gen.c

   1 #include <stdlib.h>
   2 #include <stdio.h>
   3 #include <string.h>
   4 #include "gen.h"
   5 #include "out.h"
   6 #include "tok.h"
   7
   8 #define LOC_REG         0x01
   9 #define LOC_MEM         0x02
  10 #define LOC_NUM         0x04
  11 #define LOC_SYM         0x08
  12 #define LOC_LOCAL       0x10
  13
  14 #define NREGS           16
  15
  16 #define REG_PC          15      /* program counter */
  17 #define REG_LR          14      /* link register */
  18 #define REG_SP          13      /* stack pointer */
  19 #define REG_TMP         12      /* temporary register */
  20 #define REG_FP          11      /* frame pointer register */
  21 #define REG_DP          10      /* data pointer register */
  22 #define REG_RET         0       /* returned value register */
  23
  24 #define MIN(a, b)               ((a) < (b) ? (a) : (b))
  25 #define ALIGN(x, a)             (((x) + (a) - 1) & ~((a) - 1))
  26
  27 static char cs[SECSIZE];        /* code segment */
  28 static int cslen;
  29 static char ds[SECSIZE];        /* data segment */
  30 static int dslen;
  31 static long bsslen;             /* bss segment size */
  32
  33 static int nogen;               /* don't generate code */
  34 static long sp;
  35 static long func_beg;
  36 static long maxsp;
  37
  38 #define TMP(i)          (((i) < ntmp) ? &tmps[ntmp - 1 - (i)] : NULL)
  39
  40 static struct tmp {
  41         long addr;
  42         char sym[NAMELEN];
  43         long off;       /* offset from a symbol or a local */
  44         unsigned loc;   /* variable location */
  45         unsigned bt;    /* type of address; zero when not a pointer */
  46 } tmps[MAXTMP];
  47 static int ntmp;
  48
  49 static int tmpsp;
  50
  51 static struct tmp *regs[NREGS];
  52 static int tmpregs[] = {4, 5, 6, 7, 8, 9, 0, 1, 2, 3};
  53 static int argregs[] = {0, 1, 2, 3};
  54
  55 #define I_AND           0x00
  56 #define I_EOR           0x01
  57 #define I_SUB           0x02
  58 #define I_RSB           0x03
  59 #define I_ADD           0x04
  60 #define I_TST           0x08
  61 #define I_CMP           0x0a
  62 #define I_ORR           0x0c
  63 #define I_MOV           0x0d
  64 #define I_MVN           0x0f
  65
  66 #define MAXRET                  (1 << 8)
  67
  68 static long ret[MAXRET];
  69 static int nret;
  70
  71 /* output div/mod functions */
  72 static int putdiv = 0;
  73
  74 static void os(void *s, int n)
  75 {
  76         memcpy(cs + cslen, s, n);
  77         cslen += n;
  78 }
  79
  80 static void oi(long n)
  81 {
  82         if (nogen)
  83                 return;
  84         *(int *) (cs + cslen) = n;
  85         cslen += 4;
  86 }
  87
  88 #define MAXNUMS         1024
  89
  90 /* data pool */
  91 static long num_offs[MAXNUMS];                  /* data immediate value */
  92 static char num_names[MAXNUMS][NAMELEN];        /* relocation data symbol name */
  93 static int nums;
  94
  95 static int pool_find(char *name, int off)
  96 {
  97         int i;
  98         for (i = 0; i < nums; i++)
  99                 if (!strcmp(name, num_names[i]) && off == num_offs[i])
 100                         return i;
 101         return -1;
 102 }
 103
 104 static int pool_num(long num)
 105 {
 106         int idx = pool_find("", num);
 107         if (idx < 0) {
 108                 idx = nums++;
 109                 num_offs[idx] = num;
 110                 num_names[idx][0] = '\0';
 111         }
 112         return idx << 2;
 113 }
 114
 115 static int pool_reloc(char *name, long off)
 116 {
 117         int idx = pool_find(name, off);
 118         if (idx < 0) {
 119                 idx = nums++;
 120                 num_offs[idx] = off;
 121                 strcpy(num_names[idx], name);
 122         }
 123         return idx << 2;
 124 }
 125
 126 static void pool_write(void)
 127 {
 128         int i;
 129         for (i = 0; i < nums; i++) {
 130                 if (num_names[i])
 131                         out_rel(num_names[i], OUT_CS, cslen);
 132                 oi(num_offs[i]);
 133         }
 134 }
 135
 136 /*
 137  * data processing:
 138  * +---------------------------------------+
 139  * |COND|00|I| op |S| Rn | Rd |  operand2  |
 140  * +---------------------------------------+
 141  *
 142  * S: set condition code
 143  * Rn: first operand
 144  * Rd: destination operand
 145  *
 146  * I=0 operand2=| shift  | Rm |
 147  * I=1 operand2=|rota|  imm   |
 148  */
 149 #define ADD(op, rd, rn, s, i, cond)     \
 150         (((cond) << 28) | ((i) << 25) | ((s) << 20) | \
 151                 ((op) << 21) | ((rn) << 16) | ((rd) << 12))
 152
 153 static int add_encimm(unsigned n)
 154 {
 155         int i = 0;
 156         while (i < 12 && (n >> ((4 + i) << 1)))
 157                 i++;
 158         return (n >> (i << 1)) | (((16 - i) & 0x0f) << 8);
 159 }
 160
 161 static unsigned add_decimm(int n)
 162 {
