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 void greloc_patch(unsigned char *ptr,
 108                   unsigned long addr, unsigned long val, int type)
 109 {
 110     switch(type) {
 111     case R_386_32:
 112         *(int *)ptr += val;
 113         break;
 114     case R_386_PC32:
 115         *(int *)ptr += val - addr - 4;
 116         break;
 117     }
 118 }
 119
 120 /* output a symbol and patch all calls to it */
 121 void gsym_addr(int t, int a)
 122 {
 123     int n;
 124     while (t) {
 125         n = *(int *)t; /* next value */
 126         *(int *)t = a - t - 4;
 127         t = n;
 128     }
 129 }
 130
 131 void gsym(int t)
 132 {
 133     gsym_addr(t, ind);
 134 }
 135
 136 /* psym is used to put an instruction with a data field which is a
 137    reference to a symbol. It is in fact the same as oad ! */
 138 #define psym oad
 139
 140 /* instruction + 4 bytes data. Return the address of the data */
 141 int oad(int c, int s)
 142 {
 143     o(c);
 144     *(int *)ind = s;
 145     s = ind;
 146     ind = ind + 4;
 147     return s;
 148 }
 149
 150 /* output constant with relocation if 'r & VT_SYM' is true */
 151 void gen_addr32(int r, int c)
 152 {
 153     if (!(r & VT_SYM)) {
 154         gen_le32(c);
 155     } else {
 156         greloc(cur_text_section,
 157                (Sym *)c, ind - (int)cur_text_section->data, R_386_32);
 158         gen_le32(0);
 159     }
 160 }
 161
 162 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 163    opcode bits */
 164 void gen_modrm(int op_reg, int r, int c)
 165 {
 166     op_reg = op_reg << 3;
 167     if ((r & VT_VALMASK) == VT_CONST) {
 168         /* constant memory reference */
 169         o(0x05 | op_reg);
 170         gen_addr32(r, c);
 171     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 172         /* currently, we use only ebp as base */
 173         if (c == (char)c) {
 174             /* short reference */
 175             o(0x45 | op_reg);
 176             g(c);
 177         } else {
 178             oad(0x85 | op_reg, c);
 179         }
 180     } else {
 181         g(0x00 | op_reg | (r & VT_VALMASK));
 182     }
 183 }
 184
 185
 186 /* load 'r' from value 'sv' */
 187 void load(int r, SValue *sv)
 188 {
 189     int v, t, ft, fc, fr;
 190     SValue v1;
 191
 192     fr = sv->r;
 193     ft = sv->t;
 194     fc = sv->c.ul;
 195
 196     v = fr & VT_VALMASK;
 197     if (fr & VT_LVAL) {
 198         if (v == VT_LLOCAL) {
 199             v1.t = VT_INT;
 200             v1.r = VT_LOCAL | VT_LVAL;
 201             v1.c.ul = fc;
 202             load(r, &v1);
 203             fr = r;
 204         }
 205         if ((ft & VT_BTYPE) == VT_FLOAT) {
 206             o(0xd9); /* flds */
 207             r = 0;
 208         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 209             o(0xdd); /* fldl */
 210             r = 0;
 211         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 212             o(0xdb); /* fldt */
 213             r = 5;
 214         } else if ((ft & VT_TYPE) == VT_BYTE) {
 215             o(0xbe0f);   /* movsbl */
 216         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 217             o(0xb60f);   /* movzbl */
 218         } else if ((ft & VT_TYPE) == VT_SHORT) {
 219             o(0xbf0f);   /* movswl */
 220         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 221             o(0xb70f);   /* movzwl */
 222         } else {
 223             o(0x8b);     /* movl */
 224         }
 225         gen_modrm(r, fr, fc);
 226     } else {
 227         if (v == VT_CONST) {
