gen.c

   1 #include <stdlib.h>
   2 #include <string.h>
   3 #include "gen.h"
   4 #include "tok.h"
   5
   6 #define TMP_ADDR        0x0001
   7 #define LOC_REG         0x0100
   8 #define LOC_MEM         0x0200
   9 #define LOC_NUM         0x0400
  10 #define LOC_SYM         0x0800
  11 #define LOC_LOCAL       0x1000
  12 #define LOC_MASK        0xff00
  13
  14 #define R_RAX           0x00
  15 #define R_RCX           0x01
  16 #define R_RDX           0x02
  17 #define R_RBX           0x03
  18 #define R_RSP           0x04
  19 #define R_RBP           0x05
  20 #define R_RSI           0x06
  21 #define R_RDI           0x07
  22 #define R_R8            0x08
  23 #define R_R9            0x09
  24 #define R_R10           0x0a
  25 #define R_R11           0x0b
  26 #define R_R12           0x0c
  27 #define R_R13           0x0d
  28 #define R_R14           0x0e
  29 #define R_R15           0x0f
  30 #define NREGS           0x10
  31
  32 #define MOV_M2R         0x8b
  33 #define MOV_R2X         0x89
  34 #define MOV_I2X         0xc7
  35 #define MOV_I2R         0xb8
  36 #define ADD_R2X         0x03
  37 #define SUB_R2X         0x2b
  38 #define SHX_REG         0xd3
  39 #define CMP_R2X         0x3b
  40 #define LEA_M2R         0x8d
  41 #define NEG_REG         0xf7
  42 #define NOT_REG         0xf7
  43 #define CALL_REG        0xff
  44 #define MUL_A2X         0xf7
  45 #define XOR_R2X         0x33
  46 #define AND_R2X         0x23
  47 #define OR_R2X          0x0b
  48 #define TEST_R2R        0x85
  49
  50 #define MIN(a, b)               ((a) < (b) ? (a) : (b))
  51
  52 #define R_BYTEMASK              (1 << R_RAX | 1 << R_RDX | 1 << R_RCX)
  53 #define TMP_BT(t)               ((t)->flags & TMP_ADDR ? 8 : (t)->bt)
  54 #define TMP_REG(t)              ((t)->flags & LOC_REG ? (t)->addr : reg_get(~0))
  55 #define TMP_REG2(t, r)          ((t)->flags & LOC_REG && (t)->addr != r ? \
  56                                         (t)->addr : reg_get(~(1 << r)))
  57 #define TMP_BYTEREG(t)          ((t)->flags & LOC_REG && \
  58                                  (1 << (t)->addr) & R_BYTEMASK ? \
  59                                         (t)->addr : reg_get(R_BYTEMASK))
  60 #define BT_TMPBT(bt)            (BT_SZ(bt) >= 4 ? (bt) : (bt) & BT_SIGNED | 4)
  61
  62 static char buf[SECSIZE];
  63 static char *cur;
  64 static int nogen;
  65 static long sp;
  66 static long spsub_addr;
  67 static long maxsp;
  68
  69 static struct tmp {
  70         long addr;
  71         long off;       /* used for LOC_SYM; offset from a symbol address */
  72         unsigned flags;
  73         unsigned bt;
  74 } tmp[MAXTMP];
  75 static int ntmp;
  76
  77 static int tmpsp;
  78
  79 static struct tmp *regs[NREGS];
  80 static int tmpregs[] = {R_RAX, R_RDI, R_RSI, R_RDX, R_RCX, R_R8, R_R9};
  81
  82 #define MAXRET                  (1 << 8)
  83
  84 static long ret[MAXRET];
  85 static int nret;
  86
  87 static long cmp_last;
  88 static long cmp_setl;
  89
  90 static void putint(char *s, long n, int l)
  91 {
  92         if (nogen)
  93                 return;
  94         while (l--) {
  95                 *s++ = n;
  96                 n >>= 8;
  97         }
  98 }
  99
 100 static void os(char *s, int n)
 101 {
 102         if (nogen)
 103                 return;
 104         while (n--)
 105                 *cur++ = *s++;
 106 }
 107
 108 static void oi(long n, int l)
 109 {
 110         while (l--) {
 111                 *cur++ = n;
 112                 n >>= 8;
 113         }
 114 }
 115
 116 static long codeaddr(void)
 117 {
 118         return cur - buf;
 119 }
 120
 121 static void o_op(int *op, int nop, int r1, int r2, unsigned bt)
 122 {
 123         int rex = 0;
 124         int i;
 125         if (r1 & 0x8)
 126                 rex |= 4;
 127         if (r2 & 0x8)
 128                 rex |= 1;
 129         if (BT_SZ(bt) == 8)
 130                 rex |= 8;
 131         if (rex || (bt & BT_SZMASK) == 8)
 132                 oi(0x40 | rex, 1);
 133         if ((bt & BT_SZMASK) == 2)
 134                 oi(0x66, 1);