 163         int rot = (16 - ((n >> 8) & 0x0f)) & 0x0f;
 164         return (n & 0xff) << (rot << 1);
 165 }
 166
 167 static int add_rndimm(unsigned n)
 168 {
 169         int rot = (n >> 8) & 0x0f;
 170         int num = n & 0xff;
 171         if (rot == 0)
 172                 return n;
 173         if (num == 0xff) {
 174                 num = 0;
 175                 rot = (rot + 12) & 0x0f;
 176         }
 177         return ((num + 1) & 0xff) | (rot << 8);
 178 }
 179
 180 static void i_add(int op, int rd, int rn, int rm)
 181 {
 182         oi(ADD(op, rd, rn, 0, 0, 14) | rm);
 183 }
 184
 185 static void i_add_imm(int op, int rd, int rn, long n)
 186 {
 187         oi(ADD(op, rd, rn, 0, 1, 14) | add_encimm(n));
 188 }
 189
 190 static void i_ldr(int l, int rd, int rn, int off, int bt);
 191
 192 static void i_num(int rd, long n)
 193 {
 194         int enc = add_encimm(n);
 195         if (n == add_decimm(enc)) {
 196                 oi(ADD(I_MOV, rd, 0, 0, 1, 14) | enc);
 197                 return;
 198         }
 199         enc = add_encimm(-n - 1);
 200         if (~n == add_decimm(enc)) {
 201                 oi(ADD(I_MVN, rd, 0, 0, 1, 14) | enc);
 202                 return;
 203         }
 204         if (!nogen) {
 205                 int off = pool_num(n);
 206                 i_ldr(1, rd, REG_DP, off, LONGSZ);
 207         }
 208 }
 209
 210 static void i_add_anyimm(int rd, int rn, long n)
 211 {
 212         int neg = n < 0;
 213         int imm = add_encimm(neg ? -n : n);
 214         if (imm == add_decimm(neg ? -n : n)) {
 215                 oi(ADD(neg ? I_SUB : I_ADD, rd, rn, 0, 1, 14) | imm);
 216         } else {
 217                 i_num(rd, n);
 218                 i_add(I_ADD, rd, rd, rn);
 219         }
 220 }
 221
 222 /*
 223  * multiply
 224  * +----------------------------------------+
 225  * |COND|000000|A|S| Rd | Rn | Rs |1001| Rm |
 226  * +----------------------------------------+
 227  *
 228  * Rd: destination
 229  * A: accumulate
 230  * C: set condition codes
 231  *
 232  * I=0 operand2=| shift  | Rm |
 233  * I=1 operand2=|rota|  imm   |
 234  */
 235 #define MUL(rd, rn, rs)         \
 236         ((14 << 28) | ((rd) << 16) | ((0) << 12) | ((rn) << 8) | ((9) << 4) | (rm))
 237
 238 static void i_mul(int rd, int rn, int rm)
 239 {
 240         oi(MUL(rd, rn, rm));
 241 }
 242
 243 static void i_cmp(int op, int rn, int rm)
 244 {
 245         oi(ADD(op, 0, rn, 1, 0, 14) | rm);
 246 }
 247
 248 static void i_cmp_imm(int op, int rn, long n)
 249 {
 250         oi(ADD(op, 0, rn, 1, 1, 14) | add_encimm(n));
 251 }
 252
 253 static void i_set(int cond, int rd)
 254 {
 255         oi(ADD(I_MOV, rd, 0, 0, 1, 14));
 256         oi(ADD(I_MOV, rd, 0, 0, 1, cond) | 1);
 257 }
 258
 259 #define SM_LSL          0
 260 #define SM_LSR          1
 261 #define SM_ASR          2
 262
 263 static void i_shl(int sm, int rd, int rm, int rs)
 264 {
 265         oi(ADD(I_MOV, rd, 0, 0, 0, 14) | (rs << 8) | (sm << 5) | (1 << 4) | rm);
 266 }
 267
 268 static void i_shl_imm(int sm, int rd, int rn, long n)
 269 {
 270         oi(ADD(I_MOV, rd, 0, 0, 0, 14) | (n << 7) | (sm << 5) | rn);
 271 }
 272
 273 static void i_mov(int op, int rd, int rn)
 274 {
 275         oi(ADD(op, rd, 0, 0, 0, 14) | rn);
 276 }
 277
 278 /*
 279  * single data transfer:
 280  * +------------------------------------------+
 281  * |COND|01|I|P|U|B|W|L| Rn | Rd |   offset   |
 282  * +------------------------------------------+
 283  *
 284  * I: immediate/offset
 285  * P: post/pre indexing
 286  * U: down/up
 287  * B: byte/word
 288  * W: writeback
 289  * L: store/load
 290  * Rn: base register
 291  * Rd: source/destination register
 292  *
 293  * I=0 offset=| immediate |
 294  * I=1 offset=| shift  | Rm |
 295  *
 296  * halfword and signed data transfer
 297  * +----------------------------------------------+
 298  * |COND|000|P|U|0|W|L| Rn | Rd |0000|1|S|H|1| Rm |
 299  * +----------------------------------------------+
 300  *
 301  * +----------------------------------------------+
 302  * |COND|000|P|U|1|W|L| Rn | Rd |off1|1|S|H|1|off2|
 303  * +----------------------------------------------+
 304  *
 305  * S: singed
 306  * H: halfword
 307  */
 308 #define LDR(l, rd, rn, b, u, p, w)              \
 309         ((14 << 28) | (1 << 26) | ((p) << 24) | ((b) << 22) | ((u) << 23) | \
 310         ((w) << 21) | ((l) << 20) | ((rn) << 16) | ((rd) << 12))
 311 #define LDRH(l, rd, rn, s, h, u, i)     \
 312         ((14 << 28) | (1 << 24) | ((u) << 23) | ((i) << 22) | ((l) << 20) | \
 313         ((rn) << 16) | ((rd) << 12) | ((s) << 6) | ((h) << 5) | (9 << 4))
 314
 315 static void i_ldr(int l, int rd, int rn, int off, int bt)
 316 {
 317         int b = BT_SZ(bt) == 1;
 318         int h = BT_SZ(bt) == 2;
 319         int s = l && (bt & BT_SIGNED);
 320         int half = h || (b && s);
 321         int maximm = half ? 0x100 : 0x1000;
