i386-gen.c

   1 /*
   2  *  X86 code generator for TCC
   3  *
   4  *  Copyright (c) 2001 Fabrice Bellard
   5  *
   6  *  This program is free software; you can redistribute it and/or modify
   7  *  it under the terms of the GNU General Public License as published by
   8  *  the Free Software Foundation; either version 2 of the License, or
   9  *  (at your option) any later version.
  10  *
  11  *  This program is distributed in the hope that it will be useful,
  12  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  *  GNU General Public License for more details.
  15  *
  16  *  You should have received a copy of the GNU General Public License
  17  *  along with this program; if not, write to the Free Software
  18  *  Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19  */
  20
  21 /* number of available registers */
  22 #define NB_REGS             4
  23
  24 /* a register can belong to several classes. The classes must be
  25    sorted from more general to more precise (see gv2() code which does
  26    assumptions on it). */
  27 #define RC_INT     0x0001 /* generic integer register */
  28 #define RC_FLOAT   0x0002 /* generic float register */
  29 #define RC_EAX     0x0004
  30 #define RC_ST0     0x0008
  31 #define RC_ECX     0x0010
  32 #define RC_EDX     0x0020
  33 #define RC_IRET    RC_EAX /* function return: integer register */
  34 #define RC_LRET    RC_EDX /* function return: second integer register */
  35 #define RC_FRET    RC_ST0 /* function return: float register */
  36
  37 /* pretty names for the registers */
  38 enum {
  39     REG_EAX = 0,
  40     REG_ECX,
  41     REG_EDX,
  42     REG_ST0,
  43 };
  44
  45 int reg_classes[NB_REGS] = {
  46     /* eax */ RC_INT | RC_EAX,
  47     /* ecx */ RC_INT | RC_ECX,
  48     /* edx */ RC_INT | RC_EDX,
  49     /* st0 */ RC_FLOAT | RC_ST0,
  50 };
  51
  52 /* return registers for function */
  53 #define REG_IRET REG_EAX /* single word int return register */
  54 #define REG_LRET REG_EDX /* second word return register (for long long) */
  55 #define REG_FRET REG_ST0 /* float return register */
  56
  57 /* defined if function parameters must be evaluated in reverse order */
  58 #define INVERT_FUNC_PARAMS
  59
  60 /* defined if structures are passed as pointers. Otherwise structures
  61    are directly pushed on stack. */
  62 //#define FUNC_STRUCT_PARAM_AS_PTR
  63
  64 /* pointer size, in bytes */
  65 #define PTR_SIZE 4
  66
  67 /* long double size and alignment, in bytes */
  68 #define LDOUBLE_SIZE  12
  69 #define LDOUBLE_ALIGN 4
  70
  71 /* function call context */
  72 typedef struct GFuncContext {
  73     int args_size;
  74     int func_call; /* func call type (FUNC_STDCALL or FUNC_CDECL) */
  75 } GFuncContext;
  76
  77 /******************************************************/
  78
  79 static int *func_sub_sp_ptr;
  80 static unsigned char *func_bound_ptr;
  81 static int func_ret_sub;
  82
  83 void g(int c)
  84 {
  85     *(char *)ind++ = c;
  86 }
  87
  88 void o(int c)
  89 {
  90     while (c) {
  91         g(c);
  92         c = c / 256;
  93     }
  94 }
  95
  96 void gen_le32(int c)
  97 {
  98     g(c);
  99     g(c >> 8);
 100     g(c >> 16);
 101     g(c >> 24);
 102 }
 103
 104 /* patch relocation entry with value 'val' */
 105 void greloc_patch1(Reloc *p, int val)
 106 {
 107     switch(p->type) {
 108     case RELOC_ADDR32:
 109         *(int *)p->addr = val;
 110         break;
 111     case RELOC_REL32:
 112         *(int *)p->addr = val - p->addr - 4;
 113         break;
 114     }
 115 }
 116
 117 /* output a symbol and patch all calls to it */
 118 void gsym_addr(t, a)
 119 {
 120     int n;
 121     while (t) {
 122         n = *(int *)t; /* next value */
 123         *(int *)t = a - t - 4;
 124         t = n;
 125     }
 126 }
 127
 128 void gsym(t)
 129 {
 130     gsym_addr(t, ind);
 131 }
 132
 133 /* psym is used to put an instruction with a data field which is a