 135         if ((bt & BT_SZMASK) == 1)
 136                 op[nop - 1] &= ~0x1;
 137         for (i = 0; i < nop; i++)
 138                 oi(op[i], 1);
 139 }
 140
 141 static void memop(int *op, int nop, int src, int base, int off, unsigned bt)
 142 {
 143         int dis = off == (char) off ? 1 : 4;
 144         int mod = dis == 4 ? 2 : 1;
 145         o_op(op, nop, src, base, bt);
 146         if (!off)
 147                 mod = 0;
 148         oi((mod << 6) | ((src & 0x07) << 3) | (base & 0x07), 1);
 149         if (off)
 150                 oi(off, dis);
 151 }
 152
 153 static void memop1(int op, int src, int base, int off, unsigned bt)
 154 {
 155         memop(&op, 1, src, base, off, bt);
 156 }
 157
 158 static void regop(int *op, int nop, int src, int dst, unsigned bt)
 159 {
 160         o_op(op, nop, src, dst, bt);
 161         oi((3 << 6) | (src << 3) | (dst & 0x07), 1);
 162 }
 163
 164 static void regop1(int op, int src, int dst, unsigned bt)
 165 {
 166         regop(&op, 1, src, dst, bt);
 167 }
 168
 169 static long sp_push(int size)
 170 {
 171         sp += size;
 172         if (sp > maxsp)
 173                 maxsp = sp;
 174         return sp;
 175 }
 176
 177 #define LOC_NEW(f, l)           (((f) & ~LOC_MASK) | (l))
 178
 179 static void tmp_mem(struct tmp *tmp)
 180 {
 181         int src = tmp->addr;
 182         if (!(tmp->flags & LOC_REG))
 183                 return;
 184         if (tmpsp == -1)
 185                 tmpsp = sp;
 186         tmp->addr = -sp_push(8);
 187         memop1(MOV_R2X, src, R_RBP, tmp->addr, BT_TMPBT(TMP_BT(tmp)));
 188         regs[src] = NULL;
 189         tmp->flags = LOC_NEW(tmp->flags, LOC_MEM);
 190 }
 191
 192 static int *movxx_x2r(int bt)
 193 {
 194         static int movxx[2] = {0x0f};
 195         if (BT_SZ(bt) == 1)
 196                 movxx[1] = bt & BT_SIGNED ? 0xbe : 0xb6;
 197         else
 198                 movxx[1] = bt & BT_SIGNED ? 0xbf : 0xb7;
 199         return movxx;
 200 }
 201
 202 #define MOVSXD          0x63
 203
 204 static void mov_r2r(int r1, int r2, unsigned bt1, unsigned bt2)
 205 {
 206         int s1 = bt1 & BT_SIGNED;
 207         int s2 = bt2 & BT_SIGNED;
 208         int sz1 = BT_SZ(bt1);
 209         int sz2 = BT_SZ(bt2);
 210         if (sz2 < 4 && (sz1 >= sz2 && s1 != s2)) {
 211                 regop(movxx_x2r(bt2), 2, r1, r2, 4);
 212                 return;
 213         }
 214         if (sz1 == 4 && sz2 == 8) {
 215                 regop1(MOVSXD, r2, r1, sz2);
 216                 return;
 217         }
 218         if (r1 != r2 || sz1 > sz2)
 219                 regop1(MOV_R2X, r1, r2, BT_TMPBT(bt2));
 220 }
 221
 222 static void mov_m2r(int dst, int base, int off, int bt1, int bt2)
 223 {
 224         if (BT_SZ(bt1) < 4) {
 225                 memop(movxx_x2r(bt1), 2, dst, base, off,
 226                         bt1 & BT_SIGNED && BT_SZ(bt2) == 8 ? 8 : 4);
 227                 mov_r2r(dst, dst, bt1, bt2);
 228         } else {
 229                 memop1(MOV_M2R, dst, base, off, bt1);
 230                 mov_r2r(dst, dst, bt1, bt2);
 231         }
 232 }
 233
 234 static void num_cast(struct tmp *t, unsigned bt)
 235 {
 236         if (!(bt & BT_SIGNED) && BT_SZ(bt) != 8)
 237                 t->addr &= ((1l << (long) (BT_SZ(bt) * 8)) - 1);
 238         if (bt & BT_SIGNED && BT_SZ(bt) != 8 &&
 239                                 t->addr > (1l << (BT_SZ(bt) * 8 - 1)))
 240                 t->addr = -((1l << (BT_SZ(bt) * 8)) - t->addr);
 241         t->bt = bt;
 242 }
 243
 244 static void num_reg(int reg, unsigned bt, long num)
 245 {
 246         int op = MOV_I2R + (reg & 7);
 247         if (BT_SZ(bt) == 8 && num >= 0 && num == (unsigned) num)
 248                 bt = 4;
 249         o_op(&op, 1, 0, reg, bt);
 250         oi(num, BT_SZ(bt));
 251 }
 252
 253 static void tmp_reg(struct tmp *tmp, int dst, unsigned bt, int deref)
 254 {
 255         if (!(tmp->flags & TMP_ADDR))
 256                 deref = 0;
 257         if (deref)
 258                 tmp->flags &= ~TMP_ADDR;
 259         if (tmp->flags & LOC_NUM) {
