gen.c

   1 /* neatcc code generation */
   2 #include <stdlib.h>
   3 #include <stdio.h>
   4 #include <string.h>
   5 #include "ncc.h"
   6
   7 static struct mem ds;           /* data segment */
   8 static struct mem cs;           /* code segment */
   9 static long bsslen;             /* bss segment size */
  10 static struct ic *ic;           /* current instruction stream */
  11 static long ic_n;               /* number of instructions in ic[] */
  12 static long ic_i;               /* current instruction */
  13 static long *ic_luse;           /* last instruction in which values are used */
  14
  15 static long *loc_off;           /* offset of locals on the stack */
  16 static long loc_n, loc_sz;      /* number of locals */
  17 static long loc_pos;            /* current stack position */
  18 static int *loc_mem;            /* local was accessed on the stack */
  19
  20 static char (*ds_name)[NAMELEN];/* data section symbols */
  21 static long *ds_off;            /* data section offsets */
  22 static long ds_n, ds_sz;        /* number of data section symbols */
  23
  24 static int func_argc;           /* number of arguments */
  25 static int func_varg;           /* varargs */
  26 static int func_regs;           /* used registers */
  27 static int func_maxargs;        /* the maximum number of arguments on the stack */
  28 static long *ic_bbeg;           /* whether each instruction begins a basic block */
  29
  30 static long ra_vmap[N_REGS];    /* register to intermediate value assignments */
  31 static long ra_lmap[N_REGS];    /* register to local assignments */
  32 static long *ra_gmask;          /* the mask of good registers for each value */
  33 static long ra_live[NTMPS];     /* live values */
  34 static int ra_vmax;             /* the number of values stored on the stack */
  35
  36 static long loc_add(long pos)
  37 {
  38         if (loc_n >= loc_sz) {
  39                 loc_sz = MAX(128, loc_sz * 2);
  40                 loc_off = mextend(loc_off, loc_n, loc_sz, sizeof(loc_off[0]));
  41         }
  42         loc_off[loc_n] = pos;
  43         return loc_n++;
  44 }
  45
  46 long o_mklocal(long sz)
  47 {
  48         loc_pos += ALIGN(sz, ULNG);
  49         return loc_add(loc_pos);
  50 }
  51
  52 void o_rmlocal(long addr, long sz)
  53 {
  54 }
  55
  56 long o_arg2loc(int i)
  57 {
  58         return i;
  59 }
  60
  61 void o_bsnew(char *name, long size, int global)
  62 {
  63         out_def(name, OUT_BSS | (global ? OUT_GLOB : 0), bsslen, size);
  64         bsslen += ALIGN(size, OUT_ALIGNMENT);
  65 }
  66
  67 long o_dsnew(char *name, long size, int global)
  68 {
  69         int idx;
  70         if (ds_n >= ds_sz) {
  71                 ds_sz = MAX(128, ds_sz * 2);
  72                 ds_name = mextend(ds_name, ds_n, ds_sz, sizeof(ds_name[0]));
  73                 ds_off = mextend(ds_off, ds_n, ds_sz, sizeof(ds_off[0]));
  74         }
  75         idx = ds_n++;
  76         strcpy(ds_name[idx], name);
  77         ds_off[idx] = mem_len(&ds);
  78         out_def(name, OUT_DS | (global ? OUT_GLOB : 0), mem_len(&ds), size);
  79         mem_putz(&ds, ALIGN(size, OUT_ALIGNMENT));
  80         return ds_off[idx];
  81 }
  82
  83 void o_dscpy(long addr, void *buf, long len)
  84 {
  85         mem_cpy(&ds, addr, buf, len);
  86 }
  87
  88 static int dat_off(char *name)
  89 {
  90         int i;
  91         for (i = 0; i < ds_n; i++)
  92                 if (!strcmp(name, ds_name[i]))
  93                         return ds_off[i];
  94         return 0;
  95 }
  96
  97 void o_dsset(char *name, long off, long bt)
  98 {
  99         long sym_off = dat_off(name) + off;
 100         long num, roff, rsym;
 101         if (!o_popnum(&num)) {
 102                 mem_cpy(&ds, sym_off, &num, T_SZ(bt));
 103                 return;
 104         }
 105         if (!o_popsym(&rsym, &roff)) {
 106                 out_rel(rsym, OUT_DS, sym_off);
 107                 mem_cpy(&ds, sym_off, &roff, T_SZ(bt));
 108         }
 109 }
 110
 111 static int ra_vreg(int val)
 112 {
