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 static void gadd_sp(int val)
 335 {
 336     if (val == (char)val) {
 337         o(0xc483);
 338         g(val);
 339     } else {
 340         oad(0xc481, val); /* add $xxx, %esp */
 341     }
 342 }
 343
 344 /* generate function call with address in (vtop->t, vtop->c) and free function
 345    context. Stack entry is popped */
 346 void gfunc_call(GFuncContext *c)
 347 {
 348     int r;
 349     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 350         /* constant case */
 351         if (vtop->r & VT_SYM) {
 352             /* relocation case */
 353             greloc(cur_text_section, vtop->c.sym,
 354                    ind + 1 - (int)cur_text_section->data, R_386_PC32);
 355             oad(0xe8, -4);
 356         } else {
 357             oad(0xe8, vtop->c.ul - ind - 5);
 358         }
 359     } else {
 360         /* otherwise, indirect call */
 361         r = gv(RC_INT);
 362         o(0xff); /* call *r */
 363         o(0xd0 + r);
 364     }
 365     if (c->args_size && c->func_call == FUNC_CDECL)
 366         gadd_sp(c->args_size);
 367     vtop--;
 368 }
 369
 370 /* generate function prolog of type 't' */
 371 void gfunc_prolog(int t)
 372 {
 373     int addr, align, size, u, func_call;
 374     Sym *sym;
 375
 376     sym = sym_find((unsigned)t >> VT_STRUCT_SHIFT);
 377     func_call = sym->r;
 378     addr = 8;
 379     /* if the function returns a structure, then add an
 380        implicit pointer parameter */
 381     func_vt = sym->t;
 382     if ((func_vt & VT_BTYPE) == VT_STRUCT) {
 383         func_vc = addr;
 384         addr += 4;
 385     }
 386     /* define parameters */
 387     while ((sym = sym->next) != NULL) {
 388         u = sym->t;
 389         sym_push(sym->v & ~SYM_FIELD, u,
 390                  VT_LOCAL | VT_LVAL, addr);
 391         size = type_size(u, &align);
 392         size = (size + 3) & ~3;
 393 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 394         /* structs are passed as pointer */
 395         if ((u & VT_BTYPE) == VT_STRUCT) {
 396             size = 4;
 397         }
 398 #endif
 399         addr += size;
 400     }
 401     func_ret_sub = 0;
 402     /* pascal type call ? */
 403     if (func_call == FUNC_STDCALL)
 404         func_ret_sub = addr - 8;
 405     o(0xe58955); /* push   %ebp, mov    %esp, %ebp */
 406     func_sub_sp_ptr = (int *)oad(0xec81, 0); /* sub $xxx, %esp */
 407     /* leave some room for bound checking code */
 408     if (do_bounds_check) {
 409         oad(0xb8, 0); /* lbound section pointer */
 410         oad(0xb8, 0); /* call to function */
 411         func_bound_ptr = lbounds_section->data_ptr;
 412     }
 413 }
 414
 415 /* generate function epilog */
 416 void gfunc_epilog(void)
 417 {
 418 #ifdef CONFIG_TCC_BCHECK
 419     if (do_bounds_check && func_bound_ptr != lbounds_section->data_ptr) {
 420         int saved_ind;
 421         int *bounds_ptr;
 422         Sym *sym, *sym_data;
 423         /* add end of table info */
 424         bounds_ptr = (int *)lbounds_section->data_ptr;
 425         *bounds_ptr++ = 0;
 426         lbounds_section->data_ptr = (unsigned char *)bounds_ptr;
 427         /* generate bound local allocation */
 428         saved_ind = ind;
 429         ind = (int)func_sub_sp_ptr + 4;
 430         sym_data = get_sym_ref(char_pointer_type, lbounds_section,
 431                                func_bound_ptr - lbounds_section->data,
 432                                lbounds_section->data_ptr - func_bound_ptr);
 433         greloc(cur_text_section, sym_data,
 434                ind + 1 - (int)cur_text_section->data, R_386_32);
 435         oad(0xb8, 0); /* mov %eax, xxx */
 436         sym = external_sym(TOK___bound_local_new, func_old_type, 0);
 437         greloc(cur_text_section, sym,
 438                ind + 1 - (int)cur_text_section->data, R_386_PC32);