 322         int neg = off < 0;
 323         if (neg)
 324                 off = -off;
 325         while (off >= maximm) {
 326                 int imm = add_encimm(off);
 327                 oi(ADD(neg ? I_SUB : I_ADD, REG_TMP, rn, 0, 1, 14) | imm);
 328                 rn = REG_TMP;
 329                 off -= add_decimm(imm);
 330         }
 331         if (!half)
 332                 oi(LDR(l, rd, rn, b, !neg, 1, 0) | off);
 333         else
 334                 oi(LDRH(l, rd, rn, s, h, !neg, 1) |
 335                         ((off & 0xf0) << 4) | (off & 0x0f));
 336 }
 337
 338 static void i_sym(int rd, char *sym, int off)
 339 {
 340         if (!nogen) {
 341                 int doff = pool_reloc(sym, off);
 342                 i_ldr(1, rd, REG_DP, doff, LONGSZ);
 343         }
 344 }
 345
 346 static void i_neg(int rd)
 347 {
 348         oi(ADD(I_RSB, rd, rd, 0, 1, 14));
 349 }
 350
 351 static void i_not(int rd)
 352 {
 353         oi(ADD(I_MVN, rd, 0, 0, 0, 14) | rd);
 354 }
 355
 356 static void i_lnot(int rd)
 357 {
 358         i_cmp(I_TST, rd, rd);
 359         oi(ADD(I_MOV, rd, 0, 0, 1, 14));
 360         oi(ADD(I_MOV, rd, 0, 0, 1, 0) | 1);
 361 }
 362
 363 /* rd = rd & ((1 << bits) - 1) */
 364 static void i_zx(int rd, int bits)
 365 {
 366         if (bits <= 8) {
 367                 oi(ADD(I_AND, rd, rd, 0, 1, 14) | add_encimm((1 << bits) - 1));
 368         } else {
 369                 i_shl_imm(SM_LSL, rd, rd, 32 - bits);
 370                 i_shl_imm(SM_LSR, rd, rd, 32 - bits);
 371         }
 372 }
 373
 374 static void i_sx(int rd, int bits)
 375 {
 376         i_shl_imm(SM_LSL, rd, rd, 32 - bits);
 377         i_shl_imm(SM_ASR, rd, rd, 32 - bits);
 378 }
 379
 380 /*
 381  * branch:
 382  * +-----------------------------------+
 383  * |COND|101|L|         offset         |
 384  * +-----------------------------------+
 385  *
 386  * L: link
 387  */
 388 #define BL(cond, l, o)  (((cond) << 28) | (5 << 25) | ((l) << 24) | \
 389                                 ((((o) - 8) >> 2) & 0x00ffffff))
 390
 391 static void i_b(long addr)
 392 {
 393         oi(BL(14, 0, addr - cslen));
 394 }
 395
 396 static void i_b_if(long addr, int rn, int z)
 397 {
 398         i_cmp(I_TST, rn, rn);
 399         oi(BL(z ? 0 : 1, 0, addr - cslen));
 400 }
 401
 402 static void i_b_fill(long *dst, int diff)
 403 {
 404         *dst = (*dst & 0xff000000) | (((diff - 8) >> 2) & 0x00ffffff);
 405 }
 406
 407 static void i_memcpy(int rd, int rs, int rn)
 408 {
 409         oi(ADD(I_SUB, rn, rn, 1, 1, 14) | 1);
 410         oi(BL(4, 0, 16));
 411         oi(LDR(1, REG_TMP, rs, 1, 1, 0, 0) | 1);
 412         oi(LDR(0, REG_TMP, rd, 1, 1, 0, 0) | 1);
 413         oi(BL(14, 0, -16));
 414 }
 415
 416 static void i_memset(int rd, int rs, int rn)
 417 {
 418         oi(ADD(I_SUB, rn, rn, 1, 1, 14) | 1);
 419         oi(BL(4, 0, 12));
 420         oi(LDR(0, rs, rd, 1, 1, 0, 0) | 1);
 421         oi(BL(14, 0, -12));
 422 }
 423
 424 static void i_call_reg(int rd)
 425 {
 426         i_mov(I_MOV, REG_LR, REG_PC);
 427         i_mov(I_MOV, REG_PC, rd);
 428 }
 429
 430 static void i_call(char *sym)
 431 {
 432         if (!nogen)
 433                 out_rel(sym, OUT_CS | OUT_REL24, cslen);
 434         oi(BL(14, 1, 0));
 435 }
 436
 437 static void i_prolog(void)
 438 {
 439         func_beg = cslen;
 440         nums = 0;
 441         oi(0xe1a0c00d);         /* mov   r12, sp */
 442         oi(0xe92d000f);         /* stmfd sp!, {r0-r3} */
 443         oi(0xe92d5ff0);         /* stmfd sp!, {r0-r11, r12, lr} */
 444         oi(0xe1a0b00d);         /* mov   fp, sp */
 445         oi(0xe24dd000);         /* sub   sp, sp, xx */
 446         oi(0xe28fa000);         /* add   dp, pc, xx */
 447 }
 448
 449 static void i_epilog(void)
 450 {
 451         int dpoff;
 452         oi(0xe89baff0);         /* ldmfd fp, {r4-r11, sp, pc} */
 453         dpoff = add_decimm(add_rndimm(add_encimm(cslen - func_beg - 28)));
 454         cslen = func_beg + dpoff + 28;
 455         maxsp = ALIGN(maxsp, 8);
 456         maxsp = add_decimm(add_rndimm(add_encimm(maxsp)));
 457         /* fill stack sub: sp = sp - xx */
 458         *(long *) (cs + func_beg + 16) |= add_encimm(maxsp);
 459         /* fill data ptr addition: dp = pc + xx */
 460         *(long *) (cs + func_beg + 20) |= add_encimm(dpoff);
 461         pool_write();
 462 }
 463
 464 static long sp_push(int size)
 465 {
 466         sp += size;
 467         if (sp > maxsp)
 468                 maxsp = sp;
 469         return sp;
 470 }
 471
 472 static void tmp_mem(struct tmp *tmp)
 473 {
 474         int src = tmp->addr;
 475         if (!(tmp->loc == LOC_REG))
 476                 return;
 477         if (tmpsp == -1)
 478                 tmpsp = sp;
 479         tmp->addr = -sp_push(LONGSZ);
 480         i_ldr(0, src, REG_FP, tmp->addr, LONGSZ);
 481         regs[src] = NULL;
 482         tmp->loc = LOC_MEM;
 483 }