 260                 num_cast(tmp, bt);
 261                 tmp->bt = BT_TMPBT(bt);
 262                 num_reg(dst, tmp->bt, tmp->addr);
 263                 tmp->addr = dst;
 264                 regs[dst] = tmp;
 265                 tmp->flags = LOC_NEW(tmp->flags, LOC_REG);
 266         }
 267         if (tmp->flags & LOC_SYM) {
 268                 regop1(MOV_I2X, 0, dst, TMP_BT(tmp));
 269                 if (!nogen)
 270                         out_rela(tmp->addr, codeaddr(), 0);
 271                 oi(tmp->off, 4);
 272                 tmp->addr = dst;
 273                 regs[dst] = tmp;
 274                 tmp->flags = LOC_NEW(tmp->flags, LOC_REG);
 275         }
 276         if (tmp->flags & LOC_REG) {
 277                 if (deref)
 278                         mov_m2r(dst, tmp->addr, 0, tmp->bt, bt);
 279                 else
 280                         mov_r2r(tmp->addr, dst, TMP_BT(tmp), bt);
 281                 regs[tmp->addr] = NULL;
 282                 tmp->addr = dst;
 283                 tmp->bt = BT_TMPBT(bt);
 284                 regs[dst] = tmp;
 285                 return;
 286         }
 287         if (tmp->flags & LOC_LOCAL) {
 288                 if (deref)
 289                         mov_m2r(dst, R_RBP, tmp->addr, tmp->bt, bt);
 290                 else
 291                         memop1(LEA_M2R, dst, R_RBP, tmp->addr, 8);
 292         }
 293         if (tmp->flags & LOC_MEM) {
 294                 mov_m2r(dst, R_RBP, tmp->addr,
 295                         deref ? 8 : TMP_BT(tmp), deref ? 8 : bt);
 296                 if (deref)
 297                         mov_m2r(dst, dst, 0, tmp->bt, bt);
 298         }
 299         tmp->addr = dst;
 300         tmp->bt = BT_TMPBT(bt);
 301         regs[dst] = tmp;
 302         tmp->flags = LOC_NEW(tmp->flags, LOC_REG);
 303 }
 304
 305 static void reg_free(int reg)
 306 {
 307         int i;
 308         if (!regs[reg])
 309                 return;
 310         for (i = 0; i < ARRAY_SIZE(tmpregs); i++)
 311                 if (!regs[tmpregs[i]]) {
 312                         tmp_reg(regs[reg], tmpregs[i], regs[reg]->bt, 0);
 313                         return;
 314                 }
 315         tmp_mem(regs[reg]);
 316 }
 317
 318 static void reg_for(int reg, struct tmp *t)
 319 {
 320         if (regs[reg] && regs[reg] != t)
 321                 reg_free(reg);
 322 }
 323
 324 static void tmp_pop_bt(unsigned bt, int reg)
 325 {
 326         struct tmp *t = &tmp[--ntmp];
 327         reg_for(reg, t);
 328         tmp_reg(t, reg, bt, 1);
 329         regs[reg] = NULL;
 330 }
 331
 332 static unsigned tmp_pop(int deref, int reg)
 333 {
 334         struct tmp *t = &tmp[--ntmp];
 335         reg_for(reg, t);
 336         tmp_reg(t, reg, deref ? t->bt : TMP_BT(t), deref);
 337         regs[reg] = NULL;
 338         return t->bt;
 339 }
 340
 341 static struct tmp *tmp_new(void)
 342 {
 343         cmp_last = -1;
 344         return &tmp[ntmp++];
 345 }
 346
 347 static void tmp_push_reg(unsigned bt, unsigned reg)
 348 {
 349         struct tmp *t = tmp_new();
 350         t->addr = reg;
 351         t->bt = bt;
 352         t->flags = LOC_REG;
 353         regs[reg] = t;
 354 }
 355
 356 void o_local(long addr, unsigned bt)
 357 {
 358         struct tmp *t = tmp_new();
 359         t->addr = -addr;
 360         t->bt = bt;
 361         t->flags = LOC_LOCAL | TMP_ADDR;
 362 }
 363
 364 void o_num(long num, unsigned bt)
 365 {
 366         struct tmp *t = tmp_new();
 367         t->addr = num;
 368         t->bt = bt;
 369         t->flags = LOC_NUM;
 370 }
 371
 372 void o_symaddr(long addr, unsigned bt)
 373 {
 374         struct tmp *t = tmp_new();
 375         t->bt = bt;
 376         t->addr = addr;
 377         t->flags = LOC_SYM | TMP_ADDR;
 378         t->off = 0;
 379 }
 380
 381 void o_tmpdrop(int n)
 382 {
 383         int i;
 384         cmp_last = -1;
 385         if (n == -1 || n > ntmp)
 386                 n = ntmp;
 387         ntmp -= n;
 388         for (i = ntmp; i < ntmp + n; i++)
 389                 if (tmp[i].flags & LOC_REG)
