,9,5,12,9,5,5,
 595                 11,11,5,8,9,5,12,9,5,5,5,
 596                 4,4,9,10,9,9,12,9,9,9,9,9,
 597                 7,7,5,8,9,5,12,9,5,5,5,9,5,
 598         };
 599         int t;
 600
 601         if (l < r)
 602                 t = l, l = r, r = t;
 603         switch (op) {
 604         case Omul:
 605                 if (2 <= l)
 606                 if (r == 0) {
 607                         return 3;
 608                 }
 609                 return 2;
 610         case Oadd:
 611                 return Oaddtbl[(l + l*l)/2 + r];
 612         default:
 613                 return 2;
 614         }
 615 }
 616
 617 static int
 618 refn(Ref r, Num *tn, Con *con)
 619 {
 620         int64_t n;
 621
 622         switch (rtype(r)) {
 623         case RTmp:
 624                 if (!tn[r.val].n)
 625                         tn[r.val].n = 2;
 626                 return tn[r.val].n;
 627         case RCon:
 628                 if (con[r.val].type != CBits)
 629                         return 1;
 630                 n = con[r.val].bits.i;
 631                 if (n == 8 || n == 4 || n == 2 || n == 1)
 632                         return 0;
 633                 return 1;
 634         default:
 635                 return INT_MIN;
 636         }
 637 }
 638
 639 static bits match[13] = {
 640         [4] = BIT(Pob),
 641         [5] = BIT(Pbi1),
 642         [6] = BIT(Pob) | BIT(Pois),
 643         [7] = BIT(Pob) | BIT(Pobi1),
 644         [8] = BIT(Pbi1) | BIT(Pbis),
 645         [9] = BIT(Pbi1) | BIT(Pobi1),
 646         [10] = BIT(Pbi1) | BIT(Pbis) | BIT(Pobi1) | BIT(Pobis),
 647         [11] = BIT(Pob) | BIT(Pobi1) | BIT(Pobis),
 648         [12] = BIT(Pbi1) | BIT(Pobi1) | BIT(Pobis),
 649 };
 650
 651 static uchar *matcher[] = {
 652         [Pbi1] = (uchar[]){
 653                 1,3,1,3,2,0
 654         },
 655         [Pbis] = (uchar[]){
 656                 5,1,8,5,27,1,5,1,2,5,13,3,1,1,3,3,3,2,0,1,
 657                 3,3,3,2,3,1,0,1,29
 658         },
 659         [Pob] = (uchar[]){
 660                 1,3,0,3,1,0
 661         },
 662         [Pobi1] = (uchar[]){
 663                 5,3,9,9,10,33,12,35,45,1,5,3,11,9,7,9,4,9,
 664                 17,1,3,0,3,1,3,2,0,3,1,1,3,0,34,1,37,1,5,2,
 665                 5,7,2,7,8,37,29,1,3,0,1,32
 666         },
 667         [Pobis] = (uchar[]){
 668                 5,2,10,7,11,19,49,1,1,3,3,3,2,1,3,0,3,1,0,
 669                 1,3,0,5,1,8,5,25,1,5,1,2,5,13,3,1,1,3,3,3,
 670                 2,0,1,3,3,3,2,26,1,51,1,5,1,6,5,9,1,3,0,51,
 671                 3,1,1,3,0,45
 672         },
 673         [Pois] = (uchar[]){
 674                 1,3,0,1,3,3,3,2,0
 675         },
 676 };
 677
 678 /* end of generated code */
 679
 680 static void
 681 anumber(Num *tn, Blk *b, Con *con)
 682 {
 683         Ins *i;
 684         Num *n;
 685
 686         for (i=b->ins; i<&b->ins[b->nins]; i++) {
 687                 if (rtype(i->to) != RTmp)
 688                         continue;
 689                 n = &tn[i->to.val];
 690                 n->l = i->arg[0];
 691                 n->r = i->arg[1];
 692                 n->nl = refn(n->l, tn, con);
 693                 n->nr = refn(n->r, tn, con);
 694                 n->n = opn(i->op, n->nl, n->nr);
 695         }
 696 }
 697
 698 static Ref
 699 adisp(Con *c, Num *tn, Ref r, Fn *fn, int s)
 700 {
 701         Ref v[2];
 702         int n;
 703
 704         while (!req(r, R)) {
 705                 assert(rtype(r) == RTmp);
 706                 n = refn(r, tn, fn->con);
 707                 if (!(match[n] & BIT(Pob)))
 708                         break;
 709                 runmatch(matcher[Pob], tn, r, v);
 710                 assert(rtype(v[0]) == RCon);
 711                 addcon(c, &fn->con[v[0].val], s);
 712                 r = v[1];
 713         }
 714         return r;
 715 }
 716
 717 static int
 718 amatch(Addr *a, Num *tn, Ref r, Fn *fn)
 719 {
 720         static int pat[] = {Pobis, Pobi1, Pbis, Pois, Pbi1, -1};
 721         Ref ro, rb, ri, rs, v[4];
 722         Con *c, co;
 723         int s, n, *p;
 724
 725         if (rtype(r) != RTmp)
 726                 return 0;
 727
 728         n = refn(r, tn, fn->con);
 729         memset(v, 0, sizeof v);
