i386-gen.c

   1 /*
   2  *  X86 code generator for TCC
   3  *
   4  *  Copyright (c) 2001, 2002 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 /* relocation type for 32 bit data relocation */
  72 #define R_DATA_32 R_386_32
  73
  74 /* function call context */
  75 typedef struct GFuncContext {
  76     int args_size;
  77     int func_call; /* func call type (FUNC_STDCALL or FUNC_CDECL) */
  78 } GFuncContext;
  79
  80 /******************************************************/
  81
  82 static int *func_sub_sp_ptr;
  83 static unsigned char *func_bound_ptr;
  84 static int func_ret_sub;
  85
  86 void g(int c)
  87 {
  88     *(char *)ind++ = c;
  89 }
  90
  91 void o(int c)
  92 {
  93     while (c) {
  94         g(c);
  95         c = c / 256;
  96     }
  97 }
  98
  99 void gen_le32(int c)
 100 {
 101     g(c);
 102     g(c >> 8);
 103     g(c >> 16);
 104     g(c >> 24);
 105 }
 106
 107 /* output a symbol and patch all calls to it */
 108 void gsym_addr(int t, int a)
 109 {
 110     int n;
 111     while (t) {
 112         n = *(int *)t; /* next value */
 113         *(int *)t = a - t - 4;
 114         t = n;
 115     }
 116 }
 117
 118 void gsym(int t)
 119 {
 120     gsym_addr(t, ind);
 121 }
 122
 123 /* psym is used to put an instruction with a data field which is a
 124    reference to a symbol. It is in fact the same as oad ! */
 125 #define psym oad
 126
 127 /* instruction + 4 bytes data. Return the address of the data */
 128 int oad(int c, int s)
 129 {
 130     o(c);
 131     *(int *)ind = s;
 132     s = ind;
 133     ind = ind + 4;
 134     return s;
 135 }
 136
 137 /* output constant with relocation if 'r & VT_SYM' is true */
 138 void gen_addr32(int r, int c)
 139 {
 140     if (!(r & VT_SYM)) {
 141         gen_le32(c);
 142     } else {
 143         greloc(cur_text_section,
 144                (Sym *)c, ind - (int)cur_text_section->data, R_386_32);
 145         gen_le32(0);
 146     }
 147 }
 148
 149 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 150    opcode bits */
 151 void gen_modrm(int op_reg, int r, int c)
 152 {
 153     op_reg = op_reg << 3;
 154     if ((r & VT_VALMASK) == VT_CONST) {
 155         /* constant memory reference */
 156         o(0x05 | op_reg);
 157         gen_addr32(r, c);
 158     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 159         /* currently, we use only ebp as base */
 160         if (c == (char)c) {
 161             /* short reference */
 162             o(0x45 | op_reg);
 163             g(c);
 164         } else {
 165             oad(0x85 | op_reg, c);
 166         }
 167     } else {
 168         g(0x00 | op_reg | (r & VT_VALMASK));
 169     }
 170 }
 171
 172
 173 /* load 'r' from value 'sv' */
 174 void load(int r, SValue *sv)
 175 {
 176     int v, t, ft, fc, fr;
 177     SValue v1;
 178
 179     fr = sv->r;
 180     ft = sv->t;
 181     fc = sv->c.ul;
 182
 183     v = fr & VT_VALMASK;
 184     if (fr & VT_LVAL) {
 185         if (v == VT_LLOCAL) {
 186             v1.t = VT_INT;
 187             v1.r = VT_LOCAL | VT_LVAL;
 188             v1.c.ul = fc;
 189             load(r, &v1);
 190             fr = r;
 191         }
 192         if ((ft & VT_BTYPE) == VT_FLOAT) {
 193             o(0xd9); /* flds */
 194             r = 0;
 195         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 196             o(0xdd); /* fldl */
 197             r = 0;
 198         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 199             o(0xdb); /* fldt */
 200             r = 5;
 201         } else if ((ft & VT_TYPE) == VT_BYTE) {
 202             o(0xbe0f);   /* movsbl */