 134    reference to a symbol. It is in fact the same as oad ! */
 135 #define psym oad
 136
 137 /* instruction + 4 bytes data. Return the address of the data */
 138 int oad(int c, int s)
 139 {
 140     o(c);
 141     *(int *)ind = s;
 142     s = ind;
 143     ind = ind + 4;
 144     return s;
 145 }
 146
 147 /* output constant with relocation if 'r & VT_FORWARD' is true */
 148 void gen_addr32(int r, int c)
 149 {
 150     if (!(r & VT_FORWARD)) {
 151         gen_le32(c);
 152     } else {
 153         greloc((Sym *)c, ind, RELOC_ADDR32);
 154         gen_le32(0);
 155     }
 156 }
 157
 158 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 159    opcode bits */
 160 void gen_modrm(int op_reg, int r, int c)
 161 {
 162     op_reg = op_reg << 3;
 163     if ((r & VT_VALMASK) == VT_CONST) {
 164         /* constant memory reference */
 165         o(0x05 | op_reg);
 166         gen_addr32(r, c);
 167     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 168         /* currently, we use only ebp as base */
 169         if (c == (char)c) {
 170             /* short reference */
 171             o(0x45 | op_reg);
 172             g(c);
 173         } else {
 174             oad(0x85 | op_reg, c);
 175         }
 176     } else {
 177         g(0x00 | op_reg | (r & VT_VALMASK));
 178     }
 179 }
 180
 181
 182 /* load 'r' from value 'sv' */
 183 void load(int r, SValue *sv)
 184 {
 185     int v, t, ft, fc, fr;
 186     SValue v1;
 187
 188     fr = sv->r;
 189     ft = sv->t;
 190     fc = sv->c.ul;
 191
 192     v = fr & VT_VALMASK;
 193     if (fr & VT_LVAL) {
 194         if (v == VT_LLOCAL) {
 195             v1.t = VT_INT;
 196             v1.r = VT_LOCAL | VT_LVAL;
 197             v1.c.ul = fc;
 198             load(r, &v1);
 199             fr = r;
 200         }
 201         if ((ft & VT_BTYPE) == VT_FLOAT) {
 202             o(0xd9); /* flds */
 203             r = 0;
 204         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 205             o(0xdd); /* fldl */
 206             r = 0;
 207         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 208             o(0xdb); /* fldt */
 209             r = 5;
 210         } else if ((ft & VT_TYPE) == VT_BYTE)
 211             o(0xbe0f);   /* movsbl */
 212         else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED))
 213             o(0xb60f);   /* movzbl */
 214         else if ((ft & VT_TYPE) == VT_SHORT)
 215             o(0xbf0f);   /* movswl */
 216         else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED))
 217             o(0xb70f);   /* movzwl */
 218         else
 219             o(0x8b);     /* movl */
 220         gen_modrm(r, fr, fc);
 221     } else {
 222         if (v == VT_CONST) {
 223             o(0xb8 + r); /* mov $xx, r */
 224             gen_addr32(fr, fc);
 225         } else if (v == VT_LOCAL) {
 226             o(0x8d); /* lea xxx(%ebp), r */
 227             gen_modrm(r, VT_LOCAL, fc);
 228         } else if (v == VT_CMP) {
 229             oad(0xb8 + r, 0); /* mov $0, r */
 230             o(0x0f); /* setxx %br */
 231             o(fc);
 232             o(0xc0 + r);
 233         } else if (v == VT_JMP || v == VT_JMPI) {
 234             t = v & 1;
 235             oad(0xb8 + r, t); /* mov $1, r */
 236             oad(0xe9, 5); /* jmp after */
 237             gsym(fc);
 238             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 239         } else if (v != r) {
 240             o(0x89);
 241             o(0xc0 + r + v * 8); /* mov v, r */
 242         }
 243     }
 244 }
 245
 246 /* store register 'r' in lvalue 'v' */
 247 void store(int r, SValue *v)
 248 {
 249     int fr, bt, ft, fc;
 250
 251     ft = v->t;
 252     fc = v->c.ul;
 253     fr = v->r & VT_VALMASK;
 254     bt = ft & VT_BTYPE;
 255     /* XXX: incorrect if float reg to reg */
 256     if (bt == VT_FLOAT) {
 257         o(0xd9); /* fsts */
 258         r = 2;
 259     } else if (bt == VT_DOUBLE) {