 484
 485 static void num_cast(struct tmp *t, unsigned bt)
 486 {
 487         if (!(bt & BT_SIGNED) && BT_SZ(bt) != LONGSZ)
 488                 t->addr &= ((1l << (long) (BT_SZ(bt) * 8)) - 1);
 489         if (bt & BT_SIGNED && BT_SZ(bt) != LONGSZ &&
 490                                 t->addr > (1l << (BT_SZ(bt) * 8 - 1)))
 491                 t->addr = -((1l << (BT_SZ(bt) * 8)) - t->addr);
 492 }
 493
 494 static void tmp_reg(struct tmp *tmp, int dst, int deref)
 495 {
 496         int bt = tmp->bt;
 497         if (!tmp->bt)
 498                 deref = 0;
 499         if (deref)
 500                 tmp->bt = 0;
 501         if (tmp->loc == LOC_NUM) {
 502                 i_num(dst, tmp->addr);
 503                 tmp->addr = dst;
 504                 regs[dst] = tmp;
 505                 tmp->loc = LOC_REG;
 506         }
 507         if (tmp->loc == LOC_SYM) {
 508                 i_sym(dst, tmp->sym, tmp->off);
 509                 tmp->addr = dst;
 510                 regs[dst] = tmp;
 511                 tmp->loc = LOC_REG;
 512         }
 513         if (tmp->loc == LOC_REG) {
 514                 if (deref)
 515                         i_ldr(1, dst, tmp->addr, 0, bt);
 516                 else if (dst != tmp->addr)
 517                         i_mov(I_MOV, dst, tmp->addr);
 518                 regs[tmp->addr] = NULL;
 519         }
 520         if (tmp->loc == LOC_LOCAL) {
 521                 if (deref)
 522                         i_ldr(1, dst, REG_FP, tmp->addr + tmp->off, bt);
 523                 else
 524                         i_add_anyimm(dst, REG_FP, tmp->addr + tmp->off);
 525         }
 526         if (tmp->loc == LOC_MEM) {
 527                 i_ldr(1, dst, REG_FP, tmp->addr, LONGSZ);
 528                 if (deref)
 529                         i_ldr(1, dst, dst, 0, bt);
 530         }
 531         tmp->addr = dst;
 532         regs[dst] = tmp;
 533         tmp->loc = LOC_REG;
 534 }
 535
 536 static void reg_free(int reg)
 537 {
 538         int i;
 539         if (!regs[reg])
 540                 return;
 541         for (i = 0; i < ARRAY_SIZE(tmpregs); i++)
 542                 if (!regs[tmpregs[i]]) {
 543                         tmp_reg(regs[reg], tmpregs[i], 0);
 544                         return;
 545                 }
 546         tmp_mem(regs[reg]);
 547 }
 548
 549 static void reg_for(int reg, struct tmp *t)
 550 {
 551         if (regs[reg] && regs[reg] != t)
 552                 reg_free(reg);
 553 }
 554
 555 static void tmp_mv(struct tmp *t, int reg)
 556 {
 557         reg_for(reg, t);
 558         tmp_reg(t, reg, 0);
 559 }
 560
 561 static void tmp_to(struct tmp *t, int reg)
 562 {
 563         reg_for(reg, t);
 564         tmp_reg(t, reg, 1);
 565 }
 566
 567 static void tmp_drop(int n)
 568 {
 569         int i;
 570         for (i = ntmp - n; i < ntmp; i++)
 571                 if (tmps[i].loc == LOC_REG)
 572                         regs[tmps[i].addr] = NULL;
 573         ntmp -= n;
 574 }
 575
 576 static void tmp_pop(int reg)
 577 {
 578         struct tmp *t = TMP(0);
 579         tmp_to(t, reg);
 580         tmp_drop(1);
 581 }
 582
 583 static struct tmp *tmp_new(void)
 584 {
 585         return &tmps[ntmp++];
 586 }
 587
 588 static void tmp_push(int reg)
 589 {
 590         struct tmp *t = tmp_new();
 591         t->addr = reg;
 592         t->bt = 0;
 593         t->loc = LOC_REG;
 594         regs[reg] = t;
 595 }
 596
 597 void o_local(long addr)
 598 {
 599         struct tmp *t = tmp_new();
 600         t->addr = -addr;
 601         t->loc = LOC_LOCAL;
 602         t->bt = 0;
 603         t->off = 0;
 604 }
 605
 606 void o_num(long num)
 607 {
 608         struct tmp *t = tmp_new();
 609         t->addr = num;
 610         t->bt = 0;
 611         t->loc = LOC_NUM;
 612 }
 613
 614 void o_sym(char *name)
 615 {
 616         struct tmp *t = tmp_new();
 617         strcpy(t->sym, name);
 618         t->loc = LOC_SYM;
 619         t->bt = 0;
 620         t->off = 0;
 621 }
 622
 623 void o_tmpdrop(int n)
 624 {
 625         if (n == -1 || n > ntmp)
 626                 n = ntmp;
 627         tmp_drop(n);
 628         if (!ntmp) {
 629                 if (tmpsp != -1)
 630                         sp = tmpsp;
 631                 tmpsp = -1;
 632         }
 633 }
 634
 635 #define FORK_REG                0x00
 636
 637 /* make sure tmps remain intact after a conditional expression */
 638 void o_fork(void)
 639 {
 640         int i;
 641         for (i = 0; i < ntmp - 1; i++)
 642                 tmp_mem(&tmps[i]);
 643 }
 644
 645 void o_forkpush(void)
 646 {
 647         tmp_pop(FORK_REG);
 648 }
 649
 650 void o_forkjoin(void)
 651 {
 652         tmp_push(FORK_REG);
 653 }
 654
 655 void o_tmpswap(void)
 656 {
 657         struct tmp *t1 = TMP(0);
 658         struct tmp *t2 = TMP(1);
 659         struct tmp t;
 660         memcpy(&t, t1, sizeof(t));