 203         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 204             o(0xb60f);   /* movzbl */
 205         } else if ((ft & VT_TYPE) == VT_SHORT) {
 206             o(0xbf0f);   /* movswl */
 207         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 208             o(0xb70f);   /* movzwl */
 209         } else {
 210             o(0x8b);     /* movl */
 211         }
 212         gen_modrm(r, fr, fc);
 213     } else {
 214         if (v == VT_CONST) {
 215             o(0xb8 + r); /* mov $xx, r */
 216             gen_addr32(fr, fc);
 217         } else if (v == VT_LOCAL) {
 218             o(0x8d); /* lea xxx(%ebp), r */
 219             gen_modrm(r, VT_LOCAL, fc);
 220         } else if (v == VT_CMP) {
 221             oad(0xb8 + r, 0); /* mov $0, r */
 222             o(0x0f); /* setxx %br */
 223             o(fc);
 224             o(0xc0 + r);
 225         } else if (v == VT_JMP || v == VT_JMPI) {
 226             t = v & 1;
 227             oad(0xb8 + r, t); /* mov $1, r */
 228             oad(0xe9, 5); /* jmp after */
 229             gsym(fc);
 230             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 231         } else if (v != r) {
 232             o(0x89);
 233             o(0xc0 + r + v * 8); /* mov v, r */
 234         }
 235     }
 236 }
 237
 238 /* store register 'r' in lvalue 'v' */
 239 void store(int r, SValue *v)
 240 {
 241     int fr, bt, ft, fc;
 242
 243     ft = v->t;
 244     fc = v->c.ul;
 245     fr = v->r & VT_VALMASK;
 246     bt = ft & VT_BTYPE;
 247     /* XXX: incorrect if float reg to reg */
 248     if (bt == VT_FLOAT) {
 249         o(0xd9); /* fsts */
 250         r = 2;
 251     } else if (bt == VT_DOUBLE) {
 252         o(0xdd); /* fstpl */
 253         r = 2;
 254     } else if (bt == VT_LDOUBLE) {
 255         o(0xc0d9); /* fld %st(0) */
 256         o(0xdb); /* fstpt */
 257         r = 7;
 258     } else {
 259         if (bt == VT_SHORT)
 260             o(0x66);
 261         if (bt == VT_BYTE)
 262             o(0x88);
 263         else
 264             o(0x89);
 265     }
 266     if (fr == VT_CONST ||
 267         fr == VT_LOCAL ||
 268         (v->r & VT_LVAL)) {
 269         gen_modrm(r, v->r, fc);
 270     } else if (fr != r) {
 271         o(0xc0 + fr + r * 8); /* mov r, fr */
 272     }
 273 }
 274
 275 /* start function call and return function call context */
 276 void gfunc_start(GFuncContext *c, int func_call)
 277 {
 278     c->args_size = 0;
 279     c->func_call = func_call;
 280 }
 281
 282 /* push function parameter which is in (vtop->t, vtop->c). Stack entry
 283    is then popped. */
 284 void gfunc_param(GFuncContext *c)
 285 {
 286     int size, align, r;
 287
 288     if ((vtop->t & VT_BTYPE) == VT_STRUCT) {
 289         size = type_size(vtop->t, &align);
 290         /* align to stack align size */
 291         size = (size + 3) & ~3;
 292         /* allocate the necessary size on stack */
 293         oad(0xec81, size); /* sub $xxx, %esp */
 294         /* generate structure store */
 295         r = get_reg(RC_INT);
 296         o(0x89); /* mov %esp, r */
 297         o(0xe0 + r);
 298         vset(vtop->t, r | VT_LVAL, 0);
 299         vswap();
 300         vstore();
 301         c->args_size += size;
 302     } else if (is_float(vtop->t)) {
 303         gv(RC_FLOAT); /* only one float register */
 304         if ((vtop->t & VT_BTYPE) == VT_FLOAT)
 305             size = 4;
 306         else if ((vtop->t & VT_BTYPE) == VT_DOUBLE)
 307             size = 8;
 308         else
 309             size = 12;
 310         oad(0xec81, size); /* sub $xxx, %esp */
 311         if (size == 12)
 312             o(0x7cdb);
 313         else