 113         int i;
 114         for (i = 0; i < LEN(ra_vmap); i++)
 115                 if (ra_vmap[i] == val)
 116                         return i;
 117         return -1;
 118 }
 119
 120 static int ra_lreg(int loc)
 121 {
 122         int i;
 123         for (i = 0; i < LEN(ra_lmap); i++)
 124                 if (ra_lmap[i] == loc)
 125                         return i;
 126         return -1;
 127 }
 128
 129 /* mask of registers assigned to locals */
 130 static long ra_lmask(void)
 131 {
 132         long m = 0;
 133         int i;
 134         for (i = 0; i < LEN(ra_lmap); i++)
 135                 if (ra_lmap[i] >= 0)
 136                         m |= (1 << i);
 137         return m;
 138 }
 139
 140 /* mask of registers assigned to values */
 141 static long ra_vmask(void)
 142 {
 143         long m = 0;
 144         int i;
 145         for (i = 0; i < LEN(ra_vmap); i++)
 146                 if (ra_vmap[i] >= 0)
 147                         m |= (1 << i);
 148         return m;
 149 }
 150
 151 /* find a temporary register specified in the given mask */
 152 static long ra_regscn(long mask)
 153 {
 154         int i;
 155         for (i = 0; i < N_TMPS; i++)
 156                 if ((1 << tmpregs[i]) & mask)
 157                         return tmpregs[i];
 158         return -1;
 159 }
 160
 161 /* find a register, with the given good, acceptable, and bad register masks */
 162 static long ra_regget(long iv, long gmask, long amask, long bmask)
 163 {
 164         long lmask, vmask;
 165         gmask &= ~bmask & amask;
 166         amask &= ~bmask;
 167         if (ra_vreg(iv) >= 0 && (1 << ra_vreg(iv)) & (gmask | amask))
 168                 return ra_vreg(iv);
 169         vmask = ra_vmask();
 170         lmask = ra_lmask();
 171         if (ra_regscn(gmask & ~vmask & ~lmask) >= 0)
 172                 return ra_regscn(gmask & ~vmask & ~lmask);
 173         if (ra_regscn(amask & ~vmask & ~lmask) >= 0)
 174                 return ra_regscn(amask & ~vmask & ~lmask);
 175         if (ra_regscn(gmask) >= 0)
 176                 return ra_regscn(gmask);
 177         if (ra_regscn(amask) >= 0)
 178                 return ra_regscn(amask);
 179         die("neatcc: cannot allocate an acceptable register\n");
 180         return 0;
 181 }
 182
 183 /* find a free and cheap register */
 184 static long ra_regcheap(long mask)
 185 {
 186         return ra_regscn(mask & (func_regs | (R_TMPS & ~R_PERM)) &
 187                         ~ra_lmask() & ~ra_vmask());
 188 }
 189
 190 /* allocate registers for the current instruction */
 191 static void ra_map(int *rd, int *r1, int *r2, int *r3, long *mt)
 192 {
 193         long md, m1, m2, m3;
 194         struct ic *c = &ic[ic_i];
 195         long all = 0;
 196         int n = ic_regcnt(c);
 197         int oc = O_C(c->op);
 198         int i;
 199         *rd = -1;
 200         *r1 = -1;
 201         *r2 = -1;
 202         *r3 = -1;
 203         *mt = 0;
 204         /* optimizing loading locals: point to local's register */
 205         if (oc == (O_LD | O_LOC) && ra_lreg(c->a1) >= 0 &&
 206                         ra_vmap[ra_lreg(c->a1)] < 0) {
 207                 *rd = ra_lreg(c->a1);
 208                 func_regs |= 1 << *rd;
 209                 return;
 210         }
 211         /* do not use argument registers to hold call destination */
 212         if (oc & O_CALL)
 213                 for (i = 0; i < MIN(c->a3, N_ARGS); i++)
 214                         all |= (1 << argregs[i]);
 215         /* instructions on locals can be simplified */
 216         if (oc & O_LOC) {
 217                 if (oc & O_MOV)
 218                         oc = O_ADD | O_NUM;
 219                 if (oc & (O_ST | O_LD))
 220                         oc = (oc & ~O_LOC) & O_NUM;
 221         }
 222         if (i_reg(c->op, &md, &m1, &m2, &m3, mt))
 223                 die("neatcc: instruction %08lx not supported\n", c->op);
 224         /*
 225          * the registers used in global register allocation should not
 226          * be used in the last instruction of a basic block.