 661         memcpy(t1, t2, sizeof(t));
 662         memcpy(t2, &t, sizeof(t));
 663         if (t1->loc == LOC_REG)
 664                 regs[t1->addr] = t1;
 665         if (t2->loc == LOC_REG)
 666                 regs[t2->addr] = t2;
 667 }
 668
 669 static int reg_get(int mask)
 670 {
 671         int i;
 672         for (i = 0; i < ARRAY_SIZE(tmpregs); i++)
 673                 if ((1 << tmpregs[i]) & mask && !regs[tmpregs[i]])
 674                         return tmpregs[i];
 675         for (i = 0; i < ARRAY_SIZE(tmpregs); i++)
 676                 if ((1 << tmpregs[i]) & mask) {
 677                         reg_free(tmpregs[i]);
 678                         return tmpregs[i];
 679                 }
 680         return 0;
 681 }
 682
 683 static int reg_fortmp(struct tmp *t, int notmask)
 684 {
 685        if (t->loc == LOC_REG && !(notmask & (1 << t->addr)))
 686                return t->addr;
 687        return reg_get(~notmask);
 688 }
 689
 690 static void tmp_copy(struct tmp *t1)
 691 {
 692         struct tmp *t2 = tmp_new();
 693         memcpy(t2, t1, sizeof(*t1));
 694         if (!(t1->loc & (LOC_REG | LOC_MEM)))
 695                 return;
 696         if (t1->loc == LOC_MEM) {
 697                 tmp_mv(t2, reg_get(~0));
 698         } else if (t1->loc == LOC_REG) {
 699                 t2->addr = reg_fortmp(t2, 1 << t1->addr);
 700                 i_mov(I_MOV, t2->addr, t1->addr);
 701                 regs[t2->addr] = t2;
 702         }
 703 }
 704
 705 void o_tmpcopy(void)
 706 {
 707         tmp_copy(TMP(0));
 708 }
 709
 710 void o_cast(unsigned bt)
 711 {
 712         struct tmp *t = TMP(0);
 713         if (!t->bt && t->loc == LOC_NUM) {
 714                 num_cast(t, bt);
 715                 return;
 716         }
 717         if (BT_SZ(bt) != LONGSZ) {
 718                 int reg = reg_fortmp(t, 0);
 719                 tmp_to(t, reg);
 720                 if (bt & BT_SIGNED)
 721                         i_sx(reg, BT_SZ(bt) * 8);
 722                 else
 723                         i_zx(reg, BT_SZ(bt) * 8);
 724         }
 725 }
 726
 727 void o_func_beg(char *name, int global)
 728 {
 729         out_sym(name, (global ? OUT_GLOB : 0) | OUT_CS, cslen, 0);
 730         i_prolog();
 731         sp = 0;
 732         maxsp = sp;
 733         ntmp = 0;
 734         tmpsp = -1;
 735         nret = 0;
 736         memset(regs, 0, sizeof(regs));
 737 }
 738
 739 void o_deref(unsigned bt)
 740 {
 741         struct tmp *t = TMP(0);
 742         if (t->bt)
 743                 tmp_to(t, reg_fortmp(t, 0));
 744         t->bt = bt;
 745 }
 746
 747 void o_load(void)
 748 {
 749         struct tmp *t = TMP(0);
 750         tmp_to(t, reg_fortmp(t, 0));
 751 }
 752
 753 #define TMP_NUM(t)      ((t)->loc == LOC_NUM && !(t)->bt)
 754 #define LOCAL_PTR(t)    ((t)->loc == LOC_LOCAL && !(t)->bt)
 755 #define SYM_PTR(t)      ((t)->loc == LOC_SYM && !(t)->bt)
 756
 757 int o_popnum(long *c)
 758 {
 759         struct tmp *t = TMP(0);
 760         if (!TMP_NUM(t))
 761                 return 1;
 762         *c = t->addr;
 763         tmp_drop(1);
 764         return 0;
 765 }
 766
 767 void o_ret(int rets)
 768 {
 769         if (rets)
 770                 tmp_pop(REG_RET);
 771         else
 772                 i_num(REG_RET, 0);
 773         ret[nret++] = o_jmp(0);
 774 }
 775
 776 void o_func_end(void)
 777 {
 778         int i;
 779         for (i = 0; i < nret; i++)
 780                 o_filljmp(ret[i]);
 781         i_epilog();
 782 }
 783
 784 long o_mklocal(int size)
 785 {
 786         return sp_push(ALIGN(size, LONGSZ));
 787 }
 788
 789 void o_rmlocal(long addr, int sz)
 790 {
 791         sp = addr - sz;
 792 }
 793
 794 long o_arg2loc(int i)
 795 {
 796         return -(10 + i) << 2;
 797 }
 798
 799 void o_assign(unsigned bt)
 800 {
 801         struct tmp *t1 = TMP(0);
 802         struct tmp *t2 = TMP(1);
 803         int r1 = reg_fortmp(t1, 0);
 804         int r2 = reg_fortmp(t2, 1 << r1);
 805         int off = 0;
 806         tmp_to(t1, r1);
 807         if (t2->bt)
 808                 tmp_to(t2, r2);
 809         if (t2->loc == LOC_LOCAL) {
 810                 r2 = REG_FP;
 811                 off = t2->addr + t2->off;
 812         } else {
 813                 tmp_to(t2, r2);
 814         }
 815         tmp_drop(2);
 816         i_ldr(0, r1, r2, off, bt);
 817         tmp_push(r1);
 818 }
 819
 820 static long cu(int op, long i)
 821 {
 822         switch (op & 0xff) {
 823         case O_NEG:
 824                 return -i;
 825         case O_NOT:
 826                 return ~i;
 827         case O_LNOT:
 828                 return !i;
 829         }
 830         return 0;
 831 }
 832
 833 static int c_uop(int op)
 834 {
 835         struct tmp *t1 = TMP(0);
 836         if (!TMP_NUM(t1))
 837                 return 1;
 838         tmp_drop(1);
 839         o_num(cu(op, t1->addr));
 840         return 0;
 841 }