 439         oad(0xe8, -4);
 440         ind = saved_ind;
 441         /* generate bound check local freeing */
 442         o(0x5250); /* save returned value, if any */
 443         greloc(cur_text_section, sym_data,
 444                ind + 1 - (int)cur_text_section->data, R_386_32);
 445         oad(0xb8, 0); /* mov %eax, xxx */
 446         sym = external_sym(TOK___bound_local_delete, func_old_type, 0);
 447         greloc(cur_text_section, sym,
 448                ind + 1 - (int)cur_text_section->data, R_386_PC32);
 449         oad(0xe8, -4);
 450         o(0x585a); /* restore returned value, if any */
 451     }
 452 #endif
 453     o(0xc9); /* leave */
 454     if (func_ret_sub == 0) {
 455         o(0xc3); /* ret */
 456     } else {
 457         o(0xc2); /* ret n */
 458         g(func_ret_sub);
 459         g(func_ret_sub >> 8);
 460     }
 461     /* align local size to word & save local variables */
 462     *func_sub_sp_ptr = (-loc + 3) & -4;
 463 }
 464
 465 /* generate a jump to a label */
 466 int gjmp(int t)
 467 {
 468     return psym(0xe9, t);
 469 }
 470
 471 /* generate a jump to a fixed address */
 472 void gjmp_addr(int a)
 473 {
 474     oad(0xe9, a - ind - 5);
 475 }
 476
 477 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 478 int gtst(int inv, int t)
 479 {
 480     int v, *p;
 481     v = vtop->r & VT_VALMASK;
 482     if (v == VT_CMP) {
 483         /* fast case : can jump directly since flags are set */
 484         g(0x0f);
 485         t = psym((vtop->c.i - 16) ^ inv, t);
 486     } else if (v == VT_JMP || v == VT_JMPI) {
 487         /* && or || optimization */
 488         if ((v & 1) == inv) {
 489             /* insert vtop->c jump list in t */
 490             p = &vtop->c.i;
 491             while (*p != 0)
 492                 p = (int *)*p;
 493             *p = t;
 494             t = vtop->c.i;
 495         } else {
 496             t = gjmp(t);
 497             gsym(vtop->c.i);
 498         }
 499     } else {
 500         if (is_float(vtop->t)) {
 501             vpushi(0);
 502             gen_op(TOK_NE);
 503         }
 504         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 505             /* constant jmp optimization */
 506             if ((vtop->c.i != 0) != inv)
 507                 t = gjmp(t);
 508         } else {
 509             v = gv(RC_INT);
 510             o(0x85);
 511             o(0xc0 + v * 9);
 512             g(0x0f);
 513             t = psym(0x85 ^ inv, t);
 514         }
 515     }
 516     vtop--;
 517     return t;
 518 }
 519
 520 /* generate an integer binary operation */
 521 void gen_opi(int op)
 522 {
 523     int r, fr, opc, c;
 524
 525     switch(op) {
 526     case '+':
 527     case TOK_ADDC1: /* add with carry generation */
 528         opc = 0;
 529     gen_op8:
 530         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 531             /* constant case */
 532             vswap();
 533             r = gv(RC_INT);
 534             vswap();
 535             c = vtop->c.i;
 536             if (c == (char)c) {
 537                 /* XXX: generate inc and dec for smaller code ? */
 538                 o(0x83);
 539                 o(0xc0 | (opc << 3) | r);
 540                 g(c);
 541             } else {
 542                 o(0x81);
 543                 oad(0xc0 | (opc << 3) | r, c);
 544             }
 545         } else {
 546             gv2(RC_INT, RC_INT);
 547             r = vtop[-1].r;
 548             fr = vtop[0].r;
 549             o((opc << 3) | 0x01);
 550             o(0xc0 + r + fr * 8);
 551         }
 552         vtop--;
 553         if (op >= TOK_ULT && op <= TOK_GT) {
 554             vtop--;
 555             vset(VT_INT, VT_CMP, op);
 556         }
 557         break;
 558     case '-':
 559     case TOK_SUBC1: /* sub with carry generation */
 560         opc = 5;