 228             o(0xb8 + r); /* mov $xx, r */
 229             gen_addr32(fr, fc);
 230         } else if (v == VT_LOCAL) {
 231             o(0x8d); /* lea xxx(%ebp), r */
 232             gen_modrm(r, VT_LOCAL, fc);
 233         } else if (v == VT_CMP) {
 234             oad(0xb8 + r, 0); /* mov $0, r */
 235             o(0x0f); /* setxx %br */
 236             o(fc);
 237             o(0xc0 + r);
 238         } else if (v == VT_JMP || v == VT_JMPI) {
 239             t = v & 1;
 240             oad(0xb8 + r, t); /* mov $1, r */
 241             oad(0xe9, 5); /* jmp after */
 242             gsym(fc);
 243             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 244         } else if (v != r) {
 245             o(0x89);
 246             o(0xc0 + r + v * 8); /* mov v, r */
 247         }
 248     }
 249 }
 250
 251 /* store register 'r' in lvalue 'v' */
 252 void store(int r, SValue *v)
 253 {
 254     int fr, bt, ft, fc;
 255
 256     ft = v->t;
 257     fc = v->c.ul;
 258     fr = v->r & VT_VALMASK;
 259     bt = ft & VT_BTYPE;
 260     /* XXX: incorrect if float reg to reg */
 261     if (bt == VT_FLOAT) {
 262         o(0xd9); /* fsts */
 263         r = 2;
 264     } else if (bt == VT_DOUBLE) {
 265         o(0xdd); /* fstpl */
 266         r = 2;
 267     } else if (bt == VT_LDOUBLE) {
 268         o(0xc0d9); /* fld %st(0) */
 269         o(0xdb); /* fstpt */
 270         r = 7;
 271     } else {
 272         if (bt == VT_SHORT)
 273             o(0x66);
 274         if (bt == VT_BYTE)
 275             o(0x88);
 276         else
 277             o(0x89);
 278     }
 279     if (fr == VT_CONST ||
 280         fr == VT_LOCAL ||
 281         (v->r & VT_LVAL)) {
 282         gen_modrm(r, v->r, fc);
 283     } else if (fr != r) {
 284         o(0xc0 + fr + r * 8); /* mov r, fr */
 285     }
 286 }
 287
 288 /* start function call and return function call context */
 289 void gfunc_start(GFuncContext *c, int func_call)
 290 {
 291     c->args_size = 0;
 292     c->func_call = func_call;
 293 }
 294
 295 /* push function parameter which is in (vtop->t, vtop->c). Stack entry
 296    is then popped. */
 297 void gfunc_param(GFuncContext *c)
 298 {
 299     int size, align, r;
 300
 301     if ((vtop->t & VT_BTYPE) == VT_STRUCT) {
 302         size = type_size(vtop->t, &align);
 303         /* align to stack align size */
 304         size = (size + 3) & ~3;
 305         /* allocate the necessary size on stack */
 306         oad(0xec81, size); /* sub $xxx, %esp */
 307         /* generate structure store */
 308         r = get_reg(RC_INT);
 309         o(0x89); /* mov %esp, r */
 310         o(0xe0 + r);
 311         vset(vtop->t, r | VT_LVAL, 0);
 312         vswap();
 313         vstore();
 314         c->args_size += size;
 315     } else if (is_float(vtop->t)) {
 316         gv(RC_FLOAT); /* only one float register */
 317         if ((vtop->t & VT_BTYPE) == VT_FLOAT)
 318             size = 4;
 319         else if ((vtop->t & VT_BTYPE) == VT_DOUBLE)
 320             size = 8;
 321         else
 322             size = 12;
 323         oad(0xec81, size); /* sub $xxx, %esp */
 324         if (size == 12)
 325             o(0x7cdb);
 326         else
 327             o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 328         g(0x24);
 329         g(0x00);
 330         c->args_size += size;
 331     } else {
 332         /* simple type (currently always same size) */
 333         /* XXX: implicit cast ? */
 334         r = gv(RC_INT);
 335         if ((vtop->t & VT_BTYPE) == VT_LLONG) {
 336             size = 8;
 337             o(0x50 + vtop->r2); /* push r */