 390                         regs[tmp[i].addr] = NULL;
 391         if (!ntmp) {
 392                 if (tmpsp != -1)
 393                         sp = tmpsp;
 394                 tmpsp = -1;
 395         }
 396 }
 397
 398 #define FORK_REG                R_RAX
 399
 400 void o_tmpfork(void)
 401 {
 402         struct tmp *t = &tmp[ntmp - 1];
 403         reg_for(FORK_REG, t);
 404         tmp_reg(t, FORK_REG, t->bt, 1);
 405         o_tmpdrop(1);
 406 }
 407
 408 void o_tmpjoin(void)
 409 {
 410         struct tmp *t = &tmp[ntmp - 1];
 411         reg_for(FORK_REG, t);
 412         tmp_reg(t, FORK_REG, t->bt, 1);
 413 }
 414
 415 void o_tmpswap(void)
 416 {
 417         struct tmp *t1 = &tmp[ntmp - 1];
 418         struct tmp *t2 = &tmp[ntmp - 2];
 419         struct tmp t;
 420         memcpy(&t, t1, sizeof(t));
 421         memcpy(t1, t2, sizeof(t));
 422         memcpy(t2, &t, sizeof(t));
 423 }
 424
 425 static int reg_get(int mask)
 426 {
 427         int i;
 428         for (i = 0; i < ARRAY_SIZE(tmpregs); i++)
 429                 if ((1 << tmpregs[i]) & mask && !regs[tmpregs[i]])
 430                         return tmpregs[i];
 431         for (i = 0; i < ARRAY_SIZE(tmpregs); i++)
 432                 if ((1 << tmpregs[i]) & mask) {
 433                         reg_free(tmpregs[i]);
 434                         return tmpregs[i];
 435                 }
 436         return 0;
 437 }
 438
 439 void o_tmpcopy(void)
 440 {
 441         struct tmp *t1 = &tmp[ntmp - 1];
 442         struct tmp *t2 = tmp_new();
 443         memcpy(t2, t1, sizeof(*t1));
 444         if (!(t1->flags & (LOC_REG | LOC_MEM)))
 445                 return;
 446         if (t1->flags & LOC_MEM) {
 447                 tmp_reg(t2, reg_get(~0), t2->bt, 0);
 448         } else if (t1->flags & LOC_REG) {
 449                 t2->addr = reg_get(~t1->addr);
 450                 regop1(MOV_R2X, t1->addr, t2->addr, BT_TMPBT(TMP_BT(t1)));
 451         }
 452         t2->flags = t1->flags;
 453 }
 454
 455 void o_cast(unsigned bt)
 456 {
 457         struct tmp *t = &tmp[ntmp - 1];
 458         int reg;
 459         if (t->bt == bt)
 460                 return;
 461         if (t->flags & LOC_NUM) {
 462                 num_cast(t, bt);
 463                 return;
 464         }
 465         reg = BT_SZ(bt) == 1 ? TMP_BYTEREG(t) : TMP_REG(t);
 466         tmp_pop_bt(bt, reg);
 467         tmp_push_reg(bt, reg);
 468 }
 469
 470 long o_func_beg(char *name, int global)
 471 {
 472         long addr = out_func_beg(name, global);
 473         cur = buf;
 474         os("\x55", 1);                  /* push %rbp */
 475         os("\x48\x89\xe5", 3);          /* mov %rsp, %rbp */
 476         sp = 0;
 477         maxsp = 0;
 478         ntmp = 0;
 479         tmpsp = -1;
 480         nret = 0;
 481         cmp_last = -1;
 482         memset(regs, 0, sizeof(regs));
 483         os("\x48\x81\xec", 3);          /* sub $xxx, %rsp */
 484         spsub_addr = codeaddr();
 485         oi(0, 4);
 486         return addr;
 487 }
 488
 489 void o_deref(unsigned bt)
 490 {
 491         struct tmp *t = &tmp[ntmp - 1];
 492         if (t->flags & TMP_ADDR)
 493                 tmp_reg(t, TMP_REG(t), 8, 1);
 494         t->bt = bt;
 495         t->flags |= TMP_ADDR;
 496 }
 497
 498 void o_load(void)
 499 {
 500         struct tmp *t = &tmp[ntmp - 1];
 501         tmp_reg(t, TMP_REG(t), t->bt, 1);
 502 }
 503
 504 static unsigned bt_op(unsigned bt1, unsigned bt2)
 505 {
 506         unsigned s1 = BT_SZ(bt1);
 507         unsigned s2 = BT_SZ(bt2);
 508         unsigned bt = (bt1 | bt2) & BT_SIGNED | (s1 > s2 ? s1 : s2);
 509         return BT_TMPBT(bt);
 510 }
 511
 512 #define TMP_CONST(t)    ((t)->flags & LOC_NUM && !((t)->flags & TMP_ADDR))
 513 #define LOCAL_PTR(t)    ((t)->flags & LOC_LOCAL && !((t)->flags & TMP_ADDR))
 514 #define SYM_PTR(t)      ((t)->flags & LOC_LOCAL && !((t)->flags & TMP_ADDR))
 515
 516 int o_popnum(long *c)
 517 {
 518         struct tmp *t = &tmp[ntmp - 1];
 519         if (!TMP_CONST(t))
 520                 return 1;
 521         *c = t->addr;
 522         o_tmpdrop(1);