 227          */
 228         if (c->op & (O_JZ | O_JCC))
 229                 for (i = 0; i < LEN(ra_lmap); i++)
 230                         if (reg_rmap(ic_i, i) >= 0 && ra_lmap[i] != reg_rmap(ic_i, i))
 231                                 all |= (1 << i);
 232         /* allocating registers for the operands */
 233         if (n >= 2) {
 234                 *r2 = ra_regget(c->a2, m2, m2, all);
 235                 all |= (1 << *r2);
 236         }
 237         if (n >= 1) {
 238                 *r1 = ra_regget(c->a1, m1, m1, all);
 239                 all |= (1 << *r1);
 240         }
 241         if (n >= 3) {
 242                 *r3 = ra_regget(c->a3, m3, m3, all);
 243                 all |= (1 << *r3);
 244         }
 245         if (c->op & O_OUT) {
 246                 long m4 = (md ? md : m1) & ~all;
 247                 long a4 = md ? ic_i : c->a1;
 248                 if (n >= 1 && !md)
 249                         *rd = *r1;
 250                 else if (n >= 2 && md & (1 << *r2) && ic_luse[*r2] <= ic_i)
 251                         *rd = *r2;
 252                 else if (n >= 1 && md & (1 << *r1) && ic_luse[*r1] <= ic_i)
 253                         *rd = *r1;
 254                 else
 255                         *rd = ra_regget(ic_i, ra_gmask[ic_i], md, 0);
 256                 /* if overwriting a local, use another register */
 257                 if (ra_lmap[*rd] >= 0)
 258                         if (m4 & ~(1 << *rd))
 259                                 *rd = ra_regget(a4, ra_gmask[ic_i], m4 & ~(1 << *rd), 0);
 260                 if (n >= 1 && !md)
 261                         *r1 = *rd;
 262                 all |= (1 << *rd);
 263         }
 264         func_regs |= all | *mt;
 265 }
 266
 267 static long iv_rank(long iv)
 268 {
 269         int i;
 270         for (i = 0; i < LEN(ra_live); i++)
 271                 if (ra_live[i] == iv)
 272                         return i;
 273         die("neatcc: the specified value is not live\n");
 274         return 0;
 275 }
 276
 277 static long iv_addr(long rank)
 278 {
 279         return loc_pos + rank * ULNG + ULNG;
 280 }
 281
 282 static void loc_toreg(long loc, long off, int reg, int bt)
 283 {
 284         loc_mem[loc]++;
 285         i_ins(O_MK(O_LD | O_NUM, bt), reg, REG_FP, -loc_off[loc] + off, 0);
 286 }
 287
 288 static void loc_tomem(long loc, long off, int reg, int bt)
 289 {
 290         loc_mem[loc]++;
 291         i_ins(O_MK(O_ST | O_NUM, bt), 0, reg, REG_FP, -loc_off[loc] + off);
 292 }
 293
 294 static void loc_toadd(long loc, long off, int reg)
 295 {
 296         loc_mem[loc]++;
 297         i_ins(O_ADD | O_NUM, reg, REG_FP, -loc_off[loc] + off, 0);
 298 }
 299
 300 /* return nonzero if the local is read at least once in this basic block */
 301 static int loc_isread(long loc)
 302 {
 303         int i;
 304         for (i = ic_i + 1; i < ic_n && !ic_bbeg[i]; i++)
 305                 if (ic[i].op & O_LOC)
 306                         return ic[i].op & O_LD;
 307         return 0;
 308 }
 309
 310 static void val_toreg(long val, int reg)
 311 {
 312         i_ins(O_MK(O_LD | O_NUM, ULNG), reg, REG_FP, -iv_addr(iv_rank(val)), 0);
 313 }
 314
 315 static void val_tomem(long val, int reg)
 316 {
 317         long rank = iv_rank(val);