 260         o(0xdd); /* fstpl */
 261         r = 2;
 262     } else if (bt == VT_LDOUBLE) {
 263         o(0xc0d9); /* fld %st(0) */
 264         o(0xdb); /* fstpt */
 265         r = 7;
 266     } else {
 267         if (bt == VT_SHORT)
 268             o(0x66);
 269         if (bt == VT_BYTE)
 270             o(0x88);
 271         else
 272             o(0x89);
 273     }
 274     if (fr == VT_CONST ||
 275         fr == VT_LOCAL ||
 276         (v->r & VT_LVAL)) {
 277         gen_modrm(r, v->r, fc);
 278     } else if (fr != r) {
 279         o(0xc0 + fr + r * 8); /* mov r, fr */
 280     }
 281 }
 282
 283 /* start function call and return function call context */
 284 void gfunc_start(GFuncContext *c, int func_call)
 285 {
 286     c->args_size = 0;
 287     c->func_call = func_call;
 288 }
 289
 290 /* push function parameter which is in (vtop->t, vtop->c). Stack entry
 291    is then popped. */
 292 void gfunc_param(GFuncContext *c)
 293 {
 294     int size, align, r;
 295
 296     if ((vtop->t & VT_BTYPE) == VT_STRUCT) {
 297         size = type_size(vtop->t, &align);
 298         /* align to stack align size */
 299         size = (size + 3) & ~3;
 300         /* allocate the necessary size on stack */
 301         oad(0xec81, size); /* sub $xxx, %esp */
 302         /* generate structure store */
 303         r = get_reg(RC_INT);
 304         o(0x89); /* mov %esp, r */
 305         o(0xe0 + r);
 306         vset(VT_INT, r, 0);
 307         vswap();
 308         vstore();
 309         c->args_size += size;
 310     } else if (is_float(vtop->t)) {
 311         gv(RC_FLOAT); /* only one float register */
 312         if ((vtop->t & VT_BTYPE) == VT_FLOAT)
 313             size = 4;
 314         else if ((vtop->t & VT_BTYPE) == VT_DOUBLE)
 315             size = 8;
 316         else
 317             size = 12;
 318         oad(0xec81, size); /* sub $xxx, %esp */
 319         if (size == 12)
 320             o(0x7cdb);
 321         else
 322             o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 323         g(0x24);
 324         g(0x00);
 325         c->args_size += size;
 326     } else {
 327         /* simple type (currently always same size) */
 328         /* XXX: implicit cast ? */
 329         r = gv(RC_INT);
 330         if ((vtop->t & VT_BTYPE) == VT_LLONG) {
 331             size = 8;
 332             o(0x50 + vtop->r2); /* push r */
 333         } else {
 334             size = 4;
 335         }
 336         o(0x50 + r); /* push r */
 337         c->args_size += size;
 338     }
 339     vtop--;
 340 }
 341
 342 /* generate function call with address in (vtop->t, vtop->c) and free function
 343    context. Stack entry is popped */
 344 void gfunc_call(GFuncContext *c)
 345 {
 346     int r;
 347     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 348         /* constant case */
 349         /* forward reference */
 350         if (vtop->r & VT_FORWARD) {
 351             greloc(vtop->c.sym, ind + 1, RELOC_REL32);
 352             oad(0xe8, 0);
 353         } else {
 354             oad(0xe8, vtop->c.ul - ind - 5);
 355         }
 356     } else {
 357         /* otherwise, indirect call */
 358         r = gv(RC_INT);
 359         o(0xff); /* call *r */
 360         o(0xd0 + r);
 361     }
 362     if (c->args_size && c->func_call == FUNC_CDECL)
 363         oad(0xc481, c->args_size); /* add $xxx, %esp */
 364     vtop--;
 365 }
 366
 367 /* generate function prolog of type 't' */
 368 void gfunc_prolog(int t)
 369 {
 370     int addr, align, size, u, func_call;
 371     Sym *sym;
 372
 373     sym = sym_find((unsigned)t >> VT_STRUCT_SHIFT);
 374     func_call = sym->r;
 375     addr = 8;
 376     /* if the function returns a structure, then add an
 377        implicit pointer parameter */
 378     func_vt = sym->t;
 379     if ((func_vt & VT_BTYPE) == VT_STRUCT) {
 380         func_vc = addr;
 381         addr += 4;
 382     }
 383     /* define parameters */