 338         } else {
 339             size = 4;
 340         }
 341         o(0x50 + r); /* push r */
 342         c->args_size += size;
 343     }
 344     vtop--;
 345 }
 346
 347 /* generate function call with address in (vtop->t, vtop->c) and free function
 348    context. Stack entry is popped */
 349 void gfunc_call(GFuncContext *c)
 350 {
 351     int r;
 352     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 353         /* constant case */
 354         if (vtop->r & VT_SYM) {
 355             /* relocation case */
 356             greloc(cur_text_section, vtop->c.sym,
 357                    ind + 1 - (int)cur_text_section->data, R_386_PC32);
 358             oad(0xe8, 0);
 359         } else {
 360             oad(0xe8, vtop->c.ul - ind - 5);
 361         }
 362     } else {
 363         /* otherwise, indirect call */
 364         r = gv(RC_INT);
 365         o(0xff); /* call *r */
 366         o(0xd0 + r);
 367     }
 368     if (c->args_size && c->func_call == FUNC_CDECL)
 369         oad(0xc481, c->args_size); /* add $xxx, %esp */
 370     vtop--;
 371 }
 372
 373 /* generate function prolog of type 't' */
 374 void gfunc_prolog(int t)
 375 {
 376     int addr, align, size, u, func_call;
 377     Sym *sym;
 378
 379     sym = sym_find((unsigned)t >> VT_STRUCT_SHIFT);
 380     func_call = sym->r;
 381     addr = 8;
 382     /* if the function returns a structure, then add an
 383        implicit pointer parameter */
 384     func_vt = sym->t;
 385     if ((func_vt & VT_BTYPE) == VT_STRUCT) {
 386         func_vc = addr;
 387         addr += 4;
 388     }
 389     /* define parameters */
 390     while ((sym = sym->next) != NULL) {
 391         u = sym->t;
 392         sym_push(sym->v & ~SYM_FIELD, u,
 393                  VT_LOCAL | VT_LVAL, addr);
 394         size = type_size(u, &align);
 395         size = (size + 3) & ~3;
 396 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 397         /* structs are passed as pointer */
 398         if ((u & VT_BTYPE) == VT_STRUCT) {
 399             size = 4;
 400         }
 401 #endif
 402         addr += size;
 403     }
 404     func_ret_sub = 0;
 405     /* pascal type call ? */
 406     if (func_call == FUNC_STDCALL)
 407         func_ret_sub = addr - 8;
 408     o(0xe58955); /* push   %ebp, mov    %esp, %ebp */
 409     func_sub_sp_ptr = (int *)oad(0xec81, 0); /* sub $xxx, %esp */
 410     /* leave some room for bound checking code */
 411     if (do_bounds_check) {
 412         oad(0xb8, 0); /* lbound section pointer */
 413         oad(0xb8, 0); /* call to function */
 414         func_bound_ptr = lbounds_section->data_ptr;
 415     }
 416 }
 417
 418 /* generate function epilog */
 419 void gfunc_epilog(void)
 420 {
 421 #ifdef CONFIG_TCC_BCHECK
 422     if (do_bounds_check && func_bound_ptr != lbounds_section->data_ptr) {
 423         int saved_ind;
 424         int *bounds_ptr;
 425         /* add end of table info */
 426         bounds_ptr = (int *)lbounds_section->data_ptr;
 427         *bounds_ptr++ = 0;
 428         lbounds_section->data_ptr = (unsigned char *)bounds_ptr;
 429         /* generate bound local allocation */
 430         saved_ind = ind;
 431         ind = (int)func_sub_sp_ptr + 4;
 432         oad(0xb8, (int)func_bound_ptr); /* mov %eax, xxx */
 433         oad(0xe8, (int)__bound_local_new - ind - 5);
 434         ind = saved_ind;
 435         /* generate bound check local freeing */
 436         o(0x5250); /* save returned value, if any */
 437         oad(0xb8, (int)func_bound_ptr); /* mov %eax, xxx */