 561         goto gen_op8;
 562     case TOK_ADDC2: /* add with carry use */
 563         opc = 2;
 564         goto gen_op8;
 565     case TOK_SUBC2: /* sub with carry use */
 566         opc = 3;
 567         goto gen_op8;
 568     case '&':
 569         opc = 4;
 570         goto gen_op8;
 571     case '^':
 572         opc = 6;
 573         goto gen_op8;
 574     case '|':
 575         opc = 1;
 576         goto gen_op8;
 577     case '*':
 578         gv2(RC_INT, RC_INT);
 579         r = vtop[-1].r;
 580         fr = vtop[0].r;
 581         vtop--;
 582         o(0xaf0f); /* imul fr, r */
 583         o(0xc0 + fr + r * 8);
 584         break;
 585     case TOK_SHL:
 586         opc = 4;
 587         goto gen_shift;
 588     case TOK_SHR:
 589         opc = 5;
 590         goto gen_shift;
 591     case TOK_SAR:
 592         opc = 7;
 593     gen_shift:
 594         opc = 0xc0 | (opc << 3);
 595         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 596             /* constant case */
 597             vswap();
 598             r = gv(RC_INT);
 599             vswap();
 600             c = vtop->c.i & 0x1f;
 601             o(0xc1); /* shl/shr/sar $xxx, r */
 602             o(opc | r);
 603             g(c);
 604         } else {
 605             /* we generate the shift in ecx */
 606             gv2(RC_INT, RC_ECX);
 607             r = vtop[-1].r;
 608             o(0xd3); /* shl/shr/sar %cl, r */
 609             o(opc | r);
 610         }
 611         vtop--;
 612         break;
 613     case '/':
 614     case TOK_UDIV:
 615     case TOK_PDIV:
 616     case '%':
 617     case TOK_UMOD:
 618     case TOK_UMULL:
 619         /* first operand must be in eax */
 620         /* XXX: need better constraint for second operand */
 621         gv2(RC_EAX, RC_ECX);
 622         r = vtop[-1].r;
 623         fr = vtop[0].r;
 624         vtop--;
 625         save_reg(REG_EDX);
 626         if (op == TOK_UMULL) {
 627             o(0xf7); /* mul fr */
 628             o(0xe0 + fr);
 629             vtop->r2 = REG_EDX;
 630             r = REG_EAX;
 631         } else {
 632             if (op == TOK_UDIV || op == TOK_UMOD) {
 633                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 634                 o(0xf0 + fr);
 635             } else {
 636                 o(0xf799); /* cltd, idiv fr, %eax */
 637                 o(0xf8 + fr);
 638             }
 639             if (op == '%' || op == TOK_UMOD)
 640                 r = REG_EDX;
 641             else
 642                 r = REG_EAX;
 643         }
 644         vtop->r = r;
 645         break;
 646     default:
 647         opc = 7;
 648         goto gen_op8;
 649     }
 650 }
 651
 652 /* generate a floating point operation 'v = t1 op t2' instruction. The
 653    two operands are guaranted to have the same floating point type */
 654 /* XXX: need to use ST1 too */
 655 void gen_opf(int op)
 656 {
 657     int a, ft, fc, swapped, r;
 658
 659     /* convert constants to memory references */
 660     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 661         vswap();
 662         gv(RC_FLOAT);
 663         vswap();
 664     }
 665     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 666         gv(RC_FLOAT);
 667
 668     /* must put at least one value in the floating point register */
 669     if ((vtop[-1].r & VT_LVAL) &&
 670         (vtop[0].r & VT_LVAL)) {
 671         vswap();
 672         gv(RC_FLOAT);
 673         vswap();
 674     }
 675     swapped = 0;
 676     /* swap the stack if needed so that t1 is the register and t2 is
 677        the memory reference */
 678     if (vtop[-1].r & VT_LVAL) {
 679         vswap();
 680         swapped = 1;
 681     }
 682     if (op >= TOK_ULT && op <= TOK_GT) {
 683         /* load on stack second operand */
 684         load(REG_ST0, vtop);
 685         save_reg(REG_EAX); /* eax is used by FP comparison code */