 384     while ((sym = sym->next) != NULL) {
 385         u = sym->t;
 386         sym_push(sym->v & ~SYM_FIELD, u,
 387                  VT_LOCAL | VT_LVAL, addr);
 388         size = type_size(u, &align);
 389         size = (size + 3) & ~3;
 390 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 391         /* structs are passed as pointer */
 392         if ((u & VT_BTYPE) == VT_STRUCT) {
 393             size = 4;
 394         }
 395 #endif
 396         addr += size;
 397     }
 398     func_ret_sub = 0;
 399     /* pascal type call ? */
 400     if (func_call == FUNC_STDCALL)
 401         func_ret_sub = addr - 8;
 402     o(0xe58955); /* push   %ebp, mov    %esp, %ebp */
 403     func_sub_sp_ptr = (int *)oad(0xec81, 0); /* sub $xxx, %esp */
 404     /* leave some room for bound checking code */
 405     if (do_bounds_check) {
 406         oad(0xb8, 0); /* lbound section pointer */
 407         oad(0xb8, 0); /* call to function */
 408         func_bound_ptr = lbounds_section->data_ptr;
 409     }
 410 }
 411
 412 /* generate function epilog */
 413 void gfunc_epilog(void)
 414 {
 415     if (do_bounds_check && func_bound_ptr != lbounds_section->data_ptr) {
 416         int saved_ind;
 417         int *bounds_ptr;
 418         /* add end of table info */
 419         bounds_ptr = (int *)lbounds_section->data_ptr;
 420         *bounds_ptr++ = 0;
 421         lbounds_section->data_ptr = (unsigned char *)bounds_ptr;
 422         /* generate bound local allocation */
 423         saved_ind = ind;
 424         ind = (int)func_sub_sp_ptr + 4;
 425         oad(0xb8, (int)func_bound_ptr); /* mov %eax, xxx */
 426         oad(0xe8, (int)__bound_local_new - ind - 5);
 427         ind = saved_ind;
 428         /* generate bound check local freeing */
 429         o(0x5250); /* save returned value, if any */
 430         oad(0xb8, (int)func_bound_ptr); /* mov %eax, xxx */
 431         oad(0xe8, (int)__bound_local_delete - ind - 5);
 432         o(0x585a); /* restore returned value, if any */
 433     }
 434     o(0xc9); /* leave */
 435     if (func_ret_sub == 0) {
 436         o(0xc3); /* ret */
 437     } else {
 438         o(0xc2); /* ret n */
 439         g(func_ret_sub);
 440         g(func_ret_sub >> 8);
 441     }
 442     /* align local size to word & save local variables */
 443     *func_sub_sp_ptr = (-loc + 3) & -4;
 444 }
 445
 446 /* generate a jump to a label */
 447 int gjmp(int t)
 448 {
 449     return psym(0xe9, t);
 450 }
 451
 452 /* generate a jump to a fixed address */
 453 void gjmp_addr(int a)
 454 {
 455     oad(0xe9, a - ind - 5);
 456 }
 457
 458 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 459 int gtst(int inv, int t)
 460 {
 461     int v, *p;
 462     v = vtop->r & VT_VALMASK;
 463     if (v == VT_CMP) {
 464         /* fast case : can jump directly since flags are set */
 465         g(0x0f);
 466         t = psym((vtop->c.i - 16) ^ inv, t);
 467     } else if (v == VT_JMP || v == VT_JMPI) {
 468         /* && or || optimization */
 469         if ((v & 1) == inv) {
 470             /* insert vtop->c jump list in t */
 471             p = &vtop->c.i;
 472             while (*p != 0)
 473                 p = (int *)*p;
 474             *p = t;
 475             t = vtop->c.i;
 476         } else {
 477             t = gjmp(t);
 478             gsym(vtop->c.i);
 479         }
 480     } else {
 481         if (is_float(vtop->t)) {
 482             vpushi(0);
 483             gen_op(TOK_NE);
 484         }
 485         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_FORWARD)) == VT_CONST) {
 486             /* constant jmp optimization */
 487             if ((vtop->c.i != 0) != inv)
 488                 t = gjmp(t);
 489         } else {
 490             v = gv(RC_INT);
 491             o(0x85);
 492             o(0xc0 + v * 9);
 493             g(0x0f);
 494             t = psym(0x85 ^ inv, t);
 495         }
 496     }
 497     vtop--;
 498     return t;