 314             o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 315         g(0x24);
 316         g(0x00);
 317         c->args_size += size;
 318     } else {
 319         /* simple type (currently always same size) */
 320         /* XXX: implicit cast ? */
 321         r = gv(RC_INT);
 322         if ((vtop->t & VT_BTYPE) == VT_LLONG) {
 323             size = 8;
 324             o(0x50 + vtop->r2); /* push r */
 325         } else {
 326             size = 4;
 327         }
 328         o(0x50 + r); /* push r */
 329         c->args_size += size;
 330     }
 331     vtop--;
 332 }
 333
 334 /* generate function call with address in (vtop->t, vtop->c) and free function
 335    context. Stack entry is popped */
 336 void gfunc_call(GFuncContext *c)
 337 {
 338     int r;
 339     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 340         /* constant case */
 341         if (vtop->r & VT_SYM) {
 342             /* relocation case */
 343             greloc(cur_text_section, vtop->c.sym,
 344                    ind + 1 - (int)cur_text_section->data, R_386_PC32);
 345             oad(0xe8, -4);
 346         } else {
 347             oad(0xe8, vtop->c.ul - ind - 5);
 348         }
 349     } else {
 350         /* otherwise, indirect call */
 351         r = gv(RC_INT);
 352         o(0xff); /* call *r */
 353         o(0xd0 + r);
 354     }
 355     if (c->args_size && c->func_call == FUNC_CDECL)
 356         oad(0xc481, c->args_size); /* add $xxx, %esp */
 357     vtop--;
 358 }
 359
 360 /* generate function prolog of type 't' */
 361 void gfunc_prolog(int t)
 362 {
 363     int addr, align, size, u, func_call;
 364     Sym *sym;
 365
 366     sym = sym_find((unsigned)t >> VT_STRUCT_SHIFT);
 367     func_call = sym->r;
 368     addr = 8;
 369     /* if the function returns a structure, then add an
 370        implicit pointer parameter */
 371     func_vt = sym->t;
 372     if ((func_vt & VT_BTYPE) == VT_STRUCT) {
 373         func_vc = addr;
 374         addr += 4;
 375     }
 376     /* define parameters */
 377     while ((sym = sym->next) != NULL) {
 378         u = sym->t;
 379         sym_push(sym->v & ~SYM_FIELD, u,
 380                  VT_LOCAL | VT_LVAL, addr);
 381         size = type_size(u, &align);
 382         size = (size + 3) & ~3;
 383 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 384         /* structs are passed as pointer */
 385         if ((u & VT_BTYPE) == VT_STRUCT) {
 386             size = 4;
 387         }
 388 #endif
 389         addr += size;
 390     }
 391     func_ret_sub = 0;
 392     /* pascal type call ? */
 393     if (func_call == FUNC_STDCALL)
 394         func_ret_sub = addr - 8;
 395     o(0xe58955); /* push   %ebp, mov    %esp, %ebp */
 396     func_sub_sp_ptr = (int *)oad(0xec81, 0); /* sub $xxx, %esp */
 397     /* leave some room for bound checking code */
 398     if (do_bounds_check) {
 399         oad(0xb8, 0); /* lbound section pointer */
 400         oad(0xb8, 0); /* call to function */
 401         func_bound_ptr = lbounds_section->data_ptr;
 402     }
 403 }
 404
 405 /* generate function epilog */
 406 void gfunc_epilog(void)
 407 {
 408 #ifdef CONFIG_TCC_BCHECK
 409     if (do_bounds_check && func_bound_ptr != lbounds_section->data_ptr) {
 410         int saved_ind;
 411         int *bounds_ptr;
 412         /* add end of table info */
 413         bounds_ptr = (int *)lbounds_section->data_ptr;
 414         *bounds_ptr++ = 0;
 415         lbounds_section->data_ptr = (unsigned char *)bounds_ptr;
 416         /* generate bound local allocation */
 417         saved_ind = ind;
 418         ind = (int)func_sub_sp_ptr + 4;
 419         oad(0xb8, (int)func_bound_ptr); /* mov %eax, xxx */