 686         if (op == TOK_GE || op == TOK_GT)
 687             swapped = !swapped;
 688         else if (op == TOK_EQ || op == TOK_NE)
 689             swapped = 0;
 690         if (swapped)
 691             o(0xc9d9); /* fxch %st(1) */
 692         o(0xe9da); /* fucompp */
 693         o(0xe0df); /* fnstsw %ax */
 694         if (op == TOK_EQ) {
 695             o(0x45e480); /* and $0x45, %ah */
 696             o(0x40fC80); /* cmp $0x40, %ah */
 697         } else if (op == TOK_NE) {
 698             o(0x45e480); /* and $0x45, %ah */
 699             o(0x40f480); /* xor $0x40, %ah */
 700             op = TOK_NE;
 701         } else if (op == TOK_GE || op == TOK_LE) {
 702             o(0x05c4f6); /* test $0x05, %ah */
 703             op = TOK_EQ;
 704         } else {
 705             o(0x45c4f6); /* test $0x45, %ah */
 706             op = TOK_EQ;
 707         }
 708         vtop--;
 709         vtop->r = VT_CMP;
 710         vtop->c.i = op;
 711     } else {
 712         /* no memory reference possible for long double operations */
 713         if ((vtop->t & VT_BTYPE) == VT_LDOUBLE) {
 714             load(REG_ST0, vtop);
 715             swapped = !swapped;
 716         }
 717
 718         switch(op) {
 719         default:
 720         case '+':
 721             a = 0;
 722             break;
 723         case '-':
 724             a = 4;
 725             if (swapped)
 726                 a++;
 727             break;
 728         case '*':
 729             a = 1;
 730             break;
 731         case '/':
 732             a = 6;
 733             if (swapped)
 734                 a++;
 735             break;
 736         }
 737         ft = vtop->t;
 738         fc = vtop->c.ul;
 739         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 740             o(0xde); /* fxxxp %st, %st(1) */
 741             o(0xc1 + (a << 3));
 742         } else {
 743             /* if saved lvalue, then we must reload it */
 744             r = vtop->r;
 745             if ((r & VT_VALMASK) == VT_LLOCAL) {
 746                 SValue v1;
 747                 r = get_reg(RC_INT);
 748                 v1.t = VT_INT;
 749                 v1.r = VT_LOCAL | VT_LVAL;
 750                 v1.c.ul = fc;
 751                 load(r, &v1);
 752                 fc = 0;
 753             }
 754
 755             if ((ft & VT_BTYPE) == VT_DOUBLE)
 756                 o(0xdc);
 757             else
 758                 o(0xd8);
 759             gen_modrm(a, r, fc);
 760         }
 761         vtop--;
 762     }
 763 }
 764
 765 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 766    and 'long long' cases. */
 767 void gen_cvt_itof(int t)
 768 {
 769     save_reg(REG_ST0);
 770     gv(RC_INT);
 771     if ((vtop->t & VT_BTYPE) == VT_LLONG) {
 772         /* signed long long to float/double/long double (unsigned case
 773            is handled generically) */
 774         o(0x50 + vtop->r2); /* push r2 */
 775         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 776         o(0x242cdf); /* fildll (%esp) */
 777         o(0x08c483); /* add $8, %esp */
 778     } else if ((vtop->t & (VT_BTYPE | VT_UNSIGNED)) ==
 779                (VT_INT | VT_UNSIGNED)) {
 780         /* unsigned int to float/double/long double */
 781         o(0x6a); /* push $0 */
 782         g(0x00);
 783         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 784         o(0x242cdf); /* fildll (%esp) */
 785         o(0x08c483); /* add $8, %esp */
 786     } else {
 787         /* int to float/double/long double */
 788         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 789         o(0x2404db); /* fildl (%esp) */
 790         o(0x04c483); /* add $4, %esp */
 791     }
 792     vtop->r = REG_ST0;
 793 }
 794
 795 /* convert fp to int 't' type */
 796 /* XXX: handle long long case */
 797 void gen_cvt_ftoi(int t)