 438         oad(0xe8, (int)__bound_local_delete - ind - 5);
 439         o(0x585a); /* restore returned value, if any */
 440     }
 441 #endif
 442     o(0xc9); /* leave */
 443     if (func_ret_sub == 0) {
 444         o(0xc3); /* ret */
 445     } else {
 446         o(0xc2); /* ret n */
 447         g(func_ret_sub);
 448         g(func_ret_sub >> 8);
 449     }
 450     /* align local size to word & save local variables */
 451     *func_sub_sp_ptr = (-loc + 3) & -4;
 452 }
 453
 454 /* generate a jump to a label */
 455 int gjmp(int t)
 456 {
 457     return psym(0xe9, t);
 458 }
 459
 460 /* generate a jump to a fixed address */
 461 void gjmp_addr(int a)
 462 {
 463     oad(0xe9, a - ind - 5);
 464 }
 465
 466 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 467 int gtst(int inv, int t)
 468 {
 469     int v, *p;
 470     v = vtop->r & VT_VALMASK;
 471     if (v == VT_CMP) {
 472         /* fast case : can jump directly since flags are set */
 473         g(0x0f);
 474         t = psym((vtop->c.i - 16) ^ inv, t);
 475     } else if (v == VT_JMP || v == VT_JMPI) {
 476         /* && or || optimization */
 477         if ((v & 1) == inv) {
 478             /* insert vtop->c jump list in t */
 479             p = &vtop->c.i;
 480             while (*p != 0)
 481                 p = (int *)*p;
 482             *p = t;
 483             t = vtop->c.i;
 484         } else {
 485             t = gjmp(t);
 486             gsym(vtop->c.i);
 487         }
 488     } else {
 489         if (is_float(vtop->t)) {
 490             vpushi(0);
 491             gen_op(TOK_NE);
 492         }
 493         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 494             /* constant jmp optimization */
 495             if ((vtop->c.i != 0) != inv)
 496                 t = gjmp(t);
 497         } else {
 498             v = gv(RC_INT);
 499             o(0x85);
 500             o(0xc0 + v * 9);
 501             g(0x0f);
 502             t = psym(0x85 ^ inv, t);
 503         }
 504     }
 505     vtop--;
 506     return t;
 507 }
 508
 509 /* generate an integer binary operation */
 510 void gen_opi(int op)
 511 {
 512     int r, fr, opc, c;
 513
 514     switch(op) {
 515     case '+':
 516     case TOK_ADDC1: /* add with carry generation */
 517         opc = 0;
 518     gen_op8:
 519         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 520             /* constant case */
 521             vswap();
 522             r = gv(RC_INT);
 523             vswap();
 524             c = vtop->c.i;
 525             if (c == (char)c) {
 526                 /* XXX: generate inc and dec for smaller code ? */
 527                 o(0x83);
 528                 o(0xc0 | (opc << 3) | r);
 529                 g(c);
 530             } else {
 531                 o(0x81);
 532                 oad(0xc0 | (opc << 3) | r, c);
 533             }
 534         } else {
 535             gv2(RC_INT, RC_INT);
 536             r = vtop[-1].r;
 537             fr = vtop[0].r;
 538             o((opc << 3) | 0x01);
 539             o(0xc0 + r + fr * 8);
 540         }
 541         vtop--;
 542         if (op >= TOK_ULT && op <= TOK_GT) {
 543             vtop--;
 544             vset(VT_INT, VT_CMP, op);
 545         }
 546         break;
 547     case '-':
 548     case TOK_SUBC1: /* sub with carry generation */
 549         opc = 5;
 550         goto gen_op8;
 551     case TOK_ADDC2: /* add with carry use */
 552         opc = 2;
 553         goto gen_op8;
 554     case TOK_SUBC2: /* sub with carry use */
 555         opc = 3;
 556         goto gen_op8;
 557     case '&':
 558         opc = 4;