 523         return 0;
 524 }
 525
 526 static int c_binop(long (*cop)(long a, long b, unsigned bt), unsigned bt)
 527 {
 528         struct tmp *t1 = &tmp[ntmp - 1];
 529         struct tmp *t2 = &tmp[ntmp - 2];
 530         int locals = LOCAL_PTR(t1) + LOCAL_PTR(t2);
 531         int syms = SYM_PTR(t1) + SYM_PTR(t2);
 532         int nums = TMP_CONST(t1) + TMP_CONST(t2);
 533         if (syms == 2 || syms && locals || syms + nums + locals != 2)
 534                 return 1;
 535         if (!bt) {
 536                 if (!locals)
 537                         bt = syms ? 8 : bt_op(t1->bt, t2->bt);
 538                 if (locals == 1)
 539                         bt = 8;
 540                 if (locals == 2)
 541                         bt = 4 | BT_SIGNED;
 542         }
 543         if (syms) {
 544                 long o1 = SYM_PTR(t1) ? t1->off : t1->addr;
 545                 long o2 = SYM_PTR(t2) ? t2->off : t2->addr;
 546                 long ret = cop(o2, o1, bt);
 547                 if (!SYM_PTR(t2))
 548                         o_tmpswap();
 549                 t2->off = ret;
 550                 o_tmpdrop(1);
 551         } else {
 552                 long ret = cop(t2->addr, t1->addr, locals ? 4 | BT_SIGNED : bt);
 553                 o_tmpdrop(2);
 554                 if (locals == 1) {
 555                         o_local(-ret, bt);
 556                         o_addr();
 557                 } else {
 558                         o_num(ret, bt);
 559                 }
 560         }
 561         return 0;
 562 }
 563
 564 static int c_op(long (*cop)(long a, unsigned bt), unsigned bt)
 565 {
 566         struct tmp *t1 = &tmp[ntmp - 1];
 567         long ret;
 568         if (!TMP_CONST(t1))
 569                 return 1;
 570         if (!bt)
 571                 bt = t1->bt;
 572         ret = cop(t1->addr, bt);
 573         o_tmpdrop(1);
 574         o_num(ret, bt);
 575         return 0;
 576 }
 577
 578 static void shx(int uop, int sop)
 579 {
 580         struct tmp *t = &tmp[ntmp - 2];
 581         unsigned bt;
 582         unsigned reg = TMP_REG2(t, R_RCX);
 583         tmp_pop(1, R_RCX);
 584         bt = tmp_pop(1, reg);
 585         regop1(SHX_REG, bt & BT_SIGNED ? sop : uop, reg, BT_TMPBT(bt));
 586         tmp_push_reg(bt, reg);
 587 }
 588
 589 static long c_shl(long a, long b, unsigned bt)
 590 {
 591         return a << b;
 592 }
 593
 594 void o_shl(void)
 595 {
 596         if (!c_binop(c_shl, 0))
 597                 return;
 598         shx(4, 4);
 599 }
 600
 601 static long c_shr(long a, long b, unsigned bt)
 602 {
 603         if (bt & BT_SIGNED)
 604                 return a >> b;
 605         else
 606                 return (unsigned long) a >> b;
 607 }
 608
 609 void o_shr(void)
 610 {
 611         if (!c_binop(c_shr, 0))
 612                 return;
 613         shx(5, 7);
 614 }
 615
 616 static int mulop(int uop, int sop, int reg)
 617 {
 618         struct tmp *t1 = &tmp[ntmp - 1];
 619         struct tmp *t2 = &tmp[ntmp - 2];
 620         int bt1 = TMP_BT(t1);
 621         int bt2 = TMP_BT(t2);
 622         int bt = bt_op(bt1, bt2);
 623         if (t1->flags & LOC_REG && t1->addr != R_RAX && t1->addr != R_RDX)
 624                 reg = t1->addr;
 625         reg_for(reg, t1);
 626         tmp_reg(t1, reg, bt, 1);
 627         reg_for(R_RAX, t2);
 628         tmp_reg(t2, R_RAX, bt, 1);
 629         if (reg != R_RDX)
 630                 reg_free(R_RDX);
 631         o_tmpdrop(2);
 632         regop1(MUL_A2X, bt & BT_SIGNED ? sop : uop, reg, BT_TMPBT(bt2));
 633         return bt;
 634 }
 635
 636 static long c_mul(long a, long b, unsigned bt)
 637 {
 638         return a * b;
 639 }
 640
 641 void o_mul(void)
 642 {
 643         int bt;
 644         if (!c_binop(c_mul, 0))
 645                 return;
 646         bt = mulop(4, 5, R_RDX);
 647         tmp_push_reg(bt, R_RAX);
 648 }
 649
 650 static long c_div(long a, long b, unsigned bt)
 651 {
 652         return a / b;
 653 }
 654
 655 void o_div(void)
 656 {
 657         int bt;
 658         if (!c_binop(c_div, 0))