 798 {
 799     int r, r2, size;
 800     Sym *sym;
 801
 802     gv(RC_FLOAT);
 803     if (t != VT_INT)
 804         size = 8;
 805     else
 806         size = 4;
 807
 808     o(0x2dd9); /* ldcw xxx */
 809     sym = external_sym(TOK___tcc_int_fpu_control,
 810                        VT_SHORT | VT_UNSIGNED, VT_LVAL);
 811     greloc(cur_text_section, sym,
 812            ind - (int)cur_text_section->data, R_386_32);
 813     gen_le32(0);
 814
 815     oad(0xec81, size); /* sub $xxx, %esp */
 816     if (size == 4)
 817         o(0x1cdb); /* fistpl */
 818     else
 819         o(0x3cdf); /* fistpll */
 820     o(0x24);
 821     o(0x2dd9); /* ldcw xxx */
 822     sym = external_sym(TOK___tcc_fpu_control,
 823                        VT_SHORT | VT_UNSIGNED, VT_LVAL);
 824     greloc(cur_text_section, sym,
 825            ind - (int)cur_text_section->data, R_386_32);
 826     gen_le32(0);
 827
 828     r = get_reg(RC_INT);
 829     o(0x58 + r); /* pop r */
 830     if (size == 8) {
 831         if (t == VT_LLONG) {
 832             vtop->r = r; /* mark reg as used */
 833             r2 = get_reg(RC_INT);
 834             o(0x58 + r2); /* pop r2 */
 835             vtop->r2 = r2;
 836         } else {
 837             o(0x04c483); /* add $4, %esp */
 838         }
 839     }
 840     vtop->r = r;
 841 }
 842
 843 /* convert from one floating point type to another */
 844 void gen_cvt_ftof(int t)
 845 {
 846     /* all we have to do on i386 is to put the float in a register */
 847     gv(RC_FLOAT);
 848 }
 849
 850 /* bound check support functions */
 851 #ifdef CONFIG_TCC_BCHECK
 852
 853 /* generate a bounded pointer addition */
 854 void gen_bounded_ptr_add(void)
 855 {
 856     Sym *sym;
 857
 858     /* prepare fast i386 function call (args in eax and edx) */
 859     gv2(RC_EAX, RC_EDX);
 860     /* save all temporary registers */
 861     vtop -= 2;
 862     save_regs(0);
 863     /* do a fast function call */
 864     sym = external_sym(TOK___bound_ptr_add, func_old_type, 0);
 865     greloc(cur_text_section, sym,
 866            ind + 1 - (int)cur_text_section->data, R_386_PC32);
 867     oad(0xe8, -4);
 868     /* returned pointer is in eax */
 869     vtop++;
 870     vtop->r = REG_EAX | VT_BOUNDED;
 871     /* address of bounding function call point */
 872     vtop->c.ptr = (cur_text_section->reloc->data_ptr - sizeof(Elf32_Rel));
 873 }
 874
 875 /* patch pointer addition in vtop so that pointer dereferencing is
 876    also tested */
 877 void gen_bounded_ptr_deref(void)
 878 {
 879     int func;
 880     int size, align;
 881     Elf32_Rel *rel;
 882     Sym *sym;
 883
 884     size = 0;
 885     /* XXX: put that code in generic part of tcc */
 886     if (!is_float(vtop->t)) {
 887         if (vtop->r & VT_LVAL_BYTE)
 888             size = 1;
 889         else if (vtop->r & VT_LVAL_SHORT)
 890             size = 2;
 891     }
 892     if (!size)
 893         size = type_size(vtop->t, &align);
 894     switch(size) {
 895     case  1: func = TOK___bound_ptr_indir1; break;
 896     case  2: func = TOK___bound_ptr_indir2; break;
 897     case  4: func = TOK___bound_ptr_indir4; break;
 898     case  8: func = TOK___bound_ptr_indir8; break;
 899     case 12: func = TOK___bound_ptr_indir12; break;
 900     case 16: func = TOK___bound_ptr_indir16; break;
 901     default:
 902         error("unhandled size when derefencing bounded pointer");
 903         func = 0;
 904         break;
 905     }
 906
 907     /* patch relocation */
 908     /* XXX: find a better solution ? */
 909     rel = vtop->c.ptr;
 910     sym = external_sym(func, func_old_type, 0);
 911     if (!sym->c)
 912         put_extern_sym(sym, NULL, 0);
 913     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
 914 }
 915 #endif
 916
 917 /* end of X86 code generator */
 918 /*************************************************************/
 919