 842
 843 static long cb(int op, long a, long b)
 844 {
 845         switch (op & 0xff) {
 846         case O_ADD:
 847                 return a + b;
 848         case O_SUB:
 849                 return a - b;
 850         case O_AND:
 851                 return a & b;
 852         case O_OR:
 853                 return a | b;
 854         case O_XOR:
 855                 return a ^ b;
 856         case O_MUL:
 857                 return a * b;
 858         case O_DIV:
 859                 return a / b;
 860         case O_MOD:
 861                 return a % b;
 862         case O_SHL:
 863                 return a << b;
 864         case O_SHR:
 865                 if (op & O_SIGNED)
 866                         return a >> b;
 867                 else
 868                         return (unsigned long) a >> b;
 869         case O_LT:
 870                 return a < b;
 871         case O_GT:
 872                 return a > b;
 873         case O_LE:
 874                 return a <= b;
 875         case O_GE:
 876                 return a >= b;
 877         case O_EQ:
 878                 return a == b;
 879         case O_NEQ:
 880                 return a != b;
 881         }
 882         return 0;
 883 }
 884
 885 static int c_bop(int op)
 886 {
 887         struct tmp *t1 = TMP(0);
 888         struct tmp *t2 = TMP(1);
 889         int locals = LOCAL_PTR(t1) + LOCAL_PTR(t2);
 890         int syms = SYM_PTR(t1) + SYM_PTR(t2);
 891         int nums = TMP_NUM(t1) + TMP_NUM(t2);
 892         if (syms + locals == 2 || syms + nums + locals != 2)
 893                 return 1;
 894         if (nums == 1)
 895                 if ((op & 0xff) != O_ADD && ((op & 0xff) != O_SUB || TMP_NUM(t2)))
 896                         return 1;
 897         if (nums == 1) {
 898                 long o1 = TMP_NUM(t1) ? t1->addr : t1->off;
 899                 long o2 = TMP_NUM(t2) ? t2->addr : t2->off;
 900                 long ret = cb(op, o2, o1);
 901                 if (!TMP_NUM(t1))
 902                         o_tmpswap();
 903                 t2->off = ret;
 904                 tmp_drop(1);
 905         } else {
 906                 long ret = cb(op, t2->addr, t1->addr);
 907                 tmp_drop(2);
 908                 o_num(ret);
 909         }
 910         return 0;
 911 }
 912
 913 void o_uop(int op)
 914 {
 915         int r1 = reg_fortmp(TMP(0), 0);
 916         if (!c_uop(op))
 917                 return;
 918         tmp_to(TMP(0), r1);
 919         switch (op & 0xff) {
 920         case O_NEG:
 921                 i_neg(r1);
 922                 break;
 923         case O_NOT:
 924                 i_not(r1);
 925                 break;
 926         case O_LNOT:
 927                 i_lnot(r1);
 928                 break;
 929         }
 930 }
 931
 932 static void bin_regs(int *r1, int *r2)
 933 {
 934         struct tmp *t2 = TMP(0);
 935         struct tmp *t1 = TMP(1);
 936         *r2 = reg_fortmp(t2, 0);
 937         tmp_to(t2, *r2);
 938         *r1 = reg_fortmp(t1, 1 << *r2);
 939         tmp_pop(*r2);
 940         tmp_pop(*r1);
 941 }
 942
 943 static int bop_imm(int *r1, long *n, int swap)
 944 {
 945         struct tmp *t1 = TMP(0);
 946         struct tmp *t2 = TMP(1);
 947         if (!TMP_NUM(t1) && (!swap || !TMP_NUM(t2)))
 948                 return 1;
 949         *n = TMP_NUM(t1) ? t1->addr : t2->addr;
 950         if (add_decimm(add_encimm(*n)) != *n)
 951                 return 1;
 952         if (!TMP_NUM(t1))
 953                 o_tmpswap();
 954         *r1 = reg_fortmp(t2, 0);
 955         tmp_drop(1);
 956         tmp_pop(*r1);
 957         return 0;
 958 }
 959
 960 static void bin_add(int op)
 961 {
 962         /* opcode for O_ADD, O_SUB, O_AND, O_OR, O_XOR */
 963         static int rx[] = {I_ADD, I_SUB, I_AND, I_ORR, I_EOR};
 964         int r1, r2;
 965         long n;
 966         if (!bop_imm(&r1, &n, (op & 0xff) != O_SUB)) {
 967                 i_add_imm(rx[op & 0x0f], r1, r1, n);
 968         } else {
 969                 bin_regs(&r1, &r2);
 970                 i_add(rx[op & 0x0f], r1, r1, r2);
 971         }
 972         tmp_push(r1);
 973 }
 974
 975 static void bin_shx(int op)
 976 {
 977         int sm = SM_LSL;
 978         int r1, r2;
 979         long n;
 980         if ((op & 0x0f) == 1)
 981                 sm = op & O_SIGNED ? SM_ASR : SM_LSR;
 982         if (!bop_imm(&r1, &n, 0)) {
 983                 i_shl_imm(sm, r1, r1, n);
 984         } else {
 985                 bin_regs(&r1, &r2);
 986                 i_shl(sm, r1, r1, r2);
 987         }
 988         tmp_push(r1);
 989 }
 990
 991 static int log2a(unsigned long n)
 992 {
 993         int i = 0;
 994         for (i = 0; i < LONGSZ * 8; i++)
 995                 if (n & (1u << i))
 996                         break;
 997         if (i == LONGSZ * 8 || !(n >> (i + 1)))
 998                 return i;
 999         return -1;