 659                 return;
 660         bt = mulop(6, 7, R_RCX);
 661         tmp_push_reg(bt, R_RAX);
 662 }
 663
 664 static long c_mod(long a, long b, unsigned bt)
 665 {
 666         return a % b;
 667 }
 668
 669 void o_mod(void)
 670 {
 671         int bt;
 672         if (!c_binop(c_mod, 0))
 673                 return;
 674         bt = mulop(6, 7, R_RCX);
 675         tmp_push_reg(bt, R_RDX);
 676 }
 677
 678 void o_addr(void)
 679 {
 680         tmp[ntmp - 1].flags &= ~TMP_ADDR;
 681         tmp[ntmp - 1].bt = 8;
 682 }
 683
 684 void o_ret(unsigned bt)
 685 {
 686         if (bt)
 687                 tmp_pop_bt(bt, R_RAX);
 688         else
 689                 os("\x31\xc0", 2);      /* xor %eax, %eax */
 690         ret[nret++] = o_jmp(0);
 691 }
 692
 693 static int binop(int op, int *reg)
 694 {
 695         struct tmp *t1 = &tmp[ntmp - 1];
 696         struct tmp *t2 = &tmp[ntmp - 2];
 697         int r1;
 698         unsigned bt;
 699         r1 = TMP_REG(t1);
 700         *reg = TMP_REG2(t2, r1);
 701         bt = bt_op(t1->bt, t2->bt);
 702         tmp_pop_bt(bt, r1);
 703         tmp_pop_bt(bt, *reg);
 704         regop1(op, *reg, r1, bt);
 705         return bt;
 706 }
 707
 708 static long c_add(long a, long b, unsigned bt)
 709 {
 710         return a + b;
 711 }
 712
 713 void o_add(void)
 714 {
 715         int reg;
 716         int bt;
 717         if (!c_binop(c_add, 0))
 718                 return;
 719         bt = binop(ADD_R2X, &reg);
 720         tmp_push_reg(bt, reg);
 721 }
 722
 723 static long c_xor(long a, long b, unsigned bt)
 724 {
 725         return a ^ b;
 726 }
 727
 728 void o_xor(void)
 729 {
 730         int reg;
 731         int bt;
 732         if (!c_binop(c_xor, 0))
 733                 return;
 734         bt = binop(XOR_R2X, &reg);
 735         tmp_push_reg(bt, reg);
 736 }
 737
 738 static long c_and(long a, long b, unsigned bt)
 739 {
 740         return a & b;
 741 }
 742
 743 void o_and(void)
 744 {
 745         int reg;
 746         int bt;
 747         if (!c_binop(c_and, 0))
 748                 return;
 749         bt = binop(AND_R2X, &reg);
 750         tmp_push_reg(bt, reg);
 751 }
 752
 753 static long c_or(long a, long b, unsigned bt)
 754 {
 755         return a | b;
 756 }
 757
 758 void o_or(void)
 759 {
 760         int reg;
 761         int bt;
 762         if (!c_binop(c_or, 0))
 763                 return;
 764         bt = binop(OR_R2X, &reg);
 765         tmp_push_reg(bt, reg);
 766 }
 767
 768 static long c_sub(long a, long b, unsigned bt)
 769 {
 770         return a - b;
 771 }
 772
 773 void o_sub(void)
 774 {
 775         int reg;
 776         int bt;
 777         if (!c_binop(c_sub, 0))
 778                 return;
 779         bt = binop(SUB_R2X, &reg);
 780         tmp_push_reg(bt, reg);
 781 }
 782
 783 static void o_cmp(int uop, int sop)
 784 {
 785         char set[] = "\x0f\x00\xc0";
 786         int reg;
 787         int bt = binop(CMP_R2X, &reg);
 788         set[1] = bt & BT_SIGNED ? sop : uop;
 789         reg_free(R_RAX);
 790         cmp_setl = codeaddr();
 791         os(set, 3);                     /* setl %al */
 792         os("\x0f\xb6\xc0", 3);          /* movzbl %al, %eax */
 793         tmp_push_reg(4 | BT_SIGNED, R_RAX);
 794         cmp_last = codeaddr();
 795 }
 796
 797 static long c_lt(long a, long b, unsigned bt)
 798 {
 799         return a < b;
 800 }
 801
 802 void o_lt(void)
 803 {
 804         if (!c_binop(c_lt, 4))
 805                 return;
 806         o_cmp(0x92, 0x9c);
 807 }
 808
 809 static long c_gt(long a, long b, unsigned bt)
 810 {
 811         return a > b;
 812 }
 813
 814 void o_gt(void)
 815 {
 816         if (!c_binop(c_gt, 4))
 817                 return;
 818         o_cmp(0x97, 0x9f);
 819 }
 820
 821 static long c_le(long a, long b, unsigned bt)
 822 {
 823         return a <= b;
 824 }
 825
 826 void o_le(void)
 827 {
 828         if (!c_binop(c_le, 4))
 829                 return;
 830         o_cmp(0x96, 0x9e);
 831 }
 832
 833 static long c_ge(long a, long b, unsigned bt)
 834 {
 835         return a >= b;