 499 }
 500
 501 /* generate an integer binary operation */
 502 void gen_opi(int op)
 503 {
 504     int r, fr, opc, c;
 505
 506     switch(op) {
 507     case '+':
 508     case TOK_ADDC1: /* add with carry generation */
 509         opc = 0;
 510     gen_op8:
 511         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_FORWARD)) == VT_CONST) {
 512             /* constant case */
 513             vswap();
 514             r = gv(RC_INT);
 515             vswap();
 516             c = vtop->c.i;
 517             if (c == (char)c) {
 518                 /* XXX: generate inc and dec for smaller code ? */
 519                 o(0x83);
 520                 o(0xc0 | (opc << 3) | r);
 521                 g(c);
 522             } else {
 523                 o(0x81);
 524                 oad(0xc0 | (opc << 3) | r, c);
 525             }
 526         } else {
 527             gv2(RC_INT, RC_INT);
 528             r = vtop[-1].r;
 529             fr = vtop[0].r;
 530             o((opc << 3) | 0x01);
 531             o(0xc0 + r + fr * 8);
 532         }
 533         vtop--;
 534         if (op >= TOK_ULT && op <= TOK_GT) {
 535             vtop--;
 536             vset(VT_INT, VT_CMP, op);
 537         }
 538         break;
 539     case '-':
 540     case TOK_SUBC1: /* sub with carry generation */
 541         opc = 5;
 542         goto gen_op8;
 543     case TOK_ADDC2: /* add with carry use */
 544         opc = 2;
 545         goto gen_op8;
 546     case TOK_SUBC2: /* sub with carry use */
 547         opc = 3;
 548         goto gen_op8;
 549     case '&':
 550         opc = 4;
 551         goto gen_op8;
 552     case '^':
 553         opc = 6;
 554         goto gen_op8;
 555     case '|':
 556         opc = 1;
 557         goto gen_op8;
 558     case '*':
 559         gv2(RC_INT, RC_INT);
 560         r = vtop[-1].r;
 561         fr = vtop[0].r;
 562         vtop--;
 563         o(0xaf0f); /* imul fr, r */
 564         o(0xc0 + fr + r * 8);
 565         break;
 566     case TOK_SHL:
 567         opc = 4;
 568         goto gen_shift;
 569     case TOK_SHR:
 570         opc = 5;
 571         goto gen_shift;
 572     case TOK_SAR:
 573         opc = 7;
 574     gen_shift:
 575         opc = 0xc0 | (opc << 3);
 576         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_FORWARD)) == VT_CONST) {
 577             /* constant case */
 578             vswap();
 579             r = gv(RC_INT);
 580             vswap();
 581             c = vtop->c.i & 0x1f;
 582             o(0xc1); /* shl/shr/sar $xxx, r */
 583             o(opc | r);
 584             g(c);
 585         } else {
 586             /* we generate the shift in ecx */
 587             gv2(RC_INT, RC_ECX);
 588             r = vtop[-1].r;
 589             o(0xd3); /* shl/shr/sar %cl, r */
 590             o(opc | r);
 591         }
 592         vtop--;
 593         break;
 594     case '/':
 595     case TOK_UDIV:
 596     case TOK_PDIV:
 597     case '%':
 598     case TOK_UMOD:
 599     case TOK_UMULL:
 600         /* first operand must be in eax */
 601         /* XXX: need better constraint for second operand */
 602         gv2(RC_EAX, RC_ECX);
 603         r = vtop[-1].r;
 604         fr = vtop[0].r;
 605         vtop--;
 606         save_reg(REG_EDX);
 607         if (op == TOK_UMULL) {
 608             o(0xf7); /* mul fr */
 609             o(0xe0 + fr);
 610             vtop->r2 = REG_EDX;
 611             r = REG_EAX;
 612         } else {
 613             if (op == TOK_UDIV || op == TOK_UMOD) {
 614                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 615                 o(0xf0 + fr);
 616             } else {
 617                 o(0xf799); /* cltd, idiv fr, %eax */
 618                 o(0xf8 + fr);
 619             }
 620             if (op == '%' || op == TOK_UMOD)
 621                 r = REG_EDX;
 622             else
 623                 r = REG_EAX;
 624         }
 625         vtop->r = r;
 626         break;