 730         for (p=pat; *p>=0; p++)
 731                 if (match[n] & BIT(*p)) {
 732                         runmatch(matcher[*p], tn, r, v);
 733                         break;
 734                 }
 735         if (*p < 0)
 736                 v[1] = r;
 737
 738         memset(&co, 0, sizeof co);
 739         ro = v[0];
 740         rb = adisp(&co, tn, v[1], fn, 1);
 741         ri = v[2];
 742         rs = v[3];
 743         s = 1;
 744
 745         if (*p < 0 && co.type != CUndef)
 746         if (amatch(a, tn, rb, fn))
 747                 return addcon(&a->offset, &co, 1);
 748         if (!req(ro, R)) {
 749                 assert(rtype(ro) == RCon);
 750                 c = &fn->con[ro.val];
 751                 if (!addcon(&co, c, 1))
 752                         return 0;
 753         }
 754         if (!req(rs, R)) {
 755                 assert(rtype(rs) == RCon);
 756                 c = &fn->con[rs.val];
 757                 assert(c->type = CBits);
 758                 s = c->bits.i;
 759         }
 760         ri = adisp(&co, tn, ri, fn, s);
 761         *a = (Addr){co, rb, ri, s};
 762
 763         if (rtype(ri) == RTmp)
 764         if (fn->tmp[ri.val].slot != -1) {
 765                 if (a->scale != 1
 766                 || fn->tmp[rb.val].slot != -1)
 767                         return 0;
 768                 a->base = ri;
 769                 a->index = rb;
 770         }
 771         if (!req(a->base, R)) {
 772                 assert(rtype(a->base) == RTmp);
 773                 s = fn->tmp[a->base.val].slot;
 774                 if (s != -1)
 775                         a->base = SLOT(s);
 776         }
 777         return 1;
 778 }
 779
 780 /* instruction selection
 781  * requires use counts (as given by parsing)
 782  */
 783 void
 784 amd64_isel(Fn *fn)
 785 {
 786         Blk *b, **sb;
 787         Ins *i;
 788         Phi *p;
 789         uint a;
 790         int n, al;
 791         int64_t sz;
 792         Num *num;
 793
 794         /* assign slots to fast allocs */
 795         b = fn->start;
 796         /* specific to NAlign == 3 */ /* or change n=4 and sz /= 4 below */
 797         for (al=Oalloc, n=4; al<=Oalloc1; al++, n*=2)
 798                 for (i=b->ins; i<&b->ins[b->nins]; i++)
 799                         if (i->op == al) {
 800                                 if (rtype(i->arg[0]) != RCon)
 801                                         break;
 802                                 sz = fn->con[i->arg[0].val].bits.i;
 803                                 if (sz < 0 || sz >= INT_MAX-15)
 804                                         err("invalid alloc size %"PRId64, sz);
 805                                 sz = (sz + n-1) & -n;
 806                                 sz /= 4;
 807                                 if (sz > INT_MAX - fn->slot)
 808                                         die("alloc too large");
 809                                 fn->tmp[i->to.val].slot = fn->slot;
 810                                 fn->slot += sz;
 811                                 *i = (Ins){.op = Onop};
 812                         }
 813
 814         /* process basic blocks */
 815         n = fn->ntmp;
 816         num = emalloc(n * sizeof num[0]);
 817         for (b=fn->start; b; b=b->link) {
 818                 curi = &insb[NIns];
 819                 for (sb=(Blk*[3]){b->s1, b->s2, 0}; *sb; sb++)
 820                         for (p=(*sb)->phi; p; p=p->link) {
 821                                 for (a=0; p->blk[a] != b; a++)
 822                                         assert(a+1 < p->narg);
 823                                 fixarg(&p->arg[a], p->cls, 0, fn);
 824                         }
 825                 memset(num, 0, n * sizeof num[0]);
 826                 anumber(num, b, fn->con);
 827                 seljmp(b, fn);
 828                 for (i=&b->ins[b->nins]; i!=b->ins;)
 829                         sel(*--i, num, fn);
 830                 b->nins = &insb[NIns] - curi;
 831                 idup(&b->ins, curi, b->nins);
 832         }
 833         free(num);
 834
 835         if (debug['I']) {
 836                 fprintf(stderr, "\n> After instruction selection:\n");
 837                 printfn(fn, stderr);
 838         }
 839 }