 318         ra_vmax = MAX(ra_vmax, rank + 1);
 319         i_ins(O_MK(O_ST | O_NUM, ULNG), 0, reg, REG_FP, -iv_addr(rank));
 320 }
 321
 322 /* move the value to the stack */
 323 static void ra_spill(int reg)
 324 {
 325         if (ra_vmap[reg] >= 0) {
 326                 val_tomem(ra_vmap[reg], reg);
 327                 ra_vmap[reg] = -1;
 328         }
 329         if (ra_lmap[reg] >= 0) {
 330                 if (ra_lmap[reg] == reg_rmap(ic_i, reg))
 331                         loc_tomem(ra_lmap[reg], 0, reg, ULNG);
 332                 ra_lmap[reg] = -1;
 333         }
 334 }
 335
 336 /* set the value to the given register */
 337 static void ra_vsave(long iv, int reg)
 338 {
 339         int i;
 340         ra_vmap[reg] = iv;
 341         for (i = 0; i < LEN(ra_live); i++)
 342                 if (ra_live[i] < 0)
 343                         break;
 344         if (i == LEN(ra_live))
 345                 die("neatcc: too many live values\n");
 346         ra_live[i] = iv;
 347 }
 348
 349 /* load the value into a register */
 350 static void ra_vload(long iv, int reg)
 351 {
 352         if (ra_vmap[reg] == iv)
 353                 return;
 354         if (ra_vmap[reg] >= 0 || ra_lmap[reg] >= 0)
 355                 ra_spill(reg);
 356         if (ra_vreg(iv) >= 0) {
 357                 i_ins(O_MK(O_MOV, ULNG), reg, ra_vreg(iv), 0, 0);
 358                 ra_vmap[ra_vreg(iv)] = -1;
 359         } else {
 360                 val_toreg(iv, reg);
 361         }
 362         ra_vmap[reg] = iv;
 363 }
 364
 365 /* the value is no longer needed */
 366 static void ra_vdrop(long iv)
 367 {
 368         int i;
 369         for (i = 0; i < LEN(ra_live); i++)
 370                 if (ra_live[i] == iv)
 371                         ra_live[i] = -1;
 372         if (ra_vreg(iv) >= 0)
 373                 ra_vmap[ra_vreg(iv)] = -1;
 374 }
 375
 376 /* move the given value to memory or a free register */
 377 static void ra_vmove(long iv, long mask)
 378 {
 379         int src = ra_vreg(iv);
 380         int dst = ra_regcheap(mask);
 381         if (dst >= 0 && ra_vmap[dst] < 0 && ra_lmap[dst] < 0) {
 382                 i_ins(O_MK(O_MOV, ULNG), dst, src, 0, 0);
 383                 ra_vmap[dst] = iv;
 384         } else {
 385                 val_tomem(iv, src);
 386         }
 387         ra_vmap[src] = -1;
 388 }
 389
 390 /* load the value of local loc into register reg */
 391 static void ra_lload(long loc, long off, int reg, int bt)
 392 {
 393         int greg = reg_lmap(ic_i, loc);
 394         if (greg >= 0) {
 395                 int lreg = ra_lreg(loc);
 396                 /* values using the same register */
 397                 if (ra_vmap[greg] >= 0)
 398                         ra_vmove(ra_vmap[greg],
 399                                 ra_gmask[ra_vmap[greg]] & ~(1 << reg));
 400                 if (lreg >= 0) {
 401                         ra_lmap[lreg] = -1;
 402                         if (lreg != greg)
 403                                 i_ins(O_MK(O_MOV, bt), greg, lreg, 0, 0);
 404                 } else {
 405                         loc_toreg(loc, off, greg, bt);
 406                 }
 407                 ra_lmap[greg] = loc;
 408         }
 409         if (ra_lreg(loc) < 0 && reg_safe(loc) && loc_isread(loc)) {