 627     default:
 628         opc = 7;
 629         goto gen_op8;
 630     }
 631 }
 632
 633 /* generate a floating point operation 'v = t1 op t2' instruction. The
 634    two operands are guaranted to have the same floating point type */
 635 /* XXX: need to use ST1 too */
 636 void gen_opf(int op)
 637 {
 638     int a, ft, fc, swapped, r;
 639
 640     /* convert constants to memory references */
 641     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 642         vswap();
 643         gv(RC_FLOAT);
 644         vswap();
 645     }
 646     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 647         gv(RC_FLOAT);
 648
 649     /* must put at least one value in the floating point register */
 650     if ((vtop[-1].r & VT_LVAL) &&
 651         (vtop[0].r & VT_LVAL)) {
 652         vswap();
 653         gv(RC_FLOAT);
 654         vswap();
 655     }
 656     if (op >= TOK_ULT && op <= TOK_GT) {
 657         /* load on stack second operand */
 658         load(REG_ST0, vtop);
 659         save_reg(REG_EAX); /* eax is used by FP comparison code */
 660         if (op == TOK_GE || op == TOK_GT)
 661             o(0xc9d9); /* fxch %st(1) */
 662         o(0xe9da); /* fucompp */
 663         o(0xe0df); /* fnstsw %ax */
 664         if (op == TOK_EQ) {
 665             o(0x45e480); /* and $0x45, %ah */
 666             o(0x40fC80); /* cmp $0x40, %ah */
 667         } else if (op == TOK_NE) {
 668             o(0x45e480); /* and $0x45, %ah */
 669             o(0x40f480); /* xor $0x40, %ah */
 670             op = TOK_NE;
 671         } else if (op == TOK_GE || op == TOK_LE) {
 672             o(0x05c4f6); /* test $0x05, %ah */
 673             op = TOK_EQ;
 674         } else {
 675             o(0x45c4f6); /* test $0x45, %ah */
 676             op = TOK_EQ;
 677         }
 678         vtop--;
 679         vtop->r = VT_CMP;
 680         vtop->c.i = op;
 681     } else {
 682         swapped = 0;
 683         /* swap the stack if needed so that t1 is the register and t2 is
 684            the memory reference */
 685         if (vtop[-1].r & VT_LVAL) {
 686             vswap();
 687             swapped = 1;
 688         }
 689         /* no memory reference possible for long double operations */
 690         if ((vtop->t & VT_BTYPE) == VT_LDOUBLE) {
 691             load(REG_ST0, vtop);
 692             swapped = !swapped;
 693         }
 694
 695         switch(op) {
 696         default:
 697         case '+':
 698             a = 0;
 699             break;
 700         case '-':
 701             a = 4;
 702             if (swapped)
 703                 a++;
 704             break;
 705         case '*':
 706             a = 1;
 707             break;
 708         case '/':
 709             a = 6;
 710             if (swapped)
 711                 a++;
 712             break;
 713         }
 714         ft = vtop->t;
 715         fc = vtop->c.ul;
 716         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 717             o(0xde); /* fxxxp %st, %st(1) */
 718             o(0xc1 + (a << 3));
 719         } else {
 720             /* if saved lvalue, then we must reload it */
 721             r = vtop->r;
 722             if ((r & VT_VALMASK) == VT_LLOCAL) {
 723                 SValue v1;
 724                 r = get_reg(RC_INT);
 725                 v1.t = VT_INT;
 726                 v1.r = VT_LOCAL | VT_LVAL;
 727                 v1.c.ul = fc;
 728                 load(r, &v1);
 729                 fc = 0;
 730             }
 731
 732             if ((ft & VT_BTYPE) == VT_DOUBLE)
 733                 o(0xdc);
 734             else
 735                 o(0xd8);
 736             gen_modrm(a, r, fc);
 737         }
 738         vtop--;
 739     }
 740 }
 741
 742 /* FPU control word for rounding to nearest mode */
 743 /* XXX: should move that into tcc lib support code ! */
 744 static unsigned short __tcc_fpu_control = 0x137f;
 745 /* FPU control word for round to zero mode for int convertion */