 559         goto gen_op8;
 560     case '^':
 561         opc = 6;
 562         goto gen_op8;
 563     case '|':
 564         opc = 1;
 565         goto gen_op8;
 566     case '*':
 567         gv2(RC_INT, RC_INT);
 568         r = vtop[-1].r;
 569         fr = vtop[0].r;
 570         vtop--;
 571         o(0xaf0f); /* imul fr, r */
 572         o(0xc0 + fr + r * 8);
 573         break;
 574     case TOK_SHL:
 575         opc = 4;
 576         goto gen_shift;
 577     case TOK_SHR:
 578         opc = 5;
 579         goto gen_shift;
 580     case TOK_SAR:
 581         opc = 7;
 582     gen_shift:
 583         opc = 0xc0 | (opc << 3);
 584         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 585             /* constant case */
 586             vswap();
 587             r = gv(RC_INT);
 588             vswap();
 589             c = vtop->c.i & 0x1f;
 590             o(0xc1); /* shl/shr/sar $xxx, r */
 591             o(opc | r);
 592             g(c);
 593         } else {
 594             /* we generate the shift in ecx */
 595             gv2(RC_INT, RC_ECX);
 596             r = vtop[-1].r;
 597             o(0xd3); /* shl/shr/sar %cl, r */
 598             o(opc | r);
 599         }
 600         vtop--;
 601         break;
 602     case '/':
 603     case TOK_UDIV:
 604     case TOK_PDIV:
 605     case '%':
 606     case TOK_UMOD:
 607     case TOK_UMULL:
 608         /* first operand must be in eax */
 609         /* XXX: need better constraint for second operand */
 610         gv2(RC_EAX, RC_ECX);
 611         r = vtop[-1].r;
 612         fr = vtop[0].r;
 613         vtop--;
 614         save_reg(REG_EDX);
 615         if (op == TOK_UMULL) {
 616             o(0xf7); /* mul fr */
 617             o(0xe0 + fr);
 618             vtop->r2 = REG_EDX;
 619             r = REG_EAX;
 620         } else {
 621             if (op == TOK_UDIV || op == TOK_UMOD) {
 622                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 623                 o(0xf0 + fr);
 624             } else {
 625                 o(0xf799); /* cltd, idiv fr, %eax */
 626                 o(0xf8 + fr);
 627             }
 628             if (op == '%' || op == TOK_UMOD)
 629                 r = REG_EDX;
 630             else
 631                 r = REG_EAX;
 632         }
 633         vtop->r = r;
 634         break;
 635     default:
 636         opc = 7;
 637         goto gen_op8;
 638     }
 639 }
 640
 641 /* generate a floating point operation 'v = t1 op t2' instruction. The
 642    two operands are guaranted to have the same floating point type */
 643 /* XXX: need to use ST1 too */
 644 void gen_opf(int op)
 645 {
 646     int a, ft, fc, swapped, r;
 647
 648     /* convert constants to memory references */
 649     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 650         vswap();
 651         gv(RC_FLOAT);
 652         vswap();
 653     }
 654     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 655         gv(RC_FLOAT);
 656
 657     /* must put at least one value in the floating point register */
 658     if ((vtop[-1].r & VT_LVAL) &&
 659         (vtop[0].r & VT_LVAL)) {
 660         vswap();
 661         gv(RC_FLOAT);
 662         vswap();
 663     }
 664     swapped = 0;
 665     /* swap the stack if needed so that t1 is the register and t2 is
 666        the memory reference */
 667     if (vtop[-1].r & VT_LVAL) {
 668         vswap();
 669         swapped = 1;
 670     }
 671     if (op >= TOK_ULT && op <= TOK_GT) {
 672         /* load on stack second operand */
 673         load(REG_ST0, vtop);