1000 }
1001
1002 /* optimized version of mul/div/mod for powers of two */
1003 static int mul_2(int op)
1004 {
1005         struct tmp *t1 = TMP(0);
1006         struct tmp *t2 = TMP(1);
1007         long n;
1008         int r2;
1009         int p;
1010         if ((op & 0xff) == O_MUL && t2->loc == LOC_NUM && !t2->bt)
1011                 o_tmpswap();
1012         if (t1->loc != LOC_NUM || t1->bt)
1013                 return 1;
1014         n = t1->addr;
1015         p = log2a(n);
1016         if (n && p == -1)
1017                 return 1;
1018         if ((op & 0xff) == O_MUL) {
1019                 tmp_drop(1);
1020                 if (n == 1)
1021                         return 0;
1022                 if (n == 0) {
1023                         tmp_drop(1);
1024                         o_num(0);
1025                         return 0;
1026                 }
1027                 r2 = reg_fortmp(t2, 0);
1028                 tmp_to(t2, r2);
1029                 i_shl_imm(SM_LSL, r2, r2, p);
1030                 return 0;
1031         }
1032         if (op == O_DIV) {
1033                 tmp_drop(1);
1034                 if (n == 1)
1035                         return 0;
1036                 r2 = reg_fortmp(t2, 0);
1037                 tmp_to(t2, r2);
1038                 i_shl_imm(op & O_SIGNED ? SM_ASR : SM_LSR, r2, r2, p);
1039                 return 0;
1040         }
1041         if (op == O_MOD) {
1042                 tmp_drop(1);
1043                 if (n == 1) {
1044                         tmp_drop(1);
1045                         o_num(0);
1046                         return 0;
1047                 }
1048                 r2 = reg_fortmp(t2, 0);
1049                 tmp_to(t2, r2);
1050                 i_zx(r2, p);
1051                 return 0;
1052         }
1053         return 1;
1054 }
1055
1056 static void bin_div(int op)
1057 {
1058         struct tmp *t2 = TMP(0);
1059         struct tmp *t1 = TMP(1);
1060         char *func;
1061         int i;
1062         putdiv = 1;
1063         if ((op & 0xff) == O_DIV)
1064                 func = op & O_SIGNED ? "__divdi3" : "__udivdi3";
1065         else
1066                 func = op & O_SIGNED ? "__moddi3" : "__umoddi3";
1067         for (i = 0; i < ARRAY_SIZE(argregs); i++)
1068                 if (regs[argregs[i]] && regs[argregs[i]] - tmps < ntmp - 2)
1069                         tmp_mem(regs[argregs[i]]);
1070         tmp_to(t1, argregs[0]);
1071         tmp_to(t2, argregs[1]);
1072         tmp_drop(2);
1073         i_call(func);
1074         tmp_push(REG_RET);
1075 }
1076
1077 static void bin_mul(int op)
1078 {
1079         int r1, r2;
1080         if (!mul_2(op))
1081                 return;
1082         if ((op & 0xff) == O_DIV || (op & 0xff) == O_MOD) {
1083                 bin_div(op);
1084         } else {
1085                 bin_regs(&r1, &r2);
1086                 i_mul(r1, r1, r2);
1087                 tmp_push(r1);
1088         }
1089 }
1090
1091 static void bin_cmp(int op)
1092 {
1093         /* lt, gt, le, ge, eq, neq */
1094         static int ucond[] = {3, 8, 9, 2, 0, 1};
1095         static int scond[] = {11, 12, 13, 10, 0, 1};
1096         int r1, r2;
1097         long n;
1098         if (!bop_imm(&r1, &n, (op & 0xff) == O_EQ || (op & 0xff) == O_NEQ)) {
1099                 i_cmp_imm(I_CMP, r1, n);
1100         } else {
1101                 bin_regs(&r1, &r2);
1102                 i_cmp(I_CMP, r1, r2);
1103         }
1104         i_set(op & O_SIGNED ? scond[op & 0x0f] : ucond[op & 0x0f], r1);
1105         tmp_push(r1);
1106 }
1107
1108 void o_bop(int op)
1109 {
1110         if (!c_bop(op))
1111                 return;
1112         if ((op & 0xf0) == 0x00)
1113                 bin_add(op);
1114         if ((op & 0xf0) == 0x10)
1115                 bin_shx(op);
1116         if ((op & 0xf0) == 0x20)
1117                 bin_mul(op);
1118         if ((op & 0xf0) == 0x30)
1119                 bin_cmp(op);
1120 }
1121
1122 static void load_regs2(int *r0, int *r1, int *r2)
1123 {
1124         struct tmp *t0 = TMP(0);
1125         struct tmp *t1 = TMP(1);
1126         struct tmp *t2 = TMP(2);
1127         *r0 = reg_fortmp(t0, 0);
1128         *r1 = reg_fortmp(t1, 1 << *r0);
1129         *r2 = reg_fortmp(t2, (1 << *r0) | (1 << *r1));
1130         tmp_to(t0, *r0);
1131         tmp_to(t1, *r1);
1132         tmp_to(t2, *r2);
1133 }
1134
1135 void o_memcpy(void)
1136 {
1137         int rd, rs, rn;
1138         load_regs2(&rn, &rs, &rd);
1139         i_memcpy(rd, rs, rn);
1140         tmp_drop(2);
1141 }
1142
1143 void o_memset(void)
1144 {
1145         int rd, rs, rn;
1146         load_regs2(&rn, &rs, &rd);
1147         i_memset(rd, rs, rn);
1148         tmp_drop(2);
1149 }
1150
1151 long o_mklabel(void)
1152 {
1153         return cslen;
1154 }
1155
1156 static long jxz(long addr, int z)
1157 {
1158         int r = reg_fortmp(TMP(0), 0);
1159         tmp_pop(r);