 410                 ra_lmap[reg] = loc;
 411                 loc_toreg(loc, off, reg, bt);
 412         }
 413         if (ra_lreg(loc) >= 0) {
 414                 if (ra_lreg(loc) != reg)
 415                         i_ins(O_MK(O_MOV, bt), reg, ra_lreg(loc), 0, 0);
 416         } else {
 417                 loc_toreg(loc, off, reg, bt);
 418         }
 419 }
 420
 421 /* register reg contains the value of local loc */
 422 static void ra_lsave(long loc, long off, int reg, int bt)
 423 {
 424         int lreg = ra_lreg(loc);
 425         int greg = reg_lmap(ic_i, loc);
 426         if (greg >= 0) {
 427                 if (lreg >= 0 && lreg != greg)
 428                         ra_lmap[lreg] = -1;
 429                 if (ra_vmap[greg] >= 0) /* values using the same register */
 430                         ra_vmove(ra_vmap[greg],
 431                                 ra_gmask[ra_vmap[greg]] & ~(1 << reg));
 432                 i_ins(O_MK(O_MOV, bt), greg, reg, 0, 0);
 433                 ra_lmap[greg] = loc;
 434         } else {
 435                 if (lreg >= 0)
 436                         ra_lmap[lreg] = -1;
 437                 loc_tomem(loc, off, reg, bt);
 438                 if (ra_lmap[reg] < 0 && reg_safe(loc) && loc_isread(loc))
 439                         ra_lmap[reg] = loc;
 440         }
 441 }
 442
 443 /* end of a basic block */
 444 static void ra_bbend(void)
 445 {
 446         int i;
 447         /* save values to memory */
 448         for (i = 0; i < LEN(ra_vmap); i++)
 449                 if (ra_vmap[i] >= 0)
 450                         ra_spill(i);
 451         /* dropping local caches */
 452         for (i = 0; i < LEN(ra_lmap); i++)
 453                 if (ra_lmap[i] != reg_rmap(ic_i, i) && ra_lmap[i] >= 0)
 454                         ra_spill(i);
 455         /* load global register allocations from memory */
 456         for (i = 0; i < LEN(ra_lmap); i++) {
 457                 if (ra_lmap[i] != reg_rmap(ic_i, i)) {
 458                         ra_lmap[i] = reg_rmap(ic_i, i);
 459                         loc_toreg(ra_lmap[i], 0, i, ULNG);
 460                 }
 461         }
 462 }
 463
 464 static void ra_init(struct ic *ic, long ic_n)
 465 {
 466         long md, m1, m2, m3, mt;
 467         int i, j;
 468         ic_bbeg = calloc(ic_n, sizeof(ic_bbeg[0]));
 469         ra_gmask = calloc(ic_n, sizeof(ra_gmask[0]));
 470         loc_mem = calloc(loc_n, sizeof(loc_mem[0]));
 471         /* ic_bbeg */
 472         for (i = 0; i < ic_n; i++) {
 473                 if (i + 1 < ic_n && ic[i].op & (O_JXX | O_RET))
 474                         ic_bbeg[i + 1] = 1;
 475                 if (ic[i].op & O_JXX && ic[i].a3 < ic_n)
 476                         ic_bbeg[ic[i].a3] = 1;
 477         }
 478         /* ra_gmask */
 479         for (i = 0; i < ic_n; i++) {
 480                 int n = ic_regcnt(ic + i);
 481                 int op = ic[i].op;
 482                 i_reg(op, &md, &m1, &m2, &m3, &mt);
 483                 if (n >= 1)
 484                         ra_gmask[ic[i].a1] = m1;
 485                 if (n >= 2)
 486                         ra_gmask[ic[i].a2] = m2;
 487                 if (n >= 3)
 488                         ra_gmask[ic[i].a3] = m3;
 489                 if (O_C(op) == (O_ST | O_LOC) && reg_lmap(i, ic[i].a2))