 746 static unsigned short __tcc_int_fpu_control = 0x137f | 0x0c00;
 747
 748 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 749    and 'long long' cases. */
 750 void gen_cvt_itof(int t)
 751 {
 752     save_reg(REG_ST0);
 753     gv(RC_INT);
 754     if ((vtop->t & VT_BTYPE) == VT_LLONG) {
 755         /* signed long long to float/double/long double (unsigned case
 756            is handled generically) */
 757         o(0x50 + vtop->r2); /* push r2 */
 758         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 759         o(0x242cdf); /* fildll (%esp) */
 760         o(0x08c483); /* add $8, %esp */
 761     } else if ((vtop->t & (VT_BTYPE | VT_UNSIGNED)) ==
 762                (VT_INT | VT_UNSIGNED)) {
 763         /* unsigned int to float/double/long double */
 764         o(0x6a); /* push $0 */
 765         g(0x00);
 766         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 767         o(0x242cdf); /* fildll (%esp) */
 768         o(0x08c483); /* add $8, %esp */
 769     } else {
 770         /* int to float/double/long double */
 771         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 772         o(0x2404db); /* fildl (%esp) */
 773         o(0x04c483); /* add $4, %esp */
 774     }
 775     vtop->r = REG_ST0;
 776 }
 777
 778 /* convert fp to int 't' type */
 779 /* XXX: handle long long case */
 780 void gen_cvt_ftoi(int t)
 781 {
 782     int r, r2, size;
 783
 784     gv(RC_FLOAT);
 785     if (t != VT_INT)
 786         size = 8;
 787     else
 788         size = 4;
 789
 790     oad(0x2dd9, (int)&__tcc_int_fpu_control); /* ldcw xxx */
 791     oad(0xec81, size); /* sub $xxx, %esp */
 792     if (size == 4)
 793         o(0x1cdb); /* fistpl */
 794     else
 795         o(0x3cdf); /* fistpll */
 796     o(0x24);
 797     oad(0x2dd9, (int)&__tcc_fpu_control); /* ldcw xxx */
 798     r = get_reg(RC_INT);
 799     o(0x58 + r); /* pop r */
 800     if (size == 8) {
 801         if (t == VT_LLONG) {
 802             vtop->r = r; /* mark reg as used */
 803             r2 = get_reg(RC_INT);
 804             o(0x58 + r2); /* pop r2 */
 805             vtop->r2 = r2;
 806         } else {
 807             o(0x04c483); /* add $4, %esp */
 808         }
 809     }
 810     vtop->r = r;
 811 }
 812
 813 /* convert from one floating point type to another */
 814 void gen_cvt_ftof(int t)
 815 {
 816     /* all we have to do on i386 is to put the float in a register */
 817     gv(RC_FLOAT);
 818 }
 819
 820 /* bound check support functions */
 821
 822 /* generate first part of bounded pointer addition */
 823 void gen_bounded_ptr_add1(void)
 824 {
 825     /* prepare fast i386 function call (args in eax and edx) */
 826     gv2(RC_EAX, RC_EDX);
 827     /* save all temporary registers */
 828     vtop--;
 829     vtop->r = VT_CONST;
 830     save_regs();
 831 }
 832
 833 /* if deref is true, then also test dereferencing */
 834 void gen_bounded_ptr_add2(int deref)
 835 {
 836     void *func;
 837     int size, align;
 838
 839     if (deref) {
 840         size = type_size(vtop->t, &align);
 841         switch(size) {
 842         case  1: func = __bound_ptr_indir1; break;
 843         case  2: func = __bound_ptr_indir2; break;
 844         case  4: func = __bound_ptr_indir4; break;
 845         case  8: func = __bound_ptr_indir8; break;
 846         case 12: func = __bound_ptr_indir12; break;
 847         case 16: func = __bound_ptr_indir16; break;
 848         default:
 849             error("unhandled size when derefencing bounded pointer");
 850             func = NULL;
 851             break;
 852         }
 853     } else {
 854         func = __bound_ptr_add;
 855     }
 856
 857     /* do a fast function call */
 858     oad(0xe8, (int)func - ind - 5);
 859     /* return pointer is there */
 860     vtop->r = REG_EAX;
 861 }
 862
 863 /* end of X86 code generator */
 864 /*************************************************************/
 865