1160         i_b_if(addr, r, z);
1161         return cslen - 4;
1162 }
1163
1164 long o_jz(long addr)
1165 {
1166         return jxz(addr, 1);
1167 }
1168
1169 long o_jnz(long addr)
1170 {
1171         return jxz(addr, 0);
1172 }
1173
1174 long o_jmp(long addr)
1175 {
1176         i_b(addr);
1177         return cslen - 4;
1178 }
1179
1180 void o_filljmp2(long addr, long jmpdst)
1181 {
1182         i_b_fill((void *) cs + addr, jmpdst - addr);
1183 }
1184
1185 void o_filljmp(long addr)
1186 {
1187         o_filljmp2(addr, cslen);
1188 }
1189
1190 void o_call(int argc, int rets)
1191 {
1192         struct tmp *t;
1193         int i;
1194         int aregs = MIN(ARRAY_SIZE(argregs), argc);
1195         for (i = 0; i < ARRAY_SIZE(argregs); i++)
1196                 if (regs[argregs[i]] && regs[argregs[i]] - tmps < ntmp - argc)
1197                         tmp_mem(regs[argregs[i]]);
1198         if (argc > aregs) {
1199                 sp_push(LONGSZ * (argc - aregs));
1200                 for (i = argc - 1; i >= aregs; --i) {
1201                         int reg = reg_fortmp(TMP(0), 0);
1202                         tmp_pop(reg);
1203                         i_ldr(0, reg, REG_SP, (i - aregs) * LONGSZ, LONGSZ);
1204                 }
1205         }
1206         for (i = aregs - 1; i >= 0; --i)
1207                 tmp_to(TMP(aregs - i - 1), argregs[i]);
1208         tmp_drop(aregs);
1209         t = TMP(0);
1210         if (t->loc == LOC_SYM && !t->bt) {
1211                 i_call(t->sym);
1212                 tmp_drop(1);
1213         } else {
1214                 int reg = t->loc == LOC_REG ? t->addr : REG_TMP;
1215                 tmp_pop(reg);
1216                 i_call_reg(reg);
1217         }
1218         if (rets)
1219                 tmp_push(REG_RET);
1220 }
1221
1222 void o_nogen(void)
1223 {
1224         nogen++;
1225 }
1226
1227 void o_dogen(void)
1228 {
1229         nogen--;
1230 }
1231
1232 void o_mkbss(char *name, int size, int global)
1233 {
1234         out_sym(name, OUT_BSS | (global ? OUT_GLOB : 0), bsslen, size);
1235         bsslen += ALIGN(size, LONGSZ);
1236 }
1237
1238 #define MAXDATS         (1 << 10)
1239 static char dat_names[MAXDATS][NAMELEN];
1240 static int dat_offs[MAXDATS];
1241 static int ndats;
1242
1243 void err(char *msg);
1244 void *o_mkdat(char *name, int size, int global)
1245 {
1246         void *addr = ds + dslen;
1247         int idx = ndats++;
1248         if (idx >= MAXDATS)
1249                 err("nomem: MAXDATS reached!\n");
1250         strcpy(dat_names[idx], name);
1251         dat_offs[idx] = dslen;
1252         out_sym(name, OUT_DS | (global ? OUT_GLOB : 0), dslen, size);
1253         dslen += ALIGN(size, LONGSZ);
1254         return addr;
1255 }
1256
1257 static int dat_off(char *name)
1258 {
1259         int i;
1260         for (i = 0; i < ndats; i++)
1261                 if (!strcmp(name, dat_names[i]))
1262                         return dat_offs[i];
1263         return 0;
1264 }
1265
1266 void o_datset(char *name, int off, unsigned bt)
1267 {
1268         struct tmp *t = TMP(0);
1269         int sym_off = dat_off(name) + off;
1270         if (t->loc == LOC_NUM && !t->bt) {
1271                 num_cast(t, bt);
1272                 memcpy(ds + sym_off, &t->addr, BT_SZ(bt));
1273         }
1274         if (t->loc == LOC_SYM && !t->bt) {
1275                 out_rel(t->sym, OUT_DS, sym_off);
1276                 memcpy(ds + sym_off, &t->off, BT_SZ(bt));
1277         }
1278         tmp_drop(1);
1279 }
1280
1281 /* compiled division functions; div.s contains the source */
1282 static int udivdi3[] = {
1283         0xe3a02000, 0xe3a03000, 0xe1110001, 0x0a00000a,
1284         0xe1b0c211, 0xe2822001, 0x5afffffc, 0xe3a0c001,
1285         0xe2522001, 0x4a000004, 0xe1500211, 0x3afffffb,
1286         0xe0400211, 0xe083321c, 0xeafffff8, 0xe1a01000,
1287         0xe1a00003, 0xe1a0f00e,
1288 };
1289 static int umoddi3[] = {
1290         0xe92d4000, 0xebffffeb, 0xe1a00001, 0xe8bd8000,
1291 };
1292 static int divdi3[] = {
1293         0xe92d4030, 0xe1a04000, 0xe1a05001, 0xe1100000,
1294         0x42600000, 0xe1110001, 0x42611000, 0xebffffe1,
1295         0xe1340005, 0x42600000, 0xe1140004, 0x42611000,
1296         0xe8bd8030,
1297 };
1298 static int moddi3[] = {
1299         0xe92d4000, 0xebfffff0, 0xe1a00001, 0xe8bd8000,
1300 };
1301
1302 void o_write(int fd)
1303 {
1304         if (putdiv) {
1305                 out_sym("__udivdi3", OUT_CS, cslen, 0);
1306                 os(udivdi3, sizeof(udivdi3));
1307                 out_sym("__umoddi3", OUT_CS, cslen, 0);
1308                 os(umoddi3, sizeof(umoddi3));
1309                 out_sym("__divdi3", OUT_CS, cslen, 0);
1310                 os(divdi3, sizeof(divdi3));
1311                 out_sym("__moddi3", OUT_CS, cslen, 0);
1312                 os(moddi3, sizeof(moddi3));
1313         }
1314         out_write(fd, cs, cslen, ds, dslen);
1315 }