 490                         ra_gmask[ic[i].a1] = 1 << reg_lmap(i, ic[i].a2);
 491                 if (op & O_CALL)
 492                         for (j = 0; j < MIN(N_ARGS, ic[i].a3); j++)
 493                                 ra_gmask[ic[i].args[j]] = 1 << argregs[j];
 494         }
 495         /* ra_vmap */
 496         for (i = 0; i < LEN(ra_vmap); i++)
 497                 ra_vmap[i] = -1;
 498         /* ra_lmap */
 499         for (i = 0; i < LEN(ra_lmap); i++)
 500                 ra_lmap[i] = reg_rmap(0, i);
 501         func_regs |= reg_mask();
 502         /* ra_live */
 503         for (i = 0; i < LEN(ra_live); i++)
 504                 ra_live[i] = -1;
 505         ra_vmax = 0;
 506         func_maxargs = 0;
 507 }
 508
 509 static void ra_done(void)
 510 {
 511         free(ic_bbeg);
 512         free(ra_gmask);
 513         free(loc_mem);
 514 }
 515
 516 static void ic_gencode(struct ic *ic, long ic_n)
 517 {
 518         int rd, r1, r2, r3;
 519         long mt;
 520         int i, j;
 521         /* loading arguments in their allocated registers */
 522         for (i = 0; i < LEN(ra_lmap); i++) {
 523                 int loc = ra_lmap[i];
 524                 if (loc >= 0 && loc < func_argc)
 525                         if (loc >= N_ARGS || i != argregs[loc])
 526                                 loc_toreg(loc, 0, i, ULNG);
 527         }
 528         /* generating code */
 529         for (i = 0; i < ic_n; i++) {
 530                 long op = ic[i].op;
 531                 long oc = O_C(op);
 532                 int n = ic_regcnt(ic + i);
 533                 ic_i = i;
 534                 i_label(i);
 535                 ra_map(&rd, &r1, &r2, &r3, &mt);
 536                 if (oc & O_CALL) {
 537                         int argc = ic[i].a3;
 538                         int aregs = MIN(N_ARGS, argc);
 539                         /* arguments passed via stack */
 540                         for (j = argc - 1; j >= aregs; --j) {
 541                                 int v = ic[i].args[j];
 542                                 int rx = ra_vreg(v) >= 0 ? ra_vreg(v) : rd;
 543                                 ra_vload(v, rx);
 544                                 i_ins(O_MK(O_ST | O_NUM, ULNG), 0,
 545                                         rx, REG_SP, (j - aregs) * ULNG);
 546                                 ra_vdrop(v);
 547                         }
 548                         func_maxargs = MAX(func_maxargs, argc - aregs);
 549                         /* arguments passed via registers */
 550                         for (j = aregs - 1; j >= 0; --j)
 551                                 ra_vload(ic[i].args[j], argregs[j]);
 552                 }
 553                 /* loading the operands */
 554                 if (n >= 1)
 555                         ra_vload(ic[i].a1, r1);
 556                 if (n >= 2)
 557                         ra_vload(ic[i].a2, r2);
 558                 if (n >= 3)
 559                         ra_vload(ic[i].a3, r3);
 560                 /* dropping values that are no longer used */
 561                 for (j = 0; j < LEN(ra_live); j++)
 562                         if (ra_live[j] >= 0 && ic_luse[ra_live[j]] <= i)
 563                                 ra_vdrop(ra_live[j]);