 420         oad(0xe8, (int)__bound_local_new - ind - 5);
 421         ind = saved_ind;
 422         /* generate bound check local freeing */
 423         o(0x5250); /* save returned value, if any */
 424         oad(0xb8, (int)func_bound_ptr); /* mov %eax, xxx */
 425         oad(0xe8, (int)__bound_local_delete - ind - 5);
 426         o(0x585a); /* restore returned value, if any */
 427     }
 428 #endif
 429     o(0xc9); /* leave */
 430     if (func_ret_sub == 0) {
 431         o(0xc3); /* ret */
 432     } else {
 433         o(0xc2); /* ret n */
 434         g(func_ret_sub);
 435         g(func_ret_sub >> 8);
 436     }
 437     /* align local size to word & save local variables */
 438     *func_sub_sp_ptr = (-loc + 3) & -4;
 439 }
 440
 441 /* generate a jump to a label */
 442 int gjmp(int t)
 443 {
 444     return psym(0xe9, t);
 445 }
 446
 447 /* generate a jump to a fixed address */
 448 void gjmp_addr(int a)
 449 {
 450     oad(0xe9, a - ind - 5);
 451 }
 452
 453 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 454 int gtst(int inv, int t)
 455 {
 456     int v, *p;
 457     v = vtop->r & VT_VALMASK;
 458     if (v == VT_CMP) {
 459         /* fast case : can jump directly since flags are set */
 460         g(0x0f);
 461         t = psym((vtop->c.i - 16) ^ inv, t);
 462     } else if (v == VT_JMP || v == VT_JMPI) {
 463         /* && or || optimization */
 464         if ((v & 1) == inv) {
 465             /* insert vtop->c jump list in t */
 466             p = &vtop->c.i;
 467             while (*p != 0)
 468                 p = (int *)*p;
 469             *p = t;
 470             t = vtop->c.i;
 471         } else {
 472             t = gjmp(t);
 473             gsym(vtop->c.i);
 474         }
 475     } else {
 476         if (is_float(vtop->t)) {
 477             vpushi(0);
 478             gen_op(TOK_NE);
 479         }
 480         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 481             /* constant jmp optimization */
 482             if ((vtop->c.i != 0) != inv)
 483                 t = gjmp(t);
 484         } else {
 485             v = gv(RC_INT);
 486             o(0x85);
 487             o(0xc0 + v * 9);
 488             g(0x0f);
 489             t = psym(0x85 ^ inv, t);
 490         }
 491     }
 492     vtop--;
 493     return t;
 494 }
 495
 496 /* generate an integer binary operation */
 497 void gen_opi(int op)
 498 {
 499     int r, fr, opc, c;
 500
 501     switch(op) {
 502     case '+':
 503     case TOK_ADDC1: /* add with carry generation */
 504         opc = 0;
 505     gen_op8:
 506         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 507             /* constant case */
 508             vswap();
 509             r = gv(RC_INT);
 510             vswap();
 511             c = vtop->c.i;
 512             if (c == (char)c) {
 513                 /* XXX: generate inc and dec for smaller code ? */
 514                 o(0x83);
 515                 o(0xc0 | (opc << 3) | r);
 516                 g(c);
 517             } else {
 518                 o(0x81);
 519                 oad(0xc0 | (opc << 3) | r, c);
 520             }
 521         } else {
 522             gv2(RC_INT, RC_INT);
 523             r = vtop[-1].r;
 524             fr = vtop[0].r;
 525             o((opc << 3) | 0x01);
 526             o(0xc0 + r + fr * 8);
 527         }
 528         vtop--;
 529         if (op >= TOK_ULT && op <= TOK_GT) {
 530             vtop--;
 531             vset(VT_INT, VT_CMP, op);
 532         }
 533         break;
 534     case '-':
 535     case TOK_SUBC1: /* sub with carry generation */
 536         opc = 5;
 537         goto gen_op8;