 836 }
 837
 838 void o_ge(void)
 839 {
 840         if (!c_binop(c_ge, 4))
 841                 return;
 842         o_cmp(0x93, 0x9d);
 843 }
 844
 845 static long c_eq(long a, long b, unsigned bt)
 846 {
 847         return a == b;
 848 }
 849
 850 void o_eq(void)
 851 {
 852         if (!c_binop(c_eq, 4))
 853                 return;
 854         o_cmp(0x94, 0x94);
 855 }
 856
 857 static long c_neq(long a, long b, unsigned bt)
 858 {
 859         return a != b;
 860 }
 861
 862 void o_neq(void)
 863 {
 864         if (!c_binop(c_neq, 4))
 865                 return;
 866         o_cmp(0x95, 0x95);
 867 }
 868
 869 static long c_lnot(long a, unsigned bt)
 870 {
 871         return !a;
 872 }
 873
 874 void o_lnot(void)
 875 {
 876         if (!c_op(c_lnot, 4))
 877                 return;
 878         if (cmp_last == codeaddr()) {
 879                 buf[cmp_setl + 1] ^= 0x10;
 880         } else {
 881                 o_num(0, 4 | BT_SIGNED);
 882                 o_eq();
 883         }
 884 }
 885
 886 static long c_neg(long a, unsigned bt)
 887 {
 888         return -a;
 889 }
 890
 891 void o_neg(void)
 892 {
 893         struct tmp *t = &tmp[ntmp - 1];
 894         int reg;
 895         unsigned bt = BT_TMPBT(t->bt);
 896         if (!c_op(c_neg, t->bt | BT_SIGNED))
 897                 return;
 898         reg = TMP_REG(t);
 899         tmp_pop_bt(bt, reg);
 900         regop1(NEG_REG, 3, reg, bt);
 901         tmp_push_reg(bt, reg);
 902 }
 903
 904 static long c_not(long a, unsigned bt)
 905 {
 906         return ~a;
 907 }
 908
 909 void o_not(void)
 910 {
 911         struct tmp *t = &tmp[ntmp - 1];
 912         int reg;
 913         unsigned bt = BT_TMPBT(t->bt);
 914         if (!c_op(c_not, 0))
 915                 return;
 916         reg = TMP_REG(t);
 917         tmp_pop_bt(bt, reg);
 918         regop1(NOT_REG, 2, reg, bt);
 919         tmp_push_reg(t->bt, reg);
 920 }
 921
 922 void o_func_end(void)
 923 {
 924         int i;
 925         for (i = 0; i < nret; i++)
 926                 o_filljmp(ret[i]);
 927         os("\xc9\xc3", 2);              /* leave; ret; */
 928         putint(buf + spsub_addr, (maxsp + 7) & ~0x07, 4);
 929         out_func_end(buf, cur - buf);
 930 }
 931
 932 long o_mklocal(int size)
 933 {
 934         return sp_push((size + 7) & ~0x07);
 935 }
 936
 937 void o_rmlocal(long addr, int sz)
 938 {
 939         sp = addr - sz;
 940 }
 941
 942 static int arg_regs[] = {R_RDI, R_RSI, R_RDX, R_RCX, R_R8, R_R9};
 943
 944 #define R_NARGS                 ARRAY_SIZE(arg_regs)
 945
 946 long o_arg(int i, unsigned bt)
 947 {
 948         long addr;
 949         if (i < R_NARGS) {
 950                 addr = o_mklocal(BT_SZ(bt));
 951                 memop1(MOV_R2X, arg_regs[i], R_RBP, -addr, bt);
 952         } else {
 953                 addr = -8 * (i - R_NARGS + 2);
 954         }
 955         return addr;
 956 }
 957
 958 void o_assign(unsigned bt)
 959 {
 960         struct tmp *t1 = &tmp[ntmp - 1];
 961         struct tmp *t2 = &tmp[ntmp - 2];
 962         int r1 = BT_SZ(bt) > 1 ? TMP_REG(t1) : TMP_BYTEREG(t1);
 963         int reg;
 964         int off;
 965         tmp_pop_bt(BT_TMPBT(bt), r1);
 966         if (t2->flags & LOC_LOCAL) {
 967                 reg = R_RBP;
 968                 off = t2->addr;
 969                 o_tmpdrop(1);
 970         } else {
 971                 reg = TMP_REG2(t2, r1);
 972                 off = 0;
 973                 tmp_pop(0, reg);
 974         }
 975         memop1(MOV_R2X, r1, reg, off, bt);
 976         tmp_push_reg(bt, r1);
 977 }
 978
 979 static void tmp_to(struct tmp *t, int reg)
 980 {
 981         reg_for(reg, t);
 982         tmp_reg(t, reg, t->bt, 0);
 983 }
 984
 985 void o_memcpy(int sz)
 986 {
 987         struct tmp *t0 = &tmp[ntmp];
 988         struct tmp *t1 = &tmp[ntmp - 1];
 989         struct tmp *t2 = &tmp[ntmp - 2];
 990         o_num(sz, 4);
 991         tmp_to(t0, R_RCX);
 992         tmp_to(t1, R_RSI);
 993         tmp_to(t2, R_RDI);
 994         os("\xf3\xa4", 2);              /* rep movs */
 995         o_tmpdrop(2);