 564                 /* saving values stored in registers that may change */
 565                 for (j = 0; j < N_REGS; j++)
 566                         if (mt & (1 << j))
 567                                 ra_spill(j);
 568                 /* overwriting a value that is needed later (unless loading a local to its register) */
 569                 if (oc & O_OUT)
 570                         if (oc != (O_LD | O_LOC) || ra_lmap[rd] != ic[i].a1 ||
 571                                         ra_vmap[rd] >= 0)
 572                                 ra_spill(rd);
 573                 /* before the last instruction of a basic block; for jumps */
 574                 if (i + 1 < ic_n && ic_bbeg[i + 1] && oc & O_JXX)
 575                         ra_bbend();
 576                 /* performing the instruction */
 577                 if (oc & O_BOP)
 578                         i_ins(op, rd, r1, oc & O_NUM ? ic[i].a2 : r2, 0);
 579                 if (oc & O_UOP)
 580                         i_ins(op, rd, r1, r2, 0);
 581                 if (oc == (O_LD | O_NUM))
 582                         i_ins(op, rd, r1, ic[i].a2, 0);
 583                 if (oc == (O_LD | O_LOC))
 584                         ra_lload(ic[i].a1, ic[i].a2, rd, O_T(op));
 585                 if (oc == (O_ST | O_NUM))
 586                         i_ins(op, 0, r1, r2, ic[i].a3);
 587                 if (oc == (O_ST | O_LOC))
 588                         ra_lsave(ic[i].a2, ic[i].a3, r1, O_T(op));
 589                 if (oc == O_RET)
 590                         i_ins(op, 0, r1, 0, 0);
 591                 if (oc == O_MOV)
 592                         i_ins(op, rd, r1, 0, 0);
 593                 if (oc == (O_MOV | O_NUM))
 594                         i_ins(op, rd, ic[i].a1, 0, 0);
 595                 if (oc == (O_MOV | O_LOC))
 596                         loc_toadd(ic[i].a1, ic[i].a2, rd);
 597                 if (oc == (O_MOV | O_SYM))
 598                         i_ins(op, rd, ic[i].a1, ic[i].a2, 0);
 599                 if (oc == O_CALL)
 600                         i_ins(op, rd, r1, 0, 0);
 601                 if (oc == (O_CALL | O_SYM))
 602                         i_ins(op, rd, ic[i].a1, ic[i].a2, 0);
 603                 if (oc == O_JMP)
 604                         i_ins(op, 0, 0, 0, ic[i].a3);
 605                 if (oc & O_JZ)
 606                         i_ins(op, 0, r1, 0, ic[i].a3);
 607                 if (oc & O_JCC)
 608                         i_ins(op, 0, r1, oc & O_NUM ? ic[i].a2 : r2, ic[i].a3);
 609                 if (oc == O_MSET)
 610                         i_ins(op, 0, r1, r2, r3);
 611                 if (oc == O_MCPY)
 612                         i_ins(op, 0, r1, r2, r3);
 613                 /* saving back the output register */
 614                 if (oc & O_OUT && ic_luse[i] > i)
 615                         ra_vsave(ic_i, rd);
 616                 /* after the last instruction of a basic block */
 617                 if (i + 1 < ic_n && ic_bbeg[i + 1] && !(oc & O_JXX))
 618                         ra_bbend();
 619         }
 620         i_label(ic_n);
 621 }
 622
 623 static void ic_reset(void)
 624 {
 625         o_tmpdrop(-1);
 626         o_back(0);
 627         free(loc_off);
 628         loc_off = NULL;
 629         loc_n = 0;
 630         loc_sz = 0;