 538     case TOK_ADDC2: /* add with carry use */
 539         opc = 2;
 540         goto gen_op8;
 541     case TOK_SUBC2: /* sub with carry use */
 542         opc = 3;
 543         goto gen_op8;
 544     case '&':
 545         opc = 4;
 546         goto gen_op8;
 547     case '^':
 548         opc = 6;
 549         goto gen_op8;
 550     case '|':
 551         opc = 1;
 552         goto gen_op8;
 553     case '*':
 554         gv2(RC_INT, RC_INT);
 555         r = vtop[-1].r;
 556         fr = vtop[0].r;
 557         vtop--;
 558         o(0xaf0f); /* imul fr, r */
 559         o(0xc0 + fr + r * 8);
 560         break;
 561     case TOK_SHL:
 562         opc = 4;
 563         goto gen_shift;
 564     case TOK_SHR:
 565         opc = 5;
 566         goto gen_shift;
 567     case TOK_SAR:
 568         opc = 7;
 569     gen_shift:
 570         opc = 0xc0 | (opc << 3);
 571         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 572             /* constant case */
 573             vswap();
 574             r = gv(RC_INT);
 575             vswap();
 576             c = vtop->c.i & 0x1f;
 577             o(0xc1); /* shl/shr/sar $xxx, r */
 578             o(opc | r);
 579             g(c);
 580         } else {
 581             /* we generate the shift in ecx */
 582             gv2(RC_INT, RC_ECX);
 583             r = vtop[-1].r;
 584             o(0xd3); /* shl/shr/sar %cl, r */
 585             o(opc | r);
 586         }
 587         vtop--;
 588         break;
 589     case '/':
 590     case TOK_UDIV:
 591     case TOK_PDIV:
 592     case '%':
 593     case TOK_UMOD:
 594     case TOK_UMULL:
 595         /* first operand must be in eax */
 596         /* XXX: need better constraint for second operand */
 597         gv2(RC_EAX, RC_ECX);
 598         r = vtop[-1].r;
 599         fr = vtop[0].r;
 600         vtop--;
 601         save_reg(REG_EDX);
 602         if (op == TOK_UMULL) {
 603             o(0xf7); /* mul fr */
 604             o(0xe0 + fr);
 605             vtop->r2 = REG_EDX;
 606             r = REG_EAX;
 607         } else {
 608             if (op == TOK_UDIV || op == TOK_UMOD) {
 609                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 610                 o(0xf0 + fr);
 611             } else {
 612                 o(0xf799); /* cltd, idiv fr, %eax */
 613                 o(0xf8 + fr);
 614             }
 615             if (op == '%' || op == TOK_UMOD)
 616                 r = REG_EDX;
 617             else
 618                 r = REG_EAX;
 619         }
 620         vtop->r = r;
 621         break;
 622     default:
 623         opc = 7;
 624         goto gen_op8;
 625     }
 626 }
 627
 628 /* generate a floating point operation 'v = t1 op t2' instruction. The
 629    two operands are guaranted to have the same floating point type */
 630 /* XXX: need to use ST1 too */
 631 void gen_opf(int op)
 632 {
 633     int a, ft, fc, swapped, r;
 634
 635     /* convert constants to memory references */
 636     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 637         vswap();
 638         gv(RC_FLOAT);
 639         vswap();
 640     }
 641     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 642         gv(RC_FLOAT);
 643
 644     /* must put at least one value in the floating point register */
 645     if ((vtop[-1].r & VT_LVAL) &&
 646         (vtop[0].r & VT_LVAL)) {
 647         vswap();
 648         gv(RC_FLOAT);
 649         vswap();
 650     }
 651     swapped = 0;
 652     /* swap the stack if needed so that t1 is the register and t2 is
 653        the memory reference */
 654     if (vtop[-1].r & VT_LVAL) {
 655         vswap();
 656         swapped = 1;
 657     }