 674         save_reg(REG_EAX); /* eax is used by FP comparison code */
 675         if (op == TOK_GE || op == TOK_GT)
 676             swapped = !swapped;
 677         else if (op == TOK_EQ || op == TOK_NE)
 678             swapped = 0;
 679         if (swapped)
 680             o(0xc9d9); /* fxch %st(1) */
 681         o(0xe9da); /* fucompp */
 682         o(0xe0df); /* fnstsw %ax */
 683         if (op == TOK_EQ) {
 684             o(0x45e480); /* and $0x45, %ah */
 685             o(0x40fC80); /* cmp $0x40, %ah */
 686         } else if (op == TOK_NE) {
 687             o(0x45e480); /* and $0x45, %ah */
 688             o(0x40f480); /* xor $0x40, %ah */
 689             op = TOK_NE;
 690         } else if (op == TOK_GE || op == TOK_LE) {
 691             o(0x05c4f6); /* test $0x05, %ah */
 692             op = TOK_EQ;
 693         } else {
 694             o(0x45c4f6); /* test $0x45, %ah */
 695             op = TOK_EQ;
 696         }
 697         vtop--;
 698         vtop->r = VT_CMP;
 699         vtop->c.i = op;
 700     } else {
 701         /* no memory reference possible for long double operations */
 702         if ((vtop->t & VT_BTYPE) == VT_LDOUBLE) {
 703             load(REG_ST0, vtop);
 704             swapped = !swapped;
 705         }
 706
 707         switch(op) {
 708         default:
 709         case '+':
 710             a = 0;
 711             break;
 712         case '-':
 713             a = 4;
 714             if (swapped)
 715                 a++;
 716             break;
 717         case '*':
 718             a = 1;
 719             break;
 720         case '/':
 721             a = 6;
 722             if (swapped)
 723                 a++;
 724             break;
 725         }
 726         ft = vtop->t;
 727         fc = vtop->c.ul;
 728         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 729             o(0xde); /* fxxxp %st, %st(1) */
 730             o(0xc1 + (a << 3));
 731         } else {
 732             /* if saved lvalue, then we must reload it */
 733             r = vtop->r;
 734             if ((r & VT_VALMASK) == VT_LLOCAL) {
 735                 SValue v1;
 736                 r = get_reg(RC_INT);
 737                 v1.t = VT_INT;
 738                 v1.r = VT_LOCAL | VT_LVAL;
 739                 v1.c.ul = fc;
 740                 load(r, &v1);
 741                 fc = 0;
 742             }
 743
 744             if ((ft & VT_BTYPE) == VT_DOUBLE)
 745                 o(0xdc);
 746             else
 747                 o(0xd8);
 748             gen_modrm(a, r, fc);
 749         }
 750         vtop--;
 751     }
 752 }
 753
 754 /* FPU control word for rounding to nearest mode */
 755 /* XXX: should move that into tcc lib support code ! */
 756 static unsigned short __tcc_fpu_control = 0x137f;
 757 /* FPU control word for round to zero mode for int convertion */
 758 static unsigned short __tcc_int_fpu_control = 0x137f | 0x0c00;
 759
 760 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 761    and 'long long' cases. */
 762 void gen_cvt_itof(int t)
 763 {
 764     save_reg(REG_ST0);
 765     gv(RC_INT);
 766     if ((vtop->t & VT_BTYPE) == VT_LLONG) {
 767         /* signed long long to float/double/long double (unsigned case
 768            is handled generically) */
 769         o(0x50 + vtop->r2); /* push r2 */
 770         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 771         o(0x242cdf); /* fildll (%esp) */
 772         o(0x08c483); /* add $8, %esp */
 773     } else if ((vtop->t & (VT_BTYPE | VT_UNSIGNED)) ==
 774                (VT_INT | VT_UNSIGNED)) {
 775         /* unsigned int to float/double/long double */