 631         loc_pos = I_LOC0;
 632 }
 633
 634 void o_func_beg(char *name, int argc, int global, int varg)
 635 {
 636         int i;
 637         func_argc = argc;
 638         func_varg = varg;
 639         func_regs = 0;
 640         ic_reset();
 641         for (i = 0; i < argc; i++)
 642                 loc_add(I_ARG0 + -i * ULNG);
 643         out_def(name, (global ? OUT_GLOB : 0) | OUT_CS, mem_len(&cs), 0);
 644 }
 645
 646 void o_code(char *name, char *c, long c_len)
 647 {
 648         out_def(name, OUT_CS, mem_len(&cs), 0);
 649         mem_put(&cs, c, c_len);
 650 }
 651
 652 void o_func_end(void)
 653 {
 654         long spsub;
 655         long sargs = 0;
 656         long sargs_last = -1;
 657         long sregs_pos;
 658         char *c;
 659         long c_len, *rsym, *rflg, *roff, rcnt;
 660         int leaf = 1;
 661         int locs = 0;                   /* accessing locals on the stack */
 662         int i;
 663         ic_get(&ic, &ic_n);             /* the intermediate code */
 664         reg_init(ic, ic_n);             /* global register allocation */
 665         ra_init(ic, ic_n);              /* initialize register allocation */
 666         ic_luse = ic_lastuse(ic, ic_n);
 667         ic_gencode(ic, ic_n);           /* generating machine code */
 668         free(ic_luse);
 669         /* deciding which arguments to save */
 670         for (i = 0; i < func_argc; i++)
 671                 if (loc_mem[i])
 672                         sargs_last = i + 1;
 673         for (i = 0; i < N_ARGS && (func_varg || i < sargs_last); i++)
 674                 sargs |= 1 << argregs[i];
 675         /* computing the amount of stack subtraction */
 676         for (i = 0; i < loc_n; i++)
 677                 if (loc_mem[i])
 678                         locs = 1;
 679         spsub = (locs || ra_vmax) ? loc_pos + ra_vmax * ULNG : 0;
 680         for (i = 0; i < N_TMPS; i++)
 681                 if (((1 << tmpregs[i]) & func_regs & R_PERM) != 0)
 682                         spsub += ULNG;
 683         sregs_pos = spsub;
 684         spsub += func_maxargs * ULNG;
 685         /* leaf functions */
 686         for (i = 0; i < ic_n; i++)
 687                 if (ic[i].op & O_CALL)
 688                         leaf = 0;
 689         /* adding function prologue and epilogue */
 690         i_wrap(func_argc, sargs, spsub, spsub || locs || !leaf,
 691                 func_regs & R_PERM, -sregs_pos);
 692         ra_done();
 693         i_code(&c, &c_len, &rsym, &rflg, &roff, &rcnt);
 694         for (i = 0; i < rcnt; i++)      /* adding the relocations */
 695                 out_rel(rsym[i], rflg[i], roff[i] + mem_len(&cs));
 696         mem_put(&cs, c, c_len);         /* appending function code */
 697         free(c);
 698         free(rsym);
 699         free(rflg);
 700         free(roff);
 701         for (i = 0; i < ic_n; i++)
 702                 ic_free(&ic[i]);
 703         free(ic);
 704         reg_done();
 705         ic_reset();
 706 }
 707
 708 void o_write(int fd)
 709 {
 710         i_done();
 711         out_write(fd, mem_buf(&cs), mem_len(&cs), mem_buf(&ds), mem_len(&ds));
 712         free(loc_off);
 713         free(ds_name);
 714         free(ds_off);
 715         mem_done(&cs);
 716         mem_done(&ds);
 717 }