 658     if (op >= TOK_ULT && op <= TOK_GT) {
 659         /* load on stack second operand */
 660         load(REG_ST0, vtop);
 661         save_reg(REG_EAX); /* eax is used by FP comparison code */
 662         if (op == TOK_GE || op == TOK_GT)
 663             swapped = !swapped;
 664         else if (op == TOK_EQ || op == TOK_NE)
 665             swapped = 0;
 666         if (swapped)
 667             o(0xc9d9); /* fxch %st(1) */
 668         o(0xe9da); /* fucompp */
 669         o(0xe0df); /* fnstsw %ax */
 670         if (op == TOK_EQ) {
 671             o(0x45e480); /* and $0x45, %ah */
 672             o(0x40fC80); /* cmp $0x40, %ah */
 673         } else if (op == TOK_NE) {
 674             o(0x45e480); /* and $0x45, %ah */
 675             o(0x40f480); /* xor $0x40, %ah */
 676             op = TOK_NE;
 677         } else if (op == TOK_GE || op == TOK_LE) {
 678             o(0x05c4f6); /* test $0x05, %ah */
 679             op = TOK_EQ;
 680         } else {
 681             o(0x45c4f6); /* test $0x45, %ah */
 682             op = TOK_EQ;
 683         }
 684         vtop--;
 685         vtop->r = VT_CMP;
 686         vtop->c.i = op;
 687     } else {
 688         /* no memory reference possible for long double operations */
 689         if ((vtop->t & VT_BTYPE) == VT_LDOUBLE) {
 690             load(REG_ST0, vtop);
 691             swapped = !swapped;
 692         }
 693
 694         switch(op) {
 695         default:
 696         case '+':
 697             a = 0;
 698             break;
 699         case '-':
 700             a = 4;
 701             if (swapped)
 702                 a++;
 703             break;
 704         case '*':
 705             a = 1;
 706             break;
 707         case '/':
 708             a = 6;
 709             if (swapped)
 710                 a++;
 711             break;
 712         }
 713         ft = vtop->t;
 714         fc = vtop->c.ul;
 715         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 716             o(0xde); /* fxxxp %st, %st(1) */
 717             o(0xc1 + (a << 3));
 718         } else {
 719             /* if saved lvalue, then we must reload it */
 720             r = vtop->r;
 721             if ((r & VT_VALMASK) == VT_LLOCAL) {
 722                 SValue v1;
 723                 r = get_reg(RC_INT);
 724                 v1.t = VT_INT;
 725                 v1.r = VT_LOCAL | VT_LVAL;
 726                 v1.c.ul = fc;
 727                 load(r, &v1);
 728                 fc = 0;
 729             }
 730
 731             if ((ft & VT_BTYPE) == VT_DOUBLE)
 732                 o(0xdc);
 733             else
 734                 o(0xd8);
 735             gen_modrm(a, r, fc);
 736         }
 737         vtop--;
 738     }
 739 }
 740
 741 /* FPU control word for rounding to nearest mode */
 742 /* XXX: should move that into tcc lib support code ! */
 743 static unsigned short __tcc_fpu_control = 0x137f;
 744 /* FPU control word for round to zero mode for int convertion */
 745 static unsigned short __tcc_int_fpu_control = 0x137f | 0x0c00;
 746
 747 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 748    and 'long long' cases. */
 749 void gen_cvt_itof(int t)
 750 {
 751     save_reg(REG_ST0);
 752     gv(RC_INT);
 753     if ((vtop->t & VT_BTYPE) == VT_LLONG) {
 754         /* signed long long to float/double/long double (unsigned case
 755            is handled generically) */
 756         o(0x50 + vtop->r2); /* push r2 */
 757         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 758         o(0x242cdf); /* fildll (%esp) */
 759         o(0x08c483); /* add $8, %esp */
 760     } else if ((vtop->t & (VT_BTYPE | VT_UNSIGNED)) ==