 996 }
 997
 998 void o_memset(int x, int sz)
 999 {
1000         struct tmp *t0 = &tmp[ntmp + 1];
1001         struct tmp *t1 = &tmp[ntmp];
1002         struct tmp *t2 = &tmp[ntmp - 1];
1003         o_num(sz, 4);
1004         o_num(x, 4);
1005         tmp_to(t0, R_RAX);
1006         tmp_to(t1, R_RCX);
1007         tmp_to(t2, R_RDI);
1008         os("\xf3\xaa", 2);              /* rep stosb */
1009         o_tmpdrop(2);
1010 }
1011
1012 long o_mklabel(void)
1013 {
1014         return codeaddr();
1015 }
1016
1017 static long jx(int x, long addr)
1018 {
1019         char op[2] = {0x0f};
1020         op[1] = x;
1021         os(op, 2);              /* jx $addr */
1022         oi(addr - codeaddr() - 4, 4);
1023         return codeaddr() - 4;
1024 }
1025
1026 static long jxtest(int x, long addr)
1027 {
1028         int bt = tmp_pop(1, R_RAX);
1029         regop1(TEST_R2R, R_RAX, R_RAX, bt);
1030         return jx(x, addr);
1031 }
1032
1033 static long jxcmp(long addr, int inv)
1034 {
1035         int x;
1036         if (codeaddr() != cmp_last)
1037                 return -1;
1038         o_tmpdrop(1);
1039         cur = buf + cmp_setl;
1040         x = (unsigned char) buf[cmp_setl + 1];
1041         return jx((inv ? x : x ^ 0x01) & ~0x10, addr);
1042 }
1043
1044 long o_jz(long addr)
1045 {
1046         long ret = jxcmp(addr, 0);
1047         return ret != -1 ? ret : jxtest(0x84, addr);
1048 }
1049
1050 long o_jnz(long addr)
1051 {
1052         long ret = jxcmp(addr, 1);
1053         return ret != -1 ? ret : jxtest(0x85, addr);
1054 }
1055
1056 long o_jmp(long addr)
1057 {
1058         os("\xe9", 1);                  /* jmp $addr */
1059         oi(addr - codeaddr() - 4, 4);
1060         return codeaddr() - 4;
1061 }
1062
1063 void o_filljmp2(long addr, long jmpdst)
1064 {
1065         putint(buf + addr, jmpdst - addr - 4, 4);
1066 }
1067
1068 void o_filljmp(long addr)
1069 {
1070         o_filljmp2(addr, codeaddr());
1071 }
1072
1073 void o_call(int argc, unsigned *bt, unsigned ret_bt)
1074 {
1075         int i;
1076         struct tmp *t;
1077         for (i = 0; i < ARRAY_SIZE(tmpregs); i++)
1078                 if (regs[tmpregs[i]] && regs[tmpregs[i]] - tmp < ntmp - argc)
1079                         tmp_mem(regs[tmpregs[i]]);
1080         if (argc > R_NARGS) {
1081                 long addr = sp_push(8 * (argc - R_NARGS));
1082                 for (i = argc - 1; i >= R_NARGS; --i) {
1083                         int reg = TMP_REG(&tmp[ntmp - 1]);
1084                         tmp_pop_bt(bt[i], reg);
1085                         memop1(MOV_R2X, reg, R_RBP,
1086                                 -(addr - (i - R_NARGS) * 8), BT_TMPBT(bt[i]));
1087                 }
1088         }
1089         for (i = MIN(argc, R_NARGS) - 1; i >= 0; i--)
1090                 tmp_pop_bt(bt[i], arg_regs[i]);
1091         t = &tmp[ntmp - 1];
1092         if (t->flags & LOC_SYM) {
1093                 os("\x31\xc0", 2);      /* xor %eax, %eax */
1094                 os("\xe8", 1);          /* call $x */
1095                 if (!nogen)
1096                         out_rela(t->addr, codeaddr(), 1);
1097                 oi(-4 + t->off, 4);
1098                 o_tmpdrop(1);
1099         } else {
1100                 tmp_pop(0, R_RAX);
1101                 regop1(CALL_REG, 2, R_RAX, 4);
1102         }
1103         if (ret_bt)
1104                 tmp_push_reg(ret_bt, R_RAX);
1105 }
1106
1107 int o_nogen(void)
1108 {
1109         return nogen++;
1110 }
1111
1112 void o_dogen(void)
1113 {
1114         nogen = 0;
1115 }
1116
1117 void o_datset(long addr, int off, unsigned bt)
1118 {
1119         struct tmp *t = &tmp[ntmp - 1];
1120         if (t->flags & LOC_NUM && !(t->flags & TMP_ADDR)) {
1121                 num_cast(t, bt);
1122                 out_datcpy(addr, off, (void *) &t->addr, BT_SZ(bt));
1123         }
1124         if (t->flags & LOC_SYM && !(t->flags & TMP_ADDR)) {
1125                 out_datrela(tmp->addr, addr, off);
1126                 out_datcpy(addr, off, (void *) &t->off, BT_SZ(bt));
1127         }
1128         o_tmpdrop(1);
1129 }