 776         o(0x6a); /* push $0 */
 777         g(0x00);
 778         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 779         o(0x242cdf); /* fildll (%esp) */
 780         o(0x08c483); /* add $8, %esp */
 781     } else {
 782         /* int to float/double/long double */
 783         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 784         o(0x2404db); /* fildl (%esp) */
 785         o(0x04c483); /* add $4, %esp */
 786     }
 787     vtop->r = REG_ST0;
 788 }
 789
 790 /* convert fp to int 't' type */
 791 /* XXX: handle long long case */
 792 void gen_cvt_ftoi(int t)
 793 {
 794     int r, r2, size;
 795
 796     gv(RC_FLOAT);
 797     if (t != VT_INT)
 798         size = 8;
 799     else
 800         size = 4;
 801
 802     oad(0x2dd9, (int)&__tcc_int_fpu_control); /* ldcw xxx */
 803     oad(0xec81, size); /* sub $xxx, %esp */
 804     if (size == 4)
 805         o(0x1cdb); /* fistpl */
 806     else
 807         o(0x3cdf); /* fistpll */
 808     o(0x24);
 809     oad(0x2dd9, (int)&__tcc_fpu_control); /* ldcw xxx */
 810     r = get_reg(RC_INT);
 811     o(0x58 + r); /* pop r */
 812     if (size == 8) {
 813         if (t == VT_LLONG) {
 814             vtop->r = r; /* mark reg as used */
 815             r2 = get_reg(RC_INT);
 816             o(0x58 + r2); /* pop r2 */
 817             vtop->r2 = r2;
 818         } else {
 819             o(0x04c483); /* add $4, %esp */
 820         }
 821     }
 822     vtop->r = r;
 823 }
 824
 825 /* convert from one floating point type to another */
 826 void gen_cvt_ftof(int t)
 827 {
 828     /* all we have to do on i386 is to put the float in a register */
 829     gv(RC_FLOAT);
 830 }
 831
 832 /* bound check support functions */
 833 #ifdef CONFIG_TCC_BCHECK
 834
 835 /* generate a bounded pointer addition */
 836 void gen_bounded_ptr_add(void)
 837 {
 838     int addr;
 839     /* prepare fast i386 function call (args in eax and edx) */
 840     gv2(RC_EAX, RC_EDX);
 841     /* save all temporary registers */
 842     vtop -= 2;
 843     save_regs(0);
 844     /* do a fast function call */
 845     addr = ind;
 846     oad(0xe8, (int)__bound_ptr_add - ind - 5);
 847     /* returned pointer is in eax */
 848     vtop++;
 849     vtop->r = REG_EAX | VT_BOUNDED;
 850     vtop->c.ul = addr; /* address of bounding function call point */
 851 }
 852
 853 /* patch pointer addition in vtop so that pointer dereferencing is
 854    also tested */
 855 void gen_bounded_ptr_deref(void)
 856 {
 857     void *func;
 858     int size, align, addr;
 859
 860     size = 0;
 861     /* XXX: put that code in generic part of tcc */
 862     if (!is_float(vtop->t)) {
 863         if (vtop->r & VT_LVAL_BYTE)
 864             size = 1;
 865         else if (vtop->r & VT_LVAL_SHORT)
 866             size = 2;
 867     }
 868     if (!size)
 869         size = type_size(vtop->t, &align);
 870     switch(size) {
 871     case  1: func = __bound_ptr_indir1; break;
 872     case  2: func = __bound_ptr_indir2; break;
 873     case  4: func = __bound_ptr_indir4; break;
 874     case  8: func = __bound_ptr_indir8; break;
 875     case 12: func = __bound_ptr_indir12; break;
 876     case 16: func = __bound_ptr_indir16; break;
 877     default:
 878         error("unhandled size when derefencing bounded pointer");
 879         func = NULL;
 880         break;
 881     }
 882
 883     addr = vtop->c.ul;
 884     *(int *)(addr + 1) = (int)func - addr - 5;
 885 }
 886 #endif
 887
 888 /* end of X86 code generator */
 889 /*************************************************************/
 890