 761                (VT_INT | VT_UNSIGNED)) {
 762         /* unsigned int to float/double/long double */
 763         o(0x6a); /* push $0 */
 764         g(0x00);
 765         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 766         o(0x242cdf); /* fildll (%esp) */
 767         o(0x08c483); /* add $8, %esp */
 768     } else {
 769         /* int to float/double/long double */
 770         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 771         o(0x2404db); /* fildl (%esp) */
 772         o(0x04c483); /* add $4, %esp */
 773     }
 774     vtop->r = REG_ST0;
 775 }
 776
 777 /* convert fp to int 't' type */
 778 /* XXX: handle long long case */
 779 void gen_cvt_ftoi(int t)
 780 {
 781     int r, r2, size;
 782
 783     gv(RC_FLOAT);
 784     if (t != VT_INT)
 785         size = 8;
 786     else
 787         size = 4;
 788
 789     oad(0x2dd9, (int)&__tcc_int_fpu_control); /* ldcw xxx */
 790     oad(0xec81, size); /* sub $xxx, %esp */
 791     if (size == 4)
 792         o(0x1cdb); /* fistpl */
 793     else
 794         o(0x3cdf); /* fistpll */
 795     o(0x24);
 796     oad(0x2dd9, (int)&__tcc_fpu_control); /* ldcw xxx */
 797     r = get_reg(RC_INT);
 798     o(0x58 + r); /* pop r */
 799     if (size == 8) {
 800         if (t == VT_LLONG) {
 801             vtop->r = r; /* mark reg as used */
 802             r2 = get_reg(RC_INT);
 803             o(0x58 + r2); /* pop r2 */
 804             vtop->r2 = r2;
 805         } else {
 806             o(0x04c483); /* add $4, %esp */
 807         }
 808     }
 809     vtop->r = r;
 810 }
 811
 812 /* convert from one floating point type to another */
 813 void gen_cvt_ftof(int t)
 814 {
 815     /* all we have to do on i386 is to put the float in a register */
 816     gv(RC_FLOAT);
 817 }
 818
 819 /* bound check support functions */
 820 #ifdef CONFIG_TCC_BCHECK
 821
 822 /* generate a bounded pointer addition */
 823 void gen_bounded_ptr_add(void)
 824 {
 825     int addr;
 826     /* prepare fast i386 function call (args in eax and edx) */
 827     gv2(RC_EAX, RC_EDX);
 828     /* save all temporary registers */
 829     vtop -= 2;
 830     save_regs(0);
 831     /* do a fast function call */
 832     addr = ind;
 833     oad(0xe8, (int)__bound_ptr_add - ind - 5);
 834     /* returned pointer is in eax */
 835     vtop++;
 836     vtop->r = REG_EAX | VT_BOUNDED;
 837     vtop->c.ul = addr; /* address of bounding function call point */
 838 }
 839
 840 /* patch pointer addition in vtop so that pointer dereferencing is
 841    also tested */
 842 void gen_bounded_ptr_deref(void)
 843 {
 844     void *func;
 845     int size, align, addr;
 846
 847     size = 0;
 848     /* XXX: put that code in generic part of tcc */
 849     if (!is_float(vtop->t)) {
 850         if (vtop->r & VT_LVAL_BYTE)
 851             size = 1;
 852         else if (vtop->r & VT_LVAL_SHORT)
 853             size = 2;
 854     }
 855     if (!size)
 856         size = type_size(vtop->t, &align);
 857     switch(size) {
 858     case  1: func = __bound_ptr_indir1; break;
 859     case  2: func = __bound_ptr_indir2; break;
 860     case  4: func = __bound_ptr_indir4; break;
 861     case  8: func = __bound_ptr_indir8; break;
 862     case 12: func = __bound_ptr_indir12; break;
 863     case 16: func = __bound_ptr_indir16; break;
 864     default:
 865         error("unhandled size when derefencing bounded pointer");
 866         func = NULL;
 867         break;
 868     }
 869
 870     addr = vtop->c.ul;
 871     *(int *)(addr + 1) = (int)func - addr - 5;
 872 }
 873 #endif
 874
 875 /* end of X86 code generator */
 876 /*************************************************************/
 877